Branch data Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * waiteventset.c
4 : : * ppoll()/pselect() like abstraction
5 : : *
6 : : * WaitEvents are an abstraction for waiting for one or more events at a time.
7 : : * The waiting can be done in a race free fashion, similar ppoll() or
8 : : * pselect() (as opposed to plain poll()/select()).
9 : : *
10 : : * You can wait for:
11 : : * - a latch being set from another process or from signal handler in the same
12 : : * process (WL_LATCH_SET)
13 : : * - data to become readable or writeable on a socket (WL_SOCKET_*)
14 : : * - postmaster death (WL_POSTMASTER_DEATH or WL_EXIT_ON_PM_DEATH)
15 : : * - timeout (WL_TIMEOUT)
16 : : *
17 : : * Implementation
18 : : * --------------
19 : : *
20 : : * The poll() implementation uses the so-called self-pipe trick to overcome the
21 : : * race condition involved with poll() and setting a global flag in the signal
22 : : * handler. When a latch is set and the current process is waiting for it, the
23 : : * signal handler wakes up the poll() in WaitLatch by writing a byte to a pipe.
24 : : * A signal by itself doesn't interrupt poll() on all platforms, and even on
25 : : * platforms where it does, a signal that arrives just before the poll() call
26 : : * does not prevent poll() from entering sleep. An incoming byte on a pipe
27 : : * however reliably interrupts the sleep, and causes poll() to return
28 : : * immediately even if the signal arrives before poll() begins.
29 : : *
30 : : * The epoll() implementation overcomes the race with a different technique: it
31 : : * keeps SIGURG blocked and consumes from a signalfd() descriptor instead. We
32 : : * don't need to register a signal handler or create our own self-pipe. We
33 : : * assume that any system that has Linux epoll() also has Linux signalfd().
34 : : *
35 : : * The kqueue() implementation waits for SIGURG with EVFILT_SIGNAL.
36 : : *
37 : : * The Windows implementation uses Windows events that are inherited by all
38 : : * postmaster child processes. There's no need for the self-pipe trick there.
39 : : *
40 : : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
41 : : * Portions Copyright (c) 1994, Regents of the University of California
42 : : *
43 : : * IDENTIFICATION
44 : : * src/backend/storage/ipc/waiteventset.c
45 : : *
46 : : *-------------------------------------------------------------------------
47 : : */
48 : : #include "postgres.h"
49 : :
50 : : #include <fcntl.h>
51 : : #include <limits.h>
52 : : #include <signal.h>
53 : : #include <unistd.h>
54 : : #ifdef HAVE_SYS_EPOLL_H
55 : : #include <sys/epoll.h>
56 : : #endif
57 : : #ifdef HAVE_SYS_EVENT_H
58 : : #include <sys/event.h>
59 : : #endif
60 : : #ifdef HAVE_SYS_SIGNALFD_H
61 : : #include <sys/signalfd.h>
62 : : #endif
63 : : #ifdef HAVE_POLL_H
64 : : #include <poll.h>
65 : : #endif
66 : :
67 : : #include "libpq/pqsignal.h"
68 : : #include "miscadmin.h"
69 : : #include "pgstat.h"
70 : : #include "port/atomics.h"
71 : : #include "portability/instr_time.h"
72 : : #include "postmaster/postmaster.h"
73 : : #include "storage/fd.h"
74 : : #include "storage/ipc.h"
75 : : #include "storage/pmsignal.h"
76 : : #include "storage/latch.h"
77 : : #include "storage/waiteventset.h"
78 : : #include "utils/memutils.h"
79 : : #include "utils/resowner.h"
80 : :
81 : : /*
82 : : * Select the fd readiness primitive to use. Normally the "most modern"
83 : : * primitive supported by the OS will be used, but for testing it can be
84 : : * useful to manually specify the used primitive. If desired, just add a
85 : : * define somewhere before this block.
86 : : */
87 : : #if defined(WAIT_USE_EPOLL) || defined(WAIT_USE_POLL) || \
88 : : defined(WAIT_USE_KQUEUE) || defined(WAIT_USE_WIN32)
89 : : /* don't overwrite manual choice */
90 : : #elif defined(HAVE_SYS_EPOLL_H)
91 : : #define WAIT_USE_EPOLL
92 : : #elif defined(HAVE_KQUEUE)
93 : : #define WAIT_USE_KQUEUE
94 : : #elif defined(HAVE_POLL)
95 : : #define WAIT_USE_POLL
96 : : #elif WIN32
97 : : #define WAIT_USE_WIN32
98 : : #else
99 : : #error "no wait set implementation available"
100 : : #endif
101 : :
102 : : /*
103 : : * By default, we use a self-pipe with poll() and a signalfd with epoll(), if
104 : : * available. For testing the choice can also be manually specified.
105 : : */
106 : : #if defined(WAIT_USE_POLL) || defined(WAIT_USE_EPOLL)
107 : : #if defined(WAIT_USE_SELF_PIPE) || defined(WAIT_USE_SIGNALFD)
108 : : /* don't overwrite manual choice */
109 : : #elif defined(WAIT_USE_EPOLL) && defined(HAVE_SYS_SIGNALFD_H)
110 : : #define WAIT_USE_SIGNALFD
111 : : #else
112 : : #define WAIT_USE_SELF_PIPE
113 : : #endif
114 : : #endif
115 : :
116 : : /* typedef in waiteventset.h */
117 : : struct WaitEventSet
118 : : {
119 : : ResourceOwner owner;
120 : :
121 : : int nevents; /* number of registered events */
122 : : int nevents_space; /* maximum number of events in this set */
123 : :
124 : : /*
125 : : * Array, of nevents_space length, storing the definition of events this
126 : : * set is waiting for.
127 : : */
128 : : WaitEvent *events;
129 : :
130 : : /*
131 : : * If WL_LATCH_SET is specified in any wait event, latch is a pointer to
132 : : * said latch, and latch_pos the offset in the ->events array. This is
133 : : * useful because we check the state of the latch before performing doing
134 : : * syscalls related to waiting.
135 : : */
136 : : Latch *latch;
137 : : int latch_pos;
138 : :
139 : : /*
140 : : * WL_EXIT_ON_PM_DEATH is converted to WL_POSTMASTER_DEATH, but this flag
141 : : * is set so that we'll exit immediately if postmaster death is detected,
142 : : * instead of returning.
143 : : */
144 : : bool exit_on_postmaster_death;
145 : :
146 : : #if defined(WAIT_USE_EPOLL)
147 : : int epoll_fd;
148 : : /* epoll_wait returns events in a user provided arrays, allocate once */
149 : : struct epoll_event *epoll_ret_events;
150 : : #elif defined(WAIT_USE_KQUEUE)
151 : : int kqueue_fd;
152 : : /* kevent returns events in a user provided arrays, allocate once */
153 : : struct kevent *kqueue_ret_events;
154 : : bool report_postmaster_not_running;
155 : : #elif defined(WAIT_USE_POLL)
156 : : /* poll expects events to be waited on every poll() call, prepare once */
157 : : struct pollfd *pollfds;
158 : : #elif defined(WAIT_USE_WIN32)
159 : :
160 : : /*
161 : : * Array of windows events. The first element always contains
162 : : * pgwin32_signal_event, so the remaining elements are offset by one (i.e.
163 : : * event->pos + 1).
164 : : */
165 : : HANDLE *handles;
166 : : #endif
167 : : };
168 : :
169 : : #ifndef WIN32
170 : : /* Are we currently in WaitLatch? The signal handler would like to know. */
171 : : static volatile sig_atomic_t waiting = false;
172 : : #endif
173 : :
174 : : #ifdef WAIT_USE_SIGNALFD
175 : : /* On Linux, we'll receive SIGURG via a signalfd file descriptor. */
176 : : static int signal_fd = -1;
177 : : #endif
178 : :
179 : : #ifdef WAIT_USE_SELF_PIPE
180 : : /* Read and write ends of the self-pipe */
181 : : static int selfpipe_readfd = -1;
182 : : static int selfpipe_writefd = -1;
183 : :
184 : : /* Process owning the self-pipe --- needed for checking purposes */
185 : : static int selfpipe_owner_pid = 0;
186 : :
187 : : /* Private function prototypes */
188 : : static void latch_sigurg_handler(SIGNAL_ARGS);
189 : : static void sendSelfPipeByte(void);
190 : : #endif
191 : :
192 : : #if defined(WAIT_USE_SELF_PIPE) || defined(WAIT_USE_SIGNALFD)
193 : : static void drain(void);
194 : : #endif
195 : :
196 : : #if defined(WAIT_USE_EPOLL)
197 : : static void WaitEventAdjustEpoll(WaitEventSet *set, WaitEvent *event, int action);
198 : : #elif defined(WAIT_USE_KQUEUE)
199 : : static void WaitEventAdjustKqueue(WaitEventSet *set, WaitEvent *event, int old_events);
200 : : #elif defined(WAIT_USE_POLL)
201 : : static void WaitEventAdjustPoll(WaitEventSet *set, WaitEvent *event);
202 : : #elif defined(WAIT_USE_WIN32)
203 : : static void WaitEventAdjustWin32(WaitEventSet *set, WaitEvent *event);
204 : : #endif
205 : :
206 : : static inline int WaitEventSetWaitBlock(WaitEventSet *set, int cur_timeout,
207 : : WaitEvent *occurred_events, int nevents);
208 : :
209 : : /* ResourceOwner support to hold WaitEventSets */
210 : : static void ResOwnerReleaseWaitEventSet(Datum res);
211 : :
212 : : static const ResourceOwnerDesc wait_event_set_resowner_desc =
213 : : {
214 : : .name = "WaitEventSet",
215 : : .release_phase = RESOURCE_RELEASE_AFTER_LOCKS,
216 : : .release_priority = RELEASE_PRIO_WAITEVENTSETS,
217 : : .ReleaseResource = ResOwnerReleaseWaitEventSet,
218 : : .DebugPrint = NULL
219 : : };
220 : :
221 : : /* Convenience wrappers over ResourceOwnerRemember/Forget */
222 : : static inline void
223 : 0 : ResourceOwnerRememberWaitEventSet(ResourceOwner owner, WaitEventSet *set)
224 : : {
225 : 0 : ResourceOwnerRemember(owner, PointerGetDatum(set), &wait_event_set_resowner_desc);
226 : 0 : }
227 : : static inline void
228 : 0 : ResourceOwnerForgetWaitEventSet(ResourceOwner owner, WaitEventSet *set)
229 : : {
230 : 0 : ResourceOwnerForget(owner, PointerGetDatum(set), &wait_event_set_resowner_desc);
231 : 0 : }
232 : :
233 : :
234 : : /*
235 : : * Initialize the process-local wait event infrastructure.
236 : : *
237 : : * This must be called once during startup of any process that can wait on
238 : : * latches, before it issues any InitLatch() or OwnLatch() calls.
239 : : */
240 : : void
241 : 810 : InitializeWaitEventSupport(void)
242 : : {
243 : : #if defined(WAIT_USE_SELF_PIPE)
244 : : int pipefd[2];
245 : :
246 : : if (IsUnderPostmaster)
247 : : {
248 : : /*
249 : : * We might have inherited connections to a self-pipe created by the
250 : : * postmaster. It's critical that child processes create their own
251 : : * self-pipes, of course, and we really want them to close the
252 : : * inherited FDs for safety's sake.
253 : : */
254 : : if (selfpipe_owner_pid != 0)
255 : : {
256 : : /* Assert we go through here but once in a child process */
257 : : Assert(selfpipe_owner_pid != MyProcPid);
258 : : /* Release postmaster's pipe FDs; ignore any error */
259 : : (void) close(selfpipe_readfd);
260 : : (void) close(selfpipe_writefd);
261 : : /* Clean up, just for safety's sake; we'll set these below */
262 : : selfpipe_readfd = selfpipe_writefd = -1;
263 : : selfpipe_owner_pid = 0;
264 : : /* Keep fd.c's accounting straight */
265 : : ReleaseExternalFD();
266 : : ReleaseExternalFD();
267 : : }
268 : : else
269 : : {
270 : : /*
271 : : * Postmaster didn't create a self-pipe ... or else we're in an
272 : : * EXEC_BACKEND build, in which case it doesn't matter since the
273 : : * postmaster's pipe FDs were closed by the action of FD_CLOEXEC.
274 : : * fd.c won't have state to clean up, either.
275 : : */
276 : : Assert(selfpipe_readfd == -1);
277 : : }
278 : : }
279 : : else
280 : : {
281 : : /* In postmaster or standalone backend, assert we do this but once */
282 : : Assert(selfpipe_readfd == -1);
283 : : Assert(selfpipe_owner_pid == 0);
284 : : }
285 : :
286 : : /*
287 : : * Set up the self-pipe that allows a signal handler to wake up the
288 : : * poll()/epoll_wait() in WaitLatch. Make the write-end non-blocking, so
289 : : * that SetLatch won't block if the event has already been set many times
290 : : * filling the kernel buffer. Make the read-end non-blocking too, so that
291 : : * we can easily clear the pipe by reading until EAGAIN or EWOULDBLOCK.
292 : : * Also, make both FDs close-on-exec, since we surely do not want any
293 : : * child processes messing with them.
294 : : */
295 : : if (pipe(pipefd) < 0)
296 : : elog(FATAL, "pipe() failed: %m");
297 : : if (fcntl(pipefd[0], F_SETFL, O_NONBLOCK) == -1)
298 : : elog(FATAL, "fcntl(F_SETFL) failed on read-end of self-pipe: %m");
299 : : if (fcntl(pipefd[1], F_SETFL, O_NONBLOCK) == -1)
300 : : elog(FATAL, "fcntl(F_SETFL) failed on write-end of self-pipe: %m");
301 : : if (fcntl(pipefd[0], F_SETFD, FD_CLOEXEC) == -1)
302 : : elog(FATAL, "fcntl(F_SETFD) failed on read-end of self-pipe: %m");
303 : : if (fcntl(pipefd[1], F_SETFD, FD_CLOEXEC) == -1)
304 : : elog(FATAL, "fcntl(F_SETFD) failed on write-end of self-pipe: %m");
305 : :
306 : : selfpipe_readfd = pipefd[0];
307 : : selfpipe_writefd = pipefd[1];
308 : : selfpipe_owner_pid = MyProcPid;
309 : :
310 : : /* Tell fd.c about these two long-lived FDs */
311 : : ReserveExternalFD();
312 : : ReserveExternalFD();
313 : :
314 : : pqsignal(SIGURG, latch_sigurg_handler);
315 : : #endif
316 : :
317 : : #ifdef WAIT_USE_SIGNALFD
318 : : sigset_t signalfd_mask;
319 : :
320 : : if (IsUnderPostmaster)
321 : : {
322 : : /*
323 : : * It would probably be safe to re-use the inherited signalfd since
324 : : * signalfds only see the current process's pending signals, but it
325 : : * seems less surprising to close it and create our own.
326 : : */
327 : : if (signal_fd != -1)
328 : : {
329 : : /* Release postmaster's signal FD; ignore any error */
330 : : (void) close(signal_fd);
331 : : signal_fd = -1;
332 : : ReleaseExternalFD();
333 : : }
334 : : }
335 : :
336 : : /* Block SIGURG, because we'll receive it through a signalfd. */
337 : : sigaddset(&UnBlockSig, SIGURG);
338 : :
339 : : /* Set up the signalfd to receive SIGURG notifications. */
340 : : sigemptyset(&signalfd_mask);
341 : : sigaddset(&signalfd_mask, SIGURG);
342 : : signal_fd = signalfd(-1, &signalfd_mask, SFD_NONBLOCK | SFD_CLOEXEC);
343 : : if (signal_fd < 0)
344 : : elog(FATAL, "signalfd() failed");
345 : : ReserveExternalFD();
346 : : #endif
347 : :
348 : : #ifdef WAIT_USE_KQUEUE
349 : : /* Ignore SIGURG, because we'll receive it via kqueue. */
350 : 810 : pqsignal(SIGURG, SIG_IGN);
351 : : #endif
352 : 810 : }
353 : :
354 : : /*
355 : : * Create a WaitEventSet with space for nevents different events to wait for.
356 : : *
357 : : * These events can then be efficiently waited upon together, using
358 : : * WaitEventSetWait().
359 : : *
360 : : * The WaitEventSet is tracked by the given 'resowner'. Use NULL for session
361 : : * lifetime.
362 : : */
363 : : WaitEventSet *
364 : 1127 : CreateWaitEventSet(ResourceOwner resowner, int nevents)
365 : : {
366 : 1127 : WaitEventSet *set;
367 : 1127 : char *data;
368 : 1127 : Size sz = 0;
369 : :
370 : : /*
371 : : * Use MAXALIGN size/alignment to guarantee that later uses of memory are
372 : : * aligned correctly. E.g. epoll_event might need 8 byte alignment on some
373 : : * platforms, but earlier allocations like WaitEventSet and WaitEvent
374 : : * might not be sized to guarantee that when purely using sizeof().
375 : : */
376 : 1127 : sz += MAXALIGN(sizeof(WaitEventSet));
377 : 1127 : sz += MAXALIGN(sizeof(WaitEvent) * nevents);
378 : :
379 : : #if defined(WAIT_USE_EPOLL)
380 : : sz += MAXALIGN(sizeof(struct epoll_event) * nevents);
381 : : #elif defined(WAIT_USE_KQUEUE)
382 : 1127 : sz += MAXALIGN(sizeof(struct kevent) * nevents);
383 : : #elif defined(WAIT_USE_POLL)
384 : : sz += MAXALIGN(sizeof(struct pollfd) * nevents);
385 : : #elif defined(WAIT_USE_WIN32)
386 : : /* need space for the pgwin32_signal_event */
387 : : sz += MAXALIGN(sizeof(HANDLE) * (nevents + 1));
388 : : #endif
389 : :
390 [ + - ]: 1127 : if (resowner != NULL)
391 : 0 : ResourceOwnerEnlarge(resowner);
392 : :
393 : 1127 : data = (char *) MemoryContextAllocZero(TopMemoryContext, sz);
394 : :
395 : 1127 : set = (WaitEventSet *) data;
396 : 1127 : data += MAXALIGN(sizeof(WaitEventSet));
397 : :
398 : 1127 : set->events = (WaitEvent *) data;
399 : 1127 : data += MAXALIGN(sizeof(WaitEvent) * nevents);
400 : :
401 : : #if defined(WAIT_USE_EPOLL)
402 : : set->epoll_ret_events = (struct epoll_event *) data;
403 : : data += MAXALIGN(sizeof(struct epoll_event) * nevents);
404 : : #elif defined(WAIT_USE_KQUEUE)
405 : 1127 : set->kqueue_ret_events = (struct kevent *) data;
406 : 1127 : data += MAXALIGN(sizeof(struct kevent) * nevents);
407 : : #elif defined(WAIT_USE_POLL)
408 : : set->pollfds = (struct pollfd *) data;
409 : : data += MAXALIGN(sizeof(struct pollfd) * nevents);
410 : : #elif defined(WAIT_USE_WIN32)
411 : : set->handles = (HANDLE) data;
412 : : data += MAXALIGN(sizeof(HANDLE) * nevents);
413 : : #endif
414 : :
415 : 1127 : set->latch = NULL;
416 : 1127 : set->nevents_space = nevents;
417 : 1127 : set->exit_on_postmaster_death = false;
418 : :
419 [ + - ]: 1127 : if (resowner != NULL)
420 : : {
421 : 0 : ResourceOwnerRememberWaitEventSet(resowner, set);
422 : 0 : set->owner = resowner;
423 : 0 : }
424 : :
425 : : #if defined(WAIT_USE_EPOLL)
426 : : if (!AcquireExternalFD())
427 : : elog(ERROR, "AcquireExternalFD, for epoll_create1, failed: %m");
428 : : set->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
429 : : if (set->epoll_fd < 0)
430 : : {
431 : : ReleaseExternalFD();
432 : : elog(ERROR, "epoll_create1 failed: %m");
433 : : }
434 : : #elif defined(WAIT_USE_KQUEUE)
435 [ + - ]: 1127 : if (!AcquireExternalFD())
436 [ # # # # ]: 0 : elog(ERROR, "AcquireExternalFD, for kqueue, failed: %m");
437 : 1127 : set->kqueue_fd = kqueue();
438 [ + - ]: 1127 : if (set->kqueue_fd < 0)
439 : : {
440 : 0 : ReleaseExternalFD();
441 [ # # # # ]: 0 : elog(ERROR, "kqueue failed: %m");
442 : 0 : }
443 [ + - ]: 1127 : if (fcntl(set->kqueue_fd, F_SETFD, FD_CLOEXEC) == -1)
444 : : {
445 : 0 : int save_errno = errno;
446 : :
447 : 0 : close(set->kqueue_fd);
448 : 0 : ReleaseExternalFD();
449 : 0 : errno = save_errno;
450 [ # # # # ]: 0 : elog(ERROR, "fcntl(F_SETFD) failed on kqueue descriptor: %m");
451 : 0 : }
452 : 1127 : set->report_postmaster_not_running = false;
453 : : #elif defined(WAIT_USE_WIN32)
454 : :
455 : : /*
456 : : * To handle signals while waiting, we need to add a win32 specific event.
457 : : * We accounted for the additional event at the top of this routine. See
458 : : * port/win32/signal.c for more details.
459 : : *
460 : : * Note: pgwin32_signal_event should be first to ensure that it will be
461 : : * reported when multiple events are set. We want to guarantee that
462 : : * pending signals are serviced.
463 : : */
464 : : set->handles[0] = pgwin32_signal_event;
465 : : #endif
466 : :
467 : 2254 : return set;
468 : 1127 : }
469 : :
470 : : /*
471 : : * Free a previously created WaitEventSet.
472 : : *
473 : : * Note: preferably, this shouldn't have to free any resources that could be
474 : : * inherited across an exec(). If it did, we'd likely leak those resources in
475 : : * many scenarios. For the epoll case, we ensure that by setting EPOLL_CLOEXEC
476 : : * when the FD is created. For the Windows case, we assume that the handles
477 : : * involved are non-inheritable.
478 : : */
479 : : void
480 : 2 : FreeWaitEventSet(WaitEventSet *set)
481 : : {
482 [ + - ]: 2 : if (set->owner)
483 : : {
484 : 0 : ResourceOwnerForgetWaitEventSet(set->owner, set);
485 : 0 : set->owner = NULL;
486 : 0 : }
487 : :
488 : : #if defined(WAIT_USE_EPOLL)
489 : : close(set->epoll_fd);
490 : : ReleaseExternalFD();
491 : : #elif defined(WAIT_USE_KQUEUE)
492 : 2 : close(set->kqueue_fd);
493 : 2 : ReleaseExternalFD();
494 : : #elif defined(WAIT_USE_WIN32)
495 : : for (WaitEvent *cur_event = set->events;
496 : : cur_event < (set->events + set->nevents);
497 : : cur_event++)
498 : : {
499 : : if (cur_event->events & WL_LATCH_SET)
500 : : {
501 : : /* uses the latch's HANDLE */
502 : : }
503 : : else if (cur_event->events & WL_POSTMASTER_DEATH)
504 : : {
505 : : /* uses PostmasterHandle */
506 : : }
507 : : else
508 : : {
509 : : /* Clean up the event object we created for the socket */
510 : : WSAEventSelect(cur_event->fd, NULL, 0);
511 : : WSACloseEvent(set->handles[cur_event->pos + 1]);
512 : : }
513 : : }
514 : : #endif
515 : :
516 : 2 : pfree(set);
517 : 2 : }
518 : :
519 : : /*
520 : : * Free a previously created WaitEventSet in a child process after a fork().
521 : : */
522 : : void
523 : 797 : FreeWaitEventSetAfterFork(WaitEventSet *set)
524 : : {
525 : : #if defined(WAIT_USE_EPOLL)
526 : : close(set->epoll_fd);
527 : : ReleaseExternalFD();
528 : : #elif defined(WAIT_USE_KQUEUE)
529 : : /* kqueues are not normally inherited by child processes */
530 : 797 : ReleaseExternalFD();
531 : : #endif
532 : :
533 : 797 : pfree(set);
534 : 797 : }
535 : :
536 : : /* ---
537 : : * Add an event to the set. Possible events are:
538 : : * - WL_LATCH_SET: Wait for the latch to be set
539 : : * - WL_POSTMASTER_DEATH: Wait for postmaster to die
540 : : * - WL_SOCKET_READABLE: Wait for socket to become readable,
541 : : * can be combined in one event with other WL_SOCKET_* events
542 : : * - WL_SOCKET_WRITEABLE: Wait for socket to become writeable,
543 : : * can be combined with other WL_SOCKET_* events
544 : : * - WL_SOCKET_CONNECTED: Wait for socket connection to be established,
545 : : * can be combined with other WL_SOCKET_* events (on non-Windows
546 : : * platforms, this is the same as WL_SOCKET_WRITEABLE)
547 : : * - WL_SOCKET_ACCEPT: Wait for new connection to a server socket,
548 : : * can be combined with other WL_SOCKET_* events (on non-Windows
549 : : * platforms, this is the same as WL_SOCKET_READABLE)
550 : : * - WL_SOCKET_CLOSED: Wait for socket to be closed by remote peer.
551 : : * - WL_EXIT_ON_PM_DEATH: Exit immediately if the postmaster dies
552 : : *
553 : : * Returns the offset in WaitEventSet->events (starting from 0), which can be
554 : : * used to modify previously added wait events using ModifyWaitEvent().
555 : : *
556 : : * In the WL_LATCH_SET case the latch must be owned by the current process,
557 : : * i.e. it must be a process-local latch initialized with InitLatch, or a
558 : : * shared latch associated with the current process by calling OwnLatch.
559 : : *
560 : : * In the WL_SOCKET_READABLE/WRITEABLE/CONNECTED/ACCEPT cases, EOF and error
561 : : * conditions cause the socket to be reported as readable/writable/connected,
562 : : * so that the caller can deal with the condition.
563 : : *
564 : : * The user_data pointer specified here will be set for the events returned
565 : : * by WaitEventSetWait(), allowing to easily associate additional data with
566 : : * events.
567 : : */
568 : : int
569 : 2563 : AddWaitEventToSet(WaitEventSet *set, uint32 events, pgsocket fd, Latch *latch,
570 : : void *user_data)
571 : : {
572 : 2563 : WaitEvent *event;
573 : :
574 : : /* not enough space */
575 [ + - ]: 2563 : Assert(set->nevents < set->nevents_space);
576 : :
577 [ + + ]: 2563 : if (events == WL_EXIT_ON_PM_DEATH)
578 : : {
579 : 804 : events = WL_POSTMASTER_DEATH;
580 : 804 : set->exit_on_postmaster_death = true;
581 : 804 : }
582 : :
583 [ + + ]: 2563 : if (latch)
584 : : {
585 [ + - ]: 1127 : if (latch->owner_pid != MyProcPid)
586 [ # # # # ]: 0 : elog(ERROR, "cannot wait on a latch owned by another process");
587 [ + - ]: 1127 : if (set->latch)
588 [ # # # # ]: 0 : elog(ERROR, "cannot wait on more than one latch");
589 [ + - ]: 1127 : if ((events & WL_LATCH_SET) != WL_LATCH_SET)
590 [ # # # # ]: 0 : elog(ERROR, "latch events only support being set");
591 : 1127 : }
592 : : else
593 : : {
594 [ + - ]: 1436 : if (events & WL_LATCH_SET)
595 [ # # # # ]: 0 : elog(ERROR, "cannot wait on latch without a specified latch");
596 : : }
597 : :
598 : : /* waiting for socket readiness without a socket indicates a bug */
599 [ + + + - ]: 2563 : if (fd == PGINVALID_SOCKET && (events & WL_SOCKET_MASK))
600 [ # # # # ]: 0 : elog(ERROR, "cannot wait on socket event without a socket");
601 : :
602 : 2563 : event = &set->events[set->nevents];
603 : 2563 : event->pos = set->nevents++;
604 : 2563 : event->fd = fd;
605 : 2563 : event->events = events;
606 : 2563 : event->user_data = user_data;
607 : : #ifdef WIN32
608 : : event->reset = false;
609 : : #endif
610 : :
611 [ + + ]: 2563 : if (events == WL_LATCH_SET)
612 : : {
613 : 1127 : set->latch = latch;
614 : 1127 : set->latch_pos = event->pos;
615 : : #if defined(WAIT_USE_SELF_PIPE)
616 : : event->fd = selfpipe_readfd;
617 : : #elif defined(WAIT_USE_SIGNALFD)
618 : : event->fd = signal_fd;
619 : : #else
620 : 1127 : event->fd = PGINVALID_SOCKET;
621 : : #ifdef WAIT_USE_EPOLL
622 : : return event->pos;
623 : : #endif
624 : : #endif
625 : 1127 : }
626 [ + + ]: 1436 : else if (events == WL_POSTMASTER_DEATH)
627 : : {
628 : : #ifndef WIN32
629 : 1119 : event->fd = postmaster_alive_fds[POSTMASTER_FD_WATCH];
630 : : #endif
631 : 1119 : }
632 : :
633 : : /* perform wait primitive specific initialization, if needed */
634 : : #if defined(WAIT_USE_EPOLL)
635 : : WaitEventAdjustEpoll(set, event, EPOLL_CTL_ADD);
636 : : #elif defined(WAIT_USE_KQUEUE)
637 : 2563 : WaitEventAdjustKqueue(set, event, 0);
638 : : #elif defined(WAIT_USE_POLL)
639 : : WaitEventAdjustPoll(set, event);
640 : : #elif defined(WAIT_USE_WIN32)
641 : : WaitEventAdjustWin32(set, event);
642 : : #endif
643 : :
644 : 5126 : return event->pos;
645 : 2563 : }
646 : :
647 : : /*
648 : : * Change the event mask and, in the WL_LATCH_SET case, the latch associated
649 : : * with the WaitEvent. The latch may be changed to NULL to disable the latch
650 : : * temporarily, and then set back to a latch later.
651 : : *
652 : : * 'pos' is the id returned by AddWaitEventToSet.
653 : : */
654 : : void
655 : 65690 : ModifyWaitEvent(WaitEventSet *set, int pos, uint32 events, Latch *latch)
656 : : {
657 : 65690 : WaitEvent *event;
658 : : #if defined(WAIT_USE_KQUEUE)
659 : 65690 : int old_events;
660 : : #endif
661 : :
662 [ + - ]: 65690 : Assert(pos < set->nevents);
663 : :
664 : 65690 : event = &set->events[pos];
665 : : #if defined(WAIT_USE_KQUEUE)
666 : 65690 : old_events = event->events;
667 : : #endif
668 : :
669 : : /*
670 : : * Allow switching between WL_POSTMASTER_DEATH and WL_EXIT_ON_PM_DEATH.
671 : : *
672 : : * Note that because WL_EXIT_ON_PM_DEATH is mapped to WL_POSTMASTER_DEATH
673 : : * in AddWaitEventToSet(), this needs to be checked before the fast-path
674 : : * below that checks if 'events' has changed.
675 : : */
676 [ + + ]: 65690 : if (event->events == WL_POSTMASTER_DEATH)
677 : : {
678 [ + + + - ]: 3321 : if (events != WL_POSTMASTER_DEATH && events != WL_EXIT_ON_PM_DEATH)
679 [ # # # # ]: 0 : elog(ERROR, "cannot remove postmaster death event");
680 : 3321 : set->exit_on_postmaster_death = ((events & WL_EXIT_ON_PM_DEATH) != 0);
681 : 3321 : return;
682 : : }
683 : :
684 : : /*
685 : : * If neither the event mask nor the associated latch changes, return
686 : : * early. That's an important optimization for some sockets, where
687 : : * ModifyWaitEvent is frequently used to switch from waiting for reads to
688 : : * waiting on writes.
689 : : */
690 [ + + + + ]: 66320 : if (events == event->events &&
691 [ + + ]: 61947 : (!(event->events & WL_LATCH_SET) || set->latch == latch))
692 : 61206 : return;
693 : :
694 [ + + + - ]: 1163 : if (event->events & WL_LATCH_SET && events != event->events)
695 [ # # # # ]: 0 : elog(ERROR, "cannot modify latch event");
696 : :
697 : : /* FIXME: validate event mask */
698 : 1163 : event->events = events;
699 : :
700 [ + + ]: 1163 : if (events == WL_LATCH_SET)
701 : : {
702 [ + - + - ]: 741 : if (latch && latch->owner_pid != MyProcPid)
703 [ # # # # ]: 0 : elog(ERROR, "cannot wait on a latch owned by another process");
704 : 741 : set->latch = latch;
705 : :
706 : : /*
707 : : * On Unix, we don't need to modify the kernel object because the
708 : : * underlying pipe (if there is one) is the same for all latches so we
709 : : * can return immediately. On Windows, we need to update our array of
710 : : * handles, but we leave the old one in place and tolerate spurious
711 : : * wakeups if the latch is disabled.
712 : : */
713 : : #if defined(WAIT_USE_WIN32)
714 : : if (!latch)
715 : : return;
716 : : #else
717 : 741 : return;
718 : : #endif
719 : : }
720 : :
721 : : #if defined(WAIT_USE_EPOLL)
722 : : WaitEventAdjustEpoll(set, event, EPOLL_CTL_MOD);
723 : : #elif defined(WAIT_USE_KQUEUE)
724 : 422 : WaitEventAdjustKqueue(set, event, old_events);
725 : : #elif defined(WAIT_USE_POLL)
726 : : WaitEventAdjustPoll(set, event);
727 : : #elif defined(WAIT_USE_WIN32)
728 : : WaitEventAdjustWin32(set, event);
729 : : #endif
730 [ - + ]: 65690 : }
731 : :
732 : : #if defined(WAIT_USE_EPOLL)
733 : : /*
734 : : * action can be one of EPOLL_CTL_ADD | EPOLL_CTL_MOD | EPOLL_CTL_DEL
735 : : */
736 : : static void
737 : : WaitEventAdjustEpoll(WaitEventSet *set, WaitEvent *event, int action)
738 : : {
739 : : struct epoll_event epoll_ev;
740 : : int rc;
741 : :
742 : : /* pointer to our event, returned by epoll_wait */
743 : : epoll_ev.data.ptr = event;
744 : : /* always wait for errors */
745 : : epoll_ev.events = EPOLLERR | EPOLLHUP;
746 : :
747 : : /* prepare pollfd entry once */
748 : : if (event->events == WL_LATCH_SET)
749 : : {
750 : : Assert(set->latch != NULL);
751 : : epoll_ev.events |= EPOLLIN;
752 : : }
753 : : else if (event->events == WL_POSTMASTER_DEATH)
754 : : {
755 : : epoll_ev.events |= EPOLLIN;
756 : : }
757 : : else
758 : : {
759 : : Assert(event->fd != PGINVALID_SOCKET);
760 : : Assert(event->events & (WL_SOCKET_READABLE |
761 : : WL_SOCKET_WRITEABLE |
762 : : WL_SOCKET_CLOSED));
763 : :
764 : : if (event->events & WL_SOCKET_READABLE)
765 : : epoll_ev.events |= EPOLLIN;
766 : : if (event->events & WL_SOCKET_WRITEABLE)
767 : : epoll_ev.events |= EPOLLOUT;
768 : : if (event->events & WL_SOCKET_CLOSED)
769 : : epoll_ev.events |= EPOLLRDHUP;
770 : : }
771 : :
772 : : /*
773 : : * Even though unused, we also pass epoll_ev as the data argument if
774 : : * EPOLL_CTL_DEL is passed as action. There used to be an epoll bug
775 : : * requiring that, and actually it makes the code simpler...
776 : : */
777 : : rc = epoll_ctl(set->epoll_fd, action, event->fd, &epoll_ev);
778 : :
779 : : if (rc < 0)
780 : : ereport(ERROR,
781 : : (errcode_for_socket_access(),
782 : : errmsg("%s() failed: %m",
783 : : "epoll_ctl")));
784 : : }
785 : : #endif
786 : :
787 : : #if defined(WAIT_USE_POLL)
788 : : static void
789 : : WaitEventAdjustPoll(WaitEventSet *set, WaitEvent *event)
790 : : {
791 : : struct pollfd *pollfd = &set->pollfds[event->pos];
792 : :
793 : : pollfd->revents = 0;
794 : : pollfd->fd = event->fd;
795 : :
796 : : /* prepare pollfd entry once */
797 : : if (event->events == WL_LATCH_SET)
798 : : {
799 : : Assert(set->latch != NULL);
800 : : pollfd->events = POLLIN;
801 : : }
802 : : else if (event->events == WL_POSTMASTER_DEATH)
803 : : {
804 : : pollfd->events = POLLIN;
805 : : }
806 : : else
807 : : {
808 : : Assert(event->events & (WL_SOCKET_READABLE |
809 : : WL_SOCKET_WRITEABLE |
810 : : WL_SOCKET_CLOSED));
811 : : pollfd->events = 0;
812 : : if (event->events & WL_SOCKET_READABLE)
813 : : pollfd->events |= POLLIN;
814 : : if (event->events & WL_SOCKET_WRITEABLE)
815 : : pollfd->events |= POLLOUT;
816 : : #ifdef POLLRDHUP
817 : : if (event->events & WL_SOCKET_CLOSED)
818 : : pollfd->events |= POLLRDHUP;
819 : : #endif
820 : : }
821 : :
822 : : Assert(event->fd != PGINVALID_SOCKET);
823 : : }
824 : : #endif
825 : :
826 : : #if defined(WAIT_USE_KQUEUE)
827 : :
828 : : /*
829 : : * On most BSD family systems, the udata member of struct kevent is of type
830 : : * void *, so we could directly convert to/from WaitEvent *. Unfortunately,
831 : : * NetBSD has it as intptr_t, so here we wallpaper over that difference with
832 : : * an lvalue cast.
833 : : */
834 : : #define AccessWaitEvent(k_ev) (*((WaitEvent **)(&(k_ev)->udata)))
835 : :
836 : : static inline void
837 : 1161 : WaitEventAdjustKqueueAdd(struct kevent *k_ev, int filter, int action,
838 : : WaitEvent *event)
839 : : {
840 : 1161 : k_ev->ident = event->fd;
841 : 1161 : k_ev->filter = filter;
842 : 1161 : k_ev->flags = action;
843 : 1161 : k_ev->fflags = 0;
844 : 1161 : k_ev->data = 0;
845 : 1161 : AccessWaitEvent(k_ev) = event;
846 : 1161 : }
847 : :
848 : : static inline void
849 : 1119 : WaitEventAdjustKqueueAddPostmaster(struct kevent *k_ev, WaitEvent *event)
850 : : {
851 : : /* For now postmaster death can only be added, not removed. */
852 : 1119 : k_ev->ident = PostmasterPid;
853 : 1119 : k_ev->filter = EVFILT_PROC;
854 : 1119 : k_ev->flags = EV_ADD;
855 : 1119 : k_ev->fflags = NOTE_EXIT;
856 : 1119 : k_ev->data = 0;
857 : 1119 : AccessWaitEvent(k_ev) = event;
858 : 1119 : }
859 : :
860 : : static inline void
861 : 1127 : WaitEventAdjustKqueueAddLatch(struct kevent *k_ev, WaitEvent *event)
862 : : {
863 : : /* For now latch can only be added, not removed. */
864 : 1127 : k_ev->ident = SIGURG;
865 : 1127 : k_ev->filter = EVFILT_SIGNAL;
866 : 1127 : k_ev->flags = EV_ADD;
867 : 1127 : k_ev->fflags = 0;
868 : 1127 : k_ev->data = 0;
869 : 1127 : AccessWaitEvent(k_ev) = event;
870 : 1127 : }
871 : :
872 : : /*
873 : : * old_events is the previous event mask, used to compute what has changed.
874 : : */
875 : : static void
876 : 2985 : WaitEventAdjustKqueue(WaitEventSet *set, WaitEvent *event, int old_events)
877 : : {
878 : 2985 : int rc;
879 : 2985 : struct kevent k_ev[2];
880 : 2985 : int count = 0;
881 : 2985 : bool new_filt_read = false;
882 : 2985 : bool old_filt_read = false;
883 : 2985 : bool new_filt_write = false;
884 : 2985 : bool old_filt_write = false;
885 : :
886 [ - + ]: 2985 : if (old_events == event->events)
887 : 0 : return;
888 : :
889 [ + + + - ]: 2985 : Assert(event->events != WL_LATCH_SET || set->latch != NULL);
890 [ + + + + : 2985 : Assert(event->events == WL_LATCH_SET ||
+ - ]
891 : : event->events == WL_POSTMASTER_DEATH ||
892 : : (event->events & (WL_SOCKET_READABLE |
893 : : WL_SOCKET_WRITEABLE |
894 : : WL_SOCKET_CLOSED)));
895 : :
896 [ + + ]: 2985 : if (event->events == WL_POSTMASTER_DEATH)
897 : : {
898 : : /*
899 : : * Unlike all the other implementations, we detect postmaster death
900 : : * using process notification instead of waiting on the postmaster
901 : : * alive pipe.
902 : : */
903 : 1119 : WaitEventAdjustKqueueAddPostmaster(&k_ev[count++], event);
904 : 1119 : }
905 [ + + ]: 1866 : else if (event->events == WL_LATCH_SET)
906 : : {
907 : : /* We detect latch wakeup using a signal event. */
908 : 1127 : WaitEventAdjustKqueueAddLatch(&k_ev[count++], event);
909 : 1127 : }
910 : : else
911 : : {
912 : : /*
913 : : * We need to compute the adds and deletes required to get from the
914 : : * old event mask to the new event mask, since kevent treats readable
915 : : * and writable as separate events.
916 : : */
917 [ + + ]: 739 : if (old_events & (WL_SOCKET_READABLE | WL_SOCKET_CLOSED))
918 : 54 : old_filt_read = true;
919 [ + + ]: 739 : if (event->events & (WL_SOCKET_READABLE | WL_SOCKET_CLOSED))
920 : 370 : new_filt_read = true;
921 [ + + ]: 739 : if (old_events & WL_SOCKET_WRITEABLE)
922 : 368 : old_filt_write = true;
923 [ + + ]: 739 : if (event->events & WL_SOCKET_WRITEABLE)
924 : 369 : new_filt_write = true;
925 [ + + - + ]: 739 : if (old_filt_read && !new_filt_read)
926 : 108 : WaitEventAdjustKqueueAdd(&k_ev[count++], EVFILT_READ, EV_DELETE,
927 : 54 : event);
928 [ + - + + ]: 685 : else if (!old_filt_read && new_filt_read)
929 : 740 : WaitEventAdjustKqueueAdd(&k_ev[count++], EVFILT_READ, EV_ADD,
930 : 370 : event);
931 [ + + - + ]: 739 : if (old_filt_write && !new_filt_write)
932 : 736 : WaitEventAdjustKqueueAdd(&k_ev[count++], EVFILT_WRITE, EV_DELETE,
933 : 368 : event);
934 [ + - + + ]: 371 : else if (!old_filt_write && new_filt_write)
935 : 738 : WaitEventAdjustKqueueAdd(&k_ev[count++], EVFILT_WRITE, EV_ADD,
936 : 369 : event);
937 : : }
938 : :
939 : : /* For WL_SOCKET_READ -> WL_SOCKET_CLOSED, no change needed. */
940 [ + - ]: 2985 : if (count == 0)
941 : 0 : return;
942 : :
943 [ + - ]: 2985 : Assert(count <= 2);
944 : :
945 : 2985 : rc = kevent(set->kqueue_fd, &k_ev[0], count, NULL, 0, NULL);
946 : :
947 : : /*
948 : : * When adding the postmaster's pid, we have to consider that it might
949 : : * already have exited and perhaps even been replaced by another process
950 : : * with the same pid. If so, we have to defer reporting this as an event
951 : : * until the next call to WaitEventSetWaitBlock().
952 : : */
953 : :
954 [ + - ]: 2985 : if (rc < 0)
955 : : {
956 [ # # ]: 0 : if (event->events == WL_POSTMASTER_DEATH &&
957 [ # # ]: 0 : (errno == ESRCH || errno == EACCES))
958 : 0 : set->report_postmaster_not_running = true;
959 : : else
960 [ # # # # ]: 0 : ereport(ERROR,
961 : : (errcode_for_socket_access(),
962 : : errmsg("%s() failed: %m",
963 : : "kevent")));
964 : 0 : }
965 [ + + ]: 2985 : else if (event->events == WL_POSTMASTER_DEATH &&
966 [ - + # # ]: 1119 : PostmasterPid != getppid() &&
967 : 0 : !PostmasterIsAlive())
968 : : {
969 : : /*
970 : : * The extra PostmasterIsAliveInternal() check prevents false alarms
971 : : * on systems that give a different value for getppid() while being
972 : : * traced by a debugger.
973 : : */
974 : 0 : set->report_postmaster_not_running = true;
975 : 0 : }
976 [ - + ]: 2985 : }
977 : :
978 : : #endif
979 : :
980 : : #if defined(WAIT_USE_WIN32)
981 : : StaticAssertDecl(WSA_INVALID_EVENT == NULL, "");
982 : :
983 : : static void
984 : : WaitEventAdjustWin32(WaitEventSet *set, WaitEvent *event)
985 : : {
986 : : HANDLE *handle = &set->handles[event->pos + 1];
987 : :
988 : : if (event->events == WL_LATCH_SET)
989 : : {
990 : : Assert(set->latch != NULL);
991 : : *handle = set->latch->event;
992 : : }
993 : : else if (event->events == WL_POSTMASTER_DEATH)
994 : : {
995 : : *handle = PostmasterHandle;
996 : : }
997 : : else
998 : : {
999 : : int flags = FD_CLOSE; /* always check for errors/EOF */
1000 : :
1001 : : if (event->events & WL_SOCKET_READABLE)
1002 : : flags |= FD_READ;
1003 : : if (event->events & WL_SOCKET_WRITEABLE)
1004 : : flags |= FD_WRITE;
1005 : : if (event->events & WL_SOCKET_CONNECTED)
1006 : : flags |= FD_CONNECT;
1007 : : if (event->events & WL_SOCKET_ACCEPT)
1008 : : flags |= FD_ACCEPT;
1009 : :
1010 : : if (*handle == WSA_INVALID_EVENT)
1011 : : {
1012 : : *handle = WSACreateEvent();
1013 : : if (*handle == WSA_INVALID_EVENT)
1014 : : elog(ERROR, "failed to create event for socket: error code %d",
1015 : : WSAGetLastError());
1016 : : }
1017 : : if (WSAEventSelect(event->fd, *handle, flags) != 0)
1018 : : elog(ERROR, "failed to set up event for socket: error code %d",
1019 : : WSAGetLastError());
1020 : :
1021 : : Assert(event->fd != PGINVALID_SOCKET);
1022 : : }
1023 : : }
1024 : : #endif
1025 : :
1026 : : /*
1027 : : * Wait for events added to the set to happen, or until the timeout is
1028 : : * reached. At most nevents occurred events are returned.
1029 : : *
1030 : : * If timeout = -1, block until an event occurs; if 0, check sockets for
1031 : : * readiness, but don't block; if > 0, block for at most timeout milliseconds.
1032 : : *
1033 : : * Returns the number of events occurred, or 0 if the timeout was reached.
1034 : : *
1035 : : * Returned events will have the fd, pos, user_data fields set to the
1036 : : * values associated with the registered event.
1037 : : */
1038 : : int
1039 : 62886 : WaitEventSetWait(WaitEventSet *set, long timeout,
1040 : : WaitEvent *occurred_events, int nevents,
1041 : : uint32 wait_event_info)
1042 : : {
1043 : 62886 : int returned_events = 0;
1044 : 62886 : instr_time start_time;
1045 : 62886 : instr_time cur_time;
1046 : 62886 : long cur_timeout = -1;
1047 : :
1048 [ + - ]: 62886 : Assert(nevents > 0);
1049 : :
1050 : : /*
1051 : : * Initialize timeout if requested. We must record the current time so
1052 : : * that we can determine the remaining timeout if interrupted.
1053 : : */
1054 [ + + ]: 62886 : if (timeout >= 0)
1055 : : {
1056 : 1239 : INSTR_TIME_SET_CURRENT(start_time);
1057 [ + - ]: 1239 : Assert(timeout >= 0 && timeout <= INT_MAX);
1058 : 1239 : cur_timeout = timeout;
1059 : 1239 : }
1060 : : else
1061 : 61647 : INSTR_TIME_SET_ZERO(start_time);
1062 : :
1063 : 62886 : pgstat_report_wait_start(wait_event_info);
1064 : :
1065 : : #ifndef WIN32
1066 : 62886 : waiting = true;
1067 : : #else
1068 : : /* Ensure that signals are serviced even if latch is already set */
1069 : : pgwin32_dispatch_queued_signals();
1070 : : #endif
1071 [ + + ]: 126389 : while (returned_events == 0)
1072 : : {
1073 : 65057 : int rc;
1074 : :
1075 : : /*
1076 : : * Check if the latch is set already first. If so, we either exit
1077 : : * immediately or ask the kernel for further events available right
1078 : : * now without waiting, depending on how many events the caller wants.
1079 : : *
1080 : : * If someone sets the latch between this and the
1081 : : * WaitEventSetWaitBlock() below, the setter will write a byte to the
1082 : : * pipe (or signal us and the signal handler will do that), and the
1083 : : * readiness routine will return immediately.
1084 : : *
1085 : : * On unix, If there's a pending byte in the self pipe, we'll notice
1086 : : * whenever blocking. Only clearing the pipe in that case avoids
1087 : : * having to drain it every time WaitLatchOrSocket() is used. Should
1088 : : * the pipe-buffer fill up we're still ok, because the pipe is in
1089 : : * nonblocking mode. It's unlikely for that to happen, because the
1090 : : * self pipe isn't filled unless we're blocking (waiting = true), or
1091 : : * from inside a signal handler in latch_sigurg_handler().
1092 : : *
1093 : : * On windows, we'll also notice if there's a pending event for the
1094 : : * latch when blocking, but there's no danger of anything filling up,
1095 : : * as "Setting an event that is already set has no effect.".
1096 : : *
1097 : : * Note: we assume that the kernel calls involved in latch management
1098 : : * will provide adequate synchronization on machines with weak memory
1099 : : * ordering, so that we cannot miss seeing is_set if a notification
1100 : : * has already been queued.
1101 : : */
1102 [ + - + + ]: 65057 : if (set->latch && !set->latch->is_set)
1103 : : {
1104 : : /* about to sleep on a latch */
1105 : 62782 : set->latch->maybe_sleeping = true;
1106 : 62782 : pg_memory_barrier();
1107 : : /* and recheck */
1108 : 62782 : }
1109 : :
1110 [ + - + + ]: 65057 : if (set->latch && set->latch->is_set)
1111 : : {
1112 : 2276 : occurred_events->fd = PGINVALID_SOCKET;
1113 : 2276 : occurred_events->pos = set->latch_pos;
1114 : 2276 : occurred_events->user_data =
1115 : 2276 : set->events[set->latch_pos].user_data;
1116 : 2276 : occurred_events->events = WL_LATCH_SET;
1117 : 2276 : occurred_events++;
1118 : 2276 : returned_events++;
1119 : :
1120 : : /* could have been set above */
1121 : 2276 : set->latch->maybe_sleeping = false;
1122 : :
1123 [ + + ]: 2276 : if (returned_events == nevents)
1124 : 1444 : break; /* output buffer full already */
1125 : :
1126 : : /*
1127 : : * Even though we already have an event, we'll poll just once with
1128 : : * zero timeout to see what non-latch events we can fit into the
1129 : : * output buffer at the same time.
1130 : : */
1131 : 832 : cur_timeout = 0;
1132 : 832 : timeout = 0;
1133 : 832 : }
1134 : :
1135 : : /*
1136 : : * Wait for events using the readiness primitive chosen at the top of
1137 : : * this file. If -1 is returned, a timeout has occurred, if 0 we have
1138 : : * to retry, everything >= 1 is the number of returned events.
1139 : : */
1140 : 127226 : rc = WaitEventSetWaitBlock(set, cur_timeout,
1141 : 63613 : occurred_events, nevents - returned_events);
1142 : :
1143 [ + - + + ]: 63613 : if (set->latch &&
1144 : 63613 : set->latch->maybe_sleeping)
1145 : 62781 : set->latch->maybe_sleeping = false;
1146 : :
1147 [ + + ]: 63613 : if (rc == -1)
1148 : 110 : break; /* timeout occurred */
1149 : : else
1150 : 63503 : returned_events += rc;
1151 : :
1152 : : /* If we're not done, update cur_timeout for next iteration */
1153 [ + + + + ]: 63503 : if (returned_events == 0 && timeout >= 0)
1154 : : {
1155 : 800 : INSTR_TIME_SET_CURRENT(cur_time);
1156 : 800 : INSTR_TIME_SUBTRACT(cur_time, start_time);
1157 : 800 : cur_timeout = timeout - (long) INSTR_TIME_GET_MILLISEC(cur_time);
1158 [ - + ]: 800 : if (cur_timeout <= 0)
1159 : 0 : break;
1160 : 800 : }
1161 [ - + + ]: 65057 : }
1162 : : #ifndef WIN32
1163 : 62886 : waiting = false;
1164 : : #endif
1165 : :
1166 : 62886 : pgstat_report_wait_end();
1167 : :
1168 : 125772 : return returned_events;
1169 : 62886 : }
1170 : :
1171 : :
1172 : : #if defined(WAIT_USE_EPOLL)
1173 : :
1174 : : /*
1175 : : * Wait using linux's epoll_wait(2).
1176 : : *
1177 : : * This is the preferable wait method, as several readiness notifications are
1178 : : * delivered, without having to iterate through all of set->events. The return
1179 : : * epoll_event struct contain a pointer to our events, making association
1180 : : * easy.
1181 : : */
1182 : : static inline int
1183 : : WaitEventSetWaitBlock(WaitEventSet *set, int cur_timeout,
1184 : : WaitEvent *occurred_events, int nevents)
1185 : : {
1186 : : int returned_events = 0;
1187 : : int rc;
1188 : : WaitEvent *cur_event;
1189 : : struct epoll_event *cur_epoll_event;
1190 : :
1191 : : /* Sleep */
1192 : : rc = epoll_wait(set->epoll_fd, set->epoll_ret_events,
1193 : : Min(nevents, set->nevents_space), cur_timeout);
1194 : :
1195 : : /* Check return code */
1196 : : if (rc < 0)
1197 : : {
1198 : : /* EINTR is okay, otherwise complain */
1199 : : if (errno != EINTR)
1200 : : {
1201 : : waiting = false;
1202 : : ereport(ERROR,
1203 : : (errcode_for_socket_access(),
1204 : : errmsg("%s() failed: %m",
1205 : : "epoll_wait")));
1206 : : }
1207 : : return 0;
1208 : : }
1209 : : else if (rc == 0)
1210 : : {
1211 : : /* timeout exceeded */
1212 : : return -1;
1213 : : }
1214 : :
1215 : : /*
1216 : : * At least one event occurred, iterate over the returned epoll events
1217 : : * until they're either all processed, or we've returned all the events
1218 : : * the caller desired.
1219 : : */
1220 : : for (cur_epoll_event = set->epoll_ret_events;
1221 : : cur_epoll_event < (set->epoll_ret_events + rc) &&
1222 : : returned_events < nevents;
1223 : : cur_epoll_event++)
1224 : : {
1225 : : /* epoll's data pointer is set to the associated WaitEvent */
1226 : : cur_event = (WaitEvent *) cur_epoll_event->data.ptr;
1227 : :
1228 : : occurred_events->pos = cur_event->pos;
1229 : : occurred_events->user_data = cur_event->user_data;
1230 : : occurred_events->events = 0;
1231 : :
1232 : : if (cur_event->events == WL_LATCH_SET &&
1233 : : cur_epoll_event->events & (EPOLLIN | EPOLLERR | EPOLLHUP))
1234 : : {
1235 : : /* Drain the signalfd. */
1236 : : drain();
1237 : :
1238 : : if (set->latch && set->latch->maybe_sleeping && set->latch->is_set)
1239 : : {
1240 : : occurred_events->fd = PGINVALID_SOCKET;
1241 : : occurred_events->events = WL_LATCH_SET;
1242 : : occurred_events++;
1243 : : returned_events++;
1244 : : }
1245 : : }
1246 : : else if (cur_event->events == WL_POSTMASTER_DEATH &&
1247 : : cur_epoll_event->events & (EPOLLIN | EPOLLERR | EPOLLHUP))
1248 : : {
1249 : : /*
1250 : : * We expect an EPOLLHUP when the remote end is closed, but
1251 : : * because we don't expect the pipe to become readable or to have
1252 : : * any errors either, treat those cases as postmaster death, too.
1253 : : *
1254 : : * Be paranoid about a spurious event signaling the postmaster as
1255 : : * being dead. There have been reports about that happening with
1256 : : * older primitives (select(2) to be specific), and a spurious
1257 : : * WL_POSTMASTER_DEATH event would be painful. Re-checking doesn't
1258 : : * cost much.
1259 : : */
1260 : : if (!PostmasterIsAliveInternal())
1261 : : {
1262 : : if (set->exit_on_postmaster_death)
1263 : : proc_exit(1);
1264 : : occurred_events->fd = PGINVALID_SOCKET;
1265 : : occurred_events->events = WL_POSTMASTER_DEATH;
1266 : : occurred_events++;
1267 : : returned_events++;
1268 : : }
1269 : : }
1270 : : else if (cur_event->events & (WL_SOCKET_READABLE |
1271 : : WL_SOCKET_WRITEABLE |
1272 : : WL_SOCKET_CLOSED))
1273 : : {
1274 : : Assert(cur_event->fd != PGINVALID_SOCKET);
1275 : :
1276 : : if ((cur_event->events & WL_SOCKET_READABLE) &&
1277 : : (cur_epoll_event->events & (EPOLLIN | EPOLLERR | EPOLLHUP)))
1278 : : {
1279 : : /* data available in socket, or EOF */
1280 : : occurred_events->events |= WL_SOCKET_READABLE;
1281 : : }
1282 : :
1283 : : if ((cur_event->events & WL_SOCKET_WRITEABLE) &&
1284 : : (cur_epoll_event->events & (EPOLLOUT | EPOLLERR | EPOLLHUP)))
1285 : : {
1286 : : /* writable, or EOF */
1287 : : occurred_events->events |= WL_SOCKET_WRITEABLE;
1288 : : }
1289 : :
1290 : : if ((cur_event->events & WL_SOCKET_CLOSED) &&
1291 : : (cur_epoll_event->events & (EPOLLRDHUP | EPOLLERR | EPOLLHUP)))
1292 : : {
1293 : : /* remote peer shut down, or error */
1294 : : occurred_events->events |= WL_SOCKET_CLOSED;
1295 : : }
1296 : :
1297 : : if (occurred_events->events != 0)
1298 : : {
1299 : : occurred_events->fd = cur_event->fd;
1300 : : occurred_events++;
1301 : : returned_events++;
1302 : : }
1303 : : }
1304 : : }
1305 : :
1306 : : return returned_events;
1307 : : }
1308 : :
1309 : : #elif defined(WAIT_USE_KQUEUE)
1310 : :
1311 : : /*
1312 : : * Wait using kevent(2) on BSD-family systems and macOS.
1313 : : *
1314 : : * For now this mirrors the epoll code, but in future it could modify the fd
1315 : : * set in the same call to kevent as it uses for waiting instead of doing that
1316 : : * with separate system calls.
1317 : : */
1318 : : static int
1319 : 63613 : WaitEventSetWaitBlock(WaitEventSet *set, int cur_timeout,
1320 : : WaitEvent *occurred_events, int nevents)
1321 : : {
1322 : 63613 : int returned_events = 0;
1323 : 63613 : int rc;
1324 : 63613 : WaitEvent *cur_event;
1325 : 63613 : struct kevent *cur_kqueue_event;
1326 : 63613 : struct timespec timeout;
1327 : 63613 : struct timespec *timeout_p;
1328 : :
1329 [ + + ]: 63613 : if (cur_timeout < 0)
1330 : 61594 : timeout_p = NULL;
1331 : : else
1332 : : {
1333 : 2019 : timeout.tv_sec = cur_timeout / 1000;
1334 : 2019 : timeout.tv_nsec = (cur_timeout % 1000) * 1000000;
1335 : 2019 : timeout_p = &timeout;
1336 : : }
1337 : :
1338 : : /*
1339 : : * Report postmaster events discovered by WaitEventAdjustKqueue() or an
1340 : : * earlier call to WaitEventSetWait().
1341 : : */
1342 [ - + ]: 63613 : if (unlikely(set->report_postmaster_not_running))
1343 : : {
1344 [ # # ]: 0 : if (set->exit_on_postmaster_death)
1345 : 0 : proc_exit(1);
1346 : 0 : occurred_events->fd = PGINVALID_SOCKET;
1347 : 0 : occurred_events->events = WL_POSTMASTER_DEATH;
1348 : 0 : return 1;
1349 : : }
1350 : :
1351 : : /* Sleep */
1352 : 127226 : rc = kevent(set->kqueue_fd, NULL, 0,
1353 : 63613 : set->kqueue_ret_events,
1354 [ + + ]: 63613 : Min(nevents, set->nevents_space),
1355 : 63613 : timeout_p);
1356 : :
1357 : : /* Check return code */
1358 [ + + ]: 63613 : if (rc < 0)
1359 : : {
1360 : : /* EINTR is okay, otherwise complain */
1361 [ + - ]: 1367 : if (errno != EINTR)
1362 : : {
1363 : 0 : waiting = false;
1364 [ # # # # ]: 0 : ereport(ERROR,
1365 : : (errcode_for_socket_access(),
1366 : : errmsg("%s() failed: %m",
1367 : : "kevent")));
1368 : 0 : }
1369 : 1367 : return 0;
1370 : : }
1371 [ + + ]: 62246 : else if (rc == 0)
1372 : : {
1373 : : /* timeout exceeded */
1374 : 110 : return -1;
1375 : : }
1376 : :
1377 : : /*
1378 : : * At least one event occurred, iterate over the returned kqueue events
1379 : : * until they're either all processed, or we've returned all the events
1380 : : * the caller desired.
1381 : : */
1382 [ + + ]: 248544 : for (cur_kqueue_event = set->kqueue_ret_events;
1383 [ + + ]: 124272 : cur_kqueue_event < (set->kqueue_ret_events + rc) &&
1384 : 62136 : returned_events < nevents;
1385 : 62136 : cur_kqueue_event++)
1386 : : {
1387 : : /* kevent's udata points to the associated WaitEvent */
1388 : 62136 : cur_event = AccessWaitEvent(cur_kqueue_event);
1389 : :
1390 : 62136 : occurred_events->pos = cur_event->pos;
1391 : 62136 : occurred_events->user_data = cur_event->user_data;
1392 : 62136 : occurred_events->events = 0;
1393 : :
1394 [ + + - + ]: 62136 : if (cur_event->events == WL_LATCH_SET &&
1395 : 3824 : cur_kqueue_event->filter == EVFILT_SIGNAL)
1396 : : {
1397 [ + - + + : 3824 : if (set->latch && set->latch->maybe_sleeping && set->latch->is_set)
+ + ]
1398 : : {
1399 : 2237 : occurred_events->fd = PGINVALID_SOCKET;
1400 : 2237 : occurred_events->events = WL_LATCH_SET;
1401 : 2237 : occurred_events++;
1402 : 2237 : returned_events++;
1403 : 2237 : }
1404 : 3824 : }
1405 [ - + ]: 58312 : else if (cur_event->events == WL_POSTMASTER_DEATH &&
1406 [ # # # # ]: 0 : cur_kqueue_event->filter == EVFILT_PROC &&
1407 : 0 : (cur_kqueue_event->fflags & NOTE_EXIT) != 0)
1408 : : {
1409 : : /*
1410 : : * The kernel will tell this kqueue object only once about the
1411 : : * exit of the postmaster, so let's remember that for next time so
1412 : : * that we provide level-triggered semantics.
1413 : : */
1414 : 0 : set->report_postmaster_not_running = true;
1415 : :
1416 [ # # ]: 0 : if (set->exit_on_postmaster_death)
1417 : 0 : proc_exit(1);
1418 : 0 : occurred_events->fd = PGINVALID_SOCKET;
1419 : 0 : occurred_events->events = WL_POSTMASTER_DEATH;
1420 : 0 : occurred_events++;
1421 : 0 : returned_events++;
1422 : 0 : }
1423 [ - + ]: 58312 : else if (cur_event->events & (WL_SOCKET_READABLE |
1424 : : WL_SOCKET_WRITEABLE |
1425 : : WL_SOCKET_CLOSED))
1426 : : {
1427 [ - + ]: 58312 : Assert(cur_event->fd >= 0);
1428 : :
1429 [ + + - + ]: 58312 : if ((cur_event->events & WL_SOCKET_READABLE) &&
1430 : 58219 : (cur_kqueue_event->filter == EVFILT_READ))
1431 : : {
1432 : : /* readable, or EOF */
1433 : 58219 : occurred_events->events |= WL_SOCKET_READABLE;
1434 : 58219 : }
1435 : :
1436 [ - + ]: 58312 : if ((cur_event->events & WL_SOCKET_CLOSED) &&
1437 [ # # # # ]: 0 : (cur_kqueue_event->filter == EVFILT_READ) &&
1438 : 0 : (cur_kqueue_event->flags & EV_EOF))
1439 : : {
1440 : : /* the remote peer has shut down */
1441 : 0 : occurred_events->events |= WL_SOCKET_CLOSED;
1442 : 0 : }
1443 : :
1444 [ + + - + ]: 58312 : if ((cur_event->events & WL_SOCKET_WRITEABLE) &&
1445 : 93 : (cur_kqueue_event->filter == EVFILT_WRITE))
1446 : : {
1447 : : /* writable, or EOF */
1448 : 93 : occurred_events->events |= WL_SOCKET_WRITEABLE;
1449 : 93 : }
1450 : :
1451 [ - + ]: 58312 : if (occurred_events->events != 0)
1452 : : {
1453 : 58312 : occurred_events->fd = cur_event->fd;
1454 : 58312 : occurred_events++;
1455 : 58312 : returned_events++;
1456 : 58312 : }
1457 : 58312 : }
1458 : 62136 : }
1459 : :
1460 : 62136 : return returned_events;
1461 : 63613 : }
1462 : :
1463 : : #elif defined(WAIT_USE_POLL)
1464 : :
1465 : : /*
1466 : : * Wait using poll(2).
1467 : : *
1468 : : * This allows to receive readiness notifications for several events at once,
1469 : : * but requires iterating through all of set->pollfds.
1470 : : */
1471 : : static inline int
1472 : : WaitEventSetWaitBlock(WaitEventSet *set, int cur_timeout,
1473 : : WaitEvent *occurred_events, int nevents)
1474 : : {
1475 : : int returned_events = 0;
1476 : : int rc;
1477 : : WaitEvent *cur_event;
1478 : : struct pollfd *cur_pollfd;
1479 : :
1480 : : /* Sleep */
1481 : : rc = poll(set->pollfds, set->nevents, cur_timeout);
1482 : :
1483 : : /* Check return code */
1484 : : if (rc < 0)
1485 : : {
1486 : : /* EINTR is okay, otherwise complain */
1487 : : if (errno != EINTR)
1488 : : {
1489 : : waiting = false;
1490 : : ereport(ERROR,
1491 : : (errcode_for_socket_access(),
1492 : : errmsg("%s() failed: %m",
1493 : : "poll")));
1494 : : }
1495 : : return 0;
1496 : : }
1497 : : else if (rc == 0)
1498 : : {
1499 : : /* timeout exceeded */
1500 : : return -1;
1501 : : }
1502 : :
1503 : : for (cur_event = set->events, cur_pollfd = set->pollfds;
1504 : : cur_event < (set->events + set->nevents) &&
1505 : : returned_events < nevents;
1506 : : cur_event++, cur_pollfd++)
1507 : : {
1508 : : /* no activity on this FD, skip */
1509 : : if (cur_pollfd->revents == 0)
1510 : : continue;
1511 : :
1512 : : occurred_events->pos = cur_event->pos;
1513 : : occurred_events->user_data = cur_event->user_data;
1514 : : occurred_events->events = 0;
1515 : :
1516 : : if (cur_event->events == WL_LATCH_SET &&
1517 : : (cur_pollfd->revents & (POLLIN | POLLHUP | POLLERR | POLLNVAL)))
1518 : : {
1519 : : /* There's data in the self-pipe, clear it. */
1520 : : drain();
1521 : :
1522 : : if (set->latch && set->latch->maybe_sleeping && set->latch->is_set)
1523 : : {
1524 : : occurred_events->fd = PGINVALID_SOCKET;
1525 : : occurred_events->events = WL_LATCH_SET;
1526 : : occurred_events++;
1527 : : returned_events++;
1528 : : }
1529 : : }
1530 : : else if (cur_event->events == WL_POSTMASTER_DEATH &&
1531 : : (cur_pollfd->revents & (POLLIN | POLLHUP | POLLERR | POLLNVAL)))
1532 : : {
1533 : : /*
1534 : : * We expect a POLLHUP when the remote end is closed, but because
1535 : : * we don't expect the pipe to become readable or to have any
1536 : : * errors either, treat those cases as postmaster death, too.
1537 : : *
1538 : : * Be paranoid about a spurious event signaling the postmaster as
1539 : : * being dead. There have been reports about that happening with
1540 : : * older primitives (select(2) to be specific), and a spurious
1541 : : * WL_POSTMASTER_DEATH event would be painful. Re-checking
1542 : : * doesn't cost much.
1543 : : */
1544 : : if (!PostmasterIsAliveInternal())
1545 : : {
1546 : : if (set->exit_on_postmaster_death)
1547 : : proc_exit(1);
1548 : : occurred_events->fd = PGINVALID_SOCKET;
1549 : : occurred_events->events = WL_POSTMASTER_DEATH;
1550 : : occurred_events++;
1551 : : returned_events++;
1552 : : }
1553 : : }
1554 : : else if (cur_event->events & (WL_SOCKET_READABLE |
1555 : : WL_SOCKET_WRITEABLE |
1556 : : WL_SOCKET_CLOSED))
1557 : : {
1558 : : int errflags = POLLHUP | POLLERR | POLLNVAL;
1559 : :
1560 : : Assert(cur_event->fd >= PGINVALID_SOCKET);
1561 : :
1562 : : if ((cur_event->events & WL_SOCKET_READABLE) &&
1563 : : (cur_pollfd->revents & (POLLIN | errflags)))
1564 : : {
1565 : : /* data available in socket, or EOF */
1566 : : occurred_events->events |= WL_SOCKET_READABLE;
1567 : : }
1568 : :
1569 : : if ((cur_event->events & WL_SOCKET_WRITEABLE) &&
1570 : : (cur_pollfd->revents & (POLLOUT | errflags)))
1571 : : {
1572 : : /* writeable, or EOF */
1573 : : occurred_events->events |= WL_SOCKET_WRITEABLE;
1574 : : }
1575 : :
1576 : : #ifdef POLLRDHUP
1577 : : if ((cur_event->events & WL_SOCKET_CLOSED) &&
1578 : : (cur_pollfd->revents & (POLLRDHUP | errflags)))
1579 : : {
1580 : : /* remote peer closed, or error */
1581 : : occurred_events->events |= WL_SOCKET_CLOSED;
1582 : : }
1583 : : #endif
1584 : :
1585 : : if (occurred_events->events != 0)
1586 : : {
1587 : : occurred_events->fd = cur_event->fd;
1588 : : occurred_events++;
1589 : : returned_events++;
1590 : : }
1591 : : }
1592 : : }
1593 : : return returned_events;
1594 : : }
1595 : :
1596 : : #elif defined(WAIT_USE_WIN32)
1597 : :
1598 : : /*
1599 : : * Wait using Windows' WaitForMultipleObjects(). Each call only "consumes" one
1600 : : * event, so we keep calling until we've filled up our output buffer to match
1601 : : * the behavior of the other implementations.
1602 : : *
1603 : : * https://blogs.msdn.microsoft.com/oldnewthing/20150409-00/?p=44273
1604 : : */
1605 : : static inline int
1606 : : WaitEventSetWaitBlock(WaitEventSet *set, int cur_timeout,
1607 : : WaitEvent *occurred_events, int nevents)
1608 : : {
1609 : : int returned_events = 0;
1610 : : DWORD rc;
1611 : : WaitEvent *cur_event;
1612 : :
1613 : : /* Reset any wait events that need it */
1614 : : for (cur_event = set->events;
1615 : : cur_event < (set->events + set->nevents);
1616 : : cur_event++)
1617 : : {
1618 : : if (cur_event->reset)
1619 : : {
1620 : : WaitEventAdjustWin32(set, cur_event);
1621 : : cur_event->reset = false;
1622 : : }
1623 : :
1624 : : /*
1625 : : * We associate the socket with a new event handle for each
1626 : : * WaitEventSet. FD_CLOSE is only generated once if the other end
1627 : : * closes gracefully. Therefore we might miss the FD_CLOSE
1628 : : * notification, if it was delivered to another event after we stopped
1629 : : * waiting for it. Close that race by peeking for EOF after setting
1630 : : * up this handle to receive notifications, and before entering the
1631 : : * sleep.
1632 : : *
1633 : : * XXX If we had one event handle for the lifetime of a socket, we
1634 : : * wouldn't need this.
1635 : : */
1636 : : if (cur_event->events & WL_SOCKET_READABLE)
1637 : : {
1638 : : char c;
1639 : : WSABUF buf;
1640 : : DWORD received;
1641 : : DWORD flags;
1642 : :
1643 : : buf.buf = &c;
1644 : : buf.len = 1;
1645 : : flags = MSG_PEEK;
1646 : : if (WSARecv(cur_event->fd, &buf, 1, &received, &flags, NULL, NULL) == 0)
1647 : : {
1648 : : occurred_events->pos = cur_event->pos;
1649 : : occurred_events->user_data = cur_event->user_data;
1650 : : occurred_events->events = WL_SOCKET_READABLE;
1651 : : occurred_events->fd = cur_event->fd;
1652 : : return 1;
1653 : : }
1654 : : }
1655 : :
1656 : : /*
1657 : : * Windows does not guarantee to log an FD_WRITE network event
1658 : : * indicating that more data can be sent unless the previous send()
1659 : : * failed with WSAEWOULDBLOCK. While our caller might well have made
1660 : : * such a call, we cannot assume that here. Therefore, if waiting for
1661 : : * write-ready, force the issue by doing a dummy send(). If the dummy
1662 : : * send() succeeds, assume that the socket is in fact write-ready, and
1663 : : * return immediately. Also, if it fails with something other than
1664 : : * WSAEWOULDBLOCK, return a write-ready indication to let our caller
1665 : : * deal with the error condition.
1666 : : */
1667 : : if (cur_event->events & WL_SOCKET_WRITEABLE)
1668 : : {
1669 : : char c;
1670 : : WSABUF buf;
1671 : : DWORD sent;
1672 : : int r;
1673 : :
1674 : : buf.buf = &c;
1675 : : buf.len = 0;
1676 : :
1677 : : r = WSASend(cur_event->fd, &buf, 1, &sent, 0, NULL, NULL);
1678 : : if (r == 0 || WSAGetLastError() != WSAEWOULDBLOCK)
1679 : : {
1680 : : occurred_events->pos = cur_event->pos;
1681 : : occurred_events->user_data = cur_event->user_data;
1682 : : occurred_events->events = WL_SOCKET_WRITEABLE;
1683 : : occurred_events->fd = cur_event->fd;
1684 : : return 1;
1685 : : }
1686 : : }
1687 : : }
1688 : :
1689 : : /*
1690 : : * Sleep.
1691 : : *
1692 : : * Need to wait for ->nevents + 1, because signal handle is in [0].
1693 : : */
1694 : : rc = WaitForMultipleObjects(set->nevents + 1, set->handles, FALSE,
1695 : : cur_timeout);
1696 : :
1697 : : /* Check return code */
1698 : : if (rc == WAIT_FAILED)
1699 : : elog(ERROR, "WaitForMultipleObjects() failed: error code %lu",
1700 : : GetLastError());
1701 : : else if (rc == WAIT_TIMEOUT)
1702 : : {
1703 : : /* timeout exceeded */
1704 : : return -1;
1705 : : }
1706 : :
1707 : : if (rc == WAIT_OBJECT_0)
1708 : : {
1709 : : /* Service newly-arrived signals */
1710 : : pgwin32_dispatch_queued_signals();
1711 : : return 0; /* retry */
1712 : : }
1713 : :
1714 : : /*
1715 : : * With an offset of one, due to the always present pgwin32_signal_event,
1716 : : * the handle offset directly corresponds to a wait event.
1717 : : */
1718 : : cur_event = (WaitEvent *) &set->events[rc - WAIT_OBJECT_0 - 1];
1719 : :
1720 : : for (;;)
1721 : : {
1722 : : int next_pos;
1723 : : int count;
1724 : :
1725 : : occurred_events->pos = cur_event->pos;
1726 : : occurred_events->user_data = cur_event->user_data;
1727 : : occurred_events->events = 0;
1728 : :
1729 : : if (cur_event->events == WL_LATCH_SET)
1730 : : {
1731 : : /*
1732 : : * We cannot use set->latch->event to reset the fired event if we
1733 : : * aren't waiting on this latch now.
1734 : : */
1735 : : if (!ResetEvent(set->handles[cur_event->pos + 1]))
1736 : : elog(ERROR, "ResetEvent failed: error code %lu", GetLastError());
1737 : :
1738 : : if (set->latch && set->latch->maybe_sleeping && set->latch->is_set)
1739 : : {
1740 : : occurred_events->fd = PGINVALID_SOCKET;
1741 : : occurred_events->events = WL_LATCH_SET;
1742 : : occurred_events++;
1743 : : returned_events++;
1744 : : }
1745 : : }
1746 : : else if (cur_event->events == WL_POSTMASTER_DEATH)
1747 : : {
1748 : : /*
1749 : : * Postmaster apparently died. Since the consequences of falsely
1750 : : * returning WL_POSTMASTER_DEATH could be pretty unpleasant, we
1751 : : * take the trouble to positively verify this with
1752 : : * PostmasterIsAlive(), even though there is no known reason to
1753 : : * think that the event could be falsely set on Windows.
1754 : : */
1755 : : if (!PostmasterIsAliveInternal())
1756 : : {
1757 : : if (set->exit_on_postmaster_death)
1758 : : proc_exit(1);
1759 : : occurred_events->fd = PGINVALID_SOCKET;
1760 : : occurred_events->events = WL_POSTMASTER_DEATH;
1761 : : occurred_events++;
1762 : : returned_events++;
1763 : : }
1764 : : }
1765 : : else if (cur_event->events & WL_SOCKET_MASK)
1766 : : {
1767 : : WSANETWORKEVENTS resEvents;
1768 : : HANDLE handle = set->handles[cur_event->pos + 1];
1769 : :
1770 : : Assert(cur_event->fd);
1771 : :
1772 : : occurred_events->fd = cur_event->fd;
1773 : :
1774 : : ZeroMemory(&resEvents, sizeof(resEvents));
1775 : : if (WSAEnumNetworkEvents(cur_event->fd, handle, &resEvents) != 0)
1776 : : elog(ERROR, "failed to enumerate network events: error code %d",
1777 : : WSAGetLastError());
1778 : : if ((cur_event->events & WL_SOCKET_READABLE) &&
1779 : : (resEvents.lNetworkEvents & FD_READ))
1780 : : {
1781 : : /* data available in socket */
1782 : : occurred_events->events |= WL_SOCKET_READABLE;
1783 : :
1784 : : /*------
1785 : : * WaitForMultipleObjects doesn't guarantee that a read event
1786 : : * will be returned if the latch is set at the same time. Even
1787 : : * if it did, the caller might drop that event expecting it to
1788 : : * reoccur on next call. So, we must force the event to be
1789 : : * reset if this WaitEventSet is used again in order to avoid
1790 : : * an indefinite hang.
1791 : : *
1792 : : * Refer
1793 : : * https://msdn.microsoft.com/en-us/library/windows/desktop/ms741576(v=vs.85).aspx
1794 : : * for the behavior of socket events.
1795 : : *------
1796 : : */
1797 : : cur_event->reset = true;
1798 : : }
1799 : : if ((cur_event->events & WL_SOCKET_WRITEABLE) &&
1800 : : (resEvents.lNetworkEvents & FD_WRITE))
1801 : : {
1802 : : /* writeable */
1803 : : occurred_events->events |= WL_SOCKET_WRITEABLE;
1804 : : }
1805 : : if ((cur_event->events & WL_SOCKET_CONNECTED) &&
1806 : : (resEvents.lNetworkEvents & FD_CONNECT))
1807 : : {
1808 : : /* connected */
1809 : : occurred_events->events |= WL_SOCKET_CONNECTED;
1810 : : }
1811 : : if ((cur_event->events & WL_SOCKET_ACCEPT) &&
1812 : : (resEvents.lNetworkEvents & FD_ACCEPT))
1813 : : {
1814 : : /* incoming connection could be accepted */
1815 : : occurred_events->events |= WL_SOCKET_ACCEPT;
1816 : : }
1817 : : if (resEvents.lNetworkEvents & FD_CLOSE)
1818 : : {
1819 : : /* EOF/error, so signal all caller-requested socket flags */
1820 : : occurred_events->events |= (cur_event->events & WL_SOCKET_MASK);
1821 : : }
1822 : :
1823 : : if (occurred_events->events != 0)
1824 : : {
1825 : : occurred_events++;
1826 : : returned_events++;
1827 : : }
1828 : : }
1829 : :
1830 : : /* Is the output buffer full? */
1831 : : if (returned_events == nevents)
1832 : : break;
1833 : :
1834 : : /* Have we run out of possible events? */
1835 : : next_pos = cur_event->pos + 1;
1836 : : if (next_pos == set->nevents)
1837 : : break;
1838 : :
1839 : : /*
1840 : : * Poll the rest of the event handles in the array starting at
1841 : : * next_pos being careful to skip over the initial signal handle too.
1842 : : * This time we use a zero timeout.
1843 : : */
1844 : : count = set->nevents - next_pos;
1845 : : rc = WaitForMultipleObjects(count,
1846 : : set->handles + 1 + next_pos,
1847 : : false,
1848 : : 0);
1849 : :
1850 : : /*
1851 : : * We don't distinguish between errors and WAIT_TIMEOUT here because
1852 : : * we already have events to report.
1853 : : */
1854 : : if (rc < WAIT_OBJECT_0 || rc >= WAIT_OBJECT_0 + count)
1855 : : break;
1856 : :
1857 : : /* We have another event to decode. */
1858 : : cur_event = &set->events[next_pos + (rc - WAIT_OBJECT_0)];
1859 : : }
1860 : :
1861 : : return returned_events;
1862 : : }
1863 : : #endif
1864 : :
1865 : : /*
1866 : : * Return whether the current build options can report WL_SOCKET_CLOSED.
1867 : : */
1868 : : bool
1869 : 6 : WaitEventSetCanReportClosed(void)
1870 : : {
1871 : : #if (defined(WAIT_USE_POLL) && defined(POLLRDHUP)) || \
1872 : : defined(WAIT_USE_EPOLL) || \
1873 : : defined(WAIT_USE_KQUEUE)
1874 : 6 : return true;
1875 : : #else
1876 : : return false;
1877 : : #endif
1878 : : }
1879 : :
1880 : : /*
1881 : : * Get the number of wait events registered in a given WaitEventSet.
1882 : : */
1883 : : int
1884 : 0 : GetNumRegisteredWaitEvents(WaitEventSet *set)
1885 : : {
1886 : 0 : return set->nevents;
1887 : : }
1888 : :
1889 : : #if defined(WAIT_USE_SELF_PIPE)
1890 : :
1891 : : /*
1892 : : * SetLatch uses SIGURG to wake up the process waiting on the latch.
1893 : : *
1894 : : * Wake up WaitLatch, if we're waiting.
1895 : : */
1896 : : static void
1897 : : latch_sigurg_handler(SIGNAL_ARGS)
1898 : : {
1899 : : if (waiting)
1900 : : sendSelfPipeByte();
1901 : : }
1902 : :
1903 : : /* Send one byte to the self-pipe, to wake up WaitLatch */
1904 : : static void
1905 : : sendSelfPipeByte(void)
1906 : : {
1907 : : int rc;
1908 : : char dummy = 0;
1909 : :
1910 : : retry:
1911 : : rc = write(selfpipe_writefd, &dummy, 1);
1912 : : if (rc < 0)
1913 : : {
1914 : : /* If interrupted by signal, just retry */
1915 : : if (errno == EINTR)
1916 : : goto retry;
1917 : :
1918 : : /*
1919 : : * If the pipe is full, we don't need to retry, the data that's there
1920 : : * already is enough to wake up WaitLatch.
1921 : : */
1922 : : if (errno == EAGAIN || errno == EWOULDBLOCK)
1923 : : return;
1924 : :
1925 : : /*
1926 : : * Oops, the write() failed for some other reason. We might be in a
1927 : : * signal handler, so it's not safe to elog(). We have no choice but
1928 : : * silently ignore the error.
1929 : : */
1930 : : return;
1931 : : }
1932 : : }
1933 : :
1934 : : #endif
1935 : :
1936 : : #if defined(WAIT_USE_SELF_PIPE) || defined(WAIT_USE_SIGNALFD)
1937 : :
1938 : : /*
1939 : : * Read all available data from self-pipe or signalfd.
1940 : : *
1941 : : * Note: this is only called when waiting = true. If it fails and doesn't
1942 : : * return, it must reset that flag first (though ideally, this will never
1943 : : * happen).
1944 : : */
1945 : : static void
1946 : : drain(void)
1947 : : {
1948 : : char buf[1024];
1949 : : int rc;
1950 : : int fd;
1951 : :
1952 : : #ifdef WAIT_USE_SELF_PIPE
1953 : : fd = selfpipe_readfd;
1954 : : #else
1955 : : fd = signal_fd;
1956 : : #endif
1957 : :
1958 : : for (;;)
1959 : : {
1960 : : rc = read(fd, buf, sizeof(buf));
1961 : : if (rc < 0)
1962 : : {
1963 : : if (errno == EAGAIN || errno == EWOULDBLOCK)
1964 : : break; /* the descriptor is empty */
1965 : : else if (errno == EINTR)
1966 : : continue; /* retry */
1967 : : else
1968 : : {
1969 : : waiting = false;
1970 : : #ifdef WAIT_USE_SELF_PIPE
1971 : : elog(ERROR, "read() on self-pipe failed: %m");
1972 : : #else
1973 : : elog(ERROR, "read() on signalfd failed: %m");
1974 : : #endif
1975 : : }
1976 : : }
1977 : : else if (rc == 0)
1978 : : {
1979 : : waiting = false;
1980 : : #ifdef WAIT_USE_SELF_PIPE
1981 : : elog(ERROR, "unexpected EOF on self-pipe");
1982 : : #else
1983 : : elog(ERROR, "unexpected EOF on signalfd");
1984 : : #endif
1985 : : }
1986 : : else if (rc < sizeof(buf))
1987 : : {
1988 : : /* we successfully drained the pipe; no need to read() again */
1989 : : break;
1990 : : }
1991 : : /* else buffer wasn't big enough, so read again */
1992 : : }
1993 : : }
1994 : :
1995 : : #endif
1996 : :
1997 : : static void
1998 : 0 : ResOwnerReleaseWaitEventSet(Datum res)
1999 : : {
2000 : 0 : WaitEventSet *set = (WaitEventSet *) DatumGetPointer(res);
2001 : :
2002 [ # # ]: 0 : Assert(set->owner != NULL);
2003 : 0 : set->owner = NULL;
2004 : 0 : FreeWaitEventSet(set);
2005 : 0 : }
2006 : :
2007 : : #ifndef WIN32
2008 : : /*
2009 : : * Wake up my process if it's currently sleeping in WaitEventSetWaitBlock()
2010 : : *
2011 : : * NB: be sure to save and restore errno around it. (That's standard practice
2012 : : * in most signal handlers, of course, but we used to omit it in handlers that
2013 : : * only set a flag.) XXX
2014 : : *
2015 : : * NB: this function is called from critical sections and signal handlers so
2016 : : * throwing an error is not a good idea.
2017 : : *
2018 : : * On Windows, Latch uses SetEvent directly and this is not used.
2019 : : */
2020 : : void
2021 : 1361 : WakeupMyProc(void)
2022 : : {
2023 : : #if defined(WAIT_USE_SELF_PIPE)
2024 : : if (waiting)
2025 : : sendSelfPipeByte();
2026 : : #else
2027 [ - + ]: 1361 : if (waiting)
2028 : 1361 : kill(MyProcPid, SIGURG);
2029 : : #endif
2030 : 1361 : }
2031 : :
2032 : : /* Similar to WakeupMyProc, but wake up another process */
2033 : : void
2034 : 2304 : WakeupOtherProc(int pid)
2035 : : {
2036 : 2304 : kill(pid, SIGURG);
2037 : 2304 : }
2038 : : #endif
|
