Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * procsignal.c
4 : * Routines for interprocess signaling
5 : *
6 : *
7 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
8 : * Portions Copyright (c) 1994, Regents of the University of California
9 : *
10 : * IDENTIFICATION
11 : * src/backend/storage/ipc/procsignal.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : #include "postgres.h"
16 :
17 : #include <signal.h>
18 : #include <unistd.h>
19 :
20 : #include "access/parallel.h"
21 : #include "commands/async.h"
22 : #include "commands/repack.h"
23 : #include "miscadmin.h"
24 : #include "pgstat.h"
25 : #include "port/pg_bitutils.h"
26 : #include "postmaster/datachecksum_state.h"
27 : #include "replication/logicalctl.h"
28 : #include "replication/logicalworker.h"
29 : #include "replication/slotsync.h"
30 : #include "replication/walsender.h"
31 : #include "storage/condition_variable.h"
32 : #include "storage/ipc.h"
33 : #include "storage/latch.h"
34 : #include "storage/proc.h"
35 : #include "storage/shmem.h"
36 : #include "storage/sinval.h"
37 : #include "storage/smgr.h"
38 : #include "storage/subsystems.h"
39 : #include "tcop/tcopprot.h"
40 : #include "utils/memutils.h"
41 : #include "utils/wait_event.h"
42 :
43 : /*
44 : * The SIGUSR1 signal is multiplexed to support signaling multiple event
45 : * types. The specific reason is communicated via flags in shared memory.
46 : * We keep a boolean flag for each possible "reason", so that different
47 : * reasons can be signaled to a process concurrently. (However, if the same
48 : * reason is signaled more than once nearly simultaneously, the process may
49 : * observe it only once.)
50 : *
51 : * Each process that wants to receive signals registers its process ID
52 : * in the ProcSignalSlots array. The array is indexed by ProcNumber to make
53 : * slot allocation simple, and to avoid having to search the array when you
54 : * know the ProcNumber of the process you're signaling. (We do support
55 : * signaling without ProcNumber, but it's a bit less efficient.)
56 : *
57 : * The fields in each slot are protected by a spinlock, pss_mutex. pss_pid can
58 : * also be read without holding the spinlock, as a quick preliminary check
59 : * when searching for a particular PID in the array.
60 : *
61 : * pss_signalFlags are intended to be set in cases where we don't need to
62 : * keep track of whether or not the target process has handled the signal,
63 : * but sometimes we need confirmation, as when making a global state change
64 : * that cannot be considered complete until all backends have taken notice
65 : * of it. For such use cases, we set a bit in pss_barrierCheckMask and then
66 : * increment the current "barrier generation"; when the new barrier generation
67 : * (or greater) appears in the pss_barrierGeneration flag of every process,
68 : * we know that the message has been received everywhere.
69 : */
70 : typedef struct
71 : {
72 : pg_atomic_uint32 pss_pid;
73 : int pss_cancel_key_len; /* 0 means no cancellation is possible */
74 : uint8 pss_cancel_key[MAX_CANCEL_KEY_LENGTH];
75 : volatile sig_atomic_t pss_signalFlags[NUM_PROCSIGNALS];
76 : slock_t pss_mutex; /* protects the above fields */
77 :
78 : /* Barrier-related fields (not protected by pss_mutex) */
79 : pg_atomic_uint64 pss_barrierGeneration;
80 : pg_atomic_uint32 pss_barrierCheckMask;
81 : ConditionVariable pss_barrierCV;
82 : } ProcSignalSlot;
83 :
84 : /*
85 : * Information that is global to the entire ProcSignal system can be stored
86 : * here.
87 : *
88 : * psh_barrierGeneration is the highest barrier generation in existence.
89 : */
90 : struct ProcSignalHeader
91 : {
92 : pg_atomic_uint64 psh_barrierGeneration;
93 : ProcSignalSlot psh_slot[FLEXIBLE_ARRAY_MEMBER];
94 : };
95 :
96 : /*
97 : * We reserve a slot for each possible ProcNumber, plus one for each
98 : * possible auxiliary process type. (This scheme assumes there is not
99 : * more than one of any auxiliary process type at a time, except for
100 : * IO workers.)
101 : */
102 : #define NumProcSignalSlots (MaxBackends + NUM_AUXILIARY_PROCS)
103 :
104 : /* Check whether the relevant type bit is set in the flags. */
105 : #define BARRIER_SHOULD_CHECK(flags, type) \
106 : (((flags) & (((uint32) 1) << (uint32) (type))) != 0)
107 :
108 : /* Clear the relevant type bit from the flags. */
109 : #define BARRIER_CLEAR_BIT(flags, type) \
110 : ((flags) &= ~(((uint32) 1) << (uint32) (type)))
111 :
112 : static void ProcSignalShmemRequest(void *arg);
113 : static void ProcSignalShmemInit(void *arg);
114 :
115 : const ShmemCallbacks ProcSignalShmemCallbacks = {
116 : .request_fn = ProcSignalShmemRequest,
117 : .init_fn = ProcSignalShmemInit,
118 : };
119 :
120 : NON_EXEC_STATIC ProcSignalHeader *ProcSignal = NULL;
121 :
122 : static ProcSignalSlot *MyProcSignalSlot = NULL;
123 :
124 : static bool CheckProcSignal(ProcSignalReason reason);
125 : static void CleanupProcSignalState(int status, Datum arg);
126 : static void ResetProcSignalBarrierBits(uint32 flags);
127 :
128 : /*
129 : * ProcSignalShmemRequest
130 : * Register ProcSignal's shared memory needs at postmaster startup
131 : */
132 : static void
133 1250 : ProcSignalShmemRequest(void *arg)
134 : {
135 : Size size;
136 :
137 1250 : size = mul_size(NumProcSignalSlots, sizeof(ProcSignalSlot));
138 1250 : size = add_size(size, offsetof(ProcSignalHeader, psh_slot));
139 :
140 1250 : ShmemRequestStruct(.name = "ProcSignal",
141 : .size = size,
142 : .ptr = (void **) &ProcSignal,
143 : );
144 1250 : }
145 :
146 : static void
147 1247 : ProcSignalShmemInit(void *arg)
148 : {
149 1247 : pg_atomic_init_u64(&ProcSignal->psh_barrierGeneration, 0);
150 :
151 165116 : for (int i = 0; i < NumProcSignalSlots; ++i)
152 : {
153 163869 : ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
154 :
155 163869 : SpinLockInit(&slot->pss_mutex);
156 163869 : pg_atomic_init_u32(&slot->pss_pid, 0);
157 163869 : slot->pss_cancel_key_len = 0;
158 983214 : MemSet(slot->pss_signalFlags, 0, sizeof(slot->pss_signalFlags));
159 163869 : pg_atomic_init_u64(&slot->pss_barrierGeneration, PG_UINT64_MAX);
160 163869 : pg_atomic_init_u32(&slot->pss_barrierCheckMask, 0);
161 163869 : ConditionVariableInit(&slot->pss_barrierCV);
162 : }
163 1247 : }
164 :
165 : /*
166 : * ProcSignalInit
167 : * Register the current process in the ProcSignal array
168 : */
169 : void
170 24495 : ProcSignalInit(const uint8 *cancel_key, int cancel_key_len)
171 : {
172 : ProcSignalSlot *slot;
173 : uint64 barrier_generation;
174 : uint32 old_pss_pid;
175 :
176 : Assert(cancel_key_len >= 0 && cancel_key_len <= MAX_CANCEL_KEY_LENGTH);
177 24495 : if (MyProcNumber < 0)
178 0 : elog(ERROR, "MyProcNumber not set");
179 24495 : if (MyProcNumber >= NumProcSignalSlots)
180 0 : elog(ERROR, "unexpected MyProcNumber %d in ProcSignalInit (max %d)", MyProcNumber, NumProcSignalSlots);
181 24495 : slot = &ProcSignal->psh_slot[MyProcNumber];
182 :
183 24495 : SpinLockAcquire(&slot->pss_mutex);
184 :
185 : /* Value used for sanity check below */
186 24495 : old_pss_pid = pg_atomic_read_u32(&slot->pss_pid);
187 :
188 : /* Clear out any leftover signal reasons */
189 146970 : MemSet(slot->pss_signalFlags, 0, NUM_PROCSIGNALS * sizeof(sig_atomic_t));
190 :
191 : /*
192 : * Initialize barrier state. Since we're a brand-new process, there
193 : * shouldn't be any leftover backend-private state that needs to be
194 : * updated. Therefore, we can broadcast the latest barrier generation and
195 : * disregard any previously-set check bits.
196 : *
197 : * NB: This only works if this initialization happens early enough in the
198 : * startup sequence that we haven't yet cached any state that might need
199 : * to be invalidated. That's also why we have a memory barrier here, to be
200 : * sure that any later reads of memory happen strictly after this.
201 : */
202 24495 : pg_atomic_write_u32(&slot->pss_barrierCheckMask, 0);
203 : barrier_generation =
204 24495 : pg_atomic_read_u64(&ProcSignal->psh_barrierGeneration);
205 24495 : pg_atomic_write_u64(&slot->pss_barrierGeneration, barrier_generation);
206 :
207 24495 : if (cancel_key_len > 0)
208 14788 : memcpy(slot->pss_cancel_key, cancel_key, cancel_key_len);
209 24495 : slot->pss_cancel_key_len = cancel_key_len;
210 24495 : pg_atomic_write_u32(&slot->pss_pid, MyProcPid);
211 :
212 24495 : SpinLockRelease(&slot->pss_mutex);
213 :
214 : /* Spinlock is released, do the check */
215 24495 : if (old_pss_pid != 0)
216 0 : elog(LOG, "process %d taking over ProcSignal slot %d, but it's not empty",
217 : MyProcPid, MyProcNumber);
218 :
219 : /* Remember slot location for CheckProcSignal */
220 24495 : MyProcSignalSlot = slot;
221 :
222 : /* Set up to release the slot on process exit */
223 24495 : on_shmem_exit(CleanupProcSignalState, (Datum) 0);
224 24495 : }
225 :
226 : /*
227 : * CleanupProcSignalState
228 : * Remove current process from ProcSignal mechanism
229 : *
230 : * This function is called via on_shmem_exit() during backend shutdown.
231 : */
232 : static void
233 24495 : CleanupProcSignalState(int status, Datum arg)
234 : {
235 : pid_t old_pid;
236 24495 : ProcSignalSlot *slot = MyProcSignalSlot;
237 :
238 : /*
239 : * Clear MyProcSignalSlot, so that a SIGUSR1 received after this point
240 : * won't try to access it after it's no longer ours (and perhaps even
241 : * after we've unmapped the shared memory segment).
242 : */
243 : Assert(MyProcSignalSlot != NULL);
244 24495 : MyProcSignalSlot = NULL;
245 :
246 : /* sanity check */
247 24495 : SpinLockAcquire(&slot->pss_mutex);
248 24495 : old_pid = pg_atomic_read_u32(&slot->pss_pid);
249 24495 : if (old_pid != MyProcPid)
250 : {
251 : /*
252 : * don't ERROR here. We're exiting anyway, and don't want to get into
253 : * infinite loop trying to exit
254 : */
255 0 : SpinLockRelease(&slot->pss_mutex);
256 0 : elog(LOG, "process %d releasing ProcSignal slot %d, but it contains %d",
257 : MyProcPid, (int) (slot - ProcSignal->psh_slot), (int) old_pid);
258 0 : return; /* XXX better to zero the slot anyway? */
259 : }
260 :
261 : /* Mark the slot as unused */
262 24495 : pg_atomic_write_u32(&slot->pss_pid, 0);
263 24495 : slot->pss_cancel_key_len = 0;
264 :
265 : /*
266 : * Make this slot look like it's absorbed all possible barriers, so that
267 : * no barrier waits block on it.
268 : */
269 24495 : pg_atomic_write_u64(&slot->pss_barrierGeneration, PG_UINT64_MAX);
270 :
271 24495 : SpinLockRelease(&slot->pss_mutex);
272 :
273 24495 : ConditionVariableBroadcast(&slot->pss_barrierCV);
274 : }
275 :
276 : /*
277 : * SendProcSignal
278 : * Send a signal to a Postgres process
279 : *
280 : * Providing procNumber is optional, but it will speed up the operation.
281 : *
282 : * On success (a signal was sent), zero is returned.
283 : * On error, -1 is returned, and errno is set (typically to ESRCH or EPERM).
284 : *
285 : * Not to be confused with ProcSendSignal
286 : */
287 : int
288 7759 : SendProcSignal(pid_t pid, ProcSignalReason reason, ProcNumber procNumber)
289 : {
290 : volatile ProcSignalSlot *slot;
291 :
292 7759 : if (procNumber != INVALID_PROC_NUMBER)
293 : {
294 : Assert(procNumber < NumProcSignalSlots);
295 7687 : slot = &ProcSignal->psh_slot[procNumber];
296 :
297 7687 : SpinLockAcquire(&slot->pss_mutex);
298 7687 : if (pg_atomic_read_u32(&slot->pss_pid) == pid)
299 : {
300 : /* Atomically set the proper flag */
301 7687 : slot->pss_signalFlags[reason] = true;
302 7687 : SpinLockRelease(&slot->pss_mutex);
303 : /* Send signal */
304 7687 : return kill(pid, SIGUSR1);
305 : }
306 0 : SpinLockRelease(&slot->pss_mutex);
307 : }
308 : else
309 : {
310 : /*
311 : * procNumber not provided, so search the array using pid. We search
312 : * the array back to front so as to reduce search overhead. Passing
313 : * INVALID_PROC_NUMBER means that the target is most likely an
314 : * auxiliary process, which will have a slot near the end of the
315 : * array.
316 : */
317 : int i;
318 :
319 3244 : for (i = NumProcSignalSlots - 1; i >= 0; i--)
320 : {
321 3244 : slot = &ProcSignal->psh_slot[i];
322 :
323 3244 : if (pg_atomic_read_u32(&slot->pss_pid) == pid)
324 : {
325 72 : SpinLockAcquire(&slot->pss_mutex);
326 72 : if (pg_atomic_read_u32(&slot->pss_pid) == pid)
327 : {
328 : /* Atomically set the proper flag */
329 72 : slot->pss_signalFlags[reason] = true;
330 72 : SpinLockRelease(&slot->pss_mutex);
331 : /* Send signal */
332 72 : return kill(pid, SIGUSR1);
333 : }
334 0 : SpinLockRelease(&slot->pss_mutex);
335 : }
336 : }
337 : }
338 :
339 0 : errno = ESRCH;
340 0 : return -1;
341 : }
342 :
343 : /*
344 : * EmitProcSignalBarrier
345 : * Send a signal to every Postgres process
346 : *
347 : * The return value of this function is the barrier "generation" created
348 : * by this operation. This value can be passed to WaitForProcSignalBarrier
349 : * to wait until it is known that every participant in the ProcSignal
350 : * mechanism has absorbed the signal (or started afterwards).
351 : *
352 : * Note that it would be a bad idea to use this for anything that happens
353 : * frequently, as interrupting every backend could cause a noticeable
354 : * performance hit.
355 : *
356 : * Callers are entitled to assume that this function will not throw ERROR
357 : * or FATAL.
358 : */
359 : uint64
360 657 : EmitProcSignalBarrier(ProcSignalBarrierType type)
361 : {
362 657 : uint32 flagbit = 1 << (uint32) type;
363 : uint64 generation;
364 :
365 : /*
366 : * Set all the flags.
367 : *
368 : * Note that pg_atomic_fetch_or_u32 has full barrier semantics, so this is
369 : * totally ordered with respect to anything the caller did before, and
370 : * anything that we do afterwards. (This is also true of the later call to
371 : * pg_atomic_add_fetch_u64.)
372 : */
373 70072 : for (int i = 0; i < NumProcSignalSlots; i++)
374 : {
375 69415 : volatile ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
376 :
377 69415 : pg_atomic_fetch_or_u32(&slot->pss_barrierCheckMask, flagbit);
378 : }
379 :
380 : /*
381 : * Increment the generation counter.
382 : */
383 : generation =
384 657 : pg_atomic_add_fetch_u64(&ProcSignal->psh_barrierGeneration, 1);
385 :
386 : /*
387 : * Signal all the processes, so that they update their advertised barrier
388 : * generation.
389 : *
390 : * Concurrency is not a problem here. Backends that have exited don't
391 : * matter, and new backends that have joined since we entered this
392 : * function must already have current state, since the caller is
393 : * responsible for making sure that the relevant state is entirely visible
394 : * before calling this function in the first place. We still have to wake
395 : * them up - because we can't distinguish between such backends and older
396 : * backends that need to update state - but they won't actually need to
397 : * change any state.
398 : */
399 70072 : for (int i = NumProcSignalSlots - 1; i >= 0; i--)
400 : {
401 69415 : volatile ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
402 69415 : pid_t pid = pg_atomic_read_u32(&slot->pss_pid);
403 :
404 69415 : if (pid != 0)
405 : {
406 3362 : SpinLockAcquire(&slot->pss_mutex);
407 3362 : pid = pg_atomic_read_u32(&slot->pss_pid);
408 3362 : if (pid != 0)
409 : {
410 : /* see SendProcSignal for details */
411 3362 : slot->pss_signalFlags[PROCSIG_BARRIER] = true;
412 3362 : SpinLockRelease(&slot->pss_mutex);
413 3362 : kill(pid, SIGUSR1);
414 : }
415 : else
416 0 : SpinLockRelease(&slot->pss_mutex);
417 : }
418 : }
419 :
420 657 : return generation;
421 : }
422 :
423 : /*
424 : * WaitForProcSignalBarrier - wait until it is guaranteed that all changes
425 : * requested by a specific call to EmitProcSignalBarrier() have taken effect.
426 : */
427 : void
428 641 : WaitForProcSignalBarrier(uint64 generation)
429 : {
430 : Assert(generation <= pg_atomic_read_u64(&ProcSignal->psh_barrierGeneration));
431 :
432 641 : elog(DEBUG1,
433 : "waiting for all backends to process ProcSignalBarrier generation "
434 : UINT64_FORMAT,
435 : generation);
436 :
437 68926 : for (int i = NumProcSignalSlots - 1; i >= 0; i--)
438 : {
439 68285 : ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
440 : uint64 oldval;
441 :
442 : /*
443 : * It's important that we check only pss_barrierGeneration here and
444 : * not pss_barrierCheckMask. Bits in pss_barrierCheckMask get cleared
445 : * before the barrier is actually absorbed, but pss_barrierGeneration
446 : * is updated only afterward.
447 : */
448 68285 : oldval = pg_atomic_read_u64(&slot->pss_barrierGeneration);
449 71132 : while (oldval < generation)
450 : {
451 2847 : if (ConditionVariableTimedSleep(&slot->pss_barrierCV,
452 : 5000,
453 : WAIT_EVENT_PROC_SIGNAL_BARRIER))
454 0 : ereport(LOG,
455 : (errmsg("still waiting for backend with PID %d to accept ProcSignalBarrier",
456 : (int) pg_atomic_read_u32(&slot->pss_pid))));
457 2847 : oldval = pg_atomic_read_u64(&slot->pss_barrierGeneration);
458 : }
459 68285 : ConditionVariableCancelSleep();
460 : }
461 :
462 641 : elog(DEBUG1,
463 : "finished waiting for all backends to process ProcSignalBarrier generation "
464 : UINT64_FORMAT,
465 : generation);
466 :
467 : /*
468 : * The caller is probably calling this function because it wants to read
469 : * the shared state or perform further writes to shared state once all
470 : * backends are known to have absorbed the barrier. However, the read of
471 : * pss_barrierGeneration was performed unlocked; insert a memory barrier
472 : * to separate it from whatever follows.
473 : */
474 641 : pg_memory_barrier();
475 641 : }
476 :
477 : /*
478 : * Handle receipt of an interrupt indicating a global barrier event.
479 : *
480 : * All the actual work is deferred to ProcessProcSignalBarrier(), because we
481 : * cannot safely access the barrier generation inside the signal handler as
482 : * 64bit atomics might use spinlock based emulation, even for reads. As this
483 : * routine only gets called when PROCSIG_BARRIER is sent that won't cause a
484 : * lot of unnecessary work.
485 : */
486 : static void
487 2550 : HandleProcSignalBarrierInterrupt(void)
488 : {
489 2550 : InterruptPending = true;
490 2550 : ProcSignalBarrierPending = true;
491 : /* latch will be set by procsignal_sigusr1_handler */
492 2550 : }
493 :
494 : /*
495 : * Perform global barrier related interrupt checking.
496 : *
497 : * Any backend that participates in ProcSignal signaling must arrange to
498 : * call this function periodically. It is called from CHECK_FOR_INTERRUPTS(),
499 : * which is enough for normal backends, but not necessarily for all types of
500 : * background processes.
501 : */
502 : void
503 2546 : ProcessProcSignalBarrier(void)
504 : {
505 : uint64 local_gen;
506 : uint64 shared_gen;
507 : volatile uint32 flags;
508 :
509 : Assert(MyProcSignalSlot);
510 :
511 : /* Exit quickly if there's no work to do. */
512 2546 : if (!ProcSignalBarrierPending)
513 0 : return;
514 2546 : ProcSignalBarrierPending = false;
515 :
516 : /*
517 : * It's not unlikely to process multiple barriers at once, before the
518 : * signals for all the barriers have arrived. To avoid unnecessary work in
519 : * response to subsequent signals, exit early if we already have processed
520 : * all of them.
521 : */
522 2546 : local_gen = pg_atomic_read_u64(&MyProcSignalSlot->pss_barrierGeneration);
523 2546 : shared_gen = pg_atomic_read_u64(&ProcSignal->psh_barrierGeneration);
524 :
525 : Assert(local_gen <= shared_gen);
526 :
527 2546 : if (local_gen == shared_gen)
528 0 : return;
529 :
530 : /*
531 : * Get and clear the flags that are set for this backend. Note that
532 : * pg_atomic_exchange_u32 is a full barrier, so we're guaranteed that the
533 : * read of the barrier generation above happens before we atomically
534 : * extract the flags, and that any subsequent state changes happen
535 : * afterward.
536 : *
537 : * NB: In order to avoid race conditions, we must zero
538 : * pss_barrierCheckMask first and only afterwards try to do barrier
539 : * processing. If we did it in the other order, someone could send us
540 : * another barrier of some type right after we called the
541 : * barrier-processing function but before we cleared the bit. We would
542 : * have no way of knowing that the bit needs to stay set in that case, so
543 : * the need to call the barrier-processing function again would just get
544 : * forgotten. So instead, we tentatively clear all the bits and then put
545 : * back any for which we don't manage to successfully absorb the barrier.
546 : */
547 2546 : flags = pg_atomic_exchange_u32(&MyProcSignalSlot->pss_barrierCheckMask, 0);
548 :
549 : /*
550 : * If there are no flags set, then we can skip doing any real work.
551 : * Otherwise, establish a PG_TRY block, so that we don't lose track of
552 : * which types of barrier processing are needed if an ERROR occurs.
553 : */
554 2546 : if (flags != 0)
555 : {
556 2546 : bool success = true;
557 :
558 2546 : PG_TRY();
559 : {
560 : /*
561 : * Process each type of barrier. The barrier-processing functions
562 : * should normally return true, but may return false if the
563 : * barrier can't be absorbed at the current time. This should be
564 : * rare, because it's pretty expensive. Every single
565 : * CHECK_FOR_INTERRUPTS() will return here until we manage to
566 : * absorb the barrier, and that cost will add up in a hurry.
567 : *
568 : * NB: It ought to be OK to call the barrier-processing functions
569 : * unconditionally, but it's more efficient to call only the ones
570 : * that might need us to do something based on the flags.
571 : */
572 7638 : while (flags != 0)
573 : {
574 : ProcSignalBarrierType type;
575 2546 : bool processed = true;
576 :
577 2546 : type = (ProcSignalBarrierType) pg_rightmost_one_pos32(flags);
578 2546 : switch (type)
579 : {
580 661 : case PROCSIGNAL_BARRIER_SMGRRELEASE:
581 661 : processed = ProcessBarrierSmgrRelease();
582 661 : break;
583 1625 : case PROCSIGNAL_BARRIER_UPDATE_XLOG_LOGICAL_INFO:
584 1625 : processed = ProcessBarrierUpdateXLogLogicalInfo();
585 1625 : break;
586 :
587 260 : case PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_ON:
588 : case PROCSIGNAL_BARRIER_CHECKSUM_ON:
589 : case PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_OFF:
590 : case PROCSIGNAL_BARRIER_CHECKSUM_OFF:
591 260 : processed = AbsorbDataChecksumsBarrier(type);
592 260 : break;
593 : }
594 :
595 : /*
596 : * To avoid an infinite loop, we must always unset the bit in
597 : * flags.
598 : */
599 2546 : BARRIER_CLEAR_BIT(flags, type);
600 :
601 : /*
602 : * If we failed to process the barrier, reset the shared bit
603 : * so we try again later, and set a flag so that we don't bump
604 : * our generation.
605 : */
606 2546 : if (!processed)
607 : {
608 0 : ResetProcSignalBarrierBits(((uint32) 1) << type);
609 0 : success = false;
610 : }
611 : }
612 : }
613 0 : PG_CATCH();
614 : {
615 : /*
616 : * If an ERROR occurred, we'll need to try again later to handle
617 : * that barrier type and any others that haven't been handled yet
618 : * or weren't successfully absorbed.
619 : */
620 0 : ResetProcSignalBarrierBits(flags);
621 0 : PG_RE_THROW();
622 : }
623 2546 : PG_END_TRY();
624 :
625 : /*
626 : * If some barrier types were not successfully absorbed, we will have
627 : * to try again later.
628 : */
629 2546 : if (!success)
630 0 : return;
631 : }
632 :
633 : /*
634 : * State changes related to all types of barriers that might have been
635 : * emitted have now been handled, so we can update our notion of the
636 : * generation to the one we observed before beginning the updates. If
637 : * things have changed further, it'll get fixed up when this function is
638 : * next called.
639 : */
640 2546 : pg_atomic_write_u64(&MyProcSignalSlot->pss_barrierGeneration, shared_gen);
641 2546 : ConditionVariableBroadcast(&MyProcSignalSlot->pss_barrierCV);
642 : }
643 :
644 : /*
645 : * If it turns out that we couldn't absorb one or more barrier types, either
646 : * because the barrier-processing functions returned false or due to an error,
647 : * arrange for processing to be retried later.
648 : */
649 : static void
650 0 : ResetProcSignalBarrierBits(uint32 flags)
651 : {
652 0 : pg_atomic_fetch_or_u32(&MyProcSignalSlot->pss_barrierCheckMask, flags);
653 0 : ProcSignalBarrierPending = true;
654 0 : InterruptPending = true;
655 0 : }
656 :
657 : /*
658 : * CheckProcSignal - check to see if a particular reason has been
659 : * signaled, and clear the signal flag. Should be called after receiving
660 : * SIGUSR1.
661 : */
662 : static bool
663 134680 : CheckProcSignal(ProcSignalReason reason)
664 : {
665 134680 : volatile ProcSignalSlot *slot = MyProcSignalSlot;
666 :
667 134680 : if (slot != NULL)
668 : {
669 : /*
670 : * Careful here --- don't clear flag if we haven't seen it set.
671 : * pss_signalFlags is of type "volatile sig_atomic_t" to allow us to
672 : * read it here safely, without holding the spinlock.
673 : */
674 134550 : if (slot->pss_signalFlags[reason])
675 : {
676 7942 : slot->pss_signalFlags[reason] = false;
677 7942 : return true;
678 : }
679 : }
680 :
681 126738 : return false;
682 : }
683 :
684 : /*
685 : * procsignal_sigusr1_handler - handle SIGUSR1 signal.
686 : */
687 : void
688 13468 : procsignal_sigusr1_handler(SIGNAL_ARGS)
689 : {
690 13468 : if (CheckProcSignal(PROCSIG_CATCHUP_INTERRUPT))
691 3173 : HandleCatchupInterrupt();
692 :
693 13468 : if (CheckProcSignal(PROCSIG_NOTIFY_INTERRUPT))
694 11 : HandleNotifyInterrupt();
695 :
696 13468 : if (CheckProcSignal(PROCSIG_PARALLEL_MESSAGE))
697 2105 : HandleParallelMessageInterrupt();
698 :
699 13468 : if (CheckProcSignal(PROCSIG_WALSND_INIT_STOPPING))
700 48 : HandleWalSndInitStopping();
701 :
702 13468 : if (CheckProcSignal(PROCSIG_BARRIER))
703 2550 : HandleProcSignalBarrierInterrupt();
704 :
705 13468 : if (CheckProcSignal(PROCSIG_LOG_MEMORY_CONTEXT))
706 12 : HandleLogMemoryContextInterrupt();
707 :
708 13468 : if (CheckProcSignal(PROCSIG_PARALLEL_APPLY_MESSAGE))
709 17 : HandleParallelApplyMessageInterrupt();
710 :
711 13468 : if (CheckProcSignal(PROCSIG_REPACK_MESSAGE))
712 6 : HandleRepackMessageInterrupt();
713 :
714 13468 : if (CheckProcSignal(PROCSIG_SLOTSYNC_MESSAGE))
715 1 : HandleSlotSyncMessageInterrupt();
716 :
717 13468 : if (CheckProcSignal(PROCSIG_RECOVERY_CONFLICT))
718 19 : HandleRecoveryConflictInterrupt();
719 :
720 13468 : SetLatch(MyLatch);
721 13468 : }
722 :
723 : /*
724 : * Send a query cancellation signal to backend.
725 : *
726 : * Note: This is called from a backend process before authentication. We
727 : * cannot take LWLocks yet, but that's OK; we rely on atomic reads of the
728 : * fields in the ProcSignal slots.
729 : */
730 : void
731 15 : SendCancelRequest(int backendPID, const uint8 *cancel_key, int cancel_key_len)
732 : {
733 15 : if (backendPID == 0)
734 : {
735 0 : ereport(LOG, (errmsg("invalid cancel request with PID 0")));
736 0 : return;
737 : }
738 :
739 : /*
740 : * See if we have a matching backend. Reading the pss_pid and
741 : * pss_cancel_key fields is racy, a backend might die and remove itself
742 : * from the array at any time. The probability of the cancellation key
743 : * matching wrong process is miniscule, however, so we can live with that.
744 : * PIDs are reused too, so sending the signal based on PID is inherently
745 : * racy anyway, although OS's avoid reusing PIDs too soon.
746 : */
747 218 : for (int i = 0; i < NumProcSignalSlots; i++)
748 : {
749 218 : ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
750 : bool match;
751 :
752 218 : if (pg_atomic_read_u32(&slot->pss_pid) != backendPID)
753 203 : continue;
754 :
755 : /* Acquire the spinlock and re-check */
756 15 : SpinLockAcquire(&slot->pss_mutex);
757 15 : if (pg_atomic_read_u32(&slot->pss_pid) != backendPID)
758 : {
759 0 : SpinLockRelease(&slot->pss_mutex);
760 0 : continue;
761 : }
762 : else
763 : {
764 30 : match = slot->pss_cancel_key_len == cancel_key_len &&
765 15 : timingsafe_bcmp(slot->pss_cancel_key, cancel_key, cancel_key_len) == 0;
766 :
767 15 : SpinLockRelease(&slot->pss_mutex);
768 :
769 15 : if (match)
770 : {
771 : /* Found a match; signal that backend to cancel current op */
772 15 : ereport(DEBUG2,
773 : (errmsg_internal("processing cancel request: sending SIGINT to process %d",
774 : backendPID)));
775 :
776 : /*
777 : * If we have setsid(), signal the backend's whole process
778 : * group
779 : */
780 : #ifdef HAVE_SETSID
781 15 : kill(-backendPID, SIGINT);
782 : #else
783 : kill(backendPID, SIGINT);
784 : #endif
785 : }
786 : else
787 : {
788 : /* Right PID, wrong key: no way, Jose */
789 0 : ereport(LOG,
790 : (errmsg("wrong key in cancel request for process %d",
791 : backendPID)));
792 : }
793 15 : return;
794 : }
795 : }
796 :
797 : /* No matching backend */
798 0 : ereport(LOG,
799 : (errmsg("PID %d in cancel request did not match any process",
800 : backendPID)));
801 : }
|
