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 "miscadmin.h"
23 : #include "pgstat.h"
24 : #include "port/pg_bitutils.h"
25 : #include "postmaster/datachecksum_state.h"
26 : #include "replication/logicalctl.h"
27 : #include "replication/logicalworker.h"
28 : #include "replication/walsender.h"
29 : #include "storage/condition_variable.h"
30 : #include "storage/ipc.h"
31 : #include "storage/latch.h"
32 : #include "storage/proc.h"
33 : #include "storage/shmem.h"
34 : #include "storage/sinval.h"
35 : #include "storage/smgr.h"
36 : #include "storage/subsystems.h"
37 : #include "tcop/tcopprot.h"
38 : #include "utils/memutils.h"
39 : #include "utils/wait_event.h"
40 :
41 : /*
42 : * The SIGUSR1 signal is multiplexed to support signaling multiple event
43 : * types. The specific reason is communicated via flags in shared memory.
44 : * We keep a boolean flag for each possible "reason", so that different
45 : * reasons can be signaled to a process concurrently. (However, if the same
46 : * reason is signaled more than once nearly simultaneously, the process may
47 : * observe it only once.)
48 : *
49 : * Each process that wants to receive signals registers its process ID
50 : * in the ProcSignalSlots array. The array is indexed by ProcNumber to make
51 : * slot allocation simple, and to avoid having to search the array when you
52 : * know the ProcNumber of the process you're signaling. (We do support
53 : * signaling without ProcNumber, but it's a bit less efficient.)
54 : *
55 : * The fields in each slot are protected by a spinlock, pss_mutex. pss_pid can
56 : * also be read without holding the spinlock, as a quick preliminary check
57 : * when searching for a particular PID in the array.
58 : *
59 : * pss_signalFlags are intended to be set in cases where we don't need to
60 : * keep track of whether or not the target process has handled the signal,
61 : * but sometimes we need confirmation, as when making a global state change
62 : * that cannot be considered complete until all backends have taken notice
63 : * of it. For such use cases, we set a bit in pss_barrierCheckMask and then
64 : * increment the current "barrier generation"; when the new barrier generation
65 : * (or greater) appears in the pss_barrierGeneration flag of every process,
66 : * we know that the message has been received everywhere.
67 : */
68 : typedef struct
69 : {
70 : pg_atomic_uint32 pss_pid;
71 : int pss_cancel_key_len; /* 0 means no cancellation is possible */
72 : uint8 pss_cancel_key[MAX_CANCEL_KEY_LENGTH];
73 : volatile sig_atomic_t pss_signalFlags[NUM_PROCSIGNALS];
74 : slock_t pss_mutex; /* protects the above fields */
75 :
76 : /* Barrier-related fields (not protected by pss_mutex) */
77 : pg_atomic_uint64 pss_barrierGeneration;
78 : pg_atomic_uint32 pss_barrierCheckMask;
79 : ConditionVariable pss_barrierCV;
80 : } ProcSignalSlot;
81 :
82 : /*
83 : * Information that is global to the entire ProcSignal system can be stored
84 : * here.
85 : *
86 : * psh_barrierGeneration is the highest barrier generation in existence.
87 : */
88 : struct ProcSignalHeader
89 : {
90 : pg_atomic_uint64 psh_barrierGeneration;
91 : ProcSignalSlot psh_slot[FLEXIBLE_ARRAY_MEMBER];
92 : };
93 :
94 : /*
95 : * We reserve a slot for each possible ProcNumber, plus one for each
96 : * possible auxiliary process type. (This scheme assumes there is not
97 : * more than one of any auxiliary process type at a time, except for
98 : * IO workers.)
99 : */
100 : #define NumProcSignalSlots (MaxBackends + NUM_AUXILIARY_PROCS)
101 :
102 : /* Check whether the relevant type bit is set in the flags. */
103 : #define BARRIER_SHOULD_CHECK(flags, type) \
104 : (((flags) & (((uint32) 1) << (uint32) (type))) != 0)
105 :
106 : /* Clear the relevant type bit from the flags. */
107 : #define BARRIER_CLEAR_BIT(flags, type) \
108 : ((flags) &= ~(((uint32) 1) << (uint32) (type)))
109 :
110 : static void ProcSignalShmemRequest(void *arg);
111 : static void ProcSignalShmemInit(void *arg);
112 :
113 : const ShmemCallbacks ProcSignalShmemCallbacks = {
114 : .request_fn = ProcSignalShmemRequest,
115 : .init_fn = ProcSignalShmemInit,
116 : };
117 :
118 : NON_EXEC_STATIC ProcSignalHeader *ProcSignal = NULL;
119 :
120 : static ProcSignalSlot *MyProcSignalSlot = NULL;
121 :
122 : static bool CheckProcSignal(ProcSignalReason reason);
123 : static void CleanupProcSignalState(int status, Datum arg);
124 : static void ResetProcSignalBarrierBits(uint32 flags);
125 :
126 : /*
127 : * ProcSignalShmemRequest
128 : * Register ProcSignal's shared memory needs at postmaster startup
129 : */
130 : static void
131 1233 : ProcSignalShmemRequest(void *arg)
132 : {
133 : Size size;
134 :
135 1233 : size = mul_size(NumProcSignalSlots, sizeof(ProcSignalSlot));
136 1233 : size = add_size(size, offsetof(ProcSignalHeader, psh_slot));
137 :
138 1233 : ShmemRequestStruct(.name = "ProcSignal",
139 : .size = size,
140 : .ptr = (void **) &ProcSignal,
141 : );
142 1233 : }
143 :
144 : static void
145 1230 : ProcSignalShmemInit(void *arg)
146 : {
147 1230 : pg_atomic_init_u64(&ProcSignal->psh_barrierGeneration, 0);
148 :
149 163161 : for (int i = 0; i < NumProcSignalSlots; ++i)
150 : {
151 161931 : ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
152 :
153 161931 : SpinLockInit(&slot->pss_mutex);
154 161931 : pg_atomic_init_u32(&slot->pss_pid, 0);
155 161931 : slot->pss_cancel_key_len = 0;
156 809655 : MemSet(slot->pss_signalFlags, 0, sizeof(slot->pss_signalFlags));
157 161931 : pg_atomic_init_u64(&slot->pss_barrierGeneration, PG_UINT64_MAX);
158 161931 : pg_atomic_init_u32(&slot->pss_barrierCheckMask, 0);
159 161931 : ConditionVariableInit(&slot->pss_barrierCV);
160 : }
161 1230 : }
162 :
163 : /*
164 : * ProcSignalInit
165 : * Register the current process in the ProcSignal array
166 : */
167 : void
168 24798 : ProcSignalInit(const uint8 *cancel_key, int cancel_key_len)
169 : {
170 : ProcSignalSlot *slot;
171 : uint64 barrier_generation;
172 : uint32 old_pss_pid;
173 :
174 : Assert(cancel_key_len >= 0 && cancel_key_len <= MAX_CANCEL_KEY_LENGTH);
175 24798 : if (MyProcNumber < 0)
176 0 : elog(ERROR, "MyProcNumber not set");
177 24798 : if (MyProcNumber >= NumProcSignalSlots)
178 0 : elog(ERROR, "unexpected MyProcNumber %d in ProcSignalInit (max %d)", MyProcNumber, NumProcSignalSlots);
179 24798 : slot = &ProcSignal->psh_slot[MyProcNumber];
180 :
181 24798 : SpinLockAcquire(&slot->pss_mutex);
182 :
183 : /* Value used for sanity check below */
184 24798 : old_pss_pid = pg_atomic_read_u32(&slot->pss_pid);
185 :
186 : /* Clear out any leftover signal reasons */
187 123990 : MemSet(slot->pss_signalFlags, 0, NUM_PROCSIGNALS * sizeof(sig_atomic_t));
188 :
189 : /*
190 : * Initialize barrier state. Since we're a brand-new process, there
191 : * shouldn't be any leftover backend-private state that needs to be
192 : * updated. Therefore, we can broadcast the latest barrier generation and
193 : * disregard any previously-set check bits.
194 : *
195 : * NB: This only works if this initialization happens early enough in the
196 : * startup sequence that we haven't yet cached any state that might need
197 : * to be invalidated. That's also why we have a memory barrier here, to be
198 : * sure that any later reads of memory happen strictly after this.
199 : */
200 24798 : pg_atomic_write_u32(&slot->pss_barrierCheckMask, 0);
201 : barrier_generation =
202 24798 : pg_atomic_read_u64(&ProcSignal->psh_barrierGeneration);
203 24798 : pg_atomic_write_u64(&slot->pss_barrierGeneration, barrier_generation);
204 :
205 24798 : if (cancel_key_len > 0)
206 14556 : memcpy(slot->pss_cancel_key, cancel_key, cancel_key_len);
207 24798 : slot->pss_cancel_key_len = cancel_key_len;
208 24798 : pg_atomic_write_u32(&slot->pss_pid, MyProcPid);
209 :
210 24798 : SpinLockRelease(&slot->pss_mutex);
211 :
212 : /* Spinlock is released, do the check */
213 24798 : if (old_pss_pid != 0)
214 0 : elog(LOG, "process %d taking over ProcSignal slot %d, but it's not empty",
215 : MyProcPid, MyProcNumber);
216 :
217 : /* Remember slot location for CheckProcSignal */
218 24798 : MyProcSignalSlot = slot;
219 :
220 : /* Set up to release the slot on process exit */
221 24798 : on_shmem_exit(CleanupProcSignalState, (Datum) 0);
222 24798 : }
223 :
224 : /*
225 : * CleanupProcSignalState
226 : * Remove current process from ProcSignal mechanism
227 : *
228 : * This function is called via on_shmem_exit() during backend shutdown.
229 : */
230 : static void
231 24798 : CleanupProcSignalState(int status, Datum arg)
232 : {
233 : pid_t old_pid;
234 24798 : ProcSignalSlot *slot = MyProcSignalSlot;
235 :
236 : /*
237 : * Clear MyProcSignalSlot, so that a SIGUSR1 received after this point
238 : * won't try to access it after it's no longer ours (and perhaps even
239 : * after we've unmapped the shared memory segment).
240 : */
241 : Assert(MyProcSignalSlot != NULL);
242 24798 : MyProcSignalSlot = NULL;
243 :
244 : /* sanity check */
245 24798 : SpinLockAcquire(&slot->pss_mutex);
246 24798 : old_pid = pg_atomic_read_u32(&slot->pss_pid);
247 24798 : if (old_pid != MyProcPid)
248 : {
249 : /*
250 : * don't ERROR here. We're exiting anyway, and don't want to get into
251 : * infinite loop trying to exit
252 : */
253 0 : SpinLockRelease(&slot->pss_mutex);
254 0 : elog(LOG, "process %d releasing ProcSignal slot %d, but it contains %d",
255 : MyProcPid, (int) (slot - ProcSignal->psh_slot), (int) old_pid);
256 0 : return; /* XXX better to zero the slot anyway? */
257 : }
258 :
259 : /* Mark the slot as unused */
260 24798 : pg_atomic_write_u32(&slot->pss_pid, 0);
261 24798 : slot->pss_cancel_key_len = 0;
262 :
263 : /*
264 : * Make this slot look like it's absorbed all possible barriers, so that
265 : * no barrier waits block on it.
266 : */
267 24798 : pg_atomic_write_u64(&slot->pss_barrierGeneration, PG_UINT64_MAX);
268 :
269 24798 : SpinLockRelease(&slot->pss_mutex);
270 :
271 24798 : ConditionVariableBroadcast(&slot->pss_barrierCV);
272 : }
273 :
274 : /*
275 : * SendProcSignal
276 : * Send a signal to a Postgres process
277 : *
278 : * Providing procNumber is optional, but it will speed up the operation.
279 : *
280 : * On success (a signal was sent), zero is returned.
281 : * On error, -1 is returned, and errno is set (typically to ESRCH or EPERM).
282 : *
283 : * Not to be confused with ProcSendSignal
284 : */
285 : int
286 7642 : SendProcSignal(pid_t pid, ProcSignalReason reason, ProcNumber procNumber)
287 : {
288 : volatile ProcSignalSlot *slot;
289 :
290 7642 : if (procNumber != INVALID_PROC_NUMBER)
291 : {
292 : Assert(procNumber < NumProcSignalSlots);
293 7574 : slot = &ProcSignal->psh_slot[procNumber];
294 :
295 7574 : SpinLockAcquire(&slot->pss_mutex);
296 7574 : if (pg_atomic_read_u32(&slot->pss_pid) == pid)
297 : {
298 : /* Atomically set the proper flag */
299 7574 : slot->pss_signalFlags[reason] = true;
300 7574 : SpinLockRelease(&slot->pss_mutex);
301 : /* Send signal */
302 7574 : return kill(pid, SIGUSR1);
303 : }
304 0 : SpinLockRelease(&slot->pss_mutex);
305 : }
306 : else
307 : {
308 : /*
309 : * procNumber not provided, so search the array using pid. We search
310 : * the array back to front so as to reduce search overhead. Passing
311 : * INVALID_PROC_NUMBER means that the target is most likely an
312 : * auxiliary process, which will have a slot near the end of the
313 : * array.
314 : */
315 : int i;
316 :
317 3070 : for (i = NumProcSignalSlots - 1; i >= 0; i--)
318 : {
319 3070 : slot = &ProcSignal->psh_slot[i];
320 :
321 3070 : if (pg_atomic_read_u32(&slot->pss_pid) == pid)
322 : {
323 68 : SpinLockAcquire(&slot->pss_mutex);
324 68 : if (pg_atomic_read_u32(&slot->pss_pid) == pid)
325 : {
326 : /* Atomically set the proper flag */
327 68 : slot->pss_signalFlags[reason] = true;
328 68 : SpinLockRelease(&slot->pss_mutex);
329 : /* Send signal */
330 68 : return kill(pid, SIGUSR1);
331 : }
332 0 : SpinLockRelease(&slot->pss_mutex);
333 : }
334 : }
335 : }
336 :
337 0 : errno = ESRCH;
338 0 : return -1;
339 : }
340 :
341 : /*
342 : * EmitProcSignalBarrier
343 : * Send a signal to every Postgres process
344 : *
345 : * The return value of this function is the barrier "generation" created
346 : * by this operation. This value can be passed to WaitForProcSignalBarrier
347 : * to wait until it is known that every participant in the ProcSignal
348 : * mechanism has absorbed the signal (or started afterwards).
349 : *
350 : * Note that it would be a bad idea to use this for anything that happens
351 : * frequently, as interrupting every backend could cause a noticeable
352 : * performance hit.
353 : *
354 : * Callers are entitled to assume that this function will not throw ERROR
355 : * or FATAL.
356 : */
357 : uint64
358 663 : EmitProcSignalBarrier(ProcSignalBarrierType type)
359 : {
360 663 : uint32 flagbit = 1 << (uint32) type;
361 : uint64 generation;
362 :
363 : /*
364 : * Set all the flags.
365 : *
366 : * Note that pg_atomic_fetch_or_u32 has full barrier semantics, so this is
367 : * totally ordered with respect to anything the caller did before, and
368 : * anything that we do afterwards. (This is also true of the later call to
369 : * pg_atomic_add_fetch_u64.)
370 : */
371 70820 : for (int i = 0; i < NumProcSignalSlots; i++)
372 : {
373 70157 : volatile ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
374 :
375 70157 : pg_atomic_fetch_or_u32(&slot->pss_barrierCheckMask, flagbit);
376 : }
377 :
378 : /*
379 : * Increment the generation counter.
380 : */
381 : generation =
382 663 : pg_atomic_add_fetch_u64(&ProcSignal->psh_barrierGeneration, 1);
383 :
384 : /*
385 : * Signal all the processes, so that they update their advertised barrier
386 : * generation.
387 : *
388 : * Concurrency is not a problem here. Backends that have exited don't
389 : * matter, and new backends that have joined since we entered this
390 : * function must already have current state, since the caller is
391 : * responsible for making sure that the relevant state is entirely visible
392 : * before calling this function in the first place. We still have to wake
393 : * them up - because we can't distinguish between such backends and older
394 : * backends that need to update state - but they won't actually need to
395 : * change any state.
396 : */
397 70820 : for (int i = NumProcSignalSlots - 1; i >= 0; i--)
398 : {
399 70157 : volatile ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
400 70157 : pid_t pid = pg_atomic_read_u32(&slot->pss_pid);
401 :
402 70157 : if (pid != 0)
403 : {
404 4080 : SpinLockAcquire(&slot->pss_mutex);
405 4080 : pid = pg_atomic_read_u32(&slot->pss_pid);
406 4080 : if (pid != 0)
407 : {
408 : /* see SendProcSignal for details */
409 4080 : slot->pss_signalFlags[PROCSIG_BARRIER] = true;
410 4080 : SpinLockRelease(&slot->pss_mutex);
411 4080 : kill(pid, SIGUSR1);
412 : }
413 : else
414 0 : SpinLockRelease(&slot->pss_mutex);
415 : }
416 : }
417 :
418 663 : return generation;
419 : }
420 :
421 : /*
422 : * WaitForProcSignalBarrier - wait until it is guaranteed that all changes
423 : * requested by a specific call to EmitProcSignalBarrier() have taken effect.
424 : */
425 : void
426 647 : WaitForProcSignalBarrier(uint64 generation)
427 : {
428 : Assert(generation <= pg_atomic_read_u64(&ProcSignal->psh_barrierGeneration));
429 :
430 647 : elog(DEBUG1,
431 : "waiting for all backends to process ProcSignalBarrier generation "
432 : UINT64_FORMAT,
433 : generation);
434 :
435 69674 : for (int i = NumProcSignalSlots - 1; i >= 0; i--)
436 : {
437 69027 : ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
438 : uint64 oldval;
439 :
440 : /*
441 : * It's important that we check only pss_barrierGeneration here and
442 : * not pss_barrierCheckMask. Bits in pss_barrierCheckMask get cleared
443 : * before the barrier is actually absorbed, but pss_barrierGeneration
444 : * is updated only afterward.
445 : */
446 69027 : oldval = pg_atomic_read_u64(&slot->pss_barrierGeneration);
447 72024 : while (oldval < generation)
448 : {
449 2997 : if (ConditionVariableTimedSleep(&slot->pss_barrierCV,
450 : 5000,
451 : WAIT_EVENT_PROC_SIGNAL_BARRIER))
452 0 : ereport(LOG,
453 : (errmsg("still waiting for backend with PID %d to accept ProcSignalBarrier",
454 : (int) pg_atomic_read_u32(&slot->pss_pid))));
455 2997 : oldval = pg_atomic_read_u64(&slot->pss_barrierGeneration);
456 : }
457 69027 : ConditionVariableCancelSleep();
458 : }
459 :
460 647 : elog(DEBUG1,
461 : "finished waiting for all backends to process ProcSignalBarrier generation "
462 : UINT64_FORMAT,
463 : generation);
464 :
465 : /*
466 : * The caller is probably calling this function because it wants to read
467 : * the shared state or perform further writes to shared state once all
468 : * backends are known to have absorbed the barrier. However, the read of
469 : * pss_barrierGeneration was performed unlocked; insert a memory barrier
470 : * to separate it from whatever follows.
471 : */
472 647 : pg_memory_barrier();
473 647 : }
474 :
475 : /*
476 : * Handle receipt of an interrupt indicating a global barrier event.
477 : *
478 : * All the actual work is deferred to ProcessProcSignalBarrier(), because we
479 : * cannot safely access the barrier generation inside the signal handler as
480 : * 64bit atomics might use spinlock based emulation, even for reads. As this
481 : * routine only gets called when PROCSIG_BARRIER is sent that won't cause a
482 : * lot of unnecessary work.
483 : */
484 : static void
485 3016 : HandleProcSignalBarrierInterrupt(void)
486 : {
487 3016 : InterruptPending = true;
488 3016 : ProcSignalBarrierPending = true;
489 : /* latch will be set by procsignal_sigusr1_handler */
490 3016 : }
491 :
492 : /*
493 : * Perform global barrier related interrupt checking.
494 : *
495 : * Any backend that participates in ProcSignal signaling must arrange to
496 : * call this function periodically. It is called from CHECK_FOR_INTERRUPTS(),
497 : * which is enough for normal backends, but not necessarily for all types of
498 : * background processes.
499 : */
500 : void
501 3009 : ProcessProcSignalBarrier(void)
502 : {
503 : uint64 local_gen;
504 : uint64 shared_gen;
505 : volatile uint32 flags;
506 :
507 : Assert(MyProcSignalSlot);
508 :
509 : /* Exit quickly if there's no work to do. */
510 3009 : if (!ProcSignalBarrierPending)
511 0 : return;
512 3009 : ProcSignalBarrierPending = false;
513 :
514 : /*
515 : * It's not unlikely to process multiple barriers at once, before the
516 : * signals for all the barriers have arrived. To avoid unnecessary work in
517 : * response to subsequent signals, exit early if we already have processed
518 : * all of them.
519 : */
520 3009 : local_gen = pg_atomic_read_u64(&MyProcSignalSlot->pss_barrierGeneration);
521 3009 : shared_gen = pg_atomic_read_u64(&ProcSignal->psh_barrierGeneration);
522 :
523 : Assert(local_gen <= shared_gen);
524 :
525 3009 : if (local_gen == shared_gen)
526 2 : return;
527 :
528 : /*
529 : * Get and clear the flags that are set for this backend. Note that
530 : * pg_atomic_exchange_u32 is a full barrier, so we're guaranteed that the
531 : * read of the barrier generation above happens before we atomically
532 : * extract the flags, and that any subsequent state changes happen
533 : * afterward.
534 : *
535 : * NB: In order to avoid race conditions, we must zero
536 : * pss_barrierCheckMask first and only afterwards try to do barrier
537 : * processing. If we did it in the other order, someone could send us
538 : * another barrier of some type right after we called the
539 : * barrier-processing function but before we cleared the bit. We would
540 : * have no way of knowing that the bit needs to stay set in that case, so
541 : * the need to call the barrier-processing function again would just get
542 : * forgotten. So instead, we tentatively clear all the bits and then put
543 : * back any for which we don't manage to successfully absorb the barrier.
544 : */
545 3007 : flags = pg_atomic_exchange_u32(&MyProcSignalSlot->pss_barrierCheckMask, 0);
546 :
547 : /*
548 : * If there are no flags set, then we can skip doing any real work.
549 : * Otherwise, establish a PG_TRY block, so that we don't lose track of
550 : * which types of barrier processing are needed if an ERROR occurs.
551 : */
552 3007 : if (flags != 0)
553 : {
554 3007 : bool success = true;
555 :
556 3007 : PG_TRY();
557 : {
558 : /*
559 : * Process each type of barrier. The barrier-processing functions
560 : * should normally return true, but may return false if the
561 : * barrier can't be absorbed at the current time. This should be
562 : * rare, because it's pretty expensive. Every single
563 : * CHECK_FOR_INTERRUPTS() will return here until we manage to
564 : * absorb the barrier, and that cost will add up in a hurry.
565 : *
566 : * NB: It ought to be OK to call the barrier-processing functions
567 : * unconditionally, but it's more efficient to call only the ones
568 : * that might need us to do something based on the flags.
569 : */
570 9021 : while (flags != 0)
571 : {
572 : ProcSignalBarrierType type;
573 3007 : bool processed = true;
574 :
575 3007 : type = (ProcSignalBarrierType) pg_rightmost_one_pos32(flags);
576 3007 : switch (type)
577 : {
578 829 : case PROCSIGNAL_BARRIER_SMGRRELEASE:
579 829 : processed = ProcessBarrierSmgrRelease();
580 829 : break;
581 1889 : case PROCSIGNAL_BARRIER_UPDATE_XLOG_LOGICAL_INFO:
582 1889 : processed = ProcessBarrierUpdateXLogLogicalInfo();
583 1889 : break;
584 :
585 289 : case PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_ON:
586 : case PROCSIGNAL_BARRIER_CHECKSUM_ON:
587 : case PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_OFF:
588 : case PROCSIGNAL_BARRIER_CHECKSUM_OFF:
589 289 : processed = AbsorbDataChecksumsBarrier(type);
590 289 : break;
591 : }
592 :
593 : /*
594 : * To avoid an infinite loop, we must always unset the bit in
595 : * flags.
596 : */
597 3007 : BARRIER_CLEAR_BIT(flags, type);
598 :
599 : /*
600 : * If we failed to process the barrier, reset the shared bit
601 : * so we try again later, and set a flag so that we don't bump
602 : * our generation.
603 : */
604 3007 : if (!processed)
605 : {
606 0 : ResetProcSignalBarrierBits(((uint32) 1) << type);
607 0 : success = false;
608 : }
609 : }
610 : }
611 0 : PG_CATCH();
612 : {
613 : /*
614 : * If an ERROR occurred, we'll need to try again later to handle
615 : * that barrier type and any others that haven't been handled yet
616 : * or weren't successfully absorbed.
617 : */
618 0 : ResetProcSignalBarrierBits(flags);
619 0 : PG_RE_THROW();
620 : }
621 3007 : PG_END_TRY();
622 :
623 : /*
624 : * If some barrier types were not successfully absorbed, we will have
625 : * to try again later.
626 : */
627 3007 : if (!success)
628 0 : return;
629 : }
630 :
631 : /*
632 : * State changes related to all types of barriers that might have been
633 : * emitted have now been handled, so we can update our notion of the
634 : * generation to the one we observed before beginning the updates. If
635 : * things have changed further, it'll get fixed up when this function is
636 : * next called.
637 : */
638 3007 : pg_atomic_write_u64(&MyProcSignalSlot->pss_barrierGeneration, shared_gen);
639 3007 : ConditionVariableBroadcast(&MyProcSignalSlot->pss_barrierCV);
640 : }
641 :
642 : /*
643 : * If it turns out that we couldn't absorb one or more barrier types, either
644 : * because the barrier-processing functions returned false or due to an error,
645 : * arrange for processing to be retried later.
646 : */
647 : static void
648 0 : ResetProcSignalBarrierBits(uint32 flags)
649 : {
650 0 : pg_atomic_fetch_or_u32(&MyProcSignalSlot->pss_barrierCheckMask, flags);
651 0 : ProcSignalBarrierPending = true;
652 0 : InterruptPending = true;
653 0 : }
654 :
655 : /*
656 : * CheckProcSignal - check to see if a particular reason has been
657 : * signaled, and clear the signal flag. Should be called after receiving
658 : * SIGUSR1.
659 : */
660 : static bool
661 111664 : CheckProcSignal(ProcSignalReason reason)
662 : {
663 111664 : volatile ProcSignalSlot *slot = MyProcSignalSlot;
664 :
665 111664 : if (slot != NULL)
666 : {
667 : /*
668 : * Careful here --- don't clear flag if we haven't seen it set.
669 : * pss_signalFlags is of type "volatile sig_atomic_t" to allow us to
670 : * read it here safely, without holding the spinlock.
671 : */
672 111520 : if (slot->pss_signalFlags[reason])
673 : {
674 8345 : slot->pss_signalFlags[reason] = false;
675 8345 : return true;
676 : }
677 : }
678 :
679 103319 : return false;
680 : }
681 :
682 : /*
683 : * procsignal_sigusr1_handler - handle SIGUSR1 signal.
684 : */
685 : void
686 13958 : procsignal_sigusr1_handler(SIGNAL_ARGS)
687 : {
688 13958 : if (CheckProcSignal(PROCSIG_CATCHUP_INTERRUPT))
689 3139 : HandleCatchupInterrupt();
690 :
691 13958 : if (CheckProcSignal(PROCSIG_NOTIFY_INTERRUPT))
692 11 : HandleNotifyInterrupt();
693 :
694 13958 : if (CheckProcSignal(PROCSIG_PARALLEL_MESSAGE))
695 2088 : HandleParallelMessageInterrupt();
696 :
697 13958 : if (CheckProcSignal(PROCSIG_WALSND_INIT_STOPPING))
698 47 : HandleWalSndInitStopping();
699 :
700 13958 : if (CheckProcSignal(PROCSIG_BARRIER))
701 3016 : HandleProcSignalBarrierInterrupt();
702 :
703 13958 : if (CheckProcSignal(PROCSIG_LOG_MEMORY_CONTEXT))
704 11 : HandleLogMemoryContextInterrupt();
705 :
706 13958 : if (CheckProcSignal(PROCSIG_PARALLEL_APPLY_MESSAGE))
707 15 : HandleParallelApplyMessageInterrupt();
708 :
709 13958 : if (CheckProcSignal(PROCSIG_RECOVERY_CONFLICT))
710 18 : HandleRecoveryConflictInterrupt();
711 :
712 13958 : SetLatch(MyLatch);
713 13958 : }
714 :
715 : /*
716 : * Send a query cancellation signal to backend.
717 : *
718 : * Note: This is called from a backend process before authentication. We
719 : * cannot take LWLocks yet, but that's OK; we rely on atomic reads of the
720 : * fields in the ProcSignal slots.
721 : */
722 : void
723 16 : SendCancelRequest(int backendPID, const uint8 *cancel_key, int cancel_key_len)
724 : {
725 16 : if (backendPID == 0)
726 : {
727 0 : ereport(LOG, (errmsg("invalid cancel request with PID 0")));
728 0 : return;
729 : }
730 :
731 : /*
732 : * See if we have a matching backend. Reading the pss_pid and
733 : * pss_cancel_key fields is racy, a backend might die and remove itself
734 : * from the array at any time. The probability of the cancellation key
735 : * matching wrong process is miniscule, however, so we can live with that.
736 : * PIDs are reused too, so sending the signal based on PID is inherently
737 : * racy anyway, although OS's avoid reusing PIDs too soon.
738 : */
739 276 : for (int i = 0; i < NumProcSignalSlots; i++)
740 : {
741 276 : ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
742 : bool match;
743 :
744 276 : if (pg_atomic_read_u32(&slot->pss_pid) != backendPID)
745 260 : continue;
746 :
747 : /* Acquire the spinlock and re-check */
748 16 : SpinLockAcquire(&slot->pss_mutex);
749 16 : if (pg_atomic_read_u32(&slot->pss_pid) != backendPID)
750 : {
751 0 : SpinLockRelease(&slot->pss_mutex);
752 0 : continue;
753 : }
754 : else
755 : {
756 32 : match = slot->pss_cancel_key_len == cancel_key_len &&
757 16 : timingsafe_bcmp(slot->pss_cancel_key, cancel_key, cancel_key_len) == 0;
758 :
759 16 : SpinLockRelease(&slot->pss_mutex);
760 :
761 16 : if (match)
762 : {
763 : /* Found a match; signal that backend to cancel current op */
764 16 : ereport(DEBUG2,
765 : (errmsg_internal("processing cancel request: sending SIGINT to process %d",
766 : backendPID)));
767 :
768 : /*
769 : * If we have setsid(), signal the backend's whole process
770 : * group
771 : */
772 : #ifdef HAVE_SETSID
773 16 : kill(-backendPID, SIGINT);
774 : #else
775 : kill(backendPID, SIGINT);
776 : #endif
777 : }
778 : else
779 : {
780 : /* Right PID, wrong key: no way, Jose */
781 0 : ereport(LOG,
782 : (errmsg("wrong key in cancel request for process %d",
783 : backendPID)));
784 : }
785 16 : return;
786 : }
787 : }
788 :
789 : /* No matching backend */
790 0 : ereport(LOG,
791 : (errmsg("PID %d in cancel request did not match any process",
792 : backendPID)));
793 : }
|
