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 "openvswitch/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 "openvswitch/dynamic-string.h"
  26 #include "fatal-signal.h"
  27 #include "openvswitch/list.h"
  28 #include "ovs-thread.h"
  29 #include "seq.h"
  30 #include "socket-util.h"
  31 #include "timeval.h"
  32 #include "openvswitch/vlog.h"
  33 #include "openvswitch/hmap.h"
  34 #include "hash.h"
  35
  36 VLOG_DEFINE_THIS_MODULE(poll_loop);
  37
  38 COVERAGE_DEFINE(poll_create_node);
  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 or wevent. */
  61 static struct poll_node *
  62 find_poll_node(struct poll_loop *loop, int fd, HANDLE wevent)
  63 {
  64     struct poll_node *node;
  65
  66     /* Both 'fd' and 'wevent' cannot be set. */
  67     ovs_assert(!fd != !wevent);
  68
  69     HMAP_FOR_EACH_WITH_HASH (node, hmap_node,
  70                              hash_2words(fd, (uint32_t)wevent),
  71                              &loop->poll_nodes) {
  72         if ((fd && node->pollfd.fd == fd)
  73             || (wevent && node->wevent == wevent)) {
  74             return node;
  75         }
  76     }
  77     return NULL;
  78 }
  79
  80 /* On Unix based systems:
  81  *
  82  *     Registers 'fd' as waiting for the specified 'events' (which should be
  83  *     POLLIN or POLLOUT or POLLIN | POLLOUT).  The following call to
  84  *     poll_block() will wake up when 'fd' becomes ready for one or more of the
  85  *     requested events. The 'fd's are given to poll() function later.
  86  *
  87  * On Windows system:
  88  *
  89  *     If 'fd' is specified, create a new 'wevent'. Association of 'fd' and
  90  *     'wevent' for 'events' happens in poll_block(). If 'wevent' is specified,
  91  *     it is assumed that it is unrelated to any sockets and poll_block()
  92  *     will wake up on any event on that 'wevent'. It is an error to pass
  93  *     both 'wevent' and 'fd'.
  94  *
  95  * The event registration is one-shot: only the following call to
  96  * poll_block() is affected.  The event will need to be re-registered after
  97  * poll_block() is called if it is to persist.
  98  *
  99  * ('where' is used in debug logging.  Commonly one would use poll_fd_wait() to
 100  * automatically provide the caller's source file and line number for
 101  * 'where'.) */
 102 static void
 103 poll_create_node(int fd, HANDLE wevent, short int events, const char *where)
 104 {
 105     struct poll_loop *loop = poll_loop();
 106     struct poll_node *node;
 107
 108     COVERAGE_INC(poll_create_node);
 109
 110     /* Both 'fd' and 'wevent' cannot be set. */
 111     ovs_assert(!fd != !wevent);
 112
 113     /* Check for duplicate.  If found, "or" the events. */
 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, (uint32_t)wevent));
 121         node->pollfd.fd = fd;
 122         node->pollfd.events = events;
 123 #ifdef _WIN32
 124         if (!wevent) {
 125             wevent = CreateEvent(NULL, FALSE, FALSE, NULL);
 126         }
 127 #endif
 128         node->wevent = wevent;
 129         node->where = where;
 130     }
 131 }
 132
 133 /* Registers 'fd' as waiting for the specified 'events' (which should be POLLIN
 134  * or POLLOUT or POLLIN | POLLOUT).  The following call to poll_block() will
 135  * wake up when 'fd' becomes ready for one or more of the requested events.
 136  *
 137  * On Windows, 'fd' must be a socket.
 138  *
 139  * The event registration is one-shot: only the following call to poll_block()
 140  * is affected.  The event will need to be re-registered after poll_block() is
 141  * called if it is to persist.
 142  *
 143  * ('where' is used in debug logging.  Commonly one would use poll_fd_wait() to
 144  * automatically provide the caller's source file and line number for
 145  * 'where'.) */
 146 void
 147 poll_fd_wait_at(int fd, short int events, const char *where)
 148 {
 149     poll_create_node(fd, 0, events, where);
 150 }
 151
 152 #ifdef _WIN32
 153 /* Registers for the next call to poll_block() to wake up when 'wevent' is
 154  * signaled.
 155  *
 156  * The event registration is one-shot: only the following call to poll_block()
 157  * is affected.  The event will need to be re-registered after poll_block() is
 158  * called if it is to persist.
 159  *
 160  * ('where' is used in debug logging.  Commonly one would use
 161  * poll_wevent_wait() to automatically provide the caller's source file and
 162  * line number for 'where'.) */
 163 void
 164 poll_wevent_wait_at(HANDLE wevent, const char *where)
 165 {
 166     poll_create_node(0, wevent, 0, where);
 167 }
 168 #endif /* _WIN32 */
 169
 170 /* Causes the following call to poll_block() to block for no more than 'msec'
 171  * milliseconds.  If 'msec' is nonpositive, the following call to poll_block()
 172  * will not block at all.
 173  *
 174  * The timer registration is one-shot: only the following call to poll_block()
 175  * is affected.  The timer will need to be re-registered after poll_block() is
 176  * called if it is to persist.
 177  *
 178  * ('where' is used in debug logging.  Commonly one would use poll_timer_wait()
 179  * to automatically provide the caller's source file and line number for
 180  * 'where'.) */
 181 void
 182 poll_timer_wait_at(long long int msec, const char *where)
 183 {
 184     long long int now = time_msec();
 185     long long int when;
 186
 187     if (msec <= 0) {
 188         /* Wake up immediately. */
 189         when = LLONG_MIN;
 190     } else if ((unsigned long long int) now + msec <= LLONG_MAX) {
 191         /* Normal case. */
 192         when = now + msec;
 193     } else {
 194         /* now + msec would overflow. */
 195         when = LLONG_MAX;
 196     }
 197
 198     poll_timer_wait_until_at(when, where);
 199 }
 200
 201 /* Causes the following call to poll_block() to wake up when the current time,
 202  * as returned by time_msec(), reaches 'when' or later.  If 'when' is earlier
 203  * than the current time, the following call to poll_block() will not block at
 204  * all.
 205  *
 206  * The timer registration is one-shot: only the following call to poll_block()
 207  * is affected.  The timer will need to be re-registered after poll_block() is
 208  * called if it is to persist.
 209  *
 210  * ('where' is used in debug logging.  Commonly one would use
 211  * poll_timer_wait_until() to automatically provide the caller's source file
 212  * and line number for 'where'.) */
 213 void
 214 poll_timer_wait_until_at(long long int when, const char *where)
 215 {
 216     struct poll_loop *loop = poll_loop();
 217     if (when < loop->timeout_when) {
 218         loop->timeout_when = when;
 219         loop->timeout_where = where;
 220     }
 221 }
 222
 223 /* Causes the following call to poll_block() to wake up immediately, without
 224  * blocking.
 225  *
 226  * ('where' is used in debug logging.  Commonly one would use
 227  * poll_immediate_wake() to automatically provide the caller's source file and
 228  * line number for 'where'.) */
 229 void
 230 poll_immediate_wake_at(const char *where)
 231 {
 232     poll_timer_wait_at(0, where);
 233 }
 234
 235 /* Logs, if appropriate, that the poll loop was awakened by an event
 236  * registered at 'where' (typically a source file and line number).  The other
 237  * arguments have two possible interpretations:
 238  *
 239  *   - If 'pollfd' is nonnull then it should be the "struct pollfd" that caused
 240  *     the wakeup.  'timeout' is ignored.
 241  *
 242  *   - If 'pollfd' is NULL then 'timeout' is the number of milliseconds after
 243  *     which the poll loop woke up.
 244  */
 245 static void
 246 log_wakeup(const char *where, const struct pollfd *pollfd, int timeout)
 247 {
 248     static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(10, 10);
 249     enum vlog_level level;
 250     int cpu_usage;
 251     struct ds s;
 252
 253     cpu_usage = get_cpu_usage();
 254     if (VLOG_IS_DBG_ENABLED()) {
 255         level = VLL_DBG;
 256     } else if (cpu_usage > 50
 257                && !thread_is_pmd()
 258                && !VLOG_DROP_INFO(&rl)) {
 259         level = VLL_INFO;
 260     } else {
 261         return;
 262     }
 263
 264     ds_init(&s);
 265     ds_put_cstr(&s, "wakeup due to ");
 266     if (pollfd) {
 267         char *description = describe_fd(pollfd->fd);
 268         if (pollfd->revents & POLLIN) {
 269             ds_put_cstr(&s, "[POLLIN]");
 270         }
 271         if (pollfd->revents & POLLOUT) {
 272             ds_put_cstr(&s, "[POLLOUT]");
 273         }
 274         if (pollfd->revents & POLLERR) {
 275             ds_put_cstr(&s, "[POLLERR]");
 276         }
 277         if (pollfd->revents & POLLHUP) {
 278             ds_put_cstr(&s, "[POLLHUP]");
 279         }
 280         if (pollfd->revents & POLLNVAL) {
 281             ds_put_cstr(&s, "[POLLNVAL]");
 282         }
 283         ds_put_format(&s, " on fd %d (%s)", pollfd->fd, description);
 284         free(description);
 285     } else {
 286         ds_put_format(&s, "%d-ms timeout", timeout);
 287     }
 288     if (where) {
 289         ds_put_format(&s, " at %s", where);
 290     }
 291     if (cpu_usage >= 0) {
 292         ds_put_format(&s, " (%d%% CPU usage)", cpu_usage);
 293     }
 294     VLOG(level, "%s", ds_cstr(&s));
 295     ds_destroy(&s);
 296 }
 297
 298 static void
 299 free_poll_nodes(struct poll_loop *loop)
 300 {
 301     struct poll_node *node, *next;
 302
 303     HMAP_FOR_EACH_SAFE (node, next, hmap_node, &loop->poll_nodes) {
 304         hmap_remove(&loop->poll_nodes, &node->hmap_node);
 305 #ifdef _WIN32
 306         if (node->wevent && node->pollfd.fd) {
 307             WSAEventSelect(node->pollfd.fd, NULL, 0);
 308             CloseHandle(node->wevent);
 309         }
 310 #endif
 311         free(node);
 312     }
 313 }
 314
 315 /* Blocks until one or more of the events registered with poll_fd_wait()
 316  * occurs, or until the minimum duration registered with poll_timer_wait()
 317  * elapses, or not at all if poll_immediate_wake() has been called. */
 318 void
 319 poll_block(void)
 320 {
 321     struct poll_loop *loop = poll_loop();
 322     struct poll_node *node;
 323     struct pollfd *pollfds;
 324     HANDLE *wevents = NULL;
 325     int elapsed;
 326     int retval;
 327     int i;
 328
 329     /* Register fatal signal events before actually doing any real work for
 330      * poll_block. */
 331     fatal_signal_wait();
 332
 333     if (loop->timeout_when == LLONG_MIN) {
 334         COVERAGE_INC(poll_zero_timeout);
 335     }
 336
 337     timewarp_run();
 338     pollfds = xmalloc(hmap_count(&loop->poll_nodes) * sizeof *pollfds);
 339
 340 #ifdef _WIN32
 341     wevents = xmalloc(hmap_count(&loop->poll_nodes) * sizeof *wevents);
 342 #endif
 343
 344     /* Populate with all the fds and events. */
 345     i = 0;
 346     HMAP_FOR_EACH (node, hmap_node, &loop->poll_nodes) {
 347         pollfds[i] = node->pollfd;
 348 #ifdef _WIN32
 349         wevents[i] = node->wevent;
 350         if (node->pollfd.fd && node->wevent) {
 351             short int wsa_events = 0;
 352             if (node->pollfd.events & POLLIN) {
 353                 wsa_events |= FD_READ | FD_ACCEPT | FD_CLOSE;
 354             }
 355             if (node->pollfd.events & POLLOUT) {
 356                 wsa_events |= FD_WRITE | FD_CONNECT | FD_CLOSE;
 357             }
 358             WSAEventSelect(node->pollfd.fd, node->wevent, wsa_events);
 359         }
 360 #endif
 361         i++;
 362     }
 363
 364     retval = time_poll(pollfds, hmap_count(&loop->poll_nodes), wevents,
 365                        loop->timeout_when, &elapsed);
 366     if (retval < 0) {
 367         static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
 368         VLOG_ERR_RL(&rl, "poll: %s", ovs_strerror(-retval));
 369     } else if (!retval) {
 370         log_wakeup(loop->timeout_where, NULL, elapsed);
 371     } else if (get_cpu_usage() > 50 || VLOG_IS_DBG_ENABLED()) {
 372         i = 0;
 373         HMAP_FOR_EACH (node, hmap_node, &loop->poll_nodes) {
 374             if (pollfds[i].revents) {
 375                 log_wakeup(node->where, &pollfds[i], 0);
 376             }
 377             i++;
 378         }
 379     }
 380
 381     free_poll_nodes(loop);
 382     loop->timeout_when = LLONG_MAX;
 383     loop->timeout_where = NULL;
 384     free(pollfds);
 385     free(wevents);
 386
 387     /* Handle any pending signals before doing anything else. */
 388     fatal_signal_run();
 389
 390     seq_woke();
 391 }
 392 \f
 393 static void
 394 free_poll_loop(void *loop_)
 395 {
 396     struct poll_loop *loop = loop_;
 397
 398     free_poll_nodes(loop);
 399     hmap_destroy(&loop->poll_nodes);
 400     free(loop);
 401 }
 402
 403 static struct poll_loop *
 404 poll_loop(void)
 405 {
 406     static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
 407     static pthread_key_t key;
 408     struct poll_loop *loop;
 409
 410     if (ovsthread_once_start(&once)) {
 411         xpthread_key_create(&key, free_poll_loop);
 412         ovsthread_once_done(&once);
 413     }
 414
 415     loop = pthread_getspecific(key);
 416     if (!loop) {
 417         loop = xzalloc(sizeof *loop);
 418         loop->timeout_when = LLONG_MAX;
 419         hmap_init(&loop->poll_nodes);
 420         xpthread_setspecific(key, loop);
 421     }
 422     return loop;
 423 }
 424