diff options
Diffstat (limited to 'src/core/lib/iomgr')
30 files changed, 799 insertions, 427 deletions
diff --git a/src/core/lib/iomgr/endpoint_pair.h b/src/core/lib/iomgr/endpoint_pair.h index f9de0c715e..6407a6ad3f 100644 --- a/src/core/lib/iomgr/endpoint_pair.h +++ b/src/core/lib/iomgr/endpoint_pair.h @@ -41,8 +41,7 @@ typedef struct { grpc_endpoint *server; } grpc_endpoint_pair; -grpc_endpoint_pair grpc_iomgr_create_endpoint_pair( - const char *name, grpc_resource_quota *resource_quota, - size_t read_slice_size); +grpc_endpoint_pair grpc_iomgr_create_endpoint_pair(const char *name, + grpc_channel_args *args); #endif /* GRPC_CORE_LIB_IOMGR_ENDPOINT_PAIR_H */ diff --git a/src/core/lib/iomgr/endpoint_pair_posix.c b/src/core/lib/iomgr/endpoint_pair_posix.c index b9ff969e81..5542a372d8 100644 --- a/src/core/lib/iomgr/endpoint_pair_posix.c +++ b/src/core/lib/iomgr/endpoint_pair_posix.c @@ -62,22 +62,25 @@ static void create_sockets(int sv[2]) { GPR_ASSERT(grpc_set_socket_no_sigpipe_if_possible(sv[1]) == GRPC_ERROR_NONE); } -grpc_endpoint_pair grpc_iomgr_create_endpoint_pair( - const char *name, grpc_resource_quota *resource_quota, - size_t read_slice_size) { +grpc_endpoint_pair grpc_iomgr_create_endpoint_pair(const char *name, + grpc_channel_args *args) { int sv[2]; grpc_endpoint_pair p; char *final_name; create_sockets(sv); + grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT; + gpr_asprintf(&final_name, "%s:client", name); - p.client = grpc_tcp_create(grpc_fd_create(sv[1], final_name), resource_quota, - read_slice_size, "socketpair-server"); + p.client = grpc_tcp_create(&exec_ctx, grpc_fd_create(sv[1], final_name), args, + "socketpair-server"); gpr_free(final_name); gpr_asprintf(&final_name, "%s:server", name); - p.server = grpc_tcp_create(grpc_fd_create(sv[0], final_name), resource_quota, - read_slice_size, "socketpair-client"); + p.server = grpc_tcp_create(&exec_ctx, grpc_fd_create(sv[0], final_name), args, + "socketpair-client"); gpr_free(final_name); + + grpc_exec_ctx_finish(&exec_ctx); return p; } diff --git a/src/core/lib/iomgr/endpoint_pair_uv.c b/src/core/lib/iomgr/endpoint_pair_uv.c index ff24894c6d..9718eb0523 100644 --- a/src/core/lib/iomgr/endpoint_pair_uv.c +++ b/src/core/lib/iomgr/endpoint_pair_uv.c @@ -41,9 +41,8 @@ #include "src/core/lib/iomgr/endpoint_pair.h" -grpc_endpoint_pair grpc_iomgr_create_endpoint_pair( - const char *name, grpc_resource_quota *resource_quota, - size_t read_slice_size) { +grpc_endpoint_pair grpc_iomgr_create_endpoint_pair(const char *name, + grpc_channel_args *args) { grpc_endpoint_pair endpoint_pair; // TODO(mlumish): implement this properly under libuv GPR_ASSERT(false && diff --git a/src/core/lib/iomgr/endpoint_pair_windows.c b/src/core/lib/iomgr/endpoint_pair_windows.c index 93f71b745c..25d6264dfb 100644 --- a/src/core/lib/iomgr/endpoint_pair_windows.c +++ b/src/core/lib/iomgr/endpoint_pair_windows.c @@ -83,15 +83,18 @@ static void create_sockets(SOCKET sv[2]) { } grpc_endpoint_pair grpc_iomgr_create_endpoint_pair( - const char *name, grpc_resource_quota *resource_quota, - size_t read_slice_size) { + const char *name, grpc_channel_args *channel_args) { SOCKET sv[2]; grpc_endpoint_pair p; create_sockets(sv); - p.client = grpc_tcp_create(grpc_winsocket_create(sv[1], "endpoint:client"), - resource_quota, "endpoint:server"); - p.server = grpc_tcp_create(grpc_winsocket_create(sv[0], "endpoint:server"), - resource_quota, "endpoint:client"); + grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT; + p.client = grpc_tcp_create(&exec_ctx, + grpc_winsocket_create(sv[1], "endpoint:client"), + channel_args, "endpoint:server"); + p.server = grpc_tcp_create(&exec_ctx, + grpc_winsocket_create(sv[0], "endpoint:server"), + channel_args, "endpoint:client"); + grpc_exec_ctx_finish(&exec_ctx); return p; } diff --git a/src/core/lib/iomgr/error.c b/src/core/lib/iomgr/error.c index 1dbb64e8f3..fbbca6b493 100644 --- a/src/core/lib/iomgr/error.c +++ b/src/core/lib/iomgr/error.c @@ -212,7 +212,11 @@ static uint8_t get_placement(grpc_error **err, size_t size) { GPR_ASSERT(*err); uint8_t slots = (uint8_t)(size / sizeof(intptr_t)); if ((*err)->arena_size + slots > (*err)->arena_capacity) { - (*err)->arena_capacity = (uint8_t)(3 * (*err)->arena_capacity / 2); + (*err)->arena_capacity = + (uint8_t)GPR_MIN(UINT8_MAX - 1, (3 * (*err)->arena_capacity / 2)); + if ((*err)->arena_size + slots > (*err)->arena_capacity) { + return UINT8_MAX; + } *err = gpr_realloc( *err, sizeof(grpc_error) + (*err)->arena_capacity * sizeof(intptr_t)); } @@ -223,10 +227,14 @@ static uint8_t get_placement(grpc_error **err, size_t size) { static void internal_set_int(grpc_error **err, grpc_error_ints which, intptr_t value) { - // GPR_ASSERT((*err)->ints[which] == UINT8_MAX); // TODO, enforce this uint8_t slot = (*err)->ints[which]; if (slot == UINT8_MAX) { slot = get_placement(err, sizeof(value)); + if (slot == UINT8_MAX) { + gpr_log(GPR_ERROR, "Error %p is full, dropping int {\"%s\":%" PRIiPTR "}", + *err, error_int_name(which), value); + return; + } } (*err)->ints[which] = slot; (*err)->arena[slot] = value; @@ -234,10 +242,16 @@ static void internal_set_int(grpc_error **err, grpc_error_ints which, static void internal_set_str(grpc_error **err, grpc_error_strs which, grpc_slice value) { - // GPR_ASSERT((*err)->strs[which] == UINT8_MAX); // TODO, enforce this uint8_t slot = (*err)->strs[which]; if (slot == UINT8_MAX) { slot = get_placement(err, sizeof(value)); + if (slot == UINT8_MAX) { + const char *str = grpc_slice_to_c_string(value); + gpr_log(GPR_ERROR, "Error %p is full, dropping string {\"%s\":\"%s\"}", + *err, error_str_name(which), str); + gpr_free((void *)str); + return; + } } else { unref_slice(*(grpc_slice *)((*err)->arena + slot)); } @@ -245,12 +259,19 @@ static void internal_set_str(grpc_error **err, grpc_error_strs which, memcpy((*err)->arena + slot, &value, sizeof(value)); } +static char *fmt_time(gpr_timespec tm); static void internal_set_time(grpc_error **err, grpc_error_times which, gpr_timespec value) { - // GPR_ASSERT((*err)->times[which] == UINT8_MAX); // TODO, enforce this uint8_t slot = (*err)->times[which]; if (slot == UINT8_MAX) { slot = get_placement(err, sizeof(value)); + if (slot == UINT8_MAX) { + const char *time_str = fmt_time(value); + gpr_log(GPR_ERROR, "Error %p is full, dropping \"%s\":\"%s\"}", *err, + error_time_name(which), time_str); + gpr_free((void *)time_str); + return; + } } (*err)->times[which] = slot; memcpy((*err)->arena + slot, &value, sizeof(value)); @@ -259,6 +280,12 @@ static void internal_set_time(grpc_error **err, grpc_error_times which, static void internal_add_error(grpc_error **err, grpc_error *new) { grpc_linked_error new_last = {new, UINT8_MAX}; uint8_t slot = get_placement(err, sizeof(grpc_linked_error)); + if (slot == UINT8_MAX) { + gpr_log(GPR_ERROR, "Error %p is full, dropping error %p = %s", *err, new, + grpc_error_string(new)); + GRPC_ERROR_UNREF(new); + return; + } if ((*err)->first_err == UINT8_MAX) { GPR_ASSERT((*err)->last_err == UINT8_MAX); (*err)->last_err = slot; diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c index f6372c0f3f..4a0f91391f 100644 --- a/src/core/lib/iomgr/ev_epoll_linux.c +++ b/src/core/lib/iomgr/ev_epoll_linux.c @@ -56,6 +56,8 @@ #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" @@ -140,52 +142,11 @@ struct grpc_fd { Ref/Unref by two to avoid altering the orphaned bit */ gpr_atm refst; - /* Internally stores data of type (grpc_error *). If the FD is shutdown, this - contains reason for shutdown (i.e a pointer to grpc_error) ORed with - FD_SHUTDOWN_BIT. Since address allocations are word-aligned, the lower bit - of (grpc_error *) addresses is guaranteed to be zero. Even if the - (grpc_error *), is of special types like GRPC_ERROR_NONE, GRPC_ERROR_OOM - etc, the lower bit is guaranteed to be zero. - - Once an fd is shutdown, any pending or future read/write closures on the - fd should fail */ - gpr_atm shutdown_error; - /* 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; - /* Closures to call when the fd is readable or writable respectively. These - fields contain one of the following values: - CLOSURE_READY : The fd has an I/O event of interest but there is no - closure yet to execute - - CLOSURE_NOT_READY : The fd has no I/O event of interest - - closure ptr : The closure to be executed when the fd has an I/O - event of interest - - shutdown_error | FD_SHUTDOWN_BIT : - 'shutdown_error' field ORed with FD_SHUTDOWN_BIT. - This indicates that the fd is shutdown. Since all - memory allocations are word-aligned, the lower two - bits of the shutdown_error pointer are always 0. So - it is safe to OR these with FD_SHUTDOWN_BIT - - Valid state transitions: - - <closure ptr> <-----3------ CLOSURE_NOT_READY ----1----> CLOSURE_READY - | | ^ | ^ | | - | | | | | | | - | +--------------4----------+ 6 +---------2---------------+ | - | | | - | v | - +-----5-------> [shutdown_error | FD_SHUTDOWN_BIT] <----7---------+ - - For 1, 4 : See set_ready() function - For 2, 3 : See notify_on() function - For 5,6,7: See set_shutdown() function */ gpr_atm read_closure; gpr_atm write_closure; @@ -217,11 +178,6 @@ static void fd_unref(grpc_fd *fd); static void fd_global_init(void); static void fd_global_shutdown(void); -#define CLOSURE_NOT_READY ((gpr_atm)0) -#define CLOSURE_READY ((gpr_atm)2) - -#define FD_SHUTDOWN_BIT 1 - /******************************************************************************* * Polling island Declarations */ @@ -948,10 +904,8 @@ static void unref_by(grpc_fd *fd, int n) { fd_freelist = fd; grpc_iomgr_unregister_object(&fd->iomgr_object); - grpc_error *err = (grpc_error *)gpr_atm_acq_load(&fd->shutdown_error); - /* Clear the least significant bit if it set (in case fd was shutdown) */ - err = (grpc_error *)((intptr_t)err & ~FD_SHUTDOWN_BIT); - GRPC_ERROR_UNREF(err); + grpc_lfev_destroy(&fd->read_closure); + grpc_lfev_destroy(&fd->write_closure); gpr_mu_unlock(&fd_freelist_mu); } else { @@ -1015,10 +969,9 @@ static grpc_fd *fd_create(int fd, const char *name) { gpr_atm_rel_store(&new_fd->refst, (gpr_atm)1); new_fd->fd = fd; - gpr_atm_no_barrier_store(&new_fd->shutdown_error, (gpr_atm)GRPC_ERROR_NONE); new_fd->orphaned = false; - gpr_atm_no_barrier_store(&new_fd->read_closure, CLOSURE_NOT_READY); - gpr_atm_no_barrier_store(&new_fd->write_closure, CLOSURE_NOT_READY); + 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; @@ -1104,153 +1057,6 @@ static void fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, GRPC_ERROR_UNREF(error); } -static void notify_on(grpc_exec_ctx *exec_ctx, grpc_fd *fd, gpr_atm *state, - grpc_closure *closure) { - while (true) { - gpr_atm curr = gpr_atm_no_barrier_load(state); - switch (curr) { - case CLOSURE_NOT_READY: { - /* CLOSURE_NOT_READY -> <closure>. - - We're guaranteed by API that there's an acquire barrier before here, - so there's no need to double-dip and this can be a release-only. - - The release itself pairs with the acquire half of a set_ready full - barrier. */ - if (gpr_atm_rel_cas(state, CLOSURE_NOT_READY, (gpr_atm)closure)) { - return; /* Successful. Return */ - } - - break; /* retry */ - } - - case CLOSURE_READY: { - /* Change the state to CLOSURE_NOT_READY. Schedule the closure if - successful. If not, the state most likely transitioned to shutdown. - We should retry. - - This can be a no-barrier cas since the state is being transitioned to - CLOSURE_NOT_READY; set_ready and set_shutdown do not schedule any - closure when transitioning out of CLOSURE_NO_READY state (i.e there - is no other code that needs to 'happen-after' this) */ - if (gpr_atm_no_barrier_cas(state, CLOSURE_READY, CLOSURE_NOT_READY)) { - grpc_closure_sched(exec_ctx, closure, GRPC_ERROR_NONE); - return; /* Successful. Return */ - } - - break; /* retry */ - } - - default: { - /* 'curr' is either a closure or the fd is shutdown(in which case 'curr' - contains a pointer to the shutdown-error). If the fd is shutdown, - schedule the closure with the shutdown error */ - if ((curr & FD_SHUTDOWN_BIT) > 0) { - grpc_error *shutdown_err = (grpc_error *)(curr & ~FD_SHUTDOWN_BIT); - grpc_closure_sched(exec_ctx, closure, - GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING( - "FD Shutdown", &shutdown_err, 1)); - return; - } - - /* There is already a closure!. This indicates a bug in the code */ - gpr_log(GPR_ERROR, - "notify_on called with a previous callback still pending"); - abort(); - } - } - } - - GPR_UNREACHABLE_CODE(return ); -} - -static void set_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd, gpr_atm *state, - grpc_error *shutdown_err) { - gpr_atm new_state = (gpr_atm)shutdown_err | FD_SHUTDOWN_BIT; - - while (true) { - gpr_atm curr = gpr_atm_no_barrier_load(state); - switch (curr) { - case CLOSURE_READY: - case CLOSURE_NOT_READY: - /* Need a full barrier here so that the initial load in notify_on - doesn't need a barrier */ - if (gpr_atm_full_cas(state, curr, new_state)) { - return; /* early out */ - } - break; /* retry */ - - default: { - /* 'curr' is either a closure or the fd is already shutdown */ - - /* If fd is already shutdown, we are done */ - if ((curr & FD_SHUTDOWN_BIT) > 0) { - return; - } - - /* Fd is not shutdown. Schedule the closure and move the state to - shutdown state. - Needs an acquire to pair with setting the closure (and get a - happens-after on that edge), and a release to pair with anything - loading the shutdown state. */ - if (gpr_atm_full_cas(state, curr, new_state)) { - grpc_closure_sched(exec_ctx, (grpc_closure *)curr, - GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING( - "FD Shutdown", &shutdown_err, 1)); - return; - } - - /* 'curr' was a closure but now changed to a different state. We will - have to retry */ - break; - } - } - } - - GPR_UNREACHABLE_CODE(return ); -} - -static void set_ready(grpc_exec_ctx *exec_ctx, grpc_fd *fd, gpr_atm *state) { - while (true) { - gpr_atm curr = gpr_atm_no_barrier_load(state); - - switch (curr) { - case CLOSURE_READY: { - /* Already ready. We are done here */ - return; - } - - case CLOSURE_NOT_READY: { - /* No barrier required as we're transitioning to a state that does not - involve a closure */ - if (gpr_atm_no_barrier_cas(state, CLOSURE_NOT_READY, CLOSURE_READY)) { - return; /* early out */ - } - break; /* retry */ - } - - default: { - /* 'curr' is either a closure or the fd is shutdown */ - if ((curr & FD_SHUTDOWN_BIT) > 0) { - /* The fd is shutdown. Do nothing */ - return; - } - /* Full cas: acquire pairs with this cas' release in the event of a - spurious set_ready; release pairs with this or the acquire in - notify_on (or set_shutdown) */ - else if (gpr_atm_full_cas(state, curr, CLOSURE_NOT_READY)) { - grpc_closure_sched(exec_ctx, (grpc_closure *)curr, GRPC_ERROR_NONE); - return; - } - /* else the state changed again (only possible by either a racing - set_ready or set_shutdown functions. In both these cases, the closure - would have been scheduled for execution. So we are done here */ - return; - } - } - } -} - 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); @@ -1258,33 +1064,27 @@ static grpc_pollset *fd_get_read_notifier_pollset(grpc_exec_ctx *exec_ctx, } static bool fd_is_shutdown(grpc_fd *fd) { - grpc_error *err = (grpc_error *)gpr_atm_acq_load(&fd->shutdown_error); - return (((intptr_t)err & FD_SHUTDOWN_BIT) > 0); + 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) { - /* Store the shutdown error ORed with FD_SHUTDOWN_BIT in fd->shutdown_error */ - if (gpr_atm_rel_cas(&fd->shutdown_error, (gpr_atm)GRPC_ERROR_NONE, - (gpr_atm)why | FD_SHUTDOWN_BIT)) { + if (grpc_lfev_set_shutdown(exec_ctx, &fd->read_closure, + GRPC_ERROR_REF(why))) { shutdown(fd->fd, SHUT_RDWR); - - set_shutdown(exec_ctx, fd, &fd->read_closure, why); - set_shutdown(exec_ctx, fd, &fd->write_closure, why); - } else { - /* Shutdown already called */ - GRPC_ERROR_UNREF(why); + 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) { - notify_on(exec_ctx, fd, &fd->read_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) { - notify_on(exec_ctx, fd, &fd->write_closure, closure); + grpc_lfev_notify_on(exec_ctx, &fd->write_closure, closure); } static grpc_workqueue *fd_get_workqueue(grpc_fd *fd) { @@ -1332,7 +1132,7 @@ static grpc_error *pollset_worker_kick(grpc_pollset_worker *worker) { 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, worker->pt_id); + (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"); @@ -1467,13 +1267,14 @@ static int poll_deadline_to_millis_timeout(gpr_timespec deadline, return 0; } timeout = gpr_time_sub(deadline, now); - return gpr_time_to_millis(gpr_time_add( + 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) { - set_ready(exec_ctx, fd, &fd->read_closure); + 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 @@ -1485,7 +1286,7 @@ static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd, } static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { - set_ready(exec_ctx, fd, &fd->write_closure); + grpc_lfev_set_ready(exec_ctx, &fd->write_closure); } static void pollset_release_polling_island(grpc_exec_ctx *exec_ctx, @@ -1650,6 +1451,7 @@ static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx, 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) { @@ -1714,7 +1516,7 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, worker.pt_id = pthread_self(); gpr_atm_no_barrier_store(&worker.is_kicked, (gpr_atm)0); - *worker_hdl = &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); @@ -1792,7 +1594,7 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, gpr_mu_lock(&pollset->po.mu); } - *worker_hdl = NULL; + 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); diff --git a/src/core/lib/iomgr/ev_poll_posix.c b/src/core/lib/iomgr/ev_poll_posix.c index ca6e855611..9834cdd197 100644 --- a/src/core/lib/iomgr/ev_poll_posix.c +++ b/src/core/lib/iomgr/ev_poll_posix.c @@ -52,6 +52,7 @@ #include <grpc/support/useful.h> #include "src/core/lib/iomgr/iomgr_internal.h" +#include "src/core/lib/iomgr/timer.h" #include "src/core/lib/iomgr/wakeup_fd_cv.h" #include "src/core/lib/iomgr/wakeup_fd_posix.h" #include "src/core/lib/profiling/timers.h" @@ -870,7 +871,7 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker **worker_hdl, gpr_timespec now, gpr_timespec deadline) { grpc_pollset_worker worker; - *worker_hdl = &worker; + if (worker_hdl) *worker_hdl = &worker; grpc_error *error = GRPC_ERROR_NONE; /* Avoid malloc for small number of elements. */ @@ -1006,6 +1007,7 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, } } else { if (pfds[0].revents & POLLIN_CHECK) { + grpc_timer_consume_kick(); work_combine_error(&error, grpc_wakeup_fd_consume_wakeup(&global_wakeup_fd)); } @@ -1090,7 +1092,7 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, gpr_mu_lock(&pollset->mu); } } - *worker_hdl = NULL; + if (worker_hdl) *worker_hdl = NULL; GPR_TIMER_END("pollset_work", 0); GRPC_LOG_IF_ERROR("pollset_work", GRPC_ERROR_REF(error)); return error; diff --git a/src/core/lib/iomgr/ev_posix.c b/src/core/lib/iomgr/ev_posix.c index b5be5504b9..13409a4de8 100644 --- a/src/core/lib/iomgr/ev_posix.c +++ b/src/core/lib/iomgr/ev_posix.c @@ -111,6 +111,12 @@ static void try_engine(const char *engine) { } } +/* This should be used for testing purposes ONLY */ +void grpc_set_event_engine_test_only( + const grpc_event_engine_vtable *ev_engine) { + g_event_engine = ev_engine; +} + /* Call this only after calling grpc_event_engine_init() */ const char *grpc_get_poll_strategy_name() { return g_poll_strategy_name; } diff --git a/src/core/lib/iomgr/ev_posix.h b/src/core/lib/iomgr/ev_posix.h index 1a9e5c115a..becc4d359e 100644 --- a/src/core/lib/iomgr/ev_posix.h +++ b/src/core/lib/iomgr/ev_posix.h @@ -183,4 +183,7 @@ void grpc_pollset_set_del_fd(grpc_exec_ctx *exec_ctx, typedef int (*grpc_poll_function_type)(struct pollfd *, nfds_t, int); extern grpc_poll_function_type grpc_poll_function; +/* This should be used for testing purposes ONLY */ +void grpc_set_event_engine_test_only(const grpc_event_engine_vtable *); + #endif /* GRPC_CORE_LIB_IOMGR_EV_POSIX_H */ diff --git a/src/core/lib/iomgr/lockfree_event.c b/src/core/lib/iomgr/lockfree_event.c new file mode 100644 index 0000000000..17e3bbf727 --- /dev/null +++ b/src/core/lib/iomgr/lockfree_event.c @@ -0,0 +1,238 @@ +/* + * + * 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/lockfree_event.h" + +#include <grpc/support/log.h> + +/* 'state' holds the to call when the fd is readable or writable respectively. + It can contain one of the following values: + CLOSURE_READY : The fd has an I/O event of interest but there is no + closure yet to execute + + CLOSURE_NOT_READY : The fd has no I/O event of interest + + closure ptr : The closure to be executed when the fd has an I/O + event of interest + + shutdown_error | FD_SHUTDOWN_BIT : + 'shutdown_error' field ORed with FD_SHUTDOWN_BIT. + This indicates that the fd is shutdown. Since all + memory allocations are word-aligned, the lower two + bits of the shutdown_error pointer are always 0. So + it is safe to OR these with FD_SHUTDOWN_BIT + + Valid state transitions: + + <closure ptr> <-----3------ CLOSURE_NOT_READY ----1----> CLOSURE_READY + | | ^ | ^ | | + | | | | | | | + | +--------------4----------+ 6 +---------2---------------+ | + | | | + | v | + +-----5-------> [shutdown_error | FD_SHUTDOWN_BIT] <----7---------+ + + For 1, 4 : See grpc_lfev_set_ready() function + For 2, 3 : See grpc_lfev_notify_on() function + For 5,6,7: See grpc_lfev_set_shutdown() function */ + +#define CLOSURE_NOT_READY ((gpr_atm)0) +#define CLOSURE_READY ((gpr_atm)2) + +#define FD_SHUTDOWN_BIT ((gpr_atm)1) + +void grpc_lfev_init(gpr_atm *state) { + gpr_atm_no_barrier_store(state, CLOSURE_NOT_READY); +} + +void grpc_lfev_destroy(gpr_atm *state) { + gpr_atm curr = gpr_atm_no_barrier_load(state); + if (curr & FD_SHUTDOWN_BIT) { + GRPC_ERROR_UNREF((grpc_error *)(curr & ~FD_SHUTDOWN_BIT)); + } else { + GPR_ASSERT(curr == CLOSURE_NOT_READY || curr == CLOSURE_READY); + } +} + +bool grpc_lfev_is_shutdown(gpr_atm *state) { + gpr_atm curr = gpr_atm_no_barrier_load(state); + return (curr & FD_SHUTDOWN_BIT) != 0; +} + +void grpc_lfev_notify_on(grpc_exec_ctx *exec_ctx, gpr_atm *state, + grpc_closure *closure) { + while (true) { + gpr_atm curr = gpr_atm_no_barrier_load(state); + switch (curr) { + case CLOSURE_NOT_READY: { + /* CLOSURE_NOT_READY -> <closure>. + + We're guaranteed by API that there's an acquire barrier before here, + so there's no need to double-dip and this can be a release-only. + + The release itself pairs with the acquire half of a set_ready full + barrier. */ + if (gpr_atm_rel_cas(state, CLOSURE_NOT_READY, (gpr_atm)closure)) { + return; /* Successful. Return */ + } + + break; /* retry */ + } + + case CLOSURE_READY: { + /* Change the state to CLOSURE_NOT_READY. Schedule the closure if + successful. If not, the state most likely transitioned to shutdown. + We should retry. + + This can be a no-barrier cas since the state is being transitioned to + CLOSURE_NOT_READY; set_ready and set_shutdown do not schedule any + closure when transitioning out of CLOSURE_NO_READY state (i.e there + is no other code that needs to 'happen-after' this) */ + if (gpr_atm_no_barrier_cas(state, CLOSURE_READY, CLOSURE_NOT_READY)) { + grpc_closure_sched(exec_ctx, closure, GRPC_ERROR_NONE); + return; /* Successful. Return */ + } + + break; /* retry */ + } + + default: { + /* 'curr' is either a closure or the fd is shutdown(in which case 'curr' + contains a pointer to the shutdown-error). If the fd is shutdown, + schedule the closure with the shutdown error */ + if ((curr & FD_SHUTDOWN_BIT) > 0) { + grpc_error *shutdown_err = (grpc_error *)(curr & ~FD_SHUTDOWN_BIT); + grpc_closure_sched(exec_ctx, closure, + GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING( + "FD Shutdown", &shutdown_err, 1)); + return; + } + + /* There is already a closure!. This indicates a bug in the code */ + gpr_log(GPR_ERROR, + "notify_on called with a previous callback still pending"); + abort(); + } + } + } + + GPR_UNREACHABLE_CODE(return ); +} + +bool grpc_lfev_set_shutdown(grpc_exec_ctx *exec_ctx, gpr_atm *state, + grpc_error *shutdown_err) { + gpr_atm new_state = (gpr_atm)shutdown_err | FD_SHUTDOWN_BIT; + + while (true) { + gpr_atm curr = gpr_atm_no_barrier_load(state); + switch (curr) { + case CLOSURE_READY: + case CLOSURE_NOT_READY: + /* Need a full barrier here so that the initial load in notify_on + doesn't need a barrier */ + if (gpr_atm_full_cas(state, curr, new_state)) { + return true; /* early out */ + } + break; /* retry */ + + default: { + /* 'curr' is either a closure or the fd is already shutdown */ + + /* If fd is already shutdown, we are done */ + if ((curr & FD_SHUTDOWN_BIT) > 0) { + GRPC_ERROR_UNREF(shutdown_err); + return false; + } + + /* Fd is not shutdown. Schedule the closure and move the state to + shutdown state. + Needs an acquire to pair with setting the closure (and get a + happens-after on that edge), and a release to pair with anything + loading the shutdown state. */ + if (gpr_atm_full_cas(state, curr, new_state)) { + grpc_closure_sched(exec_ctx, (grpc_closure *)curr, + GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING( + "FD Shutdown", &shutdown_err, 1)); + return true; + } + + /* 'curr' was a closure but now changed to a different state. We will + have to retry */ + break; + } + } + } + + GPR_UNREACHABLE_CODE(return false); +} + +void grpc_lfev_set_ready(grpc_exec_ctx *exec_ctx, gpr_atm *state) { + while (true) { + gpr_atm curr = gpr_atm_no_barrier_load(state); + + switch (curr) { + case CLOSURE_READY: { + /* Already ready. We are done here */ + return; + } + + case CLOSURE_NOT_READY: { + /* No barrier required as we're transitioning to a state that does not + involve a closure */ + if (gpr_atm_no_barrier_cas(state, CLOSURE_NOT_READY, CLOSURE_READY)) { + return; /* early out */ + } + break; /* retry */ + } + + default: { + /* 'curr' is either a closure or the fd is shutdown */ + if ((curr & FD_SHUTDOWN_BIT) > 0) { + /* The fd is shutdown. Do nothing */ + return; + } + /* Full cas: acquire pairs with this cas' release in the event of a + spurious set_ready; release pairs with this or the acquire in + notify_on (or set_shutdown) */ + else if (gpr_atm_full_cas(state, curr, CLOSURE_NOT_READY)) { + grpc_closure_sched(exec_ctx, (grpc_closure *)curr, GRPC_ERROR_NONE); + return; + } + /* else the state changed again (only possible by either a racing + set_ready or set_shutdown functions. In both these cases, the closure + would have been scheduled for execution. So we are done here */ + return; + } + } + } +} diff --git a/src/core/lib/iomgr/lockfree_event.h b/src/core/lib/iomgr/lockfree_event.h new file mode 100644 index 0000000000..1d9119204c --- /dev/null +++ b/src/core/lib/iomgr/lockfree_event.h @@ -0,0 +1,54 @@ +/* + * + * 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. + * + */ + +#ifndef GRPC_CORE_LIB_IOMGR_LOCKFREE_EVENT_H +#define GRPC_CORE_LIB_IOMGR_LOCKFREE_EVENT_H + +/* Lock free event notification for file descriptors */ + +#include <grpc/support/atm.h> + +#include "src/core/lib/iomgr/exec_ctx.h" + +void grpc_lfev_init(gpr_atm *state); +void grpc_lfev_destroy(gpr_atm *state); +bool grpc_lfev_is_shutdown(gpr_atm *state); + +void grpc_lfev_notify_on(grpc_exec_ctx *exec_ctx, gpr_atm *state, + grpc_closure *closure); +/* Returns true on first successful shutdown */ +bool grpc_lfev_set_shutdown(grpc_exec_ctx *exec_ctx, gpr_atm *state, + grpc_error *shutdown_err); +void grpc_lfev_set_ready(grpc_exec_ctx *exec_ctx, gpr_atm *state); + +#endif /* GRPC_CORE_LIB_IOMGR_LOCKFREE_EVENT_H */ diff --git a/src/core/lib/iomgr/pollset.h b/src/core/lib/iomgr/pollset.h index e19ce697b8..9bf3cdac89 100644 --- a/src/core/lib/iomgr/pollset.h +++ b/src/core/lib/iomgr/pollset.h @@ -75,6 +75,10 @@ void grpc_pollset_destroy(grpc_pollset *pollset); and it is guaranteed that it will not be released by grpc_pollset_work AFTER worker has been destroyed. + It's legal for worker to be NULL: in that case, this specific thread can not + be directly woken with a kick, but maybe be indirectly (with a kick against + the pollset as a whole). + Tries not to block past deadline. May call grpc_closure_list_run on grpc_closure_list, without holding the pollset diff --git a/src/core/lib/iomgr/pollset_windows.c b/src/core/lib/iomgr/pollset_windows.c index 17043c1ea1..04c6b71747 100644 --- a/src/core/lib/iomgr/pollset_windows.c +++ b/src/core/lib/iomgr/pollset_windows.c @@ -120,7 +120,7 @@ grpc_error *grpc_pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker **worker_hdl, gpr_timespec now, gpr_timespec deadline) { grpc_pollset_worker worker; - *worker_hdl = &worker; + if (worker_hdl) *worker_hdl = &worker; int added_worker = 0; worker.links[GRPC_POLLSET_WORKER_LINK_POLLSET].next = @@ -185,7 +185,7 @@ done: remove_worker(&worker, GRPC_POLLSET_WORKER_LINK_POLLSET); } gpr_cv_destroy(&worker.cv); - *worker_hdl = NULL; + if (worker_hdl) *worker_hdl = NULL; return GRPC_ERROR_NONE; } diff --git a/src/core/lib/iomgr/port.h b/src/core/lib/iomgr/port.h index 94a454c0b7..269dc35003 100644 --- a/src/core/lib/iomgr/port.h +++ b/src/core/lib/iomgr/port.h @@ -85,6 +85,10 @@ #define GRPC_LINUX_SOCKETUTILS 1 #endif #endif +#ifndef __GLIBC__ +#define GRPC_LINUX_EPOLL 1 +#define GRPC_LINUX_EVENTFD 1 +#endif #ifndef GRPC_LINUX_EVENTFD #define GRPC_POSIX_NO_SPECIAL_WAKEUP_FD 1 #endif diff --git a/src/core/lib/iomgr/resource_quota.c b/src/core/lib/iomgr/resource_quota.c index 8dcd80d001..c3ee878651 100644 --- a/src/core/lib/iomgr/resource_quota.c +++ b/src/core/lib/iomgr/resource_quota.c @@ -142,6 +142,8 @@ struct grpc_resource_quota { /* Amount of free memory in the resource quota */ int64_t free_pool; + gpr_atm last_size; + /* Has rq_step been scheduled to occur? */ bool step_scheduled; /* Are we currently reclaiming memory */ @@ -581,6 +583,7 @@ grpc_resource_quota *grpc_resource_quota_create(const char *name) { resource_quota->combiner = grpc_combiner_create(NULL); resource_quota->free_pool = INT64_MAX; resource_quota->size = INT64_MAX; + gpr_atm_no_barrier_store(&resource_quota->last_size, GPR_ATM_MAX); resource_quota->step_scheduled = false; resource_quota->reclaiming = false; gpr_atm_no_barrier_store(&resource_quota->memory_usage_estimation, 0); @@ -643,11 +646,17 @@ void grpc_resource_quota_resize(grpc_resource_quota *resource_quota, rq_resize_args *a = gpr_malloc(sizeof(*a)); a->resource_quota = grpc_resource_quota_ref_internal(resource_quota); a->size = (int64_t)size; + gpr_atm_no_barrier_store(&resource_quota->last_size, + (gpr_atm)GPR_MIN((size_t)GPR_ATM_MAX, size)); grpc_closure_init(&a->closure, rq_resize, a, grpc_schedule_on_exec_ctx); grpc_closure_sched(&exec_ctx, &a->closure, GRPC_ERROR_NONE); grpc_exec_ctx_finish(&exec_ctx); } +size_t grpc_resource_quota_peek_size(grpc_resource_quota *resource_quota) { + return (size_t)gpr_atm_no_barrier_load(&resource_quota->last_size); +} + /******************************************************************************* * grpc_resource_user channel args api */ diff --git a/src/core/lib/iomgr/resource_quota.h b/src/core/lib/iomgr/resource_quota.h index b9f62cbf83..6f99be0d51 100644 --- a/src/core/lib/iomgr/resource_quota.h +++ b/src/core/lib/iomgr/resource_quota.h @@ -90,6 +90,8 @@ grpc_resource_quota *grpc_resource_quota_from_channel_args( double grpc_resource_quota_get_memory_pressure( grpc_resource_quota *resource_quota); +size_t grpc_resource_quota_peek_size(grpc_resource_quota *resource_quota); + typedef struct grpc_resource_user grpc_resource_user; grpc_resource_user *grpc_resource_user_create( diff --git a/src/core/lib/iomgr/tcp_client.h b/src/core/lib/iomgr/tcp_client.h index 0485661316..bc367bdfa5 100644 --- a/src/core/lib/iomgr/tcp_client.h +++ b/src/core/lib/iomgr/tcp_client.h @@ -40,10 +40,6 @@ #include "src/core/lib/iomgr/pollset_set.h" #include "src/core/lib/iomgr/resolve_address.h" -/* Channel arg (integer) setting how large a slice to try and read from the wire - each time recvmsg (or equivalent) is called */ -#define GRPC_ARG_TCP_READ_CHUNK_SIZE "grpc.experimental.tcp_read_chunk_size" - /* Asynchronously connect to an address (specified as (addr, len)), and call cb with arg and the completed connection when done (or call cb with arg and NULL on failure). diff --git a/src/core/lib/iomgr/tcp_client_posix.c b/src/core/lib/iomgr/tcp_client_posix.c index a108b10da6..a2692707d9 100644 --- a/src/core/lib/iomgr/tcp_client_posix.c +++ b/src/core/lib/iomgr/tcp_client_posix.c @@ -137,29 +137,7 @@ static void tc_on_alarm(grpc_exec_ctx *exec_ctx, void *acp, grpc_error *error) { grpc_endpoint *grpc_tcp_client_create_from_fd( grpc_exec_ctx *exec_ctx, grpc_fd *fd, const grpc_channel_args *channel_args, const char *addr_str) { - size_t tcp_read_chunk_size = GRPC_TCP_DEFAULT_READ_SLICE_SIZE; - grpc_resource_quota *resource_quota = grpc_resource_quota_create(NULL); - if (channel_args != NULL) { - for (size_t i = 0; i < channel_args->num_args; i++) { - if (0 == - strcmp(channel_args->args[i].key, GRPC_ARG_TCP_READ_CHUNK_SIZE)) { - grpc_integer_options options = {(int)tcp_read_chunk_size, 1, - 8 * 1024 * 1024}; - tcp_read_chunk_size = (size_t)grpc_channel_arg_get_integer( - &channel_args->args[i], options); - } else if (0 == - strcmp(channel_args->args[i].key, GRPC_ARG_RESOURCE_QUOTA)) { - grpc_resource_quota_unref_internal(exec_ctx, resource_quota); - resource_quota = grpc_resource_quota_ref_internal( - channel_args->args[i].value.pointer.p); - } - } - } - - grpc_endpoint *ep = - grpc_tcp_create(fd, resource_quota, tcp_read_chunk_size, addr_str); - grpc_resource_quota_unref_internal(exec_ctx, resource_quota); - return ep; + return grpc_tcp_create(exec_ctx, fd, channel_args, addr_str); } static void on_writable(grpc_exec_ctx *exec_ctx, void *acp, grpc_error *error) { diff --git a/src/core/lib/iomgr/tcp_client_windows.c b/src/core/lib/iomgr/tcp_client_windows.c index a356564766..d6baca50ba 100644 --- a/src/core/lib/iomgr/tcp_client_windows.c +++ b/src/core/lib/iomgr/tcp_client_windows.c @@ -63,7 +63,7 @@ typedef struct { int refs; grpc_closure on_connect; grpc_endpoint **endpoint; - grpc_resource_quota *resource_quota; + grpc_channel_args *channel_args; } async_connect; static void async_connect_unlock_and_cleanup(grpc_exec_ctx *exec_ctx, @@ -72,7 +72,7 @@ static void async_connect_unlock_and_cleanup(grpc_exec_ctx *exec_ctx, int done = (--ac->refs == 0); gpr_mu_unlock(&ac->mu); if (done) { - grpc_resource_quota_unref_internal(exec_ctx, ac->resource_quota); + grpc_channel_args_destroy(exec_ctx, ac->channel_args); gpr_mu_destroy(&ac->mu); gpr_free(ac->addr_name); gpr_free(ac); @@ -119,7 +119,8 @@ static void on_connect(grpc_exec_ctx *exec_ctx, void *acp, grpc_error *error) { if (!wsa_success) { error = GRPC_WSA_ERROR(WSAGetLastError(), "ConnectEx"); } else { - *ep = grpc_tcp_create(socket, ac->resource_quota, ac->addr_name); + *ep = + grpc_tcp_create(exec_ctx, socket, ac->channel_args, ac->addr_name); socket = NULL; } } else { @@ -152,17 +153,6 @@ static void tcp_client_connect_impl( grpc_winsocket_callback_info *info; grpc_error *error = GRPC_ERROR_NONE; - grpc_resource_quota *resource_quota = grpc_resource_quota_create(NULL); - if (channel_args != NULL) { - for (size_t i = 0; i < channel_args->num_args; i++) { - if (0 == strcmp(channel_args->args[i].key, GRPC_ARG_RESOURCE_QUOTA)) { - grpc_resource_quota_unref_internal(exec_ctx, resource_quota); - resource_quota = grpc_resource_quota_ref_internal( - channel_args->args[i].value.pointer.p); - } - } - } - *endpoint = NULL; /* Use dualstack sockets where available. */ @@ -225,7 +215,7 @@ static void tcp_client_connect_impl( ac->refs = 2; ac->addr_name = grpc_sockaddr_to_uri(addr); ac->endpoint = endpoint; - ac->resource_quota = resource_quota; + ac->channel_args = grpc_channel_args_copy(channel_args); grpc_closure_init(&ac->on_connect, on_connect, ac, grpc_schedule_on_exec_ctx); grpc_closure_init(&ac->on_alarm, on_alarm, ac, grpc_schedule_on_exec_ctx); @@ -247,7 +237,6 @@ failure: } else if (sock != INVALID_SOCKET) { closesocket(sock); } - grpc_resource_quota_unref_internal(exec_ctx, resource_quota); grpc_closure_sched(exec_ctx, on_done, final_error); } diff --git a/src/core/lib/iomgr/tcp_posix.c b/src/core/lib/iomgr/tcp_posix.c index 4d7cf3ff51..5f4b38de2b 100644 --- a/src/core/lib/iomgr/tcp_posix.c +++ b/src/core/lib/iomgr/tcp_posix.c @@ -52,7 +52,9 @@ #include <grpc/support/string_util.h> #include <grpc/support/sync.h> #include <grpc/support/time.h> +#include <grpc/support/useful.h> +#include "src/core/lib/channel/channel_args.h" #include "src/core/lib/debug/trace.h" #include "src/core/lib/iomgr/ev_posix.h" #include "src/core/lib/profiling/timers.h" @@ -80,10 +82,14 @@ typedef struct { int fd; bool finished_edge; msg_iovlen_type iov_size; /* Number of slices to allocate per read attempt */ - size_t slice_size; + double target_length; + double bytes_read_this_round; gpr_refcount refcount; gpr_atm shutdown_count; + int min_read_chunk_size; + int max_read_chunk_size; + /* garbage after the last read */ grpc_slice_buffer last_read_buffer; @@ -108,6 +114,42 @@ typedef struct { grpc_resource_user_slice_allocator slice_allocator; } grpc_tcp; +static void add_to_estimate(grpc_tcp *tcp, size_t bytes) { + tcp->bytes_read_this_round += (double)bytes; +} + +static void finish_estimate(grpc_tcp *tcp) { + /* If we read >80% of the target buffer in one read loop, increase the size + of the target buffer to either the amount read, or twice its previous + value */ + if (tcp->bytes_read_this_round > tcp->target_length * 0.8) { + tcp->target_length = + GPR_MAX(2 * tcp->target_length, tcp->bytes_read_this_round); + } else { + tcp->target_length = + 0.99 * tcp->target_length + 0.01 * tcp->bytes_read_this_round; + } + tcp->bytes_read_this_round = 0; +} + +static size_t get_target_read_size(grpc_tcp *tcp) { + grpc_resource_quota *rq = grpc_resource_user_quota(tcp->resource_user); + double pressure = grpc_resource_quota_get_memory_pressure(rq); + double target = + tcp->target_length * (pressure > 0.8 ? (1.0 - pressure) / 0.2 : 1.0); + size_t sz = (((size_t)GPR_CLAMP(target, tcp->min_read_chunk_size, + tcp->max_read_chunk_size)) + + 255) & + ~(size_t)255; + /* don't use more than 1/16th of the overall resource quota for a single read + * alloc */ + size_t rqmax = grpc_resource_quota_peek_size(rq); + if (sz > rqmax / 16 && rqmax > 1024) { + sz = rqmax / 16; + } + return sz; +} + static grpc_error *tcp_annotate_error(grpc_error *src_error, grpc_tcp *tcp) { return grpc_error_set_str( grpc_error_set_int(src_error, GRPC_ERROR_INT_FD, tcp->fd), @@ -232,9 +274,7 @@ static void tcp_do_read(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) { /* NB: After calling call_read_cb a parallel call of the read handler may * be running. */ if (errno == EAGAIN) { - if (tcp->iov_size > 1) { - tcp->iov_size /= 2; - } + finish_estimate(tcp); /* We've consumed the edge, request a new one */ grpc_fd_notify_on_read(exec_ctx, tcp->em_fd, &tcp->read_closure); } else { @@ -253,14 +293,13 @@ static void tcp_do_read(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) { GRPC_ERROR_CREATE_FROM_STATIC_STRING("Socket closed"), tcp)); TCP_UNREF(exec_ctx, tcp, "read"); } else { + add_to_estimate(tcp, (size_t)read_bytes); GPR_ASSERT((size_t)read_bytes <= tcp->incoming_buffer->length); if ((size_t)read_bytes < tcp->incoming_buffer->length) { grpc_slice_buffer_trim_end( tcp->incoming_buffer, tcp->incoming_buffer->length - (size_t)read_bytes, &tcp->last_read_buffer); - } else if (tcp->iov_size < MAX_READ_IOVEC) { - ++tcp->iov_size; } GPR_ASSERT((size_t)read_bytes == tcp->incoming_buffer->length); call_read_cb(exec_ctx, tcp, GRPC_ERROR_NONE); @@ -285,11 +324,11 @@ static void tcp_read_allocation_done(grpc_exec_ctx *exec_ctx, void *tcpp, } static void tcp_continue_read(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) { - if (tcp->incoming_buffer->count < (size_t)tcp->iov_size) { - grpc_resource_user_alloc_slices( - exec_ctx, &tcp->slice_allocator, tcp->slice_size, - (size_t)tcp->iov_size - tcp->incoming_buffer->count, - tcp->incoming_buffer); + size_t target_read_size = get_target_read_size(tcp); + if (tcp->incoming_buffer->length < target_read_size && + tcp->incoming_buffer->count < MAX_READ_IOVEC) { + grpc_resource_user_alloc_slices(exec_ctx, &tcp->slice_allocator, + target_read_size, 1, tcp->incoming_buffer); } else { tcp_do_read(exec_ctx, tcp); } @@ -540,9 +579,50 @@ static const grpc_endpoint_vtable vtable = {tcp_read, tcp_get_peer, tcp_get_fd}; -grpc_endpoint *grpc_tcp_create(grpc_fd *em_fd, - grpc_resource_quota *resource_quota, - size_t slice_size, const char *peer_string) { +#define MAX_CHUNK_SIZE 32 * 1024 * 1024 + +grpc_endpoint *grpc_tcp_create(grpc_exec_ctx *exec_ctx, grpc_fd *em_fd, + const grpc_channel_args *channel_args, + const char *peer_string) { + int tcp_read_chunk_size = GRPC_TCP_DEFAULT_READ_SLICE_SIZE; + int tcp_max_read_chunk_size = 4 * 1024 * 1024; + int tcp_min_read_chunk_size = 256; + grpc_resource_quota *resource_quota = grpc_resource_quota_create(NULL); + if (channel_args != NULL) { + for (size_t i = 0; i < channel_args->num_args; i++) { + if (0 == + strcmp(channel_args->args[i].key, GRPC_ARG_TCP_READ_CHUNK_SIZE)) { + grpc_integer_options options = {(int)tcp_read_chunk_size, 1, + MAX_CHUNK_SIZE}; + tcp_read_chunk_size = + grpc_channel_arg_get_integer(&channel_args->args[i], options); + } else if (0 == strcmp(channel_args->args[i].key, + GRPC_ARG_TCP_MIN_READ_CHUNK_SIZE)) { + grpc_integer_options options = {(int)tcp_read_chunk_size, 1, + MAX_CHUNK_SIZE}; + tcp_min_read_chunk_size = + grpc_channel_arg_get_integer(&channel_args->args[i], options); + } else if (0 == strcmp(channel_args->args[i].key, + GRPC_ARG_TCP_MAX_READ_CHUNK_SIZE)) { + grpc_integer_options options = {(int)tcp_read_chunk_size, 1, + MAX_CHUNK_SIZE}; + tcp_max_read_chunk_size = + grpc_channel_arg_get_integer(&channel_args->args[i], options); + } else if (0 == + strcmp(channel_args->args[i].key, GRPC_ARG_RESOURCE_QUOTA)) { + grpc_resource_quota_unref_internal(exec_ctx, resource_quota); + resource_quota = grpc_resource_quota_ref_internal( + channel_args->args[i].value.pointer.p); + } + } + } + + if (tcp_min_read_chunk_size > tcp_max_read_chunk_size) { + tcp_min_read_chunk_size = tcp_max_read_chunk_size; + } + tcp_read_chunk_size = GPR_CLAMP(tcp_read_chunk_size, tcp_min_read_chunk_size, + tcp_max_read_chunk_size); + grpc_tcp *tcp = (grpc_tcp *)gpr_malloc(sizeof(grpc_tcp)); tcp->base.vtable = &vtable; tcp->peer_string = gpr_strdup(peer_string); @@ -552,7 +632,10 @@ grpc_endpoint *grpc_tcp_create(grpc_fd *em_fd, tcp->release_fd_cb = NULL; tcp->release_fd = NULL; tcp->incoming_buffer = NULL; - tcp->slice_size = slice_size; + tcp->target_length = (double)tcp_read_chunk_size; + tcp->min_read_chunk_size = tcp_min_read_chunk_size; + tcp->max_read_chunk_size = tcp_max_read_chunk_size; + tcp->bytes_read_this_round = 0; tcp->iov_size = 1; tcp->finished_edge = true; /* paired with unref in grpc_tcp_destroy */ @@ -569,6 +652,7 @@ grpc_endpoint *grpc_tcp_create(grpc_fd *em_fd, &tcp->slice_allocator, tcp->resource_user, tcp_read_allocation_done, tcp); /* Tell network status tracker about new endpoint */ grpc_network_status_register_endpoint(&tcp->base); + grpc_resource_quota_unref_internal(exec_ctx, resource_quota); return &tcp->base; } diff --git a/src/core/lib/iomgr/tcp_posix.h b/src/core/lib/iomgr/tcp_posix.h index 1c0d13f96e..1ad5788331 100644 --- a/src/core/lib/iomgr/tcp_posix.h +++ b/src/core/lib/iomgr/tcp_posix.h @@ -47,14 +47,13 @@ #include "src/core/lib/iomgr/endpoint.h" #include "src/core/lib/iomgr/ev_posix.h" -#define GRPC_TCP_DEFAULT_READ_SLICE_SIZE 8192 - extern int grpc_tcp_trace; /* Create a tcp endpoint given a file desciptor and a read slice size. Takes ownership of fd. */ -grpc_endpoint *grpc_tcp_create(grpc_fd *fd, grpc_resource_quota *resource_quota, - size_t read_slice_size, const char *peer_string); +grpc_endpoint *grpc_tcp_create(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + const grpc_channel_args *args, + const char *peer_string); /* Return the tcp endpoint's fd, or -1 if this is not available. Does not release the fd. diff --git a/src/core/lib/iomgr/tcp_server_posix.c b/src/core/lib/iomgr/tcp_server_posix.c index d6a017cf7f..e66ffc9b1c 100644 --- a/src/core/lib/iomgr/tcp_server_posix.c +++ b/src/core/lib/iomgr/tcp_server_posix.c @@ -59,6 +59,7 @@ #include <grpc/support/time.h> #include <grpc/support/useful.h> +#include "src/core/lib/channel/channel_args.h" #include "src/core/lib/iomgr/resolve_address.h" #include "src/core/lib/iomgr/sockaddr.h" #include "src/core/lib/iomgr/sockaddr_utils.h" @@ -90,7 +91,6 @@ grpc_error *grpc_tcp_server_create(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s = gpr_zalloc(sizeof(grpc_tcp_server)); s->so_reuseport = has_so_reuseport; - s->resource_quota = grpc_resource_quota_create(NULL); s->expand_wildcard_addrs = false; for (size_t i = 0; i < (args == NULL ? 0 : args->num_args); i++) { if (0 == strcmp(GRPC_ARG_ALLOW_REUSEPORT, args->args[i].key)) { @@ -98,27 +98,14 @@ grpc_error *grpc_tcp_server_create(grpc_exec_ctx *exec_ctx, s->so_reuseport = has_so_reuseport && (args->args[i].value.integer != 0); } else { - grpc_resource_quota_unref_internal(exec_ctx, s->resource_quota); gpr_free(s); return GRPC_ERROR_CREATE_FROM_STATIC_STRING(GRPC_ARG_ALLOW_REUSEPORT " must be an integer"); } - } else if (0 == strcmp(GRPC_ARG_RESOURCE_QUOTA, args->args[i].key)) { - if (args->args[i].type == GRPC_ARG_POINTER) { - grpc_resource_quota_unref_internal(exec_ctx, s->resource_quota); - s->resource_quota = - grpc_resource_quota_ref_internal(args->args[i].value.pointer.p); - } else { - grpc_resource_quota_unref_internal(exec_ctx, s->resource_quota); - gpr_free(s); - return GRPC_ERROR_CREATE_FROM_STATIC_STRING( - GRPC_ARG_RESOURCE_QUOTA " must be a pointer to a buffer pool"); - } } else if (0 == strcmp(GRPC_ARG_EXPAND_WILDCARD_ADDRS, args->args[i].key)) { if (args->args[i].type == GRPC_ARG_INTEGER) { s->expand_wildcard_addrs = (args->args[i].value.integer != 0); } else { - grpc_resource_quota_unref_internal(exec_ctx, s->resource_quota); gpr_free(s); return GRPC_ERROR_CREATE_FROM_STATIC_STRING( GRPC_ARG_EXPAND_WILDCARD_ADDRS " must be an integer"); @@ -138,6 +125,7 @@ grpc_error *grpc_tcp_server_create(grpc_exec_ctx *exec_ctx, s->head = NULL; s->tail = NULL; s->nports = 0; + s->channel_args = grpc_channel_args_copy(args); gpr_atm_no_barrier_store(&s->next_pollset_to_assign, 0); *server = s; return GRPC_ERROR_NONE; @@ -158,8 +146,7 @@ static void finish_shutdown(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s) { s->head = sp->next; gpr_free(sp); } - - grpc_resource_quota_unref_internal(exec_ctx, s->resource_quota); + grpc_channel_args_destroy(exec_ctx, s->channel_args); gpr_free(s); } @@ -286,8 +273,7 @@ static void on_read(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *err) { sp->server->on_accept_cb( exec_ctx, sp->server->on_accept_cb_arg, - grpc_tcp_create(fdobj, sp->server->resource_quota, - GRPC_TCP_DEFAULT_READ_SLICE_SIZE, addr_str), + grpc_tcp_create(exec_ctx, fdobj, sp->server->channel_args, addr_str), read_notifier_pollset, acceptor); gpr_free(name); diff --git a/src/core/lib/iomgr/tcp_server_utils_posix.h b/src/core/lib/iomgr/tcp_server_utils_posix.h index f5dc8532f9..c15e2e1493 100644 --- a/src/core/lib/iomgr/tcp_server_utils_posix.h +++ b/src/core/lib/iomgr/tcp_server_utils_posix.h @@ -103,7 +103,8 @@ struct grpc_tcp_server { /* next pollset to assign a channel to */ gpr_atm next_pollset_to_assign; - grpc_resource_quota *resource_quota; + /* channel args for this server */ + grpc_channel_args *channel_args; }; /* If successful, add a listener to \a s for \a addr, set \a dsmode for the diff --git a/src/core/lib/iomgr/tcp_server_windows.c b/src/core/lib/iomgr/tcp_server_windows.c index 12ce7d3fdd..4c17f08918 100644 --- a/src/core/lib/iomgr/tcp_server_windows.c +++ b/src/core/lib/iomgr/tcp_server_windows.c @@ -46,6 +46,7 @@ #include <grpc/support/sync.h> #include <grpc/support/time.h> +#include "src/core/lib/channel/channel_args.h" #include "src/core/lib/iomgr/iocp_windows.h" #include "src/core/lib/iomgr/pollset_windows.h" #include "src/core/lib/iomgr/resolve_address.h" @@ -102,7 +103,7 @@ struct grpc_tcp_server { /* shutdown callback */ grpc_closure *shutdown_complete; - grpc_resource_quota *resource_quota; + grpc_channel_args *channel_args; }; /* Public function. Allocates the proper data structures to hold a @@ -112,21 +113,7 @@ grpc_error *grpc_tcp_server_create(grpc_exec_ctx *exec_ctx, const grpc_channel_args *args, grpc_tcp_server **server) { grpc_tcp_server *s = gpr_malloc(sizeof(grpc_tcp_server)); - s->resource_quota = grpc_resource_quota_create(NULL); - for (size_t i = 0; i < (args == NULL ? 0 : args->num_args); i++) { - if (0 == strcmp(GRPC_ARG_RESOURCE_QUOTA, args->args[i].key)) { - if (args->args[i].type == GRPC_ARG_POINTER) { - grpc_resource_quota_unref_internal(exec_ctx, s->resource_quota); - s->resource_quota = - grpc_resource_quota_ref_internal(args->args[i].value.pointer.p); - } else { - grpc_resource_quota_unref_internal(exec_ctx, s->resource_quota); - gpr_free(s); - return GRPC_ERROR_CREATE_FROM_STATIC_STRING( - GRPC_ARG_RESOURCE_QUOTA " must be a pointer to a buffer pool"); - } - } - } + s->channel_args = grpc_channel_args_copy(args); gpr_ref_init(&s->refs, 1); gpr_mu_init(&s->mu); s->active_ports = 0; @@ -155,7 +142,7 @@ static void destroy_server(grpc_exec_ctx *exec_ctx, void *arg, grpc_winsocket_destroy(sp->socket); gpr_free(sp); } - grpc_resource_quota_unref_internal(exec_ctx, s->resource_quota); + grpc_channel_args_destroy(exec_ctx, s->channel_args); gpr_free(s); } @@ -383,8 +370,8 @@ static void on_accept(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { gpr_free(utf8_message); } gpr_asprintf(&fd_name, "tcp_server:%s", peer_name_string); - ep = grpc_tcp_create(grpc_winsocket_create(sock, fd_name), - sp->server->resource_quota, peer_name_string); + ep = grpc_tcp_create(exec_ctx, grpc_winsocket_create(sock, fd_name), + sp->server->channel_args, peer_name_string); gpr_free(fd_name); gpr_free(peer_name_string); } else { diff --git a/src/core/lib/iomgr/tcp_windows.c b/src/core/lib/iomgr/tcp_windows.c index 9134883226..f74aa68793 100644 --- a/src/core/lib/iomgr/tcp_windows.c +++ b/src/core/lib/iomgr/tcp_windows.c @@ -430,9 +430,19 @@ static grpc_endpoint_vtable vtable = {win_read, win_get_peer, win_get_fd}; -grpc_endpoint *grpc_tcp_create(grpc_winsocket *socket, - grpc_resource_quota *resource_quota, +grpc_endpoint *grpc_tcp_create(grpc_exec_ctx *exec_ctx, grpc_winsocket *socket, + grpc_channel_args *channel_args, char *peer_string) { + grpc_resource_quota *resource_quota = grpc_resource_quota_create(NULL); + if (channel_args != NULL) { + for (size_t i = 0; i < channel_args->num_args; i++) { + if (0 == strcmp(channel_args->args[i].key, GRPC_ARG_RESOURCE_QUOTA)) { + grpc_resource_quota_unref_internal(exec_ctx, resource_quota); + resource_quota = grpc_resource_quota_ref_internal( + channel_args->args[i].value.pointer.p); + } + } + } grpc_tcp *tcp = (grpc_tcp *)gpr_malloc(sizeof(grpc_tcp)); memset(tcp, 0, sizeof(grpc_tcp)); tcp->base.vtable = &vtable; diff --git a/src/core/lib/iomgr/tcp_windows.h b/src/core/lib/iomgr/tcp_windows.h index 4402de1c38..abafdb22d2 100644 --- a/src/core/lib/iomgr/tcp_windows.h +++ b/src/core/lib/iomgr/tcp_windows.h @@ -50,8 +50,8 @@ /* Create a tcp endpoint given a winsock handle. * Takes ownership of the handle. */ -grpc_endpoint *grpc_tcp_create(grpc_winsocket *socket, - grpc_resource_quota *resource_quota, +grpc_endpoint *grpc_tcp_create(grpc_exec_ctx *exec_ctx, grpc_winsocket *socket, + grpc_channel_args *channel_args, char *peer_string); grpc_error *grpc_tcp_prepare_socket(SOCKET sock); diff --git a/src/core/lib/iomgr/timer.h b/src/core/lib/iomgr/timer.h index d84a278b18..e0338f93c7 100644 --- a/src/core/lib/iomgr/timer.h +++ b/src/core/lib/iomgr/timer.h @@ -101,6 +101,9 @@ bool grpc_timer_check(grpc_exec_ctx *exec_ctx, gpr_timespec now, void grpc_timer_list_init(gpr_timespec now); void grpc_timer_list_shutdown(grpc_exec_ctx *exec_ctx); +/* Consume a kick issued by grpc_kick_poller */ +void grpc_timer_consume_kick(void); + /* the following must be implemented by each iomgr implementation */ void grpc_kick_poller(void); diff --git a/src/core/lib/iomgr/timer_generic.c b/src/core/lib/iomgr/timer_generic.c index e53c801929..d8e6068431 100644 --- a/src/core/lib/iomgr/timer_generic.c +++ b/src/core/lib/iomgr/timer_generic.c @@ -37,9 +37,13 @@ #include "src/core/lib/iomgr/timer.h" +#include <grpc/support/alloc.h> #include <grpc/support/log.h> +#include <grpc/support/string_util.h> #include <grpc/support/sync.h> +#include <grpc/support/tls.h> #include <grpc/support/useful.h> +#include "src/core/lib/debug/trace.h" #include "src/core/lib/iomgr/time_averaged_stats.h" #include "src/core/lib/iomgr/timer_heap.h" #include "src/core/lib/support/spinlock.h" @@ -52,12 +56,15 @@ #define MIN_QUEUE_WINDOW_DURATION 0.01 #define MAX_QUEUE_WINDOW_DURATION 1 +int grpc_timer_trace = 0; +int grpc_timer_check_trace = 0; + typedef struct { gpr_mu mu; grpc_time_averaged_stats stats; /* All and only timers with deadlines <= this will be in the heap. */ - gpr_timespec queue_deadline_cap; - gpr_timespec min_deadline; + gpr_atm queue_deadline_cap; + gpr_atm min_deadline; /* Index in the g_shard_queue */ uint32_t shard_queue_index; /* This holds all timers with deadlines < queue_deadline_cap. Timers in this @@ -67,38 +74,92 @@ typedef struct { grpc_timer list; } shard_type; -/* Protects g_shard_queue */ -static gpr_mu g_mu; -/* Allow only one run_some_expired_timers at once */ -static gpr_spinlock g_checker_mu = GPR_SPINLOCK_STATIC_INITIALIZER; +struct shared_mutables { + gpr_atm min_timer; + /* Allow only one run_some_expired_timers at once */ + gpr_spinlock checker_mu; + bool initialized; + /* Protects g_shard_queue */ + gpr_mu mu; +} GPR_ALIGN_STRUCT(GPR_CACHELINE_SIZE); + +static struct shared_mutables g_shared_mutables = { + .checker_mu = GPR_SPINLOCK_STATIC_INITIALIZER, .initialized = false, +}; static gpr_clock_type g_clock_type; static shard_type g_shards[NUM_SHARDS]; -/* Protected by g_mu */ +/* Protected by g_shared_mutables.mu */ static shard_type *g_shard_queue[NUM_SHARDS]; -static bool g_initialized = false; +static gpr_timespec g_start_time; + +GPR_TLS_DECL(g_last_seen_min_timer); + +static gpr_atm saturating_add(gpr_atm a, gpr_atm b) { + if (a > GPR_ATM_MAX - b) { + return GPR_ATM_MAX; + } + return a + b; +} + +static int run_some_expired_timers(grpc_exec_ctx *exec_ctx, gpr_atm now, + gpr_atm *next, grpc_error *error); + +static gpr_timespec dbl_to_ts(double d) { + gpr_timespec ts; + ts.tv_sec = (int64_t)d; + ts.tv_nsec = (int32_t)(1e9 * (d - (double)ts.tv_sec)); + ts.clock_type = GPR_TIMESPAN; + return ts; +} + +static gpr_atm timespec_to_atm_round_up(gpr_timespec ts) { + ts = gpr_time_sub(ts, g_start_time); + double x = GPR_MS_PER_SEC * (double)ts.tv_sec + + (double)ts.tv_nsec / GPR_NS_PER_MS + + (double)(GPR_NS_PER_SEC - 1) / (double)GPR_NS_PER_SEC; + if (x < 0) return 0; + if (x > GPR_ATM_MAX) return GPR_ATM_MAX; + return (gpr_atm)x; +} + +static gpr_atm timespec_to_atm_round_down(gpr_timespec ts) { + ts = gpr_time_sub(ts, g_start_time); + double x = + GPR_MS_PER_SEC * (double)ts.tv_sec + (double)ts.tv_nsec / GPR_NS_PER_MS; + if (x < 0) return 0; + if (x > GPR_ATM_MAX) return GPR_ATM_MAX; + return (gpr_atm)x; +} -static int run_some_expired_timers(grpc_exec_ctx *exec_ctx, gpr_timespec now, - gpr_timespec *next, grpc_error *error); +static gpr_timespec atm_to_timespec(gpr_atm x) { + return gpr_time_add(g_start_time, dbl_to_ts((double)x / 1000.0)); +} -static gpr_timespec compute_min_deadline(shard_type *shard) { +static gpr_atm compute_min_deadline(shard_type *shard) { return grpc_timer_heap_is_empty(&shard->heap) - ? shard->queue_deadline_cap + ? saturating_add(shard->queue_deadline_cap, 1) : grpc_timer_heap_top(&shard->heap)->deadline; } void grpc_timer_list_init(gpr_timespec now) { uint32_t i; - g_initialized = true; - gpr_mu_init(&g_mu); + g_shared_mutables.initialized = true; + gpr_mu_init(&g_shared_mutables.mu); g_clock_type = now.clock_type; + g_start_time = now; + g_shared_mutables.min_timer = timespec_to_atm_round_down(now); + gpr_tls_init(&g_last_seen_min_timer); + gpr_tls_set(&g_last_seen_min_timer, 0); + grpc_register_tracer("timer", &grpc_timer_trace); + grpc_register_tracer("timer_check", &grpc_timer_check_trace); for (i = 0; i < NUM_SHARDS; i++) { shard_type *shard = &g_shards[i]; gpr_mu_init(&shard->mu); grpc_time_averaged_stats_init(&shard->stats, 1.0 / ADD_DEADLINE_SCALE, 0.1, 0.5); - shard->queue_deadline_cap = now; + shard->queue_deadline_cap = g_shared_mutables.min_timer; shard->shard_queue_index = i; grpc_timer_heap_init(&shard->heap); shard->list.next = shard->list.prev = &shard->list; @@ -110,29 +171,23 @@ void grpc_timer_list_init(gpr_timespec now) { void grpc_timer_list_shutdown(grpc_exec_ctx *exec_ctx) { int i; run_some_expired_timers( - exec_ctx, gpr_inf_future(g_clock_type), NULL, + exec_ctx, GPR_ATM_MAX, NULL, GRPC_ERROR_CREATE_FROM_STATIC_STRING("Timer list shutdown")); for (i = 0; i < NUM_SHARDS; i++) { shard_type *shard = &g_shards[i]; gpr_mu_destroy(&shard->mu); grpc_timer_heap_destroy(&shard->heap); } - gpr_mu_destroy(&g_mu); - g_initialized = false; + gpr_mu_destroy(&g_shared_mutables.mu); + gpr_tls_destroy(&g_last_seen_min_timer); + g_shared_mutables.initialized = false; } static double ts_to_dbl(gpr_timespec ts) { return (double)ts.tv_sec + 1e-9 * ts.tv_nsec; } -static gpr_timespec dbl_to_ts(double d) { - gpr_timespec ts; - ts.tv_sec = (int64_t)d; - ts.tv_nsec = (int32_t)(1e9 * (d - (double)ts.tv_sec)); - ts.clock_type = GPR_TIMESPAN; - return ts; -} - +/* returns true if the first element in the list */ static void list_join(grpc_timer *head, grpc_timer *timer) { timer->next = head; timer->prev = head->prev; @@ -158,15 +213,13 @@ static void swap_adjacent_shards_in_queue(uint32_t first_shard_queue_index) { static void note_deadline_change(shard_type *shard) { while (shard->shard_queue_index > 0 && - gpr_time_cmp( - shard->min_deadline, - g_shard_queue[shard->shard_queue_index - 1]->min_deadline) < 0) { + shard->min_deadline < + g_shard_queue[shard->shard_queue_index - 1]->min_deadline) { swap_adjacent_shards_in_queue(shard->shard_queue_index - 1); } while (shard->shard_queue_index < NUM_SHARDS - 1 && - gpr_time_cmp( - shard->min_deadline, - g_shard_queue[shard->shard_queue_index + 1]->min_deadline) > 0) { + shard->min_deadline > + g_shard_queue[shard->shard_queue_index + 1]->min_deadline) { swap_adjacent_shards_in_queue(shard->shard_queue_index); } } @@ -179,9 +232,17 @@ void grpc_timer_init(grpc_exec_ctx *exec_ctx, grpc_timer *timer, GPR_ASSERT(deadline.clock_type == g_clock_type); GPR_ASSERT(now.clock_type == g_clock_type); timer->closure = closure; - timer->deadline = deadline; + timer->deadline = timespec_to_atm_round_up(deadline); + + if (grpc_timer_trace) { + gpr_log(GPR_DEBUG, "TIMER %p: SET %" PRId64 ".%09d [%" PRIdPTR + "] now %" PRId64 ".%09d [%" PRIdPTR "] call %p[%p]", + timer, deadline.tv_sec, deadline.tv_nsec, timer->deadline, + now.tv_sec, now.tv_nsec, timespec_to_atm_round_down(now), closure, + closure->cb); + } - if (!g_initialized) { + if (!g_shared_mutables.initialized) { timer->pending = false; grpc_closure_sched(exec_ctx, timer->closure, GRPC_ERROR_CREATE_FROM_STATIC_STRING( @@ -201,12 +262,18 @@ void grpc_timer_init(grpc_exec_ctx *exec_ctx, grpc_timer *timer, grpc_time_averaged_stats_add_sample(&shard->stats, ts_to_dbl(gpr_time_sub(deadline, now))); - if (gpr_time_cmp(deadline, shard->queue_deadline_cap) < 0) { + if (timer->deadline < shard->queue_deadline_cap) { is_first_timer = grpc_timer_heap_add(&shard->heap, timer); } else { timer->heap_index = INVALID_HEAP_INDEX; list_join(&shard->list, timer); } + if (grpc_timer_trace) { + gpr_log(GPR_DEBUG, " .. add to shard %d with queue_deadline_cap=%" PRIdPTR + " => is_first_timer=%s", + (int)(shard - g_shards), shard->queue_deadline_cap, + is_first_timer ? "true" : "false"); + } gpr_mu_unlock(&shard->mu); /* Deadline may have decreased, we need to adjust the master queue. Note @@ -221,28 +288,41 @@ void grpc_timer_init(grpc_exec_ctx *exec_ctx, grpc_timer *timer, In that case, the timer will simply have to wait for the next grpc_timer_check. */ if (is_first_timer) { - gpr_mu_lock(&g_mu); - if (gpr_time_cmp(deadline, shard->min_deadline) < 0) { - gpr_timespec old_min_deadline = g_shard_queue[0]->min_deadline; - shard->min_deadline = deadline; + gpr_mu_lock(&g_shared_mutables.mu); + if (grpc_timer_trace) { + gpr_log(GPR_DEBUG, " .. old shard min_deadline=%" PRIdPTR, + shard->min_deadline); + } + if (timer->deadline < shard->min_deadline) { + gpr_atm old_min_deadline = g_shard_queue[0]->min_deadline; + shard->min_deadline = timer->deadline; note_deadline_change(shard); - if (shard->shard_queue_index == 0 && - gpr_time_cmp(deadline, old_min_deadline) < 0) { + if (shard->shard_queue_index == 0 && timer->deadline < old_min_deadline) { + gpr_atm_no_barrier_store(&g_shared_mutables.min_timer, timer->deadline); grpc_kick_poller(); } } - gpr_mu_unlock(&g_mu); + gpr_mu_unlock(&g_shared_mutables.mu); } } +void grpc_timer_consume_kick(void) { + /* force re-evaluation of last seeen min */ + gpr_tls_set(&g_last_seen_min_timer, 0); +} + void grpc_timer_cancel(grpc_exec_ctx *exec_ctx, grpc_timer *timer) { - if (!g_initialized) { + if (!g_shared_mutables.initialized) { /* must have already been cancelled, also the shard mutex is invalid */ return; } shard_type *shard = &g_shards[GPR_HASH_POINTER(timer, NUM_SHARDS)]; gpr_mu_lock(&shard->mu); + if (grpc_timer_trace) { + gpr_log(GPR_DEBUG, "TIMER %p: CANCEL pending=%s", timer, + timer->pending ? "true" : "false"); + } if (timer->pending) { grpc_closure_sched(exec_ctx, timer->closure, GRPC_ERROR_CANCELLED); timer->pending = false; @@ -260,7 +340,7 @@ void grpc_timer_cancel(grpc_exec_ctx *exec_ctx, grpc_timer *timer) { for timers that fall at or under it. Returns true if the queue is no longer empty. REQUIRES: shard->mu locked */ -static int refill_queue(shard_type *shard, gpr_timespec now) { +static int refill_queue(shard_type *shard, gpr_atm now) { /* Compute the new queue window width and bound by the limits: */ double computed_deadline_delta = grpc_time_averaged_stats_update_average(&shard->stats) * @@ -271,12 +351,22 @@ static int refill_queue(shard_type *shard, gpr_timespec now) { grpc_timer *timer, *next; /* Compute the new cap and put all timers under it into the queue: */ - shard->queue_deadline_cap = gpr_time_add( - gpr_time_max(now, shard->queue_deadline_cap), dbl_to_ts(deadline_delta)); + shard->queue_deadline_cap = + saturating_add(GPR_MAX(now, shard->queue_deadline_cap), + (gpr_atm)(deadline_delta * 1000.0)); + + if (grpc_timer_check_trace) { + gpr_log(GPR_DEBUG, " .. shard[%d]->queue_deadline_cap --> %" PRIdPTR, + (int)(shard - g_shards), shard->queue_deadline_cap); + } for (timer = shard->list.next; timer != &shard->list; timer = next) { next = timer->next; - if (gpr_time_cmp(timer->deadline, shard->queue_deadline_cap) < 0) { + if (timer->deadline < shard->queue_deadline_cap) { + if (grpc_timer_check_trace) { + gpr_log(GPR_DEBUG, " .. add timer with deadline %" PRIdPTR " to heap", + timer->deadline); + } list_remove(timer); grpc_timer_heap_add(&shard->heap, timer); } @@ -287,15 +377,29 @@ static int refill_queue(shard_type *shard, gpr_timespec now) { /* This pops the next non-cancelled timer with deadline <= now from the queue, or returns NULL if there isn't one. REQUIRES: shard->mu locked */ -static grpc_timer *pop_one(shard_type *shard, gpr_timespec now) { +static grpc_timer *pop_one(shard_type *shard, gpr_atm now) { grpc_timer *timer; for (;;) { + if (grpc_timer_check_trace) { + gpr_log(GPR_DEBUG, " .. shard[%d]: heap_empty=%s", + (int)(shard - g_shards), + grpc_timer_heap_is_empty(&shard->heap) ? "true" : "false"); + } if (grpc_timer_heap_is_empty(&shard->heap)) { - if (gpr_time_cmp(now, shard->queue_deadline_cap) < 0) return NULL; + if (now < shard->queue_deadline_cap) return NULL; if (!refill_queue(shard, now)) return NULL; } timer = grpc_timer_heap_top(&shard->heap); - if (gpr_time_cmp(timer->deadline, now) > 0) return NULL; + if (grpc_timer_check_trace) { + gpr_log(GPR_DEBUG, + " .. check top timer deadline=%" PRIdPTR " now=%" PRIdPTR, + timer->deadline, now); + } + if (timer->deadline > now) return NULL; + if (grpc_timer_trace) { + gpr_log(GPR_DEBUG, "TIMER %p: FIRE %" PRIdPTR "ms late", timer, + now - timer->deadline); + } timer->pending = false; grpc_timer_heap_pop(&shard->heap); return timer; @@ -304,7 +408,7 @@ static grpc_timer *pop_one(shard_type *shard, gpr_timespec now) { /* REQUIRES: shard->mu unlocked */ static size_t pop_timers(grpc_exec_ctx *exec_ctx, shard_type *shard, - gpr_timespec now, gpr_timespec *new_min_deadline, + gpr_atm now, gpr_atm *new_min_deadline, grpc_error *error) { size_t n = 0; grpc_timer *timer; @@ -318,17 +422,29 @@ static size_t pop_timers(grpc_exec_ctx *exec_ctx, shard_type *shard, return n; } -static int run_some_expired_timers(grpc_exec_ctx *exec_ctx, gpr_timespec now, - gpr_timespec *next, grpc_error *error) { +static int run_some_expired_timers(grpc_exec_ctx *exec_ctx, gpr_atm now, + gpr_atm *next, grpc_error *error) { size_t n = 0; - /* TODO(ctiller): verify that there are any timers (atomically) here */ + gpr_atm min_timer = gpr_atm_no_barrier_load(&g_shared_mutables.min_timer); + gpr_tls_set(&g_last_seen_min_timer, min_timer); + if (now < min_timer) { + if (next != NULL) *next = GPR_MIN(*next, min_timer); + return 0; + } - if (gpr_spinlock_trylock(&g_checker_mu)) { - gpr_mu_lock(&g_mu); + if (gpr_spinlock_trylock(&g_shared_mutables.checker_mu)) { + gpr_mu_lock(&g_shared_mutables.mu); - while (gpr_time_cmp(g_shard_queue[0]->min_deadline, now) < 0) { - gpr_timespec new_min_deadline; + if (grpc_timer_check_trace) { + gpr_log(GPR_DEBUG, " .. shard[%d]->min_deadline = %" PRIdPTR, + (int)(g_shard_queue[0] - g_shards), + g_shard_queue[0]->min_deadline); + } + + while (g_shard_queue[0]->min_deadline < now || + (now != GPR_ATM_MAX && g_shard_queue[0]->min_deadline == now)) { + gpr_atm new_min_deadline; /* For efficiency, we pop as many available timers as we can from the shard. This may violate perfect timer deadline ordering, but that @@ -336,6 +452,14 @@ static int run_some_expired_timers(grpc_exec_ctx *exec_ctx, gpr_timespec now, n += pop_timers(exec_ctx, g_shard_queue[0], now, &new_min_deadline, error); + if (grpc_timer_check_trace) { + gpr_log(GPR_DEBUG, " .. popped --> %" PRIdPTR + ", shard[%d]->min_deadline %" PRIdPTR + " --> %" PRIdPTR ", now=%" PRIdPTR, + n, (int)(g_shard_queue[0] - g_shards), + g_shard_queue[0]->min_deadline, new_min_deadline, now); + } + /* An grpc_timer_init() on the shard could intervene here, adding a new timer that is earlier than new_min_deadline. However, grpc_timer_init() will block on the master_lock before it can call @@ -346,23 +470,24 @@ static int run_some_expired_timers(grpc_exec_ctx *exec_ctx, gpr_timespec now, } if (next) { - *next = gpr_time_min(*next, g_shard_queue[0]->min_deadline); + *next = GPR_MIN(*next, g_shard_queue[0]->min_deadline); } - gpr_mu_unlock(&g_mu); - gpr_spinlock_unlock(&g_checker_mu); + gpr_atm_no_barrier_store(&g_shared_mutables.min_timer, + g_shard_queue[0]->min_deadline); + gpr_mu_unlock(&g_shared_mutables.mu); + gpr_spinlock_unlock(&g_shared_mutables.checker_mu); } else if (next != NULL) { /* TODO(ctiller): this forces calling code to do an short poll, and then retry the timer check (because this time through the timer list was contended). - We could reduce the cost here dramatically by keeping a count of how many - currently active pollers got through the uncontended case above + We could reduce the cost here dramatically by keeping a count of how + many currently active pollers got through the uncontended case above successfully, and waking up other pollers IFF that count drops to zero. Once that count is in place, this entire else branch could disappear. */ - *next = gpr_time_min( - *next, gpr_time_add(now, gpr_time_from_millis(1, GPR_TIMESPAN))); + *next = GPR_MIN(*next, now + 1); } GRPC_ERROR_UNREF(error); @@ -372,12 +497,71 @@ static int run_some_expired_timers(grpc_exec_ctx *exec_ctx, gpr_timespec now, bool grpc_timer_check(grpc_exec_ctx *exec_ctx, gpr_timespec now, gpr_timespec *next) { + // prelude GPR_ASSERT(now.clock_type == g_clock_type); - return run_some_expired_timers( - exec_ctx, now, next, + gpr_atm now_atm = timespec_to_atm_round_down(now); + + /* fetch from a thread-local first: this avoids contention on a globally + mutable cacheline in the common case */ + gpr_atm min_timer = gpr_tls_get(&g_last_seen_min_timer); + if (now_atm < min_timer) { + if (next != NULL) { + *next = + atm_to_timespec(GPR_MIN(timespec_to_atm_round_up(*next), min_timer)); + } + if (grpc_timer_check_trace) { + gpr_log(GPR_DEBUG, + "TIMER CHECK SKIP: now_atm=%" PRIdPTR " min_timer=%" PRIdPTR, + now_atm, min_timer); + } + return 0; + } + + grpc_error *shutdown_error = gpr_time_cmp(now, gpr_inf_future(now.clock_type)) != 0 ? GRPC_ERROR_NONE - : GRPC_ERROR_CREATE_FROM_STATIC_STRING("Shutting down timer system")); + : GRPC_ERROR_CREATE_FROM_STATIC_STRING("Shutting down timer system"); + + // tracing + if (grpc_timer_check_trace) { + char *next_str; + if (next == NULL) { + next_str = gpr_strdup("NULL"); + } else { + gpr_asprintf(&next_str, "%" PRId64 ".%09d [%" PRIdPTR "]", next->tv_sec, + next->tv_nsec, timespec_to_atm_round_down(*next)); + } + gpr_log(GPR_DEBUG, "TIMER CHECK BEGIN: now=%" PRId64 ".%09d [%" PRIdPTR + "] next=%s tls_min=%" PRIdPTR " glob_min=%" PRIdPTR, + now.tv_sec, now.tv_nsec, now_atm, next_str, + gpr_tls_get(&g_last_seen_min_timer), + gpr_atm_no_barrier_load(&g_shared_mutables.min_timer)); + gpr_free(next_str); + } + // actual code + bool r; + gpr_atm next_atm; + if (next == NULL) { + r = run_some_expired_timers(exec_ctx, now_atm, NULL, shutdown_error); + } else { + next_atm = timespec_to_atm_round_down(*next); + r = run_some_expired_timers(exec_ctx, now_atm, &next_atm, shutdown_error); + *next = atm_to_timespec(next_atm); + } + // tracing + if (grpc_timer_check_trace) { + char *next_str; + if (next == NULL) { + next_str = gpr_strdup("NULL"); + } else { + gpr_asprintf(&next_str, "%" PRId64 ".%09d [%" PRIdPTR "]", next->tv_sec, + next->tv_nsec, next_atm); + } + gpr_log(GPR_DEBUG, "TIMER CHECK END: %d timers triggered; next=%s", r, + next_str); + gpr_free(next_str); + } + return r > 0; } #endif /* GRPC_TIMER_USE_GENERIC */ diff --git a/src/core/lib/iomgr/timer_generic.h b/src/core/lib/iomgr/timer_generic.h index 1608dce9fb..c79a431aa0 100644 --- a/src/core/lib/iomgr/timer_generic.h +++ b/src/core/lib/iomgr/timer_generic.h @@ -38,7 +38,7 @@ #include "src/core/lib/iomgr/exec_ctx.h" struct grpc_timer { - gpr_timespec deadline; + gpr_atm deadline; uint32_t heap_index; /* INVALID_HEAP_INDEX if not in heap */ bool pending; struct grpc_timer *next; diff --git a/src/core/lib/iomgr/timer_heap.c b/src/core/lib/iomgr/timer_heap.c index f736d335e6..03ccfe023a 100644 --- a/src/core/lib/iomgr/timer_heap.c +++ b/src/core/lib/iomgr/timer_heap.c @@ -50,7 +50,7 @@ static void adjust_upwards(grpc_timer **first, uint32_t i, grpc_timer *t) { while (i > 0) { uint32_t parent = (uint32_t)(((int)i - 1) / 2); - if (gpr_time_cmp(first[parent]->deadline, t->deadline) <= 0) break; + if (first[parent]->deadline <= t->deadline) break; first[i] = first[parent]; first[i]->heap_index = i; i = parent; @@ -68,12 +68,12 @@ static void adjust_downwards(grpc_timer **first, uint32_t i, uint32_t length, uint32_t left_child = 1u + 2u * i; if (left_child >= length) break; uint32_t right_child = left_child + 1; - uint32_t next_i = right_child < length && - gpr_time_cmp(first[left_child]->deadline, - first[right_child]->deadline) > 0 - ? right_child - : left_child; - if (gpr_time_cmp(t->deadline, first[next_i]->deadline) <= 0) break; + uint32_t next_i = + right_child < length && + first[left_child]->deadline > first[right_child]->deadline + ? right_child + : left_child; + if (t->deadline <= first[next_i]->deadline) break; first[i] = first[next_i]; first[i]->heap_index = i; i = next_i; @@ -97,7 +97,7 @@ static void maybe_shrink(grpc_timer_heap *heap) { static void note_changed_priority(grpc_timer_heap *heap, grpc_timer *timer) { uint32_t i = timer->heap_index; uint32_t parent = (uint32_t)(((int)i - 1) / 2); - if (gpr_time_cmp(heap->timers[parent]->deadline, timer->deadline) > 0) { + if (heap->timers[parent]->deadline > timer->deadline) { adjust_upwards(heap->timers, i, timer); } else { adjust_downwards(heap->timers, i, heap->timer_count, timer); |