Age Owner Branch data TLA 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
54 heikki.linnakangas@i 133 :GNC 1251 : ProcSignalShmemRequest(void *arg)
134 : : {
135 : : Size size;
136 : :
1509 rhaas@postgresql.org 137 :CBC 1251 : size = mul_size(NumProcSignalSlots, sizeof(ProcSignalSlot));
2354 138 : 1251 : size = add_size(size, offsetof(ProcSignalHeader, psh_slot));
139 : :
54 heikki.linnakangas@i 140 :GNC 1251 : ShmemRequestStruct(.name = "ProcSignal",
141 : : .size = size,
142 : : .ptr = (void **) &ProcSignal,
143 : : );
6147 tgl@sss.pgh.pa.us 144 :GIC 1251 : }
145 : :
146 : : static void
54 heikki.linnakangas@i 147 :GNC 1248 : ProcSignalShmemInit(void *arg)
148 : : {
149 : 1248 : pg_atomic_init_u64(&ProcSignal->psh_barrierGeneration, 0);
150 : :
151 [ + + ]: 164708 : for (int i = 0; i < NumProcSignalSlots; ++i)
152 : : {
153 : 163460 : ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
154 : :
155 : 163460 : SpinLockInit(&slot->pss_mutex);
156 : 163460 : pg_atomic_init_u32(&slot->pss_pid, 0);
157 : 163460 : slot->pss_cancel_key_len = 0;
158 [ + - + - : 980760 : MemSet(slot->pss_signalFlags, 0, sizeof(slot->pss_signalFlags));
+ - + - +
+ ]
159 : 163460 : pg_atomic_init_u64(&slot->pss_barrierGeneration, PG_UINT64_MAX);
160 : 163460 : pg_atomic_init_u32(&slot->pss_barrierCheckMask, 0);
161 : 163460 : ConditionVariableInit(&slot->pss_barrierCV);
162 : : }
6147 tgl@sss.pgh.pa.us 163 :CBC 1248 : }
164 : :
165 : : /*
166 : : * ProcSignalInit
167 : : * Register the current process in the ProcSignal array
168 : : */
169 : : void
387 heikki.linnakangas@i 170 : 23095 : ProcSignalInit(const uint8 *cancel_key, int cancel_key_len)
171 : : {
172 : : ProcSignalSlot *slot;
173 : : uint64 barrier_generation;
174 : : uint32 old_pss_pid;
175 : :
423 176 [ + - - + ]: 23095 : Assert(cancel_key_len >= 0 && cancel_key_len <= MAX_CANCEL_KEY_LENGTH);
818 177 [ - + ]: 23095 : if (MyProcNumber < 0)
818 heikki.linnakangas@i 178 [ # # ]:UBC 0 : elog(ERROR, "MyProcNumber not set");
818 heikki.linnakangas@i 179 [ - + ]:CBC 23095 : if (MyProcNumber >= NumProcSignalSlots)
818 heikki.linnakangas@i 180 [ # # ]:UBC 0 : elog(ERROR, "unexpected MyProcNumber %d in ProcSignalInit (max %d)", MyProcNumber, NumProcSignalSlots);
818 heikki.linnakangas@i 181 :CBC 23095 : slot = &ProcSignal->psh_slot[MyProcNumber];
182 : :
670 183 [ - + ]: 23095 : SpinLockAcquire(&slot->pss_mutex);
184 : :
185 : : /* Value used for sanity check below */
457 michael@paquier.xyz 186 : 23095 : old_pss_pid = pg_atomic_read_u32(&slot->pss_pid);
187 : :
188 : : /* Clear out any leftover signal reasons */
6147 tgl@sss.pgh.pa.us 189 [ + - + - : 138570 : MemSet(slot->pss_signalFlags, 0, NUM_PROCSIGNALS * sizeof(sig_atomic_t));
+ - + - +
+ ]
190 : :
191 : : /*
192 : : * Publish the PID before reading the global barrier generation to ensure
193 : : * that EmitProcSignalBarrier() doesn't skip us while we are grabbing an
194 : : * older generation. We need a memory barrier here to make sure that the
195 : : * update of pss_pid is ordered before the subsequent load of
196 : : * psh_barrierGeneration.
197 : : */
3 msawada@postgresql.o 198 : 23095 : pg_atomic_write_membarrier_u32(&slot->pss_pid, MyProcPid);
199 : :
200 : : /*
201 : : * Initialize barrier state. Since we're a brand-new process, there
202 : : * shouldn't be any leftover backend-private state that needs to be
203 : : * updated. Therefore, we can broadcast the latest barrier generation and
204 : : * disregard any previously-set check bits.
205 : : *
206 : : * NB: This only works if this initialization happens early enough in the
207 : : * startup sequence that we haven't yet cached any state that might need
208 : : * to be invalidated. That's also why we have a memory barrier here, to be
209 : : * sure that any later reads of memory happen strictly after this.
210 : : */
2354 rhaas@postgresql.org 211 : 23095 : pg_atomic_write_u32(&slot->pss_barrierCheckMask, 0);
212 : : barrier_generation =
213 : 23095 : pg_atomic_read_u64(&ProcSignal->psh_barrierGeneration);
214 : 23095 : pg_atomic_write_u64(&slot->pss_barrierGeneration, barrier_generation);
215 : :
423 heikki.linnakangas@i 216 [ + + ]: 23095 : if (cancel_key_len > 0)
217 : 14861 : memcpy(slot->pss_cancel_key, cancel_key, cancel_key_len);
218 : 23095 : slot->pss_cancel_key_len = cancel_key_len;
219 : :
670 220 : 23095 : SpinLockRelease(&slot->pss_mutex);
221 : :
222 : : /* Spinlock is released, do the check */
457 michael@paquier.xyz 223 [ - + ]: 23095 : if (old_pss_pid != 0)
457 michael@paquier.xyz 224 [ # # ]:UBC 0 : elog(LOG, "process %d taking over ProcSignal slot %d, but it's not empty",
225 : : MyProcPid, MyProcNumber);
226 : :
227 : : /* Remember slot location for CheckProcSignal */
6147 tgl@sss.pgh.pa.us 228 :CBC 23095 : MyProcSignalSlot = slot;
229 : :
230 : : /* Set up to release the slot on process exit */
818 heikki.linnakangas@i 231 : 23095 : on_shmem_exit(CleanupProcSignalState, (Datum) 0);
6147 tgl@sss.pgh.pa.us 232 : 23095 : }
233 : :
234 : : /*
235 : : * CleanupProcSignalState
236 : : * Remove current process from ProcSignal mechanism
237 : : *
238 : : * This function is called via on_shmem_exit() during backend shutdown.
239 : : */
240 : : static void
241 : 23095 : CleanupProcSignalState(int status, Datum arg)
242 : : {
243 : : pid_t old_pid;
818 heikki.linnakangas@i 244 : 23095 : ProcSignalSlot *slot = MyProcSignalSlot;
245 : :
246 : : /*
247 : : * Clear MyProcSignalSlot, so that a SIGUSR1 received after this point
248 : : * won't try to access it after it's no longer ours (and perhaps even
249 : : * after we've unmapped the shared memory segment).
250 : : */
251 [ - + ]: 23095 : Assert(MyProcSignalSlot != NULL);
4502 rhaas@postgresql.org 252 : 23095 : MyProcSignalSlot = NULL;
253 : :
254 : : /* sanity check */
670 heikki.linnakangas@i 255 [ - + ]: 23095 : SpinLockAcquire(&slot->pss_mutex);
256 : 23095 : old_pid = pg_atomic_read_u32(&slot->pss_pid);
257 [ - + ]: 23095 : if (old_pid != MyProcPid)
258 : : {
259 : : /*
260 : : * don't ERROR here. We're exiting anyway, and don't want to get into
261 : : * infinite loop trying to exit
262 : : */
670 heikki.linnakangas@i 263 :UBC 0 : SpinLockRelease(&slot->pss_mutex);
6147 tgl@sss.pgh.pa.us 264 [ # # ]: 0 : elog(LOG, "process %d releasing ProcSignal slot %d, but it contains %d",
265 : : MyProcPid, (int) (slot - ProcSignal->psh_slot), (int) old_pid);
266 : 0 : return; /* XXX better to zero the slot anyway? */
267 : : }
268 : :
269 : : /* Mark the slot as unused */
670 heikki.linnakangas@i 270 :CBC 23095 : pg_atomic_write_u32(&slot->pss_pid, 0);
423 271 : 23095 : slot->pss_cancel_key_len = 0;
272 : :
273 : : /*
274 : : * Make this slot look like it's absorbed all possible barriers, so that
275 : : * no barrier waits block on it.
276 : : */
2354 rhaas@postgresql.org 277 : 23095 : pg_atomic_write_u64(&slot->pss_barrierGeneration, PG_UINT64_MAX);
278 : :
670 heikki.linnakangas@i 279 : 23095 : SpinLockRelease(&slot->pss_mutex);
280 : :
281 : 23095 : ConditionVariableBroadcast(&slot->pss_barrierCV);
282 : : }
283 : :
284 : : /*
285 : : * SendProcSignal
286 : : * Send a signal to a Postgres process
287 : : *
288 : : * Providing procNumber is optional, but it will speed up the operation.
289 : : *
290 : : * On success (a signal was sent), zero is returned.
291 : : * On error, -1 is returned, and errno is set (typically to ESRCH or EPERM).
292 : : *
293 : : * Not to be confused with ProcSendSignal
294 : : */
295 : : int
818 296 : 7542 : SendProcSignal(pid_t pid, ProcSignalReason reason, ProcNumber procNumber)
297 : : {
298 : : volatile ProcSignalSlot *slot;
299 : :
300 [ + + ]: 7542 : if (procNumber != INVALID_PROC_NUMBER)
301 : : {
670 302 [ - + ]: 7474 : Assert(procNumber < NumProcSignalSlots);
818 303 : 7474 : slot = &ProcSignal->psh_slot[procNumber];
304 : :
670 305 [ + + ]: 7474 : SpinLockAcquire(&slot->pss_mutex);
306 [ + - ]: 7474 : if (pg_atomic_read_u32(&slot->pss_pid) == pid)
307 : : {
308 : : /* Atomically set the proper flag */
6147 tgl@sss.pgh.pa.us 309 : 7474 : slot->pss_signalFlags[reason] = true;
670 heikki.linnakangas@i 310 : 7474 : SpinLockRelease(&slot->pss_mutex);
311 : : /* Send signal */
6147 tgl@sss.pgh.pa.us 312 : 7474 : return kill(pid, SIGUSR1);
313 : : }
670 heikki.linnakangas@i 314 :UBC 0 : SpinLockRelease(&slot->pss_mutex);
315 : : }
316 : : else
317 : : {
318 : : /*
319 : : * procNumber not provided, so search the array using pid. We search
320 : : * the array back to front so as to reduce search overhead. Passing
321 : : * INVALID_PROC_NUMBER means that the target is most likely an
322 : : * auxiliary process, which will have a slot near the end of the
323 : : * array.
324 : : */
325 : : int i;
326 : :
1509 rhaas@postgresql.org 327 [ + - ]:CBC 3093 : for (i = NumProcSignalSlots - 1; i >= 0; i--)
328 : : {
2354 329 : 3093 : slot = &ProcSignal->psh_slot[i];
330 : :
670 heikki.linnakangas@i 331 [ + + ]: 3093 : if (pg_atomic_read_u32(&slot->pss_pid) == pid)
332 : : {
333 [ - + ]: 68 : SpinLockAcquire(&slot->pss_mutex);
334 [ + - ]: 68 : if (pg_atomic_read_u32(&slot->pss_pid) == pid)
335 : : {
336 : : /* Atomically set the proper flag */
337 : 68 : slot->pss_signalFlags[reason] = true;
338 : 68 : SpinLockRelease(&slot->pss_mutex);
339 : : /* Send signal */
340 : 68 : return kill(pid, SIGUSR1);
341 : : }
670 heikki.linnakangas@i 342 :UBC 0 : SpinLockRelease(&slot->pss_mutex);
343 : : }
344 : : }
345 : : }
346 : :
6147 tgl@sss.pgh.pa.us 347 : 0 : errno = ESRCH;
348 : 0 : return -1;
349 : : }
350 : :
351 : : /*
352 : : * EmitProcSignalBarrier
353 : : * Send a signal to every Postgres process
354 : : *
355 : : * The return value of this function is the barrier "generation" created
356 : : * by this operation. This value can be passed to WaitForProcSignalBarrier
357 : : * to wait until it is known that every participant in the ProcSignal
358 : : * mechanism has absorbed the signal (or started afterwards).
359 : : *
360 : : * Note that it would be a bad idea to use this for anything that happens
361 : : * frequently, as interrupting every backend could cause a noticeable
362 : : * performance hit.
363 : : *
364 : : * Callers are entitled to assume that this function will not throw ERROR
365 : : * or FATAL.
366 : : */
367 : : uint64
2354 rhaas@postgresql.org 368 :CBC 669 : EmitProcSignalBarrier(ProcSignalBarrierType type)
369 : : {
2175 andres@anarazel.de 370 : 669 : uint32 flagbit = 1 << (uint32) type;
371 : : uint64 generation;
372 : :
373 : : /*
374 : : * Set all the flags.
375 : : *
376 : : * Note that pg_atomic_fetch_or_u32 has full barrier semantics, so this is
377 : : * totally ordered with respect to anything the caller did before, and
378 : : * anything that we do afterwards. (This is also true of the later call to
379 : : * pg_atomic_add_fetch_u64.)
380 : : */
1509 rhaas@postgresql.org 381 [ + + ]: 71176 : for (int i = 0; i < NumProcSignalSlots; i++)
382 : : {
2354 383 : 70507 : volatile ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
384 : :
385 : 70507 : pg_atomic_fetch_or_u32(&slot->pss_barrierCheckMask, flagbit);
386 : : }
387 : :
388 : : /*
389 : : * Increment the generation counter.
390 : : */
391 : : generation =
392 : 669 : pg_atomic_add_fetch_u64(&ProcSignal->psh_barrierGeneration, 1);
393 : :
394 : : /*
395 : : * Signal all the processes, so that they update their advertised barrier
396 : : * generation.
397 : : *
398 : : * Concurrency is not a problem here. Backends that have exited don't
399 : : * matter, and new backends that have joined since we entered this
400 : : * function must already have current state, since the caller is
401 : : * responsible for making sure that the relevant state is entirely visible
402 : : * before calling this function in the first place. We still have to wake
403 : : * them up - because we can't distinguish between such backends and older
404 : : * backends that need to update state - but they won't actually need to
405 : : * change any state.
406 : : */
1509 407 [ + + ]: 71176 : for (int i = NumProcSignalSlots - 1; i >= 0; i--)
408 : : {
2354 409 : 70507 : volatile ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
670 heikki.linnakangas@i 410 : 70507 : pid_t pid = pg_atomic_read_u32(&slot->pss_pid);
411 : :
2354 rhaas@postgresql.org 412 [ + + ]: 70507 : if (pid != 0)
413 : : {
670 heikki.linnakangas@i 414 [ - + ]: 3749 : SpinLockAcquire(&slot->pss_mutex);
415 : 3749 : pid = pg_atomic_read_u32(&slot->pss_pid);
416 [ + - ]: 3749 : if (pid != 0)
417 : : {
418 : : /* see SendProcSignal for details */
419 : 3749 : slot->pss_signalFlags[PROCSIG_BARRIER] = true;
420 : 3749 : SpinLockRelease(&slot->pss_mutex);
421 : 3749 : kill(pid, SIGUSR1);
422 : : }
423 : : else
670 heikki.linnakangas@i 424 :UBC 0 : SpinLockRelease(&slot->pss_mutex);
425 : : }
426 : : }
427 : :
2354 rhaas@postgresql.org 428 :CBC 669 : return generation;
429 : : }
430 : :
431 : : /*
432 : : * WaitForProcSignalBarrier - wait until it is guaranteed that all changes
433 : : * requested by a specific call to EmitProcSignalBarrier() have taken effect.
434 : : */
435 : : void
436 : 648 : WaitForProcSignalBarrier(uint64 generation)
437 : : {
2175 andres@anarazel.de 438 [ - + ]: 648 : Assert(generation <= pg_atomic_read_u64(&ProcSignal->psh_barrierGeneration));
439 : :
1480 tmunro@postgresql.or 440 [ + + ]: 648 : elog(DEBUG1,
441 : : "waiting for all backends to process ProcSignalBarrier generation "
442 : : UINT64_FORMAT,
443 : : generation);
444 : :
1509 rhaas@postgresql.org 445 [ + + ]: 69258 : for (int i = NumProcSignalSlots - 1; i >= 0; i--)
446 : : {
1916 tmunro@postgresql.or 447 : 68610 : ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
448 : : uint64 oldval;
449 : :
450 : : /*
451 : : * It's important that we check only pss_barrierGeneration here and
452 : : * not pss_barrierCheckMask. Bits in pss_barrierCheckMask get cleared
453 : : * before the barrier is actually absorbed, but pss_barrierGeneration
454 : : * is updated only afterward.
455 : : */
2354 rhaas@postgresql.org 456 : 68610 : oldval = pg_atomic_read_u64(&slot->pss_barrierGeneration);
457 [ + + ]: 71379 : while (oldval < generation)
458 : : {
1480 tmunro@postgresql.or 459 [ - + ]: 2769 : if (ConditionVariableTimedSleep(&slot->pss_barrierCV,
460 : : 5000,
461 : : WAIT_EVENT_PROC_SIGNAL_BARRIER))
1480 tmunro@postgresql.or 462 [ # # ]:UBC 0 : ereport(LOG,
463 : : (errmsg("still waiting for backend with PID %d to accept ProcSignalBarrier",
464 : : (int) pg_atomic_read_u32(&slot->pss_pid))));
2354 rhaas@postgresql.org 465 :CBC 2769 : oldval = pg_atomic_read_u64(&slot->pss_barrierGeneration);
466 : : }
1916 tmunro@postgresql.or 467 : 68610 : ConditionVariableCancelSleep();
468 : : }
469 : :
1480 470 [ + + ]: 648 : elog(DEBUG1,
471 : : "finished waiting for all backends to process ProcSignalBarrier generation "
472 : : UINT64_FORMAT,
473 : : generation);
474 : :
475 : : /*
476 : : * The caller is probably calling this function because it wants to read
477 : : * the shared state or perform further writes to shared state once all
478 : : * backends are known to have absorbed the barrier. However, the read of
479 : : * pss_barrierGeneration was performed unlocked; insert a memory barrier
480 : : * to separate it from whatever follows.
481 : : */
2354 rhaas@postgresql.org 482 : 648 : pg_memory_barrier();
483 : 648 : }
484 : :
485 : : /*
486 : : * Handle receipt of an interrupt indicating a global barrier event.
487 : : *
488 : : * All the actual work is deferred to ProcessProcSignalBarrier(), because we
489 : : * cannot safely access the barrier generation inside the signal handler as
490 : : * 64bit atomics might use spinlock based emulation, even for reads. As this
491 : : * routine only gets called when PROCSIG_BARRIER is sent that won't cause a
492 : : * lot of unnecessary work.
493 : : */
494 : : static void
2175 andres@anarazel.de 495 : 2844 : HandleProcSignalBarrierInterrupt(void)
496 : : {
497 : 2844 : InterruptPending = true;
498 : 2844 : ProcSignalBarrierPending = true;
499 : : /* latch will be set by procsignal_sigusr1_handler */
500 : 2844 : }
501 : :
502 : : /*
503 : : * Perform global barrier related interrupt checking.
504 : : *
505 : : * Any backend that participates in ProcSignal signaling must arrange to
506 : : * call this function periodically. It is called from CHECK_FOR_INTERRUPTS(),
507 : : * which is enough for normal backends, but not necessarily for all types of
508 : : * background processes.
509 : : */
510 : : void
2354 rhaas@postgresql.org 511 : 2840 : ProcessProcSignalBarrier(void)
512 : : {
513 : : uint64 local_gen;
514 : : uint64 shared_gen;
515 : : volatile uint32 flags;
516 : :
2175 andres@anarazel.de 517 [ - + ]: 2840 : Assert(MyProcSignalSlot);
518 : :
519 : : /* Exit quickly if there's no work to do. */
2354 rhaas@postgresql.org 520 [ - + ]: 2840 : if (!ProcSignalBarrierPending)
2354 rhaas@postgresql.org 521 :UBC 0 : return;
2354 rhaas@postgresql.org 522 :CBC 2840 : ProcSignalBarrierPending = false;
523 : :
524 : : /*
525 : : * It's not unlikely to process multiple barriers at once, before the
526 : : * signals for all the barriers have arrived. To avoid unnecessary work in
527 : : * response to subsequent signals, exit early if we already have processed
528 : : * all of them.
529 : : */
2175 andres@anarazel.de 530 : 2840 : local_gen = pg_atomic_read_u64(&MyProcSignalSlot->pss_barrierGeneration);
531 : 2840 : shared_gen = pg_atomic_read_u64(&ProcSignal->psh_barrierGeneration);
532 : :
533 [ - + ]: 2840 : Assert(local_gen <= shared_gen);
534 : :
535 [ - + ]: 2840 : if (local_gen == shared_gen)
2175 andres@anarazel.de 536 :UBC 0 : return;
537 : :
538 : : /*
539 : : * Get and clear the flags that are set for this backend. Note that
540 : : * pg_atomic_exchange_u32 is a full barrier, so we're guaranteed that the
541 : : * read of the barrier generation above happens before we atomically
542 : : * extract the flags, and that any subsequent state changes happen
543 : : * afterward.
544 : : *
545 : : * NB: In order to avoid race conditions, we must zero
546 : : * pss_barrierCheckMask first and only afterwards try to do barrier
547 : : * processing. If we did it in the other order, someone could send us
548 : : * another barrier of some type right after we called the
549 : : * barrier-processing function but before we cleared the bit. We would
550 : : * have no way of knowing that the bit needs to stay set in that case, so
551 : : * the need to call the barrier-processing function again would just get
552 : : * forgotten. So instead, we tentatively clear all the bits and then put
553 : : * back any for which we don't manage to successfully absorb the barrier.
554 : : */
2354 rhaas@postgresql.org 555 :CBC 2840 : flags = pg_atomic_exchange_u32(&MyProcSignalSlot->pss_barrierCheckMask, 0);
556 : :
557 : : /*
558 : : * If there are no flags set, then we can skip doing any real work.
559 : : * Otherwise, establish a PG_TRY block, so that we don't lose track of
560 : : * which types of barrier processing are needed if an ERROR occurs.
561 : : */
1958 562 [ + + ]: 2840 : if (flags != 0)
563 : : {
1844 tgl@sss.pgh.pa.us 564 : 2839 : bool success = true;
565 : :
1958 rhaas@postgresql.org 566 [ + - ]: 2839 : PG_TRY();
567 : : {
568 : : /*
569 : : * Process each type of barrier. The barrier-processing functions
570 : : * should normally return true, but may return false if the
571 : : * barrier can't be absorbed at the current time. This should be
572 : : * rare, because it's pretty expensive. Every single
573 : : * CHECK_FOR_INTERRUPTS() will return here until we manage to
574 : : * absorb the barrier, and that cost will add up in a hurry.
575 : : *
576 : : * NB: It ought to be OK to call the barrier-processing functions
577 : : * unconditionally, but it's more efficient to call only the ones
578 : : * that might need us to do something based on the flags.
579 : : */
580 [ + + ]: 8517 : while (flags != 0)
581 : : {
582 : : ProcSignalBarrierType type;
1844 tgl@sss.pgh.pa.us 583 : 2839 : bool processed = true;
584 : :
1958 rhaas@postgresql.org 585 : 2839 : type = (ProcSignalBarrierType) pg_rightmost_one_pos32(flags);
586 [ + - ]: 2839 : switch (type)
[ + + + - ]
587 : : {
1568 tmunro@postgresql.or 588 : 643 : case PROCSIGNAL_BARRIER_SMGRRELEASE:
589 : 643 : processed = ProcessBarrierSmgrRelease();
1958 rhaas@postgresql.org 590 : 643 : break;
158 msawada@postgresql.o 591 :GNC 1920 : case PROCSIGNAL_BARRIER_UPDATE_XLOG_LOGICAL_INFO:
592 : 1920 : processed = ProcessBarrierUpdateXLogLogicalInfo();
593 : 1920 : break;
594 : :
57 dgustafsson@postgres 595 : 276 : case PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_ON:
596 : : case PROCSIGNAL_BARRIER_CHECKSUM_ON:
597 : : case PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_OFF:
598 : : case PROCSIGNAL_BARRIER_CHECKSUM_OFF:
599 : 276 : processed = AbsorbDataChecksumsBarrier(type);
600 : 276 : break;
601 : : }
602 : :
603 : : /*
604 : : * To avoid an infinite loop, we must always unset the bit in
605 : : * flags.
606 : : */
1958 rhaas@postgresql.org 607 :CBC 2839 : BARRIER_CLEAR_BIT(flags, type);
608 : :
609 : : /*
610 : : * If we failed to process the barrier, reset the shared bit
611 : : * so we try again later, and set a flag so that we don't bump
612 : : * our generation.
613 : : */
614 [ + - ]: 2839 : if (!processed)
615 : : {
1958 rhaas@postgresql.org 616 :UBC 0 : ResetProcSignalBarrierBits(((uint32) 1) << type);
617 : 0 : success = false;
618 : : }
619 : : }
620 : : }
621 : 0 : PG_CATCH();
622 : : {
623 : : /*
624 : : * If an ERROR occurred, we'll need to try again later to handle
625 : : * that barrier type and any others that haven't been handled yet
626 : : * or weren't successfully absorbed.
627 : : */
628 : 0 : ResetProcSignalBarrierBits(flags);
629 : 0 : PG_RE_THROW();
630 : : }
1958 rhaas@postgresql.org 631 [ - + ]:CBC 2839 : PG_END_TRY();
632 : :
633 : : /*
634 : : * If some barrier types were not successfully absorbed, we will have
635 : : * to try again later.
636 : : */
637 [ - + ]: 2839 : if (!success)
1958 rhaas@postgresql.org 638 :UBC 0 : return;
639 : : }
640 : :
641 : : /*
642 : : * State changes related to all types of barriers that might have been
643 : : * emitted have now been handled, so we can update our notion of the
644 : : * generation to the one we observed before beginning the updates. If
645 : : * things have changed further, it'll get fixed up when this function is
646 : : * next called.
647 : : */
2175 andres@anarazel.de 648 :CBC 2840 : pg_atomic_write_u64(&MyProcSignalSlot->pss_barrierGeneration, shared_gen);
1916 tmunro@postgresql.or 649 : 2840 : ConditionVariableBroadcast(&MyProcSignalSlot->pss_barrierCV);
650 : : }
651 : :
652 : : /*
653 : : * If it turns out that we couldn't absorb one or more barrier types, either
654 : : * because the barrier-processing functions returned false or due to an error,
655 : : * arrange for processing to be retried later.
656 : : */
657 : : static void
1958 rhaas@postgresql.org 658 :UBC 0 : ResetProcSignalBarrierBits(uint32 flags)
659 : : {
660 : 0 : pg_atomic_fetch_or_u32(&MyProcSignalSlot->pss_barrierCheckMask, flags);
661 : 0 : ProcSignalBarrierPending = true;
662 : 0 : InterruptPending = true;
663 : 0 : }
664 : :
665 : : /*
666 : : * CheckProcSignal - check to see if a particular reason has been
667 : : * signaled, and clear the signal flag. Should be called after receiving
668 : : * SIGUSR1.
669 : : */
670 : : static bool
6147 tgl@sss.pgh.pa.us 671 :CBC 129680 : CheckProcSignal(ProcSignalReason reason)
672 : : {
673 : 129680 : volatile ProcSignalSlot *slot = MyProcSignalSlot;
674 : :
675 [ + + ]: 129680 : if (slot != NULL)
676 : : {
677 : : /*
678 : : * Careful here --- don't clear flag if we haven't seen it set.
679 : : * pss_signalFlags is of type "volatile sig_atomic_t" to allow us to
680 : : * read it here safely, without holding the spinlock.
681 : : */
682 [ + + ]: 129420 : if (slot->pss_signalFlags[reason])
683 : : {
684 : 7843 : slot->pss_signalFlags[reason] = false;
685 : 7843 : return true;
686 : : }
687 : : }
688 : :
689 : 121837 : return false;
690 : : }
691 : :
692 : : /*
693 : : * procsignal_sigusr1_handler - handle SIGUSR1 signal.
694 : : */
695 : : void
696 : 12968 : procsignal_sigusr1_handler(SIGNAL_ARGS)
697 : : {
698 [ + + ]: 12968 : if (CheckProcSignal(PROCSIG_CATCHUP_INTERRUPT))
699 : 3047 : HandleCatchupInterrupt();
700 : :
701 [ + + ]: 12968 : if (CheckProcSignal(PROCSIG_NOTIFY_INTERRUPT))
702 : 6 : HandleNotifyInterrupt();
703 : :
4048 rhaas@postgresql.org 704 [ + + ]: 12968 : if (CheckProcSignal(PROCSIG_PARALLEL_MESSAGE))
705 : 1846 : HandleParallelMessageInterrupt();
706 : :
3281 andres@anarazel.de 707 [ + + ]: 12968 : if (CheckProcSignal(PROCSIG_WALSND_INIT_STOPPING))
708 : 48 : HandleWalSndInitStopping();
709 : :
2175 710 [ + + ]: 12968 : if (CheckProcSignal(PROCSIG_BARRIER))
711 : 2844 : HandleProcSignalBarrierInterrupt();
712 : :
1880 fujii@postgresql.org 713 [ + + ]: 12968 : if (CheckProcSignal(PROCSIG_LOG_MEMORY_CONTEXT))
714 : 12 : HandleLogMemoryContextInterrupt();
715 : :
1237 akapila@postgresql.o 716 [ + + ]: 12968 : if (CheckProcSignal(PROCSIG_PARALLEL_APPLY_MESSAGE))
717 : 14 : HandleParallelApplyMessageInterrupt();
718 : :
54 alvherre@kurilemu.de 719 [ + + ]:GNC 12968 : if (CheckProcSignal(PROCSIG_REPACK_MESSAGE))
720 : 7 : HandleRepackMessageInterrupt();
721 : :
52 fujii@postgresql.org 722 [ + + ]:CBC 12968 : if (CheckProcSignal(PROCSIG_SLOTSYNC_MESSAGE))
723 : 1 : HandleSlotSyncMessageInterrupt();
724 : :
109 heikki.linnakangas@i 725 [ + + ]:GNC 12968 : if (CheckProcSignal(PROCSIG_RECOVERY_CONFLICT))
726 : 18 : HandleRecoveryConflictInterrupt();
727 : :
3886 rhaas@postgresql.org 728 :CBC 12968 : SetLatch(MyLatch);
6147 tgl@sss.pgh.pa.us 729 : 12968 : }
730 : :
731 : : /*
732 : : * Send a query cancellation signal to backend.
733 : : *
734 : : * Note: This is called from a backend process before authentication. We
735 : : * cannot take LWLocks yet, but that's OK; we rely on atomic reads of the
736 : : * fields in the ProcSignal slots.
737 : : */
738 : : void
387 heikki.linnakangas@i 739 : 15 : SendCancelRequest(int backendPID, const uint8 *cancel_key, int cancel_key_len)
740 : : {
304 741 [ - + ]: 15 : if (backendPID == 0)
742 : : {
304 heikki.linnakangas@i 743 [ # # ]:UBC 0 : ereport(LOG, (errmsg("invalid cancel request with PID 0")));
744 : 0 : return;
745 : : }
746 : :
747 : : /*
748 : : * See if we have a matching backend. Reading the pss_pid and
749 : : * pss_cancel_key fields is racy, a backend might die and remove itself
750 : : * from the array at any time. The probability of the cancellation key
751 : : * matching wrong process is miniscule, however, so we can live with that.
752 : : * PIDs are reused too, so sending the signal based on PID is inherently
753 : : * racy anyway, although OS's avoid reusing PIDs too soon.
754 : : */
670 heikki.linnakangas@i 755 [ + - ]:CBC 217 : for (int i = 0; i < NumProcSignalSlots; i++)
756 : : {
757 : 217 : ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
758 : : bool match;
759 : :
760 [ + + ]: 217 : if (pg_atomic_read_u32(&slot->pss_pid) != backendPID)
761 : 202 : continue;
762 : :
763 : : /* Acquire the spinlock and re-check */
764 [ - + ]: 15 : SpinLockAcquire(&slot->pss_mutex);
765 [ - + ]: 15 : if (pg_atomic_read_u32(&slot->pss_pid) != backendPID)
766 : : {
670 heikki.linnakangas@i 767 :UBC 0 : SpinLockRelease(&slot->pss_mutex);
768 : 0 : continue;
769 : : }
770 : : else
771 : : {
423 heikki.linnakangas@i 772 [ + - + - ]:CBC 30 : match = slot->pss_cancel_key_len == cancel_key_len &&
773 : 15 : timingsafe_bcmp(slot->pss_cancel_key, cancel_key, cancel_key_len) == 0;
774 : :
670 775 : 15 : SpinLockRelease(&slot->pss_mutex);
776 : :
777 [ + - ]: 15 : if (match)
778 : : {
779 : : /* Found a match; signal that backend to cancel current op */
780 [ - + ]: 15 : ereport(DEBUG2,
781 : : (errmsg_internal("processing cancel request: sending SIGINT to process %d",
782 : : backendPID)));
783 : :
784 : : /*
785 : : * If we have setsid(), signal the backend's whole process
786 : : * group
787 : : */
788 : : #ifdef HAVE_SETSID
789 : 15 : kill(-backendPID, SIGINT);
790 : : #else
791 : : kill(backendPID, SIGINT);
792 : : #endif
793 : : }
794 : : else
795 : : {
796 : : /* Right PID, wrong key: no way, Jose */
670 heikki.linnakangas@i 797 [ # # ]:UBC 0 : ereport(LOG,
798 : : (errmsg("wrong key in cancel request for process %d",
799 : : backendPID)));
800 : : }
670 heikki.linnakangas@i 801 :CBC 15 : return;
802 : : }
803 : : }
804 : :
805 : : /* No matching backend */
670 heikki.linnakangas@i 806 [ # # ]:UBC 0 : ereport(LOG,
807 : : (errmsg("PID %d in cancel request did not match any process",
808 : : backendPID)));
809 : : }
|
