void uv__io_poll(uv_loop_t* loop, int timeout) {
struct uv__epoll_event events[1024];
struct uv__epoll_event* pe;
struct uv__epoll_event e;
QUEUE* q;
uv__io_t* w;
uint64_t base;
uint64_t diff;
int nevents;
int count;
int nfds;
int fd;
int op;
int i;
if (loop->nfds == 0) {
assert(QUEUE_EMPTY(&loop->watcher_queue));
return;
}
while (!QUEUE_EMPTY(&loop->watcher_queue)) {
q = QUEUE_HEAD(&loop->watcher_queue);
QUEUE_REMOVE(q);
QUEUE_INIT(q);
w = QUEUE_DATA(q, uv__io_t, watcher_queue);
assert(w->pevents != 0);
assert(w->fd >= 0);
assert(w->fd < (int) loop->nwatchers);
e.events = w->pevents;
e.data = w->fd;
if (w->events == 0)
op = UV__EPOLL_CTL_ADD;
else
op = UV__EPOLL_CTL_MOD;
/* XXX Future optimization: do EPOLL_CTL_MOD lazily if we stop watching
* events, skip the syscall and squelch the events after epoll_wait().
*/
if (uv__epoll_ctl(loop->backend_fd, op, w->fd, &e)) {
if (errno != EEXIST)
abort();
assert(op == UV__EPOLL_CTL_ADD);
/* We've reactivated a file descriptor that's been watched before. */
if (uv__epoll_ctl(loop->backend_fd, UV__EPOLL_CTL_MOD, w->fd, &e))
abort();
}
w->events = w->pevents;
}
assert(timeout >= -1);
base = loop->time;
count = 48; /* Benchmarks suggest this gives the best throughput. */
for (;;) {
nfds = uv__epoll_wait(loop->backend_fd,
events,
ARRAY_SIZE(events),
timeout);
/* Update loop->time unconditionally. It's tempting to skip the update when
* timeout == 0 (i.e. non-blocking poll) but there is no guarantee that the
* operating system didn't reschedule our process while in the syscall.
*/
SAVE_ERRNO(uv__update_time(loop));
if (nfds == 0) {
assert(timeout != -1);
return;
}
if (nfds == -1) {
if (errno != EINTR)
abort();
if (timeout == -1)
continue;
if (timeout == 0)
return;
/* Interrupted by a signal. Update timeout and poll again. */
goto update_timeout;
}
nevents = 0;
assert(loop->watchers != NULL);
loop->watchers[loop->nwatchers] = (void*) events;
loop->watchers[loop->nwatchers + 1] = (void*) (uintptr_t) nfds;
for (i = 0; i < nfds; i++) {
pe = events + i;
fd = pe->data;
/* Skip invalidated events, see uv__platform_invalidate_fd */
if (fd == -1)
continue;
assert(fd >= 0);
assert((unsigned) fd < loop->nwatchers);
w = loop->watchers[fd];
if (w == NULL) {
/* File descriptor that we've stopped watching, disarm it.
*
* Ignore all errors because we may be racing with another thread
* when the file descriptor is closed.
*/
uv__epoll_ctl(loop->backend_fd, UV__EPOLL_CTL_DEL, fd, pe);
continue;
}
/* Give users only events they're interested in. Prevents spurious
* callbacks when previous callback invocation in this loop has stopped
* the current watcher. Also, filters out events that users has not
* requested us to watch.
*/
pe->events &= w->pevents | UV__POLLERR | UV__POLLHUP;
/* Work around an epoll quirk where it sometimes reports just the
* EPOLLERR or EPOLLHUP event. In order to force the event loop to
* move forward, we merge in the read/write events that the watcher
* is interested in; uv__read() and uv__write() will then deal with
* the error or hangup in the usual fashion.
*
* Note to self: happens when epoll reports EPOLLIN|EPOLLHUP, the user
* reads the available data, calls uv_read_stop(), then sometime later
* calls uv_read_start() again. By then, libuv has forgotten about the
* hangup and the kernel won't report EPOLLIN again because there's
* nothing left to read. If anything, libuv is to blame here. The
* current hack is just a quick bandaid; to properly fix it, libuv
* needs to remember the error/hangup event. We should get that for
* free when we switch over to edge-triggered I/O.
*/
if (pe->events == UV__EPOLLERR || pe->events == UV__EPOLLHUP)
pe->events |= w->pevents & (UV__EPOLLIN | UV__EPOLLOUT);
if (pe->events != 0) {
w->cb(loop, w, pe->events);
nevents++;
}
}
loop->watchers[loop->nwatchers] = NULL;
loop->watchers[loop->nwatchers + 1] = NULL;
if (nevents != 0) {
if (nfds == ARRAY_SIZE(events) && --count != 0) {
/* Poll for more events but don't block this time. */
timeout = 0;
continue;
}
return;
}
if (timeout == 0)
return;
if (timeout == -1)
continue;
update_timeout:
assert(timeout > 0);
diff = loop->time - base;
if (diff >= (uint64_t) timeout)
return;
timeout -= diff;
}
}
void uv__io_poll(uv_loop_t* loop, int timeout) {
/* A bug in kernels < 2.6.37 makes timeouts larger than ~30 minutes
* effectively infinite on 32 bits architectures. To avoid blocking
* indefinitely, we cap the timeout and poll again if necessary.
*
* Note that "30 minutes" is a simplification because it depends on
* the value of CONFIG_HZ. The magic constant assumes CONFIG_HZ=1200,
* that being the largest value I have seen in the wild (and only once.)
*/
static const int max_safe_timeout = 1789569;
static int no_epoll_pwait;
static int no_epoll_wait;
struct uv__epoll_event events[1024];
struct uv__epoll_event* pe;
struct uv__epoll_event e;
int real_timeout;
QUEUE* q;
uv__io_t* w;
sigset_t sigset;
uint64_t sigmask;
uint64_t base;
int nevents;
int count;
int nfds;
int fd;
int op;
int i;
if (loop->nfds == 0) {
assert(QUEUE_EMPTY(&loop->watcher_queue));
return;
}
while (!QUEUE_EMPTY(&loop->watcher_queue)) {
q = QUEUE_HEAD(&loop->watcher_queue);
QUEUE_REMOVE(q);
QUEUE_INIT(q);
w = QUEUE_DATA(q, uv__io_t, watcher_queue);
assert(w->pevents != 0);
assert(w->fd >= 0);
assert(w->fd < (int) loop->nwatchers);
e.events = w->pevents;
e.data = w->fd;
if (w->events == 0)
op = UV__EPOLL_CTL_ADD;
else
op = UV__EPOLL_CTL_MOD;
/* XXX Future optimization: do EPOLL_CTL_MOD lazily if we stop watching
* events, skip the syscall and squelch the events after epoll_wait().
*/
if (uv__epoll_ctl(loop->backend_fd, op, w->fd, &e)) {
if (errno != EEXIST)
abort();
assert(op == UV__EPOLL_CTL_ADD);
/* We've reactivated a file descriptor that's been watched before. */
if (uv__epoll_ctl(loop->backend_fd, UV__EPOLL_CTL_MOD, w->fd, &e))
abort();
}
w->events = w->pevents;
}
sigmask = 0;
if (loop->flags & UV_LOOP_BLOCK_SIGPROF) {
sigemptyset(&sigset);
sigaddset(&sigset, SIGPROF);
sigmask |= 1 << (SIGPROF - 1);
}
assert(timeout >= -1);
base = loop->time;
count = 48; /* Benchmarks suggest this gives the best throughput. */
real_timeout = timeout;
for (;;) {
/* See the comment for max_safe_timeout for an explanation of why
* this is necessary. Executive summary: kernel bug workaround.
*/
if (sizeof(int32_t) == sizeof(long) && timeout >= max_safe_timeout)
timeout = max_safe_timeout;
if (sigmask != 0 && no_epoll_pwait != 0)
if (pthread_sigmask(SIG_BLOCK, &sigset, NULL))
abort();
if (no_epoll_wait != 0 || (sigmask != 0 && no_epoll_pwait == 0)) {
nfds = uv__epoll_pwait(loop->backend_fd,
events,
ARRAY_SIZE(events),
timeout,
sigmask);
if (nfds == -1 && errno == ENOSYS)
no_epoll_pwait = 1;
} else {
nfds = uv__epoll_wait(loop->backend_fd,
events,
ARRAY_SIZE(events),
timeout);
if (nfds == -1 && errno == ENOSYS)
no_epoll_wait = 1;
}
if (sigmask != 0 && no_epoll_pwait != 0)
if (pthread_sigmask(SIG_UNBLOCK, &sigset, NULL))
abort();
/* Update loop->time unconditionally. It's tempting to skip the update when
* timeout == 0 (i.e. non-blocking poll) but there is no guarantee that the
* operating system didn't reschedule our process while in the syscall.
*/
SAVE_ERRNO(uv__update_time(loop));
if (nfds == 0) {
assert(timeout != -1);
timeout = real_timeout - timeout;
if (timeout > 0)
continue;
return;
}
if (nfds == -1) {
if (errno == ENOSYS) {
/* epoll_wait() or epoll_pwait() failed, try the other system call. */
assert(no_epoll_wait == 0 || no_epoll_pwait == 0);
continue;
}
if (errno != EINTR)
abort();
if (timeout == -1)
continue;
if (timeout == 0)
return;
/* Interrupted by a signal. Update timeout and poll again. */
goto update_timeout;
}
nevents = 0;
assert(loop->watchers != NULL);
loop->watchers[loop->nwatchers] = (void*) events;
loop->watchers[loop->nwatchers + 1] = (void*) (uintptr_t) nfds;
for (i = 0; i < nfds; i++) {
pe = events + i;
fd = pe->data;
/* Skip invalidated events, see uv__platform_invalidate_fd */
if (fd == -1)
continue;
assert(fd >= 0);
assert((unsigned) fd < loop->nwatchers);
w = loop->watchers[fd];
if (w == NULL) {
/* File descriptor that we've stopped watching, disarm it.
*
* Ignore all errors because we may be racing with another thread
* when the file descriptor is closed.
*/
uv__epoll_ctl(loop->backend_fd, UV__EPOLL_CTL_DEL, fd, pe);
continue;
}
/* Give users only events they're interested in. Prevents spurious
* callbacks when previous callback invocation in this loop has stopped
* the current watcher. Also, filters out events that users has not
* requested us to watch.
*/
pe->events &= w->pevents | UV__POLLERR | UV__POLLHUP;
/* Work around an epoll quirk where it sometimes reports just the
* EPOLLERR or EPOLLHUP event. In order to force the event loop to
* move forward, we merge in the read/write events that the watcher
* is interested in; uv__read() and uv__write() will then deal with
* the error or hangup in the usual fashion.
*
* Note to self: happens when epoll reports EPOLLIN|EPOLLHUP, the user
* reads the available data, calls uv_read_stop(), then sometime later
* calls uv_read_start() again. By then, libuv has forgotten about the
* hangup and the kernel won't report EPOLLIN again because there's
* nothing left to read. If anything, libuv is to blame here. The
* current hack is just a quick bandaid; to properly fix it, libuv
* needs to remember the error/hangup event. We should get that for
* free when we switch over to edge-triggered I/O.
*/
if (pe->events == UV__EPOLLERR || pe->events == UV__EPOLLHUP)
pe->events |= w->pevents & (UV__EPOLLIN | UV__EPOLLOUT);
if (pe->events != 0) {
w->cb(loop, w, pe->events);
nevents++;
}
}
loop->watchers[loop->nwatchers] = NULL;
loop->watchers[loop->nwatchers + 1] = NULL;
if (nevents != 0) {
if (nfds == ARRAY_SIZE(events) && --count != 0) {
/* Poll for more events but don't block this time. */
timeout = 0;
continue;
}
return;
}
if (timeout == 0)
return;
if (timeout == -1)
continue;
update_timeout:
assert(timeout > 0);
real_timeout -= (loop->time - base);
if (real_timeout <= 0)
return;
timeout = real_timeout;
}
}
