diff options
Diffstat (limited to 'src/core/ext')
65 files changed, 1755 insertions, 7433 deletions
diff --git a/src/core/ext/census/README.md b/src/core/ext/census/README.md deleted file mode 100644 index a9826fe889..0000000000 --- a/src/core/ext/census/README.md +++ /dev/null @@ -1,61 +0,0 @@ -<!--- - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. ---> - -# Census - a resource measurement and tracing system - -This directory contains code for Census, which will ultimately provide the -following features for any gRPC-using system: -* A [dapper](http://research.google.com/pubs/pub36356.html)-like tracing - system, enabling tracing across a distributed infrastructure. -* RPC statistics and measurements for key metrics, such as latency, bytes - transferred, number of errors etc. -* Resource measurement framework which can be used for measuring custom - metrics. Through the use of [tags](#Tags), these can be broken down across - the entire distributed stack. -* Easy integration of the above with - [Google Cloud Trace](https://cloud.google.com/tools/cloud-trace) and - [Google Cloud Monitoring](https://cloud.google.com/monitoring/). - -## Concepts - -### Context - -### Operations - -### Tags - -### Metrics - -## API - -### Internal/RPC API - -### External/Client API - -### RPC API - -## Files in this directory - -Note that files and functions in this directory can be split into two -categories: -* Files that define core census library functions. Functions etc. in these - files are named census\_\*, and constitute the core census library - functionality. At some time in the future, these will become a standalone - library. -* Files that define functions etc. that provide a convenient interface between - grpc and the core census functionality. These files are all named - grpc\_\*.{c,h}, and define function names beginning with grpc\_census\_\*. - diff --git a/src/core/ext/census/aggregation.h b/src/core/ext/census/aggregation.h deleted file mode 100644 index 1ba7953ecc..0000000000 --- a/src/core/ext/census/aggregation.h +++ /dev/null @@ -1,51 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include <stddef.h> - -#ifndef GRPC_CORE_EXT_CENSUS_AGGREGATION_H -#define GRPC_CORE_EXT_CENSUS_AGGREGATION_H - -/** Structure used to describe an aggregation type. */ -struct census_aggregation_ops { - /* Create a new aggregation. The pointer returned can be used in future calls - to clone(), free(), record(), data() and reset(). */ - void *(*create)(const void *create_arg); - /* Make a copy of an aggregation created by create() */ - void *(*clone)(const void *aggregation); - /* Destroy an aggregation created by create() */ - void (*free)(void *aggregation); - /* Record a new value against aggregation. */ - void (*record)(void *aggregation, double value); - /* Return current aggregation data. The caller must cast this object into - the correct type for the aggregation result. The object returned can be - freed by using free_data(). */ - void *(*data)(const void *aggregation); - /* free data returned by data() */ - void (*free_data)(void *data); - /* Reset an aggregation to default (zero) values. */ - void (*reset)(void *aggregation); - /* Merge 'from' aggregation into 'to'. Both aggregations must be compatible */ - void (*merge)(void *to, const void *from); - /* Fill buffer with printable string version of aggregation contents. For - debugging only. Returns the number of bytes added to buffer (a value == n - implies the buffer was of insufficient size). */ - size_t (*print)(const void *aggregation, char *buffer, size_t n); -}; - -#endif /* GRPC_CORE_EXT_CENSUS_AGGREGATION_H */ diff --git a/src/core/ext/census/base_resources.cc b/src/core/ext/census/base_resources.cc deleted file mode 100644 index 3697c6f0e0..0000000000 --- a/src/core/ext/census/base_resources.cc +++ /dev/null @@ -1,56 +0,0 @@ -/* - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include "src/core/ext/census/base_resources.h" - -#include <stdio.h> -#include <string.h> - -#include <grpc/census.h> -#include <grpc/support/log.h> - -#include "src/core/ext/census/resource.h" - -// Add base RPC resource definitions for use by RPC runtime. -// -// TODO(aveitch): All of these are currently hardwired definitions encoded in -// the code in this file. These should be converted to use an external -// configuration mechanism, in which these resources are defined in a text -// file, which is compiled to .pb format and read by still-to-be-written -// configuration functions. - -// Define all base resources. This should be called by census initialization. -void define_base_resources() { - google_census_Resource_BasicUnit numerator = - google_census_Resource_BasicUnit_SECS; - resource r = {(char *)"client_rpc_latency", // name - (char *)"Client RPC latency in seconds", // description - 0, // prefix - 1, // n_numerators - &numerator, // numerators - 0, // n_denominators - NULL}; // denominators - define_resource(&r); - r = {(char *)"server_rpc_latency", // name - (char *)"Server RPC latency in seconds", // description - 0, // prefix - 1, // n_numerators - &numerator, // numerators - 0, // n_denominators - NULL}; // denominators - define_resource(&r); -} diff --git a/src/core/ext/census/base_resources.h b/src/core/ext/census/base_resources.h deleted file mode 100644 index d24923b597..0000000000 --- a/src/core/ext/census/base_resources.h +++ /dev/null @@ -1,32 +0,0 @@ -/* - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_BASE_RESOURCES_H -#define GRPC_CORE_EXT_CENSUS_BASE_RESOURCES_H - -#ifdef __cplusplus -extern "C" { -#endif - -/* Define all base resources. This should be called by census initialization. */ -void define_base_resources(); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_BASE_RESOURCES_H */ diff --git a/src/core/ext/census/census_init.cc b/src/core/ext/census/census_init.cc deleted file mode 100644 index d7f719ff8c..0000000000 --- a/src/core/ext/census/census_init.cc +++ /dev/null @@ -1,33 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include "src/core/ext/census/census_interface.h" - -#include <grpc/support/log.h> -#include "src/core/ext/census/census_rpc_stats.h" -#include "src/core/ext/census/census_tracing.h" - -void census_init(void) { - census_tracing_init(); - census_stats_store_init(); -} - -void census_shutdown(void) { - census_stats_store_shutdown(); - census_tracing_shutdown(); -} diff --git a/src/core/ext/census/census_interface.h b/src/core/ext/census/census_interface.h deleted file mode 100644 index 113c2b16ef..0000000000 --- a/src/core/ext/census/census_interface.h +++ /dev/null @@ -1,69 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_CENSUS_INTERFACE_H -#define GRPC_CORE_EXT_CENSUS_CENSUS_INTERFACE_H - -#include <grpc/support/port_platform.h> - -/* Maximum length of an individual census trace annotation. */ -#define CENSUS_MAX_ANNOTATION_LENGTH 200 - -#ifdef __cplusplus -extern "C" { -#endif - -/* Structure of a census op id. Define as structure because 64bit integer is not - available on every platform for C89. */ -typedef struct census_op_id { - uint32_t upper; - uint32_t lower; -} census_op_id; - -typedef struct census_rpc_stats census_rpc_stats; - -/* Initializes Census library. No-op if Census is already initialized. */ -void census_init(void); - -/* Shutdown Census Library. */ -void census_shutdown(void); - -/* Annotates grpc method name on a census_op_id. The method name has the format - of <full quantified rpc service name>/<rpc function name>. Returns 0 iff - op_id and method_name are all valid. op_id is valid after its creation and - before calling census_tracing_end_op(). - - TODO(hongyu): Figure out valid characters set for service name and command - name and document requirements here.*/ -int census_add_method_tag(census_op_id op_id, const char *method_name); - -/* Annotates tracing information to a specific op_id. - Up to CENSUS_MAX_ANNOTATION_LENGTH bytes are recorded. */ -void census_tracing_print(census_op_id op_id, const char *annotation); - -/* Starts tracing for an RPC. Returns a locally unique census_op_id */ -census_op_id census_tracing_start_op(void); - -/* Ends tracing. Calling this function will invalidate the input op_id. */ -void census_tracing_end_op(census_op_id op_id); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_CENSUS_INTERFACE_H */ diff --git a/src/core/ext/census/census_log.cc b/src/core/ext/census/census_log.cc deleted file mode 100644 index 100047f12b..0000000000 --- a/src/core/ext/census/census_log.cc +++ /dev/null @@ -1,588 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -/* Available log space is divided up in blocks of - CENSUS_LOG_2_MAX_RECORD_SIZE bytes. A block can be in one of the - following three data structures: - - Free blocks (free_block_list) - - Blocks with unread data (dirty_block_list) - - Blocks currently attached to cores (core_local_blocks[]) - - census_log_start_write() moves a block from core_local_blocks[] to the - end of dirty_block_list when block: - - is out-of-space OR - - has an incomplete record (an incomplete record occurs when a thread calls - census_log_start_write() and is context-switched before calling - census_log_end_write() - So, blocks in dirty_block_list are ordered, from oldest to newest, by time - when block is detached from the core. - - census_log_read_next() first iterates over dirty_block_list and then - core_local_blocks[]. It moves completely read blocks from dirty_block_list - to free_block_list. Blocks in core_local_blocks[] are not freed, even when - completely read. - - If log is configured to discard old records and free_block_list is empty, - census_log_start_write() iterates over dirty_block_list to allocate a - new block. It moves the oldest available block (no pending read/write) to - core_local_blocks[]. - - core_local_block_struct is used to implement a map from core id to the block - associated with that core. This mapping is advisory. It is possible that the - block returned by this mapping is no longer associated with that core. This - mapping is updated, lazily, by census_log_start_write(). - - Locking in block struct: - - Exclusive g_log.lock must be held before calling any functions operatong on - block structs except census_log_start_write() and - census_log_end_write(). - - Writes to a block are serialized via writer_lock. - census_log_start_write() acquires this lock and - census_log_end_write() releases it. On failure to acquire the lock, - writer allocates a new block for the current core and updates - core_local_block accordingly. - - Simultaneous read and write access is allowed. Reader can safely read up to - committed bytes (bytes_committed). - - reader_lock protects the block, currently being read, from getting recycled. - start_read() acquires reader_lock and end_read() releases the lock. - - Read/write access to a block is disabled via try_disable_access(). It returns - with both writer_lock and reader_lock held. These locks are subsequently - released by enable_access() to enable access to the block. - - A note on naming: Most function/struct names are prepended by cl_ - (shorthand for census_log). Further, functions that manipulate structures - include the name of the structure, which will be passed as the first - argument. E.g. cl_block_initialize() will initialize a cl_block. -*/ -#include "src/core/ext/census/census_log.h" -#include <grpc/support/alloc.h> -#include <grpc/support/atm.h> -#include <grpc/support/cpu.h> -#include <grpc/support/log.h> -#include <grpc/support/port_platform.h> -#include <grpc/support/sync.h> -#include <grpc/support/useful.h> -#include <string.h> - -/* End of platform specific code */ - -typedef struct census_log_block_list_struct { - struct census_log_block_list_struct *next; - struct census_log_block_list_struct *prev; - struct census_log_block *block; -} cl_block_list_struct; - -typedef struct census_log_block { - /* Pointer to underlying buffer */ - char *buffer; - gpr_atm writer_lock; - gpr_atm reader_lock; - /* Keeps completely written bytes. Declared atomic because accessed - simultaneously by reader and writer. */ - gpr_atm bytes_committed; - /* Bytes already read */ - int32_t bytes_read; - /* Links for list */ - cl_block_list_struct link; -/* We want this structure to be cacheline aligned. We assume the following - sizes for the various parts on 32/64bit systems: - type 32b size 64b size - char* 4 8 - 3x gpr_atm 12 24 - int32_t 4 8 (assumes padding) - cl_block_list_struct 12 24 - TOTAL 32 64 - - Depending on the size of our cacheline and the architecture, we - selectively add char buffering to this structure. The size is checked - via assert in census_log_initialize(). */ -#if defined(GPR_ARCH_64) -#define CL_BLOCK_PAD_SIZE (GPR_CACHELINE_SIZE - 64) -#else -#if defined(GPR_ARCH_32) -#define CL_BLOCK_PAD_SIZE (GPR_CACHELINE_SIZE - 32) -#else -#error "Unknown architecture" -#endif -#endif -#if CL_BLOCK_PAD_SIZE > 0 - char padding[CL_BLOCK_PAD_SIZE]; -#endif -} cl_block; - -/* A list of cl_blocks, doubly-linked through cl_block::link. */ -typedef struct census_log_block_list { - int32_t count; /* Number of items in list. */ - cl_block_list_struct ht; /* head/tail of linked list. */ -} cl_block_list; - -/* Cacheline aligned block pointers to avoid false sharing. Block pointer must - be initialized via set_block(), before calling other functions */ -typedef struct census_log_core_local_block { - gpr_atm block; -/* Ensure cachline alignment: we assume sizeof(gpr_atm) == 4 or 8 */ -#if defined(GPR_ARCH_64) -#define CL_CORE_LOCAL_BLOCK_PAD_SIZE (GPR_CACHELINE_SIZE - 8) -#else -#if defined(GPR_ARCH_32) -#define CL_CORE_LOCAL_BLOCK_PAD_SIZE (GPR_CACHELINE_SIZE - 4) -#else -#error "Unknown architecture" -#endif -#endif -#if CL_CORE_LOCAL_BLOCK_PAD_SIZE > 0 - char padding[CL_CORE_LOCAL_BLOCK_PAD_SIZE]; -#endif -} cl_core_local_block; - -struct census_log { - int discard_old_records; - /* Number of cores (aka hardware-contexts) */ - unsigned num_cores; - /* number of CENSUS_LOG_2_MAX_RECORD_SIZE blocks in log */ - int32_t num_blocks; - cl_block *blocks; /* Block metadata. */ - cl_core_local_block *core_local_blocks; /* Keeps core to block mappings. */ - gpr_mu lock; - int initialized; /* has log been initialized? */ - /* Keeps the state of the reader iterator. A value of 0 indicates that - iterator has reached the end. census_log_init_reader() resets the - value to num_core to restart iteration. */ - uint32_t read_iterator_state; - /* Points to the block being read. If non-NULL, the block is locked for - reading (block_being_read_->reader_lock is held). */ - cl_block *block_being_read; - /* A non-zero value indicates that log is full. */ - gpr_atm is_full; - char *buffer; - cl_block_list free_block_list; - cl_block_list dirty_block_list; - gpr_atm out_of_space_count; -}; - -/* Single internal log */ -static struct census_log g_log; - -/* Functions that operate on an atomic memory location used as a lock */ - -/* Returns non-zero if lock is acquired */ -static int cl_try_lock(gpr_atm *lock) { return gpr_atm_acq_cas(lock, 0, 1); } - -static void cl_unlock(gpr_atm *lock) { gpr_atm_rel_store(lock, 0); } - -/* Functions that operate on cl_core_local_block's */ - -static void cl_core_local_block_set_block(cl_core_local_block *clb, - cl_block *block) { - gpr_atm_rel_store(&clb->block, (gpr_atm)block); -} - -static cl_block *cl_core_local_block_get_block(cl_core_local_block *clb) { - return (cl_block *)gpr_atm_acq_load(&clb->block); -} - -/* Functions that operate on cl_block_list_struct's */ - -static void cl_block_list_struct_initialize(cl_block_list_struct *bls, - cl_block *block) { - bls->next = bls->prev = bls; - bls->block = block; -} - -/* Functions that operate on cl_block_list's */ - -static void cl_block_list_initialize(cl_block_list *list) { - list->count = 0; - cl_block_list_struct_initialize(&list->ht, NULL); -} - -/* Returns head of *this, or NULL if empty. */ -static cl_block *cl_block_list_head(cl_block_list *list) { - return list->ht.next->block; -} - -/* Insert element *e after *pos. */ -static void cl_block_list_insert(cl_block_list *list, cl_block_list_struct *pos, - cl_block_list_struct *e) { - list->count++; - e->next = pos->next; - e->prev = pos; - e->next->prev = e; - e->prev->next = e; -} - -/* Insert block at the head of the list */ -static void cl_block_list_insert_at_head(cl_block_list *list, cl_block *block) { - cl_block_list_insert(list, &list->ht, &block->link); -} - -/* Insert block at the tail of the list */ -static void cl_block_list_insert_at_tail(cl_block_list *list, cl_block *block) { - cl_block_list_insert(list, list->ht.prev, &block->link); -} - -/* Removes block *b. Requires *b be in the list. */ -static void cl_block_list_remove(cl_block_list *list, cl_block *b) { - list->count--; - b->link.next->prev = b->link.prev; - b->link.prev->next = b->link.next; -} - -/* Functions that operate on cl_block's */ - -static void cl_block_initialize(cl_block *block, char *buffer) { - block->buffer = buffer; - gpr_atm_rel_store(&block->writer_lock, 0); - gpr_atm_rel_store(&block->reader_lock, 0); - gpr_atm_rel_store(&block->bytes_committed, 0); - block->bytes_read = 0; - cl_block_list_struct_initialize(&block->link, block); -} - -/* Guards against exposing partially written buffer to the reader. */ -static void cl_block_set_bytes_committed(cl_block *block, - int32_t bytes_committed) { - gpr_atm_rel_store(&block->bytes_committed, bytes_committed); -} - -static int32_t cl_block_get_bytes_committed(cl_block *block) { - return gpr_atm_acq_load(&block->bytes_committed); -} - -/* Tries to disable future read/write access to this block. Succeeds if: - - no in-progress write AND - - no in-progress read AND - - 'discard_data' set to true OR no unread data - On success, clears the block state and returns with writer_lock_ and - reader_lock_ held. These locks are released by a subsequent - cl_block_access_enable() call. */ -static int cl_block_try_disable_access(cl_block *block, int discard_data) { - if (!cl_try_lock(&block->writer_lock)) { - return 0; - } - if (!cl_try_lock(&block->reader_lock)) { - cl_unlock(&block->writer_lock); - return 0; - } - if (!discard_data && - (block->bytes_read != cl_block_get_bytes_committed(block))) { - cl_unlock(&block->reader_lock); - cl_unlock(&block->writer_lock); - return 0; - } - cl_block_set_bytes_committed(block, 0); - block->bytes_read = 0; - return 1; -} - -static void cl_block_enable_access(cl_block *block) { - cl_unlock(&block->reader_lock); - cl_unlock(&block->writer_lock); -} - -/* Returns with writer_lock held. */ -static void *cl_block_start_write(cl_block *block, size_t size) { - int32_t bytes_committed; - if (!cl_try_lock(&block->writer_lock)) { - return NULL; - } - bytes_committed = cl_block_get_bytes_committed(block); - if (bytes_committed + size > CENSUS_LOG_MAX_RECORD_SIZE) { - cl_unlock(&block->writer_lock); - return NULL; - } - return block->buffer + bytes_committed; -} - -/* Releases writer_lock and increments committed bytes by 'bytes_written'. - 'bytes_written' must be <= 'size' specified in the corresponding - StartWrite() call. This function is thread-safe. */ -static void cl_block_end_write(cl_block *block, size_t bytes_written) { - cl_block_set_bytes_committed( - block, cl_block_get_bytes_committed(block) + bytes_written); - cl_unlock(&block->writer_lock); -} - -/* Returns a pointer to the first unread byte in buffer. The number of bytes - available are returned in 'bytes_available'. Acquires reader lock that is - released by a subsequent cl_block_end_read() call. Returns NULL if: - - read in progress - - no data available */ -static void *cl_block_start_read(cl_block *block, size_t *bytes_available) { - void *record; - if (!cl_try_lock(&block->reader_lock)) { - return NULL; - } - /* bytes_committed may change from under us. Use bytes_available to update - bytes_read below. */ - *bytes_available = cl_block_get_bytes_committed(block) - block->bytes_read; - if (*bytes_available == 0) { - cl_unlock(&block->reader_lock); - return NULL; - } - record = block->buffer + block->bytes_read; - block->bytes_read += *bytes_available; - return record; -} - -static void cl_block_end_read(cl_block *block) { - cl_unlock(&block->reader_lock); -} - -/* Internal functions operating on g_log */ - -/* Allocates a new free block (or recycles an available dirty block if log is - configured to discard old records). Returns NULL if out-of-space. */ -static cl_block *cl_allocate_block(void) { - cl_block *block = cl_block_list_head(&g_log.free_block_list); - if (block != NULL) { - cl_block_list_remove(&g_log.free_block_list, block); - return block; - } - if (!g_log.discard_old_records) { - /* No free block and log is configured to keep old records. */ - return NULL; - } - /* Recycle dirty block. Start from the oldest. */ - for (block = cl_block_list_head(&g_log.dirty_block_list); block != NULL; - block = block->link.next->block) { - if (cl_block_try_disable_access(block, 1 /* discard data */)) { - cl_block_list_remove(&g_log.dirty_block_list, block); - return block; - } - } - return NULL; -} - -/* Allocates a new block and updates core id => block mapping. 'old_block' - points to the block that the caller thinks is attached to - 'core_id'. 'old_block' may be NULL. Returns non-zero if: - - allocated a new block OR - - 'core_id' => 'old_block' mapping changed (another thread allocated a - block before lock was acquired). */ -static int cl_allocate_core_local_block(int32_t core_id, cl_block *old_block) { - /* Now that we have the lock, check if core-local mapping has changed. */ - cl_core_local_block *core_local_block = &g_log.core_local_blocks[core_id]; - cl_block *block = cl_core_local_block_get_block(core_local_block); - if ((block != NULL) && (block != old_block)) { - return 1; - } - if (block != NULL) { - cl_core_local_block_set_block(core_local_block, NULL); - cl_block_list_insert_at_tail(&g_log.dirty_block_list, block); - } - block = cl_allocate_block(); - if (block == NULL) { - gpr_atm_rel_store(&g_log.is_full, 1); - return 0; - } - cl_core_local_block_set_block(core_local_block, block); - cl_block_enable_access(block); - return 1; -} - -static cl_block *cl_get_block(void *record) { - uintptr_t p = (uintptr_t)((char *)record - g_log.buffer); - uintptr_t index = p >> CENSUS_LOG_2_MAX_RECORD_SIZE; - return &g_log.blocks[index]; -} - -/* Gets the next block to read and tries to free 'prev' block (if not NULL). - Returns NULL if reached the end. */ -static cl_block *cl_next_block_to_read(cl_block *prev) { - cl_block *block = NULL; - if (g_log.read_iterator_state == g_log.num_cores) { - /* We are traversing dirty list; find the next dirty block. */ - if (prev != NULL) { - /* Try to free the previous block if there is no unread data. This block - may have unread data if previously incomplete record completed between - read_next() calls. */ - block = prev->link.next->block; - if (cl_block_try_disable_access(prev, 0 /* do not discard data */)) { - cl_block_list_remove(&g_log.dirty_block_list, prev); - cl_block_list_insert_at_head(&g_log.free_block_list, prev); - gpr_atm_rel_store(&g_log.is_full, 0); - } - } else { - block = cl_block_list_head(&g_log.dirty_block_list); - } - if (block != NULL) { - return block; - } - /* We are done with the dirty list; moving on to core-local blocks. */ - } - while (g_log.read_iterator_state > 0) { - g_log.read_iterator_state--; - block = cl_core_local_block_get_block( - &g_log.core_local_blocks[g_log.read_iterator_state]); - if (block != NULL) { - return block; - } - } - return NULL; -} - -/* External functions: primary stats_log interface */ -void census_log_initialize(size_t size_in_mb, int discard_old_records) { - int32_t ix; - /* Check cacheline alignment. */ - GPR_ASSERT(sizeof(cl_block) % GPR_CACHELINE_SIZE == 0); - GPR_ASSERT(sizeof(cl_core_local_block) % GPR_CACHELINE_SIZE == 0); - GPR_ASSERT(!g_log.initialized); - g_log.discard_old_records = discard_old_records; - g_log.num_cores = gpr_cpu_num_cores(); - /* Ensure at least as many blocks as there are cores. */ - g_log.num_blocks = GPR_MAX( - g_log.num_cores, (size_in_mb << 20) >> CENSUS_LOG_2_MAX_RECORD_SIZE); - gpr_mu_init(&g_log.lock); - g_log.read_iterator_state = 0; - g_log.block_being_read = NULL; - gpr_atm_rel_store(&g_log.is_full, 0); - g_log.core_local_blocks = (cl_core_local_block *)gpr_malloc_aligned( - g_log.num_cores * sizeof(cl_core_local_block), GPR_CACHELINE_SIZE_LOG); - memset(g_log.core_local_blocks, 0, - g_log.num_cores * sizeof(cl_core_local_block)); - g_log.blocks = (cl_block *)gpr_malloc_aligned( - g_log.num_blocks * sizeof(cl_block), GPR_CACHELINE_SIZE_LOG); - memset(g_log.blocks, 0, g_log.num_blocks * sizeof(cl_block)); - g_log.buffer = gpr_malloc(g_log.num_blocks * CENSUS_LOG_MAX_RECORD_SIZE); - memset(g_log.buffer, 0, g_log.num_blocks * CENSUS_LOG_MAX_RECORD_SIZE); - cl_block_list_initialize(&g_log.free_block_list); - cl_block_list_initialize(&g_log.dirty_block_list); - for (ix = 0; ix < g_log.num_blocks; ++ix) { - cl_block *block = g_log.blocks + ix; - cl_block_initialize(block, - g_log.buffer + (CENSUS_LOG_MAX_RECORD_SIZE * ix)); - cl_block_try_disable_access(block, 1 /* discard data */); - cl_block_list_insert_at_tail(&g_log.free_block_list, block); - } - gpr_atm_rel_store(&g_log.out_of_space_count, 0); - g_log.initialized = 1; -} - -void census_log_shutdown(void) { - GPR_ASSERT(g_log.initialized); - gpr_mu_destroy(&g_log.lock); - gpr_free_aligned(g_log.core_local_blocks); - g_log.core_local_blocks = NULL; - gpr_free_aligned(g_log.blocks); - g_log.blocks = NULL; - gpr_free(g_log.buffer); - g_log.buffer = NULL; - g_log.initialized = 0; -} - -void *census_log_start_write(size_t size) { - /* Used to bound number of times block allocation is attempted. */ - int32_t attempts_remaining = g_log.num_blocks; - /* TODO(aveitch): move this inside the do loop when current_cpu is fixed */ - int32_t core_id = gpr_cpu_current_cpu(); - GPR_ASSERT(g_log.initialized); - if (size > CENSUS_LOG_MAX_RECORD_SIZE) { - return NULL; - } - do { - int allocated; - void *record = NULL; - cl_block *block = - cl_core_local_block_get_block(&g_log.core_local_blocks[core_id]); - if (block && (record = cl_block_start_write(block, size))) { - return record; - } - /* Need to allocate a new block. We are here if: - - No block associated with the core OR - - Write in-progress on the block OR - - block is out of space */ - if (gpr_atm_acq_load(&g_log.is_full)) { - gpr_atm_no_barrier_fetch_add(&g_log.out_of_space_count, 1); - return NULL; - } - gpr_mu_lock(&g_log.lock); - allocated = cl_allocate_core_local_block(core_id, block); - gpr_mu_unlock(&g_log.lock); - if (!allocated) { - gpr_atm_no_barrier_fetch_add(&g_log.out_of_space_count, 1); - return NULL; - } - } while (attempts_remaining--); - /* Give up. */ - gpr_atm_no_barrier_fetch_add(&g_log.out_of_space_count, 1); - return NULL; -} - -void census_log_end_write(void *record, size_t bytes_written) { - GPR_ASSERT(g_log.initialized); - cl_block_end_write(cl_get_block(record), bytes_written); -} - -void census_log_init_reader(void) { - GPR_ASSERT(g_log.initialized); - gpr_mu_lock(&g_log.lock); - /* If a block is locked for reading unlock it. */ - if (g_log.block_being_read != NULL) { - cl_block_end_read(g_log.block_being_read); - g_log.block_being_read = NULL; - } - g_log.read_iterator_state = g_log.num_cores; - gpr_mu_unlock(&g_log.lock); -} - -const void *census_log_read_next(size_t *bytes_available) { - GPR_ASSERT(g_log.initialized); - gpr_mu_lock(&g_log.lock); - if (g_log.block_being_read != NULL) { - cl_block_end_read(g_log.block_being_read); - } - do { - g_log.block_being_read = cl_next_block_to_read(g_log.block_being_read); - if (g_log.block_being_read != NULL) { - void *record = - cl_block_start_read(g_log.block_being_read, bytes_available); - if (record != NULL) { - gpr_mu_unlock(&g_log.lock); - return record; - } - } - } while (g_log.block_being_read != NULL); - gpr_mu_unlock(&g_log.lock); - return NULL; -} - -size_t census_log_remaining_space(void) { - size_t space; - GPR_ASSERT(g_log.initialized); - gpr_mu_lock(&g_log.lock); - if (g_log.discard_old_records) { - /* Remaining space is not meaningful; just return the entire log space. */ - space = g_log.num_blocks << CENSUS_LOG_2_MAX_RECORD_SIZE; - } else { - space = g_log.free_block_list.count * CENSUS_LOG_MAX_RECORD_SIZE; - } - gpr_mu_unlock(&g_log.lock); - return space; -} - -int census_log_out_of_space_count(void) { - GPR_ASSERT(g_log.initialized); - return gpr_atm_acq_load(&g_log.out_of_space_count); -} diff --git a/src/core/ext/census/census_log.h b/src/core/ext/census/census_log.h deleted file mode 100644 index ee336ae733..0000000000 --- a/src/core/ext/census/census_log.h +++ /dev/null @@ -1,84 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_CENSUS_LOG_H -#define GRPC_CORE_EXT_CENSUS_CENSUS_LOG_H - -#include <stddef.h> - -/* Maximum record size, in bytes. */ -#define CENSUS_LOG_2_MAX_RECORD_SIZE 14 /* 2^14 = 16KB */ -#define CENSUS_LOG_MAX_RECORD_SIZE (1 << CENSUS_LOG_2_MAX_RECORD_SIZE) - -#ifdef __cplusplus -extern "C" { -#endif - -/* Initialize the statistics logging subsystem with the given log size. A log - size of 0 will result in the smallest possible log for the platform - (approximately CENSUS_LOG_MAX_RECORD_SIZE * gpr_cpu_num_cores()). If - discard_old_records is non-zero, then new records will displace older ones - when the log is full. This function must be called before any other - census_log functions. -*/ -void census_log_initialize(size_t size_in_mb, int discard_old_records); - -/* Shutdown the logging subsystem. Caller must ensure that: - - no in progress or future call to any census_log functions - - no incomplete records -*/ -void census_log_shutdown(void); - -/* Allocates and returns a 'size' bytes record and marks it in use. A - subsequent census_log_end_write() marks the record complete. The - 'bytes_written' census_log_end_write() argument must be <= - 'size'. Returns NULL if out-of-space AND: - - log is configured to keep old records OR - - all blocks are pinned by incomplete records. -*/ -void *census_log_start_write(size_t size); - -void census_log_end_write(void *record, size_t bytes_written); - -/* census_log_read_next() iterates over blocks with data and for each block - returns a pointer to the first unread byte. The number of bytes that can be - read are returned in 'bytes_available'. Reader is expected to read all - available data. Reading the data consumes it i.e. it cannot be read again. - census_log_read_next() returns NULL if the end is reached i.e last block - is read. census_log_init_reader() starts the iteration or aborts the - current iteration. -*/ -void census_log_init_reader(void); -const void *census_log_read_next(size_t *bytes_available); - -/* Returns estimated remaining space across all blocks, in bytes. If log is - configured to discard old records, returns total log space. Otherwise, - returns space available in empty blocks (partially filled blocks are - treated as full). -*/ -size_t census_log_remaining_space(void); - -/* Returns the number of times grpc_stats_log_start_write() failed due to - out-of-space. */ -int census_log_out_of_space_count(void); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_CENSUS_LOG_H */ diff --git a/src/core/ext/census/census_rpc_stats.cc b/src/core/ext/census/census_rpc_stats.cc deleted file mode 100644 index 0aca1f109e..0000000000 --- a/src/core/ext/census/census_rpc_stats.cc +++ /dev/null @@ -1,238 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include <string.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/sync.h> -#include "src/core/ext/census/census_interface.h" -#include "src/core/ext/census/census_rpc_stats.h" -#include "src/core/ext/census/census_tracing.h" -#include "src/core/ext/census/hash_table.h" -#include "src/core/ext/census/window_stats.h" -#include "src/core/lib/support/murmur_hash.h" -#include "src/core/lib/support/string.h" - -#define NUM_INTERVALS 3 -#define MINUTE_INTERVAL 0 -#define HOUR_INTERVAL 1 -#define TOTAL_INTERVAL 2 - -/* for easier typing */ -typedef census_per_method_rpc_stats per_method_stats; - -/* Ensure mu is only initialized once. */ -static gpr_once g_stats_store_mu_init = GPR_ONCE_INIT; -/* Guards two stats stores. */ -static gpr_mu g_mu; -static census_ht *g_client_stats_store = NULL; -static census_ht *g_server_stats_store = NULL; - -static void init_mutex(void) { gpr_mu_init(&g_mu); } - -static void init_mutex_once(void) { - gpr_once_init(&g_stats_store_mu_init, init_mutex); -} - -static int cmp_str_keys(const void *k1, const void *k2) { - return strcmp((const char *)k1, (const char *)k2); -} - -/* TODO(hongyu): replace it with cityhash64 */ -static uint64_t simple_hash(const void *k) { - size_t len = strlen(k); - uint64_t higher = gpr_murmur_hash3((const char *)k, len / 2, 0); - return higher << 32 | - gpr_murmur_hash3((const char *)k + len / 2, len - len / 2, 0); -} - -static void delete_stats(void *stats) { - census_window_stats_destroy((struct census_window_stats *)stats); -} - -static void delete_key(void *key) { gpr_free(key); } - -static const census_ht_option ht_opt = { - CENSUS_HT_POINTER /* key type */, 1999 /* n_of_buckets */, - simple_hash /* hash function */, cmp_str_keys /* key comparator */, - delete_stats /* data deleter */, delete_key /* key deleter */ -}; - -static void init_rpc_stats(void *stats) { - memset(stats, 0, sizeof(census_rpc_stats)); -} - -static void stat_add_proportion(double p, void *base, const void *addme) { - census_rpc_stats *b = (census_rpc_stats *)base; - census_rpc_stats *a = (census_rpc_stats *)addme; - b->cnt += p * a->cnt; - b->rpc_error_cnt += p * a->rpc_error_cnt; - b->app_error_cnt += p * a->app_error_cnt; - b->elapsed_time_ms += p * a->elapsed_time_ms; - b->api_request_bytes += p * a->api_request_bytes; - b->wire_request_bytes += p * a->wire_request_bytes; - b->api_response_bytes += p * a->api_response_bytes; - b->wire_response_bytes += p * a->wire_response_bytes; -} - -static void stat_add(void *base, const void *addme) { - stat_add_proportion(1.0, base, addme); -} - -static gpr_timespec min_hour_total_intervals[3] = { - {60, 0}, {3600, 0}, {36000000, 0}}; - -static const census_window_stats_stat_info window_stats_settings = { - sizeof(census_rpc_stats), init_rpc_stats, stat_add, stat_add_proportion}; - -census_rpc_stats *census_rpc_stats_create_empty(void) { - census_rpc_stats *ret = - (census_rpc_stats *)gpr_malloc(sizeof(census_rpc_stats)); - memset(ret, 0, sizeof(census_rpc_stats)); - return ret; -} - -void census_aggregated_rpc_stats_set_empty(census_aggregated_rpc_stats *data) { - int i = 0; - for (i = 0; i < data->num_entries; i++) { - if (data->stats[i].method != NULL) { - gpr_free((void *)data->stats[i].method); - } - } - if (data->stats != NULL) { - gpr_free(data->stats); - } - data->num_entries = 0; - data->stats = NULL; -} - -static void record_stats(census_ht *store, census_op_id op_id, - const census_rpc_stats *stats) { - gpr_mu_lock(&g_mu); - if (store != NULL) { - census_trace_obj *trace = NULL; - census_internal_lock_trace_store(); - trace = census_get_trace_obj_locked(op_id); - if (trace != NULL) { - const char *method_name = census_get_trace_method_name(trace); - struct census_window_stats *window_stats = NULL; - census_ht_key key; - key.ptr = (void *)method_name; - window_stats = census_ht_find(store, key); - census_internal_unlock_trace_store(); - if (window_stats == NULL) { - window_stats = census_window_stats_create(3, min_hour_total_intervals, - 30, &window_stats_settings); - key.ptr = gpr_strdup(key.ptr); - census_ht_insert(store, key, (void *)window_stats); - } - census_window_stats_add(window_stats, gpr_now(GPR_CLOCK_REALTIME), stats); - } else { - census_internal_unlock_trace_store(); - } - } - gpr_mu_unlock(&g_mu); -} - -void census_record_rpc_client_stats(census_op_id op_id, - const census_rpc_stats *stats) { - record_stats(g_client_stats_store, op_id, stats); -} - -void census_record_rpc_server_stats(census_op_id op_id, - const census_rpc_stats *stats) { - record_stats(g_server_stats_store, op_id, stats); -} - -/* Get stats from input stats store */ -static void get_stats(census_ht *store, census_aggregated_rpc_stats *data) { - GPR_ASSERT(data != NULL); - if (data->num_entries != 0) { - census_aggregated_rpc_stats_set_empty(data); - } - gpr_mu_lock(&g_mu); - if (store != NULL) { - size_t n; - unsigned i, j; - gpr_timespec now = gpr_now(GPR_CLOCK_REALTIME); - census_ht_kv *kv = census_ht_get_all_elements(store, &n); - if (kv != NULL) { - data->num_entries = n; - data->stats = - (per_method_stats *)gpr_malloc(sizeof(per_method_stats) * n); - for (i = 0; i < n; i++) { - census_window_stats_sums sums[NUM_INTERVALS]; - for (j = 0; j < NUM_INTERVALS; j++) { - sums[j].statistic = (void *)census_rpc_stats_create_empty(); - } - data->stats[i].method = gpr_strdup(kv[i].k.ptr); - census_window_stats_get_sums(kv[i].v, now, sums); - data->stats[i].minute_stats = - *(census_rpc_stats *)sums[MINUTE_INTERVAL].statistic; - data->stats[i].hour_stats = - *(census_rpc_stats *)sums[HOUR_INTERVAL].statistic; - data->stats[i].total_stats = - *(census_rpc_stats *)sums[TOTAL_INTERVAL].statistic; - for (j = 0; j < NUM_INTERVALS; j++) { - gpr_free(sums[j].statistic); - } - } - gpr_free(kv); - } - } - gpr_mu_unlock(&g_mu); -} - -void census_get_client_stats(census_aggregated_rpc_stats *data) { - get_stats(g_client_stats_store, data); -} - -void census_get_server_stats(census_aggregated_rpc_stats *data) { - get_stats(g_server_stats_store, data); -} - -void census_stats_store_init(void) { - init_mutex_once(); - gpr_mu_lock(&g_mu); - if (g_client_stats_store == NULL && g_server_stats_store == NULL) { - g_client_stats_store = census_ht_create(&ht_opt); - g_server_stats_store = census_ht_create(&ht_opt); - } else { - gpr_log(GPR_ERROR, "Census stats store already initialized."); - } - gpr_mu_unlock(&g_mu); -} - -void census_stats_store_shutdown(void) { - init_mutex_once(); - gpr_mu_lock(&g_mu); - if (g_client_stats_store != NULL) { - census_ht_destroy(g_client_stats_store); - g_client_stats_store = NULL; - } else { - gpr_log(GPR_ERROR, "Census server stats store not initialized."); - } - if (g_server_stats_store != NULL) { - census_ht_destroy(g_server_stats_store); - g_server_stats_store = NULL; - } else { - gpr_log(GPR_ERROR, "Census client stats store not initialized."); - } - gpr_mu_unlock(&g_mu); -} diff --git a/src/core/ext/census/census_rpc_stats.h b/src/core/ext/census/census_rpc_stats.h deleted file mode 100644 index 8004ade37d..0000000000 --- a/src/core/ext/census/census_rpc_stats.h +++ /dev/null @@ -1,86 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_CENSUS_RPC_STATS_H -#define GRPC_CORE_EXT_CENSUS_CENSUS_RPC_STATS_H - -#include <grpc/support/port_platform.h> -#include "src/core/ext/census/census_interface.h" - -#ifdef __cplusplus -extern "C" { -#endif - -struct census_rpc_stats { - uint64_t cnt; - uint64_t rpc_error_cnt; - uint64_t app_error_cnt; - double elapsed_time_ms; - double api_request_bytes; - double wire_request_bytes; - double api_response_bytes; - double wire_response_bytes; -}; - -/* Creates an empty rpc stats object on heap. */ -census_rpc_stats *census_rpc_stats_create_empty(void); - -typedef struct census_per_method_rpc_stats { - const char *method; - census_rpc_stats minute_stats; /* cumulative stats in the past minute */ - census_rpc_stats hour_stats; /* cumulative stats in the past hour */ - census_rpc_stats total_stats; /* cumulative stats from last gc */ -} census_per_method_rpc_stats; - -typedef struct census_aggregated_rpc_stats { - int num_entries; - census_per_method_rpc_stats *stats; -} census_aggregated_rpc_stats; - -/* Initializes an aggregated rpc stats object to an empty state. */ -void census_aggregated_rpc_stats_set_empty(census_aggregated_rpc_stats *data); - -/* Records client side stats of a rpc. */ -void census_record_rpc_client_stats(census_op_id op_id, - const census_rpc_stats *stats); - -/* Records server side stats of a rpc. */ -void census_record_rpc_server_stats(census_op_id op_id, - const census_rpc_stats *stats); - -/* The following two functions are intended for inprocess query of - per-service per-method stats from grpc implementations. */ - -/* Populates *data_map with server side aggregated per-service per-method - stats. - DO NOT CALL from outside of grpc code. */ -void census_get_server_stats(census_aggregated_rpc_stats *data_map); - -/* Populates *data_map with client side aggregated per-service per-method - stats. - DO NOT CALL from outside of grpc code. */ -void census_get_client_stats(census_aggregated_rpc_stats *data_map); - -void census_stats_store_init(void); -void census_stats_store_shutdown(void); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_CENSUS_RPC_STATS_H */ diff --git a/src/core/ext/census/census_tracing.cc b/src/core/ext/census/census_tracing.cc deleted file mode 100644 index 199b260dd1..0000000000 --- a/src/core/ext/census/census_tracing.cc +++ /dev/null @@ -1,226 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include "src/core/ext/census/census_tracing.h" -#include "src/core/ext/census/census_interface.h" - -#include <stdio.h> -#include <string.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/port_platform.h> -#include <grpc/support/sync.h> -#include "src/core/ext/census/hash_table.h" -#include "src/core/lib/support/string.h" - -void census_trace_obj_destroy(census_trace_obj *obj) { - census_trace_annotation *p = obj->annotations; - while (p != NULL) { - census_trace_annotation *next = p->next; - gpr_free(p); - p = next; - } - gpr_free(obj->method); - gpr_free(obj); -} - -static void delete_trace_obj(void *obj) { - census_trace_obj_destroy((census_trace_obj *)obj); -} - -static const census_ht_option ht_opt = { - CENSUS_HT_UINT64 /* key type */, - 571 /* n_of_buckets */, - NULL /* hash */, - NULL /* compare_keys */, - delete_trace_obj /* delete data */, - NULL /* delete key */ -}; - -static gpr_once g_init_mutex_once = GPR_ONCE_INIT; -static gpr_mu g_mu; /* Guards following two static variables. */ -static census_ht *g_trace_store = NULL; -static uint64_t g_id = 0; - -static census_ht_key op_id_as_key(census_op_id *id) { - return *(census_ht_key *)id; -} - -static uint64_t op_id_2_uint64(census_op_id *id) { - uint64_t ret; - memcpy(&ret, id, sizeof(census_op_id)); - return ret; -} - -static void init_mutex(void) { gpr_mu_init(&g_mu); } - -static void init_mutex_once(void) { - gpr_once_init(&g_init_mutex_once, init_mutex); -} - -census_op_id census_tracing_start_op(void) { - gpr_mu_lock(&g_mu); - { - census_trace_obj *ret = gpr_malloc(sizeof(census_trace_obj)); - memset(ret, 0, sizeof(census_trace_obj)); - g_id++; - memcpy(&ret->id, &g_id, sizeof(census_op_id)); - ret->rpc_stats.cnt = 1; - ret->ts = gpr_now(GPR_CLOCK_REALTIME); - census_ht_insert(g_trace_store, op_id_as_key(&ret->id), (void *)ret); - gpr_log(GPR_DEBUG, "Start tracing for id %lu", g_id); - gpr_mu_unlock(&g_mu); - return ret->id; - } -} - -int census_add_method_tag(census_op_id op_id, const char *method) { - int ret = 0; - census_trace_obj *trace = NULL; - gpr_mu_lock(&g_mu); - trace = census_ht_find(g_trace_store, op_id_as_key(&op_id)); - if (trace == NULL) { - ret = 1; - } else { - trace->method = gpr_strdup(method); - } - gpr_mu_unlock(&g_mu); - return ret; -} - -void census_tracing_print(census_op_id op_id, const char *anno_txt) { - census_trace_obj *trace = NULL; - gpr_mu_lock(&g_mu); - trace = census_ht_find(g_trace_store, op_id_as_key(&op_id)); - if (trace != NULL) { - census_trace_annotation *anno = gpr_malloc(sizeof(census_trace_annotation)); - anno->ts = gpr_now(GPR_CLOCK_REALTIME); - { - char *d = anno->txt; - const char *s = anno_txt; - int n = 0; - for (; n < CENSUS_MAX_ANNOTATION_LENGTH && *s != '\0'; ++n) { - *d++ = *s++; - } - *d = '\0'; - } - anno->next = trace->annotations; - trace->annotations = anno; - } - gpr_mu_unlock(&g_mu); -} - -void census_tracing_end_op(census_op_id op_id) { - census_trace_obj *trace = NULL; - gpr_mu_lock(&g_mu); - trace = census_ht_find(g_trace_store, op_id_as_key(&op_id)); - if (trace != NULL) { - trace->rpc_stats.elapsed_time_ms = gpr_timespec_to_micros( - gpr_time_sub(gpr_now(GPR_CLOCK_REALTIME), trace->ts)); - gpr_log(GPR_DEBUG, "End tracing for id %lu, method %s, latency %f us", - op_id_2_uint64(&op_id), trace->method, - trace->rpc_stats.elapsed_time_ms); - census_ht_erase(g_trace_store, op_id_as_key(&op_id)); - } - gpr_mu_unlock(&g_mu); -} - -void census_tracing_init(void) { - init_mutex_once(); - gpr_mu_lock(&g_mu); - if (g_trace_store == NULL) { - g_id = 1; - g_trace_store = census_ht_create(&ht_opt); - } else { - gpr_log(GPR_ERROR, "Census trace store already initialized."); - } - gpr_mu_unlock(&g_mu); -} - -void census_tracing_shutdown(void) { - gpr_mu_lock(&g_mu); - if (g_trace_store != NULL) { - census_ht_destroy(g_trace_store); - g_trace_store = NULL; - } else { - gpr_log(GPR_ERROR, "Census trace store is not initialized."); - } - gpr_mu_unlock(&g_mu); -} - -void census_internal_lock_trace_store(void) { gpr_mu_lock(&g_mu); } - -void census_internal_unlock_trace_store(void) { gpr_mu_unlock(&g_mu); } - -census_trace_obj *census_get_trace_obj_locked(census_op_id op_id) { - if (g_trace_store == NULL) { - gpr_log(GPR_ERROR, "Census trace store is not initialized."); - return NULL; - } - return (census_trace_obj *)census_ht_find(g_trace_store, - op_id_as_key(&op_id)); -} - -const char *census_get_trace_method_name(const census_trace_obj *trace) { - return trace->method; -} - -static census_trace_annotation *dup_annotation_chain( - census_trace_annotation *from) { - census_trace_annotation *ret = NULL; - census_trace_annotation **to = &ret; - for (; from != NULL; from = from->next) { - *to = gpr_malloc(sizeof(census_trace_annotation)); - memcpy(*to, from, sizeof(census_trace_annotation)); - to = &(*to)->next; - } - return ret; -} - -static census_trace_obj *trace_obj_dup(census_trace_obj *from) { - census_trace_obj *to = NULL; - GPR_ASSERT(from != NULL); - to = gpr_malloc(sizeof(census_trace_obj)); - to->id = from->id; - to->ts = from->ts; - to->rpc_stats = from->rpc_stats; - to->method = gpr_strdup(from->method); - to->annotations = dup_annotation_chain(from->annotations); - return to; -} - -census_trace_obj **census_get_active_ops(int *num_active_ops) { - census_trace_obj **ret = NULL; - gpr_mu_lock(&g_mu); - if (g_trace_store != NULL) { - size_t n = 0; - census_ht_kv *all_kvs = census_ht_get_all_elements(g_trace_store, &n); - *num_active_ops = (int)n; - if (n != 0) { - size_t i = 0; - ret = gpr_malloc(sizeof(census_trace_obj *) * n); - for (i = 0; i < n; i++) { - ret[i] = trace_obj_dup((census_trace_obj *)all_kvs[i].v); - } - } - gpr_free(all_kvs); - } - gpr_mu_unlock(&g_mu); - return ret; -} diff --git a/src/core/ext/census/census_tracing.h b/src/core/ext/census/census_tracing.h deleted file mode 100644 index ccb767fcd0..0000000000 --- a/src/core/ext/census/census_tracing.h +++ /dev/null @@ -1,81 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_CENSUS_TRACING_H -#define GRPC_CORE_EXT_CENSUS_CENSUS_TRACING_H - -#include <grpc/support/time.h> -#include "src/core/ext/census/census_rpc_stats.h" - -/* WARNING: The data structures and APIs provided by this file are for GRPC - library's internal use ONLY. They might be changed in backward-incompatible - ways and are not subject to any deprecation policy. - They are not recommended for external use. - */ -#ifdef __cplusplus -extern "C" { -#endif - -/* Struct for a trace annotation. */ -typedef struct census_trace_annotation { - gpr_timespec ts; /* timestamp of the annotation */ - char txt[CENSUS_MAX_ANNOTATION_LENGTH + 1]; /* actual txt annotation */ - struct census_trace_annotation *next; -} census_trace_annotation; - -typedef struct census_trace_obj { - census_op_id id; - gpr_timespec ts; - census_rpc_stats rpc_stats; - char *method; - census_trace_annotation *annotations; -} census_trace_obj; - -/* Deletes trace object. */ -void census_trace_obj_destroy(census_trace_obj *obj); - -/* Initializes trace store. This function is thread safe. */ -void census_tracing_init(void); - -/* Shutsdown trace store. This function is thread safe. */ -void census_tracing_shutdown(void); - -/* Gets trace obj corresponding to the input op_id. Returns NULL if trace store - is not initialized or trace obj is not found. Requires trace store being - locked before calling this function. */ -census_trace_obj *census_get_trace_obj_locked(census_op_id op_id); - -/* The following two functions acquire and release the trace store global lock. - They are for census internal use only. */ -void census_internal_lock_trace_store(void); -void census_internal_unlock_trace_store(void); - -/* Gets method name associated with the input trace object. */ -const char *census_get_trace_method_name(const census_trace_obj *trace); - -/* Returns an array of pointers to trace objects of currently active operations - and fills in number of active operations. Returns NULL if there are no active - operations. - Caller owns the returned objects. */ -census_trace_obj **census_get_active_ops(int *num_active_ops); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_CENSUS_TRACING_H */ diff --git a/src/core/ext/census/context.cc b/src/core/ext/census/context.cc deleted file mode 100644 index 9b25a32e36..0000000000 --- a/src/core/ext/census/context.cc +++ /dev/null @@ -1,496 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include <grpc/census.h> -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/port_platform.h> -#include <grpc/support/useful.h> -#include <stdbool.h> -#include <string.h> -#include "src/core/lib/support/string.h" - -// Functions in this file support the public context API, including -// encoding/decoding as part of context propagation across RPC's. The overall -// requirements (in approximate priority order) for the -// context representation: -// 1. Efficient conversion to/from wire format -// 2. Minimal bytes used on-wire -// 3. Efficient context creation -// 4. Efficient lookup of tag value for a key -// 5. Efficient iteration over tags -// 6. Minimal memory footprint -// -// Notes on tradeoffs/decisions: -// * tag includes 1 byte length of key, as well as nil-terminating byte. These -// are to aid in efficient parsing and the ability to directly return key -// strings. This is more important than saving a single byte/tag on the wire. -// * The wire encoding uses only single byte values. This eliminates the need -// to handle endian-ness conversions. It also means there is a hard upper -// limit of 255 for both CENSUS_MAX_TAG_KV_LEN and CENSUS_MAX_PROPAGATED_TAGS. -// * Keep all tag information (keys/values/flags) in a single memory buffer, -// that can be directly copied to the wire. - -// min and max valid chars in tag keys and values. All printable ASCII is OK. -#define MIN_VALID_TAG_CHAR 32 // ' ' -#define MAX_VALID_TAG_CHAR 126 // '~' - -// Structure representing a set of tags. Essentially a count of number of tags -// present, and pointer to a chunk of memory that contains the per-tag details. -struct tag_set { - int ntags; // number of tags. - int ntags_alloc; // ntags + number of deleted tags (total number of tags - // in all of kvm). This will always be == ntags, except during the process - // of building a new tag set. - size_t kvm_size; // number of bytes allocated for key/value storage. - size_t kvm_used; // number of bytes of used key/value memory - char *kvm; // key/value memory. Consists of repeated entries of: - // Offset Size Description - // 0 1 Key length, including trailing 0. (K) - // 1 1 Value length, including trailing 0 (V) - // 2 1 Flags - // 3 K Key bytes - // 3 + K V Value bytes - // - // We refer to the first 3 entries as the 'tag header'. If extra values are - // introduced in the header, you will need to modify the TAG_HEADER_SIZE - // constant, the raw_tag structure (and everything that uses it) and the - // encode/decode functions appropriately. -}; - -// Number of bytes in tag header. -#define TAG_HEADER_SIZE 3 // key length (1) + value length (1) + flags (1) -// Offsets to tag header entries. -#define KEY_LEN_OFFSET 0 -#define VALUE_LEN_OFFSET 1 -#define FLAG_OFFSET 2 - -// raw_tag represents the raw-storage form of a tag in the kvm of a tag_set. -struct raw_tag { - uint8_t key_len; - uint8_t value_len; - uint8_t flags; - char *key; - char *value; -}; - -// Use a reserved flag bit for indication of deleted tag. -#define CENSUS_TAG_DELETED CENSUS_TAG_RESERVED -#define CENSUS_TAG_IS_DELETED(flags) (flags & CENSUS_TAG_DELETED) - -// Primary representation of a context. Composed of 2 underlying tag_set -// structs, one each for propagated and local (non-propagated) tags. This is -// to efficiently support tag encoding/decoding. -// TODO(aveitch): need to add tracing id's/structure. -struct census_context { - struct tag_set tags[2]; - census_context_status status; -}; - -// Indices into the tags member of census_context -#define PROPAGATED_TAGS 0 -#define LOCAL_TAGS 1 - -// Validate (check all characters are in range and size is less than limit) a -// key or value string. Returns 0 if the string is invalid, or the length -// (including terminator) if valid. -static size_t validate_tag(const char *kv) { - size_t len = 1; - char ch; - while ((ch = *kv++) != 0) { - if (ch < MIN_VALID_TAG_CHAR || ch > MAX_VALID_TAG_CHAR) { - return 0; - } - len++; - } - if (len > CENSUS_MAX_TAG_KV_LEN) { - return 0; - } - return len; -} - -// Extract a raw tag given a pointer (raw) to the tag header. Allow for some -// extra bytes in the tag header (see encode/decode functions for usage: this -// allows for future expansion of the tag header). -static char *decode_tag(struct raw_tag *tag, char *header, int offset) { - tag->key_len = (uint8_t)(*header++); - tag->value_len = (uint8_t)(*header++); - tag->flags = (uint8_t)(*header++); - header += offset; - tag->key = header; - header += tag->key_len; - tag->value = header; - return header + tag->value_len; -} - -// Make a copy (in 'to') of an existing tag_set. -static void tag_set_copy(struct tag_set *to, const struct tag_set *from) { - memcpy(to, from, sizeof(struct tag_set)); - to->kvm = (char *)gpr_malloc(to->kvm_size); - memcpy(to->kvm, from->kvm, from->kvm_used); -} - -// Delete a tag from a tag_set, if it exists (returns true if it did). -static bool tag_set_delete_tag(struct tag_set *tags, const char *key, - size_t key_len) { - char *kvp = tags->kvm; - for (int i = 0; i < tags->ntags_alloc; i++) { - uint8_t *flags = (uint8_t *)(kvp + FLAG_OFFSET); - struct raw_tag tag; - kvp = decode_tag(&tag, kvp, 0); - if (CENSUS_TAG_IS_DELETED(tag.flags)) continue; - if ((key_len == tag.key_len) && (memcmp(key, tag.key, key_len) == 0)) { - *flags |= CENSUS_TAG_DELETED; - tags->ntags--; - return true; - } - } - return false; -} - -// Delete a tag from a context, return true if it existed. -static bool context_delete_tag(census_context *context, const census_tag *tag, - size_t key_len) { - return ( - tag_set_delete_tag(&context->tags[LOCAL_TAGS], tag->key, key_len) || - tag_set_delete_tag(&context->tags[PROPAGATED_TAGS], tag->key, key_len)); -} - -// Add a tag to a tag_set. Return true on success, false if the tag could -// not be added because of constraints on tag set size. This function should -// not be called if the tag may already exist (in a non-deleted state) in -// the tag_set, as that would result in two tags with the same key. -static bool tag_set_add_tag(struct tag_set *tags, const census_tag *tag, - size_t key_len, size_t value_len) { - if (tags->ntags == CENSUS_MAX_PROPAGATED_TAGS) { - return false; - } - const size_t tag_size = key_len + value_len + TAG_HEADER_SIZE; - if (tags->kvm_used + tag_size > tags->kvm_size) { - // allocate new memory if needed - tags->kvm_size += 2 * CENSUS_MAX_TAG_KV_LEN + TAG_HEADER_SIZE; - char *new_kvm = (char *)gpr_malloc(tags->kvm_size); - if (tags->kvm_used > 0) memcpy(new_kvm, tags->kvm, tags->kvm_used); - gpr_free(tags->kvm); - tags->kvm = new_kvm; - } - char *kvp = tags->kvm + tags->kvm_used; - *kvp++ = (char)key_len; - *kvp++ = (char)value_len; - // ensure reserved flags are not used. - *kvp++ = (char)(tag->flags & (CENSUS_TAG_PROPAGATE | CENSUS_TAG_STATS)); - memcpy(kvp, tag->key, key_len); - kvp += key_len; - memcpy(kvp, tag->value, value_len); - tags->kvm_used += tag_size; - tags->ntags++; - tags->ntags_alloc++; - return true; -} - -// Add/modify/delete a tag to/in a context. Caller must validate that tag key -// etc. are valid. -static void context_modify_tag(census_context *context, const census_tag *tag, - size_t key_len, size_t value_len) { - // First delete the tag if it is already present. - bool deleted = context_delete_tag(context, tag, key_len); - bool added = false; - if (CENSUS_TAG_IS_PROPAGATED(tag->flags)) { - added = tag_set_add_tag(&context->tags[PROPAGATED_TAGS], tag, key_len, - value_len); - } else { - added = - tag_set_add_tag(&context->tags[LOCAL_TAGS], tag, key_len, value_len); - } - - if (deleted) { - context->status.n_modified_tags++; - } else { - if (added) { - context->status.n_added_tags++; - } else { - context->status.n_ignored_tags++; - } - } -} - -// Remove memory used for deleted tags from a tag set. Basic algorithm: -// 1) Walk through tag set to find first deleted tag. Record where it is. -// 2) Find the next not-deleted tag. Copy all of kvm from there to the end -// "over" the deleted tags -// 3) repeat #1 and #2 until we have seen all tags -// 4) if we are still looking for a not-deleted tag, then all the end portion -// of the kvm is deleted. Just reduce the used amount of memory by the -// appropriate amount. -static void tag_set_flatten(struct tag_set *tags) { - if (tags->ntags == tags->ntags_alloc) return; - bool found_deleted = false; // found a deleted tag. - char *kvp = tags->kvm; - char *dbase = NULL; // record location of deleted tag - for (int i = 0; i < tags->ntags_alloc; i++) { - struct raw_tag tag; - char *next_kvp = decode_tag(&tag, kvp, 0); - if (found_deleted) { - if (!CENSUS_TAG_IS_DELETED(tag.flags)) { - ptrdiff_t reduce = kvp - dbase; // #bytes in deleted tags - GPR_ASSERT(reduce > 0); - ptrdiff_t copy_size = tags->kvm + tags->kvm_used - kvp; - GPR_ASSERT(copy_size > 0); - memmove(dbase, kvp, (size_t)copy_size); - tags->kvm_used -= (size_t)reduce; - next_kvp -= reduce; - found_deleted = false; - } - } else { - if (CENSUS_TAG_IS_DELETED(tag.flags)) { - dbase = kvp; - found_deleted = true; - } - } - kvp = next_kvp; - } - if (found_deleted) { - GPR_ASSERT(dbase > tags->kvm); - tags->kvm_used = (size_t)(dbase - tags->kvm); - } - tags->ntags_alloc = tags->ntags; -} - -census_context *census_context_create(const census_context *base, - const census_tag *tags, int ntags, - census_context_status const **status) { - census_context *context = - (census_context *)gpr_malloc(sizeof(census_context)); - // If we are given a base, copy it into our new tag set. Otherwise set it - // to zero/NULL everything. - if (base == NULL) { - memset(context, 0, sizeof(census_context)); - } else { - tag_set_copy(&context->tags[PROPAGATED_TAGS], &base->tags[PROPAGATED_TAGS]); - tag_set_copy(&context->tags[LOCAL_TAGS], &base->tags[LOCAL_TAGS]); - memset(&context->status, 0, sizeof(context->status)); - } - // Walk over the additional tags and, for those that aren't invalid, modify - // the context to add/replace/delete as required. - for (int i = 0; i < ntags; i++) { - const census_tag *tag = &tags[i]; - size_t key_len = validate_tag(tag->key); - // ignore the tag if it is invalid or too short. - if (key_len <= 1) { - context->status.n_invalid_tags++; - } else { - if (tag->value != NULL) { - size_t value_len = validate_tag(tag->value); - if (value_len != 0) { - context_modify_tag(context, tag, key_len, value_len); - } else { - context->status.n_invalid_tags++; - } - } else { - if (context_delete_tag(context, tag, key_len)) { - context->status.n_deleted_tags++; - } - } - } - } - // Remove any deleted tags, update status if needed, and return. - tag_set_flatten(&context->tags[PROPAGATED_TAGS]); - tag_set_flatten(&context->tags[LOCAL_TAGS]); - context->status.n_propagated_tags = context->tags[PROPAGATED_TAGS].ntags; - context->status.n_local_tags = context->tags[LOCAL_TAGS].ntags; - if (status) { - *status = &context->status; - } - return context; -} - -const census_context_status *census_context_get_status( - const census_context *context) { - return &context->status; -} - -void census_context_destroy(census_context *context) { - gpr_free(context->tags[PROPAGATED_TAGS].kvm); - gpr_free(context->tags[LOCAL_TAGS].kvm); - gpr_free(context); -} - -void census_context_initialize_iterator(const census_context *context, - census_context_iterator *iterator) { - iterator->context = context; - iterator->index = 0; - if (context->tags[PROPAGATED_TAGS].ntags != 0) { - iterator->base = PROPAGATED_TAGS; - iterator->kvm = context->tags[PROPAGATED_TAGS].kvm; - } else if (context->tags[LOCAL_TAGS].ntags != 0) { - iterator->base = LOCAL_TAGS; - iterator->kvm = context->tags[LOCAL_TAGS].kvm; - } else { - iterator->base = -1; - } -} - -int census_context_next_tag(census_context_iterator *iterator, - census_tag *tag) { - if (iterator->base < 0) { - return 0; - } - struct raw_tag raw; - iterator->kvm = decode_tag(&raw, iterator->kvm, 0); - tag->key = raw.key; - tag->value = raw.value; - tag->flags = raw.flags; - if (++iterator->index == iterator->context->tags[iterator->base].ntags) { - do { - if (iterator->base == LOCAL_TAGS) { - iterator->base = -1; - return 1; - } - } while (iterator->context->tags[++iterator->base].ntags == 0); - iterator->index = 0; - iterator->kvm = iterator->context->tags[iterator->base].kvm; - } - return 1; -} - -// Find a tag in a tag_set by key. Return true if found, false otherwise. -static bool tag_set_get_tag(const struct tag_set *tags, const char *key, - size_t key_len, census_tag *tag) { - char *kvp = tags->kvm; - for (int i = 0; i < tags->ntags; i++) { - struct raw_tag raw; - kvp = decode_tag(&raw, kvp, 0); - if (key_len == raw.key_len && memcmp(raw.key, key, key_len) == 0) { - tag->key = raw.key; - tag->value = raw.value; - tag->flags = raw.flags; - return true; - } - } - return false; -} - -int census_context_get_tag(const census_context *context, const char *key, - census_tag *tag) { - size_t key_len = strlen(key) + 1; - if (key_len == 1) { - return 0; - } - if (tag_set_get_tag(&context->tags[PROPAGATED_TAGS], key, key_len, tag) || - tag_set_get_tag(&context->tags[LOCAL_TAGS], key, key_len, tag)) { - return 1; - } - return 0; -} - -// Context encoding and decoding functions. -// -// Wire format for tag_set's on the wire: -// -// First, a tag set header: -// -// offset bytes description -// 0 1 version number -// 1 1 number of bytes in this header. This allows for future -// expansion. -// 2 1 number of bytes in each tag header. -// 3 1 ntags value from tag set. -// -// This is followed by the key/value memory from struct tag_set. - -#define ENCODED_VERSION 0 // Version number -#define ENCODED_HEADER_SIZE 4 // size of tag set header - -// Encode a tag set. Returns 0 if buffer is too small. -static size_t tag_set_encode(const struct tag_set *tags, char *buffer, - size_t buf_size) { - if (buf_size < ENCODED_HEADER_SIZE + tags->kvm_used) { - return 0; - } - buf_size -= ENCODED_HEADER_SIZE; - *buffer++ = (char)ENCODED_VERSION; - *buffer++ = (char)ENCODED_HEADER_SIZE; - *buffer++ = (char)TAG_HEADER_SIZE; - *buffer++ = (char)tags->ntags; - if (tags->ntags == 0) { - return ENCODED_HEADER_SIZE; - } - memcpy(buffer, tags->kvm, tags->kvm_used); - return ENCODED_HEADER_SIZE + tags->kvm_used; -} - -size_t census_context_encode(const census_context *context, char *buffer, - size_t buf_size) { - return tag_set_encode(&context->tags[PROPAGATED_TAGS], buffer, buf_size); -} - -// Decode a tag set. -static void tag_set_decode(struct tag_set *tags, const char *buffer, - size_t size) { - uint8_t version = (uint8_t)(*buffer++); - uint8_t header_size = (uint8_t)(*buffer++); - uint8_t tag_header_size = (uint8_t)(*buffer++); - tags->ntags = tags->ntags_alloc = (int)(*buffer++); - if (tags->ntags == 0) { - tags->ntags_alloc = 0; - tags->kvm_size = 0; - tags->kvm_used = 0; - tags->kvm = NULL; - return; - } - if (header_size != ENCODED_HEADER_SIZE) { - GPR_ASSERT(version != ENCODED_VERSION); - GPR_ASSERT(ENCODED_HEADER_SIZE < header_size); - buffer += (header_size - ENCODED_HEADER_SIZE); - } - tags->kvm_used = size - header_size; - tags->kvm_size = tags->kvm_used + CENSUS_MAX_TAG_KV_LEN; - tags->kvm = (char *)gpr_malloc(tags->kvm_size); - if (tag_header_size != TAG_HEADER_SIZE) { - // something new in the tag information. I don't understand it, so - // don't copy it over. - GPR_ASSERT(version != ENCODED_VERSION); - GPR_ASSERT(tag_header_size > TAG_HEADER_SIZE); - char *kvp = tags->kvm; - for (int i = 0; i < tags->ntags; i++) { - memcpy(kvp, buffer, TAG_HEADER_SIZE); - kvp += header_size; - struct raw_tag raw; - buffer = - decode_tag(&raw, (char *)buffer, tag_header_size - TAG_HEADER_SIZE); - memcpy(kvp, raw.key, (size_t)raw.key_len + raw.value_len); - kvp += raw.key_len + raw.value_len; - } - } else { - memcpy(tags->kvm, buffer, tags->kvm_used); - } -} - -census_context *census_context_decode(const char *buffer, size_t size) { - census_context *context = - (census_context *)gpr_malloc(sizeof(census_context)); - memset(&context->tags[LOCAL_TAGS], 0, sizeof(struct tag_set)); - if (buffer == NULL) { - memset(&context->tags[PROPAGATED_TAGS], 0, sizeof(struct tag_set)); - } else { - tag_set_decode(&context->tags[PROPAGATED_TAGS], buffer, size); - } - memset(&context->status, 0, sizeof(context->status)); - context->status.n_propagated_tags = context->tags[PROPAGATED_TAGS].ntags; - return context; -} diff --git a/src/core/ext/census/gen/README.md b/src/core/ext/census/gen/README.md deleted file mode 100644 index d4612bc7c8..0000000000 --- a/src/core/ext/census/gen/README.md +++ /dev/null @@ -1,10 +0,0 @@ -Files generated for use by Census stats and trace recording subsystem. - -# Files -* census.pb.{h,c} - Generated from src/core/ext/census/census.proto, using the - script `tools/codegen/core/gen_nano_proto.sh src/proto/census/census.proto - $PWD/src/core/ext/census/gen src/core/ext/census/gen` -* trace_context.pb.{h,c} - Generated from - src/core/ext/census/trace_context.proto, using the script - `tools/codegen/core/gen_nano_proto.sh src/proto/census/trace_context.proto - $PWD/src/core/ext/census/gen src/core/ext/census/gen` diff --git a/src/core/ext/census/gen/census.pb.c b/src/core/ext/census/gen/census.pb.c deleted file mode 100644 index 88efa73661..0000000000 --- a/src/core/ext/census/gen/census.pb.c +++ /dev/null @@ -1,161 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ -/* Automatically generated nanopb constant definitions */ -/* Generated by nanopb-0.3.5-dev */ - -#include "src/core/ext/census/gen/census.pb.h" - -#if PB_PROTO_HEADER_VERSION != 30 -#error Regenerate this file with the current version of nanopb generator. -#endif - - - -const pb_field_t google_census_Duration_fields[3] = { - PB_FIELD( 1, INT64 , OPTIONAL, STATIC , FIRST, google_census_Duration, seconds, seconds, 0), - PB_FIELD( 2, INT32 , OPTIONAL, STATIC , OTHER, google_census_Duration, nanos, seconds, 0), - PB_LAST_FIELD -}; - -const pb_field_t google_census_Timestamp_fields[3] = { - PB_FIELD( 1, INT64 , OPTIONAL, STATIC , FIRST, google_census_Timestamp, seconds, seconds, 0), - PB_FIELD( 2, INT32 , OPTIONAL, STATIC , OTHER, google_census_Timestamp, nanos, seconds, 0), - PB_LAST_FIELD -}; - -const pb_field_t google_census_Resource_fields[4] = { - PB_FIELD( 1, STRING , OPTIONAL, CALLBACK, FIRST, google_census_Resource, name, name, 0), - PB_FIELD( 2, STRING , OPTIONAL, CALLBACK, OTHER, google_census_Resource, description, name, 0), - PB_FIELD( 3, MESSAGE , OPTIONAL, STATIC , OTHER, google_census_Resource, unit, description, &google_census_Resource_MeasurementUnit_fields), - PB_LAST_FIELD -}; - -const pb_field_t google_census_Resource_MeasurementUnit_fields[4] = { - PB_FIELD( 1, INT32 , OPTIONAL, STATIC , FIRST, google_census_Resource_MeasurementUnit, prefix, prefix, 0), - PB_FIELD( 2, UENUM , REPEATED, CALLBACK, OTHER, google_census_Resource_MeasurementUnit, numerator, prefix, 0), - PB_FIELD( 3, UENUM , REPEATED, CALLBACK, OTHER, google_census_Resource_MeasurementUnit, denominator, numerator, 0), - PB_LAST_FIELD -}; - -const pb_field_t google_census_AggregationDescriptor_fields[4] = { - PB_FIELD( 1, UENUM , OPTIONAL, STATIC , FIRST, google_census_AggregationDescriptor, type, type, 0), - PB_ONEOF_FIELD(options, 2, MESSAGE , ONEOF, STATIC , OTHER, google_census_AggregationDescriptor, bucket_boundaries, type, &google_census_AggregationDescriptor_BucketBoundaries_fields), - PB_ONEOF_FIELD(options, 3, MESSAGE , ONEOF, STATIC , OTHER, google_census_AggregationDescriptor, interval_boundaries, type, &google_census_AggregationDescriptor_IntervalBoundaries_fields), - PB_LAST_FIELD -}; - -const pb_field_t google_census_AggregationDescriptor_BucketBoundaries_fields[2] = { - PB_FIELD( 1, DOUBLE , REPEATED, CALLBACK, FIRST, google_census_AggregationDescriptor_BucketBoundaries, bounds, bounds, 0), - PB_LAST_FIELD -}; - -const pb_field_t google_census_AggregationDescriptor_IntervalBoundaries_fields[2] = { - PB_FIELD( 1, DOUBLE , REPEATED, CALLBACK, FIRST, google_census_AggregationDescriptor_IntervalBoundaries, window_size, window_size, 0), - PB_LAST_FIELD -}; - -const pb_field_t google_census_Distribution_fields[5] = { - PB_FIELD( 1, INT64 , OPTIONAL, STATIC , FIRST, google_census_Distribution, count, count, 0), - PB_FIELD( 2, DOUBLE , OPTIONAL, STATIC , OTHER, google_census_Distribution, mean, count, 0), - PB_FIELD( 3, MESSAGE , OPTIONAL, STATIC , OTHER, google_census_Distribution, range, mean, &google_census_Distribution_Range_fields), - PB_FIELD( 4, INT64 , REPEATED, CALLBACK, OTHER, google_census_Distribution, bucket_count, range, 0), - PB_LAST_FIELD -}; - -const pb_field_t google_census_Distribution_Range_fields[3] = { - PB_FIELD( 1, DOUBLE , OPTIONAL, STATIC , FIRST, google_census_Distribution_Range, min, min, 0), - PB_FIELD( 2, DOUBLE , OPTIONAL, STATIC , OTHER, google_census_Distribution_Range, max, min, 0), - PB_LAST_FIELD -}; - -const pb_field_t google_census_IntervalStats_fields[2] = { - PB_FIELD( 1, MESSAGE , REPEATED, CALLBACK, FIRST, google_census_IntervalStats, window, window, &google_census_IntervalStats_Window_fields), - PB_LAST_FIELD -}; - -const pb_field_t google_census_IntervalStats_Window_fields[4] = { - PB_FIELD( 1, MESSAGE , OPTIONAL, STATIC , FIRST, google_census_IntervalStats_Window, window_size, window_size, &google_census_Duration_fields), - PB_FIELD( 2, INT64 , OPTIONAL, STATIC , OTHER, google_census_IntervalStats_Window, count, window_size, 0), - PB_FIELD( 3, DOUBLE , OPTIONAL, STATIC , OTHER, google_census_IntervalStats_Window, mean, count, 0), - PB_LAST_FIELD -}; - -const pb_field_t google_census_Tag_fields[3] = { - PB_FIELD( 1, STRING , OPTIONAL, STATIC , FIRST, google_census_Tag, key, key, 0), - PB_FIELD( 2, STRING , OPTIONAL, STATIC , OTHER, google_census_Tag, value, key, 0), - PB_LAST_FIELD -}; - -const pb_field_t google_census_View_fields[6] = { - PB_FIELD( 1, STRING , OPTIONAL, CALLBACK, FIRST, google_census_View, name, name, 0), - PB_FIELD( 2, STRING , OPTIONAL, CALLBACK, OTHER, google_census_View, description, name, 0), - PB_FIELD( 3, STRING , OPTIONAL, CALLBACK, OTHER, google_census_View, resource_name, description, 0), - PB_FIELD( 4, MESSAGE , OPTIONAL, STATIC , OTHER, google_census_View, aggregation, resource_name, &google_census_AggregationDescriptor_fields), - PB_FIELD( 5, STRING , REPEATED, CALLBACK, OTHER, google_census_View, tag_key, aggregation, 0), - PB_LAST_FIELD -}; - -const pb_field_t google_census_Aggregation_fields[7] = { - PB_FIELD( 1, STRING , OPTIONAL, CALLBACK, FIRST, google_census_Aggregation, name, name, 0), - PB_FIELD( 2, STRING , OPTIONAL, CALLBACK, OTHER, google_census_Aggregation, description, name, 0), - PB_ONEOF_FIELD(data, 3, UINT64 , ONEOF, STATIC , OTHER, google_census_Aggregation, count, description, 0), - PB_ONEOF_FIELD(data, 4, MESSAGE , ONEOF, STATIC , OTHER, google_census_Aggregation, distribution, description, &google_census_Distribution_fields), - PB_ONEOF_FIELD(data, 5, MESSAGE , ONEOF, STATIC , OTHER, google_census_Aggregation, interval_stats, description, &google_census_IntervalStats_fields), - PB_FIELD( 6, MESSAGE , REPEATED, CALLBACK, OTHER, google_census_Aggregation, tag, data.interval_stats, &google_census_Tag_fields), - PB_LAST_FIELD -}; - -const pb_field_t google_census_Metric_fields[5] = { - PB_FIELD( 1, STRING , OPTIONAL, CALLBACK, FIRST, google_census_Metric, view_name, view_name, 0), - PB_FIELD( 2, MESSAGE , REPEATED, CALLBACK, OTHER, google_census_Metric, aggregation, view_name, &google_census_Aggregation_fields), - PB_FIELD( 3, MESSAGE , OPTIONAL, STATIC , OTHER, google_census_Metric, start, aggregation, &google_census_Timestamp_fields), - PB_FIELD( 4, MESSAGE , OPTIONAL, STATIC , OTHER, google_census_Metric, end, start, &google_census_Timestamp_fields), - PB_LAST_FIELD -}; - - -/* Check that field information fits in pb_field_t */ -#if !defined(PB_FIELD_32BIT) -/* If you get an error here, it means that you need to define PB_FIELD_32BIT - * compile-time option. You can do that in pb.h or on compiler command line. - * - * The reason you need to do this is that some of your messages contain tag - * numbers or field sizes that are larger than what can fit in 8 or 16 bit - * field descriptors. - */ -PB_STATIC_ASSERT((pb_membersize(google_census_Resource, unit) < 65536 && pb_membersize(google_census_AggregationDescriptor, options.bucket_boundaries) < 65536 && pb_membersize(google_census_AggregationDescriptor, options.interval_boundaries) < 65536 && pb_membersize(google_census_Resource, unit) < 65536 && pb_membersize(google_census_AggregationDescriptor, options.bucket_boundaries) < 65536 && pb_membersize(google_census_AggregationDescriptor, options.interval_boundaries) < 65536 && pb_membersize(google_census_Distribution, range) < 65536 && pb_membersize(google_census_IntervalStats, window) < 65536 && pb_membersize(google_census_IntervalStats_Window, window_size) < 65536 && pb_membersize(google_census_View, aggregation) < 65536 && pb_membersize(google_census_Aggregation, data.distribution) < 65536 && pb_membersize(google_census_Aggregation, data.interval_stats) < 65536 && pb_membersize(google_census_Resource, unit) < 65536 && pb_membersize(google_census_AggregationDescriptor, options.bucket_boundaries) < 65536 && pb_membersize(google_census_AggregationDescriptor, options.interval_boundaries) < 65536 && pb_membersize(google_census_Resource, unit) < 65536 && pb_membersize(google_census_AggregationDescriptor, options.bucket_boundaries) < 65536 && pb_membersize(google_census_AggregationDescriptor, options.interval_boundaries) < 65536 && pb_membersize(google_census_Distribution, range) < 65536 && pb_membersize(google_census_IntervalStats, window) < 65536 && pb_membersize(google_census_IntervalStats_Window, window_size) < 65536 && pb_membersize(google_census_View, aggregation) < 65536 && pb_membersize(google_census_Aggregation, data.distribution) < 65536 && pb_membersize(google_census_Aggregation, data.interval_stats) < 65536 && pb_membersize(google_census_Aggregation, tag) < 65536 && pb_membersize(google_census_Metric, aggregation) < 65536 && pb_membersize(google_census_Metric, start) < 65536 && pb_membersize(google_census_Metric, end) < 65536), YOU_MUST_DEFINE_PB_FIELD_32BIT_FOR_MESSAGES_google_census_Duration_google_census_Timestamp_google_census_Resource_google_census_Resource_MeasurementUnit_google_census_AggregationDescriptor_google_census_AggregationDescriptor_BucketBoundaries_google_census_AggregationDescriptor_IntervalBoundaries_google_census_Distribution_google_census_Distribution_Range_google_census_IntervalStats_google_census_IntervalStats_Window_google_census_Tag_google_census_View_google_census_Aggregation_google_census_Metric) -#endif - -#if !defined(PB_FIELD_16BIT) && !defined(PB_FIELD_32BIT) -/* If you get an error here, it means that you need to define PB_FIELD_16BIT - * compile-time option. You can do that in pb.h or on compiler command line. - * - * The reason you need to do this is that some of your messages contain tag - * numbers or field sizes that are larger than what can fit in the default - * 8 bit descriptors. - */ -PB_STATIC_ASSERT((pb_membersize(google_census_Resource, unit) < 256 && pb_membersize(google_census_AggregationDescriptor, options.bucket_boundaries) < 256 && pb_membersize(google_census_AggregationDescriptor, options.interval_boundaries) < 256 && pb_membersize(google_census_Resource, unit) < 256 && pb_membersize(google_census_AggregationDescriptor, options.bucket_boundaries) < 256 && pb_membersize(google_census_AggregationDescriptor, options.interval_boundaries) < 256 && pb_membersize(google_census_Distribution, range) < 256 && pb_membersize(google_census_IntervalStats, window) < 256 && pb_membersize(google_census_IntervalStats_Window, window_size) < 256 && pb_membersize(google_census_View, aggregation) < 256 && pb_membersize(google_census_Aggregation, data.distribution) < 256 && pb_membersize(google_census_Aggregation, data.interval_stats) < 256 && pb_membersize(google_census_Resource, unit) < 256 && pb_membersize(google_census_AggregationDescriptor, options.bucket_boundaries) < 256 && pb_membersize(google_census_AggregationDescriptor, options.interval_boundaries) < 256 && pb_membersize(google_census_Resource, unit) < 256 && pb_membersize(google_census_AggregationDescriptor, options.bucket_boundaries) < 256 && pb_membersize(google_census_AggregationDescriptor, options.interval_boundaries) < 256 && pb_membersize(google_census_Distribution, range) < 256 && pb_membersize(google_census_IntervalStats, window) < 256 && pb_membersize(google_census_IntervalStats_Window, window_size) < 256 && pb_membersize(google_census_View, aggregation) < 256 && pb_membersize(google_census_Aggregation, data.distribution) < 256 && pb_membersize(google_census_Aggregation, data.interval_stats) < 256 && pb_membersize(google_census_Aggregation, tag) < 256 && pb_membersize(google_census_Metric, aggregation) < 256 && pb_membersize(google_census_Metric, start) < 256 && pb_membersize(google_census_Metric, end) < 256), YOU_MUST_DEFINE_PB_FIELD_16BIT_FOR_MESSAGES_google_census_Duration_google_census_Timestamp_google_census_Resource_google_census_Resource_MeasurementUnit_google_census_AggregationDescriptor_google_census_AggregationDescriptor_BucketBoundaries_google_census_AggregationDescriptor_IntervalBoundaries_google_census_Distribution_google_census_Distribution_Range_google_census_IntervalStats_google_census_IntervalStats_Window_google_census_Tag_google_census_View_google_census_Aggregation_google_census_Metric) -#endif - - -/* On some platforms (such as AVR), double is really float. - * These are not directly supported by nanopb, but see example_avr_double. - * To get rid of this error, remove any double fields from your .proto. - */ -PB_STATIC_ASSERT(sizeof(double) == 8, DOUBLE_MUST_BE_8_BYTES) - diff --git a/src/core/ext/census/gen/census.pb.h b/src/core/ext/census/gen/census.pb.h deleted file mode 100644 index 5f28335664..0000000000 --- a/src/core/ext/census/gen/census.pb.h +++ /dev/null @@ -1,280 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ -/* Automatically generated nanopb header */ -/* Generated by nanopb-0.3.5-dev */ - -#ifndef GRPC_CORE_EXT_CENSUS_GEN_CENSUS_PB_H -#define GRPC_CORE_EXT_CENSUS_GEN_CENSUS_PB_H -#include "third_party/nanopb/pb.h" -#if PB_PROTO_HEADER_VERSION != 30 -#error Regenerate this file with the current version of nanopb generator. -#endif - -#ifdef __cplusplus -extern "C" { -#endif - -/* Enum definitions */ -typedef enum _google_census_Resource_BasicUnit { - google_census_Resource_BasicUnit_UNKNOWN = 0, - google_census_Resource_BasicUnit_BITS = 1, - google_census_Resource_BasicUnit_BYTES = 2, - google_census_Resource_BasicUnit_SECS = 3, - google_census_Resource_BasicUnit_CORES = 4, - google_census_Resource_BasicUnit_MAX_UNITS = 5 -} google_census_Resource_BasicUnit; - -typedef enum _google_census_AggregationDescriptor_AggregationType { - google_census_AggregationDescriptor_AggregationType_UNKNOWN = 0, - google_census_AggregationDescriptor_AggregationType_COUNT = 1, - google_census_AggregationDescriptor_AggregationType_DISTRIBUTION = 2, - google_census_AggregationDescriptor_AggregationType_INTERVAL = 3 -} google_census_AggregationDescriptor_AggregationType; - -/* Struct definitions */ -typedef struct _google_census_AggregationDescriptor_BucketBoundaries { - pb_callback_t bounds; -} google_census_AggregationDescriptor_BucketBoundaries; - -typedef struct _google_census_AggregationDescriptor_IntervalBoundaries { - pb_callback_t window_size; -} google_census_AggregationDescriptor_IntervalBoundaries; - -typedef struct _google_census_IntervalStats { - pb_callback_t window; -} google_census_IntervalStats; - -typedef struct _google_census_AggregationDescriptor { - bool has_type; - google_census_AggregationDescriptor_AggregationType type; - pb_size_t which_options; - union { - google_census_AggregationDescriptor_BucketBoundaries bucket_boundaries; - google_census_AggregationDescriptor_IntervalBoundaries interval_boundaries; - } options; -} google_census_AggregationDescriptor; - -typedef struct _google_census_Distribution_Range { - bool has_min; - double min; - bool has_max; - double max; -} google_census_Distribution_Range; - -typedef struct _google_census_Duration { - bool has_seconds; - int64_t seconds; - bool has_nanos; - int32_t nanos; -} google_census_Duration; - -typedef struct _google_census_Resource_MeasurementUnit { - bool has_prefix; - int32_t prefix; - pb_callback_t numerator; - pb_callback_t denominator; -} google_census_Resource_MeasurementUnit; - -typedef struct _google_census_Tag { - bool has_key; - char key[255]; - bool has_value; - char value[255]; -} google_census_Tag; - -typedef struct _google_census_Timestamp { - bool has_seconds; - int64_t seconds; - bool has_nanos; - int32_t nanos; -} google_census_Timestamp; - -typedef struct _google_census_Distribution { - bool has_count; - int64_t count; - bool has_mean; - double mean; - bool has_range; - google_census_Distribution_Range range; - pb_callback_t bucket_count; -} google_census_Distribution; - -typedef struct _google_census_IntervalStats_Window { - bool has_window_size; - google_census_Duration window_size; - bool has_count; - int64_t count; - bool has_mean; - double mean; -} google_census_IntervalStats_Window; - -typedef struct _google_census_Metric { - pb_callback_t view_name; - pb_callback_t aggregation; - bool has_start; - google_census_Timestamp start; - bool has_end; - google_census_Timestamp end; -} google_census_Metric; - -typedef struct _google_census_Resource { - pb_callback_t name; - pb_callback_t description; - bool has_unit; - google_census_Resource_MeasurementUnit unit; -} google_census_Resource; - -typedef struct _google_census_View { - pb_callback_t name; - pb_callback_t description; - pb_callback_t resource_name; - bool has_aggregation; - google_census_AggregationDescriptor aggregation; - pb_callback_t tag_key; -} google_census_View; - -typedef struct _google_census_Aggregation { - pb_callback_t name; - pb_callback_t description; - pb_size_t which_data; - union { - uint64_t count; - google_census_Distribution distribution; - google_census_IntervalStats interval_stats; - } data; - pb_callback_t tag; -} google_census_Aggregation; - -/* Default values for struct fields */ - -/* Initializer values for message structs */ -#define google_census_Duration_init_default {false, 0, false, 0} -#define google_census_Timestamp_init_default {false, 0, false, 0} -#define google_census_Resource_init_default {{{NULL}, NULL}, {{NULL}, NULL}, false, google_census_Resource_MeasurementUnit_init_default} -#define google_census_Resource_MeasurementUnit_init_default {false, 0, {{NULL}, NULL}, {{NULL}, NULL}} -#define google_census_AggregationDescriptor_init_default {false, (google_census_AggregationDescriptor_AggregationType)0, 0, {google_census_AggregationDescriptor_BucketBoundaries_init_default}} -#define google_census_AggregationDescriptor_BucketBoundaries_init_default {{{NULL}, NULL}} -#define google_census_AggregationDescriptor_IntervalBoundaries_init_default {{{NULL}, NULL}} -#define google_census_Distribution_init_default {false, 0, false, 0, false, google_census_Distribution_Range_init_default, {{NULL}, NULL}} -#define google_census_Distribution_Range_init_default {false, 0, false, 0} -#define google_census_IntervalStats_init_default {{{NULL}, NULL}} -#define google_census_IntervalStats_Window_init_default {false, google_census_Duration_init_default, false, 0, false, 0} -#define google_census_Tag_init_default {false, "", false, ""} -#define google_census_View_init_default {{{NULL}, NULL}, {{NULL}, NULL}, {{NULL}, NULL}, false, google_census_AggregationDescriptor_init_default, {{NULL}, NULL}} -#define google_census_Aggregation_init_default {{{NULL}, NULL}, {{NULL}, NULL}, 0, {0}, {{NULL}, NULL}} -#define google_census_Metric_init_default {{{NULL}, NULL}, {{NULL}, NULL}, false, google_census_Timestamp_init_default, false, google_census_Timestamp_init_default} -#define google_census_Duration_init_zero {false, 0, false, 0} -#define google_census_Timestamp_init_zero {false, 0, false, 0} -#define google_census_Resource_init_zero {{{NULL}, NULL}, {{NULL}, NULL}, false, google_census_Resource_MeasurementUnit_init_zero} -#define google_census_Resource_MeasurementUnit_init_zero {false, 0, {{NULL}, NULL}, {{NULL}, NULL}} -#define google_census_AggregationDescriptor_init_zero {false, (google_census_AggregationDescriptor_AggregationType)0, 0, {google_census_AggregationDescriptor_BucketBoundaries_init_zero}} -#define google_census_AggregationDescriptor_BucketBoundaries_init_zero {{{NULL}, NULL}} -#define google_census_AggregationDescriptor_IntervalBoundaries_init_zero {{{NULL}, NULL}} -#define google_census_Distribution_init_zero {false, 0, false, 0, false, google_census_Distribution_Range_init_zero, {{NULL}, NULL}} -#define google_census_Distribution_Range_init_zero {false, 0, false, 0} -#define google_census_IntervalStats_init_zero {{{NULL}, NULL}} -#define google_census_IntervalStats_Window_init_zero {false, google_census_Duration_init_zero, false, 0, false, 0} -#define google_census_Tag_init_zero {false, "", false, ""} -#define google_census_View_init_zero {{{NULL}, NULL}, {{NULL}, NULL}, {{NULL}, NULL}, false, google_census_AggregationDescriptor_init_zero, {{NULL}, NULL}} -#define google_census_Aggregation_init_zero {{{NULL}, NULL}, {{NULL}, NULL}, 0, {0}, {{NULL}, NULL}} -#define google_census_Metric_init_zero {{{NULL}, NULL}, {{NULL}, NULL}, false, google_census_Timestamp_init_zero, false, google_census_Timestamp_init_zero} - -/* Field tags (for use in manual encoding/decoding) */ -#define google_census_AggregationDescriptor_BucketBoundaries_bounds_tag 1 -#define google_census_AggregationDescriptor_IntervalBoundaries_window_size_tag 1 -#define google_census_IntervalStats_window_tag 1 -#define google_census_AggregationDescriptor_bucket_boundaries_tag 2 - -#define google_census_AggregationDescriptor_interval_boundaries_tag 3 -#define google_census_AggregationDescriptor_type_tag 1 -#define google_census_Distribution_Range_min_tag 1 -#define google_census_Distribution_Range_max_tag 2 -#define google_census_Duration_seconds_tag 1 -#define google_census_Duration_nanos_tag 2 -#define google_census_Resource_MeasurementUnit_prefix_tag 1 -#define google_census_Resource_MeasurementUnit_numerator_tag 2 -#define google_census_Resource_MeasurementUnit_denominator_tag 3 -#define google_census_Tag_key_tag 1 -#define google_census_Tag_value_tag 2 -#define google_census_Timestamp_seconds_tag 1 -#define google_census_Timestamp_nanos_tag 2 -#define google_census_Distribution_count_tag 1 -#define google_census_Distribution_mean_tag 2 -#define google_census_Distribution_range_tag 3 -#define google_census_Distribution_bucket_count_tag 4 -#define google_census_IntervalStats_Window_window_size_tag 1 -#define google_census_IntervalStats_Window_count_tag 2 -#define google_census_IntervalStats_Window_mean_tag 3 -#define google_census_Metric_view_name_tag 1 -#define google_census_Metric_aggregation_tag 2 -#define google_census_Metric_start_tag 3 -#define google_census_Metric_end_tag 4 -#define google_census_Resource_name_tag 1 -#define google_census_Resource_description_tag 2 -#define google_census_Resource_unit_tag 3 -#define google_census_View_name_tag 1 -#define google_census_View_description_tag 2 -#define google_census_View_resource_name_tag 3 -#define google_census_View_aggregation_tag 4 -#define google_census_View_tag_key_tag 5 -#define google_census_Aggregation_count_tag 3 - -#define google_census_Aggregation_distribution_tag 4 - -#define google_census_Aggregation_interval_stats_tag 5 -#define google_census_Aggregation_name_tag 1 -#define google_census_Aggregation_description_tag 2 -#define google_census_Aggregation_tag_tag 6 - -/* Struct field encoding specification for nanopb */ -extern const pb_field_t google_census_Duration_fields[3]; -extern const pb_field_t google_census_Timestamp_fields[3]; -extern const pb_field_t google_census_Resource_fields[4]; -extern const pb_field_t google_census_Resource_MeasurementUnit_fields[4]; -extern const pb_field_t google_census_AggregationDescriptor_fields[4]; -extern const pb_field_t google_census_AggregationDescriptor_BucketBoundaries_fields[2]; -extern const pb_field_t google_census_AggregationDescriptor_IntervalBoundaries_fields[2]; -extern const pb_field_t google_census_Distribution_fields[5]; -extern const pb_field_t google_census_Distribution_Range_fields[3]; -extern const pb_field_t google_census_IntervalStats_fields[2]; -extern const pb_field_t google_census_IntervalStats_Window_fields[4]; -extern const pb_field_t google_census_Tag_fields[3]; -extern const pb_field_t google_census_View_fields[6]; -extern const pb_field_t google_census_Aggregation_fields[7]; -extern const pb_field_t google_census_Metric_fields[5]; - -/* Maximum encoded size of messages (where known) */ -#define google_census_Duration_size 22 -#define google_census_Timestamp_size 22 -#define google_census_Distribution_Range_size 18 -#define google_census_IntervalStats_Window_size 44 -#define google_census_Tag_size 516 - -/* Message IDs (where set with "msgid" option) */ -#ifdef PB_MSGID - -#define CENSUS_MESSAGES \ - - -#endif - -#ifdef __cplusplus -} /* extern "C" */ -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_GEN_CENSUS_PB_H */ diff --git a/src/core/ext/census/gen/trace_context.pb.c b/src/core/ext/census/gen/trace_context.pb.c deleted file mode 100644 index b5c3d52a71..0000000000 --- a/src/core/ext/census/gen/trace_context.pb.c +++ /dev/null @@ -1,39 +0,0 @@ -/* - * - * Copyright 2017 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ -/* Automatically generated nanopb constant definitions */ -/* Generated by nanopb-0.3.7-dev at Fri Jan 20 16:14:22 2017. */ - -#include "src/core/ext/census/gen/trace_context.pb.h" - -/* @@protoc_insertion_point(includes) */ -#if PB_PROTO_HEADER_VERSION != 30 -#error Regenerate this file with the current version of nanopb generator. -#endif - - - -const pb_field_t google_trace_TraceContext_fields[5] = { - PB_FIELD( 1, FIXED64 , OPTIONAL, STATIC , FIRST, google_trace_TraceContext, trace_id_hi, trace_id_hi, 0), - PB_FIELD( 2, FIXED64 , OPTIONAL, STATIC , OTHER, google_trace_TraceContext, trace_id_lo, trace_id_hi, 0), - PB_FIELD( 3, FIXED64 , OPTIONAL, STATIC , OTHER, google_trace_TraceContext, span_id, trace_id_lo, 0), - PB_FIELD( 4, FIXED32 , OPTIONAL, STATIC , OTHER, google_trace_TraceContext, span_options, span_id, 0), - PB_LAST_FIELD -}; - - -/* @@protoc_insertion_point(eof) */ diff --git a/src/core/ext/census/gen/trace_context.pb.h b/src/core/ext/census/gen/trace_context.pb.h deleted file mode 100644 index 181925dc9a..0000000000 --- a/src/core/ext/census/gen/trace_context.pb.h +++ /dev/null @@ -1,78 +0,0 @@ -/* - * - * Copyright 2017 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ -/* Automatically generated nanopb header */ -/* Generated by nanopb-0.3.7-dev at Fri Jan 20 16:14:22 2017. */ - -#ifndef GRPC_CORE_EXT_CENSUS_GEN_TRACE_CONTEXT_PB_H -#define GRPC_CORE_EXT_CENSUS_GEN_TRACE_CONTEXT_PB_H -#include "third_party/nanopb/pb.h" - -/* @@protoc_insertion_point(includes) */ -#if PB_PROTO_HEADER_VERSION != 30 -#error Regenerate this file with the current version of nanopb generator. -#endif - -#ifdef __cplusplus -extern "C" { -#endif - -/* Struct definitions */ -typedef struct _google_trace_TraceContext { - bool has_trace_id_hi; - uint64_t trace_id_hi; - bool has_trace_id_lo; - uint64_t trace_id_lo; - bool has_span_id; - uint64_t span_id; - bool has_span_options; - uint32_t span_options; -/* @@protoc_insertion_point(struct:google_trace_TraceContext) */ -} google_trace_TraceContext; - -/* Default values for struct fields */ - -/* Initializer values for message structs */ -#define google_trace_TraceContext_init_default {false, 0, false, 0, false, 0, false, 0} -#define google_trace_TraceContext_init_zero {false, 0, false, 0, false, 0, false, 0} - -/* Field tags (for use in manual encoding/decoding) */ -#define google_trace_TraceContext_trace_id_hi_tag 1 -#define google_trace_TraceContext_trace_id_lo_tag 2 -#define google_trace_TraceContext_span_id_tag 3 -#define google_trace_TraceContext_span_options_tag 4 - -/* Struct field encoding specification for nanopb */ -extern const pb_field_t google_trace_TraceContext_fields[5]; - -/* Maximum encoded size of messages (where known) */ -#define google_trace_TraceContext_size 32 - -/* Message IDs (where set with "msgid" option) */ -#ifdef PB_MSGID - -#define TRACE_CONTEXT_MESSAGES \ - - -#endif - -#ifdef __cplusplus -} /* extern "C" */ -#endif -/* @@protoc_insertion_point(eof) */ - -#endif /* GRPC_CORE_EXT_CENSUS_GEN_TRACE_CONTEXT_PB_H */ diff --git a/src/core/ext/census/grpc_context.cc b/src/core/ext/census/grpc_context.cc index 0bfba63a5e..34eafcab8e 100644 --- a/src/core/ext/census/grpc_context.cc +++ b/src/core/ext/census/grpc_context.cc @@ -24,9 +24,6 @@ void grpc_census_call_set_context(grpc_call *call, census_context *context) { GRPC_API_TRACE("grpc_census_call_set_context(call=%p, census_context=%p)", 2, (call, context)); - if (census_enabled() == CENSUS_FEATURE_NONE) { - return; - } if (context != NULL) { grpc_call_context_set(call, GRPC_CONTEXT_TRACING, context, NULL); } diff --git a/src/core/ext/census/grpc_filter.cc b/src/core/ext/census/grpc_filter.cc deleted file mode 100644 index b37ab90389..0000000000 --- a/src/core/ext/census/grpc_filter.cc +++ /dev/null @@ -1,196 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include "src/core/ext/census/grpc_filter.h" - -#include <stdio.h> -#include <string.h> - -#include <grpc/census.h> -#include <grpc/slice.h> -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/time.h> - -#include "src/core/ext/census/census_interface.h" -#include "src/core/ext/census/census_rpc_stats.h" -#include "src/core/lib/channel/channel_stack.h" -#include "src/core/lib/profiling/timers.h" -#include "src/core/lib/transport/static_metadata.h" - -typedef struct call_data { - census_op_id op_id; - census_context *ctxt; - gpr_timespec start_ts; - int error; - - /* recv callback */ - grpc_metadata_batch *recv_initial_metadata; - grpc_closure *on_done_recv; - grpc_closure finish_recv; -} call_data; - -typedef struct channel_data { uint8_t unused; } channel_data; - -static void extract_and_annotate_method_tag(grpc_metadata_batch *md, - call_data *calld, - channel_data *chand) { - grpc_linked_mdelem *m; - for (m = md->list.head; m != NULL; m = m->next) { - if (grpc_slice_eq(GRPC_MDKEY(m->md), GRPC_MDSTR_PATH)) { - /* Add method tag here */ - } - } -} - -static void client_mutate_op(grpc_call_element *elem, - grpc_transport_stream_op_batch *op) { - call_data *calld = (call_data *)elem->call_data; - channel_data *chand = (channel_data *)elem->channel_data; - if (op->send_initial_metadata) { - extract_and_annotate_method_tag( - op->payload->send_initial_metadata.send_initial_metadata, calld, chand); - } -} - -static void client_start_transport_op(grpc_exec_ctx *exec_ctx, - grpc_call_element *elem, - grpc_transport_stream_op_batch *op) { - client_mutate_op(elem, op); - grpc_call_next_op(exec_ctx, elem, op); -} - -static void server_on_done_recv(grpc_exec_ctx *exec_ctx, void *ptr, - grpc_error *error) { - GPR_TIMER_BEGIN("census-server:server_on_done_recv", 0); - grpc_call_element *elem = (grpc_call_element *)ptr; - call_data *calld = (call_data *)elem->call_data; - channel_data *chand = (channel_data *)elem->channel_data; - if (error == GRPC_ERROR_NONE) { - extract_and_annotate_method_tag(calld->recv_initial_metadata, calld, chand); - } - calld->on_done_recv->cb(exec_ctx, calld->on_done_recv->cb_arg, error); - GPR_TIMER_END("census-server:server_on_done_recv", 0); -} - -static void server_mutate_op(grpc_call_element *elem, - grpc_transport_stream_op_batch *op) { - call_data *calld = (call_data *)elem->call_data; - if (op->recv_initial_metadata) { - /* substitute our callback for the op callback */ - calld->recv_initial_metadata = - op->payload->recv_initial_metadata.recv_initial_metadata; - calld->on_done_recv = - op->payload->recv_initial_metadata.recv_initial_metadata_ready; - op->payload->recv_initial_metadata.recv_initial_metadata_ready = - &calld->finish_recv; - } -} - -static void server_start_transport_op(grpc_exec_ctx *exec_ctx, - grpc_call_element *elem, - grpc_transport_stream_op_batch *op) { - /* TODO(ctiller): this code fails. I don't know why. I expect it's - incomplete, and someone should look at it soon. - - call_data *calld = elem->call_data; - GPR_ASSERT((calld->op_id.upper != 0) || (calld->op_id.lower != 0)); */ - server_mutate_op(elem, op); - grpc_call_next_op(exec_ctx, elem, op); -} - -static grpc_error *client_init_call_elem(grpc_exec_ctx *exec_ctx, - grpc_call_element *elem, - const grpc_call_element_args *args) { - call_data *d = (call_data *)elem->call_data; - GPR_ASSERT(d != NULL); - memset(d, 0, sizeof(*d)); - d->start_ts = args->start_time; - return GRPC_ERROR_NONE; -} - -static void client_destroy_call_elem(grpc_exec_ctx *exec_ctx, - grpc_call_element *elem, - const grpc_call_final_info *final_info, - grpc_closure *ignored) { - call_data *d = (call_data *)elem->call_data; - GPR_ASSERT(d != NULL); - /* TODO(hongyu): record rpc client stats and census_rpc_end_op here */ -} - -static grpc_error *server_init_call_elem(grpc_exec_ctx *exec_ctx, - grpc_call_element *elem, - const grpc_call_element_args *args) { - call_data *d = (call_data *)elem->call_data; - GPR_ASSERT(d != NULL); - memset(d, 0, sizeof(*d)); - d->start_ts = args->start_time; - /* TODO(hongyu): call census_tracing_start_op here. */ - GRPC_CLOSURE_INIT(&d->finish_recv, server_on_done_recv, elem, - grpc_schedule_on_exec_ctx); - return GRPC_ERROR_NONE; -} - -static void server_destroy_call_elem(grpc_exec_ctx *exec_ctx, - grpc_call_element *elem, - const grpc_call_final_info *final_info, - grpc_closure *ignored) { - call_data *d = (call_data *)elem->call_data; - GPR_ASSERT(d != NULL); - /* TODO(hongyu): record rpc server stats and census_tracing_end_op here */ -} - -static grpc_error *init_channel_elem(grpc_exec_ctx *exec_ctx, - grpc_channel_element *elem, - grpc_channel_element_args *args) { - channel_data *chand = (channel_data *)elem->channel_data; - GPR_ASSERT(chand != NULL); - return GRPC_ERROR_NONE; -} - -static void destroy_channel_elem(grpc_exec_ctx *exec_ctx, - grpc_channel_element *elem) { - channel_data *chand = (channel_data *)elem->channel_data; - GPR_ASSERT(chand != NULL); -} - -const grpc_channel_filter grpc_client_census_filter = { - client_start_transport_op, - grpc_channel_next_op, - sizeof(call_data), - client_init_call_elem, - grpc_call_stack_ignore_set_pollset_or_pollset_set, - client_destroy_call_elem, - sizeof(channel_data), - init_channel_elem, - destroy_channel_elem, - grpc_channel_next_get_info, - "census-client"}; - -const grpc_channel_filter grpc_server_census_filter = { - server_start_transport_op, - grpc_channel_next_op, - sizeof(call_data), - server_init_call_elem, - grpc_call_stack_ignore_set_pollset_or_pollset_set, - server_destroy_call_elem, - sizeof(channel_data), - init_channel_elem, - destroy_channel_elem, - grpc_channel_next_get_info, - "census-server"}; diff --git a/src/core/ext/census/grpc_plugin.cc b/src/core/ext/census/grpc_plugin.cc deleted file mode 100644 index 22b16c6c63..0000000000 --- a/src/core/ext/census/grpc_plugin.cc +++ /dev/null @@ -1,70 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include <grpc/support/port_platform.h> - -#include <limits.h> -#include <string.h> - -#include <grpc/census.h> - -#include "src/core/ext/census/grpc_filter.h" -#include "src/core/lib/channel/channel_stack_builder.h" -#include "src/core/lib/surface/channel_init.h" - -static bool is_census_enabled(const grpc_channel_args *a) { - size_t i; - if (a == NULL) return 0; - for (i = 0; i < a->num_args; i++) { - if (0 == strcmp(a->args[i].key, GRPC_ARG_ENABLE_CENSUS)) { - return a->args[i].value.integer != 0 && census_enabled(); - } - } - return census_enabled() && !grpc_channel_args_want_minimal_stack(a); -} - -static bool maybe_add_census_filter(grpc_exec_ctx *exec_ctx, - grpc_channel_stack_builder *builder, - void *arg) { - const grpc_channel_args *args = - grpc_channel_stack_builder_get_channel_arguments(builder); - if (is_census_enabled(args)) { - return grpc_channel_stack_builder_prepend_filter( - builder, (const grpc_channel_filter *)arg, NULL, NULL); - } - return true; -} - -extern "C" void census_grpc_plugin_init(void) { - /* Only initialize census if no one else has and some features are - * available. */ - if (census_enabled() == CENSUS_FEATURE_NONE && - census_supported() != CENSUS_FEATURE_NONE) { - if (census_initialize(census_supported())) { /* enable all features. */ - gpr_log(GPR_ERROR, "Could not initialize census."); - } - } - grpc_channel_init_register_stage(GRPC_CLIENT_CHANNEL, INT_MAX, - maybe_add_census_filter, - (void *)&grpc_client_census_filter); - grpc_channel_init_register_stage(GRPC_SERVER_CHANNEL, INT_MAX, - maybe_add_census_filter, - (void *)&grpc_server_census_filter); -} - -extern "C" void census_grpc_plugin_shutdown(void) { census_shutdown(); } diff --git a/src/core/ext/census/hash_table.cc b/src/core/ext/census/hash_table.cc deleted file mode 100644 index 545b0857c7..0000000000 --- a/src/core/ext/census/hash_table.cc +++ /dev/null @@ -1,288 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include "src/core/ext/census/hash_table.h" - -#include <stddef.h> -#include <stdio.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/port_platform.h> - -#define CENSUS_HT_NUM_BUCKETS 1999 - -/* A single hash table data entry */ -typedef struct ht_entry { - census_ht_key key; - void *data; - struct ht_entry *next; -} ht_entry; - -/* hash table bucket */ -typedef struct bucket { - /* NULL if bucket is empty */ - ht_entry *next; - /* -1 if all buckets are empty. */ - int32_t prev_non_empty_bucket; - /* -1 if all buckets are empty. */ - int32_t next_non_empty_bucket; -} bucket; - -struct unresizable_hash_table { - /* Number of entries in the table */ - size_t size; - /* Number of buckets */ - uint32_t num_buckets; - /* Array of buckets initialized at creation time. Memory consumption is - 16 bytes per bucket on a 64-bit platform. */ - bucket *buckets; - /* Index of the first non-empty bucket. -1 iff size == 0. */ - int32_t first_non_empty_bucket; - /* Index of the last non_empty bucket. -1 iff size == 0. */ - int32_t last_non_empty_bucket; - /* Immutable options of this hash table, initialized at creation time. */ - census_ht_option options; -}; - -typedef struct entry_locator { - int32_t bucket_idx; - int is_first_in_chain; - int found; - ht_entry *prev_entry; -} entry_locator; - -/* Asserts if option is not valid. */ -void check_options(const census_ht_option *option) { - GPR_ASSERT(option != NULL); - GPR_ASSERT(option->num_buckets > 0); - GPR_ASSERT(option->key_type == CENSUS_HT_UINT64 || - option->key_type == CENSUS_HT_POINTER); - if (option->key_type == CENSUS_HT_UINT64) { - GPR_ASSERT(option->hash == NULL); - } else if (option->key_type == CENSUS_HT_POINTER) { - GPR_ASSERT(option->hash != NULL); - GPR_ASSERT(option->compare_keys != NULL); - } -} - -#define REMOVE_NEXT(options, ptr) \ - do { \ - ht_entry *tmp = (ptr)->next; \ - (ptr)->next = tmp->next; \ - delete_entry(options, tmp); \ - } while (0) - -static void delete_entry(const census_ht_option *opt, ht_entry *p) { - if (opt->delete_data != NULL) { - opt->delete_data(p->data); - } - if (opt->delete_key != NULL) { - opt->delete_key(p->key.ptr); - } - gpr_free(p); -} - -static uint64_t hash(const census_ht_option *opt, census_ht_key key) { - return opt->key_type == CENSUS_HT_UINT64 ? key.val : opt->hash(key.ptr); -} - -census_ht *census_ht_create(const census_ht_option *option) { - int i; - census_ht *ret = NULL; - check_options(option); - ret = (census_ht *)gpr_malloc(sizeof(census_ht)); - ret->size = 0; - ret->num_buckets = option->num_buckets; - ret->buckets = (bucket *)gpr_malloc(sizeof(bucket) * ret->num_buckets); - ret->options = *option; - /* initialize each bucket */ - for (i = 0; i < ret->options.num_buckets; i++) { - ret->buckets[i].prev_non_empty_bucket = -1; - ret->buckets[i].next_non_empty_bucket = -1; - ret->buckets[i].next = NULL; - } - return ret; -} - -static int32_t find_bucket_idx(const census_ht *ht, census_ht_key key) { - return hash(&ht->options, key) % ht->num_buckets; -} - -static int keys_match(const census_ht_option *opt, const ht_entry *p, - const census_ht_key key) { - GPR_ASSERT(opt->key_type == CENSUS_HT_UINT64 || - opt->key_type == CENSUS_HT_POINTER); - if (opt->key_type == CENSUS_HT_UINT64) return p->key.val == key.val; - return !opt->compare_keys((p->key).ptr, key.ptr); -} - -static entry_locator ht_find(const census_ht *ht, census_ht_key key) { - entry_locator loc = {0, 0, 0, NULL}; - int32_t idx = 0; - ht_entry *ptr = NULL; - GPR_ASSERT(ht != NULL); - idx = find_bucket_idx(ht, key); - ptr = ht->buckets[idx].next; - if (ptr == NULL) { - /* bucket is empty */ - return loc; - } - if (keys_match(&ht->options, ptr, key)) { - loc.bucket_idx = idx; - loc.is_first_in_chain = 1; - loc.found = 1; - return loc; - } else { - for (; ptr->next != NULL; ptr = ptr->next) { - if (keys_match(&ht->options, ptr->next, key)) { - loc.bucket_idx = idx; - loc.is_first_in_chain = 0; - loc.found = 1; - loc.prev_entry = ptr; - return loc; - } - } - } - /* Could not find the key */ - return loc; -} - -void *census_ht_find(const census_ht *ht, census_ht_key key) { - entry_locator loc = ht_find(ht, key); - if (loc.found == 0) { - return NULL; - } - return loc.is_first_in_chain ? ht->buckets[loc.bucket_idx].next->data - : loc.prev_entry->next->data; -} - -void census_ht_insert(census_ht *ht, census_ht_key key, void *data) { - int32_t idx = find_bucket_idx(ht, key); - ht_entry *ptr = NULL; - entry_locator loc = ht_find(ht, key); - if (loc.found) { - /* Replace old value with new value. */ - ptr = loc.is_first_in_chain ? ht->buckets[loc.bucket_idx].next - : loc.prev_entry->next; - if (ht->options.delete_data != NULL) { - ht->options.delete_data(ptr->data); - } - ptr->data = data; - return; - } - - /* first entry in the table. */ - if (ht->size == 0) { - ht->buckets[idx].next_non_empty_bucket = -1; - ht->buckets[idx].prev_non_empty_bucket = -1; - ht->first_non_empty_bucket = idx; - ht->last_non_empty_bucket = idx; - } else if (ht->buckets[idx].next == NULL) { - /* first entry in the bucket. */ - ht->buckets[ht->last_non_empty_bucket].next_non_empty_bucket = idx; - ht->buckets[idx].prev_non_empty_bucket = ht->last_non_empty_bucket; - ht->buckets[idx].next_non_empty_bucket = -1; - ht->last_non_empty_bucket = idx; - } - ptr = (ht_entry *)gpr_malloc(sizeof(ht_entry)); - ptr->key = key; - ptr->data = data; - ptr->next = ht->buckets[idx].next; - ht->buckets[idx].next = ptr; - ht->size++; -} - -void census_ht_erase(census_ht *ht, census_ht_key key) { - entry_locator loc = ht_find(ht, key); - if (loc.found == 0) { - /* noop if not found */ - return; - } - ht->size--; - if (loc.is_first_in_chain) { - bucket *b = &ht->buckets[loc.bucket_idx]; - GPR_ASSERT(b->next != NULL); - /* The only entry in the bucket */ - if (b->next->next == NULL) { - int prev = b->prev_non_empty_bucket; - int next = b->next_non_empty_bucket; - if (prev != -1) { - ht->buckets[prev].next_non_empty_bucket = next; - } else { - ht->first_non_empty_bucket = next; - } - if (next != -1) { - ht->buckets[next].prev_non_empty_bucket = prev; - } else { - ht->last_non_empty_bucket = prev; - } - } - REMOVE_NEXT(&ht->options, b); - } else { - GPR_ASSERT(loc.prev_entry->next != NULL); - REMOVE_NEXT(&ht->options, loc.prev_entry); - } -} - -/* Returns NULL if input table is empty. */ -census_ht_kv *census_ht_get_all_elements(const census_ht *ht, size_t *num) { - census_ht_kv *ret = NULL; - int i = 0; - int32_t idx = -1; - GPR_ASSERT(ht != NULL && num != NULL); - *num = ht->size; - if (*num == 0) { - return NULL; - } - - ret = (census_ht_kv *)gpr_malloc(sizeof(census_ht_kv) * ht->size); - idx = ht->first_non_empty_bucket; - while (idx >= 0) { - ht_entry *ptr = ht->buckets[idx].next; - for (; ptr != NULL; ptr = ptr->next) { - ret[i].k = ptr->key; - ret[i].v = ptr->data; - i++; - } - idx = ht->buckets[idx].next_non_empty_bucket; - } - return ret; -} - -static void ht_delete_entry_chain(const census_ht_option *options, - ht_entry *first) { - if (first == NULL) { - return; - } - if (first->next != NULL) { - ht_delete_entry_chain(options, first->next); - } - delete_entry(options, first); -} - -void census_ht_destroy(census_ht *ht) { - unsigned i; - for (i = 0; i < ht->num_buckets; ++i) { - ht_delete_entry_chain(&ht->options, ht->buckets[i].next); - } - gpr_free(ht->buckets); - gpr_free(ht); -} - -size_t census_ht_get_size(const census_ht *ht) { return ht->size; } diff --git a/src/core/ext/census/hash_table.h b/src/core/ext/census/hash_table.h deleted file mode 100644 index c3ed94ea14..0000000000 --- a/src/core/ext/census/hash_table.h +++ /dev/null @@ -1,124 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_HASH_TABLE_H -#define GRPC_CORE_EXT_CENSUS_HASH_TABLE_H - -#include <stddef.h> - -#include <grpc/support/port_platform.h> - -#ifdef __cplusplus -extern "C" { -#endif - -/* A chain based hash table with fixed number of buckets. - Your probably shouldn't use this code directly. It is implemented for the - use case in census trace store and stats store, where number of entries in - the table is in the scale of upto several thousands, entries are added and - removed from the table very frequently (~100k/s), the frequency of find() - operations is roughly several times of the frequency of insert() and erase() - Comparing to find(), the insert(), erase() and get_all_entries() operations - are much less freqent (<1/s). - - Per bucket memory overhead is about (8 + sizeof(intptr_t) bytes. - Per entry memory overhead is about (8 + 2 * sizeof(intptr_t) bytes. - - All functions are not thread-safe. Synchronization will be provided in the - upper layer (in trace store and stats store). -*/ - -/* Opaque hash table struct */ -typedef struct unresizable_hash_table census_ht; - -/* Currently, the hash_table can take two types of keys. (uint64 for trace - store and const char* for stats store). */ -typedef union { - uint64_t val; - void *ptr; -} census_ht_key; - -typedef enum census_ht_key_type { - CENSUS_HT_UINT64 = 0, - CENSUS_HT_POINTER = 1 -} census_ht_key_type; - -typedef struct census_ht_option { - /* Type of hash key */ - census_ht_key_type key_type; - /* Desired number of buckets, preferably a prime number */ - int32_t num_buckets; - /* Fucntion to calculate uint64 hash value of the key. Only takes effect if - key_type is POINTER. */ - uint64_t (*hash)(const void *); - /* Function to compare two keys, returns 0 iff equal. Only takes effect if - key_type is POINTER */ - int (*compare_keys)(const void *k1, const void *k2); - /* Value deleter. NULL if no specialized delete function is needed. */ - void (*delete_data)(void *); - /* Key deleter. NULL if table does not own the key. (e.g. key is part of the - value or key is not owned by the table.) */ - void (*delete_key)(void *); -} census_ht_option; - -/* Creates a hashtable with fixed number of buckets according to the settings - specified in 'options' arg. Function pointers "hash" and "compare_keys" must - be provided if key_type is POINTER. Asserts if fail to create. */ -census_ht *census_ht_create(const census_ht_option *options); - -/* Deletes hash table instance. Frees all dynamic memory owned by ht.*/ -void census_ht_destroy(census_ht *ht); - -/* Inserts the input key-val pair into hash_table. If an entry with the same key - exists in the table, the corresponding value will be overwritten by the input - val. */ -void census_ht_insert(census_ht *ht, census_ht_key key, void *val); - -/* Returns pointer to data, returns NULL if not found. */ -void *census_ht_find(const census_ht *ht, census_ht_key key); - -/* Erase hash table entry with input key. Noop if key is not found. */ -void census_ht_erase(census_ht *ht, census_ht_key key); - -typedef struct census_ht_kv { - census_ht_key k; - void *v; -} census_ht_kv; - -/* Returns an array of pointers to all values in the hash table. Order of the - elements can be arbitrary. Sets 'num' to the size of returned array. Caller - owns returned array. */ -census_ht_kv *census_ht_get_all_elements(const census_ht *ht, size_t *num); - -/* Returns number of elements kept. */ -size_t census_ht_get_size(const census_ht *ht); - -/* Functor applied on each key-value pair while iterating through entries in the - table. The functor should not mutate data. */ -typedef void (*census_ht_itr_cb)(census_ht_key key, const void *val_ptr, - void *state); - -/* Iterates through all key-value pairs in the hash_table. The callback function - should not invalidate data entries. */ -uint64_t census_ht_for_all(const census_ht *ht, census_ht_itr_cb); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_HASH_TABLE_H */ diff --git a/src/core/ext/census/initialize.cc b/src/core/ext/census/initialize.cc deleted file mode 100644 index 165a1221d4..0000000000 --- a/src/core/ext/census/initialize.cc +++ /dev/null @@ -1,51 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include <grpc/census.h> -#include "src/core/ext/census/base_resources.h" -#include "src/core/ext/census/resource.h" - -static int features_enabled = CENSUS_FEATURE_NONE; - -int census_initialize(int features) { - if (features_enabled != CENSUS_FEATURE_NONE) { - // Must have been a previous call to census_initialize; return error - return -1; - } - features_enabled = features & CENSUS_FEATURE_ALL; - if (features & CENSUS_FEATURE_STATS) { - initialize_resources(); - define_base_resources(); - } - - return features_enabled; -} - -void census_shutdown(void) { - if (features_enabled & CENSUS_FEATURE_STATS) { - shutdown_resources(); - } - features_enabled = CENSUS_FEATURE_NONE; -} - -int census_supported(void) { - /* TODO(aveitch): improve this as we implement features... */ - return CENSUS_FEATURE_NONE; -} - -int census_enabled(void) { return features_enabled; } diff --git a/src/core/ext/census/intrusive_hash_map.cc b/src/core/ext/census/intrusive_hash_map.cc deleted file mode 100644 index 7930486963..0000000000 --- a/src/core/ext/census/intrusive_hash_map.cc +++ /dev/null @@ -1,305 +0,0 @@ -/* - * - * Copyright 2017 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include "src/core/ext/census/intrusive_hash_map.h" -#include <string.h> - -extern bool hm_index_compare(const hm_index *A, const hm_index *B); - -/* Simple hashing function that takes lower 32 bits. */ -static __inline uint32_t chunked_vector_hasher(uint64_t key) { - return (uint32_t)key; -} - -/* Vector chunks are 1MiB divided by pointer size. */ -static const size_t VECTOR_CHUNK_SIZE = (1 << 20) / sizeof(void *); - -/* Helper functions which return buckets from the chunked vector. */ -static __inline void **get_mutable_bucket(const chunked_vector *buckets, - uint32_t index) { - if (index < VECTOR_CHUNK_SIZE) { - return &buckets->first_[index]; - } - size_t rest_index = (index - VECTOR_CHUNK_SIZE) / VECTOR_CHUNK_SIZE; - return &buckets->rest_[rest_index][index % VECTOR_CHUNK_SIZE]; -} - -static __inline void *get_bucket(const chunked_vector *buckets, - uint32_t index) { - if (index < VECTOR_CHUNK_SIZE) { - return buckets->first_[index]; - } - size_t rest_index = (index - VECTOR_CHUNK_SIZE) / VECTOR_CHUNK_SIZE; - return buckets->rest_[rest_index][index % VECTOR_CHUNK_SIZE]; -} - -/* Helper function. */ -static __inline size_t RestSize(const chunked_vector *vec) { - return (vec->size_ <= VECTOR_CHUNK_SIZE) - ? 0 - : (vec->size_ - VECTOR_CHUNK_SIZE - 1) / VECTOR_CHUNK_SIZE + 1; -} - -/* Initialize chunked vector to size of 0. */ -static void chunked_vector_init(chunked_vector *vec) { - vec->size_ = 0; - vec->first_ = NULL; - vec->rest_ = NULL; -} - -/* Clear chunked vector and free all memory that has been allocated then - initialize chunked vector. */ -static void chunked_vector_clear(chunked_vector *vec) { - if (vec->first_ != NULL) { - gpr_free(vec->first_); - } - if (vec->rest_ != NULL) { - size_t rest_size = RestSize(vec); - for (size_t i = 0; i < rest_size; ++i) { - if (vec->rest_[i] != NULL) { - gpr_free(vec->rest_[i]); - } - } - gpr_free(vec->rest_); - } - chunked_vector_init(vec); -} - -/* Clear chunked vector and then resize it to n entries. Allow the first 1MB to - be read w/o an extra cache miss. The rest of the elements are stored in an - array of arrays to avoid large mallocs. */ -static void chunked_vector_reset(chunked_vector *vec, size_t n) { - chunked_vector_clear(vec); - vec->size_ = n; - if (n <= VECTOR_CHUNK_SIZE) { - vec->first_ = (void **)gpr_malloc(sizeof(void *) * n); - memset(vec->first_, 0, sizeof(void *) * n); - } else { - vec->first_ = (void **)gpr_malloc(sizeof(void *) * VECTOR_CHUNK_SIZE); - memset(vec->first_, 0, sizeof(void *) * VECTOR_CHUNK_SIZE); - size_t rest_size = RestSize(vec); - vec->rest_ = (void ***)gpr_malloc(sizeof(void **) * rest_size); - memset(vec->rest_, 0, sizeof(void **) * rest_size); - int i = 0; - n -= VECTOR_CHUNK_SIZE; - while (n > 0) { - size_t this_size = GPR_MIN(n, VECTOR_CHUNK_SIZE); - vec->rest_[i] = (void **)gpr_malloc(sizeof(void *) * this_size); - memset(vec->rest_[i], 0, sizeof(void *) * this_size); - n -= this_size; - ++i; - } - } -} - -void intrusive_hash_map_init(intrusive_hash_map *hash_map, - uint32_t initial_log2_table_size) { - hash_map->log2_num_buckets = initial_log2_table_size; - hash_map->num_items = 0; - uint32_t num_buckets = (uint32_t)1 << hash_map->log2_num_buckets; - hash_map->extend_threshold = num_buckets >> 1; - chunked_vector_init(&hash_map->buckets); - chunked_vector_reset(&hash_map->buckets, num_buckets); - hash_map->hash_mask = num_buckets - 1; -} - -bool intrusive_hash_map_empty(const intrusive_hash_map *hash_map) { - return hash_map->num_items == 0; -} - -size_t intrusive_hash_map_size(const intrusive_hash_map *hash_map) { - return hash_map->num_items; -} - -void intrusive_hash_map_end(const intrusive_hash_map *hash_map, hm_index *idx) { - idx->bucket_index = (uint32_t)hash_map->buckets.size_; - GPR_ASSERT(idx->bucket_index <= UINT32_MAX); - idx->item = NULL; -} - -void intrusive_hash_map_next(const intrusive_hash_map *hash_map, - hm_index *idx) { - idx->item = idx->item->hash_link; - while (idx->item == NULL) { - idx->bucket_index++; - if (idx->bucket_index >= hash_map->buckets.size_) { - /* Reached end of table. */ - idx->item = NULL; - return; - } - idx->item = (hm_item *)get_bucket(&hash_map->buckets, idx->bucket_index); - } -} - -void intrusive_hash_map_begin(const intrusive_hash_map *hash_map, - hm_index *idx) { - for (uint32_t i = 0; i < hash_map->buckets.size_; ++i) { - if (get_bucket(&hash_map->buckets, i) != NULL) { - idx->bucket_index = i; - idx->item = (hm_item *)get_bucket(&hash_map->buckets, i); - return; - } - } - intrusive_hash_map_end(hash_map, idx); -} - -hm_item *intrusive_hash_map_find(const intrusive_hash_map *hash_map, - uint64_t key) { - uint32_t index = chunked_vector_hasher(key) & hash_map->hash_mask; - - hm_item *p = (hm_item *)get_bucket(&hash_map->buckets, index); - while (p != NULL) { - if (key == p->key) { - return p; - } - p = p->hash_link; - } - return NULL; -} - -hm_item *intrusive_hash_map_erase(intrusive_hash_map *hash_map, uint64_t key) { - uint32_t index = chunked_vector_hasher(key) & hash_map->hash_mask; - - hm_item **slot = (hm_item **)get_mutable_bucket(&hash_map->buckets, index); - hm_item *p = *slot; - if (p == NULL) { - return NULL; - } - - if (key == p->key) { - *slot = p->hash_link; - p->hash_link = NULL; - hash_map->num_items--; - return p; - } - - hm_item *prev = p; - p = p->hash_link; - - while (p) { - if (key == p->key) { - prev->hash_link = p->hash_link; - p->hash_link = NULL; - hash_map->num_items--; - return p; - } - prev = p; - p = p->hash_link; - } - return NULL; -} - -/* Insert an hm_item* into the underlying chunked vector. hash_mask is - * array_size-1. Returns true if it is a new hm_item and false if the hm_item - * already existed. - */ -static __inline bool intrusive_hash_map_internal_insert(chunked_vector *buckets, - uint32_t hash_mask, - hm_item *item) { - const uint64_t key = item->key; - uint32_t index = chunked_vector_hasher(key) & hash_mask; - hm_item **slot = (hm_item **)get_mutable_bucket(buckets, index); - hm_item *p = *slot; - item->hash_link = p; - - /* Check to see if key already exists. */ - while (p) { - if (p->key == key) { - return false; - } - p = p->hash_link; - } - - /* Otherwise add new entry. */ - *slot = item; - return true; -} - -/* Extend the allocated number of elements in the hash map by a factor of 2. */ -void intrusive_hash_map_extend(intrusive_hash_map *hash_map) { - uint32_t new_log2_num_buckets = 1 + hash_map->log2_num_buckets; - uint32_t new_num_buckets = (uint32_t)1 << new_log2_num_buckets; - GPR_ASSERT(new_num_buckets <= UINT32_MAX && new_num_buckets > 0); - chunked_vector new_buckets; - chunked_vector_init(&new_buckets); - chunked_vector_reset(&new_buckets, new_num_buckets); - uint32_t new_hash_mask = new_num_buckets - 1; - - hm_index cur_idx; - hm_index end_idx; - intrusive_hash_map_end(hash_map, &end_idx); - intrusive_hash_map_begin(hash_map, &cur_idx); - while (!hm_index_compare(&cur_idx, &end_idx)) { - hm_item *new_item = cur_idx.item; - intrusive_hash_map_next(hash_map, &cur_idx); - intrusive_hash_map_internal_insert(&new_buckets, new_hash_mask, new_item); - } - - /* Set values for new chunked_vector. extend_threshold is set to half of - * new_num_buckets. */ - hash_map->log2_num_buckets = new_log2_num_buckets; - chunked_vector_clear(&hash_map->buckets); - hash_map->buckets = new_buckets; - hash_map->hash_mask = new_hash_mask; - hash_map->extend_threshold = new_num_buckets >> 1; -} - -/* Insert a hm_item. The hm_item must remain live until it is removed from the - table. This object does not take the ownership of hm_item. The caller must - remove this hm_item from the table and delete it before this table is - deleted. If hm_item exists already num_items is not changed. */ -bool intrusive_hash_map_insert(intrusive_hash_map *hash_map, hm_item *item) { - if (hash_map->num_items >= hash_map->extend_threshold) { - intrusive_hash_map_extend(hash_map); - } - if (intrusive_hash_map_internal_insert(&hash_map->buckets, - hash_map->hash_mask, item)) { - hash_map->num_items++; - return true; - } - return false; -} - -void intrusive_hash_map_clear(intrusive_hash_map *hash_map, - void (*free_object)(void *)) { - hm_index cur; - hm_index end; - intrusive_hash_map_end(hash_map, &end); - intrusive_hash_map_begin(hash_map, &cur); - - while (!hm_index_compare(&cur, &end)) { - hm_index next = cur; - intrusive_hash_map_next(hash_map, &next); - if (cur.item != NULL) { - hm_item *item = intrusive_hash_map_erase(hash_map, cur.item->key); - (*free_object)((void *)item); - gpr_free(item); - } - cur = next; - } -} - -void intrusive_hash_map_free(intrusive_hash_map *hash_map, - void (*free_object)(void *)) { - intrusive_hash_map_clear(hash_map, (*free_object)); - hash_map->num_items = 0; - hash_map->extend_threshold = 0; - hash_map->log2_num_buckets = 0; - hash_map->hash_mask = 0; - chunked_vector_clear(&hash_map->buckets); -} diff --git a/src/core/ext/census/intrusive_hash_map.h b/src/core/ext/census/intrusive_hash_map.h deleted file mode 100644 index 2c7baa31fb..0000000000 --- a/src/core/ext/census/intrusive_hash_map.h +++ /dev/null @@ -1,160 +0,0 @@ -/* - * - * Copyright 2017 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_INTRUSIVE_HASH_MAP_H -#define GRPC_CORE_EXT_CENSUS_INTRUSIVE_HASH_MAP_H - -#include "src/core/ext/census/intrusive_hash_map_internal.h" - -#ifdef __cplusplus -extern "C" { -#endif - -/* intrusive_hash_map is a fast chained hash table. This hash map is faster than - * a dense hash map when the application calls insert and erase more often than - * find. When the workload is dominated by find() a dense hash map may be - * faster. - * - * intrusive_hash_map uses an intrusive header placed within a user defined - * struct. The header field IHM_key MUST be set to a valid value before - * insertion into the hash map or undefined behavior may occur. The header field - * IHM_hash_link MUST to be set to NULL initially. - * - * EXAMPLE USAGE: - * - * typedef struct string_item { - * INTRUSIVE_HASH_MAP_HEADER; - * // User data. - * char *str_buf; - * uint16_t len; - * } string_item; - * - * static string_item *make_string_item(uint64_t key, const char *buf, - * uint16_t len) { - * string_item *item = (string_item *)gpr_malloc(sizeof(string_item)); - * item->IHM_key = key; - * item->IHM_hash_link = NULL; - * item->len = len; - * item->str_buf = (char *)malloc(len); - * memcpy(item->str_buf, buf, len); - * return item; - * } - * - * intrusive_hash_map hash_map; - * intrusive_hash_map_init(&hash_map, 4); - * string_item *new_item1 = make_string_item(10, "test1", 5); - * bool ok = intrusive_hash_map_insert(&hash_map, (hm_item *)new_item1); - * - * string_item *item1 = - * (string_item *)intrusive_hash_map_find(&hash_map, 10); - */ - -/* Hash map item. Stores key and a pointer to the actual object. A user defined - * version of this can be passed in provided the first 2 entries (key and - * hash_link) are the same. These entries must be first in the user defined - * struct. Pointer to struct will need to be cast as (hm_item *) when passed to - * hash map. This allows it to be intrusive. */ -typedef struct hm_item { - uint64_t key; - struct hm_item *hash_link; - /* Optional user defined data after this. */ -} hm_item; - -/* Macro provided for ease of use. This must be first in the user defined - * struct (i.e. uint64_t key and hm_item * must be the first two elements in - * that order). */ -#define INTRUSIVE_HASH_MAP_HEADER \ - uint64_t IHM_key; \ - struct hm_item *IHM_hash_link - -/* Index struct which acts as a pseudo-iterator within the hash map. */ -typedef struct hm_index { - uint32_t bucket_index; // hash map bucket index. - hm_item *item; // Pointer to hm_item within the hash map. -} hm_index; - -/* Returns true if two hm_indices point to the same object within the hash map - * and false otherwise. */ -__inline bool hm_index_compare(const hm_index *A, const hm_index *B) { - return (A->item == B->item && A->bucket_index == B->bucket_index); -} - -/* - * Helper functions for iterating over the hash map. - */ - -/* On return idx will contain an invalid index which is always equal to - * hash_map->buckets.size_ */ -void intrusive_hash_map_end(const intrusive_hash_map *hash_map, hm_index *idx); - -/* Iterates index to the next valid entry in the hash map and stores the - * index within idx. If end of table is reached, idx will contain the same - * values as if intrusive_hash_map_end() was called. */ -void intrusive_hash_map_next(const intrusive_hash_map *hash_map, hm_index *idx); - -/* On return, idx will contain the index of the first non-null entry in the hash - * map. If the hash map is empty, idx will contain the same values as if - * intrusive_hash_map_end() was called. */ -void intrusive_hash_map_begin(const intrusive_hash_map *hash_map, - hm_index *idx); - -/* Initialize intrusive hash map data structure. This must be called before - * the hash map can be used. The initial size of an intrusive hash map will be - * 2^initial_log2_map_size (valid range is [0, 31]). */ -void intrusive_hash_map_init(intrusive_hash_map *hash_map, - uint32_t initial_log2_map_size); - -/* Returns true if the hash map is empty and false otherwise. */ -bool intrusive_hash_map_empty(const intrusive_hash_map *hash_map); - -/* Returns the number of elements currently in the hash map. */ -size_t intrusive_hash_map_size(const intrusive_hash_map *hash_map); - -/* Find a hm_item within the hash map by key. Returns NULL if item was not - * found. */ -hm_item *intrusive_hash_map_find(const intrusive_hash_map *hash_map, - uint64_t key); - -/* Erase the hm_item that corresponds with key. If the hm_item is found, return - * the pointer to the hm_item. Else returns NULL. */ -hm_item *intrusive_hash_map_erase(intrusive_hash_map *hash_map, uint64_t key); - -/* Attempts to insert a new hm_item into the hash map. If an element with the - * same key already exists, it will not insert the new item and return false. - * Otherwise, it will insert the new item and return true. */ -bool intrusive_hash_map_insert(intrusive_hash_map *hash_map, hm_item *item); - -/* Clears entire contents of the hash map, but leaves internal data structure - * untouched. Second argument takes a function pointer to a function that will - * free the object designated by the user and pointed to by hash_map->value. */ -void intrusive_hash_map_clear(intrusive_hash_map *hash_map, - void (*free_object)(void *)); - -/* Erase all contents of hash map and free the memory. Hash map is invalid - * after calling this function and cannot be used until it has been - * reinitialized (intrusive_hash_map_init()). This function takes a function - * pointer to a function that will free the object designated by the user and - * pointed to by hash_map->value. */ -void intrusive_hash_map_free(intrusive_hash_map *hash_map, - void (*free_object)(void *)); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_INTRUSIVE_HASH_MAP_H */ diff --git a/src/core/ext/census/intrusive_hash_map_internal.h b/src/core/ext/census/intrusive_hash_map_internal.h deleted file mode 100644 index e9c81fc85c..0000000000 --- a/src/core/ext/census/intrusive_hash_map_internal.h +++ /dev/null @@ -1,48 +0,0 @@ -/* - * - * Copyright 2017 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_INTRUSIVE_HASH_MAP_INTERNAL_H -#define GRPC_CORE_EXT_CENSUS_INTRUSIVE_HASH_MAP_INTERNAL_H - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/useful.h> -#include <stdbool.h> - -/* The chunked vector is a data structure that allocates buckets for use in the - * hash map. ChunkedVector is logically equivalent to T*[N] (cast void* as - * T*). It's internally implemented as an array of 1MB arrays to avoid - * allocating large consecutive memory chunks. This is an internal data - * structure that should never be accessed directly. */ -typedef struct chunked_vector { - size_t size_; - void **first_; - void ***rest_; -} chunked_vector; - -/* Core intrusive hash map data structure. All internal elements are managed by - * functions and should not be altered manually. */ -typedef struct intrusive_hash_map { - uint32_t num_items; - uint32_t extend_threshold; - uint32_t log2_num_buckets; - uint32_t hash_mask; - chunked_vector buckets; -} intrusive_hash_map; - -#endif /* GRPC_CORE_EXT_CENSUS_INTRUSIVE_HASH_MAP_INTERNAL_H */ diff --git a/src/core/ext/census/mlog.cc b/src/core/ext/census/mlog.cc deleted file mode 100644 index 4b8c8466b3..0000000000 --- a/src/core/ext/census/mlog.cc +++ /dev/null @@ -1,586 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -// Implements an efficient in-memory log, optimized for multiple writers and -// a single reader. Available log space is divided up in blocks of -// CENSUS_LOG_2_MAX_RECORD_SIZE bytes. A block can be in one of the following -// three data structures: -// - Free blocks (free_block_list) -// - Blocks with unread data (dirty_block_list) -// - Blocks currently attached to cores (core_local_blocks[]) -// -// census_log_start_write() moves a block from core_local_blocks[] to the end of -// dirty_block_list when block: -// - is out-of-space OR -// - has an incomplete record (an incomplete record occurs when a thread calls -// census_log_start_write() and is context-switched before calling -// census_log_end_write() -// So, blocks in dirty_block_list are ordered, from oldest to newest, by the -// time when block is detached from the core. -// -// census_log_read_next() first iterates over dirty_block_list and then -// core_local_blocks[]. It moves completely read blocks from dirty_block_list -// to free_block_list. Blocks in core_local_blocks[] are not freed, even when -// completely read. -// -// If the log is configured to discard old records and free_block_list is empty, -// census_log_start_write() iterates over dirty_block_list to allocate a -// new block. It moves the oldest available block (no pending read/write) to -// core_local_blocks[]. -// -// core_local_block_struct is used to implement a map from core id to the block -// associated with that core. This mapping is advisory. It is possible that the -// block returned by this mapping is no longer associated with that core. This -// mapping is updated, lazily, by census_log_start_write(). -// -// Locking in block struct: -// -// Exclusive g_log.lock must be held before calling any functions operating on -// block structs except census_log_start_write() and census_log_end_write(). -// -// Writes to a block are serialized via writer_lock. census_log_start_write() -// acquires this lock and census_log_end_write() releases it. On failure to -// acquire the lock, writer allocates a new block for the current core and -// updates core_local_block accordingly. -// -// Simultaneous read and write access is allowed. Readers can safely read up to -// committed bytes (bytes_committed). -// -// reader_lock protects the block, currently being read, from getting recycled. -// start_read() acquires reader_lock and end_read() releases the lock. -// -// Read/write access to a block is disabled via try_disable_access(). It returns -// with both writer_lock and reader_lock held. These locks are subsequently -// released by enable_access() to enable access to the block. -// -// A note on naming: Most function/struct names are prepended by cl_ -// (shorthand for census_log). Further, functions that manipulate structures -// include the name of the structure, which will be passed as the first -// argument. E.g. cl_block_initialize() will initialize a cl_block. - -#include "src/core/ext/census/mlog.h" -#include <grpc/support/alloc.h> -#include <grpc/support/atm.h> -#include <grpc/support/cpu.h> -#include <grpc/support/log.h> -#include <grpc/support/sync.h> -#include <grpc/support/useful.h> -#include <stdbool.h> -#include <string.h> - -// End of platform specific code - -typedef struct census_log_block_list_struct { - struct census_log_block_list_struct* next; - struct census_log_block_list_struct* prev; - struct census_log_block* block; -} cl_block_list_struct; - -typedef struct census_log_block { - // Pointer to underlying buffer. - char* buffer; - gpr_atm writer_lock; - gpr_atm reader_lock; - // Keeps completely written bytes. Declared atomic because accessed - // simultaneously by reader and writer. - gpr_atm bytes_committed; - // Bytes already read. - size_t bytes_read; - // Links for list. - cl_block_list_struct link; -// We want this structure to be cacheline aligned. We assume the following -// sizes for the various parts on 32/64bit systems: -// type 32b size 64b size -// char* 4 8 -// 3x gpr_atm 12 24 -// size_t 4 8 -// cl_block_list_struct 12 24 -// TOTAL 32 64 -// -// Depending on the size of our cacheline and the architecture, we -// selectively add char buffering to this structure. The size is checked -// via assert in census_log_initialize(). -#if defined(GPR_ARCH_64) -#define CL_BLOCK_PAD_SIZE (GPR_CACHELINE_SIZE - 64) -#else -#if defined(GPR_ARCH_32) -#define CL_BLOCK_PAD_SIZE (GPR_CACHELINE_SIZE - 32) -#else -#error "Unknown architecture" -#endif -#endif -#if CL_BLOCK_PAD_SIZE > 0 - char padding[CL_BLOCK_PAD_SIZE]; -#endif -} cl_block; - -// A list of cl_blocks, doubly-linked through cl_block::link. -typedef struct census_log_block_list { - int32_t count; // Number of items in list. - cl_block_list_struct ht; // head/tail of linked list. -} cl_block_list; - -// Cacheline aligned block pointers to avoid false sharing. Block pointer must -// be initialized via set_block(), before calling other functions -typedef struct census_log_core_local_block { - gpr_atm block; -// Ensure cachline alignment: we assume sizeof(gpr_atm) == 4 or 8 -#if defined(GPR_ARCH_64) -#define CL_CORE_LOCAL_BLOCK_PAD_SIZE (GPR_CACHELINE_SIZE - 8) -#else -#if defined(GPR_ARCH_32) -#define CL_CORE_LOCAL_BLOCK_PAD_SIZE (GPR_CACHELINE_SIZE - 4) -#else -#error "Unknown architecture" -#endif -#endif -#if CL_CORE_LOCAL_BLOCK_PAD_SIZE > 0 - char padding[CL_CORE_LOCAL_BLOCK_PAD_SIZE]; -#endif -} cl_core_local_block; - -struct census_log { - int discard_old_records; - // Number of cores (aka hardware-contexts) - unsigned num_cores; - // number of CENSUS_LOG_2_MAX_RECORD_SIZE blocks in log - uint32_t num_blocks; - cl_block* blocks; // Block metadata. - cl_core_local_block* core_local_blocks; // Keeps core to block mappings. - gpr_mu lock; - int initialized; // has log been initialized? - // Keeps the state of the reader iterator. A value of 0 indicates that - // iterator has reached the end. census_log_init_reader() resets the value - // to num_core to restart iteration. - uint32_t read_iterator_state; - // Points to the block being read. If non-NULL, the block is locked for - // reading(block_being_read_->reader_lock is held). - cl_block* block_being_read; - char* buffer; - cl_block_list free_block_list; - cl_block_list dirty_block_list; - gpr_atm out_of_space_count; -}; - -// Single internal log. -static struct census_log g_log; - -// Functions that operate on an atomic memory location used as a lock. - -// Returns non-zero if lock is acquired. -static int cl_try_lock(gpr_atm* lock) { return gpr_atm_acq_cas(lock, 0, 1); } - -static void cl_unlock(gpr_atm* lock) { gpr_atm_rel_store(lock, 0); } - -// Functions that operate on cl_core_local_block's. - -static void cl_core_local_block_set_block(cl_core_local_block* clb, - cl_block* block) { - gpr_atm_rel_store(&clb->block, (gpr_atm)block); -} - -static cl_block* cl_core_local_block_get_block(cl_core_local_block* clb) { - return (cl_block*)gpr_atm_acq_load(&clb->block); -} - -// Functions that operate on cl_block_list_struct's. - -static void cl_block_list_struct_initialize(cl_block_list_struct* bls, - cl_block* block) { - bls->next = bls->prev = bls; - bls->block = block; -} - -// Functions that operate on cl_block_list's. - -static void cl_block_list_initialize(cl_block_list* list) { - list->count = 0; - cl_block_list_struct_initialize(&list->ht, NULL); -} - -// Returns head of *this, or NULL if empty. -static cl_block* cl_block_list_head(cl_block_list* list) { - return list->ht.next->block; -} - -// Insert element *e after *pos. -static void cl_block_list_insert(cl_block_list* list, cl_block_list_struct* pos, - cl_block_list_struct* e) { - list->count++; - e->next = pos->next; - e->prev = pos; - e->next->prev = e; - e->prev->next = e; -} - -// Insert block at the head of the list -static void cl_block_list_insert_at_head(cl_block_list* list, cl_block* block) { - cl_block_list_insert(list, &list->ht, &block->link); -} - -// Insert block at the tail of the list. -static void cl_block_list_insert_at_tail(cl_block_list* list, cl_block* block) { - cl_block_list_insert(list, list->ht.prev, &block->link); -} - -// Removes block *b. Requires *b be in the list. -static void cl_block_list_remove(cl_block_list* list, cl_block* b) { - list->count--; - b->link.next->prev = b->link.prev; - b->link.prev->next = b->link.next; -} - -// Functions that operate on cl_block's - -static void cl_block_initialize(cl_block* block, char* buffer) { - block->buffer = buffer; - gpr_atm_rel_store(&block->writer_lock, 0); - gpr_atm_rel_store(&block->reader_lock, 0); - gpr_atm_rel_store(&block->bytes_committed, 0); - block->bytes_read = 0; - cl_block_list_struct_initialize(&block->link, block); -} - -// Guards against exposing partially written buffer to the reader. -static void cl_block_set_bytes_committed(cl_block* block, - size_t bytes_committed) { - gpr_atm_rel_store(&block->bytes_committed, (gpr_atm)bytes_committed); -} - -static size_t cl_block_get_bytes_committed(cl_block* block) { - return (size_t)gpr_atm_acq_load(&block->bytes_committed); -} - -// Tries to disable future read/write access to this block. Succeeds if: -// - no in-progress write AND -// - no in-progress read AND -// - 'discard_data' set to true OR no unread data -// On success, clears the block state and returns with writer_lock_ and -// reader_lock_ held. These locks are released by a subsequent -// cl_block_access_enable() call. -static bool cl_block_try_disable_access(cl_block* block, int discard_data) { - if (!cl_try_lock(&block->writer_lock)) { - return false; - } - if (!cl_try_lock(&block->reader_lock)) { - cl_unlock(&block->writer_lock); - return false; - } - if (!discard_data && - (block->bytes_read != cl_block_get_bytes_committed(block))) { - cl_unlock(&block->reader_lock); - cl_unlock(&block->writer_lock); - return false; - } - cl_block_set_bytes_committed(block, 0); - block->bytes_read = 0; - return true; -} - -static void cl_block_enable_access(cl_block* block) { - cl_unlock(&block->reader_lock); - cl_unlock(&block->writer_lock); -} - -// Returns with writer_lock held. -static void* cl_block_start_write(cl_block* block, size_t size) { - if (!cl_try_lock(&block->writer_lock)) { - return NULL; - } - size_t bytes_committed = cl_block_get_bytes_committed(block); - if (bytes_committed + size > CENSUS_LOG_MAX_RECORD_SIZE) { - cl_unlock(&block->writer_lock); - return NULL; - } - return block->buffer + bytes_committed; -} - -// Releases writer_lock and increments committed bytes by 'bytes_written'. -// 'bytes_written' must be <= 'size' specified in the corresponding -// StartWrite() call. This function is thread-safe. -static void cl_block_end_write(cl_block* block, size_t bytes_written) { - cl_block_set_bytes_committed( - block, cl_block_get_bytes_committed(block) + bytes_written); - cl_unlock(&block->writer_lock); -} - -// Returns a pointer to the first unread byte in buffer. The number of bytes -// available are returned in 'bytes_available'. Acquires reader lock that is -// released by a subsequent cl_block_end_read() call. Returns NULL if: -// - read in progress -// - no data available -static void* cl_block_start_read(cl_block* block, size_t* bytes_available) { - if (!cl_try_lock(&block->reader_lock)) { - return NULL; - } - // bytes_committed may change from under us. Use bytes_available to update - // bytes_read below. - size_t bytes_committed = cl_block_get_bytes_committed(block); - GPR_ASSERT(bytes_committed >= block->bytes_read); - *bytes_available = bytes_committed - block->bytes_read; - if (*bytes_available == 0) { - cl_unlock(&block->reader_lock); - return NULL; - } - void* record = block->buffer + block->bytes_read; - block->bytes_read += *bytes_available; - return record; -} - -static void cl_block_end_read(cl_block* block) { - cl_unlock(&block->reader_lock); -} - -// Internal functions operating on g_log - -// Allocates a new free block (or recycles an available dirty block if log is -// configured to discard old records). Returns NULL if out-of-space. -static cl_block* cl_allocate_block(void) { - cl_block* block = cl_block_list_head(&g_log.free_block_list); - if (block != NULL) { - cl_block_list_remove(&g_log.free_block_list, block); - return block; - } - if (!g_log.discard_old_records) { - // No free block and log is configured to keep old records. - return NULL; - } - // Recycle dirty block. Start from the oldest. - for (block = cl_block_list_head(&g_log.dirty_block_list); block != NULL; - block = block->link.next->block) { - if (cl_block_try_disable_access(block, 1 /* discard data */)) { - cl_block_list_remove(&g_log.dirty_block_list, block); - return block; - } - } - return NULL; -} - -// Allocates a new block and updates core id => block mapping. 'old_block' -// points to the block that the caller thinks is attached to -// 'core_id'. 'old_block' may be NULL. Returns true if: -// - allocated a new block OR -// - 'core_id' => 'old_block' mapping changed (another thread allocated a -// block before lock was acquired). -static bool cl_allocate_core_local_block(uint32_t core_id, - cl_block* old_block) { - // Now that we have the lock, check if core-local mapping has changed. - cl_core_local_block* core_local_block = &g_log.core_local_blocks[core_id]; - cl_block* block = cl_core_local_block_get_block(core_local_block); - if ((block != NULL) && (block != old_block)) { - return true; - } - if (block != NULL) { - cl_core_local_block_set_block(core_local_block, NULL); - cl_block_list_insert_at_tail(&g_log.dirty_block_list, block); - } - block = cl_allocate_block(); - if (block == NULL) { - return false; - } - cl_core_local_block_set_block(core_local_block, block); - cl_block_enable_access(block); - return true; -} - -static cl_block* cl_get_block(void* record) { - uintptr_t p = (uintptr_t)((char*)record - g_log.buffer); - uintptr_t index = p >> CENSUS_LOG_2_MAX_RECORD_SIZE; - return &g_log.blocks[index]; -} - -// Gets the next block to read and tries to free 'prev' block (if not NULL). -// Returns NULL if reached the end. -static cl_block* cl_next_block_to_read(cl_block* prev) { - cl_block* block = NULL; - if (g_log.read_iterator_state == g_log.num_cores) { - // We are traversing dirty list; find the next dirty block. - if (prev != NULL) { - // Try to free the previous block if there is no unread data. This - // block - // may have unread data if previously incomplete record completed - // between - // read_next() calls. - block = prev->link.next->block; - if (cl_block_try_disable_access(prev, 0 /* do not discard data */)) { - cl_block_list_remove(&g_log.dirty_block_list, prev); - cl_block_list_insert_at_head(&g_log.free_block_list, prev); - } - } else { - block = cl_block_list_head(&g_log.dirty_block_list); - } - if (block != NULL) { - return block; - } - // We are done with the dirty list; moving on to core-local blocks. - } - while (g_log.read_iterator_state > 0) { - g_log.read_iterator_state--; - block = cl_core_local_block_get_block( - &g_log.core_local_blocks[g_log.read_iterator_state]); - if (block != NULL) { - return block; - } - } - return NULL; -} - -#define CL_LOG_2_MB 20 // 2^20 = 1MB - -// External functions: primary stats_log interface -void census_log_initialize(size_t size_in_mb, int discard_old_records) { - // Check cacheline alignment. - GPR_ASSERT(sizeof(cl_block) % GPR_CACHELINE_SIZE == 0); - GPR_ASSERT(sizeof(cl_core_local_block) % GPR_CACHELINE_SIZE == 0); - GPR_ASSERT(!g_log.initialized); - g_log.discard_old_records = discard_old_records; - g_log.num_cores = gpr_cpu_num_cores(); - // Ensure that we will not get any overflow in calaculating num_blocks - GPR_ASSERT(CL_LOG_2_MB >= CENSUS_LOG_2_MAX_RECORD_SIZE); - GPR_ASSERT(size_in_mb < 1000); - // Ensure at least 2x as many blocks as there are cores. - g_log.num_blocks = - (uint32_t)GPR_MAX(2 * g_log.num_cores, (size_in_mb << CL_LOG_2_MB) >> - CENSUS_LOG_2_MAX_RECORD_SIZE); - gpr_mu_init(&g_log.lock); - g_log.read_iterator_state = 0; - g_log.block_being_read = NULL; - g_log.core_local_blocks = (cl_core_local_block*)gpr_malloc_aligned( - g_log.num_cores * sizeof(cl_core_local_block), GPR_CACHELINE_SIZE_LOG); - memset(g_log.core_local_blocks, 0, - g_log.num_cores * sizeof(cl_core_local_block)); - g_log.blocks = (cl_block*)gpr_malloc_aligned( - g_log.num_blocks * sizeof(cl_block), GPR_CACHELINE_SIZE_LOG); - memset(g_log.blocks, 0, g_log.num_blocks * sizeof(cl_block)); - g_log.buffer = - (char*)gpr_malloc(g_log.num_blocks * CENSUS_LOG_MAX_RECORD_SIZE); - memset(g_log.buffer, 0, g_log.num_blocks * CENSUS_LOG_MAX_RECORD_SIZE); - cl_block_list_initialize(&g_log.free_block_list); - cl_block_list_initialize(&g_log.dirty_block_list); - for (uint32_t i = 0; i < g_log.num_blocks; ++i) { - cl_block* block = g_log.blocks + i; - cl_block_initialize(block, g_log.buffer + (CENSUS_LOG_MAX_RECORD_SIZE * i)); - cl_block_try_disable_access(block, 1 /* discard data */); - cl_block_list_insert_at_tail(&g_log.free_block_list, block); - } - gpr_atm_rel_store(&g_log.out_of_space_count, 0); - g_log.initialized = 1; -} - -void census_log_shutdown(void) { - GPR_ASSERT(g_log.initialized); - gpr_mu_destroy(&g_log.lock); - gpr_free_aligned(g_log.core_local_blocks); - g_log.core_local_blocks = NULL; - gpr_free_aligned(g_log.blocks); - g_log.blocks = NULL; - gpr_free(g_log.buffer); - g_log.buffer = NULL; - g_log.initialized = 0; -} - -void* census_log_start_write(size_t size) { - // Used to bound number of times block allocation is attempted. - GPR_ASSERT(size > 0); - GPR_ASSERT(g_log.initialized); - if (size > CENSUS_LOG_MAX_RECORD_SIZE) { - return NULL; - } - uint32_t attempts_remaining = g_log.num_blocks; - uint32_t core_id = gpr_cpu_current_cpu(); - do { - void* record = NULL; - cl_block* block = - cl_core_local_block_get_block(&g_log.core_local_blocks[core_id]); - if (block && (record = cl_block_start_write(block, size))) { - return record; - } - // Need to allocate a new block. We are here if: - // - No block associated with the core OR - // - Write in-progress on the block OR - // - block is out of space - gpr_mu_lock(&g_log.lock); - bool allocated = cl_allocate_core_local_block(core_id, block); - gpr_mu_unlock(&g_log.lock); - if (!allocated) { - gpr_atm_no_barrier_fetch_add(&g_log.out_of_space_count, 1); - return NULL; - } - } while (attempts_remaining--); - // Give up. - gpr_atm_no_barrier_fetch_add(&g_log.out_of_space_count, 1); - return NULL; -} - -void census_log_end_write(void* record, size_t bytes_written) { - GPR_ASSERT(g_log.initialized); - cl_block_end_write(cl_get_block(record), bytes_written); -} - -void census_log_init_reader(void) { - GPR_ASSERT(g_log.initialized); - gpr_mu_lock(&g_log.lock); - // If a block is locked for reading unlock it. - if (g_log.block_being_read != NULL) { - cl_block_end_read(g_log.block_being_read); - g_log.block_being_read = NULL; - } - g_log.read_iterator_state = g_log.num_cores; - gpr_mu_unlock(&g_log.lock); -} - -const void* census_log_read_next(size_t* bytes_available) { - GPR_ASSERT(g_log.initialized); - gpr_mu_lock(&g_log.lock); - if (g_log.block_being_read != NULL) { - cl_block_end_read(g_log.block_being_read); - } - do { - g_log.block_being_read = cl_next_block_to_read(g_log.block_being_read); - if (g_log.block_being_read != NULL) { - void* record = - cl_block_start_read(g_log.block_being_read, bytes_available); - if (record != NULL) { - gpr_mu_unlock(&g_log.lock); - return record; - } - } - } while (g_log.block_being_read != NULL); - gpr_mu_unlock(&g_log.lock); - return NULL; -} - -size_t census_log_remaining_space(void) { - GPR_ASSERT(g_log.initialized); - size_t space = 0; - gpr_mu_lock(&g_log.lock); - if (g_log.discard_old_records) { - // Remaining space is not meaningful; just return the entire log space. - space = g_log.num_blocks << CENSUS_LOG_2_MAX_RECORD_SIZE; - } else { - GPR_ASSERT(g_log.free_block_list.count >= 0); - space = (size_t)g_log.free_block_list.count * CENSUS_LOG_MAX_RECORD_SIZE; - } - gpr_mu_unlock(&g_log.lock); - return space; -} - -int64_t census_log_out_of_space_count(void) { - GPR_ASSERT(g_log.initialized); - return gpr_atm_acq_load(&g_log.out_of_space_count); -} diff --git a/src/core/ext/census/mlog.h b/src/core/ext/census/mlog.h deleted file mode 100644 index 8f74ba231d..0000000000 --- a/src/core/ext/census/mlog.h +++ /dev/null @@ -1,88 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -/* A very fast in-memory log, optimized for multiple writers. */ - -#ifndef GRPC_CORE_EXT_CENSUS_MLOG_H -#define GRPC_CORE_EXT_CENSUS_MLOG_H - -#include <grpc/support/port_platform.h> -#include <stddef.h> - -/* Maximum record size, in bytes. */ -#define CENSUS_LOG_2_MAX_RECORD_SIZE 14 /* 2^14 = 16KB */ -#define CENSUS_LOG_MAX_RECORD_SIZE (1 << CENSUS_LOG_2_MAX_RECORD_SIZE) - -#ifdef __cplusplus -extern "C" { -#endif - -/* Initialize the statistics logging subsystem with the given log size. A log - size of 0 will result in the smallest possible log for the platform - (approximately CENSUS_LOG_MAX_RECORD_SIZE * gpr_cpu_num_cores()). If - discard_old_records is non-zero, then new records will displace older ones - when the log is full. This function must be called before any other - census_log functions. -*/ -void census_log_initialize(size_t size_in_mb, int discard_old_records); - -/* Shutdown the logging subsystem. Caller must ensure that: - - no in progress or future call to any census_log functions - - no incomplete records -*/ -void census_log_shutdown(void); - -/* Allocates and returns a 'size' bytes record and marks it in use. A - subsequent census_log_end_write() marks the record complete. The - 'bytes_written' census_log_end_write() argument must be <= - 'size'. Returns NULL if out-of-space AND: - - log is configured to keep old records OR - - all blocks are pinned by incomplete records. -*/ -void* census_log_start_write(size_t size); - -void census_log_end_write(void* record, size_t bytes_written); - -void census_log_init_reader(void); - -/* census_log_read_next() iterates over blocks with data and for each block - returns a pointer to the first unread byte. The number of bytes that can be - read are returned in 'bytes_available'. Reader is expected to read all - available data. Reading the data consumes it i.e. it cannot be read again. - census_log_read_next() returns NULL if the end is reached i.e last block - is read. census_log_init_reader() starts the iteration or aborts the - current iteration. -*/ -const void* census_log_read_next(size_t* bytes_available); - -/* Returns estimated remaining space across all blocks, in bytes. If log is - configured to discard old records, returns total log space. Otherwise, - returns space available in empty blocks (partially filled blocks are - treated as full). -*/ -size_t census_log_remaining_space(void); - -/* Returns the number of times grpc_stats_log_start_write() failed due to - out-of-space. */ -int64_t census_log_out_of_space_count(void); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_MLOG_H */ diff --git a/src/core/ext/census/operation.cc b/src/core/ext/census/operation.cc deleted file mode 100644 index be88ac74e6..0000000000 --- a/src/core/ext/census/operation.cc +++ /dev/null @@ -1,48 +0,0 @@ -/* - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include <grpc/census.h> - -/* TODO(aveitch): These are all placeholder implementations. */ - -census_timestamp census_start_rpc_op_timestamp(void) { - census_timestamp ct; - /* TODO(aveitch): assumes gpr_timespec implementation of census_timestamp. */ - ct.ts = gpr_now(GPR_CLOCK_MONOTONIC); - return ct; -} - -census_context *census_start_client_rpc_op( - const census_context *context, int64_t rpc_name_id, - const census_rpc_name_info *rpc_name_info, const char *peer, int trace_mask, - const census_timestamp *start_time) { - return NULL; -} - -census_context *census_start_server_rpc_op( - const char *buffer, int64_t rpc_name_id, - const census_rpc_name_info *rpc_name_info, const char *peer, int trace_mask, - census_timestamp *start_time) { - return NULL; -} - -census_context *census_start_op(census_context *context, const char *family, - const char *name, int trace_mask) { - return NULL; -} - -void census_end_op(census_context *context, int status) {} diff --git a/src/core/ext/census/placeholders.cc b/src/core/ext/census/placeholders.cc deleted file mode 100644 index bed9837ee3..0000000000 --- a/src/core/ext/census/placeholders.cc +++ /dev/null @@ -1,49 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include <grpc/census.h> - -#include <grpc/support/log.h> - -/* Placeholders for the pending APIs */ - -int census_get_trace_record(census_trace_record *trace_record) { - (void)trace_record; - abort(); -} - -void census_record_values(census_context *context, census_value *values, - size_t nvalues) { - (void)context; - (void)values; - (void)nvalues; - abort(); -} - -void census_set_rpc_client_peer(census_context *context, const char *peer) { - (void)context; - (void)peer; - abort(); -} - -void census_trace_scan_end() { abort(); } - -int census_trace_scan_start(int consume) { - (void)consume; - abort(); -} diff --git a/src/core/ext/census/resource.cc b/src/core/ext/census/resource.cc deleted file mode 100644 index 44a887231c..0000000000 --- a/src/core/ext/census/resource.cc +++ /dev/null @@ -1,303 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include "src/core/ext/census/resource.h" -#include "third_party/nanopb/pb_decode.h" - -#include <grpc/census.h> -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/sync.h> - -#include <stdbool.h> -#include <string.h> - -// Protect local resource data structures. -static gpr_mu resource_lock; - -// Deleteing and creating resources are relatively rare events, and should not -// be done in the critical path of performance sensitive code. We record -// current resource id's used in a simple array, and just search it each time -// we need to assign a new id, or look up a resource. -static resource **resources = NULL; - -// Number of entries in *resources -static size_t n_resources = 0; - -// Number of defined resources -static size_t n_defined_resources = 0; - -void initialize_resources(void) { - gpr_mu_init(&resource_lock); - gpr_mu_lock(&resource_lock); - GPR_ASSERT(resources == NULL && n_resources == 0 && n_defined_resources == 0); - gpr_mu_unlock(&resource_lock); -} - -// Delete a resource given it's ID. The ID must be a valid resource ID. Must be -// called with resource_lock held. -static void delete_resource_locked(size_t rid) { - GPR_ASSERT(resources[rid] != NULL); - gpr_free(resources[rid]->name); - gpr_free(resources[rid]->description); - gpr_free(resources[rid]->numerators); - gpr_free(resources[rid]->denominators); - gpr_free(resources[rid]); - resources[rid] = NULL; - n_defined_resources--; -} - -void shutdown_resources(void) { - gpr_mu_lock(&resource_lock); - for (size_t i = 0; i < n_resources; i++) { - if (resources[i] != NULL) { - delete_resource_locked(i); - } - } - GPR_ASSERT(n_defined_resources == 0); - gpr_free(resources); - resources = NULL; - n_resources = 0; - gpr_mu_unlock(&resource_lock); -} - -// Check the contents of string fields in a resource proto. -static bool validate_string(pb_istream_t *stream, const pb_field_t *field, - void **arg) { - resource *vresource = (resource *)*arg; - switch (field->tag) { - case google_census_Resource_name_tag: - // Name must have at least one character - if (stream->bytes_left == 0) { - gpr_log(GPR_INFO, "Zero-length Resource name."); - return false; - } - vresource->name = (char *)gpr_malloc(stream->bytes_left + 1); - vresource->name[stream->bytes_left] = '\0'; - if (!pb_read(stream, (uint8_t *)vresource->name, stream->bytes_left)) { - return false; - } - // Can't have same name as an existing resource. - for (size_t i = 0; i < n_resources; i++) { - resource *compare = resources[i]; - if (compare == vresource || compare == NULL) continue; - if (strcmp(compare->name, vresource->name) == 0) { - gpr_log(GPR_INFO, "Duplicate Resource name %s.", vresource->name); - return false; - } - } - break; - case google_census_Resource_description_tag: - if (stream->bytes_left == 0) { - return true; - } - vresource->description = (char *)gpr_malloc(stream->bytes_left + 1); - vresource->description[stream->bytes_left] = '\0'; - if (!pb_read(stream, (uint8_t *)vresource->description, - stream->bytes_left)) { - return false; - } - break; - default: - // No other string fields in Resource. Print warning and skip. - gpr_log(GPR_INFO, "Unknown string field type in Resource protobuf."); - if (!pb_read(stream, NULL, stream->bytes_left)) { - return false; - } - break; - } - return true; -} - -// Decode numerators/denominators in a stream. The `count` and `bup` -// (BasicUnit pointer) are pointers to the approriate fields in a resource -// struct. -static bool validate_units_helper(pb_istream_t *stream, int *count, - google_census_Resource_BasicUnit **bup) { - while (stream->bytes_left) { - (*count)++; - // Have to allocate a new array of values. Normal case is 0 or 1, so - // this should normally not be an issue. - google_census_Resource_BasicUnit *new_bup = - (google_census_Resource_BasicUnit *)gpr_malloc( - (size_t)*count * sizeof(google_census_Resource_BasicUnit)); - if (*count != 1) { - memcpy(new_bup, *bup, - (size_t)(*count - 1) * sizeof(google_census_Resource_BasicUnit)); - gpr_free(*bup); - } - *bup = new_bup; - uint64_t value; - if (!pb_decode_varint(stream, &value)) { - return false; - } - *(*bup + *count - 1) = (google_census_Resource_BasicUnit)value; - } - return true; -} - -// Validate units field of a Resource proto. -static bool validate_units(pb_istream_t *stream, const pb_field_t *field, - void **arg) { - resource *vresource = (resource *)(*arg); - switch (field->tag) { - case google_census_Resource_MeasurementUnit_numerator_tag: - return validate_units_helper(stream, &vresource->n_numerators, - &vresource->numerators); - break; - case google_census_Resource_MeasurementUnit_denominator_tag: - return validate_units_helper(stream, &vresource->n_denominators, - &vresource->denominators); - break; - default: - gpr_log(GPR_ERROR, "Unknown field type."); - return false; - break; - } - return true; -} - -// Validate the contents of a Resource proto. `id` is the intended resource id. -static bool validate_resource_pb(const uint8_t *resource_pb, - size_t resource_pb_size, size_t id) { - GPR_ASSERT(id < n_resources); - if (resource_pb == NULL) { - return false; - } - google_census_Resource vresource; - vresource.name.funcs.decode = &validate_string; - vresource.name.arg = resources[id]; - vresource.description.funcs.decode = &validate_string; - vresource.description.arg = resources[id]; - vresource.unit.numerator.funcs.decode = &validate_units; - vresource.unit.numerator.arg = resources[id]; - vresource.unit.denominator.funcs.decode = &validate_units; - vresource.unit.denominator.arg = resources[id]; - - pb_istream_t stream = - pb_istream_from_buffer((uint8_t *)resource_pb, resource_pb_size); - if (!pb_decode(&stream, google_census_Resource_fields, &vresource)) { - return false; - } - // A Resource must have a name, a unit, with at least one numerator. - return (resources[id]->name != NULL && vresource.has_unit && - resources[id]->n_numerators > 0); -} - -// Allocate a blank resource, and return associated ID. Must be called with -// resource_lock held. -size_t allocate_resource(void) { - // use next_id to optimize expected placement of next new resource. - static size_t next_id = 0; - size_t id = n_resources; // resource ID - initialize to invalid value. - // Expand resources if needed. - if (n_resources == n_defined_resources) { - size_t new_n_resources = n_resources ? n_resources * 2 : 2; - resource **new_resources = - (resource **)gpr_malloc(new_n_resources * sizeof(resource *)); - if (n_resources != 0) { - memcpy(new_resources, resources, n_resources * sizeof(resource *)); - } - memset(new_resources + n_resources, 0, - (new_n_resources - n_resources) * sizeof(resource *)); - gpr_free(resources); - resources = new_resources; - n_resources = new_n_resources; - id = n_defined_resources; - } else { - GPR_ASSERT(n_defined_resources < n_resources); - // Find a free id. - for (size_t base = 0; base < n_resources; base++) { - id = (next_id + base) % n_resources; - if (resources[id] == NULL) break; - } - } - GPR_ASSERT(id < n_resources && resources[id] == NULL); - resources[id] = (resource *)gpr_malloc(sizeof(resource)); - memset(resources[id], 0, sizeof(resource)); - n_defined_resources++; - next_id = (id + 1) % n_resources; - return id; -} - -int32_t census_define_resource(const uint8_t *resource_pb, - size_t resource_pb_size) { - if (resource_pb == NULL) { - return -1; - } - gpr_mu_lock(&resource_lock); - size_t id = allocate_resource(); - // Validate pb, extract name. - if (!validate_resource_pb(resource_pb, resource_pb_size, id)) { - delete_resource_locked(id); - gpr_mu_unlock(&resource_lock); - return -1; - } - gpr_mu_unlock(&resource_lock); - return (int32_t)id; -} - -void census_delete_resource(int32_t rid) { - gpr_mu_lock(&resource_lock); - if (rid >= 0 && (size_t)rid < n_resources && resources[rid] != NULL) { - delete_resource_locked((size_t)rid); - } - gpr_mu_unlock(&resource_lock); -} - -int32_t census_resource_id(const char *name) { - gpr_mu_lock(&resource_lock); - for (int32_t id = 0; (size_t)id < n_resources; id++) { - if (resources[id] != NULL && strcmp(resources[id]->name, name) == 0) { - gpr_mu_unlock(&resource_lock); - return id; - } - } - gpr_mu_unlock(&resource_lock); - return -1; -} - -int32_t define_resource(const resource *base) { - GPR_ASSERT(base != NULL && base->name != NULL && base->n_numerators > 0 && - base->numerators != NULL); - gpr_mu_lock(&resource_lock); - size_t id = allocate_resource(); - size_t len = strlen(base->name) + 1; - resources[id]->name = (char *)gpr_malloc(len); - memcpy(resources[id]->name, base->name, len); - if (base->description) { - len = strlen(base->description) + 1; - resources[id]->description = (char *)gpr_malloc(len); - memcpy(resources[id]->description, base->description, len); - } - resources[id]->prefix = base->prefix; - resources[id]->n_numerators = base->n_numerators; - len = (size_t)base->n_numerators * sizeof(*base->numerators); - resources[id]->numerators = - (google_census_Resource_BasicUnit *)gpr_malloc(len); - memcpy(resources[id]->numerators, base->numerators, len); - resources[id]->n_denominators = base->n_denominators; - if (base->n_denominators != 0) { - len = (size_t)base->n_denominators * sizeof(*base->denominators); - resources[id]->denominators = - (google_census_Resource_BasicUnit *)gpr_malloc(len); - memcpy(resources[id]->denominators, base->denominators, len); - } - gpr_mu_unlock(&resource_lock); - return (int32_t)id; -} diff --git a/src/core/ext/census/resource.h b/src/core/ext/census/resource.h deleted file mode 100644 index 56aaaaf750..0000000000 --- a/src/core/ext/census/resource.h +++ /dev/null @@ -1,56 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -/* Census-internal resource definition and manipluation functions. */ -#ifndef GRPC_CORE_EXT_CENSUS_RESOURCE_H -#define GRPC_CORE_EXT_CENSUS_RESOURCE_H - -#include <grpc/grpc.h> -#include "src/core/ext/census/gen/census.pb.h" - -#ifdef __cplusplus -extern "C" { -#endif - -/* Internal representation of a resource. */ -typedef struct { - char *name; - char *description; - int32_t prefix; - int n_numerators; - google_census_Resource_BasicUnit *numerators; - int n_denominators; - google_census_Resource_BasicUnit *denominators; -} resource; - -/* Initialize and shutdown the resources subsystem. */ -void initialize_resources(void); -void shutdown_resources(void); - -/* Add a new resource, given a proposed resource structure. Returns the - resource ID, or -ve on failure. - TODO(aveitch): this function exists to support addition of the base - resources. It should be removed when we have the ability to add resources - from configuration files. */ -int32_t define_resource(const resource *base); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_RESOURCE_H */ diff --git a/src/core/ext/census/rpc_metric_id.h b/src/core/ext/census/rpc_metric_id.h deleted file mode 100644 index ea493d7288..0000000000 --- a/src/core/ext/census/rpc_metric_id.h +++ /dev/null @@ -1,36 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_RPC_METRIC_ID_H -#define GRPC_CORE_EXT_CENSUS_RPC_METRIC_ID_H - -/* Metric ID's used for RPC measurements. */ -/* Count of client requests sent. */ -#define CENSUS_METRIC_RPC_CLIENT_REQUESTS ((uint32_t)0) -/* Count of server requests sent. */ -#define CENSUS_METRIC_RPC_SERVER_REQUESTS ((uint32_t)1) -/* Client error counts. */ -#define CENSUS_METRIC_RPC_CLIENT_ERRORS ((uint32_t)2) -/* Server error counts. */ -#define CENSUS_METRIC_RPC_SERVER_ERRORS ((uint32_t)3) -/* Client side request latency. */ -#define CENSUS_METRIC_RPC_CLIENT_LATENCY ((uint32_t)4) -/* Server side request latency. */ -#define CENSUS_METRIC_RPC_SERVER_LATENCY ((uint32_t)5) - -#endif /* GRPC_CORE_EXT_CENSUS_RPC_METRIC_ID_H */ diff --git a/src/core/ext/census/trace_context.cc b/src/core/ext/census/trace_context.cc deleted file mode 100644 index af92ae6d9e..0000000000 --- a/src/core/ext/census/trace_context.cc +++ /dev/null @@ -1,71 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include "src/core/ext/census/trace_context.h" - -#include <grpc/census.h> -#include <grpc/support/log.h> -#include <stdbool.h> - -#include "third_party/nanopb/pb_decode.h" -#include "third_party/nanopb/pb_encode.h" - -// This function assumes the TraceContext is valid. -size_t encode_trace_context(google_trace_TraceContext *ctxt, uint8_t *buffer, - const size_t buf_size) { - // Create a stream that will write to our buffer. - pb_ostream_t stream = pb_ostream_from_buffer(buffer, buf_size); - - // encode message - bool status = pb_encode(&stream, google_trace_TraceContext_fields, ctxt); - - if (!status) { - gpr_log(GPR_DEBUG, "TraceContext encoding failed: %s", - PB_GET_ERROR(&stream)); - return 0; - } - - return stream.bytes_written; -} - -bool decode_trace_context(google_trace_TraceContext *ctxt, uint8_t *buffer, - const size_t nbytes) { - // Create a stream that reads nbytes from the buffer. - pb_istream_t stream = pb_istream_from_buffer(buffer, nbytes); - - // decode message - bool status = pb_decode(&stream, google_trace_TraceContext_fields, ctxt); - - if (!status) { - gpr_log(GPR_DEBUG, "TraceContext decoding failed: %s", - PB_GET_ERROR(&stream)); - return false; - } - - // check fields - if (!ctxt->has_trace_id_hi || !ctxt->has_trace_id_lo) { - gpr_log(GPR_DEBUG, "Invalid TraceContext: missing trace_id"); - return false; - } - if (!ctxt->has_span_id) { - gpr_log(GPR_DEBUG, "Invalid TraceContext: missing span_id"); - return false; - } - - return true; -} diff --git a/src/core/ext/census/trace_context.h b/src/core/ext/census/trace_context.h deleted file mode 100644 index 2b828ba4da..0000000000 --- a/src/core/ext/census/trace_context.h +++ /dev/null @@ -1,64 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -/* Functions for manipulating trace contexts as defined in - src/proto/census/trace.proto */ -#ifndef GRPC_CORE_EXT_CENSUS_TRACE_CONTEXT_H -#define GRPC_CORE_EXT_CENSUS_TRACE_CONTEXT_H - -#include "src/core/ext/census/gen/trace_context.pb.h" - -/* Span option flags. */ -#define SPAN_OPTIONS_IS_SAMPLED 0x01 - -/* Maximum number of bytes required to encode a TraceContext (31) -1 byte for trace_id field -1 byte for trace_id length -1 byte for trace_id.hi field -8 bytes for trace_id.hi (uint64_t) -1 byte for trace_id.lo field -8 bytes for trace_id.lo (uint64_t) -1 byte for span_id field -8 bytes for span_id (uint64_t) -1 byte for is_sampled field -1 byte for is_sampled (bool) */ -#define TRACE_MAX_CONTEXT_SIZE 31 - -#ifdef __cplusplus -extern "C" { -#endif - -/* Encode a trace context (ctxt) into proto format to the buffer provided. The -size of buffer must be at least TRACE_MAX_CONTEXT_SIZE. On success, returns the -number of bytes successfully encoded into buffer. On failure, returns 0. */ -size_t encode_trace_context(google_trace_TraceContext *ctxt, uint8_t *buffer, - const size_t buf_size); - -/* Decode a proto-encoded TraceContext from the provided buffer into the -TraceContext structure (ctxt). The function expects to be supplied the number -of bytes to be read from buffer (nbytes). This function will also validate that -the TraceContext has a span_id and a trace_id, and will return false if either -of these do not exist. On success, returns true and false otherwise. */ -bool decode_trace_context(google_trace_TraceContext *ctxt, uint8_t *buffer, - const size_t nbytes); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_TRACE_CONTEXT_H */ diff --git a/src/core/ext/census/trace_label.h b/src/core/ext/census/trace_label.h deleted file mode 100644 index 97ce399eb5..0000000000 --- a/src/core/ext/census/trace_label.h +++ /dev/null @@ -1,46 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_TRACE_LABEL_H -#define GRPC_CORE_EXT_CENSUS_TRACE_LABEL_H - -#include "src/core/ext/census/trace_string.h" - -/* Trace label (key/value pair) stores a label name and the label value. The - value can be one of trace_string/int64_t/bool. */ -typedef struct trace_label { - trace_string key; - enum label_type { - /* Unknown value for debugging/error purposes */ - LABEL_UNKNOWN = 0, - /* A string value */ - LABEL_STRING = 1, - /* An integer value. */ - LABEL_INT = 2, - /* A boolean value. */ - LABEL_BOOL = 3, - } value_type; - - union value { - trace_string label_str; - int64_t label_int; - bool label_bool; - } value; -} trace_label; - -#endif /* GRPC_CORE_EXT_CENSUS_TRACE_LABEL_H */ diff --git a/src/core/ext/census/trace_propagation.h b/src/core/ext/census/trace_propagation.h deleted file mode 100644 index e05fd23a1f..0000000000 --- a/src/core/ext/census/trace_propagation.h +++ /dev/null @@ -1,56 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_TRACE_PROPAGATION_H -#define GRPC_CORE_EXT_CENSUS_TRACE_PROPAGATION_H - -#include "src/core/ext/census/tracing.h" - -#ifdef __cplusplus -extern "C" { -#endif - -/* Encoding and decoding functions for receiving and sending trace contexts - over the wire. Only RPC libraries should be calling these - functions. These functions return the number of bytes encoded/decoded - (0 if a failure has occurred). buf_size indicates the size of the - input/output buffer. trace_span_context is a struct that includes the - trace ID, span ID, and a set of option flags (is_sampled, etc.). */ - -/* Converts a span context to a binary byte buffer. */ -size_t trace_span_context_to_binary(const trace_span_context *ctxt, - uint8_t *buf, size_t buf_size); - -/* Reads a binary byte buffer and populates a span context structure. */ -size_t binary_to_trace_span_context(const uint8_t *buf, size_t buf_size, - trace_span_context *ctxt); - -/* Converts a span context to an http metadata compatible string. */ -size_t trace_span_context_to_http_format(const trace_span_context *ctxt, - char *buf, size_t buf_size); - -/* Reads an http metadata compatible string and populates a span context - structure. */ -size_t http_format_to_trace_span_context(const char *buf, size_t buf_size, - trace_span_context *ctxt); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_TRACE_PROPAGATION_H */ diff --git a/src/core/ext/census/trace_status.h b/src/core/ext/census/trace_status.h deleted file mode 100644 index dd83d3f729..0000000000 --- a/src/core/ext/census/trace_status.h +++ /dev/null @@ -1,30 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_TRACE_STATUS_H -#define GRPC_CORE_EXT_CENSUS_TRACE_STATUS_H - -#include "src/core/ext/census/trace_string.h" - -/* Stores a status code and status message for a trace. */ -typedef struct trace_status { - int64_t errorCode; - trace_string errorMessage; -} trace_status; - -#endif /* GRPC_CORE_EXT_CENSUS_TRACE_STATUS_H */ diff --git a/src/core/ext/census/trace_string.h b/src/core/ext/census/trace_string.h deleted file mode 100644 index e4da3f590d..0000000000 --- a/src/core/ext/census/trace_string.h +++ /dev/null @@ -1,35 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_TRACE_STRING_H -#define GRPC_CORE_EXT_CENSUS_TRACE_STRING_H - -#include <grpc/slice.h> - -/* String struct for tracing messages. Since this is a C API, we do not have - access to a string class. This is intended for use by higher level - languages which wrap around the C API, as most of them have a string class. - This will also be more efficient when copying, as we have an explicitly - specified length. Also, grpc_slice has reference counting which allows for - interning. */ -typedef struct trace_string { - char *string; - size_t length; -} trace_string; - -#endif /* GRPC_CORE_EXT_CENSUS_TRACE_STRING_H */ diff --git a/src/core/ext/census/tracing.cc b/src/core/ext/census/tracing.cc deleted file mode 100644 index 823c681abf..0000000000 --- a/src/core/ext/census/tracing.cc +++ /dev/null @@ -1,55 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include "src/core/ext/census/tracing.h" - -#include <grpc/census.h> -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include "src/core/ext/census/mlog.h" - -void trace_start_span(const trace_span_context *span_ctxt, - const trace_string name, const start_span_options *opts, - trace_span_context *new_span_ctxt, - bool has_remote_parent) { - // Noop implementation. -} - -void trace_add_span_annotation(const trace_string description, - const trace_label *labels, const size_t n_labels, - trace_span_context *span_ctxt) { - // Noop implementation. -} - -void trace_add_span_network_event_annotation(const trace_string description, - const trace_label *labels, - const size_t n_labels, - const gpr_timespec timestamp, - bool sent, uint64_t id, - trace_span_context *span_ctxt) { - // Noop implementation. -} - -void trace_add_span_labels(const trace_label *labels, const size_t n_labels, - trace_span_context *span_ctxt) { - // Noop implementation. -} - -void trace_end_span(const trace_status *status, trace_span_context *span_ctxt) { - // Noop implementation. -} diff --git a/src/core/ext/census/tracing.h b/src/core/ext/census/tracing.h deleted file mode 100644 index 0690de8655..0000000000 --- a/src/core/ext/census/tracing.h +++ /dev/null @@ -1,117 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_TRACING_H -#define GRPC_CORE_EXT_CENSUS_TRACING_H - -#include <grpc/support/time.h> -#include <stdbool.h> -#include "src/core/ext/census/trace_context.h" -#include "src/core/ext/census/trace_label.h" -#include "src/core/ext/census/trace_status.h" - -#ifdef __cplusplus -extern "C" { -#endif - -/* This is the low level tracing API that other languages will interface with. - This is not intended to be accessed by the end-user, therefore it has been - designed with performance in mind rather than ease of use. */ - -/* The tracing level. */ -enum TraceLevel { - /* Annotations on this context will be silently discarded. */ - NO_TRACING = 0, - /* Annotations will not be saved to a persistent store. They will be - available via local APIs only. This setting is not propagated to child - spans. */ - TRANSIENT_TRACING = 1, - /* Annotations are recorded for the entire distributed trace and they are - saved to a persistent store. This setting is propagated to child spans. */ - PERSISTENT_TRACING = 2, -}; - -typedef struct trace_span_context { - /* Trace span context stores Span ID, Trace ID, and option flags. */ - /* Trace ID is 128 bits split into 2 64-bit chunks (hi and lo). */ - uint64_t trace_id_hi; - uint64_t trace_id_lo; - /* Span ID is 64 bits. */ - uint64_t span_id; - /* Span-options is 32-bit value which contains flag options. */ - uint32_t span_options; -} trace_span_context; - -typedef struct start_span_options { - /* If set, this will override the Span.local_start_time for the Span. */ - gpr_timespec local_start_timestamp; - - /* Linked spans can be used to identify spans that are linked to this span in - a different trace. This can be used (for example) in batching operations, - where a single batch handler processes multiple requests from different - traces. If set, points to a list of Spans are linked to the created Span.*/ - trace_span_context *linked_spans; - /* The number of linked spans. */ - size_t n_linked_spans; -} start_span_options; - -/* Create a new child Span (or root if parent is NULL), with parent being the - designated Span. The child span will have the provided name and starting - span options (optional). The bool has_remote_parent marks whether the - context refers to a remote parent span or not. */ -void trace_start_span(const trace_span_context *span_ctxt, - const trace_string name, const start_span_options *opts, - trace_span_context *new_span_ctxt, - bool has_remote_parent); - -/* Add a new Annotation to the Span. Annotations consist of a description - (trace_string) and a set of n labels (trace_label). This can be populated - with arbitrary user data. */ -void trace_add_span_annotation(const trace_string description, - const trace_label *labels, const size_t n_labels, - trace_span_context *span_ctxt); - -/* Add a new NetworkEvent annotation to a Span. This function is only intended - to be used by RPC systems (either client or server), not by higher level - applications. The timestamp type will be system-defined, the sent argument - designates whether this is a network send event (client request, server - reply)or receive (server request, client reply). The id argument corresponds - to Span.Annotation.NetworkEvent.id from the data model, and serves to uniquely - identify each network message. */ -void trace_add_span_network_event(const trace_string description, - const trace_label *labels, - const size_t n_labels, - const gpr_timespec timestamp, bool sent, - uint64_t id, trace_span_context *span_ctxt); - -/* Add a set of labels to the Span. These will correspond to the field -Span.labels in the data model. */ -void trace_add_span_labels(const trace_label *labels, const size_t n_labels, - trace_span_context *span_ctxt); - -/* Mark the end of Span Execution with the given status. Only the timing of the -first EndSpan call for a given Span will be recorded, and implementations are -free to ignore all further calls using the Span. EndSpanOptions can -optionally be NULL. */ -void trace_end_span(const trace_status *status, trace_span_context *span_ctxt); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_TRACING_H */ diff --git a/src/core/ext/census/window_stats.cc b/src/core/ext/census/window_stats.cc deleted file mode 100644 index 0058e4bf9c..0000000000 --- a/src/core/ext/census/window_stats.cc +++ /dev/null @@ -1,301 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include "src/core/ext/census/window_stats.h" -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/time.h> -#include <grpc/support/useful.h> -#include <math.h> -#include <stddef.h> -#include <string.h> - -/* typedefs make typing long names easier. Use cws (for census_window_stats) */ -typedef census_window_stats_stat_info cws_stat_info; -typedef struct census_window_stats_sum cws_sum; - -/* Each interval is composed of a number of buckets, which hold a count of - entries and a single statistic */ -typedef struct census_window_stats_bucket { - int64_t count; - void *statistic; -} cws_bucket; - -/* Each interval has a set of buckets, and the variables needed to keep - track of their current state */ -typedef struct census_window_stats_interval_stats { - /* The buckets. There will be 'granularity' + 1 of these. */ - cws_bucket *buckets; - /* Index of the bucket containing the smallest time interval. */ - int bottom_bucket; - /* The smallest time storable in the current window. */ - int64_t bottom; - /* The largest time storable in the current window + 1ns */ - int64_t top; - /* The width of each bucket in ns. */ - int64_t width; -} cws_interval_stats; - -typedef struct census_window_stats { - /* Number of intervals. */ - int nintervals; - /* Number of buckets in each interval. 'granularity' + 1. */ - int nbuckets; - /* Record of stat_info. */ - cws_stat_info stat_info; - /* Stats for each interval. */ - cws_interval_stats *interval_stats; - /* The time the newset stat was recorded. */ - int64_t newest_time; -} window_stats; - -/* Calculate an actual bucket index from a logical index 'IDX'. Other - parameters supply information on the interval struct and overall stats. */ -#define BUCKET_IDX(IS, IDX, WSTATS) \ - ((IS->bottom_bucket + (IDX)) % WSTATS->nbuckets) - -/* The maximum seconds value we can have in a valid timespec. More than this - will result in overflow in timespec_to_ns(). This works out to ~292 years. - TODO: consider using doubles instead of int64. */ -static int64_t max_seconds = (GPR_INT64_MAX - GPR_NS_PER_SEC) / GPR_NS_PER_SEC; - -static int64_t timespec_to_ns(const gpr_timespec ts) { - if (ts.tv_sec > max_seconds) { - return GPR_INT64_MAX - 1; - } - return ts.tv_sec * GPR_NS_PER_SEC + ts.tv_nsec; -} - -static void cws_initialize_statistic(void *statistic, - const cws_stat_info *stat_info) { - if (stat_info->stat_initialize == NULL) { - memset(statistic, 0, stat_info->stat_size); - } else { - stat_info->stat_initialize(statistic); - } -} - -/* Create and initialize a statistic */ -static void *cws_create_statistic(const cws_stat_info *stat_info) { - void *stat = gpr_malloc(stat_info->stat_size); - cws_initialize_statistic(stat, stat_info); - return stat; -} - -window_stats *census_window_stats_create(int nintervals, - const gpr_timespec intervals[], - int granularity, - const cws_stat_info *stat_info) { - window_stats *ret; - int i; - /* validate inputs */ - GPR_ASSERT(nintervals > 0 && granularity > 2 && intervals != NULL && - stat_info != NULL); - for (i = 0; i < nintervals; i++) { - int64_t ns = timespec_to_ns(intervals[i]); - GPR_ASSERT(intervals[i].tv_sec >= 0 && intervals[i].tv_nsec >= 0 && - intervals[i].tv_nsec < GPR_NS_PER_SEC && ns >= 100 && - granularity * 10 <= ns); - } - /* Allocate and initialize relevant data structures */ - ret = (window_stats *)gpr_malloc(sizeof(window_stats)); - ret->nintervals = nintervals; - ret->nbuckets = granularity + 1; - ret->stat_info = *stat_info; - ret->interval_stats = - (cws_interval_stats *)gpr_malloc(nintervals * sizeof(cws_interval_stats)); - for (i = 0; i < nintervals; i++) { - int64_t size_ns = timespec_to_ns(intervals[i]); - cws_interval_stats *is = ret->interval_stats + i; - cws_bucket *buckets = is->buckets = - (cws_bucket *)gpr_malloc(ret->nbuckets * sizeof(cws_bucket)); - int b; - for (b = 0; b < ret->nbuckets; b++) { - buckets[b].statistic = cws_create_statistic(stat_info); - buckets[b].count = 0; - } - is->bottom_bucket = 0; - is->bottom = 0; - is->width = size_ns / granularity; - /* Check for possible overflow issues, and maximize interval size if the - user requested something large enough. */ - if ((GPR_INT64_MAX - is->width) > size_ns) { - is->top = size_ns + is->width; - } else { - is->top = GPR_INT64_MAX; - is->width = GPR_INT64_MAX / (granularity + 1); - } - /* If size doesn't divide evenly, we can have a width slightly too small; - better to have it slightly large. */ - if ((size_ns - (granularity + 1) * is->width) > 0) { - is->width += 1; - } - } - ret->newest_time = 0; - return ret; -} - -/* When we try adding a measurement above the current interval range, we - need to "shift" the buckets sufficiently to cover the new range. */ -static void cws_shift_buckets(const window_stats *wstats, - cws_interval_stats *is, int64_t when_ns) { - int i; - /* number of bucket time widths to "shift" */ - int shift; - /* number of buckets to clear */ - int nclear; - GPR_ASSERT(when_ns >= is->top); - /* number of bucket time widths to "shift" */ - shift = ((when_ns - is->top) / is->width) + 1; - /* number of buckets to clear - limited by actual number of buckets */ - nclear = GPR_MIN(shift, wstats->nbuckets); - for (i = 0; i < nclear; i++) { - int b = BUCKET_IDX(is, i, wstats); - is->buckets[b].count = 0; - cws_initialize_statistic(is->buckets[b].statistic, &wstats->stat_info); - } - /* adjust top/bottom times and current bottom bucket */ - is->bottom_bucket = BUCKET_IDX(is, shift, wstats); - is->top += shift * is->width; - is->bottom += shift * is->width; -} - -void census_window_stats_add(window_stats *wstats, const gpr_timespec when, - const void *stat_value) { - int i; - int64_t when_ns = timespec_to_ns(when); - GPR_ASSERT(wstats->interval_stats != NULL); - for (i = 0; i < wstats->nintervals; i++) { - cws_interval_stats *is = wstats->interval_stats + i; - cws_bucket *bucket; - if (when_ns < is->bottom) { /* Below smallest time in interval: drop */ - continue; - } - if (when_ns >= is->top) { /* above limit: shift buckets */ - cws_shift_buckets(wstats, is, when_ns); - } - /* Add the stat. */ - GPR_ASSERT(is->bottom <= when_ns && when_ns < is->top); - bucket = is->buckets + - BUCKET_IDX(is, (when_ns - is->bottom) / is->width, wstats); - bucket->count++; - wstats->stat_info.stat_add(bucket->statistic, stat_value); - } - if (when_ns > wstats->newest_time) { - wstats->newest_time = when_ns; - } -} - -/* Add a specific bucket contents to an accumulating total. */ -static void cws_add_bucket_to_sum(cws_sum *sum, const cws_bucket *bucket, - const cws_stat_info *stat_info) { - sum->count += bucket->count; - stat_info->stat_add(sum->statistic, bucket->statistic); -} - -/* Add a proportion to an accumulating sum. */ -static void cws_add_proportion_to_sum(double p, cws_sum *sum, - const cws_bucket *bucket, - const cws_stat_info *stat_info) { - sum->count += p * bucket->count; - stat_info->stat_add_proportion(p, sum->statistic, bucket->statistic); -} - -void census_window_stats_get_sums(const window_stats *wstats, - const gpr_timespec when, cws_sum sums[]) { - int i; - int64_t when_ns = timespec_to_ns(when); - GPR_ASSERT(wstats->interval_stats != NULL); - for (i = 0; i < wstats->nintervals; i++) { - int when_bucket; - int new_bucket; - double last_proportion = 1.0; - double bottom_proportion; - cws_interval_stats *is = wstats->interval_stats + i; - cws_sum *sum = sums + i; - sum->count = 0; - cws_initialize_statistic(sum->statistic, &wstats->stat_info); - if (when_ns < is->bottom) { - continue; - } - if (when_ns >= is->top) { - cws_shift_buckets(wstats, is, when_ns); - } - /* Calculating the appropriate amount of which buckets to use can get - complicated. Essentially there are two cases: - 1) if the "top" bucket (new_bucket, where the newest additions to the - stats recorded are entered) corresponds to 'when', then we need - to take a proportion of it - (if when < newest_time) or the full - thing. We also (possibly) need to take a corresponding - proportion of the bottom bucket. - 2) Other cases, we just take a straight proportion. - */ - when_bucket = (when_ns - is->bottom) / is->width; - new_bucket = (wstats->newest_time - is->bottom) / is->width; - if (new_bucket == when_bucket) { - int64_t bottom_bucket_time = is->bottom + when_bucket * is->width; - if (when_ns < wstats->newest_time) { - last_proportion = (double)(when_ns - bottom_bucket_time) / - (double)(wstats->newest_time - bottom_bucket_time); - bottom_proportion = - (double)(is->width - (when_ns - bottom_bucket_time)) / is->width; - } else { - bottom_proportion = - (double)(is->width - (wstats->newest_time - bottom_bucket_time)) / - is->width; - } - } else { - last_proportion = - (double)(when_ns + 1 - is->bottom - when_bucket * is->width) / - is->width; - bottom_proportion = 1.0 - last_proportion; - } - cws_add_proportion_to_sum(last_proportion, sum, - is->buckets + BUCKET_IDX(is, when_bucket, wstats), - &wstats->stat_info); - if (when_bucket != 0) { /* last bucket isn't also bottom bucket */ - int b; - /* Add all of "bottom" bucket if we are looking at a subset of the - full interval, or a proportion if we are adding full interval. */ - cws_add_proportion_to_sum( - (when_bucket == wstats->nbuckets - 1 ? bottom_proportion : 1.0), sum, - is->buckets + is->bottom_bucket, &wstats->stat_info); - /* Add all the remaining buckets (everything but top and bottom). */ - for (b = 1; b < when_bucket; b++) { - cws_add_bucket_to_sum(sum, is->buckets + BUCKET_IDX(is, b, wstats), - &wstats->stat_info); - } - } - } -} - -void census_window_stats_destroy(window_stats *wstats) { - int i; - GPR_ASSERT(wstats->interval_stats != NULL); - for (i = 0; i < wstats->nintervals; i++) { - int b; - for (b = 0; b < wstats->nbuckets; b++) { - gpr_free(wstats->interval_stats[i].buckets[b].statistic); - } - gpr_free(wstats->interval_stats[i].buckets); - } - gpr_free(wstats->interval_stats); - /* Ensure any use-after free triggers assert. */ - wstats->interval_stats = NULL; - gpr_free(wstats); -} diff --git a/src/core/ext/census/window_stats.h b/src/core/ext/census/window_stats.h deleted file mode 100644 index 2a1d6d0d16..0000000000 --- a/src/core/ext/census/window_stats.h +++ /dev/null @@ -1,166 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_EXT_CENSUS_WINDOW_STATS_H -#define GRPC_CORE_EXT_CENSUS_WINDOW_STATS_H - -#include <grpc/support/time.h> - -#ifdef __cplusplus -extern "C" { -#endif - -/* Keep rolling sums of a user-defined statistic (containing a number of - measurements) over a a number of time intervals ("windows"). For example, - you can use a window_stats object to answer questions such as - "Approximately how many RPCs/s did I receive over the past minute, and - approximately how many bytes did I send out over that period?". - - The type of data to record, and the time intervals to keep are specified - when creating the object via a call to census_window_stats_create(). - - A window's interval is divided into one or more "buckets"; the interval - must be divisible by the number of buckets. Internally, these buckets - control the granularity of window_stats' measurements. Increasing the - number of buckets lets the object respond more quickly to changes in the - overall rate of data added into the object, at the cost of additional - memory usage. - - Here's some code which keeps one minute/hour measurements for two values - (latency in seconds and bytes transferred), with each interval divided into - 4 buckets. - - typedef struct my_stat { - double latency; - int bytes; - } my_stat; - - void add_my_stat(void* base, const void* addme) { - my_stat* b = (my_stat*)base; - const my_stat* a = (const my_stat*)addme; - b->latency += a->latency; - b->bytes += a->bytes; - } - - void add_proportion_my_stat(double p, void* base, const void* addme) { - (my_stat*)result->latency += p * (const my_stat*)base->latency; - (my_stat*)result->bytes += p * (const my_stat*)base->bytes; - } - - #define kNumIntervals 2 - #define kMinInterval 0 - #define kHourInterval 1 - #define kNumBuckets 4 - - const struct census_window_stats_stat_info kMyStatInfo - = { sizeof(my_stat), NULL, add_my_stat, add_proportion_my_stat }; - gpr_timespec intervals[kNumIntervals] = {{60, 0}, {3600, 0}}; - my_stat stat; - my_stat sums[kNumIntervals]; - census_window_stats_sums result[kNumIntervals]; - struct census_window_stats* stats - = census_window_stats_create(kNumIntervals, intervals, kNumBuckets, - &kMyStatInfo); - // Record a new event, taking 15.3ms, transferring 1784 bytes. - stat.latency = 0.153; - stat.bytes = 1784; - census_window_stats_add(stats, gpr_now(GPR_CLOCK_REALTIME), &stat); - // Get sums and print them out - result[kMinInterval].statistic = &sums[kMinInterval]; - result[kHourInterval].statistic = &sums[kHourInterval]; - census_window_stats_get_sums(stats, gpr_now(GPR_CLOCK_REALTIME), result); - printf("%d events/min, average time %gs, average bytes %g\n", - result[kMinInterval].count, - (my_stat*)result[kMinInterval].statistic->latency / - result[kMinInterval].count, - (my_stat*)result[kMinInterval].statistic->bytes / - result[kMinInterval].count - ); - printf("%d events/hr, average time %gs, average bytes %g\n", - result[kHourInterval].count, - (my_stat*)result[kHourInterval].statistic->latency / - result[kHourInterval].count, - (my_stat*)result[kHourInterval].statistic->bytes / - result[kHourInterval].count - ); -*/ - -/* Opaque structure for representing window_stats object */ -struct census_window_stats; - -/* Information provided by API user on the information they want to record */ -typedef struct census_window_stats_stat_info { - /* Number of bytes in user-defined object. */ - size_t stat_size; - /* Function to initialize a user-defined statistics object. If this is set - * to NULL, then the object will be zero-initialized. */ - void (*stat_initialize)(void *stat); - /* Function to add one user-defined statistics object ('addme') to 'base' */ - void (*stat_add)(void *base, const void *addme); - /* As for previous function, but only add a proportion 'p'. This API will - currently only use 'p' values in the range [0,1], but other values are - possible in the future, and should be supported. */ - void (*stat_add_proportion)(double p, void *base, const void *addme); -} census_window_stats_stat_info; - -/* Create a new window_stats object. 'nintervals' is the number of - 'intervals', and must be >=1. 'granularity' is the number of buckets, with - a larger number using more memory, but providing greater accuracy of - results. 'granularity should be > 2. We also require that each interval be - at least 10 * 'granularity' nanoseconds in size. 'stat_info' contains - information about the statistic to be gathered. Intervals greater than ~192 - years will be treated as essentially infinite in size. This function will - GPR_ASSERT() if the object cannot be created or any of the parameters have - invalid values. This function is thread-safe. */ -struct census_window_stats *census_window_stats_create( - int nintervals, const gpr_timespec intervals[], int granularity, - const census_window_stats_stat_info *stat_info); - -/* Add a new measurement (in 'stat_value'), as of a given time ('when'). - This function is thread-compatible. */ -void census_window_stats_add(struct census_window_stats *wstats, - const gpr_timespec when, const void *stat_value); - -/* Structure used to record a single intervals sum for a given statistic */ -typedef struct census_window_stats_sum { - /* Total count of samples. Note that because some internal interpolation - is performed, the count of samples returned for each interval may not be an - integral value. */ - double count; - /* Sum for statistic */ - void *statistic; -} census_window_stats_sums; - -/* Retrieve a set of all values stored in a window_stats object 'wstats'. The - number of 'sums' MUST be the same as the number 'nintervals' used in - census_window_stats_create(). This function is thread-compatible. */ -void census_window_stats_get_sums(const struct census_window_stats *wstats, - const gpr_timespec when, - struct census_window_stats_sum sums[]); - -/* Destroy a window_stats object. Once this function has been called, the - object will no longer be usable from any of the above functions (and - calling them will most likely result in a NULL-pointer dereference or - assertion failure). This function is thread-compatible. */ -void census_window_stats_destroy(struct census_window_stats *wstats); - -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_WINDOW_STATS_H */ diff --git a/src/core/ext/filters/client_channel/backup_poller.cc b/src/core/ext/filters/client_channel/backup_poller.cc new file mode 100644 index 0000000000..466bf86bc0 --- /dev/null +++ b/src/core/ext/filters/client_channel/backup_poller.cc @@ -0,0 +1,158 @@ +/* + * + * Copyright 2015 gRPC authors. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + * + */ + +#include "src/core/ext/filters/client_channel/backup_poller.h" + +#include <grpc/grpc.h> +#include <grpc/support/alloc.h> +#include <grpc/support/log.h> +#include <grpc/support/sync.h> +#include "src/core/ext/filters/client_channel/client_channel.h" +#include "src/core/lib/iomgr/error.h" +#include "src/core/lib/iomgr/pollset.h" +#include "src/core/lib/iomgr/timer.h" +#include "src/core/lib/support/env.h" +#include "src/core/lib/support/string.h" +#include "src/core/lib/surface/channel.h" +#include "src/core/lib/surface/completion_queue.h" + +#define DEFAULT_POLL_INTERVAL_MS 5000 + +typedef struct backup_poller { + grpc_timer polling_timer; + grpc_closure run_poller_closure; + grpc_closure shutdown_closure; + gpr_mu* pollset_mu; + grpc_pollset* pollset; // guarded by pollset_mu + bool shutting_down; // guarded by pollset_mu + gpr_refcount refs; + gpr_refcount shutdown_refs; +} backup_poller; + +static gpr_once g_once = GPR_ONCE_INIT; +static gpr_mu g_poller_mu; +static backup_poller* g_poller = NULL; // guarded by g_poller_mu +// g_poll_interval_ms is set only once at the first time +// grpc_client_channel_start_backup_polling() is called, after that it is +// treated as const. +static int g_poll_interval_ms = DEFAULT_POLL_INTERVAL_MS; + +static void init_globals() { + gpr_mu_init(&g_poller_mu); + char* env = gpr_getenv("GRPC_CLIENT_CHANNEL_BACKUP_POLL_INTERVAL_MS"); + if (env != NULL) { + int poll_interval_ms = gpr_parse_nonnegative_int(env); + if (poll_interval_ms == -1) { + gpr_log(GPR_ERROR, + "Invalid GRPC_CLIENT_CHANNEL_BACKUP_POLL_INTERVAL_MS: %s, " + "default value %d will be used.", + env, g_poll_interval_ms); + } else { + g_poll_interval_ms = poll_interval_ms; + } + } + gpr_free(env); +} + +static void backup_poller_shutdown_unref(grpc_exec_ctx* exec_ctx, + backup_poller* p) { + if (gpr_unref(&p->shutdown_refs)) { + grpc_pollset_destroy(exec_ctx, p->pollset); + gpr_free(p->pollset); + gpr_free(p); + } +} + +static void done_poller(grpc_exec_ctx* exec_ctx, void* arg, grpc_error* error) { + backup_poller_shutdown_unref(exec_ctx, (backup_poller*)arg); +} + +static void g_poller_unref(grpc_exec_ctx* exec_ctx) { + if (gpr_unref(&g_poller->refs)) { + gpr_mu_lock(&g_poller_mu); + backup_poller* p = g_poller; + g_poller = NULL; + gpr_mu_unlock(&g_poller_mu); + gpr_mu_lock(p->pollset_mu); + p->shutting_down = true; + grpc_pollset_shutdown(exec_ctx, p->pollset, + GRPC_CLOSURE_INIT(&p->shutdown_closure, done_poller, + p, grpc_schedule_on_exec_ctx)); + gpr_mu_unlock(p->pollset_mu); + grpc_timer_cancel(exec_ctx, &p->polling_timer); + } +} + +static void run_poller(grpc_exec_ctx* exec_ctx, void* arg, grpc_error* error) { + backup_poller* p = (backup_poller*)arg; + if (error != GRPC_ERROR_NONE) { + if (error != GRPC_ERROR_CANCELLED) { + GRPC_LOG_IF_ERROR("run_poller", GRPC_ERROR_REF(error)); + } + backup_poller_shutdown_unref(exec_ctx, p); + return; + } + gpr_mu_lock(p->pollset_mu); + if (p->shutting_down) { + gpr_mu_unlock(p->pollset_mu); + backup_poller_shutdown_unref(exec_ctx, p); + return; + } + grpc_error* err = grpc_pollset_work(exec_ctx, p->pollset, NULL, + grpc_exec_ctx_now(exec_ctx)); + gpr_mu_unlock(p->pollset_mu); + GRPC_LOG_IF_ERROR("Run client channel backup poller", err); + grpc_timer_init(exec_ctx, &p->polling_timer, + grpc_exec_ctx_now(exec_ctx) + g_poll_interval_ms, + &p->run_poller_closure); +} + +void grpc_client_channel_start_backup_polling( + grpc_exec_ctx* exec_ctx, grpc_pollset_set* interested_parties) { + gpr_once_init(&g_once, init_globals); + if (g_poll_interval_ms == 0) { + return; + } + gpr_mu_lock(&g_poller_mu); + if (g_poller == NULL) { + g_poller = (backup_poller*)gpr_zalloc(sizeof(backup_poller)); + g_poller->pollset = (grpc_pollset*)gpr_zalloc(grpc_pollset_size()); + g_poller->shutting_down = false; + grpc_pollset_init(g_poller->pollset, &g_poller->pollset_mu); + gpr_ref_init(&g_poller->refs, 0); + // one for timer cancellation, one for pollset shutdown + gpr_ref_init(&g_poller->shutdown_refs, 2); + GRPC_CLOSURE_INIT(&g_poller->run_poller_closure, run_poller, g_poller, + grpc_schedule_on_exec_ctx); + grpc_timer_init(exec_ctx, &g_poller->polling_timer, + grpc_exec_ctx_now(exec_ctx) + g_poll_interval_ms, + &g_poller->run_poller_closure); + } + gpr_ref(&g_poller->refs); + gpr_mu_unlock(&g_poller_mu); + grpc_pollset_set_add_pollset(exec_ctx, interested_parties, g_poller->pollset); +} + +void grpc_client_channel_stop_backup_polling( + grpc_exec_ctx* exec_ctx, grpc_pollset_set* interested_parties) { + if (g_poll_interval_ms == 0) { + return; + } + grpc_pollset_set_del_pollset(exec_ctx, interested_parties, g_poller->pollset); + g_poller_unref(exec_ctx); +} diff --git a/src/core/ext/census/grpc_filter.h b/src/core/ext/filters/client_channel/backup_poller.h index 7940363061..e993d50639 100644 --- a/src/core/ext/census/grpc_filter.h +++ b/src/core/ext/filters/client_channel/backup_poller.h @@ -16,22 +16,19 @@ * */ -#ifndef GRPC_CORE_EXT_CENSUS_GRPC_FILTER_H -#define GRPC_CORE_EXT_CENSUS_GRPC_FILTER_H +#ifndef GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_BACKUP_POLLER_H +#define GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_BACKUP_POLLER_H +#include <grpc/grpc.h> #include "src/core/lib/channel/channel_stack.h" +#include "src/core/lib/iomgr/exec_ctx.h" -#ifdef __cplusplus -extern "C" { -#endif +/* Start polling \a interested_parties periodically in the timer thread */ +void grpc_client_channel_start_backup_polling( + grpc_exec_ctx* exec_ctx, grpc_pollset_set* interested_parties); -/* Census filters: provides tracing and stats collection functionalities. It - needs to reside right below the surface filter in the channel stack. */ -extern const grpc_channel_filter grpc_client_census_filter; -extern const grpc_channel_filter grpc_server_census_filter; +/* Stop polling \a interested_parties */ +void grpc_client_channel_stop_backup_polling( + grpc_exec_ctx* exec_ctx, grpc_pollset_set* interested_parties); -#ifdef __cplusplus -} -#endif - -#endif /* GRPC_CORE_EXT_CENSUS_GRPC_FILTER_H */ +#endif /* GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_BACKUP_POLLER_H */ diff --git a/src/core/ext/filters/client_channel/client_channel.cc b/src/core/ext/filters/client_channel/client_channel.cc index 22c2bc8880..00c51ba543 100644 --- a/src/core/ext/filters/client_channel/client_channel.cc +++ b/src/core/ext/filters/client_channel/client_channel.cc @@ -31,6 +31,7 @@ #include <grpc/support/sync.h> #include <grpc/support/useful.h> +#include "src/core/ext/filters/client_channel/backup_poller.h" #include "src/core/ext/filters/client_channel/http_connect_handshaker.h" #include "src/core/ext/filters/client_channel/lb_policy_registry.h" #include "src/core/ext/filters/client_channel/proxy_mapper_registry.h" @@ -712,6 +713,7 @@ static grpc_error *cc_init_channel_elem(grpc_exec_ctx *exec_ctx, chand->interested_parties = grpc_pollset_set_create(); grpc_connectivity_state_init(&chand->state_tracker, GRPC_CHANNEL_IDLE, "client_channel"); + grpc_client_channel_start_backup_polling(exec_ctx, chand->interested_parties); // Record client channel factory. const grpc_arg *arg = grpc_channel_args_find(args->channel_args, GRPC_ARG_CLIENT_CHANNEL_FACTORY); @@ -790,6 +792,7 @@ static void cc_destroy_channel_elem(grpc_exec_ctx *exec_ctx, if (chand->method_params_table != NULL) { grpc_slice_hash_table_unref(exec_ctx, chand->method_params_table); } + grpc_client_channel_stop_backup_polling(exec_ctx, chand->interested_parties); grpc_connectivity_state_destroy(exec_ctx, &chand->state_tracker); grpc_pollset_set_destroy(exec_ctx, chand->interested_parties); GRPC_COMBINER_UNREF(exec_ctx, chand->combiner, "client_channel"); @@ -898,7 +901,7 @@ static void waiting_for_pick_batches_fail(grpc_exec_ctx *exec_ctx, call_data *calld = (call_data *)elem->call_data; if (GRPC_TRACER_ON(grpc_client_channel_trace)) { gpr_log(GPR_DEBUG, - "chand=%p calld=%p: failing %" PRIdPTR " pending batches: %s", + "chand=%p calld=%p: failing %" PRIuPTR " pending batches: %s", elem->channel_data, calld, calld->waiting_for_pick_batches_count, grpc_error_string(error)); } @@ -940,7 +943,7 @@ static void waiting_for_pick_batches_resume(grpc_exec_ctx *exec_ctx, channel_data *chand = (channel_data *)elem->channel_data; call_data *calld = (call_data *)elem->call_data; if (GRPC_TRACER_ON(grpc_client_channel_trace)) { - gpr_log(GPR_DEBUG, "chand=%p calld=%p: sending %" PRIdPTR + gpr_log(GPR_DEBUG, "chand=%p calld=%p: sending %" PRIuPTR " pending batches to subchannel_call=%p", chand, calld, calld->waiting_for_pick_batches_count, calld->subchannel_call); @@ -1205,6 +1208,9 @@ static void pick_after_resolver_result_cancel_locked(grpc_exec_ctx *exec_ctx, "Pick cancelled", &error, 1)); } +static void pick_after_resolver_result_start_locked(grpc_exec_ctx *exec_ctx, + grpc_call_element *elem); + static void pick_after_resolver_result_done_locked(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { @@ -1228,7 +1234,7 @@ static void pick_after_resolver_result_done_locked(grpc_exec_ctx *exec_ctx, chand, calld); } async_pick_done_locked(exec_ctx, elem, GRPC_ERROR_REF(error)); - } else { + } else if (chand->lb_policy != NULL) { if (GRPC_TRACER_ON(grpc_client_channel_trace)) { gpr_log(GPR_DEBUG, "chand=%p calld=%p: resolver returned, doing pick", chand, calld); @@ -1242,6 +1248,30 @@ static void pick_after_resolver_result_done_locked(grpc_exec_ctx *exec_ctx, async_pick_done_locked(exec_ctx, elem, GRPC_ERROR_NONE); } } + // TODO(roth): It should be impossible for chand->lb_policy to be NULL + // here, so the rest of this code should never actually be executed. + // However, we have reports of a crash on iOS that triggers this case, + // so we are temporarily adding this to restore branches that were + // removed in https://github.com/grpc/grpc/pull/12297. Need to figure + // out what is actually causing this to occur and then figure out the + // right way to deal with it. + else if (chand->resolver != NULL) { + // No LB policy, so try again. + if (GRPC_TRACER_ON(grpc_client_channel_trace)) { + gpr_log(GPR_DEBUG, + "chand=%p calld=%p: resolver returned but no LB policy, " + "trying again", + chand, calld); + } + pick_after_resolver_result_start_locked(exec_ctx, elem); + } else { + if (GRPC_TRACER_ON(grpc_client_channel_trace)) { + gpr_log(GPR_DEBUG, "chand=%p calld=%p: resolver disconnected", chand, + calld); + } + async_pick_done_locked( + exec_ctx, elem, GRPC_ERROR_CREATE_FROM_STATIC_STRING("Disconnected")); + } } static void pick_after_resolver_result_start_locked(grpc_exec_ctx *exec_ctx, @@ -1599,8 +1629,8 @@ int grpc_client_channel_num_external_connectivity_watchers( return count; } -static void on_external_watch_complete(grpc_exec_ctx *exec_ctx, void *arg, - grpc_error *error) { +static void on_external_watch_complete_locked(grpc_exec_ctx *exec_ctx, + void *arg, grpc_error *error) { external_connectivity_watcher *w = (external_connectivity_watcher *)arg; grpc_closure *follow_up = w->on_complete; grpc_polling_entity_del_from_pollset_set(exec_ctx, &w->pollent, @@ -1619,8 +1649,8 @@ static void watch_connectivity_state_locked(grpc_exec_ctx *exec_ctx, void *arg, if (w->state != NULL) { external_connectivity_watcher_list_append(w->chand, w); GRPC_CLOSURE_RUN(exec_ctx, w->watcher_timer_init, GRPC_ERROR_NONE); - GRPC_CLOSURE_INIT(&w->my_closure, on_external_watch_complete, w, - grpc_schedule_on_exec_ctx); + GRPC_CLOSURE_INIT(&w->my_closure, on_external_watch_complete_locked, w, + grpc_combiner_scheduler(w->chand->combiner)); grpc_connectivity_state_notify_on_state_change( exec_ctx, &w->chand->state_tracker, w->state, &w->my_closure); } else { diff --git a/src/core/ext/filters/client_channel/lb_policy/grpclb/grpclb.cc b/src/core/ext/filters/client_channel/lb_policy/grpclb/grpclb.cc index ffd58129c6..85e76e68b5 100644 --- a/src/core/ext/filters/client_channel/lb_policy/grpclb/grpclb.cc +++ b/src/core/ext/filters/client_channel/lb_policy/grpclb/grpclb.cc @@ -611,7 +611,6 @@ static void update_lb_connectivity_status_locked( case GRPC_CHANNEL_SHUTDOWN: GPR_ASSERT(rr_state_error != GRPC_ERROR_NONE); break; - case GRPC_CHANNEL_INIT: case GRPC_CHANNEL_IDLE: case GRPC_CHANNEL_CONNECTING: case GRPC_CHANNEL_READY: @@ -1027,15 +1026,19 @@ static void glb_shutdown_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { while (pp != NULL) { pending_pick *next = pp->next; *pp->target = NULL; - GRPC_CLOSURE_SCHED(exec_ctx, &pp->wrapped_on_complete_arg.wrapper_closure, - GRPC_ERROR_NONE); + GRPC_CLOSURE_SCHED( + exec_ctx, &pp->wrapped_on_complete_arg.wrapper_closure, + GRPC_ERROR_CREATE_FROM_STATIC_STRING("Channel Shutdown")); + gpr_free(pp); pp = next; } while (pping != NULL) { pending_ping *next = pping->next; - GRPC_CLOSURE_SCHED(exec_ctx, &pping->wrapped_notify_arg.wrapper_closure, - GRPC_ERROR_NONE); + GRPC_CLOSURE_SCHED( + exec_ctx, &pping->wrapped_notify_arg.wrapper_closure, + GRPC_ERROR_CREATE_FROM_STATIC_STRING("Channel Shutdown")); + gpr_free(pping); pping = next; } } @@ -1786,7 +1789,6 @@ static void glb_lb_channel_on_connectivity_changed_cb(grpc_exec_ctx *exec_ctx, // embedded RR policy. Note that the current RR policy, if any, will stay in // effect until an update from the new lb_call is received. switch (glb_policy->lb_channel_connectivity) { - case GRPC_CHANNEL_INIT: case GRPC_CHANNEL_CONNECTING: case GRPC_CHANNEL_TRANSIENT_FAILURE: { /* resub. */ @@ -1803,9 +1805,8 @@ static void glb_lb_channel_on_connectivity_changed_cb(grpc_exec_ctx *exec_ctx, break; } case GRPC_CHANNEL_IDLE: - // lb channel inactive (probably shutdown prior to update). Restart lb - // call to kick the lb channel into gear. - GPR_ASSERT(glb_policy->lb_call == NULL); + // lb channel inactive (probably shutdown prior to update). Restart lb + // call to kick the lb channel into gear. /* fallthrough */ case GRPC_CHANNEL_READY: if (glb_policy->lb_call != NULL) { diff --git a/src/core/ext/filters/client_channel/lb_policy/pick_first/pick_first.cc b/src/core/ext/filters/client_channel/lb_policy/pick_first/pick_first.cc index b07fc3b720..f0c66c68e1 100644 --- a/src/core/ext/filters/client_channel/lb_policy/pick_first/pick_first.cc +++ b/src/core/ext/filters/client_channel/lb_policy/pick_first/pick_first.cc @@ -20,6 +20,7 @@ #include <grpc/support/alloc.h> +#include "src/core/ext/filters/client_channel/lb_policy/subchannel_list.h" #include "src/core/ext/filters/client_channel/lb_policy_registry.h" #include "src/core/ext/filters/client_channel/subchannel.h" #include "src/core/ext/filters/client_channel/subchannel_index.h" @@ -42,99 +43,73 @@ typedef struct { /** base policy: must be first */ grpc_lb_policy base; /** all our subchannels */ - grpc_subchannel **subchannels; - grpc_subchannel **new_subchannels; - size_t num_subchannels; - size_t num_new_subchannels; - - grpc_closure connectivity_changed; - - /** remaining members are protected by the combiner */ - - /** the selected channel */ - grpc_connected_subchannel *selected; - - /** the subchannel key for \a selected, or NULL if \a selected not set */ - const grpc_subchannel_key *selected_key; - + grpc_lb_subchannel_list *subchannel_list; + /** latest pending subchannel list */ + grpc_lb_subchannel_list *latest_pending_subchannel_list; + /** selected subchannel in \a subchannel_list */ + grpc_lb_subchannel_data *selected; /** have we started picking? */ bool started_picking; /** are we shut down? */ bool shutdown; - /** are we updating the selected subchannel? */ - bool updating_selected; - /** are we updating the subchannel candidates? */ - bool updating_subchannels; - /** args from the latest update received while already updating, or NULL */ - grpc_lb_policy_args *pending_update_args; - /** which subchannel are we watching? */ - size_t checking_subchannel; - /** what is the connectivity of that channel? */ - grpc_connectivity_state checking_connectivity; /** list of picks that are waiting on connectivity */ pending_pick *pending_picks; - /** our connectivity state tracker */ grpc_connectivity_state_tracker state_tracker; } pick_first_lb_policy; static void pf_destroy(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { pick_first_lb_policy *p = (pick_first_lb_policy *)pol; + GPR_ASSERT(p->subchannel_list == NULL); + GPR_ASSERT(p->latest_pending_subchannel_list == NULL); GPR_ASSERT(p->pending_picks == NULL); - for (size_t i = 0; i < p->num_subchannels; i++) { - GRPC_SUBCHANNEL_UNREF(exec_ctx, p->subchannels[i], "pick_first_destroy"); - } - if (p->selected != NULL) { - GRPC_CONNECTED_SUBCHANNEL_UNREF(exec_ctx, p->selected, - "picked_first_destroy"); - } grpc_connectivity_state_destroy(exec_ctx, &p->state_tracker); - grpc_subchannel_index_unref(); - if (p->pending_update_args != NULL) { - grpc_channel_args_destroy(exec_ctx, p->pending_update_args->args); - gpr_free(p->pending_update_args); - } - gpr_free(p->subchannels); - gpr_free(p->new_subchannels); gpr_free(p); + grpc_subchannel_index_unref(); if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { gpr_log(GPR_DEBUG, "Pick First %p destroyed.", (void *)p); } } -static void pf_shutdown_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { - pick_first_lb_policy *p = (pick_first_lb_policy *)pol; - pending_pick *pp; - p->shutdown = true; - pp = p->pending_picks; - p->pending_picks = NULL; - grpc_connectivity_state_set( - exec_ctx, &p->state_tracker, GRPC_CHANNEL_SHUTDOWN, - GRPC_ERROR_CREATE_FROM_STATIC_STRING("Channel shutdown"), "shutdown"); - /* cancel subscription */ - if (p->selected != NULL) { - grpc_connected_subchannel_notify_on_state_change( - exec_ctx, p->selected, NULL, NULL, &p->connectivity_changed); - } else if (p->num_subchannels > 0 && p->started_picking) { - grpc_subchannel_notify_on_state_change( - exec_ctx, p->subchannels[p->checking_subchannel], NULL, NULL, - &p->connectivity_changed); +static void shutdown_locked(grpc_exec_ctx *exec_ctx, pick_first_lb_policy *p, + grpc_error *error) { + if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { + gpr_log(GPR_DEBUG, "Pick First %p Shutting down", p); } - while (pp != NULL) { - pending_pick *next = pp->next; + p->shutdown = true; + pending_pick *pp; + while ((pp = p->pending_picks) != NULL) { + p->pending_picks = pp->next; *pp->target = NULL; - GRPC_CLOSURE_SCHED(exec_ctx, pp->on_complete, GRPC_ERROR_NONE); + GRPC_CLOSURE_SCHED(exec_ctx, pp->on_complete, GRPC_ERROR_REF(error)); gpr_free(pp); - pp = next; } + grpc_connectivity_state_set(exec_ctx, &p->state_tracker, + GRPC_CHANNEL_SHUTDOWN, GRPC_ERROR_REF(error), + "shutdown"); + if (p->subchannel_list != NULL) { + grpc_lb_subchannel_list_shutdown_and_unref(exec_ctx, p->subchannel_list, + "pf_shutdown"); + p->subchannel_list = NULL; + } + if (p->latest_pending_subchannel_list != NULL) { + grpc_lb_subchannel_list_shutdown_and_unref( + exec_ctx, p->latest_pending_subchannel_list, "pf_shutdown"); + p->latest_pending_subchannel_list = NULL; + } + GRPC_ERROR_UNREF(error); +} + +static void pf_shutdown_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { + shutdown_locked(exec_ctx, (pick_first_lb_policy *)pol, + GRPC_ERROR_CREATE_FROM_STATIC_STRING("Channel shutdown")); } static void pf_cancel_pick_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, grpc_connected_subchannel **target, grpc_error *error) { pick_first_lb_policy *p = (pick_first_lb_policy *)pol; - pending_pick *pp; - pp = p->pending_picks; + pending_pick *pp = p->pending_picks; p->pending_picks = NULL; while (pp != NULL) { pending_pick *next = pp->next; @@ -158,8 +133,7 @@ static void pf_cancel_picks_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, uint32_t initial_metadata_flags_eq, grpc_error *error) { pick_first_lb_policy *p = (pick_first_lb_policy *)pol; - pending_pick *pp; - pp = p->pending_picks; + pending_pick *pp = p->pending_picks; p->pending_picks = NULL; while (pp != NULL) { pending_pick *next = pp->next; @@ -181,15 +155,12 @@ static void pf_cancel_picks_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, static void start_picking_locked(grpc_exec_ctx *exec_ctx, pick_first_lb_policy *p) { p->started_picking = true; - if (p->subchannels != NULL) { - GPR_ASSERT(p->num_subchannels > 0); - p->checking_subchannel = 0; - p->checking_connectivity = GRPC_CHANNEL_IDLE; - GRPC_LB_POLICY_WEAK_REF(&p->base, "pick_first_connectivity"); - grpc_subchannel_notify_on_state_change( - exec_ctx, p->subchannels[p->checking_subchannel], - p->base.interested_parties, &p->checking_connectivity, - &p->connectivity_changed); + if (p->subchannel_list != NULL && p->subchannel_list->num_subchannels > 0) { + p->subchannel_list->checking_subchannel = 0; + grpc_lb_subchannel_list_ref_for_connectivity_watch( + p->subchannel_list, "connectivity_watch+start_picking"); + grpc_lb_subchannel_data_start_connectivity_watch( + exec_ctx, &p->subchannel_list->subchannels[0]); } } @@ -206,19 +177,17 @@ static int pf_pick_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, grpc_call_context_element *context, void **user_data, grpc_closure *on_complete) { pick_first_lb_policy *p = (pick_first_lb_policy *)pol; - pending_pick *pp; - - /* Check atomically for a selected channel */ + // If we have a selected subchannel already, return synchronously. if (p->selected != NULL) { - *target = GRPC_CONNECTED_SUBCHANNEL_REF(p->selected, "picked"); + *target = GRPC_CONNECTED_SUBCHANNEL_REF(p->selected->connected_subchannel, + "picked"); return 1; } - - /* No subchannel selected yet, so try again */ + // No subchannel selected yet, so handle asynchronously. if (!p->started_picking) { start_picking_locked(exec_ctx, p); } - pp = (pending_pick *)gpr_malloc(sizeof(*pp)); + pending_pick *pp = (pending_pick *)gpr_malloc(sizeof(*pp)); pp->next = p->pending_picks; pp->target = target; pp->initial_metadata_flags = pick_args->initial_metadata_flags; @@ -227,19 +196,15 @@ static int pf_pick_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, return 0; } -static void destroy_subchannels_locked(grpc_exec_ctx *exec_ctx, - pick_first_lb_policy *p) { - size_t num_subchannels = p->num_subchannels; - grpc_subchannel **subchannels = p->subchannels; - - p->num_subchannels = 0; - p->subchannels = NULL; - GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &p->base, "destroy_subchannels"); - - for (size_t i = 0; i < num_subchannels; i++) { - GRPC_SUBCHANNEL_UNREF(exec_ctx, subchannels[i], "pick_first"); +static void destroy_unselected_subchannels_locked(grpc_exec_ctx *exec_ctx, + pick_first_lb_policy *p) { + for (size_t i = 0; i < p->subchannel_list->num_subchannels; ++i) { + grpc_lb_subchannel_data *sd = &p->subchannel_list->subchannels[i]; + if (p->selected != sd) { + grpc_lb_subchannel_data_unref_subchannel(exec_ctx, sd, + "selected_different_subchannel"); + } } - gpr_free(subchannels); } static grpc_connectivity_state pf_check_connectivity_locked( @@ -261,46 +226,24 @@ static void pf_ping_one_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, grpc_closure *closure) { pick_first_lb_policy *p = (pick_first_lb_policy *)pol; if (p->selected) { - grpc_connected_subchannel_ping(exec_ctx, p->selected, closure); + grpc_connected_subchannel_ping(exec_ctx, p->selected->connected_subchannel, + closure); } else { GRPC_CLOSURE_SCHED(exec_ctx, closure, GRPC_ERROR_CREATE_FROM_STATIC_STRING("Not connected")); } } -/* unsubscribe all subchannels */ -static void stop_connectivity_watchers(grpc_exec_ctx *exec_ctx, - pick_first_lb_policy *p) { - if (p->num_subchannels > 0) { - GPR_ASSERT(p->selected == NULL); - if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { - gpr_log(GPR_DEBUG, "Pick First %p unsubscribing from subchannel %p", - (void *)p, (void *)p->subchannels[p->checking_subchannel]); - } - grpc_subchannel_notify_on_state_change( - exec_ctx, p->subchannels[p->checking_subchannel], NULL, NULL, - &p->connectivity_changed); - p->updating_subchannels = true; - } else if (p->selected != NULL) { - if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { - gpr_log(GPR_DEBUG, - "Pick First %p unsubscribing from selected subchannel %p", - (void *)p, (void *)p->selected); - } - grpc_connected_subchannel_notify_on_state_change( - exec_ctx, p->selected, NULL, NULL, &p->connectivity_changed); - p->updating_selected = true; - } -} +static void pf_connectivity_changed_locked(grpc_exec_ctx *exec_ctx, void *arg, + grpc_error *error); -/* true upon success */ static void pf_update_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *policy, const grpc_lb_policy_args *args) { pick_first_lb_policy *p = (pick_first_lb_policy *)policy; const grpc_arg *arg = grpc_channel_args_find(args->args, GRPC_ARG_LB_ADDRESSES); if (arg == NULL || arg->type != GRPC_ARG_POINTER) { - if (p->subchannels == NULL) { + if (p->subchannel_list == NULL) { // If we don't have a current subchannel list, go into TRANSIENT FAILURE. grpc_connectivity_state_set( exec_ctx, &p->state_tracker, GRPC_CHANNEL_TRANSIENT_FAILURE, @@ -317,270 +260,222 @@ static void pf_update_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *policy, } const grpc_lb_addresses *addresses = (const grpc_lb_addresses *)arg->value.pointer.p; - if (addresses->num_addresses == 0) { - // Empty update. Unsubscribe from all current subchannels and put the - // channel in TRANSIENT_FAILURE. + if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { + gpr_log(GPR_INFO, "Pick First %p received update with %lu addresses", + (void *)p, (unsigned long)addresses->num_addresses); + } + grpc_lb_subchannel_list *subchannel_list = grpc_lb_subchannel_list_create( + exec_ctx, &p->base, &grpc_lb_pick_first_trace, addresses, args, + pf_connectivity_changed_locked); + if (subchannel_list->num_subchannels == 0) { + // Empty update or no valid subchannels. Unsubscribe from all current + // subchannels and put the channel in TRANSIENT_FAILURE. grpc_connectivity_state_set( exec_ctx, &p->state_tracker, GRPC_CHANNEL_TRANSIENT_FAILURE, GRPC_ERROR_CREATE_FROM_STATIC_STRING("Empty update"), "pf_update_empty"); - stop_connectivity_watchers(exec_ctx, p); + if (p->subchannel_list != NULL) { + grpc_lb_subchannel_list_shutdown_and_unref(exec_ctx, p->subchannel_list, + "sl_shutdown_empty_update"); + } + p->subchannel_list = subchannel_list; // Empty list. + p->selected = NULL; return; } - if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { - gpr_log(GPR_INFO, "Pick First %p received update with %lu addresses", - (void *)p, (unsigned long)addresses->num_addresses); - } - grpc_subchannel_args *sc_args = (grpc_subchannel_args *)gpr_zalloc( - sizeof(*sc_args) * addresses->num_addresses); - /* We remove the following keys in order for subchannel keys belonging to - * subchannels point to the same address to match. */ - static const char *keys_to_remove[] = {GRPC_ARG_SUBCHANNEL_ADDRESS, - GRPC_ARG_LB_ADDRESSES}; - size_t sc_args_count = 0; - - /* Create list of subchannel args for new addresses in \a args. */ - for (size_t i = 0; i < addresses->num_addresses; i++) { - // If there were any balancer, we would have chosen grpclb policy instead. - GPR_ASSERT(!addresses->addresses[i].is_balancer); - if (addresses->addresses[i].user_data != NULL) { - gpr_log(GPR_ERROR, - "This LB policy doesn't support user data. It will be ignored"); + if (p->selected == NULL) { + // We don't yet have a selected subchannel, so replace the current + // subchannel list immediately. + if (p->subchannel_list != NULL) { + grpc_lb_subchannel_list_shutdown_and_unref(exec_ctx, p->subchannel_list, + "pf_update_before_selected"); } - grpc_arg addr_arg = - grpc_create_subchannel_address_arg(&addresses->addresses[i].address); - grpc_channel_args *new_args = grpc_channel_args_copy_and_add_and_remove( - args->args, keys_to_remove, GPR_ARRAY_SIZE(keys_to_remove), &addr_arg, - 1); - gpr_free(addr_arg.value.string); - sc_args[sc_args_count++].args = new_args; - } - - /* Check if p->selected is amongst them. If so, we are done. */ - if (p->selected != NULL) { - GPR_ASSERT(p->selected_key != NULL); - for (size_t i = 0; i < sc_args_count; i++) { - grpc_subchannel_key *ith_sc_key = grpc_subchannel_key_create(&sc_args[i]); - const bool found_selected = - grpc_subchannel_key_compare(p->selected_key, ith_sc_key) == 0; - grpc_subchannel_key_destroy(exec_ctx, ith_sc_key); - if (found_selected) { + p->subchannel_list = subchannel_list; + } else { + // We do have a selected subchannel. + // Check if it's present in the new list. If so, we're done. + for (size_t i = 0; i < subchannel_list->num_subchannels; ++i) { + grpc_lb_subchannel_data *sd = &subchannel_list->subchannels[i]; + if (sd->subchannel == p->selected->subchannel) { // The currently selected subchannel is in the update: we are done. if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { gpr_log(GPR_INFO, - "Pick First %p found already selected subchannel %p amongst " - "updates. Update done.", - (void *)p, (void *)p->selected); + "Pick First %p found already selected subchannel %p " + "at update index %" PRIuPTR " of %" PRIuPTR "; update done", + p, p->selected->subchannel, i, + subchannel_list->num_subchannels); } - for (size_t j = 0; j < sc_args_count; j++) { - grpc_channel_args_destroy(exec_ctx, - (grpc_channel_args *)sc_args[j].args); + grpc_lb_subchannel_list_ref_for_connectivity_watch( + subchannel_list, "connectivity_watch+replace_selected"); + grpc_lb_subchannel_data_start_connectivity_watch(exec_ctx, sd); + if (p->subchannel_list != NULL) { + grpc_lb_subchannel_list_shutdown_and_unref( + exec_ctx, p->subchannel_list, "pf_update_includes_selected"); + } + p->subchannel_list = subchannel_list; + if (p->selected->connected_subchannel != NULL) { + sd->connected_subchannel = GRPC_CONNECTED_SUBCHANNEL_REF( + p->selected->connected_subchannel, "pf_update_includes_selected"); + } + p->selected = sd; + destroy_unselected_subchannels_locked(exec_ctx, p); + // If there was a previously pending update (which may or may + // not have contained the currently selected subchannel), drop + // it, so that it doesn't override what we've done here. + if (p->latest_pending_subchannel_list != NULL) { + grpc_lb_subchannel_list_shutdown_and_unref( + exec_ctx, p->latest_pending_subchannel_list, + "pf_update_includes_selected+outdated"); + p->latest_pending_subchannel_list = NULL; } - gpr_free(sc_args); return; } } - } - // We only check for already running updates here because if the previous - // steps were successful, the update can be considered done without any - // interference (ie, no callbacks were scheduled). - if (p->updating_selected || p->updating_subchannels) { - if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { - gpr_log(GPR_INFO, - "Update already in progress for pick first %p. Deferring update.", - (void *)p); - } - if (p->pending_update_args != NULL) { - grpc_channel_args_destroy(exec_ctx, p->pending_update_args->args); - gpr_free(p->pending_update_args); - } - p->pending_update_args = - (grpc_lb_policy_args *)gpr_zalloc(sizeof(*p->pending_update_args)); - p->pending_update_args->client_channel_factory = - args->client_channel_factory; - p->pending_update_args->args = grpc_channel_args_copy(args->args); - p->pending_update_args->combiner = args->combiner; - return; - } - /* Create the subchannels for the new subchannel args/addresses. */ - grpc_subchannel **new_subchannels = - (grpc_subchannel **)gpr_zalloc(sizeof(*new_subchannels) * sc_args_count); - size_t num_new_subchannels = 0; - for (size_t i = 0; i < sc_args_count; i++) { - grpc_subchannel *subchannel = grpc_client_channel_factory_create_subchannel( - exec_ctx, args->client_channel_factory, &sc_args[i]); - if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { - char *address_uri = - grpc_sockaddr_to_uri(&addresses->addresses[i].address); - gpr_log(GPR_INFO, - "Pick First %p created subchannel %p for address uri %s", - (void *)p, (void *)subchannel, address_uri); - gpr_free(address_uri); + // Not keeping the previous selected subchannel, so set the latest + // pending subchannel list to the new subchannel list. We will wait + // for it to report READY before swapping it into the current + // subchannel list. + if (p->latest_pending_subchannel_list != NULL) { + if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { + gpr_log(GPR_DEBUG, + "Pick First %p Shutting down latest pending subchannel list " + "%p, about to be replaced by newer latest %p", + (void *)p, (void *)p->latest_pending_subchannel_list, + (void *)subchannel_list); + } + grpc_lb_subchannel_list_shutdown_and_unref( + exec_ctx, p->latest_pending_subchannel_list, + "sl_outdated_dont_smash"); } - grpc_channel_args_destroy(exec_ctx, (grpc_channel_args *)sc_args[i].args); - if (subchannel != NULL) new_subchannels[num_new_subchannels++] = subchannel; + p->latest_pending_subchannel_list = subchannel_list; } - gpr_free(sc_args); - if (num_new_subchannels == 0) { - gpr_free(new_subchannels); - // Empty update. Unsubscribe from all current subchannels and put the - // channel in TRANSIENT_FAILURE. - grpc_connectivity_state_set( - exec_ctx, &p->state_tracker, GRPC_CHANNEL_TRANSIENT_FAILURE, - GRPC_ERROR_CREATE_FROM_STATIC_STRING("No valid addresses in update"), - "pf_update_no_valid_addresses"); - stop_connectivity_watchers(exec_ctx, p); - return; - } - - /* Destroy the current subchannels. Repurpose pf_shutdown/destroy. */ - stop_connectivity_watchers(exec_ctx, p); - - /* Save new subchannels. The switch over will happen in - * pf_connectivity_changed_locked */ - if (p->updating_selected || p->updating_subchannels) { - p->num_new_subchannels = num_new_subchannels; - p->new_subchannels = new_subchannels; - } else { /* nothing is updating. Get things moving from here */ - p->num_subchannels = num_new_subchannels; - p->subchannels = new_subchannels; - p->new_subchannels = NULL; - p->num_new_subchannels = 0; - if (p->started_picking) { - p->checking_subchannel = 0; - p->checking_connectivity = GRPC_CHANNEL_IDLE; - grpc_subchannel_notify_on_state_change( - exec_ctx, p->subchannels[p->checking_subchannel], - p->base.interested_parties, &p->checking_connectivity, - &p->connectivity_changed); - } + // If we've started picking, start trying to connect to the first + // subchannel in the new list. + if (p->started_picking) { + grpc_lb_subchannel_list_ref_for_connectivity_watch( + subchannel_list, "connectivity_watch+update"); + grpc_lb_subchannel_data_start_connectivity_watch( + exec_ctx, &subchannel_list->subchannels[0]); } } static void pf_connectivity_changed_locked(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { - pick_first_lb_policy *p = (pick_first_lb_policy *)arg; - grpc_subchannel *selected_subchannel; - pending_pick *pp; - + grpc_lb_subchannel_data *sd = (grpc_lb_subchannel_data *)arg; + pick_first_lb_policy *p = (pick_first_lb_policy *)sd->subchannel_list->policy; if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { - gpr_log( - GPR_DEBUG, - "Pick First %p connectivity changed. Updating selected: %d; Updating " - "subchannels: %d; Checking %lu index (%lu total); State: %d; ", - (void *)p, p->updating_selected, p->updating_subchannels, - (unsigned long)p->checking_subchannel, - (unsigned long)p->num_subchannels, p->checking_connectivity); - } - bool restart = false; - if (p->updating_selected && error != GRPC_ERROR_NONE) { - /* Captured the unsubscription for p->selected */ - GPR_ASSERT(p->selected != NULL); - GRPC_CONNECTED_SUBCHANNEL_UNREF(exec_ctx, p->selected, - "pf_update_connectivity"); - if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { - gpr_log(GPR_DEBUG, "Pick First %p unreffing selected subchannel %p", - (void *)p, (void *)p->selected); - } - p->updating_selected = false; - if (p->num_new_subchannels == 0) { - p->selected = NULL; - return; - } - restart = true; + gpr_log(GPR_DEBUG, + "Pick First %p connectivity changed for subchannel %p (%" PRIuPTR + " of %" PRIuPTR + "), subchannel_list %p: state=%s p->shutdown=%d " + "sd->subchannel_list->shutting_down=%d error=%s", + (void *)p, (void *)sd->subchannel, + sd->subchannel_list->checking_subchannel, + sd->subchannel_list->num_subchannels, (void *)sd->subchannel_list, + grpc_connectivity_state_name(sd->pending_connectivity_state_unsafe), + p->shutdown, sd->subchannel_list->shutting_down, + grpc_error_string(error)); } - if (p->updating_subchannels && error != GRPC_ERROR_NONE) { - /* Captured the unsubscription for the checking subchannel */ - GPR_ASSERT(p->selected == NULL); - for (size_t i = 0; i < p->num_subchannels; i++) { - GRPC_SUBCHANNEL_UNREF(exec_ctx, p->subchannels[i], - "pf_update_connectivity"); - if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { - gpr_log(GPR_DEBUG, "Pick First %p unreffing subchannel %p", (void *)p, - (void *)p->subchannels[i]); - } - } - gpr_free(p->subchannels); - p->subchannels = NULL; - p->num_subchannels = 0; - p->updating_subchannels = false; - if (p->num_new_subchannels == 0) return; - restart = true; - } - if (restart) { - p->selected = NULL; - p->selected_key = NULL; - GPR_ASSERT(p->new_subchannels != NULL); - GPR_ASSERT(p->num_new_subchannels > 0); - p->num_subchannels = p->num_new_subchannels; - p->subchannels = p->new_subchannels; - p->num_new_subchannels = 0; - p->new_subchannels = NULL; - if (p->started_picking) { - /* If we were picking, continue to do so over the new subchannels, - * starting from the 0th index. */ - p->checking_subchannel = 0; - p->checking_connectivity = GRPC_CHANNEL_IDLE; - /* reuses the weak ref from start_picking_locked */ - grpc_subchannel_notify_on_state_change( - exec_ctx, p->subchannels[p->checking_subchannel], - p->base.interested_parties, &p->checking_connectivity, - &p->connectivity_changed); - } - if (p->pending_update_args != NULL) { - const grpc_lb_policy_args *args = p->pending_update_args; - p->pending_update_args = NULL; - pf_update_locked(exec_ctx, &p->base, args); - } + // If the policy is shutting down, unref and return. + if (p->shutdown) { + grpc_lb_subchannel_data_stop_connectivity_watch(exec_ctx, sd); + grpc_lb_subchannel_data_unref_subchannel(exec_ctx, sd, "pf_shutdown"); + grpc_lb_subchannel_list_unref_for_connectivity_watch( + exec_ctx, sd->subchannel_list, "pf_shutdown"); return; } - GRPC_ERROR_REF(error); - if (p->shutdown) { - GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &p->base, "pick_first_connectivity"); - GRPC_ERROR_UNREF(error); + // If the subchannel list is shutting down, stop watching. + if (sd->subchannel_list->shutting_down || error == GRPC_ERROR_CANCELLED) { + grpc_lb_subchannel_data_stop_connectivity_watch(exec_ctx, sd); + grpc_lb_subchannel_data_unref_subchannel(exec_ctx, sd, "pf_sl_shutdown"); + grpc_lb_subchannel_list_unref_for_connectivity_watch( + exec_ctx, sd->subchannel_list, "pf_sl_shutdown"); return; - } else if (p->selected != NULL) { - if (p->checking_connectivity == GRPC_CHANNEL_TRANSIENT_FAILURE) { - /* if the selected channel goes bad, we're done */ - p->checking_connectivity = GRPC_CHANNEL_SHUTDOWN; - } - grpc_connectivity_state_set(exec_ctx, &p->state_tracker, - p->checking_connectivity, GRPC_ERROR_REF(error), - "selected_changed"); - if (p->checking_connectivity != GRPC_CHANNEL_SHUTDOWN) { - grpc_connected_subchannel_notify_on_state_change( - exec_ctx, p->selected, p->base.interested_parties, - &p->checking_connectivity, &p->connectivity_changed); + } + // If we're still here, the notification must be for a subchannel in + // either the current or latest pending subchannel lists. + GPR_ASSERT(sd->subchannel_list == p->subchannel_list || + sd->subchannel_list == p->latest_pending_subchannel_list); + // Update state. + sd->curr_connectivity_state = sd->pending_connectivity_state_unsafe; + // Handle updates for the currently selected subchannel. + if (p->selected == sd) { + // If the new state is anything other than READY and there is a + // pending update, switch to the pending update. + if (sd->curr_connectivity_state != GRPC_CHANNEL_READY && + p->latest_pending_subchannel_list != NULL) { + p->selected = NULL; + grpc_lb_subchannel_list_shutdown_and_unref( + exec_ctx, p->subchannel_list, "selected_not_ready+switch_to_update"); + p->subchannel_list = p->latest_pending_subchannel_list; + p->latest_pending_subchannel_list = NULL; + grpc_connectivity_state_set( + exec_ctx, &p->state_tracker, GRPC_CHANNEL_TRANSIENT_FAILURE, + GRPC_ERROR_REF(error), "selected_not_ready+switch_to_update"); } else { - GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &p->base, "pick_first_connectivity"); + if (sd->curr_connectivity_state == GRPC_CHANNEL_TRANSIENT_FAILURE) { + /* if the selected channel goes bad, we're done */ + sd->curr_connectivity_state = GRPC_CHANNEL_SHUTDOWN; + } + grpc_connectivity_state_set(exec_ctx, &p->state_tracker, + sd->curr_connectivity_state, + GRPC_ERROR_REF(error), "selected_changed"); + if (sd->curr_connectivity_state != GRPC_CHANNEL_SHUTDOWN) { + // Renew notification. + grpc_lb_subchannel_data_start_connectivity_watch(exec_ctx, sd); + } else { + grpc_lb_subchannel_data_stop_connectivity_watch(exec_ctx, sd); + grpc_lb_subchannel_list_unref_for_connectivity_watch( + exec_ctx, sd->subchannel_list, "pf_selected_shutdown"); + shutdown_locked(exec_ctx, p, GRPC_ERROR_REF(error)); + } } - } else { - loop: - switch (p->checking_connectivity) { - case GRPC_CHANNEL_INIT: - GPR_UNREACHABLE_CODE(return ); - case GRPC_CHANNEL_READY: + return; + } + // If we get here, there are two possible cases: + // 1. We do not currently have a selected subchannel, and the update is + // for a subchannel in p->subchannel_list that we're trying to + // connect to. The goal here is to find a subchannel that we can + // select. + // 2. We do currently have a selected subchannel, and the update is + // for a subchannel in p->latest_pending_subchannel_list. The + // goal here is to find a subchannel from the update that we can + // select in place of the current one. + if (sd->curr_connectivity_state == GRPC_CHANNEL_TRANSIENT_FAILURE || + sd->curr_connectivity_state == GRPC_CHANNEL_SHUTDOWN) { + grpc_lb_subchannel_data_stop_connectivity_watch(exec_ctx, sd); + } + while (true) { + switch (sd->curr_connectivity_state) { + case GRPC_CHANNEL_READY: { + // Case 2. Promote p->latest_pending_subchannel_list to + // p->subchannel_list. + if (sd->subchannel_list == p->latest_pending_subchannel_list) { + GPR_ASSERT(p->subchannel_list != NULL); + grpc_lb_subchannel_list_shutdown_and_unref( + exec_ctx, p->subchannel_list, "finish_update"); + p->subchannel_list = p->latest_pending_subchannel_list; + p->latest_pending_subchannel_list = NULL; + } + // Cases 1 and 2. grpc_connectivity_state_set(exec_ctx, &p->state_tracker, GRPC_CHANNEL_READY, GRPC_ERROR_NONE, "connecting_ready"); - selected_subchannel = p->subchannels[p->checking_subchannel]; - p->selected = GRPC_CONNECTED_SUBCHANNEL_REF( - grpc_subchannel_get_connected_subchannel(selected_subchannel), - "picked_first"); - + sd->connected_subchannel = GRPC_CONNECTED_SUBCHANNEL_REF( + grpc_subchannel_get_connected_subchannel(sd->subchannel), + "connected"); + p->selected = sd; if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { - gpr_log(GPR_INFO, - "Pick First %p selected subchannel %p (connected %p)", - (void *)p, (void *)selected_subchannel, (void *)p->selected); + gpr_log(GPR_INFO, "Pick First %p selected subchannel %p", (void *)p, + (void *)sd->subchannel); } - p->selected_key = grpc_subchannel_get_key(selected_subchannel); - /* drop the pick list: we are connected now */ - GRPC_LB_POLICY_WEAK_REF(&p->base, "destroy_subchannels"); - destroy_subchannels_locked(exec_ctx, p); - /* update any calls that were waiting for a pick */ + // Drop all other subchannels, since we are now connected. + destroy_unselected_subchannels_locked(exec_ctx, p); + // Update any calls that were waiting for a pick. + pending_pick *pp; while ((pp = p->pending_picks)) { p->pending_picks = pp->next; - *pp->target = GRPC_CONNECTED_SUBCHANNEL_REF(p->selected, "picked"); + *pp->target = GRPC_CONNECTED_SUBCHANNEL_REF( + p->selected->connected_subchannel, "picked"); if (GRPC_TRACER_ON(grpc_lb_pick_first_trace)) { gpr_log(GPR_INFO, "Servicing pending pick with selected subchannel %p", @@ -589,76 +484,86 @@ static void pf_connectivity_changed_locked(grpc_exec_ctx *exec_ctx, void *arg, GRPC_CLOSURE_SCHED(exec_ctx, pp->on_complete, GRPC_ERROR_NONE); gpr_free(pp); } - grpc_connected_subchannel_notify_on_state_change( - exec_ctx, p->selected, p->base.interested_parties, - &p->checking_connectivity, &p->connectivity_changed); - break; - case GRPC_CHANNEL_TRANSIENT_FAILURE: - p->checking_subchannel = - (p->checking_subchannel + 1) % p->num_subchannels; - if (p->checking_subchannel == 0) { - /* only trigger transient failure when we've tried all alternatives - */ + // Renew notification. + grpc_lb_subchannel_data_start_connectivity_watch(exec_ctx, sd); + return; + } + case GRPC_CHANNEL_TRANSIENT_FAILURE: { + do { + sd->subchannel_list->checking_subchannel = + (sd->subchannel_list->checking_subchannel + 1) % + sd->subchannel_list->num_subchannels; + sd = &sd->subchannel_list + ->subchannels[sd->subchannel_list->checking_subchannel]; + } while (sd->subchannel == NULL); + // Case 1: Only set state to TRANSIENT_FAILURE if we've tried + // all subchannels. + if (sd->subchannel_list->checking_subchannel == 0 && + sd->subchannel_list == p->subchannel_list) { grpc_connectivity_state_set( exec_ctx, &p->state_tracker, GRPC_CHANNEL_TRANSIENT_FAILURE, GRPC_ERROR_REF(error), "connecting_transient_failure"); } + sd->curr_connectivity_state = + grpc_subchannel_check_connectivity(sd->subchannel, &error); GRPC_ERROR_UNREF(error); - p->checking_connectivity = grpc_subchannel_check_connectivity( - p->subchannels[p->checking_subchannel], &error); - if (p->checking_connectivity == GRPC_CHANNEL_TRANSIENT_FAILURE) { - grpc_subchannel_notify_on_state_change( - exec_ctx, p->subchannels[p->checking_subchannel], - p->base.interested_parties, &p->checking_connectivity, - &p->connectivity_changed); - } else { - goto loop; + if (sd->curr_connectivity_state == GRPC_CHANNEL_TRANSIENT_FAILURE) { + // Reuses the connectivity refs from the previous watch. + grpc_lb_subchannel_data_start_connectivity_watch(exec_ctx, sd); + return; } - break; + break; // Go back to top of loop. + } case GRPC_CHANNEL_CONNECTING: - case GRPC_CHANNEL_IDLE: - grpc_connectivity_state_set( - exec_ctx, &p->state_tracker, GRPC_CHANNEL_CONNECTING, - GRPC_ERROR_REF(error), "connecting_changed"); - grpc_subchannel_notify_on_state_change( - exec_ctx, p->subchannels[p->checking_subchannel], - p->base.interested_parties, &p->checking_connectivity, - &p->connectivity_changed); - break; - case GRPC_CHANNEL_SHUTDOWN: - p->num_subchannels--; - GPR_SWAP(grpc_subchannel *, p->subchannels[p->checking_subchannel], - p->subchannels[p->num_subchannels]); - GRPC_SUBCHANNEL_UNREF(exec_ctx, p->subchannels[p->num_subchannels], - "pick_first"); - if (p->num_subchannels == 0) { + case GRPC_CHANNEL_IDLE: { + // Only update connectivity state in case 1. + if (sd->subchannel_list == p->subchannel_list) { grpc_connectivity_state_set( - exec_ctx, &p->state_tracker, GRPC_CHANNEL_SHUTDOWN, - GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING( - "Pick first exhausted channels", &error, 1), - "no_more_channels"); - while ((pp = p->pending_picks)) { - p->pending_picks = pp->next; - *pp->target = NULL; - GRPC_CLOSURE_SCHED(exec_ctx, pp->on_complete, GRPC_ERROR_NONE); - gpr_free(pp); - } - GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &p->base, - "pick_first_connectivity"); - } else { + exec_ctx, &p->state_tracker, GRPC_CHANNEL_CONNECTING, + GRPC_ERROR_REF(error), "connecting_changed"); + } + // Renew notification. + grpc_lb_subchannel_data_start_connectivity_watch(exec_ctx, sd); + return; + } + case GRPC_CHANNEL_SHUTDOWN: { + grpc_lb_subchannel_data_unref_subchannel(exec_ctx, sd, + "pf_candidate_shutdown"); + // Advance to next subchannel and check its state. + grpc_lb_subchannel_data *original_sd = sd; + do { + sd->subchannel_list->checking_subchannel = + (sd->subchannel_list->checking_subchannel + 1) % + sd->subchannel_list->num_subchannels; + sd = &sd->subchannel_list + ->subchannels[sd->subchannel_list->checking_subchannel]; + } while (sd->subchannel == NULL && sd != original_sd); + if (sd == original_sd) { + grpc_lb_subchannel_list_unref_for_connectivity_watch( + exec_ctx, sd->subchannel_list, "pf_candidate_shutdown"); + shutdown_locked(exec_ctx, p, + GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING( + "Pick first exhausted channels", &error, 1)); + return; + } + if (sd->subchannel_list == p->subchannel_list) { grpc_connectivity_state_set( exec_ctx, &p->state_tracker, GRPC_CHANNEL_TRANSIENT_FAILURE, GRPC_ERROR_REF(error), "subchannel_failed"); - p->checking_subchannel %= p->num_subchannels; - GRPC_ERROR_UNREF(error); - p->checking_connectivity = grpc_subchannel_check_connectivity( - p->subchannels[p->checking_subchannel], &error); - goto loop; } + sd->curr_connectivity_state = + grpc_subchannel_check_connectivity(sd->subchannel, &error); + GRPC_ERROR_UNREF(error); + if (sd->curr_connectivity_state == GRPC_CHANNEL_TRANSIENT_FAILURE) { + // Reuses the connectivity refs from the previous watch. + grpc_lb_subchannel_data_start_connectivity_watch(exec_ctx, sd); + return; + } + // For any other state, go back to top of loop. + // We will reuse the connectivity refs from the previous watch. + } } } - - GRPC_ERROR_UNREF(error); } static const grpc_lb_policy_vtable pick_first_lb_policy_vtable = { @@ -688,8 +593,6 @@ static grpc_lb_policy *create_pick_first(grpc_exec_ctx *exec_ctx, pf_update_locked(exec_ctx, &p->base, args); grpc_lb_policy_init(&p->base, &pick_first_lb_policy_vtable, args->combiner); grpc_subchannel_index_ref(); - GRPC_CLOSURE_INIT(&p->connectivity_changed, pf_connectivity_changed_locked, p, - grpc_combiner_scheduler(args->combiner)); return &p->base; } diff --git a/src/core/ext/filters/client_channel/lb_policy/round_robin/round_robin.cc b/src/core/ext/filters/client_channel/lb_policy/round_robin/round_robin.cc index 6812bb50cd..8f29c80130 100644 --- a/src/core/ext/filters/client_channel/lb_policy/round_robin/round_robin.cc +++ b/src/core/ext/filters/client_channel/lb_policy/round_robin/round_robin.cc @@ -28,6 +28,7 @@ #include <grpc/support/alloc.h> +#include "src/core/ext/filters/client_channel/lb_policy/subchannel_list.h" #include "src/core/ext/filters/client_channel/lb_policy_registry.h" #include "src/core/ext/filters/client_channel/subchannel.h" #include "src/core/ext/filters/client_channel/subchannel_index.h" @@ -64,12 +65,11 @@ typedef struct pending_pick { grpc_closure *on_complete; } pending_pick; -typedef struct rr_subchannel_list rr_subchannel_list; typedef struct round_robin_lb_policy { /** base policy: must be first */ grpc_lb_policy base; - rr_subchannel_list *subchannel_list; + grpc_lb_subchannel_list *subchannel_list; /** have we started picking? */ bool started_picking; @@ -89,157 +89,9 @@ typedef struct round_robin_lb_policy { * lists if they equal \a latest_pending_subchannel_list. In other words, * racing callbacks that reference outdated subchannel lists won't perform any * update. */ - rr_subchannel_list *latest_pending_subchannel_list; + grpc_lb_subchannel_list *latest_pending_subchannel_list; } round_robin_lb_policy; -typedef struct { - /** backpointer to owning subchannel list */ - rr_subchannel_list *subchannel_list; - /** subchannel itself */ - grpc_subchannel *subchannel; - /** notification that connectivity has changed on subchannel */ - grpc_closure connectivity_changed_closure; - /** last observed connectivity. Not updated by - * \a grpc_subchannel_notify_on_state_change. Used to determine the previous - * state while processing the new state in \a rr_connectivity_changed */ - grpc_connectivity_state prev_connectivity_state; - /** current connectivity state. Updated by \a - * grpc_subchannel_notify_on_state_change */ - grpc_connectivity_state curr_connectivity_state; - /** connectivity state to be updated by the watcher, not guarded by - * the combiner. Will be moved to curr_connectivity_state inside of - * the combiner by rr_connectivity_changed_locked(). */ - grpc_connectivity_state pending_connectivity_state_unsafe; - /** the subchannel's target user data */ - void *user_data; - /** vtable to operate over \a user_data */ - const grpc_lb_user_data_vtable *user_data_vtable; -} subchannel_data; - -struct rr_subchannel_list { - /** backpointer to owning policy */ - round_robin_lb_policy *policy; - - /** all our subchannels */ - size_t num_subchannels; - subchannel_data *subchannels; - - /** how many subchannels are in state READY */ - size_t num_ready; - /** how many subchannels are in state TRANSIENT_FAILURE */ - size_t num_transient_failures; - /** how many subchannels are in state SHUTDOWN */ - size_t num_shutdown; - /** how many subchannels are in state IDLE */ - size_t num_idle; - - /** There will be one ref for each entry in subchannels for which there is a - * pending connectivity state watcher callback. */ - gpr_refcount refcount; - - /** Is this list shutting down? This may be true due to the shutdown of the - * policy itself or because a newer update has arrived while this one hadn't - * finished processing. */ - bool shutting_down; -}; - -static rr_subchannel_list *rr_subchannel_list_create(round_robin_lb_policy *p, - size_t num_subchannels) { - rr_subchannel_list *subchannel_list = - (rr_subchannel_list *)gpr_zalloc(sizeof(*subchannel_list)); - subchannel_list->policy = p; - subchannel_list->subchannels = - (subchannel_data *)gpr_zalloc(sizeof(subchannel_data) * num_subchannels); - subchannel_list->num_subchannels = num_subchannels; - gpr_ref_init(&subchannel_list->refcount, 1); - if (GRPC_TRACER_ON(grpc_lb_round_robin_trace)) { - gpr_log(GPR_INFO, "[RR %p] Created subchannel list %p for %lu subchannels", - (void *)p, (void *)subchannel_list, (unsigned long)num_subchannels); - } - return subchannel_list; -} - -static void rr_subchannel_list_destroy(grpc_exec_ctx *exec_ctx, - rr_subchannel_list *subchannel_list) { - GPR_ASSERT(subchannel_list->shutting_down); - if (GRPC_TRACER_ON(grpc_lb_round_robin_trace)) { - gpr_log(GPR_INFO, "[RR %p] Destroying subchannel_list %p", - (void *)subchannel_list->policy, (void *)subchannel_list); - } - for (size_t i = 0; i < subchannel_list->num_subchannels; i++) { - subchannel_data *sd = &subchannel_list->subchannels[i]; - if (sd->subchannel != NULL) { - GRPC_SUBCHANNEL_UNREF(exec_ctx, sd->subchannel, - "rr_subchannel_list_destroy"); - } - sd->subchannel = NULL; - if (sd->user_data != NULL) { - GPR_ASSERT(sd->user_data_vtable != NULL); - sd->user_data_vtable->destroy(exec_ctx, sd->user_data); - sd->user_data = NULL; - } - } - gpr_free(subchannel_list->subchannels); - gpr_free(subchannel_list); -} - -static void rr_subchannel_list_ref(rr_subchannel_list *subchannel_list, - const char *reason) { - gpr_ref_non_zero(&subchannel_list->refcount); - if (GRPC_TRACER_ON(grpc_lb_round_robin_trace)) { - const gpr_atm count = gpr_atm_acq_load(&subchannel_list->refcount.count); - gpr_log(GPR_INFO, "[RR %p] subchannel_list %p REF %lu->%lu (%s)", - (void *)subchannel_list->policy, (void *)subchannel_list, - (unsigned long)(count - 1), (unsigned long)count, reason); - } -} - -static void rr_subchannel_list_unref(grpc_exec_ctx *exec_ctx, - rr_subchannel_list *subchannel_list, - const char *reason) { - const bool done = gpr_unref(&subchannel_list->refcount); - if (GRPC_TRACER_ON(grpc_lb_round_robin_trace)) { - const gpr_atm count = gpr_atm_acq_load(&subchannel_list->refcount.count); - gpr_log(GPR_INFO, "[RR %p] subchannel_list %p UNREF %lu->%lu (%s)", - (void *)subchannel_list->policy, (void *)subchannel_list, - (unsigned long)(count + 1), (unsigned long)count, reason); - } - if (done) { - rr_subchannel_list_destroy(exec_ctx, subchannel_list); - } -} - -/** Mark \a subchannel_list as discarded. Unsubscribes all its subchannels. The - * watcher's callback will ultimately unref \a subchannel_list. */ -static void rr_subchannel_list_shutdown_and_unref( - grpc_exec_ctx *exec_ctx, rr_subchannel_list *subchannel_list, - const char *reason) { - GPR_ASSERT(!subchannel_list->shutting_down); - if (GRPC_TRACER_ON(grpc_lb_round_robin_trace)) { - gpr_log(GPR_DEBUG, "[RR %p] Shutting down subchannel_list %p (%s)", - (void *)subchannel_list->policy, (void *)subchannel_list, reason); - } - GPR_ASSERT(!subchannel_list->shutting_down); - subchannel_list->shutting_down = true; - for (size_t i = 0; i < subchannel_list->num_subchannels; i++) { - subchannel_data *sd = &subchannel_list->subchannels[i]; - if (sd->subchannel != NULL) { // if subchannel isn't shutdown, unsubscribe. - if (GRPC_TRACER_ON(grpc_lb_round_robin_trace)) { - gpr_log( - GPR_DEBUG, - "[RR %p] Unsubscribing from subchannel %p as part of shutting down " - "subchannel_list %p", - (void *)subchannel_list->policy, (void *)sd->subchannel, - (void *)subchannel_list); - } - grpc_subchannel_notify_on_state_change(exec_ctx, sd->subchannel, NULL, - NULL, - &sd->connectivity_changed_closure); - } - } - rr_subchannel_list_unref(exec_ctx, subchannel_list, reason); -} - /** Returns the index into p->subchannel_list->subchannels of the next * subchannel in READY state, or p->subchannel_list->num_subchannels if no * subchannel is READY. @@ -299,8 +151,8 @@ static void update_last_ready_subchannel_index_locked(round_robin_lb_policy *p, "[RR %p] setting last_ready_subchannel_index=%lu (SC %p, CSC %p)", (void *)p, (unsigned long)last_ready_index, (void *)p->subchannel_list->subchannels[last_ready_index].subchannel, - (void *)grpc_subchannel_get_connected_subchannel( - p->subchannel_list->subchannels[last_ready_index].subchannel)); + (void *)p->subchannel_list->subchannels[last_ready_index] + .connected_subchannel); } } @@ -310,42 +162,47 @@ static void rr_destroy(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { gpr_log(GPR_DEBUG, "[RR %p] Destroying Round Robin policy at %p", (void *)pol, (void *)pol); } + GPR_ASSERT(p->subchannel_list == NULL); + GPR_ASSERT(p->latest_pending_subchannel_list == NULL); grpc_connectivity_state_destroy(exec_ctx, &p->state_tracker); grpc_subchannel_index_unref(); gpr_free(p); } -static void rr_shutdown_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { - round_robin_lb_policy *p = (round_robin_lb_policy *)pol; +static void shutdown_locked(grpc_exec_ctx *exec_ctx, round_robin_lb_policy *p, + grpc_error *error) { if (GRPC_TRACER_ON(grpc_lb_round_robin_trace)) { - gpr_log(GPR_DEBUG, "[RR %p] Shutting down Round Robin policy at %p", - (void *)pol, (void *)pol); + gpr_log(GPR_DEBUG, "[RR %p] Shutting down", p); } p->shutdown = true; pending_pick *pp; - while ((pp = p->pending_picks)) { + while ((pp = p->pending_picks) != NULL) { p->pending_picks = pp->next; *pp->target = NULL; - GRPC_CLOSURE_SCHED( - exec_ctx, pp->on_complete, - GRPC_ERROR_CREATE_FROM_STATIC_STRING("Channel Shutdown")); + GRPC_CLOSURE_SCHED(exec_ctx, pp->on_complete, GRPC_ERROR_REF(error)); gpr_free(pp); } - grpc_connectivity_state_set( - exec_ctx, &p->state_tracker, GRPC_CHANNEL_SHUTDOWN, - GRPC_ERROR_CREATE_FROM_STATIC_STRING("Channel Shutdown"), "rr_shutdown"); - const bool latest_is_current = - p->subchannel_list == p->latest_pending_subchannel_list; - rr_subchannel_list_shutdown_and_unref(exec_ctx, p->subchannel_list, - "sl_shutdown_rr_shutdown"); - p->subchannel_list = NULL; - if (!latest_is_current && p->latest_pending_subchannel_list != NULL && - !p->latest_pending_subchannel_list->shutting_down) { - rr_subchannel_list_shutdown_and_unref(exec_ctx, - p->latest_pending_subchannel_list, - "sl_shutdown_pending_rr_shutdown"); + grpc_connectivity_state_set(exec_ctx, &p->state_tracker, + GRPC_CHANNEL_SHUTDOWN, GRPC_ERROR_REF(error), + "rr_shutdown"); + if (p->subchannel_list != NULL) { + grpc_lb_subchannel_list_shutdown_and_unref(exec_ctx, p->subchannel_list, + "sl_shutdown_rr_shutdown"); + p->subchannel_list = NULL; + } + if (p->latest_pending_subchannel_list != NULL) { + grpc_lb_subchannel_list_shutdown_and_unref( + exec_ctx, p->latest_pending_subchannel_list, + "sl_shutdown_pending_rr_shutdown"); p->latest_pending_subchannel_list = NULL; } + GRPC_ERROR_UNREF(error); +} + +static void rr_shutdown_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { + round_robin_lb_policy *p = (round_robin_lb_policy *)pol; + shutdown_locked(exec_ctx, p, + GRPC_ERROR_CREATE_FROM_STATIC_STRING("Channel Shutdown")); } static void rr_cancel_pick_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, @@ -400,13 +257,10 @@ static void start_picking_locked(grpc_exec_ctx *exec_ctx, round_robin_lb_policy *p) { p->started_picking = true; for (size_t i = 0; i < p->subchannel_list->num_subchannels; i++) { - subchannel_data *sd = &p->subchannel_list->subchannels[i]; - GRPC_LB_POLICY_WEAK_REF(&p->base, "start_picking_locked"); - rr_subchannel_list_ref(sd->subchannel_list, "started_picking"); - grpc_subchannel_notify_on_state_change( - exec_ctx, sd->subchannel, p->base.interested_parties, - &sd->pending_connectivity_state_unsafe, - &sd->connectivity_changed_closure); + grpc_lb_subchannel_list_ref_for_connectivity_watch(p->subchannel_list, + "connectivity_watch"); + grpc_lb_subchannel_data_start_connectivity_watch( + exec_ctx, &p->subchannel_list->subchannels[i]); } } @@ -431,10 +285,10 @@ static int rr_pick_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, const size_t next_ready_index = get_next_ready_subchannel_index_locked(p); if (next_ready_index < p->subchannel_list->num_subchannels) { /* readily available, report right away */ - subchannel_data *sd = &p->subchannel_list->subchannels[next_ready_index]; - *target = GRPC_CONNECTED_SUBCHANNEL_REF( - grpc_subchannel_get_connected_subchannel(sd->subchannel), - "rr_picked"); + grpc_lb_subchannel_data *sd = + &p->subchannel_list->subchannels[next_ready_index]; + *target = + GRPC_CONNECTED_SUBCHANNEL_REF(sd->connected_subchannel, "rr_picked"); if (user_data != NULL) { *user_data = sd->user_data; } @@ -465,8 +319,8 @@ static int rr_pick_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, return 0; } -static void update_state_counters_locked(subchannel_data *sd) { - rr_subchannel_list *subchannel_list = sd->subchannel_list; +static void update_state_counters_locked(grpc_lb_subchannel_data *sd) { + grpc_lb_subchannel_list *subchannel_list = sd->subchannel_list; if (sd->prev_connectivity_state == GRPC_CHANNEL_READY) { GPR_ASSERT(subchannel_list->num_ready > 0); --subchannel_list->num_ready; @@ -480,6 +334,7 @@ static void update_state_counters_locked(subchannel_data *sd) { GPR_ASSERT(subchannel_list->num_idle > 0); --subchannel_list->num_idle; } + sd->prev_connectivity_state = sd->curr_connectivity_state; if (sd->curr_connectivity_state == GRPC_CHANNEL_READY) { ++subchannel_list->num_ready; } else if (sd->curr_connectivity_state == GRPC_CHANNEL_TRANSIENT_FAILURE) { @@ -492,12 +347,12 @@ static void update_state_counters_locked(subchannel_data *sd) { } /** Sets the policy's connectivity status based on that of the passed-in \a sd - * (the subchannel_data associted with the updated subchannel) and the + * (the grpc_lb_subchannel_data associted with the updated subchannel) and the * subchannel list \a sd belongs to (sd->subchannel_list). \a error will only be * used upon policy transition to TRANSIENT_FAILURE or SHUTDOWN. Returns the * connectivity status set. */ static grpc_connectivity_state update_lb_connectivity_status_locked( - grpc_exec_ctx *exec_ctx, subchannel_data *sd, grpc_error *error) { + grpc_exec_ctx *exec_ctx, grpc_lb_subchannel_data *sd, grpc_error *error) { /* In priority order. The first rule to match terminates the search (ie, if we * are on rule n, all previous rules were unfulfilled). * @@ -519,8 +374,8 @@ static grpc_connectivity_state update_lb_connectivity_status_locked( * CHECK: p->num_idle == p->subchannel_list->num_subchannels. */ grpc_connectivity_state new_state = sd->curr_connectivity_state; - rr_subchannel_list *subchannel_list = sd->subchannel_list; - round_robin_lb_policy *p = subchannel_list->policy; + grpc_lb_subchannel_list *subchannel_list = sd->subchannel_list; + round_robin_lb_policy *p = (round_robin_lb_policy *)subchannel_list->policy; if (subchannel_list->num_ready > 0) { /* 1) READY */ grpc_connectivity_state_set(exec_ctx, &p->state_tracker, GRPC_CHANNEL_READY, GRPC_ERROR_NONE, "rr_ready"); @@ -556,8 +411,9 @@ static grpc_connectivity_state update_lb_connectivity_status_locked( static void rr_connectivity_changed_locked(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { - subchannel_data *sd = (subchannel_data *)arg; - round_robin_lb_policy *p = sd->subchannel_list->policy; + grpc_lb_subchannel_data *sd = (grpc_lb_subchannel_data *)arg; + round_robin_lb_policy *p = + (round_robin_lb_policy *)sd->subchannel_list->policy; if (GRPC_TRACER_ON(grpc_lb_round_robin_trace)) { gpr_log( GPR_DEBUG, @@ -572,71 +428,50 @@ static void rr_connectivity_changed_locked(grpc_exec_ctx *exec_ctx, void *arg, } // If the policy is shutting down, unref and return. if (p->shutdown) { - rr_subchannel_list_unref(exec_ctx, sd->subchannel_list, - "pol_shutdown+started_picking"); - GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &p->base, "pol_shutdown"); + grpc_lb_subchannel_data_stop_connectivity_watch(exec_ctx, sd); + grpc_lb_subchannel_data_unref_subchannel(exec_ctx, sd, "rr_shutdown"); + grpc_lb_subchannel_list_unref_for_connectivity_watch( + exec_ctx, sd->subchannel_list, "rr_shutdown"); return; } - if (sd->subchannel_list->shutting_down && error == GRPC_ERROR_CANCELLED) { - // the subchannel list associated with sd has been discarded. This callback - // corresponds to the unsubscription. The unrefs correspond to the picking - // ref (start_picking_locked or update_started_picking). - rr_subchannel_list_unref(exec_ctx, sd->subchannel_list, - "sl_shutdown+started_picking"); - GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &p->base, "sl_shutdown+picking"); - return; - } - // Dispose of outdated subchannel lists. - if (sd->subchannel_list != p->subchannel_list && - sd->subchannel_list != p->latest_pending_subchannel_list) { - const char *reason = NULL; - if (sd->subchannel_list->shutting_down) { - reason = "sl_outdated_straggler"; - rr_subchannel_list_unref(exec_ctx, sd->subchannel_list, reason); - } else { - reason = "sl_outdated"; - rr_subchannel_list_shutdown_and_unref(exec_ctx, sd->subchannel_list, - reason); - } - GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &p->base, reason); + // If the subchannel list is shutting down, stop watching. + if (sd->subchannel_list->shutting_down || error == GRPC_ERROR_CANCELLED) { + grpc_lb_subchannel_data_stop_connectivity_watch(exec_ctx, sd); + grpc_lb_subchannel_data_unref_subchannel(exec_ctx, sd, "rr_sl_shutdown"); + grpc_lb_subchannel_list_unref_for_connectivity_watch( + exec_ctx, sd->subchannel_list, "rr_sl_shutdown"); return; } + // If we're still here, the notification must be for a subchannel in + // either the current or latest pending subchannel lists. + GPR_ASSERT(sd->subchannel_list == p->subchannel_list || + sd->subchannel_list == p->latest_pending_subchannel_list); // Now that we're inside the combiner, copy the pending connectivity // state (which was set by the connectivity state watcher) to // curr_connectivity_state, which is what we use inside of the combiner. sd->curr_connectivity_state = sd->pending_connectivity_state_unsafe; // Update state counters and determine new overall state. update_state_counters_locked(sd); - sd->prev_connectivity_state = sd->curr_connectivity_state; const grpc_connectivity_state new_policy_connectivity_state = update_lb_connectivity_status_locked(exec_ctx, sd, GRPC_ERROR_REF(error)); // If the sd's new state is SHUTDOWN, unref the subchannel, and if the new // policy's state is SHUTDOWN, clean up. if (sd->curr_connectivity_state == GRPC_CHANNEL_SHUTDOWN) { - GRPC_SUBCHANNEL_UNREF(exec_ctx, sd->subchannel, "rr_subchannel_shutdown"); - sd->subchannel = NULL; - if (sd->user_data != NULL) { - GPR_ASSERT(sd->user_data_vtable != NULL); - sd->user_data_vtable->destroy(exec_ctx, sd->user_data); - sd->user_data = NULL; - } + grpc_lb_subchannel_data_stop_connectivity_watch(exec_ctx, sd); + grpc_lb_subchannel_data_unref_subchannel(exec_ctx, sd, + "rr_connectivity_shutdown"); + grpc_lb_subchannel_list_unref_for_connectivity_watch( + exec_ctx, sd->subchannel_list, "rr_connectivity_shutdown"); if (new_policy_connectivity_state == GRPC_CHANNEL_SHUTDOWN) { - // the policy is shutting down. Flush all the pending picks... - pending_pick *pp; - while ((pp = p->pending_picks)) { - p->pending_picks = pp->next; - *pp->target = NULL; - GRPC_CLOSURE_SCHED(exec_ctx, pp->on_complete, GRPC_ERROR_NONE); - gpr_free(pp); - } + shutdown_locked(exec_ctx, p, GRPC_ERROR_REF(error)); } - rr_subchannel_list_unref(exec_ctx, sd->subchannel_list, - "sd_shutdown+started_picking"); - // unref the "rr_connectivity_update" weak ref from start_picking. - GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &p->base, - "rr_connectivity_sd_shutdown"); } else { // sd not in SHUTDOWN if (sd->curr_connectivity_state == GRPC_CHANNEL_READY) { + if (sd->connected_subchannel == NULL) { + sd->connected_subchannel = GRPC_CONNECTED_SUBCHANNEL_REF( + grpc_subchannel_get_connected_subchannel(sd->subchannel), + "connected"); + } if (sd->subchannel_list != p->subchannel_list) { // promote sd->subchannel_list to p->subchannel_list. // sd->subchannel_list must be equal to @@ -657,8 +492,8 @@ static void rr_connectivity_changed_locked(grpc_exec_ctx *exec_ctx, void *arg, } if (p->subchannel_list != NULL) { // dispose of the current subchannel_list - rr_subchannel_list_shutdown_and_unref(exec_ctx, p->subchannel_list, - "sl_phase_out_shutdown"); + grpc_lb_subchannel_list_shutdown_and_unref( + exec_ctx, p->subchannel_list, "sl_phase_out_shutdown"); } p->subchannel_list = p->latest_pending_subchannel_list; p->latest_pending_subchannel_list = NULL; @@ -668,7 +503,7 @@ static void rr_connectivity_changed_locked(grpc_exec_ctx *exec_ctx, void *arg, * p->pending_picks. This preemtively replicates rr_pick()'s actions. */ const size_t next_ready_index = get_next_ready_subchannel_index_locked(p); GPR_ASSERT(next_ready_index < p->subchannel_list->num_subchannels); - subchannel_data *selected = + grpc_lb_subchannel_data *selected = &p->subchannel_list->subchannels[next_ready_index]; if (p->pending_picks != NULL) { // if the selected subchannel is going to be used for the pending @@ -679,8 +514,7 @@ static void rr_connectivity_changed_locked(grpc_exec_ctx *exec_ctx, void *arg, while ((pp = p->pending_picks)) { p->pending_picks = pp->next; *pp->target = GRPC_CONNECTED_SUBCHANNEL_REF( - grpc_subchannel_get_connected_subchannel(selected->subchannel), - "rr_picked"); + selected->connected_subchannel, "rr_picked"); if (pp->user_data != NULL) { *pp->user_data = selected->user_data; } @@ -695,12 +529,8 @@ static void rr_connectivity_changed_locked(grpc_exec_ctx *exec_ctx, void *arg, gpr_free(pp); } } - /* renew notification: reuses the "rr_connectivity_update" weak ref on the - * policy as well as the sd->subchannel_list ref. */ - grpc_subchannel_notify_on_state_change( - exec_ctx, sd->subchannel, p->base.interested_parties, - &sd->pending_connectivity_state_unsafe, - &sd->connectivity_changed_closure); + // Renew notification. + grpc_lb_subchannel_data_start_connectivity_watch(exec_ctx, sd); } } @@ -724,13 +554,12 @@ static void rr_ping_one_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, round_robin_lb_policy *p = (round_robin_lb_policy *)pol; const size_t next_ready_index = get_next_ready_subchannel_index_locked(p); if (next_ready_index < p->subchannel_list->num_subchannels) { - subchannel_data *selected = + grpc_lb_subchannel_data *selected = &p->subchannel_list->subchannels[next_ready_index]; grpc_connected_subchannel *target = GRPC_CONNECTED_SUBCHANNEL_REF( - grpc_subchannel_get_connected_subchannel(selected->subchannel), - "rr_picked"); + selected->connected_subchannel, "rr_ping"); grpc_connected_subchannel_ping(exec_ctx, target, closure); - GRPC_CONNECTED_SUBCHANNEL_UNREF(exec_ctx, target, "rr_picked"); + GRPC_CONNECTED_SUBCHANNEL_UNREF(exec_ctx, target, "rr_ping"); } else { GRPC_CLOSURE_SCHED(exec_ctx, closure, GRPC_ERROR_CREATE_FROM_STATIC_STRING( "Round Robin not connected")); @@ -743,130 +572,68 @@ static void rr_update_locked(grpc_exec_ctx *exec_ctx, grpc_lb_policy *policy, const grpc_arg *arg = grpc_channel_args_find(args->args, GRPC_ARG_LB_ADDRESSES); if (arg == NULL || arg->type != GRPC_ARG_POINTER) { + gpr_log(GPR_ERROR, "[RR %p] update provided no addresses; ignoring", p); + // If we don't have a current subchannel list, go into TRANSIENT_FAILURE. + // Otherwise, keep using the current subchannel list (ignore this update). if (p->subchannel_list == NULL) { - // If we don't have a current subchannel list, go into TRANSIENT FAILURE. grpc_connectivity_state_set( exec_ctx, &p->state_tracker, GRPC_CHANNEL_TRANSIENT_FAILURE, GRPC_ERROR_CREATE_FROM_STATIC_STRING("Missing update in args"), "rr_update_missing"); - } else { - // otherwise, keep using the current subchannel list (ignore this update). - gpr_log(GPR_ERROR, - "[RR %p] No valid LB addresses channel arg for update, ignoring.", - (void *)p); } return; } grpc_lb_addresses *addresses = (grpc_lb_addresses *)arg->value.pointer.p; - rr_subchannel_list *subchannel_list = - rr_subchannel_list_create(p, addresses->num_addresses); - if (addresses->num_addresses == 0) { + if (GRPC_TRACER_ON(grpc_lb_round_robin_trace)) { + gpr_log(GPR_DEBUG, "[RR %p] received update with %" PRIuPTR " addresses", p, + addresses->num_addresses); + } + grpc_lb_subchannel_list *subchannel_list = grpc_lb_subchannel_list_create( + exec_ctx, &p->base, &grpc_lb_round_robin_trace, addresses, args, + rr_connectivity_changed_locked); + if (subchannel_list->num_subchannels == 0) { grpc_connectivity_state_set( exec_ctx, &p->state_tracker, GRPC_CHANNEL_TRANSIENT_FAILURE, GRPC_ERROR_CREATE_FROM_STATIC_STRING("Empty update"), "rr_update_empty"); if (p->subchannel_list != NULL) { - rr_subchannel_list_shutdown_and_unref(exec_ctx, p->subchannel_list, - "sl_shutdown_empty_update"); + grpc_lb_subchannel_list_shutdown_and_unref(exec_ctx, p->subchannel_list, + "sl_shutdown_empty_update"); } p->subchannel_list = subchannel_list; // empty list return; } - size_t subchannel_index = 0; - if (p->latest_pending_subchannel_list != NULL && p->started_picking) { - if (GRPC_TRACER_ON(grpc_lb_round_robin_trace)) { - gpr_log(GPR_DEBUG, - "[RR %p] Shutting down latest pending subchannel list %p, about " - "to be replaced by newer latest %p", - (void *)p, (void *)p->latest_pending_subchannel_list, - (void *)subchannel_list); - } - rr_subchannel_list_shutdown_and_unref( - exec_ctx, p->latest_pending_subchannel_list, "sl_outdated_dont_smash"); - } - p->latest_pending_subchannel_list = subchannel_list; - grpc_subchannel_args sc_args; - /* We need to remove the LB addresses in order to be able to compare the - * subchannel keys of subchannels from a different batch of addresses. */ - static const char *keys_to_remove[] = {GRPC_ARG_SUBCHANNEL_ADDRESS, - GRPC_ARG_LB_ADDRESSES}; - /* Create subchannels for addresses in the update. */ - for (size_t i = 0; i < addresses->num_addresses; i++) { - // If there were any balancer, we would have chosen grpclb policy instead. - GPR_ASSERT(!addresses->addresses[i].is_balancer); - memset(&sc_args, 0, sizeof(grpc_subchannel_args)); - grpc_arg addr_arg = - grpc_create_subchannel_address_arg(&addresses->addresses[i].address); - grpc_channel_args *new_args = grpc_channel_args_copy_and_add_and_remove( - args->args, keys_to_remove, GPR_ARRAY_SIZE(keys_to_remove), &addr_arg, - 1); - gpr_free(addr_arg.value.string); - sc_args.args = new_args; - grpc_subchannel *subchannel = grpc_client_channel_factory_create_subchannel( - exec_ctx, args->client_channel_factory, &sc_args); - grpc_channel_args_destroy(exec_ctx, new_args); - grpc_error *error; - // Get the connectivity state of the subchannel. Already existing ones may - // be in a state other than INIT. - const grpc_connectivity_state subchannel_connectivity_state = - grpc_subchannel_check_connectivity(subchannel, &error); - if (error != GRPC_ERROR_NONE) { - // The subchannel is in error (e.g. shutting down). Ignore it. - GRPC_SUBCHANNEL_UNREF(exec_ctx, subchannel, "new_sc_connectivity_error"); - GRPC_ERROR_UNREF(error); - continue; - } - if (GRPC_TRACER_ON(grpc_lb_round_robin_trace)) { - char *address_uri = - grpc_sockaddr_to_uri(&addresses->addresses[i].address); - gpr_log( - GPR_DEBUG, - "[RR %p] index %lu: Created subchannel %p for address uri %s into " - "subchannel_list %p. Connectivity state %s", - (void *)p, (unsigned long)subchannel_index, (void *)subchannel, - address_uri, (void *)subchannel_list, - grpc_connectivity_state_name(subchannel_connectivity_state)); - gpr_free(address_uri); - } - subchannel_data *sd = &subchannel_list->subchannels[subchannel_index++]; - sd->subchannel_list = subchannel_list; - sd->subchannel = subchannel; - GRPC_CLOSURE_INIT(&sd->connectivity_changed_closure, - rr_connectivity_changed_locked, sd, - grpc_combiner_scheduler(args->combiner)); - /* use some sentinel value outside of the range of - * grpc_connectivity_state to signal an undefined previous state. We - * won't be referring to this value again and it'll be overwritten after - * the first call to rr_connectivity_changed_locked */ - sd->prev_connectivity_state = GRPC_CHANNEL_INIT; - sd->curr_connectivity_state = subchannel_connectivity_state; - sd->user_data_vtable = addresses->user_data_vtable; - if (sd->user_data_vtable != NULL) { - sd->user_data = - sd->user_data_vtable->copy(addresses->addresses[i].user_data); + if (p->started_picking) { + if (p->latest_pending_subchannel_list != NULL) { + if (GRPC_TRACER_ON(grpc_lb_round_robin_trace)) { + gpr_log(GPR_DEBUG, + "[RR %p] Shutting down latest pending subchannel list %p, " + "about to be replaced by newer latest %p", + (void *)p, (void *)p->latest_pending_subchannel_list, + (void *)subchannel_list); + } + grpc_lb_subchannel_list_shutdown_and_unref( + exec_ctx, p->latest_pending_subchannel_list, "sl_outdated"); } - if (p->started_picking) { - rr_subchannel_list_ref(sd->subchannel_list, "update_started_picking"); - GRPC_LB_POLICY_WEAK_REF(&p->base, "rr_connectivity_update"); - /* 2. Watch every new subchannel. A subchannel list becomes active the + p->latest_pending_subchannel_list = subchannel_list; + for (size_t i = 0; i < subchannel_list->num_subchannels; ++i) { + /* Watch every new subchannel. A subchannel list becomes active the * moment one of its subchannels is READY. At that moment, we swap * p->subchannel_list for sd->subchannel_list, provided the subchannel * list is still valid (ie, isn't shutting down) */ - grpc_subchannel_notify_on_state_change( - exec_ctx, sd->subchannel, p->base.interested_parties, - &sd->pending_connectivity_state_unsafe, - &sd->connectivity_changed_closure); + grpc_lb_subchannel_list_ref_for_connectivity_watch(subchannel_list, + "connectivity_watch"); + grpc_lb_subchannel_data_start_connectivity_watch( + exec_ctx, &subchannel_list->subchannels[i]); } - } - if (!p->started_picking) { + } else { // The policy isn't picking yet. Save the update for later, disposing of // previous version if any. if (p->subchannel_list != NULL) { - rr_subchannel_list_shutdown_and_unref(exec_ctx, p->subchannel_list, - "rr_update_before_started_picking"); + grpc_lb_subchannel_list_shutdown_and_unref( + exec_ctx, p->subchannel_list, "rr_update_before_started_picking"); } p->subchannel_list = subchannel_list; - p->latest_pending_subchannel_list = NULL; } } diff --git a/src/core/ext/filters/client_channel/lb_policy/subchannel_list.cc b/src/core/ext/filters/client_channel/lb_policy/subchannel_list.cc new file mode 100644 index 0000000000..08ea4f480b --- /dev/null +++ b/src/core/ext/filters/client_channel/lb_policy/subchannel_list.cc @@ -0,0 +1,265 @@ +/* + * + * Copyright 2015 gRPC authors. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + * + */ + +#include <string.h> + +#include <grpc/support/alloc.h> + +#include "src/core/ext/filters/client_channel/lb_policy/subchannel_list.h" +#include "src/core/lib/channel/channel_args.h" +#include "src/core/lib/debug/trace.h" +#include "src/core/lib/iomgr/closure.h" +#include "src/core/lib/iomgr/combiner.h" +#include "src/core/lib/iomgr/sockaddr_utils.h" +#include "src/core/lib/transport/connectivity_state.h" + +void grpc_lb_subchannel_data_unref_subchannel(grpc_exec_ctx *exec_ctx, + grpc_lb_subchannel_data *sd, + const char *reason) { + if (sd->subchannel != NULL) { + if (GRPC_TRACER_ON(*sd->subchannel_list->tracer)) { + gpr_log( + GPR_DEBUG, "[%s %p] subchannel list %p index %" PRIuPTR + " of %" PRIuPTR " (subchannel %p): unreffing subchannel", + sd->subchannel_list->tracer->name, sd->subchannel_list->policy, + sd->subchannel_list, (size_t)(sd - sd->subchannel_list->subchannels), + sd->subchannel_list->num_subchannels, sd->subchannel); + } + GRPC_SUBCHANNEL_UNREF(exec_ctx, sd->subchannel, reason); + sd->subchannel = NULL; + if (sd->connected_subchannel != NULL) { + GRPC_CONNECTED_SUBCHANNEL_UNREF(exec_ctx, sd->connected_subchannel, + reason); + sd->connected_subchannel = NULL; + } + if (sd->user_data != NULL) { + GPR_ASSERT(sd->user_data_vtable != NULL); + sd->user_data_vtable->destroy(exec_ctx, sd->user_data); + sd->user_data = NULL; + } + } +} + +void grpc_lb_subchannel_data_start_connectivity_watch( + grpc_exec_ctx *exec_ctx, grpc_lb_subchannel_data *sd) { + if (GRPC_TRACER_ON(*sd->subchannel_list->tracer)) { + gpr_log(GPR_DEBUG, + "[%s %p] subchannel list %p index %" PRIuPTR " of %" PRIuPTR + " (subchannel %p): requesting connectivity change notification", + sd->subchannel_list->tracer->name, sd->subchannel_list->policy, + sd->subchannel_list, + (size_t)(sd - sd->subchannel_list->subchannels), + sd->subchannel_list->num_subchannels, sd->subchannel); + } + sd->connectivity_notification_pending = true; + grpc_subchannel_notify_on_state_change( + exec_ctx, sd->subchannel, sd->subchannel_list->policy->interested_parties, + &sd->pending_connectivity_state_unsafe, + &sd->connectivity_changed_closure); +} + +void grpc_lb_subchannel_data_stop_connectivity_watch( + grpc_exec_ctx *exec_ctx, grpc_lb_subchannel_data *sd) { + if (GRPC_TRACER_ON(*sd->subchannel_list->tracer)) { + gpr_log( + GPR_DEBUG, "[%s %p] subchannel list %p index %" PRIuPTR " of %" PRIuPTR + " (subchannel %p): stopping connectivity watch", + sd->subchannel_list->tracer->name, sd->subchannel_list->policy, + sd->subchannel_list, (size_t)(sd - sd->subchannel_list->subchannels), + sd->subchannel_list->num_subchannels, sd->subchannel); + } + GPR_ASSERT(sd->connectivity_notification_pending); + sd->connectivity_notification_pending = false; +} + +grpc_lb_subchannel_list *grpc_lb_subchannel_list_create( + grpc_exec_ctx *exec_ctx, grpc_lb_policy *p, grpc_tracer_flag *tracer, + const grpc_lb_addresses *addresses, const grpc_lb_policy_args *args, + grpc_iomgr_cb_func connectivity_changed_cb) { + grpc_lb_subchannel_list *subchannel_list = + (grpc_lb_subchannel_list *)gpr_zalloc(sizeof(*subchannel_list)); + if (GRPC_TRACER_ON(*tracer)) { + gpr_log(GPR_DEBUG, + "[%s %p] Creating subchannel list %p for %" PRIuPTR " subchannels", + tracer->name, p, subchannel_list, addresses->num_addresses); + } + subchannel_list->policy = p; + subchannel_list->tracer = tracer; + gpr_ref_init(&subchannel_list->refcount, 1); + subchannel_list->subchannels = (grpc_lb_subchannel_data *)gpr_zalloc( + sizeof(grpc_lb_subchannel_data) * addresses->num_addresses); + // We need to remove the LB addresses in order to be able to compare the + // subchannel keys of subchannels from a different batch of addresses. + static const char *keys_to_remove[] = {GRPC_ARG_SUBCHANNEL_ADDRESS, + GRPC_ARG_LB_ADDRESSES}; + // Create a subchannel for each address. + grpc_subchannel_args sc_args; + size_t subchannel_index = 0; + for (size_t i = 0; i < addresses->num_addresses; i++) { + // If there were any balancer, we would have chosen grpclb policy instead. + GPR_ASSERT(!addresses->addresses[i].is_balancer); + memset(&sc_args, 0, sizeof(grpc_subchannel_args)); + grpc_arg addr_arg = + grpc_create_subchannel_address_arg(&addresses->addresses[i].address); + grpc_channel_args *new_args = grpc_channel_args_copy_and_add_and_remove( + args->args, keys_to_remove, GPR_ARRAY_SIZE(keys_to_remove), &addr_arg, + 1); + gpr_free(addr_arg.value.string); + sc_args.args = new_args; + grpc_subchannel *subchannel = grpc_client_channel_factory_create_subchannel( + exec_ctx, args->client_channel_factory, &sc_args); + grpc_channel_args_destroy(exec_ctx, new_args); + if (subchannel == NULL) { + // Subchannel could not be created. + if (GRPC_TRACER_ON(*tracer)) { + char *address_uri = + grpc_sockaddr_to_uri(&addresses->addresses[i].address); + gpr_log(GPR_DEBUG, + "[%s %p] could not create subchannel for address uri %s, " + "ignoring", + tracer->name, subchannel_list->policy, address_uri); + gpr_free(address_uri); + } + continue; + } + if (GRPC_TRACER_ON(*tracer)) { + char *address_uri = + grpc_sockaddr_to_uri(&addresses->addresses[i].address); + gpr_log(GPR_DEBUG, "[%s %p] subchannel list %p index %" PRIuPTR + ": Created subchannel %p for address uri %s", + tracer->name, p, subchannel_list, subchannel_index, subchannel, + address_uri); + gpr_free(address_uri); + } + grpc_lb_subchannel_data *sd = + &subchannel_list->subchannels[subchannel_index++]; + sd->subchannel_list = subchannel_list; + sd->subchannel = subchannel; + GRPC_CLOSURE_INIT(&sd->connectivity_changed_closure, + connectivity_changed_cb, sd, + grpc_combiner_scheduler(args->combiner)); + // We assume that the current state is IDLE. If not, we'll get a + // callback telling us that. + sd->prev_connectivity_state = GRPC_CHANNEL_IDLE; + sd->curr_connectivity_state = GRPC_CHANNEL_IDLE; + sd->pending_connectivity_state_unsafe = GRPC_CHANNEL_IDLE; + sd->user_data_vtable = addresses->user_data_vtable; + if (sd->user_data_vtable != NULL) { + sd->user_data = + sd->user_data_vtable->copy(addresses->addresses[i].user_data); + } + } + subchannel_list->num_subchannels = subchannel_index; + subchannel_list->num_idle = subchannel_index; + return subchannel_list; +} + +static void subchannel_list_destroy(grpc_exec_ctx *exec_ctx, + grpc_lb_subchannel_list *subchannel_list) { + if (GRPC_TRACER_ON(*subchannel_list->tracer)) { + gpr_log(GPR_DEBUG, "[%s %p] Destroying subchannel_list %p", + subchannel_list->tracer->name, subchannel_list->policy, + subchannel_list); + } + for (size_t i = 0; i < subchannel_list->num_subchannels; i++) { + grpc_lb_subchannel_data *sd = &subchannel_list->subchannels[i]; + grpc_lb_subchannel_data_unref_subchannel(exec_ctx, sd, + "subchannel_list_destroy"); + } + gpr_free(subchannel_list->subchannels); + gpr_free(subchannel_list); +} + +void grpc_lb_subchannel_list_ref(grpc_lb_subchannel_list *subchannel_list, + const char *reason) { + gpr_ref_non_zero(&subchannel_list->refcount); + if (GRPC_TRACER_ON(*subchannel_list->tracer)) { + const gpr_atm count = gpr_atm_acq_load(&subchannel_list->refcount.count); + gpr_log(GPR_DEBUG, "[%s %p] subchannel_list %p REF %lu->%lu (%s)", + subchannel_list->tracer->name, subchannel_list->policy, + subchannel_list, (unsigned long)(count - 1), (unsigned long)count, + reason); + } +} + +void grpc_lb_subchannel_list_unref(grpc_exec_ctx *exec_ctx, + grpc_lb_subchannel_list *subchannel_list, + const char *reason) { + const bool done = gpr_unref(&subchannel_list->refcount); + if (GRPC_TRACER_ON(*subchannel_list->tracer)) { + const gpr_atm count = gpr_atm_acq_load(&subchannel_list->refcount.count); + gpr_log(GPR_DEBUG, "[%s %p] subchannel_list %p UNREF %lu->%lu (%s)", + subchannel_list->tracer->name, subchannel_list->policy, + subchannel_list, (unsigned long)(count + 1), (unsigned long)count, + reason); + } + if (done) { + subchannel_list_destroy(exec_ctx, subchannel_list); + } +} + +void grpc_lb_subchannel_list_ref_for_connectivity_watch( + grpc_lb_subchannel_list *subchannel_list, const char *reason) { + GRPC_LB_POLICY_WEAK_REF(subchannel_list->policy, reason); + grpc_lb_subchannel_list_ref(subchannel_list, reason); +} + +void grpc_lb_subchannel_list_unref_for_connectivity_watch( + grpc_exec_ctx *exec_ctx, grpc_lb_subchannel_list *subchannel_list, + const char *reason) { + GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, subchannel_list->policy, reason); + grpc_lb_subchannel_list_unref(exec_ctx, subchannel_list, reason); +} + +static void subchannel_data_cancel_connectivity_watch( + grpc_exec_ctx *exec_ctx, grpc_lb_subchannel_data *sd, const char *reason) { + if (GRPC_TRACER_ON(*sd->subchannel_list->tracer)) { + gpr_log( + GPR_DEBUG, "[%s %p] subchannel list %p index %" PRIuPTR " of %" PRIuPTR + " (subchannel %p): canceling connectivity watch (%s)", + sd->subchannel_list->tracer->name, sd->subchannel_list->policy, + sd->subchannel_list, (size_t)(sd - sd->subchannel_list->subchannels), + sd->subchannel_list->num_subchannels, sd->subchannel, reason); + } + grpc_subchannel_notify_on_state_change(exec_ctx, sd->subchannel, NULL, NULL, + &sd->connectivity_changed_closure); +} + +void grpc_lb_subchannel_list_shutdown_and_unref( + grpc_exec_ctx *exec_ctx, grpc_lb_subchannel_list *subchannel_list, + const char *reason) { + if (GRPC_TRACER_ON(*subchannel_list->tracer)) { + gpr_log(GPR_DEBUG, "[%s %p] Shutting down subchannel_list %p (%s)", + subchannel_list->tracer->name, subchannel_list->policy, + subchannel_list, reason); + } + GPR_ASSERT(!subchannel_list->shutting_down); + subchannel_list->shutting_down = true; + for (size_t i = 0; i < subchannel_list->num_subchannels; i++) { + grpc_lb_subchannel_data *sd = &subchannel_list->subchannels[i]; + // If there's a pending notification for this subchannel, cancel it; + // the callback is responsible for unreffing the subchannel. + // Otherwise, unref the subchannel directly. + if (sd->connectivity_notification_pending) { + subchannel_data_cancel_connectivity_watch(exec_ctx, sd, reason); + } else if (sd->subchannel != NULL) { + grpc_lb_subchannel_data_unref_subchannel(exec_ctx, sd, reason); + } + } + grpc_lb_subchannel_list_unref(exec_ctx, subchannel_list, reason); +} diff --git a/src/core/ext/filters/client_channel/lb_policy/subchannel_list.h b/src/core/ext/filters/client_channel/lb_policy/subchannel_list.h new file mode 100644 index 0000000000..9d5984260f --- /dev/null +++ b/src/core/ext/filters/client_channel/lb_policy/subchannel_list.h @@ -0,0 +1,153 @@ +/* + * + * Copyright 2015 gRPC authors. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + * + */ + +#ifndef GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_LB_POLICY_SUBCHANNEL_LIST_H +#define GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_LB_POLICY_SUBCHANNEL_LIST_H + +#include "src/core/ext/filters/client_channel/lb_policy_registry.h" +#include "src/core/ext/filters/client_channel/subchannel.h" +#include "src/core/lib/debug/trace.h" +#include "src/core/lib/transport/connectivity_state.h" + +// TODO(roth): This code is intended to be shared between pick_first and +// round_robin. However, the interface needs more work to provide clean +// encapsulation. For example, the structs here have some fields that are +// only used in one of the two (e.g., the state counters in +// grpc_lb_subchannel_list and the prev_connectivity_state field in +// grpc_lb_subchannel_data are only used in round_robin, and the +// checking_subchannel field in grpc_lb_subchannel_list is only used by +// pick_first). Also, there is probably some code duplication between the +// connectivity state notification callback code in both pick_first and +// round_robin that could be refactored and moved here. In a future PR, +// need to clean this up. + +#ifdef __cplusplus +extern "C" { +#endif + +typedef struct grpc_lb_subchannel_list grpc_lb_subchannel_list; + +typedef struct { + /** backpointer to owning subchannel list */ + grpc_lb_subchannel_list *subchannel_list; + /** subchannel itself */ + grpc_subchannel *subchannel; + grpc_connected_subchannel *connected_subchannel; + /** Is a connectivity notification pending? */ + bool connectivity_notification_pending; + /** notification that connectivity has changed on subchannel */ + grpc_closure connectivity_changed_closure; + /** previous and current connectivity states. Updated by \a + * \a connectivity_changed_closure based on + * \a pending_connectivity_state_unsafe. */ + grpc_connectivity_state prev_connectivity_state; + grpc_connectivity_state curr_connectivity_state; + /** connectivity state to be updated by + * grpc_subchannel_notify_on_state_change(), not guarded by + * the combiner. To be copied to \a curr_connectivity_state by + * \a connectivity_changed_closure. */ + grpc_connectivity_state pending_connectivity_state_unsafe; + /** the subchannel's target user data */ + void *user_data; + /** vtable to operate over \a user_data */ + const grpc_lb_user_data_vtable *user_data_vtable; +} grpc_lb_subchannel_data; + +/// Unrefs the subchannel contained in sd. +void grpc_lb_subchannel_data_unref_subchannel(grpc_exec_ctx *exec_ctx, + grpc_lb_subchannel_data *sd, + const char *reason); + +/// Starts watching the connectivity state of the subchannel. +/// The connectivity_changed_cb callback must invoke either +/// grpc_lb_subchannel_data_stop_connectivity_watch() or again call +/// grpc_lb_subchannel_data_start_connectivity_watch(). +void grpc_lb_subchannel_data_start_connectivity_watch( + grpc_exec_ctx *exec_ctx, grpc_lb_subchannel_data *sd); + +/// Stops watching the connectivity state of the subchannel. +void grpc_lb_subchannel_data_stop_connectivity_watch( + grpc_exec_ctx *exec_ctx, grpc_lb_subchannel_data *sd); + +struct grpc_lb_subchannel_list { + /** backpointer to owning policy */ + grpc_lb_policy *policy; + + grpc_tracer_flag *tracer; + + /** all our subchannels */ + size_t num_subchannels; + grpc_lb_subchannel_data *subchannels; + + /** Index into subchannels of the one we're currently checking. + * Used when connecting to subchannels serially instead of in parallel. */ + // TODO(roth): When we have time, we can probably make this go away + // and compute the index dynamically by subtracting + // subchannel_list->subchannels from the subchannel_data pointer. + size_t checking_subchannel; + + /** how many subchannels are in state READY */ + size_t num_ready; + /** how many subchannels are in state TRANSIENT_FAILURE */ + size_t num_transient_failures; + /** how many subchannels are in state SHUTDOWN */ + size_t num_shutdown; + /** how many subchannels are in state IDLE */ + size_t num_idle; + + /** There will be one ref for each entry in subchannels for which there is a + * pending connectivity state watcher callback. */ + gpr_refcount refcount; + + /** Is this list shutting down? This may be true due to the shutdown of the + * policy itself or because a newer update has arrived while this one hadn't + * finished processing. */ + bool shutting_down; +}; + +grpc_lb_subchannel_list *grpc_lb_subchannel_list_create( + grpc_exec_ctx *exec_ctx, grpc_lb_policy *p, grpc_tracer_flag *tracer, + const grpc_lb_addresses *addresses, const grpc_lb_policy_args *args, + grpc_iomgr_cb_func connectivity_changed_cb); + +void grpc_lb_subchannel_list_ref(grpc_lb_subchannel_list *subchannel_list, + const char *reason); + +void grpc_lb_subchannel_list_unref(grpc_exec_ctx *exec_ctx, + grpc_lb_subchannel_list *subchannel_list, + const char *reason); + +/// Takes and releases refs needed for a connectivity notification. +/// This includes a ref to subchannel_list and a weak ref to the LB policy. +void grpc_lb_subchannel_list_ref_for_connectivity_watch( + grpc_lb_subchannel_list *subchannel_list, const char *reason); +void grpc_lb_subchannel_list_unref_for_connectivity_watch( + grpc_exec_ctx *exec_ctx, grpc_lb_subchannel_list *subchannel_list, + const char *reason); + +/// Mark subchannel_list as discarded. Unsubscribes all its subchannels. The +/// connectivity state notification callback will ultimately unref it. +void grpc_lb_subchannel_list_shutdown_and_unref( + grpc_exec_ctx *exec_ctx, grpc_lb_subchannel_list *subchannel_list, + const char *reason); + +#ifdef __cplusplus +} +#endif + +#endif /* GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_LB_POLICY_SUBCHANNEL_LIST_H */ diff --git a/src/core/ext/filters/client_channel/subchannel.h b/src/core/ext/filters/client_channel/subchannel.h index 46b29f1fe0..1cd73f3ff4 100644 --- a/src/core/ext/filters/client_channel/subchannel.h +++ b/src/core/ext/filters/client_channel/subchannel.h @@ -127,8 +127,8 @@ void grpc_connected_subchannel_process_transport_op( grpc_connectivity_state grpc_subchannel_check_connectivity( grpc_subchannel *channel, grpc_error **error); -/** call notify when the connectivity state of a channel changes from *state. - Updates *state with the new state of the channel */ +/** Calls notify when the connectivity state of a channel becomes different + from *state. Updates *state with the new state of the channel. */ void grpc_subchannel_notify_on_state_change( grpc_exec_ctx *exec_ctx, grpc_subchannel *channel, grpc_pollset_set *interested_parties, grpc_connectivity_state *state, diff --git a/src/core/ext/transport/chttp2/client/chttp2_connector.cc b/src/core/ext/transport/chttp2/client/chttp2_connector.cc index 202bcd47f5..74839f2156 100644 --- a/src/core/ext/transport/chttp2/client/chttp2_connector.cc +++ b/src/core/ext/transport/chttp2/client/chttp2_connector.cc @@ -115,6 +115,8 @@ static void on_handshake_done(grpc_exec_ctx *exec_ctx, void *arg, } memset(c->result, 0, sizeof(*c->result)); } else { + grpc_endpoint_delete_from_pollset_set(exec_ctx, args->endpoint, + c->args.interested_parties); c->result->transport = grpc_create_chttp2_transport(exec_ctx, args->args, args->endpoint, 1); GPR_ASSERT(c->result->transport); @@ -136,6 +138,8 @@ static void start_handshake_locked(grpc_exec_ctx *exec_ctx, c->handshake_mgr = grpc_handshake_manager_create(); grpc_handshakers_add(exec_ctx, HANDSHAKER_CLIENT, c->args.channel_args, c->handshake_mgr); + grpc_endpoint_add_to_pollset_set(exec_ctx, c->endpoint, + c->args.interested_parties); grpc_handshake_manager_do_handshake( exec_ctx, c->handshake_mgr, c->endpoint, c->args.channel_args, c->args.deadline, NULL /* acceptor */, on_handshake_done, c); diff --git a/src/core/ext/transport/chttp2/transport/chttp2_transport.cc b/src/core/ext/transport/chttp2/transport/chttp2_transport.cc index 9462d1085e..02fc53122d 100644 --- a/src/core/ext/transport/chttp2/transport/chttp2_transport.cc +++ b/src/core/ext/transport/chttp2/transport/chttp2_transport.cc @@ -54,7 +54,6 @@ #include "src/core/lib/transport/transport.h" #include "src/core/lib/transport/transport_impl.h" -#define DEFAULT_WINDOW 65535 #define DEFAULT_CONNECTION_WINDOW_TARGET (1024 * 1024) #define MAX_WINDOW 0x7fffffffu #define MAX_WRITE_BUFFER_SIZE (64 * 1024 * 1024) @@ -222,7 +221,7 @@ static void destruct_transport(grpc_exec_ctx *exec_ctx, t->write_cb_pool = next; } - t->flow_control.bdp_estimator.Destroy(); + t->flow_control.Destroy(); GRPC_ERROR_UNREF(t->closed_with_error); gpr_free(t->ping_acks); @@ -282,10 +281,6 @@ static void init_transport(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t, t->endpoint_reading = 1; t->next_stream_id = is_client ? 1 : 2; t->is_client = is_client; - t->flow_control.remote_window = DEFAULT_WINDOW; - t->flow_control.announced_window = DEFAULT_WINDOW; - t->flow_control.target_initial_window_size = DEFAULT_WINDOW; - t->flow_control.t = t; t->deframe_state = is_client ? GRPC_DTS_FH_0 : GRPC_DTS_CLIENT_PREFIX_0; t->is_first_frame = true; grpc_connectivity_state_init( @@ -325,8 +320,6 @@ static void init_transport(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t, keepalive_watchdog_fired_locked, t, grpc_combiner_scheduler(t->combiner)); - t->flow_control.bdp_estimator.Init(t->peer_string); - grpc_chttp2_goaway_parser_init(&t->goaway_parser); grpc_chttp2_hpack_parser_init(exec_ctx, &t->hpack_parser); @@ -350,8 +343,7 @@ static void init_transport(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t, window -- this should by rights be 0 */ t->force_send_settings = 1 << GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE; t->sent_local_settings = 0; - t->write_buffer_size = DEFAULT_WINDOW; - t->flow_control.enable_bdp_probe = true; + t->write_buffer_size = grpc_core::chttp2::kDefaultWindow; if (is_client) { grpc_slice_buffer_add(&t->outbuf, grpc_slice_from_copied_string( @@ -396,6 +388,8 @@ static void init_transport(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t, t->opt_target = GRPC_CHTTP2_OPTIMIZE_FOR_LATENCY; + bool enable_bdp = true; + if (channel_args) { for (i = 0; i < channel_args->num_args; i++) { if (0 == strcmp(channel_args->args[i].key, @@ -456,8 +450,7 @@ static void init_transport(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t, &channel_args->args[i], {0, 0, MAX_WRITE_BUFFER_SIZE}); } else if (0 == strcmp(channel_args->args[i].key, GRPC_ARG_HTTP2_BDP_PROBE)) { - t->flow_control.enable_bdp_probe = - grpc_channel_arg_get_integer(&channel_args->args[i], {1, 0, 1}); + enable_bdp = grpc_channel_arg_get_bool(&channel_args->args[i], true); } else if (0 == strcmp(channel_args->args[i].key, GRPC_ARG_KEEPALIVE_TIME_MS)) { const int value = grpc_channel_arg_get_integer( @@ -552,6 +545,8 @@ static void init_transport(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t, } } + t->flow_control.Init(exec_ctx, t, enable_bdp); + /* No pings allowed before receiving a header or data frame. */ t->ping_state.pings_before_data_required = 0; t->ping_state.is_delayed_ping_timer_set = false; @@ -572,15 +567,13 @@ static void init_transport(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t, t->keepalive_state = GRPC_CHTTP2_KEEPALIVE_STATE_DISABLED; } - if (t->flow_control.enable_bdp_probe) { + if (enable_bdp) { GRPC_CHTTP2_REF_TRANSPORT(t, "bdp_ping"); schedule_bdp_ping_locked(exec_ctx, t); - } - grpc_chttp2_act_on_flowctl_action( - exec_ctx, - grpc_chttp2_flowctl_get_action(exec_ctx, &t->flow_control, NULL), t, - NULL); + grpc_chttp2_act_on_flowctl_action( + exec_ctx, t->flow_control->PeriodicUpdate(exec_ctx), t, NULL); + } grpc_chttp2_initiate_write(exec_ctx, t, GRPC_CHTTP2_INITIATE_WRITE_INITIAL_WRITE); @@ -718,7 +711,7 @@ static int init_stream(grpc_exec_ctx *exec_ctx, grpc_transport *gt, post_destructive_reclaimer(exec_ctx, t); } - s->flow_control.s = s; + s->flow_control.Init(t->flow_control.get(), s); GPR_TIMER_END("init_stream", 0); return 0; @@ -769,7 +762,7 @@ static void destroy_stream_locked(grpc_exec_ctx *exec_ctx, void *sp, GRPC_ERROR_UNREF(s->write_closed_error); GRPC_ERROR_UNREF(s->byte_stream_error); - grpc_chttp2_flowctl_destroy_stream(&t->flow_control, &s->flow_control); + s->flow_control.Destroy(); GRPC_CHTTP2_UNREF_TRANSPORT(exec_ctx, t, "stream"); @@ -1638,13 +1631,10 @@ static void perform_stream_op_locked(grpc_exec_ctx *exec_ctx, void *stream_op, if (s->id != 0) { if (!s->read_closed) { already_received = s->frame_storage.length; - grpc_chttp2_flowctl_incoming_bs_update( - &t->flow_control, &s->flow_control, GRPC_HEADER_SIZE_IN_BYTES, - already_received); - grpc_chttp2_act_on_flowctl_action( - exec_ctx, grpc_chttp2_flowctl_get_action(exec_ctx, &t->flow_control, - &s->flow_control), - t, s); + s->flow_control->IncomingByteStreamUpdate(GRPC_HEADER_SIZE_IN_BYTES, + already_received); + grpc_chttp2_act_on_flowctl_action(exec_ctx, + s->flow_control->MakeAction(), t, s); } } grpc_chttp2_maybe_complete_recv_message(exec_ctx, t, s); @@ -2420,49 +2410,44 @@ static void end_all_the_calls(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t, * INPUT PROCESSING - PARSING */ -void grpc_chttp2_act_on_flowctl_action(grpc_exec_ctx *exec_ctx, - grpc_chttp2_flowctl_action action, - grpc_chttp2_transport *t, - grpc_chttp2_stream *s) { - switch (action.send_stream_update) { - case GRPC_CHTTP2_FLOWCTL_NO_ACTION_NEEDED: - break; - case GRPC_CHTTP2_FLOWCTL_UPDATE_IMMEDIATELY: - grpc_chttp2_mark_stream_writable(exec_ctx, t, s); - grpc_chttp2_initiate_write( - exec_ctx, t, GRPC_CHTTP2_INITIATE_WRITE_STREAM_FLOW_CONTROL); +template <class F> +static void WithUrgency(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t, + grpc_core::chttp2::FlowControlAction::Urgency urgency, + grpc_chttp2_initiate_write_reason reason, F action) { + switch (urgency) { + case grpc_core::chttp2::FlowControlAction::Urgency::NO_ACTION_NEEDED: break; - case GRPC_CHTTP2_FLOWCTL_QUEUE_UPDATE: - grpc_chttp2_mark_stream_writable(exec_ctx, t, s); + case grpc_core::chttp2::FlowControlAction::Urgency::UPDATE_IMMEDIATELY: + grpc_chttp2_initiate_write(exec_ctx, t, reason); + // fallthrough + case grpc_core::chttp2::FlowControlAction::Urgency::QUEUE_UPDATE: + action(); break; } - switch (action.send_transport_update) { - case GRPC_CHTTP2_FLOWCTL_NO_ACTION_NEEDED: - break; - case GRPC_CHTTP2_FLOWCTL_UPDATE_IMMEDIATELY: - grpc_chttp2_initiate_write( - exec_ctx, t, GRPC_CHTTP2_INITIATE_WRITE_TRANSPORT_FLOW_CONTROL); - break; - // this is the same as no action b/c every time the transport enters the - // writing path it will maybe do an update - case GRPC_CHTTP2_FLOWCTL_QUEUE_UPDATE: - break; - } - if (action.send_setting_update != GRPC_CHTTP2_FLOWCTL_NO_ACTION_NEEDED) { - if (action.initial_window_size > 0) { - queue_setting_update(exec_ctx, t, - GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE, - (uint32_t)action.initial_window_size); - } - if (action.max_frame_size > 0) { - queue_setting_update(exec_ctx, t, GRPC_CHTTP2_SETTINGS_MAX_FRAME_SIZE, - (uint32_t)action.max_frame_size); - } - if (action.send_setting_update == GRPC_CHTTP2_FLOWCTL_UPDATE_IMMEDIATELY) { - grpc_chttp2_initiate_write(exec_ctx, t, - GRPC_CHTTP2_INITIATE_WRITE_SEND_SETTINGS); - } - } +} + +void grpc_chttp2_act_on_flowctl_action( + grpc_exec_ctx *exec_ctx, const grpc_core::chttp2::FlowControlAction &action, + grpc_chttp2_transport *t, grpc_chttp2_stream *s) { + WithUrgency( + exec_ctx, t, action.send_stream_update(), + GRPC_CHTTP2_INITIATE_WRITE_STREAM_FLOW_CONTROL, + [exec_ctx, t, s]() { grpc_chttp2_mark_stream_writable(exec_ctx, t, s); }); + WithUrgency(exec_ctx, t, action.send_transport_update(), + GRPC_CHTTP2_INITIATE_WRITE_TRANSPORT_FLOW_CONTROL, []() {}); + WithUrgency(exec_ctx, t, action.send_initial_window_update(), + GRPC_CHTTP2_INITIATE_WRITE_SEND_SETTINGS, + [exec_ctx, t, &action]() { + queue_setting_update(exec_ctx, t, + GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE, + action.initial_window_size()); + }); + WithUrgency( + exec_ctx, t, action.send_max_frame_size_update(), + GRPC_CHTTP2_INITIATE_WRITE_SEND_SETTINGS, [exec_ctx, t, &action]() { + queue_setting_update(exec_ctx, t, GRPC_CHTTP2_SETTINGS_MAX_FRAME_SIZE, + action.max_frame_size()); + }); } static grpc_error *try_http_parsing(grpc_exec_ctx *exec_ctx, @@ -2518,7 +2503,7 @@ static void read_action_locked(grpc_exec_ctx *exec_ctx, void *tp, grpc_error *errors[3] = {GRPC_ERROR_REF(error), GRPC_ERROR_NONE, GRPC_ERROR_NONE}; for (; i < t->read_buffer.count && errors[1] == GRPC_ERROR_NONE; i++) { - t->flow_control.bdp_estimator->AddIncomingBytes( + t->flow_control->bdp_estimator()->AddIncomingBytes( (int64_t)GRPC_SLICE_LENGTH(t->read_buffer.slices[i])); errors[1] = grpc_chttp2_perform_read(exec_ctx, t, t->read_buffer.slices[i]); @@ -2535,8 +2520,8 @@ static void read_action_locked(grpc_exec_ctx *exec_ctx, void *tp, GPR_TIMER_END("reading_action.parse", 0); GPR_TIMER_BEGIN("post_parse_locked", 0); - if (t->flow_control.initial_window_update != 0) { - if (t->flow_control.initial_window_update > 0) { + if (t->initial_window_update != 0) { + if (t->initial_window_update > 0) { grpc_chttp2_stream *s; while (grpc_chttp2_list_pop_stalled_by_stream(t, &s)) { grpc_chttp2_mark_stream_writable(exec_ctx, t, s); @@ -2545,7 +2530,7 @@ static void read_action_locked(grpc_exec_ctx *exec_ctx, void *tp, GRPC_CHTTP2_INITIATE_WRITE_FLOW_CONTROL_UNSTALLED_BY_SETTING); } } - t->flow_control.initial_window_update = 0; + t->initial_window_update = 0; } GPR_TIMER_END("post_parse_locked", 0); } @@ -2568,10 +2553,8 @@ static void read_action_locked(grpc_exec_ctx *exec_ctx, void *tp, if (keep_reading) { grpc_endpoint_read(exec_ctx, t->ep, &t->read_buffer, &t->read_action_locked); - grpc_chttp2_act_on_flowctl_action( - exec_ctx, - grpc_chttp2_flowctl_get_action(exec_ctx, &t->flow_control, NULL), t, - NULL); + grpc_chttp2_act_on_flowctl_action(exec_ctx, t->flow_control->MakeAction(), + t, NULL); GRPC_CHTTP2_UNREF_TRANSPORT(exec_ctx, t, "keep_reading"); } else { GRPC_CHTTP2_UNREF_TRANSPORT(exec_ctx, t, "reading_action"); @@ -2588,7 +2571,7 @@ static void read_action_locked(grpc_exec_ctx *exec_ctx, void *tp, // that kicks off finishes, it's unreffed static void schedule_bdp_ping_locked(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t) { - t->flow_control.bdp_estimator->SchedulePing(); + t->flow_control->bdp_estimator()->SchedulePing(); send_ping_locked(exec_ctx, t, &t->start_bdp_ping_locked, &t->finish_bdp_ping_locked); } @@ -2604,7 +2587,7 @@ static void start_bdp_ping_locked(grpc_exec_ctx *exec_ctx, void *tp, if (t->keepalive_state == GRPC_CHTTP2_KEEPALIVE_STATE_WAITING) { grpc_timer_cancel(exec_ctx, &t->keepalive_ping_timer); } - t->flow_control.bdp_estimator->StartPing(); + t->flow_control->bdp_estimator()->StartPing(); } static void finish_bdp_ping_locked(grpc_exec_ctx *exec_ctx, void *tp, @@ -2618,7 +2601,10 @@ static void finish_bdp_ping_locked(grpc_exec_ctx *exec_ctx, void *tp, GRPC_CHTTP2_UNREF_TRANSPORT(exec_ctx, t, "bdp_ping"); return; } - grpc_millis next_ping = t->flow_control.bdp_estimator->CompletePing(exec_ctx); + grpc_millis next_ping = + t->flow_control->bdp_estimator()->CompletePing(exec_ctx); + grpc_chttp2_act_on_flowctl_action( + exec_ctx, t->flow_control->PeriodicUpdate(exec_ctx), t, nullptr); GPR_ASSERT(!t->have_next_bdp_ping_timer); t->have_next_bdp_ping_timer = true; grpc_timer_init(exec_ctx, &t->next_bdp_ping_timer, next_ping, @@ -2844,13 +2830,10 @@ static void incoming_byte_stream_next_locked(grpc_exec_ctx *exec_ctx, size_t cur_length = s->frame_storage.length; if (!s->read_closed) { - grpc_chttp2_flowctl_incoming_bs_update(&t->flow_control, &s->flow_control, - bs->next_action.max_size_hint, - cur_length); - grpc_chttp2_act_on_flowctl_action( - exec_ctx, grpc_chttp2_flowctl_get_action(exec_ctx, &t->flow_control, - &s->flow_control), - t, s); + s->flow_control->IncomingByteStreamUpdate(bs->next_action.max_size_hint, + cur_length); + grpc_chttp2_act_on_flowctl_action(exec_ctx, s->flow_control->MakeAction(), + t, s); } GPR_ASSERT(s->unprocessed_incoming_frames_buffer.length == 0); if (s->frame_storage.length > 0) { diff --git a/src/core/ext/transport/chttp2/transport/flow_control.cc b/src/core/ext/transport/chttp2/transport/flow_control.cc index 716cd71490..40545bc74b 100644 --- a/src/core/ext/transport/chttp2/transport/flow_control.cc +++ b/src/core/ext/transport/chttp2/transport/flow_control.cc @@ -16,7 +16,7 @@ * */ -#include "src/core/ext/transport/chttp2/transport/internal.h" +#include "src/core/ext/transport/chttp2/transport/flow_control.h" #include <inttypes.h> #include <limits.h> @@ -28,38 +28,15 @@ #include <grpc/support/string_util.h> #include <grpc/support/useful.h> +#include "src/core/ext/transport/chttp2/transport/internal.h" #include "src/core/lib/support/string.h" -static uint32_t grpc_chttp2_target_announced_window( - const grpc_chttp2_transport_flowctl* tfc); - -#ifndef NDEBUG - -typedef struct { - int64_t remote_window; - int64_t target_window; - int64_t announced_window; - int64_t remote_window_delta; - int64_t local_window_delta; - int64_t announced_window_delta; - uint32_t local_init_window; - uint32_t local_max_frame; -} shadow_flow_control; - -static void pretrace(shadow_flow_control* shadow_fc, - grpc_chttp2_transport_flowctl* tfc, - grpc_chttp2_stream_flowctl* sfc) { - shadow_fc->remote_window = tfc->remote_window; - shadow_fc->target_window = grpc_chttp2_target_announced_window(tfc); - shadow_fc->announced_window = tfc->announced_window; - if (sfc != NULL) { - shadow_fc->remote_window_delta = sfc->remote_window_delta; - shadow_fc->local_window_delta = sfc->local_window_delta; - shadow_fc->announced_window_delta = sfc->announced_window_delta; - } -} +namespace grpc_core { +namespace chttp2 { + +namespace { -#define TRACE_PADDING 30 +static constexpr const int kTracePadding = 30; static char* fmt_int64_diff_str(int64_t old_val, int64_t new_val) { char* str; @@ -68,7 +45,7 @@ static char* fmt_int64_diff_str(int64_t old_val, int64_t new_val) { } else { gpr_asprintf(&str, "%" PRId64 "", old_val); } - char* str_lp = gpr_leftpad(str, ' ', TRACE_PADDING); + char* str_lp = gpr_leftpad(str, ' ', kTracePadding); gpr_free(str); return str_lp; } @@ -80,47 +57,58 @@ static char* fmt_uint32_diff_str(uint32_t old_val, uint32_t new_val) { } else { gpr_asprintf(&str, "%" PRIu32 "", old_val); } - char* str_lp = gpr_leftpad(str, ' ', TRACE_PADDING); + char* str_lp = gpr_leftpad(str, ' ', kTracePadding); gpr_free(str); return str_lp; } +} // namespace + +void FlowControlTrace::Init(const char* reason, TransportFlowControl* tfc, + StreamFlowControl* sfc) { + tfc_ = tfc; + sfc_ = sfc; + reason_ = reason; + remote_window_ = tfc->remote_window(); + target_window_ = tfc->target_window(); + announced_window_ = tfc->announced_window(); + if (sfc != nullptr) { + remote_window_delta_ = sfc->remote_window_delta(); + local_window_delta_ = sfc->local_window_delta(); + announced_window_delta_ = sfc->announced_window_delta(); + } +} -static void posttrace(shadow_flow_control* shadow_fc, - grpc_chttp2_transport_flowctl* tfc, - grpc_chttp2_stream_flowctl* sfc, const char* reason) { +void FlowControlTrace::Finish() { uint32_t acked_local_window = - tfc->t->settings[GRPC_SENT_SETTINGS] - [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]; + tfc_->transport()->settings[GRPC_SENT_SETTINGS] + [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]; uint32_t remote_window = - tfc->t->settings[GRPC_PEER_SETTINGS] - [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]; - char* trw_str = - fmt_int64_diff_str(shadow_fc->remote_window, tfc->remote_window); - char* tlw_str = fmt_int64_diff_str(shadow_fc->target_window, - grpc_chttp2_target_announced_window(tfc)); + tfc_->transport()->settings[GRPC_PEER_SETTINGS] + [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]; + char* trw_str = fmt_int64_diff_str(remote_window_, tfc_->remote_window()); + char* tlw_str = fmt_int64_diff_str(target_window_, tfc_->target_window()); char* taw_str = - fmt_int64_diff_str(shadow_fc->announced_window, tfc->announced_window); + fmt_int64_diff_str(announced_window_, tfc_->announced_window()); char* srw_str; char* slw_str; char* saw_str; - if (sfc != NULL) { - srw_str = fmt_int64_diff_str(shadow_fc->remote_window_delta + remote_window, - sfc->remote_window_delta + remote_window); - slw_str = - fmt_int64_diff_str(shadow_fc->local_window_delta + acked_local_window, - sfc->local_window_delta + acked_local_window); - saw_str = fmt_int64_diff_str( - shadow_fc->announced_window_delta + acked_local_window, - sfc->announced_window_delta + acked_local_window); + if (sfc_ != nullptr) { + srw_str = fmt_int64_diff_str(remote_window_delta_ + remote_window, + sfc_->remote_window_delta() + remote_window); + slw_str = fmt_int64_diff_str(local_window_delta_ + acked_local_window, + local_window_delta_ + acked_local_window); + saw_str = fmt_int64_diff_str(announced_window_delta_ + acked_local_window, + announced_window_delta_ + acked_local_window); } else { - srw_str = gpr_leftpad("", ' ', TRACE_PADDING); - slw_str = gpr_leftpad("", ' ', TRACE_PADDING); - saw_str = gpr_leftpad("", ' ', TRACE_PADDING); + srw_str = gpr_leftpad("", ' ', kTracePadding); + slw_str = gpr_leftpad("", ' ', kTracePadding); + saw_str = gpr_leftpad("", ' ', kTracePadding); } gpr_log(GPR_DEBUG, "%p[%u][%s] | %s | trw:%s, ttw:%s, taw:%s, srw:%s, slw:%s, saw:%s", - tfc, sfc != NULL ? sfc->s->id : 0, tfc->t->is_client ? "cli" : "svr", - reason, trw_str, tlw_str, taw_str, srw_str, slw_str, saw_str); + tfc_, sfc_ != nullptr ? sfc_->stream()->id : 0, + tfc_->transport()->is_client ? "cli" : "svr", reason_, trw_str, + tlw_str, taw_str, srw_str, slw_str, saw_str); gpr_free(trw_str); gpr_free(tlw_str); gpr_free(taw_str); @@ -129,13 +117,13 @@ static void posttrace(shadow_flow_control* shadow_fc, gpr_free(saw_str); } -static const char* urgency_to_string(grpc_chttp2_flowctl_urgency urgency) { - switch (urgency) { - case GRPC_CHTTP2_FLOWCTL_NO_ACTION_NEEDED: +const char* FlowControlAction::UrgencyString(Urgency u) { + switch (u) { + case Urgency::NO_ACTION_NEEDED: return "no action"; - case GRPC_CHTTP2_FLOWCTL_UPDATE_IMMEDIATELY: + case Urgency::UPDATE_IMMEDIATELY: return "update immediately"; - case GRPC_CHTTP2_FLOWCTL_QUEUE_UPDATE: + case Urgency::QUEUE_UPDATE: return "queue update"; default: GPR_UNREACHABLE_CODE(return "unknown"); @@ -143,209 +131,132 @@ static const char* urgency_to_string(grpc_chttp2_flowctl_urgency urgency) { GPR_UNREACHABLE_CODE(return "unknown"); } -static void trace_action(grpc_chttp2_transport_flowctl* tfc, - grpc_chttp2_flowctl_action action) { +void FlowControlAction::Trace(grpc_chttp2_transport* t) const { char* iw_str = fmt_uint32_diff_str( - tfc->t->settings[GRPC_SENT_SETTINGS] - [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE], - action.initial_window_size); + t->settings[GRPC_SENT_SETTINGS][GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE], + initial_window_size_); char* mf_str = fmt_uint32_diff_str( - tfc->t->settings[GRPC_SENT_SETTINGS][GRPC_CHTTP2_SETTINGS_MAX_FRAME_SIZE], - action.max_frame_size); - gpr_log(GPR_DEBUG, "t[%s], s[%s], settings[%s] iw:%s mf:%s", - urgency_to_string(action.send_transport_update), - urgency_to_string(action.send_stream_update), - urgency_to_string(action.send_setting_update), iw_str, mf_str); + t->settings[GRPC_SENT_SETTINGS][GRPC_CHTTP2_SETTINGS_MAX_FRAME_SIZE], + max_frame_size_); + gpr_log(GPR_DEBUG, "t[%s], s[%s], iw:%s:%s mf:%s:%s", + UrgencyString(send_transport_update_), + UrgencyString(send_stream_update_), + UrgencyString(send_initial_window_update_), iw_str, + UrgencyString(send_max_frame_size_update_), mf_str); gpr_free(iw_str); gpr_free(mf_str); } -#define PRETRACE(tfc, sfc) \ - shadow_flow_control shadow_fc; \ - GRPC_FLOW_CONTROL_IF_TRACING(pretrace(&shadow_fc, tfc, sfc)) -#define POSTTRACE(tfc, sfc, reason) \ - GRPC_FLOW_CONTROL_IF_TRACING(posttrace(&shadow_fc, tfc, sfc, reason)) -#define TRACEACTION(tfc, action) \ - GRPC_FLOW_CONTROL_IF_TRACING(trace_action(tfc, action)) -#else -#define PRETRACE(tfc, sfc) -#define POSTTRACE(tfc, sfc, reason) -#define TRACEACTION(tfc, action) -#endif - -/* How many bytes of incoming flow control would we like to advertise */ -static uint32_t grpc_chttp2_target_announced_window( - const grpc_chttp2_transport_flowctl* tfc) { - return (uint32_t)GPR_MIN((int64_t)((1u << 31) - 1), - tfc->announced_stream_total_over_incoming_window + - tfc->target_initial_window_size); -} - -// we have sent data on the wire, we must track this in our bookkeeping for the -// remote peer's flow control. -void grpc_chttp2_flowctl_sent_data(grpc_chttp2_transport_flowctl* tfc, - grpc_chttp2_stream_flowctl* sfc, - int64_t size) { - PRETRACE(tfc, sfc); - tfc->remote_window -= size; - sfc->remote_window_delta -= size; - POSTTRACE(tfc, sfc, " data sent"); -} - -static void announced_window_delta_preupdate(grpc_chttp2_transport_flowctl* tfc, - grpc_chttp2_stream_flowctl* sfc) { - if (sfc->announced_window_delta > 0) { - tfc->announced_stream_total_over_incoming_window -= - sfc->announced_window_delta; - } else { - tfc->announced_stream_total_under_incoming_window += - -sfc->announced_window_delta; - } -} - -static void announced_window_delta_postupdate( - grpc_chttp2_transport_flowctl* tfc, grpc_chttp2_stream_flowctl* sfc) { - if (sfc->announced_window_delta > 0) { - tfc->announced_stream_total_over_incoming_window += - sfc->announced_window_delta; - } else { - tfc->announced_stream_total_under_incoming_window -= - -sfc->announced_window_delta; +TransportFlowControl::TransportFlowControl(grpc_exec_ctx* exec_ctx, + const grpc_chttp2_transport* t, + bool enable_bdp_probe) + : t_(t), + enable_bdp_probe_(enable_bdp_probe), + bdp_estimator_(t->peer_string), + pid_controller_(grpc_core::PidController::Args() + .set_gain_p(4) + .set_gain_i(8) + .set_gain_d(0) + .set_initial_control_value(TargetLogBdp()) + .set_min_control_value(-1) + .set_max_control_value(25) + .set_integral_range(10)), + last_pid_update_(grpc_exec_ctx_now(exec_ctx)) {} + +uint32_t TransportFlowControl::MaybeSendUpdate(bool writing_anyway) { + FlowControlTrace trace("t updt sent", this, nullptr); + const uint32_t target_announced_window = (const uint32_t)target_window(); + if ((writing_anyway || announced_window_ <= target_announced_window / 2) && + announced_window_ != target_announced_window) { + const uint32_t announce = (uint32_t)GPR_CLAMP( + target_announced_window - announced_window_, 0, UINT32_MAX); + announced_window_ += announce; + return announce; } + return 0; } -// We have received data from the wire. We must track this in our own flow -// control bookkeeping. -// Returns an error if the incoming frame violates our flow control. -grpc_error* grpc_chttp2_flowctl_recv_data(grpc_chttp2_transport_flowctl* tfc, - grpc_chttp2_stream_flowctl* sfc, - int64_t incoming_frame_size) { - uint32_t sent_init_window = - tfc->t->settings[GRPC_SENT_SETTINGS] - [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]; - uint32_t acked_init_window = - tfc->t->settings[GRPC_ACKED_SETTINGS] - [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]; - PRETRACE(tfc, sfc); - if (incoming_frame_size > tfc->announced_window) { +grpc_error* TransportFlowControl::ValidateRecvData( + int64_t incoming_frame_size) { + if (incoming_frame_size > announced_window_) { char* msg; gpr_asprintf(&msg, "frame of size %" PRId64 " overflows local window of %" PRId64, - incoming_frame_size, tfc->announced_window); + incoming_frame_size, announced_window_); grpc_error* err = GRPC_ERROR_CREATE_FROM_COPIED_STRING(msg); gpr_free(msg); return err; } + return GRPC_ERROR_NONE; +} - if (sfc != NULL) { - int64_t acked_stream_window = - sfc->announced_window_delta + acked_init_window; - int64_t sent_stream_window = sfc->announced_window_delta + sent_init_window; - if (incoming_frame_size > acked_stream_window) { - if (incoming_frame_size <= sent_stream_window) { - gpr_log( - GPR_ERROR, - "Incoming frame of size %" PRId64 - " exceeds local window size of %" PRId64 - ".\n" - "The (un-acked, future) window size would be %" PRId64 - " which is not exceeded.\n" - "This would usually cause a disconnection, but allowing it due to" - "broken HTTP2 implementations in the wild.\n" - "See (for example) https://github.com/netty/netty/issues/6520.", - incoming_frame_size, acked_stream_window, sent_stream_window); - } else { - char* msg; - gpr_asprintf(&msg, "frame of size %" PRId64 - " overflows local window of %" PRId64, - incoming_frame_size, acked_stream_window); - grpc_error* err = GRPC_ERROR_CREATE_FROM_COPIED_STRING(msg); - gpr_free(msg); - return err; - } - } +StreamFlowControl::StreamFlowControl(TransportFlowControl* tfc, + const grpc_chttp2_stream* s) + : tfc_(tfc), s_(s) {} - announced_window_delta_preupdate(tfc, sfc); - sfc->announced_window_delta -= incoming_frame_size; - announced_window_delta_postupdate(tfc, sfc); - sfc->local_window_delta -= incoming_frame_size; - } +grpc_error* StreamFlowControl::RecvData(int64_t incoming_frame_size) { + FlowControlTrace trace(" data recv", tfc_, this); - tfc->announced_window -= incoming_frame_size; + grpc_error* error = GRPC_ERROR_NONE; + error = tfc_->ValidateRecvData(incoming_frame_size); + if (error != GRPC_ERROR_NONE) return error; - POSTTRACE(tfc, sfc, " data recv"); - return GRPC_ERROR_NONE; -} - -// Returns a non zero announce integer if we should send a transport window -// update -uint32_t grpc_chttp2_flowctl_maybe_send_transport_update( - grpc_chttp2_transport_flowctl* tfc, bool writing_anyway) { - PRETRACE(tfc, NULL); - uint32_t target_announced_window = grpc_chttp2_target_announced_window(tfc); - uint32_t threshold_to_send_transport_window_update = - tfc->t->outbuf.count > 0 ? 3 * target_announced_window / 4 - : target_announced_window / 2; - if ((writing_anyway || - tfc->announced_window <= threshold_to_send_transport_window_update) && - tfc->announced_window != target_announced_window) { - uint32_t announce = (uint32_t)GPR_CLAMP( - target_announced_window - tfc->announced_window, 0, UINT32_MAX); - tfc->announced_window += announce; - POSTTRACE(tfc, NULL, "t updt sent"); - return announce; + uint32_t sent_init_window = + tfc_->transport()->settings[GRPC_SENT_SETTINGS] + [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]; + uint32_t acked_init_window = + tfc_->transport()->settings[GRPC_ACKED_SETTINGS] + [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]; + + int64_t acked_stream_window = announced_window_delta_ + acked_init_window; + int64_t sent_stream_window = announced_window_delta_ + sent_init_window; + if (incoming_frame_size > acked_stream_window) { + if (incoming_frame_size <= sent_stream_window) { + gpr_log(GPR_ERROR, + "Incoming frame of size %" PRId64 + " exceeds local window size of %" PRId64 + ".\n" + "The (un-acked, future) window size would be %" PRId64 + " which is not exceeded.\n" + "This would usually cause a disconnection, but allowing it due to" + "broken HTTP2 implementations in the wild.\n" + "See (for example) https://github.com/netty/netty/issues/6520.", + incoming_frame_size, acked_stream_window, sent_stream_window); + } else { + char* msg; + gpr_asprintf(&msg, "frame of size %" PRId64 + " overflows local window of %" PRId64, + incoming_frame_size, acked_stream_window); + grpc_error* err = GRPC_ERROR_CREATE_FROM_COPIED_STRING(msg); + gpr_free(msg); + return err; + } } - GRPC_FLOW_CONTROL_IF_TRACING( - gpr_log(GPR_DEBUG, "%p[0][%s] will not send transport update", tfc, - tfc->t->is_client ? "cli" : "svr")); - return 0; + + UpdateAnnouncedWindowDelta(tfc_, -incoming_frame_size); + local_window_delta_ -= incoming_frame_size; + tfc_->CommitRecvData(incoming_frame_size); + return GRPC_ERROR_NONE; } -// Returns a non zero announce integer if we should send a stream window update -uint32_t grpc_chttp2_flowctl_maybe_send_stream_update( - grpc_chttp2_transport_flowctl* tfc, grpc_chttp2_stream_flowctl* sfc) { - PRETRACE(tfc, sfc); - if (sfc->local_window_delta > sfc->announced_window_delta) { +uint32_t StreamFlowControl::MaybeSendUpdate() { + FlowControlTrace trace("s updt sent", tfc_, this); + if (local_window_delta_ > announced_window_delta_) { uint32_t announce = (uint32_t)GPR_CLAMP( - sfc->local_window_delta - sfc->announced_window_delta, 0, UINT32_MAX); - announced_window_delta_preupdate(tfc, sfc); - sfc->announced_window_delta += announce; - announced_window_delta_postupdate(tfc, sfc); - POSTTRACE(tfc, sfc, "s updt sent"); + local_window_delta_ - announced_window_delta_, 0, UINT32_MAX); + UpdateAnnouncedWindowDelta(tfc_, announce); return announce; } - GRPC_FLOW_CONTROL_IF_TRACING( - gpr_log(GPR_DEBUG, "%p[%u][%s] will not send stream update", tfc, - sfc->s->id, tfc->t->is_client ? "cli" : "svr")); return 0; } -// we have received a WINDOW_UPDATE frame for a transport -void grpc_chttp2_flowctl_recv_transport_update( - grpc_chttp2_transport_flowctl* tfc, uint32_t size) { - PRETRACE(tfc, NULL); - tfc->remote_window += size; - POSTTRACE(tfc, NULL, "t updt recv"); -} - -// we have received a WINDOW_UPDATE frame for a stream -void grpc_chttp2_flowctl_recv_stream_update(grpc_chttp2_transport_flowctl* tfc, - grpc_chttp2_stream_flowctl* sfc, - uint32_t size) { - PRETRACE(tfc, sfc); - sfc->remote_window_delta += size; - POSTTRACE(tfc, sfc, "s updt recv"); -} - -void grpc_chttp2_flowctl_incoming_bs_update(grpc_chttp2_transport_flowctl* tfc, - grpc_chttp2_stream_flowctl* sfc, - size_t max_size_hint, - size_t have_already) { - PRETRACE(tfc, sfc); +void StreamFlowControl::IncomingByteStreamUpdate(size_t max_size_hint, + size_t have_already) { + FlowControlTrace trace("app st recv", tfc_, this); uint32_t max_recv_bytes; uint32_t sent_init_window = - tfc->t->settings[GRPC_SENT_SETTINGS] - [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]; + tfc_->transport()->settings[GRPC_SENT_SETTINGS] + [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]; /* clamp max recv hint to an allowable size */ if (max_size_hint >= UINT32_MAX - sent_init_window) { @@ -363,65 +274,18 @@ void grpc_chttp2_flowctl_incoming_bs_update(grpc_chttp2_transport_flowctl* tfc, /* add some small lookahead to keep pipelines flowing */ GPR_ASSERT(max_recv_bytes <= UINT32_MAX - sent_init_window); - if (sfc->local_window_delta < max_recv_bytes) { + if (local_window_delta_ < max_recv_bytes) { uint32_t add_max_recv_bytes = - (uint32_t)(max_recv_bytes - sfc->local_window_delta); - sfc->local_window_delta += add_max_recv_bytes; - } - POSTTRACE(tfc, sfc, "app st recv"); -} - -void grpc_chttp2_flowctl_destroy_stream(grpc_chttp2_transport_flowctl* tfc, - grpc_chttp2_stream_flowctl* sfc) { - announced_window_delta_preupdate(tfc, sfc); -} - -// Returns an urgency with which to make an update -static grpc_chttp2_flowctl_urgency delta_is_significant( - const grpc_chttp2_transport_flowctl* tfc, int32_t value, - grpc_chttp2_setting_id setting_id) { - int64_t delta = (int64_t)value - - (int64_t)tfc->t->settings[GRPC_LOCAL_SETTINGS][setting_id]; - // TODO(ncteisen): tune this - if (delta != 0 && (delta <= -value / 5 || delta >= value / 5)) { - return GRPC_CHTTP2_FLOWCTL_QUEUE_UPDATE; - } else { - return GRPC_CHTTP2_FLOWCTL_NO_ACTION_NEEDED; - } -} - -// Takes in a target and uses the pid controller to return a stabilized -// guess at the new bdp. -static double get_pid_controller_guess(grpc_exec_ctx* exec_ctx, - grpc_chttp2_transport_flowctl* tfc, - double target) { - grpc_millis now = grpc_exec_ctx_now(exec_ctx); - if (!tfc->pid_controller_initialized) { - tfc->last_pid_update = now; - tfc->pid_controller_initialized = true; - tfc->pid_controller.Init(grpc_core::PidController::Args() - .set_gain_p(4) - .set_gain_i(8) - .set_gain_d(0) - .set_initial_control_value(target) - .set_min_control_value(-1) - .set_max_control_value(25) - .set_integral_range(10)); - return pow(2, target); + (uint32_t)(max_recv_bytes - local_window_delta_); + local_window_delta_ += add_max_recv_bytes; } - double bdp_error = target - tfc->pid_controller->last_control_value(); - double dt = (double)(now - tfc->last_pid_update) * 1e-3; - double log2_bdp_guess = tfc->pid_controller->Update(bdp_error, dt); - tfc->last_pid_update = now; - return pow(2, log2_bdp_guess); } // Take in a target and modifies it based on the memory pressure of the system -static double get_target_under_memory_pressure( - grpc_chttp2_transport_flowctl* tfc, double target) { +static double AdjustForMemoryPressure(grpc_resource_quota* quota, + double target) { // do not increase window under heavy memory pressure. - double memory_pressure = grpc_resource_quota_get_memory_pressure( - grpc_resource_user_quota(grpc_endpoint_get_resource_user(tfc->t->ep))); + double memory_pressure = grpc_resource_quota_get_memory_pressure(quota); static const double kLowMemPressure = 0.1; static const double kZeroTarget = 22; static const double kHighMemPressure = 0.8; @@ -436,75 +300,82 @@ static double get_target_under_memory_pressure( return target; } -grpc_chttp2_flowctl_action grpc_chttp2_flowctl_get_action( - grpc_exec_ctx* exec_ctx, grpc_chttp2_transport_flowctl* tfc, - grpc_chttp2_stream_flowctl* sfc) { - grpc_chttp2_flowctl_action action; - memset(&action, 0, sizeof(action)); +double TransportFlowControl::TargetLogBdp() { + return AdjustForMemoryPressure( + grpc_resource_user_quota(grpc_endpoint_get_resource_user(t_->ep)), + 1 + log2(bdp_estimator_.EstimateBdp())); +} + +double TransportFlowControl::SmoothLogBdp(grpc_exec_ctx* exec_ctx, + double value) { + grpc_millis now = grpc_exec_ctx_now(exec_ctx); + double bdp_error = value - pid_controller_.last_control_value(); + const double dt = (double)(now - last_pid_update_) * 1e-3; + last_pid_update_ = now; + return pid_controller_.Update(bdp_error, dt); +} + +FlowControlAction::Urgency TransportFlowControl::DeltaUrgency( + int32_t value, grpc_chttp2_setting_id setting_id) { + int64_t delta = + (int64_t)value - (int64_t)t_->settings[GRPC_LOCAL_SETTINGS][setting_id]; // TODO(ncteisen): tune this - if (sfc != NULL && !sfc->s->read_closed) { - uint32_t sent_init_window = - tfc->t->settings[GRPC_SENT_SETTINGS] - [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]; - if ((int64_t)sfc->local_window_delta > - (int64_t)sfc->announced_window_delta && - (int64_t)sfc->announced_window_delta + sent_init_window <= - sent_init_window / 2) { - action.send_stream_update = GRPC_CHTTP2_FLOWCTL_UPDATE_IMMEDIATELY; - } else if (sfc->local_window_delta > sfc->announced_window_delta) { - action.send_stream_update = GRPC_CHTTP2_FLOWCTL_QUEUE_UPDATE; - } + if (delta != 0 && (delta <= -value / 5 || delta >= value / 5)) { + return FlowControlAction::Urgency::QUEUE_UPDATE; + } else { + return FlowControlAction::Urgency::NO_ACTION_NEEDED; } - if (tfc->enable_bdp_probe) { +} + +FlowControlAction TransportFlowControl::PeriodicUpdate( + grpc_exec_ctx* exec_ctx) { + FlowControlAction action; + if (enable_bdp_probe_) { // get bdp estimate and update initial_window accordingly. - int64_t estimate = -1; - if (tfc->bdp_estimator->EstimateBdp(&estimate)) { - double target = 1 + log2((double)estimate); - - // target might change based on how much memory pressure we are under - // TODO(ncteisen): experiment with setting target to be huge under low - // memory pressure. - target = get_target_under_memory_pressure(tfc, target); - - // run our target through the pid controller to stabilize change. - // TODO(ncteisen): experiment with other controllers here. - double bdp_guess = get_pid_controller_guess(exec_ctx, tfc, target); - - // Though initial window 'could' drop to 0, we keep the floor at 128 - tfc->target_initial_window_size = - (int32_t)GPR_CLAMP(bdp_guess, 128, INT32_MAX); - - grpc_chttp2_flowctl_urgency init_window_update_urgency = - delta_is_significant(tfc, tfc->target_initial_window_size, - GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE); - if (init_window_update_urgency != GRPC_CHTTP2_FLOWCTL_NO_ACTION_NEEDED) { - action.send_setting_update = init_window_update_urgency; - action.initial_window_size = (uint32_t)tfc->target_initial_window_size; - } - } + // target might change based on how much memory pressure we are under + // TODO(ncteisen): experiment with setting target to be huge under low + // memory pressure. + const double target = pow(2, SmoothLogBdp(exec_ctx, TargetLogBdp())); + + // Though initial window 'could' drop to 0, we keep the floor at 128 + target_initial_window_size_ = (int32_t)GPR_CLAMP(target, 128, INT32_MAX); + + action.set_send_initial_window_update( + DeltaUrgency(target_initial_window_size_, + GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE), + target_initial_window_size_); // get bandwidth estimate and update max_frame accordingly. - double bw_dbl = -1; - if (tfc->bdp_estimator->EstimateBandwidth(&bw_dbl)) { - // we target the max of BDP or bandwidth in microseconds. - int32_t frame_size = (int32_t)GPR_CLAMP( - GPR_MAX((int32_t)GPR_CLAMP(bw_dbl, 0, INT_MAX) / 1000, - tfc->target_initial_window_size), - 16384, 16777215); - grpc_chttp2_flowctl_urgency frame_size_urgency = delta_is_significant( - tfc, frame_size, GRPC_CHTTP2_SETTINGS_MAX_FRAME_SIZE); - if (frame_size_urgency != GRPC_CHTTP2_FLOWCTL_NO_ACTION_NEEDED) { - if (frame_size_urgency > action.send_setting_update) { - action.send_setting_update = frame_size_urgency; - } - action.max_frame_size = (uint32_t)frame_size; - } - } + double bw_dbl = bdp_estimator_.EstimateBandwidth(); + // we target the max of BDP or bandwidth in microseconds. + int32_t frame_size = (int32_t)GPR_CLAMP( + GPR_MAX((int32_t)GPR_CLAMP(bw_dbl, 0, INT_MAX) / 1000, + target_initial_window_size_), + 16384, 16777215); + action.set_send_max_frame_size_update( + DeltaUrgency(frame_size, GRPC_CHTTP2_SETTINGS_MAX_FRAME_SIZE), + frame_size); } - uint32_t target_announced_window = grpc_chttp2_target_announced_window(tfc); - if (tfc->announced_window < target_announced_window / 2) { - action.send_transport_update = GRPC_CHTTP2_FLOWCTL_UPDATE_IMMEDIATELY; + return UpdateAction(action); +} + +FlowControlAction StreamFlowControl::UpdateAction(FlowControlAction action) { + // TODO(ncteisen): tune this + if (!s_->read_closed) { + uint32_t sent_init_window = + tfc_->transport()->settings[GRPC_SENT_SETTINGS] + [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]; + if (local_window_delta_ > announced_window_delta_ && + announced_window_delta_ + sent_init_window <= sent_init_window / 2) { + action.set_send_stream_update( + FlowControlAction::Urgency::UPDATE_IMMEDIATELY); + } else if (local_window_delta_ > announced_window_delta_) { + action.set_send_stream_update(FlowControlAction::Urgency::QUEUE_UPDATE); + } } - TRACEACTION(tfc, action); + return action; } + +} // namespace chttp2 +} // namespace grpc_core diff --git a/src/core/ext/transport/chttp2/transport/flow_control.h b/src/core/ext/transport/chttp2/transport/flow_control.h new file mode 100644 index 0000000000..7dd348ed5f --- /dev/null +++ b/src/core/ext/transport/chttp2/transport/flow_control.h @@ -0,0 +1,336 @@ +/* + * + * Copyright 2017 gRPC authors. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + * + */ + +#ifndef GRPC_CORE_EXT_TRANSPORT_CHTTP2_TRANSPORT_FLOW_CONTROL_H +#define GRPC_CORE_EXT_TRANSPORT_CHTTP2_TRANSPORT_FLOW_CONTROL_H + +#include <stdint.h> + +#include <grpc/support/useful.h> +#include "src/core/ext/transport/chttp2/transport/http2_settings.h" +#include "src/core/lib/support/manual_constructor.h" +#include "src/core/lib/transport/bdp_estimator.h" +#include "src/core/lib/transport/pid_controller.h" + +struct grpc_chttp2_transport; +struct grpc_chttp2_stream; + +extern "C" grpc_tracer_flag grpc_flowctl_trace; + +namespace grpc { +namespace testing { +class TrickledCHTTP2; // to make this a friend +} // namespace testing +} // namespace grpc + +namespace grpc_core { +namespace chttp2 { + +static constexpr uint32_t kDefaultWindow = 65535; + +class TransportFlowControl; +class StreamFlowControl; + +class FlowControlAction { + public: + enum class Urgency : uint8_t { + // Nothing to be done. + NO_ACTION_NEEDED = 0, + // Initiate a write to update the initial window immediately. + UPDATE_IMMEDIATELY, + // Push the flow control update into a send buffer, to be sent + // out the next time a write is initiated. + QUEUE_UPDATE, + }; + + Urgency send_stream_update() const { return send_stream_update_; } + Urgency send_transport_update() const { return send_transport_update_; } + Urgency send_initial_window_update() const { + return send_initial_window_update_; + } + Urgency send_max_frame_size_update() const { + return send_max_frame_size_update_; + } + uint32_t initial_window_size() const { return initial_window_size_; } + uint32_t max_frame_size() const { return max_frame_size_; } + + FlowControlAction& set_send_stream_update(Urgency u) { + send_stream_update_ = u; + return *this; + } + FlowControlAction& set_send_transport_update(Urgency u) { + send_transport_update_ = u; + return *this; + } + FlowControlAction& set_send_initial_window_update(Urgency u, + uint32_t update) { + send_initial_window_update_ = u; + initial_window_size_ = update; + return *this; + } + FlowControlAction& set_send_max_frame_size_update(Urgency u, + uint32_t update) { + send_max_frame_size_update_ = u; + max_frame_size_ = update; + return *this; + } + + static const char* UrgencyString(Urgency u); + void Trace(grpc_chttp2_transport* t) const; + + private: + Urgency send_stream_update_ = Urgency::NO_ACTION_NEEDED; + Urgency send_transport_update_ = Urgency::NO_ACTION_NEEDED; + Urgency send_initial_window_update_ = Urgency::NO_ACTION_NEEDED; + Urgency send_max_frame_size_update_ = Urgency::NO_ACTION_NEEDED; + uint32_t initial_window_size_ = 0; + uint32_t max_frame_size_ = 0; +}; + +class FlowControlTrace { + public: + FlowControlTrace(const char* reason, TransportFlowControl* tfc, + StreamFlowControl* sfc) { + if (enabled_) Init(reason, tfc, sfc); + } + + ~FlowControlTrace() { + if (enabled_) Finish(); + } + + private: + void Init(const char* reason, TransportFlowControl* tfc, + StreamFlowControl* sfc); + void Finish(); + + const bool enabled_ = GRPC_TRACER_ON(grpc_flowctl_trace); + + TransportFlowControl* tfc_; + StreamFlowControl* sfc_; + const char* reason_; + int64_t remote_window_; + int64_t target_window_; + int64_t announced_window_; + int64_t remote_window_delta_; + int64_t local_window_delta_; + int64_t announced_window_delta_; +}; + +class TransportFlowControl { + public: + TransportFlowControl(grpc_exec_ctx* exec_ctx, const grpc_chttp2_transport* t, + bool enable_bdp_probe); + ~TransportFlowControl() {} + + bool bdp_probe() const { return enable_bdp_probe_; } + + // returns an announce if we should send a transport update to our peer, + // else returns zero; writing_anyway indicates if a write would happen + // regardless of the send - if it is false and this function returns non-zero, + // this announce will cause a write to occur + uint32_t MaybeSendUpdate(bool writing_anyway); + + // Reads the flow control data and returns and actionable struct that will + // tell chttp2 exactly what it needs to do + FlowControlAction MakeAction() { return UpdateAction(FlowControlAction()); } + + // Call periodically (at a low-ish rate, 100ms - 10s makes sense) + // to perform more complex flow control calculations and return an action + // to let chttp2 change its parameters + FlowControlAction PeriodicUpdate(grpc_exec_ctx* exec_ctx); + + void StreamSentData(int64_t size) { remote_window_ -= size; } + + grpc_error* ValidateRecvData(int64_t incoming_frame_size); + void CommitRecvData(int64_t incoming_frame_size) { + announced_window_ -= incoming_frame_size; + } + + grpc_error* RecvData(int64_t incoming_frame_size) { + FlowControlTrace trace(" data recv", this, nullptr); + grpc_error* error = ValidateRecvData(incoming_frame_size); + if (error != GRPC_ERROR_NONE) return error; + CommitRecvData(incoming_frame_size); + return GRPC_ERROR_NONE; + } + + // we have received a WINDOW_UPDATE frame for a transport + void RecvUpdate(uint32_t size) { + FlowControlTrace trace("t updt recv", this, nullptr); + remote_window_ += size; + } + + int64_t remote_window() const { return remote_window_; } + int64_t target_window() const { + return (uint32_t)GPR_MIN((int64_t)((1u << 31) - 1), + announced_stream_total_over_incoming_window_ + + target_initial_window_size_); + } + int64_t announced_window() const { return announced_window_; } + + const grpc_chttp2_transport* transport() const { return t_; } + + void PreUpdateAnnouncedWindowOverIncomingWindow(int64_t delta) { + if (delta > 0) { + announced_stream_total_over_incoming_window_ -= delta; + } else { + announced_stream_total_under_incoming_window_ += -delta; + } + } + + void PostUpdateAnnouncedWindowOverIncomingWindow(int64_t delta) { + if (delta > 0) { + announced_stream_total_over_incoming_window_ += delta; + } else { + announced_stream_total_under_incoming_window_ -= -delta; + } + } + + BdpEstimator* bdp_estimator() { return &bdp_estimator_; } + + void TestOnlyForceHugeWindow() { + announced_window_ = 1024 * 1024 * 1024; + remote_window_ = 1024 * 1024 * 1024; + } + + private: + friend class ::grpc::testing::TrickledCHTTP2; + double TargetLogBdp(); + double SmoothLogBdp(grpc_exec_ctx* exec_ctx, double value); + FlowControlAction::Urgency DeltaUrgency(int32_t value, + grpc_chttp2_setting_id setting_id); + + FlowControlAction UpdateAction(FlowControlAction action) { + if (announced_window_ < target_window() / 2) { + action.set_send_transport_update( + FlowControlAction::Urgency::UPDATE_IMMEDIATELY); + } + return action; + } + + const grpc_chttp2_transport* const t_; + + /** Our bookkeeping for the remote peer's available window */ + int64_t remote_window_ = kDefaultWindow; + + /** calculating what we should give for local window: + we track the total amount of flow control over initial window size + across all streams: this is data that we want to receive right now (it + has an outstanding read) + and the total amount of flow control under initial window size across all + streams: this is data we've read early + we want to adjust incoming_window such that: + incoming_window = total_over - max(bdp - total_under, 0) */ + int64_t announced_stream_total_over_incoming_window_ = 0; + int64_t announced_stream_total_under_incoming_window_ = 0; + + /** This is out window according to what we have sent to our remote peer. The + * difference between this and target window is what we use to decide when + * to send WINDOW_UPDATE frames. */ + int64_t announced_window_ = kDefaultWindow; + + int32_t target_initial_window_size_ = kDefaultWindow; + + /** should we probe bdp? */ + const bool enable_bdp_probe_; + + /* bdp estimation */ + grpc_core::BdpEstimator bdp_estimator_; + + /* pid controller */ + grpc_core::PidController pid_controller_; + grpc_millis last_pid_update_ = 0; +}; + +class StreamFlowControl { + public: + StreamFlowControl(TransportFlowControl* tfc, const grpc_chttp2_stream* s); + ~StreamFlowControl() { + tfc_->PreUpdateAnnouncedWindowOverIncomingWindow(announced_window_delta_); + } + + FlowControlAction UpdateAction(FlowControlAction action); + FlowControlAction MakeAction() { return UpdateAction(tfc_->MakeAction()); } + + // we have sent data on the wire, we must track this in our bookkeeping for + // the remote peer's flow control. + void SentData(int64_t outgoing_frame_size) { + FlowControlTrace tracer(" data sent", tfc_, this); + tfc_->StreamSentData(outgoing_frame_size); + remote_window_delta_ -= outgoing_frame_size; + } + + // we have received data from the wire + grpc_error* RecvData(int64_t incoming_frame_size); + + // returns an announce if we should send a stream update to our peer, else + // returns zero + uint32_t MaybeSendUpdate(); + + // we have received a WINDOW_UPDATE frame for a stream + void RecvUpdate(uint32_t size) { + FlowControlTrace trace("s updt recv", tfc_, this); + remote_window_delta_ += size; + } + + // the application is asking for a certain amount of bytes + void IncomingByteStreamUpdate(size_t max_size_hint, size_t have_already); + + int64_t remote_window_delta() const { return remote_window_delta_; } + int64_t local_window_delta() const { return local_window_delta_; } + int64_t announced_window_delta() const { return announced_window_delta_; } + + const grpc_chttp2_stream* stream() const { return s_; } + + void TestOnlyForceHugeWindow() { + announced_window_delta_ = 1024 * 1024 * 1024; + local_window_delta_ = 1024 * 1024 * 1024; + remote_window_delta_ = 1024 * 1024 * 1024; + } + + private: + friend class ::grpc::testing::TrickledCHTTP2; + TransportFlowControl* const tfc_; + const grpc_chttp2_stream* const s_; + + void UpdateAnnouncedWindowDelta(TransportFlowControl* tfc, int64_t change) { + tfc->PreUpdateAnnouncedWindowOverIncomingWindow(announced_window_delta_); + announced_window_delta_ += change; + tfc->PostUpdateAnnouncedWindowOverIncomingWindow(announced_window_delta_); + } + + /** window available for us to send to peer, over or under the initial + * window + * size of the transport... ie: + * remote_window = remote_window_delta + transport.initial_window_size */ + int64_t remote_window_delta_ = 0; + + /** window available for peer to send to us (as a delta on + * transport.initial_window_size) + * local_window = local_window_delta + transport.initial_window_size */ + int64_t local_window_delta_ = 0; + + /** window available for peer to send to us over this stream that we have + * announced to the peer */ + int64_t announced_window_delta_ = 0; +}; + +} // namespace chttp2 +} // namespace grpc_core + +#endif diff --git a/src/core/ext/transport/chttp2/transport/frame_settings.cc b/src/core/ext/transport/chttp2/transport/frame_settings.cc index 2995bf7310..db0245bb57 100644 --- a/src/core/ext/transport/chttp2/transport/frame_settings.cc +++ b/src/core/ext/transport/chttp2/transport/frame_settings.cc @@ -202,13 +202,13 @@ grpc_error *grpc_chttp2_settings_parser_parse(grpc_exec_ctx *exec_ctx, void *p, } if (id == GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE && parser->incoming_settings[id] != parser->value) { - t->flow_control.initial_window_update += + t->initial_window_update += (int64_t)parser->value - parser->incoming_settings[id]; if (GRPC_TRACER_ON(grpc_http_trace) || GRPC_TRACER_ON(grpc_flowctl_trace)) { gpr_log(GPR_DEBUG, "%p[%s] adding %d for initial_window change", t, t->is_client ? "cli" : "svr", - (int)t->flow_control.initial_window_update); + (int)t->initial_window_update); } } parser->incoming_settings[id] = parser->value; diff --git a/src/core/ext/transport/chttp2/transport/frame_window_update.cc b/src/core/ext/transport/chttp2/transport/frame_window_update.cc index c9ab8d1b50..15eaf59285 100644 --- a/src/core/ext/transport/chttp2/transport/frame_window_update.cc +++ b/src/core/ext/transport/chttp2/transport/frame_window_update.cc @@ -96,8 +96,7 @@ grpc_error *grpc_chttp2_window_update_parser_parse( if (t->incoming_stream_id != 0) { if (s != NULL) { - grpc_chttp2_flowctl_recv_stream_update( - &t->flow_control, &s->flow_control, received_update); + s->flow_control->RecvUpdate(received_update); if (grpc_chttp2_list_remove_stalled_by_stream(t, s)) { grpc_chttp2_mark_stream_writable(exec_ctx, t, s); grpc_chttp2_initiate_write( @@ -106,10 +105,9 @@ grpc_error *grpc_chttp2_window_update_parser_parse( } } } else { - bool was_zero = t->flow_control.remote_window <= 0; - grpc_chttp2_flowctl_recv_transport_update(&t->flow_control, - received_update); - bool is_zero = t->flow_control.remote_window <= 0; + bool was_zero = t->flow_control->remote_window() <= 0; + t->flow_control->RecvUpdate(received_update); + bool is_zero = t->flow_control->remote_window() <= 0; if (was_zero && !is_zero) { grpc_chttp2_initiate_write( exec_ctx, t, diff --git a/src/core/ext/transport/chttp2/transport/hpack_parser.cc b/src/core/ext/transport/chttp2/transport/hpack_parser.cc index 3d1df19bc3..7c17229122 100644 --- a/src/core/ext/transport/chttp2/transport/hpack_parser.cc +++ b/src/core/ext/transport/chttp2/transport/hpack_parser.cc @@ -33,6 +33,7 @@ #include "src/core/lib/debug/stats.h" #include "src/core/lib/profiling/timers.h" #include "src/core/lib/slice/slice_internal.h" +#include "src/core/lib/slice/slice_string_helpers.h" #include "src/core/lib/support/string.h" #include "src/core/lib/transport/http2_errors.h" @@ -650,9 +651,14 @@ static const uint8_t inverse_base64[256] = { /* emission helpers */ static grpc_error *on_hdr(grpc_exec_ctx *exec_ctx, grpc_chttp2_hpack_parser *p, grpc_mdelem md, int add_to_table) { - if (GRPC_TRACER_ON(grpc_http_trace) && !GRPC_MDELEM_IS_INTERNED(md)) { + if (GRPC_TRACER_ON(grpc_http_trace)) { char *k = grpc_slice_to_c_string(GRPC_MDKEY(md)); - char *v = grpc_slice_to_c_string(GRPC_MDVALUE(md)); + char *v = NULL; + if (grpc_is_binary_header(GRPC_MDKEY(md))) { + v = grpc_dump_slice(GRPC_MDVALUE(md), GPR_DUMP_HEX); + } else { + v = grpc_slice_to_c_string(GRPC_MDVALUE(md)); + } gpr_log( GPR_DEBUG, "Decode: '%s: %s', elem_interned=%d [%d], k_interned=%d, v_interned=%d", diff --git a/src/core/ext/transport/chttp2/transport/internal.h b/src/core/ext/transport/chttp2/transport/internal.h index c75f813393..9e0796e820 100644 --- a/src/core/ext/transport/chttp2/transport/internal.h +++ b/src/core/ext/transport/chttp2/transport/internal.h @@ -22,6 +22,7 @@ #include <assert.h> #include <stdbool.h> +#include "src/core/ext/transport/chttp2/transport/flow_control.h" #include "src/core/ext/transport/chttp2/transport/frame.h" #include "src/core/ext/transport/chttp2/transport/frame_data.h" #include "src/core/ext/transport/chttp2/transport/frame_goaway.h" @@ -38,9 +39,7 @@ #include "src/core/lib/iomgr/endpoint.h" #include "src/core/lib/iomgr/timer.h" #include "src/core/lib/support/manual_constructor.h" -#include "src/core/lib/transport/bdp_estimator.h" #include "src/core/lib/transport/connectivity_state.h" -#include "src/core/lib/transport/pid_controller.h" #include "src/core/lib/transport/transport_impl.h" #ifdef __cplusplus @@ -238,48 +237,6 @@ typedef enum { GRPC_CHTTP2_KEEPALIVE_STATE_DISABLED, } grpc_chttp2_keepalive_state; -typedef struct { - /** initial window change. This is tracked as we parse settings frames from - * the remote peer. If there is a positive delta, then we will make all - * streams readable since they may have become unstalled */ - int64_t initial_window_update; - - /** Our bookkeeping for the remote peer's available window */ - int64_t remote_window; - - /** calculating what we should give for local window: - we track the total amount of flow control over initial window size - across all streams: this is data that we want to receive right now (it - has an outstanding read) - and the total amount of flow control under initial window size across all - streams: this is data we've read early - we want to adjust incoming_window such that: - incoming_window = total_over - max(bdp - total_under, 0) */ - int64_t announced_stream_total_over_incoming_window; - int64_t announced_stream_total_under_incoming_window; - - /** This is out window according to what we have sent to our remote peer. The - * difference between this and target window is what we use to decide when - * to send WINDOW_UPDATE frames. */ - int64_t announced_window; - - int32_t target_initial_window_size; - - /** should we probe bdp? */ - bool enable_bdp_probe; - - /* bdp estimation */ - grpc_core::ManualConstructor<grpc_core::BdpEstimator> bdp_estimator; - - /* pid controller */ - bool pid_controller_initialized; - grpc_core::ManualConstructor<grpc_core::PidController> pid_controller; - grpc_millis last_pid_update; - - // pointer back to transport for tracing - const grpc_chttp2_transport *t; -} grpc_chttp2_transport_flowctl; - struct grpc_chttp2_transport { grpc_transport base; /* must be first */ gpr_refcount refs; @@ -395,7 +352,12 @@ struct grpc_chttp2_transport { /** parser for goaway frames */ grpc_chttp2_goaway_parser goaway_parser; - grpc_chttp2_transport_flowctl flow_control; + grpc_core::ManualConstructor<grpc_core::chttp2::TransportFlowControl> + flow_control; + /** initial window change. This is tracked as we parse settings frames from + * the remote peer. If there is a positive delta, then we will make all + * streams readable since they may have become unstalled */ + int64_t initial_window_update = 0; /* deframing */ grpc_chttp2_deframe_transport_state deframe_state; @@ -477,25 +439,6 @@ typedef enum { GPRC_METADATA_PUBLISHED_AT_CLOSE } grpc_published_metadata_method; -typedef struct { - /** window available for us to send to peer, over or under the initial window - * size of the transport... ie: - * remote_window = remote_window_delta + transport.initial_window_size */ - int64_t remote_window_delta; - - /** window available for peer to send to us (as a delta on - * transport.initial_window_size) - * local_window = local_window_delta + transport.initial_window_size */ - int64_t local_window_delta; - - /** window available for peer to send to us over this stream that we have - * announced to the peer */ - int64_t announced_window_delta; - - // read only pointer back to stream for data - const grpc_chttp2_stream *s; -} grpc_chttp2_stream_flowctl; - struct grpc_chttp2_stream { grpc_chttp2_transport *t; grpc_stream_refcount *refcount; @@ -589,7 +532,8 @@ struct grpc_chttp2_stream { bool sent_initial_metadata; bool sent_trailing_metadata; - grpc_chttp2_stream_flowctl flow_control; + grpc_core::ManualConstructor<grpc_core::chttp2::StreamFlowControl> + flow_control; grpc_slice_buffer flow_controlled_buffer; @@ -700,73 +644,10 @@ bool grpc_chttp2_list_remove_stalled_by_stream(grpc_chttp2_transport *t, /********* Flow Control ***************/ -// we have sent data on the wire -void grpc_chttp2_flowctl_sent_data(grpc_chttp2_transport_flowctl *tfc, - grpc_chttp2_stream_flowctl *sfc, - int64_t size); - -// we have received data from the wire -grpc_error *grpc_chttp2_flowctl_recv_data(grpc_chttp2_transport_flowctl *tfc, - grpc_chttp2_stream_flowctl *sfc, - int64_t incoming_frame_size); - -// returns an announce if we should send a transport update to our peer, -// else returns zero -uint32_t grpc_chttp2_flowctl_maybe_send_transport_update( - grpc_chttp2_transport_flowctl *tfc, bool writing_anyway); - -// returns an announce if we should send a stream update to our peer, else -// returns zero -uint32_t grpc_chttp2_flowctl_maybe_send_stream_update( - grpc_chttp2_transport_flowctl *tfc, grpc_chttp2_stream_flowctl *sfc); - -// we have received a WINDOW_UPDATE frame for a transport -void grpc_chttp2_flowctl_recv_transport_update( - grpc_chttp2_transport_flowctl *tfc, uint32_t size); - -// we have received a WINDOW_UPDATE frame for a stream -void grpc_chttp2_flowctl_recv_stream_update(grpc_chttp2_transport_flowctl *tfc, - grpc_chttp2_stream_flowctl *sfc, - uint32_t size); - -// the application is asking for a certain amount of bytes -void grpc_chttp2_flowctl_incoming_bs_update(grpc_chttp2_transport_flowctl *tfc, - grpc_chttp2_stream_flowctl *sfc, - size_t max_size_hint, - size_t have_already); - -void grpc_chttp2_flowctl_destroy_stream(grpc_chttp2_transport_flowctl *tfc, - grpc_chttp2_stream_flowctl *sfc); - -typedef enum { - // Nothing to be done. - GRPC_CHTTP2_FLOWCTL_NO_ACTION_NEEDED = 0, - // Initiate a write to update the initial window immediately. - GRPC_CHTTP2_FLOWCTL_UPDATE_IMMEDIATELY, - // Push the flow control update into a send buffer, to be sent - // out the next time a write is initiated. - GRPC_CHTTP2_FLOWCTL_QUEUE_UPDATE, -} grpc_chttp2_flowctl_urgency; - -typedef struct { - grpc_chttp2_flowctl_urgency send_stream_update; - grpc_chttp2_flowctl_urgency send_transport_update; - grpc_chttp2_flowctl_urgency send_setting_update; - uint32_t initial_window_size; - uint32_t max_frame_size; -} grpc_chttp2_flowctl_action; - -// Reads the flow control data and returns and actionable struct that will tell -// chttp2 exactly what it needs to do -grpc_chttp2_flowctl_action grpc_chttp2_flowctl_get_action( - grpc_exec_ctx *exec_ctx, grpc_chttp2_transport_flowctl *tfc, - grpc_chttp2_stream_flowctl *sfc); - // Takes in a flow control action and performs all the needed operations. -void grpc_chttp2_act_on_flowctl_action(grpc_exec_ctx *exec_ctx, - grpc_chttp2_flowctl_action action, - grpc_chttp2_transport *t, - grpc_chttp2_stream *s); +void grpc_chttp2_act_on_flowctl_action( + grpc_exec_ctx *exec_ctx, const grpc_core::chttp2::FlowControlAction &action, + grpc_chttp2_transport *t, grpc_chttp2_stream *s); /********* End of Flow Control ***************/ @@ -800,16 +681,6 @@ void grpc_chttp2_complete_closure_step(grpc_exec_ctx *exec_ctx, extern grpc_tracer_flag grpc_http_trace; extern grpc_tracer_flag grpc_flowctl_trace; -#ifndef NDEBUG -#define GRPC_FLOW_CONTROL_IF_TRACING(stmt) \ - if (!(GRPC_TRACER_ON(grpc_flowctl_trace))) \ - ; \ - else \ - stmt -#else -#define GRPC_FLOW_CONTROL_IF_TRACING(stmt) -#endif - #define GRPC_CHTTP2_IF_TRACING(stmt) \ if (!(GRPC_TRACER_ON(grpc_http_trace))) \ ; \ diff --git a/src/core/ext/transport/chttp2/transport/parsing.cc b/src/core/ext/transport/chttp2/transport/parsing.cc index 78886b497a..efa5791d3f 100644 --- a/src/core/ext/transport/chttp2/transport/parsing.cc +++ b/src/core/ext/transport/chttp2/transport/parsing.cc @@ -355,14 +355,15 @@ static grpc_error *init_data_frame_parser(grpc_exec_ctx *exec_ctx, grpc_chttp2_stream *s = grpc_chttp2_parsing_lookup_stream(t, t->incoming_stream_id); grpc_error *err = GRPC_ERROR_NONE; - err = grpc_chttp2_flowctl_recv_data(&t->flow_control, - s == NULL ? NULL : &s->flow_control, - t->incoming_frame_size); - grpc_chttp2_act_on_flowctl_action( - exec_ctx, - grpc_chttp2_flowctl_get_action(exec_ctx, &t->flow_control, - s == NULL ? NULL : &s->flow_control), - t, s); + grpc_core::chttp2::FlowControlAction action; + if (s == nullptr) { + err = t->flow_control->RecvData(t->incoming_frame_size); + action = t->flow_control->MakeAction(); + } else { + err = s->flow_control->RecvData(t->incoming_frame_size); + action = s->flow_control->MakeAction(); + } + grpc_chttp2_act_on_flowctl_action(exec_ctx, action, t, s); if (err != GRPC_ERROR_NONE) { goto error_handler; } @@ -434,19 +435,21 @@ static void on_initial_header(grpc_exec_ctx *exec_ctx, void *tp, grpc_millis *cached_timeout = static_cast<grpc_millis *>(grpc_mdelem_get_user_data(md, free_timeout)); grpc_millis timeout; - if (cached_timeout == NULL) { - /* not already parsed: parse it now, and store the result away */ - cached_timeout = (grpc_millis *)gpr_malloc(sizeof(grpc_millis)); - if (!grpc_http2_decode_timeout(GRPC_MDVALUE(md), cached_timeout)) { + if (cached_timeout != NULL) { + timeout = *cached_timeout; + } else { + if (!grpc_http2_decode_timeout(GRPC_MDVALUE(md), &timeout)) { char *val = grpc_slice_to_c_string(GRPC_MDVALUE(md)); gpr_log(GPR_ERROR, "Ignoring bad timeout value '%s'", val); gpr_free(val); - *cached_timeout = GRPC_MILLIS_INF_FUTURE; + timeout = GRPC_MILLIS_INF_FUTURE; + } + if (GRPC_MDELEM_IS_INTERNED(md)) { + /* store the result */ + cached_timeout = (grpc_millis *)gpr_malloc(sizeof(grpc_millis)); + *cached_timeout = timeout; + grpc_mdelem_set_user_data(md, free_timeout, cached_timeout); } - timeout = *cached_timeout; - grpc_mdelem_set_user_data(md, free_timeout, cached_timeout); - } else { - timeout = *cached_timeout; } if (timeout != GRPC_MILLIS_INF_FUTURE) { grpc_chttp2_incoming_metadata_buffer_set_deadline( diff --git a/src/core/ext/transport/chttp2/transport/writing.cc b/src/core/ext/transport/chttp2/transport/writing.cc index c6fecf2ee9..ff76a5fcdb 100644 --- a/src/core/ext/transport/chttp2/transport/writing.cc +++ b/src/core/ext/transport/chttp2/transport/writing.cc @@ -146,13 +146,13 @@ static void report_stall(grpc_chttp2_transport *t, grpc_chttp2_stream *s, s->flow_controlled_bytes_flowed, t->settings[GRPC_ACKED_SETTINGS] [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE], - t->flow_control.remote_window, + t->flow_control->remote_window(), (uint32_t)GPR_MAX( 0, - s->flow_control.remote_window_delta + + s->flow_control->remote_window_delta() + (int64_t)t->settings[GRPC_PEER_SETTINGS] [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]), - s->flow_control.remote_window_delta); + s->flow_control->remote_window_delta()); } static bool stream_ref_if_not_destroyed(gpr_refcount *r) { @@ -216,8 +216,7 @@ class WriteContext { void FlushWindowUpdates(grpc_exec_ctx *exec_ctx) { uint32_t transport_announce = - grpc_chttp2_flowctl_maybe_send_transport_update(&t_->flow_control, - t_->outbuf.count > 0); + t_->flow_control->MaybeSendUpdate(t_->outbuf.count > 0); if (transport_announce) { grpc_transport_one_way_stats throwaway_stats; grpc_slice_buffer_add( @@ -312,7 +311,7 @@ class DataSendContext { uint32_t stream_remote_window() const { return (uint32_t)GPR_MAX( - 0, s_->flow_control.remote_window_delta + + 0, s_->flow_control->remote_window_delta() + (int64_t)t_->settings[GRPC_PEER_SETTINGS] [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]); } @@ -320,7 +319,7 @@ class DataSendContext { uint32_t max_outgoing() const { return (uint32_t)GPR_MIN( t_->settings[GRPC_PEER_SETTINGS][GRPC_CHTTP2_SETTINGS_MAX_FRAME_SIZE], - GPR_MIN(stream_remote_window(), t_->flow_control.remote_window)); + GPR_MIN(stream_remote_window(), t_->flow_control->remote_window())); } bool AnyOutgoing() const { return max_outgoing() != 0; } @@ -352,8 +351,7 @@ class DataSendContext { grpc_metadata_batch_is_empty(s_->send_trailing_metadata); grpc_chttp2_encode_data(s_->id, &s_->compressed_data_buffer, send_bytes, is_last_frame_, &s_->stats.outgoing, &t_->outbuf); - grpc_chttp2_flowctl_sent_data(&t_->flow_control, &s_->flow_control, - send_bytes); + s_->flow_control->SentData(send_bytes); if (s_->compressed_data_buffer.length == 0) { s_->sending_bytes += s_->uncompressed_data_size; } @@ -400,8 +398,8 @@ class StreamWriteContext { gpr_log(GPR_DEBUG, "W:%p %s[%d] im-(sent,send)=(%d,%d) announce=%d", t_, t_->is_client ? "CLIENT" : "SERVER", s->id, s->sent_initial_metadata, s->send_initial_metadata != NULL, - (int)(s->flow_control.local_window_delta - - s->flow_control.announced_window_delta))); + (int)(s->flow_control->local_window_delta() - + s->flow_control->announced_window_delta()))); } void FlushInitialMetadata(grpc_exec_ctx *exec_ctx) { @@ -447,8 +445,7 @@ class StreamWriteContext { void FlushWindowUpdates(grpc_exec_ctx *exec_ctx) { /* send any window updates */ - uint32_t stream_announce = grpc_chttp2_flowctl_maybe_send_stream_update( - &t_->flow_control, &s_->flow_control); + const uint32_t stream_announce = s_->flow_control->MaybeSendUpdate(); if (stream_announce == 0) return; grpc_slice_buffer_add( @@ -469,10 +466,10 @@ class StreamWriteContext { DataSendContext data_send_context(write_context_, t_, s_); if (!data_send_context.AnyOutgoing()) { - if (t_->flow_control.remote_window == 0) { + if (t_->flow_control->remote_window() <= 0) { report_stall(t_, s_, "transport"); grpc_chttp2_list_add_stalled_by_transport(t_, s_); - } else if (data_send_context.stream_remote_window() == 0) { + } else if (data_send_context.stream_remote_window() <= 0) { report_stall(t_, s_, "stream"); grpc_chttp2_list_add_stalled_by_stream(t_, s_); } @@ -588,7 +585,7 @@ grpc_chttp2_begin_write_result grpc_chttp2_begin_write( ctx.FlushQueuedBuffers(exec_ctx); ctx.EnactHpackSettings(exec_ctx); - if (t->flow_control.remote_window > 0) { + if (t->flow_control->remote_window() > 0) { ctx.UpdateStreamsNoLongerStalled(); } diff --git a/src/core/ext/transport/cronet/transport/cronet_transport.cc b/src/core/ext/transport/cronet/transport/cronet_transport.cc index ff1367fb28..97e4f7d72b 100644 --- a/src/core/ext/transport/cronet/transport/cronet_transport.cc +++ b/src/core/ext/transport/cronet/transport/cronet_transport.cc @@ -692,7 +692,7 @@ static void create_grpc_frame(grpc_exec_ctx *exec_ctx, uint8_t *p = (uint8_t *)write_buffer; /* Append 5 byte header */ /* Compressed flag */ - *p++ = (flags & GRPC_WRITE_INTERNAL_COMPRESS) ? 1 : 0; + *p++ = (uint8_t)((flags & GRPC_WRITE_INTERNAL_COMPRESS) ? 1 : 0); /* Message length */ *p++ = (uint8_t)(length >> 24); *p++ = (uint8_t)(length >> 16); diff --git a/src/core/ext/transport/inproc/inproc_transport.cc b/src/core/ext/transport/inproc/inproc_transport.cc index 67a8358927..1551f5e988 100644 --- a/src/core/ext/transport/inproc/inproc_transport.cc +++ b/src/core/ext/transport/inproc/inproc_transport.cc @@ -623,7 +623,7 @@ static void op_state_machine(grpc_exec_ctx *exec_ctx, void *arg, fail_helper_locked(exec_ctx, s, GRPC_ERROR_REF(new_err)); goto done; } else { - if (other && !other->closed) { + if (!other || !other->closed) { fill_in_metadata(exec_ctx, s, s->send_trailing_md_op->payload->send_trailing_metadata .send_trailing_metadata, @@ -925,7 +925,7 @@ static void perform_stream_op(grpc_exec_ctx *exec_ctx, grpc_transport *gt, INPROC_LOG(GPR_DEBUG, "Extra initial metadata %p", s); error = GRPC_ERROR_CREATE_FROM_STATIC_STRING("Extra initial metadata"); } else { - if (!other->closed) { + if (!other || !other->closed) { fill_in_metadata( exec_ctx, s, op->payload->send_initial_metadata.send_initial_metadata, @@ -976,7 +976,7 @@ static void perform_stream_op(grpc_exec_ctx *exec_ctx, grpc_transport *gt, (other->recv_trailing_md_op != NULL))) || (op->send_trailing_metadata && !op->send_message) || (op->recv_initial_metadata && s->to_read_initial_md_filled) || - (op->recv_message && (other && other->send_message_op != NULL)) || + (op->recv_message && other && (other->send_message_op != NULL)) || (s->to_read_trailing_md_filled || s->trailing_md_recvd)) { if (!s->op_closure_scheduled) { GRPC_CLOSURE_SCHED(exec_ctx, &s->op_closure, GRPC_ERROR_NONE); |