PostgreSQL Source Code  git master
sinvaladt.c
Go to the documentation of this file.
1 /*-------------------------------------------------------------------------
2  *
3  * sinvaladt.c
4  * POSTGRES shared cache invalidation data manager.
5  *
6  * Portions Copyright (c) 1996-2022, PostgreSQL Global Development Group
7  * Portions Copyright (c) 1994, Regents of the University of California
8  *
9  *
10  * IDENTIFICATION
11  * src/backend/storage/ipc/sinvaladt.c
12  *
13  *-------------------------------------------------------------------------
14  */
15 #include "postgres.h"
16 
17 #include <signal.h>
18 #include <unistd.h>
19 
20 #include "access/transam.h"
21 #include "miscadmin.h"
22 #include "storage/backendid.h"
23 #include "storage/ipc.h"
24 #include "storage/proc.h"
25 #include "storage/procsignal.h"
26 #include "storage/shmem.h"
27 #include "storage/sinvaladt.h"
28 #include "storage/spin.h"
29 
30 /*
31  * Conceptually, the shared cache invalidation messages are stored in an
32  * infinite array, where maxMsgNum is the next array subscript to store a
33  * submitted message in, minMsgNum is the smallest array subscript containing
34  * a message not yet read by all backends, and we always have maxMsgNum >=
35  * minMsgNum. (They are equal when there are no messages pending.) For each
36  * active backend, there is a nextMsgNum pointer indicating the next message it
37  * needs to read; we have maxMsgNum >= nextMsgNum >= minMsgNum for every
38  * backend.
39  *
40  * (In the current implementation, minMsgNum is a lower bound for the
41  * per-process nextMsgNum values, but it isn't rigorously kept equal to the
42  * smallest nextMsgNum --- it may lag behind. We only update it when
43  * SICleanupQueue is called, and we try not to do that often.)
44  *
45  * In reality, the messages are stored in a circular buffer of MAXNUMMESSAGES
46  * entries. We translate MsgNum values into circular-buffer indexes by
47  * computing MsgNum % MAXNUMMESSAGES (this should be fast as long as
48  * MAXNUMMESSAGES is a constant and a power of 2). As long as maxMsgNum
49  * doesn't exceed minMsgNum by more than MAXNUMMESSAGES, we have enough space
50  * in the buffer. If the buffer does overflow, we recover by setting the
51  * "reset" flag for each backend that has fallen too far behind. A backend
52  * that is in "reset" state is ignored while determining minMsgNum. When
53  * it does finally attempt to receive inval messages, it must discard all
54  * its invalidatable state, since it won't know what it missed.
55  *
56  * To reduce the probability of needing resets, we send a "catchup" interrupt
57  * to any backend that seems to be falling unreasonably far behind. The
58  * normal behavior is that at most one such interrupt is in flight at a time;
59  * when a backend completes processing a catchup interrupt, it executes
60  * SICleanupQueue, which will signal the next-furthest-behind backend if
61  * needed. This avoids undue contention from multiple backends all trying
62  * to catch up at once. However, the furthest-back backend might be stuck
63  * in a state where it can't catch up. Eventually it will get reset, so it
64  * won't cause any more problems for anyone but itself. But we don't want
65  * to find that a bunch of other backends are now too close to the reset
66  * threshold to be saved. So SICleanupQueue is designed to occasionally
67  * send extra catchup interrupts as the queue gets fuller, to backends that
68  * are far behind and haven't gotten one yet. As long as there aren't a lot
69  * of "stuck" backends, we won't need a lot of extra interrupts, since ones
70  * that aren't stuck will propagate their interrupts to the next guy.
71  *
72  * We would have problems if the MsgNum values overflow an integer, so
73  * whenever minMsgNum exceeds MSGNUMWRAPAROUND, we subtract MSGNUMWRAPAROUND
74  * from all the MsgNum variables simultaneously. MSGNUMWRAPAROUND can be
75  * large so that we don't need to do this often. It must be a multiple of
76  * MAXNUMMESSAGES so that the existing circular-buffer entries don't need
77  * to be moved when we do it.
78  *
79  * Access to the shared sinval array is protected by two locks, SInvalReadLock
80  * and SInvalWriteLock. Readers take SInvalReadLock in shared mode; this
81  * authorizes them to modify their own ProcState but not to modify or even
82  * look at anyone else's. When we need to perform array-wide updates,
83  * such as in SICleanupQueue, we take SInvalReadLock in exclusive mode to
84  * lock out all readers. Writers take SInvalWriteLock (always in exclusive
85  * mode) to serialize adding messages to the queue. Note that a writer
86  * can operate in parallel with one or more readers, because the writer
87  * has no need to touch anyone's ProcState, except in the infrequent cases
88  * when SICleanupQueue is needed. The only point of overlap is that
89  * the writer wants to change maxMsgNum while readers need to read it.
90  * We deal with that by having a spinlock that readers must take for just
91  * long enough to read maxMsgNum, while writers take it for just long enough
92  * to write maxMsgNum. (The exact rule is that you need the spinlock to
93  * read maxMsgNum if you are not holding SInvalWriteLock, and you need the
94  * spinlock to write maxMsgNum unless you are holding both locks.)
95  *
96  * Note: since maxMsgNum is an int and hence presumably atomically readable/
97  * writable, the spinlock might seem unnecessary. The reason it is needed
98  * is to provide a memory barrier: we need to be sure that messages written
99  * to the array are actually there before maxMsgNum is increased, and that
100  * readers will see that data after fetching maxMsgNum. Multiprocessors
101  * that have weak memory-ordering guarantees can fail without the memory
102  * barrier instructions that are included in the spinlock sequences.
103  */
104 
105 
106 /*
107  * Configurable parameters.
108  *
109  * MAXNUMMESSAGES: max number of shared-inval messages we can buffer.
110  * Must be a power of 2 for speed.
111  *
112  * MSGNUMWRAPAROUND: how often to reduce MsgNum variables to avoid overflow.
113  * Must be a multiple of MAXNUMMESSAGES. Should be large.
114  *
115  * CLEANUP_MIN: the minimum number of messages that must be in the buffer
116  * before we bother to call SICleanupQueue.
117  *
118  * CLEANUP_QUANTUM: how often (in messages) to call SICleanupQueue once
119  * we exceed CLEANUP_MIN. Should be a power of 2 for speed.
120  *
121  * SIG_THRESHOLD: the minimum number of messages a backend must have fallen
122  * behind before we'll send it PROCSIG_CATCHUP_INTERRUPT.
123  *
124  * WRITE_QUANTUM: the max number of messages to push into the buffer per
125  * iteration of SIInsertDataEntries. Noncritical but should be less than
126  * CLEANUP_QUANTUM, because we only consider calling SICleanupQueue once
127  * per iteration.
128  */
129 
130 #define MAXNUMMESSAGES 4096
131 #define MSGNUMWRAPAROUND (MAXNUMMESSAGES * 262144)
132 #define CLEANUP_MIN (MAXNUMMESSAGES / 2)
133 #define CLEANUP_QUANTUM (MAXNUMMESSAGES / 16)
134 #define SIG_THRESHOLD (MAXNUMMESSAGES / 2)
135 #define WRITE_QUANTUM 64
136 
137 /* Per-backend state in shared invalidation structure */
138 typedef struct ProcState
139 {
140  /* procPid is zero in an inactive ProcState array entry. */
141  pid_t procPid; /* PID of backend, for signaling */
142  PGPROC *proc; /* PGPROC of backend */
143  /* nextMsgNum is meaningless if procPid == 0 or resetState is true. */
144  int nextMsgNum; /* next message number to read */
145  bool resetState; /* backend needs to reset its state */
146  bool signaled; /* backend has been sent catchup signal */
147  bool hasMessages; /* backend has unread messages */
148 
149  /*
150  * Backend only sends invalidations, never receives them. This only makes
151  * sense for Startup process during recovery because it doesn't maintain a
152  * relcache, yet it fires inval messages to allow query backends to see
153  * schema changes.
154  */
155  bool sendOnly; /* backend only sends, never receives */
156 
157  /*
158  * Next LocalTransactionId to use for each idle backend slot. We keep
159  * this here because it is indexed by BackendId and it is convenient to
160  * copy the value to and from local memory when MyBackendId is set. It's
161  * meaningless in an active ProcState entry.
162  */
165 
166 /* Shared cache invalidation memory segment */
167 typedef struct SISeg
168 {
169  /*
170  * General state information
171  */
172  int minMsgNum; /* oldest message still needed */
173  int maxMsgNum; /* next message number to be assigned */
174  int nextThreshold; /* # of messages to call SICleanupQueue */
175  int lastBackend; /* index of last active procState entry, +1 */
176  int maxBackends; /* size of procState array */
177 
178  slock_t msgnumLock; /* spinlock protecting maxMsgNum */
179 
180  /*
181  * Circular buffer holding shared-inval messages
182  */
184 
185  /*
186  * Per-backend invalidation state info (has MaxBackends entries).
187  */
190 
191 static SISeg *shmInvalBuffer; /* pointer to the shared inval buffer */
192 
193 
195 
196 static void CleanupInvalidationState(int status, Datum arg);
197 
198 
199 /*
200  * SInvalShmemSize --- return shared-memory space needed
201  */
202 Size
204 {
205  Size size;
206 
207  size = offsetof(SISeg, procState);
208 
209  /*
210  * In Hot Standby mode, the startup process requests a procState array
211  * slot using InitRecoveryTransactionEnvironment(). Even though
212  * MaxBackends doesn't account for the startup process, it is guaranteed
213  * to get a free slot. This is because the autovacuum launcher and worker
214  * processes, which are included in MaxBackends, are not started in Hot
215  * Standby mode.
216  */
217  size = add_size(size, mul_size(sizeof(ProcState), MaxBackends));
218 
219  return size;
220 }
221 
222 /*
223  * CreateSharedInvalidationState
224  * Create and initialize the SI message buffer
225  */
226 void
228 {
229  int i;
230  bool found;
231 
232  /* Allocate space in shared memory */
233  shmInvalBuffer = (SISeg *)
234  ShmemInitStruct("shmInvalBuffer", SInvalShmemSize(), &found);
235  if (found)
236  return;
237 
238  /* Clear message counters, save size of procState array, init spinlock */
245 
246  /* The buffer[] array is initially all unused, so we need not fill it */
247 
248  /* Mark all backends inactive, and initialize nextLXID */
249  for (i = 0; i < shmInvalBuffer->maxBackends; i++)
250  {
251  shmInvalBuffer->procState[i].procPid = 0; /* inactive */
252  shmInvalBuffer->procState[i].proc = NULL;
253  shmInvalBuffer->procState[i].nextMsgNum = 0; /* meaningless */
258  }
259 }
260 
261 /*
262  * SharedInvalBackendInit
263  * Initialize a new backend to operate on the sinval buffer
264  */
265 void
267 {
268  int index;
269  ProcState *stateP = NULL;
270  SISeg *segP = shmInvalBuffer;
271 
272  /*
273  * This can run in parallel with read operations, but not with write
274  * operations, since SIInsertDataEntries relies on lastBackend to set
275  * hasMessages appropriately.
276  */
277  LWLockAcquire(SInvalWriteLock, LW_EXCLUSIVE);
278 
279  /* Look for a free entry in the procState array */
280  for (index = 0; index < segP->lastBackend; index++)
281  {
282  if (segP->procState[index].procPid == 0) /* inactive slot? */
283  {
284  stateP = &segP->procState[index];
285  break;
286  }
287  }
288 
289  if (stateP == NULL)
290  {
291  if (segP->lastBackend < segP->maxBackends)
292  {
293  stateP = &segP->procState[segP->lastBackend];
294  Assert(stateP->procPid == 0);
295  segP->lastBackend++;
296  }
297  else
298  {
299  /*
300  * out of procState slots: MaxBackends exceeded -- report normally
301  */
303  LWLockRelease(SInvalWriteLock);
304  ereport(FATAL,
305  (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
306  errmsg("sorry, too many clients already")));
307  }
308  }
309 
310  MyBackendId = (stateP - &segP->procState[0]) + 1;
311 
312  /* Advertise assigned backend ID in MyProc */
314 
315  /* Fetch next local transaction ID into local memory */
317 
318  /* mark myself active, with all extant messages already read */
319  stateP->procPid = MyProcPid;
320  stateP->proc = MyProc;
321  stateP->nextMsgNum = segP->maxMsgNum;
322  stateP->resetState = false;
323  stateP->signaled = false;
324  stateP->hasMessages = false;
325  stateP->sendOnly = sendOnly;
326 
327  LWLockRelease(SInvalWriteLock);
328 
329  /* register exit routine to mark my entry inactive at exit */
331 
332  elog(DEBUG4, "my backend ID is %d", MyBackendId);
333 }
334 
335 /*
336  * CleanupInvalidationState
337  * Mark the current backend as no longer active.
338  *
339  * This function is called via on_shmem_exit() during backend shutdown.
340  *
341  * arg is really of type "SISeg*".
342  */
343 static void
345 {
346  SISeg *segP = (SISeg *) DatumGetPointer(arg);
347  ProcState *stateP;
348  int i;
349 
350  Assert(PointerIsValid(segP));
351 
352  LWLockAcquire(SInvalWriteLock, LW_EXCLUSIVE);
353 
354  stateP = &segP->procState[MyBackendId - 1];
355 
356  /* Update next local transaction ID for next holder of this backendID */
358 
359  /* Mark myself inactive */
360  stateP->procPid = 0;
361  stateP->proc = NULL;
362  stateP->nextMsgNum = 0;
363  stateP->resetState = false;
364  stateP->signaled = false;
365 
366  /* Recompute index of last active backend */
367  for (i = segP->lastBackend; i > 0; i--)
368  {
369  if (segP->procState[i - 1].procPid != 0)
370  break;
371  }
372  segP->lastBackend = i;
373 
374  LWLockRelease(SInvalWriteLock);
375 }
376 
377 /*
378  * BackendIdGetProc
379  * Get the PGPROC structure for a backend, given the backend ID.
380  * The result may be out of date arbitrarily quickly, so the caller
381  * must be careful about how this information is used. NULL is
382  * returned if the backend is not active.
383  */
384 PGPROC *
385 BackendIdGetProc(int backendID)
386 {
387  PGPROC *result = NULL;
388  SISeg *segP = shmInvalBuffer;
389 
390  /* Need to lock out additions/removals of backends */
391  LWLockAcquire(SInvalWriteLock, LW_SHARED);
392 
393  if (backendID > 0 && backendID <= segP->lastBackend)
394  {
395  ProcState *stateP = &segP->procState[backendID - 1];
396 
397  result = stateP->proc;
398  }
399 
400  LWLockRelease(SInvalWriteLock);
401 
402  return result;
403 }
404 
405 /*
406  * BackendIdGetTransactionIds
407  * Get the xid and xmin of the backend. The result may be out of date
408  * arbitrarily quickly, so the caller must be careful about how this
409  * information is used.
410  */
411 void
413 {
414  SISeg *segP = shmInvalBuffer;
415 
416  *xid = InvalidTransactionId;
417  *xmin = InvalidTransactionId;
418 
419  /* Need to lock out additions/removals of backends */
420  LWLockAcquire(SInvalWriteLock, LW_SHARED);
421 
422  if (backendID > 0 && backendID <= segP->lastBackend)
423  {
424  ProcState *stateP = &segP->procState[backendID - 1];
425  PGPROC *proc = stateP->proc;
426 
427  if (proc != NULL)
428  {
429  *xid = proc->xid;
430  *xmin = proc->xmin;
431  }
432  }
433 
434  LWLockRelease(SInvalWriteLock);
435 }
436 
437 /*
438  * SIInsertDataEntries
439  * Add new invalidation message(s) to the buffer.
440  */
441 void
443 {
444  SISeg *segP = shmInvalBuffer;
445 
446  /*
447  * N can be arbitrarily large. We divide the work into groups of no more
448  * than WRITE_QUANTUM messages, to be sure that we don't hold the lock for
449  * an unreasonably long time. (This is not so much because we care about
450  * letting in other writers, as that some just-caught-up backend might be
451  * trying to do SICleanupQueue to pass on its signal, and we don't want it
452  * to have to wait a long time.) Also, we need to consider calling
453  * SICleanupQueue every so often.
454  */
455  while (n > 0)
456  {
457  int nthistime = Min(n, WRITE_QUANTUM);
458  int numMsgs;
459  int max;
460  int i;
461 
462  n -= nthistime;
463 
464  LWLockAcquire(SInvalWriteLock, LW_EXCLUSIVE);
465 
466  /*
467  * If the buffer is full, we *must* acquire some space. Clean the
468  * queue and reset anyone who is preventing space from being freed.
469  * Otherwise, clean the queue only when it's exceeded the next
470  * fullness threshold. We have to loop and recheck the buffer state
471  * after any call of SICleanupQueue.
472  */
473  for (;;)
474  {
475  numMsgs = segP->maxMsgNum - segP->minMsgNum;
476  if (numMsgs + nthistime > MAXNUMMESSAGES ||
477  numMsgs >= segP->nextThreshold)
478  SICleanupQueue(true, nthistime);
479  else
480  break;
481  }
482 
483  /*
484  * Insert new message(s) into proper slot of circular buffer
485  */
486  max = segP->maxMsgNum;
487  while (nthistime-- > 0)
488  {
489  segP->buffer[max % MAXNUMMESSAGES] = *data++;
490  max++;
491  }
492 
493  /* Update current value of maxMsgNum using spinlock */
494  SpinLockAcquire(&segP->msgnumLock);
495  segP->maxMsgNum = max;
496  SpinLockRelease(&segP->msgnumLock);
497 
498  /*
499  * Now that the maxMsgNum change is globally visible, we give everyone
500  * a swift kick to make sure they read the newly added messages.
501  * Releasing SInvalWriteLock will enforce a full memory barrier, so
502  * these (unlocked) changes will be committed to memory before we exit
503  * the function.
504  */
505  for (i = 0; i < segP->lastBackend; i++)
506  {
507  ProcState *stateP = &segP->procState[i];
508 
509  stateP->hasMessages = true;
510  }
511 
512  LWLockRelease(SInvalWriteLock);
513  }
514 }
515 
516 /*
517  * SIGetDataEntries
518  * get next SI message(s) for current backend, if there are any
519  *
520  * Possible return values:
521  * 0: no SI message available
522  * n>0: next n SI messages have been extracted into data[]
523  * -1: SI reset message extracted
524  *
525  * If the return value is less than the array size "datasize", the caller
526  * can assume that there are no more SI messages after the one(s) returned.
527  * Otherwise, another call is needed to collect more messages.
528  *
529  * NB: this can run in parallel with other instances of SIGetDataEntries
530  * executing on behalf of other backends, since each instance will modify only
531  * fields of its own backend's ProcState, and no instance will look at fields
532  * of other backends' ProcStates. We express this by grabbing SInvalReadLock
533  * in shared mode. Note that this is not exactly the normal (read-only)
534  * interpretation of a shared lock! Look closely at the interactions before
535  * allowing SInvalReadLock to be grabbed in shared mode for any other reason!
536  *
537  * NB: this can also run in parallel with SIInsertDataEntries. It is not
538  * guaranteed that we will return any messages added after the routine is
539  * entered.
540  *
541  * Note: we assume that "datasize" is not so large that it might be important
542  * to break our hold on SInvalReadLock into segments.
543  */
544 int
546 {
547  SISeg *segP;
548  ProcState *stateP;
549  int max;
550  int n;
551 
552  segP = shmInvalBuffer;
553  stateP = &segP->procState[MyBackendId - 1];
554 
555  /*
556  * Before starting to take locks, do a quick, unlocked test to see whether
557  * there can possibly be anything to read. On a multiprocessor system,
558  * it's possible that this load could migrate backwards and occur before
559  * we actually enter this function, so we might miss a sinval message that
560  * was just added by some other processor. But they can't migrate
561  * backwards over a preceding lock acquisition, so it should be OK. If we
562  * haven't acquired a lock preventing against further relevant
563  * invalidations, any such occurrence is not much different than if the
564  * invalidation had arrived slightly later in the first place.
565  */
566  if (!stateP->hasMessages)
567  return 0;
568 
569  LWLockAcquire(SInvalReadLock, LW_SHARED);
570 
571  /*
572  * We must reset hasMessages before determining how many messages we're
573  * going to read. That way, if new messages arrive after we have
574  * determined how many we're reading, the flag will get reset and we'll
575  * notice those messages part-way through.
576  *
577  * Note that, if we don't end up reading all of the messages, we had
578  * better be certain to reset this flag before exiting!
579  */
580  stateP->hasMessages = false;
581 
582  /* Fetch current value of maxMsgNum using spinlock */
583  SpinLockAcquire(&segP->msgnumLock);
584  max = segP->maxMsgNum;
585  SpinLockRelease(&segP->msgnumLock);
586 
587  if (stateP->resetState)
588  {
589  /*
590  * Force reset. We can say we have dealt with any messages added
591  * since the reset, as well; and that means we should clear the
592  * signaled flag, too.
593  */
594  stateP->nextMsgNum = max;
595  stateP->resetState = false;
596  stateP->signaled = false;
597  LWLockRelease(SInvalReadLock);
598  return -1;
599  }
600 
601  /*
602  * Retrieve messages and advance backend's counter, until data array is
603  * full or there are no more messages.
604  *
605  * There may be other backends that haven't read the message(s), so we
606  * cannot delete them here. SICleanupQueue() will eventually remove them
607  * from the queue.
608  */
609  n = 0;
610  while (n < datasize && stateP->nextMsgNum < max)
611  {
612  data[n++] = segP->buffer[stateP->nextMsgNum % MAXNUMMESSAGES];
613  stateP->nextMsgNum++;
614  }
615 
616  /*
617  * If we have caught up completely, reset our "signaled" flag so that
618  * we'll get another signal if we fall behind again.
619  *
620  * If we haven't caught up completely, reset the hasMessages flag so that
621  * we see the remaining messages next time.
622  */
623  if (stateP->nextMsgNum >= max)
624  stateP->signaled = false;
625  else
626  stateP->hasMessages = true;
627 
628  LWLockRelease(SInvalReadLock);
629  return n;
630 }
631 
632 /*
633  * SICleanupQueue
634  * Remove messages that have been consumed by all active backends
635  *
636  * callerHasWriteLock is true if caller is holding SInvalWriteLock.
637  * minFree is the minimum number of message slots to make free.
638  *
639  * Possible side effects of this routine include marking one or more
640  * backends as "reset" in the array, and sending PROCSIG_CATCHUP_INTERRUPT
641  * to some backend that seems to be getting too far behind. We signal at
642  * most one backend at a time, for reasons explained at the top of the file.
643  *
644  * Caution: because we transiently release write lock when we have to signal
645  * some other backend, it is NOT guaranteed that there are still minFree
646  * free message slots at exit. Caller must recheck and perhaps retry.
647  */
648 void
649 SICleanupQueue(bool callerHasWriteLock, int minFree)
650 {
651  SISeg *segP = shmInvalBuffer;
652  int min,
653  minsig,
654  lowbound,
655  numMsgs,
656  i;
657  ProcState *needSig = NULL;
658 
659  /* Lock out all writers and readers */
660  if (!callerHasWriteLock)
661  LWLockAcquire(SInvalWriteLock, LW_EXCLUSIVE);
662  LWLockAcquire(SInvalReadLock, LW_EXCLUSIVE);
663 
664  /*
665  * Recompute minMsgNum = minimum of all backends' nextMsgNum, identify the
666  * furthest-back backend that needs signaling (if any), and reset any
667  * backends that are too far back. Note that because we ignore sendOnly
668  * backends here it is possible for them to keep sending messages without
669  * a problem even when they are the only active backend.
670  */
671  min = segP->maxMsgNum;
672  minsig = min - SIG_THRESHOLD;
673  lowbound = min - MAXNUMMESSAGES + minFree;
674 
675  for (i = 0; i < segP->lastBackend; i++)
676  {
677  ProcState *stateP = &segP->procState[i];
678  int n = stateP->nextMsgNum;
679 
680  /* Ignore if inactive or already in reset state */
681  if (stateP->procPid == 0 || stateP->resetState || stateP->sendOnly)
682  continue;
683 
684  /*
685  * If we must free some space and this backend is preventing it, force
686  * him into reset state and then ignore until he catches up.
687  */
688  if (n < lowbound)
689  {
690  stateP->resetState = true;
691  /* no point in signaling him ... */
692  continue;
693  }
694 
695  /* Track the global minimum nextMsgNum */
696  if (n < min)
697  min = n;
698 
699  /* Also see who's furthest back of the unsignaled backends */
700  if (n < minsig && !stateP->signaled)
701  {
702  minsig = n;
703  needSig = stateP;
704  }
705  }
706  segP->minMsgNum = min;
707 
708  /*
709  * When minMsgNum gets really large, decrement all message counters so as
710  * to forestall overflow of the counters. This happens seldom enough that
711  * folding it into the previous loop would be a loser.
712  */
713  if (min >= MSGNUMWRAPAROUND)
714  {
715  segP->minMsgNum -= MSGNUMWRAPAROUND;
716  segP->maxMsgNum -= MSGNUMWRAPAROUND;
717  for (i = 0; i < segP->lastBackend; i++)
718  {
719  /* we don't bother skipping inactive entries here */
721  }
722  }
723 
724  /*
725  * Determine how many messages are still in the queue, and set the
726  * threshold at which we should repeat SICleanupQueue().
727  */
728  numMsgs = segP->maxMsgNum - segP->minMsgNum;
729  if (numMsgs < CLEANUP_MIN)
730  segP->nextThreshold = CLEANUP_MIN;
731  else
732  segP->nextThreshold = (numMsgs / CLEANUP_QUANTUM + 1) * CLEANUP_QUANTUM;
733 
734  /*
735  * Lastly, signal anyone who needs a catchup interrupt. Since
736  * SendProcSignal() might not be fast, we don't want to hold locks while
737  * executing it.
738  */
739  if (needSig)
740  {
741  pid_t his_pid = needSig->procPid;
742  BackendId his_backendId = (needSig - &segP->procState[0]) + 1;
743 
744  needSig->signaled = true;
745  LWLockRelease(SInvalReadLock);
746  LWLockRelease(SInvalWriteLock);
747  elog(DEBUG4, "sending sinval catchup signal to PID %d", (int) his_pid);
748  SendProcSignal(his_pid, PROCSIG_CATCHUP_INTERRUPT, his_backendId);
749  if (callerHasWriteLock)
750  LWLockAcquire(SInvalWriteLock, LW_EXCLUSIVE);
751  }
752  else
753  {
754  LWLockRelease(SInvalReadLock);
755  if (!callerHasWriteLock)
756  LWLockRelease(SInvalWriteLock);
757  }
758 }
759 
760 
761 /*
762  * GetNextLocalTransactionId --- allocate a new LocalTransactionId
763  *
764  * We split VirtualTransactionIds into two parts so that it is possible
765  * to allocate a new one without any contention for shared memory, except
766  * for a bit of additional overhead during backend startup/shutdown.
767  * The high-order part of a VirtualTransactionId is a BackendId, and the
768  * low-order part is a LocalTransactionId, which we assign from a local
769  * counter. To avoid the risk of a VirtualTransactionId being reused
770  * within a short interval, successive procs occupying the same backend ID
771  * slot should use a consecutive sequence of local IDs, which is implemented
772  * by copying nextLocalTransactionId as seen above.
773  */
776 {
777  LocalTransactionId result;
778 
779  /* loop to avoid returning InvalidLocalTransactionId at wraparound */
780  do
781  {
782  result = nextLocalTransactionId++;
783  } while (!LocalTransactionIdIsValid(result));
784 
785  return result;
786 }
int BackendId
Definition: backendid.h:21
#define InvalidBackendId
Definition: backendid.h:23
#define Min(x, y)
Definition: c.h:937
#define PointerIsValid(pointer)
Definition: c.h:699
#define FLEXIBLE_ARRAY_MEMBER
Definition: c.h:362
uint32 LocalTransactionId
Definition: c.h:590
uint32 TransactionId
Definition: c.h:588
size_t Size
Definition: c.h:541
int errcode(int sqlerrcode)
Definition: elog.c:695
int errmsg(const char *fmt,...)
Definition: elog.c:906
#define FATAL
Definition: elog.h:37
#define ereport(elevel,...)
Definition: elog.h:145
#define DEBUG4
Definition: elog.h:23
int MyProcPid
Definition: globals.c:44
BackendId MyBackendId
Definition: globals.c:85
int MaxBackends
Definition: globals.c:140
void on_shmem_exit(pg_on_exit_callback function, Datum arg)
Definition: ipc.c:361
int i
Definition: isn.c:73
Assert(fmt[strlen(fmt) - 1] !='\n')
#define InvalidLocalTransactionId
Definition: lock.h:70
#define LocalTransactionIdIsValid(lxid)
Definition: lock.h:71
bool LWLockAcquire(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:1194
void LWLockRelease(LWLock *lock)
Definition: lwlock.c:1802
@ LW_SHARED
Definition: lwlock.h:113
@ LW_EXCLUSIVE
Definition: lwlock.h:112
void * arg
const void * data
static void static void status(const char *fmt,...) pg_attribute_printf(1
Definition: pg_regress.c:225
static Datum PointerGetDatum(const void *X)
Definition: postgres.h:670
uintptr_t Datum
Definition: postgres.h:412
static Pointer DatumGetPointer(Datum X)
Definition: postgres.h:660
int SendProcSignal(pid_t pid, ProcSignalReason reason, BackendId backendId)
Definition: procsignal.c:261
@ PROCSIG_CATCHUP_INTERRUPT
Definition: procsignal.h:32
int slock_t
Definition: s_lock.h:754
Size add_size(Size s1, Size s2)
Definition: shmem.c:502
void * ShmemInitStruct(const char *name, Size size, bool *foundPtr)
Definition: shmem.c:396
Size mul_size(Size s1, Size s2)
Definition: shmem.c:519
#define CLEANUP_QUANTUM
Definition: sinvaladt.c:133
struct ProcState ProcState
static SISeg * shmInvalBuffer
Definition: sinvaladt.c:191
#define WRITE_QUANTUM
Definition: sinvaladt.c:135
void SICleanupQueue(bool callerHasWriteLock, int minFree)
Definition: sinvaladt.c:649
#define MAXNUMMESSAGES
Definition: sinvaladt.c:130
void BackendIdGetTransactionIds(int backendID, TransactionId *xid, TransactionId *xmin)
Definition: sinvaladt.c:412
#define MSGNUMWRAPAROUND
Definition: sinvaladt.c:131
static void CleanupInvalidationState(int status, Datum arg)
Definition: sinvaladt.c:344
Size SInvalShmemSize(void)
Definition: sinvaladt.c:203
#define SIG_THRESHOLD
Definition: sinvaladt.c:134
void SharedInvalBackendInit(bool sendOnly)
Definition: sinvaladt.c:266
int SIGetDataEntries(SharedInvalidationMessage *data, int datasize)
Definition: sinvaladt.c:545
struct SISeg SISeg
void CreateSharedInvalidationState(void)
Definition: sinvaladt.c:227
void SIInsertDataEntries(const SharedInvalidationMessage *data, int n)
Definition: sinvaladt.c:442
LocalTransactionId GetNextLocalTransactionId(void)
Definition: sinvaladt.c:775
static LocalTransactionId nextLocalTransactionId
Definition: sinvaladt.c:194
PGPROC * BackendIdGetProc(int backendID)
Definition: sinvaladt.c:385
#define CLEANUP_MIN
Definition: sinvaladt.c:132
#define SpinLockInit(lock)
Definition: spin.h:60
#define SpinLockRelease(lock)
Definition: spin.h:64
#define SpinLockAcquire(lock)
Definition: spin.h:62
PGPROC * MyProc
Definition: proc.c:68
Definition: proc.h:162
TransactionId xmin
Definition: proc.h:178
BackendId backendId
Definition: proc.h:197
TransactionId xid
Definition: proc.h:173
int nextMsgNum
Definition: sinvaladt.c:144
PGPROC * proc
Definition: sinvaladt.c:142
bool signaled
Definition: sinvaladt.c:146
LocalTransactionId nextLXID
Definition: sinvaladt.c:163
pid_t procPid
Definition: sinvaladt.c:141
bool hasMessages
Definition: sinvaladt.c:147
bool sendOnly
Definition: sinvaladt.c:155
bool resetState
Definition: sinvaladt.c:145
SharedInvalidationMessage buffer[MAXNUMMESSAGES]
Definition: sinvaladt.c:183
int minMsgNum
Definition: sinvaladt.c:172
int maxMsgNum
Definition: sinvaladt.c:173
int lastBackend
Definition: sinvaladt.c:175
slock_t msgnumLock
Definition: sinvaladt.c:178
int nextThreshold
Definition: sinvaladt.c:174
int maxBackends
Definition: sinvaladt.c:176
ProcState procState[FLEXIBLE_ARRAY_MEMBER]
Definition: sinvaladt.c:188
Definition: type.h:95
#define InvalidTransactionId
Definition: transam.h:31