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) 1996-2019, 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 "utils/memutils.h"
34
35
36/*
37 * One edge in the waits-for graph.
38 *
39 * waiter and blocker may or may not be members of a lock group, but if either
40 * is, it will be the leader rather than any other member of the lock group.
41 * The group leaders act as representatives of the whole group even though
42 * those particular processes need not be waiting at all. There will be at
43 * least one member of the waiter's lock group on the wait queue for the given
44 * lock, maybe more.
45 */
46typedef struct
47{
48 PGPROC *waiter; /* the leader of the waiting lock group */
49 PGPROC *blocker; /* the leader of the group it is waiting for */
50 LOCK *lock; /* the lock being waited for */
51 int pred; /* workspace for TopoSort */
52 int link; /* workspace for TopoSort */
53} EDGE;
54
55/* One potential reordering of a lock's wait queue */
56typedef struct
57{
58 LOCK *lock; /* the lock whose wait queue is described */
59 PGPROC **procs; /* array of PGPROC *'s in new wait order */
60 int nProcs;
61} WAIT_ORDER;
62
63/*
64 * Information saved about each edge in a detected deadlock cycle. This
65 * is used to print a diagnostic message upon failure.
66 *
67 * Note: because we want to examine this info after releasing the lock
68 * manager's partition locks, we can't just store LOCK and PGPROC pointers;
69 * we must extract out all the info we want to be able to print.
70 */
71typedef struct
72{
73 LOCKTAG locktag; /* ID of awaited lock object */
74 LOCKMODE lockmode; /* type of lock we're waiting for */
75 int pid; /* PID of blocked backend */
76} DEADLOCK_INFO;
77
78
79static bool DeadLockCheckRecurse(PGPROC *proc);
80static int TestConfiguration(PGPROC *startProc);
81static bool FindLockCycle(PGPROC *checkProc,
82 EDGE *softEdges, int *nSoftEdges);
83static bool FindLockCycleRecurse(PGPROC *checkProc, int depth,
84 EDGE *softEdges, int *nSoftEdges);
85static bool FindLockCycleRecurseMember(PGPROC *checkProc,
86 PGPROC *checkProcLeader,
87 int depth, EDGE *softEdges, int *nSoftEdges);
88static bool ExpandConstraints(EDGE *constraints, int nConstraints);
89static bool TopoSort(LOCK *lock, EDGE *constraints, int nConstraints,
90 PGPROC **ordering);
91
92#ifdef DEBUG_DEADLOCK
93static void PrintLockQueue(LOCK *lock, const char *info);
94#endif
95
96
97/*
98 * Working space for the deadlock detector
99 */
100
101/* Workspace for FindLockCycle */
102static PGPROC **visitedProcs; /* Array of visited procs */
103static int nVisitedProcs;
104
105/* Workspace for TopoSort */
106static PGPROC **topoProcs; /* Array of not-yet-output procs */
107static int *beforeConstraints; /* Counts of remaining before-constraints */
108static int *afterConstraints; /* List head for after-constraints */
109
110/* Output area for ExpandConstraints */
111static WAIT_ORDER *waitOrders; /* Array of proposed queue rearrangements */
112static int nWaitOrders;
113static PGPROC **waitOrderProcs; /* Space for waitOrders queue contents */
114
115/* Current list of constraints being considered */
116static EDGE *curConstraints;
117static int nCurConstraints;
118static int maxCurConstraints;
119
120/* Storage space for results from FindLockCycle */
121static EDGE *possibleConstraints;
122static int nPossibleConstraints;
123static int maxPossibleConstraints;
124static DEADLOCK_INFO *deadlockDetails;
125static int nDeadlockDetails;
126
127/* PGPROC pointer of any blocking autovacuum worker found */
128static PGPROC *blocking_autovacuum_proc = NULL;
129
130
131/*
132 * InitDeadLockChecking -- initialize deadlock checker during backend startup
133 *
134 * This does per-backend initialization of the deadlock checker; primarily,
135 * allocation of working memory for DeadLockCheck. We do this per-backend
136 * since there's no percentage in making the kernel do copy-on-write
137 * inheritance of workspace from the postmaster. We want to allocate the
138 * space at startup because (a) the deadlock checker might be invoked when
139 * there's no free memory left, and (b) the checker is normally run inside a
140 * signal handler, which is a very dangerous place to invoke palloc from.
141 */
142void
143InitDeadLockChecking(void)
144{
145 MemoryContext oldcxt;
146
147 /* Make sure allocations are permanent */
148 oldcxt = MemoryContextSwitchTo(TopMemoryContext);
149
150 /*
151 * FindLockCycle needs at most MaxBackends entries in visitedProcs[] and
152 * deadlockDetails[].
153 */
154 visitedProcs = (PGPROC **) palloc(MaxBackends * sizeof(PGPROC *));
155 deadlockDetails = (DEADLOCK_INFO *) palloc(MaxBackends * sizeof(DEADLOCK_INFO));
156
157 /*
158 * TopoSort needs to consider at most MaxBackends wait-queue entries, and
159 * it needn't run concurrently with FindLockCycle.
160 */
161 topoProcs = visitedProcs; /* re-use this space */
162 beforeConstraints = (int *) palloc(MaxBackends * sizeof(int));
163 afterConstraints = (int *) palloc(MaxBackends * sizeof(int));
164
165 /*
166 * We need to consider rearranging at most MaxBackends/2 wait queues
167 * (since it takes at least two waiters in a queue to create a soft edge),
168 * and the expanded form of the wait queues can't involve more than
169 * MaxBackends total waiters.
170 */
171 waitOrders = (WAIT_ORDER *)
172 palloc((MaxBackends / 2) * sizeof(WAIT_ORDER));
173 waitOrderProcs = (PGPROC **) palloc(MaxBackends * sizeof(PGPROC *));
174
175 /*
176 * Allow at most MaxBackends distinct constraints in a configuration. (Is
177 * this enough? In practice it seems it should be, but I don't quite see
178 * how to prove it. If we run out, we might fail to find a workable wait
179 * queue rearrangement even though one exists.) NOTE that this number
180 * limits the maximum recursion depth of DeadLockCheckRecurse. Making it
181 * really big might potentially allow a stack-overflow problem.
182 */
183 maxCurConstraints = MaxBackends;
184 curConstraints = (EDGE *) palloc(maxCurConstraints * sizeof(EDGE));
185
186 /*
187 * Allow up to 3*MaxBackends constraints to be saved without having to
188 * re-run TestConfiguration. (This is probably more than enough, but we
189 * can survive if we run low on space by doing excess runs of
190 * TestConfiguration to re-compute constraint lists each time needed.) The
191 * last MaxBackends entries in possibleConstraints[] are reserved as
192 * output workspace for FindLockCycle.
193 */
194 maxPossibleConstraints = MaxBackends * 4;
195 possibleConstraints =
196 (EDGE *) palloc(maxPossibleConstraints * sizeof(EDGE));
197
198 MemoryContextSwitchTo(oldcxt);
199}
200
201/*
202 * DeadLockCheck -- Checks for deadlocks for a given process
203 *
204 * This code looks for deadlocks involving the given process. If any
205 * are found, it tries to rearrange lock wait queues to resolve the
206 * deadlock. If resolution is impossible, return DS_HARD_DEADLOCK ---
207 * the caller is then expected to abort the given proc's transaction.
208 *
209 * Caller must already have locked all partitions of the lock tables.
210 *
211 * On failure, deadlock details are recorded in deadlockDetails[] for
212 * subsequent printing by DeadLockReport(). That activity is separate
213 * because (a) we don't want to do it while holding all those LWLocks,
214 * and (b) we are typically invoked inside a signal handler.
215 */
216DeadLockState
217DeadLockCheck(PGPROC *proc)
218{
219 int i,
220 j;
221
222 /* Initialize to "no constraints" */
223 nCurConstraints = 0;
224 nPossibleConstraints = 0;
225 nWaitOrders = 0;
226
227 /* Initialize to not blocked by an autovacuum worker */
228 blocking_autovacuum_proc = NULL;
229
230 /* Search for deadlocks and possible fixes */
231 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 nWaitOrders = 0;
242 if (!FindLockCycle(proc, possibleConstraints, &nSoftEdges))
243 elog(FATAL, "deadlock seems to have disappeared");
244
245 return DS_HARD_DEADLOCK; /* cannot find a non-deadlocked state */
246 }
247
248 /* Apply any needed rearrangements of wait queues */
249 for (i = 0; i < nWaitOrders; i++)
250 {
251 LOCK *lock = waitOrders[i].lock;
252 PGPROC **procs = waitOrders[i].procs;
253 int nProcs = waitOrders[i].nProcs;
254 PROC_QUEUE *waitQueue = &(lock->waitProcs);
255
256 Assert(nProcs == waitQueue->size);
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 ProcQueueInit(waitQueue);
264 for (j = 0; j < nProcs; j++)
265 {
266 SHMQueueInsertBefore(&(waitQueue->links), &(procs[j]->links));
267 waitQueue->size++;
268 }
269
270#ifdef DEBUG_DEADLOCK
271 PrintLockQueue(lock, "rearranged to:");
272#endif
273
274 /* See if any waiters for the lock can be woken up now */
275 ProcLockWakeup(GetLocksMethodTable(lock), lock);
276 }
277
278 /* Return code tells caller if we had to escape a deadlock or not */
279 if (nWaitOrders > 0)
280 return DS_SOFT_DEADLOCK;
281 else if (blocking_autovacuum_proc != NULL)
282 return DS_BLOCKED_BY_AUTOVACUUM;
283 else
284 return DS_NO_DEADLOCK;
285}
286
287/*
288 * Return the PGPROC of the autovacuum that's blocking a process.
289 *
290 * We reset the saved pointer as soon as we pass it back.
291 */
292PGPROC *
293GetBlockingAutoVacuumPgproc(void)
294{
295 PGPROC *ptr;
296
297 ptr = blocking_autovacuum_proc;
298 blocking_autovacuum_proc = NULL;
299
300 return ptr;
301}
302
303/*
304 * DeadLockCheckRecurse -- recursively search for valid orderings
305 *
306 * curConstraints[] holds the current set of constraints being considered
307 * by an outer level of recursion. Add to this each possible solution
308 * constraint for any cycle detected at this level.
309 *
310 * Returns true if no solution exists. Returns false if a deadlock-free
311 * state is attainable, in which case waitOrders[] shows the required
312 * rearrangements of lock wait queues (if any).
313 */
314static bool
315DeadLockCheckRecurse(PGPROC *proc)
316{
317 int nEdges;
318 int oldPossibleConstraints;
319 bool savedList;
320 int i;
321
322 nEdges = TestConfiguration(proc);
323 if (nEdges < 0)
324 return true; /* hard deadlock --- no solution */
325 if (nEdges == 0)
326 return false; /* good configuration found */
327 if (nCurConstraints >= maxCurConstraints)
328 return true; /* out of room for active constraints? */
329 oldPossibleConstraints = nPossibleConstraints;
330 if (nPossibleConstraints + nEdges + MaxBackends <= maxPossibleConstraints)
331 {
332 /* We can save the edge list in possibleConstraints[] */
333 nPossibleConstraints += nEdges;
334 savedList = true;
335 }
336 else
337 {
338 /* Not room; will need to regenerate the edges on-the-fly */
339 savedList = false;
340 }
341
342 /*
343 * Try each available soft edge as an addition to the configuration.
344 */
345 for (i = 0; i < nEdges; i++)
346 {
347 if (!savedList && i > 0)
348 {
349 /* Regenerate the list of possible added constraints */
350 if (nEdges != TestConfiguration(proc))
351 elog(FATAL, "inconsistent results during deadlock check");
352 }
353 curConstraints[nCurConstraints] =
354 possibleConstraints[oldPossibleConstraints + i];
355 nCurConstraints++;
356 if (!DeadLockCheckRecurse(proc))
357 return false; /* found a valid solution! */
358 /* give up on that added constraint, try again */
359 nCurConstraints--;
360 }
361 nPossibleConstraints = oldPossibleConstraints;
362 return true; /* no solution found */
363}
364
365
366/*--------------------
367 * Test a configuration (current set of constraints) for validity.
368 *
369 * Returns:
370 * 0: the configuration is good (no deadlocks)
371 * -1: the configuration has a hard deadlock or is not self-consistent
372 * >0: the configuration has one or more soft deadlocks
373 *
374 * In the soft-deadlock case, one of the soft cycles is chosen arbitrarily
375 * and a list of its soft edges is returned beginning at
376 * possibleConstraints+nPossibleConstraints. The return value is the
377 * number of soft edges.
378 *--------------------
379 */
380static int
381TestConfiguration(PGPROC *startProc)
382{
383 int softFound = 0;
384 EDGE *softEdges = possibleConstraints + nPossibleConstraints;
385 int nSoftEdges;
386 int i;
387
388 /*
389 * Make sure we have room for FindLockCycle's output.
390 */
391 if (nPossibleConstraints + MaxBackends > maxPossibleConstraints)
392 return -1;
393
394 /*
395 * Expand current constraint set into wait orderings. Fail if the
396 * constraint set is not self-consistent.
397 */
398 if (!ExpandConstraints(curConstraints, nCurConstraints))
399 return -1;
400
401 /*
402 * Check for cycles involving startProc or any of the procs mentioned in
403 * constraints. We check startProc last because if it has a soft cycle
404 * still to be dealt with, we want to deal with that first.
405 */
406 for (i = 0; i < nCurConstraints; i++)
407 {
408 if (FindLockCycle(curConstraints[i].waiter, softEdges, &nSoftEdges))
409 {
410 if (nSoftEdges == 0)
411 return -1; /* hard deadlock detected */
412 softFound = nSoftEdges;
413 }
414 if (FindLockCycle(curConstraints[i].blocker, softEdges, &nSoftEdges))
415 {
416 if (nSoftEdges == 0)
417 return -1; /* hard deadlock detected */
418 softFound = nSoftEdges;
419 }
420 }
421 if (FindLockCycle(startProc, softEdges, &nSoftEdges))
422 {
423 if (nSoftEdges == 0)
424 return -1; /* hard deadlock detected */
425 softFound = nSoftEdges;
426 }
427 return softFound;
428}
429
430
431/*
432 * FindLockCycle -- basic check for deadlock cycles
433 *
434 * Scan outward from the given proc to see if there is a cycle in the
435 * waits-for graph that includes this proc. Return true if a cycle
436 * is found, else false. If a cycle is found, we return a list of
437 * the "soft edges", if any, included in the cycle. These edges could
438 * potentially be eliminated by rearranging wait queues. We also fill
439 * deadlockDetails[] with information about the detected cycle; this info
440 * is not used by the deadlock algorithm itself, only to print a useful
441 * message after failing.
442 *
443 * Since we need to be able to check hypothetical configurations that would
444 * exist after wait queue rearrangement, the routine pays attention to the
445 * table of hypothetical queue orders in waitOrders[]. These orders will
446 * be believed in preference to the actual ordering seen in the locktable.
447 */
448static bool
449FindLockCycle(PGPROC *checkProc,
450 EDGE *softEdges, /* output argument */
451 int *nSoftEdges) /* output argument */
452{
453 nVisitedProcs = 0;
454 nDeadlockDetails = 0;
455 *nSoftEdges = 0;
456 return FindLockCycleRecurse(checkProc, 0, softEdges, nSoftEdges);
457}
458
459static bool
460FindLockCycleRecurse(PGPROC *checkProc,
461 int depth,
462 EDGE *softEdges, /* output argument */
463 int *nSoftEdges) /* output argument */
464{
465 int i;
466 dlist_iter iter;
467
468 /*
469 * If this process is a lock group member, check the leader instead. (Note
470 * that we might be the leader, in which case this is a no-op.)
471 */
472 if (checkProc->lockGroupLeader != NULL)
473 checkProc = checkProc->lockGroupLeader;
474
475 /*
476 * Have we already seen this proc?
477 */
478 for (i = 0; i < nVisitedProcs; i++)
479 {
480 if (visitedProcs[i] == checkProc)
481 {
482 /* If we return to starting point, we have a deadlock cycle */
483 if (i == 0)
484 {
485 /*
486 * record total length of cycle --- outer levels will now fill
487 * deadlockDetails[]
488 */
489 Assert(depth <= MaxBackends);
490 nDeadlockDetails = depth;
491
492 return true;
493 }
494
495 /*
496 * Otherwise, we have a cycle but it does not include the start
497 * point, so say "no deadlock".
498 */
499 return false;
500 }
501 }
502 /* Mark proc as seen */
503 Assert(nVisitedProcs < MaxBackends);
504 visitedProcs[nVisitedProcs++] = checkProc;
505
506 /*
507 * If the process is waiting, there is an outgoing waits-for edge to each
508 * process that blocks it.
509 */
510 if (checkProc->links.next != NULL && checkProc->waitLock != NULL &&
511 FindLockCycleRecurseMember(checkProc, checkProc, depth, softEdges,
512 nSoftEdges))
513 return true;
514
515 /*
516 * If the process is not waiting, there could still be outgoing waits-for
517 * edges if it is part of a lock group, because other members of the lock
518 * group might be waiting even though this process is not. (Given lock
519 * groups {A1, A2} and {B1, B2}, if A1 waits for B1 and B2 waits for A2,
520 * that is a deadlock even neither of B1 and A2 are waiting for anything.)
521 */
522 dlist_foreach(iter, &checkProc->lockGroupMembers)
523 {
524 PGPROC *memberProc;
525
526 memberProc = dlist_container(PGPROC, lockGroupLink, iter.cur);
527
528 if (memberProc->links.next != NULL && memberProc->waitLock != NULL &&
529 memberProc != checkProc &&
530 FindLockCycleRecurseMember(memberProc, checkProc, depth, softEdges,
531 nSoftEdges))
532 return true;
533 }
534
535 return false;
536}
537
538static bool
539FindLockCycleRecurseMember(PGPROC *checkProc,
540 PGPROC *checkProcLeader,
541 int depth,
542 EDGE *softEdges, /* output argument */
543 int *nSoftEdges) /* output argument */
544{
545 PGPROC *proc;
546 LOCK *lock = checkProc->waitLock;
547 PGXACT *pgxact;
548 PROCLOCK *proclock;
549 SHM_QUEUE *procLocks;
550 LockMethod lockMethodTable;
551 PROC_QUEUE *waitQueue;
552 int queue_size;
553 int conflictMask;
554 int i;
555 int numLockModes,
556 lm;
557
558 lockMethodTable = GetLocksMethodTable(lock);
559 numLockModes = lockMethodTable->numLockModes;
560 conflictMask = lockMethodTable->conflictTab[checkProc->waitLockMode];
561
562 /*
563 * Scan for procs that already hold conflicting locks. These are "hard"
564 * edges in the waits-for graph.
565 */
566 procLocks = &(lock->procLocks);
567
568 proclock = (PROCLOCK *) SHMQueueNext(procLocks, procLocks,
569 offsetof(PROCLOCK, lockLink));
570
571 while (proclock)
572 {
573 PGPROC *leader;
574
575 proc = proclock->tag.myProc;
576 pgxact = &ProcGlobal->allPgXact[proc->pgprocno];
577 leader = proc->lockGroupLeader == NULL ? proc : proc->lockGroupLeader;
578
579 /* A proc never blocks itself or any other lock group member */
580 if (leader != checkProcLeader)
581 {
582 for (lm = 1; lm <= numLockModes; lm++)
583 {
584 if ((proclock->holdMask & LOCKBIT_ON(lm)) &&
585 (conflictMask & LOCKBIT_ON(lm)))
586 {
587 /* This proc hard-blocks checkProc */
588 if (FindLockCycleRecurse(proc, depth + 1,
589 softEdges, nSoftEdges))
590 {
591 /* fill deadlockDetails[] */
592 DEADLOCK_INFO *info = &deadlockDetails[depth];
593
594 info->locktag = lock->tag;
595 info->lockmode = checkProc->waitLockMode;
596 info->pid = checkProc->pid;
597
598 return true;
599 }
600
601 /*
602 * No deadlock here, but see if this proc is an autovacuum
603 * that is directly hard-blocking our own proc. If so,
604 * report it so that the caller can send a cancel signal
605 * to it, if appropriate. If there's more than one such
606 * proc, it's indeterminate which one will be reported.
607 *
608 * We don't touch autovacuums that are indirectly blocking
609 * us; it's up to the direct blockee to take action. This
610 * rule simplifies understanding the behavior and ensures
611 * that an autovacuum won't be canceled with less than
612 * deadlock_timeout grace period.
613 *
614 * Note we read vacuumFlags without any locking. This is
615 * OK only for checking the PROC_IS_AUTOVACUUM flag,
616 * because that flag is set at process start and never
617 * reset. There is logic elsewhere to avoid canceling an
618 * autovacuum that is working to prevent XID wraparound
619 * problems (which needs to read a different vacuumFlag
620 * bit), but we don't do that here to avoid grabbing
621 * ProcArrayLock.
622 */
623 if (checkProc == MyProc &&
624 pgxact->vacuumFlags & PROC_IS_AUTOVACUUM)
625 blocking_autovacuum_proc = proc;
626
627 /* We're done looking at this proclock */
628 break;
629 }
630 }
631 }
632
633 proclock = (PROCLOCK *) SHMQueueNext(procLocks, &proclock->lockLink,
634 offsetof(PROCLOCK, lockLink));
635 }
636
637 /*
638 * Scan for procs that are ahead of this one in the lock's wait queue.
639 * Those that have conflicting requests soft-block this one. This must be
640 * done after the hard-block search, since if another proc both hard- and
641 * soft-blocks this one, we want to call it a hard edge.
642 *
643 * If there is a proposed re-ordering of the lock's wait order, use that
644 * rather than the current wait order.
645 */
646 for (i = 0; i < nWaitOrders; i++)
647 {
648 if (waitOrders[i].lock == lock)
649 break;
650 }
651
652 if (i < nWaitOrders)
653 {
654 /* Use the given hypothetical wait queue order */
655 PGPROC **procs = waitOrders[i].procs;
656
657 queue_size = waitOrders[i].nProcs;
658
659 for (i = 0; i < queue_size; i++)
660 {
661 PGPROC *leader;
662
663 proc = procs[i];
664 leader = proc->lockGroupLeader == NULL ? proc :
665 proc->lockGroupLeader;
666
667 /*
668 * TopoSort will always return an ordering with group members
669 * adjacent to each other in the wait queue (see comments
670 * therein). So, as soon as we reach a process in the same lock
671 * group as checkProc, we know we've found all the conflicts that
672 * precede any member of the lock group lead by checkProcLeader.
673 */
674 if (leader == checkProcLeader)
675 break;
676
677 /* Is there a conflict with this guy's request? */
678 if ((LOCKBIT_ON(proc->waitLockMode) & conflictMask) != 0)
679 {
680 /* This proc soft-blocks checkProc */
681 if (FindLockCycleRecurse(proc, depth + 1,
682 softEdges, nSoftEdges))
683 {
684 /* fill deadlockDetails[] */
685 DEADLOCK_INFO *info = &deadlockDetails[depth];
686
687 info->locktag = lock->tag;
688 info->lockmode = checkProc->waitLockMode;
689 info->pid = checkProc->pid;
690
691 /*
692 * Add this edge to the list of soft edges in the cycle
693 */
694 Assert(*nSoftEdges < MaxBackends);
695 softEdges[*nSoftEdges].waiter = checkProcLeader;
696 softEdges[*nSoftEdges].blocker = leader;
697 softEdges[*nSoftEdges].lock = lock;
698 (*nSoftEdges)++;
699 return true;
700 }
701 }
702 }
703 }
704 else
705 {
706 PGPROC *lastGroupMember = NULL;
707
708 /* Use the true lock wait queue order */
709 waitQueue = &(lock->waitProcs);
710
711 /*
712 * Find the last member of the lock group that is present in the wait
713 * queue. Anything after this is not a soft lock conflict. If group
714 * locking is not in use, then we know immediately which process we're
715 * looking for, but otherwise we've got to search the wait queue to
716 * find the last process actually present.
717 */
718 if (checkProc->lockGroupLeader == NULL)
719 lastGroupMember = checkProc;
720 else
721 {
722 proc = (PGPROC *) waitQueue->links.next;
723 queue_size = waitQueue->size;
724 while (queue_size-- > 0)
725 {
726 if (proc->lockGroupLeader == checkProcLeader)
727 lastGroupMember = proc;
728 proc = (PGPROC *) proc->links.next;
729 }
730 Assert(lastGroupMember != NULL);
731 }
732
733 /*
734 * OK, now rescan (or scan) the queue to identify the soft conflicts.
735 */
736 queue_size = waitQueue->size;
737 proc = (PGPROC *) waitQueue->links.next;
738 while (queue_size-- > 0)
739 {
740 PGPROC *leader;
741
742 leader = proc->lockGroupLeader == NULL ? proc :
743 proc->lockGroupLeader;
744
745 /* Done when we reach the target proc */
746 if (proc == lastGroupMember)
747 break;
748
749 /* Is there a conflict with this guy's request? */
750 if ((LOCKBIT_ON(proc->waitLockMode) & conflictMask) != 0 &&
751 leader != checkProcLeader)
752 {
753 /* This proc soft-blocks checkProc */
754 if (FindLockCycleRecurse(proc, depth + 1,
755 softEdges, nSoftEdges))
756 {
757 /* fill deadlockDetails[] */
758 DEADLOCK_INFO *info = &deadlockDetails[depth];
759
760 info->locktag = lock->tag;
761 info->lockmode = checkProc->waitLockMode;
762 info->pid = checkProc->pid;
763
764 /*
765 * Add this edge to the list of soft edges in the cycle
766 */
767 Assert(*nSoftEdges < MaxBackends);
768 softEdges[*nSoftEdges].waiter = checkProcLeader;
769 softEdges[*nSoftEdges].blocker = leader;
770 softEdges[*nSoftEdges].lock = lock;
771 (*nSoftEdges)++;
772 return true;
773 }
774 }
775
776 proc = (PGPROC *) proc->links.next;
777 }
778 }
779
780 /*
781 * No conflict detected here.
782 */
783 return false;
784}
785
786
787/*
788 * ExpandConstraints -- expand a list of constraints into a set of
789 * specific new orderings for affected wait queues
790 *
791 * Input is a list of soft edges to be reversed. The output is a list
792 * of nWaitOrders WAIT_ORDER structs in waitOrders[], with PGPROC array
793 * workspace in waitOrderProcs[].
794 *
795 * Returns true if able to build an ordering that satisfies all the
796 * constraints, false if not (there are contradictory constraints).
797 */
798static bool
799ExpandConstraints(EDGE *constraints,
800 int nConstraints)
801{
802 int nWaitOrderProcs = 0;
803 int i,
804 j;
805
806 nWaitOrders = 0;
807
808 /*
809 * Scan constraint list backwards. This is because the last-added
810 * constraint is the only one that could fail, and so we want to test it
811 * for inconsistency first.
812 */
813 for (i = nConstraints; --i >= 0;)
814 {
815 LOCK *lock = constraints[i].lock;
816
817 /* Did we already make a list for this lock? */
818 for (j = nWaitOrders; --j >= 0;)
819 {
820 if (waitOrders[j].lock == lock)
821 break;
822 }
823 if (j >= 0)
824 continue;
825 /* No, so allocate a new list */
826 waitOrders[nWaitOrders].lock = lock;
827 waitOrders[nWaitOrders].procs = waitOrderProcs + nWaitOrderProcs;
828 waitOrders[nWaitOrders].nProcs = lock->waitProcs.size;
829 nWaitOrderProcs += lock->waitProcs.size;
830 Assert(nWaitOrderProcs <= MaxBackends);
831
832 /*
833 * Do the topo sort. TopoSort need not examine constraints after this
834 * one, since they must be for different locks.
835 */
836 if (!TopoSort(lock, constraints, i + 1,
837 waitOrders[nWaitOrders].procs))
838 return false;
839 nWaitOrders++;
840 }
841 return true;
842}
843
844
845/*
846 * TopoSort -- topological sort of a wait queue
847 *
848 * Generate a re-ordering of a lock's wait queue that satisfies given
849 * constraints about certain procs preceding others. (Each such constraint
850 * is a fact of a partial ordering.) Minimize rearrangement of the queue
851 * not needed to achieve the partial ordering.
852 *
853 * This is a lot simpler and slower than, for example, the topological sort
854 * algorithm shown in Knuth's Volume 1. However, Knuth's method doesn't
855 * try to minimize the damage to the existing order. In practice we are
856 * not likely to be working with more than a few constraints, so the apparent
857 * slowness of the algorithm won't really matter.
858 *
859 * The initial queue ordering is taken directly from the lock's wait queue.
860 * The output is an array of PGPROC pointers, of length equal to the lock's
861 * wait queue length (the caller is responsible for providing this space).
862 * The partial order is specified by an array of EDGE structs. Each EDGE
863 * is one that we need to reverse, therefore the "waiter" must appear before
864 * the "blocker" in the output array. The EDGE array may well contain
865 * edges associated with other locks; these should be ignored.
866 *
867 * Returns true if able to build an ordering that satisfies all the
868 * constraints, false if not (there are contradictory constraints).
869 */
870static bool
871TopoSort(LOCK *lock,
872 EDGE *constraints,
873 int nConstraints,
874 PGPROC **ordering) /* output argument */
875{
876 PROC_QUEUE *waitQueue = &(lock->waitProcs);
877 int queue_size = waitQueue->size;
878 PGPROC *proc;
879 int i,
880 j,
881 jj,
882 k,
883 kk,
884 last;
885
886 /* First, fill topoProcs[] array with the procs in their current order */
887 proc = (PGPROC *) waitQueue->links.next;
888 for (i = 0; i < queue_size; i++)
889 {
890 topoProcs[i] = proc;
891 proc = (PGPROC *) proc->links.next;
892 }
893
894 /*
895 * Scan the constraints, and for each proc in the array, generate a count
896 * of the number of constraints that say it must be before something else,
897 * plus a list of the constraints that say it must be after something
898 * else. The count for the j'th proc is stored in beforeConstraints[j],
899 * and the head of its list in afterConstraints[j]. Each constraint
900 * stores its list link in constraints[i].link (note any constraint will
901 * be in just one list). The array index for the before-proc of the i'th
902 * constraint is remembered in constraints[i].pred.
903 *
904 * Note that it's not necessarily the case that every constraint affects
905 * this particular wait queue. Prior to group locking, a process could be
906 * waiting for at most one lock. But a lock group can be waiting for
907 * zero, one, or multiple locks. Since topoProcs[] is an array of the
908 * processes actually waiting, while constraints[] is an array of group
909 * leaders, we've got to scan through topoProcs[] for each constraint,
910 * checking whether both a waiter and a blocker for that group are
911 * present. If so, the constraint is relevant to this wait queue; if not,
912 * it isn't.
913 */
914 MemSet(beforeConstraints, 0, queue_size * sizeof(int));
915 MemSet(afterConstraints, 0, queue_size * sizeof(int));
916 for (i = 0; i < nConstraints; i++)
917 {
918 /*
919 * Find a representative process that is on the lock queue and part of
920 * the waiting lock group. This may or may not be the leader, which
921 * may or may not be waiting at all. If there are any other processes
922 * in the same lock group on the queue, set their number of
923 * beforeConstraints to -1 to indicate that they should be emitted
924 * with their groupmates rather than considered separately.
925 *
926 * In this loop and the similar one just below, it's critical that we
927 * consistently select the same representative member of any one lock
928 * group, so that all the constraints are associated with the same
929 * proc, and the -1's are only associated with not-representative
930 * members. We select the last one in the topoProcs array.
931 */
932 proc = constraints[i].waiter;
933 Assert(proc != NULL);
934 jj = -1;
935 for (j = queue_size; --j >= 0;)
936 {
937 PGPROC *waiter = topoProcs[j];
938
939 if (waiter == proc || waiter->lockGroupLeader == proc)
940 {
941 Assert(waiter->waitLock == lock);
942 if (jj == -1)
943 jj = j;
944 else
945 {
946 Assert(beforeConstraints[j] <= 0);
947 beforeConstraints[j] = -1;
948 }
949 }
950 }
951
952 /* If no matching waiter, constraint is not relevant to this lock. */
953 if (jj < 0)
954 continue;
955
956 /*
957 * Similarly, find a representative process that is on the lock queue
958 * and waiting for the blocking lock group. Again, this could be the
959 * leader but does not need to be.
960 */
961 proc = constraints[i].blocker;
962 Assert(proc != NULL);
963 kk = -1;
964 for (k = queue_size; --k >= 0;)
965 {
966 PGPROC *blocker = topoProcs[k];
967
968 if (blocker == proc || blocker->lockGroupLeader == proc)
969 {
970 Assert(blocker->waitLock == lock);
971 if (kk == -1)
972 kk = k;
973 else
974 {
975 Assert(beforeConstraints[k] <= 0);
976 beforeConstraints[k] = -1;
977 }
978 }
979 }
980
981 /* If no matching blocker, constraint is not relevant to this lock. */
982 if (kk < 0)
983 continue;
984
985 Assert(beforeConstraints[jj] >= 0);
986 beforeConstraints[jj]++; /* waiter must come before */
987 /* add this constraint to list of after-constraints for blocker */
988 constraints[i].pred = jj;
989 constraints[i].link = afterConstraints[kk];
990 afterConstraints[kk] = i + 1;
991 }
992
993 /*--------------------
994 * Now scan the topoProcs array backwards. At each step, output the
995 * last proc that has no remaining before-constraints plus any other
996 * members of the same lock group; then decrease the beforeConstraints
997 * count of each of the procs it was constrained against.
998 * i = index of ordering[] entry we want to output this time
999 * j = search index for topoProcs[]
1000 * k = temp for scanning constraint list for proc j
1001 * last = last non-null index in topoProcs (avoid redundant searches)
1002 *--------------------
1003 */
1004 last = queue_size - 1;
1005 for (i = queue_size - 1; i >= 0;)
1006 {
1007 int c;
1008 int nmatches = 0;
1009
1010 /* Find next candidate to output */
1011 while (topoProcs[last] == NULL)
1012 last--;
1013 for (j = last; j >= 0; j--)
1014 {
1015 if (topoProcs[j] != NULL && beforeConstraints[j] == 0)
1016 break;
1017 }
1018
1019 /* If no available candidate, topological sort fails */
1020 if (j < 0)
1021 return false;
1022
1023 /*
1024 * Output everything in the lock group. There's no point in
1025 * outputting an ordering where members of the same lock group are not
1026 * consecutive on the wait queue: if some other waiter is between two
1027 * requests that belong to the same group, then either it conflicts
1028 * with both of them and is certainly not a solution; or it conflicts
1029 * with at most one of them and is thus isomorphic to an ordering
1030 * where the group members are consecutive.
1031 */
1032 proc = topoProcs[j];
1033 if (proc->lockGroupLeader != NULL)
1034 proc = proc->lockGroupLeader;
1035 Assert(proc != NULL);
1036 for (c = 0; c <= last; ++c)
1037 {
1038 if (topoProcs[c] == proc || (topoProcs[c] != NULL &&
1039 topoProcs[c]->lockGroupLeader == proc))
1040 {
1041 ordering[i - nmatches] = topoProcs[c];
1042 topoProcs[c] = NULL;
1043 ++nmatches;
1044 }
1045 }
1046 Assert(nmatches > 0);
1047 i -= nmatches;
1048
1049 /* Update beforeConstraints counts of its predecessors */
1050 for (k = afterConstraints[j]; k > 0; k = constraints[k - 1].link)
1051 beforeConstraints[constraints[k - 1].pred]--;
1052 }
1053
1054 /* Done */
1055 return true;
1056}
1057
1058#ifdef DEBUG_DEADLOCK
1059static void
1060PrintLockQueue(LOCK *lock, const char *info)
1061{
1062 PROC_QUEUE *waitQueue = &(lock->waitProcs);
1063 int queue_size = waitQueue->size;
1064 PGPROC *proc;
1065 int i;
1066
1067 printf("%s lock %p queue ", info, lock);
1068 proc = (PGPROC *) waitQueue->links.next;
1069 for (i = 0; i < queue_size; i++)
1070 {
1071 printf(" %d", proc->pid);
1072 proc = (PGPROC *) proc->links.next;
1073 }
1074 printf("\n");
1075 fflush(stdout);
1076}
1077#endif
1078
1079/*
1080 * Report a detected deadlock, with available details.
1081 */
1082void
1083DeadLockReport(void)
1084{
1085 StringInfoData clientbuf; /* errdetail for client */
1086 StringInfoData logbuf; /* errdetail for server log */
1087 StringInfoData locktagbuf;
1088 int i;
1089
1090 initStringInfo(&clientbuf);
1091 initStringInfo(&logbuf);
1092 initStringInfo(&locktagbuf);
1093
1094 /* Generate the "waits for" lines sent to the client */
1095 for (i = 0; i < nDeadlockDetails; i++)
1096 {
1097 DEADLOCK_INFO *info = &deadlockDetails[i];
1098 int nextpid;
1099
1100 /* The last proc waits for the first one... */
1101 if (i < nDeadlockDetails - 1)
1102 nextpid = info[1].pid;
1103 else
1104 nextpid = deadlockDetails[0].pid;
1105
1106 /* reset locktagbuf to hold next object description */
1107 resetStringInfo(&locktagbuf);
1108
1109 DescribeLockTag(&locktagbuf, &info->locktag);
1110
1111 if (i > 0)
1112 appendStringInfoChar(&clientbuf, '\n');
1113
1114 appendStringInfo(&clientbuf,
1115 _("Process %d waits for %s on %s; blocked by process %d."),
1116 info->pid,
1117 GetLockmodeName(info->locktag.locktag_lockmethodid,
1118 info->lockmode),
1119 locktagbuf.data,
1120 nextpid);
1121 }
1122
1123 /* Duplicate all the above for the server ... */
1124 appendStringInfoString(&logbuf, clientbuf.data);
1125
1126 /* ... and add info about query strings */
1127 for (i = 0; i < nDeadlockDetails; i++)
1128 {
1129 DEADLOCK_INFO *info = &deadlockDetails[i];
1130
1131 appendStringInfoChar(&logbuf, '\n');
1132
1133 appendStringInfo(&logbuf,
1134 _("Process %d: %s"),
1135 info->pid,
1136 pgstat_get_backend_current_activity(info->pid, false));
1137 }
1138
1139 pgstat_report_deadlock();
1140
1141 ereport(ERROR,
1142 (errcode(ERRCODE_T_R_DEADLOCK_DETECTED),
1143 errmsg("deadlock detected"),
1144 errdetail_internal("%s", clientbuf.data),
1145 errdetail_log("%s", logbuf.data),
1146 errhint("See server log for query details.")));
1147}
1148
1149/*
1150 * RememberSimpleDeadLock: set up info for DeadLockReport when ProcSleep
1151 * detects a trivial (two-way) deadlock. proc1 wants to block for lockmode
1152 * on lock, but proc2 is already waiting and would be blocked by proc1.
1153 */
1154void
1155RememberSimpleDeadLock(PGPROC *proc1,
1156 LOCKMODE lockmode,
1157 LOCK *lock,
1158 PGPROC *proc2)
1159{
1160 DEADLOCK_INFO *info = &deadlockDetails[0];
1161
1162 info->locktag = lock->tag;
1163 info->lockmode = lockmode;
1164 info->pid = proc1->pid;
1165 info++;
1166 info->locktag = proc2->waitLock->tag;
1167 info->lockmode = proc2->waitLockMode;
1168 info->pid = proc2->pid;
1169 nDeadlockDetails = 2;
1170}
1171