lib/poll-loop.c

   1 /*
   2  * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014 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 "seq.h"
  30 #include "socket-util.h"
  31 #include "timeval.h"
  32 #include "vlog.h"
  33 #include "hmap.h"
  34 #include "hash.h"
  35
  36 VLOG_DEFINE_THIS_MODULE(poll_loop);
  37
  38 COVERAGE_DEFINE(poll_fd_wait);
  39 COVERAGE_DEFINE(poll_zero_timeout);
  40
  41 struct poll_node {
  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. */
  46 };
  47
  48 struct poll_loop {
  49     /* All active poll waiters. */
  50     struct hmap poll_nodes;
  51
  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. */
  56 };
  57
  58 static struct poll_loop *poll_loop(void);
  59
  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)
  63 {
  64     struct poll_node *node;
  65
  66     HMAP_FOR_EACH_WITH_HASH (node, hmap_node, hash_2words(fd, wevent),
  67                              &loop->poll_nodes) {
  68         if (node->pollfd.fd == fd && node->wevent == wevent) {
  69             return node;
  70         }
  71     }
  72     return NULL;
  73 }
  74
  75 /* On Unix based systems:
  76  *
  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.
  81  *
  82  * On Windows system:
  83  *
  84  *     If both 'wevent' handle and 'fd' is specified, associate the 'fd' with
  85  *     with that 'wevent' for 'events' (implemented in poll_block()).
  86  *     In case of no 'fd' specified, wake up on any event on that 'wevent'.
  87  *     These wevents are given to the WaitForMultipleObjects() to be polled.
  88  *     The event registration is one-shot: only the following call to
  89  *     poll_block() is affected.  The event will need to be re-registered after
  90  *     poll_block() is called if it is to persist.
  91  *
  92  * ('where' is used in debug logging.  Commonly one would use poll_fd_wait() to
  93  * automatically provide the caller's source file and line number for
  94  * 'where'.) */
  95 void
  96 poll_fd_wait_at(int fd, HANDLE wevent, short int events, const char *where)
  97 {
  98     struct poll_loop *loop = poll_loop();
  99     struct poll_node *node;
 100
 101     COVERAGE_INC(poll_fd_wait);
 102
 103 #ifdef _WIN32
 104     /* Null event cannot be polled. */
 105     if (wevent == 0) {
 106         VLOG_ERR("No event to wait fd %d", fd);
 107         return;
 108     }
 109 #else
 110     wevent = 0;
 111 #endif
 112
 113     /* Check for duplicate.  If found, "or" the event. */
 114     node = find_poll_node(loop, fd, wevent);
 115     if (node) {
 116         node->pollfd.events |= events;
 117     } else {
 118         node = xzalloc(sizeof *node);
 119         hmap_insert(&loop->poll_nodes, &node->hmap_node,
 120                     hash_2words(fd, wevent));
 121         node->pollfd.fd = fd;
 122         node->pollfd.events = events;
 123         node->wevent = wevent;
 124         node->where = where;
 125     }
 126 }
 127
 128 /* Causes the following call to poll_block() to block for no more than 'msec'
 129  * milliseconds.  If 'msec' is nonpositive, the following call to poll_block()
 130  * will not block at all.
 131  *
 132  * The timer registration is one-shot: only the following call to poll_block()
 133  * is affected.  The timer will need to be re-registered after poll_block() is
 134  * called if it is to persist.
 135  *
 136  * ('where' is used in debug logging.  Commonly one would use poll_timer_wait()
 137  * to automatically provide the caller's source file and line number for
 138  * 'where'.) */
 139 void
 140 poll_timer_wait_at(long long int msec, const char *where)
 141 {
 142     long long int now = time_msec();
 143     long long int when;
 144
 145     if (msec <= 0) {
 146         /* Wake up immediately. */
 147         when = LLONG_MIN;
 148     } else if ((unsigned long long int) now + msec <= LLONG_MAX) {
 149         /* Normal case. */
 150         when = now + msec;
 151     } else {
 152         /* now + msec would overflow. */
 153         when = LLONG_MAX;
 154     }
 155
 156     poll_timer_wait_until_at(when, where);
 157 }
 158
 159 /* Causes the following call to poll_block() to wake up when the current time,
 160  * as returned by time_msec(), reaches 'when' or later.  If 'when' is earlier
 161  * than the current time, the following call to poll_block() will not block at
 162  * all.
 163  *
 164  * The timer registration is one-shot: only the following call to poll_block()
 165  * is affected.  The timer will need to be re-registered after poll_block() is
 166  * called if it is to persist.
 167  *
 168  * ('where' is used in debug logging.  Commonly one would use
 169  * poll_timer_wait_until() to automatically provide the caller's source file
 170  * and line number for 'where'.) */
 171 void
 172 poll_timer_wait_until_at(long long int when, const char *where)
 173 {
 174     struct poll_loop *loop = poll_loop();
 175     if (when < loop->timeout_when) {
 176         loop->timeout_when = when;
 177         loop->timeout_where = where;
 178     }
 179 }
 180
 181 /* Causes the following call to poll_block() to wake up immediately, without
 182  * blocking.
 183  *
 184  * ('where' is used in debug logging.  Commonly one would use
 185  * poll_immediate_wake() to automatically provide the caller's source file and
 186  * line number for 'where'.) */
 187 void
 188 poll_immediate_wake_at(const char *where)
 189 {
 190     poll_timer_wait_at(0, where);
 191 }
 192
 193 /* Logs, if appropriate, that the poll loop was awakened by an event
 194  * registered at 'where' (typically a source file and line number).  The other
 195  * arguments have two possible interpretations:
 196  *
 197  *   - If 'pollfd' is nonnull then it should be the "struct pollfd" that caused
 198  *     the wakeup.  'timeout' is ignored.
 199  *
 200  *   - If 'pollfd' is NULL then 'timeout' is the number of milliseconds after
 201  *     which the poll loop woke up.
 202  */
 203 static void
 204 log_wakeup(const char *where, const struct pollfd *pollfd, int timeout)
 205 {
 206     static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(10, 10);
 207     enum vlog_level level;
 208     int cpu_usage;
 209     struct ds s;
 210
 211     cpu_usage = get_cpu_usage();
 212     if (VLOG_IS_DBG_ENABLED()) {
 213         level = VLL_DBG;
 214     } else if (cpu_usage > 50 && !VLOG_DROP_INFO(&rl)) {
 215         level = VLL_INFO;
 216     } else {
 217         return;
 218     }
 219
 220     ds_init(&s);
 221     ds_put_cstr(&s, "wakeup due to ");
 222     if (pollfd) {
 223         char *description = describe_fd(pollfd->fd);
 224         if (pollfd->revents & POLLIN) {
 225             ds_put_cstr(&s, "[POLLIN]");
 226         }
 227         if (pollfd->revents & POLLOUT) {
 228             ds_put_cstr(&s, "[POLLOUT]");
 229         }
 230         if (pollfd->revents & POLLERR) {
 231             ds_put_cstr(&s, "[POLLERR]");
 232         }
 233         if (pollfd->revents & POLLHUP) {
 234             ds_put_cstr(&s, "[POLLHUP]");
 235         }
 236         if (pollfd->revents & POLLNVAL) {
 237             ds_put_cstr(&s, "[POLLNVAL]");
 238         }
 239         ds_put_format(&s, " on fd %d (%s)", pollfd->fd, description);
 240         free(description);
 241     } else {
 242         ds_put_format(&s, "%d-ms timeout", timeout);
 243     }
 244     if (where) {
 245         ds_put_format(&s, " at %s", where);
 246     }
 247     if (cpu_usage >= 0) {
 248         ds_put_format(&s, " (%d%% CPU usage)", cpu_usage);
 249     }
 250     VLOG(level, "%s", ds_cstr(&s));
 251     ds_destroy(&s);
 252 }
 253
 254 static void
 255 free_poll_nodes(struct poll_loop *loop)
 256 {
 257     struct poll_node *node, *next;
 258
 259     HMAP_FOR_EACH_SAFE (node, next, hmap_node, &loop->poll_nodes) {
 260         hmap_remove(&loop->poll_nodes, &node->hmap_node);
 261         free(node);
 262     }
 263 }
 264
 265 /* Blocks until one or more of the events registered with poll_fd_wait()
 266  * occurs, or until the minimum duration registered with poll_timer_wait()
 267  * elapses, or not at all if poll_immediate_wake() has been called. */
 268 void
 269 poll_block(void)
 270 {
 271     struct poll_loop *loop = poll_loop();
 272     struct poll_node *node;
 273     struct pollfd *pollfds;
 274     HANDLE *wevents = NULL;
 275     int elapsed;
 276     int retval;
 277     int i;
 278
 279     /* Register fatal signal events before actually doing any real work for
 280      * poll_block. */
 281     fatal_signal_wait();
 282
 283     if (loop->timeout_when == LLONG_MIN) {
 284         COVERAGE_INC(poll_zero_timeout);
 285     }
 286
 287     timewarp_run();
 288     pollfds = xmalloc(hmap_count(&loop->poll_nodes) * sizeof *pollfds);
 289
 290 #ifdef _WIN32
 291     wevents = xmalloc(hmap_count(&loop->poll_nodes) * sizeof *wevents);
 292 #endif
 293
 294     /* Populate with all the fds and events. */
 295     i = 0;
 296     HMAP_FOR_EACH (node, hmap_node, &loop->poll_nodes) {
 297         pollfds[i] = node->pollfd;
 298 #ifdef _WIN32
 299         wevents[i] = node->wevent;
 300         if (node->pollfd.fd && node->wevent) {
 301             short int wsa_events = 0;
 302             if (node->pollfd.events & POLLIN) {
 303                 wsa_events |= FD_READ | FD_ACCEPT | FD_CLOSE;
 304             }
 305             if (node->pollfd.events & POLLOUT) {
 306                 wsa_events |= FD_WRITE | FD_CONNECT | FD_CLOSE;
 307             }
 308             WSAEventSelect(node->pollfd.fd, node->wevent, wsa_events);
 309         }
 310 #endif
 311         i++;
 312     }
 313
 314     retval = time_poll(pollfds, hmap_count(&loop->poll_nodes), wevents,
 315                        loop->timeout_when, &elapsed);
 316     if (retval < 0) {
 317         static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
 318         VLOG_ERR_RL(&rl, "poll: %s", ovs_strerror(-retval));
 319     } else if (!retval) {
 320         log_wakeup(loop->timeout_where, NULL, elapsed);
 321     } else if (get_cpu_usage() > 50 || VLOG_IS_DBG_ENABLED()) {
 322         i = 0;
 323         HMAP_FOR_EACH (node, hmap_node, &loop->poll_nodes) {
 324             if (pollfds[i].revents) {
 325                 log_wakeup(node->where, &pollfds[i], 0);
 326             }
 327             i++;
 328         }
 329     }
 330
 331     free_poll_nodes(loop);
 332     loop->timeout_when = LLONG_MAX;
 333     loop->timeout_where = NULL;
 334     free(pollfds);
 335     free(wevents);
 336
 337     /* Handle any pending signals before doing anything else. */
 338     fatal_signal_run();
 339
 340     seq_woke();
 341 }
 342 \f
 343 static void
 344 free_poll_loop(void *loop_)
 345 {
 346     struct poll_loop *loop = loop_;
 347
 348     free_poll_nodes(loop);
 349     hmap_destroy(&loop->poll_nodes);
 350     free(loop);
 351 }
 352
 353 static struct poll_loop *
 354 poll_loop(void)
 355 {
 356     static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
 357     static pthread_key_t key;
 358     struct poll_loop *loop;
 359
 360     if (ovsthread_once_start(&once)) {
 361         xpthread_key_create(&key, free_poll_loop);
 362         ovsthread_once_done(&once);
 363     }
 364
 365     loop = pthread_getspecific(key);
 366     if (!loop) {
 367         loop = xzalloc(sizeof *loop);
 368         hmap_init(&loop->poll_nodes);
 369         xpthread_setspecific(key, loop);
 370     }
 371     return loop;
 372 }
 373