diff options
Diffstat (limited to 'src/core/lib/iomgr')
-rw-r--r-- | src/core/lib/iomgr/endpoint.c | 6 | ||||
-rw-r--r-- | src/core/lib/iomgr/endpoint.h | 17 | ||||
-rw-r--r-- | src/core/lib/iomgr/ev_epoll_linux.c | 2 | ||||
-rw-r--r-- | src/core/lib/iomgr/ev_poll_and_epoll_posix.c | 2076 | ||||
-rw-r--r-- | src/core/lib/iomgr/ev_poll_and_epoll_posix.h | 41 | ||||
-rw-r--r-- | src/core/lib/iomgr/ev_posix.c | 2 | ||||
-rw-r--r-- | src/core/lib/iomgr/load_file.c | 6 | ||||
-rw-r--r-- | src/core/lib/iomgr/load_file.h | 4 | ||||
-rw-r--r-- | src/core/lib/iomgr/resource_quota.c | 217 | ||||
-rw-r--r-- | src/core/lib/iomgr/resource_quota.h | 98 | ||||
-rw-r--r-- | src/core/lib/iomgr/socket_mutator.c | 98 | ||||
-rw-r--r-- | src/core/lib/iomgr/socket_mutator.h | 80 | ||||
-rw-r--r-- | src/core/lib/iomgr/socket_utils_common_posix.c | 9 | ||||
-rw-r--r-- | src/core/lib/iomgr/socket_utils_posix.h | 5 | ||||
-rw-r--r-- | src/core/lib/iomgr/tcp_client.h | 1 | ||||
-rw-r--r-- | src/core/lib/iomgr/tcp_client_posix.c | 23 | ||||
-rw-r--r-- | src/core/lib/iomgr/tcp_client_windows.c | 2 | ||||
-rw-r--r-- | src/core/lib/iomgr/tcp_posix.c | 100 | ||||
-rw-r--r-- | src/core/lib/iomgr/tcp_uv.c | 107 | ||||
-rw-r--r-- | src/core/lib/iomgr/tcp_windows.c | 92 | ||||
-rw-r--r-- | src/core/lib/iomgr/wakeup_fd_pipe.c | 2 |
21 files changed, 523 insertions, 2465 deletions
diff --git a/src/core/lib/iomgr/endpoint.c b/src/core/lib/iomgr/endpoint.c index 74fa9c45df..2d300f4560 100644 --- a/src/core/lib/iomgr/endpoint.c +++ b/src/core/lib/iomgr/endpoint.c @@ -34,12 +34,12 @@ #include "src/core/lib/iomgr/endpoint.h" void grpc_endpoint_read(grpc_exec_ctx* exec_ctx, grpc_endpoint* ep, - gpr_slice_buffer* slices, grpc_closure* cb) { + grpc_slice_buffer* slices, grpc_closure* cb) { ep->vtable->read(exec_ctx, ep, slices, cb); } void grpc_endpoint_write(grpc_exec_ctx* exec_ctx, grpc_endpoint* ep, - gpr_slice_buffer* slices, grpc_closure* cb) { + grpc_slice_buffer* slices, grpc_closure* cb) { ep->vtable->write(exec_ctx, ep, slices, cb); } @@ -66,6 +66,8 @@ char* grpc_endpoint_get_peer(grpc_endpoint* ep) { return ep->vtable->get_peer(ep); } +int grpc_endpoint_get_fd(grpc_endpoint* ep) { return ep->vtable->get_fd(ep); } + grpc_workqueue* grpc_endpoint_get_workqueue(grpc_endpoint* ep) { return ep->vtable->get_workqueue(ep); } diff --git a/src/core/lib/iomgr/endpoint.h b/src/core/lib/iomgr/endpoint.h index 0ac5486ff5..1609b64f2b 100644 --- a/src/core/lib/iomgr/endpoint.h +++ b/src/core/lib/iomgr/endpoint.h @@ -34,8 +34,8 @@ #ifndef GRPC_CORE_LIB_IOMGR_ENDPOINT_H #define GRPC_CORE_LIB_IOMGR_ENDPOINT_H -#include <grpc/support/slice.h> -#include <grpc/support/slice_buffer.h> +#include <grpc/slice.h> +#include <grpc/slice_buffer.h> #include <grpc/support/time.h> #include "src/core/lib/iomgr/pollset.h" #include "src/core/lib/iomgr/pollset_set.h" @@ -49,9 +49,9 @@ typedef struct grpc_endpoint_vtable grpc_endpoint_vtable; struct grpc_endpoint_vtable { void (*read)(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, - gpr_slice_buffer *slices, grpc_closure *cb); + grpc_slice_buffer *slices, grpc_closure *cb); void (*write)(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, - gpr_slice_buffer *slices, grpc_closure *cb); + grpc_slice_buffer *slices, grpc_closure *cb); grpc_workqueue *(*get_workqueue)(grpc_endpoint *ep); void (*add_to_pollset)(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, grpc_pollset *pollset); @@ -61,6 +61,7 @@ struct grpc_endpoint_vtable { void (*destroy)(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep); grpc_resource_user *(*get_resource_user)(grpc_endpoint *ep); char *(*get_peer)(grpc_endpoint *ep); + int (*get_fd)(grpc_endpoint *ep); }; /* When data is available on the connection, calls the callback with slices. @@ -69,10 +70,14 @@ struct grpc_endpoint_vtable { Valid slices may be placed into \a slices even when the callback is invoked with error != GRPC_ERROR_NONE. */ void grpc_endpoint_read(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, - gpr_slice_buffer *slices, grpc_closure *cb); + grpc_slice_buffer *slices, grpc_closure *cb); char *grpc_endpoint_get_peer(grpc_endpoint *ep); +/* Get the file descriptor used by \a ep. Return -1 if \a ep is not using an fd. + */ +int grpc_endpoint_get_fd(grpc_endpoint *ep); + /* Retrieve a reference to the workqueue associated with this endpoint */ grpc_workqueue *grpc_endpoint_get_workqueue(grpc_endpoint *ep); @@ -87,7 +92,7 @@ grpc_workqueue *grpc_endpoint_get_workqueue(grpc_endpoint *ep); it is a valid slice buffer. */ void grpc_endpoint_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, - gpr_slice_buffer *slices, grpc_closure *cb); + grpc_slice_buffer *slices, grpc_closure *cb); /* Causes any pending and future read/write callbacks to run immediately with success==0 */ diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c index db51ec4939..91041a7c28 100644 --- a/src/core/lib/iomgr/ev_epoll_linux.c +++ b/src/core/lib/iomgr/ev_epoll_linux.c @@ -163,7 +163,7 @@ static void fd_global_shutdown(void); #define PI_ADD_REF(p, r) pi_add_ref((p)) #define PI_UNREF(exec_ctx, p, r) pi_unref((exec_ctx), (p)) -#endif /* !defined(GPRC_PI_REF_COUNT_DEBUG) */ +#endif /* !defined(GRPC_PI_REF_COUNT_DEBUG) */ /* This is also used as grpc_workqueue (by directly casing it) */ typedef struct polling_island { diff --git a/src/core/lib/iomgr/ev_poll_and_epoll_posix.c b/src/core/lib/iomgr/ev_poll_and_epoll_posix.c deleted file mode 100644 index bf51404203..0000000000 --- a/src/core/lib/iomgr/ev_poll_and_epoll_posix.c +++ /dev/null @@ -1,2076 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -/* This file will be removed shortly: it's here to keep refactoring - * steps simple and auditable. - * It's the combination of the old files: - * - fd_posix.{h,c} - * - pollset_posix.{h,c} - * - pullset_multipoller_with_{poll,epoll}.{h,c} - * The new version will be split into: - * - ev_poll_posix.{h,c} - * - ev_epoll_posix.{h,c} - */ - -#include "src/core/lib/iomgr/port.h" - -#ifdef GRPC_POSIX_SOCKET - -#include "src/core/lib/iomgr/ev_poll_and_epoll_posix.h" - -#include <assert.h> -#include <errno.h> -#include <poll.h> -#include <string.h> -#include <sys/socket.h> -#include <unistd.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/string_util.h> -#include <grpc/support/tls.h> -#include <grpc/support/useful.h> - -#include "src/core/lib/iomgr/iomgr_internal.h" -#include "src/core/lib/iomgr/wakeup_fd_posix.h" -#include "src/core/lib/profiling/timers.h" -#include "src/core/lib/support/block_annotate.h" - -/******************************************************************************* - * FD declarations - */ - -typedef struct grpc_fd_watcher { - struct grpc_fd_watcher *next; - struct grpc_fd_watcher *prev; - grpc_pollset *pollset; - grpc_pollset_worker *worker; - grpc_fd *fd; -} grpc_fd_watcher; - -struct grpc_fd { - int fd; - /* refst format: - bit0: 1=active/0=orphaned - bit1-n: refcount - meaning that mostly we ref by two to avoid altering the orphaned bit, - and just unref by 1 when we're ready to flag the object as orphaned */ - gpr_atm refst; - - gpr_mu mu; - int shutdown; - int closed; - int released; - - /* The watcher list. - - The following watcher related fields are protected by watcher_mu. - - An fd_watcher is an ephemeral object created when an fd wants to - begin polling, and destroyed after the poll. - - It denotes the fd's interest in whether to read poll or write poll - or both or neither on this fd. - - If a watcher is asked to poll for reads or writes, the read_watcher - or write_watcher fields are set respectively. A watcher may be asked - to poll for both, in which case both fields will be set. - - read_watcher and write_watcher may be NULL if no watcher has been - asked to poll for reads or writes. - - If an fd_watcher is not asked to poll for reads or writes, it's added - to a linked list of inactive watchers, rooted at inactive_watcher_root. - If at a later time there becomes need of a poller to poll, one of - the inactive pollers may be kicked out of their poll loops to take - that responsibility. */ - grpc_fd_watcher inactive_watcher_root; - grpc_fd_watcher *read_watcher; - grpc_fd_watcher *write_watcher; - - grpc_closure *read_closure; - grpc_closure *write_closure; - - struct grpc_fd *freelist_next; - - grpc_closure *on_done_closure; - - grpc_iomgr_object iomgr_object; - - /* The pollset that last noticed and notified that the fd is readable */ - grpc_pollset *read_notifier_pollset; -}; - -/* Begin polling on an fd. - Registers that the given pollset is interested in this fd - so that if read - or writability interest changes, the pollset can be kicked to pick up that - new interest. - Return value is: - (fd_needs_read? read_mask : 0) | (fd_needs_write? write_mask : 0) - i.e. a combination of read_mask and write_mask determined by the fd's current - interest in said events. - Polling strategies that do not need to alter their behavior depending on the - fd's current interest (such as epoll) do not need to call this function. - MUST NOT be called with a pollset lock taken */ -static uint32_t fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset, - grpc_pollset_worker *worker, uint32_t read_mask, - uint32_t write_mask, grpc_fd_watcher *rec); -/* Complete polling previously started with fd_begin_poll - MUST NOT be called with a pollset lock taken - if got_read or got_write are 1, also does the become_{readable,writable} as - appropriate. */ -static void fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *rec, - int got_read, int got_write, - grpc_pollset *read_notifier_pollset); - -/* Return 1 if this fd is orphaned, 0 otherwise */ -static bool fd_is_orphaned(grpc_fd *fd); - -/* Reference counting for fds */ -/*#define GRPC_FD_REF_COUNT_DEBUG*/ -#ifdef GRPC_FD_REF_COUNT_DEBUG -static void fd_ref(grpc_fd *fd, const char *reason, const char *file, int line); -static void fd_unref(grpc_fd *fd, const char *reason, const char *file, - int line); -#define GRPC_FD_REF(fd, reason) fd_ref(fd, reason, __FILE__, __LINE__) -#define GRPC_FD_UNREF(fd, reason) fd_unref(fd, reason, __FILE__, __LINE__) -#else -static void fd_ref(grpc_fd *fd); -static void fd_unref(grpc_fd *fd); -#define GRPC_FD_REF(fd, reason) fd_ref(fd) -#define GRPC_FD_UNREF(fd, reason) fd_unref(fd) -#endif - -static void fd_global_init(void); -static void fd_global_shutdown(void); - -#define CLOSURE_NOT_READY ((grpc_closure *)0) -#define CLOSURE_READY ((grpc_closure *)1) - -/******************************************************************************* - * pollset declarations - */ - -typedef struct grpc_pollset_vtable grpc_pollset_vtable; - -typedef struct grpc_cached_wakeup_fd { - grpc_wakeup_fd fd; - struct grpc_cached_wakeup_fd *next; -} grpc_cached_wakeup_fd; - -struct grpc_pollset_worker { - grpc_cached_wakeup_fd *wakeup_fd; - int reevaluate_polling_on_wakeup; - int kicked_specifically; - struct grpc_pollset_worker *next; - struct grpc_pollset_worker *prev; -}; - -struct grpc_pollset { - /* pollsets under posix can mutate representation as fds are added and - removed. - For example, we may choose a poll() based implementation on linux for - few fds, and an epoll() based implementation for many fds */ - const grpc_pollset_vtable *vtable; - gpr_mu mu; - grpc_pollset_worker root_worker; - int in_flight_cbs; - int shutting_down; - int called_shutdown; - int kicked_without_pollers; - grpc_closure *shutdown_done; - grpc_closure_list idle_jobs; - union { - int fd; - void *ptr; - } data; - /* Local cache of eventfds for workers */ - grpc_cached_wakeup_fd *local_wakeup_cache; -}; - -struct grpc_pollset_vtable { - void (*add_fd)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - struct grpc_fd *fd, int and_unlock_pollset); - grpc_error *(*maybe_work_and_unlock)(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, - grpc_pollset_worker *worker, - gpr_timespec deadline, gpr_timespec now); - void (*finish_shutdown)(grpc_pollset *pollset); - void (*destroy)(grpc_pollset *pollset); -}; - -/* Add an fd to a pollset */ -static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - struct grpc_fd *fd); - -static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, grpc_fd *fd); - -/* Convert a timespec to milliseconds: - - very small or negative poll times are clamped to zero to do a - non-blocking poll (which becomes spin polling) - - other small values are rounded up to one millisecond - - longer than a millisecond polls are rounded up to the next nearest - millisecond to avoid spinning - - infinite timeouts are converted to -1 */ -static int poll_deadline_to_millis_timeout(gpr_timespec deadline, - gpr_timespec now); - -/* Allow kick to wakeup the currently polling worker */ -#define GRPC_POLLSET_CAN_KICK_SELF 1 -/* Force the wakee to repoll when awoken */ -#define GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP 2 -/* As per pollset_kick, with an extended set of flags (defined above) - -- mostly for fd_posix's use. */ -static grpc_error *pollset_kick_ext(grpc_pollset *p, - grpc_pollset_worker *specific_worker, - uint32_t flags) GRPC_MUST_USE_RESULT; - -/* turn a pollset into a multipoller: platform specific */ -typedef void (*platform_become_multipoller_type)(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, - struct grpc_fd **fds, - size_t fd_count); -static platform_become_multipoller_type platform_become_multipoller; - -/* Return 1 if the pollset has active threads in pollset_work (pollset must - * be locked) */ -static int pollset_has_workers(grpc_pollset *pollset); - -static void remove_fd_from_all_epoll_sets(int fd); - -/******************************************************************************* - * pollset_set definitions - */ - -struct grpc_pollset_set { - gpr_mu mu; - - size_t pollset_count; - size_t pollset_capacity; - grpc_pollset **pollsets; - - size_t pollset_set_count; - size_t pollset_set_capacity; - struct grpc_pollset_set **pollset_sets; - - size_t fd_count; - size_t fd_capacity; - grpc_fd **fds; -}; - -/******************************************************************************* - * fd_posix.c - */ - -/* We need to keep a freelist not because of any concerns of malloc performance - * but instead so that implementations with multiple threads in (for example) - * epoll_wait deal with the race between pollset removal and incoming poll - * notifications. - * - * The problem is that the poller ultimately holds a reference to this - * object, so it is very difficult to know when is safe to free it, at least - * without some expensive synchronization. - * - * If we keep the object freelisted, in the worst case losing this race just - * becomes a spurious read notification on a reused fd. - */ -/* TODO(klempner): We could use some form of polling generation count to know - * when these are safe to free. */ -/* TODO(klempner): Consider disabling freelisting if we don't have multiple - * threads in poll on the same fd */ -/* TODO(klempner): Batch these allocations to reduce fragmentation */ -static grpc_fd *fd_freelist = NULL; -static gpr_mu fd_freelist_mu; - -static void freelist_fd(grpc_fd *fd) { - gpr_mu_lock(&fd_freelist_mu); - fd->freelist_next = fd_freelist; - fd_freelist = fd; - grpc_iomgr_unregister_object(&fd->iomgr_object); - gpr_mu_unlock(&fd_freelist_mu); -} - -static grpc_fd *alloc_fd(int fd) { - grpc_fd *r = NULL; - gpr_mu_lock(&fd_freelist_mu); - if (fd_freelist != NULL) { - r = fd_freelist; - fd_freelist = fd_freelist->freelist_next; - } - gpr_mu_unlock(&fd_freelist_mu); - if (r == NULL) { - r = gpr_malloc(sizeof(grpc_fd)); - gpr_mu_init(&r->mu); - } - - gpr_mu_lock(&r->mu); - gpr_atm_rel_store(&r->refst, 1); - r->shutdown = 0; - r->read_closure = CLOSURE_NOT_READY; - r->write_closure = CLOSURE_NOT_READY; - r->fd = fd; - r->inactive_watcher_root.next = r->inactive_watcher_root.prev = - &r->inactive_watcher_root; - r->freelist_next = NULL; - r->read_watcher = r->write_watcher = NULL; - r->on_done_closure = NULL; - r->closed = 0; - r->released = 0; - r->read_notifier_pollset = NULL; - gpr_mu_unlock(&r->mu); - return r; -} - -static void destroy(grpc_fd *fd) { - gpr_mu_destroy(&fd->mu); - gpr_free(fd); -} - -#ifdef GRPC_FD_REF_COUNT_DEBUG -#define REF_BY(fd, n, reason) ref_by(fd, n, reason, __FILE__, __LINE__) -#define UNREF_BY(fd, n, reason) unref_by(fd, n, reason, __FILE__, __LINE__) -static void ref_by(grpc_fd *fd, int n, const char *reason, const char *file, - int line) { - gpr_log(GPR_DEBUG, "FD %d %p ref %d %d -> %d [%s; %s:%d]", fd->fd, fd, n, - gpr_atm_no_barrier_load(&fd->refst), - gpr_atm_no_barrier_load(&fd->refst) + n, reason, file, line); -#else -#define REF_BY(fd, n, reason) ref_by(fd, n) -#define UNREF_BY(fd, n, reason) unref_by(fd, n) -static void ref_by(grpc_fd *fd, int n) { -#endif - GPR_ASSERT(gpr_atm_no_barrier_fetch_add(&fd->refst, n) > 0); -} - -#ifdef GRPC_FD_REF_COUNT_DEBUG -static void unref_by(grpc_fd *fd, int n, const char *reason, const char *file, - int line) { - gpr_atm old; - gpr_log(GPR_DEBUG, "FD %d %p unref %d %d -> %d [%s; %s:%d]", fd->fd, fd, n, - gpr_atm_no_barrier_load(&fd->refst), - gpr_atm_no_barrier_load(&fd->refst) - n, reason, file, line); -#else -static void unref_by(grpc_fd *fd, int n) { - gpr_atm old; -#endif - old = gpr_atm_full_fetch_add(&fd->refst, -n); - if (old == n) { - freelist_fd(fd); - } else { - GPR_ASSERT(old > n); - } -} - -static void fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); } - -static void fd_global_shutdown(void) { - gpr_mu_lock(&fd_freelist_mu); - gpr_mu_unlock(&fd_freelist_mu); - while (fd_freelist != NULL) { - grpc_fd *fd = fd_freelist; - fd_freelist = fd_freelist->freelist_next; - destroy(fd); - } - gpr_mu_destroy(&fd_freelist_mu); -} - -static grpc_fd *fd_create(int fd, const char *name) { - grpc_fd *r = alloc_fd(fd); - char *name2; - gpr_asprintf(&name2, "%s fd=%d", name, fd); - grpc_iomgr_register_object(&r->iomgr_object, name2); - gpr_free(name2); -#ifdef GRPC_FD_REF_COUNT_DEBUG - gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, r, name); -#endif - return r; -} - -static bool fd_is_orphaned(grpc_fd *fd) { - return (gpr_atm_acq_load(&fd->refst) & 1) == 0; -} - -static grpc_error *pollset_kick_locked(grpc_fd_watcher *watcher) { - gpr_mu_lock(&watcher->pollset->mu); - GPR_ASSERT(watcher->worker); - grpc_error *err = pollset_kick_ext(watcher->pollset, watcher->worker, - GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP); - gpr_mu_unlock(&watcher->pollset->mu); - return err; -} - -static void maybe_wake_one_watcher_locked(grpc_fd *fd) { - if (fd->inactive_watcher_root.next != &fd->inactive_watcher_root) { - pollset_kick_locked(fd->inactive_watcher_root.next); - } else if (fd->read_watcher) { - pollset_kick_locked(fd->read_watcher); - } else if (fd->write_watcher) { - pollset_kick_locked(fd->write_watcher); - } -} - -static void wake_all_watchers_locked(grpc_fd *fd) { - grpc_fd_watcher *watcher; - for (watcher = fd->inactive_watcher_root.next; - watcher != &fd->inactive_watcher_root; watcher = watcher->next) { - pollset_kick_locked(watcher); - } - if (fd->read_watcher) { - pollset_kick_locked(fd->read_watcher); - } - if (fd->write_watcher && fd->write_watcher != fd->read_watcher) { - pollset_kick_locked(fd->write_watcher); - } -} - -static int has_watchers(grpc_fd *fd) { - return fd->read_watcher != NULL || fd->write_watcher != NULL || - fd->inactive_watcher_root.next != &fd->inactive_watcher_root; -} - -static void close_fd_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { - fd->closed = 1; - if (!fd->released) { - close(fd->fd); - } else { - remove_fd_from_all_epoll_sets(fd->fd); - } - grpc_exec_ctx_sched(exec_ctx, fd->on_done_closure, GRPC_ERROR_NONE, NULL); -} - -static int fd_wrapped_fd(grpc_fd *fd) { - if (fd->released || fd->closed) { - return -1; - } else { - return fd->fd; - } -} - -static void fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure *on_done, int *release_fd, - const char *reason) { - fd->on_done_closure = on_done; - fd->released = release_fd != NULL; - if (!fd->released) { - shutdown(fd->fd, SHUT_RDWR); - } else { - *release_fd = fd->fd; - } - gpr_mu_lock(&fd->mu); - REF_BY(fd, 1, reason); /* remove active status, but keep referenced */ - if (!has_watchers(fd)) { - close_fd_locked(exec_ctx, fd); - } else { - wake_all_watchers_locked(fd); - } - gpr_mu_unlock(&fd->mu); - UNREF_BY(fd, 2, reason); /* drop the reference */ -} - -/* increment refcount by two to avoid changing the orphan bit */ -#ifdef GRPC_FD_REF_COUNT_DEBUG -static void fd_ref(grpc_fd *fd, const char *reason, const char *file, - int line) { - ref_by(fd, 2, reason, file, line); -} - -static void fd_unref(grpc_fd *fd, const char *reason, const char *file, - int line) { - unref_by(fd, 2, reason, file, line); -} -#else -static void fd_ref(grpc_fd *fd) { ref_by(fd, 2); } - -static void fd_unref(grpc_fd *fd) { unref_by(fd, 2); } -#endif - -static grpc_error *fd_shutdown_error(bool shutdown) { - if (!shutdown) { - return GRPC_ERROR_NONE; - } else { - return GRPC_ERROR_CREATE("FD shutdown"); - } -} - -static void notify_on_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure **st, grpc_closure *closure) { - if (fd->shutdown) { - grpc_exec_ctx_sched(exec_ctx, closure, GRPC_ERROR_CREATE("FD shutdown"), - NULL); - } else if (*st == CLOSURE_NOT_READY) { - /* not ready ==> switch to a waiting state by setting the closure */ - *st = closure; - } else if (*st == CLOSURE_READY) { - /* already ready ==> queue the closure to run immediately */ - *st = CLOSURE_NOT_READY; - grpc_exec_ctx_sched(exec_ctx, closure, fd_shutdown_error(fd->shutdown), - NULL); - maybe_wake_one_watcher_locked(fd); - } else { - /* upcallptr was set to a different closure. This is an error! */ - gpr_log(GPR_ERROR, - "User called a notify_on function with a previous callback still " - "pending"); - abort(); - } -} - -/* returns 1 if state becomes not ready */ -static int set_ready_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure **st) { - if (*st == CLOSURE_READY) { - /* duplicate ready ==> ignore */ - return 0; - } else if (*st == CLOSURE_NOT_READY) { - /* not ready, and not waiting ==> flag ready */ - *st = CLOSURE_READY; - return 0; - } else { - /* waiting ==> queue closure */ - grpc_exec_ctx_sched(exec_ctx, *st, fd_shutdown_error(fd->shutdown), NULL); - *st = CLOSURE_NOT_READY; - return 1; - } -} - -static void set_read_notifier_pollset_locked( - grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_pollset *read_notifier_pollset) { - fd->read_notifier_pollset = read_notifier_pollset; -} - -static void fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { - gpr_mu_lock(&fd->mu); - /* only shutdown once */ - if (!fd->shutdown) { - fd->shutdown = 1; - /* signal read/write closed to OS so that future operations fail */ - shutdown(fd->fd, SHUT_RDWR); - set_ready_locked(exec_ctx, fd, &fd->read_closure); - set_ready_locked(exec_ctx, fd, &fd->write_closure); - } - gpr_mu_unlock(&fd->mu); -} - -static bool fd_is_shutdown(grpc_fd *fd) { - gpr_mu_lock(&fd->mu); - bool r = fd->shutdown; - gpr_mu_unlock(&fd->mu); - return r; -} - -static void fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure *closure) { - gpr_mu_lock(&fd->mu); - notify_on_locked(exec_ctx, fd, &fd->read_closure, closure); - gpr_mu_unlock(&fd->mu); -} - -static void fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure *closure) { - gpr_mu_lock(&fd->mu); - notify_on_locked(exec_ctx, fd, &fd->write_closure, closure); - gpr_mu_unlock(&fd->mu); -} - -/* Return the read-notifier pollset */ -static grpc_pollset *fd_get_read_notifier_pollset(grpc_exec_ctx *exec_ctx, - grpc_fd *fd) { - grpc_pollset *notifier = NULL; - - gpr_mu_lock(&fd->mu); - notifier = fd->read_notifier_pollset; - gpr_mu_unlock(&fd->mu); - - return notifier; -} - -static uint32_t fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset, - grpc_pollset_worker *worker, uint32_t read_mask, - uint32_t write_mask, grpc_fd_watcher *watcher) { - uint32_t mask = 0; - grpc_closure *cur; - int requested; - /* keep track of pollers that have requested our events, in case they change - */ - GRPC_FD_REF(fd, "poll"); - - gpr_mu_lock(&fd->mu); - - /* if we are shutdown, then don't add to the watcher set */ - if (fd->shutdown) { - watcher->fd = NULL; - watcher->pollset = NULL; - watcher->worker = NULL; - gpr_mu_unlock(&fd->mu); - GRPC_FD_UNREF(fd, "poll"); - return 0; - } - - /* if there is nobody polling for read, but we need to, then start doing so */ - cur = fd->read_closure; - requested = cur != CLOSURE_READY; - if (read_mask && fd->read_watcher == NULL && requested) { - fd->read_watcher = watcher; - mask |= read_mask; - } - /* if there is nobody polling for write, but we need to, then start doing so - */ - cur = fd->write_closure; - requested = cur != CLOSURE_READY; - if (write_mask && fd->write_watcher == NULL && requested) { - fd->write_watcher = watcher; - mask |= write_mask; - } - /* if not polling, remember this watcher in case we need someone to later */ - if (mask == 0 && worker != NULL) { - watcher->next = &fd->inactive_watcher_root; - watcher->prev = watcher->next->prev; - watcher->next->prev = watcher->prev->next = watcher; - } - watcher->pollset = pollset; - watcher->worker = worker; - watcher->fd = fd; - gpr_mu_unlock(&fd->mu); - - return mask; -} - -static void fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *watcher, - int got_read, int got_write, - grpc_pollset *read_notifier_pollset) { - int was_polling = 0; - int kick = 0; - grpc_fd *fd = watcher->fd; - - if (fd == NULL) { - return; - } - - gpr_mu_lock(&fd->mu); - - if (watcher == fd->read_watcher) { - /* remove read watcher, kick if we still need a read */ - was_polling = 1; - if (!got_read) { - kick = 1; - } - fd->read_watcher = NULL; - } - if (watcher == fd->write_watcher) { - /* remove write watcher, kick if we still need a write */ - was_polling = 1; - if (!got_write) { - kick = 1; - } - fd->write_watcher = NULL; - } - if (!was_polling && watcher->worker != NULL) { - /* remove from inactive list */ - watcher->next->prev = watcher->prev; - watcher->prev->next = watcher->next; - } - if (got_read) { - if (set_ready_locked(exec_ctx, fd, &fd->read_closure)) { - kick = 1; - } - - if (read_notifier_pollset != NULL) { - set_read_notifier_pollset_locked(exec_ctx, fd, read_notifier_pollset); - } - } - if (got_write) { - if (set_ready_locked(exec_ctx, fd, &fd->write_closure)) { - kick = 1; - } - } - if (kick) { - maybe_wake_one_watcher_locked(fd); - } - if (fd_is_orphaned(fd) && !has_watchers(fd) && !fd->closed) { - close_fd_locked(exec_ctx, fd); - } - gpr_mu_unlock(&fd->mu); - - GRPC_FD_UNREF(fd, "poll"); -} - -static grpc_workqueue *fd_get_workqueue(grpc_fd *fd) { return NULL; } - -/******************************************************************************* - * pollset_posix.c - */ - -GPR_TLS_DECL(g_current_thread_poller); -GPR_TLS_DECL(g_current_thread_worker); - -/** The alarm system needs to be able to wakeup 'some poller' sometimes - * (specifically when a new alarm needs to be triggered earlier than the next - * alarm 'epoch'). - * This wakeup_fd gives us something to alert on when such a case occurs. */ -grpc_wakeup_fd grpc_global_wakeup_fd; - -static void remove_worker(grpc_pollset *p, grpc_pollset_worker *worker) { - worker->prev->next = worker->next; - worker->next->prev = worker->prev; -} - -static int pollset_has_workers(grpc_pollset *p) { - return p->root_worker.next != &p->root_worker; -} - -static grpc_pollset_worker *pop_front_worker(grpc_pollset *p) { - if (pollset_has_workers(p)) { - grpc_pollset_worker *w = p->root_worker.next; - remove_worker(p, w); - return w; - } else { - return NULL; - } -} - -static void push_back_worker(grpc_pollset *p, grpc_pollset_worker *worker) { - worker->next = &p->root_worker; - worker->prev = worker->next->prev; - worker->prev->next = worker->next->prev = worker; -} - -static void push_front_worker(grpc_pollset *p, grpc_pollset_worker *worker) { - worker->prev = &p->root_worker; - worker->next = worker->prev->next; - worker->prev->next = worker->next->prev = worker; -} - -static void kick_append_error(grpc_error **composite, grpc_error *error) { - if (error == GRPC_ERROR_NONE) return; - if (*composite == GRPC_ERROR_NONE) { - *composite = GRPC_ERROR_CREATE("Kick Failure"); - } - *composite = grpc_error_add_child(*composite, error); -} - -static grpc_error *pollset_kick_ext(grpc_pollset *p, - grpc_pollset_worker *specific_worker, - uint32_t flags) { - GPR_TIMER_BEGIN("pollset_kick_ext", 0); - grpc_error *error = GRPC_ERROR_NONE; - - /* pollset->mu already held */ - if (specific_worker != NULL) { - if (specific_worker == GRPC_POLLSET_KICK_BROADCAST) { - GPR_TIMER_BEGIN("pollset_kick_ext.broadcast", 0); - GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0); - for (specific_worker = p->root_worker.next; - specific_worker != &p->root_worker; - specific_worker = specific_worker->next) { - kick_append_error( - &error, grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd)); - } - p->kicked_without_pollers = true; - GPR_TIMER_END("pollset_kick_ext.broadcast", 0); - } else if (gpr_tls_get(&g_current_thread_worker) != - (intptr_t)specific_worker) { - GPR_TIMER_MARK("different_thread_worker", 0); - if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) { - specific_worker->reevaluate_polling_on_wakeup = true; - } - specific_worker->kicked_specifically = true; - kick_append_error(&error, - grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd)); - } else if ((flags & GRPC_POLLSET_CAN_KICK_SELF) != 0) { - GPR_TIMER_MARK("kick_yoself", 0); - if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) { - specific_worker->reevaluate_polling_on_wakeup = true; - } - specific_worker->kicked_specifically = true; - kick_append_error(&error, - grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd)); - } - } else if (gpr_tls_get(&g_current_thread_poller) != (intptr_t)p) { - GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0); - GPR_TIMER_MARK("kick_anonymous", 0); - specific_worker = pop_front_worker(p); - if (specific_worker != NULL) { - if (gpr_tls_get(&g_current_thread_worker) == (intptr_t)specific_worker) { - GPR_TIMER_MARK("kick_anonymous_not_self", 0); - push_back_worker(p, specific_worker); - specific_worker = pop_front_worker(p); - if ((flags & GRPC_POLLSET_CAN_KICK_SELF) == 0 && - gpr_tls_get(&g_current_thread_worker) == - (intptr_t)specific_worker) { - push_back_worker(p, specific_worker); - specific_worker = NULL; - } - } - if (specific_worker != NULL) { - GPR_TIMER_MARK("finally_kick", 0); - push_back_worker(p, specific_worker); - kick_append_error( - &error, grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd)); - } - } else { - GPR_TIMER_MARK("kicked_no_pollers", 0); - p->kicked_without_pollers = true; - } - } - - GPR_TIMER_END("pollset_kick_ext", 0); - return error; -} - -static grpc_error *pollset_kick(grpc_pollset *p, - grpc_pollset_worker *specific_worker) { - return pollset_kick_ext(p, specific_worker, 0); -} - -/* global state management */ - -static grpc_error *pollset_global_init(void) { - gpr_tls_init(&g_current_thread_poller); - gpr_tls_init(&g_current_thread_worker); - return grpc_wakeup_fd_init(&grpc_global_wakeup_fd); -} - -static void pollset_global_shutdown(void) { - grpc_wakeup_fd_destroy(&grpc_global_wakeup_fd); - gpr_tls_destroy(&g_current_thread_poller); - gpr_tls_destroy(&g_current_thread_worker); -} - -static grpc_error *kick_poller(void) { - return grpc_wakeup_fd_wakeup(&grpc_global_wakeup_fd); -} - -/* main interface */ - -static void become_basic_pollset(grpc_pollset *pollset, grpc_fd *fd_or_null); - -static void pollset_init(grpc_pollset *pollset, gpr_mu **mu) { - gpr_mu_init(&pollset->mu); - *mu = &pollset->mu; - pollset->root_worker.next = pollset->root_worker.prev = &pollset->root_worker; - pollset->in_flight_cbs = 0; - pollset->shutting_down = 0; - pollset->called_shutdown = 0; - pollset->kicked_without_pollers = 0; - pollset->idle_jobs.head = pollset->idle_jobs.tail = NULL; - pollset->local_wakeup_cache = NULL; - pollset->kicked_without_pollers = 0; - become_basic_pollset(pollset, NULL); -} - -static void pollset_destroy(grpc_pollset *pollset) { - GPR_ASSERT(pollset->in_flight_cbs == 0); - GPR_ASSERT(!pollset_has_workers(pollset)); - GPR_ASSERT(pollset->idle_jobs.head == pollset->idle_jobs.tail); - pollset->vtable->destroy(pollset); - while (pollset->local_wakeup_cache) { - grpc_cached_wakeup_fd *next = pollset->local_wakeup_cache->next; - grpc_wakeup_fd_destroy(&pollset->local_wakeup_cache->fd); - gpr_free(pollset->local_wakeup_cache); - pollset->local_wakeup_cache = next; - } - gpr_mu_destroy(&pollset->mu); -} - -static void pollset_reset(grpc_pollset *pollset) { - GPR_ASSERT(pollset->shutting_down); - GPR_ASSERT(pollset->in_flight_cbs == 0); - GPR_ASSERT(!pollset_has_workers(pollset)); - GPR_ASSERT(pollset->idle_jobs.head == pollset->idle_jobs.tail); - pollset->vtable->destroy(pollset); - pollset->shutting_down = 0; - pollset->called_shutdown = 0; - pollset->kicked_without_pollers = 0; - become_basic_pollset(pollset, NULL); -} - -static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_fd *fd) { - gpr_mu_lock(&pollset->mu); - pollset->vtable->add_fd(exec_ctx, pollset, fd, 1); -/* the following (enabled only in debug) will reacquire and then release - our lock - meaning that if the unlocking flag passed to add_fd above is - not respected, the code will deadlock (in a way that we have a chance of - debugging) */ -#ifndef NDEBUG - gpr_mu_lock(&pollset->mu); - gpr_mu_unlock(&pollset->mu); -#endif -} - -static void finish_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset) { - GPR_ASSERT(grpc_closure_list_empty(pollset->idle_jobs)); - pollset->vtable->finish_shutdown(pollset); - grpc_exec_ctx_sched(exec_ctx, pollset->shutdown_done, GRPC_ERROR_NONE, NULL); -} - -static void work_combine_error(grpc_error **composite, grpc_error *error) { - if (error == GRPC_ERROR_NONE) return; - if (*composite == GRPC_ERROR_NONE) { - *composite = GRPC_ERROR_CREATE("pollset_work"); - } - *composite = grpc_error_add_child(*composite, error); -} - -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; - grpc_error *error = GRPC_ERROR_NONE; - - /* pollset->mu already held */ - int added_worker = 0; - int locked = 1; - int queued_work = 0; - int keep_polling = 0; - GPR_TIMER_BEGIN("pollset_work", 0); - /* this must happen before we (potentially) drop pollset->mu */ - worker.next = worker.prev = NULL; - worker.reevaluate_polling_on_wakeup = 0; - if (pollset->local_wakeup_cache != NULL) { - worker.wakeup_fd = pollset->local_wakeup_cache; - pollset->local_wakeup_cache = worker.wakeup_fd->next; - } else { - worker.wakeup_fd = gpr_malloc(sizeof(*worker.wakeup_fd)); - error = grpc_wakeup_fd_init(&worker.wakeup_fd->fd); - if (error != GRPC_ERROR_NONE) { - return error; - } - } - worker.kicked_specifically = 0; - /* If there's work waiting for the pollset to be idle, and the - pollset is idle, then do that work */ - if (!pollset_has_workers(pollset) && - !grpc_closure_list_empty(pollset->idle_jobs)) { - GPR_TIMER_MARK("pollset_work.idle_jobs", 0); - grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); - goto done; - } - /* If we're shutting down then we don't execute any extended work */ - if (pollset->shutting_down) { - GPR_TIMER_MARK("pollset_work.shutting_down", 0); - goto done; - } - /* Give do_promote priority so we don't starve it out */ - if (pollset->in_flight_cbs) { - GPR_TIMER_MARK("pollset_work.in_flight_cbs", 0); - gpr_mu_unlock(&pollset->mu); - locked = 0; - goto done; - } - /* Start polling, and keep doing so while we're being asked to - re-evaluate our pollers (this allows poll() based pollers to - ensure they don't miss wakeups) */ - keep_polling = 1; - while (keep_polling) { - keep_polling = 0; - if (!pollset->kicked_without_pollers) { - if (!added_worker) { - push_front_worker(pollset, &worker); - added_worker = 1; - gpr_tls_set(&g_current_thread_worker, (intptr_t)&worker); - } - gpr_tls_set(&g_current_thread_poller, (intptr_t)pollset); - GPR_TIMER_BEGIN("maybe_work_and_unlock", 0); - work_combine_error(&error, - pollset->vtable->maybe_work_and_unlock( - exec_ctx, pollset, &worker, deadline, now)); - GPR_TIMER_END("maybe_work_and_unlock", 0); - locked = 0; - gpr_tls_set(&g_current_thread_poller, 0); - } else { - GPR_TIMER_MARK("pollset_work.kicked_without_pollers", 0); - pollset->kicked_without_pollers = 0; - } - /* Finished execution - start cleaning up. - Note that we may arrive here from outside the enclosing while() loop. - In that case we won't loop though as we haven't added worker to the - worker list, which means nobody could ask us to re-evaluate polling). */ - done: - if (!locked) { - queued_work |= grpc_exec_ctx_flush(exec_ctx); - gpr_mu_lock(&pollset->mu); - locked = 1; - } - /* If we're forced to re-evaluate polling (via pollset_kick with - GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) then we land here and force - a loop */ - if (worker.reevaluate_polling_on_wakeup) { - worker.reevaluate_polling_on_wakeup = 0; - pollset->kicked_without_pollers = 0; - if (queued_work || worker.kicked_specifically) { - /* If there's queued work on the list, then set the deadline to be - immediate so we get back out of the polling loop quickly */ - deadline = gpr_inf_past(GPR_CLOCK_MONOTONIC); - } - keep_polling = 1; - } - } - if (added_worker) { - remove_worker(pollset, &worker); - gpr_tls_set(&g_current_thread_worker, 0); - } - /* release wakeup fd to the local pool */ - worker.wakeup_fd->next = pollset->local_wakeup_cache; - pollset->local_wakeup_cache = worker.wakeup_fd; - /* check shutdown conditions */ - if (pollset->shutting_down) { - if (pollset_has_workers(pollset)) { - pollset_kick(pollset, NULL); - } else if (!pollset->called_shutdown && pollset->in_flight_cbs == 0) { - pollset->called_shutdown = 1; - gpr_mu_unlock(&pollset->mu); - finish_shutdown(exec_ctx, pollset); - grpc_exec_ctx_flush(exec_ctx); - /* Continuing to access pollset here is safe -- it is the caller's - * responsibility to not destroy when it has outstanding calls to - * pollset_work. - * TODO(dklempner): Can we refactor the shutdown logic to avoid this? */ - gpr_mu_lock(&pollset->mu); - } else if (!grpc_closure_list_empty(pollset->idle_jobs)) { - grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); - gpr_mu_unlock(&pollset->mu); - grpc_exec_ctx_flush(exec_ctx); - gpr_mu_lock(&pollset->mu); - } - } - *worker_hdl = NULL; - GPR_TIMER_END("pollset_work", 0); - return error; -} - -static void pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_closure *closure) { - GPR_ASSERT(!pollset->shutting_down); - pollset->shutting_down = 1; - pollset->shutdown_done = closure; - pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); - if (!pollset_has_workers(pollset)) { - grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); - } - if (!pollset->called_shutdown && pollset->in_flight_cbs == 0 && - !pollset_has_workers(pollset)) { - pollset->called_shutdown = 1; - finish_shutdown(exec_ctx, pollset); - } -} - -static int poll_deadline_to_millis_timeout(gpr_timespec deadline, - gpr_timespec now) { - gpr_timespec timeout; - static const int64_t max_spin_polling_us = 10; - if (gpr_time_cmp(deadline, gpr_inf_future(deadline.clock_type)) == 0) { - return -1; - } - if (gpr_time_cmp(deadline, gpr_time_add(now, gpr_time_from_micros( - max_spin_polling_us, - GPR_TIMESPAN))) <= 0) { - return 0; - } - timeout = gpr_time_sub(deadline, now); - return gpr_time_to_millis(gpr_time_add( - timeout, gpr_time_from_nanos(GPR_NS_PER_MS - 1, GPR_TIMESPAN))); -} - -/* - * basic_pollset - a vtable that provides polling for zero or one file - * descriptor via poll() - */ - -typedef struct grpc_unary_promote_args { - const grpc_pollset_vtable *original_vtable; - grpc_pollset *pollset; - grpc_fd *fd; - grpc_closure promotion_closure; -} grpc_unary_promote_args; - -static void basic_do_promote(grpc_exec_ctx *exec_ctx, void *args, - grpc_error *error) { - grpc_unary_promote_args *up_args = args; - const grpc_pollset_vtable *original_vtable = up_args->original_vtable; - grpc_pollset *pollset = up_args->pollset; - grpc_fd *fd = up_args->fd; - - /* - * This is quite tricky. There are a number of cases to keep in mind here: - * 1. fd may have been orphaned - * 2. The pollset may no longer be a unary poller (and we can't let case #1 - * leak to other pollset types!) - * 3. pollset's fd (which may have changed) may have been orphaned - * 4. The pollset may be shutting down. - */ - - gpr_mu_lock(&pollset->mu); - /* First we need to ensure that nobody is polling concurrently */ - GPR_ASSERT(!pollset_has_workers(pollset)); - - gpr_free(up_args); - /* At this point the pollset may no longer be a unary poller. In that case - * we should just call the right add function and be done. */ - /* TODO(klempner): If we're not careful this could cause infinite recursion. - * That's not a problem for now because empty_pollset has a trivial poller - * and we don't have any mechanism to unbecome multipoller. */ - pollset->in_flight_cbs--; - if (pollset->shutting_down) { - /* We don't care about this pollset anymore. */ - if (pollset->in_flight_cbs == 0 && !pollset->called_shutdown) { - pollset->called_shutdown = 1; - finish_shutdown(exec_ctx, pollset); - } - } else if (fd_is_orphaned(fd)) { - /* Don't try to add it to anything, we'll drop our ref on it below */ - } else if (pollset->vtable != original_vtable) { - pollset->vtable->add_fd(exec_ctx, pollset, fd, 0); - } else if (fd != pollset->data.ptr) { - grpc_fd *fds[2]; - fds[0] = pollset->data.ptr; - fds[1] = fd; - - if (fds[0] && !fd_is_orphaned(fds[0])) { - platform_become_multipoller(exec_ctx, pollset, fds, GPR_ARRAY_SIZE(fds)); - GRPC_FD_UNREF(fds[0], "basicpoll"); - } else { - /* old fd is orphaned and we haven't cleaned it up until now, so remain a - * unary poller */ - /* Note that it is possible that fds[1] is also orphaned at this point. - * That's okay, we'll correct it at the next add or poll. */ - if (fds[0]) GRPC_FD_UNREF(fds[0], "basicpoll"); - pollset->data.ptr = fd; - GRPC_FD_REF(fd, "basicpoll"); - } - } - - gpr_mu_unlock(&pollset->mu); - - /* Matching ref in basic_pollset_add_fd */ - GRPC_FD_UNREF(fd, "basicpoll_add"); -} - -static void basic_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_fd *fd, int and_unlock_pollset) { - grpc_unary_promote_args *up_args; - GPR_ASSERT(fd); - if (fd == pollset->data.ptr) goto exit; - - if (!pollset_has_workers(pollset)) { - /* Fast path -- no in flight cbs */ - /* TODO(klempner): Comment this out and fix any test failures or establish - * they are due to timing issues */ - grpc_fd *fds[2]; - fds[0] = pollset->data.ptr; - fds[1] = fd; - - if (fds[0] == NULL) { - pollset->data.ptr = fd; - GRPC_FD_REF(fd, "basicpoll"); - } else if (!fd_is_orphaned(fds[0])) { - platform_become_multipoller(exec_ctx, pollset, fds, GPR_ARRAY_SIZE(fds)); - GRPC_FD_UNREF(fds[0], "basicpoll"); - } else { - /* old fd is orphaned and we haven't cleaned it up until now, so remain a - * unary poller */ - GRPC_FD_UNREF(fds[0], "basicpoll"); - pollset->data.ptr = fd; - GRPC_FD_REF(fd, "basicpoll"); - } - goto exit; - } - - /* Now we need to promote. This needs to happen when we're not polling. Since - * this may be called from poll, the wait needs to happen asynchronously. */ - GRPC_FD_REF(fd, "basicpoll_add"); - pollset->in_flight_cbs++; - up_args = gpr_malloc(sizeof(*up_args)); - up_args->fd = fd; - up_args->original_vtable = pollset->vtable; - up_args->pollset = pollset; - up_args->promotion_closure.cb = basic_do_promote; - up_args->promotion_closure.cb_arg = up_args; - - grpc_closure_list_append(&pollset->idle_jobs, &up_args->promotion_closure, - GRPC_ERROR_NONE); - pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); - -exit: - if (and_unlock_pollset) { - gpr_mu_unlock(&pollset->mu); - } -} - -static grpc_error *basic_pollset_maybe_work_and_unlock( - grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker, - gpr_timespec deadline, gpr_timespec now) { -#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR) -#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR) - - struct pollfd pfd[3]; - grpc_fd *fd; - grpc_fd_watcher fd_watcher; - int timeout; - int r; - nfds_t nfds; - grpc_error *error = GRPC_ERROR_NONE; - - fd = pollset->data.ptr; - if (fd && fd_is_orphaned(fd)) { - GRPC_FD_UNREF(fd, "basicpoll"); - fd = pollset->data.ptr = NULL; - } - timeout = poll_deadline_to_millis_timeout(deadline, now); - pfd[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd); - pfd[0].events = POLLIN; - pfd[0].revents = 0; - pfd[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); - pfd[1].events = POLLIN; - pfd[1].revents = 0; - nfds = 2; - if (fd) { - pfd[2].fd = fd->fd; - pfd[2].revents = 0; - GRPC_FD_REF(fd, "basicpoll_begin"); - gpr_mu_unlock(&pollset->mu); - pfd[2].events = - (short)fd_begin_poll(fd, pollset, worker, POLLIN, POLLOUT, &fd_watcher); - if (pfd[2].events != 0) { - nfds++; - } - } else { - gpr_mu_unlock(&pollset->mu); - } - - /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid - even going into the blocking annotation if possible */ - /* poll fd count (argument 2) is shortened by one if we have no events - to poll on - such that it only includes the kicker */ - GPR_TIMER_BEGIN("poll", 0); - GRPC_SCHEDULING_START_BLOCKING_REGION; - r = grpc_poll_function(pfd, nfds, timeout); - GRPC_SCHEDULING_END_BLOCKING_REGION; - GPR_TIMER_END("poll", 0); - - if (r < 0) { - if (errno != EINTR) { - work_combine_error(&error, GRPC_OS_ERROR(errno, "poll")); - } - if (fd) { - fd_end_poll(exec_ctx, &fd_watcher, 0, 0, NULL); - } - } else if (r == 0) { - if (fd) { - fd_end_poll(exec_ctx, &fd_watcher, 0, 0, NULL); - } - } else { - if (pfd[0].revents & POLLIN_CHECK) { - work_combine_error(&error, - grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd)); - } - if (pfd[1].revents & POLLIN_CHECK) { - work_combine_error(&error, - grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd)); - } - if (nfds > 2) { - fd_end_poll(exec_ctx, &fd_watcher, pfd[2].revents & POLLIN_CHECK, - pfd[2].revents & POLLOUT_CHECK, pollset); - } else if (fd) { - fd_end_poll(exec_ctx, &fd_watcher, 0, 0, NULL); - } - } - - if (fd) { - GRPC_FD_UNREF(fd, "basicpoll_begin"); - } - - return error; -} - -static void basic_pollset_destroy(grpc_pollset *pollset) { - if (pollset->data.ptr != NULL) { - GRPC_FD_UNREF(pollset->data.ptr, "basicpoll"); - pollset->data.ptr = NULL; - } -} - -static const grpc_pollset_vtable basic_pollset = { - basic_pollset_add_fd, basic_pollset_maybe_work_and_unlock, - basic_pollset_destroy, basic_pollset_destroy}; - -static void become_basic_pollset(grpc_pollset *pollset, grpc_fd *fd_or_null) { - pollset->vtable = &basic_pollset; - pollset->data.ptr = fd_or_null; - if (fd_or_null != NULL) { - GRPC_FD_REF(fd_or_null, "basicpoll"); - } -} - -/******************************************************************************* - * pollset_multipoller_with_poll_posix.c - */ - -#ifndef GRPC_LINUX_MULTIPOLL_WITH_EPOLL - -typedef struct { - /* all polled fds */ - size_t fd_count; - size_t fd_capacity; - grpc_fd **fds; - /* fds that have been removed from the pollset explicitly */ - size_t del_count; - size_t del_capacity; - grpc_fd **dels; -} poll_hdr; - -static void multipoll_with_poll_pollset_add_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, - grpc_fd *fd, - int and_unlock_pollset) { - size_t i; - poll_hdr *h = pollset->data.ptr; - /* TODO(ctiller): this is O(num_fds^2); maybe switch to a hash set here */ - for (i = 0; i < h->fd_count; i++) { - if (h->fds[i] == fd) goto exit; - } - if (h->fd_count == h->fd_capacity) { - h->fd_capacity = GPR_MAX(h->fd_capacity + 8, h->fd_count * 3 / 2); - h->fds = gpr_realloc(h->fds, sizeof(grpc_fd *) * h->fd_capacity); - } - h->fds[h->fd_count++] = fd; - GRPC_FD_REF(fd, "multipoller"); -exit: - if (and_unlock_pollset) { - gpr_mu_unlock(&pollset->mu); - } -} - -static grpc_error *multipoll_with_poll_pollset_maybe_work_and_unlock( - grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker, - gpr_timespec deadline, gpr_timespec now) { -#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR) -#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR) - - int timeout; - int r; - size_t i, j, fd_count; - nfds_t pfd_count; - poll_hdr *h; - /* TODO(ctiller): inline some elements to avoid an allocation */ - grpc_fd_watcher *watchers; - struct pollfd *pfds; - grpc_error *error = GRPC_ERROR_NONE; - - h = pollset->data.ptr; - timeout = poll_deadline_to_millis_timeout(deadline, now); - /* TODO(ctiller): perform just one malloc here if we exceed the inline case */ - pfds = gpr_malloc(sizeof(*pfds) * (h->fd_count + 2)); - watchers = gpr_malloc(sizeof(*watchers) * (h->fd_count + 2)); - fd_count = 0; - pfd_count = 2; - pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd); - pfds[0].events = POLLIN; - pfds[0].revents = 0; - pfds[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); - pfds[1].events = POLLIN; - pfds[1].revents = 0; - for (i = 0; i < h->fd_count; i++) { - int remove = fd_is_orphaned(h->fds[i]); - for (j = 0; !remove && j < h->del_count; j++) { - if (h->fds[i] == h->dels[j]) remove = 1; - } - if (remove) { - GRPC_FD_UNREF(h->fds[i], "multipoller"); - } else { - h->fds[fd_count++] = h->fds[i]; - watchers[pfd_count].fd = h->fds[i]; - GRPC_FD_REF(watchers[pfd_count].fd, "multipoller_start"); - pfds[pfd_count].fd = h->fds[i]->fd; - pfds[pfd_count].revents = 0; - pfd_count++; - } - } - for (j = 0; j < h->del_count; j++) { - GRPC_FD_UNREF(h->dels[j], "multipoller_del"); - } - h->del_count = 0; - h->fd_count = fd_count; - gpr_mu_unlock(&pollset->mu); - - for (i = 2; i < pfd_count; i++) { - grpc_fd *fd = watchers[i].fd; - pfds[i].events = (short)fd_begin_poll(fd, pollset, worker, POLLIN, POLLOUT, - &watchers[i]); - GRPC_FD_UNREF(fd, "multipoller_start"); - } - - /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid - even going into the blocking annotation if possible */ - GRPC_SCHEDULING_START_BLOCKING_REGION; - r = grpc_poll_function(pfds, pfd_count, timeout); - GRPC_SCHEDULING_END_BLOCKING_REGION; - - if (r < 0) { - if (errno != EINTR) { - work_combine_error(&error, GRPC_OS_ERROR(errno, "poll")); - } - for (i = 2; i < pfd_count; i++) { - fd_end_poll(exec_ctx, &watchers[i], 0, 0, NULL); - } - } else if (r == 0) { - for (i = 2; i < pfd_count; i++) { - fd_end_poll(exec_ctx, &watchers[i], 0, 0, NULL); - } - } else { - if (pfds[0].revents & POLLIN_CHECK) { - work_combine_error(&error, - grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd)); - } - if (pfds[1].revents & POLLIN_CHECK) { - work_combine_error(&error, - grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd)); - } - for (i = 2; i < pfd_count; i++) { - if (watchers[i].fd == NULL) { - fd_end_poll(exec_ctx, &watchers[i], 0, 0, NULL); - continue; - } - fd_end_poll(exec_ctx, &watchers[i], pfds[i].revents & POLLIN_CHECK, - pfds[i].revents & POLLOUT_CHECK, pollset); - } - } - - gpr_free(pfds); - gpr_free(watchers); - - return error; -} - -static void multipoll_with_poll_pollset_finish_shutdown(grpc_pollset *pollset) { - size_t i; - poll_hdr *h = pollset->data.ptr; - for (i = 0; i < h->fd_count; i++) { - GRPC_FD_UNREF(h->fds[i], "multipoller"); - } - for (i = 0; i < h->del_count; i++) { - GRPC_FD_UNREF(h->dels[i], "multipoller_del"); - } - h->fd_count = 0; - h->del_count = 0; -} - -static void multipoll_with_poll_pollset_destroy(grpc_pollset *pollset) { - poll_hdr *h = pollset->data.ptr; - multipoll_with_poll_pollset_finish_shutdown(pollset); - gpr_free(h->fds); - gpr_free(h->dels); - gpr_free(h); -} - -static const grpc_pollset_vtable multipoll_with_poll_pollset = { - multipoll_with_poll_pollset_add_fd, - multipoll_with_poll_pollset_maybe_work_and_unlock, - multipoll_with_poll_pollset_finish_shutdown, - multipoll_with_poll_pollset_destroy}; - -static void poll_become_multipoller(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, grpc_fd **fds, - size_t nfds) { - size_t i; - poll_hdr *h = gpr_malloc(sizeof(poll_hdr)); - pollset->vtable = &multipoll_with_poll_pollset; - pollset->data.ptr = h; - h->fd_count = nfds; - h->fd_capacity = nfds; - h->fds = gpr_malloc(nfds * sizeof(grpc_fd *)); - h->del_count = 0; - h->del_capacity = 0; - h->dels = NULL; - for (i = 0; i < nfds; i++) { - h->fds[i] = fds[i]; - GRPC_FD_REF(fds[i], "multipoller"); - } -} - -#endif /* !GRPC_LINUX_MULTIPOLL_WITH_EPOLL */ - -/******************************************************************************* - * pollset_multipoller_with_epoll_posix.c - */ - -#ifdef GRPC_LINUX_MULTIPOLL_WITH_EPOLL - -#include <errno.h> -#include <poll.h> -#include <string.h> -#include <sys/epoll.h> -#include <unistd.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/useful.h> - -#include "src/core/lib/iomgr/ev_posix.h" -#include "src/core/lib/profiling/timers.h" -#include "src/core/lib/support/block_annotate.h" - -static void set_ready(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure **st, - grpc_pollset *read_notifier_pollset) { - /* only one set_ready can be active at once (but there may be a racing - notify_on) */ - gpr_mu_lock(&fd->mu); - set_ready_locked(exec_ctx, fd, st); - - /* A non-NULL read_notifier_pollset means that the fd is readable. */ - if (read_notifier_pollset != NULL) { - /* Note: Since the fd might be a part of multiple pollsets, this might be - * called multiple times (for each time the fd becomes readable) and it is - * okay to set the fd's read-notifier pollset to anyone of these pollsets */ - set_read_notifier_pollset_locked(exec_ctx, fd, read_notifier_pollset); - } - - gpr_mu_unlock(&fd->mu); -} - -static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_pollset *notifier_pollset) { - set_ready(exec_ctx, fd, &fd->read_closure, notifier_pollset); -} - -static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { - set_ready(exec_ctx, fd, &fd->write_closure, NULL); -} - -struct epoll_fd_list { - int *epoll_fds; - size_t count; - size_t capacity; -}; - -static struct epoll_fd_list epoll_fd_global_list; -static gpr_once init_epoll_fd_list_mu = GPR_ONCE_INIT; -static gpr_mu epoll_fd_list_mu; - -static void init_mu(void) { gpr_mu_init(&epoll_fd_list_mu); } - -static void add_epoll_fd_to_global_list(int epoll_fd) { - gpr_once_init(&init_epoll_fd_list_mu, init_mu); - - gpr_mu_lock(&epoll_fd_list_mu); - if (epoll_fd_global_list.count == epoll_fd_global_list.capacity) { - epoll_fd_global_list.capacity = - GPR_MAX((size_t)8, epoll_fd_global_list.capacity * 2); - epoll_fd_global_list.epoll_fds = - gpr_realloc(epoll_fd_global_list.epoll_fds, - epoll_fd_global_list.capacity * sizeof(int)); - } - epoll_fd_global_list.epoll_fds[epoll_fd_global_list.count++] = epoll_fd; - gpr_mu_unlock(&epoll_fd_list_mu); -} - -static void remove_epoll_fd_from_global_list(int epoll_fd) { - gpr_mu_lock(&epoll_fd_list_mu); - GPR_ASSERT(epoll_fd_global_list.count > 0); - for (size_t i = 0; i < epoll_fd_global_list.count; i++) { - if (epoll_fd == epoll_fd_global_list.epoll_fds[i]) { - epoll_fd_global_list.epoll_fds[i] = - epoll_fd_global_list.epoll_fds[--(epoll_fd_global_list.count)]; - break; - } - } - gpr_mu_unlock(&epoll_fd_list_mu); -} - -static void remove_fd_from_all_epoll_sets(int fd) { - int err; - gpr_once_init(&init_epoll_fd_list_mu, init_mu); - gpr_mu_lock(&epoll_fd_list_mu); - if (epoll_fd_global_list.count == 0) { - gpr_mu_unlock(&epoll_fd_list_mu); - return; - } - for (size_t i = 0; i < epoll_fd_global_list.count; i++) { - err = epoll_ctl(epoll_fd_global_list.epoll_fds[i], EPOLL_CTL_DEL, fd, NULL); - if (err < 0 && errno != ENOENT) { - gpr_log(GPR_ERROR, "epoll_ctl del for %d failed: %s", fd, - strerror(errno)); - } - } - gpr_mu_unlock(&epoll_fd_list_mu); -} - -typedef struct { - grpc_pollset *pollset; - grpc_fd *fd; - grpc_closure closure; -} delayed_add; - -typedef struct { int epoll_fd; } epoll_hdr; - -static void finally_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_fd *fd) { - epoll_hdr *h = pollset->data.ptr; - struct epoll_event ev; - int err; - grpc_fd_watcher watcher; - - /* We pretend to be polling whilst adding an fd to keep the fd from being - closed during the add. This may result in a spurious wakeup being assigned - to this pollset whilst adding, but that should be benign. */ - GPR_ASSERT(fd_begin_poll(fd, pollset, NULL, 0, 0, &watcher) == 0); - if (watcher.fd != NULL) { - ev.events = (uint32_t)(EPOLLIN | EPOLLOUT | EPOLLET); - ev.data.ptr = fd; - err = epoll_ctl(h->epoll_fd, EPOLL_CTL_ADD, fd->fd, &ev); - if (err < 0) { - /* FDs may be added to a pollset multiple times, so EEXIST is normal. */ - if (errno != EEXIST) { - gpr_log(GPR_ERROR, "epoll_ctl add for %d failed: %s", fd->fd, - strerror(errno)); - } - } - } - fd_end_poll(exec_ctx, &watcher, 0, 0, NULL); -} - -static void perform_delayed_add(grpc_exec_ctx *exec_ctx, void *arg, - grpc_error *error) { - delayed_add *da = arg; - - if (!fd_is_orphaned(da->fd)) { - finally_add_fd(exec_ctx, da->pollset, da->fd); - } - - gpr_mu_lock(&da->pollset->mu); - da->pollset->in_flight_cbs--; - if (da->pollset->shutting_down) { - /* We don't care about this pollset anymore. */ - if (da->pollset->in_flight_cbs == 0 && !da->pollset->called_shutdown) { - da->pollset->called_shutdown = 1; - grpc_exec_ctx_sched(exec_ctx, da->pollset->shutdown_done, GRPC_ERROR_NONE, - NULL); - } - } - gpr_mu_unlock(&da->pollset->mu); - - GRPC_FD_UNREF(da->fd, "delayed_add"); - - gpr_free(da); -} - -static void multipoll_with_epoll_pollset_add_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, - grpc_fd *fd, - int and_unlock_pollset) { - if (and_unlock_pollset) { - gpr_mu_unlock(&pollset->mu); - finally_add_fd(exec_ctx, pollset, fd); - } else { - delayed_add *da = gpr_malloc(sizeof(*da)); - da->pollset = pollset; - da->fd = fd; - GRPC_FD_REF(fd, "delayed_add"); - grpc_closure_init(&da->closure, perform_delayed_add, da); - pollset->in_flight_cbs++; - grpc_exec_ctx_sched(exec_ctx, &da->closure, GRPC_ERROR_NONE, NULL); - } -} - -/* TODO(klempner): We probably want to turn this down a bit */ -#define GRPC_EPOLL_MAX_EVENTS 1000 - -static grpc_error *multipoll_with_epoll_pollset_maybe_work_and_unlock( - grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker, - gpr_timespec deadline, gpr_timespec now) { - struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS]; - int ep_rv; - int poll_rv; - epoll_hdr *h = pollset->data.ptr; - int timeout_ms; - struct pollfd pfds[2]; - grpc_error *error = GRPC_ERROR_NONE; - - /* If you want to ignore epoll's ability to sanely handle parallel pollers, - * for a more apples-to-apples performance comparison with poll, add a - * if (pollset->counter != 0) { return 0; } - * here. - */ - - gpr_mu_unlock(&pollset->mu); - - timeout_ms = poll_deadline_to_millis_timeout(deadline, now); - - pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); - pfds[0].events = POLLIN; - pfds[0].revents = 0; - pfds[1].fd = h->epoll_fd; - pfds[1].events = POLLIN; - pfds[1].revents = 0; - - /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid - even going into the blocking annotation if possible */ - GPR_TIMER_BEGIN("poll", 0); - GRPC_SCHEDULING_START_BLOCKING_REGION; - poll_rv = grpc_poll_function(pfds, 2, timeout_ms); - GRPC_SCHEDULING_END_BLOCKING_REGION; - GPR_TIMER_END("poll", 0); - - if (poll_rv < 0) { - if (errno != EINTR) { - work_combine_error(&error, GRPC_OS_ERROR(errno, "poll")); - } - } else if (poll_rv == 0) { - /* do nothing */ - } else { - if (pfds[0].revents) { - work_combine_error(&error, - grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd)); - } - if (pfds[1].revents) { - do { - /* The following epoll_wait never blocks; it has a timeout of 0 */ - ep_rv = epoll_wait(h->epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, 0); - if (ep_rv < 0) { - if (errno != EINTR) { - work_combine_error(&error, GRPC_OS_ERROR(errno, "epoll_wait")); - } - } else { - int i; - for (i = 0; i < ep_rv; ++i) { - grpc_fd *fd = ep_ev[i].data.ptr; - /* TODO(klempner): We might want to consider making err and pri - * separate events */ - int cancel = ep_ev[i].events & (EPOLLERR | EPOLLHUP); - int read_ev = ep_ev[i].events & (EPOLLIN | EPOLLPRI); - int write_ev = ep_ev[i].events & EPOLLOUT; - if (fd == NULL) { - work_combine_error(&error, grpc_wakeup_fd_consume_wakeup( - &grpc_global_wakeup_fd)); - } else { - if (read_ev || cancel) { - fd_become_readable(exec_ctx, fd, pollset); - } - if (write_ev || cancel) { - fd_become_writable(exec_ctx, fd); - } - } - } - } - } while (ep_rv == GRPC_EPOLL_MAX_EVENTS); - } - } - return error; -} - -static void multipoll_with_epoll_pollset_finish_shutdown( - grpc_pollset *pollset) {} - -static void multipoll_with_epoll_pollset_destroy(grpc_pollset *pollset) { - epoll_hdr *h = pollset->data.ptr; - close(h->epoll_fd); - remove_epoll_fd_from_global_list(h->epoll_fd); - gpr_free(h); -} - -static const grpc_pollset_vtable multipoll_with_epoll_pollset = { - multipoll_with_epoll_pollset_add_fd, - multipoll_with_epoll_pollset_maybe_work_and_unlock, - multipoll_with_epoll_pollset_finish_shutdown, - multipoll_with_epoll_pollset_destroy}; - -static void epoll_become_multipoller(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, grpc_fd **fds, - size_t nfds) { - size_t i; - epoll_hdr *h = gpr_malloc(sizeof(epoll_hdr)); - struct epoll_event ev; - int err; - - pollset->vtable = &multipoll_with_epoll_pollset; - pollset->data.ptr = h; - h->epoll_fd = epoll_create1(EPOLL_CLOEXEC); - if (h->epoll_fd < 0) { - /* TODO(klempner): Fall back to poll here, especially on ENOSYS */ - gpr_log(GPR_ERROR, "epoll_create1 failed: %s", strerror(errno)); - abort(); - } - add_epoll_fd_to_global_list(h->epoll_fd); - - ev.events = (uint32_t)(EPOLLIN | EPOLLET); - ev.data.ptr = NULL; - err = epoll_ctl(h->epoll_fd, EPOLL_CTL_ADD, - GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), &ev); - if (err < 0) { - gpr_log(GPR_ERROR, "epoll_ctl add for %d failed: %s", - GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), - strerror(errno)); - } - - for (i = 0; i < nfds; i++) { - multipoll_with_epoll_pollset_add_fd(exec_ctx, pollset, fds[i], 0); - } -} - -#else /* GRPC_LINUX_MULTIPOLL_WITH_EPOLL */ - -static void remove_fd_from_all_epoll_sets(int fd) {} - -#endif /* GRPC_LINUX_MULTIPOLL_WITH_EPOLL */ - -/******************************************************************************* - * pollset_set_posix.c - */ - -static grpc_pollset_set *pollset_set_create(void) { - grpc_pollset_set *pollset_set = gpr_malloc(sizeof(*pollset_set)); - memset(pollset_set, 0, sizeof(*pollset_set)); - gpr_mu_init(&pollset_set->mu); - return pollset_set; -} - -static void pollset_set_destroy(grpc_pollset_set *pollset_set) { - size_t i; - gpr_mu_destroy(&pollset_set->mu); - for (i = 0; i < pollset_set->fd_count; i++) { - GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set"); - } - gpr_free(pollset_set->pollsets); - gpr_free(pollset_set->pollset_sets); - gpr_free(pollset_set->fds); - gpr_free(pollset_set); -} - -static void pollset_set_add_pollset(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, - grpc_pollset *pollset) { - size_t i, j; - gpr_mu_lock(&pollset_set->mu); - if (pollset_set->pollset_count == pollset_set->pollset_capacity) { - pollset_set->pollset_capacity = - GPR_MAX(8, 2 * pollset_set->pollset_capacity); - pollset_set->pollsets = - gpr_realloc(pollset_set->pollsets, pollset_set->pollset_capacity * - sizeof(*pollset_set->pollsets)); - } - pollset_set->pollsets[pollset_set->pollset_count++] = pollset; - for (i = 0, j = 0; i < pollset_set->fd_count; i++) { - if (fd_is_orphaned(pollset_set->fds[i])) { - GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set"); - } else { - pollset_add_fd(exec_ctx, pollset, pollset_set->fds[i]); - pollset_set->fds[j++] = pollset_set->fds[i]; - } - } - pollset_set->fd_count = j; - gpr_mu_unlock(&pollset_set->mu); -} - -static void pollset_set_del_pollset(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, - grpc_pollset *pollset) { - size_t i; - gpr_mu_lock(&pollset_set->mu); - for (i = 0; i < pollset_set->pollset_count; i++) { - if (pollset_set->pollsets[i] == pollset) { - pollset_set->pollset_count--; - GPR_SWAP(grpc_pollset *, pollset_set->pollsets[i], - pollset_set->pollsets[pollset_set->pollset_count]); - break; - } - } - gpr_mu_unlock(&pollset_set->mu); -} - -static void pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *bag, - grpc_pollset_set *item) { - size_t i, j; - gpr_mu_lock(&bag->mu); - if (bag->pollset_set_count == bag->pollset_set_capacity) { - bag->pollset_set_capacity = GPR_MAX(8, 2 * bag->pollset_set_capacity); - bag->pollset_sets = - gpr_realloc(bag->pollset_sets, - bag->pollset_set_capacity * sizeof(*bag->pollset_sets)); - } - bag->pollset_sets[bag->pollset_set_count++] = item; - for (i = 0, j = 0; i < bag->fd_count; i++) { - if (fd_is_orphaned(bag->fds[i])) { - GRPC_FD_UNREF(bag->fds[i], "pollset_set"); - } else { - pollset_set_add_fd(exec_ctx, item, bag->fds[i]); - bag->fds[j++] = bag->fds[i]; - } - } - bag->fd_count = j; - gpr_mu_unlock(&bag->mu); -} - -static void pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *bag, - grpc_pollset_set *item) { - size_t i; - gpr_mu_lock(&bag->mu); - for (i = 0; i < bag->pollset_set_count; i++) { - if (bag->pollset_sets[i] == item) { - bag->pollset_set_count--; - GPR_SWAP(grpc_pollset_set *, bag->pollset_sets[i], - bag->pollset_sets[bag->pollset_set_count]); - break; - } - } - gpr_mu_unlock(&bag->mu); -} - -static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, grpc_fd *fd) { - size_t i; - gpr_mu_lock(&pollset_set->mu); - if (pollset_set->fd_count == pollset_set->fd_capacity) { - pollset_set->fd_capacity = GPR_MAX(8, 2 * pollset_set->fd_capacity); - pollset_set->fds = gpr_realloc( - pollset_set->fds, pollset_set->fd_capacity * sizeof(*pollset_set->fds)); - } - GRPC_FD_REF(fd, "pollset_set"); - pollset_set->fds[pollset_set->fd_count++] = fd; - for (i = 0; i < pollset_set->pollset_count; i++) { - pollset_add_fd(exec_ctx, pollset_set->pollsets[i], fd); - } - for (i = 0; i < pollset_set->pollset_set_count; i++) { - pollset_set_add_fd(exec_ctx, pollset_set->pollset_sets[i], fd); - } - gpr_mu_unlock(&pollset_set->mu); -} - -static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, grpc_fd *fd) { - size_t i; - gpr_mu_lock(&pollset_set->mu); - for (i = 0; i < pollset_set->fd_count; i++) { - if (pollset_set->fds[i] == fd) { - pollset_set->fd_count--; - GPR_SWAP(grpc_fd *, pollset_set->fds[i], - pollset_set->fds[pollset_set->fd_count]); - GRPC_FD_UNREF(fd, "pollset_set"); - break; - } - } - for (i = 0; i < pollset_set->pollset_set_count; i++) { - pollset_set_del_fd(exec_ctx, pollset_set->pollset_sets[i], fd); - } - gpr_mu_unlock(&pollset_set->mu); -} - -/******************************************************************************* - * workqueue stubs - */ - -#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG -static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue, - const char *file, int line, - const char *reason) { - return workqueue; -} -static void workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, - const char *file, int line, const char *reason) {} -#else -static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue) { - return workqueue; -} -static void workqueue_unref(grpc_exec_ctx *exec_ctx, - grpc_workqueue *workqueue) {} -#endif - -static void workqueue_enqueue(grpc_exec_ctx *exec_ctx, - grpc_workqueue *workqueue, grpc_closure *closure, - grpc_error *error) { - grpc_exec_ctx_sched(exec_ctx, closure, error, NULL); -} - -/******************************************************************************* - * event engine binding - */ - -static void shutdown_engine(void) { - fd_global_shutdown(); - pollset_global_shutdown(); -} - -static const grpc_event_engine_vtable vtable = { - .pollset_size = sizeof(grpc_pollset), - - .fd_create = fd_create, - .fd_wrapped_fd = fd_wrapped_fd, - .fd_orphan = fd_orphan, - .fd_shutdown = fd_shutdown, - .fd_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_reset = pollset_reset, - .pollset_destroy = pollset_destroy, - .pollset_work = pollset_work, - .pollset_kick = pollset_kick, - .pollset_add_fd = pollset_add_fd, - - .pollset_set_create = pollset_set_create, - .pollset_set_destroy = pollset_set_destroy, - .pollset_set_add_pollset = pollset_set_add_pollset, - .pollset_set_del_pollset = pollset_set_del_pollset, - .pollset_set_add_pollset_set = pollset_set_add_pollset_set, - .pollset_set_del_pollset_set = pollset_set_del_pollset_set, - .pollset_set_add_fd = pollset_set_add_fd, - .pollset_set_del_fd = pollset_set_del_fd, - - .kick_poller = kick_poller, - - .workqueue_ref = workqueue_ref, - .workqueue_unref = workqueue_unref, - .workqueue_enqueue = workqueue_enqueue, - - .shutdown_engine = shutdown_engine, -}; - -const grpc_event_engine_vtable *grpc_init_poll_and_epoll_posix(void) { -#ifdef GRPC_LINUX_MULTIPOLL_WITH_EPOLL - platform_become_multipoller = epoll_become_multipoller; -#else - platform_become_multipoller = poll_become_multipoller; -#endif - fd_global_init(); - pollset_global_init(); - return &vtable; -} - -#endif diff --git a/src/core/lib/iomgr/ev_poll_and_epoll_posix.h b/src/core/lib/iomgr/ev_poll_and_epoll_posix.h deleted file mode 100644 index 06d6dbf29d..0000000000 --- a/src/core/lib/iomgr/ev_poll_and_epoll_posix.h +++ /dev/null @@ -1,41 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H -#define GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H - -#include "src/core/lib/iomgr/ev_posix.h" - -const grpc_event_engine_vtable *grpc_init_poll_and_epoll_posix(void); - -#endif /* GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H */ diff --git a/src/core/lib/iomgr/ev_posix.c b/src/core/lib/iomgr/ev_posix.c index ef36ba89b2..ab139895fd 100644 --- a/src/core/lib/iomgr/ev_posix.c +++ b/src/core/lib/iomgr/ev_posix.c @@ -45,7 +45,6 @@ #include <grpc/support/useful.h> #include "src/core/lib/iomgr/ev_epoll_linux.h" -#include "src/core/lib/iomgr/ev_poll_and_epoll_posix.h" #include "src/core/lib/iomgr/ev_poll_posix.h" #include "src/core/lib/support/env.h" @@ -67,7 +66,6 @@ static const event_engine_factory g_factories[] = { {"epoll", grpc_init_epoll_linux}, {"poll", grpc_init_poll_posix}, {"poll-cv", grpc_init_poll_cv_posix}, - {"legacy", grpc_init_poll_and_epoll_posix}, }; static void add(const char *beg, const char *end, char ***ss, size_t *ns) { diff --git a/src/core/lib/iomgr/load_file.c b/src/core/lib/iomgr/load_file.c index b62ecbc534..217bc5da59 100644 --- a/src/core/lib/iomgr/load_file.c +++ b/src/core/lib/iomgr/load_file.c @@ -44,10 +44,10 @@ #include "src/core/lib/support/string.h" grpc_error *grpc_load_file(const char *filename, int add_null_terminator, - gpr_slice *output) { + grpc_slice *output) { unsigned char *contents = NULL; size_t contents_size = 0; - gpr_slice result = gpr_empty_slice(); + grpc_slice result = gpr_empty_slice(); FILE *file; size_t bytes_read = 0; grpc_error *error = GRPC_ERROR_NONE; @@ -72,7 +72,7 @@ grpc_error *grpc_load_file(const char *filename, int add_null_terminator, if (add_null_terminator) { contents[contents_size++] = 0; } - result = gpr_slice_new(contents, contents_size, gpr_free); + result = grpc_slice_new(contents, contents_size, gpr_free); end: *output = result; diff --git a/src/core/lib/iomgr/load_file.h b/src/core/lib/iomgr/load_file.h index 9aac2225d1..73ee8c3abf 100644 --- a/src/core/lib/iomgr/load_file.h +++ b/src/core/lib/iomgr/load_file.h @@ -36,7 +36,7 @@ #include <stdio.h> -#include <grpc/support/slice.h> +#include <grpc/slice.h> #include "src/core/lib/iomgr/error.h" @@ -47,7 +47,7 @@ extern "C" { /* Loads the content of a file into a slice. add_null_terminator will add a NULL terminator if non-zero. */ grpc_error *grpc_load_file(const char *filename, int add_null_terminator, - gpr_slice *slice); + grpc_slice *slice); #ifdef __cplusplus } diff --git a/src/core/lib/iomgr/resource_quota.c b/src/core/lib/iomgr/resource_quota.c index 8a06443d58..051a30baa3 100644 --- a/src/core/lib/iomgr/resource_quota.c +++ b/src/core/lib/iomgr/resource_quota.c @@ -44,6 +44,81 @@ int grpc_resource_quota_trace = 0; +/* Internal linked list pointers for a resource user */ +typedef struct { + grpc_resource_user *next; + grpc_resource_user *prev; +} grpc_resource_user_link; + +/* Resource users are kept in (potentially) several intrusive linked lists + at once. These are the list names. */ +typedef enum { + /* Resource users that are waiting for an allocation */ + GRPC_RULIST_AWAITING_ALLOCATION, + /* Resource users that have free memory available for internal reclamation */ + GRPC_RULIST_NON_EMPTY_FREE_POOL, + /* Resource users that have published a benign reclamation is available */ + GRPC_RULIST_RECLAIMER_BENIGN, + /* Resource users that have published a destructive reclamation is + available */ + GRPC_RULIST_RECLAIMER_DESTRUCTIVE, + /* Number of lists: must be last */ + GRPC_RULIST_COUNT +} grpc_rulist; + +struct grpc_resource_user { + /* The quota this resource user consumes from */ + grpc_resource_quota *resource_quota; + + /* Closure to schedule an allocation under the resource quota combiner lock */ + grpc_closure allocate_closure; + /* Closure to publish a non empty free pool under the resource quota combiner + lock */ + grpc_closure add_to_free_pool_closure; + + /* one ref for each ref call (released by grpc_resource_user_unref), and one + ref for each byte allocated (released by grpc_resource_user_free) */ + gpr_atm refs; + /* is this resource user unlocked? starts at 0, increases for each shutdown + call */ + gpr_atm shutdown; + + gpr_mu mu; + /* The amount of memory (in bytes) this user has cached for its own use: to + avoid quota contention, each resource user can keep some memory in + addition to what it is immediately using (e.g., for caching), and the quota + can pull it back under memory pressure. + This value can become negative if more memory has been requested than + existed in the free pool, at which point the quota is consulted to bring + this value non-negative (asynchronously). */ + int64_t free_pool; + /* A list of closures to call once free_pool becomes non-negative - ie when + all outstanding allocations have been granted. */ + grpc_closure_list on_allocated; + /* True if we are currently trying to allocate from the quota, false if not */ + bool allocating; + /* True if we are currently trying to add ourselves to the non-free quota + list, false otherwise */ + bool added_to_free_pool; + + /* Reclaimers: index 0 is the benign reclaimer, 1 is the destructive reclaimer + */ + grpc_closure *reclaimers[2]; + /* Trampoline closures to finish reclamation and re-enter the quota combiner + lock */ + grpc_closure post_reclaimer_closure[2]; + + /* Closure to execute under the quota combiner to de-register and shutdown the + resource user */ + grpc_closure destroy_closure; + + /* Links in the various grpc_rulist lists */ + grpc_resource_user_link links[GRPC_RULIST_COUNT]; + + /* The name of this resource user, for debugging/tracing */ + char *name; +}; + struct grpc_resource_quota { /* refcount */ gpr_refcount refs; @@ -272,7 +347,7 @@ static bool rq_reclaim(grpc_exec_ctx *exec_ctx, */ typedef struct { - gpr_slice_refcount base; + grpc_slice_refcount base; gpr_refcount refs; grpc_resource_user *resource_user; size_t size; @@ -289,7 +364,7 @@ static void ru_slice_unref(void *p) { /* TODO(ctiller): this is dangerous, but I think safe for now: we have no guarantee here that we're at a safe point for creating an execution context, but we have no way of writing this code otherwise. - In the future: consider lifting gpr_slice to grpc, and offering an + In the future: consider lifting grpc_slice to grpc, and offering an internal_{ref,unref} pair that is execution context aware. Alternatively, make exec_ctx be thread local and 'do the right thing' (whatever that @@ -302,15 +377,15 @@ static void ru_slice_unref(void *p) { } } -static gpr_slice ru_slice_create(grpc_resource_user *resource_user, - size_t size) { +static grpc_slice ru_slice_create(grpc_resource_user *resource_user, + size_t size) { ru_slice_refcount *rc = gpr_malloc(sizeof(ru_slice_refcount) + size); rc->base.ref = ru_slice_ref; rc->base.unref = ru_slice_unref; gpr_ref_init(&rc->refs, 1); rc->resource_user = resource_user; rc->size = size; - gpr_slice slice; + grpc_slice slice; slice.refcount = &rc->base; slice.data.refcounted.bytes = (uint8_t *)(rc + 1); slice.data.refcounted.length = size; @@ -373,9 +448,19 @@ static void ru_post_destructive_reclaimer(grpc_exec_ctx *exec_ctx, void *ru, rulist_add_tail(resource_user, GRPC_RULIST_RECLAIMER_DESTRUCTIVE); } +static void ru_shutdown(grpc_exec_ctx *exec_ctx, void *ru, grpc_error *error) { + grpc_resource_user *resource_user = ru; + grpc_exec_ctx_sched(exec_ctx, resource_user->reclaimers[0], + GRPC_ERROR_CANCELLED, NULL); + grpc_exec_ctx_sched(exec_ctx, resource_user->reclaimers[1], + GRPC_ERROR_CANCELLED, NULL); + resource_user->reclaimers[0] = NULL; + resource_user->reclaimers[1] = NULL; +} + static void ru_destroy(grpc_exec_ctx *exec_ctx, void *ru, grpc_error *error) { grpc_resource_user *resource_user = ru; - GPR_ASSERT(resource_user->allocated == 0); + GPR_ASSERT(gpr_atm_no_barrier_load(&resource_user->refs) == 0); for (int i = 0; i < GRPC_RULIST_COUNT; i++) { rulist_remove(resource_user, (grpc_rulist)i); } @@ -383,13 +468,14 @@ static void ru_destroy(grpc_exec_ctx *exec_ctx, void *ru, grpc_error *error) { GRPC_ERROR_CANCELLED, NULL); grpc_exec_ctx_sched(exec_ctx, resource_user->reclaimers[1], GRPC_ERROR_CANCELLED, NULL); - grpc_exec_ctx_sched(exec_ctx, (grpc_closure *)gpr_atm_no_barrier_load( - &resource_user->on_done_destroy_closure), - GRPC_ERROR_NONE, NULL); if (resource_user->free_pool != 0) { resource_user->resource_quota->free_pool += resource_user->free_pool; rq_step_sched(exec_ctx, resource_user->resource_quota); } + grpc_resource_quota_internal_unref(exec_ctx, resource_user->resource_quota); + gpr_mu_destroy(&resource_user->mu); + gpr_free(resource_user->name); + gpr_free(resource_user); } static void ru_allocated_slices(grpc_exec_ctx *exec_ctx, void *arg, @@ -397,7 +483,7 @@ static void ru_allocated_slices(grpc_exec_ctx *exec_ctx, void *arg, grpc_resource_user_slice_allocator *slice_allocator = arg; if (error == GRPC_ERROR_NONE) { for (size_t i = 0; i < slice_allocator->count; i++) { - gpr_slice_buffer_add_indexed( + grpc_slice_buffer_add_indexed( slice_allocator->dest, ru_slice_create(slice_allocator->resource_user, slice_allocator->length)); } @@ -539,9 +625,9 @@ const grpc_arg_pointer_vtable *grpc_resource_quota_arg_vtable(void) { * grpc_resource_user api */ -void grpc_resource_user_init(grpc_resource_user *resource_user, - grpc_resource_quota *resource_quota, - const char *name) { +grpc_resource_user *grpc_resource_user_create( + grpc_resource_quota *resource_quota, const char *name) { + grpc_resource_user *resource_user = gpr_malloc(sizeof(*resource_user)); resource_user->resource_quota = grpc_resource_quota_internal_ref(resource_quota); grpc_closure_init(&resource_user->allocate_closure, &ru_allocate, @@ -555,12 +641,12 @@ void grpc_resource_user_init(grpc_resource_user *resource_user, grpc_closure_init(&resource_user->destroy_closure, &ru_destroy, resource_user); gpr_mu_init(&resource_user->mu); - resource_user->allocated = 0; + gpr_atm_rel_store(&resource_user->refs, 1); + gpr_atm_rel_store(&resource_user->shutdown, 0); resource_user->free_pool = 0; grpc_closure_list_init(&resource_user->on_allocated); resource_user->allocating = false; resource_user->added_to_free_pool = false; - gpr_atm_no_barrier_store(&resource_user->on_done_destroy_closure, 0); resource_user->reclaimers[0] = NULL; resource_user->reclaimers[1] = NULL; for (int i = 0; i < GRPC_RULIST_COUNT; i++) { @@ -572,56 +658,54 @@ void grpc_resource_user_init(grpc_resource_user *resource_user, gpr_asprintf(&resource_user->name, "anonymous_resource_user_%" PRIxPTR, (intptr_t)resource_user); } + return resource_user; } -void grpc_resource_user_shutdown(grpc_exec_ctx *exec_ctx, - grpc_resource_user *resource_user, - grpc_closure *on_done) { - gpr_mu_lock(&resource_user->mu); - GPR_ASSERT(gpr_atm_no_barrier_load(&resource_user->on_done_destroy_closure) == - 0); - gpr_atm_no_barrier_store(&resource_user->on_done_destroy_closure, - (gpr_atm)on_done); - if (resource_user->allocated == 0) { +static void ru_ref_by(grpc_resource_user *resource_user, gpr_atm amount) { + GPR_ASSERT(amount > 0); + GPR_ASSERT(gpr_atm_no_barrier_fetch_add(&resource_user->refs, amount) != 0); +} + +static void ru_unref_by(grpc_exec_ctx *exec_ctx, + grpc_resource_user *resource_user, gpr_atm amount) { + GPR_ASSERT(amount > 0); + gpr_atm old = gpr_atm_full_fetch_add(&resource_user->refs, -amount); + GPR_ASSERT(old >= amount); + if (old == amount) { grpc_combiner_execute(exec_ctx, resource_user->resource_quota->combiner, &resource_user->destroy_closure, GRPC_ERROR_NONE, false); } - gpr_mu_unlock(&resource_user->mu); } -void grpc_resource_user_destroy(grpc_exec_ctx *exec_ctx, - grpc_resource_user *resource_user) { - grpc_resource_quota_internal_unref(exec_ctx, resource_user->resource_quota); - gpr_mu_destroy(&resource_user->mu); - gpr_free(resource_user->name); +void grpc_resource_user_ref(grpc_resource_user *resource_user) { + ru_ref_by(resource_user, 1); +} + +void grpc_resource_user_unref(grpc_exec_ctx *exec_ctx, + grpc_resource_user *resource_user) { + ru_unref_by(exec_ctx, resource_user, 1); +} + +void grpc_resource_user_shutdown(grpc_exec_ctx *exec_ctx, + grpc_resource_user *resource_user) { + if (gpr_atm_full_fetch_add(&resource_user->shutdown, 1) == 0) { + grpc_combiner_execute(exec_ctx, resource_user->resource_quota->combiner, + grpc_closure_create(ru_shutdown, resource_user), + GRPC_ERROR_NONE, false); + } } void grpc_resource_user_alloc(grpc_exec_ctx *exec_ctx, grpc_resource_user *resource_user, size_t size, grpc_closure *optional_on_done) { gpr_mu_lock(&resource_user->mu); - grpc_closure *on_done_destroy = (grpc_closure *)gpr_atm_no_barrier_load( - &resource_user->on_done_destroy_closure); - if (on_done_destroy != NULL) { - /* already shutdown */ - if (grpc_resource_quota_trace) { - gpr_log(GPR_DEBUG, "RQ %s %s: alloc %" PRIdPTR " after shutdown", - resource_user->resource_quota->name, resource_user->name, size); - } - grpc_exec_ctx_sched( - exec_ctx, optional_on_done, - GRPC_ERROR_CREATE("Buffer pool user is already shutdown"), NULL); - gpr_mu_unlock(&resource_user->mu); - return; - } - resource_user->allocated += (int64_t)size; + ru_ref_by(resource_user, (gpr_atm)size); resource_user->free_pool -= (int64_t)size; if (grpc_resource_quota_trace) { - gpr_log(GPR_DEBUG, "RQ %s %s: alloc %" PRIdPTR "; allocated -> %" PRId64 - ", free_pool -> %" PRId64, + gpr_log(GPR_DEBUG, "RQ %s %s: alloc %" PRIdPTR "; free_pool -> %" PRId64, resource_user->resource_quota->name, resource_user->name, size, - resource_user->allocated, resource_user->free_pool); + resource_user->free_pool); } if (resource_user->free_pool < 0) { grpc_closure_list_append(&resource_user->on_allocated, optional_on_done, @@ -641,15 +725,12 @@ void grpc_resource_user_alloc(grpc_exec_ctx *exec_ctx, void grpc_resource_user_free(grpc_exec_ctx *exec_ctx, grpc_resource_user *resource_user, size_t size) { gpr_mu_lock(&resource_user->mu); - GPR_ASSERT(resource_user->allocated >= (int64_t)size); bool was_zero_or_negative = resource_user->free_pool <= 0; resource_user->free_pool += (int64_t)size; - resource_user->allocated -= (int64_t)size; if (grpc_resource_quota_trace) { - gpr_log(GPR_DEBUG, "RQ %s %s: free %" PRIdPTR "; allocated -> %" PRId64 - ", free_pool -> %" PRId64, + gpr_log(GPR_DEBUG, "RQ %s %s: free %" PRIdPTR "; free_pool -> %" PRId64, resource_user->resource_quota->name, resource_user->name, size, - resource_user->allocated, resource_user->free_pool); + resource_user->free_pool); } bool is_bigger_than_zero = resource_user->free_pool > 0; if (is_bigger_than_zero && was_zero_or_negative && @@ -659,29 +740,23 @@ void grpc_resource_user_free(grpc_exec_ctx *exec_ctx, &resource_user->add_to_free_pool_closure, GRPC_ERROR_NONE, false); } - grpc_closure *on_done_destroy = (grpc_closure *)gpr_atm_no_barrier_load( - &resource_user->on_done_destroy_closure); - if (on_done_destroy != NULL && resource_user->allocated == 0) { - grpc_combiner_execute(exec_ctx, resource_user->resource_quota->combiner, - &resource_user->destroy_closure, GRPC_ERROR_NONE, - false); - } gpr_mu_unlock(&resource_user->mu); + ru_unref_by(exec_ctx, resource_user, (gpr_atm)size); } void grpc_resource_user_post_reclaimer(grpc_exec_ctx *exec_ctx, grpc_resource_user *resource_user, bool destructive, grpc_closure *closure) { - if (gpr_atm_acq_load(&resource_user->on_done_destroy_closure) == 0) { - GPR_ASSERT(resource_user->reclaimers[destructive] == NULL); - resource_user->reclaimers[destructive] = closure; - grpc_combiner_execute(exec_ctx, resource_user->resource_quota->combiner, - &resource_user->post_reclaimer_closure[destructive], - GRPC_ERROR_NONE, false); - } else { + GPR_ASSERT(resource_user->reclaimers[destructive] == NULL); + if (gpr_atm_acq_load(&resource_user->shutdown) > 0) { grpc_exec_ctx_sched(exec_ctx, closure, GRPC_ERROR_CANCELLED, NULL); + return; } + resource_user->reclaimers[destructive] = closure; + grpc_combiner_execute(exec_ctx, resource_user->resource_quota->combiner, + &resource_user->post_reclaimer_closure[destructive], + GRPC_ERROR_NONE, false); } void grpc_resource_user_finish_reclamation(grpc_exec_ctx *exec_ctx, @@ -708,7 +783,7 @@ void grpc_resource_user_slice_allocator_init( void grpc_resource_user_alloc_slices( grpc_exec_ctx *exec_ctx, grpc_resource_user_slice_allocator *slice_allocator, size_t length, - size_t count, gpr_slice_buffer *dest) { + size_t count, grpc_slice_buffer *dest) { slice_allocator->length = length; slice_allocator->count = count; slice_allocator->dest = dest; @@ -716,9 +791,9 @@ void grpc_resource_user_alloc_slices( count * length, &slice_allocator->on_allocated); } -gpr_slice grpc_resource_user_slice_malloc(grpc_exec_ctx *exec_ctx, - grpc_resource_user *resource_user, - size_t size) { +grpc_slice grpc_resource_user_slice_malloc(grpc_exec_ctx *exec_ctx, + grpc_resource_user *resource_user, + size_t size) { grpc_resource_user_alloc(exec_ctx, resource_user, size, NULL); return ru_slice_create(resource_user, size); } diff --git a/src/core/lib/iomgr/resource_quota.h b/src/core/lib/iomgr/resource_quota.h index da68f21a2c..0181fd978b 100644 --- a/src/core/lib/iomgr/resource_quota.h +++ b/src/core/lib/iomgr/resource_quota.h @@ -84,91 +84,15 @@ void grpc_resource_quota_internal_unref(grpc_exec_ctx *exec_ctx, grpc_resource_quota *grpc_resource_quota_from_channel_args( const grpc_channel_args *channel_args); -/* Resource users are kept in (potentially) several intrusive linked lists - at once. These are the list names. */ -typedef enum { - /* Resource users that are waiting for an allocation */ - GRPC_RULIST_AWAITING_ALLOCATION, - /* Resource users that have free memory available for internal reclamation */ - GRPC_RULIST_NON_EMPTY_FREE_POOL, - /* Resource users that have published a benign reclamation is available */ - GRPC_RULIST_RECLAIMER_BENIGN, - /* Resource users that have published a destructive reclamation is - available */ - GRPC_RULIST_RECLAIMER_DESTRUCTIVE, - /* Number of lists: must be last */ - GRPC_RULIST_COUNT -} grpc_rulist; - typedef struct grpc_resource_user grpc_resource_user; -/* Internal linked list pointers for a resource user */ -typedef struct { - grpc_resource_user *next; - grpc_resource_user *prev; -} grpc_resource_user_link; - -struct grpc_resource_user { - /* The quota this resource user consumes from */ - grpc_resource_quota *resource_quota; - - /* Closure to schedule an allocation under the resource quota combiner lock */ - grpc_closure allocate_closure; - /* Closure to publish a non empty free pool under the resource quota combiner - lock */ - grpc_closure add_to_free_pool_closure; - - gpr_mu mu; - /* Total allocated memory outstanding by this resource user in bytes; - always positive */ - int64_t allocated; - /* The amount of memory (in bytes) this user has cached for its own use: to - avoid quota contention, each resource user can keep some memory in - addition to what it is immediately using (e.g., for caching), and the quota - can pull it back under memory pressure. - This value can become negative if more memory has been requested than - existed in the free pool, at which point the quota is consulted to bring - this value non-negative (asynchronously). */ - int64_t free_pool; - /* A list of closures to call once free_pool becomes non-negative - ie when - all outstanding allocations have been granted. */ - grpc_closure_list on_allocated; - /* True if we are currently trying to allocate from the quota, false if not */ - bool allocating; - /* True if we are currently trying to add ourselves to the non-free quota - list, false otherwise */ - bool added_to_free_pool; - - /* Reclaimers: index 0 is the benign reclaimer, 1 is the destructive reclaimer - */ - grpc_closure *reclaimers[2]; - /* Trampoline closures to finish reclamation and re-enter the quota combiner - lock */ - grpc_closure post_reclaimer_closure[2]; - - /* Closure to execute under the quota combiner to de-register and shutdown the - resource user */ - grpc_closure destroy_closure; - /* User supplied closure to call once the user has finished shutting down AND - all outstanding allocations have been freed. Real type is grpc_closure*, - but it's stored as an atomic to avoid a mutex on some fast paths. */ - gpr_atm on_done_destroy_closure; - - /* Links in the various grpc_rulist lists */ - grpc_resource_user_link links[GRPC_RULIST_COUNT]; - - /* The name of this resource user, for debugging/tracing */ - char *name; -}; - -void grpc_resource_user_init(grpc_resource_user *resource_user, - grpc_resource_quota *resource_quota, - const char *name); +grpc_resource_user *grpc_resource_user_create( + grpc_resource_quota *resource_quota, const char *name); +void grpc_resource_user_ref(grpc_resource_user *resource_user); +void grpc_resource_user_unref(grpc_exec_ctx *exec_ctx, + grpc_resource_user *resource_user); void grpc_resource_user_shutdown(grpc_exec_ctx *exec_ctx, - grpc_resource_user *resource_user, - grpc_closure *on_done); -void grpc_resource_user_destroy(grpc_exec_ctx *exec_ctx, - grpc_resource_user *resource_user); + grpc_resource_user *resource_user); /* Allocate from the resource user (and its quota). If optional_on_done is NULL, then allocate immediately. This may push the @@ -203,7 +127,7 @@ typedef struct grpc_resource_user_slice_allocator { /* Number of slices to allocate on the current request */ size_t count; /* Destination for slices to allocate on the current request */ - gpr_slice_buffer *dest; + grpc_slice_buffer *dest; /* Parent resource user */ grpc_resource_user *resource_user; } grpc_resource_user_slice_allocator; @@ -219,11 +143,11 @@ void grpc_resource_user_slice_allocator_init( void grpc_resource_user_alloc_slices( grpc_exec_ctx *exec_ctx, grpc_resource_user_slice_allocator *slice_allocator, size_t length, - size_t count, gpr_slice_buffer *dest); + size_t count, grpc_slice_buffer *dest); /* Allocate one slice of length \a size synchronously. */ -gpr_slice grpc_resource_user_slice_malloc(grpc_exec_ctx *exec_ctx, - grpc_resource_user *resource_user, - size_t size); +grpc_slice grpc_resource_user_slice_malloc(grpc_exec_ctx *exec_ctx, + grpc_resource_user *resource_user, + size_t size); #endif /* GRPC_CORE_LIB_IOMGR_RESOURCE_QUOTA_H */ diff --git a/src/core/lib/iomgr/socket_mutator.c b/src/core/lib/iomgr/socket_mutator.c new file mode 100644 index 0000000000..8b1efb6bab --- /dev/null +++ b/src/core/lib/iomgr/socket_mutator.c @@ -0,0 +1,98 @@ +/* + * + * Copyright 2015, Google Inc. + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are + * met: + * + * * Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * * Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following disclaimer + * in the documentation and/or other materials provided with the + * distribution. + * * Neither the name of Google Inc. nor the names of its + * contributors may be used to endorse or promote products derived from + * this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + */ + +#include "src/core/lib/iomgr/socket_mutator.h" + +#include <grpc/impl/codegen/grpc_types.h> +#include <grpc/support/sync.h> +#include <grpc/support/useful.h> + +void grpc_socket_mutator_init(grpc_socket_mutator *mutator, + const grpc_socket_mutator_vtable *vtable) { + mutator->vtable = vtable; + gpr_ref_init(&mutator->refcount, 1); +} + +grpc_socket_mutator *grpc_socket_mutator_ref(grpc_socket_mutator *mutator) { + gpr_ref(&mutator->refcount); + return mutator; +} + +bool grpc_socket_mutator_mutate_fd(grpc_socket_mutator *mutator, int fd) { + return mutator->vtable->mutate_fd(fd, mutator); +} + +int grpc_socket_mutator_compare(grpc_socket_mutator *a, + grpc_socket_mutator *b) { + int c = GPR_ICMP(a, b); + if (c != 0) { + grpc_socket_mutator *sma = a; + grpc_socket_mutator *smb = b; + c = GPR_ICMP(sma->vtable, smb->vtable); + if (c == 0) { + c = sma->vtable->compare(sma, smb); + } + } + return c; +} + +void grpc_socket_mutator_unref(grpc_socket_mutator *mutator) { + if (gpr_unref(&mutator->refcount)) { + mutator->vtable->destory(mutator); + } +} + +static void *socket_mutator_arg_copy(void *p) { + return grpc_socket_mutator_ref(p); +} + +static void socket_mutator_arg_destroy(void *p) { + grpc_socket_mutator_unref(p); +} + +static int socket_mutator_cmp(void *a, void *b) { + return grpc_socket_mutator_compare((grpc_socket_mutator *)a, + (grpc_socket_mutator *)b); +} + +static const grpc_arg_pointer_vtable socket_mutator_arg_vtable = { + socket_mutator_arg_copy, socket_mutator_arg_destroy, socket_mutator_cmp}; + +grpc_arg grpc_socket_mutator_to_arg(grpc_socket_mutator *mutator) { + grpc_arg arg; + arg.type = GRPC_ARG_POINTER; + arg.key = GRPC_ARG_SOCKET_MUTATOR; + arg.value.pointer.vtable = &socket_mutator_arg_vtable; + arg.value.pointer.p = mutator; + return arg; +} diff --git a/src/core/lib/iomgr/socket_mutator.h b/src/core/lib/iomgr/socket_mutator.h new file mode 100644 index 0000000000..2f5b6c248e --- /dev/null +++ b/src/core/lib/iomgr/socket_mutator.h @@ -0,0 +1,80 @@ +/* + * + * 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_SOCKET_MUTATOR_H +#define GRPC_CORE_LIB_IOMGR_SOCKET_MUTATOR_H + +#include <grpc/impl/codegen/grpc_types.h> +#include <grpc/support/sync.h> + +#ifdef __cplusplus +extern "C" { +#endif + +/** The virtual table of grpc_socket_mutator */ +typedef struct { + /** Mutates the socket opitons of \a fd */ + bool (*mutate_fd)(int fd, grpc_socket_mutator *mutator); + /** Compare socket mutator \a a and \a b */ + int (*compare)(grpc_socket_mutator *a, grpc_socket_mutator *b); + /** Destroys the socket mutator instance */ + void (*destory)(grpc_socket_mutator *mutator); +} grpc_socket_mutator_vtable; + +/** The Socket Mutator interface allows changes on socket options */ +struct grpc_socket_mutator { + const grpc_socket_mutator_vtable *vtable; + gpr_refcount refcount; +}; + +/** called by concrete implementations to initialize the base struct */ +void grpc_socket_mutator_init(grpc_socket_mutator *mutator, + const grpc_socket_mutator_vtable *vtable); + +/** Wrap \a mutator as a grpc_arg */ +grpc_arg grpc_socket_mutator_to_arg(grpc_socket_mutator *mutator); + +/** Perform the file descriptor mutation operation of \a mutator on \a fd */ +bool grpc_socket_mutator_mutate_fd(grpc_socket_mutator *mutator, int fd); + +/** Compare if \a a and \a b are the same mutator or have same settings */ +int grpc_socket_mutator_compare(grpc_socket_mutator *a, grpc_socket_mutator *b); + +grpc_socket_mutator *grpc_socket_mutator_ref(grpc_socket_mutator *mutator); +void grpc_socket_mutator_unref(grpc_socket_mutator *mutator); + +#ifdef __cplusplus +} +#endif + +#endif /* GRPC_CORE_LIB_IOMGR_SOCKET_MUTATOR_H */ diff --git a/src/core/lib/iomgr/socket_utils_common_posix.c b/src/core/lib/iomgr/socket_utils_common_posix.c index bc28bbe316..88e9ade253 100644 --- a/src/core/lib/iomgr/socket_utils_common_posix.c +++ b/src/core/lib/iomgr/socket_utils_common_posix.c @@ -209,6 +209,15 @@ grpc_error *grpc_set_socket_low_latency(int fd, int low_latency) { return GRPC_ERROR_NONE; } +/* set a socket using a grpc_socket_mutator */ +grpc_error *grpc_set_socket_with_mutator(int fd, grpc_socket_mutator *mutator) { + GPR_ASSERT(mutator); + if (!grpc_socket_mutator_mutate_fd(mutator, fd)) { + return GRPC_ERROR_CREATE("grpc_socket_mutator failed."); + } + return GRPC_ERROR_NONE; +} + static gpr_once g_probe_ipv6_once = GPR_ONCE_INIT; static int g_ipv6_loopback_available; diff --git a/src/core/lib/iomgr/socket_utils_posix.h b/src/core/lib/iomgr/socket_utils_posix.h index 175fb2b717..e84d3781a1 100644 --- a/src/core/lib/iomgr/socket_utils_posix.h +++ b/src/core/lib/iomgr/socket_utils_posix.h @@ -39,7 +39,9 @@ #include <sys/socket.h> #include <unistd.h> +#include <grpc/impl/codegen/grpc_types.h> #include "src/core/lib/iomgr/error.h" +#include "src/core/lib/iomgr/socket_mutator.h" /* a wrapper for accept or accept4 */ int grpc_accept4(int sockfd, grpc_resolved_address *resolved_addr, int nonblock, @@ -88,6 +90,9 @@ grpc_error *grpc_set_socket_sndbuf(int fd, int buffer_size_bytes); /* Tries to set the socket's receive buffer to given size. */ grpc_error *grpc_set_socket_rcvbuf(int fd, int buffer_size_bytes); +/* Tries to set the socket using a grpc_socket_mutator */ +grpc_error *grpc_set_socket_with_mutator(int fd, grpc_socket_mutator *mutator); + /* An enum to keep track of IPv4/IPv6 socket modes. Currently, this information is only used when a socket is first created, but diff --git a/src/core/lib/iomgr/tcp_client.h b/src/core/lib/iomgr/tcp_client.h index 18e6e60ebc..0485661316 100644 --- a/src/core/lib/iomgr/tcp_client.h +++ b/src/core/lib/iomgr/tcp_client.h @@ -34,6 +34,7 @@ #ifndef GRPC_CORE_LIB_IOMGR_TCP_CLIENT_H #define GRPC_CORE_LIB_IOMGR_TCP_CLIENT_H +#include <grpc/impl/codegen/grpc_types.h> #include <grpc/support/time.h> #include "src/core/lib/iomgr/endpoint.h" #include "src/core/lib/iomgr/pollset_set.h" diff --git a/src/core/lib/iomgr/tcp_client_posix.c b/src/core/lib/iomgr/tcp_client_posix.c index bc08c94ee0..a3a70a8ed7 100644 --- a/src/core/lib/iomgr/tcp_client_posix.c +++ b/src/core/lib/iomgr/tcp_client_posix.c @@ -51,6 +51,7 @@ #include "src/core/lib/iomgr/ev_posix.h" #include "src/core/lib/iomgr/iomgr_posix.h" #include "src/core/lib/iomgr/sockaddr_utils.h" +#include "src/core/lib/iomgr/socket_mutator.h" #include "src/core/lib/iomgr/socket_utils_posix.h" #include "src/core/lib/iomgr/tcp_posix.h" #include "src/core/lib/iomgr/timer.h" @@ -73,7 +74,8 @@ typedef struct { grpc_channel_args *channel_args; } async_connect; -static grpc_error *prepare_socket(const grpc_resolved_address *addr, int fd) { +static grpc_error *prepare_socket(const grpc_resolved_address *addr, int fd, + const grpc_channel_args *channel_args) { grpc_error *err = GRPC_ERROR_NONE; GPR_ASSERT(fd >= 0); @@ -88,6 +90,16 @@ static grpc_error *prepare_socket(const grpc_resolved_address *addr, int fd) { } err = grpc_set_socket_no_sigpipe_if_possible(fd); if (err != GRPC_ERROR_NONE) goto error; + if (channel_args) { + for (size_t i = 0; i < channel_args->num_args; i++) { + if (0 == strcmp(channel_args->args[i].key, GRPC_ARG_SOCKET_MUTATOR)) { + GPR_ASSERT(channel_args->args[i].type == GRPC_ARG_POINTER); + grpc_socket_mutator *mutator = channel_args->args[i].value.pointer.p; + err = grpc_set_socket_with_mutator(fd, mutator); + if (err != GRPC_ERROR_NONE) goto error; + } + } + } goto done; error: @@ -239,8 +251,11 @@ finish: done = (--ac->refs == 0); gpr_mu_unlock(&ac->mu); if (error != GRPC_ERROR_NONE) { - error = grpc_error_set_str(error, GRPC_ERROR_STR_DESCRIPTION, - "Failed to connect to remote host"); + char *error_descr; + gpr_asprintf(&error_descr, "Failed to connect to remote host: %s", + grpc_error_get_str(error, GRPC_ERROR_STR_DESCRIPTION)); + error = grpc_error_set_str(error, GRPC_ERROR_STR_DESCRIPTION, error_descr); + gpr_free(error_descr); error = grpc_error_set_str(error, GRPC_ERROR_STR_TARGET_ADDRESS, ac->addr_str); } @@ -287,7 +302,7 @@ static void tcp_client_connect_impl(grpc_exec_ctx *exec_ctx, GPR_ASSERT(grpc_sockaddr_is_v4mapped(addr, &addr4_copy)); addr = &addr4_copy; } - if ((error = prepare_socket(addr, fd)) != GRPC_ERROR_NONE) { + if ((error = prepare_socket(addr, fd, channel_args)) != GRPC_ERROR_NONE) { grpc_exec_ctx_sched(exec_ctx, closure, error, NULL); return; } diff --git a/src/core/lib/iomgr/tcp_client_windows.c b/src/core/lib/iomgr/tcp_client_windows.c index 30f7c66f15..4d1e809872 100644 --- a/src/core/lib/iomgr/tcp_client_windows.c +++ b/src/core/lib/iomgr/tcp_client_windows.c @@ -37,10 +37,10 @@ #include "src/core/lib/iomgr/sockaddr_windows.h" +#include <grpc/slice_buffer.h> #include <grpc/support/alloc.h> #include <grpc/support/log.h> #include <grpc/support/log_windows.h> -#include <grpc/support/slice_buffer.h> #include <grpc/support/useful.h> #include "src/core/lib/channel/channel_args.h" diff --git a/src/core/lib/iomgr/tcp_posix.c b/src/core/lib/iomgr/tcp_posix.c index 880af93ee1..12a4797e6f 100644 --- a/src/core/lib/iomgr/tcp_posix.c +++ b/src/core/lib/iomgr/tcp_posix.c @@ -46,9 +46,9 @@ #include <sys/types.h> #include <unistd.h> +#include <grpc/slice.h> #include <grpc/support/alloc.h> #include <grpc/support/log.h> -#include <grpc/support/slice.h> #include <grpc/support/string_util.h> #include <grpc/support/sync.h> #include <grpc/support/time.h> @@ -56,6 +56,7 @@ #include "src/core/lib/debug/trace.h" #include "src/core/lib/iomgr/ev_posix.h" #include "src/core/lib/profiling/timers.h" +#include "src/core/lib/slice/slice_string_helpers.h" #include "src/core/lib/support/string.h" #ifdef GRPC_HAVE_MSG_NOSIGNAL @@ -83,10 +84,10 @@ typedef struct { gpr_atm shutdown_count; /* garbage after the last read */ - gpr_slice_buffer last_read_buffer; + grpc_slice_buffer last_read_buffer; - gpr_slice_buffer *incoming_buffer; - gpr_slice_buffer *outgoing_buffer; + grpc_slice_buffer *incoming_buffer; + grpc_slice_buffer *outgoing_buffer; /** slice within outgoing_buffer to write next */ size_t outgoing_slice_idx; /** byte within outgoing_buffer->slices[outgoing_slice_idx] to write next */ @@ -102,7 +103,7 @@ typedef struct { char *peer_string; - grpc_resource_user resource_user; + grpc_resource_user *resource_user; grpc_resource_user_slice_allocator slice_allocator; } grpc_tcp; @@ -110,28 +111,18 @@ static void tcp_handle_read(grpc_exec_ctx *exec_ctx, void *arg /* grpc_tcp */, grpc_error *error); static void tcp_handle_write(grpc_exec_ctx *exec_ctx, void *arg /* grpc_tcp */, grpc_error *error); -static void tcp_unref_closure(grpc_exec_ctx *exec_ctx, void *arg /* grpc_tcp */, - grpc_error *error); - -static void tcp_maybe_shutdown_resource_user(grpc_exec_ctx *exec_ctx, - grpc_tcp *tcp) { - if (gpr_atm_full_fetch_add(&tcp->shutdown_count, 1) == 0) { - grpc_resource_user_shutdown(exec_ctx, &tcp->resource_user, - grpc_closure_create(tcp_unref_closure, tcp)); - } -} static void tcp_shutdown(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep) { grpc_tcp *tcp = (grpc_tcp *)ep; - tcp_maybe_shutdown_resource_user(exec_ctx, tcp); grpc_fd_shutdown(exec_ctx, tcp->em_fd); + grpc_resource_user_shutdown(exec_ctx, tcp->resource_user); } static void tcp_free(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) { grpc_fd_orphan(exec_ctx, tcp->em_fd, tcp->release_fd_cb, tcp->release_fd, "tcp_unref_orphan"); - gpr_slice_buffer_destroy(&tcp->last_read_buffer); - grpc_resource_user_destroy(exec_ctx, &tcp->resource_user); + grpc_slice_buffer_destroy(&tcp->last_read_buffer); + grpc_resource_user_unref(exec_ctx, tcp->resource_user); gpr_free(tcp->peer_string); gpr_free(tcp); } @@ -168,16 +159,10 @@ static void tcp_unref(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) { static void tcp_ref(grpc_tcp *tcp) { gpr_ref(&tcp->refcount); } #endif -static void tcp_unref_closure(grpc_exec_ctx *exec_ctx, void *arg, - grpc_error *error) { - TCP_UNREF(exec_ctx, arg, "resource_user"); -} - static void tcp_destroy(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep) { grpc_network_status_unregister_endpoint(ep); grpc_tcp *tcp = (grpc_tcp *)ep; - tcp_maybe_shutdown_resource_user(exec_ctx, tcp); - gpr_slice_buffer_reset_and_unref(&tcp->last_read_buffer); + grpc_slice_buffer_reset_and_unref(&tcp->last_read_buffer); TCP_UNREF(exec_ctx, tcp, "destroy"); } @@ -191,8 +176,8 @@ static void call_read_cb(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp, gpr_log(GPR_DEBUG, "read: error=%s", str); grpc_error_free_string(str); for (i = 0; i < tcp->incoming_buffer->count; i++) { - char *dump = gpr_dump_slice(tcp->incoming_buffer->slices[i], - GPR_DUMP_HEX | GPR_DUMP_ASCII); + char *dump = grpc_dump_slice(tcp->incoming_buffer->slices[i], + GPR_DUMP_HEX | GPR_DUMP_ASCII); gpr_log(GPR_DEBUG, "READ %p (peer=%s): %s", tcp, tcp->peer_string, dump); gpr_free(dump); } @@ -216,8 +201,8 @@ static void tcp_do_read(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) { GPR_TIMER_BEGIN("tcp_continue_read", 0); for (i = 0; i < tcp->incoming_buffer->count; i++) { - iov[i].iov_base = GPR_SLICE_START_PTR(tcp->incoming_buffer->slices[i]); - iov[i].iov_len = GPR_SLICE_LENGTH(tcp->incoming_buffer->slices[i]); + iov[i].iov_base = GRPC_SLICE_START_PTR(tcp->incoming_buffer->slices[i]); + iov[i].iov_len = GRPC_SLICE_LENGTH(tcp->incoming_buffer->slices[i]); } msg.msg_name = NULL; @@ -244,19 +229,19 @@ static void tcp_do_read(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) { /* We've consumed the edge, request a new one */ grpc_fd_notify_on_read(exec_ctx, tcp->em_fd, &tcp->read_closure); } else { - gpr_slice_buffer_reset_and_unref(tcp->incoming_buffer); + grpc_slice_buffer_reset_and_unref(tcp->incoming_buffer); call_read_cb(exec_ctx, tcp, GRPC_OS_ERROR(errno, "recvmsg")); TCP_UNREF(exec_ctx, tcp, "read"); } } else if (read_bytes == 0) { /* 0 read size ==> end of stream */ - gpr_slice_buffer_reset_and_unref(tcp->incoming_buffer); + grpc_slice_buffer_reset_and_unref(tcp->incoming_buffer); call_read_cb(exec_ctx, tcp, GRPC_ERROR_CREATE("Socket closed")); TCP_UNREF(exec_ctx, tcp, "read"); } else { GPR_ASSERT((size_t)read_bytes <= tcp->incoming_buffer->length); if ((size_t)read_bytes < tcp->incoming_buffer->length) { - gpr_slice_buffer_trim_end( + grpc_slice_buffer_trim_end( tcp->incoming_buffer, tcp->incoming_buffer->length - (size_t)read_bytes, &tcp->last_read_buffer); @@ -275,8 +260,8 @@ static void tcp_read_allocation_done(grpc_exec_ctx *exec_ctx, void *tcpp, grpc_error *error) { grpc_tcp *tcp = tcpp; if (error != GRPC_ERROR_NONE) { - gpr_slice_buffer_reset_and_unref(tcp->incoming_buffer); - gpr_slice_buffer_reset_and_unref(&tcp->last_read_buffer); + grpc_slice_buffer_reset_and_unref(tcp->incoming_buffer); + grpc_slice_buffer_reset_and_unref(&tcp->last_read_buffer); call_read_cb(exec_ctx, tcp, GRPC_ERROR_REF(error)); TCP_UNREF(exec_ctx, tcp, "read"); } else { @@ -301,8 +286,8 @@ static void tcp_handle_read(grpc_exec_ctx *exec_ctx, void *arg /* grpc_tcp */, GPR_ASSERT(!tcp->finished_edge); if (error != GRPC_ERROR_NONE) { - gpr_slice_buffer_reset_and_unref(tcp->incoming_buffer); - gpr_slice_buffer_reset_and_unref(&tcp->last_read_buffer); + grpc_slice_buffer_reset_and_unref(tcp->incoming_buffer); + grpc_slice_buffer_reset_and_unref(&tcp->last_read_buffer); call_read_cb(exec_ctx, tcp, GRPC_ERROR_REF(error)); TCP_UNREF(exec_ctx, tcp, "read"); } else { @@ -311,13 +296,13 @@ static void tcp_handle_read(grpc_exec_ctx *exec_ctx, void *arg /* grpc_tcp */, } static void tcp_read(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, - gpr_slice_buffer *incoming_buffer, grpc_closure *cb) { + grpc_slice_buffer *incoming_buffer, grpc_closure *cb) { grpc_tcp *tcp = (grpc_tcp *)ep; GPR_ASSERT(tcp->read_cb == NULL); tcp->read_cb = cb; tcp->incoming_buffer = incoming_buffer; - gpr_slice_buffer_reset_and_unref(incoming_buffer); - gpr_slice_buffer_swap(incoming_buffer, &tcp->last_read_buffer); + grpc_slice_buffer_reset_and_unref(incoming_buffer); + grpc_slice_buffer_swap(incoming_buffer, &tcp->last_read_buffer); TCP_REF(tcp, "read"); if (tcp->finished_edge) { tcp->finished_edge = false; @@ -347,11 +332,11 @@ static bool tcp_flush(grpc_tcp *tcp, grpc_error **error) { iov_size != MAX_WRITE_IOVEC; iov_size++) { iov[iov_size].iov_base = - GPR_SLICE_START_PTR( + GRPC_SLICE_START_PTR( tcp->outgoing_buffer->slices[tcp->outgoing_slice_idx]) + tcp->outgoing_byte_idx; iov[iov_size].iov_len = - GPR_SLICE_LENGTH( + GRPC_SLICE_LENGTH( tcp->outgoing_buffer->slices[tcp->outgoing_slice_idx]) - tcp->outgoing_byte_idx; sending_length += iov[iov_size].iov_len; @@ -392,7 +377,7 @@ static bool tcp_flush(grpc_tcp *tcp, grpc_error **error) { size_t slice_length; tcp->outgoing_slice_idx--; - slice_length = GPR_SLICE_LENGTH( + slice_length = GRPC_SLICE_LENGTH( tcp->outgoing_buffer->slices[tcp->outgoing_slice_idx]); if (slice_length > trailing) { tcp->outgoing_byte_idx = slice_length - trailing; @@ -442,7 +427,7 @@ static void tcp_handle_write(grpc_exec_ctx *exec_ctx, void *arg /* grpc_tcp */, } static void tcp_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, - gpr_slice_buffer *buf, grpc_closure *cb) { + grpc_slice_buffer *buf, grpc_closure *cb) { grpc_tcp *tcp = (grpc_tcp *)ep; grpc_error *error = GRPC_ERROR_NONE; @@ -451,7 +436,7 @@ static void tcp_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, for (i = 0; i < buf->count; i++) { char *data = - gpr_dump_slice(buf->slices[i], GPR_DUMP_HEX | GPR_DUMP_ASCII); + grpc_dump_slice(buf->slices[i], GPR_DUMP_HEX | GPR_DUMP_ASCII); gpr_log(GPR_DEBUG, "WRITE %p (peer=%s): %s", tcp, tcp->peer_string, data); gpr_free(data); } @@ -508,6 +493,11 @@ static char *tcp_get_peer(grpc_endpoint *ep) { return gpr_strdup(tcp->peer_string); } +static int tcp_get_fd(grpc_endpoint *ep) { + grpc_tcp *tcp = (grpc_tcp *)ep; + return tcp->fd; +} + static grpc_workqueue *tcp_get_workqueue(grpc_endpoint *ep) { grpc_tcp *tcp = (grpc_tcp *)ep; return grpc_fd_get_workqueue(tcp->em_fd); @@ -515,7 +505,7 @@ static grpc_workqueue *tcp_get_workqueue(grpc_endpoint *ep) { static grpc_resource_user *tcp_get_resource_user(grpc_endpoint *ep) { grpc_tcp *tcp = (grpc_tcp *)ep; - return &tcp->resource_user; + return tcp->resource_user; } static const grpc_endpoint_vtable vtable = {tcp_read, @@ -526,7 +516,8 @@ static const grpc_endpoint_vtable vtable = {tcp_read, tcp_shutdown, tcp_destroy, tcp_get_resource_user, - tcp_get_peer}; + tcp_get_peer, + tcp_get_fd}; grpc_endpoint *grpc_tcp_create(grpc_fd *em_fd, grpc_resource_quota *resource_quota, @@ -543,20 +534,18 @@ grpc_endpoint *grpc_tcp_create(grpc_fd *em_fd, tcp->slice_size = slice_size; tcp->iov_size = 1; tcp->finished_edge = true; - /* paired with unref in grpc_tcp_destroy, and with the shutdown for our - * resource_user */ - gpr_ref_init(&tcp->refcount, 2); + /* paired with unref in grpc_tcp_destroy */ + gpr_ref_init(&tcp->refcount, 1); gpr_atm_no_barrier_store(&tcp->shutdown_count, 0); tcp->em_fd = em_fd; tcp->read_closure.cb = tcp_handle_read; tcp->read_closure.cb_arg = tcp; tcp->write_closure.cb = tcp_handle_write; tcp->write_closure.cb_arg = tcp; - gpr_slice_buffer_init(&tcp->last_read_buffer); - grpc_resource_user_init(&tcp->resource_user, resource_quota, peer_string); - grpc_resource_user_slice_allocator_init(&tcp->slice_allocator, - &tcp->resource_user, - tcp_read_allocation_done, tcp); + grpc_slice_buffer_init(&tcp->last_read_buffer); + tcp->resource_user = grpc_resource_user_create(resource_quota, peer_string); + grpc_resource_user_slice_allocator_init( + &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); @@ -576,8 +565,7 @@ void grpc_tcp_destroy_and_release_fd(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, GPR_ASSERT(ep->vtable == &vtable); tcp->release_fd = fd; tcp->release_fd_cb = done; - tcp_maybe_shutdown_resource_user(exec_ctx, tcp); - gpr_slice_buffer_reset_and_unref(&tcp->last_read_buffer); + grpc_slice_buffer_reset_and_unref(&tcp->last_read_buffer); TCP_UNREF(exec_ctx, tcp, "destroy"); } diff --git a/src/core/lib/iomgr/tcp_uv.c b/src/core/lib/iomgr/tcp_uv.c index 8e74c9e863..6e2ad1dbe9 100644 --- a/src/core/lib/iomgr/tcp_uv.c +++ b/src/core/lib/iomgr/tcp_uv.c @@ -38,14 +38,17 @@ #include <limits.h> #include <string.h> +#include <grpc/slice_buffer.h> + #include <grpc/support/alloc.h> #include <grpc/support/log.h> -#include <grpc/support/slice_buffer.h> #include <grpc/support/string_util.h> #include "src/core/lib/iomgr/error.h" #include "src/core/lib/iomgr/network_status_tracker.h" +#include "src/core/lib/iomgr/resource_quota.h" #include "src/core/lib/iomgr/tcp_uv.h" +#include "src/core/lib/slice/slice_string_helpers.h" #include "src/core/lib/support/string.h" int grpc_tcp_trace = 0; @@ -62,15 +65,14 @@ typedef struct { grpc_closure *read_cb; grpc_closure *write_cb; - gpr_slice read_slice; - gpr_slice_buffer *read_slices; - gpr_slice_buffer *write_slices; + grpc_slice read_slice; + grpc_slice_buffer *read_slices; + grpc_slice_buffer *write_slices; uv_buf_t *write_buffers; - grpc_resource_user resource_user; + grpc_resource_user *resource_user; bool shutting_down; - bool resource_user_shutting_down; char *peer_string; grpc_pollset *pollset; @@ -78,23 +80,23 @@ typedef struct { static void uv_close_callback(uv_handle_t *handle) { gpr_free(handle); } -static void tcp_free(grpc_tcp *tcp) { - grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT; - grpc_resource_user_destroy(&exec_ctx, &tcp->resource_user); +static void tcp_free(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) { + grpc_resource_user_unref(exec_ctx, tcp->resource_user); gpr_free(tcp); - grpc_exec_ctx_finish(&exec_ctx); } /*#define GRPC_TCP_REFCOUNT_DEBUG*/ #ifdef GRPC_TCP_REFCOUNT_DEBUG -#define TCP_UNREF(tcp, reason) tcp_unref((tcp), (reason), __FILE__, __LINE__) -#define TCP_REF(tcp, reason) tcp_ref((tcp), (reason), __FILE__, __LINE__) -static void tcp_unref(grpc_tcp *tcp, const char *reason, const char *file, - int line) { +#define TCP_UNREF(exec_ctx, tcp, reason) \ + tcp_unref((exec_ctx), (tcp), (reason), __FILE__, __LINE__) +#define TCP_REF(tcp, reason) \ + tcp_ref((exec_ctx), (tcp), (reason), __FILE__, __LINE__) +static void tcp_unref(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp, + const char *reason, const char *file, int line) { gpr_log(file, line, GPR_LOG_SEVERITY_DEBUG, "TCP unref %p : %s %d -> %d", tcp, reason, tcp->refcount.count, tcp->refcount.count - 1); if (gpr_unref(&tcp->refcount)) { - tcp_free(tcp); + tcp_free(exec_ctx, tcp); } } @@ -105,11 +107,11 @@ static void tcp_ref(grpc_tcp *tcp, const char *reason, const char *file, gpr_ref(&tcp->refcount); } #else -#define TCP_UNREF(tcp, reason) tcp_unref((tcp)) +#define TCP_UNREF(exec_ctx, tcp, reason) tcp_unref((exec_ctx), (tcp)) #define TCP_REF(tcp, reason) tcp_ref((tcp)) -static void tcp_unref(grpc_tcp *tcp) { +static void tcp_unref(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) { if (gpr_unref(&tcp->refcount)) { - tcp_free(tcp); + tcp_free(exec_ctx, tcp); } } @@ -122,15 +124,15 @@ static void alloc_uv_buf(uv_handle_t *handle, size_t suggested_size, grpc_tcp *tcp = handle->data; (void)suggested_size; tcp->read_slice = grpc_resource_user_slice_malloc( - &exec_ctx, &tcp->resource_user, GRPC_TCP_DEFAULT_READ_SLICE_SIZE); - buf->base = (char *)GPR_SLICE_START_PTR(tcp->read_slice); - buf->len = GPR_SLICE_LENGTH(tcp->read_slice); + &exec_ctx, tcp->resource_user, GRPC_TCP_DEFAULT_READ_SLICE_SIZE); + buf->base = (char *)GRPC_SLICE_START_PTR(tcp->read_slice); + buf->len = GRPC_SLICE_LENGTH(tcp->read_slice); grpc_exec_ctx_finish(&exec_ctx); } static void read_callback(uv_stream_t *stream, ssize_t nread, const uv_buf_t *buf) { - gpr_slice sub; + grpc_slice sub; grpc_error *error; grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT; grpc_tcp *tcp = stream->data; @@ -139,7 +141,7 @@ static void read_callback(uv_stream_t *stream, ssize_t nread, // Nothing happened. Wait for the next callback return; } - TCP_UNREF(tcp, "read"); + TCP_UNREF(&exec_ctx, tcp, "read"); tcp->read_cb = NULL; // TODO(murgatroid99): figure out what the return value here means uv_read_stop(stream); @@ -147,8 +149,8 @@ static void read_callback(uv_stream_t *stream, ssize_t nread, error = GRPC_ERROR_CREATE("EOF"); } else if (nread > 0) { // Successful read - sub = gpr_slice_sub_no_ref(tcp->read_slice, 0, (size_t)nread); - gpr_slice_buffer_add(tcp->read_slices, sub); + sub = grpc_slice_sub_no_ref(tcp->read_slice, 0, (size_t)nread); + grpc_slice_buffer_add(tcp->read_slices, sub); error = GRPC_ERROR_NONE; if (grpc_tcp_trace) { size_t i; @@ -156,8 +158,8 @@ static void read_callback(uv_stream_t *stream, ssize_t nread, gpr_log(GPR_DEBUG, "read: error=%s", str); grpc_error_free_string(str); for (i = 0; i < tcp->read_slices->count; i++) { - char *dump = gpr_dump_slice(tcp->read_slices->slices[i], - GPR_DUMP_HEX | GPR_DUMP_ASCII); + char *dump = grpc_dump_slice(tcp->read_slices->slices[i], + GPR_DUMP_HEX | GPR_DUMP_ASCII); gpr_log(GPR_DEBUG, "READ %p (peer=%s): %s", tcp, tcp->peer_string, dump); gpr_free(dump); @@ -172,14 +174,14 @@ static void read_callback(uv_stream_t *stream, ssize_t nread, } static void uv_endpoint_read(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, - gpr_slice_buffer *read_slices, grpc_closure *cb) { + grpc_slice_buffer *read_slices, grpc_closure *cb) { grpc_tcp *tcp = (grpc_tcp *)ep; int status; grpc_error *error = GRPC_ERROR_NONE; GPR_ASSERT(tcp->read_cb == NULL); tcp->read_cb = cb; tcp->read_slices = read_slices; - gpr_slice_buffer_reset_and_unref(read_slices); + grpc_slice_buffer_reset_and_unref(read_slices); TCP_REF(tcp, "read"); // TODO(murgatroid99): figure out what the return value here means status = @@ -202,7 +204,7 @@ static void write_callback(uv_write_t *req, int status) { grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT; grpc_closure *cb = tcp->write_cb; tcp->write_cb = NULL; - TCP_UNREF(tcp, "write"); + TCP_UNREF(&exec_ctx, tcp, "write"); if (status == 0) { error = GRPC_ERROR_NONE; } else { @@ -213,28 +215,28 @@ static void write_callback(uv_write_t *req, int status) { gpr_log(GPR_DEBUG, "write complete on %p: error=%s", tcp, str); } gpr_free(tcp->write_buffers); - grpc_resource_user_free(&exec_ctx, &tcp->resource_user, + grpc_resource_user_free(&exec_ctx, tcp->resource_user, sizeof(uv_buf_t) * tcp->write_slices->count); grpc_exec_ctx_sched(&exec_ctx, cb, error, NULL); grpc_exec_ctx_finish(&exec_ctx); } static void uv_endpoint_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, - gpr_slice_buffer *write_slices, + grpc_slice_buffer *write_slices, grpc_closure *cb) { grpc_tcp *tcp = (grpc_tcp *)ep; uv_buf_t *buffers; unsigned int buffer_count; unsigned int i; - gpr_slice *slice; + grpc_slice *slice; uv_write_t *write_req; if (grpc_tcp_trace) { size_t j; for (j = 0; j < write_slices->count; j++) { - char *data = gpr_dump_slice(write_slices->slices[j], - GPR_DUMP_HEX | GPR_DUMP_ASCII); + char *data = grpc_dump_slice(write_slices->slices[j], + GPR_DUMP_HEX | GPR_DUMP_ASCII); gpr_log(GPR_DEBUG, "WRITE %p (peer=%s): %s", tcp, tcp->peer_string, data); gpr_free(data); } @@ -259,12 +261,12 @@ static void uv_endpoint_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, tcp->write_cb = cb; buffer_count = (unsigned int)tcp->write_slices->count; buffers = gpr_malloc(sizeof(uv_buf_t) * buffer_count); - grpc_resource_user_alloc(exec_ctx, &tcp->resource_user, + grpc_resource_user_alloc(exec_ctx, tcp->resource_user, sizeof(uv_buf_t) * buffer_count, NULL); for (i = 0; i < buffer_count; i++) { slice = &tcp->write_slices->slices[i]; - buffers[i].base = (char *)GPR_SLICE_START_PTR(*slice); - buffers[i].len = GPR_SLICE_LENGTH(*slice); + buffers[i].base = (char *)GRPC_SLICE_START_PTR(*slice); + buffers[i].len = GRPC_SLICE_LENGTH(*slice); } tcp->write_buffers = buffers; write_req = &tcp->write_req; @@ -295,22 +297,6 @@ static void uv_add_to_pollset_set(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, static void shutdown_callback(uv_shutdown_t *req, int status) {} -static void resource_user_shutdown_done(grpc_exec_ctx *exec_ctx, void *arg, - grpc_error *error) { - TCP_UNREF(arg, "resource_user"); -} - -static void uv_resource_user_maybe_shutdown(grpc_exec_ctx *exec_ctx, - grpc_tcp *tcp) { - if (!tcp->resource_user_shutting_down) { - tcp->resource_user_shutting_down = true; - TCP_REF(tcp, "resource_user"); - grpc_resource_user_shutdown( - exec_ctx, &tcp->resource_user, - grpc_closure_create(resource_user_shutdown_done, tcp)); - } -} - static void uv_endpoint_shutdown(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep) { grpc_tcp *tcp = (grpc_tcp *)ep; if (!tcp->shutting_down) { @@ -324,8 +310,7 @@ static void uv_destroy(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep) { grpc_network_status_unregister_endpoint(ep); grpc_tcp *tcp = (grpc_tcp *)ep; uv_close((uv_handle_t *)tcp->handle, uv_close_callback); - uv_resource_user_maybe_shutdown(exec_ctx, tcp); - TCP_UNREF(tcp, "destroy"); + TCP_UNREF(exec_ctx, tcp, "destroy"); } static char *uv_get_peer(grpc_endpoint *ep) { @@ -335,15 +320,18 @@ static char *uv_get_peer(grpc_endpoint *ep) { static grpc_resource_user *uv_get_resource_user(grpc_endpoint *ep) { grpc_tcp *tcp = (grpc_tcp *)ep; - return &tcp->resource_user; + return tcp->resource_user; } static grpc_workqueue *uv_get_workqueue(grpc_endpoint *ep) { return NULL; } +static int uv_get_fd(grpc_endpoint *ep) { return -1; } + static grpc_endpoint_vtable vtable = { uv_endpoint_read, uv_endpoint_write, uv_get_workqueue, uv_add_to_pollset, uv_add_to_pollset_set, uv_endpoint_shutdown, - uv_destroy, uv_get_resource_user, uv_get_peer}; + uv_destroy, uv_get_resource_user, uv_get_peer, + uv_get_fd}; grpc_endpoint *grpc_tcp_create(uv_tcp_t *handle, grpc_resource_quota *resource_quota, @@ -364,8 +352,7 @@ grpc_endpoint *grpc_tcp_create(uv_tcp_t *handle, gpr_ref_init(&tcp->refcount, 1); tcp->peer_string = gpr_strdup(peer_string); tcp->shutting_down = false; - tcp->resource_user_shutting_down = false; - grpc_resource_user_init(&tcp->resource_user, resource_quota, peer_string); + tcp->resource_user = grpc_resource_user_create(resource_quota, peer_string); /* Tell network status tracking code about the new endpoint */ grpc_network_status_register_endpoint(&tcp->base); diff --git a/src/core/lib/iomgr/tcp_windows.c b/src/core/lib/iomgr/tcp_windows.c index 46f0491d10..62afbcef51 100644 --- a/src/core/lib/iomgr/tcp_windows.c +++ b/src/core/lib/iomgr/tcp_windows.c @@ -40,10 +40,10 @@ #include "src/core/lib/iomgr/network_status_tracker.h" #include "src/core/lib/iomgr/sockaddr_windows.h" +#include <grpc/slice_buffer.h> #include <grpc/support/alloc.h> #include <grpc/support/log.h> #include <grpc/support/log_windows.h> -#include <grpc/support/slice_buffer.h> #include <grpc/support/string_util.h> #include <grpc/support/useful.h> @@ -105,50 +105,39 @@ typedef struct grpc_tcp { grpc_closure *read_cb; grpc_closure *write_cb; - gpr_slice read_slice; - gpr_slice_buffer *write_slices; - gpr_slice_buffer *read_slices; + grpc_slice read_slice; + grpc_slice_buffer *write_slices; + grpc_slice_buffer *read_slices; - grpc_resource_user resource_user; + grpc_resource_user *resource_user; /* The IO Completion Port runs from another thread. We need some mechanism to protect ourselves when requesting a shutdown. */ gpr_mu mu; int shutting_down; - gpr_atm resource_user_shutdown_count; - char *peer_string; } grpc_tcp; -static void win_unref_closure(grpc_exec_ctx *exec_ctx, void *arg /* grpc_tcp */, - grpc_error *error); - -static void win_maybe_shutdown_resource_user(grpc_exec_ctx *exec_ctx, - grpc_tcp *tcp) { - if (gpr_atm_full_fetch_add(&tcp->resource_user_shutdown_count, 1) == 0) { - grpc_resource_user_shutdown(exec_ctx, &tcp->resource_user, - grpc_closure_create(win_unref_closure, tcp)); - } -} - -static void tcp_free(grpc_tcp *tcp) { +static void tcp_free(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) { grpc_winsocket_destroy(tcp->socket); gpr_mu_destroy(&tcp->mu); gpr_free(tcp->peer_string); + grpc_resource_user_unref(exec_ctx, tcp->resource_user); gpr_free(tcp); } /*#define GRPC_TCP_REFCOUNT_DEBUG*/ #ifdef GRPC_TCP_REFCOUNT_DEBUG -#define TCP_UNREF(tcp, reason) tcp_unref((tcp), (reason), __FILE__, __LINE__) +#define TCP_UNREF(exec_ctx, tcp, reason) \ + tcp_unref((exec_ctx), (tcp), (reason), __FILE__, __LINE__) #define TCP_REF(tcp, reason) tcp_ref((tcp), (reason), __FILE__, __LINE__) -static void tcp_unref(grpc_tcp *tcp, const char *reason, const char *file, - int line) { +static void tcp_unref(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp, + const char *reason, const char *file, int line) { gpr_log(file, line, GPR_LOG_SEVERITY_DEBUG, "TCP unref %p : %s %d -> %d", tcp, reason, tcp->refcount.count, tcp->refcount.count - 1); if (gpr_unref(&tcp->refcount)) { - tcp_free(tcp); + tcp_free(exec_ctx, tcp); } } @@ -159,28 +148,23 @@ static void tcp_ref(grpc_tcp *tcp, const char *reason, const char *file, gpr_ref(&tcp->refcount); } #else -#define TCP_UNREF(tcp, reason) tcp_unref((tcp)) +#define TCP_UNREF(exec_ctx, tcp, reason) tcp_unref((exec_ctx), (tcp)) #define TCP_REF(tcp, reason) tcp_ref((tcp)) -static void tcp_unref(grpc_tcp *tcp) { +static void tcp_unref(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) { if (gpr_unref(&tcp->refcount)) { - tcp_free(tcp); + tcp_free(exec_ctx, tcp); } } static void tcp_ref(grpc_tcp *tcp) { gpr_ref(&tcp->refcount); } #endif -static void win_unref_closure(grpc_exec_ctx *exec_ctx, void *arg, - grpc_error *error) { - TCP_UNREF(arg, "resource_user"); -} - /* Asynchronous callback from the IOCP, or the background thread. */ static void on_read(grpc_exec_ctx *exec_ctx, void *tcpp, grpc_error *error) { grpc_tcp *tcp = tcpp; grpc_closure *cb = tcp->read_cb; grpc_winsocket *socket = tcp->socket; - gpr_slice sub; + grpc_slice sub; grpc_winsocket_callback_info *info = &socket->read_info; GRPC_ERROR_REF(error); @@ -190,25 +174,25 @@ static void on_read(grpc_exec_ctx *exec_ctx, void *tcpp, grpc_error *error) { char *utf8_message = gpr_format_message(info->wsa_error); error = GRPC_ERROR_CREATE(utf8_message); gpr_free(utf8_message); - gpr_slice_unref(tcp->read_slice); + grpc_slice_unref(tcp->read_slice); } else { if (info->bytes_transfered != 0 && !tcp->shutting_down) { - sub = gpr_slice_sub_no_ref(tcp->read_slice, 0, info->bytes_transfered); - gpr_slice_buffer_add(tcp->read_slices, sub); + sub = grpc_slice_sub_no_ref(tcp->read_slice, 0, info->bytes_transfered); + grpc_slice_buffer_add(tcp->read_slices, sub); } else { - gpr_slice_unref(tcp->read_slice); + grpc_slice_unref(tcp->read_slice); error = GRPC_ERROR_CREATE("End of TCP stream"); } } } tcp->read_cb = NULL; - TCP_UNREF(tcp, "read"); + TCP_UNREF(exec_ctx, tcp, "read"); grpc_exec_ctx_sched(exec_ctx, cb, error, NULL); } static void win_read(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, - gpr_slice_buffer *read_slices, grpc_closure *cb) { + grpc_slice_buffer *read_slices, grpc_closure *cb) { grpc_tcp *tcp = (grpc_tcp *)ep; grpc_winsocket *handle = tcp->socket; grpc_winsocket_callback_info *info = &handle->read_info; @@ -225,13 +209,13 @@ static void win_read(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, tcp->read_cb = cb; tcp->read_slices = read_slices; - gpr_slice_buffer_reset_and_unref(read_slices); + grpc_slice_buffer_reset_and_unref(read_slices); - tcp->read_slice = gpr_slice_malloc(8192); + tcp->read_slice = grpc_slice_malloc(8192); - buffer.len = (ULONG)GPR_SLICE_LENGTH( + buffer.len = (ULONG)GRPC_SLICE_LENGTH( tcp->read_slice); // we know slice size fits in 32bit. - buffer.buf = (char *)GPR_SLICE_START_PTR(tcp->read_slice); + buffer.buf = (char *)GRPC_SLICE_START_PTR(tcp->read_slice); TCP_REF(tcp, "read"); @@ -287,13 +271,13 @@ static void on_write(grpc_exec_ctx *exec_ctx, void *tcpp, grpc_error *error) { } } - TCP_UNREF(tcp, "write"); + TCP_UNREF(exec_ctx, tcp, "write"); grpc_exec_ctx_sched(exec_ctx, cb, error, NULL); } /* Initiates a write. */ static void win_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, - gpr_slice_buffer *slices, grpc_closure *cb) { + grpc_slice_buffer *slices, grpc_closure *cb) { grpc_tcp *tcp = (grpc_tcp *)ep; grpc_winsocket *socket = tcp->socket; grpc_winsocket_callback_info *info = &socket->write_info; @@ -320,10 +304,10 @@ static void win_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, } for (i = 0; i < tcp->write_slices->count; i++) { - len = GPR_SLICE_LENGTH(tcp->write_slices->slices[i]); + len = GRPC_SLICE_LENGTH(tcp->write_slices->slices[i]); GPR_ASSERT(len <= ULONG_MAX); buffers[i].len = (ULONG)len; - buffers[i].buf = (char *)GPR_SLICE_START_PTR(tcp->write_slices->slices[i]); + buffers[i].buf = (char *)GRPC_SLICE_START_PTR(tcp->write_slices->slices[i]); } /* First, let's try a synchronous, non-blocking write. */ @@ -355,7 +339,7 @@ static void win_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep, if (status != 0) { int wsa_error = WSAGetLastError(); if (wsa_error != WSA_IO_PENDING) { - TCP_UNREF(tcp, "write"); + TCP_UNREF(exec_ctx, tcp, "write"); grpc_exec_ctx_sched(exec_ctx, cb, GRPC_WSA_ERROR(wsa_error, "WSASend"), NULL); return; @@ -396,15 +380,14 @@ static void win_shutdown(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep) { callback. See the comments in on_read and on_write. */ tcp->shutting_down = 1; grpc_winsocket_shutdown(tcp->socket); - win_maybe_shutdown_resource_user(exec_ctx, tcp); gpr_mu_unlock(&tcp->mu); + grpc_resource_user_shutdown(exec_ctx, tcp->resource_user); } static void win_destroy(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep) { grpc_network_status_unregister_endpoint(ep); grpc_tcp *tcp = (grpc_tcp *)ep; - win_maybe_shutdown_resource_user(exec_ctx, tcp); - TCP_UNREF(tcp, "destroy"); + TCP_UNREF(exec_ctx, tcp, "destroy"); } static char *win_get_peer(grpc_endpoint *ep) { @@ -416,9 +399,11 @@ static grpc_workqueue *win_get_workqueue(grpc_endpoint *ep) { return NULL; } static grpc_resource_user *win_get_resource_user(grpc_endpoint *ep) { grpc_tcp *tcp = (grpc_tcp *)ep; - return &tcp->resource_user; + return tcp->resource_user; } +static int win_get_fd(grpc_endpoint *ep) { return -1; } + static grpc_endpoint_vtable vtable = {win_read, win_write, win_get_workqueue, @@ -427,7 +412,8 @@ static grpc_endpoint_vtable vtable = {win_read, win_shutdown, win_destroy, win_get_resource_user, - win_get_peer}; + win_get_peer, + win_get_fd}; grpc_endpoint *grpc_tcp_create(grpc_winsocket *socket, grpc_resource_quota *resource_quota, @@ -441,7 +427,7 @@ grpc_endpoint *grpc_tcp_create(grpc_winsocket *socket, grpc_closure_init(&tcp->on_read, on_read, tcp); grpc_closure_init(&tcp->on_write, on_write, tcp); tcp->peer_string = gpr_strdup(peer_string); - grpc_resource_user_init(&tcp->resource_user, resource_quota, peer_string); + tcp->resource_user = grpc_resource_user_create(resource_quota, peer_string); /* Tell network status tracking code about the new endpoint */ grpc_network_status_register_endpoint(&tcp->base); diff --git a/src/core/lib/iomgr/wakeup_fd_pipe.c b/src/core/lib/iomgr/wakeup_fd_pipe.c index 183f0eb930..e186d63683 100644 --- a/src/core/lib/iomgr/wakeup_fd_pipe.c +++ b/src/core/lib/iomgr/wakeup_fd_pipe.c @@ -95,6 +95,8 @@ static void pipe_destroy(grpc_wakeup_fd* fd_info) { static int pipe_check_availability(void) { grpc_wakeup_fd fd; + fd.read_fd = fd.write_fd = -1; + if (pipe_init(&fd) == GRPC_ERROR_NONE) { pipe_destroy(&fd); return 1; |