diff options
Diffstat (limited to 'src/core')
36 files changed, 2316 insertions, 92 deletions
diff --git a/src/core/ext/census/grpc_filter.c b/src/core/ext/census/grpc_filter.c index 72e4e5427e..f51d850e01 100644 --- a/src/core/ext/census/grpc_filter.c +++ b/src/core/ext/census/grpc_filter.c @@ -45,6 +45,7 @@ #include "src/core/ext/census/census_interface.h" #include "src/core/ext/census/census_rpc_stats.h" #include "src/core/lib/channel/channel_stack.h" +#include "src/core/lib/profiling/timers.h" #include "src/core/lib/transport/static_metadata.h" typedef struct call_data { @@ -92,6 +93,7 @@ static void client_start_transport_op(grpc_exec_ctx *exec_ctx, static void server_on_done_recv(grpc_exec_ctx *exec_ctx, void *ptr, grpc_error *error) { + GPR_TIMER_BEGIN("census-server:server_on_done_recv", 0); grpc_call_element *elem = ptr; call_data *calld = elem->call_data; channel_data *chand = elem->channel_data; @@ -99,6 +101,7 @@ static void server_on_done_recv(grpc_exec_ctx *exec_ctx, void *ptr, extract_and_annotate_method_tag(calld->recv_initial_metadata, calld, chand); } calld->on_done_recv->cb(exec_ctx, calld->on_done_recv->cb_arg, error); + GPR_TIMER_END("census-server:server_on_done_recv", 0); } static void server_mutate_op(grpc_call_element *elem, diff --git a/src/core/ext/client_config/channel_connectivity.c b/src/core/ext/client_config/channel_connectivity.c index 20c01a9a7c..c1220e3a8c 100644 --- a/src/core/ext/client_config/channel_connectivity.c +++ b/src/core/ext/client_config/channel_connectivity.c @@ -189,10 +189,11 @@ void grpc_channel_watch_connectivity_state( GRPC_API_TRACE( "grpc_channel_watch_connectivity_state(" "channel=%p, last_observed_state=%d, " - "deadline=gpr_timespec { tv_sec: %lld, tv_nsec: %d, clock_type: %d }, " + "deadline=gpr_timespec { tv_sec: %" PRId64 + ", tv_nsec: %d, clock_type: %d }, " "cq=%p, tag=%p)", - 7, (channel, (int)last_observed_state, (long long)deadline.tv_sec, - (int)deadline.tv_nsec, (int)deadline.clock_type, cq, tag)); + 7, (channel, (int)last_observed_state, deadline.tv_sec, deadline.tv_nsec, + (int)deadline.clock_type, cq, tag)); grpc_cq_begin_op(cq, tag); diff --git a/src/core/ext/client_config/client_channel.c b/src/core/ext/client_config/client_channel.c index 1d5a7d5224..a096435c98 100644 --- a/src/core/ext/client_config/client_channel.c +++ b/src/core/ext/client_config/client_channel.c @@ -367,6 +367,8 @@ static int cc_pick_subchannel(grpc_exec_ctx *exec_ctx, void *elemp, uint32_t initial_metadata_flags, grpc_connected_subchannel **connected_subchannel, grpc_closure *on_ready) { + GPR_TIMER_BEGIN("cc_pick_subchannel", 0); + grpc_call_element *elem = elemp; channel_data *chand = elem->channel_data; call_data *calld = elem->call_data; @@ -391,6 +393,7 @@ static int cc_pick_subchannel(grpc_exec_ctx *exec_ctx, void *elemp, } } gpr_mu_unlock(&chand->mu_config); + GPR_TIMER_END("cc_pick_subchannel", 0); return 1; } if (chand->lb_policy != NULL) { @@ -402,6 +405,7 @@ static int cc_pick_subchannel(grpc_exec_ctx *exec_ctx, void *elemp, initial_metadata, initial_metadata_flags, connected_subchannel, on_ready); GRPC_LB_POLICY_UNREF(exec_ctx, lb_policy, "cc_pick_subchannel"); + GPR_TIMER_END("cc_pick_subchannel", 0); return r; } if (chand->resolver != NULL && !chand->started_resolving) { @@ -426,6 +430,8 @@ static int cc_pick_subchannel(grpc_exec_ctx *exec_ctx, void *elemp, NULL); } gpr_mu_unlock(&chand->mu_config); + + GPR_TIMER_END("cc_pick_subchannel", 0); return 0; } diff --git a/src/core/ext/client_config/subchannel.c b/src/core/ext/client_config/subchannel.c index 468067ea57..d089cd4399 100644 --- a/src/core/ext/client_config/subchannel.c +++ b/src/core/ext/client_config/subchannel.c @@ -690,9 +690,11 @@ char *grpc_subchannel_call_get_peer(grpc_exec_ctx *exec_ctx, void grpc_subchannel_call_process_op(grpc_exec_ctx *exec_ctx, grpc_subchannel_call *call, grpc_transport_stream_op *op) { + GPR_TIMER_BEGIN("grpc_subchannel_call_process_op", 0); grpc_call_stack *call_stack = SUBCHANNEL_CALL_TO_CALL_STACK(call); grpc_call_element *top_elem = grpc_call_stack_element(call_stack, 0); top_elem->filter->start_transport_stream_op(exec_ctx, top_elem, op); + GPR_TIMER_END("grpc_subchannel_call_process_op", 0); } grpc_connected_subchannel *grpc_subchannel_get_connected_subchannel( diff --git a/src/core/ext/transport/chttp2/server/insecure/server_chttp2.c b/src/core/ext/transport/chttp2/server/insecure/server_chttp2.c index 9bae3a94f9..e5c987925c 100644 --- a/src/core/ext/transport/chttp2/server/insecure/server_chttp2.c +++ b/src/core/ext/transport/chttp2/server/insecure/server_chttp2.c @@ -97,7 +97,8 @@ int grpc_server_add_insecure_http2_port(grpc_server *server, const char *addr) { goto error; } - err = grpc_tcp_server_create(NULL, &tcp); + err = + grpc_tcp_server_create(NULL, grpc_server_get_channel_args(server), &tcp); if (err != GRPC_ERROR_NONE) { goto error; } diff --git a/src/core/ext/transport/chttp2/server/secure/server_secure_chttp2.c b/src/core/ext/transport/chttp2/server/secure/server_secure_chttp2.c index ead8a4d566..c42810e913 100644 --- a/src/core/ext/transport/chttp2/server/secure/server_secure_chttp2.c +++ b/src/core/ext/transport/chttp2/server/secure/server_secure_chttp2.c @@ -216,7 +216,8 @@ int grpc_server_add_secure_http2_port(grpc_server *server, const char *addr, state = gpr_malloc(sizeof(*state)); memset(state, 0, sizeof(*state)); grpc_closure_init(&state->destroy_closure, destroy_done, state); - err = grpc_tcp_server_create(&state->destroy_closure, &tcp); + err = grpc_tcp_server_create(&state->destroy_closure, + grpc_server_get_channel_args(server), &tcp); if (err != GRPC_ERROR_NONE) { goto error; } diff --git a/src/core/ext/transport/chttp2/transport/chttp2_transport.c b/src/core/ext/transport/chttp2/transport/chttp2_transport.c index 9c99ff883a..b710bbcb3c 100644 --- a/src/core/ext/transport/chttp2/transport/chttp2_transport.c +++ b/src/core/ext/transport/chttp2/transport/chttp2_transport.c @@ -642,6 +642,8 @@ static void finish_global_actions(grpc_exec_ctx *exec_ctx, grpc_chttp2_executor_action_header *hdr; grpc_chttp2_executor_action_header *next; + GPR_TIMER_BEGIN("finish_global_actions", 0); + for (;;) { if (!t->executor.writing_active && !t->closed && grpc_chttp2_unlocking_check_writes(exec_ctx, &t->global, &t->writing)) { @@ -659,7 +661,9 @@ static void finish_global_actions(grpc_exec_ctx *exec_ctx, NULL; gpr_mu_unlock(&t->executor.mu); while (hdr != NULL) { + GPR_TIMER_BEGIN("chttp2:locked_action", 0); hdr->action(exec_ctx, t, hdr->stream, hdr->arg); + GPR_TIMER_END("chttp2:locked_action", 0); next = hdr->next; gpr_free(hdr); UNREF_TRANSPORT(exec_ctx, t, "pending_action"); @@ -672,6 +676,8 @@ static void finish_global_actions(grpc_exec_ctx *exec_ctx, gpr_mu_unlock(&t->executor.mu); break; } + + GPR_TIMER_END("finish_global_actions", 0); } void grpc_chttp2_run_with_global_lock(grpc_exec_ctx *exec_ctx, @@ -681,6 +687,8 @@ void grpc_chttp2_run_with_global_lock(grpc_exec_ctx *exec_ctx, void *arg, size_t sizeof_arg) { grpc_chttp2_executor_action_header *hdr; + GPR_TIMER_BEGIN("grpc_chttp2_run_with_global_lock", 0); + REF_TRANSPORT(t, "run_global"); gpr_mu_lock(&t->executor.mu); @@ -689,7 +697,9 @@ void grpc_chttp2_run_with_global_lock(grpc_exec_ctx *exec_ctx, t->executor.global_active = 1; gpr_mu_unlock(&t->executor.mu); + GPR_TIMER_BEGIN("chttp2:locked_action", 0); action(exec_ctx, t, optional_stream, arg); + GPR_TIMER_END("chttp2:locked_action", 0); finish_global_actions(exec_ctx, t); } else { @@ -726,6 +736,8 @@ void grpc_chttp2_run_with_global_lock(grpc_exec_ctx *exec_ctx, } UNREF_TRANSPORT(exec_ctx, t, "run_global"); + + GPR_TIMER_END("grpc_chttp2_run_with_global_lock", 0); } /******************************************************************************* @@ -1557,17 +1569,15 @@ static void close_from_api(grpc_exec_ctx *exec_ctx, gpr_slice_buffer_add(&transport_global->qbuf, gpr_slice_ref(*optional_message)); } - gpr_slice_buffer_add( &transport_global->qbuf, grpc_chttp2_rst_stream_create(stream_global->id, GRPC_CHTTP2_NO_ERROR, &stream_global->stats.outgoing)); - - if (optional_message) { - gpr_slice_ref(*optional_message); - } } + if (optional_message) { + gpr_slice_ref(*optional_message); + } grpc_chttp2_fake_status(exec_ctx, transport_global, stream_global, status, optional_message); grpc_error *err = GRPC_ERROR_CREATE("Stream closed"); diff --git a/src/core/lib/iomgr/error.c b/src/core/lib/iomgr/error.c index 540fb4fa7e..6f24200b77 100644 --- a/src/core/lib/iomgr/error.c +++ b/src/core/lib/iomgr/error.c @@ -47,6 +47,8 @@ #include <grpc/support/log_windows.h> #endif +#include "src/core/lib/profiling/timers.h" + static void destroy_integer(void *key) {} static void *copy_integer(void *key) { return key; } @@ -213,6 +215,7 @@ void grpc_error_unref(grpc_error *err) { grpc_error *grpc_error_create(const char *file, int line, const char *desc, grpc_error **referencing, size_t num_referencing) { + GPR_TIMER_BEGIN("grpc_error_create", 0); grpc_error *err = gpr_malloc(sizeof(*err)); if (err == NULL) { // TODO(ctiller): make gpr_malloc return NULL return GRPC_ERROR_OOM; @@ -238,57 +241,71 @@ grpc_error *grpc_error_create(const char *file, int line, const char *desc, (void *)(uintptr_t)GRPC_ERROR_TIME_CREATED, box_time(gpr_now(GPR_CLOCK_REALTIME))); gpr_ref_init(&err->refs, 1); + GPR_TIMER_END("grpc_error_create", 0); return err; } static grpc_error *copy_error_and_unref(grpc_error *in) { + GPR_TIMER_BEGIN("copy_error_and_unref", 0); + grpc_error *out; if (is_special(in)) { if (in == GRPC_ERROR_NONE) return GRPC_ERROR_CREATE("no error"); if (in == GRPC_ERROR_OOM) return GRPC_ERROR_CREATE("oom"); if (in == GRPC_ERROR_CANCELLED) return GRPC_ERROR_CREATE("cancelled"); - return GRPC_ERROR_CREATE("unknown"); - } - grpc_error *out = gpr_malloc(sizeof(*out)); + out = GRPC_ERROR_CREATE("unknown"); + } else { + out = gpr_malloc(sizeof(*out)); #ifdef GRPC_ERROR_REFCOUNT_DEBUG - gpr_log(GPR_DEBUG, "%p create copying", out); + gpr_log(GPR_DEBUG, "%p create copying", out); #endif - out->ints = gpr_avl_ref(in->ints); - out->strs = gpr_avl_ref(in->strs); - out->errs = gpr_avl_ref(in->errs); - out->times = gpr_avl_ref(in->times); - out->next_err = in->next_err; - gpr_ref_init(&out->refs, 1); - GRPC_ERROR_UNREF(in); + out->ints = gpr_avl_ref(in->ints); + out->strs = gpr_avl_ref(in->strs); + out->errs = gpr_avl_ref(in->errs); + out->times = gpr_avl_ref(in->times); + out->next_err = in->next_err; + gpr_ref_init(&out->refs, 1); + GRPC_ERROR_UNREF(in); + } + GPR_TIMER_END("copy_error_and_unref", 0); return out; } grpc_error *grpc_error_set_int(grpc_error *src, grpc_error_ints which, intptr_t value) { + GPR_TIMER_BEGIN("grpc_error_set_int", 0); grpc_error *new = copy_error_and_unref(src); new->ints = gpr_avl_add(new->ints, (void *)(uintptr_t)which, (void *)value); + GPR_TIMER_END("grpc_error_set_int", 0); return new; } bool grpc_error_get_int(grpc_error *err, grpc_error_ints which, intptr_t *p) { void *pp; + GPR_TIMER_BEGIN("grpc_error_get_int", 0); if (gpr_avl_maybe_get(err->ints, (void *)(uintptr_t)which, &pp)) { if (p != NULL) *p = (intptr_t)pp; + GPR_TIMER_END("grpc_error_get_int", 0); return true; } + GPR_TIMER_END("grpc_error_get_int", 0); return false; } grpc_error *grpc_error_set_str(grpc_error *src, grpc_error_strs which, const char *value) { + GPR_TIMER_BEGIN("grpc_error_set_str", 0); grpc_error *new = copy_error_and_unref(src); new->strs = gpr_avl_add(new->strs, (void *)(uintptr_t)which, gpr_strdup(value)); + GPR_TIMER_END("grpc_error_set_str", 0); return new; } grpc_error *grpc_error_add_child(grpc_error *src, grpc_error *child) { + GPR_TIMER_BEGIN("grpc_error_add_child", 0); grpc_error *new = copy_error_and_unref(src); new->errs = gpr_avl_add(new->errs, (void *)(new->next_err++), child); + GPR_TIMER_END("grpc_error_add_child", 0); return new; } @@ -480,6 +497,7 @@ void grpc_error_free_string(const char *str) { } const char *grpc_error_string(grpc_error *err) { + GPR_TIMER_BEGIN("grpc_error_string", 0); if (err == GRPC_ERROR_NONE) return no_error_string; if (err == GRPC_ERROR_OOM) return oom_error_string; if (err == GRPC_ERROR_CANCELLED) return cancelled_error_string; @@ -496,7 +514,9 @@ const char *grpc_error_string(grpc_error *err) { qsort(kvs.kvs, kvs.num_kvs, sizeof(kv_pair), cmp_kvs); - return finish_kvs(&kvs); + const char *out = finish_kvs(&kvs); + GPR_TIMER_END("grpc_error_string", 0); + return out; } grpc_error *grpc_os_error(const char *file, int line, int err, diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c new file mode 100644 index 0000000000..475bd3c1df --- /dev/null +++ b/src/core/lib/iomgr/ev_epoll_linux.c @@ -0,0 +1,1818 @@ +/* + * + * 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/grpc_posix.h> +#include <grpc/support/port_platform.h> + +/* This polling engine is only relevant on linux kernels supporting epoll() */ +#ifdef GPR_LINUX_EPOLL + +#include "src/core/lib/iomgr/ev_epoll_linux.h" + +#include <assert.h> +#include <errno.h> +#include <poll.h> +#include <signal.h> +#include <string.h> +#include <sys/epoll.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/ev_posix.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" + +static int grpc_wakeup_signal = -1; +static bool is_grpc_wakeup_signal_initialized = false; + +/* Implements the function defined in grpc_posix.h. This function might be + * called before even calling grpc_init() to set either a different signal to + * use. If signum == -1, then the use of signals is disabled */ +void grpc_use_signal(int signum) { + grpc_wakeup_signal = signum; + is_grpc_wakeup_signal_initialized = true; + + if (grpc_wakeup_signal < 0) { + gpr_log(GPR_INFO, + "Use of signals is disabled. Epoll engine will not be used"); + } else { + gpr_log(GPR_INFO, "epoll engine will be using signal: %d", + grpc_wakeup_signal); + } +} + +struct polling_island; + +/******************************************************************************* + * Fd Declarations + */ +struct grpc_fd { + int fd; + /* refst format: + bit 0 : 1=Active / 0=Orphaned + bits 1-n : refcount + Ref/Unref by two to avoid altering the orphaned bit */ + gpr_atm refst; + + gpr_mu mu; + + /* Indicates that the fd is shutdown and that any pending read/write closures + should fail */ + bool shutdown; + + /* The fd is either closed or we relinquished control of it. In either cases, + this indicates that the 'fd' on this structure is no longer valid */ + bool orphaned; + + /* TODO: sreek - Move this to a lockfree implementation */ + grpc_closure *read_closure; + grpc_closure *write_closure; + + /* The polling island to which this fd belongs to and the mutex protecting the + the field */ + gpr_mu pi_mu; + struct polling_island *polling_island; + + struct grpc_fd *freelist_next; + grpc_closure *on_done_closure; + + /* The pollset that last noticed that the fd is readable */ + grpc_pollset *read_notifier_pollset; + + grpc_iomgr_object iomgr_object; +}; + +/* 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) + +/******************************************************************************* + * Polling island Declarations + */ + +// #define GRPC_PI_REF_COUNT_DEBUG +#ifdef GRPC_PI_REF_COUNT_DEBUG + +#define PI_ADD_REF(p, r) pi_add_ref_dbg((p), (r), __FILE__, __LINE__) +#define PI_UNREF(p, r) pi_unref_dbg((p), (r), __FILE__, __LINE__) + +#else /* defined(GRPC_PI_REF_COUNT_DEBUG) */ + +#define PI_ADD_REF(p, r) pi_add_ref((p)) +#define PI_UNREF(p, r) pi_unref((p)) + +#endif /* !defined(GPRC_PI_REF_COUNT_DEBUG) */ + +typedef struct polling_island { + gpr_mu mu; + /* Ref count. Use PI_ADD_REF() and PI_UNREF() macros to increment/decrement + the refcount. + Once the ref count becomes zero, this structure is destroyed which means + we should ensure that there is never a scenario where a PI_ADD_REF() is + racing with a PI_UNREF() that just made the ref_count zero. */ + gpr_refcount ref_count; + + /* Pointer to the polling_island this merged into. + * merged_to value is only set once in polling_island's lifetime (and that too + * only if the island is merged with another island). Because of this, we can + * use gpr_atm type here so that we can do atomic access on this and reduce + * lock contention on 'mu' mutex. + * + * Note that if this field is not NULL (i.e not 0), all the remaining fields + * (except mu and ref_count) are invalid and must be ignored. */ + gpr_atm merged_to; + + /* The fd of the underlying epoll set */ + int epoll_fd; + + /* The file descriptors in the epoll set */ + size_t fd_cnt; + size_t fd_capacity; + grpc_fd **fds; + + /* Polling islands that are no longer needed are kept in a freelist so that + they can be reused. This field points to the next polling island in the + free list */ + struct polling_island *next_free; +} polling_island; + +/******************************************************************************* + * Pollset Declarations + */ +struct grpc_pollset_worker { + pthread_t pt_id; /* Thread id of this worker */ + struct grpc_pollset_worker *next; + struct grpc_pollset_worker *prev; +}; + +struct grpc_pollset { + gpr_mu mu; + grpc_pollset_worker root_worker; + bool kicked_without_pollers; + + bool shutting_down; /* Is the pollset shutting down ? */ + bool finish_shutdown_called; /* Is the 'finish_shutdown_locked()' called ? */ + grpc_closure *shutdown_done; /* Called after after shutdown is complete */ + + /* The polling island to which this pollset belongs to */ + struct polling_island *polling_island; +}; + +/******************************************************************************* + * Pollset-set Declarations + */ +/* TODO: sreek - Change the pollset_set implementation such that a pollset_set + * directly points to a polling_island (and adding an fd/pollset/pollset_set to + * the current pollset_set would result in polling island merges. This would + * remove the need to maintain fd_count here. This will also significantly + * simplify the grpc_fd structure since we would no longer need to explicitly + * maintain the orphaned state */ +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; +}; + +/******************************************************************************* + * Common helpers + */ + +static void append_error(grpc_error **composite, grpc_error *error, + const char *desc) { + if (error == GRPC_ERROR_NONE) return; + if (*composite == GRPC_ERROR_NONE) { + *composite = GRPC_ERROR_CREATE(desc); + } + *composite = grpc_error_add_child(*composite, error); +} + +/******************************************************************************* + * Polling island Definitions + */ + +/* The wakeup fd that is used to wake up all threads in a Polling island. This + is useful in the polling island merge operation where we need to wakeup all + the threads currently polling the smaller polling island (so that they can + start polling the new/merged polling island) + + NOTE: This fd is initialized to be readable and MUST NOT be consumed i.e the + threads that woke up MUST NOT call grpc_wakeup_fd_consume_wakeup() */ +static grpc_wakeup_fd polling_island_wakeup_fd; + +/* Polling island freelist */ +static gpr_mu g_pi_freelist_mu; +static polling_island *g_pi_freelist = NULL; + +static void polling_island_delete(); /* Forward declaration */ + +#ifdef GRPC_TSAN +/* Currently TSAN may incorrectly flag data races between epoll_ctl and + epoll_wait for any grpc_fd structs that are added to the epoll set via + epoll_ctl and are returned (within a very short window) via epoll_wait(). + + To work-around this race, we establish a happens-before relation between + the code just-before epoll_ctl() and the code after epoll_wait() by using + this atomic */ +gpr_atm g_epoll_sync; +#endif /* defined(GRPC_TSAN) */ + +#ifdef GRPC_PI_REF_COUNT_DEBUG +void pi_add_ref(polling_island *pi); +void pi_unref(polling_island *pi); + +void pi_add_ref_dbg(polling_island *pi, char *reason, char *file, int line) { + long old_cnt = gpr_atm_acq_load(&(pi->ref_count.count)); + pi_add_ref(pi); + gpr_log(GPR_DEBUG, "Add ref pi: %p, old: %ld -> new:%ld (%s) - (%s, %d)", + (void *)pi, old_cnt, old_cnt + 1, reason, file, line); +} + +void pi_unref_dbg(polling_island *pi, char *reason, char *file, int line) { + long old_cnt = gpr_atm_acq_load(&(pi->ref_count.count)); + pi_unref(pi); + gpr_log(GPR_DEBUG, "Unref pi: %p, old:%ld -> new:%ld (%s) - (%s, %d)", + (void *)pi, old_cnt, (old_cnt - 1), reason, file, line); +} +#endif + +void pi_add_ref(polling_island *pi) { gpr_ref(&pi->ref_count); } + +void pi_unref(polling_island *pi) { + /* If ref count went to zero, delete the polling island. + Note that this deletion not be done under a lock. Once the ref count goes + to zero, we are guaranteed that no one else holds a reference to the + polling island (and that there is no racing pi_add_ref() call either). + + Also, if we are deleting the polling island and the merged_to field is + non-empty, we should remove a ref to the merged_to polling island + */ + if (gpr_unref(&pi->ref_count)) { + polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); + polling_island_delete(pi); + if (next != NULL) { + PI_UNREF(next, "pi_delete"); /* Recursive call */ + } + } +} + +/* The caller is expected to hold pi->mu lock before calling this function */ +static void polling_island_add_fds_locked(polling_island *pi, grpc_fd **fds, + size_t fd_count, bool add_fd_refs, + grpc_error **error) { + int err; + size_t i; + struct epoll_event ev; + char *err_msg; + const char *err_desc = "polling_island_add_fds"; + +#ifdef GRPC_TSAN + /* See the definition of g_epoll_sync for more context */ + gpr_atm_rel_store(&g_epoll_sync, (gpr_atm)0); +#endif /* defined(GRPC_TSAN) */ + + for (i = 0; i < fd_count; i++) { + ev.events = (uint32_t)(EPOLLIN | EPOLLOUT | EPOLLET); + ev.data.ptr = fds[i]; + err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_ADD, fds[i]->fd, &ev); + + if (err < 0) { + if (errno != EEXIST) { + gpr_asprintf( + &err_msg, + "epoll_ctl (epoll_fd: %d) add fd: %d failed with error: %d (%s)", + pi->epoll_fd, fds[i]->fd, errno, strerror(errno)); + append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); + gpr_free(err_msg); + } + + continue; + } + + if (pi->fd_cnt == pi->fd_capacity) { + pi->fd_capacity = GPR_MAX(pi->fd_capacity + 8, pi->fd_cnt * 3 / 2); + pi->fds = gpr_realloc(pi->fds, sizeof(grpc_fd *) * pi->fd_capacity); + } + + pi->fds[pi->fd_cnt++] = fds[i]; + if (add_fd_refs) { + GRPC_FD_REF(fds[i], "polling_island"); + } + } +} + +/* The caller is expected to hold pi->mu before calling this */ +static void polling_island_add_wakeup_fd_locked(polling_island *pi, + grpc_wakeup_fd *wakeup_fd, + grpc_error **error) { + struct epoll_event ev; + int err; + char *err_msg; + const char *err_desc = "polling_island_add_wakeup_fd"; + + ev.events = (uint32_t)(EPOLLIN | EPOLLET); + ev.data.ptr = wakeup_fd; + err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_ADD, + GRPC_WAKEUP_FD_GET_READ_FD(wakeup_fd), &ev); + if (err < 0 && errno != EEXIST) { + gpr_asprintf(&err_msg, + "epoll_ctl (epoll_fd: %d) add wakeup fd: %d failed with " + "error: %d (%s)", + pi->epoll_fd, + GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), errno, + strerror(errno)); + append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); + gpr_free(err_msg); + } +} + +/* The caller is expected to hold pi->mu lock before calling this function */ +static void polling_island_remove_all_fds_locked(polling_island *pi, + bool remove_fd_refs, + grpc_error **error) { + int err; + size_t i; + char *err_msg; + const char *err_desc = "polling_island_remove_fds"; + + for (i = 0; i < pi->fd_cnt; i++) { + err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_DEL, pi->fds[i]->fd, NULL); + if (err < 0 && errno != ENOENT) { + gpr_asprintf(&err_msg, + "epoll_ctl (epoll_fd: %d) delete fds[%zu]: %d failed with " + "error: %d (%s)", + pi->epoll_fd, i, pi->fds[i]->fd, errno, strerror(errno)); + append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); + gpr_free(err_msg); + } + + if (remove_fd_refs) { + GRPC_FD_UNREF(pi->fds[i], "polling_island"); + } + } + + pi->fd_cnt = 0; +} + +/* The caller is expected to hold pi->mu lock before calling this function */ +static void polling_island_remove_fd_locked(polling_island *pi, grpc_fd *fd, + bool is_fd_closed, + grpc_error **error) { + int err; + size_t i; + char *err_msg; + const char *err_desc = "polling_island_remove_fd"; + + /* If fd is already closed, then it would have been automatically been removed + from the epoll set */ + if (!is_fd_closed) { + err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_DEL, fd->fd, NULL); + if (err < 0 && errno != ENOENT) { + gpr_asprintf( + &err_msg, + "epoll_ctl (epoll_fd: %d) del fd: %d failed with error: %d (%s)", + pi->epoll_fd, fd->fd, errno, strerror(errno)); + append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); + gpr_free(err_msg); + } + } + + for (i = 0; i < pi->fd_cnt; i++) { + if (pi->fds[i] == fd) { + pi->fds[i] = pi->fds[--pi->fd_cnt]; + GRPC_FD_UNREF(fd, "polling_island"); + break; + } + } +} + +/* Might return NULL in case of an error */ +static polling_island *polling_island_create(grpc_fd *initial_fd, + grpc_error **error) { + polling_island *pi = NULL; + char *err_msg; + const char *err_desc = "polling_island_create"; + + /* Try to get one from the polling island freelist */ + gpr_mu_lock(&g_pi_freelist_mu); + if (g_pi_freelist != NULL) { + pi = g_pi_freelist; + g_pi_freelist = g_pi_freelist->next_free; + pi->next_free = NULL; + } + gpr_mu_unlock(&g_pi_freelist_mu); + + /* Create new polling island if we could not get one from the free list */ + if (pi == NULL) { + pi = gpr_malloc(sizeof(*pi)); + gpr_mu_init(&pi->mu); + pi->fd_cnt = 0; + pi->fd_capacity = 0; + pi->fds = NULL; + } + + gpr_ref_init(&pi->ref_count, 0); + gpr_atm_rel_store(&pi->merged_to, (gpr_atm)NULL); + + pi->epoll_fd = epoll_create1(EPOLL_CLOEXEC); + + if (pi->epoll_fd < 0) { + gpr_asprintf(&err_msg, "epoll_create1 failed with error %d (%s)", errno, + strerror(errno)); + append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); + gpr_free(err_msg); + } else { + polling_island_add_wakeup_fd_locked(pi, &grpc_global_wakeup_fd, error); + pi->next_free = NULL; + + if (initial_fd != NULL) { + /* Lock the polling island here just in case we got this structure from + the freelist and the polling island lock was not released yet (by the + code that adds the polling island to the freelist) */ + gpr_mu_lock(&pi->mu); + polling_island_add_fds_locked(pi, &initial_fd, 1, true, error); + gpr_mu_unlock(&pi->mu); + } + } + + return pi; +} + +static void polling_island_delete(polling_island *pi) { + GPR_ASSERT(pi->fd_cnt == 0); + + gpr_atm_rel_store(&pi->merged_to, (gpr_atm)NULL); + + close(pi->epoll_fd); + pi->epoll_fd = -1; + + gpr_mu_lock(&g_pi_freelist_mu); + pi->next_free = g_pi_freelist; + g_pi_freelist = pi; + gpr_mu_unlock(&g_pi_freelist_mu); +} + +/* Attempts to gets the last polling island in the linked list (liked by the + * 'merged_to' field). Since this does not lock the polling island, there are no + * guarantees that the island returned is the last island */ +static polling_island *polling_island_maybe_get_latest(polling_island *pi) { + polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); + while (next != NULL) { + pi = next; + next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); + } + + return pi; +} + +/* Gets the lock on the *latest* polling island i.e the last polling island in + the linked list (linked by the 'merged_to' field). Call gpr_mu_unlock on the + returned polling island's mu. + Usage: To lock/unlock polling island "pi", do the following: + polling_island *pi_latest = polling_island_lock(pi); + ... + ... critical section .. + ... + gpr_mu_unlock(&pi_latest->mu); // NOTE: use pi_latest->mu. NOT pi->mu */ +static polling_island *polling_island_lock(polling_island *pi) { + polling_island *next = NULL; + + while (true) { + next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); + if (next == NULL) { + /* Looks like 'pi' is the last node in the linked list but unless we check + this by holding the pi->mu lock, we cannot be sure (i.e without the + pi->mu lock, we don't prevent island merges). + To be absolutely sure, check once more by holding the pi->mu lock */ + gpr_mu_lock(&pi->mu); + next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); + if (next == NULL) { + /* pi is infact the last node and we have the pi->mu lock. we're done */ + break; + } + + /* pi->merged_to is not NULL i.e pi isn't the last node anymore. pi->mu + * isn't the lock we are interested in. Continue traversing the list */ + gpr_mu_unlock(&pi->mu); + } + + pi = next; + } + + return pi; +} + +/* Gets the lock on the *latest* polling islands in the linked lists pointed by + *p and *q (and also updates *p and *q to point to the latest polling islands) + + This function is needed because calling the following block of code to obtain + locks on polling islands (*p and *q) is prone to deadlocks. + { + polling_island_lock(*p, true); + polling_island_lock(*q, true); + } + + Usage/example: + polling_island *p1; + polling_island *p2; + .. + polling_island_lock_pair(&p1, &p2); + .. + .. Critical section with both p1 and p2 locked + .. + // Release locks: Always call polling_island_unlock_pair() to release locks + polling_island_unlock_pair(p1, p2); +*/ +static void polling_island_lock_pair(polling_island **p, polling_island **q) { + polling_island *pi_1 = *p; + polling_island *pi_2 = *q; + polling_island *next_1 = NULL; + polling_island *next_2 = NULL; + + /* The algorithm is simple: + - Go to the last polling islands in the linked lists *pi_1 and *pi_2 (and + keep updating pi_1 and pi_2) + - Then obtain locks on the islands by following a lock order rule of + locking polling_island with lower address first + Special case: Before obtaining the locks, check if pi_1 and pi_2 are + pointing to the same island. If that is the case, we can just call + polling_island_lock() + - After obtaining both the locks, double check that the polling islands + are still the last polling islands in their respective linked lists + (this is because there might have been polling island merges before + we got the lock) + - If the polling islands are the last islands, we are done. If not, + release the locks and continue the process from the first step */ + while (true) { + next_1 = (polling_island *)gpr_atm_acq_load(&pi_1->merged_to); + while (next_1 != NULL) { + pi_1 = next_1; + next_1 = (polling_island *)gpr_atm_acq_load(&pi_1->merged_to); + } + + next_2 = (polling_island *)gpr_atm_acq_load(&pi_2->merged_to); + while (next_2 != NULL) { + pi_2 = next_2; + next_2 = (polling_island *)gpr_atm_acq_load(&pi_2->merged_to); + } + + if (pi_1 == pi_2) { + pi_1 = pi_2 = polling_island_lock(pi_1); + break; + } + + if (pi_1 < pi_2) { + gpr_mu_lock(&pi_1->mu); + gpr_mu_lock(&pi_2->mu); + } else { + gpr_mu_lock(&pi_2->mu); + gpr_mu_lock(&pi_1->mu); + } + + next_1 = (polling_island *)gpr_atm_acq_load(&pi_1->merged_to); + next_2 = (polling_island *)gpr_atm_acq_load(&pi_2->merged_to); + if (next_1 == NULL && next_2 == NULL) { + break; + } + + gpr_mu_unlock(&pi_1->mu); + gpr_mu_unlock(&pi_2->mu); + } + + *p = pi_1; + *q = pi_2; +} + +static void polling_island_unlock_pair(polling_island *p, polling_island *q) { + if (p == q) { + gpr_mu_unlock(&p->mu); + } else { + gpr_mu_unlock(&p->mu); + gpr_mu_unlock(&q->mu); + } +} + +static polling_island *polling_island_merge(polling_island *p, + polling_island *q, + grpc_error **error) { + /* Get locks on both the polling islands */ + polling_island_lock_pair(&p, &q); + + if (p != q) { + /* Make sure that p points to the polling island with fewer fds than q */ + if (p->fd_cnt > q->fd_cnt) { + GPR_SWAP(polling_island *, p, q); + } + + /* Merge p with q i.e move all the fds from p (The one with fewer fds) to q + Note that the refcounts on the fds being moved will not change here. + This is why the last param in the following two functions is 'false') */ + polling_island_add_fds_locked(q, p->fds, p->fd_cnt, false, error); + polling_island_remove_all_fds_locked(p, false, error); + + /* Wakeup all the pollers (if any) on p so that they pickup this change */ + polling_island_add_wakeup_fd_locked(p, &polling_island_wakeup_fd, error); + + /* Add the 'merged_to' link from p --> q */ + gpr_atm_rel_store(&p->merged_to, (gpr_atm)q); + PI_ADD_REF(q, "pi_merge"); /* To account for the new incoming ref from p */ + } + /* else if p == q, nothing needs to be done */ + + polling_island_unlock_pair(p, q); + + /* Return the merged polling island (Note that no merge would have happened + if p == q which is ok) */ + return q; +} + +static grpc_error *polling_island_global_init() { + grpc_error *error = GRPC_ERROR_NONE; + + gpr_mu_init(&g_pi_freelist_mu); + g_pi_freelist = NULL; + + error = grpc_wakeup_fd_init(&polling_island_wakeup_fd); + if (error == GRPC_ERROR_NONE) { + error = grpc_wakeup_fd_wakeup(&polling_island_wakeup_fd); + } + + return error; +} + +static void polling_island_global_shutdown() { + polling_island *next; + gpr_mu_lock(&g_pi_freelist_mu); + gpr_mu_unlock(&g_pi_freelist_mu); + while (g_pi_freelist != NULL) { + next = g_pi_freelist->next_free; + gpr_mu_destroy(&g_pi_freelist->mu); + gpr_free(g_pi_freelist->fds); + gpr_free(g_pi_freelist); + g_pi_freelist = next; + } + gpr_mu_destroy(&g_pi_freelist_mu); + + grpc_wakeup_fd_destroy(&polling_island_wakeup_fd); +} + +/******************************************************************************* + * Fd Definitions + */ + +/* 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. + */ + +/* 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. */ + +/* TODO: sreek: Right now, this wakes up all pollers. In future we should make + * sure to wake up one polling thread (which can wake up other threads if + * needed) */ +grpc_wakeup_fd grpc_global_wakeup_fd; + +static grpc_fd *fd_freelist = NULL; +static gpr_mu fd_freelist_mu; + +#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 %ld -> %ld [%s; %s:%d]", fd->fd, + (void *)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 %ld -> %ld [%s; %s:%d]", fd->fd, + (void *)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) { + /* Add the fd to the freelist */ + 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); + } else { + GPR_ASSERT(old > n); + } +} + +/* 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 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; + gpr_mu_destroy(&fd->mu); + gpr_free(fd); + } + gpr_mu_destroy(&fd_freelist_mu); +} + +static grpc_fd *fd_create(int fd, const char *name) { + grpc_fd *new_fd = NULL; + + gpr_mu_lock(&fd_freelist_mu); + if (fd_freelist != NULL) { + new_fd = fd_freelist; + fd_freelist = fd_freelist->freelist_next; + } + gpr_mu_unlock(&fd_freelist_mu); + + if (new_fd == NULL) { + new_fd = gpr_malloc(sizeof(grpc_fd)); + gpr_mu_init(&new_fd->mu); + gpr_mu_init(&new_fd->pi_mu); + } + + /* Note: It is not really needed to get the new_fd->mu lock here. If this is a + newly created fd (or an fd we got from the freelist), no one else would be + holding a lock to it anyway. */ + gpr_mu_lock(&new_fd->mu); + + gpr_atm_rel_store(&new_fd->refst, (gpr_atm)1); + new_fd->fd = fd; + new_fd->shutdown = false; + new_fd->orphaned = false; + new_fd->read_closure = CLOSURE_NOT_READY; + new_fd->write_closure = CLOSURE_NOT_READY; + new_fd->polling_island = NULL; + new_fd->freelist_next = NULL; + new_fd->on_done_closure = NULL; + new_fd->read_notifier_pollset = NULL; + + gpr_mu_unlock(&new_fd->mu); + + char *fd_name; + gpr_asprintf(&fd_name, "%s fd=%d", name, fd); + grpc_iomgr_register_object(&new_fd->iomgr_object, fd_name); +#ifdef GRPC_FD_REF_COUNT_DEBUG + gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, (void *)new_fd, fd_name); +#endif + gpr_free(fd_name); + return new_fd; +} + +static bool fd_is_orphaned(grpc_fd *fd) { + return (gpr_atm_acq_load(&fd->refst) & 1) == 0; +} + +static int fd_wrapped_fd(grpc_fd *fd) { + int ret_fd = -1; + gpr_mu_lock(&fd->mu); + if (!fd->orphaned) { + ret_fd = fd->fd; + } + gpr_mu_unlock(&fd->mu); + + return ret_fd; +} + +static void fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *on_done, int *release_fd, + const char *reason) { + bool is_fd_closed = false; + grpc_error *error = GRPC_ERROR_NONE; + + gpr_mu_lock(&fd->mu); + fd->on_done_closure = on_done; + + /* If release_fd is not NULL, we should be relinquishing control of the file + descriptor fd->fd (but we still own the grpc_fd structure). */ + if (release_fd != NULL) { + *release_fd = fd->fd; + } else { + close(fd->fd); + is_fd_closed = true; + } + + fd->orphaned = true; + + /* Remove the active status but keep referenced. We want this grpc_fd struct + to be alive (and not added to freelist) until the end of this function */ + REF_BY(fd, 1, reason); + + /* Remove the fd from the polling island: + - Get a lock on the latest polling island (i.e the last island in the + linked list pointed by fd->polling_island). This is the island that + would actually contain the fd + - Remove the fd from the latest polling island + - Unlock the latest polling island + - Set fd->polling_island to NULL (but remove the ref on the polling island + before doing this.) */ + gpr_mu_lock(&fd->pi_mu); + if (fd->polling_island != NULL) { + polling_island *pi_latest = polling_island_lock(fd->polling_island); + polling_island_remove_fd_locked(pi_latest, fd, is_fd_closed, &error); + gpr_mu_unlock(&pi_latest->mu); + + PI_UNREF(fd->polling_island, "fd_orphan"); + fd->polling_island = NULL; + } + gpr_mu_unlock(&fd->pi_mu); + + grpc_exec_ctx_sched(exec_ctx, fd->on_done_closure, error, NULL); + + gpr_mu_unlock(&fd->mu); + UNREF_BY(fd, 2, reason); /* Drop the reference */ + GRPC_LOG_IF_ERROR("fd_orphan", GRPC_ERROR_REF(error)); +} + +static grpc_error *fd_shutdown_error(bool shutdown) { + if (!shutdown) { + return GRPC_ERROR_NONE; + } else { + return GRPC_ERROR_CREATE("FD shutdown"); + } +} + +static void notify_on_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure **st, grpc_closure *closure) { + if (fd->shutdown) { + grpc_exec_ctx_sched(exec_ctx, closure, GRPC_ERROR_CREATE("FD shutdown"), + NULL); + } else 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_sched(exec_ctx, closure, fd_shutdown_error(fd->shutdown), + NULL); + } 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_sched(exec_ctx, *st, fd_shutdown_error(fd->shutdown), NULL); + *st = CLOSURE_NOT_READY; + return 1; + } +} + +static grpc_pollset *fd_get_read_notifier_pollset(grpc_exec_ctx *exec_ctx, + grpc_fd *fd) { + grpc_pollset *notifier = NULL; + + gpr_mu_lock(&fd->mu); + notifier = fd->read_notifier_pollset; + gpr_mu_unlock(&fd->mu); + + return notifier; +} + +static bool fd_is_shutdown(grpc_fd *fd) { + gpr_mu_lock(&fd->mu); + const bool r = fd->shutdown; + gpr_mu_unlock(&fd->mu); + return r; +} + +/* Might be called multiple times */ +static void fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + gpr_mu_lock(&fd->mu); + /* Do the actual shutdown only once */ + if (!fd->shutdown) { + fd->shutdown = true; + + shutdown(fd->fd, SHUT_RDWR); + /* Flush any pending read and write closures. Since fd->shutdown is 'true' + at this point, the closures would be called with 'success = false' */ + 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); +} + +/******************************************************************************* + * Pollset Definitions + */ +GPR_TLS_DECL(g_current_thread_pollset); +GPR_TLS_DECL(g_current_thread_worker); +static __thread bool g_initialized_sigmask; +static __thread sigset_t g_orig_sigmask; + +static void sig_handler(int sig_num) { +#ifdef GRPC_EPOLL_DEBUG + gpr_log(GPR_INFO, "Received signal %d", sig_num); +#endif +} + +static void poller_kick_init() { signal(grpc_wakeup_signal, sig_handler); } + +/* Global state management */ +static grpc_error *pollset_global_init(void) { + gpr_tls_init(&g_current_thread_pollset); + gpr_tls_init(&g_current_thread_worker); + poller_kick_init(); + return 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_pollset); + gpr_tls_destroy(&g_current_thread_worker); +} + +static grpc_error *pollset_worker_kick(grpc_pollset_worker *worker) { + grpc_error *err = GRPC_ERROR_NONE; + int err_num = pthread_kill(worker->pt_id, grpc_wakeup_signal); + if (err_num != 0) { + err = GRPC_OS_ERROR(err_num, "pthread_kill"); + } + return err; +} + +/* Return 1 if the pollset has active threads in pollset_work (pollset must + * be locked) */ +static int pollset_has_workers(grpc_pollset *p) { + return p->root_worker.next != &p->root_worker; +} + +static void remove_worker(grpc_pollset *p, grpc_pollset_worker *worker) { + worker->prev->next = worker->next; + worker->next->prev = worker->prev; +} + +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; +} + +/* p->mu must be held before calling this function */ +static grpc_error *pollset_kick(grpc_pollset *p, + grpc_pollset_worker *specific_worker) { + GPR_TIMER_BEGIN("pollset_kick", 0); + grpc_error *error = GRPC_ERROR_NONE; + const char *err_desc = "Kick Failure"; + + grpc_pollset_worker *worker = specific_worker; + if (worker != NULL) { + if (worker == GRPC_POLLSET_KICK_BROADCAST) { + if (pollset_has_workers(p)) { + GPR_TIMER_BEGIN("pollset_kick.broadcast", 0); + for (worker = p->root_worker.next; worker != &p->root_worker; + worker = worker->next) { + if (gpr_tls_get(&g_current_thread_worker) != (intptr_t)worker) { + append_error(&error, pollset_worker_kick(worker), err_desc); + } + } + GPR_TIMER_END("pollset_kick.broadcast", 0); + } else { + p->kicked_without_pollers = true; + } + } else { + GPR_TIMER_MARK("kicked_specifically", 0); + if (gpr_tls_get(&g_current_thread_worker) != (intptr_t)worker) { + append_error(&error, pollset_worker_kick(worker), err_desc); + } + } + } else if (gpr_tls_get(&g_current_thread_pollset) != (intptr_t)p) { + /* Since worker == NULL, it means that we can kick "any" worker on this + pollset 'p'. If 'p' happens to be the same pollset this thread is + currently polling (i.e in pollset_work() function), then there is no need + to kick any other worker since the current thread can just absorb the + kick. This is the reason why we enter this case only when + g_current_thread_pollset is != p */ + + GPR_TIMER_MARK("kick_anonymous", 0); + worker = pop_front_worker(p); + if (worker != NULL) { + GPR_TIMER_MARK("finally_kick", 0); + push_back_worker(p, worker); + append_error(&error, pollset_worker_kick(worker), err_desc); + } else { + GPR_TIMER_MARK("kicked_no_pollers", 0); + p->kicked_without_pollers = true; + } + } + + GPR_TIMER_END("pollset_kick", 0); + GRPC_LOG_IF_ERROR("pollset_kick", GRPC_ERROR_REF(error)); + return error; +} + +static grpc_error *kick_poller(void) { + return grpc_wakeup_fd_wakeup(&grpc_global_wakeup_fd); +} + +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->kicked_without_pollers = false; + + pollset->shutting_down = false; + pollset->finish_shutdown_called = false; + pollset->shutdown_done = NULL; + + pollset->polling_island = NULL; +} + +/* 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) { + 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))); +} + +static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_pollset *notifier) { + /* Need the fd->mu since we might be racing with fd_notify_on_read */ + gpr_mu_lock(&fd->mu); + set_ready_locked(exec_ctx, fd, &fd->read_closure); + fd->read_notifier_pollset = notifier; + gpr_mu_unlock(&fd->mu); +} + +static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + /* Need the fd->mu since we might be racing with fd_notify_on_write */ + gpr_mu_lock(&fd->mu); + set_ready_locked(exec_ctx, fd, &fd->write_closure); + gpr_mu_unlock(&fd->mu); +} + +static void pollset_release_polling_island(grpc_pollset *ps, char *reason) { + if (ps->polling_island != NULL) { + PI_UNREF(ps->polling_island, reason); + } + ps->polling_island = NULL; +} + +static void finish_shutdown_locked(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset) { + /* The pollset cannot have any workers if we are at this stage */ + GPR_ASSERT(!pollset_has_workers(pollset)); + + pollset->finish_shutdown_called = true; + + /* Release the ref and set pollset->polling_island to NULL */ + pollset_release_polling_island(pollset, "ps_shutdown"); + grpc_exec_ctx_sched(exec_ctx, pollset->shutdown_done, GRPC_ERROR_NONE, NULL); +} + +/* pollset->mu lock must be held by the caller before calling this */ +static void pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_closure *closure) { + GPR_TIMER_BEGIN("pollset_shutdown", 0); + GPR_ASSERT(!pollset->shutting_down); + pollset->shutting_down = true; + pollset->shutdown_done = closure; + pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); + + /* If the pollset has any workers, we cannot call finish_shutdown_locked() + because it would release the underlying polling island. In such a case, we + let the last worker call finish_shutdown_locked() from pollset_work() */ + if (!pollset_has_workers(pollset)) { + GPR_ASSERT(!pollset->finish_shutdown_called); + GPR_TIMER_MARK("pollset_shutdown.finish_shutdown_locked", 0); + finish_shutdown_locked(exec_ctx, pollset); + } + GPR_TIMER_END("pollset_shutdown", 0); +} + +/* pollset_shutdown is guaranteed to be called before pollset_destroy. So other + * than destroying the mutexes, there is nothing special that needs to be done + * here */ +static void pollset_destroy(grpc_pollset *pollset) { + GPR_ASSERT(!pollset_has_workers(pollset)); + gpr_mu_destroy(&pollset->mu); +} + +static void pollset_reset(grpc_pollset *pollset) { + GPR_ASSERT(pollset->shutting_down); + GPR_ASSERT(!pollset_has_workers(pollset)); + pollset->shutting_down = false; + pollset->finish_shutdown_called = false; + pollset->kicked_without_pollers = false; + pollset->shutdown_done = NULL; + pollset_release_polling_island(pollset, "ps_reset"); +} + +#define GRPC_EPOLL_MAX_EVENTS 1000 +/* Note: sig_mask contains the signal mask to use *during* epoll_wait() */ +static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, int timeout_ms, + sigset_t *sig_mask, grpc_error **error) { + struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS]; + int epoll_fd = -1; + int ep_rv; + polling_island *pi = NULL; + char *err_msg; + const char *err_desc = "pollset_work_and_unlock"; + GPR_TIMER_BEGIN("pollset_work_and_unlock", 0); + + /* We need to get the epoll_fd to wait on. The epoll_fd is in inside the + latest polling island pointed by pollset->polling_island. + + Since epoll_fd is immutable, we can read it without obtaining the polling + island lock. There is however a possibility that the polling island (from + which we got the epoll_fd) got merged with another island while we are + in this function. This is still okay because in such a case, we will wakeup + right-away from epoll_wait() and pick up the latest polling_island the next + this function (i.e pollset_work_and_unlock()) is called */ + + if (pollset->polling_island == NULL) { + pollset->polling_island = polling_island_create(NULL, error); + if (pollset->polling_island == NULL) { + GPR_TIMER_END("pollset_work_and_unlock", 0); + return; /* Fatal error. We cannot continue */ + } + + PI_ADD_REF(pollset->polling_island, "ps"); + } + + pi = polling_island_maybe_get_latest(pollset->polling_island); + epoll_fd = pi->epoll_fd; + + /* Update the pollset->polling_island since the island being pointed by + pollset->polling_island maybe older than the one pointed by pi) */ + if (pollset->polling_island != pi) { + /* Always do PI_ADD_REF before PI_UNREF because PI_UNREF may cause the + polling island to be deleted */ + PI_ADD_REF(pi, "ps"); + PI_UNREF(pollset->polling_island, "ps"); + pollset->polling_island = pi; + } + + /* Add an extra ref so that the island does not get destroyed (which means + the epoll_fd won't be closed) while we are are doing an epoll_wait() on the + epoll_fd */ + PI_ADD_REF(pi, "ps_work"); + gpr_mu_unlock(&pollset->mu); + + do { + ep_rv = epoll_pwait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms, + sig_mask); + if (ep_rv < 0) { + if (errno != EINTR) { + gpr_asprintf(&err_msg, + "epoll_wait() epoll fd: %d failed with error: %d (%s)", + epoll_fd, errno, strerror(errno)); + append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); + } else { + /* We were interrupted. Save an interation by doing a zero timeout + epoll_wait to see if there are any other events of interest */ + ep_rv = epoll_wait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, 0); + } + } + +#ifdef GRPC_TSAN + /* See the definition of g_poll_sync for more details */ + gpr_atm_acq_load(&g_epoll_sync); +#endif /* defined(GRPC_TSAN) */ + + for (int i = 0; i < ep_rv; ++i) { + void *data_ptr = ep_ev[i].data.ptr; + if (data_ptr == &grpc_global_wakeup_fd) { + append_error(error, + grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd), + err_desc); + } else if (data_ptr == &polling_island_wakeup_fd) { + /* This means that our polling island is merged with a different + island. We do not have to do anything here since the subsequent call + to the function pollset_work_and_unlock() will pick up the correct + epoll_fd */ + } else { + grpc_fd *fd = data_ptr; + 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 (read_ev || cancel) { + fd_become_readable(exec_ctx, fd, pollset); + } + if (write_ev || cancel) { + fd_become_writable(exec_ctx, fd); + } + } + } + } while (ep_rv == GRPC_EPOLL_MAX_EVENTS); + + GPR_ASSERT(pi != NULL); + + /* Before leaving, release the extra ref we added to the polling island. It + is important to use "pi" here (i.e our old copy of pollset->polling_island + that we got before releasing the polling island lock). This is because + pollset->polling_island pointer might get udpated in other parts of the + code when there is an island merge while we are doing epoll_wait() above */ + PI_UNREF(pi, "ps_work"); + + GPR_TIMER_END("pollset_work_and_unlock", 0); +} + +/* pollset->mu lock must be held by the caller before calling this. + The function pollset_work() may temporarily release the lock (pollset->mu) + during the course of its execution but it will always re-acquire the lock and + ensure that it is held by the time the function returns */ +static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_pollset_worker **worker_hdl, + gpr_timespec now, gpr_timespec deadline) { + GPR_TIMER_BEGIN("pollset_work", 0); + grpc_error *error = GRPC_ERROR_NONE; + int timeout_ms = poll_deadline_to_millis_timeout(deadline, now); + + sigset_t new_mask; + + grpc_pollset_worker worker; + worker.next = worker.prev = NULL; + worker.pt_id = pthread_self(); + + *worker_hdl = &worker; + + gpr_tls_set(&g_current_thread_pollset, (intptr_t)pollset); + gpr_tls_set(&g_current_thread_worker, (intptr_t)&worker); + + if (pollset->kicked_without_pollers) { + /* If the pollset was kicked without pollers, pretend that the current + worker got the kick and skip polling. A kick indicates that there is some + work that needs attention like an event on the completion queue or an + alarm */ + GPR_TIMER_MARK("pollset_work.kicked_without_pollers", 0); + pollset->kicked_without_pollers = 0; + } else if (!pollset->shutting_down) { + /* We use the posix-signal with number 'grpc_wakeup_signal' for waking up + (i.e 'kicking') a worker in the pollset. + A 'kick' is a way to inform that worker that there is some pending work + that needs immediate attention (like an event on the completion queue, + or a polling island merge that results in a new epoll-fd to wait on) and + that the worker should not spend time waiting in epoll_pwait(). + + A kick can come at anytime (i.e before/during or after the worker calls + epoll_pwait()) but in all cases we have to make sure that when a worker + gets a kick, it does not spend time in epoll_pwait(). In other words, one + kick should result in skipping/exiting of one epoll_pwait(); + + To accomplish this, we mask 'grpc_wakeup_signal' on this worker at all + times *except* when it is in epoll_pwait(). This way, the worker never + misses acting on a kick */ + + if (!g_initialized_sigmask) { + sigemptyset(&new_mask); + sigaddset(&new_mask, grpc_wakeup_signal); + pthread_sigmask(SIG_BLOCK, &new_mask, &g_orig_sigmask); + sigdelset(&g_orig_sigmask, grpc_wakeup_signal); + g_initialized_sigmask = true; + /* new_mask: The new thread mask which blocks 'grpc_wakeup_signal'. + This is the mask used at all times *except during + epoll_wait()*" + g_orig_sigmask: The thread mask which allows 'grpc_wakeup_signal' and + this is the mask to use *during epoll_wait()* + + The new_mask is set on the worker before it is added to the pollset + (i.e before it can be kicked) */ + } + + push_front_worker(pollset, &worker); /* Add worker to pollset */ + + pollset_work_and_unlock(exec_ctx, pollset, timeout_ms, &g_orig_sigmask, + &error); + grpc_exec_ctx_flush(exec_ctx); + + gpr_mu_lock(&pollset->mu); + remove_worker(pollset, &worker); + } + + /* If we are the last worker on the pollset (i.e pollset_has_workers() is + false at this point) and the pollset is shutting down, we may have to + finish the shutdown process by calling finish_shutdown_locked(). + See pollset_shutdown() for more details. + + Note: Continuing to access pollset here is safe; it is the caller's + responsibility to not destroy a pollset when it has outstanding calls to + pollset_work() */ + if (pollset->shutting_down && !pollset_has_workers(pollset) && + !pollset->finish_shutdown_called) { + GPR_TIMER_MARK("pollset_work.finish_shutdown_locked", 0); + finish_shutdown_locked(exec_ctx, pollset); + + gpr_mu_unlock(&pollset->mu); + grpc_exec_ctx_flush(exec_ctx); + gpr_mu_lock(&pollset->mu); + } + + *worker_hdl = NULL; + + gpr_tls_set(&g_current_thread_pollset, (intptr_t)0); + gpr_tls_set(&g_current_thread_worker, (intptr_t)0); + + GPR_TIMER_END("pollset_work", 0); + + GRPC_LOG_IF_ERROR("pollset_work", GRPC_ERROR_REF(error)); + return error; +} + +static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_fd *fd) { + grpc_error *error = GRPC_ERROR_NONE; + + gpr_mu_lock(&pollset->mu); + gpr_mu_lock(&fd->pi_mu); + + polling_island *pi_new = NULL; + + /* 1) If fd->polling_island and pollset->polling_island are both non-NULL and + * equal, do nothing. + * 2) If fd->polling_island and pollset->polling_island are both NULL, create + * a new polling island (with a refcount of 2) and make the polling_island + * fields in both fd and pollset to point to the new island + * 3) If one of fd->polling_island or pollset->polling_island is NULL, update + * the NULL polling_island field to point to the non-NULL polling_island + * field (ensure that the refcount on the polling island is incremented by + * 1 to account for the newly added reference) + * 4) Finally, if fd->polling_island and pollset->polling_island are non-NULL + * and different, merge both the polling islands and update the + * polling_island fields in both fd and pollset to point to the merged + * polling island. + */ + if (fd->polling_island == pollset->polling_island) { + pi_new = fd->polling_island; + if (pi_new == NULL) { + pi_new = polling_island_create(fd, &error); + } + } else if (fd->polling_island == NULL) { + pi_new = polling_island_lock(pollset->polling_island); + polling_island_add_fds_locked(pi_new, &fd, 1, true, &error); + gpr_mu_unlock(&pi_new->mu); + } else if (pollset->polling_island == NULL) { + pi_new = polling_island_lock(fd->polling_island); + gpr_mu_unlock(&pi_new->mu); + } else { + pi_new = polling_island_merge(fd->polling_island, pollset->polling_island, + &error); + } + + /* At this point, pi_new is the polling island that both fd->polling_island + and pollset->polling_island must be pointing to */ + + if (fd->polling_island != pi_new) { + PI_ADD_REF(pi_new, "fd"); + if (fd->polling_island != NULL) { + PI_UNREF(fd->polling_island, "fd"); + } + fd->polling_island = pi_new; + } + + if (pollset->polling_island != pi_new) { + PI_ADD_REF(pi_new, "ps"); + if (pollset->polling_island != NULL) { + PI_UNREF(pollset->polling_island, "ps"); + } + pollset->polling_island = pi_new; + } + + gpr_mu_unlock(&fd->pi_mu); + gpr_mu_unlock(&pollset->mu); +} + +/******************************************************************************* + * Pollset-set Definitions + */ + +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_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); +} + +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); +} + +/* Test helper functions + * */ +void *grpc_fd_get_polling_island(grpc_fd *fd) { + polling_island *pi; + + gpr_mu_lock(&fd->pi_mu); + pi = fd->polling_island; + gpr_mu_unlock(&fd->pi_mu); + + return pi; +} + +void *grpc_pollset_get_polling_island(grpc_pollset *ps) { + polling_island *pi; + + gpr_mu_lock(&ps->mu); + pi = ps->polling_island; + gpr_mu_unlock(&ps->mu); + + return pi; +} + +bool grpc_are_polling_islands_equal(void *p, void *q) { + polling_island *p1 = p; + polling_island *p2 = q; + + /* Note: polling_island_lock_pair() may change p1 and p2 to point to the + latest polling islands in their respective linked lists */ + polling_island_lock_pair(&p1, &p2); + polling_island_unlock_pair(p1, p2); + + return p1 == p2; +} + +/******************************************************************************* + * Event engine binding + */ + +static void shutdown_engine(void) { + fd_global_shutdown(); + pollset_global_shutdown(); + polling_island_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_is_shutdown = fd_is_shutdown, + .fd_notify_on_read = fd_notify_on_read, + .fd_notify_on_write = fd_notify_on_write, + .fd_get_read_notifier_pollset = fd_get_read_notifier_pollset, + + .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, +}; + +/* It is possible that GLIBC has epoll but the underlying kernel doesn't. + * Create a dummy epoll_fd to make sure epoll support is available */ +static bool is_epoll_available() { + int fd = epoll_create1(EPOLL_CLOEXEC); + if (fd < 0) { + gpr_log( + GPR_ERROR, + "epoll_create1 failed with error: %d. Not using epoll polling engine", + fd); + return false; + } + close(fd); + return true; +} + +const grpc_event_engine_vtable *grpc_init_epoll_linux(void) { + /* If use of signals is disabled, we cannot use epoll engine*/ + if (is_grpc_wakeup_signal_initialized && grpc_wakeup_signal < 0) { + return NULL; + } + + if (!is_epoll_available()) { + return NULL; + } + + if (!is_grpc_wakeup_signal_initialized) { + grpc_use_signal(SIGRTMIN + 2); + } + + fd_global_init(); + + if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) { + return NULL; + } + + if (!GRPC_LOG_IF_ERROR("polling_island_global_init", + polling_island_global_init())) { + return NULL; + } + + return &vtable; +} + +#else /* defined(GPR_LINUX_EPOLL) */ +#if defined(GPR_POSIX_SOCKET) +#include "src/core/lib/iomgr/ev_posix.h" +/* If GPR_LINUX_EPOLL is not defined, it means epoll is not available. Return + * NULL */ +const grpc_event_engine_vtable *grpc_init_epoll_linux(void) { return NULL; } +#endif /* defined(GPR_POSIX_SOCKET) */ + +void grpc_use_signal(int signum) {} +#endif /* !defined(GPR_LINUX_EPOLL) */ diff --git a/src/core/lib/iomgr/ev_epoll_linux.h b/src/core/lib/iomgr/ev_epoll_linux.h new file mode 100644 index 0000000000..7a494aba19 --- /dev/null +++ b/src/core/lib/iomgr/ev_epoll_linux.h @@ -0,0 +1,47 @@ +/* + * + * 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_EV_EPOLL_LINUX_H +#define GRPC_CORE_LIB_IOMGR_EV_EPOLL_LINUX_H + +#include "src/core/lib/iomgr/ev_posix.h" + +const grpc_event_engine_vtable *grpc_init_epoll_linux(void); + +#ifdef GPR_LINUX_EPOLL +void *grpc_fd_get_polling_island(grpc_fd *fd); +void *grpc_pollset_get_polling_island(grpc_pollset *ps); +bool grpc_are_polling_islands_equal(void *p, void *q); +#endif /* defined(GPR_LINUX_EPOLL) */ + +#endif /* GRPC_CORE_LIB_IOMGR_EV_EPOLL_LINUX_H */ diff --git a/src/core/lib/iomgr/ev_posix.c b/src/core/lib/iomgr/ev_posix.c index 4c360f13a3..a3c1e9db9a 100644 --- a/src/core/lib/iomgr/ev_posix.c +++ b/src/core/lib/iomgr/ev_posix.c @@ -44,6 +44,7 @@ #include <grpc/support/string_util.h> #include <grpc/support/useful.h> +#include "src/core/lib/iomgr/ev_epoll_linux.h" #include "src/core/lib/iomgr/ev_poll_and_epoll_posix.h" #include "src/core/lib/iomgr/ev_poll_posix.h" #include "src/core/lib/support/env.h" @@ -53,6 +54,7 @@ grpc_poll_function_type grpc_poll_function = poll; static const grpc_event_engine_vtable *g_event_engine; +static const char *g_poll_strategy_name = NULL; typedef const grpc_event_engine_vtable *(*event_engine_factory_fn)(void); @@ -62,7 +64,9 @@ typedef struct { } event_engine_factory; static const event_engine_factory g_factories[] = { - {"poll", grpc_init_poll_posix}, {"legacy", grpc_init_poll_and_epoll_posix}, + {"epoll", grpc_init_epoll_linux}, + {"poll", grpc_init_poll_posix}, + {"legacy", grpc_init_poll_and_epoll_posix}, }; static void add(const char *beg, const char *end, char ***ss, size_t *ns) { @@ -98,6 +102,7 @@ static void try_engine(const char *engine) { for (size_t i = 0; i < GPR_ARRAY_SIZE(g_factories); i++) { if (is(engine, g_factories[i].name)) { if ((g_event_engine = g_factories[i].factory())) { + g_poll_strategy_name = g_factories[i].name; gpr_log(GPR_DEBUG, "Using polling engine: %s", g_factories[i].name); return; } @@ -105,6 +110,9 @@ static void try_engine(const char *engine) { } } +/* Call this only after calling grpc_event_engine_init() */ +const char *grpc_get_poll_strategy_name() { return g_poll_strategy_name; } + void grpc_event_engine_init(void) { char *s = gpr_getenv("GRPC_POLL_STRATEGY"); if (s == NULL) { @@ -167,11 +175,6 @@ void grpc_fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, g_event_engine->fd_notify_on_write(exec_ctx, fd, closure); } -grpc_pollset *grpc_fd_get_read_notifier_pollset(grpc_exec_ctx *exec_ctx, - grpc_fd *fd) { - return g_event_engine->fd_get_read_notifier_pollset(exec_ctx, fd); -} - size_t grpc_pollset_size(void) { return g_event_engine->pollset_size; } void grpc_pollset_init(grpc_pollset *pollset, gpr_mu **mu) { diff --git a/src/core/lib/iomgr/ev_posix.h b/src/core/lib/iomgr/ev_posix.h index 87e9bc4d46..579c84ef70 100644 --- a/src/core/lib/iomgr/ev_posix.h +++ b/src/core/lib/iomgr/ev_posix.h @@ -99,6 +99,9 @@ typedef struct grpc_event_engine_vtable { void grpc_event_engine_init(void); void grpc_event_engine_shutdown(void); +/* Return the name of the poll strategy */ +const char *grpc_get_poll_strategy_name(); + /* Create a wrapped file descriptor. Requires fd is a non-blocking file descriptor. This takes ownership of closing fd. */ diff --git a/src/core/lib/iomgr/network_status_tracker.c b/src/core/lib/iomgr/network_status_tracker.c new file mode 100644 index 0000000000..38a1c9b7d4 --- /dev/null +++ b/src/core/lib/iomgr/network_status_tracker.c @@ -0,0 +1,121 @@ +/* + * + * 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/alloc.h> +#include <grpc/support/log.h> +#include "src/core/lib/iomgr/endpoint.h" + +typedef struct endpoint_ll_node { + grpc_endpoint *ep; + struct endpoint_ll_node *next; +} endpoint_ll_node; + +static endpoint_ll_node *head = NULL; +static gpr_mu g_endpoint_mutex; +static bool g_init_done = false; + +void grpc_initialize_network_status_monitor() { + g_init_done = true; + gpr_mu_init(&g_endpoint_mutex); + // TODO(makarandd): Install callback with OS to monitor network status. +} + +void grpc_destroy_network_status_monitor() { + for (endpoint_ll_node *curr = head; curr != NULL;) { + endpoint_ll_node *next = curr->next; + gpr_free(curr); + curr = next; + } + gpr_mu_destroy(&g_endpoint_mutex); +} + +void grpc_network_status_register_endpoint(grpc_endpoint *ep) { + if (!g_init_done) { + grpc_initialize_network_status_monitor(); + } + gpr_mu_lock(&g_endpoint_mutex); + if (head == NULL) { + head = (endpoint_ll_node *)gpr_malloc(sizeof(endpoint_ll_node)); + head->ep = ep; + head->next = NULL; + } else { + endpoint_ll_node *prev_head = head; + head = (endpoint_ll_node *)gpr_malloc(sizeof(endpoint_ll_node)); + head->ep = ep; + head->next = prev_head; + } + gpr_mu_unlock(&g_endpoint_mutex); +} + +void grpc_network_status_unregister_endpoint(grpc_endpoint *ep) { + gpr_mu_lock(&g_endpoint_mutex); + GPR_ASSERT(head); + bool found = false; + endpoint_ll_node *prev = head; + // if we're unregistering the head, just move head to the next + if (ep == head->ep) { + head = head->next; + gpr_free(prev); + found = true; + } else { + for (endpoint_ll_node *curr = head->next; curr != NULL; curr = curr->next) { + if (ep == curr->ep) { + prev->next = curr->next; + gpr_free(curr); + found = true; + break; + } + prev = curr; + } + } + gpr_mu_unlock(&g_endpoint_mutex); + GPR_ASSERT(found); +} + +// Walk the linked-list from head and execute shutdown. It is possible that +// other threads might be in the process of shutdown as well, but that has +// no side effect since endpoint shutdown is idempotent. +void grpc_network_status_shutdown_all_endpoints() { + gpr_mu_lock(&g_endpoint_mutex); + if (head == NULL) { + gpr_mu_unlock(&g_endpoint_mutex); + return; + } + grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT; + + for (endpoint_ll_node *curr = head; curr != NULL; curr = curr->next) { + curr->ep->vtable->shutdown(&exec_ctx, curr->ep); + } + gpr_mu_unlock(&g_endpoint_mutex); + grpc_exec_ctx_finish(&exec_ctx); +} diff --git a/src/core/lib/iomgr/network_status_tracker.h b/src/core/lib/iomgr/network_status_tracker.h new file mode 100644 index 0000000000..74a1aa8135 --- /dev/null +++ b/src/core/lib/iomgr/network_status_tracker.h @@ -0,0 +1,41 @@ +/* + * + * 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. + * + */ + +#ifndef GRPC_CORE_LIB_IOMGR_NETWORK_STATUS_TRACKER_H +#define GRPC_CORE_LIB_IOMGR_NETWORK_STATUS_TRACKER_H +#include "src/core/lib/iomgr/endpoint.h" + +void grpc_network_status_register_endpoint(grpc_endpoint *ep); +void grpc_network_status_unregister_endpoint(grpc_endpoint *ep); +void grpc_network_status_shutdown_all_endpoints(); +#endif /* GRPC_CORE_LIB_IOMGR_NETWORK_STATUS_TRACKER_H */ diff --git a/src/core/lib/iomgr/socket_utils_common_posix.c b/src/core/lib/iomgr/socket_utils_common_posix.c index 3a1371617e..d2f6261e2a 100644 --- a/src/core/lib/iomgr/socket_utils_common_posix.c +++ b/src/core/lib/iomgr/socket_utils_common_posix.c @@ -169,6 +169,28 @@ grpc_error *grpc_set_socket_reuse_addr(int fd, int reuse) { return GRPC_ERROR_NONE; } +/* set a socket to reuse old addresses */ +grpc_error *grpc_set_socket_reuse_port(int fd, int reuse) { +#ifndef SO_REUSEPORT + return GRPC_ERROR_CREATE("SO_REUSEPORT unavailable on compiling system"); +#else + int val = (reuse != 0); + int newval; + socklen_t intlen = sizeof(newval); + if (0 != setsockopt(fd, SOL_SOCKET, SO_REUSEPORT, &val, sizeof(val))) { + return GRPC_OS_ERROR(errno, "setsockopt(SO_REUSEPORT)"); + } + if (0 != getsockopt(fd, SOL_SOCKET, SO_REUSEPORT, &newval, &intlen)) { + return GRPC_OS_ERROR(errno, "getsockopt(SO_REUSEPORT)"); + } + if ((newval != 0) != val) { + return GRPC_ERROR_CREATE("Failed to set SO_REUSEPORT"); + } + + return GRPC_ERROR_NONE; +#endif +} + /* disable nagle */ grpc_error *grpc_set_socket_low_latency(int fd, int low_latency) { int val = (low_latency != 0); diff --git a/src/core/lib/iomgr/socket_utils_posix.h b/src/core/lib/iomgr/socket_utils_posix.h index 30ff39dfa3..7bcc2219ae 100644 --- a/src/core/lib/iomgr/socket_utils_posix.h +++ b/src/core/lib/iomgr/socket_utils_posix.h @@ -55,6 +55,9 @@ grpc_error *grpc_set_socket_reuse_addr(int fd, int reuse); /* disable nagle */ grpc_error *grpc_set_socket_low_latency(int fd, int low_latency); +/* set SO_REUSEPORT */ +grpc_error *grpc_set_socket_reuse_port(int fd, int reuse); + /* Returns true if this system can create AF_INET6 sockets bound to ::1. The value is probed once, and cached for the life of the process. diff --git a/src/core/lib/iomgr/tcp_posix.c b/src/core/lib/iomgr/tcp_posix.c index 017f52c367..2ab45e33ce 100644 --- a/src/core/lib/iomgr/tcp_posix.c +++ b/src/core/lib/iomgr/tcp_posix.c @@ -35,6 +35,7 @@ #ifdef GPR_POSIX_SOCKET +#include "src/core/lib/iomgr/network_status_tracker.h" #include "src/core/lib/iomgr/tcp_posix.h" #include <errno.h> @@ -152,6 +153,7 @@ static void tcp_ref(grpc_tcp *tcp) { gpr_ref(&tcp->refcount); } #endif static void tcp_destroy(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep) { + grpc_network_status_unregister_endpoint(ep); grpc_tcp *tcp = (grpc_tcp *)ep; TCP_UNREF(exec_ctx, tcp, "destroy"); } @@ -160,7 +162,7 @@ static void call_read_cb(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp, grpc_error *error) { grpc_closure *cb = tcp->read_cb; - if (false && grpc_tcp_trace) { + if (grpc_tcp_trace) { size_t i; const char *str = grpc_error_string(error); gpr_log(GPR_DEBUG, "read: error=%s", str); @@ -394,7 +396,7 @@ static void tcp_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, grpc_tcp *tcp = (grpc_tcp *)ep; grpc_error *error = GRPC_ERROR_NONE; - if (false && grpc_tcp_trace) { + if (grpc_tcp_trace) { size_t i; for (i = 0; i < buf->count; i++) { @@ -474,6 +476,8 @@ grpc_endpoint *grpc_tcp_create(grpc_fd *em_fd, size_t slice_size, tcp->write_closure.cb = tcp_handle_write; tcp->write_closure.cb_arg = tcp; gpr_slice_buffer_init(&tcp->last_read_buffer); + /* Tell network status tracker about new endpoint */ + grpc_network_status_register_endpoint(&tcp->base); return &tcp->base; } diff --git a/src/core/lib/iomgr/tcp_server.h b/src/core/lib/iomgr/tcp_server.h index 1d6e70dbe9..875a6ca14a 100644 --- a/src/core/lib/iomgr/tcp_server.h +++ b/src/core/lib/iomgr/tcp_server.h @@ -34,6 +34,8 @@ #ifndef GRPC_CORE_LIB_IOMGR_TCP_SERVER_H #define GRPC_CORE_LIB_IOMGR_TCP_SERVER_H +#include <grpc/grpc.h> + #include "src/core/lib/iomgr/closure.h" #include "src/core/lib/iomgr/endpoint.h" @@ -59,6 +61,7 @@ typedef void (*grpc_tcp_server_cb)(grpc_exec_ctx *exec_ctx, void *arg, If shutdown_complete is not NULL, it will be used by grpc_tcp_server_unref() when the ref count reaches zero. */ grpc_error *grpc_tcp_server_create(grpc_closure *shutdown_complete, + const grpc_channel_args *args, grpc_tcp_server **server); /* Start listening to bound ports */ diff --git a/src/core/lib/iomgr/tcp_server_posix.c b/src/core/lib/iomgr/tcp_server_posix.c index 6fb547bb36..a1a463550a 100644 --- a/src/core/lib/iomgr/tcp_server_posix.c +++ b/src/core/lib/iomgr/tcp_server_posix.c @@ -112,8 +112,10 @@ struct grpc_tcp_server { /* destroyed port count: how many ports are completely destroyed */ size_t destroyed_ports; - /* is this server shutting down? (boolean) */ - int shutdown; + /* is this server shutting down? */ + bool shutdown; + /* use SO_REUSEPORT */ + bool so_reuseport; /* linked list of server ports */ grpc_tcp_listener *head; @@ -135,14 +137,42 @@ struct grpc_tcp_server { size_t next_pollset_to_assign; }; +static gpr_once check_init = GPR_ONCE_INIT; +static bool has_so_reuseport; + +static void init(void) { + int s = socket(AF_INET, SOCK_STREAM, 0); + if (s >= 0) { + has_so_reuseport = GRPC_LOG_IF_ERROR("check for SO_REUSEPORT", + grpc_set_socket_reuse_port(s, 1)); + close(s); + } +} + grpc_error *grpc_tcp_server_create(grpc_closure *shutdown_complete, + const grpc_channel_args *args, grpc_tcp_server **server) { + gpr_once_init(&check_init, init); + grpc_tcp_server *s = gpr_malloc(sizeof(grpc_tcp_server)); + s->so_reuseport = has_so_reuseport; + for (size_t i = 0; i < (args == NULL ? 0 : args->num_args); i++) { + if (0 == strcmp(GRPC_ARG_ALLOW_REUSEPORT, args->args[i].key)) { + if (args->args[i].type == GRPC_ARG_INTEGER) { + s->so_reuseport = + has_so_reuseport && (args->args[i].value.integer != 0); + } else { + gpr_free(s); + return GRPC_ERROR_CREATE(GRPC_ARG_ALLOW_REUSEPORT + " must be an integer"); + } + } + } gpr_ref_init(&s->refs, 1); gpr_mu_init(&s->mu); s->active_ports = 0; s->destroyed_ports = 0; - s->shutdown = 0; + s->shutdown = false; s->shutdown_starting.head = NULL; s->shutdown_starting.tail = NULL; s->shutdown_complete = shutdown_complete; @@ -218,7 +248,7 @@ static void tcp_server_destroy(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s) { gpr_mu_lock(&s->mu); GPR_ASSERT(!s->shutdown); - s->shutdown = 1; + s->shutdown = true; /* shutdown all fd's */ if (s->active_ports) { @@ -268,13 +298,19 @@ static int get_max_accept_queue_size(void) { /* Prepare a recently-created socket for listening. */ static grpc_error *prepare_socket(int fd, const struct sockaddr *addr, - size_t addr_len, int *port) { + size_t addr_len, bool so_reuseport, + int *port) { struct sockaddr_storage sockname_temp; socklen_t sockname_len; grpc_error *err = GRPC_ERROR_NONE; GPR_ASSERT(fd >= 0); + if (so_reuseport) { + err = grpc_set_socket_reuse_port(fd, 1); + if (err != GRPC_ERROR_NONE) goto error; + } + err = grpc_set_socket_nonblocking(fd, 1); if (err != GRPC_ERROR_NONE) goto error; err = grpc_set_socket_cloexec(fd, 1); @@ -407,7 +443,7 @@ static grpc_error *add_socket_to_server(grpc_tcp_server *s, int fd, char *addr_str; char *name; - grpc_error *err = prepare_socket(fd, addr, addr_len, &port); + grpc_error *err = prepare_socket(fd, addr, addr_len, s->so_reuseport, &port); if (err == GRPC_ERROR_NONE) { GPR_ASSERT(port > 0); grpc_sockaddr_to_string(&addr_str, (struct sockaddr *)&addr, 1); @@ -443,6 +479,52 @@ static grpc_error *add_socket_to_server(grpc_tcp_server *s, int fd, return err; } +static grpc_error *clone_port(grpc_tcp_listener *listener, unsigned count) { + grpc_tcp_listener *sp = NULL; + char *addr_str; + char *name; + grpc_error *err; + + for (grpc_tcp_listener *l = listener->next; l && l->is_sibling; l = l->next) { + l->fd_index += count; + } + + for (unsigned i = 0; i < count; i++) { + int fd, port; + grpc_dualstack_mode dsmode; + err = grpc_create_dualstack_socket(&listener->addr.sockaddr, SOCK_STREAM, 0, + &dsmode, &fd); + if (err != GRPC_ERROR_NONE) return err; + err = prepare_socket(fd, &listener->addr.sockaddr, listener->addr_len, true, + &port); + if (err != GRPC_ERROR_NONE) return err; + listener->server->nports++; + grpc_sockaddr_to_string(&addr_str, &listener->addr.sockaddr, 1); + gpr_asprintf(&name, "tcp-server-listener:%s/clone-%d", addr_str, i); + sp = gpr_malloc(sizeof(grpc_tcp_listener)); + sp->next = listener->next; + listener->next = sp; + sp->server = listener->server; + sp->fd = fd; + sp->emfd = grpc_fd_create(fd, name); + memcpy(sp->addr.untyped, listener->addr.untyped, listener->addr_len); + sp->addr_len = listener->addr_len; + sp->port = port; + sp->port_index = listener->port_index; + sp->fd_index = listener->fd_index + count - i; + sp->is_sibling = 1; + sp->sibling = listener->is_sibling ? listener->sibling : listener; + GPR_ASSERT(sp->emfd); + while (listener->server->tail->next != NULL) { + listener->server->tail = listener->server->tail->next; + } + gpr_free(addr_str); + gpr_free(name); + } + + return GRPC_ERROR_NONE; +} + grpc_error *grpc_tcp_server_add_port(grpc_tcp_server *s, const void *addr, size_t addr_len, int *out_port) { grpc_tcp_listener *sp; @@ -599,14 +681,29 @@ void grpc_tcp_server_start(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s, s->on_accept_cb_arg = on_accept_cb_arg; s->pollsets = pollsets; s->pollset_count = pollset_count; - for (sp = s->head; sp; sp = sp->next) { - for (i = 0; i < pollset_count; i++) { - grpc_pollset_add_fd(exec_ctx, pollsets[i], sp->emfd); + sp = s->head; + while (sp != NULL) { + if (s->so_reuseport && pollset_count > 1) { + GPR_ASSERT(GRPC_LOG_IF_ERROR( + "clone_port", clone_port(sp, (unsigned)(pollset_count - 1)))); + for (i = 0; i < pollset_count; i++) { + grpc_pollset_add_fd(exec_ctx, pollsets[i], sp->emfd); + sp->read_closure.cb = on_read; + sp->read_closure.cb_arg = sp; + grpc_fd_notify_on_read(exec_ctx, sp->emfd, &sp->read_closure); + s->active_ports++; + sp = sp->next; + } + } else { + for (i = 0; i < pollset_count; i++) { + grpc_pollset_add_fd(exec_ctx, pollsets[i], sp->emfd); + } + sp->read_closure.cb = on_read; + sp->read_closure.cb_arg = sp; + grpc_fd_notify_on_read(exec_ctx, sp->emfd, &sp->read_closure); + s->active_ports++; + sp = sp->next; } - sp->read_closure.cb = on_read; - sp->read_closure.cb_arg = sp; - grpc_fd_notify_on_read(exec_ctx, sp->emfd, &sp->read_closure); - s->active_ports++; } gpr_mu_unlock(&s->mu); } diff --git a/src/core/lib/iomgr/tcp_server_windows.c b/src/core/lib/iomgr/tcp_server_windows.c index 86982bc183..7b0966704c 100644 --- a/src/core/lib/iomgr/tcp_server_windows.c +++ b/src/core/lib/iomgr/tcp_server_windows.c @@ -103,6 +103,7 @@ struct grpc_tcp_server { /* Public function. Allocates the proper data structures to hold a grpc_tcp_server. */ grpc_error *grpc_tcp_server_create(grpc_closure *shutdown_complete, + const grpc_channel_args *args, grpc_tcp_server **server) { grpc_tcp_server *s = gpr_malloc(sizeof(grpc_tcp_server)); gpr_ref_init(&s->refs, 1); diff --git a/src/core/lib/iomgr/tcp_windows.c b/src/core/lib/iomgr/tcp_windows.c index b2af8030aa..37ab59021e 100644 --- a/src/core/lib/iomgr/tcp_windows.c +++ b/src/core/lib/iomgr/tcp_windows.c @@ -37,6 +37,7 @@ #include <limits.h> +#include "src/core/lib/iomgr/network_status_tracker.h" #include "src/core/lib/iomgr/sockaddr_windows.h" #include <grpc/support/alloc.h> @@ -378,6 +379,7 @@ static void win_shutdown(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep) { } static void win_destroy(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep) { + grpc_network_status_unregister_endpoint(ep); grpc_tcp *tcp = (grpc_tcp *)ep; TCP_UNREF(tcp, "destroy"); } @@ -401,6 +403,9 @@ grpc_endpoint *grpc_tcp_create(grpc_winsocket *socket, char *peer_string) { grpc_closure_init(&tcp->on_read, on_read, tcp); grpc_closure_init(&tcp->on_write, on_write, tcp); tcp->peer_string = gpr_strdup(peer_string); + /* Tell network status tracking code about the new endpoint */ + grpc_network_status_register_endpoint(&tcp->base); + return &tcp->base; } diff --git a/src/core/lib/iomgr/udp_server.c b/src/core/lib/iomgr/udp_server.c index 16150687d3..1ebccf2ee2 100644 --- a/src/core/lib/iomgr/udp_server.c +++ b/src/core/lib/iomgr/udp_server.c @@ -243,13 +243,13 @@ static int prepare_socket(int fd, const struct sockaddr *addr, if (!grpc_set_socket_sndbuf(fd, buffer_size_bytes)) { gpr_log(GPR_ERROR, "Failed to set send buffer size to %d bytes", - buf_size_bytes); + buffer_size_bytes); goto error; } if (!grpc_set_socket_rcvbuf(fd, buffer_size_bytes)) { gpr_log(GPR_ERROR, "Failed to set receive buffer size to %d bytes", - buf_size_bytes); + buffer_size_bytes); goto error; } diff --git a/src/core/lib/profiling/basic_timers.c b/src/core/lib/profiling/basic_timers.c index 50082cd7ee..51813d0461 100644 --- a/src/core/lib/profiling/basic_timers.c +++ b/src/core/lib/profiling/basic_timers.c @@ -141,11 +141,11 @@ static void write_log(gpr_timer_log *log) { entry->tm = gpr_time_0(entry->tm.clock_type); } fprintf(output_file, - "{\"t\": %lld.%09d, \"thd\": \"%d\", \"type\": \"%c\", \"tag\": " + "{\"t\": %" PRId64 + ".%09d, \"thd\": \"%d\", \"type\": \"%c\", \"tag\": " "\"%s\", \"file\": \"%s\", \"line\": %d, \"imp\": %d}\n", - (long long)entry->tm.tv_sec, (int)entry->tm.tv_nsec, entry->thd, - entry->type, entry->tagstr, entry->file, entry->line, - entry->important); + entry->tm.tv_sec, entry->tm.tv_nsec, entry->thd, entry->type, + entry->tagstr, entry->file, entry->line, entry->important); } } diff --git a/src/core/lib/security/credentials/composite/composite_credentials.c b/src/core/lib/security/credentials/composite/composite_credentials.c index 07db8bfd75..850e41e646 100644 --- a/src/core/lib/security/credentials/composite/composite_credentials.c +++ b/src/core/lib/security/credentials/composite/composite_credentials.c @@ -72,11 +72,12 @@ static void composite_call_md_context_destroy( static void composite_call_metadata_cb(grpc_exec_ctx *exec_ctx, void *user_data, grpc_credentials_md *md_elems, size_t num_md, - grpc_credentials_status status) { + grpc_credentials_status status, + const char *error_details) { grpc_composite_call_credentials_metadata_context *ctx = (grpc_composite_call_credentials_metadata_context *)user_data; if (status != GRPC_CREDENTIALS_OK) { - ctx->cb(exec_ctx, ctx->user_data, NULL, 0, status); + ctx->cb(exec_ctx, ctx->user_data, NULL, 0, status, error_details); return; } @@ -101,7 +102,7 @@ static void composite_call_metadata_cb(grpc_exec_ctx *exec_ctx, void *user_data, /* We're done!. */ ctx->cb(exec_ctx, ctx->user_data, ctx->md_elems->entries, - ctx->md_elems->num_entries, GRPC_CREDENTIALS_OK); + ctx->md_elems->num_entries, GRPC_CREDENTIALS_OK, NULL); composite_call_md_context_destroy(ctx); } diff --git a/src/core/lib/security/credentials/credentials.c b/src/core/lib/security/credentials/credentials.c index 0eadaec191..029a357261 100644 --- a/src/core/lib/security/credentials/credentials.c +++ b/src/core/lib/security/credentials/credentials.c @@ -117,7 +117,7 @@ void grpc_call_credentials_get_request_metadata( grpc_credentials_metadata_cb cb, void *user_data) { if (creds == NULL || creds->vtable->get_request_metadata == NULL) { if (cb != NULL) { - cb(exec_ctx, user_data, NULL, 0, GRPC_CREDENTIALS_OK); + cb(exec_ctx, user_data, NULL, 0, GRPC_CREDENTIALS_OK, NULL); } return; } diff --git a/src/core/lib/security/credentials/credentials.h b/src/core/lib/security/credentials/credentials.h index ce235e3a1d..8e9d842ead 100644 --- a/src/core/lib/security/credentials/credentials.h +++ b/src/core/lib/security/credentials/credentials.h @@ -156,11 +156,10 @@ void grpc_credentials_md_store_unref(grpc_credentials_md_store *store); /* --- grpc_call_credentials. --- */ -typedef void (*grpc_credentials_metadata_cb)(grpc_exec_ctx *exec_ctx, - void *user_data, - grpc_credentials_md *md_elems, - size_t num_md, - grpc_credentials_status status); +/* error_details must be NULL if status is GRPC_CREDENTIALS_OK. */ +typedef void (*grpc_credentials_metadata_cb)( + grpc_exec_ctx *exec_ctx, void *user_data, grpc_credentials_md *md_elems, + size_t num_md, grpc_credentials_status status, const char *error_details); typedef struct { void (*destruct)(grpc_call_credentials *c); diff --git a/src/core/lib/security/credentials/fake/fake_credentials.c b/src/core/lib/security/credentials/fake/fake_credentials.c index ee6d964de1..51cafd986f 100644 --- a/src/core/lib/security/credentials/fake/fake_credentials.c +++ b/src/core/lib/security/credentials/fake/fake_credentials.c @@ -100,7 +100,7 @@ static void on_simulated_token_fetch_done(grpc_exec_ctx *exec_ctx, (grpc_credentials_metadata_request *)user_data; grpc_md_only_test_credentials *c = (grpc_md_only_test_credentials *)r->creds; r->cb(exec_ctx, r->user_data, c->md_store->entries, c->md_store->num_entries, - GRPC_CREDENTIALS_OK); + GRPC_CREDENTIALS_OK, NULL); grpc_credentials_metadata_request_destroy(r); } @@ -117,7 +117,7 @@ static void md_only_test_get_request_metadata( grpc_closure_create(on_simulated_token_fetch_done, cb_arg), GRPC_ERROR_NONE); } else { - cb(exec_ctx, user_data, c->md_store->entries, 1, GRPC_CREDENTIALS_OK); + cb(exec_ctx, user_data, c->md_store->entries, 1, GRPC_CREDENTIALS_OK, NULL); } } diff --git a/src/core/lib/security/credentials/iam/iam_credentials.c b/src/core/lib/security/credentials/iam/iam_credentials.c index 64d5871844..370a384d0e 100644 --- a/src/core/lib/security/credentials/iam/iam_credentials.c +++ b/src/core/lib/security/credentials/iam/iam_credentials.c @@ -55,7 +55,7 @@ static void iam_get_request_metadata(grpc_exec_ctx *exec_ctx, void *user_data) { grpc_google_iam_credentials *c = (grpc_google_iam_credentials *)creds; cb(exec_ctx, user_data, c->iam_md->entries, c->iam_md->num_entries, - GRPC_CREDENTIALS_OK); + GRPC_CREDENTIALS_OK, NULL); } static grpc_call_credentials_vtable iam_vtable = {iam_destruct, diff --git a/src/core/lib/security/credentials/jwt/jwt_credentials.c b/src/core/lib/security/credentials/jwt/jwt_credentials.c index 973fb75eaa..f87ba0ce8d 100644 --- a/src/core/lib/security/credentials/jwt/jwt_credentials.c +++ b/src/core/lib/security/credentials/jwt/jwt_credentials.c @@ -113,10 +113,11 @@ static void jwt_get_request_metadata(grpc_exec_ctx *exec_ctx, if (jwt_md != NULL) { cb(exec_ctx, user_data, jwt_md->entries, jwt_md->num_entries, - GRPC_CREDENTIALS_OK); + GRPC_CREDENTIALS_OK, NULL); grpc_credentials_md_store_unref(jwt_md); } else { - cb(exec_ctx, user_data, NULL, 0, GRPC_CREDENTIALS_ERROR); + cb(exec_ctx, user_data, NULL, 0, GRPC_CREDENTIALS_ERROR, + "Could not generate JWT."); } } @@ -149,11 +150,11 @@ grpc_call_credentials *grpc_service_account_jwt_access_credentials_create( "grpc_service_account_jwt_access_credentials_create(" "json_key=%s, " "token_lifetime=" - "gpr_timespec { tv_sec: %lld, tv_nsec: %d, clock_type: %d }, " + "gpr_timespec { tv_sec: %" PRId64 + ", tv_nsec: %d, clock_type: %d }, " "reserved=%p)", - 5, - (json_key, (long long)token_lifetime.tv_sec, (int)token_lifetime.tv_nsec, - (int)token_lifetime.clock_type, reserved)); + 5, (json_key, token_lifetime.tv_sec, token_lifetime.tv_nsec, + (int)token_lifetime.clock_type, reserved)); GPR_ASSERT(reserved == NULL); return grpc_service_account_jwt_access_credentials_create_from_auth_json_key( grpc_auth_json_key_create_from_string(json_key), token_lifetime); diff --git a/src/core/lib/security/credentials/oauth2/oauth2_credentials.c b/src/core/lib/security/credentials/oauth2/oauth2_credentials.c index 1102553dd3..c22ea5c468 100644 --- a/src/core/lib/security/credentials/oauth2/oauth2_credentials.c +++ b/src/core/lib/security/credentials/oauth2/oauth2_credentials.c @@ -235,10 +235,11 @@ static void on_oauth2_token_fetcher_http_response(grpc_exec_ctx *exec_ctx, c->token_expiration = gpr_time_add(gpr_now(GPR_CLOCK_REALTIME), token_lifetime); r->cb(exec_ctx, r->user_data, c->access_token_md->entries, - c->access_token_md->num_entries, status); + c->access_token_md->num_entries, GRPC_CREDENTIALS_OK, NULL); } else { c->token_expiration = gpr_inf_past(GPR_CLOCK_REALTIME); - r->cb(exec_ctx, r->user_data, NULL, 0, status); + r->cb(exec_ctx, r->user_data, NULL, 0, status, + "Error occured when fetching oauth2 token."); } gpr_mu_unlock(&c->mu); grpc_credentials_metadata_request_destroy(r); @@ -266,7 +267,7 @@ static void oauth2_token_fetcher_get_request_metadata( } if (cached_access_token_md != NULL) { cb(exec_ctx, user_data, cached_access_token_md->entries, - cached_access_token_md->num_entries, GRPC_CREDENTIALS_OK); + cached_access_token_md->num_entries, GRPC_CREDENTIALS_OK, NULL); grpc_credentials_md_store_unref(cached_access_token_md); } else { c->fetch_func( @@ -404,7 +405,8 @@ static void access_token_get_request_metadata( grpc_polling_entity *pollent, grpc_auth_metadata_context context, grpc_credentials_metadata_cb cb, void *user_data) { grpc_access_token_credentials *c = (grpc_access_token_credentials *)creds; - cb(exec_ctx, user_data, c->access_token_md->entries, 1, GRPC_CREDENTIALS_OK); + cb(exec_ctx, user_data, c->access_token_md->entries, 1, GRPC_CREDENTIALS_OK, + NULL); } static grpc_call_credentials_vtable access_token_vtable = { diff --git a/src/core/lib/security/credentials/plugin/plugin_credentials.c b/src/core/lib/security/credentials/plugin/plugin_credentials.c index 9fb55e8466..824ff081dc 100644 --- a/src/core/lib/security/credentials/plugin/plugin_credentials.c +++ b/src/core/lib/security/credentials/plugin/plugin_credentials.c @@ -67,7 +67,8 @@ static void plugin_md_request_metadata_ready(void *request, gpr_log(GPR_ERROR, "Getting metadata from plugin failed with error: %s", error_details); } - r->cb(&exec_ctx, r->user_data, NULL, 0, GRPC_CREDENTIALS_ERROR); + r->cb(&exec_ctx, r->user_data, NULL, 0, GRPC_CREDENTIALS_ERROR, + error_details); } else { size_t i; grpc_credentials_md *md_array = NULL; @@ -79,7 +80,7 @@ static void plugin_md_request_metadata_ready(void *request, gpr_slice_from_copied_buffer(md[i].value, md[i].value_length); } } - r->cb(&exec_ctx, r->user_data, md_array, num_md, GRPC_CREDENTIALS_OK); + r->cb(&exec_ctx, r->user_data, md_array, num_md, GRPC_CREDENTIALS_OK, NULL); if (md_array != NULL) { for (i = 0; i < num_md; i++) { gpr_slice_unref(md_array[i].key); @@ -107,7 +108,7 @@ static void plugin_get_request_metadata(grpc_exec_ctx *exec_ctx, c->plugin.get_metadata(c->plugin.state, context, plugin_md_request_metadata_ready, request); } else { - cb(exec_ctx, user_data, NULL, 0, GRPC_CREDENTIALS_OK); + cb(exec_ctx, user_data, NULL, 0, GRPC_CREDENTIALS_OK, NULL); } } diff --git a/src/core/lib/security/transport/client_auth_filter.c b/src/core/lib/security/transport/client_auth_filter.c index 76be2acd72..399b92c8e1 100644 --- a/src/core/lib/security/transport/client_auth_filter.c +++ b/src/core/lib/security/transport/client_auth_filter.c @@ -91,14 +91,16 @@ static void bubble_up_error(grpc_exec_ctx *exec_ctx, grpc_call_element *elem, grpc_status_code status, const char *error_msg) { call_data *calld = elem->call_data; gpr_log(GPR_ERROR, "Client side authentication failure: %s", error_msg); - grpc_transport_stream_op_add_cancellation(&calld->op, status); + gpr_slice error_slice = gpr_slice_from_copied_string(error_msg); + grpc_transport_stream_op_add_close(&calld->op, status, &error_slice); grpc_call_next_op(exec_ctx, elem, &calld->op); } static void on_credentials_metadata(grpc_exec_ctx *exec_ctx, void *user_data, grpc_credentials_md *md_elems, size_t num_md, - grpc_credentials_status status) { + grpc_credentials_status status, + const char *error_details) { grpc_call_element *elem = (grpc_call_element *)user_data; call_data *calld = elem->call_data; grpc_transport_stream_op *op = &calld->op; @@ -107,7 +109,9 @@ static void on_credentials_metadata(grpc_exec_ctx *exec_ctx, void *user_data, reset_auth_metadata_context(&calld->auth_md_context); if (status != GRPC_CREDENTIALS_OK) { bubble_up_error(exec_ctx, elem, GRPC_STATUS_UNAUTHENTICATED, - "Credentials failed to get metadata."); + (error_details != NULL && strlen(error_details) > 0) + ? error_details + : "Credentials failed to get metadata."); return; } GPR_ASSERT(num_md <= MAX_CREDENTIALS_METADATA_COUNT); diff --git a/src/core/lib/surface/channel.c b/src/core/lib/surface/channel.c index 8936e7164f..2cf6d8890a 100644 --- a/src/core/lib/surface/channel.c +++ b/src/core/lib/surface/channel.c @@ -223,11 +223,12 @@ grpc_call *grpc_channel_create_call(grpc_channel *channel, "grpc_channel_create_call(" "channel=%p, parent_call=%p, propagation_mask=%x, cq=%p, method=%s, " "host=%s, " - "deadline=gpr_timespec { tv_sec: %lld, tv_nsec: %d, clock_type: %d }, " + "deadline=gpr_timespec { tv_sec: %" PRId64 + ", tv_nsec: %d, clock_type: %d }, " "reserved=%p)", - 10, (channel, parent_call, (unsigned)propagation_mask, cq, method, host, - (long long)deadline.tv_sec, (int)deadline.tv_nsec, - (int)deadline.clock_type, reserved)); + 10, + (channel, parent_call, (unsigned)propagation_mask, cq, method, host, + deadline.tv_sec, deadline.tv_nsec, (int)deadline.clock_type, reserved)); GPR_ASSERT(!reserved); return grpc_channel_create_call_internal( channel, parent_call, propagation_mask, cq, NULL, @@ -282,11 +283,12 @@ grpc_call *grpc_channel_create_registered_call( "grpc_channel_create_registered_call(" "channel=%p, parent_call=%p, propagation_mask=%x, completion_queue=%p, " "registered_call_handle=%p, " - "deadline=gpr_timespec { tv_sec: %lld, tv_nsec: %d, clock_type: %d }, " + "deadline=gpr_timespec { tv_sec: %" PRId64 + ", tv_nsec: %d, clock_type: %d }, " "reserved=%p)", 9, (channel, parent_call, (unsigned)propagation_mask, completion_queue, - registered_call_handle, (long long)deadline.tv_sec, - (int)deadline.tv_nsec, (int)deadline.clock_type, reserved)); + registered_call_handle, deadline.tv_sec, deadline.tv_nsec, + (int)deadline.clock_type, reserved)); GPR_ASSERT(!reserved); return grpc_channel_create_call_internal( channel, parent_call, propagation_mask, completion_queue, NULL, diff --git a/src/core/lib/surface/completion_queue.c b/src/core/lib/surface/completion_queue.c index 71495fa937..1eeacf81a0 100644 --- a/src/core/lib/surface/completion_queue.c +++ b/src/core/lib/surface/completion_queue.c @@ -326,10 +326,11 @@ grpc_event grpc_completion_queue_next(grpc_completion_queue *cc, GRPC_API_TRACE( "grpc_completion_queue_next(" "cc=%p, " - "deadline=gpr_timespec { tv_sec: %lld, tv_nsec: %d, clock_type: %d }, " + "deadline=gpr_timespec { tv_sec: %" PRId64 + ", tv_nsec: %d, clock_type: %d }, " "reserved=%p)", - 5, (cc, (long long)deadline.tv_sec, (int)deadline.tv_nsec, - (int)deadline.clock_type, reserved)); + 5, (cc, deadline.tv_sec, deadline.tv_nsec, (int)deadline.clock_type, + reserved)); GPR_ASSERT(!reserved); deadline = gpr_convert_clock_type(deadline, GPR_CLOCK_MONOTONIC); @@ -439,9 +440,10 @@ grpc_event grpc_completion_queue_pluck(grpc_completion_queue *cc, void *tag, GRPC_API_TRACE( "grpc_completion_queue_pluck(" "cc=%p, tag=%p, " - "deadline=gpr_timespec { tv_sec: %lld, tv_nsec: %d, clock_type: %d }, " + "deadline=gpr_timespec { tv_sec: %" PRId64 + ", tv_nsec: %d, clock_type: %d }, " "reserved=%p)", - 6, (cc, tag, (long long)deadline.tv_sec, (int)deadline.tv_nsec, + 6, (cc, tag, deadline.tv_sec, deadline.tv_nsec, (int)deadline.clock_type, reserved)); } GPR_ASSERT(!reserved); diff --git a/src/core/lib/surface/version.c b/src/core/lib/surface/version.c index aca76d2bb7..53f3c43854 100644 --- a/src/core/lib/surface/version.c +++ b/src/core/lib/surface/version.c @@ -36,4 +36,4 @@ #include <grpc/grpc.h> -const char *grpc_version_string(void) { return "0.15.0-dev"; } +const char *grpc_version_string(void) { return "0.16.0-dev"; } diff --git a/src/core/lib/transport/transport_op_string.c b/src/core/lib/transport/transport_op_string.c index df04c61127..aeaba5339f 100644 --- a/src/core/lib/transport/transport_op_string.c +++ b/src/core/lib/transport/transport_op_string.c @@ -63,8 +63,8 @@ static void put_metadata_list(gpr_strvec *b, grpc_metadata_batch md) { } if (gpr_time_cmp(md.deadline, gpr_inf_future(md.deadline.clock_type)) != 0) { char *tmp; - gpr_asprintf(&tmp, " deadline=%lld.%09d", (long long)md.deadline.tv_sec, - (int)md.deadline.tv_nsec); + gpr_asprintf(&tmp, " deadline=%" PRId64 ".%09d", md.deadline.tv_sec, + md.deadline.tv_nsec); gpr_strvec_add(b, tmp); } } |