2 * Copyright (c) 2008, 2009, 2010, 2011 Nicira Networks.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "poll-loop.h"
26 #include "dynamic-string.h"
27 #include "fatal-signal.h"
29 #include "socket-util.h"
34 #undef poll_timer_wait
35 #undef poll_timer_wait_until
36 #undef poll_immediate_wake
38 VLOG_DEFINE_THIS_MODULE(poll_loop);
40 COVERAGE_DEFINE(poll_fd_wait);
41 COVERAGE_DEFINE(poll_zero_timeout);
43 /* An event that will wake the following call to poll_block(). */
45 /* Set when the waiter is created. */
46 struct list node; /* Element in global waiters list. */
47 int fd; /* File descriptor. */
48 short int events; /* Events to wait for (POLLIN, POLLOUT). */
49 const char *where; /* Where the waiter was created. */
51 /* Set only when poll_block() is called. */
52 struct pollfd *pollfd; /* Pointer to element of the pollfds array. */
55 /* All active poll waiters. */
56 static struct list waiters = LIST_INITIALIZER(&waiters);
58 /* Max time to wait in next call to poll_block(), in milliseconds, or -1 to
60 static int timeout = -1;
62 /* Location where waiter created. */
63 static const char *timeout_where;
65 /* Array of file descriptors from last run of poll_block(). */
66 static struct pollfd *pollfds;
68 /* Allocated size of pollfds. */
69 static size_t max_pollfds;
71 /* Current number of elements in pollfds. */
74 static struct poll_waiter *new_waiter(int fd, short int events,
77 /* Registers 'fd' as waiting for the specified 'events' (which should be POLLIN
78 * or POLLOUT or POLLIN | POLLOUT). The following call to poll_block() will
79 * wake up when 'fd' becomes ready for one or more of the requested events.
81 * The event registration is one-shot: only the following call to poll_block()
82 * is affected. The event will need to be re-registered after poll_block() is
83 * called if it is to persist.
85 * Ordinarily the 'where' argument is supplied automatically; see poll-loop.h
86 * for more information. */
88 poll_fd_wait(int fd, short int events, const char *where)
90 COVERAGE_INC(poll_fd_wait);
91 return new_waiter(fd, events, where);
94 /* The caller must ensure that 'msec' is not negative. */
96 poll_timer_wait__(int msec, const char *where)
98 if (timeout < 0 || msec < timeout) {
100 timeout_where = where;
104 /* Causes the following call to poll_block() to block for no more than 'msec'
105 * milliseconds. If 'msec' is nonpositive, the following call to poll_block()
106 * will not block at all.
108 * The timer registration is one-shot: only the following call to poll_block()
109 * is affected. The timer will need to be re-registered after poll_block() is
110 * called if it is to persist.
112 * Ordinarily the 'where' argument is supplied automatically; see poll-loop.h
113 * for more information. */
115 poll_timer_wait(long long int msec, const char *where)
117 poll_timer_wait__((msec < 0 ? 0
118 : msec > INT_MAX ? INT_MAX
123 /* Causes the following call to poll_block() to wake up when the current time,
124 * as returned by time_msec(), reaches 'msec' or later. If 'msec' is earlier
125 * than the current time, the following call to poll_block() will not block at
128 * The timer registration is one-shot: only the following call to poll_block()
129 * is affected. The timer will need to be re-registered after poll_block() is
130 * called if it is to persist.
132 * Ordinarily the 'where' argument is supplied automatically; see poll-loop.h
133 * for more information. */
135 poll_timer_wait_until(long long int msec, const char *where)
137 long long int now = time_msec();
138 poll_timer_wait__((msec <= now ? 0
139 : msec < now + INT_MAX ? msec - now
144 /* Causes the following call to poll_block() to wake up immediately, without
147 * Ordinarily the 'where' argument is supplied automatically; see poll-loop.h
148 * for more information. */
150 poll_immediate_wake(const char *where)
152 poll_timer_wait(0, where);
155 /* Logs, if appropriate, that the poll loop was awakened by an event
156 * registered at 'where' (typically a source file and line number). The other
157 * arguments have two possible interpretations:
159 * - If 'pollfd' is nonnull then it should be the "struct pollfd" that caused
160 * the wakeup. 'timeout' is ignored.
162 * - If 'pollfd' is NULL then 'timeout' is the number of milliseconds after
163 * which the poll loop woke up.
166 log_wakeup(const char *where, const struct pollfd *pollfd, int timeout)
168 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(120, 120);
169 enum vlog_level level;
173 cpu_usage = get_cpu_usage();
174 if (VLOG_IS_DBG_ENABLED()) {
176 } else if (cpu_usage > 50 && !VLOG_DROP_WARN(&rl)) {
183 ds_put_cstr(&s, "wakeup due to ");
185 char *description = describe_fd(pollfd->fd);
186 if (pollfd->revents & POLLIN) {
187 ds_put_cstr(&s, "[POLLIN]");
189 if (pollfd->revents & POLLOUT) {
190 ds_put_cstr(&s, "[POLLOUT]");
192 if (pollfd->revents & POLLERR) {
193 ds_put_cstr(&s, "[POLLERR]");
195 if (pollfd->revents & POLLHUP) {
196 ds_put_cstr(&s, "[POLLHUP]");
198 if (pollfd->revents & POLLNVAL) {
199 ds_put_cstr(&s, "[POLLNVAL]");
201 ds_put_format(&s, " on fd %d (%s)", pollfd->fd, description);
204 ds_put_format(&s, "%d-ms timeout", timeout);
207 ds_put_format(&s, " at %s", where);
209 if (cpu_usage >= 0) {
210 ds_put_format(&s, " (%d%% CPU usage)", cpu_usage);
212 VLOG(level, "%s", ds_cstr(&s));
216 /* Blocks until one or more of the events registered with poll_fd_wait()
217 * occurs, or until the minimum duration registered with poll_timer_wait()
218 * elapses, or not at all if poll_immediate_wake() has been called. */
222 struct poll_waiter *pw, *next;
226 /* Register fatal signal events before actually doing any real work for
230 n_waiters = list_size(&waiters);
231 if (max_pollfds < n_waiters) {
232 max_pollfds = n_waiters;
233 pollfds = xrealloc(pollfds, max_pollfds * sizeof *pollfds);
237 LIST_FOR_EACH (pw, node, &waiters) {
238 pw->pollfd = &pollfds[n_pollfds];
239 pollfds[n_pollfds].fd = pw->fd;
240 pollfds[n_pollfds].events = pw->events;
241 pollfds[n_pollfds].revents = 0;
246 COVERAGE_INC(poll_zero_timeout);
248 retval = time_poll(pollfds, n_pollfds, timeout);
250 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
251 VLOG_ERR_RL(&rl, "poll: %s", strerror(-retval));
252 } else if (!retval) {
253 log_wakeup(timeout_where, NULL, timeout);
256 LIST_FOR_EACH_SAFE (pw, next, node, &waiters) {
257 if (pw->pollfd->revents) {
258 log_wakeup(pw->where, pw->pollfd, 0);
264 timeout_where = NULL;
266 /* Handle any pending signals before doing anything else. */
270 /* Cancels the file descriptor event registered with poll_fd_wait() using 'pw',
271 * the struct poll_waiter returned by that function.
273 * An event registered with poll_fd_wait() may be canceled from its time of
274 * registration until the next call to poll_block(). At that point, the event
275 * is automatically canceled by the system and its poll_waiter is freed. */
277 poll_cancel(struct poll_waiter *pw)
280 list_remove(&pw->node);
285 /* Checks whether the given file descriptor caused the poll loop to wake up
286 * in the previous iteration. If it did, returns a bitmask of the events
287 * that caused the wakeup. Otherwise returns 0;
293 short int events = 0;
295 for (i = 0; i < n_pollfds; i++) {
296 if (pollfds[i].fd == fd) {
297 events |= pollfds[i].revents;
304 /* Creates and returns a new poll_waiter for 'fd' and 'events'. */
305 static struct poll_waiter *
306 new_waiter(int fd, short int events, const char *where)
308 struct poll_waiter *waiter = xzalloc(sizeof *waiter);
311 waiter->events = events;
312 waiter->where = where;
313 list_push_back(&waiters, &waiter->node);