From 1898a29f7ecf2ac4f706351d68dedd5767f02ec9 Mon Sep 17 00:00:00 2001 From: Sree Kuchibhotla Date: Wed, 19 Apr 2017 19:37:39 -0700 Subject: [PATCH 1/7] Fixed pollers --- include/grpc++/server_builder.h | 2 +- src/core/lib/iomgr/ev_epoll_linux.c | 362 +++++++++++++++++++------- test/core/iomgr/ev_epoll_linux_test.c | 46 +++- 3 files changed, 320 insertions(+), 90 deletions(-) diff --git a/include/grpc++/server_builder.h b/include/grpc++/server_builder.h index 7ac23349c84..68c5344e2a1 100644 --- a/include/grpc++/server_builder.h +++ b/include/grpc++/server_builder.h @@ -195,7 +195,7 @@ class ServerBuilder { struct SyncServerSettings { SyncServerSettings() - : num_cqs(1), min_pollers(1), max_pollers(2), cq_timeout_msec(10000) {} + : num_cqs(gpr_cpu_num_cores()), min_pollers(1), max_pollers(2), cq_timeout_msec(10000) {} // Number of server completion queues to create to listen to incoming RPCs. int num_cqs; diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c index e603a755937..16a9199ab92 100644 --- a/src/core/lib/iomgr/ev_epoll_linux.c +++ b/src/core/lib/iomgr/ev_epoll_linux.c @@ -97,6 +97,9 @@ void grpc_use_signal(int signum) { } } +/* The maximum number of polling threads per polling island */ +#define GRPC_MAX_POLLERS_PER_ISLAND 1 + struct polling_island; typedef enum { @@ -195,6 +198,11 @@ static void fd_global_shutdown(void); #endif /* !defined(GRPC_PI_REF_COUNT_DEBUG) */ +typedef struct worker_node { + struct worker_node *next; + struct worker_node *prev; +} worker_node; + /* This is also used as grpc_workqueue (by directly casing it) */ typedef struct polling_island { grpc_closure_scheduler workqueue_scheduler; @@ -229,6 +237,9 @@ typedef struct polling_island { /* Wakeup fd used to wake pollers to check the contents of workqueue_items */ grpc_wakeup_fd workqueue_wakeup_fd; + gpr_mu worker_list_mu; + worker_node worker_list_head; + /* The fd of the underlying epoll set */ int epoll_fd; @@ -241,14 +252,21 @@ typedef struct polling_island { /******************************************************************************* * Pollset Declarations */ +#define WORKER_FROM_WORKER_LIST_NODE(p) \ + (struct grpc_pollset_worker *)(((char *)(p)) - \ + offsetof(grpc_pollset_worker, pi_list_link)) struct grpc_pollset_worker { /* Thread id of this worker */ pthread_t pt_id; /* Used to prevent a worker from getting kicked multiple times */ gpr_atm is_kicked; + struct grpc_pollset_worker *next; struct grpc_pollset_worker *prev; + + gpr_atm is_polling_turn; + worker_node pi_list_link; }; struct grpc_pollset { @@ -392,7 +410,47 @@ static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi) { } } -/* The caller is expected to hold pi->mu lock before calling this function */ +static void worker_node_init(worker_node *node) { + node->next = node->prev = node; +} + +/* Not thread safe. Do under a list-level lock */ +static void push_back_worker_node(worker_node *head, worker_node *node) { + node->next = head; + node->prev = head->prev; + head->prev->next = node; + head->prev = node; +} + +/* Not thread safe. Do under a list-level lock */ +static void remove_worker_node(worker_node *node) { + node->next->prev = node->prev; + node->prev->next = node->next; + /* If node's next and prev point to itself, the node is considered detached + * from the list*/ + node->next = node->prev = node; +} + +/* Not thread safe. Do under a list-level lock */ +static worker_node *pop_front_worker_node(worker_node *head) { + worker_node *node = head->next; + if (node != head) { + remove_worker_node(node); + } else { + node = NULL; + } + + return node; +} + +/* Returns true if the node's next and prev are pointing to itself (which + indicates that the node is not in the list */ +static bool is_worker_node_detached(worker_node *node) { + return (node->next == node->prev && node->next == node); +} + +/* 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) { @@ -546,6 +604,9 @@ static polling_island *polling_island_create(grpc_exec_ctx *exec_ctx, gpr_atm_rel_store(&pi->poller_count, 0); gpr_atm_rel_store(&pi->merged_to, (gpr_atm)NULL); + gpr_mu_init(&pi->worker_list_mu); + worker_node_init(&pi->worker_list_head); + if (!append_error(error, grpc_wakeup_fd_init(&pi->workqueue_wakeup_fd), err_desc)) { goto done; @@ -584,6 +645,9 @@ static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi) { gpr_mpscq_destroy(&pi->workqueue_items); gpr_mu_destroy(&pi->mu); grpc_wakeup_fd_destroy(&pi->workqueue_wakeup_fd); + gpr_mu_destroy(&pi->worker_list_mu); + GPR_ASSERT(is_worker_node_detached(&pi->worker_list_head)); + gpr_free(pi->fds); gpr_free(pi); } @@ -1102,6 +1166,7 @@ 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 __thread sigset_t g_wakeup_sig_set; static void sig_handler(int sig_num) { #ifdef GRPC_EPOLL_DEBUG @@ -1109,6 +1174,14 @@ static void sig_handler(int sig_num) { #endif } +static void pollset_worker_init(grpc_pollset_worker *worker) { + worker->pt_id = pthread_self(); + worker->next = worker->prev = NULL; + gpr_atm_no_barrier_store(&worker->is_kicked, (gpr_atm)0); + gpr_atm_no_barrier_store(&worker->is_polling_turn, (gpr_atm)0); + worker_node_init(&worker->pi_list_link); +} + static void poller_kick_init() { signal(grpc_wakeup_signal, sig_handler); } /* Global state management */ @@ -1125,11 +1198,12 @@ static void pollset_global_shutdown(void) { gpr_tls_destroy(&g_current_thread_worker); } -static grpc_error *pollset_worker_kick(grpc_pollset_worker *worker) { +static grpc_error *worker_kick(grpc_pollset_worker *worker, + gpr_atm *is_kicked) { grpc_error *err = GRPC_ERROR_NONE; /* Kick the worker only if it was not already kicked */ - if (gpr_atm_no_barrier_cas(&worker->is_kicked, (gpr_atm)0, (gpr_atm)1)) { + if (gpr_atm_no_barrier_cas(is_kicked, (gpr_atm)0, (gpr_atm)1)) { GRPC_POLLING_TRACE( "pollset_worker_kick: Kicking worker: %p (thread id: %ld)", (void *)worker, (long int)worker->pt_id); @@ -1141,6 +1215,14 @@ static grpc_error *pollset_worker_kick(grpc_pollset_worker *worker) { return err; } +static grpc_error *pollset_worker_kick(grpc_pollset_worker *worker) { + return worker_kick(worker, &worker->is_kicked); +} + +static grpc_error *poller_kick(grpc_pollset_worker *worker) { + return worker_kick(worker, &worker->is_polling_turn); +} + /* Return 1 if the pollset has active threads in pollset_work (pollset must * be locked) */ static int pollset_has_workers(grpc_pollset *p) { @@ -1246,6 +1328,22 @@ static void pollset_init(grpc_pollset *pollset, gpr_mu **mu) { pollset->shutdown_done = NULL; } +/* Convert millis to timespec (clock-type is assumed to be GPR_TIMESPAN) */ +static struct timespec millis_to_timespec(int millis) { + struct timespec linux_ts; + gpr_timespec gpr_ts; + + if (millis == -1) { + gpr_ts = gpr_inf_future(GPR_TIMESPAN); + } else { + gpr_ts = gpr_time_from_millis(millis, GPR_TIMESPAN); + } + + linux_ts.tv_sec = (time_t)gpr_ts.tv_sec; + linux_ts.tv_nsec = gpr_ts.tv_nsec; + return linux_ts; +} + /* 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) @@ -1364,35 +1462,190 @@ static bool maybe_do_workqueue_work(grpc_exec_ctx *exec_ctx, return false; } +/* NOTE: May modify 'now' */ +static bool acquire_polling_lease(grpc_pollset_worker *worker, + polling_island *pi, gpr_timespec deadline, + gpr_timespec *now) { + bool is_lease_acquired = false; + + gpr_mu_lock(&pi->worker_list_mu); // Lock + long num_pollers = gpr_atm_no_barrier_load(&pi->poller_count); + + if (num_pollers >= GRPC_MAX_POLLERS_PER_ISLAND) { + push_back_worker_node(&pi->worker_list_head, &worker->pi_list_link); + gpr_mu_unlock(&pi->worker_list_mu); // Unlock + + bool is_timeout = false; + int ret; + int timeout_ms = poll_deadline_to_millis_timeout(deadline, *now); + if (timeout_ms == -1) { + ret = sigwaitinfo(&g_wakeup_sig_set, NULL); + } else { + struct timespec sigwait_timeout = millis_to_timespec(timeout_ms); + ret = sigtimedwait(&g_wakeup_sig_set, NULL, &sigwait_timeout); + } + + if (ret == -1) { + if (errno == EAGAIN) { + // gpr_log(GPR_INFO, "timeout"); // TODO: sreek remove this + // log-line + } else { + gpr_log(GPR_ERROR, "Failed with retcode: %d (timeout_ms: %d)", errno, + timeout_ms); + } + is_timeout = true; + } + + bool is_polling_turn = gpr_atm_acq_load(&worker->is_polling_turn); + /* + if (is_polling_turn) { +/ gpr_log(GPR_ERROR, "do epoll is true (timeout_ms:%d)", + timeout_ms); // TODO: sreek remove this logline + } + */ + + bool is_kicked = gpr_atm_no_barrier_load(&worker->is_kicked); + if (is_kicked || is_timeout) { + *now = deadline; + } else if (is_polling_turn) { + *now = gpr_now(GPR_CLOCK_MONOTONIC); + } + + gpr_mu_lock(&pi->worker_list_mu); // Lock + /* The node might have already been removed from the list by the poller + that kicked this. However it is safe to call 'remove_worker_node' on + an already detached node */ + remove_worker_node(&worker->pi_list_link); + num_pollers = gpr_atm_no_barrier_load(&pi->poller_count); + } + + if (num_pollers < GRPC_MAX_POLLERS_PER_ISLAND) { + gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1); // Add a poller + is_lease_acquired = true; + } + + gpr_mu_unlock(&pi->worker_list_mu); // Unlock + return is_lease_acquired; +} + +static void release_polling_lease(polling_island *pi, grpc_error **error) { + gpr_mu_lock(&pi->worker_list_mu); // Lock + + gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1); // Remove poller + worker_node *node = pop_front_worker_node(&pi->worker_list_head); + if (node != NULL) { + grpc_pollset_worker *next_worker = WORKER_FROM_WORKER_LIST_NODE(node); + append_error(error, poller_kick(next_worker), "poller kick error"); + } + + gpr_mu_unlock(&pi->worker_list_mu); +} + #define GRPC_EPOLL_MAX_EVENTS 100 +static void pollset_do_epoll_pwait(grpc_exec_ctx *exec_ctx, int epoll_fd, + grpc_pollset *pollset, polling_island *pi, + grpc_pollset_worker *worker, + gpr_timespec now, gpr_timespec deadline, + sigset_t *sig_mask, grpc_error **error) { + if (!acquire_polling_lease(worker, pi, deadline, &now)) { + return; + } + + struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS]; + int ep_rv; + char *err_msg; + const char *err_desc = "pollset_work_and_unlock"; + + int timeout_ms = poll_deadline_to_millis_timeout(deadline, now); + + GRPC_SCHEDULING_START_BLOCKING_REGION; + ep_rv = + epoll_pwait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms, sig_mask); + GRPC_SCHEDULING_END_BLOCKING_REGION; + + release_polling_lease(pi, error); + + 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 */ + GRPC_POLLING_TRACE("pollset_work: pollset: %p, worker: %p received kick", + (void *)pollset, (void *)worker); + 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 == &global_wakeup_fd) { + grpc_timer_consume_kick(); + append_error(error, grpc_wakeup_fd_consume_wakeup(&global_wakeup_fd), + err_desc); + } else if (data_ptr == &pi->workqueue_wakeup_fd) { + append_error(error, + grpc_wakeup_fd_consume_wakeup(&pi->workqueue_wakeup_fd), + err_desc); + maybe_do_workqueue_work(exec_ctx, pi); + } else if (data_ptr == &polling_island_wakeup_fd) { + GRPC_POLLING_TRACE( + "pollset_work: pollset: %p, worker: %p polling island (epoll_fd: " + "%d) got merged", + (void *)pollset, (void *)worker, epoll_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); + } + } + } +} + /* 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, - grpc_pollset_worker *worker, int timeout_ms, + grpc_pollset_worker *worker, + gpr_timespec now, gpr_timespec deadline, 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->po.pi - 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 */ + Since epoll_fd is immutable, it is safe to read it without a lock on the + polling island. There is however a possibility that the polling island from + which we got the epoll_fd, got merged with another island in the meantime. + This is okay because in such a case, we will wakeup right-away from + epoll_pwait() (because any merge will poison the old polling island's epoll + set 'polling_island_wakeup_fd') and then pick up the latest polling_island + the next time this function - pollset_work_and_unlock()) is called */ if (pollset->po.pi == NULL) { pollset->po.pi = polling_island_create(exec_ctx, NULL, error); if (pollset->po.pi == NULL) { GPR_TIMER_END("pollset_work_and_unlock", 0); - return; /* Fatal error. We cannot continue */ + return; /* Fatal error. Cannot continue */ } PI_ADD_REF(pollset->po.pi, "ps"); @@ -1423,70 +1676,10 @@ static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx, the completion queue, so there's no need to poll... so we skip that and redo the complete loop to verify */ if (!maybe_do_workqueue_work(exec_ctx, pi)) { - gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1); g_current_thread_polling_island = pi; - - GRPC_SCHEDULING_START_BLOCKING_REGION; - ep_rv = epoll_pwait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms, - sig_mask); - GRPC_SCHEDULING_END_BLOCKING_REGION; - 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 */ - GRPC_POLLING_TRACE( - "pollset_work: pollset: %p, worker: %p received kick", - (void *)pollset, (void *)worker); - 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 == &global_wakeup_fd) { - grpc_timer_consume_kick(); - append_error(error, grpc_wakeup_fd_consume_wakeup(&global_wakeup_fd), - err_desc); - } else if (data_ptr == &pi->workqueue_wakeup_fd) { - append_error(error, - grpc_wakeup_fd_consume_wakeup(&pi->workqueue_wakeup_fd), - err_desc); - maybe_do_workqueue_work(exec_ctx, pi); - } else if (data_ptr == &polling_island_wakeup_fd) { - GRPC_POLLING_TRACE( - "pollset_work: pollset: %p, worker: %p polling island (epoll_fd: " - "%d) got merged", - (void *)pollset, (void *)worker, epoll_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); - } - } - } - + pollset_do_epoll_pwait(exec_ctx, epoll_fd, pollset, pi, worker, now, + deadline, sig_mask, error); g_current_thread_polling_island = NULL; - gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1); } GPR_ASSERT(pi != NULL); @@ -1510,14 +1703,9 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, 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(); - gpr_atm_no_barrier_store(&worker.is_kicked, (gpr_atm)0); + pollset_worker_init(&worker); if (worker_hdl) *worker_hdl = &worker; @@ -1551,9 +1739,9 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, 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); + sigemptyset(&g_wakeup_sig_set); + sigaddset(&g_wakeup_sig_set, grpc_wakeup_signal); + pthread_sigmask(SIG_BLOCK, &g_wakeup_sig_set, &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'. @@ -1568,7 +1756,7 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, push_front_worker(pollset, &worker); /* Add worker to pollset */ - pollset_work_and_unlock(exec_ctx, pollset, &worker, timeout_ms, + pollset_work_and_unlock(exec_ctx, pollset, &worker, now, deadline, &g_orig_sigmask, &error); grpc_exec_ctx_flush(exec_ctx); diff --git a/test/core/iomgr/ev_epoll_linux_test.c b/test/core/iomgr/ev_epoll_linux_test.c index 0856023b14f..f7745cddb4e 100644 --- a/test/core/iomgr/ev_epoll_linux_test.c +++ b/test/core/iomgr/ev_epoll_linux_test.c @@ -38,7 +38,10 @@ #include "src/core/lib/iomgr/ev_posix.h" #include +#include +#include #include +#include #include #include @@ -327,7 +330,7 @@ static __thread int thread_wakeups = 0; static void test_threading_loop(void *arg) { threading_shared *shared = arg; - while (thread_wakeups < 1000000) { + while (thread_wakeups < 20000) { grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT; grpc_pollset_worker *worker; gpr_mu_lock(shared->mu); @@ -360,7 +363,7 @@ static void test_threading(void) { shared.pollset = gpr_zalloc(grpc_pollset_size()); grpc_pollset_init(shared.pollset, &shared.mu); - gpr_thd_id thds[10]; + gpr_thd_id thds[20]; for (size_t i = 0; i < GPR_ARRAY_SIZE(thds); i++) { gpr_thd_options opt = gpr_thd_options_default(); gpr_thd_options_set_joinable(&opt); @@ -399,6 +402,44 @@ static void test_threading(void) { gpr_free(shared.pollset); } +/* Convert milliseconds into 'struct timespec' struct. millis == -1 is + * * considered as an infinity-time in future */ +static struct timespec millis_to_timespec(int millis) { + struct timespec linux_ts; + gpr_timespec gpr_ts; + + if (millis == -1) { + gpr_ts = gpr_inf_future(GPR_TIMESPAN); + } else { + gpr_ts = gpr_time_from_millis(millis, GPR_TIMESPAN); + } + + linux_ts.tv_sec = (time_t)gpr_ts.tv_sec; + linux_ts.tv_nsec = gpr_ts.tv_nsec; + return linux_ts; +} + +void test_sigwait() { + sigset_t wakeup_sig_set; + sigemptyset(&wakeup_sig_set); + sigaddset(&wakeup_sig_set, SIGRTMIN + 6); + int timeout_ms[] = {10, 1400}; + + for (size_t i = 0; i < GPR_ARRAY_SIZE(timeout_ms); i++) { + struct timespec sigwait_timeout = millis_to_timespec(timeout_ms[i]); + gpr_log(GPR_ERROR, "sigwait_timeout: %ld, %ld", sigwait_timeout.tv_sec, + sigwait_timeout.tv_nsec); + + gpr_log(GPR_ERROR, "Waiting for %d ms...", timeout_ms[i]); + gpr_timespec bef = gpr_now(GPR_CLOCK_REALTIME); + sigtimedwait(&wakeup_sig_set, NULL, &sigwait_timeout); + gpr_timespec af = gpr_now(GPR_CLOCK_REALTIME); + + gpr_log(GPR_ERROR, "Bef: %ld, %d", bef.tv_sec, bef.tv_nsec); + gpr_log(GPR_ERROR, "Aft: %ld, %d", af.tv_sec, af.tv_nsec); + } +} + int main(int argc, char **argv) { const char *poll_strategy = NULL; grpc_test_init(argc, argv); @@ -409,6 +450,7 @@ int main(int argc, char **argv) { test_add_fd_to_pollset(); test_pollset_queue_merge_items(); test_threading(); + test_sigwait(); } else { gpr_log(GPR_INFO, "Skipping the test. The test is only relevant for 'epoll' " From 92290167d63fd8b591e0028ba89a14546783b6b1 Mon Sep 17 00:00:00 2001 From: Sree Kuchibhotla Date: Fri, 21 Apr 2017 15:40:27 -0700 Subject: [PATCH 2/7] Minor changes --- include/grpc++/server_builder.h | 5 +++- src/core/lib/iomgr/ev_epoll_linux.c | 41 +++++++++++++++++------------ 2 files changed, 28 insertions(+), 18 deletions(-) diff --git a/include/grpc++/server_builder.h b/include/grpc++/server_builder.h index 68c5344e2a1..a56f81dc4b3 100644 --- a/include/grpc++/server_builder.h +++ b/include/grpc++/server_builder.h @@ -195,7 +195,10 @@ class ServerBuilder { struct SyncServerSettings { SyncServerSettings() - : num_cqs(gpr_cpu_num_cores()), min_pollers(1), max_pollers(2), cq_timeout_msec(10000) {} + : num_cqs(gpr_cpu_num_cores()), + min_pollers(1), + max_pollers(2), + cq_timeout_msec(10000) {} // Number of server completion queues to create to listen to incoming RPCs. int num_cqs; diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c index 16a9199ab92..a686d9e0ac3 100644 --- a/src/core/lib/iomgr/ev_epoll_linux.c +++ b/src/core/lib/iomgr/ev_epoll_linux.c @@ -237,6 +237,7 @@ typedef struct polling_island { /* Wakeup fd used to wake pollers to check the contents of workqueue_items */ grpc_wakeup_fd workqueue_wakeup_fd; + /* The list of workers waiting to do polling on this polling island */ gpr_mu worker_list_mu; worker_node worker_list_head; @@ -265,7 +266,10 @@ struct grpc_pollset_worker { struct grpc_pollset_worker *next; struct grpc_pollset_worker *prev; + /* Indicates if it is this worker's turn to do epoll */ gpr_atm is_polling_turn; + + /* Node in the polling island's worker list. */ worker_node pi_list_link; }; @@ -1468,12 +1472,12 @@ static bool acquire_polling_lease(grpc_pollset_worker *worker, gpr_timespec *now) { bool is_lease_acquired = false; - gpr_mu_lock(&pi->worker_list_mu); // Lock + gpr_mu_lock(&pi->worker_list_mu); // LOCK long num_pollers = gpr_atm_no_barrier_load(&pi->poller_count); if (num_pollers >= GRPC_MAX_POLLERS_PER_ISLAND) { push_back_worker_node(&pi->worker_list_head, &worker->pi_list_link); - gpr_mu_unlock(&pi->worker_list_mu); // Unlock + gpr_mu_unlock(&pi->worker_list_mu); // UNLOCK bool is_timeout = false; int ret; @@ -1487,51 +1491,54 @@ static bool acquire_polling_lease(grpc_pollset_worker *worker, if (ret == -1) { if (errno == EAGAIN) { - // gpr_log(GPR_INFO, "timeout"); // TODO: sreek remove this - // log-line + is_timeout = true; } else { + /* TODO: sreek This should not happen. If we see these log messages, it + * means we are most likely something incorrect in the setup needed for + * sigwaitinfo/sigtimedwait */ gpr_log(GPR_ERROR, "Failed with retcode: %d (timeout_ms: %d)", errno, timeout_ms); } - is_timeout = true; } + /* Did the worker come out of sigtimedwait due to a thread that just + * exited epoll and kicking it (in release_polling_lease function). */ bool is_polling_turn = gpr_atm_acq_load(&worker->is_polling_turn); - /* - if (is_polling_turn) { -/ gpr_log(GPR_ERROR, "do epoll is true (timeout_ms:%d)", - timeout_ms); // TODO: sreek remove this logline - } - */ + /* Did the worker come out of sigtimedwait due to a thread alerting it that + * some completion event was (likely) available in the completion queue */ bool is_kicked = gpr_atm_no_barrier_load(&worker->is_kicked); + if (is_kicked || is_timeout) { - *now = deadline; + *now = deadline; /* Essentially make the epoll timeout = 0 */ } else if (is_polling_turn) { - *now = gpr_now(GPR_CLOCK_MONOTONIC); + *now = gpr_now(GPR_CLOCK_MONOTONIC); /* Reduce the epoll timeout */ } - gpr_mu_lock(&pi->worker_list_mu); // Lock + gpr_mu_lock(&pi->worker_list_mu); // LOCK /* The node might have already been removed from the list by the poller that kicked this. However it is safe to call 'remove_worker_node' on an already detached node */ remove_worker_node(&worker->pi_list_link); + /* It is important to read the num_pollers again under the lock so that we + * have the latest num_pollers value that doesn't change while we are doing + * the "(num_pollers < GPRC_MAX_POLLERS_PER_ISLAND)" a a few lines below */ num_pollers = gpr_atm_no_barrier_load(&pi->poller_count); } if (num_pollers < GRPC_MAX_POLLERS_PER_ISLAND) { - gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1); // Add a poller + gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1); is_lease_acquired = true; } - gpr_mu_unlock(&pi->worker_list_mu); // Unlock + gpr_mu_unlock(&pi->worker_list_mu); // UNLOCK return is_lease_acquired; } static void release_polling_lease(polling_island *pi, grpc_error **error) { gpr_mu_lock(&pi->worker_list_mu); // Lock - gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1); // Remove poller + gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1); worker_node *node = pop_front_worker_node(&pi->worker_list_head); if (node != NULL) { grpc_pollset_worker *next_worker = WORKER_FROM_WORKER_LIST_NODE(node); From f3e94ecaa13433bb5310df716420da6624238f9d Mon Sep 17 00:00:00 2001 From: Sree Kuchibhotla Date: Fri, 21 Apr 2017 15:48:33 -0700 Subject: [PATCH 3/7] Add TODOs to ev_epoll_linux_test --- src/core/lib/iomgr/ev_epoll_linux.c | 4 ++-- test/core/iomgr/ev_epoll_linux_test.c | 4 +++- 2 files changed, 5 insertions(+), 3 deletions(-) diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c index a686d9e0ac3..3e13e657d35 100644 --- a/src/core/lib/iomgr/ev_epoll_linux.c +++ b/src/core/lib/iomgr/ev_epoll_linux.c @@ -1472,12 +1472,12 @@ static bool acquire_polling_lease(grpc_pollset_worker *worker, gpr_timespec *now) { bool is_lease_acquired = false; - gpr_mu_lock(&pi->worker_list_mu); // LOCK + gpr_mu_lock(&pi->worker_list_mu); // LOCK long num_pollers = gpr_atm_no_barrier_load(&pi->poller_count); if (num_pollers >= GRPC_MAX_POLLERS_PER_ISLAND) { push_back_worker_node(&pi->worker_list_head, &worker->pi_list_link); - gpr_mu_unlock(&pi->worker_list_mu); // UNLOCK + gpr_mu_unlock(&pi->worker_list_mu); // UNLOCK bool is_timeout = false; int ret; diff --git a/test/core/iomgr/ev_epoll_linux_test.c b/test/core/iomgr/ev_epoll_linux_test.c index f7745cddb4e..3813fb9cb07 100644 --- a/test/core/iomgr/ev_epoll_linux_test.c +++ b/test/core/iomgr/ev_epoll_linux_test.c @@ -419,11 +419,13 @@ static struct timespec millis_to_timespec(int millis) { return linux_ts; } +/* TODO (sreek) - Remove this test before merging. This is written just to + * understand the functionality of sigtimedwait and serves no other purpose */ void test_sigwait() { sigset_t wakeup_sig_set; sigemptyset(&wakeup_sig_set); sigaddset(&wakeup_sig_set, SIGRTMIN + 6); - int timeout_ms[] = {10, 1400}; + int timeout_ms[] = {10, 100}; for (size_t i = 0; i < GPR_ARRAY_SIZE(timeout_ms); i++) { struct timespec sigwait_timeout = millis_to_timespec(timeout_ms[i]); From 22b8bb1e0824eb21b3f244e543589639a85fb36d Mon Sep 17 00:00:00 2001 From: Sree Kuchibhotla Date: Wed, 26 Apr 2017 20:08:30 -0700 Subject: [PATCH 4/7] Make max-pollers configurable via GRPC_MAX_POLLERS_PER_PI env variable --- src/core/lib/iomgr/ev_epoll_linux.c | 53 +++++++++++++++++++++-------- 1 file changed, 39 insertions(+), 14 deletions(-) diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c index 3e13e657d35..3a5f48cb8f0 100644 --- a/src/core/lib/iomgr/ev_epoll_linux.c +++ b/src/core/lib/iomgr/ev_epoll_linux.c @@ -40,6 +40,7 @@ #include #include +#include #include #include #include @@ -62,6 +63,7 @@ #include "src/core/lib/iomgr/workqueue.h" #include "src/core/lib/profiling/timers.h" #include "src/core/lib/support/block_annotate.h" +#include "src/core/lib/support/env.h" /* TODO: sreek - Move this to init.c and initialize this like other tracers. */ static int grpc_polling_trace = 0; /* Disabled by default */ @@ -73,6 +75,10 @@ static int grpc_polling_trace = 0; /* Disabled by default */ /* Uncomment the following to enable extra checks on poll_object operations */ /* #define PO_DEBUG */ +/* The maximum number of polling threads per polling island. By default no + limit */ +static int g_max_pollers_per_pi = INT_MAX; + static int grpc_wakeup_signal = -1; static bool is_grpc_wakeup_signal_initialized = false; @@ -97,9 +103,6 @@ void grpc_use_signal(int signum) { } } -/* The maximum number of polling threads per polling island */ -#define GRPC_MAX_POLLERS_PER_ISLAND 1 - struct polling_island; typedef enum { @@ -1466,7 +1469,7 @@ static bool maybe_do_workqueue_work(grpc_exec_ctx *exec_ctx, return false; } -/* NOTE: May modify 'now' */ +/* NOTE: This function may modify 'now' */ static bool acquire_polling_lease(grpc_pollset_worker *worker, polling_island *pi, gpr_timespec deadline, gpr_timespec *now) { @@ -1475,7 +1478,7 @@ static bool acquire_polling_lease(grpc_pollset_worker *worker, gpr_mu_lock(&pi->worker_list_mu); // LOCK long num_pollers = gpr_atm_no_barrier_load(&pi->poller_count); - if (num_pollers >= GRPC_MAX_POLLERS_PER_ISLAND) { + if (num_pollers >= g_max_pollers_per_pi) { push_back_worker_node(&pi->worker_list_head, &worker->pi_list_link); gpr_mu_unlock(&pi->worker_list_mu); // UNLOCK @@ -1493,20 +1496,21 @@ static bool acquire_polling_lease(grpc_pollset_worker *worker, if (errno == EAGAIN) { is_timeout = true; } else { - /* TODO: sreek This should not happen. If we see these log messages, it - * means we are most likely something incorrect in the setup needed for - * sigwaitinfo/sigtimedwait */ - gpr_log(GPR_ERROR, "Failed with retcode: %d (timeout_ms: %d)", errno, + /* NOTE: This should not happen. If we see these log messages, it means + we are most likely doing something incorrect in the setup * needed + for sigwaitinfo/sigtimedwait */ + gpr_log(GPR_ERROR, + "sigtimedwait failed with retcode: %d (timeout_ms: %d)", errno, timeout_ms); } } /* Did the worker come out of sigtimedwait due to a thread that just - * exited epoll and kicking it (in release_polling_lease function). */ + exited epoll and kicking it (in release_polling_lease function). */ bool is_polling_turn = gpr_atm_acq_load(&worker->is_polling_turn); /* Did the worker come out of sigtimedwait due to a thread alerting it that - * some completion event was (likely) available in the completion queue */ + some completion event was (likely) available in the completion queue */ bool is_kicked = gpr_atm_no_barrier_load(&worker->is_kicked); if (is_kicked || is_timeout) { @@ -1522,11 +1526,11 @@ static bool acquire_polling_lease(grpc_pollset_worker *worker, remove_worker_node(&worker->pi_list_link); /* It is important to read the num_pollers again under the lock so that we * have the latest num_pollers value that doesn't change while we are doing - * the "(num_pollers < GPRC_MAX_POLLERS_PER_ISLAND)" a a few lines below */ + * the "(num_pollers < g_max_pollers_per_pi)" a a few lines below */ num_pollers = gpr_atm_no_barrier_load(&pi->poller_count); } - if (num_pollers < GRPC_MAX_POLLERS_PER_ISLAND) { + if (num_pollers < g_max_pollers_per_pi) { gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1); is_lease_acquired = true; } @@ -1536,7 +1540,7 @@ static bool acquire_polling_lease(grpc_pollset_worker *worker, } static void release_polling_lease(polling_island *pi, grpc_error **error) { - gpr_mu_lock(&pi->worker_list_mu); // Lock + gpr_mu_lock(&pi->worker_list_mu); gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1); worker_node *node = pop_front_worker_node(&pi->worker_list_head); @@ -1554,6 +1558,8 @@ static void pollset_do_epoll_pwait(grpc_exec_ctx *exec_ctx, int epoll_fd, grpc_pollset_worker *worker, gpr_timespec now, gpr_timespec deadline, sigset_t *sig_mask, grpc_error **error) { + /* Only g_max_pollers_per_pi threads can be doing polling in parallel. + If we cannot get a lease, we cannot continue to do epoll_pwait() */ if (!acquire_polling_lease(worker, pi, deadline, &now)) { return; } @@ -1563,6 +1569,7 @@ static void pollset_do_epoll_pwait(grpc_exec_ctx *exec_ctx, int epoll_fd, char *err_msg; const char *err_desc = "pollset_work_and_unlock"; + /* timeout_ms is the time between 'now' and 'deadline' */ int timeout_ms = poll_deadline_to_millis_timeout(deadline, now); GRPC_SCHEDULING_START_BLOCKING_REGION; @@ -1570,6 +1577,7 @@ static void pollset_do_epoll_pwait(grpc_exec_ctx *exec_ctx, int epoll_fd, epoll_pwait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms, sig_mask); GRPC_SCHEDULING_END_BLOCKING_REGION; + /* Give back the lease right away so that some other thread can enter */ release_polling_lease(pi, error); if (ep_rv < 0) { @@ -2116,6 +2124,21 @@ static bool is_epoll_available() { return true; } +/* This is mainly for testing purposes. Checks to see if environment variable + * GRPC_MAX_POLLERS_PER_PI is set and if so, assigns that value to the */ +static void set_max_pollers_per_island() { + char *s = gpr_getenv("GRPC_MAX_POLLERS_PER_PI"); + if (s) { + int max_pollers = (int)strtol(s, NULL, 10); + if (max_pollers > 0) { + g_max_pollers_per_pi = max_pollers; + } + } + + gpr_log(GPR_INFO, "Max number of pollers per polling island: %d", + g_max_pollers_per_pi); +} + 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) { @@ -2134,6 +2157,8 @@ const grpc_event_engine_vtable *grpc_init_epoll_linux(void) { grpc_use_signal(SIGRTMIN + 6); } + set_max_pollers_per_island(); + fd_global_init(); if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) { From 93478976b7bc72e1cf44c843fa5505b0e5caf97d Mon Sep 17 00:00:00 2001 From: Sree Kuchibhotla Date: Wed, 26 Apr 2017 20:39:03 -0700 Subject: [PATCH 5/7] Cleanup ev_epoll_linux_test and add some log statements in ev_epoll_linux.c --- src/core/lib/iomgr/ev_epoll_linux.c | 11 +++--- test/core/iomgr/ev_epoll_linux_test.c | 48 ++------------------------- 2 files changed, 9 insertions(+), 50 deletions(-) diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c index 3a5f48cb8f0..1234828063b 100644 --- a/src/core/lib/iomgr/ev_epoll_linux.c +++ b/src/core/lib/iomgr/ev_epoll_linux.c @@ -2125,14 +2125,17 @@ static bool is_epoll_available() { } /* This is mainly for testing purposes. Checks to see if environment variable - * GRPC_MAX_POLLERS_PER_PI is set and if so, assigns that value to the */ + * GRPC_MAX_POLLERS_PER_PI is set and if so, assigns that value to + * g_max_pollers_per_pi (any negative value is considered INT_MAX) */ static void set_max_pollers_per_island() { char *s = gpr_getenv("GRPC_MAX_POLLERS_PER_PI"); if (s) { - int max_pollers = (int)strtol(s, NULL, 10); - if (max_pollers > 0) { - g_max_pollers_per_pi = max_pollers; + g_max_pollers_per_pi = (int)strtol(s, NULL, 10); + if (g_max_pollers_per_pi < 0) { + g_max_pollers_per_pi = INT_MAX; } + } else { + g_max_pollers_per_pi = INT_MAX; } gpr_log(GPR_INFO, "Max number of pollers per polling island: %d", diff --git a/test/core/iomgr/ev_epoll_linux_test.c b/test/core/iomgr/ev_epoll_linux_test.c index 3813fb9cb07..0856023b14f 100644 --- a/test/core/iomgr/ev_epoll_linux_test.c +++ b/test/core/iomgr/ev_epoll_linux_test.c @@ -38,10 +38,7 @@ #include "src/core/lib/iomgr/ev_posix.h" #include -#include -#include #include -#include #include #include @@ -330,7 +327,7 @@ static __thread int thread_wakeups = 0; static void test_threading_loop(void *arg) { threading_shared *shared = arg; - while (thread_wakeups < 20000) { + while (thread_wakeups < 1000000) { grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT; grpc_pollset_worker *worker; gpr_mu_lock(shared->mu); @@ -363,7 +360,7 @@ static void test_threading(void) { shared.pollset = gpr_zalloc(grpc_pollset_size()); grpc_pollset_init(shared.pollset, &shared.mu); - gpr_thd_id thds[20]; + gpr_thd_id thds[10]; for (size_t i = 0; i < GPR_ARRAY_SIZE(thds); i++) { gpr_thd_options opt = gpr_thd_options_default(); gpr_thd_options_set_joinable(&opt); @@ -402,46 +399,6 @@ static void test_threading(void) { gpr_free(shared.pollset); } -/* Convert milliseconds into 'struct timespec' struct. millis == -1 is - * * considered as an infinity-time in future */ -static struct timespec millis_to_timespec(int millis) { - struct timespec linux_ts; - gpr_timespec gpr_ts; - - if (millis == -1) { - gpr_ts = gpr_inf_future(GPR_TIMESPAN); - } else { - gpr_ts = gpr_time_from_millis(millis, GPR_TIMESPAN); - } - - linux_ts.tv_sec = (time_t)gpr_ts.tv_sec; - linux_ts.tv_nsec = gpr_ts.tv_nsec; - return linux_ts; -} - -/* TODO (sreek) - Remove this test before merging. This is written just to - * understand the functionality of sigtimedwait and serves no other purpose */ -void test_sigwait() { - sigset_t wakeup_sig_set; - sigemptyset(&wakeup_sig_set); - sigaddset(&wakeup_sig_set, SIGRTMIN + 6); - int timeout_ms[] = {10, 100}; - - for (size_t i = 0; i < GPR_ARRAY_SIZE(timeout_ms); i++) { - struct timespec sigwait_timeout = millis_to_timespec(timeout_ms[i]); - gpr_log(GPR_ERROR, "sigwait_timeout: %ld, %ld", sigwait_timeout.tv_sec, - sigwait_timeout.tv_nsec); - - gpr_log(GPR_ERROR, "Waiting for %d ms...", timeout_ms[i]); - gpr_timespec bef = gpr_now(GPR_CLOCK_REALTIME); - sigtimedwait(&wakeup_sig_set, NULL, &sigwait_timeout); - gpr_timespec af = gpr_now(GPR_CLOCK_REALTIME); - - gpr_log(GPR_ERROR, "Bef: %ld, %d", bef.tv_sec, bef.tv_nsec); - gpr_log(GPR_ERROR, "Aft: %ld, %d", af.tv_sec, af.tv_nsec); - } -} - int main(int argc, char **argv) { const char *poll_strategy = NULL; grpc_test_init(argc, argv); @@ -452,7 +409,6 @@ int main(int argc, char **argv) { test_add_fd_to_pollset(); test_pollset_queue_merge_items(); test_threading(); - test_sigwait(); } else { gpr_log(GPR_INFO, "Skipping the test. The test is only relevant for 'epoll' " From 2f128f9a0b304ad740ddf26c7b25494d5a9cb6c6 Mon Sep 17 00:00:00 2001 From: Sree Kuchibhotla Date: Thu, 27 Apr 2017 16:40:43 -0700 Subject: [PATCH 6/7] add blocking region annotations around sigtimedwait --- src/core/lib/iomgr/ev_epoll_linux.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c index 1234828063b..64ea7b192ac 100644 --- a/src/core/lib/iomgr/ev_epoll_linux.c +++ b/src/core/lib/iomgr/ev_epoll_linux.c @@ -1489,7 +1489,9 @@ static bool acquire_polling_lease(grpc_pollset_worker *worker, ret = sigwaitinfo(&g_wakeup_sig_set, NULL); } else { struct timespec sigwait_timeout = millis_to_timespec(timeout_ms); + GRPC_SCHEDULING_START_BLOCKING_REGION; ret = sigtimedwait(&g_wakeup_sig_set, NULL, &sigwait_timeout); + GRPC_SCHEDULING_END_BLOCKING_REGION; } if (ret == -1) { From 5a1a3b49f47afeeb1b7eb2e4ea84432cce98c314 Mon Sep 17 00:00:00 2001 From: Sree Kuchibhotla Date: Fri, 5 May 2017 13:14:45 -0700 Subject: [PATCH 7/7] Separate this into a new poller --- CMakeLists.txt | 7 + Makefile | 7 + binding.gyp | 1 + build.yaml | 2 + config.m4 | 1 + gRPC-Core.podspec | 3 + grpc.gemspec | 2 + package.xml | 2 + .../iomgr/ev_epoll_limited_pollers_linux.c | 2156 +++++++++++++++++ .../iomgr/ev_epoll_limited_pollers_linux.h | 42 + src/core/lib/iomgr/ev_epoll_linux.c | 399 +-- src/core/lib/iomgr/ev_posix.c | 2 + src/python/grpcio/grpc_core_dependencies.py | 1 + tools/doxygen/Doxyfile.c++.internal | 2 + tools/doxygen/Doxyfile.core.internal | 2 + .../generated/sources_and_headers.json | 3 + vsprojects/vcxproj/grpc++/grpc++.vcxproj | 3 + .../vcxproj/grpc++/grpc++.vcxproj.filters | 6 + .../grpc++_unsecure/grpc++_unsecure.vcxproj | 3 + .../grpc++_unsecure.vcxproj.filters | 6 + vsprojects/vcxproj/grpc/grpc.vcxproj | 3 + vsprojects/vcxproj/grpc/grpc.vcxproj.filters | 6 + .../grpc_test_util/grpc_test_util.vcxproj | 3 + .../grpc_test_util.vcxproj.filters | 6 + .../grpc_unsecure/grpc_unsecure.vcxproj | 3 + .../grpc_unsecure.vcxproj.filters | 6 + 26 files changed, 2365 insertions(+), 312 deletions(-) create mode 100644 src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c create mode 100644 src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h diff --git a/CMakeLists.txt b/CMakeLists.txt index f6c5f9e56d5..68c58470ebc 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -938,6 +938,7 @@ add_library(grpc src/core/lib/iomgr/endpoint_pair_uv.c src/core/lib/iomgr/endpoint_pair_windows.c src/core/lib/iomgr/error.c + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c src/core/lib/iomgr/ev_epoll_linux.c src/core/lib/iomgr/ev_poll_posix.c src/core/lib/iomgr/ev_posix.c @@ -1266,6 +1267,7 @@ add_library(grpc_cronet src/core/lib/iomgr/endpoint_pair_uv.c src/core/lib/iomgr/endpoint_pair_windows.c src/core/lib/iomgr/error.c + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c src/core/lib/iomgr/ev_epoll_linux.c src/core/lib/iomgr/ev_poll_posix.c src/core/lib/iomgr/ev_posix.c @@ -1577,6 +1579,7 @@ add_library(grpc_test_util src/core/lib/iomgr/endpoint_pair_uv.c src/core/lib/iomgr/endpoint_pair_windows.c src/core/lib/iomgr/error.c + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c src/core/lib/iomgr/ev_epoll_linux.c src/core/lib/iomgr/ev_poll_posix.c src/core/lib/iomgr/ev_posix.c @@ -1833,6 +1836,7 @@ add_library(grpc_unsecure src/core/lib/iomgr/endpoint_pair_uv.c src/core/lib/iomgr/endpoint_pair_windows.c src/core/lib/iomgr/error.c + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c src/core/lib/iomgr/ev_epoll_linux.c src/core/lib/iomgr/ev_poll_posix.c src/core/lib/iomgr/ev_posix.c @@ -2254,6 +2258,7 @@ add_library(grpc++ src/core/lib/iomgr/endpoint_pair_uv.c src/core/lib/iomgr/endpoint_pair_windows.c src/core/lib/iomgr/error.c + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c src/core/lib/iomgr/ev_epoll_linux.c src/core/lib/iomgr/ev_poll_posix.c src/core/lib/iomgr/ev_posix.c @@ -2579,6 +2584,7 @@ add_library(grpc++_cronet src/core/lib/iomgr/endpoint_pair_uv.c src/core/lib/iomgr/endpoint_pair_windows.c src/core/lib/iomgr/error.c + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c src/core/lib/iomgr/ev_epoll_linux.c src/core/lib/iomgr/ev_poll_posix.c src/core/lib/iomgr/ev_posix.c @@ -3348,6 +3354,7 @@ add_library(grpc++_unsecure src/core/lib/iomgr/endpoint_pair_uv.c src/core/lib/iomgr/endpoint_pair_windows.c src/core/lib/iomgr/error.c + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c src/core/lib/iomgr/ev_epoll_linux.c src/core/lib/iomgr/ev_poll_posix.c src/core/lib/iomgr/ev_posix.c diff --git a/Makefile b/Makefile index 1a8ee553ca3..1fdf64d03fd 100644 --- a/Makefile +++ b/Makefile @@ -2921,6 +2921,7 @@ LIBGRPC_SRC = \ src/core/lib/iomgr/endpoint_pair_uv.c \ src/core/lib/iomgr/endpoint_pair_windows.c \ src/core/lib/iomgr/error.c \ + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c \ src/core/lib/iomgr/ev_epoll_linux.c \ src/core/lib/iomgr/ev_poll_posix.c \ src/core/lib/iomgr/ev_posix.c \ @@ -3247,6 +3248,7 @@ LIBGRPC_CRONET_SRC = \ src/core/lib/iomgr/endpoint_pair_uv.c \ src/core/lib/iomgr/endpoint_pair_windows.c \ src/core/lib/iomgr/error.c \ + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c \ src/core/lib/iomgr/ev_epoll_linux.c \ src/core/lib/iomgr/ev_poll_posix.c \ src/core/lib/iomgr/ev_posix.c \ @@ -3557,6 +3559,7 @@ LIBGRPC_TEST_UTIL_SRC = \ src/core/lib/iomgr/endpoint_pair_uv.c \ src/core/lib/iomgr/endpoint_pair_windows.c \ src/core/lib/iomgr/error.c \ + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c \ src/core/lib/iomgr/ev_epoll_linux.c \ src/core/lib/iomgr/ev_poll_posix.c \ src/core/lib/iomgr/ev_posix.c \ @@ -3785,6 +3788,7 @@ LIBGRPC_UNSECURE_SRC = \ src/core/lib/iomgr/endpoint_pair_uv.c \ src/core/lib/iomgr/endpoint_pair_windows.c \ src/core/lib/iomgr/error.c \ + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c \ src/core/lib/iomgr/ev_epoll_linux.c \ src/core/lib/iomgr/ev_poll_posix.c \ src/core/lib/iomgr/ev_posix.c \ @@ -4183,6 +4187,7 @@ LIBGRPC++_SRC = \ src/core/lib/iomgr/endpoint_pair_uv.c \ src/core/lib/iomgr/endpoint_pair_windows.c \ src/core/lib/iomgr/error.c \ + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c \ src/core/lib/iomgr/ev_epoll_linux.c \ src/core/lib/iomgr/ev_poll_posix.c \ src/core/lib/iomgr/ev_posix.c \ @@ -4516,6 +4521,7 @@ LIBGRPC++_CRONET_SRC = \ src/core/lib/iomgr/endpoint_pair_uv.c \ src/core/lib/iomgr/endpoint_pair_windows.c \ src/core/lib/iomgr/error.c \ + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c \ src/core/lib/iomgr/ev_epoll_linux.c \ src/core/lib/iomgr/ev_poll_posix.c \ src/core/lib/iomgr/ev_posix.c \ @@ -5275,6 +5281,7 @@ LIBGRPC++_UNSECURE_SRC = \ src/core/lib/iomgr/endpoint_pair_uv.c \ src/core/lib/iomgr/endpoint_pair_windows.c \ src/core/lib/iomgr/error.c \ + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c \ src/core/lib/iomgr/ev_epoll_linux.c \ src/core/lib/iomgr/ev_poll_posix.c \ src/core/lib/iomgr/ev_posix.c \ diff --git a/binding.gyp b/binding.gyp index c8dd5c1cfdd..bf7c6a15290 100644 --- a/binding.gyp +++ b/binding.gyp @@ -674,6 +674,7 @@ 'src/core/lib/iomgr/endpoint_pair_uv.c', 'src/core/lib/iomgr/endpoint_pair_windows.c', 'src/core/lib/iomgr/error.c', + 'src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c', 'src/core/lib/iomgr/ev_epoll_linux.c', 'src/core/lib/iomgr/ev_poll_posix.c', 'src/core/lib/iomgr/ev_posix.c', diff --git a/build.yaml b/build.yaml index b2930602346..2ae2f1b4081 100644 --- a/build.yaml +++ b/build.yaml @@ -197,6 +197,7 @@ filegroups: - src/core/lib/iomgr/endpoint_pair.h - src/core/lib/iomgr/error.h - src/core/lib/iomgr/error_internal.h + - src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h - src/core/lib/iomgr/ev_epoll_linux.h - src/core/lib/iomgr/ev_poll_posix.h - src/core/lib/iomgr/ev_posix.h @@ -304,6 +305,7 @@ filegroups: - src/core/lib/iomgr/endpoint_pair_uv.c - src/core/lib/iomgr/endpoint_pair_windows.c - src/core/lib/iomgr/error.c + - src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c - src/core/lib/iomgr/ev_epoll_linux.c - src/core/lib/iomgr/ev_poll_posix.c - src/core/lib/iomgr/ev_posix.c diff --git a/config.m4 b/config.m4 index 1c0c6d92fc4..f54033c0c22 100644 --- a/config.m4 +++ b/config.m4 @@ -108,6 +108,7 @@ if test "$PHP_GRPC" != "no"; then src/core/lib/iomgr/endpoint_pair_uv.c \ src/core/lib/iomgr/endpoint_pair_windows.c \ src/core/lib/iomgr/error.c \ + src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c \ src/core/lib/iomgr/ev_epoll_linux.c \ src/core/lib/iomgr/ev_poll_posix.c \ src/core/lib/iomgr/ev_posix.c \ diff --git a/gRPC-Core.podspec b/gRPC-Core.podspec index 241eba01dd2..cdc5e5837da 100644 --- a/gRPC-Core.podspec +++ b/gRPC-Core.podspec @@ -279,6 +279,7 @@ Pod::Spec.new do |s| 'src/core/lib/iomgr/endpoint_pair.h', 'src/core/lib/iomgr/error.h', 'src/core/lib/iomgr/error_internal.h', + 'src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h', 'src/core/lib/iomgr/ev_epoll_linux.h', 'src/core/lib/iomgr/ev_poll_posix.h', 'src/core/lib/iomgr/ev_posix.h', @@ -487,6 +488,7 @@ Pod::Spec.new do |s| 'src/core/lib/iomgr/endpoint_pair_uv.c', 'src/core/lib/iomgr/endpoint_pair_windows.c', 'src/core/lib/iomgr/error.c', + 'src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c', 'src/core/lib/iomgr/ev_epoll_linux.c', 'src/core/lib/iomgr/ev_poll_posix.c', 'src/core/lib/iomgr/ev_posix.c', @@ -744,6 +746,7 @@ Pod::Spec.new do |s| 'src/core/lib/iomgr/endpoint_pair.h', 'src/core/lib/iomgr/error.h', 'src/core/lib/iomgr/error_internal.h', + 'src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h', 'src/core/lib/iomgr/ev_epoll_linux.h', 'src/core/lib/iomgr/ev_poll_posix.h', 'src/core/lib/iomgr/ev_posix.h', diff --git a/grpc.gemspec b/grpc.gemspec index b96f3cbab54..22151fb29ba 100755 --- a/grpc.gemspec +++ b/grpc.gemspec @@ -195,6 +195,7 @@ Gem::Specification.new do |s| s.files += %w( src/core/lib/iomgr/endpoint_pair.h ) s.files += %w( src/core/lib/iomgr/error.h ) s.files += %w( src/core/lib/iomgr/error_internal.h ) + s.files += %w( src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h ) s.files += %w( src/core/lib/iomgr/ev_epoll_linux.h ) s.files += %w( src/core/lib/iomgr/ev_poll_posix.h ) s.files += %w( src/core/lib/iomgr/ev_posix.h ) @@ -403,6 +404,7 @@ Gem::Specification.new do |s| s.files += %w( src/core/lib/iomgr/endpoint_pair_uv.c ) s.files += %w( src/core/lib/iomgr/endpoint_pair_windows.c ) s.files += %w( src/core/lib/iomgr/error.c ) + s.files += %w( src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c ) s.files += %w( src/core/lib/iomgr/ev_epoll_linux.c ) s.files += %w( src/core/lib/iomgr/ev_poll_posix.c ) s.files += %w( src/core/lib/iomgr/ev_posix.c ) diff --git a/package.xml b/package.xml index d47708efa4e..15c573bd48a 100644 --- a/package.xml +++ b/package.xml @@ -204,6 +204,7 @@ + @@ -412,6 +413,7 @@ + diff --git a/src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c b/src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c new file mode 100644 index 00000000000..1b1c597d827 --- /dev/null +++ b/src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c @@ -0,0 +1,2156 @@ +/* + * + * Copyright 2017, Google Inc. + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are + * met: + * + * * Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * * Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following disclaimer + * in the documentation and/or other materials provided with the + * distribution. + * * Neither the name of Google Inc. nor the names of its + * contributors may be used to endorse or promote products derived from + * this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + */ + +#include "src/core/lib/iomgr/port.h" + +/* This polling engine is only relevant on linux kernels supporting epoll() */ +#ifdef GRPC_LINUX_EPOLL + +#include "src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include + +#include "src/core/lib/iomgr/ev_posix.h" +#include "src/core/lib/iomgr/iomgr_internal.h" +#include "src/core/lib/iomgr/lockfree_event.h" +#include "src/core/lib/iomgr/timer.h" +#include "src/core/lib/iomgr/wakeup_fd_posix.h" +#include "src/core/lib/iomgr/workqueue.h" +#include "src/core/lib/profiling/timers.h" +#include "src/core/lib/support/block_annotate.h" +#include "src/core/lib/support/env.h" + +/* TODO: sreek - Move this to init.c and initialize this like other tracers. */ +static int grpc_polling_trace = 0; /* Disabled by default */ +#define GRPC_POLLING_TRACE(fmt, ...) \ + if (grpc_polling_trace) { \ + gpr_log(GPR_INFO, (fmt), __VA_ARGS__); \ + } + +/* Uncomment the following to enable extra checks on poll_object operations */ +/* #define PO_DEBUG */ + +/* The maximum number of polling threads per polling island. By default no + limit */ +static int g_max_pollers_per_pi = INT_MAX; + +static int grpc_wakeup_signal = -1; +static bool is_grpc_wakeup_signal_initialized = false; + +/* 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) */ +static grpc_wakeup_fd global_wakeup_fd; + +/* 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 */ +static 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; + +typedef enum { + POLL_OBJ_FD, + POLL_OBJ_POLLSET, + POLL_OBJ_POLLSET_SET +} poll_obj_type; + +typedef struct poll_obj { +#ifdef PO_DEBUG + poll_obj_type obj_type; +#endif + gpr_mu mu; + struct polling_island *pi; +} poll_obj; + +static const char *poll_obj_string(poll_obj_type po_type) { + switch (po_type) { + case POLL_OBJ_FD: + return "fd"; + case POLL_OBJ_POLLSET: + return "pollset"; + case POLL_OBJ_POLLSET_SET: + return "pollset_set"; + } + + GPR_UNREACHABLE_CODE(return "UNKNOWN"); +} + +/******************************************************************************* + * Fd Declarations + */ + +#define FD_FROM_PO(po) ((grpc_fd *)(po)) + +struct grpc_fd { + poll_obj po; + + 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; + + /* 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; + + gpr_atm read_closure; + gpr_atm write_closure; + + struct grpc_fd *freelist_next; + grpc_closure *on_done_closure; + + /* The pollset that last noticed that the fd is readable. The actual type + * stored in this is (grpc_pollset *) */ + gpr_atm 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); + +/******************************************************************************* + * Polling island Declarations + */ + +#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG + +#define PI_ADD_REF(p, r) pi_add_ref_dbg((p), (r), __FILE__, __LINE__) +#define PI_UNREF(exec_ctx, p, r) \ + pi_unref_dbg((exec_ctx), (p), (r), __FILE__, __LINE__) + +#else /* defined(GRPC_WORKQUEUE_REFCOUNT_DEBUG) */ + +#define PI_ADD_REF(p, r) pi_add_ref((p)) +#define PI_UNREF(exec_ctx, p, r) pi_unref((exec_ctx), (p)) + +#endif /* !defined(GRPC_PI_REF_COUNT_DEBUG) */ + +typedef struct worker_node { + struct worker_node *next; + struct worker_node *prev; +} worker_node; + +/* This is also used as grpc_workqueue (by directly casing it) */ +typedef struct polling_island { + grpc_closure_scheduler workqueue_scheduler; + + 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_atm 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; + + /* Number of threads currently polling on this island */ + gpr_atm poller_count; + /* Mutex guarding the read end of the workqueue (must be held to pop from + * workqueue_items) */ + gpr_mu workqueue_read_mu; + /* Queue of closures to be executed */ + gpr_mpscq workqueue_items; + /* Count of items in workqueue_items */ + gpr_atm workqueue_item_count; + /* Wakeup fd used to wake pollers to check the contents of workqueue_items */ + grpc_wakeup_fd workqueue_wakeup_fd; + + /* The list of workers waiting to do polling on this polling island */ + gpr_mu worker_list_mu; + worker_node worker_list_head; + + /* 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_island; + +/******************************************************************************* + * Pollset Declarations + */ +#define WORKER_FROM_WORKER_LIST_NODE(p) \ + (struct grpc_pollset_worker *)(((char *)(p)) - \ + offsetof(grpc_pollset_worker, pi_list_link)) +struct grpc_pollset_worker { + /* Thread id of this worker */ + pthread_t pt_id; + + /* Used to prevent a worker from getting kicked multiple times */ + gpr_atm is_kicked; + + struct grpc_pollset_worker *next; + struct grpc_pollset_worker *prev; + + /* Indicates if it is this worker's turn to do epoll */ + gpr_atm is_polling_turn; + + /* Node in the polling island's worker list. */ + worker_node pi_list_link; +}; + +struct grpc_pollset { + poll_obj po; + + 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 */ +}; + +/******************************************************************************* + * Pollset-set Declarations + */ +struct grpc_pollset_set { + poll_obj po; +}; + +/******************************************************************************* + * Common helpers + */ + +static bool append_error(grpc_error **composite, grpc_error *error, + const char *desc) { + if (error == GRPC_ERROR_NONE) return true; + if (*composite == GRPC_ERROR_NONE) { + *composite = GRPC_ERROR_CREATE_FROM_COPIED_STRING(desc); + } + *composite = grpc_error_add_child(*composite, error); + return false; +} + +/******************************************************************************* + * 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; + +/* The polling island being polled right now. + See comments in workqueue_maybe_wakeup for why this is tracked. */ +static __thread polling_island *g_current_thread_polling_island; + +/* Forward declaration */ +static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi); +static void workqueue_enqueue(grpc_exec_ctx *exec_ctx, grpc_closure *closure, + grpc_error *error); + +#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) */ + +static const grpc_closure_scheduler_vtable workqueue_scheduler_vtable = { + workqueue_enqueue, workqueue_enqueue, "workqueue"}; + +static void pi_add_ref(polling_island *pi); +static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi); + +#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG +static void pi_add_ref_dbg(polling_island *pi, const char *reason, + const char *file, int line) { + long old_cnt = gpr_atm_acq_load(&pi->ref_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); +} + +static void pi_unref_dbg(grpc_exec_ctx *exec_ctx, polling_island *pi, + const char *reason, const char *file, int line) { + long old_cnt = gpr_atm_acq_load(&pi->ref_count); + pi_unref(exec_ctx, 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); +} + +static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue, + const char *file, int line, + const char *reason) { + if (workqueue != NULL) { + pi_add_ref_dbg((polling_island *)workqueue, reason, file, line); + } + return workqueue; +} + +static void workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, + const char *file, int line, const char *reason) { + if (workqueue != NULL) { + pi_unref_dbg(exec_ctx, (polling_island *)workqueue, reason, file, line); + } +} +#else +static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue) { + if (workqueue != NULL) { + pi_add_ref((polling_island *)workqueue); + } + return workqueue; +} + +static void workqueue_unref(grpc_exec_ctx *exec_ctx, + grpc_workqueue *workqueue) { + if (workqueue != NULL) { + pi_unref(exec_ctx, (polling_island *)workqueue); + } +} +#endif + +static void pi_add_ref(polling_island *pi) { + gpr_atm_no_barrier_fetch_add(&pi->ref_count, 1); +} + +static void pi_unref(grpc_exec_ctx *exec_ctx, 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 (1 == gpr_atm_full_fetch_add(&pi->ref_count, -1)) { + polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); + polling_island_delete(exec_ctx, pi); + if (next != NULL) { + PI_UNREF(exec_ctx, next, "pi_delete"); /* Recursive call */ + } + } +} + +static void worker_node_init(worker_node *node) { + node->next = node->prev = node; +} + +/* Not thread safe. Do under a list-level lock */ +static void push_back_worker_node(worker_node *head, worker_node *node) { + node->next = head; + node->prev = head->prev; + head->prev->next = node; + head->prev = node; +} + +/* Not thread safe. Do under a list-level lock */ +static void remove_worker_node(worker_node *node) { + node->next->prev = node->prev; + node->prev->next = node->next; + /* If node's next and prev point to itself, the node is considered detached + * from the list*/ + node->next = node->prev = node; +} + +/* Not thread safe. Do under a list-level lock */ +static worker_node *pop_front_worker_node(worker_node *head) { + worker_node *node = head->next; + if (node != head) { + remove_worker_node(node); + } else { + node = NULL; + } + + return node; +} + +/* Returns true if the node's next and prev are pointing to itself (which + indicates that the node is not in the list */ +static bool is_worker_node_detached(worker_node *node) { + return (node->next == node->prev && node->next == node); +} + +/* 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(&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_exec_ctx *exec_ctx, + grpc_fd *initial_fd, + grpc_error **error) { + polling_island *pi = NULL; + const char *err_desc = "polling_island_create"; + + *error = GRPC_ERROR_NONE; + + pi = gpr_malloc(sizeof(*pi)); + pi->workqueue_scheduler.vtable = &workqueue_scheduler_vtable; + gpr_mu_init(&pi->mu); + pi->fd_cnt = 0; + pi->fd_capacity = 0; + pi->fds = NULL; + pi->epoll_fd = -1; + + gpr_mu_init(&pi->workqueue_read_mu); + gpr_mpscq_init(&pi->workqueue_items); + gpr_atm_rel_store(&pi->workqueue_item_count, 0); + + gpr_atm_rel_store(&pi->ref_count, 0); + gpr_atm_rel_store(&pi->poller_count, 0); + gpr_atm_rel_store(&pi->merged_to, (gpr_atm)NULL); + + gpr_mu_init(&pi->worker_list_mu); + worker_node_init(&pi->worker_list_head); + + if (!append_error(error, grpc_wakeup_fd_init(&pi->workqueue_wakeup_fd), + err_desc)) { + goto done; + } + + pi->epoll_fd = epoll_create1(EPOLL_CLOEXEC); + + if (pi->epoll_fd < 0) { + append_error(error, GRPC_OS_ERROR(errno, "epoll_create1"), err_desc); + goto done; + } + + polling_island_add_wakeup_fd_locked(pi, &global_wakeup_fd, error); + polling_island_add_wakeup_fd_locked(pi, &pi->workqueue_wakeup_fd, error); + + if (initial_fd != NULL) { + polling_island_add_fds_locked(pi, &initial_fd, 1, true, error); + } + +done: + if (*error != GRPC_ERROR_NONE) { + polling_island_delete(exec_ctx, pi); + pi = NULL; + } + return pi; +} + +static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi) { + GPR_ASSERT(pi->fd_cnt == 0); + + if (pi->epoll_fd >= 0) { + close(pi->epoll_fd); + } + GPR_ASSERT(gpr_atm_no_barrier_load(&pi->workqueue_item_count) == 0); + gpr_mu_destroy(&pi->workqueue_read_mu); + gpr_mpscq_destroy(&pi->workqueue_items); + gpr_mu_destroy(&pi->mu); + grpc_wakeup_fd_destroy(&pi->workqueue_wakeup_fd); + gpr_mu_destroy(&pi->worker_list_mu); + GPR_ASSERT(is_worker_node_detached(&pi->worker_list_head)); + + gpr_free(pi->fds); + gpr_free(pi); +} + +/* 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 void workqueue_maybe_wakeup(polling_island *pi) { + /* If this thread is the current poller, then it may be that it's about to + decrement the current poller count, so we need to look past this thread */ + bool is_current_poller = (g_current_thread_polling_island == pi); + gpr_atm min_current_pollers_for_wakeup = is_current_poller ? 1 : 0; + gpr_atm current_pollers = gpr_atm_no_barrier_load(&pi->poller_count); + /* Only issue a wakeup if it's likely that some poller could come in and take + it right now. Note that since we do an anticipatory mpscq_pop every poll + loop, it's ok if we miss the wakeup here, as we'll get the work item when + the next poller enters anyway. */ + if (current_pollers > min_current_pollers_for_wakeup) { + GRPC_LOG_IF_ERROR("workqueue_wakeup_fd", + grpc_wakeup_fd_wakeup(&pi->workqueue_wakeup_fd)); + } +} + +static void workqueue_move_items_to_parent(polling_island *q) { + polling_island *p = (polling_island *)gpr_atm_no_barrier_load(&q->merged_to); + if (p == NULL) { + return; + } + gpr_mu_lock(&q->workqueue_read_mu); + int num_added = 0; + while (gpr_atm_no_barrier_load(&q->workqueue_item_count) > 0) { + gpr_mpscq_node *n = gpr_mpscq_pop(&q->workqueue_items); + if (n != NULL) { + gpr_atm_no_barrier_fetch_add(&q->workqueue_item_count, -1); + gpr_atm_no_barrier_fetch_add(&p->workqueue_item_count, 1); + gpr_mpscq_push(&p->workqueue_items, n); + num_added++; + } + } + gpr_mu_unlock(&q->workqueue_read_mu); + if (num_added > 0) { + workqueue_maybe_wakeup(p); + } + workqueue_move_items_to_parent(p); +} + +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 */ + + workqueue_move_items_to_parent(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 void workqueue_enqueue(grpc_exec_ctx *exec_ctx, grpc_closure *closure, + grpc_error *error) { + GPR_TIMER_BEGIN("workqueue.enqueue", 0); + grpc_workqueue *workqueue = (grpc_workqueue *)closure->scheduler; + /* take a ref to the workqueue: otherwise it can happen that whatever events + * this kicks off ends up destroying the workqueue before this function + * completes */ + GRPC_WORKQUEUE_REF(workqueue, "enqueue"); + polling_island *pi = (polling_island *)workqueue; + gpr_atm last = gpr_atm_no_barrier_fetch_add(&pi->workqueue_item_count, 1); + closure->error_data.error = error; + gpr_mpscq_push(&pi->workqueue_items, &closure->next_data.atm_next); + if (last == 0) { + workqueue_maybe_wakeup(pi); + } + workqueue_move_items_to_parent(pi); + GRPC_WORKQUEUE_UNREF(exec_ctx, workqueue, "enqueue"); + GPR_TIMER_END("workqueue.enqueue", 0); +} + +static grpc_closure_scheduler *workqueue_scheduler(grpc_workqueue *workqueue) { + polling_island *pi = (polling_island *)workqueue; + return workqueue == NULL ? grpc_schedule_on_exec_ctx + : &pi->workqueue_scheduler; +} + +static grpc_error *polling_island_global_init() { + grpc_error *error = GRPC_ERROR_NONE; + + 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() { + 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. */ + +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); + + grpc_lfev_destroy(&fd->read_closure); + grpc_lfev_destroy(&fd->write_closure); + + 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->po.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->po.mu); + } + + /* Note: It is not really needed to get the new_fd->po.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->po.mu); + new_fd->po.pi = NULL; +#ifdef PO_DEBUG + new_fd->po.obj_type = POLL_OBJ_FD; +#endif + + gpr_atm_rel_store(&new_fd->refst, (gpr_atm)1); + new_fd->fd = fd; + new_fd->orphaned = false; + grpc_lfev_init(&new_fd->read_closure); + grpc_lfev_init(&new_fd->write_closure); + gpr_atm_no_barrier_store(&new_fd->read_notifier_pollset, (gpr_atm)NULL); + + new_fd->freelist_next = NULL; + new_fd->on_done_closure = NULL; + + gpr_mu_unlock(&new_fd->po.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 int fd_wrapped_fd(grpc_fd *fd) { + int ret_fd = -1; + gpr_mu_lock(&fd->po.mu); + if (!fd->orphaned) { + ret_fd = fd->fd; + } + gpr_mu_unlock(&fd->po.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; + polling_island *unref_pi = NULL; + + gpr_mu_lock(&fd->po.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->po.pi). 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->po.pi to NULL (but remove the ref on the polling island + before doing this.) */ + if (fd->po.pi != NULL) { + polling_island *pi_latest = polling_island_lock(fd->po.pi); + polling_island_remove_fd_locked(pi_latest, fd, is_fd_closed, &error); + gpr_mu_unlock(&pi_latest->mu); + + unref_pi = fd->po.pi; + fd->po.pi = NULL; + } + + grpc_closure_sched(exec_ctx, fd->on_done_closure, GRPC_ERROR_REF(error)); + + gpr_mu_unlock(&fd->po.mu); + UNREF_BY(fd, 2, reason); /* Drop the reference */ + if (unref_pi != NULL) { + /* Unref stale polling island here, outside the fd lock above. + The polling island owns a workqueue which owns an fd, and unreffing + inside the lock can cause an eventual lock loop that makes TSAN very + unhappy. */ + PI_UNREF(exec_ctx, unref_pi, "fd_orphan"); + } + GRPC_LOG_IF_ERROR("fd_orphan", GRPC_ERROR_REF(error)); + GRPC_ERROR_UNREF(error); +} + +static grpc_pollset *fd_get_read_notifier_pollset(grpc_exec_ctx *exec_ctx, + grpc_fd *fd) { + gpr_atm notifier = gpr_atm_acq_load(&fd->read_notifier_pollset); + return (grpc_pollset *)notifier; +} + +static bool fd_is_shutdown(grpc_fd *fd) { + return grpc_lfev_is_shutdown(&fd->read_closure); +} + +/* Might be called multiple times */ +static void fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_error *why) { + if (grpc_lfev_set_shutdown(exec_ctx, &fd->read_closure, + GRPC_ERROR_REF(why))) { + shutdown(fd->fd, SHUT_RDWR); + grpc_lfev_set_shutdown(exec_ctx, &fd->write_closure, GRPC_ERROR_REF(why)); + } + GRPC_ERROR_UNREF(why); +} + +static void fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure) { + grpc_lfev_notify_on(exec_ctx, &fd->read_closure, closure); +} + +static void fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure) { + grpc_lfev_notify_on(exec_ctx, &fd->write_closure, closure); +} + +static grpc_workqueue *fd_get_workqueue(grpc_fd *fd) { + gpr_mu_lock(&fd->po.mu); + grpc_workqueue *workqueue = + GRPC_WORKQUEUE_REF((grpc_workqueue *)fd->po.pi, "fd_get_workqueue"); + gpr_mu_unlock(&fd->po.mu); + return workqueue; +} + +/******************************************************************************* + * 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 __thread sigset_t g_wakeup_sig_set; + +static void sig_handler(int sig_num) { +#ifdef GRPC_EPOLL_DEBUG + gpr_log(GPR_INFO, "Received signal %d", sig_num); +#endif +} + +static void pollset_worker_init(grpc_pollset_worker *worker) { + worker->pt_id = pthread_self(); + worker->next = worker->prev = NULL; + gpr_atm_no_barrier_store(&worker->is_kicked, (gpr_atm)0); + gpr_atm_no_barrier_store(&worker->is_polling_turn, (gpr_atm)0); + worker_node_init(&worker->pi_list_link); +} + +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(&global_wakeup_fd); +} + +static void pollset_global_shutdown(void) { + grpc_wakeup_fd_destroy(&global_wakeup_fd); + gpr_tls_destroy(&g_current_thread_pollset); + gpr_tls_destroy(&g_current_thread_worker); +} + +static grpc_error *worker_kick(grpc_pollset_worker *worker, + gpr_atm *is_kicked) { + grpc_error *err = GRPC_ERROR_NONE; + + /* Kick the worker only if it was not already kicked */ + if (gpr_atm_no_barrier_cas(is_kicked, (gpr_atm)0, (gpr_atm)1)) { + GRPC_POLLING_TRACE( + "pollset_worker_kick: Kicking worker: %p (thread id: %ld)", + (void *)worker, (long int)worker->pt_id); + 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; +} + +static grpc_error *pollset_worker_kick(grpc_pollset_worker *worker) { + return worker_kick(worker, &worker->is_kicked); +} + +static grpc_error *poller_kick(grpc_pollset_worker *worker) { + return worker_kick(worker, &worker->is_polling_turn); +} + +/* 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(&global_wakeup_fd); +} + +static void pollset_init(grpc_pollset *pollset, gpr_mu **mu) { + gpr_mu_init(&pollset->po.mu); + *mu = &pollset->po.mu; + pollset->po.pi = NULL; +#ifdef PO_DEBUG + pollset->po.obj_type = POLL_OBJ_POLLSET; +#endif + + 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; +} + +/* Convert millis to timespec (clock-type is assumed to be GPR_TIMESPAN) */ +static struct timespec millis_to_timespec(int millis) { + struct timespec linux_ts; + gpr_timespec gpr_ts; + + if (millis == -1) { + gpr_ts = gpr_inf_future(GPR_TIMESPAN); + } else { + gpr_ts = gpr_time_from_millis(millis, GPR_TIMESPAN); + } + + linux_ts.tv_sec = (time_t)gpr_ts.tv_sec; + linux_ts.tv_nsec = gpr_ts.tv_nsec; + return linux_ts; +} + +/* 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); + int millis = gpr_time_to_millis(gpr_time_add( + timeout, gpr_time_from_nanos(GPR_NS_PER_MS - 1, GPR_TIMESPAN))); + return millis >= 1 ? millis : 1; +} + +static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_pollset *notifier) { + grpc_lfev_set_ready(exec_ctx, &fd->read_closure); + + /* Note, it is possible that fd_become_readable might be called twice with + different 'notifier's when an fd becomes readable and it is in two epoll + sets (This can happen briefly during polling island merges). In such cases + it does not really matter which notifer is set as the read_notifier_pollset + (They would both point to the same polling island anyway) */ + /* Use release store to match with acquire load in fd_get_read_notifier */ + gpr_atm_rel_store(&fd->read_notifier_pollset, (gpr_atm)notifier); +} + +static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + grpc_lfev_set_ready(exec_ctx, &fd->write_closure); +} + +static void pollset_release_polling_island(grpc_exec_ctx *exec_ctx, + grpc_pollset *ps, char *reason) { + if (ps->po.pi != NULL) { + PI_UNREF(exec_ctx, ps->po.pi, reason); + } + ps->po.pi = 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->po.pi to NULL */ + pollset_release_polling_island(exec_ctx, pollset, "ps_shutdown"); + grpc_closure_sched(exec_ctx, pollset->shutdown_done, GRPC_ERROR_NONE); +} + +/* pollset->po.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->po.mu); +} + +static bool maybe_do_workqueue_work(grpc_exec_ctx *exec_ctx, + polling_island *pi) { + if (gpr_mu_trylock(&pi->workqueue_read_mu)) { + gpr_mpscq_node *n = gpr_mpscq_pop(&pi->workqueue_items); + gpr_mu_unlock(&pi->workqueue_read_mu); + if (n != NULL) { + if (gpr_atm_full_fetch_add(&pi->workqueue_item_count, -1) > 1) { + workqueue_maybe_wakeup(pi); + } + grpc_closure *c = (grpc_closure *)n; + grpc_error *error = c->error_data.error; +#ifndef NDEBUG + c->scheduled = false; +#endif + c->cb(exec_ctx, c->cb_arg, error); + GRPC_ERROR_UNREF(error); + return true; + } else if (gpr_atm_no_barrier_load(&pi->workqueue_item_count) > 0) { + /* n == NULL might mean there's work but it's not available to be popped + * yet - try to ensure another workqueue wakes up to check shortly if so + */ + workqueue_maybe_wakeup(pi); + } + } + return false; +} + +/* NOTE: This function may modify 'now' */ +static bool acquire_polling_lease(grpc_pollset_worker *worker, + polling_island *pi, gpr_timespec deadline, + gpr_timespec *now) { + bool is_lease_acquired = false; + + gpr_mu_lock(&pi->worker_list_mu); // LOCK + long num_pollers = gpr_atm_no_barrier_load(&pi->poller_count); + + if (num_pollers >= g_max_pollers_per_pi) { + push_back_worker_node(&pi->worker_list_head, &worker->pi_list_link); + gpr_mu_unlock(&pi->worker_list_mu); // UNLOCK + + bool is_timeout = false; + int ret; + int timeout_ms = poll_deadline_to_millis_timeout(deadline, *now); + if (timeout_ms == -1) { + ret = sigwaitinfo(&g_wakeup_sig_set, NULL); + } else { + struct timespec sigwait_timeout = millis_to_timespec(timeout_ms); + GRPC_SCHEDULING_START_BLOCKING_REGION; + ret = sigtimedwait(&g_wakeup_sig_set, NULL, &sigwait_timeout); + GRPC_SCHEDULING_END_BLOCKING_REGION; + } + + if (ret == -1) { + if (errno == EAGAIN) { + is_timeout = true; + } else { + /* NOTE: This should not happen. If we see these log messages, it means + we are most likely doing something incorrect in the setup * needed + for sigwaitinfo/sigtimedwait */ + gpr_log(GPR_ERROR, + "sigtimedwait failed with retcode: %d (timeout_ms: %d)", errno, + timeout_ms); + } + } + + /* Did the worker come out of sigtimedwait due to a thread that just + exited epoll and kicking it (in release_polling_lease function). */ + bool is_polling_turn = gpr_atm_acq_load(&worker->is_polling_turn); + + /* Did the worker come out of sigtimedwait due to a thread alerting it that + some completion event was (likely) available in the completion queue */ + bool is_kicked = gpr_atm_no_barrier_load(&worker->is_kicked); + + if (is_kicked || is_timeout) { + *now = deadline; /* Essentially make the epoll timeout = 0 */ + } else if (is_polling_turn) { + *now = gpr_now(GPR_CLOCK_MONOTONIC); /* Reduce the epoll timeout */ + } + + gpr_mu_lock(&pi->worker_list_mu); // LOCK + /* The node might have already been removed from the list by the poller + that kicked this. However it is safe to call 'remove_worker_node' on + an already detached node */ + remove_worker_node(&worker->pi_list_link); + /* It is important to read the num_pollers again under the lock so that we + * have the latest num_pollers value that doesn't change while we are doing + * the "(num_pollers < g_max_pollers_per_pi)" a a few lines below */ + num_pollers = gpr_atm_no_barrier_load(&pi->poller_count); + } + + if (num_pollers < g_max_pollers_per_pi) { + gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1); + is_lease_acquired = true; + } + + gpr_mu_unlock(&pi->worker_list_mu); // UNLOCK + return is_lease_acquired; +} + +static void release_polling_lease(polling_island *pi, grpc_error **error) { + gpr_mu_lock(&pi->worker_list_mu); + + gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1); + worker_node *node = pop_front_worker_node(&pi->worker_list_head); + if (node != NULL) { + grpc_pollset_worker *next_worker = WORKER_FROM_WORKER_LIST_NODE(node); + append_error(error, poller_kick(next_worker), "poller kick error"); + } + + gpr_mu_unlock(&pi->worker_list_mu); +} + +#define GRPC_EPOLL_MAX_EVENTS 100 +static void pollset_do_epoll_pwait(grpc_exec_ctx *exec_ctx, int epoll_fd, + grpc_pollset *pollset, polling_island *pi, + grpc_pollset_worker *worker, + gpr_timespec now, gpr_timespec deadline, + sigset_t *sig_mask, grpc_error **error) { + /* Only g_max_pollers_per_pi threads can be doing polling in parallel. + If we cannot get a lease, we cannot continue to do epoll_pwait() */ + if (!acquire_polling_lease(worker, pi, deadline, &now)) { + return; + } + + struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS]; + int ep_rv; + char *err_msg; + const char *err_desc = "pollset_work_and_unlock"; + + /* timeout_ms is the time between 'now' and 'deadline' */ + int timeout_ms = poll_deadline_to_millis_timeout(deadline, now); + + GRPC_SCHEDULING_START_BLOCKING_REGION; + ep_rv = + epoll_pwait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms, sig_mask); + GRPC_SCHEDULING_END_BLOCKING_REGION; + + /* Give back the lease right away so that some other thread can enter */ + release_polling_lease(pi, error); + + 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 */ + GRPC_POLLING_TRACE("pollset_work: pollset: %p, worker: %p received kick", + (void *)pollset, (void *)worker); + 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 == &global_wakeup_fd) { + grpc_timer_consume_kick(); + append_error(error, grpc_wakeup_fd_consume_wakeup(&global_wakeup_fd), + err_desc); + } else if (data_ptr == &pi->workqueue_wakeup_fd) { + append_error(error, + grpc_wakeup_fd_consume_wakeup(&pi->workqueue_wakeup_fd), + err_desc); + maybe_do_workqueue_work(exec_ctx, pi); + } else if (data_ptr == &polling_island_wakeup_fd) { + GRPC_POLLING_TRACE( + "pollset_work: pollset: %p, worker: %p polling island (epoll_fd: " + "%d) got merged", + (void *)pollset, (void *)worker, epoll_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); + } + } + } +} + +/* 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, + grpc_pollset_worker *worker, + gpr_timespec now, gpr_timespec deadline, + sigset_t *sig_mask, grpc_error **error) { + int epoll_fd = -1; + polling_island *pi = NULL; + 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->po.pi + + Since epoll_fd is immutable, it is safe to read it without a lock on the + polling island. There is however a possibility that the polling island from + which we got the epoll_fd, got merged with another island in the meantime. + This is okay because in such a case, we will wakeup right-away from + epoll_pwait() (because any merge will poison the old polling island's epoll + set 'polling_island_wakeup_fd') and then pick up the latest polling_island + the next time this function - pollset_work_and_unlock()) is called */ + + if (pollset->po.pi == NULL) { + pollset->po.pi = polling_island_create(exec_ctx, NULL, error); + if (pollset->po.pi == NULL) { + GPR_TIMER_END("pollset_work_and_unlock", 0); + return; /* Fatal error. Cannot continue */ + } + + PI_ADD_REF(pollset->po.pi, "ps"); + GRPC_POLLING_TRACE("pollset_work: pollset: %p created new pi: %p", + (void *)pollset, (void *)pollset->po.pi); + } + + pi = polling_island_maybe_get_latest(pollset->po.pi); + epoll_fd = pi->epoll_fd; + + /* Update the pollset->po.pi since the island being pointed by + pollset->po.pi maybe older than the one pointed by pi) */ + if (pollset->po.pi != 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(exec_ctx, pollset->po.pi, "ps"); + pollset->po.pi = 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->po.mu); + + /* If we get some workqueue work to do, it might end up completing an item on + the completion queue, so there's no need to poll... so we skip that and + redo the complete loop to verify */ + if (!maybe_do_workqueue_work(exec_ctx, pi)) { + g_current_thread_polling_island = pi; + pollset_do_epoll_pwait(exec_ctx, epoll_fd, pollset, pi, worker, now, + deadline, sig_mask, error); + g_current_thread_polling_island = NULL; + } + + 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->po.pi + that we got before releasing the polling island lock). This is because + pollset->po.pi 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(exec_ctx, pi, "ps_work"); + + GPR_TIMER_END("pollset_work_and_unlock", 0); +} + +/* pollset->po.mu lock must be held by the caller before calling this. + The function pollset_work() may temporarily release the lock (pollset->po.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; + + grpc_pollset_worker worker; + pollset_worker_init(&worker); + + if (worker_hdl) *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 the + 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 worker can be kicked anytime from the point it is added to the pollset + via push_front_worker() (or push_back_worker()) to the point it is + removed via remove_worker(). + If the worker is kicked before/during it calls epoll_pwait(), it should + immediately exit from epoll_wait(). If the worker is kicked after it + returns from epoll_wait(), then nothing really needs to be done. + + To accomplish this, we mask 'grpc_wakeup_signal' on this thread 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(&g_wakeup_sig_set); + sigaddset(&g_wakeup_sig_set, grpc_wakeup_signal); + pthread_sigmask(SIG_BLOCK, &g_wakeup_sig_set, &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, &worker, now, deadline, + &g_orig_sigmask, &error); + grpc_exec_ctx_flush(exec_ctx); + + gpr_mu_lock(&pollset->po.mu); + + /* Note: There is no need to reset worker.is_kicked to 0 since we are no + longer going to use this worker */ + 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->po.mu); + grpc_exec_ctx_flush(exec_ctx); + gpr_mu_lock(&pollset->po.mu); + } + + if (worker_hdl) *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 add_poll_object(grpc_exec_ctx *exec_ctx, poll_obj *bag, + poll_obj_type bag_type, poll_obj *item, + poll_obj_type item_type) { + GPR_TIMER_BEGIN("add_poll_object", 0); + +#ifdef PO_DEBUG + GPR_ASSERT(item->obj_type == item_type); + GPR_ASSERT(bag->obj_type == bag_type); +#endif + + grpc_error *error = GRPC_ERROR_NONE; + polling_island *pi_new = NULL; + + gpr_mu_lock(&bag->mu); + gpr_mu_lock(&item->mu); + +retry: + /* + * 1) If item->pi and bag->pi are both non-NULL and equal, do nothing + * 2) If item->pi and bag->pi are both NULL, create a new polling island (with + * a refcount of 2) and point item->pi and bag->pi to the new island + * 3) If exactly one of item->pi or bag->pi is NULL, update it to point to + * the other's non-NULL pi + * 4) Finally if item->pi and bag-pi are non-NULL and not-equal, merge the + * polling islands and update item->pi and bag->pi to point to the new + * island + */ + + /* Early out if we are trying to add an 'fd' to a 'bag' but the fd is already + * orphaned */ + if (item_type == POLL_OBJ_FD && (FD_FROM_PO(item))->orphaned) { + gpr_mu_unlock(&item->mu); + gpr_mu_unlock(&bag->mu); + return; + } + + if (item->pi == bag->pi) { + pi_new = item->pi; + if (pi_new == NULL) { + /* GPR_ASSERT(item->pi == bag->pi == NULL) */ + + /* If we are adding an fd to a bag (i.e pollset or pollset_set), then + * we need to do some extra work to make TSAN happy */ + if (item_type == POLL_OBJ_FD) { + /* Unlock before creating a new polling island: the polling island will + create a workqueue which creates a file descriptor, and holding an fd + lock here can eventually cause a loop to appear to TSAN (making it + unhappy). We don't think it's a real loop (there's an epoch point + where that loop possibility disappears), but the advantages of + keeping TSAN happy outweigh any performance advantage we might have + by keeping the lock held. */ + gpr_mu_unlock(&item->mu); + pi_new = polling_island_create(exec_ctx, FD_FROM_PO(item), &error); + gpr_mu_lock(&item->mu); + + /* Need to reverify any assumptions made between the initial lock and + getting to this branch: if they've changed, we need to throw away our + work and figure things out again. */ + if (item->pi != NULL) { + GRPC_POLLING_TRACE( + "add_poll_object: Raced creating new polling island. pi_new: %p " + "(fd: %d, %s: %p)", + (void *)pi_new, FD_FROM_PO(item)->fd, poll_obj_string(bag_type), + (void *)bag); + /* No need to lock 'pi_new' here since this is a new polling island + and no one has a reference to it yet */ + polling_island_remove_all_fds_locked(pi_new, true, &error); + + /* Ref and unref so that the polling island gets deleted during unref + */ + PI_ADD_REF(pi_new, "dance_of_destruction"); + PI_UNREF(exec_ctx, pi_new, "dance_of_destruction"); + goto retry; + } + } else { + pi_new = polling_island_create(exec_ctx, NULL, &error); + } + + GRPC_POLLING_TRACE( + "add_poll_object: Created new polling island. pi_new: %p (%s: %p, " + "%s: %p)", + (void *)pi_new, poll_obj_string(item_type), (void *)item, + poll_obj_string(bag_type), (void *)bag); + } else { + GRPC_POLLING_TRACE( + "add_poll_object: Same polling island. pi: %p (%s, %s)", + (void *)pi_new, poll_obj_string(item_type), + poll_obj_string(bag_type)); + } + } else if (item->pi == NULL) { + /* GPR_ASSERT(bag->pi != NULL) */ + /* Make pi_new point to latest pi*/ + pi_new = polling_island_lock(bag->pi); + + if (item_type == POLL_OBJ_FD) { + grpc_fd *fd = FD_FROM_PO(item); + polling_island_add_fds_locked(pi_new, &fd, 1, true, &error); + } + + gpr_mu_unlock(&pi_new->mu); + GRPC_POLLING_TRACE( + "add_poll_obj: item->pi was NULL. pi_new: %p (item(%s): %p, " + "bag(%s): %p)", + (void *)pi_new, poll_obj_string(item_type), (void *)item, + poll_obj_string(bag_type), (void *)bag); + } else if (bag->pi == NULL) { + /* GPR_ASSERT(item->pi != NULL) */ + /* Make pi_new to point to latest pi */ + pi_new = polling_island_lock(item->pi); + gpr_mu_unlock(&pi_new->mu); + GRPC_POLLING_TRACE( + "add_poll_obj: bag->pi was NULL. pi_new: %p (item(%s): %p, " + "bag(%s): %p)", + (void *)pi_new, poll_obj_string(item_type), (void *)item, + poll_obj_string(bag_type), (void *)bag); + } else { + pi_new = polling_island_merge(item->pi, bag->pi, &error); + GRPC_POLLING_TRACE( + "add_poll_obj: polling islands merged. pi_new: %p (item(%s): %p, " + "bag(%s): %p)", + (void *)pi_new, poll_obj_string(item_type), (void *)item, + poll_obj_string(bag_type), (void *)bag); + } + + /* At this point, pi_new is the polling island that both item->pi and bag->pi + MUST be pointing to */ + + if (item->pi != pi_new) { + PI_ADD_REF(pi_new, poll_obj_string(item_type)); + if (item->pi != NULL) { + PI_UNREF(exec_ctx, item->pi, poll_obj_string(item_type)); + } + item->pi = pi_new; + } + + if (bag->pi != pi_new) { + PI_ADD_REF(pi_new, poll_obj_string(bag_type)); + if (bag->pi != NULL) { + PI_UNREF(exec_ctx, bag->pi, poll_obj_string(bag_type)); + } + bag->pi = pi_new; + } + + gpr_mu_unlock(&item->mu); + gpr_mu_unlock(&bag->mu); + + GRPC_LOG_IF_ERROR("add_poll_object", error); + GPR_TIMER_END("add_poll_object", 0); +} + +static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_fd *fd) { + add_poll_object(exec_ctx, &pollset->po, POLL_OBJ_POLLSET, &fd->po, + POLL_OBJ_FD); +} + +/******************************************************************************* + * Pollset-set Definitions + */ + +static grpc_pollset_set *pollset_set_create(void) { + grpc_pollset_set *pss = gpr_malloc(sizeof(*pss)); + gpr_mu_init(&pss->po.mu); + pss->po.pi = NULL; +#ifdef PO_DEBUG + pss->po.obj_type = POLL_OBJ_POLLSET_SET; +#endif + return pss; +} + +static void pollset_set_destroy(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pss) { + gpr_mu_destroy(&pss->po.mu); + + if (pss->po.pi != NULL) { + PI_UNREF(exec_ctx, pss->po.pi, "pss_destroy"); + } + + gpr_free(pss); +} + +static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset_set *pss, + grpc_fd *fd) { + add_poll_object(exec_ctx, &pss->po, POLL_OBJ_POLLSET_SET, &fd->po, + POLL_OBJ_FD); +} + +static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx, grpc_pollset_set *pss, + grpc_fd *fd) { + /* Nothing to do */ +} + +static void pollset_set_add_pollset(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pss, grpc_pollset *ps) { + add_poll_object(exec_ctx, &pss->po, POLL_OBJ_POLLSET_SET, &ps->po, + POLL_OBJ_POLLSET); +} + +static void pollset_set_del_pollset(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pss, grpc_pollset *ps) { + /* Nothing to do */ +} + +static void pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item) { + add_poll_object(exec_ctx, &bag->po, POLL_OBJ_POLLSET_SET, &item->po, + POLL_OBJ_POLLSET_SET); +} + +static void pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item) { + /* Nothing to do */ +} + +/******************************************************************************* + * 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, + .fd_get_workqueue = fd_get_workqueue, + + .pollset_init = pollset_init, + .pollset_shutdown = pollset_shutdown, + .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, + + .workqueue_ref = workqueue_ref, + .workqueue_unref = workqueue_unref, + .workqueue_scheduler = workqueue_scheduler, + + .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; +} + +/* This is mainly for testing purposes. Checks to see if environment variable + * GRPC_MAX_POLLERS_PER_PI is set and if so, assigns that value to + * g_max_pollers_per_pi (any negative value is considered INT_MAX) */ +static void set_max_pollers_per_island() { + char *s = gpr_getenv("GRPC_MAX_POLLERS_PER_PI"); + if (s) { + g_max_pollers_per_pi = (int)strtol(s, NULL, 10); + if (g_max_pollers_per_pi < 0) { + g_max_pollers_per_pi = INT_MAX; + } + } else { + g_max_pollers_per_pi = INT_MAX; + } + + gpr_log(GPR_INFO, "Max number of pollers per polling island: %d", + g_max_pollers_per_pi); +} + +const grpc_event_engine_vtable *grpc_init_epoll_limited_pollers_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 (!grpc_has_wakeup_fd()) { + return NULL; + } + + if (!is_epoll_available()) { + return NULL; + } + + if (!is_grpc_wakeup_signal_initialized) { + grpc_use_signal(SIGRTMIN + 6); + } + + set_max_pollers_per_island(); + + 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(GRPC_LINUX_EPOLL) */ +#if defined(GRPC_POSIX_SOCKET) +#include "src/core/lib/iomgr/ev_posix.h" +/* If GRPC_LINUX_EPOLL is not defined, it means epoll is not available. Return + * NULL */ +const grpc_event_engine_vtable *grpc_init_epoll_limited_pollers_linux(void) { + return NULL; +} +#endif /* defined(GRPC_POSIX_SOCKET) */ +#endif /* !defined(GRPC_LINUX_EPOLL) */ diff --git a/src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h b/src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h new file mode 100644 index 00000000000..23ff8570c1b --- /dev/null +++ b/src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h @@ -0,0 +1,42 @@ +/* + * + * 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_LIMITED_POLLERS_LINUX_H +#define GRPC_CORE_LIB_IOMGR_EV_EPOLL_LIMITED_POLLERS_LINUX_H + +#include "src/core/lib/iomgr/ev_posix.h" +#include "src/core/lib/iomgr/port.h" + +const grpc_event_engine_vtable *grpc_init_epoll_limited_pollers_linux(void); + +#endif /* GRPC_CORE_LIB_IOMGR_EV_EPOLL_LIMITED_POLLERS_LINUX_H */ diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c index 64ea7b192ac..e603a755937 100644 --- a/src/core/lib/iomgr/ev_epoll_linux.c +++ b/src/core/lib/iomgr/ev_epoll_linux.c @@ -40,7 +40,6 @@ #include #include -#include #include #include #include @@ -63,7 +62,6 @@ #include "src/core/lib/iomgr/workqueue.h" #include "src/core/lib/profiling/timers.h" #include "src/core/lib/support/block_annotate.h" -#include "src/core/lib/support/env.h" /* TODO: sreek - Move this to init.c and initialize this like other tracers. */ static int grpc_polling_trace = 0; /* Disabled by default */ @@ -75,10 +73,6 @@ static int grpc_polling_trace = 0; /* Disabled by default */ /* Uncomment the following to enable extra checks on poll_object operations */ /* #define PO_DEBUG */ -/* The maximum number of polling threads per polling island. By default no - limit */ -static int g_max_pollers_per_pi = INT_MAX; - static int grpc_wakeup_signal = -1; static bool is_grpc_wakeup_signal_initialized = false; @@ -201,11 +195,6 @@ static void fd_global_shutdown(void); #endif /* !defined(GRPC_PI_REF_COUNT_DEBUG) */ -typedef struct worker_node { - struct worker_node *next; - struct worker_node *prev; -} worker_node; - /* This is also used as grpc_workqueue (by directly casing it) */ typedef struct polling_island { grpc_closure_scheduler workqueue_scheduler; @@ -240,10 +229,6 @@ typedef struct polling_island { /* Wakeup fd used to wake pollers to check the contents of workqueue_items */ grpc_wakeup_fd workqueue_wakeup_fd; - /* The list of workers waiting to do polling on this polling island */ - gpr_mu worker_list_mu; - worker_node worker_list_head; - /* The fd of the underlying epoll set */ int epoll_fd; @@ -256,24 +241,14 @@ typedef struct polling_island { /******************************************************************************* * Pollset Declarations */ -#define WORKER_FROM_WORKER_LIST_NODE(p) \ - (struct grpc_pollset_worker *)(((char *)(p)) - \ - offsetof(grpc_pollset_worker, pi_list_link)) struct grpc_pollset_worker { /* Thread id of this worker */ pthread_t pt_id; /* Used to prevent a worker from getting kicked multiple times */ gpr_atm is_kicked; - struct grpc_pollset_worker *next; struct grpc_pollset_worker *prev; - - /* Indicates if it is this worker's turn to do epoll */ - gpr_atm is_polling_turn; - - /* Node in the polling island's worker list. */ - worker_node pi_list_link; }; struct grpc_pollset { @@ -417,47 +392,7 @@ static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi) { } } -static void worker_node_init(worker_node *node) { - node->next = node->prev = node; -} - -/* Not thread safe. Do under a list-level lock */ -static void push_back_worker_node(worker_node *head, worker_node *node) { - node->next = head; - node->prev = head->prev; - head->prev->next = node; - head->prev = node; -} - -/* Not thread safe. Do under a list-level lock */ -static void remove_worker_node(worker_node *node) { - node->next->prev = node->prev; - node->prev->next = node->next; - /* If node's next and prev point to itself, the node is considered detached - * from the list*/ - node->next = node->prev = node; -} - -/* Not thread safe. Do under a list-level lock */ -static worker_node *pop_front_worker_node(worker_node *head) { - worker_node *node = head->next; - if (node != head) { - remove_worker_node(node); - } else { - node = NULL; - } - - return node; -} - -/* Returns true if the node's next and prev are pointing to itself (which - indicates that the node is not in the list */ -static bool is_worker_node_detached(worker_node *node) { - return (node->next == node->prev && node->next == node); -} - -/* The caller is expected to hold pi->mu lock before calling this function - */ +/* 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) { @@ -611,9 +546,6 @@ static polling_island *polling_island_create(grpc_exec_ctx *exec_ctx, gpr_atm_rel_store(&pi->poller_count, 0); gpr_atm_rel_store(&pi->merged_to, (gpr_atm)NULL); - gpr_mu_init(&pi->worker_list_mu); - worker_node_init(&pi->worker_list_head); - if (!append_error(error, grpc_wakeup_fd_init(&pi->workqueue_wakeup_fd), err_desc)) { goto done; @@ -652,9 +584,6 @@ static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi) { gpr_mpscq_destroy(&pi->workqueue_items); gpr_mu_destroy(&pi->mu); grpc_wakeup_fd_destroy(&pi->workqueue_wakeup_fd); - gpr_mu_destroy(&pi->worker_list_mu); - GPR_ASSERT(is_worker_node_detached(&pi->worker_list_head)); - gpr_free(pi->fds); gpr_free(pi); } @@ -1173,7 +1102,6 @@ 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 __thread sigset_t g_wakeup_sig_set; static void sig_handler(int sig_num) { #ifdef GRPC_EPOLL_DEBUG @@ -1181,14 +1109,6 @@ static void sig_handler(int sig_num) { #endif } -static void pollset_worker_init(grpc_pollset_worker *worker) { - worker->pt_id = pthread_self(); - worker->next = worker->prev = NULL; - gpr_atm_no_barrier_store(&worker->is_kicked, (gpr_atm)0); - gpr_atm_no_barrier_store(&worker->is_polling_turn, (gpr_atm)0); - worker_node_init(&worker->pi_list_link); -} - static void poller_kick_init() { signal(grpc_wakeup_signal, sig_handler); } /* Global state management */ @@ -1205,12 +1125,11 @@ static void pollset_global_shutdown(void) { gpr_tls_destroy(&g_current_thread_worker); } -static grpc_error *worker_kick(grpc_pollset_worker *worker, - gpr_atm *is_kicked) { +static grpc_error *pollset_worker_kick(grpc_pollset_worker *worker) { grpc_error *err = GRPC_ERROR_NONE; /* Kick the worker only if it was not already kicked */ - if (gpr_atm_no_barrier_cas(is_kicked, (gpr_atm)0, (gpr_atm)1)) { + if (gpr_atm_no_barrier_cas(&worker->is_kicked, (gpr_atm)0, (gpr_atm)1)) { GRPC_POLLING_TRACE( "pollset_worker_kick: Kicking worker: %p (thread id: %ld)", (void *)worker, (long int)worker->pt_id); @@ -1222,14 +1141,6 @@ static grpc_error *worker_kick(grpc_pollset_worker *worker, return err; } -static grpc_error *pollset_worker_kick(grpc_pollset_worker *worker) { - return worker_kick(worker, &worker->is_kicked); -} - -static grpc_error *poller_kick(grpc_pollset_worker *worker) { - return worker_kick(worker, &worker->is_polling_turn); -} - /* Return 1 if the pollset has active threads in pollset_work (pollset must * be locked) */ static int pollset_has_workers(grpc_pollset *p) { @@ -1335,22 +1246,6 @@ static void pollset_init(grpc_pollset *pollset, gpr_mu **mu) { pollset->shutdown_done = NULL; } -/* Convert millis to timespec (clock-type is assumed to be GPR_TIMESPAN) */ -static struct timespec millis_to_timespec(int millis) { - struct timespec linux_ts; - gpr_timespec gpr_ts; - - if (millis == -1) { - gpr_ts = gpr_inf_future(GPR_TIMESPAN); - } else { - gpr_ts = gpr_time_from_millis(millis, GPR_TIMESPAN); - } - - linux_ts.tv_sec = (time_t)gpr_ts.tv_sec; - linux_ts.tv_nsec = gpr_ts.tv_nsec; - return linux_ts; -} - /* 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) @@ -1469,200 +1364,35 @@ static bool maybe_do_workqueue_work(grpc_exec_ctx *exec_ctx, return false; } -/* NOTE: This function may modify 'now' */ -static bool acquire_polling_lease(grpc_pollset_worker *worker, - polling_island *pi, gpr_timespec deadline, - gpr_timespec *now) { - bool is_lease_acquired = false; - - gpr_mu_lock(&pi->worker_list_mu); // LOCK - long num_pollers = gpr_atm_no_barrier_load(&pi->poller_count); - - if (num_pollers >= g_max_pollers_per_pi) { - push_back_worker_node(&pi->worker_list_head, &worker->pi_list_link); - gpr_mu_unlock(&pi->worker_list_mu); // UNLOCK - - bool is_timeout = false; - int ret; - int timeout_ms = poll_deadline_to_millis_timeout(deadline, *now); - if (timeout_ms == -1) { - ret = sigwaitinfo(&g_wakeup_sig_set, NULL); - } else { - struct timespec sigwait_timeout = millis_to_timespec(timeout_ms); - GRPC_SCHEDULING_START_BLOCKING_REGION; - ret = sigtimedwait(&g_wakeup_sig_set, NULL, &sigwait_timeout); - GRPC_SCHEDULING_END_BLOCKING_REGION; - } - - if (ret == -1) { - if (errno == EAGAIN) { - is_timeout = true; - } else { - /* NOTE: This should not happen. If we see these log messages, it means - we are most likely doing something incorrect in the setup * needed - for sigwaitinfo/sigtimedwait */ - gpr_log(GPR_ERROR, - "sigtimedwait failed with retcode: %d (timeout_ms: %d)", errno, - timeout_ms); - } - } - - /* Did the worker come out of sigtimedwait due to a thread that just - exited epoll and kicking it (in release_polling_lease function). */ - bool is_polling_turn = gpr_atm_acq_load(&worker->is_polling_turn); - - /* Did the worker come out of sigtimedwait due to a thread alerting it that - some completion event was (likely) available in the completion queue */ - bool is_kicked = gpr_atm_no_barrier_load(&worker->is_kicked); - - if (is_kicked || is_timeout) { - *now = deadline; /* Essentially make the epoll timeout = 0 */ - } else if (is_polling_turn) { - *now = gpr_now(GPR_CLOCK_MONOTONIC); /* Reduce the epoll timeout */ - } - - gpr_mu_lock(&pi->worker_list_mu); // LOCK - /* The node might have already been removed from the list by the poller - that kicked this. However it is safe to call 'remove_worker_node' on - an already detached node */ - remove_worker_node(&worker->pi_list_link); - /* It is important to read the num_pollers again under the lock so that we - * have the latest num_pollers value that doesn't change while we are doing - * the "(num_pollers < g_max_pollers_per_pi)" a a few lines below */ - num_pollers = gpr_atm_no_barrier_load(&pi->poller_count); - } - - if (num_pollers < g_max_pollers_per_pi) { - gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1); - is_lease_acquired = true; - } - - gpr_mu_unlock(&pi->worker_list_mu); // UNLOCK - return is_lease_acquired; -} - -static void release_polling_lease(polling_island *pi, grpc_error **error) { - gpr_mu_lock(&pi->worker_list_mu); - - gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1); - worker_node *node = pop_front_worker_node(&pi->worker_list_head); - if (node != NULL) { - grpc_pollset_worker *next_worker = WORKER_FROM_WORKER_LIST_NODE(node); - append_error(error, poller_kick(next_worker), "poller kick error"); - } - - gpr_mu_unlock(&pi->worker_list_mu); -} - #define GRPC_EPOLL_MAX_EVENTS 100 -static void pollset_do_epoll_pwait(grpc_exec_ctx *exec_ctx, int epoll_fd, - grpc_pollset *pollset, polling_island *pi, - grpc_pollset_worker *worker, - gpr_timespec now, gpr_timespec deadline, - sigset_t *sig_mask, grpc_error **error) { - /* Only g_max_pollers_per_pi threads can be doing polling in parallel. - If we cannot get a lease, we cannot continue to do epoll_pwait() */ - if (!acquire_polling_lease(worker, pi, deadline, &now)) { - return; - } - - struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS]; - int ep_rv; - char *err_msg; - const char *err_desc = "pollset_work_and_unlock"; - - /* timeout_ms is the time between 'now' and 'deadline' */ - int timeout_ms = poll_deadline_to_millis_timeout(deadline, now); - - GRPC_SCHEDULING_START_BLOCKING_REGION; - ep_rv = - epoll_pwait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms, sig_mask); - GRPC_SCHEDULING_END_BLOCKING_REGION; - - /* Give back the lease right away so that some other thread can enter */ - release_polling_lease(pi, error); - - 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 */ - GRPC_POLLING_TRACE("pollset_work: pollset: %p, worker: %p received kick", - (void *)pollset, (void *)worker); - 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 == &global_wakeup_fd) { - grpc_timer_consume_kick(); - append_error(error, grpc_wakeup_fd_consume_wakeup(&global_wakeup_fd), - err_desc); - } else if (data_ptr == &pi->workqueue_wakeup_fd) { - append_error(error, - grpc_wakeup_fd_consume_wakeup(&pi->workqueue_wakeup_fd), - err_desc); - maybe_do_workqueue_work(exec_ctx, pi); - } else if (data_ptr == &polling_island_wakeup_fd) { - GRPC_POLLING_TRACE( - "pollset_work: pollset: %p, worker: %p polling island (epoll_fd: " - "%d) got merged", - (void *)pollset, (void *)worker, epoll_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); - } - } - } -} - /* 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, - grpc_pollset_worker *worker, - gpr_timespec now, gpr_timespec deadline, + grpc_pollset_worker *worker, 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->po.pi - Since epoll_fd is immutable, it is safe to read it without a lock on the - polling island. There is however a possibility that the polling island from - which we got the epoll_fd, got merged with another island in the meantime. - This is okay because in such a case, we will wakeup right-away from - epoll_pwait() (because any merge will poison the old polling island's epoll - set 'polling_island_wakeup_fd') and then pick up the latest polling_island - the next time this function - pollset_work_and_unlock()) is called */ + 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->po.pi == NULL) { pollset->po.pi = polling_island_create(exec_ctx, NULL, error); if (pollset->po.pi == NULL) { GPR_TIMER_END("pollset_work_and_unlock", 0); - return; /* Fatal error. Cannot continue */ + return; /* Fatal error. We cannot continue */ } PI_ADD_REF(pollset->po.pi, "ps"); @@ -1693,10 +1423,70 @@ static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx, the completion queue, so there's no need to poll... so we skip that and redo the complete loop to verify */ if (!maybe_do_workqueue_work(exec_ctx, pi)) { + gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1); g_current_thread_polling_island = pi; - pollset_do_epoll_pwait(exec_ctx, epoll_fd, pollset, pi, worker, now, - deadline, sig_mask, error); + + GRPC_SCHEDULING_START_BLOCKING_REGION; + ep_rv = epoll_pwait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms, + sig_mask); + GRPC_SCHEDULING_END_BLOCKING_REGION; + 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 */ + GRPC_POLLING_TRACE( + "pollset_work: pollset: %p, worker: %p received kick", + (void *)pollset, (void *)worker); + 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 == &global_wakeup_fd) { + grpc_timer_consume_kick(); + append_error(error, grpc_wakeup_fd_consume_wakeup(&global_wakeup_fd), + err_desc); + } else if (data_ptr == &pi->workqueue_wakeup_fd) { + append_error(error, + grpc_wakeup_fd_consume_wakeup(&pi->workqueue_wakeup_fd), + err_desc); + maybe_do_workqueue_work(exec_ctx, pi); + } else if (data_ptr == &polling_island_wakeup_fd) { + GRPC_POLLING_TRACE( + "pollset_work: pollset: %p, worker: %p polling island (epoll_fd: " + "%d) got merged", + (void *)pollset, (void *)worker, epoll_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); + } + } + } + g_current_thread_polling_island = NULL; + gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1); } GPR_ASSERT(pi != NULL); @@ -1720,9 +1510,14 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, 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; - pollset_worker_init(&worker); + worker.next = worker.prev = NULL; + worker.pt_id = pthread_self(); + gpr_atm_no_barrier_store(&worker.is_kicked, (gpr_atm)0); if (worker_hdl) *worker_hdl = &worker; @@ -1756,9 +1551,9 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, misses acting on a kick */ if (!g_initialized_sigmask) { - sigemptyset(&g_wakeup_sig_set); - sigaddset(&g_wakeup_sig_set, grpc_wakeup_signal); - pthread_sigmask(SIG_BLOCK, &g_wakeup_sig_set, &g_orig_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'. @@ -1773,7 +1568,7 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, push_front_worker(pollset, &worker); /* Add worker to pollset */ - pollset_work_and_unlock(exec_ctx, pollset, &worker, now, deadline, + pollset_work_and_unlock(exec_ctx, pollset, &worker, timeout_ms, &g_orig_sigmask, &error); grpc_exec_ctx_flush(exec_ctx); @@ -2126,24 +1921,6 @@ static bool is_epoll_available() { return true; } -/* This is mainly for testing purposes. Checks to see if environment variable - * GRPC_MAX_POLLERS_PER_PI is set and if so, assigns that value to - * g_max_pollers_per_pi (any negative value is considered INT_MAX) */ -static void set_max_pollers_per_island() { - char *s = gpr_getenv("GRPC_MAX_POLLERS_PER_PI"); - if (s) { - g_max_pollers_per_pi = (int)strtol(s, NULL, 10); - if (g_max_pollers_per_pi < 0) { - g_max_pollers_per_pi = INT_MAX; - } - } else { - g_max_pollers_per_pi = INT_MAX; - } - - gpr_log(GPR_INFO, "Max number of pollers per polling island: %d", - g_max_pollers_per_pi); -} - 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) { @@ -2162,8 +1939,6 @@ const grpc_event_engine_vtable *grpc_init_epoll_linux(void) { grpc_use_signal(SIGRTMIN + 6); } - set_max_pollers_per_island(); - fd_global_init(); if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) { diff --git a/src/core/lib/iomgr/ev_posix.c b/src/core/lib/iomgr/ev_posix.c index 13409a4de83..07a9ed570a1 100644 --- a/src/core/lib/iomgr/ev_posix.c +++ b/src/core/lib/iomgr/ev_posix.c @@ -44,6 +44,7 @@ #include #include +#include "src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h" #include "src/core/lib/iomgr/ev_epoll_linux.h" #include "src/core/lib/iomgr/ev_poll_posix.h" #include "src/core/lib/support/env.h" @@ -68,6 +69,7 @@ static const event_engine_factory g_factories[] = { {"epoll", grpc_init_epoll_linux}, {"poll", grpc_init_poll_posix}, {"poll-cv", grpc_init_poll_cv_posix}, + {"epoll-limited", grpc_init_epoll_limited_pollers_linux}, }; static void add(const char *beg, const char *end, char ***ss, size_t *ns) { diff --git a/src/python/grpcio/grpc_core_dependencies.py b/src/python/grpcio/grpc_core_dependencies.py index 42c13b2cb63..c3053612e48 100644 --- a/src/python/grpcio/grpc_core_dependencies.py +++ b/src/python/grpcio/grpc_core_dependencies.py @@ -97,6 +97,7 @@ CORE_SOURCE_FILES = [ 'src/core/lib/iomgr/endpoint_pair_uv.c', 'src/core/lib/iomgr/endpoint_pair_windows.c', 'src/core/lib/iomgr/error.c', + 'src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c', 'src/core/lib/iomgr/ev_epoll_linux.c', 'src/core/lib/iomgr/ev_poll_posix.c', 'src/core/lib/iomgr/ev_posix.c', diff --git a/tools/doxygen/Doxyfile.c++.internal b/tools/doxygen/Doxyfile.c++.internal index 9664234f9f4..5ce89e7a8e1 100644 --- a/tools/doxygen/Doxyfile.c++.internal +++ b/tools/doxygen/Doxyfile.c++.internal @@ -939,6 +939,8 @@ src/core/lib/iomgr/endpoint_pair_windows.c \ src/core/lib/iomgr/error.c \ src/core/lib/iomgr/error.h \ src/core/lib/iomgr/error_internal.h \ +src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c \ +src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h \ src/core/lib/iomgr/ev_epoll_linux.c \ src/core/lib/iomgr/ev_epoll_linux.h \ src/core/lib/iomgr/ev_poll_posix.c \ diff --git a/tools/doxygen/Doxyfile.core.internal b/tools/doxygen/Doxyfile.core.internal index 51d77c2b521..54a7ed4e971 100644 --- a/tools/doxygen/Doxyfile.core.internal +++ b/tools/doxygen/Doxyfile.core.internal @@ -1078,6 +1078,8 @@ src/core/lib/iomgr/endpoint_pair_windows.c \ src/core/lib/iomgr/error.c \ src/core/lib/iomgr/error.h \ src/core/lib/iomgr/error_internal.h \ +src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c \ +src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h \ src/core/lib/iomgr/ev_epoll_linux.c \ src/core/lib/iomgr/ev_epoll_linux.h \ src/core/lib/iomgr/ev_poll_posix.c \ diff --git a/tools/run_tests/generated/sources_and_headers.json b/tools/run_tests/generated/sources_and_headers.json index c9628234379..f2c467e4ea1 100644 --- a/tools/run_tests/generated/sources_and_headers.json +++ b/tools/run_tests/generated/sources_and_headers.json @@ -7750,6 +7750,7 @@ "src/core/lib/iomgr/endpoint_pair.h", "src/core/lib/iomgr/error.h", "src/core/lib/iomgr/error_internal.h", + "src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h", "src/core/lib/iomgr/ev_epoll_linux.h", "src/core/lib/iomgr/ev_poll_posix.h", "src/core/lib/iomgr/ev_posix.h", @@ -7891,6 +7892,8 @@ "src/core/lib/iomgr/error.c", "src/core/lib/iomgr/error.h", "src/core/lib/iomgr/error_internal.h", + "src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c", + "src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h", "src/core/lib/iomgr/ev_epoll_linux.c", "src/core/lib/iomgr/ev_epoll_linux.h", "src/core/lib/iomgr/ev_poll_posix.c", diff --git a/vsprojects/vcxproj/grpc++/grpc++.vcxproj b/vsprojects/vcxproj/grpc++/grpc++.vcxproj index 32d2e09a588..4d44d450a21 100644 --- a/vsprojects/vcxproj/grpc++/grpc++.vcxproj +++ b/vsprojects/vcxproj/grpc++/grpc++.vcxproj @@ -396,6 +396,7 @@ + @@ -610,6 +611,8 @@ + + diff --git a/vsprojects/vcxproj/grpc++/grpc++.vcxproj.filters b/vsprojects/vcxproj/grpc++/grpc++.vcxproj.filters index a3346bc2979..b2afed1e504 100644 --- a/vsprojects/vcxproj/grpc++/grpc++.vcxproj.filters +++ b/vsprojects/vcxproj/grpc++/grpc++.vcxproj.filters @@ -184,6 +184,9 @@ src\core\lib\iomgr + + src\core\lib\iomgr + src\core\lib\iomgr @@ -908,6 +911,9 @@ src\core\lib\iomgr + + src\core\lib\iomgr + src\core\lib\iomgr diff --git a/vsprojects/vcxproj/grpc++_unsecure/grpc++_unsecure.vcxproj b/vsprojects/vcxproj/grpc++_unsecure/grpc++_unsecure.vcxproj index 28ccefc651c..fd207a2c794 100644 --- a/vsprojects/vcxproj/grpc++_unsecure/grpc++_unsecure.vcxproj +++ b/vsprojects/vcxproj/grpc++_unsecure/grpc++_unsecure.vcxproj @@ -390,6 +390,7 @@ + @@ -594,6 +595,8 @@ + + diff --git a/vsprojects/vcxproj/grpc++_unsecure/grpc++_unsecure.vcxproj.filters b/vsprojects/vcxproj/grpc++_unsecure/grpc++_unsecure.vcxproj.filters index 83f869dab36..531ae520474 100644 --- a/vsprojects/vcxproj/grpc++_unsecure/grpc++_unsecure.vcxproj.filters +++ b/vsprojects/vcxproj/grpc++_unsecure/grpc++_unsecure.vcxproj.filters @@ -169,6 +169,9 @@ src\core\lib\iomgr + + src\core\lib\iomgr + src\core\lib\iomgr @@ -875,6 +878,9 @@ src\core\lib\iomgr + + src\core\lib\iomgr + src\core\lib\iomgr diff --git a/vsprojects/vcxproj/grpc/grpc.vcxproj b/vsprojects/vcxproj/grpc/grpc.vcxproj index 86b5856d68a..9e0fb5de8d4 100644 --- a/vsprojects/vcxproj/grpc/grpc.vcxproj +++ b/vsprojects/vcxproj/grpc/grpc.vcxproj @@ -320,6 +320,7 @@ + @@ -551,6 +552,8 @@ + + diff --git a/vsprojects/vcxproj/grpc/grpc.vcxproj.filters b/vsprojects/vcxproj/grpc/grpc.vcxproj.filters index 943ad521f75..4abf85b10f0 100644 --- a/vsprojects/vcxproj/grpc/grpc.vcxproj.filters +++ b/vsprojects/vcxproj/grpc/grpc.vcxproj.filters @@ -64,6 +64,9 @@ src\core\lib\iomgr + + src\core\lib\iomgr + src\core\lib\iomgr @@ -875,6 +878,9 @@ src\core\lib\iomgr + + src\core\lib\iomgr + src\core\lib\iomgr diff --git a/vsprojects/vcxproj/grpc_test_util/grpc_test_util.vcxproj b/vsprojects/vcxproj/grpc_test_util/grpc_test_util.vcxproj index df89932a97c..60deba9243f 100644 --- a/vsprojects/vcxproj/grpc_test_util/grpc_test_util.vcxproj +++ b/vsprojects/vcxproj/grpc_test_util/grpc_test_util.vcxproj @@ -215,6 +215,7 @@ + @@ -383,6 +384,8 @@ + + diff --git a/vsprojects/vcxproj/grpc_test_util/grpc_test_util.vcxproj.filters b/vsprojects/vcxproj/grpc_test_util/grpc_test_util.vcxproj.filters index 22cfbe14d4d..44a29fadb77 100644 --- a/vsprojects/vcxproj/grpc_test_util/grpc_test_util.vcxproj.filters +++ b/vsprojects/vcxproj/grpc_test_util/grpc_test_util.vcxproj.filters @@ -121,6 +121,9 @@ src\core\lib\iomgr + + src\core\lib\iomgr + src\core\lib\iomgr @@ -623,6 +626,9 @@ src\core\lib\iomgr + + src\core\lib\iomgr + src\core\lib\iomgr diff --git a/vsprojects/vcxproj/grpc_unsecure/grpc_unsecure.vcxproj b/vsprojects/vcxproj/grpc_unsecure/grpc_unsecure.vcxproj index 2db0ffa6920..6fdf6bb6903 100644 --- a/vsprojects/vcxproj/grpc_unsecure/grpc_unsecure.vcxproj +++ b/vsprojects/vcxproj/grpc_unsecure/grpc_unsecure.vcxproj @@ -310,6 +310,7 @@ + @@ -518,6 +519,8 @@ + + diff --git a/vsprojects/vcxproj/grpc_unsecure/grpc_unsecure.vcxproj.filters b/vsprojects/vcxproj/grpc_unsecure/grpc_unsecure.vcxproj.filters index c7d6670db1b..707ab9e098e 100644 --- a/vsprojects/vcxproj/grpc_unsecure/grpc_unsecure.vcxproj.filters +++ b/vsprojects/vcxproj/grpc_unsecure/grpc_unsecure.vcxproj.filters @@ -67,6 +67,9 @@ src\core\lib\iomgr + + src\core\lib\iomgr + src\core\lib\iomgr @@ -785,6 +788,9 @@ src\core\lib\iomgr + + src\core\lib\iomgr + src\core\lib\iomgr