lib/poll-loop.c

   1 /*
   2  * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013 Nicira, Inc.
   3  *
   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:
   7  *
   8  *     http://www.apache.org/licenses/LICENSE-2.0
   9  *
  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.
  15  */
  16
  17 #include <config.h>
  18 #include "poll-loop.h"
  19 #include <errno.h>
  20 #include <inttypes.h>
  21 #include <poll.h>
  22 #include <stdlib.h>
  23 #include <string.h>
  24 #include "coverage.h"
  25 #include "dynamic-string.h"
  26 #include "fatal-signal.h"
  27 #include "list.h"
  28 #include "ovs-thread.h"
  29 #include "socket-util.h"
  30 #include "timeval.h"
  31 #include "vlog.h"
  32
  33 VLOG_DEFINE_THIS_MODULE(poll_loop);
  34
  35 COVERAGE_DEFINE(poll_fd_wait);
  36 COVERAGE_DEFINE(poll_zero_timeout);
  37
  38 struct poll_loop {
  39     /* All active poll waiters. */
  40     struct pollfd *pollfds;     /* Events to pass to poll(). */
  41     const char **where;         /* Where each pollfd was created. */
  42     size_t n_waiters;           /* Number of elems in 'where' and 'pollfds'. */
  43     size_t allocated_waiters;   /* Allocated elems in 'where' and 'pollfds'. */
  44
  45     /* Time at which to wake up the next call to poll_block(), LLONG_MIN to
  46      * wake up immediately, or LLONG_MAX to wait forever. */
  47     long long int timeout_when; /* In msecs as returned by time_msec(). */
  48     const char *timeout_where;  /* Where 'timeout_when' was set. */
  49 };
  50
  51 static struct poll_loop *poll_loop(void);
  52
  53 /* Registers 'fd' as waiting for the specified 'events' (which should be POLLIN
  54  * or POLLOUT or POLLIN | POLLOUT).  The following call to poll_block() will
  55  * wake up when 'fd' becomes ready for one or more of the requested events.
  56  *
  57  * The event registration is one-shot: only the following call to poll_block()
  58  * is affected.  The event will need to be re-registered after poll_block() is
  59  * called if it is to persist.
  60  *
  61  * ('where' is used in debug logging.  Commonly one would use poll_fd_wait() to
  62  * automatically provide the caller's source file and line number for
  63  * 'where'.) */
  64 void
  65 poll_fd_wait_at(int fd, short int events, const char *where)
  66 {
  67     struct poll_loop *loop = poll_loop();
  68
  69     COVERAGE_INC(poll_fd_wait);
  70     if (loop->n_waiters >= loop->allocated_waiters) {
  71         loop->where = x2nrealloc(loop->where, &loop->allocated_waiters,
  72                                  sizeof *loop->where);
  73         loop->pollfds = xrealloc(loop->pollfds,
  74                                  (loop->allocated_waiters
  75                                   * sizeof *loop->pollfds));
  76     }
  77
  78     loop->where[loop->n_waiters] = where;
  79     loop->pollfds[loop->n_waiters].fd = fd;
  80     loop->pollfds[loop->n_waiters].events = events;
  81     loop->n_waiters++;
  82 }
  83
  84 /* Causes the following call to poll_block() to block for no more than 'msec'
  85  * milliseconds.  If 'msec' is nonpositive, the following call to poll_block()
  86  * will not block at all.
  87  *
  88  * The timer registration is one-shot: only the following call to poll_block()
  89  * is affected.  The timer will need to be re-registered after poll_block() is
  90  * called if it is to persist.
  91  *
  92  * ('where' is used in debug logging.  Commonly one would use poll_timer_wait()
  93  * to automatically provide the caller's source file and line number for
  94  * 'where'.) */
  95 void
  96 poll_timer_wait_at(long long int msec, const char *where)
  97 {
  98     long long int now = time_msec();
  99     long long int when;
 100
 101     if (msec <= 0) {
 102         /* Wake up immediately. */
 103         when = LLONG_MIN;
 104     } else if ((unsigned long long int) now + msec <= LLONG_MAX) {
 105         /* Normal case. */
 106         when = now + msec;
 107     } else {
 108         /* now + msec would overflow. */
 109         when = LLONG_MAX;
 110     }
 111
 112     poll_timer_wait_until_at(when, where);
 113 }
 114
 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
 118  * all.
 119  *
 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.
 123  *
 124  * ('where' is used in debug logging.  Commonly one would use
 125  * poll_timer_wait_until() to automatically provide the caller's source file
 126  * and line number for 'where'.) */
 127 void
 128 poll_timer_wait_until_at(long long int when, const char *where)
 129 {
 130     struct poll_loop *loop = poll_loop();
 131     if (when < loop->timeout_when) {
 132         loop->timeout_when = when;
 133         loop->timeout_where = where;
 134     }
 135 }
 136
 137 /* Causes the following call to poll_block() to wake up immediately, without
 138  * blocking.
 139  *
 140  * ('where' is used in debug logging.  Commonly one would use
 141  * poll_immediate_wake() to automatically provide the caller's source file and
 142  * line number for 'where'.) */
 143 void
 144 poll_immediate_wake_at(const char *where)
 145 {
 146     poll_timer_wait_at(0, where);
 147 }
 148
 149 /* Logs, if appropriate, that the poll loop was awakened by an event
 150  * registered at 'where' (typically a source file and line number).  The other
 151  * arguments have two possible interpretations:
 152  *
 153  *   - If 'pollfd' is nonnull then it should be the "struct pollfd" that caused
 154  *     the wakeup.  'timeout' is ignored.
 155  *
 156  *   - If 'pollfd' is NULL then 'timeout' is the number of milliseconds after
 157  *     which the poll loop woke up.
 158  */
 159 static void
 160 log_wakeup(const char *where, const struct pollfd *pollfd, int timeout)
 161 {
 162     static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(10, 10);
 163     enum vlog_level level;
 164     int cpu_usage;
 165     struct ds s;
 166
 167     cpu_usage = get_cpu_usage();
 168     if (VLOG_IS_DBG_ENABLED()) {
 169         level = VLL_DBG;
 170     } else if (cpu_usage > 50 && !VLOG_DROP_INFO(&rl)) {
 171         level = VLL_INFO;
 172     } else {
 173         return;
 174     }
 175
 176     ds_init(&s);
 177     ds_put_cstr(&s, "wakeup due to ");
 178     if (pollfd) {
 179         char *description = describe_fd(pollfd->fd);
 180         if (pollfd->revents & POLLIN) {
 181             ds_put_cstr(&s, "[POLLIN]");
 182         }
 183         if (pollfd->revents & POLLOUT) {
 184             ds_put_cstr(&s, "[POLLOUT]");
 185         }
 186         if (pollfd->revents & POLLERR) {
 187             ds_put_cstr(&s, "[POLLERR]");
 188         }
 189         if (pollfd->revents & POLLHUP) {
 190             ds_put_cstr(&s, "[POLLHUP]");
 191         }
 192         if (pollfd->revents & POLLNVAL) {
 193             ds_put_cstr(&s, "[POLLNVAL]");
 194         }
 195         ds_put_format(&s, " on fd %d (%s)", pollfd->fd, description);
 196         free(description);
 197     } else {
 198         ds_put_format(&s, "%d-ms timeout", timeout);
 199     }
 200     if (where) {
 201         ds_put_format(&s, " at %s", where);
 202     }
 203     if (cpu_usage >= 0) {
 204         ds_put_format(&s, " (%d%% CPU usage)", cpu_usage);
 205     }
 206     VLOG(level, "%s", ds_cstr(&s));
 207     ds_destroy(&s);
 208 }
 209
 210 /* Blocks until one or more of the events registered with poll_fd_wait()
 211  * occurs, or until the minimum duration registered with poll_timer_wait()
 212  * elapses, or not at all if poll_immediate_wake() has been called. */
 213 void
 214 poll_block(void)
 215 {
 216     struct poll_loop *loop = poll_loop();
 217     int elapsed;
 218     int retval;
 219
 220     /* Register fatal signal events before actually doing any real work for
 221      * poll_block. */
 222     fatal_signal_wait();
 223
 224     if (loop->timeout_when == LLONG_MIN) {
 225         COVERAGE_INC(poll_zero_timeout);
 226     }
 227
 228     retval = time_poll(loop->pollfds, loop->n_waiters,
 229                        loop->timeout_when, &elapsed);
 230     if (retval < 0) {
 231         static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
 232         VLOG_ERR_RL(&rl, "poll: %s", ovs_strerror(-retval));
 233     } else if (!retval) {
 234         log_wakeup(loop->timeout_where, NULL, elapsed);
 235     } else if (get_cpu_usage() > 50 || VLOG_IS_DBG_ENABLED()) {
 236         size_t i;
 237
 238         for (i = 0; i < loop->n_waiters; i++) {
 239             if (loop->pollfds[i].revents) {
 240                 log_wakeup(loop->where[i], &loop->pollfds[i], 0);
 241             }
 242         }
 243     }
 244
 245     loop->timeout_when = LLONG_MAX;
 246     loop->timeout_where = NULL;
 247     loop->n_waiters = 0;
 248
 249     /* Handle any pending signals before doing anything else. */
 250     fatal_signal_run();
 251 }
 252 \f
 253 static void
 254 free_poll_loop(void *loop_)
 255 {
 256     struct poll_loop *loop = loop_;
 257
 258     free(loop->pollfds);
 259     free(loop->where);
 260     free(loop);
 261 }
 262
 263 static struct poll_loop *
 264 poll_loop(void)
 265 {
 266     static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
 267     static pthread_key_t key;
 268     struct poll_loop *loop;
 269
 270     if (ovsthread_once_start(&once)) {
 271         xpthread_key_create(&key, free_poll_loop);
 272         ovsthread_once_done(&once);
 273     }
 274
 275     loop = pthread_getspecific(key);
 276     if (!loop) {
 277         loop = xzalloc(sizeof *loop);
 278         xpthread_setspecific(key, loop);
 279     }
 280     return loop;
 281 }
 282