2 * Copyright (c) 2008, 2009, 2010, 2011, 2012 Nicira, Inc.
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"
25 #include "dynamic-string.h"
26 #include "fatal-signal.h"
28 #include "socket-util.h"
33 #undef poll_timer_wait
34 #undef poll_timer_wait_until
35 #undef poll_immediate_wake
37 VLOG_DEFINE_THIS_MODULE(poll_loop);
39 COVERAGE_DEFINE(poll_fd_wait);
40 COVERAGE_DEFINE(poll_zero_timeout);
42 /* An event that will wake the following call to poll_block(). */
44 /* Set when the waiter is created. */
45 struct list node; /* Element in global waiters list. */
46 int fd; /* File descriptor. */
47 short int events; /* Events to wait for (POLLIN, POLLOUT). */
48 const char *where; /* Where the waiter was created. */
50 /* Set only when poll_block() is called. */
51 struct pollfd *pollfd; /* Pointer to element of the pollfds array. */
54 /* All active poll waiters. */
55 static struct list waiters = LIST_INITIALIZER(&waiters);
57 /* Time at which to wake up the next call to poll_block(), in milliseconds as
58 * returned by time_msec(), LLONG_MIN to wake up immediately, or LLONG_MAX to
60 static long long int timeout_when = LLONG_MAX;
62 /* Location where waiter created. */
63 static const char *timeout_where;
65 static struct poll_waiter *new_waiter(int fd, short int events,
68 /* Registers 'fd' as waiting for the specified 'events' (which should be POLLIN
69 * or POLLOUT or POLLIN | POLLOUT). The following call to poll_block() will
70 * wake up when 'fd' becomes ready for one or more of the requested events.
72 * The event registration is one-shot: only the following call to poll_block()
73 * is affected. The event will need to be re-registered after poll_block() is
74 * called if it is to persist.
76 * Ordinarily the 'where' argument is supplied automatically; see poll-loop.h
77 * for more information. */
79 poll_fd_wait(int fd, short int events, const char *where)
81 COVERAGE_INC(poll_fd_wait);
82 return new_waiter(fd, events, where);
85 /* Causes the following call to poll_block() to block for no more than 'msec'
86 * milliseconds. If 'msec' is nonpositive, the following call to poll_block()
87 * will not block at all.
89 * The timer registration is one-shot: only the following call to poll_block()
90 * is affected. The timer will need to be re-registered after poll_block() is
91 * called if it is to persist.
93 * Ordinarily the 'where' argument is supplied automatically; see poll-loop.h
94 * for more information. */
96 poll_timer_wait(long long int msec, const char *where)
98 long long int now = time_msec();
102 /* Wake up immediately. */
104 } else if ((unsigned long long int) now + msec <= LLONG_MAX) {
108 /* now + msec would overflow. */
112 poll_timer_wait_until(when, where);
115 /* Causes the following call to poll_block() to wake up when the current time,
116 * as returned by time_msec(), reaches 'when' or later. If 'when' is earlier
117 * than the current time, the following call to poll_block() will not block at
120 * The timer registration is one-shot: only the following call to poll_block()
121 * is affected. The timer will need to be re-registered after poll_block() is
122 * called if it is to persist.
124 * Ordinarily the 'where' argument is supplied automatically; see poll-loop.h
125 * for more information. */
127 poll_timer_wait_until(long long int when, const char *where)
129 if (when < timeout_when) {
131 timeout_where = where;
135 /* Causes the following call to poll_block() to wake up immediately, without
138 * Ordinarily the 'where' argument is supplied automatically; see poll-loop.h
139 * for more information. */
141 poll_immediate_wake(const char *where)
143 poll_timer_wait(0, where);
146 /* Logs, if appropriate, that the poll loop was awakened by an event
147 * registered at 'where' (typically a source file and line number). The other
148 * arguments have two possible interpretations:
150 * - If 'pollfd' is nonnull then it should be the "struct pollfd" that caused
151 * the wakeup. 'timeout' is ignored.
153 * - If 'pollfd' is NULL then 'timeout' is the number of milliseconds after
154 * which the poll loop woke up.
157 log_wakeup(const char *where, const struct pollfd *pollfd, int timeout)
159 static struct vlog_rate_limit trace_rl = VLOG_RATE_LIMIT_INIT(1, 1);
160 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(10, 10);
161 enum vlog_level level;
165 cpu_usage = get_cpu_usage();
166 if (VLOG_IS_DBG_ENABLED()) {
168 } else if (cpu_usage > 50 && !VLOG_DROP_INFO(&rl)) {
175 ds_put_cstr(&s, "wakeup due to ");
177 char *description = describe_fd(pollfd->fd);
178 if (pollfd->revents & POLLIN) {
179 ds_put_cstr(&s, "[POLLIN]");
181 if (pollfd->revents & POLLOUT) {
182 ds_put_cstr(&s, "[POLLOUT]");
184 if (pollfd->revents & POLLERR) {
185 ds_put_cstr(&s, "[POLLERR]");
187 if (pollfd->revents & POLLHUP) {
188 ds_put_cstr(&s, "[POLLHUP]");
190 if (pollfd->revents & POLLNVAL) {
191 ds_put_cstr(&s, "[POLLNVAL]");
193 ds_put_format(&s, " on fd %d (%s)", pollfd->fd, description);
196 ds_put_format(&s, "%d-ms timeout", timeout);
199 ds_put_format(&s, " at %s", where);
201 if (cpu_usage >= 0) {
202 ds_put_format(&s, " (%d%% CPU usage)", cpu_usage);
204 if (!vlog_should_drop(THIS_MODULE, level, &trace_rl)) {
205 ds_put_char(&s, '\n');
206 format_backtraces(&s, 2);
209 VLOG(level, "%s", ds_cstr(&s));
213 /* Blocks until one or more of the events registered with poll_fd_wait()
214 * occurs, or until the minimum duration registered with poll_timer_wait()
215 * elapses, or not at all if poll_immediate_wake() has been called. */
219 static struct pollfd *pollfds;
220 static size_t max_pollfds;
222 struct poll_waiter *pw, *next;
223 int n_waiters, n_pollfds;
227 /* Register fatal signal events before actually doing any real work for
231 n_waiters = list_size(&waiters);
232 if (max_pollfds < n_waiters) {
233 max_pollfds = n_waiters;
234 pollfds = xrealloc(pollfds, max_pollfds * sizeof *pollfds);
238 LIST_FOR_EACH (pw, node, &waiters) {
239 pw->pollfd = &pollfds[n_pollfds];
240 pollfds[n_pollfds].fd = pw->fd;
241 pollfds[n_pollfds].events = pw->events;
242 pollfds[n_pollfds].revents = 0;
246 if (timeout_when == LLONG_MIN) {
247 COVERAGE_INC(poll_zero_timeout);
249 retval = time_poll(pollfds, n_pollfds, timeout_when, &elapsed);
251 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
252 VLOG_ERR_RL(&rl, "poll: %s", strerror(-retval));
253 } else if (!retval) {
254 log_wakeup(timeout_where, NULL, elapsed);
257 LIST_FOR_EACH_SAFE (pw, next, node, &waiters) {
258 if (pw->pollfd->revents) {
259 log_wakeup(pw->where, pw->pollfd, 0);
264 timeout_when = LLONG_MAX;
265 timeout_where = NULL;
267 /* Handle any pending signals before doing anything else. */
271 /* Cancels the file descriptor event registered with poll_fd_wait() using 'pw',
272 * the struct poll_waiter returned by that function.
274 * An event registered with poll_fd_wait() may be canceled from its time of
275 * registration until the next call to poll_block(). At that point, the event
276 * is automatically canceled by the system and its poll_waiter is freed. */
278 poll_cancel(struct poll_waiter *pw)
281 list_remove(&pw->node);
286 /* Creates and returns a new poll_waiter for 'fd' and 'events'. */
287 static struct poll_waiter *
288 new_waiter(int fd, short int events, const char *where)
290 struct poll_waiter *waiter = xzalloc(sizeof *waiter);
293 waiter->events = events;
294 waiter->where = where;
295 list_push_back(&waiters, &waiter->node);