diff options
Diffstat (limited to 'src/core/lib')
60 files changed, 2304 insertions, 8047 deletions
diff --git a/src/core/lib/census/README.md b/src/core/lib/census/README.md deleted file mode 100644 index fb615a2194..0000000000 --- a/src/core/lib/census/README.md +++ /dev/null @@ -1,76 +0,0 @@ -<!--- - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. ---> - -# 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/lib/census/aggregation.h b/src/core/lib/census/aggregation.h deleted file mode 100644 index 088ad808ef..0000000000 --- a/src/core/lib/census/aggregation.h +++ /dev/null @@ -1,66 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <stddef.h> - -#ifndef GRPC_CORE_LIB_CENSUS_AGGREGATION_H -#define GRPC_CORE_LIB_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_LIB_CENSUS_AGGREGATION_H */ diff --git a/src/core/lib/census/context.c b/src/core/lib/census/context.c deleted file mode 100644 index 0dfc4ecbf1..0000000000 --- a/src/core/lib/census/context.c +++ /dev/null @@ -1,509 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#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 = 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 = gpr_malloc(tags->kvm_size); - 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 = 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 = 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 = 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/lib/census/grpc_context.c b/src/core/lib/census/grpc_context.c deleted file mode 100644 index 98285ab2d5..0000000000 --- a/src/core/lib/census/grpc_context.c +++ /dev/null @@ -1,53 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/census.h> -#include <grpc/grpc.h> -#include "src/core/lib/surface/api_trace.h" -#include "src/core/lib/surface/call.h" - -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); - } -} - -census_context *grpc_census_call_get_context(grpc_call *call) { - GRPC_API_TRACE("grpc_census_call_get_context(call=%p)", 1, (call)); - return (census_context *)grpc_call_context_get(call, GRPC_CONTEXT_TRACING); -} diff --git a/src/core/lib/census/grpc_filter.c b/src/core/lib/census/grpc_filter.c deleted file mode 100644 index 9bca2d434d..0000000000 --- a/src/core/lib/census/grpc_filter.c +++ /dev/null @@ -1,198 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include "src/core/lib/census/grpc_filter.h" - -#include <stdio.h> -#include <string.h> - -#include <grpc/census.h> -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/slice.h> -#include <grpc/support/time.h> - -#include "src/core/lib/channel/channel_stack.h" -#include "src/core/lib/statistics/census_interface.h" -#include "src/core/lib/statistics/census_rpc_stats.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 (m->md->key == GRPC_MDSTR_PATH) { - gpr_log(GPR_DEBUG, "%s", - (const char *)GPR_SLICE_START_PTR(m->md->value->slice)); - /* Add method tag here */ - } - } -} - -static void client_mutate_op(grpc_call_element *elem, - grpc_transport_stream_op *op) { - call_data *calld = elem->call_data; - channel_data *chand = elem->channel_data; - if (op->send_initial_metadata) { - extract_and_annotate_method_tag(op->send_initial_metadata, calld, chand); - } -} - -static void client_start_transport_op(grpc_exec_ctx *exec_ctx, - grpc_call_element *elem, - grpc_transport_stream_op *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, - bool success) { - grpc_call_element *elem = ptr; - call_data *calld = elem->call_data; - channel_data *chand = elem->channel_data; - if (success) { - extract_and_annotate_method_tag(calld->recv_initial_metadata, calld, chand); - } - calld->on_done_recv->cb(exec_ctx, calld->on_done_recv->cb_arg, success); -} - -static void server_mutate_op(grpc_call_element *elem, - grpc_transport_stream_op *op) { - call_data *calld = elem->call_data; - if (op->recv_initial_metadata) { - /* substitute our callback for the op callback */ - calld->recv_initial_metadata = op->recv_initial_metadata; - calld->on_done_recv = op->recv_initial_metadata_ready; - op->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 *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 void client_init_call_elem(grpc_exec_ctx *exec_ctx, - grpc_call_element *elem, - grpc_call_element_args *args) { - call_data *d = elem->call_data; - GPR_ASSERT(d != NULL); - memset(d, 0, sizeof(*d)); - d->start_ts = gpr_now(GPR_CLOCK_REALTIME); -} - -static void client_destroy_call_elem(grpc_exec_ctx *exec_ctx, - grpc_call_element *elem) { - call_data *d = elem->call_data; - GPR_ASSERT(d != NULL); - /* TODO(hongyu): record rpc client stats and census_rpc_end_op here */ -} - -static void server_init_call_elem(grpc_exec_ctx *exec_ctx, - grpc_call_element *elem, - grpc_call_element_args *args) { - call_data *d = elem->call_data; - GPR_ASSERT(d != NULL); - memset(d, 0, sizeof(*d)); - d->start_ts = gpr_now(GPR_CLOCK_REALTIME); - /* TODO(hongyu): call census_tracing_start_op here. */ - grpc_closure_init(&d->finish_recv, server_on_done_recv, elem); -} - -static void server_destroy_call_elem(grpc_exec_ctx *exec_ctx, - grpc_call_element *elem) { - call_data *d = elem->call_data; - GPR_ASSERT(d != NULL); - /* TODO(hongyu): record rpc server stats and census_tracing_end_op here */ -} - -static void init_channel_elem(grpc_exec_ctx *exec_ctx, - grpc_channel_element *elem, - grpc_channel_element_args *args) { - channel_data *chand = elem->channel_data; - GPR_ASSERT(chand != NULL); -} - -static void destroy_channel_elem(grpc_exec_ctx *exec_ctx, - grpc_channel_element *elem) { - channel_data *chand = 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, - client_destroy_call_elem, - sizeof(channel_data), - init_channel_elem, - destroy_channel_elem, - grpc_call_next_get_peer, - "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, - server_destroy_call_elem, - sizeof(channel_data), - init_channel_elem, - destroy_channel_elem, - grpc_call_next_get_peer, - "census-server"}; diff --git a/src/core/lib/census/grpc_filter.h b/src/core/lib/census/grpc_filter.h deleted file mode 100644 index e747fe6590..0000000000 --- a/src/core/lib/census/grpc_filter.h +++ /dev/null @@ -1,44 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_CENSUS_GRPC_FILTER_H -#define GRPC_CORE_LIB_CENSUS_GRPC_FILTER_H - -#include "src/core/lib/channel/channel_stack.h" - -/* 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; - -#endif /* GRPC_CORE_LIB_CENSUS_GRPC_FILTER_H */ diff --git a/src/core/lib/census/grpc_plugin.c b/src/core/lib/census/grpc_plugin.c deleted file mode 100644 index 7f6df11f00..0000000000 --- a/src/core/lib/census/grpc_plugin.c +++ /dev/null @@ -1,70 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include "src/core/lib/census/grpc_plugin.h" - -#include <limits.h> - -#include <grpc/census.h> - -#include "src/core/lib/census/grpc_filter.h" -#include "src/core/lib/channel/channel_stack_builder.h" -#include "src/core/lib/surface/channel_init.h" - -static bool maybe_add_census_filter(grpc_channel_stack_builder *builder, - void *arg_must_be_null) { - const grpc_channel_args *args = - grpc_channel_stack_builder_get_channel_arguments(builder); - if (grpc_channel_args_is_census_enabled(args)) { - return grpc_channel_stack_builder_prepend_filter( - builder, &grpc_client_census_filter, NULL, NULL); - } - return true; -} - -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, NULL); - grpc_channel_init_register_stage(GRPC_SERVER_CHANNEL, INT_MAX, - maybe_add_census_filter, NULL); -} - -void census_grpc_plugin_destroy(void) { census_shutdown(); } diff --git a/src/core/lib/census/initialize.c b/src/core/lib/census/initialize.c deleted file mode 100644 index 896276e44a..0000000000 --- a/src/core/lib/census/initialize.c +++ /dev/null @@ -1,54 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/census.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; - return 0; -} - -void census_shutdown(void) { 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/lib/census/mlog.c b/src/core/lib/census/mlog.c deleted file mode 100644 index ce060a2d85..0000000000 --- a/src/core/lib/census/mlog.c +++ /dev/null @@ -1,600 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -// 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/lib/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 = 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/lib/census/mlog.h b/src/core/lib/census/mlog.h deleted file mode 100644 index dea43de725..0000000000 --- a/src/core/lib/census/mlog.h +++ /dev/null @@ -1,95 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -/* A very fast in-memory log, optimized for multiple writers. */ - -#ifndef GRPC_CORE_LIB_CENSUS_MLOG_H -#define GRPC_CORE_LIB_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) - -/* 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 gprc_stats_log_start_write() failed due to - out-of-space. */ -int64_t census_log_out_of_space_count(void); - -#endif /* GRPC_CORE_LIB_CENSUS_MLOG_H */ diff --git a/src/core/lib/census/operation.c b/src/core/lib/census/operation.c deleted file mode 100644 index 5c58704372..0000000000 --- a/src/core/lib/census/operation.c +++ /dev/null @@ -1,63 +0,0 @@ -/* - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#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/lib/census/placeholders.c b/src/core/lib/census/placeholders.c deleted file mode 100644 index fe23d13971..0000000000 --- a/src/core/lib/census/placeholders.c +++ /dev/null @@ -1,109 +0,0 @@ -/* - * - * Copyright 2016, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#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(); -} - -const census_aggregation *census_view_aggregrations(const census_view *view) { - (void)view; - abort(); -} - -census_view *census_view_create(uint32_t metric_id, const census_context *tags, - const census_aggregation *aggregations, - size_t naggregations) { - (void)metric_id; - (void)tags; - (void)aggregations; - (void)naggregations; - abort(); -} - -const census_context *census_view_tags(const census_view *view) { - (void)view; - abort(); -} - -void census_view_delete(census_view *view) { - (void)view; - abort(); -} - -const census_view_data *census_view_get_data(const census_view *view) { - (void)view; - abort(); -} - -size_t census_view_metric(const census_view *view) { - (void)view; - abort(); -} - -size_t census_view_naggregations(const census_view *view) { - (void)view; - abort(); -} - -void census_view_reset(census_view *view) { - (void)view; - abort(); -} diff --git a/src/core/lib/census/rpc_metric_id.h b/src/core/lib/census/rpc_metric_id.h deleted file mode 100644 index 0f05b7de8a..0000000000 --- a/src/core/lib/census/rpc_metric_id.h +++ /dev/null @@ -1,51 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_CENSUS_RPC_METRIC_ID_H -#define GRPC_CORE_LIB_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_LIB_CENSUS_RPC_METRIC_ID_H */ diff --git a/src/core/lib/census/tracing.c b/src/core/lib/census/tracing.c deleted file mode 100644 index 3b5d6dab2b..0000000000 --- a/src/core/lib/census/tracing.c +++ /dev/null @@ -1,45 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/census.h> - -/* TODO(aveitch): These are all placeholder implementations. */ - -int census_trace_mask(const census_context *context) { - return CENSUS_TRACE_MASK_NONE; -} - -void census_set_trace_mask(int trace_mask) {} - -void census_trace_print(census_context *context, uint32_t type, - const char *buffer, size_t n) {} diff --git a/src/core/lib/client_config/resolver_registry.c b/src/core/lib/client_config/resolver_registry.c index e8432d165a..5f3db273b5 100644 --- a/src/core/lib/client_config/resolver_registry.c +++ b/src/core/lib/client_config/resolver_registry.c @@ -70,14 +70,11 @@ void grpc_register_resolver_type(grpc_resolver_factory *factory) { g_all_of_the_resolvers[g_number_of_resolvers++] = factory; } -static grpc_resolver_factory *lookup_factory(grpc_uri *uri) { +static grpc_resolver_factory *lookup_factory(const char *name) { int i; - /* handling NULL uri's here simplifies grpc_resolver_create */ - if (!uri) return NULL; - for (i = 0; i < g_number_of_resolvers; i++) { - if (0 == strcmp(uri->scheme, g_all_of_the_resolvers[i]->vtable->scheme)) { + if (0 == strcmp(name, g_all_of_the_resolvers[i]->vtable->scheme)) { return g_all_of_the_resolvers[i]; } } @@ -85,6 +82,17 @@ static grpc_resolver_factory *lookup_factory(grpc_uri *uri) { return NULL; } +grpc_resolver_factory *grpc_resolver_factory_lookup(const char *name) { + grpc_resolver_factory *f = lookup_factory(name); + if (f) grpc_resolver_factory_ref(f); + return f; +} + +static grpc_resolver_factory *lookup_factory_by_uri(grpc_uri *uri) { + if (!uri) return NULL; + return lookup_factory(uri->scheme); +} + static grpc_resolver_factory *resolve_factory(const char *target, grpc_uri **uri) { char *tmp; @@ -92,13 +100,13 @@ static grpc_resolver_factory *resolve_factory(const char *target, GPR_ASSERT(uri != NULL); *uri = grpc_uri_parse(target, 1); - factory = lookup_factory(*uri); + factory = lookup_factory_by_uri(*uri); if (factory == NULL) { if (g_default_resolver_prefix != NULL) { grpc_uri_destroy(*uri); gpr_asprintf(&tmp, "%s%s", g_default_resolver_prefix, target); *uri = grpc_uri_parse(tmp, 1); - factory = lookup_factory(*uri); + factory = lookup_factory_by_uri(*uri); if (factory == NULL) { grpc_uri_destroy(grpc_uri_parse(target, 0)); grpc_uri_destroy(grpc_uri_parse(tmp, 0)); diff --git a/src/core/lib/client_config/resolver_registry.h b/src/core/lib/client_config/resolver_registry.h index 330cb7f544..36c4f2fe03 100644 --- a/src/core/lib/client_config/resolver_registry.h +++ b/src/core/lib/client_config/resolver_registry.h @@ -58,6 +58,10 @@ void grpc_register_resolver_type(grpc_resolver_factory *factory); grpc_resolver *grpc_resolver_create( const char *target, grpc_subchannel_factory *subchannel_factory); +/** Find a resolver factory given a name and return an (owned-by-the-caller) + * reference to it */ +grpc_resolver_factory *grpc_resolver_factory_lookup(const char *name); + /** Given a target, return a (freshly allocated with gpr_malloc) string representing the default authority to pass from a client. */ char *grpc_get_default_authority(const char *target); diff --git a/src/core/lib/client_config/resolvers/dns_resolver.c b/src/core/lib/client_config/resolvers/dns_resolver.c deleted file mode 100644 index ad586aa7e4..0000000000 --- a/src/core/lib/client_config/resolvers/dns_resolver.c +++ /dev/null @@ -1,294 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include "src/core/lib/client_config/resolvers/dns_resolver.h" - -#include <string.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/host_port.h> -#include <grpc/support/string_util.h> - -#include "src/core/lib/client_config/lb_policy_registry.h" -#include "src/core/lib/iomgr/resolve_address.h" -#include "src/core/lib/iomgr/timer.h" -#include "src/core/lib/support/backoff.h" -#include "src/core/lib/support/string.h" - -#define BACKOFF_MULTIPLIER 1.6 -#define BACKOFF_JITTER 0.2 -#define BACKOFF_MIN_SECONDS 1 -#define BACKOFF_MAX_SECONDS 120 - -typedef struct { - /** base class: must be first */ - grpc_resolver base; - /** refcount */ - gpr_refcount refs; - /** name to resolve */ - char *name; - /** default port to use */ - char *default_port; - /** subchannel factory */ - grpc_subchannel_factory *subchannel_factory; - /** load balancing policy name */ - char *lb_policy_name; - - /** mutex guarding the rest of the state */ - gpr_mu mu; - /** are we currently resolving? */ - int resolving; - /** which version of resolved_config have we published? */ - int published_version; - /** which version of resolved_config is current? */ - int resolved_version; - /** pending next completion, or NULL */ - grpc_closure *next_completion; - /** target config address for next completion */ - grpc_client_config **target_config; - /** current (fully resolved) config */ - grpc_client_config *resolved_config; - /** retry timer */ - bool have_retry_timer; - grpc_timer retry_timer; - /** retry backoff state */ - gpr_backoff backoff_state; -} dns_resolver; - -static void dns_destroy(grpc_exec_ctx *exec_ctx, grpc_resolver *r); - -static void dns_start_resolving_locked(dns_resolver *r); -static void dns_maybe_finish_next_locked(grpc_exec_ctx *exec_ctx, - dns_resolver *r); - -static void dns_shutdown(grpc_exec_ctx *exec_ctx, grpc_resolver *r); -static void dns_channel_saw_error(grpc_exec_ctx *exec_ctx, grpc_resolver *r); -static void dns_next(grpc_exec_ctx *exec_ctx, grpc_resolver *r, - grpc_client_config **target_config, - grpc_closure *on_complete); - -static const grpc_resolver_vtable dns_resolver_vtable = { - dns_destroy, dns_shutdown, dns_channel_saw_error, dns_next}; - -static void dns_shutdown(grpc_exec_ctx *exec_ctx, grpc_resolver *resolver) { - dns_resolver *r = (dns_resolver *)resolver; - gpr_mu_lock(&r->mu); - if (r->have_retry_timer) { - grpc_timer_cancel(exec_ctx, &r->retry_timer); - } - if (r->next_completion != NULL) { - *r->target_config = NULL; - grpc_exec_ctx_enqueue(exec_ctx, r->next_completion, true, NULL); - r->next_completion = NULL; - } - gpr_mu_unlock(&r->mu); -} - -static void dns_channel_saw_error(grpc_exec_ctx *exec_ctx, - grpc_resolver *resolver) { - dns_resolver *r = (dns_resolver *)resolver; - gpr_mu_lock(&r->mu); - if (!r->resolving) { - gpr_backoff_reset(&r->backoff_state); - dns_start_resolving_locked(r); - } - gpr_mu_unlock(&r->mu); -} - -static void dns_next(grpc_exec_ctx *exec_ctx, grpc_resolver *resolver, - grpc_client_config **target_config, - grpc_closure *on_complete) { - dns_resolver *r = (dns_resolver *)resolver; - gpr_mu_lock(&r->mu); - GPR_ASSERT(!r->next_completion); - r->next_completion = on_complete; - r->target_config = target_config; - if (r->resolved_version == 0 && !r->resolving) { - gpr_backoff_reset(&r->backoff_state); - dns_start_resolving_locked(r); - } else { - dns_maybe_finish_next_locked(exec_ctx, r); - } - gpr_mu_unlock(&r->mu); -} - -static void dns_on_retry_timer(grpc_exec_ctx *exec_ctx, void *arg, - bool success) { - dns_resolver *r = arg; - - gpr_mu_lock(&r->mu); - r->have_retry_timer = false; - if (success) { - if (!r->resolving) { - dns_start_resolving_locked(r); - } - } - gpr_mu_unlock(&r->mu); - - GRPC_RESOLVER_UNREF(exec_ctx, &r->base, "retry-timer"); -} - -static void dns_on_resolved(grpc_exec_ctx *exec_ctx, void *arg, - grpc_resolved_addresses *addresses) { - dns_resolver *r = arg; - grpc_client_config *config = NULL; - grpc_lb_policy *lb_policy; - gpr_mu_lock(&r->mu); - GPR_ASSERT(r->resolving); - r->resolving = 0; - if (addresses != NULL) { - grpc_lb_policy_args lb_policy_args; - config = grpc_client_config_create(); - memset(&lb_policy_args, 0, sizeof(lb_policy_args)); - lb_policy_args.addresses = addresses; - lb_policy_args.subchannel_factory = r->subchannel_factory; - lb_policy = - grpc_lb_policy_create(exec_ctx, r->lb_policy_name, &lb_policy_args); - if (lb_policy != NULL) { - grpc_client_config_set_lb_policy(config, lb_policy); - GRPC_LB_POLICY_UNREF(exec_ctx, lb_policy, "construction"); - } - grpc_resolved_addresses_destroy(addresses); - } else { - gpr_timespec now = gpr_now(GPR_CLOCK_MONOTONIC); - gpr_timespec next_try = gpr_backoff_step(&r->backoff_state, now); - gpr_timespec timeout = gpr_time_sub(next_try, now); - gpr_log(GPR_DEBUG, "dns resolution failed: retrying in %d.%09d seconds", - timeout.tv_sec, timeout.tv_nsec); - GPR_ASSERT(!r->have_retry_timer); - r->have_retry_timer = true; - GRPC_RESOLVER_REF(&r->base, "retry-timer"); - grpc_timer_init(exec_ctx, &r->retry_timer, next_try, dns_on_retry_timer, r, - now); - } - if (r->resolved_config) { - grpc_client_config_unref(exec_ctx, r->resolved_config); - } - r->resolved_config = config; - r->resolved_version++; - dns_maybe_finish_next_locked(exec_ctx, r); - gpr_mu_unlock(&r->mu); - - GRPC_RESOLVER_UNREF(exec_ctx, &r->base, "dns-resolving"); -} - -static void dns_start_resolving_locked(dns_resolver *r) { - GRPC_RESOLVER_REF(&r->base, "dns-resolving"); - GPR_ASSERT(!r->resolving); - r->resolving = 1; - grpc_resolve_address(r->name, r->default_port, dns_on_resolved, r); -} - -static void dns_maybe_finish_next_locked(grpc_exec_ctx *exec_ctx, - dns_resolver *r) { - if (r->next_completion != NULL && - r->resolved_version != r->published_version) { - *r->target_config = r->resolved_config; - if (r->resolved_config) { - grpc_client_config_ref(r->resolved_config); - } - grpc_exec_ctx_enqueue(exec_ctx, r->next_completion, true, NULL); - r->next_completion = NULL; - r->published_version = r->resolved_version; - } -} - -static void dns_destroy(grpc_exec_ctx *exec_ctx, grpc_resolver *gr) { - dns_resolver *r = (dns_resolver *)gr; - gpr_mu_destroy(&r->mu); - if (r->resolved_config) { - grpc_client_config_unref(exec_ctx, r->resolved_config); - } - grpc_subchannel_factory_unref(exec_ctx, r->subchannel_factory); - gpr_free(r->name); - gpr_free(r->default_port); - gpr_free(r->lb_policy_name); - gpr_free(r); -} - -static grpc_resolver *dns_create(grpc_resolver_args *args, - const char *default_port, - const char *lb_policy_name) { - dns_resolver *r; - const char *path = args->uri->path; - - if (0 != strcmp(args->uri->authority, "")) { - gpr_log(GPR_ERROR, "authority based dns uri's not supported"); - return NULL; - } - - if (path[0] == '/') ++path; - - r = gpr_malloc(sizeof(dns_resolver)); - memset(r, 0, sizeof(*r)); - gpr_ref_init(&r->refs, 1); - gpr_mu_init(&r->mu); - grpc_resolver_init(&r->base, &dns_resolver_vtable); - r->name = gpr_strdup(path); - r->default_port = gpr_strdup(default_port); - r->subchannel_factory = args->subchannel_factory; - gpr_backoff_init(&r->backoff_state, BACKOFF_MULTIPLIER, BACKOFF_JITTER, - BACKOFF_MIN_SECONDS * 1000, BACKOFF_MAX_SECONDS * 1000); - grpc_subchannel_factory_ref(r->subchannel_factory); - r->lb_policy_name = gpr_strdup(lb_policy_name); - return &r->base; -} - -/* - * FACTORY - */ - -static void dns_factory_ref(grpc_resolver_factory *factory) {} - -static void dns_factory_unref(grpc_resolver_factory *factory) {} - -static grpc_resolver *dns_factory_create_resolver( - grpc_resolver_factory *factory, grpc_resolver_args *args) { - return dns_create(args, "https", "pick_first"); -} - -char *dns_factory_get_default_host_name(grpc_resolver_factory *factory, - grpc_uri *uri) { - const char *path = uri->path; - if (path[0] == '/') ++path; - return gpr_strdup(path); -} - -static const grpc_resolver_factory_vtable dns_factory_vtable = { - dns_factory_ref, dns_factory_unref, dns_factory_create_resolver, - dns_factory_get_default_host_name, "dns"}; -static grpc_resolver_factory dns_resolver_factory = {&dns_factory_vtable}; - -grpc_resolver_factory *grpc_dns_resolver_factory_create() { - return &dns_resolver_factory; -} diff --git a/src/core/lib/client_config/resolvers/dns_resolver.h b/src/core/lib/client_config/resolvers/dns_resolver.h deleted file mode 100644 index de41961ea9..0000000000 --- a/src/core/lib/client_config/resolvers/dns_resolver.h +++ /dev/null @@ -1,42 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_DNS_RESOLVER_H -#define GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_DNS_RESOLVER_H - -#include "src/core/lib/client_config/resolver_factory.h" - -/** Create a dns resolver factory */ -grpc_resolver_factory *grpc_dns_resolver_factory_create(void); - -#endif /* GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_DNS_RESOLVER_H */ diff --git a/src/core/lib/client_config/resolvers/sockaddr_resolver.c b/src/core/lib/client_config/resolvers/sockaddr_resolver.c deleted file mode 100644 index 0e2656d033..0000000000 --- a/src/core/lib/client_config/resolvers/sockaddr_resolver.c +++ /dev/null @@ -1,354 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/support/port_platform.h> - -#include "src/core/lib/client_config/resolvers/sockaddr_resolver.h" - -#include <stdio.h> -#include <string.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/host_port.h> -#include <grpc/support/string_util.h> - -#include "src/core/lib/client_config/lb_policy_registry.h" -#include "src/core/lib/iomgr/resolve_address.h" -#include "src/core/lib/iomgr/unix_sockets_posix.h" -#include "src/core/lib/support/string.h" - -typedef struct { - /** base class: must be first */ - grpc_resolver base; - /** refcount */ - gpr_refcount refs; - /** subchannel factory */ - grpc_subchannel_factory *subchannel_factory; - /** load balancing policy name */ - char *lb_policy_name; - - /** the addresses that we've 'resolved' */ - grpc_resolved_addresses *addresses; - - /** mutex guarding the rest of the state */ - gpr_mu mu; - /** have we published? */ - int published; - /** pending next completion, or NULL */ - grpc_closure *next_completion; - /** target config address for next completion */ - grpc_client_config **target_config; -} sockaddr_resolver; - -static void sockaddr_destroy(grpc_exec_ctx *exec_ctx, grpc_resolver *r); - -static void sockaddr_maybe_finish_next_locked(grpc_exec_ctx *exec_ctx, - sockaddr_resolver *r); - -static void sockaddr_shutdown(grpc_exec_ctx *exec_ctx, grpc_resolver *r); -static void sockaddr_channel_saw_error(grpc_exec_ctx *exec_ctx, - grpc_resolver *r); -static void sockaddr_next(grpc_exec_ctx *exec_ctx, grpc_resolver *r, - grpc_client_config **target_config, - grpc_closure *on_complete); - -static const grpc_resolver_vtable sockaddr_resolver_vtable = { - sockaddr_destroy, sockaddr_shutdown, sockaddr_channel_saw_error, - sockaddr_next}; - -static void sockaddr_shutdown(grpc_exec_ctx *exec_ctx, - grpc_resolver *resolver) { - sockaddr_resolver *r = (sockaddr_resolver *)resolver; - gpr_mu_lock(&r->mu); - if (r->next_completion != NULL) { - *r->target_config = NULL; - grpc_exec_ctx_enqueue(exec_ctx, r->next_completion, true, NULL); - r->next_completion = NULL; - } - gpr_mu_unlock(&r->mu); -} - -static void sockaddr_channel_saw_error(grpc_exec_ctx *exec_ctx, - grpc_resolver *resolver) { - sockaddr_resolver *r = (sockaddr_resolver *)resolver; - gpr_mu_lock(&r->mu); - r->published = 0; - sockaddr_maybe_finish_next_locked(exec_ctx, r); - gpr_mu_unlock(&r->mu); -} - -static void sockaddr_next(grpc_exec_ctx *exec_ctx, grpc_resolver *resolver, - grpc_client_config **target_config, - grpc_closure *on_complete) { - sockaddr_resolver *r = (sockaddr_resolver *)resolver; - gpr_mu_lock(&r->mu); - GPR_ASSERT(!r->next_completion); - r->next_completion = on_complete; - r->target_config = target_config; - sockaddr_maybe_finish_next_locked(exec_ctx, r); - gpr_mu_unlock(&r->mu); -} - -static void sockaddr_maybe_finish_next_locked(grpc_exec_ctx *exec_ctx, - sockaddr_resolver *r) { - if (r->next_completion != NULL && !r->published) { - grpc_client_config *cfg = grpc_client_config_create(); - grpc_lb_policy_args lb_policy_args; - memset(&lb_policy_args, 0, sizeof(lb_policy_args)); - lb_policy_args.addresses = r->addresses; - lb_policy_args.subchannel_factory = r->subchannel_factory; - grpc_lb_policy *lb_policy = - grpc_lb_policy_create(exec_ctx, r->lb_policy_name, &lb_policy_args); - grpc_client_config_set_lb_policy(cfg, lb_policy); - GRPC_LB_POLICY_UNREF(exec_ctx, lb_policy, "sockaddr"); - r->published = 1; - *r->target_config = cfg; - grpc_exec_ctx_enqueue(exec_ctx, r->next_completion, true, NULL); - r->next_completion = NULL; - } -} - -static void sockaddr_destroy(grpc_exec_ctx *exec_ctx, grpc_resolver *gr) { - sockaddr_resolver *r = (sockaddr_resolver *)gr; - gpr_mu_destroy(&r->mu); - grpc_subchannel_factory_unref(exec_ctx, r->subchannel_factory); - grpc_resolved_addresses_destroy(r->addresses); - gpr_free(r->lb_policy_name); - gpr_free(r); -} - -static char *ip_get_default_authority(grpc_uri *uri) { - const char *path = uri->path; - if (path[0] == '/') ++path; - return gpr_strdup(path); -} - -static char *ipv4_get_default_authority(grpc_resolver_factory *factory, - grpc_uri *uri) { - return ip_get_default_authority(uri); -} - -static char *ipv6_get_default_authority(grpc_resolver_factory *factory, - grpc_uri *uri) { - return ip_get_default_authority(uri); -} - -static int parse_ipv4(grpc_uri *uri, struct sockaddr_storage *addr, - size_t *len) { - const char *host_port = uri->path; - char *host; - char *port; - int port_num; - int result = 0; - struct sockaddr_in *in = (struct sockaddr_in *)addr; - - if (*host_port == '/') ++host_port; - if (!gpr_split_host_port(host_port, &host, &port)) { - return 0; - } - - memset(in, 0, sizeof(*in)); - *len = sizeof(*in); - in->sin_family = AF_INET; - if (inet_pton(AF_INET, host, &in->sin_addr) == 0) { - gpr_log(GPR_ERROR, "invalid ipv4 address: '%s'", host); - goto done; - } - - if (port != NULL) { - if (sscanf(port, "%d", &port_num) != 1 || port_num < 0 || - port_num > 65535) { - gpr_log(GPR_ERROR, "invalid ipv4 port: '%s'", port); - goto done; - } - in->sin_port = htons((uint16_t)port_num); - } else { - gpr_log(GPR_ERROR, "no port given for ipv4 scheme"); - goto done; - } - - result = 1; -done: - gpr_free(host); - gpr_free(port); - return result; -} - -static int parse_ipv6(grpc_uri *uri, struct sockaddr_storage *addr, - size_t *len) { - const char *host_port = uri->path; - char *host; - char *port; - int port_num; - int result = 0; - struct sockaddr_in6 *in6 = (struct sockaddr_in6 *)addr; - - if (*host_port == '/') ++host_port; - if (!gpr_split_host_port(host_port, &host, &port)) { - return 0; - } - - memset(in6, 0, sizeof(*in6)); - *len = sizeof(*in6); - in6->sin6_family = AF_INET6; - if (inet_pton(AF_INET6, host, &in6->sin6_addr) == 0) { - gpr_log(GPR_ERROR, "invalid ipv6 address: '%s'", host); - goto done; - } - - if (port != NULL) { - if (sscanf(port, "%d", &port_num) != 1 || port_num < 0 || - port_num > 65535) { - gpr_log(GPR_ERROR, "invalid ipv6 port: '%s'", port); - goto done; - } - in6->sin6_port = htons((uint16_t)port_num); - } else { - gpr_log(GPR_ERROR, "no port given for ipv6 scheme"); - goto done; - } - - result = 1; -done: - gpr_free(host); - gpr_free(port); - return result; -} - -static void do_nothing(void *ignored) {} - -static grpc_resolver *sockaddr_create( - grpc_resolver_args *args, const char *default_lb_policy_name, - int parse(grpc_uri *uri, struct sockaddr_storage *dst, size_t *len)) { - int errors_found = 0; /* GPR_FALSE */ - sockaddr_resolver *r; - gpr_slice path_slice; - gpr_slice_buffer path_parts; - - if (0 != strcmp(args->uri->authority, "")) { - gpr_log(GPR_ERROR, "authority based uri's not supported by the %s scheme", - args->uri->scheme); - return NULL; - } - - r = gpr_malloc(sizeof(sockaddr_resolver)); - memset(r, 0, sizeof(*r)); - - r->lb_policy_name = NULL; - if (0 != strcmp(args->uri->query, "")) { - gpr_slice query_slice; - gpr_slice_buffer query_parts; - - query_slice = - gpr_slice_new(args->uri->query, strlen(args->uri->query), do_nothing); - gpr_slice_buffer_init(&query_parts); - gpr_slice_split(query_slice, "=", &query_parts); - GPR_ASSERT(query_parts.count == 2); - if (0 == gpr_slice_str_cmp(query_parts.slices[0], "lb_policy")) { - r->lb_policy_name = gpr_dump_slice(query_parts.slices[1], GPR_DUMP_ASCII); - } - gpr_slice_buffer_destroy(&query_parts); - gpr_slice_unref(query_slice); - } - if (r->lb_policy_name == NULL) { - r->lb_policy_name = gpr_strdup(default_lb_policy_name); - } - - path_slice = - gpr_slice_new(args->uri->path, strlen(args->uri->path), do_nothing); - gpr_slice_buffer_init(&path_parts); - - gpr_slice_split(path_slice, ",", &path_parts); - r->addresses = gpr_malloc(sizeof(grpc_resolved_addresses)); - r->addresses->naddrs = path_parts.count; - r->addresses->addrs = - gpr_malloc(sizeof(grpc_resolved_address) * r->addresses->naddrs); - - for (size_t i = 0; i < r->addresses->naddrs; i++) { - grpc_uri ith_uri = *args->uri; - char *part_str = gpr_dump_slice(path_parts.slices[i], GPR_DUMP_ASCII); - ith_uri.path = part_str; - if (!parse(&ith_uri, - (struct sockaddr_storage *)(&r->addresses->addrs[i].addr), - &r->addresses->addrs[i].len)) { - errors_found = 1; /* GPR_TRUE */ - } - gpr_free(part_str); - if (errors_found) break; - } - - gpr_slice_buffer_destroy(&path_parts); - gpr_slice_unref(path_slice); - if (errors_found) { - gpr_free(r->lb_policy_name); - grpc_resolved_addresses_destroy(r->addresses); - gpr_free(r); - return NULL; - } - - gpr_ref_init(&r->refs, 1); - gpr_mu_init(&r->mu); - grpc_resolver_init(&r->base, &sockaddr_resolver_vtable); - r->subchannel_factory = args->subchannel_factory; - grpc_subchannel_factory_ref(r->subchannel_factory); - - return &r->base; -} - -/* - * FACTORY - */ - -static void sockaddr_factory_ref(grpc_resolver_factory *factory) {} - -static void sockaddr_factory_unref(grpc_resolver_factory *factory) {} - -#define DECL_FACTORY(name, prefix) \ - static grpc_resolver *name##_factory_create_resolver( \ - grpc_resolver_factory *factory, grpc_resolver_args *args) { \ - return sockaddr_create(args, "pick_first", prefix##parse_##name); \ - } \ - static const grpc_resolver_factory_vtable name##_factory_vtable = { \ - sockaddr_factory_ref, sockaddr_factory_unref, \ - name##_factory_create_resolver, prefix##name##_get_default_authority, \ - #name}; \ - static grpc_resolver_factory name##_resolver_factory = { \ - &name##_factory_vtable}; \ - grpc_resolver_factory *grpc_##name##_resolver_factory_create() { \ - return &name##_resolver_factory; \ - } - -#ifdef GPR_HAVE_UNIX_SOCKET -DECL_FACTORY(unix, grpc_) -#endif -DECL_FACTORY(ipv4, ) DECL_FACTORY(ipv6, ) diff --git a/src/core/lib/client_config/resolvers/sockaddr_resolver.h b/src/core/lib/client_config/resolvers/sockaddr_resolver.h deleted file mode 100644 index a8b208176c..0000000000 --- a/src/core/lib/client_config/resolvers/sockaddr_resolver.h +++ /dev/null @@ -1,50 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_SOCKADDR_RESOLVER_H -#define GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_SOCKADDR_RESOLVER_H - -#include <grpc/support/port_platform.h> - -#include "src/core/lib/client_config/resolver_factory.h" - -grpc_resolver_factory *grpc_ipv4_resolver_factory_create(void); - -grpc_resolver_factory *grpc_ipv6_resolver_factory_create(void); - -#ifdef GPR_POSIX_SOCKET -/** Create a unix resolver factory */ -grpc_resolver_factory *grpc_unix_resolver_factory_create(void); -#endif - -#endif /* GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_SOCKADDR_RESOLVER_H */ diff --git a/src/core/lib/client_config/resolvers/zookeeper_resolver.c b/src/core/lib/client_config/resolvers/zookeeper_resolver.c deleted file mode 100644 index a66c9925f0..0000000000 --- a/src/core/lib/client_config/resolvers/zookeeper_resolver.c +++ /dev/null @@ -1,514 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include "src/core/lib/client_config/resolvers/zookeeper_resolver.h" - -#include <string.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/string_util.h> - -#include <grpc/grpc_zookeeper.h> -#include <zookeeper/zookeeper.h> - -#include "src/core/lib/client_config/lb_policy_registry.h" -#include "src/core/lib/client_config/resolver_registry.h" -#include "src/core/lib/iomgr/resolve_address.h" -#include "src/core/lib/json/json.h" -#include "src/core/lib/support/string.h" -#include "src/core/lib/surface/api_trace.h" - -/** Zookeeper session expiration time in milliseconds */ -#define GRPC_ZOOKEEPER_SESSION_TIMEOUT 15000 - -typedef struct { - /** base class: must be first */ - grpc_resolver base; - /** refcount */ - gpr_refcount refs; - /** name to resolve */ - char *name; - /** subchannel factory */ - grpc_subchannel_factory *subchannel_factory; - /** load balancing policy name */ - char *lb_policy_name; - - /** mutex guarding the rest of the state */ - gpr_mu mu; - /** are we currently resolving? */ - int resolving; - /** which version of resolved_config have we published? */ - int published_version; - /** which version of resolved_config is current? */ - int resolved_version; - /** pending next completion, or NULL */ - grpc_closure *next_completion; - /** target config address for next completion */ - grpc_client_config **target_config; - /** current (fully resolved) config */ - grpc_client_config *resolved_config; - - /** zookeeper handle */ - zhandle_t *zookeeper_handle; - /** zookeeper resolved addresses */ - grpc_resolved_addresses *resolved_addrs; - /** total number of addresses to be resolved */ - int resolved_total; - /** number of addresses resolved */ - int resolved_num; -} zookeeper_resolver; - -static void zookeeper_destroy(grpc_exec_ctx *exec_ctx, grpc_resolver *r); - -static void zookeeper_start_resolving_locked(zookeeper_resolver *r); -static void zookeeper_maybe_finish_next_locked(grpc_exec_ctx *exec_ctx, - zookeeper_resolver *r); - -static void zookeeper_shutdown(grpc_exec_ctx *exec_ctx, grpc_resolver *r); -static void zookeeper_channel_saw_error(grpc_exec_ctx *exec_ctx, - grpc_resolver *r); -static void zookeeper_next(grpc_exec_ctx *exec_ctx, grpc_resolver *r, - grpc_client_config **target_config, - grpc_closure *on_complete); - -static const grpc_resolver_vtable zookeeper_resolver_vtable = { - zookeeper_destroy, zookeeper_shutdown, zookeeper_channel_saw_error, - zookeeper_next}; - -static void zookeeper_shutdown(grpc_exec_ctx *exec_ctx, - grpc_resolver *resolver) { - zookeeper_resolver *r = (zookeeper_resolver *)resolver; - grpc_closure *call = NULL; - gpr_mu_lock(&r->mu); - if (r->next_completion != NULL) { - *r->target_config = NULL; - call = r->next_completion; - r->next_completion = NULL; - } - zookeeper_close(r->zookeeper_handle); - gpr_mu_unlock(&r->mu); - if (call != NULL) { - call->cb(exec_ctx, call->cb_arg, 1); - } -} - -static void zookeeper_channel_saw_error(grpc_exec_ctx *exec_ctx, - grpc_resolver *resolver) { - zookeeper_resolver *r = (zookeeper_resolver *)resolver; - gpr_mu_lock(&r->mu); - if (r->resolving == 0) { - zookeeper_start_resolving_locked(r); - } - gpr_mu_unlock(&r->mu); -} - -static void zookeeper_next(grpc_exec_ctx *exec_ctx, grpc_resolver *resolver, - grpc_client_config **target_config, - grpc_closure *on_complete) { - zookeeper_resolver *r = (zookeeper_resolver *)resolver; - gpr_mu_lock(&r->mu); - GPR_ASSERT(r->next_completion == NULL); - r->next_completion = on_complete; - r->target_config = target_config; - if (r->resolved_version == 0 && r->resolving == 0) { - zookeeper_start_resolving_locked(r); - } else { - zookeeper_maybe_finish_next_locked(exec_ctx, r); - } - gpr_mu_unlock(&r->mu); -} - -/** Zookeeper global watcher for connection management - TODO: better connection management besides logs */ -static void zookeeper_global_watcher(zhandle_t *zookeeper_handle, int type, - int state, const char *path, - void *watcher_ctx) { - if (type == ZOO_SESSION_EVENT) { - if (state == ZOO_EXPIRED_SESSION_STATE) { - gpr_log(GPR_ERROR, "Zookeeper session expired"); - } else if (state == ZOO_AUTH_FAILED_STATE) { - gpr_log(GPR_ERROR, "Zookeeper authentication failed"); - } - } -} - -/** Zookeeper watcher triggered by changes to watched nodes - Once triggered, it tries to resolve again to get updated addresses */ -static void zookeeper_watcher(zhandle_t *zookeeper_handle, int type, int state, - const char *path, void *watcher_ctx) { - if (watcher_ctx != NULL) { - zookeeper_resolver *r = (zookeeper_resolver *)watcher_ctx; - if (state == ZOO_CONNECTED_STATE) { - gpr_mu_lock(&r->mu); - if (r->resolving == 0) { - zookeeper_start_resolving_locked(r); - } - gpr_mu_unlock(&r->mu); - } - } -} - -/** Callback function after getting all resolved addresses - Creates a subchannel for each address */ -static void zookeeper_on_resolved(grpc_exec_ctx *exec_ctx, void *arg, - grpc_resolved_addresses *addresses) { - zookeeper_resolver *r = arg; - grpc_client_config *config = NULL; - grpc_lb_policy *lb_policy; - - if (addresses != NULL) { - grpc_lb_policy_args lb_policy_args; - config = grpc_client_config_create(); - - lb_policy_args.addresses = addresses; - lb_policy_args.subchannel_factory = r->subchannel_factory; - lb_policy = - grpc_lb_policy_create(exec_ctx, r->lb_policy_name, &lb_policy_args); - - if (lb_policy != NULL) { - grpc_client_config_set_lb_policy(config, lb_policy); - GRPC_LB_POLICY_UNREF(exec_ctx, lb_policy, "construction"); - } - grpc_resolved_addresses_destroy(addresses); - } - gpr_mu_lock(&r->mu); - GPR_ASSERT(r->resolving == 1); - r->resolving = 0; - if (r->resolved_config != NULL) { - grpc_client_config_unref(exec_ctx, r->resolved_config); - } - r->resolved_config = config; - r->resolved_version++; - zookeeper_maybe_finish_next_locked(exec_ctx, r); - gpr_mu_unlock(&r->mu); - - GRPC_RESOLVER_UNREF(exec_ctx, &r->base, "zookeeper-resolving"); -} - -/** Callback function for each DNS resolved address */ -static void zookeeper_dns_resolved(grpc_exec_ctx *exec_ctx, void *arg, - grpc_resolved_addresses *addresses) { - size_t i; - zookeeper_resolver *r = arg; - int resolve_done = 0; - - gpr_mu_lock(&r->mu); - r->resolved_num++; - r->resolved_addrs->addrs = - gpr_realloc(r->resolved_addrs->addrs, - sizeof(grpc_resolved_address) * - (r->resolved_addrs->naddrs + addresses->naddrs)); - for (i = 0; i < addresses->naddrs; i++) { - memcpy(r->resolved_addrs->addrs[i + r->resolved_addrs->naddrs].addr, - addresses->addrs[i].addr, addresses->addrs[i].len); - r->resolved_addrs->addrs[i + r->resolved_addrs->naddrs].len = - addresses->addrs[i].len; - } - - r->resolved_addrs->naddrs += addresses->naddrs; - grpc_resolved_addresses_destroy(addresses); - - /** Wait for all addresses to be resolved */ - resolve_done = (r->resolved_num == r->resolved_total); - gpr_mu_unlock(&r->mu); - if (resolve_done) { - zookeeper_on_resolved(exec_ctx, r, r->resolved_addrs); - } -} - -/** Parses JSON format address of a zookeeper node */ -static char *zookeeper_parse_address(const char *value, size_t value_len) { - grpc_json *json; - grpc_json *cur; - const char *host; - const char *port; - char *buffer; - char *address = NULL; - - buffer = gpr_malloc(value_len); - memcpy(buffer, value, value_len); - json = grpc_json_parse_string_with_len(buffer, value_len); - if (json != NULL) { - host = NULL; - port = NULL; - for (cur = json->child; cur != NULL; cur = cur->next) { - if (!strcmp(cur->key, "host")) { - host = cur->value; - if (port != NULL) { - break; - } - } else if (!strcmp(cur->key, "port")) { - port = cur->value; - if (host != NULL) { - break; - } - } - } - if (host != NULL && port != NULL) { - gpr_asprintf(&address, "%s:%s", host, port); - } - grpc_json_destroy(json); - } - gpr_free(buffer); - - return address; -} - -static void zookeeper_get_children_node_completion(int rc, const char *value, - int value_len, - const struct Stat *stat, - const void *arg) { - char *address = NULL; - zookeeper_resolver *r = (zookeeper_resolver *)arg; - int resolve_done = 0; - grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT; - - if (rc != 0) { - gpr_log(GPR_ERROR, "Error in getting a child node of %s", r->name); - grpc_exec_ctx_finish(&exec_ctx); - return; - } - - address = zookeeper_parse_address(value, (size_t)value_len); - if (address != NULL) { - /** Further resolves address by DNS */ - grpc_resolve_address(address, NULL, zookeeper_dns_resolved, r); - gpr_free(address); - } else { - gpr_log(GPR_ERROR, "Error in resolving a child node of %s", r->name); - gpr_mu_lock(&r->mu); - r->resolved_total--; - resolve_done = (r->resolved_num == r->resolved_total); - gpr_mu_unlock(&r->mu); - if (resolve_done) { - zookeeper_on_resolved(&exec_ctx, r, r->resolved_addrs); - } - } - - grpc_exec_ctx_finish(&exec_ctx); -} - -static void zookeeper_get_children_completion( - int rc, const struct String_vector *children, const void *arg) { - char *path; - int status; - int i; - zookeeper_resolver *r = (zookeeper_resolver *)arg; - - if (rc != 0) { - gpr_log(GPR_ERROR, "Error in getting zookeeper children of %s", r->name); - return; - } - - if (children->count == 0) { - gpr_log(GPR_ERROR, "Error in resolving zookeeper address %s", r->name); - return; - } - - r->resolved_addrs = gpr_malloc(sizeof(grpc_resolved_addresses)); - r->resolved_addrs->addrs = NULL; - r->resolved_addrs->naddrs = 0; - r->resolved_total = children->count; - - /** TODO: Replace expensive heap allocation with stack - if we can get maximum length of zookeeper path */ - for (i = 0; i < children->count; i++) { - gpr_asprintf(&path, "%s/%s", r->name, children->data[i]); - status = zoo_awget(r->zookeeper_handle, path, zookeeper_watcher, r, - zookeeper_get_children_node_completion, r); - gpr_free(path); - if (status != 0) { - gpr_log(GPR_ERROR, "Error in getting zookeeper node %s", path); - } - } -} - -static void zookeeper_get_node_completion(int rc, const char *value, - int value_len, - const struct Stat *stat, - const void *arg) { - int status; - char *address = NULL; - zookeeper_resolver *r = (zookeeper_resolver *)arg; - r->resolved_addrs = NULL; - r->resolved_total = 0; - r->resolved_num = 0; - - if (rc != 0) { - gpr_log(GPR_ERROR, "Error in getting zookeeper node %s", r->name); - return; - } - - /** If zookeeper node of path r->name does not have address - (i.e. service node), get its children */ - address = zookeeper_parse_address(value, (size_t)value_len); - if (address != NULL) { - r->resolved_addrs = gpr_malloc(sizeof(grpc_resolved_addresses)); - r->resolved_addrs->addrs = NULL; - r->resolved_addrs->naddrs = 0; - r->resolved_total = 1; - /** Further resolves address by DNS */ - grpc_resolve_address(address, NULL, zookeeper_dns_resolved, r); - gpr_free(address); - return; - } - - status = zoo_awget_children(r->zookeeper_handle, r->name, zookeeper_watcher, - r, zookeeper_get_children_completion, r); - if (status != 0) { - gpr_log(GPR_ERROR, "Error in getting zookeeper children of %s", r->name); - } -} - -static void zookeeper_resolve_address(zookeeper_resolver *r) { - int status; - status = zoo_awget(r->zookeeper_handle, r->name, zookeeper_watcher, r, - zookeeper_get_node_completion, r); - if (status != 0) { - gpr_log(GPR_ERROR, "Error in getting zookeeper node %s", r->name); - } -} - -static void zookeeper_start_resolving_locked(zookeeper_resolver *r) { - GRPC_RESOLVER_REF(&r->base, "zookeeper-resolving"); - GPR_ASSERT(r->resolving == 0); - r->resolving = 1; - zookeeper_resolve_address(r); -} - -static void zookeeper_maybe_finish_next_locked(grpc_exec_ctx *exec_ctx, - zookeeper_resolver *r) { - if (r->next_completion != NULL && - r->resolved_version != r->published_version) { - *r->target_config = r->resolved_config; - if (r->resolved_config != NULL) { - grpc_client_config_ref(r->resolved_config); - } - grpc_exec_ctx_enqueue(exec_ctx, r->next_completion, true, NULL); - r->next_completion = NULL; - r->published_version = r->resolved_version; - } -} - -static void zookeeper_destroy(grpc_exec_ctx *exec_ctx, grpc_resolver *gr) { - zookeeper_resolver *r = (zookeeper_resolver *)gr; - gpr_mu_destroy(&r->mu); - if (r->resolved_config != NULL) { - grpc_client_config_unref(exec_ctx, r->resolved_config); - } - grpc_subchannel_factory_unref(exec_ctx, r->subchannel_factory); - gpr_free(r->name); - gpr_free(r->lb_policy_name); - gpr_free(r); -} - -static grpc_resolver *zookeeper_create(grpc_resolver_args *args, - const char *lb_policy_name) { - zookeeper_resolver *r; - size_t length; - char *path = args->uri->path; - - if (0 == strcmp(args->uri->authority, "")) { - gpr_log(GPR_ERROR, "No authority specified in zookeeper uri"); - return NULL; - } - - /** Removes the trailing slash if exists */ - length = strlen(path); - if (length > 1 && path[length - 1] == '/') { - path[length - 1] = 0; - } - - r = gpr_malloc(sizeof(zookeeper_resolver)); - memset(r, 0, sizeof(*r)); - gpr_ref_init(&r->refs, 1); - gpr_mu_init(&r->mu); - grpc_resolver_init(&r->base, &zookeeper_resolver_vtable); - r->name = gpr_strdup(path); - - r->subchannel_factory = args->subchannel_factory; - grpc_subchannel_factory_ref(r->subchannel_factory); - - r->lb_policy_name = gpr_strdup(lb_policy_name); - - /** Initializes zookeeper client */ - zoo_set_debug_level(ZOO_LOG_LEVEL_WARN); - r->zookeeper_handle = - zookeeper_init(args->uri->authority, zookeeper_global_watcher, - GRPC_ZOOKEEPER_SESSION_TIMEOUT, 0, 0, 0); - if (r->zookeeper_handle == NULL) { - gpr_log(GPR_ERROR, "Unable to connect to zookeeper server"); - return NULL; - } - - return &r->base; -} - -static void zookeeper_plugin_init() { - grpc_register_resolver_type(grpc_zookeeper_resolver_factory_create()); -} - -void grpc_zookeeper_register() { - GRPC_API_TRACE("grpc_zookeeper_register(void)", 0, ()); - grpc_register_plugin(zookeeper_plugin_init, NULL); -} - -/* - * FACTORY - */ - -static void zookeeper_factory_ref(grpc_resolver_factory *factory) {} - -static void zookeeper_factory_unref(grpc_resolver_factory *factory) {} - -static char *zookeeper_factory_get_default_hostname( - grpc_resolver_factory *factory, grpc_uri *uri) { - return NULL; -} - -static grpc_resolver *zookeeper_factory_create_resolver( - grpc_resolver_factory *factory, grpc_resolver_args *args) { - return zookeeper_create(args, "pick_first"); -} - -static const grpc_resolver_factory_vtable zookeeper_factory_vtable = { - zookeeper_factory_ref, zookeeper_factory_unref, - zookeeper_factory_create_resolver, zookeeper_factory_get_default_hostname, - "zookeeper"}; - -static grpc_resolver_factory zookeeper_resolver_factory = { - &zookeeper_factory_vtable}; - -grpc_resolver_factory *grpc_zookeeper_resolver_factory_create() { - return &zookeeper_resolver_factory; -} diff --git a/src/core/lib/client_config/resolvers/zookeeper_resolver.h b/src/core/lib/client_config/resolvers/zookeeper_resolver.h deleted file mode 100644 index a067bd0bdb..0000000000 --- a/src/core/lib/client_config/resolvers/zookeeper_resolver.h +++ /dev/null @@ -1,42 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_ZOOKEEPER_RESOLVER_H -#define GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_ZOOKEEPER_RESOLVER_H - -#include "src/core/lib/client_config/resolver_factory.h" - -/** Create a zookeeper resolver factory */ -grpc_resolver_factory *grpc_zookeeper_resolver_factory_create(void); - -#endif /* GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_ZOOKEEPER_RESOLVER_H */ diff --git a/src/core/lib/iomgr/ev_poll_and_epoll_posix.c b/src/core/lib/iomgr/ev_poll_and_epoll_posix.c new file mode 100644 index 0000000000..dba335490b --- /dev/null +++ b/src/core/lib/iomgr/ev_poll_and_epoll_posix.c @@ -0,0 +1,1936 @@ +/* + * + * Copyright 2015-2016, Google Inc. + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are + * met: + * + * * Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * * Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following disclaimer + * in the documentation and/or other materials provided with the + * distribution. + * * Neither the name of Google Inc. nor the names of its + * contributors may be used to endorse or promote products derived from + * this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + */ + +/* This file will be removed shortly: it's here to keep refactoring + * steps simple and auditable. + * It's the combination of the old files: + * - fd_posix.{h,c} + * - pollset_posix.{h,c} + * - pullset_multipoller_with_{poll,epoll}.{h,c} + * The new version will be split into: + * - ev_poll_posix.{h,c} + * - ev_epoll_posix.{h,c} + */ + +#include <grpc/support/port_platform.h> + +#ifdef GPR_POSIX_SOCKET + +#include "src/core/lib/iomgr/ev_poll_and_epoll_posix.h" + +#include <assert.h> +#include <errno.h> +#include <poll.h> +#include <string.h> +#include <sys/socket.h> +#include <unistd.h> + +#include <grpc/support/alloc.h> +#include <grpc/support/log.h> +#include <grpc/support/string_util.h> +#include <grpc/support/tls.h> +#include <grpc/support/useful.h> + +#include "src/core/lib/iomgr/iomgr_internal.h" +#include "src/core/lib/iomgr/wakeup_fd_posix.h" +#include "src/core/lib/profiling/timers.h" +#include "src/core/lib/support/block_annotate.h" + +/******************************************************************************* + * FD declarations + */ + +typedef struct grpc_fd_watcher { + struct grpc_fd_watcher *next; + struct grpc_fd_watcher *prev; + grpc_pollset *pollset; + grpc_pollset_worker *worker; + grpc_fd *fd; +} grpc_fd_watcher; + +struct grpc_fd { + int fd; + /* refst format: + bit0: 1=active/0=orphaned + bit1-n: refcount + meaning that mostly we ref by two to avoid altering the orphaned bit, + and just unref by 1 when we're ready to flag the object as orphaned */ + gpr_atm refst; + + gpr_mu mu; + int shutdown; + int closed; + int released; + + /* The watcher list. + + The following watcher related fields are protected by watcher_mu. + + An fd_watcher is an ephemeral object created when an fd wants to + begin polling, and destroyed after the poll. + + It denotes the fd's interest in whether to read poll or write poll + or both or neither on this fd. + + If a watcher is asked to poll for reads or writes, the read_watcher + or write_watcher fields are set respectively. A watcher may be asked + to poll for both, in which case both fields will be set. + + read_watcher and write_watcher may be NULL if no watcher has been + asked to poll for reads or writes. + + If an fd_watcher is not asked to poll for reads or writes, it's added + to a linked list of inactive watchers, rooted at inactive_watcher_root. + If at a later time there becomes need of a poller to poll, one of + the inactive pollers may be kicked out of their poll loops to take + that responsibility. */ + grpc_fd_watcher inactive_watcher_root; + grpc_fd_watcher *read_watcher; + grpc_fd_watcher *write_watcher; + + grpc_closure *read_closure; + grpc_closure *write_closure; + + struct grpc_fd *freelist_next; + + grpc_closure *on_done_closure; + + grpc_iomgr_object iomgr_object; +}; + +/* Begin polling on an fd. + Registers that the given pollset is interested in this fd - so that if read + or writability interest changes, the pollset can be kicked to pick up that + new interest. + Return value is: + (fd_needs_read? read_mask : 0) | (fd_needs_write? write_mask : 0) + i.e. a combination of read_mask and write_mask determined by the fd's current + interest in said events. + Polling strategies that do not need to alter their behavior depending on the + fd's current interest (such as epoll) do not need to call this function. + MUST NOT be called with a pollset lock taken */ +static uint32_t fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset, + grpc_pollset_worker *worker, uint32_t read_mask, + uint32_t write_mask, grpc_fd_watcher *rec); +/* Complete polling previously started with fd_begin_poll + MUST NOT be called with a pollset lock taken + if got_read or got_write are 1, also does the become_{readable,writable} as + appropriate. */ +static void fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *rec, + int got_read, int got_write); + +/* Return 1 if this fd is orphaned, 0 otherwise */ +static bool fd_is_orphaned(grpc_fd *fd); + +/* Reference counting for fds */ +/*#define GRPC_FD_REF_COUNT_DEBUG*/ +#ifdef GRPC_FD_REF_COUNT_DEBUG +static void fd_ref(grpc_fd *fd, const char *reason, const char *file, int line); +static void fd_unref(grpc_fd *fd, const char *reason, const char *file, + int line); +#define GRPC_FD_REF(fd, reason) fd_ref(fd, reason, __FILE__, __LINE__) +#define GRPC_FD_UNREF(fd, reason) fd_unref(fd, reason, __FILE__, __LINE__) +#else +static void fd_ref(grpc_fd *fd); +static void fd_unref(grpc_fd *fd); +#define GRPC_FD_REF(fd, reason) fd_ref(fd) +#define GRPC_FD_UNREF(fd, reason) fd_unref(fd) +#endif + +static void fd_global_init(void); +static void fd_global_shutdown(void); + +#define CLOSURE_NOT_READY ((grpc_closure *)0) +#define CLOSURE_READY ((grpc_closure *)1) + +/******************************************************************************* + * pollset declarations + */ + +typedef struct grpc_pollset_vtable grpc_pollset_vtable; + +typedef struct grpc_cached_wakeup_fd { + grpc_wakeup_fd fd; + struct grpc_cached_wakeup_fd *next; +} grpc_cached_wakeup_fd; + +struct grpc_pollset_worker { + grpc_cached_wakeup_fd *wakeup_fd; + int reevaluate_polling_on_wakeup; + int kicked_specifically; + struct grpc_pollset_worker *next; + struct grpc_pollset_worker *prev; +}; + +struct grpc_pollset { + /* pollsets under posix can mutate representation as fds are added and + removed. + For example, we may choose a poll() based implementation on linux for + few fds, and an epoll() based implementation for many fds */ + const grpc_pollset_vtable *vtable; + gpr_mu mu; + grpc_pollset_worker root_worker; + int in_flight_cbs; + int shutting_down; + int called_shutdown; + int kicked_without_pollers; + grpc_closure *shutdown_done; + grpc_closure_list idle_jobs; + union { + int fd; + void *ptr; + } data; + /* Local cache of eventfds for workers */ + grpc_cached_wakeup_fd *local_wakeup_cache; +}; + +struct grpc_pollset_vtable { + void (*add_fd)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + struct grpc_fd *fd, int and_unlock_pollset); + void (*maybe_work_and_unlock)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_pollset_worker *worker, + gpr_timespec deadline, gpr_timespec now); + void (*finish_shutdown)(grpc_pollset *pollset); + void (*destroy)(grpc_pollset *pollset); +}; + +/* Add an fd to a pollset */ +static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + struct grpc_fd *fd); + +static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd); + +/* Convert a timespec to milliseconds: + - very small or negative poll times are clamped to zero to do a + non-blocking poll (which becomes spin polling) + - other small values are rounded up to one millisecond + - longer than a millisecond polls are rounded up to the next nearest + millisecond to avoid spinning + - infinite timeouts are converted to -1 */ +static int poll_deadline_to_millis_timeout(gpr_timespec deadline, + gpr_timespec now); + +/* Allow kick to wakeup the currently polling worker */ +#define GRPC_POLLSET_CAN_KICK_SELF 1 +/* Force the wakee to repoll when awoken */ +#define GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP 2 +/* As per pollset_kick, with an extended set of flags (defined above) + -- mostly for fd_posix's use. */ +static void pollset_kick_ext(grpc_pollset *p, + grpc_pollset_worker *specific_worker, + uint32_t flags); + +/* turn a pollset into a multipoller: platform specific */ +typedef void (*platform_become_multipoller_type)(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, + struct grpc_fd **fds, + size_t fd_count); +static platform_become_multipoller_type platform_become_multipoller; + +/* Return 1 if the pollset has active threads in pollset_work (pollset must + * be locked) */ +static int pollset_has_workers(grpc_pollset *pollset); + +static void remove_fd_from_all_epoll_sets(int fd); + +/******************************************************************************* + * pollset_set definitions + */ + +struct grpc_pollset_set { + gpr_mu mu; + + size_t pollset_count; + size_t pollset_capacity; + grpc_pollset **pollsets; + + size_t pollset_set_count; + size_t pollset_set_capacity; + struct grpc_pollset_set **pollset_sets; + + size_t fd_count; + size_t fd_capacity; + grpc_fd **fds; +}; + +/******************************************************************************* + * fd_posix.c + */ + +/* We need to keep a freelist not because of any concerns of malloc performance + * but instead so that implementations with multiple threads in (for example) + * epoll_wait deal with the race between pollset removal and incoming poll + * notifications. + * + * The problem is that the poller ultimately holds a reference to this + * object, so it is very difficult to know when is safe to free it, at least + * without some expensive synchronization. + * + * If we keep the object freelisted, in the worst case losing this race just + * becomes a spurious read notification on a reused fd. + */ +/* TODO(klempner): We could use some form of polling generation count to know + * when these are safe to free. */ +/* TODO(klempner): Consider disabling freelisting if we don't have multiple + * threads in poll on the same fd */ +/* TODO(klempner): Batch these allocations to reduce fragmentation */ +static grpc_fd *fd_freelist = NULL; +static gpr_mu fd_freelist_mu; + +static void freelist_fd(grpc_fd *fd) { + gpr_mu_lock(&fd_freelist_mu); + fd->freelist_next = fd_freelist; + fd_freelist = fd; + grpc_iomgr_unregister_object(&fd->iomgr_object); + gpr_mu_unlock(&fd_freelist_mu); +} + +static grpc_fd *alloc_fd(int fd) { + grpc_fd *r = NULL; + gpr_mu_lock(&fd_freelist_mu); + if (fd_freelist != NULL) { + r = fd_freelist; + fd_freelist = fd_freelist->freelist_next; + } + gpr_mu_unlock(&fd_freelist_mu); + if (r == NULL) { + r = gpr_malloc(sizeof(grpc_fd)); + gpr_mu_init(&r->mu); + } + + gpr_mu_lock(&r->mu); + gpr_atm_rel_store(&r->refst, 1); + r->shutdown = 0; + r->read_closure = CLOSURE_NOT_READY; + r->write_closure = CLOSURE_NOT_READY; + r->fd = fd; + r->inactive_watcher_root.next = r->inactive_watcher_root.prev = + &r->inactive_watcher_root; + r->freelist_next = NULL; + r->read_watcher = r->write_watcher = NULL; + r->on_done_closure = NULL; + r->closed = 0; + r->released = 0; + gpr_mu_unlock(&r->mu); + return r; +} + +static void destroy(grpc_fd *fd) { + gpr_mu_destroy(&fd->mu); + gpr_free(fd); +} + +#ifdef GRPC_FD_REF_COUNT_DEBUG +#define REF_BY(fd, n, reason) ref_by(fd, n, reason, __FILE__, __LINE__) +#define UNREF_BY(fd, n, reason) unref_by(fd, n, reason, __FILE__, __LINE__) +static void ref_by(grpc_fd *fd, int n, const char *reason, const char *file, + int line) { + gpr_log(GPR_DEBUG, "FD %d %p ref %d %d -> %d [%s; %s:%d]", fd->fd, fd, n, + gpr_atm_no_barrier_load(&fd->refst), + gpr_atm_no_barrier_load(&fd->refst) + n, reason, file, line); +#else +#define REF_BY(fd, n, reason) ref_by(fd, n) +#define UNREF_BY(fd, n, reason) unref_by(fd, n) +static void ref_by(grpc_fd *fd, int n) { +#endif + GPR_ASSERT(gpr_atm_no_barrier_fetch_add(&fd->refst, n) > 0); +} + +#ifdef GRPC_FD_REF_COUNT_DEBUG +static void unref_by(grpc_fd *fd, int n, const char *reason, const char *file, + int line) { + gpr_atm old; + gpr_log(GPR_DEBUG, "FD %d %p unref %d %d -> %d [%s; %s:%d]", fd->fd, fd, n, + gpr_atm_no_barrier_load(&fd->refst), + gpr_atm_no_barrier_load(&fd->refst) - n, reason, file, line); +#else +static void unref_by(grpc_fd *fd, int n) { + gpr_atm old; +#endif + old = gpr_atm_full_fetch_add(&fd->refst, -n); + if (old == n) { + freelist_fd(fd); + } else { + GPR_ASSERT(old > n); + } +} + +static void fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); } + +static void fd_global_shutdown(void) { + gpr_mu_lock(&fd_freelist_mu); + gpr_mu_unlock(&fd_freelist_mu); + while (fd_freelist != NULL) { + grpc_fd *fd = fd_freelist; + fd_freelist = fd_freelist->freelist_next; + destroy(fd); + } + gpr_mu_destroy(&fd_freelist_mu); +} + +static grpc_fd *fd_create(int fd, const char *name) { + grpc_fd *r = alloc_fd(fd); + char *name2; + gpr_asprintf(&name2, "%s fd=%d", name, fd); + grpc_iomgr_register_object(&r->iomgr_object, name2); + gpr_free(name2); +#ifdef GRPC_FD_REF_COUNT_DEBUG + gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, r, name); +#endif + return r; +} + +static bool fd_is_orphaned(grpc_fd *fd) { + return (gpr_atm_acq_load(&fd->refst) & 1) == 0; +} + +static void pollset_kick_locked(grpc_fd_watcher *watcher) { + gpr_mu_lock(&watcher->pollset->mu); + GPR_ASSERT(watcher->worker); + pollset_kick_ext(watcher->pollset, watcher->worker, + GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP); + gpr_mu_unlock(&watcher->pollset->mu); +} + +static void maybe_wake_one_watcher_locked(grpc_fd *fd) { + if (fd->inactive_watcher_root.next != &fd->inactive_watcher_root) { + pollset_kick_locked(fd->inactive_watcher_root.next); + } else if (fd->read_watcher) { + pollset_kick_locked(fd->read_watcher); + } else if (fd->write_watcher) { + pollset_kick_locked(fd->write_watcher); + } +} + +static void wake_all_watchers_locked(grpc_fd *fd) { + grpc_fd_watcher *watcher; + for (watcher = fd->inactive_watcher_root.next; + watcher != &fd->inactive_watcher_root; watcher = watcher->next) { + pollset_kick_locked(watcher); + } + if (fd->read_watcher) { + pollset_kick_locked(fd->read_watcher); + } + if (fd->write_watcher && fd->write_watcher != fd->read_watcher) { + pollset_kick_locked(fd->write_watcher); + } +} + +static int has_watchers(grpc_fd *fd) { + return fd->read_watcher != NULL || fd->write_watcher != NULL || + fd->inactive_watcher_root.next != &fd->inactive_watcher_root; +} + +static void close_fd_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + fd->closed = 1; + if (!fd->released) { + close(fd->fd); + } else { + remove_fd_from_all_epoll_sets(fd->fd); + } + grpc_exec_ctx_enqueue(exec_ctx, fd->on_done_closure, true, NULL); +} + +static int fd_wrapped_fd(grpc_fd *fd) { + if (fd->released || fd->closed) { + return -1; + } else { + return fd->fd; + } +} + +static void fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *on_done, int *release_fd, + const char *reason) { + fd->on_done_closure = on_done; + fd->released = release_fd != NULL; + if (!fd->released) { + shutdown(fd->fd, SHUT_RDWR); + } else { + *release_fd = fd->fd; + } + gpr_mu_lock(&fd->mu); + REF_BY(fd, 1, reason); /* remove active status, but keep referenced */ + if (!has_watchers(fd)) { + close_fd_locked(exec_ctx, fd); + } else { + wake_all_watchers_locked(fd); + } + gpr_mu_unlock(&fd->mu); + UNREF_BY(fd, 2, reason); /* drop the reference */ +} + +/* increment refcount by two to avoid changing the orphan bit */ +#ifdef GRPC_FD_REF_COUNT_DEBUG +static void fd_ref(grpc_fd *fd, const char *reason, const char *file, + int line) { + ref_by(fd, 2, reason, file, line); +} + +static void fd_unref(grpc_fd *fd, const char *reason, const char *file, + int line) { + unref_by(fd, 2, reason, file, line); +} +#else +static void fd_ref(grpc_fd *fd) { ref_by(fd, 2); } + +static void fd_unref(grpc_fd *fd) { unref_by(fd, 2); } +#endif + +static void notify_on_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure **st, grpc_closure *closure) { + if (*st == CLOSURE_NOT_READY) { + /* not ready ==> switch to a waiting state by setting the closure */ + *st = closure; + } else if (*st == CLOSURE_READY) { + /* already ready ==> queue the closure to run immediately */ + *st = CLOSURE_NOT_READY; + grpc_exec_ctx_enqueue(exec_ctx, closure, !fd->shutdown, NULL); + maybe_wake_one_watcher_locked(fd); + } else { + /* upcallptr was set to a different closure. This is an error! */ + gpr_log(GPR_ERROR, + "User called a notify_on function with a previous callback still " + "pending"); + abort(); + } +} + +/* returns 1 if state becomes not ready */ +static int set_ready_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure **st) { + if (*st == CLOSURE_READY) { + /* duplicate ready ==> ignore */ + return 0; + } else if (*st == CLOSURE_NOT_READY) { + /* not ready, and not waiting ==> flag ready */ + *st = CLOSURE_READY; + return 0; + } else { + /* waiting ==> queue closure */ + grpc_exec_ctx_enqueue(exec_ctx, *st, !fd->shutdown, NULL); + *st = CLOSURE_NOT_READY; + return 1; + } +} + +static void fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + gpr_mu_lock(&fd->mu); + GPR_ASSERT(!fd->shutdown); + fd->shutdown = 1; + set_ready_locked(exec_ctx, fd, &fd->read_closure); + set_ready_locked(exec_ctx, fd, &fd->write_closure); + gpr_mu_unlock(&fd->mu); +} + +static void fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure) { + gpr_mu_lock(&fd->mu); + notify_on_locked(exec_ctx, fd, &fd->read_closure, closure); + gpr_mu_unlock(&fd->mu); +} + +static void fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure) { + gpr_mu_lock(&fd->mu); + notify_on_locked(exec_ctx, fd, &fd->write_closure, closure); + gpr_mu_unlock(&fd->mu); +} + +static uint32_t fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset, + grpc_pollset_worker *worker, uint32_t read_mask, + uint32_t write_mask, grpc_fd_watcher *watcher) { + uint32_t mask = 0; + grpc_closure *cur; + int requested; + /* keep track of pollers that have requested our events, in case they change + */ + GRPC_FD_REF(fd, "poll"); + + gpr_mu_lock(&fd->mu); + + /* if we are shutdown, then don't add to the watcher set */ + if (fd->shutdown) { + watcher->fd = NULL; + watcher->pollset = NULL; + watcher->worker = NULL; + gpr_mu_unlock(&fd->mu); + GRPC_FD_UNREF(fd, "poll"); + return 0; + } + + /* if there is nobody polling for read, but we need to, then start doing so */ + cur = fd->read_closure; + requested = cur != CLOSURE_READY; + if (read_mask && fd->read_watcher == NULL && requested) { + fd->read_watcher = watcher; + mask |= read_mask; + } + /* if there is nobody polling for write, but we need to, then start doing so + */ + cur = fd->write_closure; + requested = cur != CLOSURE_READY; + if (write_mask && fd->write_watcher == NULL && requested) { + fd->write_watcher = watcher; + mask |= write_mask; + } + /* if not polling, remember this watcher in case we need someone to later */ + if (mask == 0 && worker != NULL) { + watcher->next = &fd->inactive_watcher_root; + watcher->prev = watcher->next->prev; + watcher->next->prev = watcher->prev->next = watcher; + } + watcher->pollset = pollset; + watcher->worker = worker; + watcher->fd = fd; + gpr_mu_unlock(&fd->mu); + + return mask; +} + +static void fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *watcher, + int got_read, int got_write) { + int was_polling = 0; + int kick = 0; + grpc_fd *fd = watcher->fd; + + if (fd == NULL) { + return; + } + + gpr_mu_lock(&fd->mu); + + if (watcher == fd->read_watcher) { + /* remove read watcher, kick if we still need a read */ + was_polling = 1; + if (!got_read) { + kick = 1; + } + fd->read_watcher = NULL; + } + if (watcher == fd->write_watcher) { + /* remove write watcher, kick if we still need a write */ + was_polling = 1; + if (!got_write) { + kick = 1; + } + fd->write_watcher = NULL; + } + if (!was_polling && watcher->worker != NULL) { + /* remove from inactive list */ + watcher->next->prev = watcher->prev; + watcher->prev->next = watcher->next; + } + if (got_read) { + if (set_ready_locked(exec_ctx, fd, &fd->read_closure)) { + kick = 1; + } + } + if (got_write) { + if (set_ready_locked(exec_ctx, fd, &fd->write_closure)) { + kick = 1; + } + } + if (kick) { + maybe_wake_one_watcher_locked(fd); + } + if (fd_is_orphaned(fd) && !has_watchers(fd) && !fd->closed) { + close_fd_locked(exec_ctx, fd); + } + gpr_mu_unlock(&fd->mu); + + GRPC_FD_UNREF(fd, "poll"); +} + +/******************************************************************************* + * pollset_posix.c + */ + +GPR_TLS_DECL(g_current_thread_poller); +GPR_TLS_DECL(g_current_thread_worker); + +/** The alarm system needs to be able to wakeup 'some poller' sometimes + * (specifically when a new alarm needs to be triggered earlier than the next + * alarm 'epoch'). + * This wakeup_fd gives us something to alert on when such a case occurs. */ +grpc_wakeup_fd grpc_global_wakeup_fd; + +static void remove_worker(grpc_pollset *p, grpc_pollset_worker *worker) { + worker->prev->next = worker->next; + worker->next->prev = worker->prev; +} + +static int pollset_has_workers(grpc_pollset *p) { + return p->root_worker.next != &p->root_worker; +} + +static grpc_pollset_worker *pop_front_worker(grpc_pollset *p) { + if (pollset_has_workers(p)) { + grpc_pollset_worker *w = p->root_worker.next; + remove_worker(p, w); + return w; + } else { + return NULL; + } +} + +static void push_back_worker(grpc_pollset *p, grpc_pollset_worker *worker) { + worker->next = &p->root_worker; + worker->prev = worker->next->prev; + worker->prev->next = worker->next->prev = worker; +} + +static void push_front_worker(grpc_pollset *p, grpc_pollset_worker *worker) { + worker->prev = &p->root_worker; + worker->next = worker->prev->next; + worker->prev->next = worker->next->prev = worker; +} + +static void pollset_kick_ext(grpc_pollset *p, + grpc_pollset_worker *specific_worker, + uint32_t flags) { + GPR_TIMER_BEGIN("pollset_kick_ext", 0); + + /* pollset->mu already held */ + if (specific_worker != NULL) { + if (specific_worker == GRPC_POLLSET_KICK_BROADCAST) { + GPR_TIMER_BEGIN("pollset_kick_ext.broadcast", 0); + GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0); + for (specific_worker = p->root_worker.next; + specific_worker != &p->root_worker; + specific_worker = specific_worker->next) { + grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); + } + p->kicked_without_pollers = 1; + GPR_TIMER_END("pollset_kick_ext.broadcast", 0); + } else if (gpr_tls_get(&g_current_thread_worker) != + (intptr_t)specific_worker) { + GPR_TIMER_MARK("different_thread_worker", 0); + if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) { + specific_worker->reevaluate_polling_on_wakeup = 1; + } + specific_worker->kicked_specifically = 1; + grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); + } else if ((flags & GRPC_POLLSET_CAN_KICK_SELF) != 0) { + GPR_TIMER_MARK("kick_yoself", 0); + if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) { + specific_worker->reevaluate_polling_on_wakeup = 1; + } + specific_worker->kicked_specifically = 1; + grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); + } + } else if (gpr_tls_get(&g_current_thread_poller) != (intptr_t)p) { + GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0); + GPR_TIMER_MARK("kick_anonymous", 0); + specific_worker = pop_front_worker(p); + if (specific_worker != NULL) { + if (gpr_tls_get(&g_current_thread_worker) == (intptr_t)specific_worker) { + GPR_TIMER_MARK("kick_anonymous_not_self", 0); + push_back_worker(p, specific_worker); + specific_worker = pop_front_worker(p); + if ((flags & GRPC_POLLSET_CAN_KICK_SELF) == 0 && + gpr_tls_get(&g_current_thread_worker) == + (intptr_t)specific_worker) { + push_back_worker(p, specific_worker); + specific_worker = NULL; + } + } + if (specific_worker != NULL) { + GPR_TIMER_MARK("finally_kick", 0); + push_back_worker(p, specific_worker); + grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); + } + } else { + GPR_TIMER_MARK("kicked_no_pollers", 0); + p->kicked_without_pollers = 1; + } + } + + GPR_TIMER_END("pollset_kick_ext", 0); +} + +static void pollset_kick(grpc_pollset *p, + grpc_pollset_worker *specific_worker) { + pollset_kick_ext(p, specific_worker, 0); +} + +/* global state management */ + +static void pollset_global_init(void) { + gpr_tls_init(&g_current_thread_poller); + gpr_tls_init(&g_current_thread_worker); + grpc_wakeup_fd_global_init(); + grpc_wakeup_fd_init(&grpc_global_wakeup_fd); +} + +static void pollset_global_shutdown(void) { + grpc_wakeup_fd_destroy(&grpc_global_wakeup_fd); + gpr_tls_destroy(&g_current_thread_poller); + gpr_tls_destroy(&g_current_thread_worker); + grpc_wakeup_fd_global_destroy(); +} + +static void kick_poller(void) { grpc_wakeup_fd_wakeup(&grpc_global_wakeup_fd); } + +/* main interface */ + +static void become_basic_pollset(grpc_pollset *pollset, grpc_fd *fd_or_null); + +static void pollset_init(grpc_pollset *pollset, gpr_mu **mu) { + gpr_mu_init(&pollset->mu); + *mu = &pollset->mu; + pollset->root_worker.next = pollset->root_worker.prev = &pollset->root_worker; + pollset->in_flight_cbs = 0; + pollset->shutting_down = 0; + pollset->called_shutdown = 0; + pollset->kicked_without_pollers = 0; + pollset->idle_jobs.head = pollset->idle_jobs.tail = NULL; + pollset->local_wakeup_cache = NULL; + pollset->kicked_without_pollers = 0; + become_basic_pollset(pollset, NULL); +} + +static void pollset_destroy(grpc_pollset *pollset) { + GPR_ASSERT(pollset->in_flight_cbs == 0); + GPR_ASSERT(!pollset_has_workers(pollset)); + GPR_ASSERT(pollset->idle_jobs.head == pollset->idle_jobs.tail); + pollset->vtable->destroy(pollset); + while (pollset->local_wakeup_cache) { + grpc_cached_wakeup_fd *next = pollset->local_wakeup_cache->next; + grpc_wakeup_fd_destroy(&pollset->local_wakeup_cache->fd); + gpr_free(pollset->local_wakeup_cache); + pollset->local_wakeup_cache = next; + } + gpr_mu_destroy(&pollset->mu); +} + +static void pollset_reset(grpc_pollset *pollset) { + GPR_ASSERT(pollset->shutting_down); + GPR_ASSERT(pollset->in_flight_cbs == 0); + GPR_ASSERT(!pollset_has_workers(pollset)); + GPR_ASSERT(pollset->idle_jobs.head == pollset->idle_jobs.tail); + pollset->vtable->destroy(pollset); + pollset->shutting_down = 0; + pollset->called_shutdown = 0; + pollset->kicked_without_pollers = 0; + become_basic_pollset(pollset, NULL); +} + +static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_fd *fd) { + gpr_mu_lock(&pollset->mu); + pollset->vtable->add_fd(exec_ctx, pollset, fd, 1); +/* the following (enabled only in debug) will reacquire and then release + our lock - meaning that if the unlocking flag passed to add_fd above is + not respected, the code will deadlock (in a way that we have a chance of + debugging) */ +#ifndef NDEBUG + gpr_mu_lock(&pollset->mu); + gpr_mu_unlock(&pollset->mu); +#endif +} + +static void finish_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset) { + GPR_ASSERT(grpc_closure_list_empty(pollset->idle_jobs)); + pollset->vtable->finish_shutdown(pollset); + grpc_exec_ctx_enqueue(exec_ctx, pollset->shutdown_done, true, NULL); +} + +static void pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_pollset_worker **worker_hdl, gpr_timespec now, + gpr_timespec deadline) { + grpc_pollset_worker worker; + *worker_hdl = &worker; + + /* pollset->mu already held */ + int added_worker = 0; + int locked = 1; + int queued_work = 0; + int keep_polling = 0; + GPR_TIMER_BEGIN("pollset_work", 0); + /* this must happen before we (potentially) drop pollset->mu */ + worker.next = worker.prev = NULL; + worker.reevaluate_polling_on_wakeup = 0; + if (pollset->local_wakeup_cache != NULL) { + worker.wakeup_fd = pollset->local_wakeup_cache; + pollset->local_wakeup_cache = worker.wakeup_fd->next; + } else { + worker.wakeup_fd = gpr_malloc(sizeof(*worker.wakeup_fd)); + grpc_wakeup_fd_init(&worker.wakeup_fd->fd); + } + worker.kicked_specifically = 0; + /* If there's work waiting for the pollset to be idle, and the + pollset is idle, then do that work */ + if (!pollset_has_workers(pollset) && + !grpc_closure_list_empty(pollset->idle_jobs)) { + GPR_TIMER_MARK("pollset_work.idle_jobs", 0); + grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); + goto done; + } + /* If we're shutting down then we don't execute any extended work */ + if (pollset->shutting_down) { + GPR_TIMER_MARK("pollset_work.shutting_down", 0); + goto done; + } + /* Give do_promote priority so we don't starve it out */ + if (pollset->in_flight_cbs) { + GPR_TIMER_MARK("pollset_work.in_flight_cbs", 0); + gpr_mu_unlock(&pollset->mu); + locked = 0; + goto done; + } + /* Start polling, and keep doing so while we're being asked to + re-evaluate our pollers (this allows poll() based pollers to + ensure they don't miss wakeups) */ + keep_polling = 1; + while (keep_polling) { + keep_polling = 0; + if (!pollset->kicked_without_pollers) { + if (!added_worker) { + push_front_worker(pollset, &worker); + added_worker = 1; + gpr_tls_set(&g_current_thread_worker, (intptr_t)&worker); + } + gpr_tls_set(&g_current_thread_poller, (intptr_t)pollset); + GPR_TIMER_BEGIN("maybe_work_and_unlock", 0); + pollset->vtable->maybe_work_and_unlock(exec_ctx, pollset, &worker, + deadline, now); + GPR_TIMER_END("maybe_work_and_unlock", 0); + locked = 0; + gpr_tls_set(&g_current_thread_poller, 0); + } else { + GPR_TIMER_MARK("pollset_work.kicked_without_pollers", 0); + pollset->kicked_without_pollers = 0; + } + /* Finished execution - start cleaning up. + Note that we may arrive here from outside the enclosing while() loop. + In that case we won't loop though as we haven't added worker to the + worker list, which means nobody could ask us to re-evaluate polling). */ + done: + if (!locked) { + queued_work |= grpc_exec_ctx_flush(exec_ctx); + gpr_mu_lock(&pollset->mu); + locked = 1; + } + /* If we're forced to re-evaluate polling (via pollset_kick with + GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) then we land here and force + a loop */ + if (worker.reevaluate_polling_on_wakeup) { + worker.reevaluate_polling_on_wakeup = 0; + pollset->kicked_without_pollers = 0; + if (queued_work || worker.kicked_specifically) { + /* If there's queued work on the list, then set the deadline to be + immediate so we get back out of the polling loop quickly */ + deadline = gpr_inf_past(GPR_CLOCK_MONOTONIC); + } + keep_polling = 1; + } + } + if (added_worker) { + remove_worker(pollset, &worker); + gpr_tls_set(&g_current_thread_worker, 0); + } + /* release wakeup fd to the local pool */ + worker.wakeup_fd->next = pollset->local_wakeup_cache; + pollset->local_wakeup_cache = worker.wakeup_fd; + /* check shutdown conditions */ + if (pollset->shutting_down) { + if (pollset_has_workers(pollset)) { + pollset_kick(pollset, NULL); + } else if (!pollset->called_shutdown && pollset->in_flight_cbs == 0) { + pollset->called_shutdown = 1; + gpr_mu_unlock(&pollset->mu); + finish_shutdown(exec_ctx, pollset); + grpc_exec_ctx_flush(exec_ctx); + /* Continuing to access pollset here is safe -- it is the caller's + * responsibility to not destroy when it has outstanding calls to + * pollset_work. + * TODO(dklempner): Can we refactor the shutdown logic to avoid this? */ + gpr_mu_lock(&pollset->mu); + } else if (!grpc_closure_list_empty(pollset->idle_jobs)) { + grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); + gpr_mu_unlock(&pollset->mu); + grpc_exec_ctx_flush(exec_ctx); + gpr_mu_lock(&pollset->mu); + } + } + *worker_hdl = NULL; + GPR_TIMER_END("pollset_work", 0); +} + +static void pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_closure *closure) { + GPR_ASSERT(!pollset->shutting_down); + pollset->shutting_down = 1; + pollset->shutdown_done = closure; + pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); + if (!pollset_has_workers(pollset)) { + grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); + } + if (!pollset->called_shutdown && pollset->in_flight_cbs == 0 && + !pollset_has_workers(pollset)) { + pollset->called_shutdown = 1; + finish_shutdown(exec_ctx, pollset); + } +} + +static int poll_deadline_to_millis_timeout(gpr_timespec deadline, + gpr_timespec now) { + gpr_timespec timeout; + static const int64_t max_spin_polling_us = 10; + if (gpr_time_cmp(deadline, gpr_inf_future(deadline.clock_type)) == 0) { + return -1; + } + if (gpr_time_cmp(deadline, gpr_time_add(now, gpr_time_from_micros( + max_spin_polling_us, + GPR_TIMESPAN))) <= 0) { + return 0; + } + timeout = gpr_time_sub(deadline, now); + return gpr_time_to_millis(gpr_time_add( + timeout, gpr_time_from_nanos(GPR_NS_PER_MS - 1, GPR_TIMESPAN))); +} + +/* + * basic_pollset - a vtable that provides polling for zero or one file + * descriptor via poll() + */ + +typedef struct grpc_unary_promote_args { + const grpc_pollset_vtable *original_vtable; + grpc_pollset *pollset; + grpc_fd *fd; + grpc_closure promotion_closure; +} grpc_unary_promote_args; + +static void basic_do_promote(grpc_exec_ctx *exec_ctx, void *args, + bool success) { + grpc_unary_promote_args *up_args = args; + const grpc_pollset_vtable *original_vtable = up_args->original_vtable; + grpc_pollset *pollset = up_args->pollset; + grpc_fd *fd = up_args->fd; + + /* + * This is quite tricky. There are a number of cases to keep in mind here: + * 1. fd may have been orphaned + * 2. The pollset may no longer be a unary poller (and we can't let case #1 + * leak to other pollset types!) + * 3. pollset's fd (which may have changed) may have been orphaned + * 4. The pollset may be shutting down. + */ + + gpr_mu_lock(&pollset->mu); + /* First we need to ensure that nobody is polling concurrently */ + GPR_ASSERT(!pollset_has_workers(pollset)); + + gpr_free(up_args); + /* At this point the pollset may no longer be a unary poller. In that case + * we should just call the right add function and be done. */ + /* TODO(klempner): If we're not careful this could cause infinite recursion. + * That's not a problem for now because empty_pollset has a trivial poller + * and we don't have any mechanism to unbecome multipoller. */ + pollset->in_flight_cbs--; + if (pollset->shutting_down) { + /* We don't care about this pollset anymore. */ + if (pollset->in_flight_cbs == 0 && !pollset->called_shutdown) { + pollset->called_shutdown = 1; + finish_shutdown(exec_ctx, pollset); + } + } else if (fd_is_orphaned(fd)) { + /* Don't try to add it to anything, we'll drop our ref on it below */ + } else if (pollset->vtable != original_vtable) { + pollset->vtable->add_fd(exec_ctx, pollset, fd, 0); + } else if (fd != pollset->data.ptr) { + grpc_fd *fds[2]; + fds[0] = pollset->data.ptr; + fds[1] = fd; + + if (fds[0] && !fd_is_orphaned(fds[0])) { + platform_become_multipoller(exec_ctx, pollset, fds, GPR_ARRAY_SIZE(fds)); + GRPC_FD_UNREF(fds[0], "basicpoll"); + } else { + /* old fd is orphaned and we haven't cleaned it up until now, so remain a + * unary poller */ + /* Note that it is possible that fds[1] is also orphaned at this point. + * That's okay, we'll correct it at the next add or poll. */ + if (fds[0]) GRPC_FD_UNREF(fds[0], "basicpoll"); + pollset->data.ptr = fd; + GRPC_FD_REF(fd, "basicpoll"); + } + } + + gpr_mu_unlock(&pollset->mu); + + /* Matching ref in basic_pollset_add_fd */ + GRPC_FD_UNREF(fd, "basicpoll_add"); +} + +static void basic_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_fd *fd, int and_unlock_pollset) { + grpc_unary_promote_args *up_args; + GPR_ASSERT(fd); + if (fd == pollset->data.ptr) goto exit; + + if (!pollset_has_workers(pollset)) { + /* Fast path -- no in flight cbs */ + /* TODO(klempner): Comment this out and fix any test failures or establish + * they are due to timing issues */ + grpc_fd *fds[2]; + fds[0] = pollset->data.ptr; + fds[1] = fd; + + if (fds[0] == NULL) { + pollset->data.ptr = fd; + GRPC_FD_REF(fd, "basicpoll"); + } else if (!fd_is_orphaned(fds[0])) { + platform_become_multipoller(exec_ctx, pollset, fds, GPR_ARRAY_SIZE(fds)); + GRPC_FD_UNREF(fds[0], "basicpoll"); + } else { + /* old fd is orphaned and we haven't cleaned it up until now, so remain a + * unary poller */ + GRPC_FD_UNREF(fds[0], "basicpoll"); + pollset->data.ptr = fd; + GRPC_FD_REF(fd, "basicpoll"); + } + goto exit; + } + + /* Now we need to promote. This needs to happen when we're not polling. Since + * this may be called from poll, the wait needs to happen asynchronously. */ + GRPC_FD_REF(fd, "basicpoll_add"); + pollset->in_flight_cbs++; + up_args = gpr_malloc(sizeof(*up_args)); + up_args->fd = fd; + up_args->original_vtable = pollset->vtable; + up_args->pollset = pollset; + up_args->promotion_closure.cb = basic_do_promote; + up_args->promotion_closure.cb_arg = up_args; + + grpc_closure_list_add(&pollset->idle_jobs, &up_args->promotion_closure, 1); + pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); + +exit: + if (and_unlock_pollset) { + gpr_mu_unlock(&pollset->mu); + } +} + +static void basic_pollset_maybe_work_and_unlock(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, + grpc_pollset_worker *worker, + gpr_timespec deadline, + gpr_timespec now) { +#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR) +#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR) + + struct pollfd pfd[3]; + grpc_fd *fd; + grpc_fd_watcher fd_watcher; + int timeout; + int r; + nfds_t nfds; + + fd = pollset->data.ptr; + if (fd && fd_is_orphaned(fd)) { + GRPC_FD_UNREF(fd, "basicpoll"); + fd = pollset->data.ptr = NULL; + } + timeout = poll_deadline_to_millis_timeout(deadline, now); + pfd[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd); + pfd[0].events = POLLIN; + pfd[0].revents = 0; + pfd[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); + pfd[1].events = POLLIN; + pfd[1].revents = 0; + nfds = 2; + if (fd) { + pfd[2].fd = fd->fd; + pfd[2].revents = 0; + GRPC_FD_REF(fd, "basicpoll_begin"); + gpr_mu_unlock(&pollset->mu); + pfd[2].events = + (short)fd_begin_poll(fd, pollset, worker, POLLIN, POLLOUT, &fd_watcher); + if (pfd[2].events != 0) { + nfds++; + } + } else { + gpr_mu_unlock(&pollset->mu); + } + + /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid + even going into the blocking annotation if possible */ + /* poll fd count (argument 2) is shortened by one if we have no events + to poll on - such that it only includes the kicker */ + GPR_TIMER_BEGIN("poll", 0); + GRPC_SCHEDULING_START_BLOCKING_REGION; + r = grpc_poll_function(pfd, nfds, timeout); + GRPC_SCHEDULING_END_BLOCKING_REGION; + GPR_TIMER_END("poll", 0); + + if (r < 0) { + if (errno != EINTR) { + gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno)); + } + if (fd) { + fd_end_poll(exec_ctx, &fd_watcher, 0, 0); + } + } else if (r == 0) { + if (fd) { + fd_end_poll(exec_ctx, &fd_watcher, 0, 0); + } + } else { + if (pfd[0].revents & POLLIN_CHECK) { + grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd); + } + if (pfd[1].revents & POLLIN_CHECK) { + grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd); + } + if (nfds > 2) { + fd_end_poll(exec_ctx, &fd_watcher, pfd[2].revents & POLLIN_CHECK, + pfd[2].revents & POLLOUT_CHECK); + } else if (fd) { + fd_end_poll(exec_ctx, &fd_watcher, 0, 0); + } + } + + if (fd) { + GRPC_FD_UNREF(fd, "basicpoll_begin"); + } +} + +static void basic_pollset_destroy(grpc_pollset *pollset) { + if (pollset->data.ptr != NULL) { + GRPC_FD_UNREF(pollset->data.ptr, "basicpoll"); + pollset->data.ptr = NULL; + } +} + +static const grpc_pollset_vtable basic_pollset = { + basic_pollset_add_fd, basic_pollset_maybe_work_and_unlock, + basic_pollset_destroy, basic_pollset_destroy}; + +static void become_basic_pollset(grpc_pollset *pollset, grpc_fd *fd_or_null) { + pollset->vtable = &basic_pollset; + pollset->data.ptr = fd_or_null; + if (fd_or_null != NULL) { + GRPC_FD_REF(fd_or_null, "basicpoll"); + } +} + +/******************************************************************************* + * pollset_multipoller_with_poll_posix.c + */ + +#ifndef GPR_LINUX_MULTIPOLL_WITH_EPOLL + +typedef struct { + /* all polled fds */ + size_t fd_count; + size_t fd_capacity; + grpc_fd **fds; + /* fds that have been removed from the pollset explicitly */ + size_t del_count; + size_t del_capacity; + grpc_fd **dels; +} poll_hdr; + +static void multipoll_with_poll_pollset_add_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, + grpc_fd *fd, + int and_unlock_pollset) { + size_t i; + poll_hdr *h = pollset->data.ptr; + /* TODO(ctiller): this is O(num_fds^2); maybe switch to a hash set here */ + for (i = 0; i < h->fd_count; i++) { + if (h->fds[i] == fd) goto exit; + } + if (h->fd_count == h->fd_capacity) { + h->fd_capacity = GPR_MAX(h->fd_capacity + 8, h->fd_count * 3 / 2); + h->fds = gpr_realloc(h->fds, sizeof(grpc_fd *) * h->fd_capacity); + } + h->fds[h->fd_count++] = fd; + GRPC_FD_REF(fd, "multipoller"); +exit: + if (and_unlock_pollset) { + gpr_mu_unlock(&pollset->mu); + } +} + +static void multipoll_with_poll_pollset_maybe_work_and_unlock( + grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker, + gpr_timespec deadline, gpr_timespec now) { +#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR) +#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR) + + int timeout; + int r; + size_t i, j, fd_count; + nfds_t pfd_count; + poll_hdr *h; + /* TODO(ctiller): inline some elements to avoid an allocation */ + grpc_fd_watcher *watchers; + struct pollfd *pfds; + + h = pollset->data.ptr; + timeout = poll_deadline_to_millis_timeout(deadline, now); + /* TODO(ctiller): perform just one malloc here if we exceed the inline case */ + pfds = gpr_malloc(sizeof(*pfds) * (h->fd_count + 2)); + watchers = gpr_malloc(sizeof(*watchers) * (h->fd_count + 2)); + fd_count = 0; + pfd_count = 2; + pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd); + pfds[0].events = POLLIN; + pfds[0].revents = 0; + pfds[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); + pfds[1].events = POLLIN; + pfds[1].revents = 0; + for (i = 0; i < h->fd_count; i++) { + int remove = fd_is_orphaned(h->fds[i]); + for (j = 0; !remove && j < h->del_count; j++) { + if (h->fds[i] == h->dels[j]) remove = 1; + } + if (remove) { + GRPC_FD_UNREF(h->fds[i], "multipoller"); + } else { + h->fds[fd_count++] = h->fds[i]; + watchers[pfd_count].fd = h->fds[i]; + GRPC_FD_REF(watchers[pfd_count].fd, "multipoller_start"); + pfds[pfd_count].fd = h->fds[i]->fd; + pfds[pfd_count].revents = 0; + pfd_count++; + } + } + for (j = 0; j < h->del_count; j++) { + GRPC_FD_UNREF(h->dels[j], "multipoller_del"); + } + h->del_count = 0; + h->fd_count = fd_count; + gpr_mu_unlock(&pollset->mu); + + for (i = 2; i < pfd_count; i++) { + grpc_fd *fd = watchers[i].fd; + pfds[i].events = (short)fd_begin_poll(fd, pollset, worker, POLLIN, POLLOUT, + &watchers[i]); + GRPC_FD_UNREF(fd, "multipoller_start"); + } + + /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid + even going into the blocking annotation if possible */ + GRPC_SCHEDULING_START_BLOCKING_REGION; + r = grpc_poll_function(pfds, pfd_count, timeout); + GRPC_SCHEDULING_END_BLOCKING_REGION; + + if (r < 0) { + if (errno != EINTR) { + gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno)); + } + for (i = 2; i < pfd_count; i++) { + fd_end_poll(exec_ctx, &watchers[i], 0, 0); + } + } else if (r == 0) { + for (i = 2; i < pfd_count; i++) { + fd_end_poll(exec_ctx, &watchers[i], 0, 0); + } + } else { + if (pfds[0].revents & POLLIN_CHECK) { + grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd); + } + if (pfds[1].revents & POLLIN_CHECK) { + grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd); + } + for (i = 2; i < pfd_count; i++) { + if (watchers[i].fd == NULL) { + fd_end_poll(exec_ctx, &watchers[i], 0, 0); + continue; + } + fd_end_poll(exec_ctx, &watchers[i], pfds[i].revents & POLLIN_CHECK, + pfds[i].revents & POLLOUT_CHECK); + } + } + + gpr_free(pfds); + gpr_free(watchers); +} + +static void multipoll_with_poll_pollset_finish_shutdown(grpc_pollset *pollset) { + size_t i; + poll_hdr *h = pollset->data.ptr; + for (i = 0; i < h->fd_count; i++) { + GRPC_FD_UNREF(h->fds[i], "multipoller"); + } + for (i = 0; i < h->del_count; i++) { + GRPC_FD_UNREF(h->dels[i], "multipoller_del"); + } + h->fd_count = 0; + h->del_count = 0; +} + +static void multipoll_with_poll_pollset_destroy(grpc_pollset *pollset) { + poll_hdr *h = pollset->data.ptr; + multipoll_with_poll_pollset_finish_shutdown(pollset); + gpr_free(h->fds); + gpr_free(h->dels); + gpr_free(h); +} + +static const grpc_pollset_vtable multipoll_with_poll_pollset = { + multipoll_with_poll_pollset_add_fd, + multipoll_with_poll_pollset_maybe_work_and_unlock, + multipoll_with_poll_pollset_finish_shutdown, + multipoll_with_poll_pollset_destroy}; + +static void poll_become_multipoller(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, grpc_fd **fds, + size_t nfds) { + size_t i; + poll_hdr *h = gpr_malloc(sizeof(poll_hdr)); + pollset->vtable = &multipoll_with_poll_pollset; + pollset->data.ptr = h; + h->fd_count = nfds; + h->fd_capacity = nfds; + h->fds = gpr_malloc(nfds * sizeof(grpc_fd *)); + h->del_count = 0; + h->del_capacity = 0; + h->dels = NULL; + for (i = 0; i < nfds; i++) { + h->fds[i] = fds[i]; + GRPC_FD_REF(fds[i], "multipoller"); + } +} + +#endif /* !GPR_LINUX_MULTIPOLL_WITH_EPOLL */ + +/******************************************************************************* + * pollset_multipoller_with_epoll_posix.c + */ + +#ifdef GPR_LINUX_MULTIPOLL_WITH_EPOLL + +#include <errno.h> +#include <poll.h> +#include <string.h> +#include <sys/epoll.h> +#include <unistd.h> + +#include <grpc/support/alloc.h> +#include <grpc/support/log.h> +#include <grpc/support/useful.h> + +#include "src/core/lib/iomgr/ev_posix.h" +#include "src/core/lib/profiling/timers.h" +#include "src/core/lib/support/block_annotate.h" + +static void set_ready(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure **st) { + /* only one set_ready can be active at once (but there may be a racing + notify_on) */ + gpr_mu_lock(&fd->mu); + set_ready_locked(exec_ctx, fd, st); + gpr_mu_unlock(&fd->mu); +} + +static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + set_ready(exec_ctx, fd, &fd->read_closure); +} + +static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + set_ready(exec_ctx, fd, &fd->write_closure); +} + +struct epoll_fd_list { + int *epoll_fds; + size_t count; + size_t capacity; +}; + +static struct epoll_fd_list epoll_fd_global_list; +static gpr_once init_epoll_fd_list_mu = GPR_ONCE_INIT; +static gpr_mu epoll_fd_list_mu; + +static void init_mu(void) { gpr_mu_init(&epoll_fd_list_mu); } + +static void add_epoll_fd_to_global_list(int epoll_fd) { + gpr_once_init(&init_epoll_fd_list_mu, init_mu); + + gpr_mu_lock(&epoll_fd_list_mu); + if (epoll_fd_global_list.count == epoll_fd_global_list.capacity) { + epoll_fd_global_list.capacity = + GPR_MAX((size_t)8, epoll_fd_global_list.capacity * 2); + epoll_fd_global_list.epoll_fds = + gpr_realloc(epoll_fd_global_list.epoll_fds, + epoll_fd_global_list.capacity * sizeof(int)); + } + epoll_fd_global_list.epoll_fds[epoll_fd_global_list.count++] = epoll_fd; + gpr_mu_unlock(&epoll_fd_list_mu); +} + +static void remove_epoll_fd_from_global_list(int epoll_fd) { + gpr_mu_lock(&epoll_fd_list_mu); + GPR_ASSERT(epoll_fd_global_list.count > 0); + for (size_t i = 0; i < epoll_fd_global_list.count; i++) { + if (epoll_fd == epoll_fd_global_list.epoll_fds[i]) { + epoll_fd_global_list.epoll_fds[i] = + epoll_fd_global_list.epoll_fds[--(epoll_fd_global_list.count)]; + break; + } + } + gpr_mu_unlock(&epoll_fd_list_mu); +} + +static void remove_fd_from_all_epoll_sets(int fd) { + int err; + gpr_once_init(&init_epoll_fd_list_mu, init_mu); + gpr_mu_lock(&epoll_fd_list_mu); + if (epoll_fd_global_list.count == 0) { + gpr_mu_unlock(&epoll_fd_list_mu); + return; + } + for (size_t i = 0; i < epoll_fd_global_list.count; i++) { + err = epoll_ctl(epoll_fd_global_list.epoll_fds[i], EPOLL_CTL_DEL, fd, NULL); + if (err < 0 && errno != ENOENT) { + gpr_log(GPR_ERROR, "epoll_ctl del for %d failed: %s", fd, + strerror(errno)); + } + } + gpr_mu_unlock(&epoll_fd_list_mu); +} + +typedef struct { + grpc_pollset *pollset; + grpc_fd *fd; + grpc_closure closure; +} delayed_add; + +typedef struct { int epoll_fd; } epoll_hdr; + +static void finally_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_fd *fd) { + epoll_hdr *h = pollset->data.ptr; + struct epoll_event ev; + int err; + grpc_fd_watcher watcher; + + /* We pretend to be polling whilst adding an fd to keep the fd from being + closed during the add. This may result in a spurious wakeup being assigned + to this pollset whilst adding, but that should be benign. */ + GPR_ASSERT(fd_begin_poll(fd, pollset, NULL, 0, 0, &watcher) == 0); + if (watcher.fd != NULL) { + ev.events = (uint32_t)(EPOLLIN | EPOLLOUT | EPOLLET); + ev.data.ptr = fd; + err = epoll_ctl(h->epoll_fd, EPOLL_CTL_ADD, fd->fd, &ev); + if (err < 0) { + /* FDs may be added to a pollset multiple times, so EEXIST is normal. */ + if (errno != EEXIST) { + gpr_log(GPR_ERROR, "epoll_ctl add for %d failed: %s", fd->fd, + strerror(errno)); + } + } + } + fd_end_poll(exec_ctx, &watcher, 0, 0); +} + +static void perform_delayed_add(grpc_exec_ctx *exec_ctx, void *arg, + bool iomgr_status) { + delayed_add *da = arg; + + if (!fd_is_orphaned(da->fd)) { + finally_add_fd(exec_ctx, da->pollset, da->fd); + } + + gpr_mu_lock(&da->pollset->mu); + da->pollset->in_flight_cbs--; + if (da->pollset->shutting_down) { + /* We don't care about this pollset anymore. */ + if (da->pollset->in_flight_cbs == 0 && !da->pollset->called_shutdown) { + da->pollset->called_shutdown = 1; + grpc_exec_ctx_enqueue(exec_ctx, da->pollset->shutdown_done, true, NULL); + } + } + gpr_mu_unlock(&da->pollset->mu); + + GRPC_FD_UNREF(da->fd, "delayed_add"); + + gpr_free(da); +} + +static void multipoll_with_epoll_pollset_add_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, + grpc_fd *fd, + int and_unlock_pollset) { + if (and_unlock_pollset) { + gpr_mu_unlock(&pollset->mu); + finally_add_fd(exec_ctx, pollset, fd); + } else { + delayed_add *da = gpr_malloc(sizeof(*da)); + da->pollset = pollset; + da->fd = fd; + GRPC_FD_REF(fd, "delayed_add"); + grpc_closure_init(&da->closure, perform_delayed_add, da); + pollset->in_flight_cbs++; + grpc_exec_ctx_enqueue(exec_ctx, &da->closure, true, NULL); + } +} + +/* TODO(klempner): We probably want to turn this down a bit */ +#define GRPC_EPOLL_MAX_EVENTS 1000 + +static void multipoll_with_epoll_pollset_maybe_work_and_unlock( + grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker, + gpr_timespec deadline, gpr_timespec now) { + struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS]; + int ep_rv; + int poll_rv; + epoll_hdr *h = pollset->data.ptr; + int timeout_ms; + struct pollfd pfds[2]; + + /* If you want to ignore epoll's ability to sanely handle parallel pollers, + * for a more apples-to-apples performance comparison with poll, add a + * if (pollset->counter != 0) { return 0; } + * here. + */ + + gpr_mu_unlock(&pollset->mu); + + timeout_ms = poll_deadline_to_millis_timeout(deadline, now); + + pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); + pfds[0].events = POLLIN; + pfds[0].revents = 0; + pfds[1].fd = h->epoll_fd; + pfds[1].events = POLLIN; + pfds[1].revents = 0; + + /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid + even going into the blocking annotation if possible */ + GPR_TIMER_BEGIN("poll", 0); + GRPC_SCHEDULING_START_BLOCKING_REGION; + poll_rv = grpc_poll_function(pfds, 2, timeout_ms); + GRPC_SCHEDULING_END_BLOCKING_REGION; + GPR_TIMER_END("poll", 0); + + if (poll_rv < 0) { + if (errno != EINTR) { + gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno)); + } + } else if (poll_rv == 0) { + /* do nothing */ + } else { + if (pfds[0].revents) { + grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd); + } + if (pfds[1].revents) { + do { + /* The following epoll_wait never blocks; it has a timeout of 0 */ + ep_rv = epoll_wait(h->epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, 0); + if (ep_rv < 0) { + if (errno != EINTR) { + gpr_log(GPR_ERROR, "epoll_wait() failed: %s", strerror(errno)); + } + } else { + int i; + for (i = 0; i < ep_rv; ++i) { + grpc_fd *fd = ep_ev[i].data.ptr; + /* TODO(klempner): We might want to consider making err and pri + * separate events */ + int cancel = ep_ev[i].events & (EPOLLERR | EPOLLHUP); + int read_ev = ep_ev[i].events & (EPOLLIN | EPOLLPRI); + int write_ev = ep_ev[i].events & EPOLLOUT; + if (fd == NULL) { + grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd); + } else { + if (read_ev || cancel) { + fd_become_readable(exec_ctx, fd); + } + if (write_ev || cancel) { + fd_become_writable(exec_ctx, fd); + } + } + } + } + } while (ep_rv == GRPC_EPOLL_MAX_EVENTS); + } + } +} + +static void multipoll_with_epoll_pollset_finish_shutdown( + grpc_pollset *pollset) {} + +static void multipoll_with_epoll_pollset_destroy(grpc_pollset *pollset) { + epoll_hdr *h = pollset->data.ptr; + close(h->epoll_fd); + remove_epoll_fd_from_global_list(h->epoll_fd); + gpr_free(h); +} + +static const grpc_pollset_vtable multipoll_with_epoll_pollset = { + multipoll_with_epoll_pollset_add_fd, + multipoll_with_epoll_pollset_maybe_work_and_unlock, + multipoll_with_epoll_pollset_finish_shutdown, + multipoll_with_epoll_pollset_destroy}; + +static void epoll_become_multipoller(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, grpc_fd **fds, + size_t nfds) { + size_t i; + epoll_hdr *h = gpr_malloc(sizeof(epoll_hdr)); + struct epoll_event ev; + int err; + + pollset->vtable = &multipoll_with_epoll_pollset; + pollset->data.ptr = h; + h->epoll_fd = epoll_create1(EPOLL_CLOEXEC); + if (h->epoll_fd < 0) { + /* TODO(klempner): Fall back to poll here, especially on ENOSYS */ + gpr_log(GPR_ERROR, "epoll_create1 failed: %s", strerror(errno)); + abort(); + } + add_epoll_fd_to_global_list(h->epoll_fd); + + ev.events = (uint32_t)(EPOLLIN | EPOLLET); + ev.data.ptr = NULL; + err = epoll_ctl(h->epoll_fd, EPOLL_CTL_ADD, + GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), &ev); + if (err < 0) { + gpr_log(GPR_ERROR, "epoll_ctl add for %d failed: %s", + GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), + strerror(errno)); + } + + for (i = 0; i < nfds; i++) { + multipoll_with_epoll_pollset_add_fd(exec_ctx, pollset, fds[i], 0); + } +} + +#else /* GPR_LINUX_MULTIPOLL_WITH_EPOLL */ + +static void remove_fd_from_all_epoll_sets(int fd) {} + +#endif /* GPR_LINUX_MULTIPOLL_WITH_EPOLL */ + +/******************************************************************************* + * pollset_set_posix.c + */ + +static grpc_pollset_set *pollset_set_create(void) { + grpc_pollset_set *pollset_set = gpr_malloc(sizeof(*pollset_set)); + memset(pollset_set, 0, sizeof(*pollset_set)); + gpr_mu_init(&pollset_set->mu); + return pollset_set; +} + +static void pollset_set_destroy(grpc_pollset_set *pollset_set) { + size_t i; + gpr_mu_destroy(&pollset_set->mu); + for (i = 0; i < pollset_set->fd_count; i++) { + GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set"); + } + gpr_free(pollset_set->pollsets); + gpr_free(pollset_set->pollset_sets); + gpr_free(pollset_set->fds); + gpr_free(pollset_set); +} + +static void pollset_set_add_pollset(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, + grpc_pollset *pollset) { + size_t i, j; + gpr_mu_lock(&pollset_set->mu); + if (pollset_set->pollset_count == pollset_set->pollset_capacity) { + pollset_set->pollset_capacity = + GPR_MAX(8, 2 * pollset_set->pollset_capacity); + pollset_set->pollsets = + gpr_realloc(pollset_set->pollsets, pollset_set->pollset_capacity * + sizeof(*pollset_set->pollsets)); + } + pollset_set->pollsets[pollset_set->pollset_count++] = pollset; + for (i = 0, j = 0; i < pollset_set->fd_count; i++) { + if (fd_is_orphaned(pollset_set->fds[i])) { + GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set"); + } else { + pollset_add_fd(exec_ctx, pollset, pollset_set->fds[i]); + pollset_set->fds[j++] = pollset_set->fds[i]; + } + } + pollset_set->fd_count = j; + gpr_mu_unlock(&pollset_set->mu); +} + +static void pollset_set_del_pollset(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, + grpc_pollset *pollset) { + size_t i; + gpr_mu_lock(&pollset_set->mu); + for (i = 0; i < pollset_set->pollset_count; i++) { + if (pollset_set->pollsets[i] == pollset) { + pollset_set->pollset_count--; + GPR_SWAP(grpc_pollset *, pollset_set->pollsets[i], + pollset_set->pollsets[pollset_set->pollset_count]); + break; + } + } + gpr_mu_unlock(&pollset_set->mu); +} + +static void pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item) { + size_t i, j; + gpr_mu_lock(&bag->mu); + if (bag->pollset_set_count == bag->pollset_set_capacity) { + bag->pollset_set_capacity = GPR_MAX(8, 2 * bag->pollset_set_capacity); + bag->pollset_sets = + gpr_realloc(bag->pollset_sets, + bag->pollset_set_capacity * sizeof(*bag->pollset_sets)); + } + bag->pollset_sets[bag->pollset_set_count++] = item; + for (i = 0, j = 0; i < bag->fd_count; i++) { + if (fd_is_orphaned(bag->fds[i])) { + GRPC_FD_UNREF(bag->fds[i], "pollset_set"); + } else { + pollset_set_add_fd(exec_ctx, item, bag->fds[i]); + bag->fds[j++] = bag->fds[i]; + } + } + bag->fd_count = j; + gpr_mu_unlock(&bag->mu); +} + +static void pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item) { + size_t i; + gpr_mu_lock(&bag->mu); + for (i = 0; i < bag->pollset_set_count; i++) { + if (bag->pollset_sets[i] == item) { + bag->pollset_set_count--; + GPR_SWAP(grpc_pollset_set *, bag->pollset_sets[i], + bag->pollset_sets[bag->pollset_set_count]); + break; + } + } + gpr_mu_unlock(&bag->mu); +} + +static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd) { + size_t i; + gpr_mu_lock(&pollset_set->mu); + if (pollset_set->fd_count == pollset_set->fd_capacity) { + pollset_set->fd_capacity = GPR_MAX(8, 2 * pollset_set->fd_capacity); + pollset_set->fds = gpr_realloc( + pollset_set->fds, pollset_set->fd_capacity * sizeof(*pollset_set->fds)); + } + GRPC_FD_REF(fd, "pollset_set"); + pollset_set->fds[pollset_set->fd_count++] = fd; + for (i = 0; i < pollset_set->pollset_count; i++) { + pollset_add_fd(exec_ctx, pollset_set->pollsets[i], fd); + } + for (i = 0; i < pollset_set->pollset_set_count; i++) { + pollset_set_add_fd(exec_ctx, pollset_set->pollset_sets[i], fd); + } + gpr_mu_unlock(&pollset_set->mu); +} + +static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd) { + size_t i; + gpr_mu_lock(&pollset_set->mu); + for (i = 0; i < pollset_set->fd_count; i++) { + if (pollset_set->fds[i] == fd) { + pollset_set->fd_count--; + GPR_SWAP(grpc_fd *, pollset_set->fds[i], + pollset_set->fds[pollset_set->fd_count]); + GRPC_FD_UNREF(fd, "pollset_set"); + break; + } + } + for (i = 0; i < pollset_set->pollset_set_count; i++) { + pollset_set_del_fd(exec_ctx, pollset_set->pollset_sets[i], fd); + } + gpr_mu_unlock(&pollset_set->mu); +} + +/******************************************************************************* + * event engine binding + */ + +static void shutdown_engine(void) { + fd_global_shutdown(); + pollset_global_shutdown(); +} + +static const grpc_event_engine_vtable vtable = { + .pollset_size = sizeof(grpc_pollset), + + .fd_create = fd_create, + .fd_wrapped_fd = fd_wrapped_fd, + .fd_orphan = fd_orphan, + .fd_shutdown = fd_shutdown, + .fd_notify_on_read = fd_notify_on_read, + .fd_notify_on_write = fd_notify_on_write, + + .pollset_init = pollset_init, + .pollset_shutdown = pollset_shutdown, + .pollset_reset = pollset_reset, + .pollset_destroy = pollset_destroy, + .pollset_work = pollset_work, + .pollset_kick = pollset_kick, + .pollset_add_fd = pollset_add_fd, + + .pollset_set_create = pollset_set_create, + .pollset_set_destroy = pollset_set_destroy, + .pollset_set_add_pollset = pollset_set_add_pollset, + .pollset_set_del_pollset = pollset_set_del_pollset, + .pollset_set_add_pollset_set = pollset_set_add_pollset_set, + .pollset_set_del_pollset_set = pollset_set_del_pollset_set, + .pollset_set_add_fd = pollset_set_add_fd, + .pollset_set_del_fd = pollset_set_del_fd, + + .kick_poller = kick_poller, + + .shutdown_engine = shutdown_engine, +}; + +const grpc_event_engine_vtable *grpc_init_poll_and_epoll_posix(void) { +#ifdef GPR_LINUX_MULTIPOLL_WITH_EPOLL + platform_become_multipoller = epoll_become_multipoller; +#else + platform_become_multipoller = poll_become_multipoller; +#endif + fd_global_init(); + pollset_global_init(); + return &vtable; +} + +#endif diff --git a/src/core/lib/census/grpc_plugin.h b/src/core/lib/iomgr/ev_poll_and_epoll_posix.h index 4d642a65c7..06d6dbf29d 100644 --- a/src/core/lib/census/grpc_plugin.h +++ b/src/core/lib/iomgr/ev_poll_and_epoll_posix.h @@ -31,10 +31,11 @@ * */ -#ifndef GRPC_CORE_LIB_CENSUS_GRPC_PLUGIN_H -#define GRPC_CORE_LIB_CENSUS_GRPC_PLUGIN_H +#ifndef GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H +#define GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H -void census_grpc_plugin_init(void); -void census_grpc_plugin_destroy(void); +#include "src/core/lib/iomgr/ev_posix.h" -#endif /* GRPC_CORE_LIB_CENSUS_GRPC_PLUGIN_H */ +const grpc_event_engine_vtable *grpc_init_poll_and_epoll_posix(void); + +#endif /* GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H */ diff --git a/src/core/lib/iomgr/ev_posix.c b/src/core/lib/iomgr/ev_posix.c new file mode 100644 index 0000000000..235a7df08d --- /dev/null +++ b/src/core/lib/iomgr/ev_posix.c @@ -0,0 +1,165 @@ +/* + * + * Copyright 2015-2016, Google Inc. + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are + * met: + * + * * Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * * Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following disclaimer + * in the documentation and/or other materials provided with the + * distribution. + * * Neither the name of Google Inc. nor the names of its + * contributors may be used to endorse or promote products derived from + * this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + */ + +#include <grpc/support/port_platform.h> + +#ifdef GPR_POSIX_SOCKET + +#include "src/core/lib/iomgr/ev_posix.h" + +#include <grpc/support/log.h> + +#include "src/core/lib/iomgr/ev_poll_and_epoll_posix.h" + +static const grpc_event_engine_vtable *g_event_engine; + +grpc_poll_function_type grpc_poll_function = poll; +grpc_wakeup_fd grpc_global_wakeup_fd; + +void grpc_event_engine_init(void) { + if ((g_event_engine = grpc_init_poll_and_epoll_posix())) { + return; + } + gpr_log(GPR_ERROR, "No event engine could be initialized"); + abort(); +} + +void grpc_event_engine_shutdown(void) { g_event_engine->shutdown_engine(); } + +grpc_fd *grpc_fd_create(int fd, const char *name) { + return g_event_engine->fd_create(fd, name); +} + +int grpc_fd_wrapped_fd(grpc_fd *fd) { + return g_event_engine->fd_wrapped_fd(fd); +} + +void grpc_fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done, + int *release_fd, const char *reason) { + g_event_engine->fd_orphan(exec_ctx, fd, on_done, release_fd, reason); +} + +void grpc_fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + g_event_engine->fd_shutdown(exec_ctx, fd); +} + +void grpc_fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure) { + g_event_engine->fd_notify_on_read(exec_ctx, fd, closure); +} + +void grpc_fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure) { + g_event_engine->fd_notify_on_write(exec_ctx, fd, closure); +} + +size_t grpc_pollset_size(void) { return g_event_engine->pollset_size; } + +void grpc_pollset_init(grpc_pollset *pollset, gpr_mu **mu) { + g_event_engine->pollset_init(pollset, mu); +} + +void grpc_pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_closure *closure) { + g_event_engine->pollset_shutdown(exec_ctx, pollset, closure); +} + +void grpc_pollset_reset(grpc_pollset *pollset) { + g_event_engine->pollset_reset(pollset); +} + +void grpc_pollset_destroy(grpc_pollset *pollset) { + g_event_engine->pollset_destroy(pollset); +} + +void grpc_pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_pollset_worker **worker, gpr_timespec now, + gpr_timespec deadline) { + g_event_engine->pollset_work(exec_ctx, pollset, worker, now, deadline); +} + +void grpc_pollset_kick(grpc_pollset *pollset, + grpc_pollset_worker *specific_worker) { + g_event_engine->pollset_kick(pollset, specific_worker); +} + +void grpc_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + struct grpc_fd *fd) { + g_event_engine->pollset_add_fd(exec_ctx, pollset, fd); +} + +grpc_pollset_set *grpc_pollset_set_create(void) { + return g_event_engine->pollset_set_create(); +} + +void grpc_pollset_set_destroy(grpc_pollset_set *pollset_set) { + g_event_engine->pollset_set_destroy(pollset_set); +} + +void grpc_pollset_set_add_pollset(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, + grpc_pollset *pollset) { + g_event_engine->pollset_set_add_pollset(exec_ctx, pollset_set, pollset); +} + +void grpc_pollset_set_del_pollset(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, + grpc_pollset *pollset) { + g_event_engine->pollset_set_del_pollset(exec_ctx, pollset_set, pollset); +} + +void grpc_pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item) { + g_event_engine->pollset_set_add_pollset_set(exec_ctx, bag, item); +} + +void grpc_pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item) { + g_event_engine->pollset_set_del_pollset_set(exec_ctx, bag, item); +} + +void grpc_pollset_set_add_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd) { + g_event_engine->pollset_set_add_fd(exec_ctx, pollset_set, fd); +} + +void grpc_pollset_set_del_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd) { + g_event_engine->pollset_set_del_fd(exec_ctx, pollset_set, fd); +} + +void grpc_kick_poller(void) { g_event_engine->kick_poller(); } + +#endif // GPR_POSIX_SOCKET diff --git a/src/core/lib/iomgr/ev_posix.h b/src/core/lib/iomgr/ev_posix.h new file mode 100644 index 0000000000..9d27b2bcda --- /dev/null +++ b/src/core/lib/iomgr/ev_posix.h @@ -0,0 +1,158 @@ +/* + * + * Copyright 2015-2016, Google Inc. + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are + * met: + * + * * Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * * Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following disclaimer + * in the documentation and/or other materials provided with the + * distribution. + * * Neither the name of Google Inc. nor the names of its + * contributors may be used to endorse or promote products derived from + * this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + */ + +#ifndef GRPC_CORE_LIB_IOMGR_EV_POSIX_H +#define GRPC_CORE_LIB_IOMGR_EV_POSIX_H + +#include <poll.h> + +#include "src/core/lib/iomgr/exec_ctx.h" +#include "src/core/lib/iomgr/pollset.h" +#include "src/core/lib/iomgr/pollset_set.h" +#include "src/core/lib/iomgr/wakeup_fd_posix.h" + +typedef struct grpc_fd grpc_fd; + +typedef struct grpc_event_engine_vtable { + size_t pollset_size; + + grpc_fd *(*fd_create)(int fd, const char *name); + int (*fd_wrapped_fd)(grpc_fd *fd); + void (*fd_orphan)(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done, + int *release_fd, const char *reason); + void (*fd_shutdown)(grpc_exec_ctx *exec_ctx, grpc_fd *fd); + void (*fd_notify_on_read)(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure); + void (*fd_notify_on_write)(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure); + + void (*pollset_init)(grpc_pollset *pollset, gpr_mu **mu); + void (*pollset_shutdown)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_closure *closure); + void (*pollset_reset)(grpc_pollset *pollset); + void (*pollset_destroy)(grpc_pollset *pollset); + void (*pollset_work)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_pollset_worker **worker, gpr_timespec now, + gpr_timespec deadline); + void (*pollset_kick)(grpc_pollset *pollset, + grpc_pollset_worker *specific_worker); + void (*pollset_add_fd)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + struct grpc_fd *fd); + + grpc_pollset_set *(*pollset_set_create)(void); + void (*pollset_set_destroy)(grpc_pollset_set *pollset_set); + void (*pollset_set_add_pollset)(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, + grpc_pollset *pollset); + void (*pollset_set_del_pollset)(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, + grpc_pollset *pollset); + void (*pollset_set_add_pollset_set)(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item); + void (*pollset_set_del_pollset_set)(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item); + void (*pollset_set_add_fd)(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd); + void (*pollset_set_del_fd)(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd); + + void (*kick_poller)(void); + + void (*shutdown_engine)(void); +} grpc_event_engine_vtable; + +void grpc_event_engine_init(void); +void grpc_event_engine_shutdown(void); + +/* Create a wrapped file descriptor. + Requires fd is a non-blocking file descriptor. + This takes ownership of closing fd. */ +grpc_fd *grpc_fd_create(int fd, const char *name); + +/* Return the wrapped fd, or -1 if it has been released or closed. */ +int grpc_fd_wrapped_fd(grpc_fd *fd); + +/* Releases fd to be asynchronously destroyed. + on_done is called when the underlying file descriptor is definitely close()d. + If on_done is NULL, no callback will be made. + If release_fd is not NULL, it's set to fd and fd will not be closed. + Requires: *fd initialized; no outstanding notify_on_read or + notify_on_write. + MUST NOT be called with a pollset lock taken */ +void grpc_fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done, + int *release_fd, const char *reason); + +/* Cause any current callbacks to error out with GRPC_CALLBACK_CANCELLED. */ +void grpc_fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd); + +/* Register read interest, causing read_cb to be called once when fd becomes + readable, on deadline specified by deadline, or on shutdown triggered by + grpc_fd_shutdown. + read_cb will be called with read_cb_arg when *fd becomes readable. + read_cb is Called with status of GRPC_CALLBACK_SUCCESS if readable, + GRPC_CALLBACK_TIMED_OUT if the call timed out, + and CANCELLED if the call was cancelled. + + Requires:This method must not be called before the read_cb for any previous + call runs. Edge triggered events are used whenever they are supported by the + underlying platform. This means that users must drain fd in read_cb before + calling notify_on_read again. Users are also expected to handle spurious + events, i.e read_cb is called while nothing can be readable from fd */ +void grpc_fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure); + +/* Exactly the same semantics as above, except based on writable events. */ +void grpc_fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure); + +/* pollset_posix functions */ + +/* Add an fd to a pollset */ +void grpc_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + struct grpc_fd *fd); + +/* pollset_set_posix functions */ + +void grpc_pollset_set_add_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd); +void grpc_pollset_set_del_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd); + +/* override to allow tests to hook poll() usage */ +typedef int (*grpc_poll_function_type)(struct pollfd *, nfds_t, int); +extern grpc_poll_function_type grpc_poll_function; +extern grpc_wakeup_fd grpc_global_wakeup_fd; + +#endif /* GRPC_CORE_LIB_IOMGR_EV_POSIX_H */ diff --git a/src/core/lib/iomgr/fd_posix.c b/src/core/lib/iomgr/fd_posix.c deleted file mode 100644 index 1397302658..0000000000 --- a/src/core/lib/iomgr/fd_posix.c +++ /dev/null @@ -1,454 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/support/port_platform.h> - -#ifdef GPR_POSIX_SOCKET - -#include "src/core/lib/iomgr/fd_posix.h" - -#include <assert.h> -#include <sys/socket.h> -#include <unistd.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/string_util.h> -#include <grpc/support/useful.h> - -#include "src/core/lib/iomgr/pollset_posix.h" - -#define CLOSURE_NOT_READY ((grpc_closure *)0) -#define CLOSURE_READY ((grpc_closure *)1) - -/* We need to keep a freelist not because of any concerns of malloc performance - * but instead so that implementations with multiple threads in (for example) - * epoll_wait deal with the race between pollset removal and incoming poll - * notifications. - * - * The problem is that the poller ultimately holds a reference to this - * object, so it is very difficult to know when is safe to free it, at least - * without some expensive synchronization. - * - * If we keep the object freelisted, in the worst case losing this race just - * becomes a spurious read notification on a reused fd. - */ -/* TODO(klempner): We could use some form of polling generation count to know - * when these are safe to free. */ -/* TODO(klempner): Consider disabling freelisting if we don't have multiple - * threads in poll on the same fd */ -/* TODO(klempner): Batch these allocations to reduce fragmentation */ -static grpc_fd *fd_freelist = NULL; -static gpr_mu fd_freelist_mu; - -static void freelist_fd(grpc_fd *fd) { - gpr_mu_lock(&fd_freelist_mu); - fd->freelist_next = fd_freelist; - fd_freelist = fd; - grpc_iomgr_unregister_object(&fd->iomgr_object); - gpr_mu_unlock(&fd_freelist_mu); -} - -static grpc_fd *alloc_fd(int fd) { - grpc_fd *r = NULL; - gpr_mu_lock(&fd_freelist_mu); - if (fd_freelist != NULL) { - r = fd_freelist; - fd_freelist = fd_freelist->freelist_next; - } - gpr_mu_unlock(&fd_freelist_mu); - if (r == NULL) { - r = gpr_malloc(sizeof(grpc_fd)); - gpr_mu_init(&r->mu); - } - - gpr_mu_lock(&r->mu); - r->shutdown = 0; - r->read_closure = CLOSURE_NOT_READY; - r->write_closure = CLOSURE_NOT_READY; - r->fd = fd; - r->inactive_watcher_root.next = r->inactive_watcher_root.prev = - &r->inactive_watcher_root; - r->freelist_next = NULL; - r->read_watcher = r->write_watcher = NULL; - r->on_done_closure = NULL; - r->closed = 0; - r->released = 0; - gpr_atm_rel_store(&r->refst, 1); - gpr_mu_unlock(&r->mu); - - return r; -} - -static void destroy(grpc_fd *fd) { - gpr_mu_destroy(&fd->mu); - gpr_free(fd); -} - -#ifdef GRPC_FD_REF_COUNT_DEBUG -#define REF_BY(fd, n, reason) ref_by(fd, n, reason, __FILE__, __LINE__) -#define UNREF_BY(fd, n, reason) unref_by(fd, n, reason, __FILE__, __LINE__) -static void ref_by(grpc_fd *fd, int n, const char *reason, const char *file, - int line) { - gpr_log(GPR_DEBUG, "FD %d %p ref %d %d -> %d [%s; %s:%d]", fd->fd, fd, n, - gpr_atm_no_barrier_load(&fd->refst), - gpr_atm_no_barrier_load(&fd->refst) + n, reason, file, line); -#else -#define REF_BY(fd, n, reason) ref_by(fd, n) -#define UNREF_BY(fd, n, reason) unref_by(fd, n) -static void ref_by(grpc_fd *fd, int n) { -#endif - GPR_ASSERT(gpr_atm_no_barrier_fetch_add(&fd->refst, n) > 0); -} - -#ifdef GRPC_FD_REF_COUNT_DEBUG -static void unref_by(grpc_fd *fd, int n, const char *reason, const char *file, - int line) { - gpr_atm old; - gpr_log(GPR_DEBUG, "FD %d %p unref %d %d -> %d [%s; %s:%d]", fd->fd, fd, n, - gpr_atm_no_barrier_load(&fd->refst), - gpr_atm_no_barrier_load(&fd->refst) - n, reason, file, line); -#else -static void unref_by(grpc_fd *fd, int n) { - gpr_atm old; -#endif - old = gpr_atm_full_fetch_add(&fd->refst, -n); - if (old == n) { - freelist_fd(fd); - } else { - GPR_ASSERT(old > n); - } -} - -void grpc_fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); } - -void grpc_fd_global_shutdown(void) { - gpr_mu_lock(&fd_freelist_mu); - gpr_mu_unlock(&fd_freelist_mu); - while (fd_freelist != NULL) { - grpc_fd *fd = fd_freelist; - fd_freelist = fd_freelist->freelist_next; - destroy(fd); - } - gpr_mu_destroy(&fd_freelist_mu); -} - -grpc_fd *grpc_fd_create(int fd, const char *name) { - grpc_fd *r = alloc_fd(fd); - char *name2; - gpr_asprintf(&name2, "%s fd=%d", name, fd); - grpc_iomgr_register_object(&r->iomgr_object, name2); - gpr_free(name2); -#ifdef GRPC_FD_REF_COUNT_DEBUG - gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, r, name); -#endif - return r; -} - -int grpc_fd_is_orphaned(grpc_fd *fd) { - return (gpr_atm_acq_load(&fd->refst) & 1) == 0; -} - -static void pollset_kick_locked(grpc_fd_watcher *watcher) { - gpr_mu_lock(&watcher->pollset->mu); - GPR_ASSERT(watcher->worker); - grpc_pollset_kick_ext(watcher->pollset, watcher->worker, - GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP); - gpr_mu_unlock(&watcher->pollset->mu); -} - -static void maybe_wake_one_watcher_locked(grpc_fd *fd) { - if (fd->inactive_watcher_root.next != &fd->inactive_watcher_root) { - pollset_kick_locked(fd->inactive_watcher_root.next); - } else if (fd->read_watcher) { - pollset_kick_locked(fd->read_watcher); - } else if (fd->write_watcher) { - pollset_kick_locked(fd->write_watcher); - } -} - -static void wake_all_watchers_locked(grpc_fd *fd) { - grpc_fd_watcher *watcher; - for (watcher = fd->inactive_watcher_root.next; - watcher != &fd->inactive_watcher_root; watcher = watcher->next) { - pollset_kick_locked(watcher); - } - if (fd->read_watcher) { - pollset_kick_locked(fd->read_watcher); - } - if (fd->write_watcher && fd->write_watcher != fd->read_watcher) { - pollset_kick_locked(fd->write_watcher); - } -} - -static int has_watchers(grpc_fd *fd) { - return fd->read_watcher != NULL || fd->write_watcher != NULL || - fd->inactive_watcher_root.next != &fd->inactive_watcher_root; -} - -static void close_fd_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { - fd->closed = 1; - if (!fd->released) { - close(fd->fd); - } else { - grpc_remove_fd_from_all_epoll_sets(fd->fd); - } - grpc_exec_ctx_enqueue(exec_ctx, fd->on_done_closure, true, NULL); -} - -int grpc_fd_wrapped_fd(grpc_fd *fd) { - if (fd->released || fd->closed) { - return -1; - } else { - return fd->fd; - } -} - -void grpc_fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done, - int *release_fd, const char *reason) { - fd->on_done_closure = on_done; - fd->released = release_fd != NULL; - if (!fd->released) { - shutdown(fd->fd, SHUT_RDWR); - } else { - *release_fd = fd->fd; - } - gpr_mu_lock(&fd->mu); - REF_BY(fd, 1, reason); /* remove active status, but keep referenced */ - if (!has_watchers(fd)) { - close_fd_locked(exec_ctx, fd); - } else { - wake_all_watchers_locked(fd); - } - gpr_mu_unlock(&fd->mu); - UNREF_BY(fd, 2, reason); /* drop the reference */ -} - -/* increment refcount by two to avoid changing the orphan bit */ -#ifdef GRPC_FD_REF_COUNT_DEBUG -void grpc_fd_ref(grpc_fd *fd, const char *reason, const char *file, int line) { - ref_by(fd, 2, reason, file, line); -} - -void grpc_fd_unref(grpc_fd *fd, const char *reason, const char *file, - int line) { - unref_by(fd, 2, reason, file, line); -} -#else -void grpc_fd_ref(grpc_fd *fd) { ref_by(fd, 2); } - -void grpc_fd_unref(grpc_fd *fd) { unref_by(fd, 2); } -#endif - -static void notify_on_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure **st, grpc_closure *closure) { - if (*st == CLOSURE_NOT_READY) { - /* not ready ==> switch to a waiting state by setting the closure */ - *st = closure; - } else if (*st == CLOSURE_READY) { - /* already ready ==> queue the closure to run immediately */ - *st = CLOSURE_NOT_READY; - grpc_exec_ctx_enqueue(exec_ctx, closure, !fd->shutdown, NULL); - maybe_wake_one_watcher_locked(fd); - } else { - /* upcallptr was set to a different closure. This is an error! */ - gpr_log(GPR_ERROR, - "User called a notify_on function with a previous callback still " - "pending"); - abort(); - } -} - -/* returns 1 if state becomes not ready */ -static int set_ready_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure **st) { - if (*st == CLOSURE_READY) { - /* duplicate ready ==> ignore */ - return 0; - } else if (*st == CLOSURE_NOT_READY) { - /* not ready, and not waiting ==> flag ready */ - *st = CLOSURE_READY; - return 0; - } else { - /* waiting ==> queue closure */ - grpc_exec_ctx_enqueue(exec_ctx, *st, !fd->shutdown, NULL); - *st = CLOSURE_NOT_READY; - return 1; - } -} - -static void set_ready(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure **st) { - /* only one set_ready can be active at once (but there may be a racing - notify_on) */ - gpr_mu_lock(&fd->mu); - set_ready_locked(exec_ctx, fd, st); - gpr_mu_unlock(&fd->mu); -} - -void grpc_fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { - gpr_mu_lock(&fd->mu); - GPR_ASSERT(!fd->shutdown); - fd->shutdown = 1; - set_ready_locked(exec_ctx, fd, &fd->read_closure); - set_ready_locked(exec_ctx, fd, &fd->write_closure); - gpr_mu_unlock(&fd->mu); -} - -void grpc_fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure *closure) { - gpr_mu_lock(&fd->mu); - notify_on_locked(exec_ctx, fd, &fd->read_closure, closure); - gpr_mu_unlock(&fd->mu); -} - -void grpc_fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure *closure) { - gpr_mu_lock(&fd->mu); - notify_on_locked(exec_ctx, fd, &fd->write_closure, closure); - gpr_mu_unlock(&fd->mu); -} - -uint32_t grpc_fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset, - grpc_pollset_worker *worker, uint32_t read_mask, - uint32_t write_mask, grpc_fd_watcher *watcher) { - uint32_t mask = 0; - grpc_closure *cur; - int requested; - /* keep track of pollers that have requested our events, in case they change - */ - GRPC_FD_REF(fd, "poll"); - - gpr_mu_lock(&fd->mu); - - /* if we are shutdown, then don't add to the watcher set */ - if (fd->shutdown) { - watcher->fd = NULL; - watcher->pollset = NULL; - watcher->worker = NULL; - gpr_mu_unlock(&fd->mu); - GRPC_FD_UNREF(fd, "poll"); - return 0; - } - - /* if there is nobody polling for read, but we need to, then start doing so */ - cur = fd->read_closure; - requested = cur != CLOSURE_READY; - if (read_mask && fd->read_watcher == NULL && requested) { - fd->read_watcher = watcher; - mask |= read_mask; - } - /* if there is nobody polling for write, but we need to, then start doing so - */ - cur = fd->write_closure; - requested = cur != CLOSURE_READY; - if (write_mask && fd->write_watcher == NULL && requested) { - fd->write_watcher = watcher; - mask |= write_mask; - } - /* if not polling, remember this watcher in case we need someone to later */ - if (mask == 0 && worker != NULL) { - watcher->next = &fd->inactive_watcher_root; - watcher->prev = watcher->next->prev; - watcher->next->prev = watcher->prev->next = watcher; - } - watcher->pollset = pollset; - watcher->worker = worker; - watcher->fd = fd; - gpr_mu_unlock(&fd->mu); - - return mask; -} - -void grpc_fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *watcher, - int got_read, int got_write) { - int was_polling = 0; - int kick = 0; - grpc_fd *fd = watcher->fd; - - if (fd == NULL) { - return; - } - - gpr_mu_lock(&fd->mu); - - if (watcher == fd->read_watcher) { - /* remove read watcher, kick if we still need a read */ - was_polling = 1; - if (!got_read) { - kick = 1; - } - fd->read_watcher = NULL; - } - if (watcher == fd->write_watcher) { - /* remove write watcher, kick if we still need a write */ - was_polling = 1; - if (!got_write) { - kick = 1; - } - fd->write_watcher = NULL; - } - if (!was_polling && watcher->worker != NULL) { - /* remove from inactive list */ - watcher->next->prev = watcher->prev; - watcher->prev->next = watcher->next; - } - if (got_read) { - if (set_ready_locked(exec_ctx, fd, &fd->read_closure)) { - kick = 1; - } - } - if (got_write) { - if (set_ready_locked(exec_ctx, fd, &fd->write_closure)) { - kick = 1; - } - } - if (kick) { - maybe_wake_one_watcher_locked(fd); - } - if (grpc_fd_is_orphaned(fd) && !has_watchers(fd) && !fd->closed) { - close_fd_locked(exec_ctx, fd); - } - gpr_mu_unlock(&fd->mu); - - GRPC_FD_UNREF(fd, "poll"); -} - -void grpc_fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { - set_ready(exec_ctx, fd, &fd->read_closure); -} - -void grpc_fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { - set_ready(exec_ctx, fd, &fd->write_closure); -} - -#endif diff --git a/src/core/lib/iomgr/fd_posix.h b/src/core/lib/iomgr/fd_posix.h deleted file mode 100644 index dc389ba743..0000000000 --- a/src/core/lib/iomgr/fd_posix.h +++ /dev/null @@ -1,192 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_IOMGR_FD_POSIX_H -#define GRPC_CORE_LIB_IOMGR_FD_POSIX_H - -#include <grpc/support/atm.h> -#include <grpc/support/sync.h> -#include <grpc/support/time.h> -#include "src/core/lib/iomgr/iomgr_internal.h" -#include "src/core/lib/iomgr/pollset.h" - -typedef struct grpc_fd grpc_fd; - -typedef struct grpc_fd_watcher { - struct grpc_fd_watcher *next; - struct grpc_fd_watcher *prev; - grpc_pollset *pollset; - grpc_pollset_worker *worker; - grpc_fd *fd; -} grpc_fd_watcher; - -struct grpc_fd { - int fd; - /* refst format: - bit0: 1=active/0=orphaned - bit1-n: refcount - meaning that mostly we ref by two to avoid altering the orphaned bit, - and just unref by 1 when we're ready to flag the object as orphaned */ - gpr_atm refst; - - gpr_mu mu; - int shutdown; - int closed; - int released; - - /* The watcher list. - - The following watcher related fields are protected by watcher_mu. - - An fd_watcher is an ephemeral object created when an fd wants to - begin polling, and destroyed after the poll. - - It denotes the fd's interest in whether to read poll or write poll - or both or neither on this fd. - - If a watcher is asked to poll for reads or writes, the read_watcher - or write_watcher fields are set respectively. A watcher may be asked - to poll for both, in which case both fields will be set. - - read_watcher and write_watcher may be NULL if no watcher has been - asked to poll for reads or writes. - - If an fd_watcher is not asked to poll for reads or writes, it's added - to a linked list of inactive watchers, rooted at inactive_watcher_root. - If at a later time there becomes need of a poller to poll, one of - the inactive pollers may be kicked out of their poll loops to take - that responsibility. */ - grpc_fd_watcher inactive_watcher_root; - grpc_fd_watcher *read_watcher; - grpc_fd_watcher *write_watcher; - - grpc_closure *read_closure; - grpc_closure *write_closure; - - struct grpc_fd *freelist_next; - - grpc_closure *on_done_closure; - - grpc_iomgr_object iomgr_object; -}; - -/* Create a wrapped file descriptor. - Requires fd is a non-blocking file descriptor. - This takes ownership of closing fd. */ -grpc_fd *grpc_fd_create(int fd, const char *name); - -/* Return the wrapped fd, or -1 if it has been released or closed. */ -int grpc_fd_wrapped_fd(grpc_fd *fd); - -/* Releases fd to be asynchronously destroyed. - on_done is called when the underlying file descriptor is definitely close()d. - If on_done is NULL, no callback will be made. - If release_fd is not NULL, it's set to fd and fd will not be closed. - Requires: *fd initialized; no outstanding notify_on_read or - notify_on_write. - MUST NOT be called with a pollset lock taken */ -void grpc_fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done, - int *release_fd, const char *reason); - -/* Begin polling on an fd. - Registers that the given pollset is interested in this fd - so that if read - or writability interest changes, the pollset can be kicked to pick up that - new interest. - Return value is: - (fd_needs_read? read_mask : 0) | (fd_needs_write? write_mask : 0) - i.e. a combination of read_mask and write_mask determined by the fd's current - interest in said events. - Polling strategies that do not need to alter their behavior depending on the - fd's current interest (such as epoll) do not need to call this function. - MUST NOT be called with a pollset lock taken */ -uint32_t grpc_fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset, - grpc_pollset_worker *worker, uint32_t read_mask, - uint32_t write_mask, grpc_fd_watcher *rec); -/* Complete polling previously started with grpc_fd_begin_poll - MUST NOT be called with a pollset lock taken - if got_read or got_write are 1, also does the become_{readable,writable} as - appropriate. */ -void grpc_fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *rec, - int got_read, int got_write); - -/* Return 1 if this fd is orphaned, 0 otherwise */ -int grpc_fd_is_orphaned(grpc_fd *fd); - -/* Cause any current callbacks to error out with GRPC_CALLBACK_CANCELLED. */ -void grpc_fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd); - -/* Register read interest, causing read_cb to be called once when fd becomes - readable, on deadline specified by deadline, or on shutdown triggered by - grpc_fd_shutdown. - read_cb will be called with read_cb_arg when *fd becomes readable. - read_cb is Called with status of GRPC_CALLBACK_SUCCESS if readable, - GRPC_CALLBACK_TIMED_OUT if the call timed out, - and CANCELLED if the call was cancelled. - - Requires:This method must not be called before the read_cb for any previous - call runs. Edge triggered events are used whenever they are supported by the - underlying platform. This means that users must drain fd in read_cb before - calling notify_on_read again. Users are also expected to handle spurious - events, i.e read_cb is called while nothing can be readable from fd */ -void grpc_fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure *closure); - -/* Exactly the same semantics as above, except based on writable events. */ -void grpc_fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure *closure); - -/* Notification from the poller to an fd that it has become readable or - writable. - If allow_synchronous_callback is 1, allow running the fd callback inline - in this callstack, otherwise register an asynchronous callback and return */ -void grpc_fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd); -void grpc_fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd); - -/* Reference counting for fds */ -/*#define GRPC_FD_REF_COUNT_DEBUG*/ -#ifdef GRPC_FD_REF_COUNT_DEBUG -void grpc_fd_ref(grpc_fd *fd, const char *reason, const char *file, int line); -void grpc_fd_unref(grpc_fd *fd, const char *reason, const char *file, int line); -#define GRPC_FD_REF(fd, reason) grpc_fd_ref(fd, reason, __FILE__, __LINE__) -#define GRPC_FD_UNREF(fd, reason) grpc_fd_unref(fd, reason, __FILE__, __LINE__) -#else -void grpc_fd_ref(grpc_fd *fd); -void grpc_fd_unref(grpc_fd *fd); -#define GRPC_FD_REF(fd, reason) grpc_fd_ref(fd) -#define GRPC_FD_UNREF(fd, reason) grpc_fd_unref(fd) -#endif - -void grpc_fd_global_init(void); -void grpc_fd_global_shutdown(void); - -#endif /* GRPC_CORE_LIB_IOMGR_FD_POSIX_H */ diff --git a/src/core/lib/iomgr/iomgr.c b/src/core/lib/iomgr/iomgr.c index 5ac47f42f2..146663984d 100644 --- a/src/core/lib/iomgr/iomgr.c +++ b/src/core/lib/iomgr/iomgr.c @@ -63,7 +63,6 @@ void grpc_iomgr_init(void) { g_root_object.next = g_root_object.prev = &g_root_object; g_root_object.name = "root"; grpc_iomgr_platform_init(); - grpc_pollset_global_init(); } static size_t count_objects(void) { @@ -138,7 +137,6 @@ void grpc_iomgr_shutdown(void) { gpr_mu_lock(&g_mu); gpr_mu_unlock(&g_mu); - grpc_pollset_global_shutdown(); grpc_iomgr_platform_shutdown(); grpc_exec_ctx_global_shutdown(); gpr_mu_destroy(&g_mu); diff --git a/src/core/lib/iomgr/iomgr_internal.h b/src/core/lib/iomgr/iomgr_internal.h index 87d5a3d672..805be1f1e4 100644 --- a/src/core/lib/iomgr/iomgr_internal.h +++ b/src/core/lib/iomgr/iomgr_internal.h @@ -36,7 +36,6 @@ #include <stdbool.h> -#include <grpc/support/sync.h> #include "src/core/lib/iomgr/iomgr.h" typedef struct grpc_iomgr_object { @@ -45,9 +44,6 @@ typedef struct grpc_iomgr_object { struct grpc_iomgr_object *prev; } grpc_iomgr_object; -void grpc_pollset_global_init(void); -void grpc_pollset_global_shutdown(void); - void grpc_iomgr_register_object(grpc_iomgr_object *obj, const char *name); void grpc_iomgr_unregister_object(grpc_iomgr_object *obj); diff --git a/src/core/lib/iomgr/iomgr_posix.c b/src/core/lib/iomgr/iomgr_posix.c index 5d909bd220..016c501f75 100644 --- a/src/core/lib/iomgr/iomgr_posix.c +++ b/src/core/lib/iomgr/iomgr_posix.c @@ -36,17 +36,17 @@ #ifdef GPR_POSIX_SOCKET #include "src/core/lib/debug/trace.h" -#include "src/core/lib/iomgr/fd_posix.h" +#include "src/core/lib/iomgr/ev_posix.h" #include "src/core/lib/iomgr/iomgr_posix.h" #include "src/core/lib/iomgr/tcp_posix.h" void grpc_iomgr_platform_init(void) { - grpc_fd_global_init(); + grpc_event_engine_init(); grpc_register_tracer("tcp", &grpc_tcp_trace); } void grpc_iomgr_platform_flush(void) {} -void grpc_iomgr_platform_shutdown(void) { grpc_fd_global_shutdown(); } +void grpc_iomgr_platform_shutdown(void) { grpc_event_engine_shutdown(); } #endif /* GRPC_POSIX_SOCKET */ diff --git a/src/core/lib/iomgr/iomgr_windows.c b/src/core/lib/iomgr/iomgr_windows.c index 25c88425c4..398517fc75 100644 --- a/src/core/lib/iomgr/iomgr_windows.c +++ b/src/core/lib/iomgr/iomgr_windows.c @@ -41,6 +41,7 @@ #include "src/core/lib/iomgr/iocp_windows.h" #include "src/core/lib/iomgr/iomgr.h" +#include "src/core/lib/iomgr/pollset_windows.h" #include "src/core/lib/iomgr/socket_windows.h" /* Windows' io manager is going to be fully designed using IO completion @@ -61,11 +62,13 @@ static void winsock_shutdown(void) { void grpc_iomgr_platform_init(void) { winsock_init(); grpc_iocp_init(); + grpc_pollset_global_init(); } void grpc_iomgr_platform_flush(void) { grpc_iocp_flush(); } void grpc_iomgr_platform_shutdown(void) { + grpc_pollset_global_shutdown(); grpc_iocp_shutdown(); winsock_shutdown(); } diff --git a/src/core/lib/iomgr/pollset_multipoller_with_epoll.c b/src/core/lib/iomgr/pollset_multipoller_with_epoll.c deleted file mode 100644 index 04d912e94d..0000000000 --- a/src/core/lib/iomgr/pollset_multipoller_with_epoll.c +++ /dev/null @@ -1,324 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/support/port_platform.h> - -#ifdef GPR_LINUX_MULTIPOLL_WITH_EPOLL - -#include <errno.h> -#include <poll.h> -#include <string.h> -#include <sys/epoll.h> -#include <unistd.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/useful.h> -#include "src/core/lib/iomgr/fd_posix.h" -#include "src/core/lib/iomgr/pollset_posix.h" -#include "src/core/lib/profiling/timers.h" -#include "src/core/lib/support/block_annotate.h" - -struct epoll_fd_list { - int *epoll_fds; - size_t count; - size_t capacity; -}; - -static struct epoll_fd_list epoll_fd_global_list; -static gpr_once init_epoll_fd_list_mu = GPR_ONCE_INIT; -static gpr_mu epoll_fd_list_mu; - -static void init_mu(void) { gpr_mu_init(&epoll_fd_list_mu); } - -static void add_epoll_fd_to_global_list(int epoll_fd) { - gpr_once_init(&init_epoll_fd_list_mu, init_mu); - - gpr_mu_lock(&epoll_fd_list_mu); - if (epoll_fd_global_list.count == epoll_fd_global_list.capacity) { - epoll_fd_global_list.capacity = - GPR_MAX((size_t)8, epoll_fd_global_list.capacity * 2); - epoll_fd_global_list.epoll_fds = - gpr_realloc(epoll_fd_global_list.epoll_fds, - epoll_fd_global_list.capacity * sizeof(int)); - } - epoll_fd_global_list.epoll_fds[epoll_fd_global_list.count++] = epoll_fd; - gpr_mu_unlock(&epoll_fd_list_mu); -} - -static void remove_epoll_fd_from_global_list(int epoll_fd) { - gpr_mu_lock(&epoll_fd_list_mu); - GPR_ASSERT(epoll_fd_global_list.count > 0); - for (size_t i = 0; i < epoll_fd_global_list.count; i++) { - if (epoll_fd == epoll_fd_global_list.epoll_fds[i]) { - epoll_fd_global_list.epoll_fds[i] = - epoll_fd_global_list.epoll_fds[--(epoll_fd_global_list.count)]; - break; - } - } - gpr_mu_unlock(&epoll_fd_list_mu); -} - -void grpc_remove_fd_from_all_epoll_sets(int fd) { - int err; - gpr_once_init(&init_epoll_fd_list_mu, init_mu); - gpr_mu_lock(&epoll_fd_list_mu); - if (epoll_fd_global_list.count == 0) { - gpr_mu_unlock(&epoll_fd_list_mu); - return; - } - for (size_t i = 0; i < epoll_fd_global_list.count; i++) { - err = epoll_ctl(epoll_fd_global_list.epoll_fds[i], EPOLL_CTL_DEL, fd, NULL); - if (err < 0 && errno != ENOENT) { - gpr_log(GPR_ERROR, "epoll_ctl del for %d failed: %s", fd, - strerror(errno)); - } - } - gpr_mu_unlock(&epoll_fd_list_mu); -} - -typedef struct { - grpc_pollset *pollset; - grpc_fd *fd; - grpc_closure closure; -} delayed_add; - -typedef struct { int epoll_fd; } pollset_hdr; - -static void finally_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_fd *fd) { - pollset_hdr *h = pollset->data.ptr; - struct epoll_event ev; - int err; - grpc_fd_watcher watcher; - - /* We pretend to be polling whilst adding an fd to keep the fd from being - closed during the add. This may result in a spurious wakeup being assigned - to this pollset whilst adding, but that should be benign. */ - GPR_ASSERT(grpc_fd_begin_poll(fd, pollset, NULL, 0, 0, &watcher) == 0); - if (watcher.fd != NULL) { - ev.events = (uint32_t)(EPOLLIN | EPOLLOUT | EPOLLET); - ev.data.ptr = fd; - err = epoll_ctl(h->epoll_fd, EPOLL_CTL_ADD, fd->fd, &ev); - if (err < 0) { - /* FDs may be added to a pollset multiple times, so EEXIST is normal. */ - if (errno != EEXIST) { - gpr_log(GPR_ERROR, "epoll_ctl add for %d failed: %s", fd->fd, - strerror(errno)); - } - } - } - grpc_fd_end_poll(exec_ctx, &watcher, 0, 0); -} - -static void perform_delayed_add(grpc_exec_ctx *exec_ctx, void *arg, - bool iomgr_status) { - delayed_add *da = arg; - - if (!grpc_fd_is_orphaned(da->fd)) { - finally_add_fd(exec_ctx, da->pollset, da->fd); - } - - gpr_mu_lock(&da->pollset->mu); - da->pollset->in_flight_cbs--; - if (da->pollset->shutting_down) { - /* We don't care about this pollset anymore. */ - if (da->pollset->in_flight_cbs == 0 && !da->pollset->called_shutdown) { - da->pollset->called_shutdown = 1; - grpc_exec_ctx_enqueue(exec_ctx, da->pollset->shutdown_done, true, NULL); - } - } - gpr_mu_unlock(&da->pollset->mu); - - GRPC_FD_UNREF(da->fd, "delayed_add"); - - gpr_free(da); -} - -static void multipoll_with_epoll_pollset_add_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, - grpc_fd *fd, - int and_unlock_pollset) { - if (and_unlock_pollset) { - gpr_mu_unlock(&pollset->mu); - finally_add_fd(exec_ctx, pollset, fd); - } else { - delayed_add *da = gpr_malloc(sizeof(*da)); - da->pollset = pollset; - da->fd = fd; - GRPC_FD_REF(fd, "delayed_add"); - grpc_closure_init(&da->closure, perform_delayed_add, da); - pollset->in_flight_cbs++; - grpc_exec_ctx_enqueue(exec_ctx, &da->closure, true, NULL); - } -} - -/* TODO(klempner): We probably want to turn this down a bit */ -#define GRPC_EPOLL_MAX_EVENTS 1000 - -static void multipoll_with_epoll_pollset_maybe_work_and_unlock( - grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker, - gpr_timespec deadline, gpr_timespec now) { - struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS]; - int ep_rv; - int poll_rv; - pollset_hdr *h = pollset->data.ptr; - int timeout_ms; - struct pollfd pfds[2]; - - /* If you want to ignore epoll's ability to sanely handle parallel pollers, - * for a more apples-to-apples performance comparison with poll, add a - * if (pollset->counter != 0) { return 0; } - * here. - */ - - gpr_mu_unlock(&pollset->mu); - - timeout_ms = grpc_poll_deadline_to_millis_timeout(deadline, now); - - pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); - pfds[0].events = POLLIN; - pfds[0].revents = 0; - pfds[1].fd = h->epoll_fd; - pfds[1].events = POLLIN; - pfds[1].revents = 0; - - /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid - even going into the blocking annotation if possible */ - GPR_TIMER_BEGIN("poll", 0); - GRPC_SCHEDULING_START_BLOCKING_REGION; - poll_rv = grpc_poll_function(pfds, 2, timeout_ms); - GRPC_SCHEDULING_END_BLOCKING_REGION; - GPR_TIMER_END("poll", 0); - - if (poll_rv < 0) { - if (errno != EINTR) { - gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno)); - } - } else if (poll_rv == 0) { - /* do nothing */ - } else { - if (pfds[0].revents) { - grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd); - } - if (pfds[1].revents) { - do { - /* The following epoll_wait never blocks; it has a timeout of 0 */ - ep_rv = epoll_wait(h->epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, 0); - if (ep_rv < 0) { - if (errno != EINTR) { - gpr_log(GPR_ERROR, "epoll_wait() failed: %s", strerror(errno)); - } - } else { - int i; - for (i = 0; i < ep_rv; ++i) { - grpc_fd *fd = ep_ev[i].data.ptr; - /* TODO(klempner): We might want to consider making err and pri - * separate events */ - int cancel = ep_ev[i].events & (EPOLLERR | EPOLLHUP); - int read_ev = ep_ev[i].events & (EPOLLIN | EPOLLPRI); - int write_ev = ep_ev[i].events & EPOLLOUT; - if (fd == NULL) { - grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd); - } else { - if (read_ev || cancel) { - grpc_fd_become_readable(exec_ctx, fd); - } - if (write_ev || cancel) { - grpc_fd_become_writable(exec_ctx, fd); - } - } - } - } - } while (ep_rv == GRPC_EPOLL_MAX_EVENTS); - } - } -} - -static void multipoll_with_epoll_pollset_finish_shutdown( - grpc_pollset *pollset) {} - -static void multipoll_with_epoll_pollset_destroy(grpc_pollset *pollset) { - pollset_hdr *h = pollset->data.ptr; - close(h->epoll_fd); - remove_epoll_fd_from_global_list(h->epoll_fd); - gpr_free(h); -} - -static const grpc_pollset_vtable multipoll_with_epoll_pollset = { - multipoll_with_epoll_pollset_add_fd, - multipoll_with_epoll_pollset_maybe_work_and_unlock, - multipoll_with_epoll_pollset_finish_shutdown, - multipoll_with_epoll_pollset_destroy}; - -static void epoll_become_multipoller(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, grpc_fd **fds, - size_t nfds) { - size_t i; - pollset_hdr *h = gpr_malloc(sizeof(pollset_hdr)); - struct epoll_event ev; - int err; - - pollset->vtable = &multipoll_with_epoll_pollset; - pollset->data.ptr = h; - h->epoll_fd = epoll_create1(EPOLL_CLOEXEC); - if (h->epoll_fd < 0) { - /* TODO(klempner): Fall back to poll here, especially on ENOSYS */ - gpr_log(GPR_ERROR, "epoll_create1 failed: %s", strerror(errno)); - abort(); - } - add_epoll_fd_to_global_list(h->epoll_fd); - - ev.events = (uint32_t)(EPOLLIN | EPOLLET); - ev.data.ptr = NULL; - err = epoll_ctl(h->epoll_fd, EPOLL_CTL_ADD, - GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), &ev); - if (err < 0) { - gpr_log(GPR_ERROR, "epoll_ctl add for %d failed: %s", - GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), - strerror(errno)); - } - - for (i = 0; i < nfds; i++) { - multipoll_with_epoll_pollset_add_fd(exec_ctx, pollset, fds[i], 0); - } -} - -grpc_platform_become_multipoller_type grpc_platform_become_multipoller = - epoll_become_multipoller; - -#else /* GPR_LINUX_MULTIPOLL_WITH_EPOLL */ - -void grpc_remove_fd_from_all_epoll_sets(int fd) {} - -#endif /* GPR_LINUX_MULTIPOLL_WITH_EPOLL */ diff --git a/src/core/lib/iomgr/pollset_multipoller_with_poll_posix.c b/src/core/lib/iomgr/pollset_multipoller_with_poll_posix.c deleted file mode 100644 index 7668bf63f5..0000000000 --- a/src/core/lib/iomgr/pollset_multipoller_with_poll_posix.c +++ /dev/null @@ -1,234 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/support/port_platform.h> - -#ifdef GPR_POSIX_SOCKET - -#include "src/core/lib/iomgr/pollset_posix.h" - -#include <errno.h> -#include <poll.h> -#include <stdlib.h> -#include <string.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/useful.h> - -#include "src/core/lib/iomgr/fd_posix.h" -#include "src/core/lib/iomgr/iomgr_internal.h" -#include "src/core/lib/iomgr/pollset_posix.h" -#include "src/core/lib/support/block_annotate.h" - -typedef struct { - /* all polled fds */ - size_t fd_count; - size_t fd_capacity; - grpc_fd **fds; - /* fds that have been removed from the pollset explicitly */ - size_t del_count; - size_t del_capacity; - grpc_fd **dels; -} pollset_hdr; - -static void multipoll_with_poll_pollset_add_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, - grpc_fd *fd, - int and_unlock_pollset) { - size_t i; - pollset_hdr *h = pollset->data.ptr; - /* TODO(ctiller): this is O(num_fds^2); maybe switch to a hash set here */ - for (i = 0; i < h->fd_count; i++) { - if (h->fds[i] == fd) goto exit; - } - if (h->fd_count == h->fd_capacity) { - h->fd_capacity = GPR_MAX(h->fd_capacity + 8, h->fd_count * 3 / 2); - h->fds = gpr_realloc(h->fds, sizeof(grpc_fd *) * h->fd_capacity); - } - h->fds[h->fd_count++] = fd; - GRPC_FD_REF(fd, "multipoller"); -exit: - if (and_unlock_pollset) { - gpr_mu_unlock(&pollset->mu); - } -} - -static void multipoll_with_poll_pollset_maybe_work_and_unlock( - grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker, - gpr_timespec deadline, gpr_timespec now) { -#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR) -#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR) - - int timeout; - int r; - size_t i, j, fd_count; - nfds_t pfd_count; - pollset_hdr *h; - /* TODO(ctiller): inline some elements to avoid an allocation */ - grpc_fd_watcher *watchers; - struct pollfd *pfds; - - h = pollset->data.ptr; - timeout = grpc_poll_deadline_to_millis_timeout(deadline, now); - /* TODO(ctiller): perform just one malloc here if we exceed the inline case */ - pfds = gpr_malloc(sizeof(*pfds) * (h->fd_count + 2)); - watchers = gpr_malloc(sizeof(*watchers) * (h->fd_count + 2)); - fd_count = 0; - pfd_count = 2; - pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd); - pfds[0].events = POLLIN; - pfds[0].revents = 0; - pfds[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); - pfds[1].events = POLLIN; - pfds[1].revents = 0; - for (i = 0; i < h->fd_count; i++) { - int remove = grpc_fd_is_orphaned(h->fds[i]); - for (j = 0; !remove && j < h->del_count; j++) { - if (h->fds[i] == h->dels[j]) remove = 1; - } - if (remove) { - GRPC_FD_UNREF(h->fds[i], "multipoller"); - } else { - h->fds[fd_count++] = h->fds[i]; - watchers[pfd_count].fd = h->fds[i]; - GRPC_FD_REF(watchers[pfd_count].fd, "multipoller_start"); - pfds[pfd_count].fd = h->fds[i]->fd; - pfds[pfd_count].revents = 0; - pfd_count++; - } - } - for (j = 0; j < h->del_count; j++) { - GRPC_FD_UNREF(h->dels[j], "multipoller_del"); - } - h->del_count = 0; - h->fd_count = fd_count; - gpr_mu_unlock(&pollset->mu); - - for (i = 2; i < pfd_count; i++) { - grpc_fd *fd = watchers[i].fd; - pfds[i].events = (short)grpc_fd_begin_poll(fd, pollset, worker, POLLIN, - POLLOUT, &watchers[i]); - GRPC_FD_UNREF(fd, "multipoller_start"); - } - - /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid - even going into the blocking annotation if possible */ - GRPC_SCHEDULING_START_BLOCKING_REGION; - r = grpc_poll_function(pfds, pfd_count, timeout); - GRPC_SCHEDULING_END_BLOCKING_REGION; - - if (r < 0) { - if (errno != EINTR) { - gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno)); - } - for (i = 2; i < pfd_count; i++) { - grpc_fd_end_poll(exec_ctx, &watchers[i], 0, 0); - } - } else if (r == 0) { - for (i = 2; i < pfd_count; i++) { - grpc_fd_end_poll(exec_ctx, &watchers[i], 0, 0); - } - } else { - if (pfds[0].revents & POLLIN_CHECK) { - grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd); - } - if (pfds[1].revents & POLLIN_CHECK) { - grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd); - } - for (i = 2; i < pfd_count; i++) { - if (watchers[i].fd == NULL) { - grpc_fd_end_poll(exec_ctx, &watchers[i], 0, 0); - continue; - } - grpc_fd_end_poll(exec_ctx, &watchers[i], pfds[i].revents & POLLIN_CHECK, - pfds[i].revents & POLLOUT_CHECK); - } - } - - gpr_free(pfds); - gpr_free(watchers); -} - -static void multipoll_with_poll_pollset_finish_shutdown(grpc_pollset *pollset) { - size_t i; - pollset_hdr *h = pollset->data.ptr; - for (i = 0; i < h->fd_count; i++) { - GRPC_FD_UNREF(h->fds[i], "multipoller"); - } - for (i = 0; i < h->del_count; i++) { - GRPC_FD_UNREF(h->dels[i], "multipoller_del"); - } - h->fd_count = 0; - h->del_count = 0; -} - -static void multipoll_with_poll_pollset_destroy(grpc_pollset *pollset) { - pollset_hdr *h = pollset->data.ptr; - multipoll_with_poll_pollset_finish_shutdown(pollset); - gpr_free(h->fds); - gpr_free(h->dels); - gpr_free(h); -} - -static const grpc_pollset_vtable multipoll_with_poll_pollset = { - multipoll_with_poll_pollset_add_fd, - multipoll_with_poll_pollset_maybe_work_and_unlock, - multipoll_with_poll_pollset_finish_shutdown, - multipoll_with_poll_pollset_destroy}; - -void grpc_poll_become_multipoller(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, grpc_fd **fds, - size_t nfds) { - size_t i; - pollset_hdr *h = gpr_malloc(sizeof(pollset_hdr)); - pollset->vtable = &multipoll_with_poll_pollset; - pollset->data.ptr = h; - h->fd_count = nfds; - h->fd_capacity = nfds; - h->fds = gpr_malloc(nfds * sizeof(grpc_fd *)); - h->del_count = 0; - h->del_capacity = 0; - h->dels = NULL; - for (i = 0; i < nfds; i++) { - h->fds[i] = fds[i]; - GRPC_FD_REF(fds[i], "multipoller"); - } -} - -#endif /* GPR_POSIX_SOCKET */ - -#ifdef GPR_POSIX_MULTIPOLL_WITH_POLL -grpc_platform_become_multipoller_type grpc_platform_become_multipoller = - grpc_poll_become_multipoller; -#endif diff --git a/src/core/lib/iomgr/pollset_posix.c b/src/core/lib/iomgr/pollset_posix.c deleted file mode 100644 index 84a9153034..0000000000 --- a/src/core/lib/iomgr/pollset_posix.c +++ /dev/null @@ -1,633 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/support/port_platform.h> - -#ifdef GPR_POSIX_SOCKET - -#include "src/core/lib/iomgr/pollset_posix.h" - -#include <errno.h> -#include <stdlib.h> -#include <string.h> -#include <unistd.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/thd.h> -#include <grpc/support/tls.h> -#include <grpc/support/useful.h> -#include "src/core/lib/iomgr/fd_posix.h" -#include "src/core/lib/iomgr/iomgr_internal.h" -#include "src/core/lib/iomgr/socket_utils_posix.h" -#include "src/core/lib/profiling/timers.h" -#include "src/core/lib/support/block_annotate.h" - -GPR_TLS_DECL(g_current_thread_poller); -GPR_TLS_DECL(g_current_thread_worker); - -/** Default poll() function - a pointer so that it can be overridden by some - * tests */ -grpc_poll_function_type grpc_poll_function = poll; - -/** The alarm system needs to be able to wakeup 'some poller' sometimes - * (specifically when a new alarm needs to be triggered earlier than the next - * alarm 'epoch'). - * This wakeup_fd gives us something to alert on when such a case occurs. */ -grpc_wakeup_fd grpc_global_wakeup_fd; - -static void remove_worker(grpc_pollset *p, grpc_pollset_worker *worker) { - worker->prev->next = worker->next; - worker->next->prev = worker->prev; -} - -int grpc_pollset_has_workers(grpc_pollset *p) { - return p->root_worker.next != &p->root_worker; -} - -static grpc_pollset_worker *pop_front_worker(grpc_pollset *p) { - if (grpc_pollset_has_workers(p)) { - grpc_pollset_worker *w = p->root_worker.next; - remove_worker(p, w); - return w; - } else { - return NULL; - } -} - -static void push_back_worker(grpc_pollset *p, grpc_pollset_worker *worker) { - worker->next = &p->root_worker; - worker->prev = worker->next->prev; - worker->prev->next = worker->next->prev = worker; -} - -static void push_front_worker(grpc_pollset *p, grpc_pollset_worker *worker) { - worker->prev = &p->root_worker; - worker->next = worker->prev->next; - worker->prev->next = worker->next->prev = worker; -} - -size_t grpc_pollset_size(void) { return sizeof(grpc_pollset); } - -void grpc_pollset_kick_ext(grpc_pollset *p, - grpc_pollset_worker *specific_worker, - uint32_t flags) { - GPR_TIMER_BEGIN("grpc_pollset_kick_ext", 0); - - /* pollset->mu already held */ - if (specific_worker != NULL) { - if (specific_worker == GRPC_POLLSET_KICK_BROADCAST) { - GPR_TIMER_BEGIN("grpc_pollset_kick_ext.broadcast", 0); - GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0); - for (specific_worker = p->root_worker.next; - specific_worker != &p->root_worker; - specific_worker = specific_worker->next) { - grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); - } - p->kicked_without_pollers = 1; - GPR_TIMER_END("grpc_pollset_kick_ext.broadcast", 0); - } else if (gpr_tls_get(&g_current_thread_worker) != - (intptr_t)specific_worker) { - GPR_TIMER_MARK("different_thread_worker", 0); - if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) { - specific_worker->reevaluate_polling_on_wakeup = 1; - } - specific_worker->kicked_specifically = 1; - grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); - } else if ((flags & GRPC_POLLSET_CAN_KICK_SELF) != 0) { - GPR_TIMER_MARK("kick_yoself", 0); - if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) { - specific_worker->reevaluate_polling_on_wakeup = 1; - } - specific_worker->kicked_specifically = 1; - grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); - } - } else if (gpr_tls_get(&g_current_thread_poller) != (intptr_t)p) { - GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0); - GPR_TIMER_MARK("kick_anonymous", 0); - specific_worker = pop_front_worker(p); - if (specific_worker != NULL) { - if (gpr_tls_get(&g_current_thread_worker) == (intptr_t)specific_worker) { - GPR_TIMER_MARK("kick_anonymous_not_self", 0); - push_back_worker(p, specific_worker); - specific_worker = pop_front_worker(p); - if ((flags & GRPC_POLLSET_CAN_KICK_SELF) == 0 && - gpr_tls_get(&g_current_thread_worker) == - (intptr_t)specific_worker) { - push_back_worker(p, specific_worker); - specific_worker = NULL; - } - } - if (specific_worker != NULL) { - GPR_TIMER_MARK("finally_kick", 0); - push_back_worker(p, specific_worker); - grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); - } - } else { - GPR_TIMER_MARK("kicked_no_pollers", 0); - p->kicked_without_pollers = 1; - } - } - - GPR_TIMER_END("grpc_pollset_kick_ext", 0); -} - -void grpc_pollset_kick(grpc_pollset *p, grpc_pollset_worker *specific_worker) { - grpc_pollset_kick_ext(p, specific_worker, 0); -} - -/* global state management */ - -void grpc_pollset_global_init(void) { - gpr_tls_init(&g_current_thread_poller); - gpr_tls_init(&g_current_thread_worker); - grpc_wakeup_fd_global_init(); - grpc_wakeup_fd_init(&grpc_global_wakeup_fd); -} - -void grpc_pollset_global_shutdown(void) { - grpc_wakeup_fd_destroy(&grpc_global_wakeup_fd); - gpr_tls_destroy(&g_current_thread_poller); - gpr_tls_destroy(&g_current_thread_worker); - grpc_wakeup_fd_global_destroy(); -} - -void grpc_kick_poller(void) { grpc_wakeup_fd_wakeup(&grpc_global_wakeup_fd); } - -/* main interface */ - -static void become_basic_pollset(grpc_pollset *pollset, grpc_fd *fd_or_null); - -void grpc_pollset_init(grpc_pollset *pollset, gpr_mu **mu) { - gpr_mu_init(&pollset->mu); - *mu = &pollset->mu; - pollset->root_worker.next = pollset->root_worker.prev = &pollset->root_worker; - pollset->in_flight_cbs = 0; - pollset->shutting_down = 0; - pollset->called_shutdown = 0; - pollset->kicked_without_pollers = 0; - pollset->idle_jobs.head = pollset->idle_jobs.tail = NULL; - pollset->local_wakeup_cache = NULL; - pollset->kicked_without_pollers = 0; - become_basic_pollset(pollset, NULL); -} - -void grpc_pollset_destroy(grpc_pollset *pollset) { - GPR_ASSERT(pollset->in_flight_cbs == 0); - GPR_ASSERT(!grpc_pollset_has_workers(pollset)); - GPR_ASSERT(pollset->idle_jobs.head == pollset->idle_jobs.tail); - pollset->vtable->destroy(pollset); - while (pollset->local_wakeup_cache) { - grpc_cached_wakeup_fd *next = pollset->local_wakeup_cache->next; - grpc_wakeup_fd_destroy(&pollset->local_wakeup_cache->fd); - gpr_free(pollset->local_wakeup_cache); - pollset->local_wakeup_cache = next; - } -} - -void grpc_pollset_reset(grpc_pollset *pollset) { - GPR_ASSERT(pollset->shutting_down); - GPR_ASSERT(pollset->in_flight_cbs == 0); - GPR_ASSERT(!grpc_pollset_has_workers(pollset)); - GPR_ASSERT(pollset->idle_jobs.head == pollset->idle_jobs.tail); - pollset->vtable->destroy(pollset); - pollset->shutting_down = 0; - pollset->called_shutdown = 0; - pollset->kicked_without_pollers = 0; - become_basic_pollset(pollset, NULL); -} - -void grpc_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_fd *fd) { - gpr_mu_lock(&pollset->mu); - pollset->vtable->add_fd(exec_ctx, pollset, fd, 1); -/* the following (enabled only in debug) will reacquire and then release - our lock - meaning that if the unlocking flag passed to add_fd above is - not respected, the code will deadlock (in a way that we have a chance of - debugging) */ -#ifndef NDEBUG - gpr_mu_lock(&pollset->mu); - gpr_mu_unlock(&pollset->mu); -#endif -} - -static void finish_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset) { - GPR_ASSERT(grpc_closure_list_empty(pollset->idle_jobs)); - pollset->vtable->finish_shutdown(pollset); - grpc_exec_ctx_enqueue(exec_ctx, pollset->shutdown_done, true, NULL); -} - -void grpc_pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_pollset_worker **worker_hdl, gpr_timespec now, - gpr_timespec deadline) { - grpc_pollset_worker worker; - *worker_hdl = &worker; - - /* pollset->mu already held */ - int added_worker = 0; - int locked = 1; - int queued_work = 0; - int keep_polling = 0; - GPR_TIMER_BEGIN("grpc_pollset_work", 0); - /* this must happen before we (potentially) drop pollset->mu */ - worker.next = worker.prev = NULL; - worker.reevaluate_polling_on_wakeup = 0; - if (pollset->local_wakeup_cache != NULL) { - worker.wakeup_fd = pollset->local_wakeup_cache; - pollset->local_wakeup_cache = worker.wakeup_fd->next; - } else { - worker.wakeup_fd = gpr_malloc(sizeof(*worker.wakeup_fd)); - grpc_wakeup_fd_init(&worker.wakeup_fd->fd); - } - worker.kicked_specifically = 0; - /* If there's work waiting for the pollset to be idle, and the - pollset is idle, then do that work */ - if (!grpc_pollset_has_workers(pollset) && - !grpc_closure_list_empty(pollset->idle_jobs)) { - GPR_TIMER_MARK("grpc_pollset_work.idle_jobs", 0); - grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); - goto done; - } - /* If we're shutting down then we don't execute any extended work */ - if (pollset->shutting_down) { - GPR_TIMER_MARK("grpc_pollset_work.shutting_down", 0); - goto done; - } - /* Give do_promote priority so we don't starve it out */ - if (pollset->in_flight_cbs) { - GPR_TIMER_MARK("grpc_pollset_work.in_flight_cbs", 0); - gpr_mu_unlock(&pollset->mu); - locked = 0; - goto done; - } - /* Start polling, and keep doing so while we're being asked to - re-evaluate our pollers (this allows poll() based pollers to - ensure they don't miss wakeups) */ - keep_polling = 1; - while (keep_polling) { - keep_polling = 0; - if (!pollset->kicked_without_pollers) { - if (!added_worker) { - push_front_worker(pollset, &worker); - added_worker = 1; - gpr_tls_set(&g_current_thread_worker, (intptr_t)&worker); - } - gpr_tls_set(&g_current_thread_poller, (intptr_t)pollset); - GPR_TIMER_BEGIN("maybe_work_and_unlock", 0); - pollset->vtable->maybe_work_and_unlock(exec_ctx, pollset, &worker, - deadline, now); - GPR_TIMER_END("maybe_work_and_unlock", 0); - locked = 0; - gpr_tls_set(&g_current_thread_poller, 0); - } else { - GPR_TIMER_MARK("grpc_pollset_work.kicked_without_pollers", 0); - pollset->kicked_without_pollers = 0; - } - /* Finished execution - start cleaning up. - Note that we may arrive here from outside the enclosing while() loop. - In that case we won't loop though as we haven't added worker to the - worker list, which means nobody could ask us to re-evaluate polling). */ - done: - if (!locked) { - queued_work |= grpc_exec_ctx_flush(exec_ctx); - gpr_mu_lock(&pollset->mu); - locked = 1; - } - /* If we're forced to re-evaluate polling (via grpc_pollset_kick with - GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) then we land here and force - a loop */ - if (worker.reevaluate_polling_on_wakeup) { - worker.reevaluate_polling_on_wakeup = 0; - pollset->kicked_without_pollers = 0; - if (queued_work || worker.kicked_specifically) { - /* If there's queued work on the list, then set the deadline to be - immediate so we get back out of the polling loop quickly */ - deadline = gpr_inf_past(GPR_CLOCK_MONOTONIC); - } - keep_polling = 1; - } - } - if (added_worker) { - remove_worker(pollset, &worker); - gpr_tls_set(&g_current_thread_worker, 0); - } - /* release wakeup fd to the local pool */ - worker.wakeup_fd->next = pollset->local_wakeup_cache; - pollset->local_wakeup_cache = worker.wakeup_fd; - /* check shutdown conditions */ - if (pollset->shutting_down) { - if (grpc_pollset_has_workers(pollset)) { - grpc_pollset_kick(pollset, NULL); - } else if (!pollset->called_shutdown && pollset->in_flight_cbs == 0) { - pollset->called_shutdown = 1; - gpr_mu_unlock(&pollset->mu); - finish_shutdown(exec_ctx, pollset); - grpc_exec_ctx_flush(exec_ctx); - /* Continuing to access pollset here is safe -- it is the caller's - * responsibility to not destroy when it has outstanding calls to - * grpc_pollset_work. - * TODO(dklempner): Can we refactor the shutdown logic to avoid this? */ - gpr_mu_lock(&pollset->mu); - } else if (!grpc_closure_list_empty(pollset->idle_jobs)) { - grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); - gpr_mu_unlock(&pollset->mu); - grpc_exec_ctx_flush(exec_ctx); - gpr_mu_lock(&pollset->mu); - } - } - *worker_hdl = NULL; - GPR_TIMER_END("grpc_pollset_work", 0); -} - -void grpc_pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_closure *closure) { - GPR_ASSERT(!pollset->shutting_down); - pollset->shutting_down = 1; - pollset->shutdown_done = closure; - grpc_pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); - if (!grpc_pollset_has_workers(pollset)) { - grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); - } - if (!pollset->called_shutdown && pollset->in_flight_cbs == 0 && - !grpc_pollset_has_workers(pollset)) { - pollset->called_shutdown = 1; - finish_shutdown(exec_ctx, pollset); - } -} - -int grpc_poll_deadline_to_millis_timeout(gpr_timespec deadline, - gpr_timespec now) { - gpr_timespec timeout; - static const int64_t max_spin_polling_us = 10; - if (gpr_time_cmp(deadline, gpr_inf_future(deadline.clock_type)) == 0) { - return -1; - } - if (gpr_time_cmp(deadline, gpr_time_add(now, gpr_time_from_micros( - max_spin_polling_us, - GPR_TIMESPAN))) <= 0) { - return 0; - } - timeout = gpr_time_sub(deadline, now); - return gpr_time_to_millis(gpr_time_add( - timeout, gpr_time_from_nanos(GPR_NS_PER_MS - 1, GPR_TIMESPAN))); -} - -/* - * basic_pollset - a vtable that provides polling for zero or one file - * descriptor via poll() - */ - -typedef struct grpc_unary_promote_args { - const grpc_pollset_vtable *original_vtable; - grpc_pollset *pollset; - grpc_fd *fd; - grpc_closure promotion_closure; -} grpc_unary_promote_args; - -static void basic_do_promote(grpc_exec_ctx *exec_ctx, void *args, - bool success) { - grpc_unary_promote_args *up_args = args; - const grpc_pollset_vtable *original_vtable = up_args->original_vtable; - grpc_pollset *pollset = up_args->pollset; - grpc_fd *fd = up_args->fd; - - /* - * This is quite tricky. There are a number of cases to keep in mind here: - * 1. fd may have been orphaned - * 2. The pollset may no longer be a unary poller (and we can't let case #1 - * leak to other pollset types!) - * 3. pollset's fd (which may have changed) may have been orphaned - * 4. The pollset may be shutting down. - */ - - gpr_mu_lock(&pollset->mu); - /* First we need to ensure that nobody is polling concurrently */ - GPR_ASSERT(!grpc_pollset_has_workers(pollset)); - - gpr_free(up_args); - /* At this point the pollset may no longer be a unary poller. In that case - * we should just call the right add function and be done. */ - /* TODO(klempner): If we're not careful this could cause infinite recursion. - * That's not a problem for now because empty_pollset has a trivial poller - * and we don't have any mechanism to unbecome multipoller. */ - pollset->in_flight_cbs--; - if (pollset->shutting_down) { - /* We don't care about this pollset anymore. */ - if (pollset->in_flight_cbs == 0 && !pollset->called_shutdown) { - pollset->called_shutdown = 1; - finish_shutdown(exec_ctx, pollset); - } - } else if (grpc_fd_is_orphaned(fd)) { - /* Don't try to add it to anything, we'll drop our ref on it below */ - } else if (pollset->vtable != original_vtable) { - pollset->vtable->add_fd(exec_ctx, pollset, fd, 0); - } else if (fd != pollset->data.ptr) { - grpc_fd *fds[2]; - fds[0] = pollset->data.ptr; - fds[1] = fd; - - if (fds[0] && !grpc_fd_is_orphaned(fds[0])) { - grpc_platform_become_multipoller(exec_ctx, pollset, fds, - GPR_ARRAY_SIZE(fds)); - GRPC_FD_UNREF(fds[0], "basicpoll"); - } else { - /* old fd is orphaned and we haven't cleaned it up until now, so remain a - * unary poller */ - /* Note that it is possible that fds[1] is also orphaned at this point. - * That's okay, we'll correct it at the next add or poll. */ - if (fds[0]) GRPC_FD_UNREF(fds[0], "basicpoll"); - pollset->data.ptr = fd; - GRPC_FD_REF(fd, "basicpoll"); - } - } - - gpr_mu_unlock(&pollset->mu); - - /* Matching ref in basic_pollset_add_fd */ - GRPC_FD_UNREF(fd, "basicpoll_add"); -} - -static void basic_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_fd *fd, int and_unlock_pollset) { - grpc_unary_promote_args *up_args; - GPR_ASSERT(fd); - if (fd == pollset->data.ptr) goto exit; - - if (!grpc_pollset_has_workers(pollset)) { - /* Fast path -- no in flight cbs */ - /* TODO(klempner): Comment this out and fix any test failures or establish - * they are due to timing issues */ - grpc_fd *fds[2]; - fds[0] = pollset->data.ptr; - fds[1] = fd; - - if (fds[0] == NULL) { - pollset->data.ptr = fd; - GRPC_FD_REF(fd, "basicpoll"); - } else if (!grpc_fd_is_orphaned(fds[0])) { - grpc_platform_become_multipoller(exec_ctx, pollset, fds, - GPR_ARRAY_SIZE(fds)); - GRPC_FD_UNREF(fds[0], "basicpoll"); - } else { - /* old fd is orphaned and we haven't cleaned it up until now, so remain a - * unary poller */ - GRPC_FD_UNREF(fds[0], "basicpoll"); - pollset->data.ptr = fd; - GRPC_FD_REF(fd, "basicpoll"); - } - goto exit; - } - - /* Now we need to promote. This needs to happen when we're not polling. Since - * this may be called from poll, the wait needs to happen asynchronously. */ - GRPC_FD_REF(fd, "basicpoll_add"); - pollset->in_flight_cbs++; - up_args = gpr_malloc(sizeof(*up_args)); - up_args->fd = fd; - up_args->original_vtable = pollset->vtable; - up_args->pollset = pollset; - up_args->promotion_closure.cb = basic_do_promote; - up_args->promotion_closure.cb_arg = up_args; - - grpc_closure_list_add(&pollset->idle_jobs, &up_args->promotion_closure, 1); - grpc_pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); - -exit: - if (and_unlock_pollset) { - gpr_mu_unlock(&pollset->mu); - } -} - -static void basic_pollset_maybe_work_and_unlock(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, - grpc_pollset_worker *worker, - gpr_timespec deadline, - gpr_timespec now) { -#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR) -#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR) - - struct pollfd pfd[3]; - grpc_fd *fd; - grpc_fd_watcher fd_watcher; - int timeout; - int r; - nfds_t nfds; - - fd = pollset->data.ptr; - if (fd && grpc_fd_is_orphaned(fd)) { - GRPC_FD_UNREF(fd, "basicpoll"); - fd = pollset->data.ptr = NULL; - } - timeout = grpc_poll_deadline_to_millis_timeout(deadline, now); - pfd[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd); - pfd[0].events = POLLIN; - pfd[0].revents = 0; - pfd[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); - pfd[1].events = POLLIN; - pfd[1].revents = 0; - nfds = 2; - if (fd) { - pfd[2].fd = fd->fd; - pfd[2].revents = 0; - GRPC_FD_REF(fd, "basicpoll_begin"); - gpr_mu_unlock(&pollset->mu); - pfd[2].events = (short)grpc_fd_begin_poll(fd, pollset, worker, POLLIN, - POLLOUT, &fd_watcher); - if (pfd[2].events != 0) { - nfds++; - } - } else { - gpr_mu_unlock(&pollset->mu); - } - - /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid - even going into the blocking annotation if possible */ - /* poll fd count (argument 2) is shortened by one if we have no events - to poll on - such that it only includes the kicker */ - GPR_TIMER_BEGIN("poll", 0); - GRPC_SCHEDULING_START_BLOCKING_REGION; - r = grpc_poll_function(pfd, nfds, timeout); - GRPC_SCHEDULING_END_BLOCKING_REGION; - GPR_TIMER_END("poll", 0); - - if (r < 0) { - if (errno != EINTR) { - gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno)); - } - if (fd) { - grpc_fd_end_poll(exec_ctx, &fd_watcher, 0, 0); - } - } else if (r == 0) { - if (fd) { - grpc_fd_end_poll(exec_ctx, &fd_watcher, 0, 0); - } - } else { - if (pfd[0].revents & POLLIN_CHECK) { - grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd); - } - if (pfd[1].revents & POLLIN_CHECK) { - grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd); - } - if (nfds > 2) { - grpc_fd_end_poll(exec_ctx, &fd_watcher, pfd[2].revents & POLLIN_CHECK, - pfd[2].revents & POLLOUT_CHECK); - } else if (fd) { - grpc_fd_end_poll(exec_ctx, &fd_watcher, 0, 0); - } - } - - if (fd) { - GRPC_FD_UNREF(fd, "basicpoll_begin"); - } -} - -static void basic_pollset_destroy(grpc_pollset *pollset) { - if (pollset->data.ptr != NULL) { - GRPC_FD_UNREF(pollset->data.ptr, "basicpoll"); - pollset->data.ptr = NULL; - } -} - -static const grpc_pollset_vtable basic_pollset = { - basic_pollset_add_fd, basic_pollset_maybe_work_and_unlock, - basic_pollset_destroy, basic_pollset_destroy}; - -static void become_basic_pollset(grpc_pollset *pollset, grpc_fd *fd_or_null) { - pollset->vtable = &basic_pollset; - pollset->data.ptr = fd_or_null; - if (fd_or_null != NULL) { - GRPC_FD_REF(fd_or_null, "basicpoll"); - } -} - -#endif /* GPR_POSIX_POLLSET */ diff --git a/src/core/lib/iomgr/pollset_posix.h b/src/core/lib/iomgr/pollset_posix.h deleted file mode 100644 index f53ef8855c..0000000000 --- a/src/core/lib/iomgr/pollset_posix.h +++ /dev/null @@ -1,153 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_IOMGR_POLLSET_POSIX_H -#define GRPC_CORE_LIB_IOMGR_POLLSET_POSIX_H - -#include <poll.h> - -#include <grpc/support/sync.h> - -#include "src/core/lib/iomgr/exec_ctx.h" -#include "src/core/lib/iomgr/iomgr.h" -#include "src/core/lib/iomgr/pollset.h" -#include "src/core/lib/iomgr/wakeup_fd_posix.h" - -typedef struct grpc_pollset_vtable grpc_pollset_vtable; - -/* forward declare only in this file to avoid leaking impl details via - pollset.h; real users of grpc_fd should always include 'fd_posix.h' and not - use the struct tag */ -struct grpc_fd; - -typedef struct grpc_cached_wakeup_fd { - grpc_wakeup_fd fd; - struct grpc_cached_wakeup_fd *next; -} grpc_cached_wakeup_fd; - -struct grpc_pollset_worker { - grpc_cached_wakeup_fd *wakeup_fd; - int reevaluate_polling_on_wakeup; - int kicked_specifically; - struct grpc_pollset_worker *next; - struct grpc_pollset_worker *prev; -}; - -struct grpc_pollset { - /* pollsets under posix can mutate representation as fds are added and - removed. - For example, we may choose a poll() based implementation on linux for - few fds, and an epoll() based implementation for many fds */ - const grpc_pollset_vtable *vtable; - gpr_mu mu; - grpc_pollset_worker root_worker; - int in_flight_cbs; - int shutting_down; - int called_shutdown; - int kicked_without_pollers; - grpc_closure *shutdown_done; - grpc_closure_list idle_jobs; - union { - int fd; - void *ptr; - } data; - /* Local cache of eventfds for workers */ - grpc_cached_wakeup_fd *local_wakeup_cache; -}; - -struct grpc_pollset_vtable { - void (*add_fd)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - struct grpc_fd *fd, int and_unlock_pollset); - void (*maybe_work_and_unlock)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_pollset_worker *worker, - gpr_timespec deadline, gpr_timespec now); - void (*finish_shutdown)(grpc_pollset *pollset); - void (*destroy)(grpc_pollset *pollset); -}; - -/* Add an fd to a pollset */ -void grpc_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - struct grpc_fd *fd); - -/* Returns the fd to listen on for kicks */ -int grpc_kick_read_fd(grpc_pollset *p); -/* Call after polling has been kicked to leave the kicked state */ -void grpc_kick_drain(grpc_pollset *p); - -/* Convert a timespec to milliseconds: - - very small or negative poll times are clamped to zero to do a - non-blocking poll (which becomes spin polling) - - other small values are rounded up to one millisecond - - longer than a millisecond polls are rounded up to the next nearest - millisecond to avoid spinning - - infinite timeouts are converted to -1 */ -int grpc_poll_deadline_to_millis_timeout(gpr_timespec deadline, - gpr_timespec now); - -/* Allow kick to wakeup the currently polling worker */ -#define GRPC_POLLSET_CAN_KICK_SELF 1 -/* Force the wakee to repoll when awoken */ -#define GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP 2 -/* As per grpc_pollset_kick, with an extended set of flags (defined above) - -- mostly for fd_posix's use. */ -void grpc_pollset_kick_ext(grpc_pollset *p, - grpc_pollset_worker *specific_worker, - uint32_t flags); - -/* turn a pollset into a multipoller: platform specific */ -typedef void (*grpc_platform_become_multipoller_type)(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, - struct grpc_fd **fds, - size_t fd_count); -extern grpc_platform_become_multipoller_type grpc_platform_become_multipoller; - -void grpc_poll_become_multipoller(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, struct grpc_fd **fds, - size_t fd_count); - -/* Return 1 if the pollset has active threads in grpc_pollset_work (pollset must - * be locked) */ -int grpc_pollset_has_workers(grpc_pollset *pollset); - -void grpc_remove_fd_from_all_epoll_sets(int fd); - -/* override to allow tests to hook poll() usage */ -/* NOTE: Any changes to grpc_poll_function must take place when the gRPC - is certainly not doing any polling anywhere. - Otherwise, there might be a race between changing the variable and actually - doing a polling operation */ -typedef int (*grpc_poll_function_type)(struct pollfd *, nfds_t, int); -extern grpc_poll_function_type grpc_poll_function; -extern grpc_wakeup_fd grpc_global_wakeup_fd; - -#endif /* GRPC_CORE_LIB_IOMGR_POLLSET_POSIX_H */ diff --git a/src/core/lib/iomgr/pollset_set_posix.c b/src/core/lib/iomgr/pollset_set_posix.c deleted file mode 100644 index ef85075f38..0000000000 --- a/src/core/lib/iomgr/pollset_set_posix.c +++ /dev/null @@ -1,202 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/support/port_platform.h> - -#ifdef GPR_POSIX_SOCKET - -#include <stdlib.h> -#include <string.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/useful.h> - -#include "src/core/lib/iomgr/pollset_posix.h" -#include "src/core/lib/iomgr/pollset_set_posix.h" - -struct grpc_pollset_set { - gpr_mu mu; - - size_t pollset_count; - size_t pollset_capacity; - grpc_pollset **pollsets; - - size_t pollset_set_count; - size_t pollset_set_capacity; - struct grpc_pollset_set **pollset_sets; - - size_t fd_count; - size_t fd_capacity; - grpc_fd **fds; -}; - -grpc_pollset_set *grpc_pollset_set_create(void) { - grpc_pollset_set *pollset_set = gpr_malloc(sizeof(*pollset_set)); - memset(pollset_set, 0, sizeof(*pollset_set)); - gpr_mu_init(&pollset_set->mu); - return pollset_set; -} - -void grpc_pollset_set_destroy(grpc_pollset_set *pollset_set) { - size_t i; - gpr_mu_destroy(&pollset_set->mu); - for (i = 0; i < pollset_set->fd_count; i++) { - GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set"); - } - gpr_free(pollset_set->pollsets); - gpr_free(pollset_set->pollset_sets); - gpr_free(pollset_set->fds); - gpr_free(pollset_set); -} - -void grpc_pollset_set_add_pollset(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, - grpc_pollset *pollset) { - size_t i, j; - gpr_mu_lock(&pollset_set->mu); - if (pollset_set->pollset_count == pollset_set->pollset_capacity) { - pollset_set->pollset_capacity = - GPR_MAX(8, 2 * pollset_set->pollset_capacity); - pollset_set->pollsets = - gpr_realloc(pollset_set->pollsets, pollset_set->pollset_capacity * - sizeof(*pollset_set->pollsets)); - } - pollset_set->pollsets[pollset_set->pollset_count++] = pollset; - for (i = 0, j = 0; i < pollset_set->fd_count; i++) { - if (grpc_fd_is_orphaned(pollset_set->fds[i])) { - GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set"); - } else { - grpc_pollset_add_fd(exec_ctx, pollset, pollset_set->fds[i]); - pollset_set->fds[j++] = pollset_set->fds[i]; - } - } - pollset_set->fd_count = j; - gpr_mu_unlock(&pollset_set->mu); -} - -void grpc_pollset_set_del_pollset(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, - grpc_pollset *pollset) { - size_t i; - gpr_mu_lock(&pollset_set->mu); - for (i = 0; i < pollset_set->pollset_count; i++) { - if (pollset_set->pollsets[i] == pollset) { - pollset_set->pollset_count--; - GPR_SWAP(grpc_pollset *, pollset_set->pollsets[i], - pollset_set->pollsets[pollset_set->pollset_count]); - break; - } - } - gpr_mu_unlock(&pollset_set->mu); -} - -void grpc_pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *bag, - grpc_pollset_set *item) { - size_t i, j; - gpr_mu_lock(&bag->mu); - if (bag->pollset_set_count == bag->pollset_set_capacity) { - bag->pollset_set_capacity = GPR_MAX(8, 2 * bag->pollset_set_capacity); - bag->pollset_sets = - gpr_realloc(bag->pollset_sets, - bag->pollset_set_capacity * sizeof(*bag->pollset_sets)); - } - bag->pollset_sets[bag->pollset_set_count++] = item; - for (i = 0, j = 0; i < bag->fd_count; i++) { - if (grpc_fd_is_orphaned(bag->fds[i])) { - GRPC_FD_UNREF(bag->fds[i], "pollset_set"); - } else { - grpc_pollset_set_add_fd(exec_ctx, item, bag->fds[i]); - bag->fds[j++] = bag->fds[i]; - } - } - bag->fd_count = j; - gpr_mu_unlock(&bag->mu); -} - -void grpc_pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *bag, - grpc_pollset_set *item) { - size_t i; - gpr_mu_lock(&bag->mu); - for (i = 0; i < bag->pollset_set_count; i++) { - if (bag->pollset_sets[i] == item) { - bag->pollset_set_count--; - GPR_SWAP(grpc_pollset_set *, bag->pollset_sets[i], - bag->pollset_sets[bag->pollset_set_count]); - break; - } - } - gpr_mu_unlock(&bag->mu); -} - -void grpc_pollset_set_add_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, grpc_fd *fd) { - size_t i; - gpr_mu_lock(&pollset_set->mu); - if (pollset_set->fd_count == pollset_set->fd_capacity) { - pollset_set->fd_capacity = GPR_MAX(8, 2 * pollset_set->fd_capacity); - pollset_set->fds = gpr_realloc( - pollset_set->fds, pollset_set->fd_capacity * sizeof(*pollset_set->fds)); - } - GRPC_FD_REF(fd, "pollset_set"); - pollset_set->fds[pollset_set->fd_count++] = fd; - for (i = 0; i < pollset_set->pollset_count; i++) { - grpc_pollset_add_fd(exec_ctx, pollset_set->pollsets[i], fd); - } - for (i = 0; i < pollset_set->pollset_set_count; i++) { - grpc_pollset_set_add_fd(exec_ctx, pollset_set->pollset_sets[i], fd); - } - gpr_mu_unlock(&pollset_set->mu); -} - -void grpc_pollset_set_del_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, grpc_fd *fd) { - size_t i; - gpr_mu_lock(&pollset_set->mu); - for (i = 0; i < pollset_set->fd_count; i++) { - if (pollset_set->fds[i] == fd) { - pollset_set->fd_count--; - GPR_SWAP(grpc_fd *, pollset_set->fds[i], - pollset_set->fds[pollset_set->fd_count]); - GRPC_FD_UNREF(fd, "pollset_set"); - break; - } - } - for (i = 0; i < pollset_set->pollset_set_count; i++) { - grpc_pollset_set_del_fd(exec_ctx, pollset_set->pollset_sets[i], fd); - } - gpr_mu_unlock(&pollset_set->mu); -} - -#endif /* GPR_POSIX_SOCKET */ diff --git a/src/core/lib/iomgr/pollset_set_posix.h b/src/core/lib/iomgr/pollset_set_posix.h deleted file mode 100644 index e23e974921..0000000000 --- a/src/core/lib/iomgr/pollset_set_posix.h +++ /dev/null @@ -1,45 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_IOMGR_POLLSET_SET_POSIX_H -#define GRPC_CORE_LIB_IOMGR_POLLSET_SET_POSIX_H - -#include "src/core/lib/iomgr/fd_posix.h" -#include "src/core/lib/iomgr/pollset_set.h" - -void grpc_pollset_set_add_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, grpc_fd *fd); -void grpc_pollset_set_del_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, grpc_fd *fd); - -#endif /* GRPC_CORE_LIB_IOMGR_POLLSET_SET_POSIX_H */ diff --git a/src/core/lib/iomgr/pollset_windows.c b/src/core/lib/iomgr/pollset_windows.c index c2e7f647b5..bff5c586f8 100644 --- a/src/core/lib/iomgr/pollset_windows.c +++ b/src/core/lib/iomgr/pollset_windows.c @@ -47,7 +47,7 @@ gpr_mu grpc_polling_mu; static grpc_pollset_worker *g_active_poller; static grpc_pollset_worker g_global_root_worker; -void grpc_pollset_global_init() { +void grpc_pollset_global_init(void) { gpr_mu_init(&grpc_polling_mu); g_active_poller = NULL; g_global_root_worker.links[GRPC_POLLSET_WORKER_LINK_GLOBAL].next = @@ -55,7 +55,7 @@ void grpc_pollset_global_init() { &g_global_root_worker; } -void grpc_pollset_global_shutdown() { gpr_mu_destroy(&grpc_polling_mu); } +void grpc_pollset_global_shutdown(void) { gpr_mu_destroy(&grpc_polling_mu); } static void remove_worker(grpc_pollset_worker *worker, grpc_pollset_worker_link_type type) { diff --git a/src/core/lib/iomgr/pollset_windows.h b/src/core/lib/iomgr/pollset_windows.h index fef0e87b71..2642013afc 100644 --- a/src/core/lib/iomgr/pollset_windows.h +++ b/src/core/lib/iomgr/pollset_windows.h @@ -72,4 +72,7 @@ struct grpc_pollset { grpc_closure *on_shutdown; }; +void grpc_pollset_global_init(void); +void grpc_pollset_global_shutdown(void); + #endif /* GRPC_CORE_LIB_IOMGR_POLLSET_WINDOWS_H */ diff --git a/src/core/lib/iomgr/tcp_client_posix.c b/src/core/lib/iomgr/tcp_client_posix.c index 77718a5b5c..6430cb629f 100644 --- a/src/core/lib/iomgr/tcp_client_posix.c +++ b/src/core/lib/iomgr/tcp_client_posix.c @@ -47,9 +47,8 @@ #include <grpc/support/string_util.h> #include <grpc/support/time.h> +#include "src/core/lib/iomgr/ev_posix.h" #include "src/core/lib/iomgr/iomgr_posix.h" -#include "src/core/lib/iomgr/pollset_posix.h" -#include "src/core/lib/iomgr/pollset_set_posix.h" #include "src/core/lib/iomgr/sockaddr_utils.h" #include "src/core/lib/iomgr/socket_utils_posix.h" #include "src/core/lib/iomgr/tcp_posix.h" @@ -140,7 +139,8 @@ static void on_writable(grpc_exec_ctx *exec_ctx, void *acp, bool success) { if (success) { do { so_error_size = sizeof(so_error); - err = getsockopt(fd->fd, SOL_SOCKET, SO_ERROR, &so_error, &so_error_size); + err = getsockopt(grpc_fd_wrapped_fd(fd), SOL_SOCKET, SO_ERROR, &so_error, + &so_error_size); } while (err < 0 && errno == EINTR); if (err < 0) { gpr_log(GPR_ERROR, "failed to connect to '%s': getsockopt(ERROR): %s", diff --git a/src/core/lib/iomgr/tcp_posix.c b/src/core/lib/iomgr/tcp_posix.c index 33ec29a411..7210aef5d5 100644 --- a/src/core/lib/iomgr/tcp_posix.c +++ b/src/core/lib/iomgr/tcp_posix.c @@ -52,8 +52,7 @@ #include <grpc/support/time.h> #include "src/core/lib/debug/trace.h" -#include "src/core/lib/iomgr/pollset_posix.h" -#include "src/core/lib/iomgr/pollset_set_posix.h" +#include "src/core/lib/iomgr/ev_posix.h" #include "src/core/lib/profiling/timers.h" #include "src/core/lib/support/string.h" @@ -454,7 +453,7 @@ grpc_endpoint *grpc_tcp_create(grpc_fd *em_fd, size_t slice_size, grpc_tcp *tcp = (grpc_tcp *)gpr_malloc(sizeof(grpc_tcp)); tcp->base.vtable = &vtable; tcp->peer_string = gpr_strdup(peer_string); - tcp->fd = em_fd->fd; + tcp->fd = grpc_fd_wrapped_fd(em_fd); tcp->read_cb = NULL; tcp->write_cb = NULL; tcp->release_fd_cb = NULL; diff --git a/src/core/lib/iomgr/tcp_posix.h b/src/core/lib/iomgr/tcp_posix.h index acce56467a..99125836d6 100644 --- a/src/core/lib/iomgr/tcp_posix.h +++ b/src/core/lib/iomgr/tcp_posix.h @@ -45,7 +45,7 @@ */ #include "src/core/lib/iomgr/endpoint.h" -#include "src/core/lib/iomgr/fd_posix.h" +#include "src/core/lib/iomgr/ev_posix.h" #define GRPC_TCP_DEFAULT_READ_SLICE_SIZE 8192 diff --git a/src/core/lib/iomgr/tcp_server_posix.c b/src/core/lib/iomgr/tcp_server_posix.c index 76aeb1c2f0..cfb5251684 100644 --- a/src/core/lib/iomgr/tcp_server_posix.c +++ b/src/core/lib/iomgr/tcp_server_posix.c @@ -59,7 +59,6 @@ #include <grpc/support/string_util.h> #include <grpc/support/sync.h> #include <grpc/support/time.h> -#include "src/core/lib/iomgr/pollset_posix.h" #include "src/core/lib/iomgr/resolve_address.h" #include "src/core/lib/iomgr/sockaddr_utils.h" #include "src/core/lib/iomgr/socket_utils_posix.h" diff --git a/src/core/lib/iomgr/udp_server.c b/src/core/lib/iomgr/udp_server.c index 2d8220a9b1..a0b9709be5 100644 --- a/src/core/lib/iomgr/udp_server.c +++ b/src/core/lib/iomgr/udp_server.c @@ -60,12 +60,10 @@ #include <grpc/support/string_util.h> #include <grpc/support/sync.h> #include <grpc/support/time.h> -#include "src/core/lib/iomgr/fd_posix.h" -#include "src/core/lib/iomgr/pollset_posix.h" +#include "src/core/lib/iomgr/ev_posix.h" #include "src/core/lib/iomgr/resolve_address.h" #include "src/core/lib/iomgr/sockaddr_utils.h" #include "src/core/lib/iomgr/socket_utils_posix.h" -#include "src/core/lib/iomgr/unix_sockets_posix.h" #include "src/core/lib/support/string.h" #define INIT_PORT_CAP 2 diff --git a/src/core/lib/iomgr/udp_server.h b/src/core/lib/iomgr/udp_server.h index b0ca15101f..d8cf957a22 100644 --- a/src/core/lib/iomgr/udp_server.h +++ b/src/core/lib/iomgr/udp_server.h @@ -35,7 +35,7 @@ #define GRPC_CORE_LIB_IOMGR_UDP_SERVER_H #include "src/core/lib/iomgr/endpoint.h" -#include "src/core/lib/iomgr/fd_posix.h" +#include "src/core/lib/iomgr/ev_posix.h" /* Forward decl of struct grpc_server */ /* This is not typedef'ed to avoid a typedef-redefinition error */ diff --git a/src/core/lib/iomgr/workqueue_posix.c b/src/core/lib/iomgr/workqueue_posix.c index 9d68d9e16a..80e7a0b206 100644 --- a/src/core/lib/iomgr/workqueue_posix.c +++ b/src/core/lib/iomgr/workqueue_posix.c @@ -43,8 +43,7 @@ #include <grpc/support/log.h> #include <grpc/support/useful.h> -#include "src/core/lib/iomgr/fd_posix.h" -#include "src/core/lib/iomgr/pollset_posix.h" +#include "src/core/lib/iomgr/ev_posix.h" static void on_readable(grpc_exec_ctx *exec_ctx, void *arg, bool success); diff --git a/src/core/lib/statistics/census_init.c b/src/core/lib/statistics/census_init.c deleted file mode 100644 index 7a8b8b660a..0000000000 --- a/src/core/lib/statistics/census_init.c +++ /dev/null @@ -1,48 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include "src/core/lib/statistics/census_interface.h" - -#include <grpc/support/log.h> -#include "src/core/lib/statistics/census_rpc_stats.h" -#include "src/core/lib/statistics/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/lib/statistics/census_interface.h b/src/core/lib/statistics/census_interface.h deleted file mode 100644 index 4dce81b0c7..0000000000 --- a/src/core/lib/statistics/census_interface.h +++ /dev/null @@ -1,76 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_STATISTICS_CENSUS_INTERFACE_H -#define GRPC_CORE_LIB_STATISTICS_CENSUS_INTERFACE_H - -#include <grpc/support/port_platform.h> - -/* Maximum length of an individual census trace annotation. */ -#define CENSUS_MAX_ANNOTATION_LENGTH 200 - -/* 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); - -#endif /* GRPC_CORE_LIB_STATISTICS_CENSUS_INTERFACE_H */ diff --git a/src/core/lib/statistics/census_log.c b/src/core/lib/statistics/census_log.c deleted file mode 100644 index 782b34e3a1..0000000000 --- a/src/core/lib/statistics/census_log.c +++ /dev/null @@ -1,603 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -/* 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/lib/statistics/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/lib/statistics/census_log.h b/src/core/lib/statistics/census_log.h deleted file mode 100644 index 6efcef49a4..0000000000 --- a/src/core/lib/statistics/census_log.h +++ /dev/null @@ -1,91 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_STATISTICS_CENSUS_LOG_H -#define GRPC_CORE_LIB_STATISTICS_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) - -/* 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 gprc_stats_log_start_write() failed due to - out-of-space. */ -int census_log_out_of_space_count(void); - -#endif /* GRPC_CORE_LIB_STATISTICS_CENSUS_LOG_H */ diff --git a/src/core/lib/statistics/census_rpc_stats.c b/src/core/lib/statistics/census_rpc_stats.c deleted file mode 100644 index 33522e0f40..0000000000 --- a/src/core/lib/statistics/census_rpc_stats.c +++ /dev/null @@ -1,253 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <string.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/sync.h> -#include "src/core/lib/statistics/census_interface.h" -#include "src/core/lib/statistics/census_rpc_stats.h" -#include "src/core/lib/statistics/census_tracing.h" -#include "src/core/lib/statistics/hash_table.h" -#include "src/core/lib/statistics/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/lib/statistics/census_rpc_stats.h b/src/core/lib/statistics/census_rpc_stats.h deleted file mode 100644 index 3853642362..0000000000 --- a/src/core/lib/statistics/census_rpc_stats.h +++ /dev/null @@ -1,101 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_STATISTICS_CENSUS_RPC_STATS_H -#define GRPC_CORE_LIB_STATISTICS_CENSUS_RPC_STATS_H - -#include <grpc/support/port_platform.h> -#include "src/core/lib/statistics/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_LIB_STATISTICS_CENSUS_RPC_STATS_H */ diff --git a/src/core/lib/statistics/census_tracing.c b/src/core/lib/statistics/census_tracing.c deleted file mode 100644 index ff71e796ad..0000000000 --- a/src/core/lib/statistics/census_tracing.c +++ /dev/null @@ -1,241 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include "src/core/lib/statistics/census_tracing.h" -#include "src/core/lib/statistics/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/lib/statistics/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/lib/statistics/census_tracing.h b/src/core/lib/statistics/census_tracing.h deleted file mode 100644 index 7b2db1e008..0000000000 --- a/src/core/lib/statistics/census_tracing.h +++ /dev/null @@ -1,96 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_STATISTICS_CENSUS_TRACING_H -#define GRPC_CORE_LIB_STATISTICS_CENSUS_TRACING_H - -#include <grpc/support/time.h> -#include "src/core/lib/statistics/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_LIB_STATISTICS_CENSUS_TRACING_H */ diff --git a/src/core/lib/statistics/hash_table.c b/src/core/lib/statistics/hash_table.c deleted file mode 100644 index 79481c3eb1..0000000000 --- a/src/core/lib/statistics/hash_table.c +++ /dev/null @@ -1,303 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include "src/core/lib/statistics/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/lib/statistics/hash_table.h b/src/core/lib/statistics/hash_table.h deleted file mode 100644 index 45984a1c41..0000000000 --- a/src/core/lib/statistics/hash_table.h +++ /dev/null @@ -1,131 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_STATISTICS_HASH_TABLE_H -#define GRPC_CORE_LIB_STATISTICS_HASH_TABLE_H - -#include <stddef.h> - -#include <grpc/support/port_platform.h> - -/* 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); - -#endif /* GRPC_CORE_LIB_STATISTICS_HASH_TABLE_H */ diff --git a/src/core/lib/statistics/window_stats.c b/src/core/lib/statistics/window_stats.c deleted file mode 100644 index 9185e98d60..0000000000 --- a/src/core/lib/statistics/window_stats.c +++ /dev/null @@ -1,316 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include "src/core/lib/statistics/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/lib/statistics/window_stats.h b/src/core/lib/statistics/window_stats.h deleted file mode 100644 index 629c08b8d5..0000000000 --- a/src/core/lib/statistics/window_stats.h +++ /dev/null @@ -1,173 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_STATISTICS_WINDOW_STATS_H -#define GRPC_CORE_LIB_STATISTICS_WINDOW_STATS_H - -#include <grpc/support/time.h> - -/* 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); - -#endif /* GRPC_CORE_LIB_STATISTICS_WINDOW_STATS_H */ diff --git a/src/core/lib/surface/init.c b/src/core/lib/surface/init.c index 566b030912..aca4ce9d07 100644 --- a/src/core/lib/surface/init.c +++ b/src/core/lib/surface/init.c @@ -39,9 +39,7 @@ #include <grpc/grpc.h> #include <grpc/support/alloc.h> #include <grpc/support/time.h> -/* TODO(ctiller): find another way? - better not to include census here */ #include "src/core/ext/transport/chttp2/transport/chttp2_transport.h" -#include "src/core/lib/census/grpc_plugin.h" #include "src/core/lib/channel/channel_stack.h" #include "src/core/lib/channel/client_channel.h" #include "src/core/lib/channel/compress_filter.h" @@ -50,8 +48,6 @@ #include "src/core/lib/channel/http_server_filter.h" #include "src/core/lib/client_config/lb_policy_registry.h" #include "src/core/lib/client_config/resolver_registry.h" -#include "src/core/lib/client_config/resolvers/dns_resolver.h" -#include "src/core/lib/client_config/resolvers/sockaddr_resolver.h" #include "src/core/lib/client_config/subchannel.h" #include "src/core/lib/client_config/subchannel_index.h" #include "src/core/lib/debug/trace.h" @@ -85,8 +81,6 @@ static int g_initializations; static void do_basic_init(void) { gpr_mu_init(&g_init_mu); grpc_register_built_in_plugins(); - /* TODO(ctiller): ideally remove this strict linkage */ - grpc_register_plugin(census_grpc_plugin_init, census_grpc_plugin_destroy); g_initializations = 0; } @@ -169,12 +163,6 @@ void grpc_init(void) { grpc_channel_init_init(); grpc_lb_policy_registry_init(); grpc_resolver_registry_init(GRPC_DEFAULT_NAME_PREFIX); - grpc_register_resolver_type(grpc_dns_resolver_factory_create()); - grpc_register_resolver_type(grpc_ipv4_resolver_factory_create()); - grpc_register_resolver_type(grpc_ipv6_resolver_factory_create()); -#ifdef GPR_HAVE_UNIX_SOCKET - grpc_register_resolver_type(grpc_unix_resolver_factory_create()); -#endif grpc_register_tracer("api", &grpc_api_trace); grpc_register_tracer("channel", &grpc_trace_channel); grpc_register_tracer("http", &grpc_http_trace); |