2 * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014 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 "ovs-thread.h"
30 #include "socket-util.h"
36 VLOG_DEFINE_THIS_MODULE(poll_loop);
38 COVERAGE_DEFINE(poll_fd_wait);
39 COVERAGE_DEFINE(poll_zero_timeout);
42 struct hmap_node hmap_node;
43 struct pollfd pollfd; /* Events to pass to time_poll(). */
44 HANDLE wevent; /* Events for WaitForMultipleObjects(). */
45 const char *where; /* Where poll_node was created. */
49 /* All active poll waiters. */
50 struct hmap poll_nodes;
52 /* Time at which to wake up the next call to poll_block(), LLONG_MIN to
53 * wake up immediately, or LLONG_MAX to wait forever. */
54 long long int timeout_when; /* In msecs as returned by time_msec(). */
55 const char *timeout_where; /* Where 'timeout_when' was set. */
58 static struct poll_loop *poll_loop(void);
60 /* Look up the node with same fd and wevent. */
61 static struct poll_node *
62 find_poll_node(struct poll_loop *loop, int fd, uint32_t wevent)
64 struct poll_node *node;
66 HMAP_FOR_EACH_WITH_HASH (node, hmap_node, hash_2words(fd, wevent),
68 if (node->pollfd.fd == fd && node->wevent == wevent) {
75 /* On Unix based systems:
77 * Registers 'fd' as waiting for the specified 'events' (which should be
78 * POLLIN or POLLOUT or POLLIN | POLLOUT). The following call to
79 * poll_block() will wake up when 'fd' becomes ready for one or more of the
80 * requested events. the 'fd's are given to poll() function later.
84 * Register 'wevent' handle for the specified 'events'. These wevents are
85 * given to the handleMultipleObjects() to be polled. The event
86 * registration is one-shot: only the following call to poll_block() is
87 * affected. The event will need to be re-registered after poll_block() is
88 * called if it is to persist.
90 * ('where' is used in debug logging. Commonly one would use poll_fd_wait() to
91 * automatically provide the caller's source file and line number for
94 poll_fd_wait_at(int fd, HANDLE wevent, short int events, const char *where)
96 struct poll_loop *loop = poll_loop();
97 struct poll_node *node;
99 COVERAGE_INC(poll_fd_wait);
102 /* Null event cannot be polled. */
104 VLOG_ERR("No event to wait fd %d", fd);
109 /* Check for duplicate. If found, "or" the event. */
110 node = find_poll_node(loop, fd, wevent);
112 node->pollfd.events |= events;
114 node = xzalloc(sizeof *node);
115 hmap_insert(&loop->poll_nodes, &node->hmap_node,
116 hash_2words(fd, wevent));
117 node->pollfd.fd = fd;
118 node->pollfd.events = events;
119 node->wevent = wevent;
124 /* Causes the following call to poll_block() to block for no more than 'msec'
125 * milliseconds. If 'msec' is nonpositive, the following call to poll_block()
126 * will not block at all.
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 * ('where' is used in debug logging. Commonly one would use poll_timer_wait()
133 * to automatically provide the caller's source file and line number for
136 poll_timer_wait_at(long long int msec, const char *where)
138 long long int now = time_msec();
142 /* Wake up immediately. */
144 } else if ((unsigned long long int) now + msec <= LLONG_MAX) {
148 /* now + msec would overflow. */
152 poll_timer_wait_until_at(when, where);
155 /* Causes the following call to poll_block() to wake up when the current time,
156 * as returned by time_msec(), reaches 'when' or later. If 'when' is earlier
157 * than the current time, the following call to poll_block() will not block at
160 * The timer registration is one-shot: only the following call to poll_block()
161 * is affected. The timer will need to be re-registered after poll_block() is
162 * called if it is to persist.
164 * ('where' is used in debug logging. Commonly one would use
165 * poll_timer_wait_until() to automatically provide the caller's source file
166 * and line number for 'where'.) */
168 poll_timer_wait_until_at(long long int when, const char *where)
170 struct poll_loop *loop = poll_loop();
171 if (when < loop->timeout_when) {
172 loop->timeout_when = when;
173 loop->timeout_where = where;
177 /* Causes the following call to poll_block() to wake up immediately, without
180 * ('where' is used in debug logging. Commonly one would use
181 * poll_immediate_wake() to automatically provide the caller's source file and
182 * line number for 'where'.) */
184 poll_immediate_wake_at(const char *where)
186 poll_timer_wait_at(0, where);
189 /* Logs, if appropriate, that the poll loop was awakened by an event
190 * registered at 'where' (typically a source file and line number). The other
191 * arguments have two possible interpretations:
193 * - If 'pollfd' is nonnull then it should be the "struct pollfd" that caused
194 * the wakeup. 'timeout' is ignored.
196 * - If 'pollfd' is NULL then 'timeout' is the number of milliseconds after
197 * which the poll loop woke up.
200 log_wakeup(const char *where, const struct pollfd *pollfd, int timeout)
202 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(10, 10);
203 enum vlog_level level;
207 cpu_usage = get_cpu_usage();
208 if (VLOG_IS_DBG_ENABLED()) {
210 } else if (cpu_usage > 50 && !VLOG_DROP_INFO(&rl)) {
217 ds_put_cstr(&s, "wakeup due to ");
219 char *description = describe_fd(pollfd->fd);
220 if (pollfd->revents & POLLIN) {
221 ds_put_cstr(&s, "[POLLIN]");
223 if (pollfd->revents & POLLOUT) {
224 ds_put_cstr(&s, "[POLLOUT]");
226 if (pollfd->revents & POLLERR) {
227 ds_put_cstr(&s, "[POLLERR]");
229 if (pollfd->revents & POLLHUP) {
230 ds_put_cstr(&s, "[POLLHUP]");
232 if (pollfd->revents & POLLNVAL) {
233 ds_put_cstr(&s, "[POLLNVAL]");
235 ds_put_format(&s, " on fd %d (%s)", pollfd->fd, description);
238 ds_put_format(&s, "%d-ms timeout", timeout);
241 ds_put_format(&s, " at %s", where);
243 if (cpu_usage >= 0) {
244 ds_put_format(&s, " (%d%% CPU usage)", cpu_usage);
246 VLOG(level, "%s", ds_cstr(&s));
251 free_poll_nodes(struct poll_loop *loop)
253 struct poll_node *node, *next;
255 HMAP_FOR_EACH_SAFE (node, next, hmap_node, &loop->poll_nodes) {
256 hmap_remove(&loop->poll_nodes, &node->hmap_node);
261 /* Blocks until one or more of the events registered with poll_fd_wait()
262 * occurs, or until the minimum duration registered with poll_timer_wait()
263 * elapses, or not at all if poll_immediate_wake() has been called. */
267 struct poll_loop *loop = poll_loop();
268 struct poll_node *node;
269 struct pollfd *pollfds;
270 HANDLE *wevents = NULL;
275 /* Register fatal signal events before actually doing any real work for
279 if (loop->timeout_when == LLONG_MIN) {
280 COVERAGE_INC(poll_zero_timeout);
284 pollfds = xmalloc(hmap_count(&loop->poll_nodes) * sizeof *pollfds);
287 wevents = xmalloc(hmap_count(&loop->poll_nodes) * sizeof *wevents);
290 /* Populate with all the fds and events. */
292 HMAP_FOR_EACH (node, hmap_node, &loop->poll_nodes) {
293 pollfds[i] = node->pollfd;
295 wevents[i] = node->wevent;
300 retval = time_poll(pollfds, hmap_count(&loop->poll_nodes), wevents,
301 loop->timeout_when, &elapsed);
303 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
304 VLOG_ERR_RL(&rl, "poll: %s", ovs_strerror(-retval));
305 } else if (!retval) {
306 log_wakeup(loop->timeout_where, NULL, elapsed);
307 } else if (get_cpu_usage() > 50 || VLOG_IS_DBG_ENABLED()) {
309 HMAP_FOR_EACH (node, hmap_node, &loop->poll_nodes) {
310 if (pollfds[i].revents) {
311 log_wakeup(node->where, &pollfds[i], 0);
317 free_poll_nodes(loop);
318 loop->timeout_when = LLONG_MAX;
319 loop->timeout_where = NULL;
323 /* Handle any pending signals before doing anything else. */
330 free_poll_loop(void *loop_)
332 struct poll_loop *loop = loop_;
334 free_poll_nodes(loop);
335 hmap_destroy(&loop->poll_nodes);
339 static struct poll_loop *
342 static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
343 static pthread_key_t key;
344 struct poll_loop *loop;
346 if (ovsthread_once_start(&once)) {
347 xpthread_key_create(&key, free_poll_loop);
348 ovsthread_once_done(&once);
351 loop = pthread_getspecific(key);
353 loop = xzalloc(sizeof *loop);
354 hmap_init(&loop->poll_nodes);
355 xpthread_setspecific(key, loop);