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) 1996-2020, 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  */
164 } ProcState;
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  */
189 } SISeg;
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  size = add_size(size, mul_size(sizeof(ProcState), MaxBackends));
209 
210  return size;
211 }
212 
213 /*
214  * CreateSharedInvalidationState
215  * Create and initialize the SI message buffer
216  */
217 void
219 {
220  int i;
221  bool found;
222 
223  /* Allocate space in shared memory */
224  shmInvalBuffer = (SISeg *)
225  ShmemInitStruct("shmInvalBuffer", SInvalShmemSize(), &found);
226  if (found)
227  return;
228 
229  /* Clear message counters, save size of procState array, init spinlock */
230  shmInvalBuffer->minMsgNum = 0;
231  shmInvalBuffer->maxMsgNum = 0;
232  shmInvalBuffer->nextThreshold = CLEANUP_MIN;
233  shmInvalBuffer->lastBackend = 0;
234  shmInvalBuffer->maxBackends = MaxBackends;
235  SpinLockInit(&shmInvalBuffer->msgnumLock);
236 
237  /* The buffer[] array is initially all unused, so we need not fill it */
238 
239  /* Mark all backends inactive, and initialize nextLXID */
240  for (i = 0; i < shmInvalBuffer->maxBackends; i++)
241  {
242  shmInvalBuffer->procState[i].procPid = 0; /* inactive */
243  shmInvalBuffer->procState[i].proc = NULL;
244  shmInvalBuffer->procState[i].nextMsgNum = 0; /* meaningless */
245  shmInvalBuffer->procState[i].resetState = false;
246  shmInvalBuffer->procState[i].signaled = false;
247  shmInvalBuffer->procState[i].hasMessages = false;
248  shmInvalBuffer->procState[i].nextLXID = InvalidLocalTransactionId;
249  }
250 }
251 
252 /*
253  * SharedInvalBackendInit
254  * Initialize a new backend to operate on the sinval buffer
255  */
256 void
258 {
259  int index;
260  ProcState *stateP = NULL;
261  SISeg *segP = shmInvalBuffer;
262 
263  /*
264  * This can run in parallel with read operations, but not with write
265  * operations, since SIInsertDataEntries relies on lastBackend to set
266  * hasMessages appropriately.
267  */
268  LWLockAcquire(SInvalWriteLock, LW_EXCLUSIVE);
269 
270  /* Look for a free entry in the procState array */
271  for (index = 0; index < segP->lastBackend; index++)
272  {
273  if (segP->procState[index].procPid == 0) /* inactive slot? */
274  {
275  stateP = &segP->procState[index];
276  break;
277  }
278  }
279 
280  if (stateP == NULL)
281  {
282  if (segP->lastBackend < segP->maxBackends)
283  {
284  stateP = &segP->procState[segP->lastBackend];
285  Assert(stateP->procPid == 0);
286  segP->lastBackend++;
287  }
288  else
289  {
290  /*
291  * out of procState slots: MaxBackends exceeded -- report normally
292  */
294  LWLockRelease(SInvalWriteLock);
295  ereport(FATAL,
296  (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
297  errmsg("sorry, too many clients already")));
298  }
299  }
300 
301  MyBackendId = (stateP - &segP->procState[0]) + 1;
302 
303  /* Advertise assigned backend ID in MyProc */
305 
306  /* Fetch next local transaction ID into local memory */
308 
309  /* mark myself active, with all extant messages already read */
310  stateP->procPid = MyProcPid;
311  stateP->proc = MyProc;
312  stateP->nextMsgNum = segP->maxMsgNum;
313  stateP->resetState = false;
314  stateP->signaled = false;
315  stateP->hasMessages = false;
316  stateP->sendOnly = sendOnly;
317 
318  LWLockRelease(SInvalWriteLock);
319 
320  /* register exit routine to mark my entry inactive at exit */
322 
323  elog(DEBUG4, "my backend ID is %d", MyBackendId);
324 }
325 
326 /*
327  * CleanupInvalidationState
328  * Mark the current backend as no longer active.
329  *
330  * This function is called via on_shmem_exit() during backend shutdown.
331  *
332  * arg is really of type "SISeg*".
333  */
334 static void
336 {
337  SISeg *segP = (SISeg *) DatumGetPointer(arg);
338  ProcState *stateP;
339  int i;
340 
341  Assert(PointerIsValid(segP));
342 
343  LWLockAcquire(SInvalWriteLock, LW_EXCLUSIVE);
344 
345  stateP = &segP->procState[MyBackendId - 1];
346 
347  /* Update next local transaction ID for next holder of this backendID */
349 
350  /* Mark myself inactive */
351  stateP->procPid = 0;
352  stateP->proc = NULL;
353  stateP->nextMsgNum = 0;
354  stateP->resetState = false;
355  stateP->signaled = false;
356 
357  /* Recompute index of last active backend */
358  for (i = segP->lastBackend; i > 0; i--)
359  {
360  if (segP->procState[i - 1].procPid != 0)
361  break;
362  }
363  segP->lastBackend = i;
364 
365  LWLockRelease(SInvalWriteLock);
366 }
367 
368 /*
369  * BackendIdGetProc
370  * Get the PGPROC structure for a backend, given the backend ID.
371  * The result may be out of date arbitrarily quickly, so the caller
372  * must be careful about how this information is used. NULL is
373  * returned if the backend is not active.
374  */
375 PGPROC *
376 BackendIdGetProc(int backendID)
377 {
378  PGPROC *result = NULL;
379  SISeg *segP = shmInvalBuffer;
380 
381  /* Need to lock out additions/removals of backends */
382  LWLockAcquire(SInvalWriteLock, LW_SHARED);
383 
384  if (backendID > 0 && backendID <= segP->lastBackend)
385  {
386  ProcState *stateP = &segP->procState[backendID - 1];
387 
388  result = stateP->proc;
389  }
390 
391  LWLockRelease(SInvalWriteLock);
392 
393  return result;
394 }
395 
396 /*
397  * BackendIdGetTransactionIds
398  * Get the xid and xmin of the backend. The result may be out of date
399  * arbitrarily quickly, so the caller must be careful about how this
400  * information is used.
401  */
402 void
404 {
405  SISeg *segP = shmInvalBuffer;
406 
407  *xid = InvalidTransactionId;
408  *xmin = InvalidTransactionId;
409 
410  /* Need to lock out additions/removals of backends */
411  LWLockAcquire(SInvalWriteLock, LW_SHARED);
412 
413  if (backendID > 0 && backendID <= segP->lastBackend)
414  {
415  ProcState *stateP = &segP->procState[backendID - 1];
416  PGPROC *proc = stateP->proc;
417 
418  if (proc != NULL)
419  {
420  *xid = proc->xid;
421  *xmin = proc->xmin;
422  }
423  }
424 
425  LWLockRelease(SInvalWriteLock);
426 }
427 
428 /*
429  * SIInsertDataEntries
430  * Add new invalidation message(s) to the buffer.
431  */
432 void
434 {
435  SISeg *segP = shmInvalBuffer;
436 
437  /*
438  * N can be arbitrarily large. We divide the work into groups of no more
439  * than WRITE_QUANTUM messages, to be sure that we don't hold the lock for
440  * an unreasonably long time. (This is not so much because we care about
441  * letting in other writers, as that some just-caught-up backend might be
442  * trying to do SICleanupQueue to pass on its signal, and we don't want it
443  * to have to wait a long time.) Also, we need to consider calling
444  * SICleanupQueue every so often.
445  */
446  while (n > 0)
447  {
448  int nthistime = Min(n, WRITE_QUANTUM);
449  int numMsgs;
450  int max;
451  int i;
452 
453  n -= nthistime;
454 
455  LWLockAcquire(SInvalWriteLock, LW_EXCLUSIVE);
456 
457  /*
458  * If the buffer is full, we *must* acquire some space. Clean the
459  * queue and reset anyone who is preventing space from being freed.
460  * Otherwise, clean the queue only when it's exceeded the next
461  * fullness threshold. We have to loop and recheck the buffer state
462  * after any call of SICleanupQueue.
463  */
464  for (;;)
465  {
466  numMsgs = segP->maxMsgNum - segP->minMsgNum;
467  if (numMsgs + nthistime > MAXNUMMESSAGES ||
468  numMsgs >= segP->nextThreshold)
469  SICleanupQueue(true, nthistime);
470  else
471  break;
472  }
473 
474  /*
475  * Insert new message(s) into proper slot of circular buffer
476  */
477  max = segP->maxMsgNum;
478  while (nthistime-- > 0)
479  {
480  segP->buffer[max % MAXNUMMESSAGES] = *data++;
481  max++;
482  }
483 
484  /* Update current value of maxMsgNum using spinlock */
485  SpinLockAcquire(&segP->msgnumLock);
486  segP->maxMsgNum = max;
487  SpinLockRelease(&segP->msgnumLock);
488 
489  /*
490  * Now that the maxMsgNum change is globally visible, we give everyone
491  * a swift kick to make sure they read the newly added messages.
492  * Releasing SInvalWriteLock will enforce a full memory barrier, so
493  * these (unlocked) changes will be committed to memory before we exit
494  * the function.
495  */
496  for (i = 0; i < segP->lastBackend; i++)
497  {
498  ProcState *stateP = &segP->procState[i];
499 
500  stateP->hasMessages = true;
501  }
502 
503  LWLockRelease(SInvalWriteLock);
504  }
505 }
506 
507 /*
508  * SIGetDataEntries
509  * get next SI message(s) for current backend, if there are any
510  *
511  * Possible return values:
512  * 0: no SI message available
513  * n>0: next n SI messages have been extracted into data[]
514  * -1: SI reset message extracted
515  *
516  * If the return value is less than the array size "datasize", the caller
517  * can assume that there are no more SI messages after the one(s) returned.
518  * Otherwise, another call is needed to collect more messages.
519  *
520  * NB: this can run in parallel with other instances of SIGetDataEntries
521  * executing on behalf of other backends, since each instance will modify only
522  * fields of its own backend's ProcState, and no instance will look at fields
523  * of other backends' ProcStates. We express this by grabbing SInvalReadLock
524  * in shared mode. Note that this is not exactly the normal (read-only)
525  * interpretation of a shared lock! Look closely at the interactions before
526  * allowing SInvalReadLock to be grabbed in shared mode for any other reason!
527  *
528  * NB: this can also run in parallel with SIInsertDataEntries. It is not
529  * guaranteed that we will return any messages added after the routine is
530  * entered.
531  *
532  * Note: we assume that "datasize" is not so large that it might be important
533  * to break our hold on SInvalReadLock into segments.
534  */
535 int
537 {
538  SISeg *segP;
539  ProcState *stateP;
540  int max;
541  int n;
542 
543  segP = shmInvalBuffer;
544  stateP = &segP->procState[MyBackendId - 1];
545 
546  /*
547  * Before starting to take locks, do a quick, unlocked test to see whether
548  * there can possibly be anything to read. On a multiprocessor system,
549  * it's possible that this load could migrate backwards and occur before
550  * we actually enter this function, so we might miss a sinval message that
551  * was just added by some other processor. But they can't migrate
552  * backwards over a preceding lock acquisition, so it should be OK. If we
553  * haven't acquired a lock preventing against further relevant
554  * invalidations, any such occurrence is not much different than if the
555  * invalidation had arrived slightly later in the first place.
556  */
557  if (!stateP->hasMessages)
558  return 0;
559 
560  LWLockAcquire(SInvalReadLock, LW_SHARED);
561 
562  /*
563  * We must reset hasMessages before determining how many messages we're
564  * going to read. That way, if new messages arrive after we have
565  * determined how many we're reading, the flag will get reset and we'll
566  * notice those messages part-way through.
567  *
568  * Note that, if we don't end up reading all of the messages, we had
569  * better be certain to reset this flag before exiting!
570  */
571  stateP->hasMessages = false;
572 
573  /* Fetch current value of maxMsgNum using spinlock */
574  SpinLockAcquire(&segP->msgnumLock);
575  max = segP->maxMsgNum;
576  SpinLockRelease(&segP->msgnumLock);
577 
578  if (stateP->resetState)
579  {
580  /*
581  * Force reset. We can say we have dealt with any messages added
582  * since the reset, as well; and that means we should clear the
583  * signaled flag, too.
584  */
585  stateP->nextMsgNum = max;
586  stateP->resetState = false;
587  stateP->signaled = false;
588  LWLockRelease(SInvalReadLock);
589  return -1;
590  }
591 
592  /*
593  * Retrieve messages and advance backend's counter, until data array is
594  * full or there are no more messages.
595  *
596  * There may be other backends that haven't read the message(s), so we
597  * cannot delete them here. SICleanupQueue() will eventually remove them
598  * from the queue.
599  */
600  n = 0;
601  while (n < datasize && stateP->nextMsgNum < max)
602  {
603  data[n++] = segP->buffer[stateP->nextMsgNum % MAXNUMMESSAGES];
604  stateP->nextMsgNum++;
605  }
606 
607  /*
608  * If we have caught up completely, reset our "signaled" flag so that
609  * we'll get another signal if we fall behind again.
610  *
611  * If we haven't caught up completely, reset the hasMessages flag so that
612  * we see the remaining messages next time.
613  */
614  if (stateP->nextMsgNum >= max)
615  stateP->signaled = false;
616  else
617  stateP->hasMessages = true;
618 
619  LWLockRelease(SInvalReadLock);
620  return n;
621 }
622 
623 /*
624  * SICleanupQueue
625  * Remove messages that have been consumed by all active backends
626  *
627  * callerHasWriteLock is true if caller is holding SInvalWriteLock.
628  * minFree is the minimum number of message slots to make free.
629  *
630  * Possible side effects of this routine include marking one or more
631  * backends as "reset" in the array, and sending PROCSIG_CATCHUP_INTERRUPT
632  * to some backend that seems to be getting too far behind. We signal at
633  * most one backend at a time, for reasons explained at the top of the file.
634  *
635  * Caution: because we transiently release write lock when we have to signal
636  * some other backend, it is NOT guaranteed that there are still minFree
637  * free message slots at exit. Caller must recheck and perhaps retry.
638  */
639 void
640 SICleanupQueue(bool callerHasWriteLock, int minFree)
641 {
642  SISeg *segP = shmInvalBuffer;
643  int min,
644  minsig,
645  lowbound,
646  numMsgs,
647  i;
648  ProcState *needSig = NULL;
649 
650  /* Lock out all writers and readers */
651  if (!callerHasWriteLock)
652  LWLockAcquire(SInvalWriteLock, LW_EXCLUSIVE);
653  LWLockAcquire(SInvalReadLock, LW_EXCLUSIVE);
654 
655  /*
656  * Recompute minMsgNum = minimum of all backends' nextMsgNum, identify the
657  * furthest-back backend that needs signaling (if any), and reset any
658  * backends that are too far back. Note that because we ignore sendOnly
659  * backends here it is possible for them to keep sending messages without
660  * a problem even when they are the only active backend.
661  */
662  min = segP->maxMsgNum;
663  minsig = min - SIG_THRESHOLD;
664  lowbound = min - MAXNUMMESSAGES + minFree;
665 
666  for (i = 0; i < segP->lastBackend; i++)
667  {
668  ProcState *stateP = &segP->procState[i];
669  int n = stateP->nextMsgNum;
670 
671  /* Ignore if inactive or already in reset state */
672  if (stateP->procPid == 0 || stateP->resetState || stateP->sendOnly)
673  continue;
674 
675  /*
676  * If we must free some space and this backend is preventing it, force
677  * him into reset state and then ignore until he catches up.
678  */
679  if (n < lowbound)
680  {
681  stateP->resetState = true;
682  /* no point in signaling him ... */
683  continue;
684  }
685 
686  /* Track the global minimum nextMsgNum */
687  if (n < min)
688  min = n;
689 
690  /* Also see who's furthest back of the unsignaled backends */
691  if (n < minsig && !stateP->signaled)
692  {
693  minsig = n;
694  needSig = stateP;
695  }
696  }
697  segP->minMsgNum = min;
698 
699  /*
700  * When minMsgNum gets really large, decrement all message counters so as
701  * to forestall overflow of the counters. This happens seldom enough that
702  * folding it into the previous loop would be a loser.
703  */
704  if (min >= MSGNUMWRAPAROUND)
705  {
706  segP->minMsgNum -= MSGNUMWRAPAROUND;
707  segP->maxMsgNum -= MSGNUMWRAPAROUND;
708  for (i = 0; i < segP->lastBackend; i++)
709  {
710  /* we don't bother skipping inactive entries here */
712  }
713  }
714 
715  /*
716  * Determine how many messages are still in the queue, and set the
717  * threshold at which we should repeat SICleanupQueue().
718  */
719  numMsgs = segP->maxMsgNum - segP->minMsgNum;
720  if (numMsgs < CLEANUP_MIN)
721  segP->nextThreshold = CLEANUP_MIN;
722  else
723  segP->nextThreshold = (numMsgs / CLEANUP_QUANTUM + 1) * CLEANUP_QUANTUM;
724 
725  /*
726  * Lastly, signal anyone who needs a catchup interrupt. Since
727  * SendProcSignal() might not be fast, we don't want to hold locks while
728  * executing it.
729  */
730  if (needSig)
731  {
732  pid_t his_pid = needSig->procPid;
733  BackendId his_backendId = (needSig - &segP->procState[0]) + 1;
734 
735  needSig->signaled = true;
736  LWLockRelease(SInvalReadLock);
737  LWLockRelease(SInvalWriteLock);
738  elog(DEBUG4, "sending sinval catchup signal to PID %d", (int) his_pid);
739  SendProcSignal(his_pid, PROCSIG_CATCHUP_INTERRUPT, his_backendId);
740  if (callerHasWriteLock)
741  LWLockAcquire(SInvalWriteLock, LW_EXCLUSIVE);
742  }
743  else
744  {
745  LWLockRelease(SInvalReadLock);
746  if (!callerHasWriteLock)
747  LWLockRelease(SInvalWriteLock);
748  }
749 }
750 
751 
752 /*
753  * GetNextLocalTransactionId --- allocate a new LocalTransactionId
754  *
755  * We split VirtualTransactionIds into two parts so that it is possible
756  * to allocate a new one without any contention for shared memory, except
757  * for a bit of additional overhead during backend startup/shutdown.
758  * The high-order part of a VirtualTransactionId is a BackendId, and the
759  * low-order part is a LocalTransactionId, which we assign from a local
760  * counter. To avoid the risk of a VirtualTransactionId being reused
761  * within a short interval, successive procs occupying the same backend ID
762  * slot should use a consecutive sequence of local IDs, which is implemented
763  * by copying nextLocalTransactionId as seen above.
764  */
767 {
768  LocalTransactionId result;
769 
770  /* loop to avoid returning InvalidLocalTransactionId at wraparound */
771  do
772  {
773  result = nextLocalTransactionId++;
774  } while (!LocalTransactionIdIsValid(result));
775 
776  return result;
777 }
int slock_t
Definition: s_lock.h:934
Size SInvalShmemSize(void)
Definition: sinvaladt.c:203
#define CLEANUP_MIN
Definition: sinvaladt.c:132
#define CLEANUP_QUANTUM
Definition: sinvaladt.c:133
int lastBackend
Definition: sinvaladt.c:175
int MyProcPid
Definition: globals.c:40
BackendId MyBackendId
Definition: globals.c:81
bool signaled
Definition: sinvaladt.c:146
BackendId backendId
Definition: proc.h:144
uint32 TransactionId
Definition: c.h:521
void SharedInvalBackendInit(bool sendOnly)
Definition: sinvaladt.c:257
PGPROC * MyProc
Definition: proc.c:67
#define PointerGetDatum(X)
Definition: postgres.h:556
void CreateSharedInvalidationState(void)
Definition: sinvaladt.c:218
ProcState procState[FLEXIBLE_ARRAY_MEMBER]
Definition: sinvaladt.c:188
#define SpinLockInit(lock)
Definition: spin.h:60
PGPROC * proc
Definition: sinvaladt.c:142
slock_t msgnumLock
Definition: sinvaladt.c:178
#define Min(x, y)
Definition: c.h:928
#define FLEXIBLE_ARRAY_MEMBER
Definition: c.h:284
static void CleanupInvalidationState(int status, Datum arg)
Definition: sinvaladt.c:335
int errcode(int sqlerrcode)
Definition: elog.c:610
#define DEBUG4
Definition: elog.h:22
#define SIG_THRESHOLD
Definition: sinvaladt.c:134
bool resetState
Definition: sinvaladt.c:145
Definition: type.h:89
void LWLockRelease(LWLock *lock)
Definition: lwlock.c:1812
#define SpinLockAcquire(lock)
Definition: spin.h:62
int SendProcSignal(pid_t pid, ProcSignalReason reason, BackendId backendId)
Definition: procsignal.c:250
void * ShmemInitStruct(const char *name, Size size, bool *foundPtr)
Definition: shmem.c:392
#define FATAL
Definition: elog.h:52
int SIGetDataEntries(SharedInvalidationMessage *data, int datasize)
Definition: sinvaladt.c:536
LocalTransactionId GetNextLocalTransactionId(void)
Definition: sinvaladt.c:766
int MaxBackends
Definition: globals.c:136
TransactionId xmin
Definition: proc.h:129
bool sendOnly
Definition: sinvaladt.c:155
void on_shmem_exit(pg_on_exit_callback function, Datum arg)
Definition: ipc.c:361
#define InvalidTransactionId
Definition: transam.h:31
void SICleanupQueue(bool callerHasWriteLock, int minFree)
Definition: sinvaladt.c:640
SharedInvalidationMessage buffer[MAXNUMMESSAGES]
Definition: sinvaladt.c:183
uint32 LocalTransactionId
Definition: c.h:523
#define WRITE_QUANTUM
Definition: sinvaladt.c:135
int maxMsgNum
Definition: sinvaladt.c:173
#define MAXNUMMESSAGES
Definition: sinvaladt.c:130
#define SpinLockRelease(lock)
Definition: spin.h:64
Size mul_size(Size s1, Size s2)
Definition: shmem.c:515
#define InvalidBackendId
Definition: backendid.h:23
uintptr_t Datum
Definition: postgres.h:367
void BackendIdGetTransactionIds(int backendID, TransactionId *xid, TransactionId *xmin)
Definition: sinvaladt.c:403
Size add_size(Size s1, Size s2)
Definition: shmem.c:498
int BackendId
Definition: backendid.h:21
LocalTransactionId nextLXID
Definition: sinvaladt.c:163
#define ereport(elevel,...)
Definition: elog.h:144
#define Assert(condition)
Definition: c.h:746
struct ProcState ProcState
size_t Size
Definition: c.h:474
static LocalTransactionId nextLocalTransactionId
Definition: sinvaladt.c:194
bool LWLockAcquire(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:1208
static SISeg * shmInvalBuffer
Definition: sinvaladt.c:191
struct SISeg SISeg
#define DatumGetPointer(X)
Definition: postgres.h:549
TransactionId xid
Definition: proc.h:124
int minMsgNum
Definition: sinvaladt.c:172
#define MSGNUMWRAPAROUND
Definition: sinvaladt.c:131
int errmsg(const char *fmt,...)
Definition: elog.c:821
int maxBackends
Definition: sinvaladt.c:176
#define elog(elevel,...)
Definition: elog.h:214
#define InvalidLocalTransactionId
Definition: lock.h:68
int i
int nextThreshold
Definition: sinvaladt.c:174
#define LocalTransactionIdIsValid(lxid)
Definition: lock.h:69
void * arg
void SIInsertDataEntries(const SharedInvalidationMessage *data, int n)
Definition: sinvaladt.c:433
bool hasMessages
Definition: sinvaladt.c:147
pid_t procPid
Definition: sinvaladt.c:141
static void static void status(const char *fmt,...) pg_attribute_printf(1
Definition: pg_regress.c:227
Definition: proc.h:112
#define PointerIsValid(pointer)
Definition: c.h:640
#define offsetof(type, field)
Definition: c.h:669
int nextMsgNum
Definition: sinvaladt.c:144
PGPROC * BackendIdGetProc(int backendID)
Definition: sinvaladt.c:376