1 /*-------------------------------------------------------------------------
4 * routines to manage per-process shared memory data structure
6 * Portions Copyright (c) 1996-2000, PostgreSQL, Inc
7 * Portions Copyright (c) 1994, Regents of the University of California
11 * $Header: /cvsroot/pgsql/src/backend/storage/lmgr/proc.c,v 1.67 2000/01/26 05:57:02 momjian Exp $
13 *-------------------------------------------------------------------------
16 * Each postgres backend gets one of these. We'll use it to
17 * clean up after the process should the process suddenly die.
21 * ProcSleep(), ProcWakeup(), ProcWakeupNext(),
22 * ProcQueueAlloc() -- create a shm queue for sleeping processes
23 * ProcQueueInit() -- create a queue without allocing memory
25 * Locking and waiting for buffers can cause the backend to be
26 * put to sleep. Whoever releases the lock, etc. wakes the
27 * process up again (and gives it an error code so it knows
28 * whether it was awoken on an error condition).
32 * ProcReleaseLocks -- frees the locks associated with this process,
33 * ProcKill -- destroys the shared memory state (and locks)
34 * associated with the process.
36 * 5/15/91 -- removed the buffer pool based lock chain in favor
37 * of a shared memory lock chain. The write-protection is
38 * more expensive if the lock chain is in the buffer pool.
39 * The only reason I kept the lock chain in the buffer pool
40 * in the first place was to allow the lock table to grow larger
41 * than available shared memory and that isn't going to work
42 * without a lot of unimplemented support anyway.
44 * 4/7/95 -- instead of allocating a set of 1 semaphore per process, we
45 * allocate a semaphore from a set of PROC_NSEMS_PER_SET semaphores
46 * shared among backends (we keep a few sets of semaphores around).
47 * This is so that we can support more backends. (system-wide semaphore
48 * sets run out pretty fast.) -ay 4/95
50 * $Header: /cvsroot/pgsql/src/backend/storage/lmgr/proc.c,v 1.67 2000/01/26 05:57:02 momjian Exp $
55 #include <sys/types.h>
57 #if defined(solaris_sparc) || defined(__CYGWIN__)
63 #include "miscadmin.h"
64 #include "libpq/pqsignal.h"
67 #include "storage/ipc.h"
68 /* In Ultrix and QNX, sem.h must be included after ipc.h */
71 #include "storage/lmgr.h"
72 #include "storage/proc.h"
73 #include "utils/trace.h"
75 void HandleDeadLock(SIGNAL_ARGS);
76 static void ProcFreeAllSemaphores(void);
78 #define DeadlockCheckTimer pg_options[OPT_DEADLOCKTIMEOUT]
80 /* --------------------
81 * Spin lock for manipulating the shared process data structure:
82 * ProcGlobal.... Adding an extra spin lock seemed like the smallest
83 * hack to get around reading and updating this structure in shared
84 * memory. -mer 17 July 1991
85 * --------------------
87 SPINLOCK ProcStructLock;
89 static PROC_HDR *ProcGlobal = NULL;
93 static void ProcKill(int exitStatus, int pid);
94 static void ProcGetNewSemKeyAndNum(IPCKey *key, int *semNum);
95 static void ProcFreeSem(IpcSemaphoreKey semKey, int semNum);
97 static char *DeadLockMessage = "Deadlock detected -- See the lock(l) manual page for a possible cause.";
101 * initializes the global process table. We put it here so that
102 * the postmaster can do this initialization. (ProcFreeAllSemaphores needs
103 * to read this table on exiting the postmaster. If we have the first
104 * backend do this, starting up and killing the postmaster without
105 * starting any backends will be a problem.)
107 * We also allocate all the per-process semaphores we will need to support
108 * the requested number of backends. We used to allocate semaphores
109 * only when backends were actually started up, but that is bad because
110 * it lets Postgres fail under load --- a lot of Unix systems are
111 * (mis)configured with small limits on the number of semaphores, and
112 * running out when trying to start another backend is a common failure.
113 * So, now we grab enough semaphores to support the desired max number
114 * of backends immediately at initialization --- if the sysadmin has set
115 * MaxBackends higher than his kernel will support, he'll find out sooner
119 InitProcGlobal(IPCKey key, int maxBackends)
123 /* attach to the free list */
124 ProcGlobal = (PROC_HDR *)
125 ShmemInitStruct("Proc Header", (unsigned) sizeof(PROC_HDR), &found);
127 /* --------------------
128 * We're the first - initialize.
129 * XXX if found should ever be true, it is a sign of impending doom ...
130 * ought to complain if so?
131 * --------------------
137 ProcGlobal->freeProcs = INVALID_OFFSET;
138 ProcGlobal->currKey = IPCGetProcessSemaphoreInitKey(key);
139 for (i = 0; i < MAX_PROC_SEMS / PROC_NSEMS_PER_SET; i++)
140 ProcGlobal->freeSemMap[i] = 0;
143 * Arrange to delete semas on exit --- set this up now so that we
144 * will clean up if pre-allocation fails...
146 on_shmem_exit(ProcFreeAllSemaphores, NULL);
149 * Pre-create the semaphores for the first maxBackends processes,
150 * unless we are running as a standalone backend.
152 if (key != PrivateIPCKey)
155 i < (maxBackends + PROC_NSEMS_PER_SET - 1) / PROC_NSEMS_PER_SET;
158 IPCKey semKey = ProcGlobal->currKey + i;
161 semId = IpcSemaphoreCreate(semKey,
164 IpcSemaphoreDefaultStartValue,
167 elog(FATAL, "InitProcGlobal: IpcSemaphoreCreate failed");
168 /* mark this sema set allocated */
169 ProcGlobal->freeSemMap[i] = (1 << PROC_NSEMS_PER_SET);
175 /* ------------------------
176 * InitProc -- create a per-process data structure for this process
177 * used by the lock manager on semaphore queues.
178 * ------------------------
181 InitProcess(IPCKey key)
184 unsigned long location,
187 SpinAcquire(ProcStructLock);
189 /* attach to the free list */
190 ProcGlobal = (PROC_HDR *)
191 ShmemInitStruct("Proc Header", (unsigned) sizeof(PROC_HDR), &found);
194 /* this should not happen. InitProcGlobal() is called before this. */
195 elog(STOP, "InitProcess: Proc Header uninitialized");
200 SpinRelease(ProcStructLock);
201 elog(ERROR, "ProcInit: you already exist");
205 /* try to get a proc from the free list first */
207 myOffset = ProcGlobal->freeProcs;
209 if (myOffset != INVALID_OFFSET)
211 MyProc = (PROC *) MAKE_PTR(myOffset);
212 ProcGlobal->freeProcs = MyProc->links.next;
218 * have to allocate one. We can't use the normal shmem index
219 * table mechanism because the proc structure is stored by PID
220 * instead of by a global name (need to look it up by PID when we
221 * cleanup dead processes).
224 MyProc = (PROC *) ShmemAlloc((unsigned) sizeof(PROC));
227 SpinRelease(ProcStructLock);
228 elog(FATAL, "cannot create new proc: out of memory");
231 /* this cannot be initialized until after the buffer pool */
232 SHMQueueInit(&(MyProc->lockQueue));
236 * zero out the spin lock counts and set the sLocks field for
237 * ProcStructLock to 1 as we have acquired this spinlock above but
238 * didn't record it since we didn't have MyProc until now.
240 MemSet(MyProc->sLocks, 0, sizeof(MyProc->sLocks));
241 MyProc->sLocks[ProcStructLock] = 1;
244 if (IsUnderPostmaster)
251 ProcGetNewSemKeyAndNum(&semKey, &semNum);
254 * Note: because of the pre-allocation done in InitProcGlobal,
255 * this call should always attach to an existing semaphore. It
256 * will (try to) create a new group of semaphores only if the
257 * postmaster tries to start more backends than it said it would.
259 semId = IpcSemaphoreCreate(semKey,
262 IpcSemaphoreDefaultStartValue,
266 * we might be reusing a semaphore that belongs to a dead backend.
267 * So be careful and reinitialize its value here.
269 semun.val = IpcSemaphoreDefaultStartValue;
270 semctl(semId, semNum, SETVAL, semun);
272 IpcSemaphoreLock(semId, semNum, IpcExclusiveLock);
273 MyProc->sem.semId = semId;
274 MyProc->sem.semNum = semNum;
275 MyProc->sem.semKey = semKey;
278 MyProc->sem.semId = -1;
280 /* ----------------------
282 * ----------------------
284 SpinRelease(ProcStructLock);
286 MyProc->pid = MyProcPid;
287 MyProc->databaseId = MyDatabaseId;
288 MyProc->xid = InvalidTransactionId;
289 MyProc->xmin = InvalidTransactionId;
292 * Start keeping spin lock stats from here on. Any botch before
293 * this initialization is forever botched
296 MemSet(MyProc->sLocks, 0, MAX_SPINS * sizeof(*MyProc->sLocks));
298 /* -------------------------
299 * Install ourselves in the shmem index table. The name to
300 * use is determined by the OS-assigned process id. That
301 * allows the cleanup process to find us after any untimely
303 * -------------------------
305 location = MAKE_OFFSET(MyProc);
306 if ((!ShmemPIDLookup(MyProcPid, &location)) || (location != MAKE_OFFSET(MyProc)))
307 elog(STOP, "InitProc: ShmemPID table broken");
309 MyProc->errType = NO_ERROR;
310 SHMQueueElemInit(&(MyProc->links));
312 on_shmem_exit(ProcKill, (caddr_t) MyProcPid);
316 * ProcReleaseLocks() -- release all locks associated with this process
324 LockReleaseAll(1, &MyProc->lockQueue);
329 * used by the postmaster to clean up the global tables. This also frees
330 * up the semaphore used for the lmgr of the process. (We have to do
331 * this is the postmaster instead of doing a IpcSemaphoreKill on exiting
332 * the process because the semaphore set is shared among backends and
333 * we don't want to remove other's semaphores on exit.)
338 SHMEM_OFFSET location;
341 location = INVALID_OFFSET;
343 location = ShmemPIDDestroy(pid);
344 if (location == INVALID_OFFSET)
346 proc = (PROC *) MAKE_PTR(location);
348 SpinAcquire(ProcStructLock);
350 ProcFreeSem(proc->sem.semKey, proc->sem.semNum);
352 proc->links.next = ProcGlobal->freeProcs;
353 ProcGlobal->freeProcs = MAKE_OFFSET(proc);
355 SpinRelease(ProcStructLock);
361 * ProcKill() -- Destroy the per-proc data structure for
362 * this process. Release any of its held spin locks.
365 ProcKill(int exitStatus, int pid)
368 SHMEM_OFFSET location;
370 /* --------------------
371 * If this is a FATAL exit the postmaster will have to kill all the
372 * existing backends and reinitialize shared memory. So all we don't
373 * need to do anything here.
374 * --------------------
379 ShmemPIDLookup(MyProcPid, &location);
380 if (location == INVALID_OFFSET)
383 proc = (PROC *) MAKE_PTR(location);
385 Assert(proc == MyProc || pid != MyProcPid);
390 * Assume one lock table.
393 ProcReleaseSpins(proc);
394 LockReleaseAll(DEFAULT_LOCKMETHOD, &proc->lockQueue);
399 * Assume we have a second lock table.
401 LockReleaseAll(USER_LOCKMETHOD, &proc->lockQueue);
405 * get off the wait queue
409 if (proc->links.next != INVALID_OFFSET)
411 Assert(proc->waitLock->waitProcs.size > 0);
412 SHMQueueDelete(&(proc->links));
413 --proc->waitLock->waitProcs.size;
415 SHMQueueElemInit(&(proc->links));
422 * ProcQueue package: routines for putting processes to sleep
427 * ProcQueueAlloc -- alloc/attach to a shared memory process queue
429 * Returns: a pointer to the queue or NULL
430 * Side Effects: Initializes the queue if we allocated one
434 ProcQueueAlloc(char *name)
437 PROC_QUEUE *queue = (PROC_QUEUE *)
438 ShmemInitStruct(name, (unsigned) sizeof(PROC_QUEUE), &found);
443 ProcQueueInit(queue);
450 * ProcQueueInit -- initialize a shared memory process queue
453 ProcQueueInit(PROC_QUEUE *queue)
455 SHMQueueInit(&(queue->links));
462 * ProcSleep -- put a process to sleep
464 * P() on the semaphore should put us to sleep. The process
465 * semaphore is cleared by default, so the first time we try
466 * to acquire it, we sleep.
468 * ASSUME: that no one will fiddle with the queue until after
469 * we release the spin lock.
471 * NOTES: The process queue is now a priority queue for locking.
474 ProcSleep(PROC_QUEUE *waitQueue,/* lock->waitProcs */
475 LOCKMETHODCTL *lockctl,
476 int token, /* lockmode */
480 SPINLOCK spinlock = lockctl->masterLock;
482 int myMask = (1 << token);
483 int waitMask = lock->waitMask;
484 int aheadHolders[MAX_LOCKMODES];
485 bool selfConflict = (lockctl->conflictTab[token] & myMask),
487 bool deadlock_checked = false;
488 struct itimerval timeval,
491 MyProc->token = token;
492 MyProc->waitLock = lock;
494 proc = (PROC *) MAKE_PTR(waitQueue->links.prev);
496 /* if we don't conflict with any waiter - be first in queue */
497 if (!(lockctl->conflictTab[token] & waitMask))
500 for (i = 1; i < MAX_LOCKMODES; i++)
501 aheadHolders[i] = lock->activeHolders[i];
502 (aheadHolders[token])++;
504 for (i = 0; i < waitQueue->size; i++)
506 /* am I waiting for him ? */
507 if (lockctl->conflictTab[token] & proc->holdLock)
509 /* is he waiting for me ? */
510 if (lockctl->conflictTab[proc->token] & MyProc->holdLock)
512 MyProc->errType = STATUS_ERROR;
513 elog(NOTICE, DeadLockMessage);
516 /* being waiting for him - go past */
518 /* if he waits for me */
519 else if (lockctl->conflictTab[proc->token] & MyProc->holdLock)
521 /* if conflicting locks requested */
522 else if (lockctl->conflictTab[proc->token] & myMask)
526 * If I request non self-conflicting lock and there are others
527 * requesting the same lock just before me - stay here.
529 if (!selfConflict && prevSame)
534 * Last attempt to don't move any more: if we don't conflict with
535 * rest waiters in queue.
537 else if (!(lockctl->conflictTab[token] & waitMask))
540 prevSame = (proc->token == token);
541 (aheadHolders[proc->token])++;
542 if (aheadHolders[proc->token] == lock->holders[proc->token])
543 waitMask &= ~(1 << proc->token);
544 proc = (PROC *) MAKE_PTR(proc->links.prev);
548 /* -------------------
549 * assume that these two operations are atomic (because
551 * -------------------
553 SHMQueueInsertTL(&(proc->links), &(MyProc->links));
556 lock->waitMask |= myMask;
557 SpinRelease(spinlock);
560 * We set this so we can wake up periodically and check for a deadlock.
561 * If a deadlock is detected, the handler releases the processes
562 * semaphore and aborts the current transaction.
564 * Need to zero out struct to set the interval and the micro seconds fields
568 MemSet(&timeval, 0, sizeof(struct itimerval));
569 timeval.it_value.tv_sec = \
570 (DeadlockCheckTimer ? DeadlockCheckTimer : DEADLOCK_CHECK_TIMER);
574 MyProc->errType = NO_ERROR; /* reset flag after deadlock check */
576 if (!deadlock_checked)
577 if (setitimer(ITIMER_REAL, &timeval, &dummy))
578 elog(FATAL, "ProcSleep: Unable to set timer for process wakeup");
579 deadlock_checked = true;
582 * if someone wakes us between SpinRelease and IpcSemaphoreLock,
583 * IpcSemaphoreLock will not block. The wakeup is "saved" by
584 * the semaphore implementation.
587 IpcSemaphoreLock(MyProc->sem.semId, MyProc->sem.semNum,
589 } while (MyProc->errType == STATUS_NOT_FOUND); /* sleep after deadlock
593 * We were awoken before a timeout - now disable the timer
596 timeval.it_value.tv_sec = 0;
597 if (setitimer(ITIMER_REAL, &timeval, &dummy))
598 elog(FATAL, "ProcSleep: Unable to diable timer for process wakeup");
601 * We were assumed to be in a critical section when we went
605 SpinAcquire(spinlock);
609 #ifdef LOCK_MGR_DEBUG
610 /* Just to get meaningful debug messages from DumpLocks() */
611 MyProc->waitLock = (LOCK *) NULL;
614 return MyProc->errType;
619 * ProcWakeup -- wake up a process by releasing its private semaphore.
621 * remove the process from the wait queue and set its links invalid.
622 * RETURN: the next process in the wait queue.
625 ProcWakeup(PROC *proc, int errType)
629 /* assume that spinlock has been acquired */
631 if (proc->links.prev == INVALID_OFFSET ||
632 proc->links.next == INVALID_OFFSET)
633 return (PROC *) NULL;
635 retProc = (PROC *) MAKE_PTR(proc->links.prev);
637 /* you have to update waitLock->waitProcs.size yourself */
638 SHMQueueDelete(&(proc->links));
639 SHMQueueElemInit(&(proc->links));
641 proc->errType = errType;
643 IpcSemaphoreUnlock(proc->sem.semId, proc->sem.semNum, IpcExclusiveLock);
649 * ProcLockWakeup -- routine for waking up processes when a lock is
653 ProcLockWakeup(PROC_QUEUE *queue, LOCKMETHOD lockmethod, LOCK *lock)
658 int last_locktype = 0;
659 int queue_size = queue->size;
661 Assert(queue->size >= 0);
664 return STATUS_NOT_FOUND;
666 proc = (PROC *) MAKE_PTR(queue->links.prev);
667 while ((queue_size--) && (proc))
671 * This proc will conflict as the previous one did, don't even
674 if (proc->token == last_locktype)
678 * Does this proc conflict with locks held by others ?
680 if (LockResolveConflicts(lockmethod,
684 (XIDLookupEnt *) NULL) != STATUS_OK)
688 last_locktype = proc->token;
693 * there was a waiting process, grant it the lock before waking it
694 * up. This will prevent another process from seizing the lock
695 * between the time we release the lock master (spinlock) and the
696 * time that the awoken process begins executing again.
698 GrantLock(lock, proc->token);
701 * ProcWakeup removes proc from the lock waiting process queue and
702 * returns the next proc in chain.
707 proc = ProcWakeup(proc, NO_ERROR);
710 Assert(queue->size >= 0);
716 /* Something is still blocking us. May have deadlocked. */
717 trace_flag = (lock->tag.lockmethod == USER_LOCKMETHOD) ? \
718 TRACE_USERLOCKS : TRACE_LOCKS;
720 "ProcLockWakeup: lock(%x) can't wake up any process",
722 #ifdef DEADLOCK_DEBUG
723 if (pg_options[trace_flag] >= 2)
726 return STATUS_NOT_FOUND;
731 ProcAddLock(SHM_QUEUE *elem)
733 SHMQueueInsertTL(&MyProc->lockQueue, elem);
736 /* --------------------
737 * We only get to this routine if we got SIGALRM after DEADLOCK_CHECK_TIMER
738 * while waiting for a lock to be released by some other process. If we have
739 * a real deadlock, we must also indicate that I'm no longer waiting
740 * on a lock so that other processes don't try to wake me up and screw
742 * --------------------
745 HandleDeadLock(SIGNAL_ARGS)
751 /* ---------------------
752 * Check to see if we've been awoken by anyone in the interim.
754 * If we have we can return and resume our transaction -- happy day.
755 * Before we are awoken the process releasing the lock grants it to
756 * us so we know that we don't have to wait anymore.
758 * Damn these names are LONG! -mer
759 * ---------------------
761 if (IpcSemaphoreGetCount(MyProc->sem.semId, MyProc->sem.semNum) ==
762 IpcSemaphoreDefaultStartValue)
769 * you would think this would be unnecessary, but...
771 * this also means we've been removed already. in some ports (e.g.,
772 * sparc and aix) the semop(2) implementation is such that we can
773 * actually end up in this handler after someone has removed us from
774 * the queue and bopped the semaphore *but the test above fails to
775 * detect the semaphore update* (presumably something weird having to
776 * do with the order in which the semaphore wakeup signal and SIGALRM
779 if (MyProc->links.prev == INVALID_OFFSET ||
780 MyProc->links.next == INVALID_OFFSET)
786 #ifdef DEADLOCK_DEBUG
790 MyProc->errType = STATUS_NOT_FOUND;
791 if (!DeadLockCheck(MyProc, MyProc->waitLock))
797 mywaitlock = MyProc->waitLock;
799 /* ------------------------
800 * Get this process off the lock's wait queue
801 * ------------------------
803 Assert(mywaitlock->waitProcs.size > 0);
804 --mywaitlock->waitProcs.size;
805 SHMQueueDelete(&(MyProc->links));
806 SHMQueueElemInit(&(MyProc->links));
808 /* ------------------
809 * Unlock my semaphore so that the count is right for next time.
810 * I was awoken by a signal, not by someone unlocking my semaphore.
813 IpcSemaphoreUnlock(MyProc->sem.semId, MyProc->sem.semNum,
817 * Set MyProc->errType to STATUS_ERROR so that we abort after
818 * returning from this handler.
821 MyProc->errType = STATUS_ERROR;
824 * if this doesn't follow the IpcSemaphoreUnlock then we get lock
825 * table corruption ("LockReplace: xid table corrupted") due to race
826 * conditions. i don't claim to understand this...
830 elog(NOTICE, DeadLockMessage);
835 ProcReleaseSpins(PROC *proc)
844 for (i = 0; i < (int) MAX_SPINS; i++)
848 Assert(proc->sLocks[i] == 1);
855 /*****************************************************************************
857 *****************************************************************************/
860 * ProcGetNewSemKeyAndNum -
861 * scan the free semaphore bitmap and allocate a single semaphore from
862 * a semaphore set. (If the semaphore set doesn't exist yet,
863 * IpcSemaphoreCreate will create it. Otherwise, we use the existing
867 ProcGetNewSemKeyAndNum(IPCKey *key, int *semNum)
870 int32 *freeSemMap = ProcGlobal->freeSemMap;
871 int32 fullmask = (1 << (PROC_NSEMS_PER_SET + 1)) - 1;
874 * we hold ProcStructLock when entering this routine. We scan through
875 * the bitmap to look for a free semaphore.
878 for (i = 0; i < MAX_PROC_SEMS / PROC_NSEMS_PER_SET; i++)
883 if (freeSemMap[i] == fullmask)
884 continue; /* this set is fully allocated */
886 for (j = 0; j < PROC_NSEMS_PER_SET; j++)
888 if ((freeSemMap[i] & mask) == 0)
892 * a free semaphore found. Mark it as allocated. Also set
893 * the bit indicating whole set is allocated.
895 freeSemMap[i] |= mask + (1 << PROC_NSEMS_PER_SET);
897 *key = ProcGlobal->currKey + i;
905 /* if we reach here, all the semaphores are in use. */
906 elog(ERROR, "InitProc: cannot allocate a free semaphore");
911 * free up our semaphore in the semaphore set.
914 ProcFreeSem(IpcSemaphoreKey semKey, int semNum)
918 int32 *freeSemMap = ProcGlobal->freeSemMap;
920 i = semKey - ProcGlobal->currKey;
921 mask = ~(1 << semNum);
922 freeSemMap[i] &= mask;
925 * Formerly we'd release a semaphore set if it was now completely
926 * unused, but now we keep the semaphores to ensure we won't run out
927 * when starting new backends --- cf. InitProcGlobal. Note that the
928 * PROC_NSEMS_PER_SET+1'st bit of the freeSemMap entry remains set to
929 * indicate it is still allocated; ProcFreeAllSemaphores() needs that.
934 * ProcFreeAllSemaphores -
935 * called at shmem_exit time, ie when exiting the postmaster or
936 * destroying shared state for a failed set of backends.
937 * Free up all the semaphores allocated to the lmgrs of the backends.
940 ProcFreeAllSemaphores()
943 int32 *freeSemMap = ProcGlobal->freeSemMap;
945 for (i = 0; i < MAX_PROC_SEMS / PROC_NSEMS_PER_SET; i++)
947 if (freeSemMap[i] != 0)
948 IpcSemaphoreKill(ProcGlobal->currKey + i);