Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * deadlock.c
4 : * POSTGRES deadlock detection code
5 : *
6 : * See src/backend/storage/lmgr/README for a description of the deadlock
7 : * detection and resolution algorithms.
8 : *
9 : *
10 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
11 : * Portions Copyright (c) 1994, Regents of the University of California
12 : *
13 : *
14 : * IDENTIFICATION
15 : * src/backend/storage/lmgr/deadlock.c
16 : *
17 : * Interface:
18 : *
19 : * DeadLockCheck()
20 : * DeadLockReport()
21 : * RememberSimpleDeadLock()
22 : * InitDeadLockChecking()
23 : *
24 : *-------------------------------------------------------------------------
25 : */
26 : #include "postgres.h"
27 :
28 : #include "miscadmin.h"
29 : #include "pg_trace.h"
30 : #include "pgstat.h"
31 : #include "storage/lmgr.h"
32 : #include "storage/proc.h"
33 : #include "storage/procnumber.h"
34 : #include "utils/memutils.h"
35 :
36 :
37 : /*
38 : * One edge in the waits-for graph.
39 : *
40 : * waiter and blocker may or may not be members of a lock group, but if either
41 : * is, it will be the leader rather than any other member of the lock group.
42 : * The group leaders act as representatives of the whole group even though
43 : * those particular processes need not be waiting at all. There will be at
44 : * least one member of the waiter's lock group on the wait queue for the given
45 : * lock, maybe more.
46 : */
47 : typedef struct
48 : {
49 : PGPROC *waiter; /* the leader of the waiting lock group */
50 : PGPROC *blocker; /* the leader of the group it is waiting for */
51 : LOCK *lock; /* the lock being waited for */
52 : int pred; /* workspace for TopoSort */
53 : int link; /* workspace for TopoSort */
54 : } EDGE;
55 :
56 : /* One potential reordering of a lock's wait queue */
57 : typedef struct
58 : {
59 : LOCK *lock; /* the lock whose wait queue is described */
60 : PGPROC **procs; /* array of PGPROC *'s in new wait order */
61 : int nProcs;
62 : } WAIT_ORDER;
63 :
64 : /*
65 : * Information saved about each edge in a detected deadlock cycle. This
66 : * is used to print a diagnostic message upon failure.
67 : *
68 : * Note: because we want to examine this info after releasing the lock
69 : * manager's partition locks, we can't just store LOCK and PGPROC pointers;
70 : * we must extract out all the info we want to be able to print.
71 : */
72 : typedef struct
73 : {
74 : LOCKTAG locktag; /* ID of awaited lock object */
75 : LOCKMODE lockmode; /* type of lock we're waiting for */
76 : int pid; /* PID of blocked backend */
77 : } DEADLOCK_INFO;
78 :
79 :
80 : static bool DeadLockCheckRecurse(PGPROC *proc);
81 : static int TestConfiguration(PGPROC *startProc);
82 : static bool FindLockCycle(PGPROC *checkProc,
83 : EDGE *softEdges, int *nSoftEdges);
84 : static bool FindLockCycleRecurse(PGPROC *checkProc, int depth,
85 : EDGE *softEdges, int *nSoftEdges);
86 : static bool FindLockCycleRecurseMember(PGPROC *checkProc,
87 : PGPROC *checkProcLeader,
88 : int depth, EDGE *softEdges, int *nSoftEdges);
89 : static bool ExpandConstraints(EDGE *constraints, int nConstraints);
90 : static bool TopoSort(LOCK *lock, EDGE *constraints, int nConstraints,
91 : PGPROC **ordering);
92 :
93 : #ifdef DEBUG_DEADLOCK
94 : static void PrintLockQueue(LOCK *lock, const char *info);
95 : #endif
96 :
97 :
98 : /*
99 : * Working space for the deadlock detector
100 : */
101 :
102 : /* Workspace for FindLockCycle */
103 : static PGPROC **visitedProcs; /* Array of visited procs */
104 : static int nVisitedProcs;
105 :
106 : /* Workspace for TopoSort */
107 : static PGPROC **topoProcs; /* Array of not-yet-output procs */
108 : static int *beforeConstraints; /* Counts of remaining before-constraints */
109 : static int *afterConstraints; /* List head for after-constraints */
110 :
111 : /* Output area for ExpandConstraints */
112 : static WAIT_ORDER *waitOrders; /* Array of proposed queue rearrangements */
113 : static int nWaitOrders;
114 : static PGPROC **waitOrderProcs; /* Space for waitOrders queue contents */
115 :
116 : /* Current list of constraints being considered */
117 : static EDGE *curConstraints;
118 : static int nCurConstraints;
119 : static int maxCurConstraints;
120 :
121 : /* Storage space for results from FindLockCycle */
122 : static EDGE *possibleConstraints;
123 : static int nPossibleConstraints;
124 : static int maxPossibleConstraints;
125 : static DEADLOCK_INFO *deadlockDetails;
126 : static int nDeadlockDetails;
127 :
128 : /* PGPROC pointer of any blocking autovacuum worker found */
129 : static PGPROC *blocking_autovacuum_proc = NULL;
130 :
131 :
132 : /*
133 : * InitDeadLockChecking -- initialize deadlock checker during backend startup
134 : *
135 : * This does per-backend initialization of the deadlock checker; primarily,
136 : * allocation of working memory for DeadLockCheck. We do this per-backend
137 : * since there's no percentage in making the kernel do copy-on-write
138 : * inheritance of workspace from the postmaster. We want to allocate the
139 : * space at startup because (a) the deadlock checker might be invoked when
140 : * there's no free memory left, and (b) the checker is normally run inside a
141 : * signal handler, which is a very dangerous place to invoke palloc from.
142 : */
143 : void
144 34976 : InitDeadLockChecking(void)
145 : {
146 : MemoryContext oldcxt;
147 :
148 : /* Make sure allocations are permanent */
149 34976 : oldcxt = MemoryContextSwitchTo(TopMemoryContext);
150 :
151 : /*
152 : * FindLockCycle needs at most MaxBackends entries in visitedProcs[] and
153 : * deadlockDetails[].
154 : */
155 34976 : visitedProcs = (PGPROC **) palloc(MaxBackends * sizeof(PGPROC *));
156 34976 : deadlockDetails = (DEADLOCK_INFO *) palloc(MaxBackends * sizeof(DEADLOCK_INFO));
157 :
158 : /*
159 : * TopoSort needs to consider at most MaxBackends wait-queue entries, and
160 : * it needn't run concurrently with FindLockCycle.
161 : */
162 34976 : topoProcs = visitedProcs; /* re-use this space */
163 34976 : beforeConstraints = (int *) palloc(MaxBackends * sizeof(int));
164 34976 : afterConstraints = (int *) palloc(MaxBackends * sizeof(int));
165 :
166 : /*
167 : * We need to consider rearranging at most MaxBackends/2 wait queues
168 : * (since it takes at least two waiters in a queue to create a soft edge),
169 : * and the expanded form of the wait queues can't involve more than
170 : * MaxBackends total waiters.
171 : */
172 34976 : waitOrders = (WAIT_ORDER *)
173 34976 : palloc((MaxBackends / 2) * sizeof(WAIT_ORDER));
174 34976 : waitOrderProcs = (PGPROC **) palloc(MaxBackends * sizeof(PGPROC *));
175 :
176 : /*
177 : * Allow at most MaxBackends distinct constraints in a configuration. (Is
178 : * this enough? In practice it seems it should be, but I don't quite see
179 : * how to prove it. If we run out, we might fail to find a workable wait
180 : * queue rearrangement even though one exists.) NOTE that this number
181 : * limits the maximum recursion depth of DeadLockCheckRecurse. Making it
182 : * really big might potentially allow a stack-overflow problem.
183 : */
184 34976 : maxCurConstraints = MaxBackends;
185 34976 : curConstraints = (EDGE *) palloc(maxCurConstraints * sizeof(EDGE));
186 :
187 : /*
188 : * Allow up to 3*MaxBackends constraints to be saved without having to
189 : * re-run TestConfiguration. (This is probably more than enough, but we
190 : * can survive if we run low on space by doing excess runs of
191 : * TestConfiguration to re-compute constraint lists each time needed.) The
192 : * last MaxBackends entries in possibleConstraints[] are reserved as
193 : * output workspace for FindLockCycle.
194 : */
195 : StaticAssertStmt(MAX_BACKENDS_BITS <= (32 - 3),
196 : "MAX_BACKENDS_BITS too big for * 4");
197 34976 : maxPossibleConstraints = MaxBackends * 4;
198 34976 : possibleConstraints =
199 34976 : (EDGE *) palloc(maxPossibleConstraints * sizeof(EDGE));
200 :
201 34976 : MemoryContextSwitchTo(oldcxt);
202 34976 : }
203 :
204 : /*
205 : * DeadLockCheck -- Checks for deadlocks for a given process
206 : *
207 : * This code looks for deadlocks involving the given process. If any
208 : * are found, it tries to rearrange lock wait queues to resolve the
209 : * deadlock. If resolution is impossible, return DS_HARD_DEADLOCK ---
210 : * the caller is then expected to abort the given proc's transaction.
211 : *
212 : * Caller must already have locked all partitions of the lock tables.
213 : *
214 : * On failure, deadlock details are recorded in deadlockDetails[] for
215 : * subsequent printing by DeadLockReport(). That activity is separate
216 : * because (a) we don't want to do it while holding all those LWLocks,
217 : * and (b) we are typically invoked inside a signal handler.
218 : */
219 : DeadLockState
220 56 : DeadLockCheck(PGPROC *proc)
221 : {
222 : /* Initialize to "no constraints" */
223 56 : nCurConstraints = 0;
224 56 : nPossibleConstraints = 0;
225 56 : nWaitOrders = 0;
226 :
227 : /* Initialize to not blocked by an autovacuum worker */
228 56 : blocking_autovacuum_proc = NULL;
229 :
230 : /* Search for deadlocks and possible fixes */
231 56 : if (DeadLockCheckRecurse(proc))
232 : {
233 : /*
234 : * Call FindLockCycle one more time, to record the correct
235 : * deadlockDetails[] for the basic state with no rearrangements.
236 : */
237 : int nSoftEdges;
238 :
239 : TRACE_POSTGRESQL_DEADLOCK_FOUND();
240 :
241 10 : nWaitOrders = 0;
242 10 : if (!FindLockCycle(proc, possibleConstraints, &nSoftEdges))
243 0 : elog(FATAL, "deadlock seems to have disappeared");
244 :
245 10 : return DS_HARD_DEADLOCK; /* cannot find a non-deadlocked state */
246 : }
247 :
248 : /* Apply any needed rearrangements of wait queues */
249 52 : for (int i = 0; i < nWaitOrders; i++)
250 : {
251 6 : LOCK *lock = waitOrders[i].lock;
252 6 : PGPROC **procs = waitOrders[i].procs;
253 6 : int nProcs = waitOrders[i].nProcs;
254 6 : dclist_head *waitQueue = &lock->waitProcs;
255 :
256 : Assert(nProcs == dclist_count(waitQueue));
257 :
258 : #ifdef DEBUG_DEADLOCK
259 : PrintLockQueue(lock, "DeadLockCheck:");
260 : #endif
261 :
262 : /* Reset the queue and re-add procs in the desired order */
263 6 : dclist_init(waitQueue);
264 24 : for (int j = 0; j < nProcs; j++)
265 18 : dclist_push_tail(waitQueue, &procs[j]->links);
266 :
267 : #ifdef DEBUG_DEADLOCK
268 : PrintLockQueue(lock, "rearranged to:");
269 : #endif
270 :
271 : /* See if any waiters for the lock can be woken up now */
272 6 : ProcLockWakeup(GetLocksMethodTable(lock), lock);
273 : }
274 :
275 : /* Return code tells caller if we had to escape a deadlock or not */
276 46 : if (nWaitOrders > 0)
277 6 : return DS_SOFT_DEADLOCK;
278 40 : else if (blocking_autovacuum_proc != NULL)
279 0 : return DS_BLOCKED_BY_AUTOVACUUM;
280 : else
281 40 : return DS_NO_DEADLOCK;
282 : }
283 :
284 : /*
285 : * Return the PGPROC of the autovacuum that's blocking a process.
286 : *
287 : * We reset the saved pointer as soon as we pass it back.
288 : */
289 : PGPROC *
290 0 : GetBlockingAutoVacuumPgproc(void)
291 : {
292 : PGPROC *ptr;
293 :
294 0 : ptr = blocking_autovacuum_proc;
295 0 : blocking_autovacuum_proc = NULL;
296 :
297 0 : return ptr;
298 : }
299 :
300 : /*
301 : * DeadLockCheckRecurse -- recursively search for valid orderings
302 : *
303 : * curConstraints[] holds the current set of constraints being considered
304 : * by an outer level of recursion. Add to this each possible solution
305 : * constraint for any cycle detected at this level.
306 : *
307 : * Returns true if no solution exists. Returns false if a deadlock-free
308 : * state is attainable, in which case waitOrders[] shows the required
309 : * rearrangements of lock wait queues (if any).
310 : */
311 : static bool
312 62 : DeadLockCheckRecurse(PGPROC *proc)
313 : {
314 : int nEdges;
315 : int oldPossibleConstraints;
316 : bool savedList;
317 : int i;
318 :
319 62 : nEdges = TestConfiguration(proc);
320 62 : if (nEdges < 0)
321 10 : return true; /* hard deadlock --- no solution */
322 52 : if (nEdges == 0)
323 46 : return false; /* good configuration found */
324 6 : if (nCurConstraints >= maxCurConstraints)
325 0 : return true; /* out of room for active constraints? */
326 6 : oldPossibleConstraints = nPossibleConstraints;
327 6 : if (nPossibleConstraints + nEdges + MaxBackends <= maxPossibleConstraints)
328 : {
329 : /* We can save the edge list in possibleConstraints[] */
330 6 : nPossibleConstraints += nEdges;
331 6 : savedList = true;
332 : }
333 : else
334 : {
335 : /* Not room; will need to regenerate the edges on-the-fly */
336 0 : savedList = false;
337 : }
338 :
339 : /*
340 : * Try each available soft edge as an addition to the configuration.
341 : */
342 6 : for (i = 0; i < nEdges; i++)
343 : {
344 6 : if (!savedList && i > 0)
345 : {
346 : /* Regenerate the list of possible added constraints */
347 0 : if (nEdges != TestConfiguration(proc))
348 0 : elog(FATAL, "inconsistent results during deadlock check");
349 : }
350 6 : curConstraints[nCurConstraints] =
351 6 : possibleConstraints[oldPossibleConstraints + i];
352 6 : nCurConstraints++;
353 6 : if (!DeadLockCheckRecurse(proc))
354 6 : return false; /* found a valid solution! */
355 : /* give up on that added constraint, try again */
356 0 : nCurConstraints--;
357 : }
358 0 : nPossibleConstraints = oldPossibleConstraints;
359 0 : return true; /* no solution found */
360 : }
361 :
362 :
363 : /*--------------------
364 : * Test a configuration (current set of constraints) for validity.
365 : *
366 : * Returns:
367 : * 0: the configuration is good (no deadlocks)
368 : * -1: the configuration has a hard deadlock or is not self-consistent
369 : * >0: the configuration has one or more soft deadlocks
370 : *
371 : * In the soft-deadlock case, one of the soft cycles is chosen arbitrarily
372 : * and a list of its soft edges is returned beginning at
373 : * possibleConstraints+nPossibleConstraints. The return value is the
374 : * number of soft edges.
375 : *--------------------
376 : */
377 : static int
378 62 : TestConfiguration(PGPROC *startProc)
379 : {
380 62 : int softFound = 0;
381 62 : EDGE *softEdges = possibleConstraints + nPossibleConstraints;
382 : int nSoftEdges;
383 : int i;
384 :
385 : /*
386 : * Make sure we have room for FindLockCycle's output.
387 : */
388 62 : if (nPossibleConstraints + MaxBackends > maxPossibleConstraints)
389 0 : return -1;
390 :
391 : /*
392 : * Expand current constraint set into wait orderings. Fail if the
393 : * constraint set is not self-consistent.
394 : */
395 62 : if (!ExpandConstraints(curConstraints, nCurConstraints))
396 0 : return -1;
397 :
398 : /*
399 : * Check for cycles involving startProc or any of the procs mentioned in
400 : * constraints. We check startProc last because if it has a soft cycle
401 : * still to be dealt with, we want to deal with that first.
402 : */
403 68 : for (i = 0; i < nCurConstraints; i++)
404 : {
405 6 : if (FindLockCycle(curConstraints[i].waiter, softEdges, &nSoftEdges))
406 : {
407 0 : if (nSoftEdges == 0)
408 0 : return -1; /* hard deadlock detected */
409 0 : softFound = nSoftEdges;
410 : }
411 6 : if (FindLockCycle(curConstraints[i].blocker, softEdges, &nSoftEdges))
412 : {
413 0 : if (nSoftEdges == 0)
414 0 : return -1; /* hard deadlock detected */
415 0 : softFound = nSoftEdges;
416 : }
417 : }
418 62 : if (FindLockCycle(startProc, softEdges, &nSoftEdges))
419 : {
420 16 : if (nSoftEdges == 0)
421 10 : return -1; /* hard deadlock detected */
422 6 : softFound = nSoftEdges;
423 : }
424 52 : return softFound;
425 : }
426 :
427 :
428 : /*
429 : * FindLockCycle -- basic check for deadlock cycles
430 : *
431 : * Scan outward from the given proc to see if there is a cycle in the
432 : * waits-for graph that includes this proc. Return true if a cycle
433 : * is found, else false. If a cycle is found, we return a list of
434 : * the "soft edges", if any, included in the cycle. These edges could
435 : * potentially be eliminated by rearranging wait queues. We also fill
436 : * deadlockDetails[] with information about the detected cycle; this info
437 : * is not used by the deadlock algorithm itself, only to print a useful
438 : * message after failing.
439 : *
440 : * Since we need to be able to check hypothetical configurations that would
441 : * exist after wait queue rearrangement, the routine pays attention to the
442 : * table of hypothetical queue orders in waitOrders[]. These orders will
443 : * be believed in preference to the actual ordering seen in the locktable.
444 : */
445 : static bool
446 84 : FindLockCycle(PGPROC *checkProc,
447 : EDGE *softEdges, /* output argument */
448 : int *nSoftEdges) /* output argument */
449 : {
450 84 : nVisitedProcs = 0;
451 84 : nDeadlockDetails = 0;
452 84 : *nSoftEdges = 0;
453 84 : return FindLockCycleRecurse(checkProc, 0, softEdges, nSoftEdges);
454 : }
455 :
456 : static bool
457 256 : FindLockCycleRecurse(PGPROC *checkProc,
458 : int depth,
459 : EDGE *softEdges, /* output argument */
460 : int *nSoftEdges) /* output argument */
461 : {
462 : int i;
463 : dlist_iter iter;
464 :
465 : /*
466 : * If this process is a lock group member, check the leader instead. (Note
467 : * that we might be the leader, in which case this is a no-op.)
468 : */
469 256 : if (checkProc->lockGroupLeader != NULL)
470 52 : checkProc = checkProc->lockGroupLeader;
471 :
472 : /*
473 : * Have we already seen this proc?
474 : */
475 546 : for (i = 0; i < nVisitedProcs; i++)
476 : {
477 328 : if (visitedProcs[i] == checkProc)
478 : {
479 : /* If we return to starting point, we have a deadlock cycle */
480 38 : if (i == 0)
481 : {
482 : /*
483 : * record total length of cycle --- outer levels will now fill
484 : * deadlockDetails[]
485 : */
486 : Assert(depth <= MaxBackends);
487 26 : nDeadlockDetails = depth;
488 :
489 26 : return true;
490 : }
491 :
492 : /*
493 : * Otherwise, we have a cycle but it does not include the start
494 : * point, so say "no deadlock".
495 : */
496 12 : return false;
497 : }
498 : }
499 : /* Mark proc as seen */
500 : Assert(nVisitedProcs < MaxBackends);
501 218 : visitedProcs[nVisitedProcs++] = checkProc;
502 :
503 : /*
504 : * If the process is waiting, there is an outgoing waits-for edge to each
505 : * process that blocks it.
506 : */
507 368 : if (checkProc->links.next != NULL && checkProc->waitLock != NULL &&
508 150 : FindLockCycleRecurseMember(checkProc, checkProc, depth, softEdges,
509 : nSoftEdges))
510 82 : return true;
511 :
512 : /*
513 : * If the process is not waiting, there could still be outgoing waits-for
514 : * edges if it is part of a lock group, because other members of the lock
515 : * group might be waiting even though this process is not. (Given lock
516 : * groups {A1, A2} and {B1, B2}, if A1 waits for B1 and B2 waits for A2,
517 : * that is a deadlock even neither of B1 and A2 are waiting for anything.)
518 : */
519 214 : dlist_foreach(iter, &checkProc->lockGroupMembers)
520 : {
521 : PGPROC *memberProc;
522 :
523 86 : memberProc = dlist_container(PGPROC, lockGroupLink, iter.cur);
524 :
525 86 : if (memberProc->links.next != NULL && memberProc->waitLock != NULL &&
526 42 : memberProc != checkProc &&
527 42 : FindLockCycleRecurseMember(memberProc, checkProc, depth, softEdges,
528 : nSoftEdges))
529 8 : return true;
530 : }
531 :
532 128 : return false;
533 : }
534 :
535 : static bool
536 192 : FindLockCycleRecurseMember(PGPROC *checkProc,
537 : PGPROC *checkProcLeader,
538 : int depth,
539 : EDGE *softEdges, /* output argument */
540 : int *nSoftEdges) /* output argument */
541 : {
542 : PGPROC *proc;
543 192 : LOCK *lock = checkProc->waitLock;
544 : dlist_iter proclock_iter;
545 : LockMethod lockMethodTable;
546 : int conflictMask;
547 : int i;
548 : int numLockModes,
549 : lm;
550 :
551 : /*
552 : * The relation extension lock can never participate in actual deadlock
553 : * cycle. See Assert in LockAcquireExtended. So, there is no advantage
554 : * in checking wait edges from it.
555 : */
556 192 : if (LOCK_LOCKTAG(*lock) == LOCKTAG_RELATION_EXTEND)
557 0 : return false;
558 :
559 192 : lockMethodTable = GetLocksMethodTable(lock);
560 192 : numLockModes = lockMethodTable->numLockModes;
561 192 : conflictMask = lockMethodTable->conflictTab[checkProc->waitLockMode];
562 :
563 : /*
564 : * Scan for procs that already hold conflicting locks. These are "hard"
565 : * edges in the waits-for graph.
566 : */
567 554 : dlist_foreach(proclock_iter, &lock->procLocks)
568 : {
569 442 : PROCLOCK *proclock = dlist_container(PROCLOCK, lockLink, proclock_iter.cur);
570 : PGPROC *leader;
571 :
572 442 : proc = proclock->tag.myProc;
573 442 : leader = proc->lockGroupLeader == NULL ? proc : proc->lockGroupLeader;
574 :
575 : /* A proc never blocks itself or any other lock group member */
576 442 : if (leader != checkProcLeader)
577 : {
578 2188 : for (lm = 1; lm <= numLockModes; lm++)
579 : {
580 2042 : if ((proclock->holdMask & LOCKBIT_ON(lm)) &&
581 : (conflictMask & LOCKBIT_ON(lm)))
582 : {
583 : /* This proc hard-blocks checkProc */
584 140 : if (FindLockCycleRecurse(proc, depth + 1,
585 : softEdges, nSoftEdges))
586 : {
587 : /* fill deadlockDetails[] */
588 80 : DEADLOCK_INFO *info = &deadlockDetails[depth];
589 :
590 80 : info->locktag = lock->tag;
591 80 : info->lockmode = checkProc->waitLockMode;
592 80 : info->pid = checkProc->pid;
593 :
594 80 : return true;
595 : }
596 :
597 : /*
598 : * No deadlock here, but see if this proc is an autovacuum
599 : * that is directly hard-blocking our own proc. If so,
600 : * report it so that the caller can send a cancel signal
601 : * to it, if appropriate. If there's more than one such
602 : * proc, it's indeterminate which one will be reported.
603 : *
604 : * We don't touch autovacuums that are indirectly blocking
605 : * us; it's up to the direct blockee to take action. This
606 : * rule simplifies understanding the behavior and ensures
607 : * that an autovacuum won't be canceled with less than
608 : * deadlock_timeout grace period.
609 : *
610 : * Note we read statusFlags without any locking. This is
611 : * OK only for checking the PROC_IS_AUTOVACUUM flag,
612 : * because that flag is set at process start and never
613 : * reset. There is logic elsewhere to avoid canceling an
614 : * autovacuum that is working to prevent XID wraparound
615 : * problems (which needs to read a different statusFlags
616 : * bit), but we don't do that here to avoid grabbing
617 : * ProcArrayLock.
618 : */
619 60 : if (checkProc == MyProc &&
620 38 : proc->statusFlags & PROC_IS_AUTOVACUUM)
621 0 : blocking_autovacuum_proc = proc;
622 :
623 : /* We're done looking at this proclock */
624 60 : break;
625 : }
626 : }
627 : }
628 : }
629 :
630 : /*
631 : * Scan for procs that are ahead of this one in the lock's wait queue.
632 : * Those that have conflicting requests soft-block this one. This must be
633 : * done after the hard-block search, since if another proc both hard- and
634 : * soft-blocks this one, we want to call it a hard edge.
635 : *
636 : * If there is a proposed re-ordering of the lock's wait order, use that
637 : * rather than the current wait order.
638 : */
639 130 : for (i = 0; i < nWaitOrders; i++)
640 : {
641 54 : if (waitOrders[i].lock == lock)
642 36 : break;
643 : }
644 :
645 112 : if (i < nWaitOrders)
646 : {
647 : /* Use the given hypothetical wait queue order */
648 36 : PGPROC **procs = waitOrders[i].procs;
649 36 : int queue_size = waitOrders[i].nProcs;
650 :
651 46 : for (i = 0; i < queue_size; i++)
652 : {
653 : PGPROC *leader;
654 :
655 46 : proc = procs[i];
656 46 : leader = proc->lockGroupLeader == NULL ? proc :
657 : proc->lockGroupLeader;
658 :
659 : /*
660 : * TopoSort will always return an ordering with group members
661 : * adjacent to each other in the wait queue (see comments
662 : * therein). So, as soon as we reach a process in the same lock
663 : * group as checkProc, we know we've found all the conflicts that
664 : * precede any member of the lock group lead by checkProcLeader.
665 : */
666 46 : if (leader == checkProcLeader)
667 36 : break;
668 :
669 : /* Is there a conflict with this guy's request? */
670 10 : if ((LOCKBIT_ON(proc->waitLockMode) & conflictMask) != 0)
671 : {
672 : /* This proc soft-blocks checkProc */
673 10 : if (FindLockCycleRecurse(proc, depth + 1,
674 : softEdges, nSoftEdges))
675 : {
676 : /* fill deadlockDetails[] */
677 0 : DEADLOCK_INFO *info = &deadlockDetails[depth];
678 :
679 0 : info->locktag = lock->tag;
680 0 : info->lockmode = checkProc->waitLockMode;
681 0 : info->pid = checkProc->pid;
682 :
683 : /*
684 : * Add this edge to the list of soft edges in the cycle
685 : */
686 : Assert(*nSoftEdges < MaxBackends);
687 0 : softEdges[*nSoftEdges].waiter = checkProcLeader;
688 0 : softEdges[*nSoftEdges].blocker = leader;
689 0 : softEdges[*nSoftEdges].lock = lock;
690 0 : (*nSoftEdges)++;
691 0 : return true;
692 : }
693 : }
694 : }
695 : }
696 : else
697 : {
698 76 : PGPROC *lastGroupMember = NULL;
699 : dlist_iter proc_iter;
700 : dclist_head *waitQueue;
701 :
702 : /* Use the true lock wait queue order */
703 76 : waitQueue = &lock->waitProcs;
704 :
705 : /*
706 : * Find the last member of the lock group that is present in the wait
707 : * queue. Anything after this is not a soft lock conflict. If group
708 : * locking is not in use, then we know immediately which process we're
709 : * looking for, but otherwise we've got to search the wait queue to
710 : * find the last process actually present.
711 : */
712 76 : if (checkProc->lockGroupLeader == NULL)
713 58 : lastGroupMember = checkProc;
714 : else
715 : {
716 64 : dclist_foreach(proc_iter, waitQueue)
717 : {
718 46 : proc = dlist_container(PGPROC, links, proc_iter.cur);
719 :
720 46 : if (proc->lockGroupLeader == checkProcLeader)
721 26 : lastGroupMember = proc;
722 : }
723 : Assert(lastGroupMember != NULL);
724 : }
725 :
726 : /*
727 : * OK, now rescan (or scan) the queue to identify the soft conflicts.
728 : */
729 98 : dclist_foreach(proc_iter, waitQueue)
730 : {
731 : PGPROC *leader;
732 :
733 98 : proc = dlist_container(PGPROC, links, proc_iter.cur);
734 :
735 98 : leader = proc->lockGroupLeader == NULL ? proc :
736 : proc->lockGroupLeader;
737 :
738 : /* Done when we reach the target proc */
739 98 : if (proc == lastGroupMember)
740 66 : break;
741 :
742 : /* Is there a conflict with this guy's request? */
743 32 : if ((LOCKBIT_ON(proc->waitLockMode) & conflictMask) != 0 &&
744 : leader != checkProcLeader)
745 : {
746 : /* This proc soft-blocks checkProc */
747 22 : if (FindLockCycleRecurse(proc, depth + 1,
748 : softEdges, nSoftEdges))
749 : {
750 : /* fill deadlockDetails[] */
751 10 : DEADLOCK_INFO *info = &deadlockDetails[depth];
752 :
753 10 : info->locktag = lock->tag;
754 10 : info->lockmode = checkProc->waitLockMode;
755 10 : info->pid = checkProc->pid;
756 :
757 : /*
758 : * Add this edge to the list of soft edges in the cycle
759 : */
760 : Assert(*nSoftEdges < MaxBackends);
761 10 : softEdges[*nSoftEdges].waiter = checkProcLeader;
762 10 : softEdges[*nSoftEdges].blocker = leader;
763 10 : softEdges[*nSoftEdges].lock = lock;
764 10 : (*nSoftEdges)++;
765 10 : return true;
766 : }
767 : }
768 : }
769 : }
770 :
771 : /*
772 : * No conflict detected here.
773 : */
774 102 : return false;
775 : }
776 :
777 :
778 : /*
779 : * ExpandConstraints -- expand a list of constraints into a set of
780 : * specific new orderings for affected wait queues
781 : *
782 : * Input is a list of soft edges to be reversed. The output is a list
783 : * of nWaitOrders WAIT_ORDER structs in waitOrders[], with PGPROC array
784 : * workspace in waitOrderProcs[].
785 : *
786 : * Returns true if able to build an ordering that satisfies all the
787 : * constraints, false if not (there are contradictory constraints).
788 : */
789 : static bool
790 62 : ExpandConstraints(EDGE *constraints,
791 : int nConstraints)
792 : {
793 62 : int nWaitOrderProcs = 0;
794 : int i,
795 : j;
796 :
797 62 : nWaitOrders = 0;
798 :
799 : /*
800 : * Scan constraint list backwards. This is because the last-added
801 : * constraint is the only one that could fail, and so we want to test it
802 : * for inconsistency first.
803 : */
804 68 : for (i = nConstraints; --i >= 0;)
805 : {
806 6 : LOCK *lock = constraints[i].lock;
807 :
808 : /* Did we already make a list for this lock? */
809 6 : for (j = nWaitOrders; --j >= 0;)
810 : {
811 0 : if (waitOrders[j].lock == lock)
812 0 : break;
813 : }
814 6 : if (j >= 0)
815 0 : continue;
816 : /* No, so allocate a new list */
817 6 : waitOrders[nWaitOrders].lock = lock;
818 6 : waitOrders[nWaitOrders].procs = waitOrderProcs + nWaitOrderProcs;
819 6 : waitOrders[nWaitOrders].nProcs = dclist_count(&lock->waitProcs);
820 6 : nWaitOrderProcs += dclist_count(&lock->waitProcs);
821 : Assert(nWaitOrderProcs <= MaxBackends);
822 :
823 : /*
824 : * Do the topo sort. TopoSort need not examine constraints after this
825 : * one, since they must be for different locks.
826 : */
827 6 : if (!TopoSort(lock, constraints, i + 1,
828 6 : waitOrders[nWaitOrders].procs))
829 0 : return false;
830 6 : nWaitOrders++;
831 : }
832 62 : return true;
833 : }
834 :
835 :
836 : /*
837 : * TopoSort -- topological sort of a wait queue
838 : *
839 : * Generate a re-ordering of a lock's wait queue that satisfies given
840 : * constraints about certain procs preceding others. (Each such constraint
841 : * is a fact of a partial ordering.) Minimize rearrangement of the queue
842 : * not needed to achieve the partial ordering.
843 : *
844 : * This is a lot simpler and slower than, for example, the topological sort
845 : * algorithm shown in Knuth's Volume 1. However, Knuth's method doesn't
846 : * try to minimize the damage to the existing order. In practice we are
847 : * not likely to be working with more than a few constraints, so the apparent
848 : * slowness of the algorithm won't really matter.
849 : *
850 : * The initial queue ordering is taken directly from the lock's wait queue.
851 : * The output is an array of PGPROC pointers, of length equal to the lock's
852 : * wait queue length (the caller is responsible for providing this space).
853 : * The partial order is specified by an array of EDGE structs. Each EDGE
854 : * is one that we need to reverse, therefore the "waiter" must appear before
855 : * the "blocker" in the output array. The EDGE array may well contain
856 : * edges associated with other locks; these should be ignored.
857 : *
858 : * Returns true if able to build an ordering that satisfies all the
859 : * constraints, false if not (there are contradictory constraints).
860 : */
861 : static bool
862 6 : TopoSort(LOCK *lock,
863 : EDGE *constraints,
864 : int nConstraints,
865 : PGPROC **ordering) /* output argument */
866 : {
867 6 : dclist_head *waitQueue = &lock->waitProcs;
868 6 : int queue_size = dclist_count(waitQueue);
869 : PGPROC *proc;
870 : int i,
871 : j,
872 : jj,
873 : k,
874 : kk,
875 : last;
876 : dlist_iter proc_iter;
877 :
878 : /* First, fill topoProcs[] array with the procs in their current order */
879 6 : i = 0;
880 24 : dclist_foreach(proc_iter, waitQueue)
881 : {
882 18 : proc = dlist_container(PGPROC, links, proc_iter.cur);
883 18 : topoProcs[i++] = proc;
884 : }
885 : Assert(i == queue_size);
886 :
887 : /*
888 : * Scan the constraints, and for each proc in the array, generate a count
889 : * of the number of constraints that say it must be before something else,
890 : * plus a list of the constraints that say it must be after something
891 : * else. The count for the j'th proc is stored in beforeConstraints[j],
892 : * and the head of its list in afterConstraints[j]. Each constraint
893 : * stores its list link in constraints[i].link (note any constraint will
894 : * be in just one list). The array index for the before-proc of the i'th
895 : * constraint is remembered in constraints[i].pred.
896 : *
897 : * Note that it's not necessarily the case that every constraint affects
898 : * this particular wait queue. Prior to group locking, a process could be
899 : * waiting for at most one lock. But a lock group can be waiting for
900 : * zero, one, or multiple locks. Since topoProcs[] is an array of the
901 : * processes actually waiting, while constraints[] is an array of group
902 : * leaders, we've got to scan through topoProcs[] for each constraint,
903 : * checking whether both a waiter and a blocker for that group are
904 : * present. If so, the constraint is relevant to this wait queue; if not,
905 : * it isn't.
906 : */
907 12 : MemSet(beforeConstraints, 0, queue_size * sizeof(int));
908 12 : MemSet(afterConstraints, 0, queue_size * sizeof(int));
909 12 : for (i = 0; i < nConstraints; i++)
910 : {
911 : /*
912 : * Find a representative process that is on the lock queue and part of
913 : * the waiting lock group. This may or may not be the leader, which
914 : * may or may not be waiting at all. If there are any other processes
915 : * in the same lock group on the queue, set their number of
916 : * beforeConstraints to -1 to indicate that they should be emitted
917 : * with their groupmates rather than considered separately.
918 : *
919 : * In this loop and the similar one just below, it's critical that we
920 : * consistently select the same representative member of any one lock
921 : * group, so that all the constraints are associated with the same
922 : * proc, and the -1's are only associated with not-representative
923 : * members. We select the last one in the topoProcs array.
924 : */
925 6 : proc = constraints[i].waiter;
926 : Assert(proc != NULL);
927 6 : jj = -1;
928 24 : for (j = queue_size; --j >= 0;)
929 : {
930 18 : PGPROC *waiter = topoProcs[j];
931 :
932 18 : if (waiter == proc || waiter->lockGroupLeader == proc)
933 : {
934 : Assert(waiter->waitLock == lock);
935 10 : if (jj == -1)
936 6 : jj = j;
937 : else
938 : {
939 : Assert(beforeConstraints[j] <= 0);
940 4 : beforeConstraints[j] = -1;
941 : }
942 : }
943 : }
944 :
945 : /* If no matching waiter, constraint is not relevant to this lock. */
946 6 : if (jj < 0)
947 0 : continue;
948 :
949 : /*
950 : * Similarly, find a representative process that is on the lock queue
951 : * and waiting for the blocking lock group. Again, this could be the
952 : * leader but does not need to be.
953 : */
954 6 : proc = constraints[i].blocker;
955 : Assert(proc != NULL);
956 6 : kk = -1;
957 24 : for (k = queue_size; --k >= 0;)
958 : {
959 18 : PGPROC *blocker = topoProcs[k];
960 :
961 18 : if (blocker == proc || blocker->lockGroupLeader == proc)
962 : {
963 : Assert(blocker->waitLock == lock);
964 6 : if (kk == -1)
965 6 : kk = k;
966 : else
967 : {
968 : Assert(beforeConstraints[k] <= 0);
969 0 : beforeConstraints[k] = -1;
970 : }
971 : }
972 : }
973 :
974 : /* If no matching blocker, constraint is not relevant to this lock. */
975 6 : if (kk < 0)
976 0 : continue;
977 :
978 : Assert(beforeConstraints[jj] >= 0);
979 6 : beforeConstraints[jj]++; /* waiter must come before */
980 : /* add this constraint to list of after-constraints for blocker */
981 6 : constraints[i].pred = jj;
982 6 : constraints[i].link = afterConstraints[kk];
983 6 : afterConstraints[kk] = i + 1;
984 : }
985 :
986 : /*--------------------
987 : * Now scan the topoProcs array backwards. At each step, output the
988 : * last proc that has no remaining before-constraints plus any other
989 : * members of the same lock group; then decrease the beforeConstraints
990 : * count of each of the procs it was constrained against.
991 : * i = index of ordering[] entry we want to output this time
992 : * j = search index for topoProcs[]
993 : * k = temp for scanning constraint list for proc j
994 : * last = last non-null index in topoProcs (avoid redundant searches)
995 : *--------------------
996 : */
997 6 : last = queue_size - 1;
998 20 : for (i = queue_size - 1; i >= 0;)
999 : {
1000 : int c;
1001 14 : int nmatches = 0;
1002 :
1003 : /* Find next candidate to output */
1004 14 : while (topoProcs[last] == NULL)
1005 0 : last--;
1006 28 : for (j = last; j >= 0; j--)
1007 : {
1008 28 : if (topoProcs[j] != NULL && beforeConstraints[j] == 0)
1009 14 : break;
1010 : }
1011 :
1012 : /* If no available candidate, topological sort fails */
1013 14 : if (j < 0)
1014 0 : return false;
1015 :
1016 : /*
1017 : * Output everything in the lock group. There's no point in
1018 : * outputting an ordering where members of the same lock group are not
1019 : * consecutive on the wait queue: if some other waiter is between two
1020 : * requests that belong to the same group, then either it conflicts
1021 : * with both of them and is certainly not a solution; or it conflicts
1022 : * with at most one of them and is thus isomorphic to an ordering
1023 : * where the group members are consecutive.
1024 : */
1025 14 : proc = topoProcs[j];
1026 14 : if (proc->lockGroupLeader != NULL)
1027 4 : proc = proc->lockGroupLeader;
1028 : Assert(proc != NULL);
1029 56 : for (c = 0; c <= last; ++c)
1030 : {
1031 42 : if (topoProcs[c] == proc || (topoProcs[c] != NULL &&
1032 22 : topoProcs[c]->lockGroupLeader == proc))
1033 : {
1034 18 : ordering[i - nmatches] = topoProcs[c];
1035 18 : topoProcs[c] = NULL;
1036 18 : ++nmatches;
1037 : }
1038 : }
1039 : Assert(nmatches > 0);
1040 14 : i -= nmatches;
1041 :
1042 : /* Update beforeConstraints counts of its predecessors */
1043 20 : for (k = afterConstraints[j]; k > 0; k = constraints[k - 1].link)
1044 6 : beforeConstraints[constraints[k - 1].pred]--;
1045 : }
1046 :
1047 : /* Done */
1048 6 : return true;
1049 : }
1050 :
1051 : #ifdef DEBUG_DEADLOCK
1052 : static void
1053 : PrintLockQueue(LOCK *lock, const char *info)
1054 : {
1055 : dclist_head *waitQueue = &lock->waitProcs;
1056 : dlist_iter proc_iter;
1057 :
1058 : printf("%s lock %p queue ", info, lock);
1059 :
1060 : dclist_foreach(proc_iter, waitQueue)
1061 : {
1062 : PGPROC *proc = dlist_container(PGPROC, links, proc_iter.cur);
1063 :
1064 : printf(" %d", proc->pid);
1065 : }
1066 : printf("\n");
1067 : fflush(stdout);
1068 : }
1069 : #endif
1070 :
1071 : /*
1072 : * Report a detected deadlock, with available details.
1073 : */
1074 : void
1075 12 : DeadLockReport(void)
1076 : {
1077 : StringInfoData clientbuf; /* errdetail for client */
1078 : StringInfoData logbuf; /* errdetail for server log */
1079 : StringInfoData locktagbuf;
1080 : int i;
1081 :
1082 12 : initStringInfo(&clientbuf);
1083 12 : initStringInfo(&logbuf);
1084 12 : initStringInfo(&locktagbuf);
1085 :
1086 : /* Generate the "waits for" lines sent to the client */
1087 50 : for (i = 0; i < nDeadlockDetails; i++)
1088 : {
1089 38 : DEADLOCK_INFO *info = &deadlockDetails[i];
1090 : int nextpid;
1091 :
1092 : /* The last proc waits for the first one... */
1093 38 : if (i < nDeadlockDetails - 1)
1094 26 : nextpid = info[1].pid;
1095 : else
1096 12 : nextpid = deadlockDetails[0].pid;
1097 :
1098 : /* reset locktagbuf to hold next object description */
1099 38 : resetStringInfo(&locktagbuf);
1100 :
1101 38 : DescribeLockTag(&locktagbuf, &info->locktag);
1102 :
1103 38 : if (i > 0)
1104 26 : appendStringInfoChar(&clientbuf, '\n');
1105 :
1106 76 : appendStringInfo(&clientbuf,
1107 38 : _("Process %d waits for %s on %s; blocked by process %d."),
1108 : info->pid,
1109 38 : GetLockmodeName(info->locktag.locktag_lockmethodid,
1110 : info->lockmode),
1111 : locktagbuf.data,
1112 : nextpid);
1113 : }
1114 :
1115 : /* Duplicate all the above for the server ... */
1116 12 : appendBinaryStringInfo(&logbuf, clientbuf.data, clientbuf.len);
1117 :
1118 : /* ... and add info about query strings */
1119 50 : for (i = 0; i < nDeadlockDetails; i++)
1120 : {
1121 38 : DEADLOCK_INFO *info = &deadlockDetails[i];
1122 :
1123 38 : appendStringInfoChar(&logbuf, '\n');
1124 :
1125 38 : appendStringInfo(&logbuf,
1126 38 : _("Process %d: %s"),
1127 : info->pid,
1128 : pgstat_get_backend_current_activity(info->pid, false));
1129 : }
1130 :
1131 12 : pgstat_report_deadlock();
1132 :
1133 12 : ereport(ERROR,
1134 : (errcode(ERRCODE_T_R_DEADLOCK_DETECTED),
1135 : errmsg("deadlock detected"),
1136 : errdetail_internal("%s", clientbuf.data),
1137 : errdetail_log("%s", logbuf.data),
1138 : errhint("See server log for query details.")));
1139 : }
1140 :
1141 : /*
1142 : * RememberSimpleDeadLock: set up info for DeadLockReport when ProcSleep
1143 : * detects a trivial (two-way) deadlock. proc1 wants to block for lockmode
1144 : * on lock, but proc2 is already waiting and would be blocked by proc1.
1145 : */
1146 : void
1147 2 : RememberSimpleDeadLock(PGPROC *proc1,
1148 : LOCKMODE lockmode,
1149 : LOCK *lock,
1150 : PGPROC *proc2)
1151 : {
1152 2 : DEADLOCK_INFO *info = &deadlockDetails[0];
1153 :
1154 2 : info->locktag = lock->tag;
1155 2 : info->lockmode = lockmode;
1156 2 : info->pid = proc1->pid;
1157 2 : info++;
1158 2 : info->locktag = proc2->waitLock->tag;
1159 2 : info->lockmode = proc2->waitLockMode;
1160 2 : info->pid = proc2->pid;
1161 2 : nDeadlockDetails = 2;
1162 2 : }
|
