proc.c 24.2 KB
Newer Older
1 2
/*-------------------------------------------------------------------------
 *
3
 * proc.c
4
 *	  routines to manage per-process shared memory data structure
5 6 7 8 9
 *
 * Copyright (c) 1994, Regents of the University of California
 *
 *
 * IDENTIFICATION
10
 *	  $Header: /cvsroot/pgsql/src/backend/storage/lmgr/proc.c,v 1.65 1999/12/16 01:25:08 momjian Exp $
11 12 13 14
 *
 *-------------------------------------------------------------------------
 */
/*
15 16
 *	Each postgres backend gets one of these.  We'll use it to
 *	clean up after the process should the process suddenly die.
17 18 19
 *
 *
 * Interface (a):
20 21 22
 *		ProcSleep(), ProcWakeup(), ProcWakeupNext(),
 *		ProcQueueAlloc() -- create a shm queue for sleeping processes
 *		ProcQueueInit() -- create a queue without allocing memory
23 24 25 26 27 28 29 30 31 32
 *
 * Locking and waiting for buffers can cause the backend to be
 * put to sleep.  Whoever releases the lock, etc. wakes the
 * process up again (and gives it an error code so it knows
 * whether it was awoken on an error condition).
 *
 * Interface (b):
 *
 * ProcReleaseLocks -- frees the locks associated with this process,
 * ProcKill -- destroys the shared memory state (and locks)
33
 *		associated with the process.
34 35
 *
 * 5/15/91 -- removed the buffer pool based lock chain in favor
36 37 38 39 40 41
 *		of a shared memory lock chain.	The write-protection is
 *		more expensive if the lock chain is in the buffer pool.
 *		The only reason I kept the lock chain in the buffer pool
 *		in the first place was to allow the lock table to grow larger
 *		than available shared memory and that isn't going to work
 *		without a lot of unimplemented support anyway.
42 43
 *
 * 4/7/95 -- instead of allocating a set of 1 semaphore per process, we
44 45 46 47
 *		allocate a semaphore from a set of PROC_NSEMS_PER_SET semaphores
 *		shared among backends (we keep a few sets of semaphores around).
 *		This is so that we can support more backends. (system-wide semaphore
 *		sets run out pretty fast.)				  -ay 4/95
48
 *
49
 * $Header: /cvsroot/pgsql/src/backend/storage/lmgr/proc.c,v 1.65 1999/12/16 01:25:08 momjian Exp $
50 51 52
 */
#include <sys/time.h>
#include <unistd.h>
53
#include <signal.h>
54
#include <sys/types.h>
M
Marc G. Fournier 已提交
55

56
#if defined(solaris_sparc) || defined(__CYGWIN__)
57 58 59 60
#include <sys/ipc.h>
#include <sys/sem.h>
#endif

M
Marc G. Fournier 已提交
61
#include "postgres.h"
62
#include "miscadmin.h"
63
#include "libpq/pqsignal.h"
64 65


66 67
#include "storage/ipc.h"
/* In Ultrix and QNX, sem.h must be included after ipc.h */
68
#include <sys/sem.h>
B
Bruce Momjian 已提交
69

B
Bruce Momjian 已提交
70
#include "storage/lmgr.h"
71
#include "storage/proc.h"
M
 
Marc G. Fournier 已提交
72
#include "utils/trace.h"
73

74
void HandleDeadLock(SIGNAL_ARGS);
75
static void ProcFreeAllSemaphores(void);
76

M
 
Marc G. Fournier 已提交
77 78
#define DeadlockCheckTimer pg_options[OPT_DEADLOCKTIMEOUT]

79 80 81 82 83 84 85
/* --------------------
 * Spin lock for manipulating the shared process data structure:
 * ProcGlobal.... Adding an extra spin lock seemed like the smallest
 * hack to get around reading and updating this structure in shared
 * memory. -mer 17 July 1991
 * --------------------
 */
86
SPINLOCK	ProcStructLock;
87 88 89

static PROC_HDR *ProcGlobal = NULL;

90
PROC	   *MyProc = NULL;
91

92
static void ProcKill(int exitStatus, int pid);
93
static void ProcGetNewSemKeyAndNum(IPCKey *key, int *semNum);
94
static void ProcFreeSem(IpcSemaphoreKey semKey, int semNum);
95

V
Vadim B. Mikheev 已提交
96 97
static char *DeadLockMessage = "Deadlock detected -- See the lock(l) manual page for a possible cause.";

98 99
/*
 * InitProcGlobal -
100
 *	  initializes the global process table. We put it here so that
101
 *	  the postmaster can do this initialization. (ProcFreeAllSemaphores needs
102 103 104
 *	  to read this table on exiting the postmaster. If we have the first
 *	  backend do this, starting up and killing the postmaster without
 *	  starting any backends will be a problem.)
105 106 107 108 109 110 111 112 113 114 115
 *
 *	  We also allocate all the per-process semaphores we will need to support
 *	  the requested number of backends.  We used to allocate semaphores
 *	  only when backends were actually started up, but that is bad because
 *	  it lets Postgres fail under load --- a lot of Unix systems are
 *	  (mis)configured with small limits on the number of semaphores, and
 *	  running out when trying to start another backend is a common failure.
 *	  So, now we grab enough semaphores to support the desired max number
 *	  of backends immediately at initialization --- if the sysadmin has set
 *	  MaxBackends higher than his kernel will support, he'll find out sooner
 *	  rather than later.
116 117
 */
void
118
InitProcGlobal(IPCKey key, int maxBackends)
119
{
120
	bool		found = false;
121

122 123 124
	/* attach to the free list */
	ProcGlobal = (PROC_HDR *)
		ShmemInitStruct("Proc Header", (unsigned) sizeof(PROC_HDR), &found);
125

126 127
	/* --------------------
	 * We're the first - initialize.
128 129
	 * XXX if found should ever be true, it is a sign of impending doom ...
	 * ought to complain if so?
130 131 132
	 * --------------------
	 */
	if (!found)
133
	{
134
		int			i;
135

136 137 138 139
		ProcGlobal->freeProcs = INVALID_OFFSET;
		ProcGlobal->currKey = IPCGetProcessSemaphoreInitKey(key);
		for (i = 0; i < MAX_PROC_SEMS / PROC_NSEMS_PER_SET; i++)
			ProcGlobal->freeSemMap[i] = 0;
140

B
Bruce Momjian 已提交
141 142 143
		/*
		 * Arrange to delete semas on exit --- set this up now so that we
		 * will clean up if pre-allocation fails...
144 145 146
		 */
		on_shmem_exit(ProcFreeAllSemaphores, NULL);

B
Bruce Momjian 已提交
147 148
		/*
		 * Pre-create the semaphores for the first maxBackends processes,
149 150 151
		 * unless we are running as a standalone backend.
		 */
		if (key != PrivateIPCKey)
152
		{
153
			for (i = 0;
B
Bruce Momjian 已提交
154
				 i < (maxBackends + PROC_NSEMS_PER_SET - 1) / PROC_NSEMS_PER_SET;
155 156 157 158 159 160 161 162 163
				 i++)
			{
				IPCKey		semKey = ProcGlobal->currKey + i;
				int			semId;

				semId = IpcSemaphoreCreate(semKey,
										   PROC_NSEMS_PER_SET,
										   IPCProtection,
										   IpcSemaphoreDefaultStartValue,
164 165 166
										   0);
				if (semId < 0)
					elog(FATAL, "InitProcGlobal: IpcSemaphoreCreate failed");
167 168 169
				/* mark this sema set allocated */
				ProcGlobal->freeSemMap[i] = (1 << PROC_NSEMS_PER_SET);
			}
170
		}
171 172 173 174 175 176 177 178 179 180 181
	}
}

/* ------------------------
 * InitProc -- create a per-process data structure for this process
 * used by the lock manager on semaphore queues.
 * ------------------------
 */
void
InitProcess(IPCKey key)
{
182 183 184
	bool		found = false;
	unsigned long location,
				myOffset;
185 186 187 188 189 190 191

	SpinAcquire(ProcStructLock);

	/* attach to the free list */
	ProcGlobal = (PROC_HDR *)
		ShmemInitStruct("Proc Header", (unsigned) sizeof(PROC_HDR), &found);
	if (!found)
192
	{
193
		/* this should not happen. InitProcGlobal() is called before this. */
194
		elog(STOP, "InitProcess: Proc Header uninitialized");
195
	}
196 197

	if (MyProc != NULL)
198
	{
199
		SpinRelease(ProcStructLock);
200
		elog(ERROR, "ProcInit: you already exist");
201
		return;
202
	}
203 204 205 206 207 208

	/* try to get a proc from the free list first */

	myOffset = ProcGlobal->freeProcs;

	if (myOffset != INVALID_OFFSET)
209
	{
210 211 212 213 214 215 216
		MyProc = (PROC *) MAKE_PTR(myOffset);
		ProcGlobal->freeProcs = MyProc->links.next;
	}
	else
	{

		/*
217 218 219 220
		 * have to allocate one.  We can't use the normal shmem index
		 * table mechanism because the proc structure is stored by PID
		 * instead of by a global name (need to look it up by PID when we
		 * cleanup dead processes).
221 222 223 224
		 */

		MyProc = (PROC *) ShmemAlloc((unsigned) sizeof(PROC));
		if (!MyProc)
225
		{
226 227
			SpinRelease(ProcStructLock);
			elog(FATAL, "cannot create new proc: out of memory");
228
		}
229 230 231

		/* this cannot be initialized until after the buffer pool */
		SHMQueueInit(&(MyProc->lockQueue));
232
	}
233

234
	/*
235 236 237
	 * zero out the spin lock counts and set the sLocks field for
	 * ProcStructLock to 1 as we have acquired this spinlock above but
	 * didn't record it since we didn't have MyProc until now.
238
	 */
B
Bruce Momjian 已提交
239
	MemSet(MyProc->sLocks, 0, sizeof(MyProc->sLocks));
240 241 242 243 244
	MyProc->sLocks[ProcStructLock] = 1;


	if (IsUnderPostmaster)
	{
245 246 247 248
		IPCKey		semKey;
		int			semNum;
		int			semId;
		union semun semun;
249 250 251

		ProcGetNewSemKeyAndNum(&semKey, &semNum);

B
Bruce Momjian 已提交
252 253 254 255 256
		/*
		 * Note: because of the pre-allocation done in InitProcGlobal,
		 * this call should always attach to an existing semaphore. It
		 * will (try to) create a new group of semaphores only if the
		 * postmaster tries to start more backends than it said it would.
257
		 */
258 259 260 261
		semId = IpcSemaphoreCreate(semKey,
								   PROC_NSEMS_PER_SET,
								   IPCProtection,
								   IpcSemaphoreDefaultStartValue,
262
								   0);
263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284

		/*
		 * we might be reusing a semaphore that belongs to a dead backend.
		 * So be careful and reinitialize its value here.
		 */
		semun.val = IpcSemaphoreDefaultStartValue;
		semctl(semId, semNum, SETVAL, semun);

		IpcSemaphoreLock(semId, semNum, IpcExclusiveLock);
		MyProc->sem.semId = semId;
		MyProc->sem.semNum = semNum;
		MyProc->sem.semKey = semKey;
	}
	else
		MyProc->sem.semId = -1;

	/* ----------------------
	 * Release the lock.
	 * ----------------------
	 */
	SpinRelease(ProcStructLock);

B
Bruce Momjian 已提交
285
	MyProc->pid = MyProcPid;
286
	MyProc->databaseId = MyDatabaseId;
287
	MyProc->xid = InvalidTransactionId;
288
	MyProc->xmin = InvalidTransactionId;
289 290 291 292 293 294

	/* ----------------
	 * Start keeping spin lock stats from here on.	Any botch before
	 * this initialization is forever botched
	 * ----------------
	 */
B
Bruce Momjian 已提交
295
	MemSet(MyProc->sLocks, 0, MAX_SPINS * sizeof(*MyProc->sLocks));
296 297

	/* -------------------------
298
	 * Install ourselves in the shmem index table.	The name to
299 300 301 302 303 304
	 * use is determined by the OS-assigned process id.  That
	 * allows the cleanup process to find us after any untimely
	 * exit.
	 * -------------------------
	 */
	location = MAKE_OFFSET(MyProc);
B
Bruce Momjian 已提交
305
	if ((!ShmemPIDLookup(MyProcPid, &location)) || (location != MAKE_OFFSET(MyProc)))
306
		elog(STOP, "InitProc: ShmemPID table broken");
307 308 309 310

	MyProc->errType = NO_ERROR;
	SHMQueueElemInit(&(MyProc->links));

311
	on_shmem_exit(ProcKill, (caddr_t) MyProcPid);
312 313 314 315 316 317 318 319 320
}

/*
 * ProcReleaseLocks() -- release all locks associated with this process
 *
 */
void
ProcReleaseLocks()
{
321 322 323
	if (!MyProc)
		return;
	LockReleaseAll(1, &MyProc->lockQueue);
324 325 326 327
}

/*
 * ProcRemove -
328 329 330 331 332
 *	  used by the postmaster to clean up the global tables. This also frees
 *	  up the semaphore used for the lmgr of the process. (We have to do
 *	  this is the postmaster instead of doing a IpcSemaphoreKill on exiting
 *	  the process because the semaphore set is shared among backends and
 *	  we don't want to remove other's semaphores on exit.)
333 334 335 336
 */
bool
ProcRemove(int pid)
{
337 338
	SHMEM_OFFSET location;
	PROC	   *proc;
339 340 341 342 343

	location = INVALID_OFFSET;

	location = ShmemPIDDestroy(pid);
	if (location == INVALID_OFFSET)
344
		return FALSE;
345 346 347 348 349 350 351 352 353 354 355
	proc = (PROC *) MAKE_PTR(location);

	SpinAcquire(ProcStructLock);

	ProcFreeSem(proc->sem.semKey, proc->sem.semNum);

	proc->links.next = ProcGlobal->freeProcs;
	ProcGlobal->freeProcs = MAKE_OFFSET(proc);

	SpinRelease(ProcStructLock);

356
	return TRUE;
357 358 359 360
}

/*
 * ProcKill() -- Destroy the per-proc data structure for
361
 *		this process. Release any of its held spin locks.
362 363 364 365
 */
static void
ProcKill(int exitStatus, int pid)
{
366 367
	PROC	   *proc;
	SHMEM_OFFSET location;
368 369 370 371 372 373 374 375 376 377

	/* --------------------
	 * If this is a FATAL exit the postmaster will have to kill all the
	 * existing backends and reinitialize shared memory.  So all we don't
	 * need to do anything here.
	 * --------------------
	 */
	if (exitStatus != 0)
		return;

B
Bruce Momjian 已提交
378
	ShmemPIDLookup(MyProcPid, &location);
379 380 381 382 383
	if (location == INVALID_OFFSET)
		return;

	proc = (PROC *) MAKE_PTR(location);

384 385 386
	Assert(proc == MyProc || pid != MyProcPid);

	MyProc = NULL;
387 388 389 390 391 392

	/* ---------------
	 * Assume one lock table.
	 * ---------------
	 */
	ProcReleaseSpins(proc);
M
 
Marc G. Fournier 已提交
393
	LockReleaseAll(DEFAULT_LOCKMETHOD, &proc->lockQueue);
394

395
#ifdef USER_LOCKS
396

M
 
Marc G. Fournier 已提交
397 398 399 400
	/*
	 * Assume we have a second lock table.
	 */
	LockReleaseAll(USER_LOCKMETHOD, &proc->lockQueue);
401 402
#endif

403 404 405 406 407 408 409 410 411 412 413 414 415 416 417
	/* ----------------
	 * get off the wait queue
	 * ----------------
	 */
	LockLockTable();
	if (proc->links.next != INVALID_OFFSET)
	{
		Assert(proc->waitLock->waitProcs.size > 0);
		SHMQueueDelete(&(proc->links));
		--proc->waitLock->waitProcs.size;
	}
	SHMQueueElemInit(&(proc->links));
	UnlockLockTable();

	return;
418 419 420 421
}

/*
 * ProcQueue package: routines for putting processes to sleep
422
 *		and  waking them up
423 424 425 426 427 428 429 430
 */

/*
 * ProcQueueAlloc -- alloc/attach to a shared memory process queue
 *
 * Returns: a pointer to the queue or NULL
 * Side Effects: Initializes the queue if we allocated one
 */
431
#ifdef NOT_USED
432
PROC_QUEUE *
433 434
ProcQueueAlloc(char *name)
{
435 436
	bool		found;
	PROC_QUEUE *queue = (PROC_QUEUE *)
437 438 439
	ShmemInitStruct(name, (unsigned) sizeof(PROC_QUEUE), &found);

	if (!queue)
440
		return NULL;
441 442
	if (!found)
		ProcQueueInit(queue);
443
	return queue;
444
}
445

446
#endif
447 448 449 450 451

/*
 * ProcQueueInit -- initialize a shared memory process queue
 */
void
452
ProcQueueInit(PROC_QUEUE *queue)
453
{
454 455
	SHMQueueInit(&(queue->links));
	queue->size = 0;
456 457 458 459 460 461 462 463 464 465 466 467
}



/*
 * ProcSleep -- put a process to sleep
 *
 * P() on the semaphore should put us to sleep.  The process
 * semaphore is cleared by default, so the first time we try
 * to acquire it, we sleep.
 *
 * ASSUME: that no one will fiddle with the queue until after
468
 *		we release the spin lock.
469 470 471 472
 *
 * NOTES: The process queue is now a priority queue for locking.
 */
int
473
ProcSleep(PROC_QUEUE *waitQueue,/* lock->waitProcs */
474
		  LOCKMETHODCTL *lockctl,
475
		  int token,			/* lockmode */
V
Vadim B. Mikheev 已提交
476
		  LOCK *lock)
477
{
478
	int			i;
V
Vadim B. Mikheev 已提交
479
	SPINLOCK	spinlock = lockctl->masterLock;
480
	PROC	   *proc;
V
Vadim B. Mikheev 已提交
481 482 483 484 485
	int			myMask = (1 << token);
	int			waitMask = lock->waitMask;
	int			aheadHolders[MAX_LOCKMODES];
	bool		selfConflict = (lockctl->conflictTab[token] & myMask),
				prevSame = false;
B
Bruce Momjian 已提交
486 487 488
	bool		deadlock_checked = false;
	struct itimerval timeval,
				dummy;
489

V
Vadim B. Mikheev 已提交
490 491 492
	MyProc->token = token;
	MyProc->waitLock = lock;

B
Bruce Momjian 已提交
493
	proc = (PROC *) MAKE_PTR(waitQueue->links.prev);
494

V
Vadim B. Mikheev 已提交
495 496 497
	/* if we don't conflict with any waiter - be first in queue */
	if (!(lockctl->conflictTab[token] & waitMask))
		goto ins;
498

V
Vadim B. Mikheev 已提交
499 500 501
	for (i = 1; i < MAX_LOCKMODES; i++)
		aheadHolders[i] = lock->activeHolders[i];
	(aheadHolders[token])++;
502

V
Vadim B. Mikheev 已提交
503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522
	for (i = 0; i < waitQueue->size; i++)
	{
		/* am I waiting for him ? */
		if (lockctl->conflictTab[token] & proc->holdLock)
		{
			/* is he waiting for me ? */
			if (lockctl->conflictTab[proc->token] & MyProc->holdLock)
			{
				MyProc->errType = STATUS_ERROR;
				elog(NOTICE, DeadLockMessage);
				goto rt;
			}
			/* being waiting for him - go past */
		}
		/* if he waits for me */
		else if (lockctl->conflictTab[proc->token] & MyProc->holdLock)
			break;
		/* if conflicting locks requested */
		else if (lockctl->conflictTab[proc->token] & myMask)
		{
B
Bruce Momjian 已提交
523

V
Vadim B. Mikheev 已提交
524
			/*
B
Bruce Momjian 已提交
525 526
			 * If I request non self-conflicting lock and there are others
			 * requesting the same lock just before me - stay here.
V
Vadim B. Mikheev 已提交
527 528 529 530
			 */
			if (!selfConflict && prevSame)
				break;
		}
B
Bruce Momjian 已提交
531

V
Vadim B. Mikheev 已提交
532
		/*
B
Bruce Momjian 已提交
533 534
		 * Last attempt to don't move any more: if we don't conflict with
		 * rest waiters in queue.
V
Vadim B. Mikheev 已提交
535 536 537
		 */
		else if (!(lockctl->conflictTab[token] & waitMask))
			break;
538

V
Vadim B. Mikheev 已提交
539 540 541
		prevSame = (proc->token == token);
		(aheadHolders[proc->token])++;
		if (aheadHolders[proc->token] == lock->holders[proc->token])
B
Bruce Momjian 已提交
542
			waitMask &= ~(1 << proc->token);
V
Vadim B. Mikheev 已提交
543 544
		proc = (PROC *) MAKE_PTR(proc->links.prev);
	}
545

V
Vadim B. Mikheev 已提交
546
ins:;
547 548 549 550 551 552
	/* -------------------
	 * assume that these two operations are atomic (because
	 * of the spinlock).
	 * -------------------
	 */
	SHMQueueInsertTL(&(proc->links), &(MyProc->links));
B
Bruce Momjian 已提交
553
	waitQueue->size++;
554

V
Vadim B. Mikheev 已提交
555
	lock->waitMask |= myMask;
556 557 558
	SpinRelease(spinlock);

	/* --------------
B
Bruce Momjian 已提交
559
	 * We set this so we can wake up periodically and check for a deadlock.
B
Bruce Momjian 已提交
560 561
	 * If a deadlock is detected, the handler releases the processes
	 * semaphore and aborts the current transaction.
B
Bruce Momjian 已提交
562 563 564
	 *
	 * Need to zero out struct to set the interval and the micro seconds fields
	 * to 0.
565 566
	 * --------------
	 */
B
Bruce Momjian 已提交
567 568 569
	MemSet(&timeval, 0, sizeof(struct itimerval));
	timeval.it_value.tv_sec = \
		(DeadlockCheckTimer ? DeadlockCheckTimer : DEADLOCK_CHECK_TIMER);
570

B
Bruce Momjian 已提交
571 572
	do
	{
573
		MyProc->errType = NO_ERROR;		/* reset flag after deadlock check */
574

B
Bruce Momjian 已提交
575 576 577 578 579
		if (!deadlock_checked)
			if (setitimer(ITIMER_REAL, &timeval, &dummy))
				elog(FATAL, "ProcSleep: Unable to set timer for process wakeup");
		deadlock_checked = true;

B
Bruce Momjian 已提交
580 581 582 583 584 585
		/* --------------
		 * if someone wakes us between SpinRelease and IpcSemaphoreLock,
		 * IpcSemaphoreLock will not block.  The wakeup is "saved" by
		 * the semaphore implementation.
		 * --------------
		 */
M
 
Marc G. Fournier 已提交
586 587
		IpcSemaphoreLock(MyProc->sem.semId, MyProc->sem.semNum,
						 IpcExclusiveLock);
588 589 590
	} while (MyProc->errType == STATUS_NOT_FOUND);		/* sleep after deadlock
														 * check */

B
Bruce Momjian 已提交
591 592 593 594 595 596 597 598
	/* ---------------
	 * We were awoken before a timeout - now disable the timer
	 * ---------------
	 */
	timeval.it_value.tv_sec = 0;
	if (setitimer(ITIMER_REAL, &timeval, &dummy))
		elog(FATAL, "ProcSleep: Unable to diable timer for process wakeup");

599 600 601 602 603 604 605
	/* ----------------
	 * We were assumed to be in a critical section when we went
	 * to sleep.
	 * ----------------
	 */
	SpinAcquire(spinlock);

V
Vadim B. Mikheev 已提交
606 607
rt:;

M
 
Marc G. Fournier 已提交
608 609
#ifdef LOCK_MGR_DEBUG
	/* Just to get meaningful debug messages from DumpLocks() */
610
	MyProc->waitLock = (LOCK *) NULL;
M
 
Marc G. Fournier 已提交
611 612
#endif

613
	return MyProc->errType;
614 615 616 617 618 619
}


/*
 * ProcWakeup -- wake up a process by releasing its private semaphore.
 *
620 621
 *	 remove the process from the wait queue and set its links invalid.
 *	 RETURN: the next process in the wait queue.
622
 */
B
Bruce Momjian 已提交
623
PROC *
624
ProcWakeup(PROC *proc, int errType)
625
{
626
	PROC	   *retProc;
627 628 629 630 631

	/* assume that spinlock has been acquired */

	if (proc->links.prev == INVALID_OFFSET ||
		proc->links.next == INVALID_OFFSET)
632
		return (PROC *) NULL;
633 634 635 636 637 638 639 640 641 642 643 644

	retProc = (PROC *) MAKE_PTR(proc->links.prev);

	/* you have to update waitLock->waitProcs.size yourself */
	SHMQueueDelete(&(proc->links));
	SHMQueueElemInit(&(proc->links));

	proc->errType = errType;

	IpcSemaphoreUnlock(proc->sem.semId, proc->sem.semNum, IpcExclusiveLock);

	return retProc;
645 646 647 648
}

/*
 * ProcLockWakeup -- routine for waking up processes when a lock is
649
 *		released.
650 651
 */
int
652
ProcLockWakeup(PROC_QUEUE *queue, LOCKMETHOD lockmethod, LOCK *lock)
653
{
654
	PROC	   *proc;
V
Vadim B. Mikheev 已提交
655
	int			count = 0;
M
 
Marc G. Fournier 已提交
656
	int			trace_flag;
V
Vadim B. Mikheev 已提交
657
	int			last_locktype = 0;
M
 
Marc G. Fournier 已提交
658 659 660
	int			queue_size = queue->size;

	Assert(queue->size >= 0);
661 662

	if (!queue->size)
663
		return STATUS_NOT_FOUND;
664 665

	proc = (PROC *) MAKE_PTR(queue->links.prev);
M
 
Marc G. Fournier 已提交
666 667
	while ((queue_size--) && (proc))
	{
668

M
 
Marc G. Fournier 已提交
669
		/*
670 671
		 * This proc will conflict as the previous one did, don't even
		 * try.
M
 
Marc G. Fournier 已提交
672 673 674 675 676
		 */
		if (proc->token == last_locktype)
			continue;

		/*
V
Vadim B. Mikheev 已提交
677
		 * Does this proc conflict with locks held by others ?
M
 
Marc G. Fournier 已提交
678 679
		 */
		if (LockResolveConflicts(lockmethod,
680
								 lock,
681
								 proc->token,
M
 
Marc G. Fournier 已提交
682 683 684
								 proc->xid,
								 (XIDLookupEnt *) NULL) != STATUS_OK)
		{
V
Vadim B. Mikheev 已提交
685 686
			if (count != 0)
				break;
M
 
Marc G. Fournier 已提交
687 688 689
			last_locktype = proc->token;
			continue;
		}
690 691 692 693 694 695 696

		/*
		 * there was a waiting process, grant it the lock before waking it
		 * up.	This will prevent another process from seizing the lock
		 * between the time we release the lock master (spinlock) and the
		 * time that the awoken process begins executing again.
		 */
697
		GrantLock(lock, proc->token);
698 699 700

		/*
		 * ProcWakeup removes proc from the lock waiting process queue and
701
		 * returns the next proc in chain.
702 703 704
		 */

		count++;
M
 
Marc G. Fournier 已提交
705 706
		queue->size--;
		proc = ProcWakeup(proc, NO_ERROR);
707
	}
708

M
 
Marc G. Fournier 已提交
709 710
	Assert(queue->size >= 0);

711
	if (count)
712
		return STATUS_OK;
713 714
	else
	{
715
		/* Something is still blocking us.	May have deadlocked. */
M
 
Marc G. Fournier 已提交
716 717 718 719 720 721 722 723 724
		trace_flag = (lock->tag.lockmethod == USER_LOCKMETHOD) ? \
			TRACE_USERLOCKS : TRACE_LOCKS;
		TPRINTF(trace_flag,
				"ProcLockWakeup: lock(%x) can't wake up any process",
				MAKE_OFFSET(lock));
#ifdef DEADLOCK_DEBUG
		if (pg_options[trace_flag] >= 2)
			DumpAllLocks();
#endif
725
		return STATUS_NOT_FOUND;
M
 
Marc G. Fournier 已提交
726
	}
727 728 729
}

void
730
ProcAddLock(SHM_QUEUE *elem)
731
{
732
	SHMQueueInsertTL(&MyProc->lockQueue, elem);
733 734 735
}

/* --------------------
B
Bruce Momjian 已提交
736 737 738
 * We only get to this routine if we got SIGALRM after DEADLOCK_CHECK_TIMER
 * while waiting for a lock to be released by some other process.  If we have
 * a real deadlock, we must also indicate that I'm no longer waiting
739
 * on a lock so that other processes don't try to wake me up and screw
740 741 742
 * up my semaphore.
 * --------------------
 */
743 744
void
HandleDeadLock(SIGNAL_ARGS)
745
{
B
Bruce Momjian 已提交
746
	LOCK	   *mywaitlock;
747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784

	LockLockTable();

	/* ---------------------
	 * Check to see if we've been awoken by anyone in the interim.
	 *
	 * If we have we can return and resume our transaction -- happy day.
	 * Before we are awoken the process releasing the lock grants it to
	 * us so we know that we don't have to wait anymore.
	 *
	 * Damn these names are LONG! -mer
	 * ---------------------
	 */
	if (IpcSemaphoreGetCount(MyProc->sem.semId, MyProc->sem.semNum) ==
		IpcSemaphoreDefaultStartValue)
	{
		UnlockLockTable();
		return;
	}

	/*
	 * you would think this would be unnecessary, but...
	 *
	 * this also means we've been removed already.  in some ports (e.g.,
	 * sparc and aix) the semop(2) implementation is such that we can
	 * actually end up in this handler after someone has removed us from
	 * the queue and bopped the semaphore *but the test above fails to
	 * detect the semaphore update* (presumably something weird having to
	 * do with the order in which the semaphore wakeup signal and SIGALRM
	 * get handled).
	 */
	if (MyProc->links.prev == INVALID_OFFSET ||
		MyProc->links.next == INVALID_OFFSET)
	{
		UnlockLockTable();
		return;
	}

785
#ifdef DEADLOCK_DEBUG
M
 
Marc G. Fournier 已提交
786
	DumpAllLocks();
787 788
#endif

B
Bruce Momjian 已提交
789 790
	MyProc->errType = STATUS_NOT_FOUND;
	if (!DeadLockCheck(MyProc, MyProc->waitLock))
B
Bruce Momjian 已提交
791 792 793 794 795 796 797
	{
		UnlockLockTable();
		return;
	}

	mywaitlock = MyProc->waitLock;

798 799 800 801
	/* ------------------------
	 * Get this process off the lock's wait queue
	 * ------------------------
	 */
B
Bruce Momjian 已提交
802 803
	Assert(mywaitlock->waitProcs.size > 0);
	--mywaitlock->waitProcs.size;
804 805 806 807 808 809 810 811
	SHMQueueDelete(&(MyProc->links));
	SHMQueueElemInit(&(MyProc->links));

	/* ------------------
	 * Unlock my semaphore so that the count is right for next time.
	 * I was awoken by a signal, not by someone unlocking my semaphore.
	 * ------------------
	 */
M
 
Marc G. Fournier 已提交
812 813
	IpcSemaphoreUnlock(MyProc->sem.semId, MyProc->sem.semNum,
					   IpcExclusiveLock);
814 815 816 817 818 819 820 821 822 823 824 825 826 827 828

	/* -------------
	 * Set MyProc->errType to STATUS_ERROR so that we abort after
	 * returning from this handler.
	 * -------------
	 */
	MyProc->errType = STATUS_ERROR;

	/*
	 * if this doesn't follow the IpcSemaphoreUnlock then we get lock
	 * table corruption ("LockReplace: xid table corrupted") due to race
	 * conditions.	i don't claim to understand this...
	 */
	UnlockLockTable();

V
Vadim B. Mikheev 已提交
829
	elog(NOTICE, DeadLockMessage);
830
	return;
831 832 833
}

void
834
ProcReleaseSpins(PROC *proc)
835
{
836
	int			i;
837 838 839 840 841 842 843

	if (!proc)
		proc = MyProc;

	if (!proc)
		return;
	for (i = 0; i < (int) MAX_SPINS; i++)
844
	{
845
		if (proc->sLocks[i])
846
		{
847 848
			Assert(proc->sLocks[i] == 1);
			SpinRelease(i);
849 850 851 852 853
		}
	}
}

/*****************************************************************************
854
 *
855 856 857 858
 *****************************************************************************/

/*
 * ProcGetNewSemKeyAndNum -
859 860 861 862
 *	  scan the free semaphore bitmap and allocate a single semaphore from
 *	  a semaphore set. (If the semaphore set doesn't exist yet,
 *	  IpcSemaphoreCreate will create it. Otherwise, we use the existing
 *	  semaphore set.)
863 864
 */
static void
865
ProcGetNewSemKeyAndNum(IPCKey *key, int *semNum)
866
{
867 868
	int			i;
	int32	   *freeSemMap = ProcGlobal->freeSemMap;
B
Bruce Momjian 已提交
869
	int32		fullmask = (1 << (PROC_NSEMS_PER_SET + 1)) - 1;
870

871 872 873 874
	/*
	 * we hold ProcStructLock when entering this routine. We scan through
	 * the bitmap to look for a free semaphore.
	 */
875

876 877
	for (i = 0; i < MAX_PROC_SEMS / PROC_NSEMS_PER_SET; i++)
	{
878 879
		int			mask = 1;
		int			j;
880 881

		if (freeSemMap[i] == fullmask)
882
			continue;			/* this set is fully allocated */
883 884 885 886 887 888 889

		for (j = 0; j < PROC_NSEMS_PER_SET; j++)
		{
			if ((freeSemMap[i] & mask) == 0)
			{

				/*
B
Bruce Momjian 已提交
890 891
				 * a free semaphore found. Mark it as allocated. Also set
				 * the bit indicating whole set is allocated.
892
				 */
893
				freeSemMap[i] |= mask + (1 << PROC_NSEMS_PER_SET);
894 895 896 897 898 899 900

				*key = ProcGlobal->currKey + i;
				*semNum = j;
				return;
			}
			mask <<= 1;
		}
901 902
	}

903
	/* if we reach here, all the semaphores are in use. */
904
	elog(ERROR, "InitProc: cannot allocate a free semaphore");
905 906 907 908
}

/*
 * ProcFreeSem -
909
 *	  free up our semaphore in the semaphore set.
910 911 912 913
 */
static void
ProcFreeSem(IpcSemaphoreKey semKey, int semNum)
{
914 915 916
	int			mask;
	int			i;
	int32	   *freeSemMap = ProcGlobal->freeSemMap;
917

918 919 920
	i = semKey - ProcGlobal->currKey;
	mask = ~(1 << semNum);
	freeSemMap[i] &= mask;
921

B
Bruce Momjian 已提交
922 923 924 925
	/*
	 * Formerly we'd release a semaphore set if it was now completely
	 * unused, but now we keep the semaphores to ensure we won't run out
	 * when starting new backends --- cf. InitProcGlobal.  Note that the
926 927 928
	 * PROC_NSEMS_PER_SET+1'st bit of the freeSemMap entry remains set to
	 * indicate it is still allocated; ProcFreeAllSemaphores() needs that.
	 */
929 930 931 932
}

/*
 * ProcFreeAllSemaphores -
933 934 935
 *	  called at shmem_exit time, ie when exiting the postmaster or
 *	  destroying shared state for a failed set of backends.
 *	  Free up all the semaphores allocated to the lmgrs of the backends.
936
 */
937
static void
938 939
ProcFreeAllSemaphores()
{
940 941
	int			i;
	int32	   *freeSemMap = ProcGlobal->freeSemMap;
942

943 944 945 946 947
	for (i = 0; i < MAX_PROC_SEMS / PROC_NSEMS_PER_SET; i++)
	{
		if (freeSemMap[i] != 0)
			IpcSemaphoreKill(ProcGlobal->currKey + i);
	}
948
}