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) 1996-2025, 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 "replication/logicalworker.h"
26 : : #include "replication/walsender.h"
27 : : #include "storage/condition_variable.h"
28 : : #include "storage/ipc.h"
29 : : #include "storage/latch.h"
30 : : #include "storage/shmem.h"
31 : : #include "storage/sinval.h"
32 : : #include "storage/smgr.h"
33 : : #include "tcop/tcopprot.h"
34 : : #include "utils/memutils.h"
35 : :
36 : : /*
37 : : * The SIGUSR1 signal is multiplexed to support signaling multiple event
38 : : * types. The specific reason is communicated via flags in shared memory.
39 : : * We keep a boolean flag for each possible "reason", so that different
40 : : * reasons can be signaled to a process concurrently. (However, if the same
41 : : * reason is signaled more than once nearly simultaneously, the process may
42 : : * observe it only once.)
43 : : *
44 : : * Each process that wants to receive signals registers its process ID
45 : : * in the ProcSignalSlots array. The array is indexed by ProcNumber to make
46 : : * slot allocation simple, and to avoid having to search the array when you
47 : : * know the ProcNumber of the process you're signaling. (We do support
48 : : * signaling without ProcNumber, but it's a bit less efficient.)
49 : : *
50 : : * The fields in each slot are protected by a spinlock, pss_mutex. pss_pid can
51 : : * also be read without holding the spinlock, as a quick preliminary check
52 : : * when searching for a particular PID in the array.
53 : : *
54 : : * pss_signalFlags are intended to be set in cases where we don't need to
55 : : * keep track of whether or not the target process has handled the signal,
56 : : * but sometimes we need confirmation, as when making a global state change
57 : : * that cannot be considered complete until all backends have taken notice
58 : : * of it. For such use cases, we set a bit in pss_barrierCheckMask and then
59 : : * increment the current "barrier generation"; when the new barrier generation
60 : : * (or greater) appears in the pss_barrierGeneration flag of every process,
61 : : * we know that the message has been received everywhere.
62 : : */
63 : : typedef struct
64 : : {
65 : : pg_atomic_uint32 pss_pid;
66 : : int pss_cancel_key_len; /* 0 means no cancellation is possible */
67 : : uint8 pss_cancel_key[MAX_CANCEL_KEY_LENGTH];
68 : : volatile sig_atomic_t pss_signalFlags[NUM_PROCSIGNALS];
69 : : slock_t pss_mutex; /* protects the above fields */
70 : :
71 : : /* Barrier-related fields (not protected by pss_mutex) */
72 : : pg_atomic_uint64 pss_barrierGeneration;
73 : : pg_atomic_uint32 pss_barrierCheckMask;
74 : : ConditionVariable pss_barrierCV;
75 : : } ProcSignalSlot;
76 : :
77 : : /*
78 : : * Information that is global to the entire ProcSignal system can be stored
79 : : * here.
80 : : *
81 : : * psh_barrierGeneration is the highest barrier generation in existence.
82 : : */
83 : : struct ProcSignalHeader
84 : : {
85 : : pg_atomic_uint64 psh_barrierGeneration;
86 : : ProcSignalSlot psh_slot[FLEXIBLE_ARRAY_MEMBER];
87 : : };
88 : :
89 : : /*
90 : : * We reserve a slot for each possible ProcNumber, plus one for each
91 : : * possible auxiliary process type. (This scheme assumes there is not
92 : : * more than one of any auxiliary process type at a time, except for
93 : : * IO workers.)
94 : : */
95 : : #define NumProcSignalSlots (MaxBackends + NUM_AUXILIARY_PROCS)
96 : :
97 : : /* Check whether the relevant type bit is set in the flags. */
98 : : #define BARRIER_SHOULD_CHECK(flags, type) \
99 : : (((flags) & (((uint32) 1) << (uint32) (type))) != 0)
100 : :
101 : : /* Clear the relevant type bit from the flags. */
102 : : #define BARRIER_CLEAR_BIT(flags, type) \
103 : : ((flags) &= ~(((uint32) 1) << (uint32) (type)))
104 : :
105 : : NON_EXEC_STATIC ProcSignalHeader *ProcSignal = NULL;
106 : : static ProcSignalSlot *MyProcSignalSlot = NULL;
107 : :
108 : : static bool CheckProcSignal(ProcSignalReason reason);
109 : : static void CleanupProcSignalState(int status, Datum arg);
110 : : static void ResetProcSignalBarrierBits(uint32 flags);
111 : :
112 : : /*
113 : : * ProcSignalShmemSize
114 : : * Compute space needed for ProcSignal's shared memory
115 : : */
116 : : Size
5881 tgl@sss.pgh.pa.us 117 :CBC 2938 : ProcSignalShmemSize(void)
118 : : {
119 : : Size size;
120 : :
1243 rhaas@postgresql.org 121 : 2938 : size = mul_size(NumProcSignalSlots, sizeof(ProcSignalSlot));
2088 122 : 2938 : size = add_size(size, offsetof(ProcSignalHeader, psh_slot));
123 : 2938 : return size;
124 : : }
125 : :
126 : : /*
127 : : * ProcSignalShmemInit
128 : : * Allocate and initialize ProcSignal's shared memory
129 : : */
130 : : void
5881 tgl@sss.pgh.pa.us 131 : 1029 : ProcSignalShmemInit(void)
132 : : {
133 : 1029 : Size size = ProcSignalShmemSize();
134 : : bool found;
135 : :
2088 rhaas@postgresql.org 136 : 1029 : ProcSignal = (ProcSignalHeader *)
137 : 1029 : ShmemInitStruct("ProcSignal", size, &found);
138 : :
139 : : /* If we're first, initialize. */
5881 tgl@sss.pgh.pa.us 140 [ + - ]: 1029 : if (!found)
141 : : {
142 : : int i;
143 : :
2088 rhaas@postgresql.org 144 : 1029 : pg_atomic_init_u64(&ProcSignal->psh_barrierGeneration, 0);
145 : :
1243 146 [ + + ]: 134389 : for (i = 0; i < NumProcSignalSlots; ++i)
147 : : {
2088 148 : 133360 : ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
149 : :
404 heikki.linnakangas@i 150 : 133360 : SpinLockInit(&slot->pss_mutex);
151 : 133360 : pg_atomic_init_u32(&slot->pss_pid, 0);
157 152 : 133360 : slot->pss_cancel_key_len = 0;
2088 rhaas@postgresql.org 153 [ + - + - : 1066880 : MemSet(slot->pss_signalFlags, 0, sizeof(slot->pss_signalFlags));
+ - + - +
+ ]
154 : 133360 : pg_atomic_init_u64(&slot->pss_barrierGeneration, PG_UINT64_MAX);
155 : 133360 : pg_atomic_init_u32(&slot->pss_barrierCheckMask, 0);
1650 tmunro@postgresql.or 156 : 133360 : ConditionVariableInit(&slot->pss_barrierCV);
157 : : }
158 : : }
5881 tgl@sss.pgh.pa.us 159 : 1029 : }
160 : :
161 : : /*
162 : : * ProcSignalInit
163 : : * Register the current process in the ProcSignal array
164 : : */
165 : : void
121 heikki.linnakangas@i 166 : 18757 : ProcSignalInit(const uint8 *cancel_key, int cancel_key_len)
167 : : {
168 : : ProcSignalSlot *slot;
169 : : uint64 barrier_generation;
170 : : uint32 old_pss_pid;
171 : :
157 172 [ + - - + ]: 18757 : Assert(cancel_key_len >= 0 && cancel_key_len <= MAX_CANCEL_KEY_LENGTH);
552 173 [ - + ]: 18757 : if (MyProcNumber < 0)
552 heikki.linnakangas@i 174 [ # # ]:UBC 0 : elog(ERROR, "MyProcNumber not set");
552 heikki.linnakangas@i 175 [ - + ]:CBC 18757 : if (MyProcNumber >= NumProcSignalSlots)
552 heikki.linnakangas@i 176 [ # # ]:UBC 0 : elog(ERROR, "unexpected MyProcNumber %d in ProcSignalInit (max %d)", MyProcNumber, NumProcSignalSlots);
552 heikki.linnakangas@i 177 :CBC 18757 : slot = &ProcSignal->psh_slot[MyProcNumber];
178 : :
404 179 [ - + ]: 18757 : SpinLockAcquire(&slot->pss_mutex);
180 : :
181 : : /* Value used for sanity check below */
191 michael@paquier.xyz 182 : 18757 : old_pss_pid = pg_atomic_read_u32(&slot->pss_pid);
183 : :
184 : : /* Clear out any leftover signal reasons */
5881 tgl@sss.pgh.pa.us 185 [ + - + - : 150056 : MemSet(slot->pss_signalFlags, 0, NUM_PROCSIGNALS * sizeof(sig_atomic_t));
+ - + - +
+ ]
186 : :
187 : : /*
188 : : * Initialize barrier state. Since we're a brand-new process, there
189 : : * shouldn't be any leftover backend-private state that needs to be
190 : : * updated. Therefore, we can broadcast the latest barrier generation and
191 : : * disregard any previously-set check bits.
192 : : *
193 : : * NB: This only works if this initialization happens early enough in the
194 : : * startup sequence that we haven't yet cached any state that might need
195 : : * to be invalidated. That's also why we have a memory barrier here, to be
196 : : * sure that any later reads of memory happen strictly after this.
197 : : */
2088 rhaas@postgresql.org 198 : 18757 : pg_atomic_write_u32(&slot->pss_barrierCheckMask, 0);
199 : : barrier_generation =
200 : 18757 : pg_atomic_read_u64(&ProcSignal->psh_barrierGeneration);
201 : 18757 : pg_atomic_write_u64(&slot->pss_barrierGeneration, barrier_generation);
202 : :
157 heikki.linnakangas@i 203 [ + + ]: 18757 : if (cancel_key_len > 0)
204 : 12029 : memcpy(slot->pss_cancel_key, cancel_key, cancel_key_len);
205 : 18757 : slot->pss_cancel_key_len = cancel_key_len;
404 206 : 18757 : pg_atomic_write_u32(&slot->pss_pid, MyProcPid);
207 : :
208 : 18757 : SpinLockRelease(&slot->pss_mutex);
209 : :
210 : : /* Spinlock is released, do the check */
191 michael@paquier.xyz 211 [ - + ]: 18757 : if (old_pss_pid != 0)
191 michael@paquier.xyz 212 [ # # ]:UBC 0 : elog(LOG, "process %d taking over ProcSignal slot %d, but it's not empty",
213 : : MyProcPid, MyProcNumber);
214 : :
215 : : /* Remember slot location for CheckProcSignal */
5881 tgl@sss.pgh.pa.us 216 :CBC 18757 : MyProcSignalSlot = slot;
217 : :
218 : : /* Set up to release the slot on process exit */
552 heikki.linnakangas@i 219 : 18757 : on_shmem_exit(CleanupProcSignalState, (Datum) 0);
5881 tgl@sss.pgh.pa.us 220 : 18757 : }
221 : :
222 : : /*
223 : : * CleanupProcSignalState
224 : : * Remove current process from ProcSignal mechanism
225 : : *
226 : : * This function is called via on_shmem_exit() during backend shutdown.
227 : : */
228 : : static void
229 : 18757 : CleanupProcSignalState(int status, Datum arg)
230 : : {
231 : : pid_t old_pid;
552 heikki.linnakangas@i 232 : 18757 : ProcSignalSlot *slot = MyProcSignalSlot;
233 : :
234 : : /*
235 : : * Clear MyProcSignalSlot, so that a SIGUSR1 received after this point
236 : : * won't try to access it after it's no longer ours (and perhaps even
237 : : * after we've unmapped the shared memory segment).
238 : : */
239 [ - + ]: 18757 : Assert(MyProcSignalSlot != NULL);
4236 rhaas@postgresql.org 240 : 18757 : MyProcSignalSlot = NULL;
241 : :
242 : : /* sanity check */
404 heikki.linnakangas@i 243 [ - + ]: 18757 : SpinLockAcquire(&slot->pss_mutex);
244 : 18757 : old_pid = pg_atomic_read_u32(&slot->pss_pid);
245 [ - + ]: 18757 : if (old_pid != MyProcPid)
246 : : {
247 : : /*
248 : : * don't ERROR here. We're exiting anyway, and don't want to get into
249 : : * infinite loop trying to exit
250 : : */
404 heikki.linnakangas@i 251 :UBC 0 : SpinLockRelease(&slot->pss_mutex);
5881 tgl@sss.pgh.pa.us 252 [ # # ]: 0 : elog(LOG, "process %d releasing ProcSignal slot %d, but it contains %d",
253 : : MyProcPid, (int) (slot - ProcSignal->psh_slot), (int) old_pid);
254 : 0 : return; /* XXX better to zero the slot anyway? */
255 : : }
256 : :
257 : : /* Mark the slot as unused */
404 heikki.linnakangas@i 258 :CBC 18757 : pg_atomic_write_u32(&slot->pss_pid, 0);
157 259 : 18757 : slot->pss_cancel_key_len = 0;
260 : :
261 : : /*
262 : : * Make this slot look like it's absorbed all possible barriers, so that
263 : : * no barrier waits block on it.
264 : : */
2088 rhaas@postgresql.org 265 : 18757 : pg_atomic_write_u64(&slot->pss_barrierGeneration, PG_UINT64_MAX);
266 : :
404 heikki.linnakangas@i 267 : 18757 : SpinLockRelease(&slot->pss_mutex);
268 : :
269 : 18757 : ConditionVariableBroadcast(&slot->pss_barrierCV);
270 : : }
271 : :
272 : : /*
273 : : * SendProcSignal
274 : : * Send a signal to a Postgres process
275 : : *
276 : : * Providing procNumber is optional, but it will speed up the operation.
277 : : *
278 : : * On success (a signal was sent), zero is returned.
279 : : * On error, -1 is returned, and errno is set (typically to ESRCH or EPERM).
280 : : *
281 : : * Not to be confused with ProcSendSignal
282 : : */
283 : : int
552 284 : 5427 : SendProcSignal(pid_t pid, ProcSignalReason reason, ProcNumber procNumber)
285 : : {
286 : : volatile ProcSignalSlot *slot;
287 : :
288 [ + + ]: 5427 : if (procNumber != INVALID_PROC_NUMBER)
289 : : {
404 290 [ - + ]: 5380 : Assert(procNumber < NumProcSignalSlots);
552 291 : 5380 : slot = &ProcSignal->psh_slot[procNumber];
292 : :
404 293 [ + + ]: 5380 : SpinLockAcquire(&slot->pss_mutex);
294 [ + - ]: 5380 : if (pg_atomic_read_u32(&slot->pss_pid) == pid)
295 : : {
296 : : /* Atomically set the proper flag */
5881 tgl@sss.pgh.pa.us 297 : 5380 : slot->pss_signalFlags[reason] = true;
404 heikki.linnakangas@i 298 : 5380 : SpinLockRelease(&slot->pss_mutex);
299 : : /* Send signal */
5881 tgl@sss.pgh.pa.us 300 : 5380 : return kill(pid, SIGUSR1);
301 : : }
404 heikki.linnakangas@i 302 :UBC 0 : SpinLockRelease(&slot->pss_mutex);
303 : : }
304 : : else
305 : : {
306 : : /*
307 : : * procNumber not provided, so search the array using pid. We search
308 : : * the array back to front so as to reduce search overhead. Passing
309 : : * INVALID_PROC_NUMBER means that the target is most likely an
310 : : * auxiliary process, which will have a slot near the end of the
311 : : * array.
312 : : */
313 : : int i;
314 : :
1243 rhaas@postgresql.org 315 [ + - ]:CBC 2119 : for (i = NumProcSignalSlots - 1; i >= 0; i--)
316 : : {
2088 317 : 2119 : slot = &ProcSignal->psh_slot[i];
318 : :
404 heikki.linnakangas@i 319 [ + + ]: 2119 : if (pg_atomic_read_u32(&slot->pss_pid) == pid)
320 : : {
321 [ - + ]: 47 : SpinLockAcquire(&slot->pss_mutex);
322 [ + - ]: 47 : if (pg_atomic_read_u32(&slot->pss_pid) == pid)
323 : : {
324 : : /* Atomically set the proper flag */
325 : 47 : slot->pss_signalFlags[reason] = true;
326 : 47 : SpinLockRelease(&slot->pss_mutex);
327 : : /* Send signal */
328 : 47 : return kill(pid, SIGUSR1);
329 : : }
404 heikki.linnakangas@i 330 :UBC 0 : SpinLockRelease(&slot->pss_mutex);
331 : : }
332 : : }
333 : : }
334 : :
5881 tgl@sss.pgh.pa.us 335 : 0 : errno = ESRCH;
336 : 0 : return -1;
337 : : }
338 : :
339 : : /*
340 : : * EmitProcSignalBarrier
341 : : * Send a signal to every Postgres process
342 : : *
343 : : * The return value of this function is the barrier "generation" created
344 : : * by this operation. This value can be passed to WaitForProcSignalBarrier
345 : : * to wait until it is known that every participant in the ProcSignal
346 : : * mechanism has absorbed the signal (or started afterwards).
347 : : *
348 : : * Note that it would be a bad idea to use this for anything that happens
349 : : * frequently, as interrupting every backend could cause a noticeable
350 : : * performance hit.
351 : : *
352 : : * Callers are entitled to assume that this function will not throw ERROR
353 : : * or FATAL.
354 : : */
355 : : uint64
2088 rhaas@postgresql.org 356 :CBC 88 : EmitProcSignalBarrier(ProcSignalBarrierType type)
357 : : {
1909 andres@anarazel.de 358 : 88 : uint32 flagbit = 1 << (uint32) type;
359 : : uint64 generation;
360 : :
361 : : /*
362 : : * Set all the flags.
363 : : *
364 : : * Note that pg_atomic_fetch_or_u32 has full barrier semantics, so this is
365 : : * totally ordered with respect to anything the caller did before, and
366 : : * anything that we do afterwards. (This is also true of the later call to
367 : : * pg_atomic_add_fetch_u64.)
368 : : */
1243 rhaas@postgresql.org 369 [ + + ]: 9976 : for (int i = 0; i < NumProcSignalSlots; i++)
370 : : {
2088 371 : 9888 : volatile ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
372 : :
373 : 9888 : pg_atomic_fetch_or_u32(&slot->pss_barrierCheckMask, flagbit);
374 : : }
375 : :
376 : : /*
377 : : * Increment the generation counter.
378 : : */
379 : : generation =
380 : 88 : pg_atomic_add_fetch_u64(&ProcSignal->psh_barrierGeneration, 1);
381 : :
382 : : /*
383 : : * Signal all the processes, so that they update their advertised barrier
384 : : * generation.
385 : : *
386 : : * Concurrency is not a problem here. Backends that have exited don't
387 : : * matter, and new backends that have joined since we entered this
388 : : * function must already have current state, since the caller is
389 : : * responsible for making sure that the relevant state is entirely visible
390 : : * before calling this function in the first place. We still have to wake
391 : : * them up - because we can't distinguish between such backends and older
392 : : * backends that need to update state - but they won't actually need to
393 : : * change any state.
394 : : */
1243 395 [ + + ]: 9976 : for (int i = NumProcSignalSlots - 1; i >= 0; i--)
396 : : {
2088 397 : 9888 : volatile ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
404 heikki.linnakangas@i 398 : 9888 : pid_t pid = pg_atomic_read_u32(&slot->pss_pid);
399 : :
2088 rhaas@postgresql.org 400 [ + + ]: 9888 : if (pid != 0)
401 : : {
404 heikki.linnakangas@i 402 [ - + ]: 778 : SpinLockAcquire(&slot->pss_mutex);
403 : 778 : pid = pg_atomic_read_u32(&slot->pss_pid);
404 [ + - ]: 778 : if (pid != 0)
405 : : {
406 : : /* see SendProcSignal for details */
407 : 778 : slot->pss_signalFlags[PROCSIG_BARRIER] = true;
408 : 778 : SpinLockRelease(&slot->pss_mutex);
409 : 778 : kill(pid, SIGUSR1);
410 : : }
411 : : else
404 heikki.linnakangas@i 412 :UBC 0 : SpinLockRelease(&slot->pss_mutex);
413 : : }
414 : : }
415 : :
2088 rhaas@postgresql.org 416 :CBC 88 : return generation;
417 : : }
418 : :
419 : : /*
420 : : * WaitForProcSignalBarrier - wait until it is guaranteed that all changes
421 : : * requested by a specific call to EmitProcSignalBarrier() have taken effect.
422 : : */
423 : : void
424 : 88 : WaitForProcSignalBarrier(uint64 generation)
425 : : {
1909 andres@anarazel.de 426 [ - + ]: 88 : Assert(generation <= pg_atomic_read_u64(&ProcSignal->psh_barrierGeneration));
427 : :
1214 tmunro@postgresql.or 428 [ + + ]: 88 : elog(DEBUG1,
429 : : "waiting for all backends to process ProcSignalBarrier generation "
430 : : UINT64_FORMAT,
431 : : generation);
432 : :
1243 rhaas@postgresql.org 433 [ + + ]: 9976 : for (int i = NumProcSignalSlots - 1; i >= 0; i--)
434 : : {
1650 tmunro@postgresql.or 435 : 9888 : ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
436 : : uint64 oldval;
437 : :
438 : : /*
439 : : * It's important that we check only pss_barrierGeneration here and
440 : : * not pss_barrierCheckMask. Bits in pss_barrierCheckMask get cleared
441 : : * before the barrier is actually absorbed, but pss_barrierGeneration
442 : : * is updated only afterward.
443 : : */
2088 rhaas@postgresql.org 444 : 9888 : oldval = pg_atomic_read_u64(&slot->pss_barrierGeneration);
445 [ + + ]: 10306 : while (oldval < generation)
446 : : {
1214 tmunro@postgresql.or 447 [ - + ]: 418 : if (ConditionVariableTimedSleep(&slot->pss_barrierCV,
448 : : 5000,
449 : : WAIT_EVENT_PROC_SIGNAL_BARRIER))
1214 tmunro@postgresql.or 450 [ # # ]:UBC 0 : ereport(LOG,
451 : : (errmsg("still waiting for backend with PID %d to accept ProcSignalBarrier",
452 : : (int) pg_atomic_read_u32(&slot->pss_pid))));
2088 rhaas@postgresql.org 453 :CBC 418 : oldval = pg_atomic_read_u64(&slot->pss_barrierGeneration);
454 : : }
1650 tmunro@postgresql.or 455 : 9888 : ConditionVariableCancelSleep();
456 : : }
457 : :
1214 458 [ + + ]: 88 : elog(DEBUG1,
459 : : "finished waiting for all backends to process ProcSignalBarrier generation "
460 : : UINT64_FORMAT,
461 : : generation);
462 : :
463 : : /*
464 : : * The caller is probably calling this function because it wants to read
465 : : * the shared state or perform further writes to shared state once all
466 : : * backends are known to have absorbed the barrier. However, the read of
467 : : * pss_barrierGeneration was performed unlocked; insert a memory barrier
468 : : * to separate it from whatever follows.
469 : : */
2088 rhaas@postgresql.org 470 : 88 : pg_memory_barrier();
471 : 88 : }
472 : :
473 : : /*
474 : : * Handle receipt of an interrupt indicating a global barrier event.
475 : : *
476 : : * All the actual work is deferred to ProcessProcSignalBarrier(), because we
477 : : * cannot safely access the barrier generation inside the signal handler as
478 : : * 64bit atomics might use spinlock based emulation, even for reads. As this
479 : : * routine only gets called when PROCSIG_BARRIER is sent that won't cause a
480 : : * lot of unnecessary work.
481 : : */
482 : : static void
1909 andres@anarazel.de 483 : 597 : HandleProcSignalBarrierInterrupt(void)
484 : : {
485 : 597 : InterruptPending = true;
486 : 597 : ProcSignalBarrierPending = true;
487 : : /* latch will be set by procsignal_sigusr1_handler */
488 : 597 : }
489 : :
490 : : /*
491 : : * Perform global barrier related interrupt checking.
492 : : *
493 : : * Any backend that participates in ProcSignal signaling must arrange to
494 : : * call this function periodically. It is called from CHECK_FOR_INTERRUPTS(),
495 : : * which is enough for normal backends, but not necessarily for all types of
496 : : * background processes.
497 : : */
498 : : void
2088 rhaas@postgresql.org 499 : 597 : ProcessProcSignalBarrier(void)
500 : : {
501 : : uint64 local_gen;
502 : : uint64 shared_gen;
503 : : volatile uint32 flags;
504 : :
1909 andres@anarazel.de 505 [ - + ]: 597 : Assert(MyProcSignalSlot);
506 : :
507 : : /* Exit quickly if there's no work to do. */
2088 rhaas@postgresql.org 508 [ - + ]: 597 : if (!ProcSignalBarrierPending)
2088 rhaas@postgresql.org 509 :UBC 0 : return;
2088 rhaas@postgresql.org 510 :CBC 597 : ProcSignalBarrierPending = false;
511 : :
512 : : /*
513 : : * It's not unlikely to process multiple barriers at once, before the
514 : : * signals for all the barriers have arrived. To avoid unnecessary work in
515 : : * response to subsequent signals, exit early if we already have processed
516 : : * all of them.
517 : : */
1909 andres@anarazel.de 518 : 597 : local_gen = pg_atomic_read_u64(&MyProcSignalSlot->pss_barrierGeneration);
519 : 597 : shared_gen = pg_atomic_read_u64(&ProcSignal->psh_barrierGeneration);
520 : :
521 [ - + ]: 597 : Assert(local_gen <= shared_gen);
522 : :
523 [ - + ]: 597 : if (local_gen == shared_gen)
1909 andres@anarazel.de 524 :UBC 0 : return;
525 : :
526 : : /*
527 : : * Get and clear the flags that are set for this backend. Note that
528 : : * pg_atomic_exchange_u32 is a full barrier, so we're guaranteed that the
529 : : * read of the barrier generation above happens before we atomically
530 : : * extract the flags, and that any subsequent state changes happen
531 : : * afterward.
532 : : *
533 : : * NB: In order to avoid race conditions, we must zero
534 : : * pss_barrierCheckMask first and only afterwards try to do barrier
535 : : * processing. If we did it in the other order, someone could send us
536 : : * another barrier of some type right after we called the
537 : : * barrier-processing function but before we cleared the bit. We would
538 : : * have no way of knowing that the bit needs to stay set in that case, so
539 : : * the need to call the barrier-processing function again would just get
540 : : * forgotten. So instead, we tentatively clear all the bits and then put
541 : : * back any for which we don't manage to successfully absorb the barrier.
542 : : */
2088 rhaas@postgresql.org 543 :CBC 597 : flags = pg_atomic_exchange_u32(&MyProcSignalSlot->pss_barrierCheckMask, 0);
544 : :
545 : : /*
546 : : * If there are no flags set, then we can skip doing any real work.
547 : : * Otherwise, establish a PG_TRY block, so that we don't lose track of
548 : : * which types of barrier processing are needed if an ERROR occurs.
549 : : */
1692 550 [ + - ]: 597 : if (flags != 0)
551 : : {
1578 tgl@sss.pgh.pa.us 552 : 597 : bool success = true;
553 : :
1692 rhaas@postgresql.org 554 [ + - ]: 597 : PG_TRY();
555 : : {
556 : : /*
557 : : * Process each type of barrier. The barrier-processing functions
558 : : * should normally return true, but may return false if the
559 : : * barrier can't be absorbed at the current time. This should be
560 : : * rare, because it's pretty expensive. Every single
561 : : * CHECK_FOR_INTERRUPTS() will return here until we manage to
562 : : * absorb the barrier, and that cost will add up in a hurry.
563 : : *
564 : : * NB: It ought to be OK to call the barrier-processing functions
565 : : * unconditionally, but it's more efficient to call only the ones
566 : : * that might need us to do something based on the flags.
567 : : */
568 [ + + ]: 1791 : while (flags != 0)
569 : : {
570 : : ProcSignalBarrierType type;
1578 tgl@sss.pgh.pa.us 571 : 597 : bool processed = true;
572 : :
1692 rhaas@postgresql.org 573 : 597 : type = (ProcSignalBarrierType) pg_rightmost_one_pos32(flags);
574 [ + - ]: 597 : switch (type)
575 : : {
1302 tmunro@postgresql.or 576 : 597 : case PROCSIGNAL_BARRIER_SMGRRELEASE:
577 : 597 : processed = ProcessBarrierSmgrRelease();
1692 rhaas@postgresql.org 578 : 597 : break;
579 : : }
580 : :
581 : : /*
582 : : * To avoid an infinite loop, we must always unset the bit in
583 : : * flags.
584 : : */
585 : 597 : BARRIER_CLEAR_BIT(flags, type);
586 : :
587 : : /*
588 : : * If we failed to process the barrier, reset the shared bit
589 : : * so we try again later, and set a flag so that we don't bump
590 : : * our generation.
591 : : */
592 [ + - ]: 597 : if (!processed)
593 : : {
1692 rhaas@postgresql.org 594 :UBC 0 : ResetProcSignalBarrierBits(((uint32) 1) << type);
595 : 0 : success = false;
596 : : }
597 : : }
598 : : }
599 : 0 : PG_CATCH();
600 : : {
601 : : /*
602 : : * If an ERROR occurred, we'll need to try again later to handle
603 : : * that barrier type and any others that haven't been handled yet
604 : : * or weren't successfully absorbed.
605 : : */
606 : 0 : ResetProcSignalBarrierBits(flags);
607 : 0 : PG_RE_THROW();
608 : : }
1692 rhaas@postgresql.org 609 [ - + ]:CBC 597 : PG_END_TRY();
610 : :
611 : : /*
612 : : * If some barrier types were not successfully absorbed, we will have
613 : : * to try again later.
614 : : */
615 [ - + ]: 597 : if (!success)
1692 rhaas@postgresql.org 616 :UBC 0 : return;
617 : : }
618 : :
619 : : /*
620 : : * State changes related to all types of barriers that might have been
621 : : * emitted have now been handled, so we can update our notion of the
622 : : * generation to the one we observed before beginning the updates. If
623 : : * things have changed further, it'll get fixed up when this function is
624 : : * next called.
625 : : */
1909 andres@anarazel.de 626 :CBC 597 : pg_atomic_write_u64(&MyProcSignalSlot->pss_barrierGeneration, shared_gen);
1650 tmunro@postgresql.or 627 : 597 : ConditionVariableBroadcast(&MyProcSignalSlot->pss_barrierCV);
628 : : }
629 : :
630 : : /*
631 : : * If it turns out that we couldn't absorb one or more barrier types, either
632 : : * because the barrier-processing functions returned false or due to an error,
633 : : * arrange for processing to be retried later.
634 : : */
635 : : static void
1692 rhaas@postgresql.org 636 :UBC 0 : ResetProcSignalBarrierBits(uint32 flags)
637 : : {
638 : 0 : pg_atomic_fetch_or_u32(&MyProcSignalSlot->pss_barrierCheckMask, flags);
639 : 0 : ProcSignalBarrierPending = true;
640 : 0 : InterruptPending = true;
641 : 0 : }
642 : :
643 : : /*
644 : : * CheckProcSignal - check to see if a particular reason has been
645 : : * signaled, and clear the signal flag. Should be called after receiving
646 : : * SIGUSR1.
647 : : */
648 : : static bool
5881 tgl@sss.pgh.pa.us 649 :CBC 116284 : CheckProcSignal(ProcSignalReason reason)
650 : : {
651 : 116284 : volatile ProcSignalSlot *slot = MyProcSignalSlot;
652 : :
653 [ + + ]: 116284 : if (slot != NULL)
654 : : {
655 : : /*
656 : : * Careful here --- don't clear flag if we haven't seen it set.
657 : : * pss_signalFlags is of type "volatile sig_atomic_t" to allow us to
658 : : * read it here safely, without holding the spinlock.
659 : : */
660 [ + + ]: 116130 : if (slot->pss_signalFlags[reason])
661 : : {
662 : 4482 : slot->pss_signalFlags[reason] = false;
663 : 4482 : return true;
664 : : }
665 : : }
666 : :
667 : 111802 : return false;
668 : : }
669 : :
670 : : /*
671 : : * procsignal_sigusr1_handler - handle SIGUSR1 signal.
672 : : */
673 : : void
674 : 8306 : procsignal_sigusr1_handler(SIGNAL_ARGS)
675 : : {
676 [ + + ]: 8306 : if (CheckProcSignal(PROCSIG_CATCHUP_INTERRUPT))
677 : 2563 : HandleCatchupInterrupt();
678 : :
679 [ + + ]: 8306 : if (CheckProcSignal(PROCSIG_NOTIFY_INTERRUPT))
680 : 20 : HandleNotifyInterrupt();
681 : :
3782 rhaas@postgresql.org 682 [ + + ]: 8306 : if (CheckProcSignal(PROCSIG_PARALLEL_MESSAGE))
683 : 1236 : HandleParallelMessageInterrupt();
684 : :
3015 andres@anarazel.de 685 [ + + ]: 8306 : if (CheckProcSignal(PROCSIG_WALSND_INIT_STOPPING))
686 : 32 : HandleWalSndInitStopping();
687 : :
1909 688 [ + + ]: 8306 : if (CheckProcSignal(PROCSIG_BARRIER))
689 : 597 : HandleProcSignalBarrierInterrupt();
690 : :
1614 fujii@postgresql.org 691 [ + + ]: 8306 : if (CheckProcSignal(PROCSIG_LOG_MEMORY_CONTEXT))
692 : 9 : HandleLogMemoryContextInterrupt();
693 : :
971 akapila@postgresql.o 694 [ + + ]: 8306 : if (CheckProcSignal(PROCSIG_PARALLEL_APPLY_MESSAGE))
695 : 13 : HandleParallelApplyMessageInterrupt();
696 : :
5712 simon@2ndQuadrant.co 697 [ + + ]: 8306 : if (CheckProcSignal(PROCSIG_RECOVERY_CONFLICT_DATABASE))
730 tmunro@postgresql.or 698 : 1 : HandleRecoveryConflictInterrupt(PROCSIG_RECOVERY_CONFLICT_DATABASE);
699 : :
5712 simon@2ndQuadrant.co 700 [ + + ]: 8306 : if (CheckProcSignal(PROCSIG_RECOVERY_CONFLICT_TABLESPACE))
730 tmunro@postgresql.or 701 : 1 : HandleRecoveryConflictInterrupt(PROCSIG_RECOVERY_CONFLICT_TABLESPACE);
702 : :
5712 simon@2ndQuadrant.co 703 [ + + ]: 8306 : if (CheckProcSignal(PROCSIG_RECOVERY_CONFLICT_LOCK))
730 tmunro@postgresql.or 704 : 1 : HandleRecoveryConflictInterrupt(PROCSIG_RECOVERY_CONFLICT_LOCK);
705 : :
5712 simon@2ndQuadrant.co 706 [ + + ]: 8306 : if (CheckProcSignal(PROCSIG_RECOVERY_CONFLICT_SNAPSHOT))
730 tmunro@postgresql.or 707 : 1 : HandleRecoveryConflictInterrupt(PROCSIG_RECOVERY_CONFLICT_SNAPSHOT);
708 : :
883 andres@anarazel.de 709 [ - + ]: 8306 : if (CheckProcSignal(PROCSIG_RECOVERY_CONFLICT_LOGICALSLOT))
730 tmunro@postgresql.or 710 :UBC 0 : HandleRecoveryConflictInterrupt(PROCSIG_RECOVERY_CONFLICT_LOGICALSLOT);
711 : :
5684 simon@2ndQuadrant.co 712 [ + + ]:CBC 8306 : if (CheckProcSignal(PROCSIG_RECOVERY_CONFLICT_STARTUP_DEADLOCK))
730 tmunro@postgresql.or 713 : 7 : HandleRecoveryConflictInterrupt(PROCSIG_RECOVERY_CONFLICT_STARTUP_DEADLOCK);
714 : :
5705 simon@2ndQuadrant.co 715 [ + + ]: 8306 : if (CheckProcSignal(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN))
730 tmunro@postgresql.or 716 : 1 : HandleRecoveryConflictInterrupt(PROCSIG_RECOVERY_CONFLICT_BUFFERPIN);
717 : :
3620 rhaas@postgresql.org 718 : 8306 : SetLatch(MyLatch);
5881 tgl@sss.pgh.pa.us 719 : 8306 : }
720 : :
721 : : /*
722 : : * Send a query cancellation signal to backend.
723 : : *
724 : : * Note: This is called from a backend process before authentication. We
725 : : * cannot take LWLocks yet, but that's OK; we rely on atomic reads of the
726 : : * fields in the ProcSignal slots.
727 : : */
728 : : void
121 heikki.linnakangas@i 729 : 16 : SendCancelRequest(int backendPID, const uint8 *cancel_key, int cancel_key_len)
730 : : {
38 731 [ - + ]: 16 : if (backendPID == 0)
732 : : {
38 heikki.linnakangas@i 733 [ # # ]:UBC 0 : ereport(LOG, (errmsg("invalid cancel request with PID 0")));
734 : 0 : return;
735 : : }
736 : :
737 : : /*
738 : : * See if we have a matching backend. Reading the pss_pid and
739 : : * pss_cancel_key fields is racy, a backend might die and remove itself
740 : : * from the array at any time. The probability of the cancellation key
741 : : * matching wrong process is miniscule, however, so we can live with that.
742 : : * PIDs are reused too, so sending the signal based on PID is inherently
743 : : * racy anyway, although OS's avoid reusing PIDs too soon.
744 : : */
404 heikki.linnakangas@i 745 [ + - ]:CBC 269 : for (int i = 0; i < NumProcSignalSlots; i++)
746 : : {
747 : 269 : ProcSignalSlot *slot = &ProcSignal->psh_slot[i];
748 : : bool match;
749 : :
750 [ + + ]: 269 : if (pg_atomic_read_u32(&slot->pss_pid) != backendPID)
751 : 253 : continue;
752 : :
753 : : /* Acquire the spinlock and re-check */
754 [ - + ]: 16 : SpinLockAcquire(&slot->pss_mutex);
755 [ - + ]: 16 : if (pg_atomic_read_u32(&slot->pss_pid) != backendPID)
756 : : {
404 heikki.linnakangas@i 757 :UBC 0 : SpinLockRelease(&slot->pss_mutex);
758 : 0 : continue;
759 : : }
760 : : else
761 : : {
157 heikki.linnakangas@i 762 [ + - + - ]:CBC 32 : match = slot->pss_cancel_key_len == cancel_key_len &&
763 : 16 : timingsafe_bcmp(slot->pss_cancel_key, cancel_key, cancel_key_len) == 0;
764 : :
404 765 : 16 : SpinLockRelease(&slot->pss_mutex);
766 : :
767 [ + - ]: 16 : if (match)
768 : : {
769 : : /* Found a match; signal that backend to cancel current op */
770 [ - + ]: 16 : ereport(DEBUG2,
771 : : (errmsg_internal("processing cancel request: sending SIGINT to process %d",
772 : : backendPID)));
773 : :
774 : : /*
775 : : * If we have setsid(), signal the backend's whole process
776 : : * group
777 : : */
778 : : #ifdef HAVE_SETSID
779 : 16 : kill(-backendPID, SIGINT);
780 : : #else
781 : : kill(backendPID, SIGINT);
782 : : #endif
783 : : }
784 : : else
785 : : {
786 : : /* Right PID, wrong key: no way, Jose */
404 heikki.linnakangas@i 787 [ # # ]:UBC 0 : ereport(LOG,
788 : : (errmsg("wrong key in cancel request for process %d",
789 : : backendPID)));
790 : : }
404 heikki.linnakangas@i 791 :CBC 16 : return;
792 : : }
793 : : }
794 : :
795 : : /* No matching backend */
404 heikki.linnakangas@i 796 [ # # ]:UBC 0 : ereport(LOG,
797 : : (errmsg("PID %d in cancel request did not match any process",
798 : : backendPID)));
799 : : }
|
