TLA Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * procarray.c
4 : * POSTGRES process array code.
5 : *
6 : *
7 : * This module maintains arrays of PGPROC substructures, as well as associated
8 : * arrays in ProcGlobal, for all active backends. Although there are several
9 : * uses for this, the principal one is as a means of determining the set of
10 : * currently running transactions.
11 : *
12 : * Because of various subtle race conditions it is critical that a backend
13 : * hold the correct locks while setting or clearing its xid (in
14 : * ProcGlobal->xids[]/MyProc->xid). See notes in
15 : * src/backend/access/transam/README.
16 : *
17 : * The process arrays now also include structures representing prepared
18 : * transactions. The xid and subxids fields of these are valid, as are the
19 : * myProcLocks lists. They can be distinguished from regular backend PGPROCs
20 : * at need by checking for pid == 0.
21 : *
22 : * During hot standby, we also keep a list of XIDs representing transactions
23 : * that are known to be running on the primary (or more precisely, were running
24 : * as of the current point in the WAL stream). This list is kept in the
25 : * KnownAssignedXids array, and is updated by watching the sequence of
26 : * arriving XIDs. This is necessary because if we leave those XIDs out of
27 : * snapshots taken for standby queries, then they will appear to be already
28 : * complete, leading to MVCC failures. Note that in hot standby, the PGPROC
29 : * array represents standby processes, which by definition are not running
30 : * transactions that have XIDs.
31 : *
32 : * It is perhaps possible for a backend on the primary to terminate without
33 : * writing an abort record for its transaction. While that shouldn't really
34 : * happen, it would tie up KnownAssignedXids indefinitely, so we protect
35 : * ourselves by pruning the array when a valid list of running XIDs arrives.
36 : *
37 : * Portions Copyright (c) 1996-2023, PostgreSQL Global Development Group
38 : * Portions Copyright (c) 1994, Regents of the University of California
39 : *
40 : *
41 : * IDENTIFICATION
42 : * src/backend/storage/ipc/procarray.c
43 : *
44 : *-------------------------------------------------------------------------
45 : */
46 : #include "postgres.h"
47 :
48 : #include <signal.h>
49 :
50 : #include "access/clog.h"
51 : #include "access/subtrans.h"
52 : #include "access/transam.h"
53 : #include "access/twophase.h"
54 : #include "access/xact.h"
55 : #include "access/xlogutils.h"
56 : #include "catalog/catalog.h"
57 : #include "catalog/pg_authid.h"
58 : #include "commands/dbcommands.h"
59 : #include "miscadmin.h"
60 : #include "pgstat.h"
61 : #include "port/pg_lfind.h"
62 : #include "storage/proc.h"
63 : #include "storage/procarray.h"
64 : #include "storage/spin.h"
65 : #include "utils/acl.h"
66 : #include "utils/builtins.h"
67 : #include "utils/rel.h"
68 : #include "utils/snapmgr.h"
69 :
70 : #define UINT32_ACCESS_ONCE(var) ((uint32)(*((volatile uint32 *)&(var))))
71 :
72 : /* Our shared memory area */
73 : typedef struct ProcArrayStruct
74 : {
75 : int numProcs; /* number of valid procs entries */
76 : int maxProcs; /* allocated size of procs array */
77 :
78 : /*
79 : * Known assigned XIDs handling
80 : */
81 : int maxKnownAssignedXids; /* allocated size of array */
82 : int numKnownAssignedXids; /* current # of valid entries */
83 : int tailKnownAssignedXids; /* index of oldest valid element */
84 : int headKnownAssignedXids; /* index of newest element, + 1 */
85 : slock_t known_assigned_xids_lck; /* protects head/tail pointers */
86 :
87 : /*
88 : * Highest subxid that has been removed from KnownAssignedXids array to
89 : * prevent overflow; or InvalidTransactionId if none. We track this for
90 : * similar reasons to tracking overflowing cached subxids in PGPROC
91 : * entries. Must hold exclusive ProcArrayLock to change this, and shared
92 : * lock to read it.
93 : */
94 : TransactionId lastOverflowedXid;
95 :
96 : /* oldest xmin of any replication slot */
97 : TransactionId replication_slot_xmin;
98 : /* oldest catalog xmin of any replication slot */
99 : TransactionId replication_slot_catalog_xmin;
100 :
101 : /* indexes into allProcs[], has PROCARRAY_MAXPROCS entries */
102 : int pgprocnos[FLEXIBLE_ARRAY_MEMBER];
103 : } ProcArrayStruct;
104 :
105 : /*
106 : * State for the GlobalVisTest* family of functions. Those functions can
107 : * e.g. be used to decide if a deleted row can be removed without violating
108 : * MVCC semantics: If the deleted row's xmax is not considered to be running
109 : * by anyone, the row can be removed.
110 : *
111 : * To avoid slowing down GetSnapshotData(), we don't calculate a precise
112 : * cutoff XID while building a snapshot (looking at the frequently changing
113 : * xmins scales badly). Instead we compute two boundaries while building the
114 : * snapshot:
115 : *
116 : * 1) definitely_needed, indicating that rows deleted by XIDs >=
117 : * definitely_needed are definitely still visible.
118 : *
119 : * 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can
120 : * definitely be removed
121 : *
122 : * When testing an XID that falls in between the two (i.e. XID >= maybe_needed
123 : * && XID < definitely_needed), the boundaries can be recomputed (using
124 : * ComputeXidHorizons()) to get a more accurate answer. This is cheaper than
125 : * maintaining an accurate value all the time.
126 : *
127 : * As it is not cheap to compute accurate boundaries, we limit the number of
128 : * times that happens in short succession. See GlobalVisTestShouldUpdate().
129 : *
130 : *
131 : * There are three backend lifetime instances of this struct, optimized for
132 : * different types of relations. As e.g. a normal user defined table in one
133 : * database is inaccessible to backends connected to another database, a test
134 : * specific to a relation can be more aggressive than a test for a shared
135 : * relation. Currently we track four different states:
136 : *
137 : * 1) GlobalVisSharedRels, which only considers an XID's
138 : * effects visible-to-everyone if neither snapshots in any database, nor a
139 : * replication slot's xmin, nor a replication slot's catalog_xmin might
140 : * still consider XID as running.
141 : *
142 : * 2) GlobalVisCatalogRels, which only considers an XID's
143 : * effects visible-to-everyone if neither snapshots in the current
144 : * database, nor a replication slot's xmin, nor a replication slot's
145 : * catalog_xmin might still consider XID as running.
146 : *
147 : * I.e. the difference to GlobalVisSharedRels is that
148 : * snapshot in other databases are ignored.
149 : *
150 : * 3) GlobalVisDataRels, which only considers an XID's
151 : * effects visible-to-everyone if neither snapshots in the current
152 : * database, nor a replication slot's xmin consider XID as running.
153 : *
154 : * I.e. the difference to GlobalVisCatalogRels is that
155 : * replication slot's catalog_xmin is not taken into account.
156 : *
157 : * 4) GlobalVisTempRels, which only considers the current session, as temp
158 : * tables are not visible to other sessions.
159 : *
160 : * GlobalVisTestFor(relation) returns the appropriate state
161 : * for the relation.
162 : *
163 : * The boundaries are FullTransactionIds instead of TransactionIds to avoid
164 : * wraparound dangers. There e.g. would otherwise exist no procarray state to
165 : * prevent maybe_needed to become old enough after the GetSnapshotData()
166 : * call.
167 : *
168 : * The typedef is in the header.
169 : */
170 : struct GlobalVisState
171 : {
172 : /* XIDs >= are considered running by some backend */
173 : FullTransactionId definitely_needed;
174 :
175 : /* XIDs < are not considered to be running by any backend */
176 : FullTransactionId maybe_needed;
177 : };
178 :
179 : /*
180 : * Result of ComputeXidHorizons().
181 : */
182 : typedef struct ComputeXidHorizonsResult
183 : {
184 : /*
185 : * The value of ShmemVariableCache->latestCompletedXid when
186 : * ComputeXidHorizons() held ProcArrayLock.
187 : */
188 : FullTransactionId latest_completed;
189 :
190 : /*
191 : * The same for procArray->replication_slot_xmin and.
192 : * procArray->replication_slot_catalog_xmin.
193 : */
194 : TransactionId slot_xmin;
195 : TransactionId slot_catalog_xmin;
196 :
197 : /*
198 : * Oldest xid that any backend might still consider running. This needs to
199 : * include processes running VACUUM, in contrast to the normal visibility
200 : * cutoffs, as vacuum needs to be able to perform pg_subtrans lookups when
201 : * determining visibility, but doesn't care about rows above its xmin to
202 : * be removed.
203 : *
204 : * This likely should only be needed to determine whether pg_subtrans can
205 : * be truncated. It currently includes the effects of replication slots,
206 : * for historical reasons. But that could likely be changed.
207 : */
208 : TransactionId oldest_considered_running;
209 :
210 : /*
211 : * Oldest xid for which deleted tuples need to be retained in shared
212 : * tables.
213 : *
214 : * This includes the effects of replication slots. If that's not desired,
215 : * look at shared_oldest_nonremovable_raw;
216 : */
217 : TransactionId shared_oldest_nonremovable;
218 :
219 : /*
220 : * Oldest xid that may be necessary to retain in shared tables. This is
221 : * the same as shared_oldest_nonremovable, except that is not affected by
222 : * replication slot's catalog_xmin.
223 : *
224 : * This is mainly useful to be able to send the catalog_xmin to upstream
225 : * streaming replication servers via hot_standby_feedback, so they can
226 : * apply the limit only when accessing catalog tables.
227 : */
228 : TransactionId shared_oldest_nonremovable_raw;
229 :
230 : /*
231 : * Oldest xid for which deleted tuples need to be retained in non-shared
232 : * catalog tables.
233 : */
234 : TransactionId catalog_oldest_nonremovable;
235 :
236 : /*
237 : * Oldest xid for which deleted tuples need to be retained in normal user
238 : * defined tables.
239 : */
240 : TransactionId data_oldest_nonremovable;
241 :
242 : /*
243 : * Oldest xid for which deleted tuples need to be retained in this
244 : * session's temporary tables.
245 : */
246 : TransactionId temp_oldest_nonremovable;
247 : } ComputeXidHorizonsResult;
248 :
249 : /*
250 : * Return value for GlobalVisHorizonKindForRel().
251 : */
252 : typedef enum GlobalVisHorizonKind
253 : {
254 : VISHORIZON_SHARED,
255 : VISHORIZON_CATALOG,
256 : VISHORIZON_DATA,
257 : VISHORIZON_TEMP
258 : } GlobalVisHorizonKind;
259 :
260 : /*
261 : * Reason codes for KnownAssignedXidsCompress().
262 : */
263 : typedef enum KAXCompressReason
264 : {
265 : KAX_NO_SPACE, /* need to free up space at array end */
266 : KAX_PRUNE, /* we just pruned old entries */
267 : KAX_TRANSACTION_END, /* we just committed/removed some XIDs */
268 : KAX_STARTUP_PROCESS_IDLE /* startup process is about to sleep */
269 : } KAXCompressReason;
270 :
271 :
272 : static ProcArrayStruct *procArray;
273 :
274 : static PGPROC *allProcs;
275 :
276 : /*
277 : * Cache to reduce overhead of repeated calls to TransactionIdIsInProgress()
278 : */
279 : static TransactionId cachedXidIsNotInProgress = InvalidTransactionId;
280 :
281 : /*
282 : * Bookkeeping for tracking emulated transactions in recovery
283 : */
284 : static TransactionId *KnownAssignedXids;
285 : static bool *KnownAssignedXidsValid;
286 : static TransactionId latestObservedXid = InvalidTransactionId;
287 :
288 : /*
289 : * If we're in STANDBY_SNAPSHOT_PENDING state, standbySnapshotPendingXmin is
290 : * the highest xid that might still be running that we don't have in
291 : * KnownAssignedXids.
292 : */
293 : static TransactionId standbySnapshotPendingXmin;
294 :
295 : /*
296 : * State for visibility checks on different types of relations. See struct
297 : * GlobalVisState for details. As shared, catalog, normal and temporary
298 : * relations can have different horizons, one such state exists for each.
299 : */
300 : static GlobalVisState GlobalVisSharedRels;
301 : static GlobalVisState GlobalVisCatalogRels;
302 : static GlobalVisState GlobalVisDataRels;
303 : static GlobalVisState GlobalVisTempRels;
304 :
305 : /*
306 : * This backend's RecentXmin at the last time the accurate xmin horizon was
307 : * recomputed, or InvalidTransactionId if it has not. Used to limit how many
308 : * times accurate horizons are recomputed. See GlobalVisTestShouldUpdate().
309 : */
310 : static TransactionId ComputeXidHorizonsResultLastXmin;
311 :
312 : #ifdef XIDCACHE_DEBUG
313 :
314 : /* counters for XidCache measurement */
315 : static long xc_by_recent_xmin = 0;
316 : static long xc_by_known_xact = 0;
317 : static long xc_by_my_xact = 0;
318 : static long xc_by_latest_xid = 0;
319 : static long xc_by_main_xid = 0;
320 : static long xc_by_child_xid = 0;
321 : static long xc_by_known_assigned = 0;
322 : static long xc_no_overflow = 0;
323 : static long xc_slow_answer = 0;
324 :
325 : #define xc_by_recent_xmin_inc() (xc_by_recent_xmin++)
326 : #define xc_by_known_xact_inc() (xc_by_known_xact++)
327 : #define xc_by_my_xact_inc() (xc_by_my_xact++)
328 : #define xc_by_latest_xid_inc() (xc_by_latest_xid++)
329 : #define xc_by_main_xid_inc() (xc_by_main_xid++)
330 : #define xc_by_child_xid_inc() (xc_by_child_xid++)
331 : #define xc_by_known_assigned_inc() (xc_by_known_assigned++)
332 : #define xc_no_overflow_inc() (xc_no_overflow++)
333 : #define xc_slow_answer_inc() (xc_slow_answer++)
334 :
335 : static void DisplayXidCache(void);
336 : #else /* !XIDCACHE_DEBUG */
337 :
338 : #define xc_by_recent_xmin_inc() ((void) 0)
339 : #define xc_by_known_xact_inc() ((void) 0)
340 : #define xc_by_my_xact_inc() ((void) 0)
341 : #define xc_by_latest_xid_inc() ((void) 0)
342 : #define xc_by_main_xid_inc() ((void) 0)
343 : #define xc_by_child_xid_inc() ((void) 0)
344 : #define xc_by_known_assigned_inc() ((void) 0)
345 : #define xc_no_overflow_inc() ((void) 0)
346 : #define xc_slow_answer_inc() ((void) 0)
347 : #endif /* XIDCACHE_DEBUG */
348 :
349 : /* Primitives for KnownAssignedXids array handling for standby */
350 : static void KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock);
351 : static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
352 : bool exclusive_lock);
353 : static bool KnownAssignedXidsSearch(TransactionId xid, bool remove);
354 : static bool KnownAssignedXidExists(TransactionId xid);
355 : static void KnownAssignedXidsRemove(TransactionId xid);
356 : static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
357 : TransactionId *subxids);
358 : static void KnownAssignedXidsRemovePreceding(TransactionId removeXid);
359 : static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax);
360 : static int KnownAssignedXidsGetAndSetXmin(TransactionId *xarray,
361 : TransactionId *xmin,
362 : TransactionId xmax);
363 : static TransactionId KnownAssignedXidsGetOldestXmin(void);
364 : static void KnownAssignedXidsDisplay(int trace_level);
365 : static void KnownAssignedXidsReset(void);
366 : static inline void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid);
367 : static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid);
368 : static void MaintainLatestCompletedXid(TransactionId latestXid);
369 : static void MaintainLatestCompletedXidRecovery(TransactionId latestXid);
370 : static void TransactionIdRetreatSafely(TransactionId *xid,
371 : int retreat_by,
372 : FullTransactionId rel);
373 :
374 : static inline FullTransactionId FullXidRelativeTo(FullTransactionId rel,
375 : TransactionId xid);
376 : static void GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons);
377 :
378 : /*
379 : * Report shared-memory space needed by CreateSharedProcArray.
380 : */
381 : Size
382 GIC 2738 : ProcArrayShmemSize(void)
383 ECB : {
384 : Size size;
385 :
386 : /* Size of the ProcArray structure itself */
387 : #define PROCARRAY_MAXPROCS (MaxBackends + max_prepared_xacts)
388 :
389 GIC 2738 : size = offsetof(ProcArrayStruct, pgprocnos);
390 CBC 2738 : size = add_size(size, mul_size(sizeof(int), PROCARRAY_MAXPROCS));
391 ECB :
392 : /*
393 : * During Hot Standby processing we have a data structure called
394 : * KnownAssignedXids, created in shared memory. Local data structures are
395 : * also created in various backends during GetSnapshotData(),
396 : * TransactionIdIsInProgress() and GetRunningTransactionData(). All of the
397 : * main structures created in those functions must be identically sized,
398 : * since we may at times copy the whole of the data structures around. We
399 : * refer to this size as TOTAL_MAX_CACHED_SUBXIDS.
400 : *
401 : * Ideally we'd only create this structure if we were actually doing hot
402 : * standby in the current run, but we don't know that yet at the time
403 : * shared memory is being set up.
404 : */
405 : #define TOTAL_MAX_CACHED_SUBXIDS \
406 : ((PGPROC_MAX_CACHED_SUBXIDS + 1) * PROCARRAY_MAXPROCS)
407 :
408 GIC 2738 : if (EnableHotStandby)
409 ECB : {
410 GIC 2728 : size = add_size(size,
411 ECB : mul_size(sizeof(TransactionId),
412 GIC 2728 : TOTAL_MAX_CACHED_SUBXIDS));
413 CBC 2728 : size = add_size(size,
414 2728 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS));
415 ECB : }
416 :
417 GIC 2738 : return size;
418 ECB : }
419 :
420 : /*
421 : * Initialize the shared PGPROC array during postmaster startup.
422 : */
423 : void
424 GIC 1826 : CreateSharedProcArray(void)
425 ECB : {
426 : bool found;
427 :
428 : /* Create or attach to the ProcArray shared structure */
429 GIC 1826 : procArray = (ProcArrayStruct *)
430 CBC 1826 : ShmemInitStruct("Proc Array",
431 ECB : add_size(offsetof(ProcArrayStruct, pgprocnos),
432 : mul_size(sizeof(int),
433 GIC 1826 : PROCARRAY_MAXPROCS)),
434 ECB : &found);
435 :
436 GIC 1826 : if (!found)
437 ECB : {
438 : /*
439 : * We're the first - initialize.
440 : */
441 GIC 1826 : procArray->numProcs = 0;
442 CBC 1826 : procArray->maxProcs = PROCARRAY_MAXPROCS;
443 1826 : procArray->maxKnownAssignedXids = TOTAL_MAX_CACHED_SUBXIDS;
444 1826 : procArray->numKnownAssignedXids = 0;
445 1826 : procArray->tailKnownAssignedXids = 0;
446 1826 : procArray->headKnownAssignedXids = 0;
447 1826 : SpinLockInit(&procArray->known_assigned_xids_lck);
448 1826 : procArray->lastOverflowedXid = InvalidTransactionId;
449 1826 : procArray->replication_slot_xmin = InvalidTransactionId;
450 1826 : procArray->replication_slot_catalog_xmin = InvalidTransactionId;
451 1826 : ShmemVariableCache->xactCompletionCount = 1;
452 ECB : }
453 :
454 GIC 1826 : allProcs = ProcGlobal->allProcs;
455 ECB :
456 : /* Create or attach to the KnownAssignedXids arrays too, if needed */
457 GIC 1826 : if (EnableHotStandby)
458 ECB : {
459 GIC 1821 : KnownAssignedXids = (TransactionId *)
460 CBC 1821 : ShmemInitStruct("KnownAssignedXids",
461 ECB : mul_size(sizeof(TransactionId),
462 GIC 1821 : TOTAL_MAX_CACHED_SUBXIDS),
463 ECB : &found);
464 GIC 1821 : KnownAssignedXidsValid = (bool *)
465 CBC 1821 : ShmemInitStruct("KnownAssignedXidsValid",
466 1821 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS),
467 ECB : &found);
468 : }
469 GIC 1826 : }
470 ECB :
471 : /*
472 : * Add the specified PGPROC to the shared array.
473 : */
474 : void
475 GIC 11890 : ProcArrayAdd(PGPROC *proc)
476 ECB : {
477 GIC 11890 : ProcArrayStruct *arrayP = procArray;
478 ECB : int index;
479 : int movecount;
480 :
481 : /* See ProcGlobal comment explaining why both locks are held */
482 GIC 11890 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
483 CBC 11890 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
484 ECB :
485 GIC 11890 : if (arrayP->numProcs >= arrayP->maxProcs)
486 ECB : {
487 : /*
488 : * Oops, no room. (This really shouldn't happen, since there is a
489 : * fixed supply of PGPROC structs too, and so we should have failed
490 : * earlier.)
491 : */
492 UIC 0 : ereport(FATAL,
493 EUB : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
494 : errmsg("sorry, too many clients already")));
495 : }
496 :
497 : /*
498 : * Keep the procs array sorted by (PGPROC *) so that we can utilize
499 : * locality of references much better. This is useful while traversing the
500 : * ProcArray because there is an increased likelihood of finding the next
501 : * PGPROC structure in the cache.
502 : *
503 : * Since the occurrence of adding/removing a proc is much lower than the
504 : * access to the ProcArray itself, the overhead should be marginal
505 : */
506 GIC 23473 : for (index = 0; index < arrayP->numProcs; index++)
507 ECB : {
508 GIC 21448 : int procno PG_USED_FOR_ASSERTS_ONLY = arrayP->pgprocnos[index];
509 ECB :
510 GIC 21448 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
511 CBC 21448 : Assert(allProcs[procno].pgxactoff == index);
512 ECB :
513 : /* If we have found our right position in the array, break */
514 GIC 21448 : if (arrayP->pgprocnos[index] > proc->pgprocno)
515 CBC 9865 : break;
516 ECB : }
517 :
518 GIC 11890 : movecount = arrayP->numProcs - index;
519 CBC 11890 : memmove(&arrayP->pgprocnos[index + 1],
520 11890 : &arrayP->pgprocnos[index],
521 ECB : movecount * sizeof(*arrayP->pgprocnos));
522 GIC 11890 : memmove(&ProcGlobal->xids[index + 1],
523 CBC 11890 : &ProcGlobal->xids[index],
524 ECB : movecount * sizeof(*ProcGlobal->xids));
525 GIC 11890 : memmove(&ProcGlobal->subxidStates[index + 1],
526 CBC 11890 : &ProcGlobal->subxidStates[index],
527 ECB : movecount * sizeof(*ProcGlobal->subxidStates));
528 GIC 11890 : memmove(&ProcGlobal->statusFlags[index + 1],
529 CBC 11890 : &ProcGlobal->statusFlags[index],
530 ECB : movecount * sizeof(*ProcGlobal->statusFlags));
531 :
532 GIC 11890 : arrayP->pgprocnos[index] = proc->pgprocno;
533 CBC 11890 : proc->pgxactoff = index;
534 11890 : ProcGlobal->xids[index] = proc->xid;
535 11890 : ProcGlobal->subxidStates[index] = proc->subxidStatus;
536 11890 : ProcGlobal->statusFlags[index] = proc->statusFlags;
537 ECB :
538 GIC 11890 : arrayP->numProcs++;
539 ECB :
540 : /* adjust pgxactoff for all following PGPROCs */
541 GIC 11890 : index++;
542 CBC 38685 : for (; index < arrayP->numProcs; index++)
543 ECB : {
544 GIC 26795 : int procno = arrayP->pgprocnos[index];
545 ECB :
546 GIC 26795 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
547 CBC 26795 : Assert(allProcs[procno].pgxactoff == index - 1);
548 ECB :
549 GIC 26795 : allProcs[procno].pgxactoff = index;
550 ECB : }
551 :
552 : /*
553 : * Release in reversed acquisition order, to reduce frequency of having to
554 : * wait for XidGenLock while holding ProcArrayLock.
555 : */
556 GIC 11890 : LWLockRelease(XidGenLock);
557 CBC 11890 : LWLockRelease(ProcArrayLock);
558 11890 : }
559 ECB :
560 : /*
561 : * Remove the specified PGPROC from the shared array.
562 : *
563 : * When latestXid is a valid XID, we are removing a live 2PC gxact from the
564 : * array, and thus causing it to appear as "not running" anymore. In this
565 : * case we must advance latestCompletedXid. (This is essentially the same
566 : * as ProcArrayEndTransaction followed by removal of the PGPROC, but we take
567 : * the ProcArrayLock only once, and don't damage the content of the PGPROC;
568 : * twophase.c depends on the latter.)
569 : */
570 : void
571 GIC 11864 : ProcArrayRemove(PGPROC *proc, TransactionId latestXid)
572 ECB : {
573 GIC 11864 : ProcArrayStruct *arrayP = procArray;
574 ECB : int myoff;
575 : int movecount;
576 :
577 : #ifdef XIDCACHE_DEBUG
578 : /* dump stats at backend shutdown, but not prepared-xact end */
579 : if (proc->pid != 0)
580 : DisplayXidCache();
581 : #endif
582 :
583 : /* See ProcGlobal comment explaining why both locks are held */
584 GIC 11864 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
585 CBC 11864 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
586 ECB :
587 GIC 11864 : myoff = proc->pgxactoff;
588 ECB :
589 GIC 11864 : Assert(myoff >= 0 && myoff < arrayP->numProcs);
590 CBC 11864 : Assert(ProcGlobal->allProcs[arrayP->pgprocnos[myoff]].pgxactoff == myoff);
591 ECB :
592 GIC 11864 : if (TransactionIdIsValid(latestXid))
593 ECB : {
594 GIC 365 : Assert(TransactionIdIsValid(ProcGlobal->xids[myoff]));
595 ECB :
596 : /* Advance global latestCompletedXid while holding the lock */
597 GIC 365 : MaintainLatestCompletedXid(latestXid);
598 ECB :
599 : /* Same with xactCompletionCount */
600 GIC 365 : ShmemVariableCache->xactCompletionCount++;
601 ECB :
602 GIC 365 : ProcGlobal->xids[myoff] = InvalidTransactionId;
603 CBC 365 : ProcGlobal->subxidStates[myoff].overflowed = false;
604 365 : ProcGlobal->subxidStates[myoff].count = 0;
605 ECB : }
606 : else
607 : {
608 : /* Shouldn't be trying to remove a live transaction here */
609 GIC 11499 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
610 ECB : }
611 :
612 GIC 11864 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
613 CBC 11864 : Assert(ProcGlobal->subxidStates[myoff].count == 0);
614 11864 : Assert(ProcGlobal->subxidStates[myoff].overflowed == false);
615 ECB :
616 GIC 11864 : ProcGlobal->statusFlags[myoff] = 0;
617 ECB :
618 : /* Keep the PGPROC array sorted. See notes above */
619 GIC 11864 : movecount = arrayP->numProcs - myoff - 1;
620 CBC 11864 : memmove(&arrayP->pgprocnos[myoff],
621 11864 : &arrayP->pgprocnos[myoff + 1],
622 ECB : movecount * sizeof(*arrayP->pgprocnos));
623 GIC 11864 : memmove(&ProcGlobal->xids[myoff],
624 CBC 11864 : &ProcGlobal->xids[myoff + 1],
625 ECB : movecount * sizeof(*ProcGlobal->xids));
626 GIC 11864 : memmove(&ProcGlobal->subxidStates[myoff],
627 CBC 11864 : &ProcGlobal->subxidStates[myoff + 1],
628 ECB : movecount * sizeof(*ProcGlobal->subxidStates));
629 GIC 11864 : memmove(&ProcGlobal->statusFlags[myoff],
630 CBC 11864 : &ProcGlobal->statusFlags[myoff + 1],
631 ECB : movecount * sizeof(*ProcGlobal->statusFlags));
632 :
633 GIC 11864 : arrayP->pgprocnos[arrayP->numProcs - 1] = -1; /* for debugging */
634 CBC 11864 : arrayP->numProcs--;
635 ECB :
636 : /*
637 : * Adjust pgxactoff of following procs for removed PGPROC (note that
638 : * numProcs already has been decremented).
639 : */
640 GIC 36833 : for (int index = myoff; index < arrayP->numProcs; index++)
641 ECB : {
642 GIC 24969 : int procno = arrayP->pgprocnos[index];
643 ECB :
644 GIC 24969 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
645 CBC 24969 : Assert(allProcs[procno].pgxactoff - 1 == index);
646 ECB :
647 GIC 24969 : allProcs[procno].pgxactoff = index;
648 ECB : }
649 :
650 : /*
651 : * Release in reversed acquisition order, to reduce frequency of having to
652 : * wait for XidGenLock while holding ProcArrayLock.
653 : */
654 GIC 11864 : LWLockRelease(XidGenLock);
655 CBC 11864 : LWLockRelease(ProcArrayLock);
656 11864 : }
657 ECB :
658 :
659 : /*
660 : * ProcArrayEndTransaction -- mark a transaction as no longer running
661 : *
662 : * This is used interchangeably for commit and abort cases. The transaction
663 : * commit/abort must already be reported to WAL and pg_xact.
664 : *
665 : * proc is currently always MyProc, but we pass it explicitly for flexibility.
666 : * latestXid is the latest Xid among the transaction's main XID and
667 : * subtransactions, or InvalidTransactionId if it has no XID. (We must ask
668 : * the caller to pass latestXid, instead of computing it from the PGPROC's
669 : * contents, because the subxid information in the PGPROC might be
670 : * incomplete.)
671 : */
672 : void
673 GIC 485806 : ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid)
674 ECB : {
675 GIC 485806 : if (TransactionIdIsValid(latestXid))
676 ECB : {
677 : /*
678 : * We must lock ProcArrayLock while clearing our advertised XID, so
679 : * that we do not exit the set of "running" transactions while someone
680 : * else is taking a snapshot. See discussion in
681 : * src/backend/access/transam/README.
682 : */
683 GIC 297704 : Assert(TransactionIdIsValid(proc->xid));
684 ECB :
685 : /*
686 : * If we can immediately acquire ProcArrayLock, we clear our own XID
687 : * and release the lock. If not, use group XID clearing to improve
688 : * efficiency.
689 : */
690 GIC 297704 : if (LWLockConditionalAcquire(ProcArrayLock, LW_EXCLUSIVE))
691 ECB : {
692 GIC 297529 : ProcArrayEndTransactionInternal(proc, latestXid);
693 CBC 297529 : LWLockRelease(ProcArrayLock);
694 ECB : }
695 : else
696 GIC 175 : ProcArrayGroupClearXid(proc, latestXid);
697 ECB : }
698 : else
699 : {
700 : /*
701 : * If we have no XID, we don't need to lock, since we won't affect
702 : * anyone else's calculation of a snapshot. We might change their
703 : * estimate of global xmin, but that's OK.
704 : */
705 GIC 188102 : Assert(!TransactionIdIsValid(proc->xid));
706 CBC 188102 : Assert(proc->subxidStatus.count == 0);
707 188102 : Assert(!proc->subxidStatus.overflowed);
708 ECB :
709 GIC 188102 : proc->lxid = InvalidLocalTransactionId;
710 CBC 188102 : proc->xmin = InvalidTransactionId;
711 ECB :
712 : /* be sure this is cleared in abort */
713 GIC 188102 : proc->delayChkptFlags = 0;
714 ECB :
715 GIC 188102 : proc->recoveryConflictPending = false;
716 ECB :
717 : /* must be cleared with xid/xmin: */
718 : /* avoid unnecessarily dirtying shared cachelines */
719 GIC 188102 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
720 ECB : {
721 GIC 37331 : Assert(!LWLockHeldByMe(ProcArrayLock));
722 CBC 37331 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
723 37331 : Assert(proc->statusFlags == ProcGlobal->statusFlags[proc->pgxactoff]);
724 37331 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
725 37331 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
726 37331 : LWLockRelease(ProcArrayLock);
727 ECB : }
728 : }
729 GIC 485806 : }
730 ECB :
731 : /*
732 : * Mark a write transaction as no longer running.
733 : *
734 : * We don't do any locking here; caller must handle that.
735 : */
736 : static inline void
737 GIC 297704 : ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid)
738 ECB : {
739 GIC 297704 : int pgxactoff = proc->pgxactoff;
740 ECB :
741 : /*
742 : * Note: we need exclusive lock here because we're going to change other
743 : * processes' PGPROC entries.
744 : */
745 GIC 297704 : Assert(LWLockHeldByMeInMode(ProcArrayLock, LW_EXCLUSIVE));
746 CBC 297704 : Assert(TransactionIdIsValid(ProcGlobal->xids[pgxactoff]));
747 297704 : Assert(ProcGlobal->xids[pgxactoff] == proc->xid);
748 ECB :
749 GIC 297704 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
750 CBC 297704 : proc->xid = InvalidTransactionId;
751 297704 : proc->lxid = InvalidLocalTransactionId;
752 297704 : proc->xmin = InvalidTransactionId;
753 ECB :
754 : /* be sure this is cleared in abort */
755 GIC 297704 : proc->delayChkptFlags = 0;
756 ECB :
757 GIC 297704 : proc->recoveryConflictPending = false;
758 ECB :
759 : /* must be cleared with xid/xmin: */
760 : /* avoid unnecessarily dirtying shared cachelines */
761 GIC 297704 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
762 ECB : {
763 GIC 534 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
764 CBC 534 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
765 ECB : }
766 :
767 : /* Clear the subtransaction-XID cache too while holding the lock */
768 GIC 297704 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
769 ECB : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
770 GIC 297704 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
771 ECB : {
772 GIC 466 : ProcGlobal->subxidStates[pgxactoff].count = 0;
773 CBC 466 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
774 466 : proc->subxidStatus.count = 0;
775 466 : proc->subxidStatus.overflowed = false;
776 ECB : }
777 :
778 : /* Also advance global latestCompletedXid while holding the lock */
779 GIC 297704 : MaintainLatestCompletedXid(latestXid);
780 ECB :
781 : /* Same with xactCompletionCount */
782 GIC 297704 : ShmemVariableCache->xactCompletionCount++;
783 CBC 297704 : }
784 ECB :
785 : /*
786 : * ProcArrayGroupClearXid -- group XID clearing
787 : *
788 : * When we cannot immediately acquire ProcArrayLock in exclusive mode at
789 : * commit time, add ourselves to a list of processes that need their XIDs
790 : * cleared. The first process to add itself to the list will acquire
791 : * ProcArrayLock in exclusive mode and perform ProcArrayEndTransactionInternal
792 : * on behalf of all group members. This avoids a great deal of contention
793 : * around ProcArrayLock when many processes are trying to commit at once,
794 : * since the lock need not be repeatedly handed off from one committing
795 : * process to the next.
796 : */
797 : static void
798 GIC 175 : ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid)
799 ECB : {
800 GIC 175 : PROC_HDR *procglobal = ProcGlobal;
801 ECB : uint32 nextidx;
802 : uint32 wakeidx;
803 :
804 : /* We should definitely have an XID to clear. */
805 GIC 175 : Assert(TransactionIdIsValid(proc->xid));
806 ECB :
807 : /* Add ourselves to the list of processes needing a group XID clear. */
808 GIC 175 : proc->procArrayGroupMember = true;
809 CBC 175 : proc->procArrayGroupMemberXid = latestXid;
810 175 : nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);
811 ECB : while (true)
812 : {
813 GIC 175 : pg_atomic_write_u32(&proc->procArrayGroupNext, nextidx);
814 ECB :
815 GIC 175 : if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst,
816 ECB : &nextidx,
817 GIC 175 : (uint32) proc->pgprocno))
818 CBC 175 : break;
819 ECB : }
820 :
821 : /*
822 : * If the list was not empty, the leader will clear our XID. It is
823 : * impossible to have followers without a leader because the first process
824 : * that has added itself to the list will always have nextidx as
825 : * INVALID_PGPROCNO.
826 : */
827 GIC 175 : if (nextidx != INVALID_PGPROCNO)
828 ECB : {
829 GIC 11 : int extraWaits = 0;
830 ECB :
831 : /* Sleep until the leader clears our XID. */
832 GIC 11 : pgstat_report_wait_start(WAIT_EVENT_PROCARRAY_GROUP_UPDATE);
833 ECB : for (;;)
834 : {
835 : /* acts as a read barrier */
836 GIC 11 : PGSemaphoreLock(proc->sem);
837 CBC 11 : if (!proc->procArrayGroupMember)
838 11 : break;
839 LBC 0 : extraWaits++;
840 EUB : }
841 GIC 11 : pgstat_report_wait_end();
842 ECB :
843 GIC 11 : Assert(pg_atomic_read_u32(&proc->procArrayGroupNext) == INVALID_PGPROCNO);
844 ECB :
845 : /* Fix semaphore count for any absorbed wakeups */
846 GIC 11 : while (extraWaits-- > 0)
847 LBC 0 : PGSemaphoreUnlock(proc->sem);
848 GBC 11 : return;
849 ECB : }
850 :
851 : /* We are the leader. Acquire the lock on behalf of everyone. */
852 GIC 164 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
853 ECB :
854 : /*
855 : * Now that we've got the lock, clear the list of processes waiting for
856 : * group XID clearing, saving a pointer to the head of the list. Trying
857 : * to pop elements one at a time could lead to an ABA problem.
858 : */
859 GIC 164 : nextidx = pg_atomic_exchange_u32(&procglobal->procArrayGroupFirst,
860 ECB : INVALID_PGPROCNO);
861 :
862 : /* Remember head of list so we can perform wakeups after dropping lock. */
863 GIC 164 : wakeidx = nextidx;
864 ECB :
865 : /* Walk the list and clear all XIDs. */
866 GIC 339 : while (nextidx != INVALID_PGPROCNO)
867 ECB : {
868 GIC 175 : PGPROC *nextproc = &allProcs[nextidx];
869 ECB :
870 GIC 175 : ProcArrayEndTransactionInternal(nextproc, nextproc->procArrayGroupMemberXid);
871 ECB :
872 : /* Move to next proc in list. */
873 GIC 175 : nextidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
874 ECB : }
875 :
876 : /* We're done with the lock now. */
877 GIC 164 : LWLockRelease(ProcArrayLock);
878 ECB :
879 : /*
880 : * Now that we've released the lock, go back and wake everybody up. We
881 : * don't do this under the lock so as to keep lock hold times to a
882 : * minimum. The system calls we need to perform to wake other processes
883 : * up are probably much slower than the simple memory writes we did while
884 : * holding the lock.
885 : */
886 GIC 339 : while (wakeidx != INVALID_PGPROCNO)
887 ECB : {
888 GIC 175 : PGPROC *nextproc = &allProcs[wakeidx];
889 ECB :
890 GIC 175 : wakeidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
891 CBC 175 : pg_atomic_write_u32(&nextproc->procArrayGroupNext, INVALID_PGPROCNO);
892 ECB :
893 : /* ensure all previous writes are visible before follower continues. */
894 GIC 175 : pg_write_barrier();
895 ECB :
896 GIC 175 : nextproc->procArrayGroupMember = false;
897 ECB :
898 GIC 175 : if (nextproc != MyProc)
899 CBC 11 : PGSemaphoreUnlock(nextproc->sem);
900 ECB : }
901 : }
902 :
903 : /*
904 : * ProcArrayClearTransaction -- clear the transaction fields
905 : *
906 : * This is used after successfully preparing a 2-phase transaction. We are
907 : * not actually reporting the transaction's XID as no longer running --- it
908 : * will still appear as running because the 2PC's gxact is in the ProcArray
909 : * too. We just have to clear out our own PGPROC.
910 : */
911 : void
912 GIC 361 : ProcArrayClearTransaction(PGPROC *proc)
913 ECB : {
914 : int pgxactoff;
915 :
916 : /*
917 : * Currently we need to lock ProcArrayLock exclusively here, as we
918 : * increment xactCompletionCount below. We also need it at least in shared
919 : * mode for pgproc->pgxactoff to stay the same below.
920 : *
921 : * We could however, as this action does not actually change anyone's view
922 : * of the set of running XIDs (our entry is duplicate with the gxact that
923 : * has already been inserted into the ProcArray), lower the lock level to
924 : * shared if we were to make xactCompletionCount an atomic variable. But
925 : * that doesn't seem worth it currently, as a 2PC commit is heavyweight
926 : * enough for this not to be the bottleneck. If it ever becomes a
927 : * bottleneck it may also be worth considering to combine this with the
928 : * subsequent ProcArrayRemove()
929 : */
930 GIC 361 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
931 ECB :
932 GIC 361 : pgxactoff = proc->pgxactoff;
933 ECB :
934 GIC 361 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
935 CBC 361 : proc->xid = InvalidTransactionId;
936 ECB :
937 GIC 361 : proc->lxid = InvalidLocalTransactionId;
938 CBC 361 : proc->xmin = InvalidTransactionId;
939 361 : proc->recoveryConflictPending = false;
940 ECB :
941 GIC 361 : Assert(!(proc->statusFlags & PROC_VACUUM_STATE_MASK));
942 CBC 361 : Assert(!proc->delayChkptFlags);
943 ECB :
944 : /*
945 : * Need to increment completion count even though transaction hasn't
946 : * really committed yet. The reason for that is that GetSnapshotData()
947 : * omits the xid of the current transaction, thus without the increment we
948 : * otherwise could end up reusing the snapshot later. Which would be bad,
949 : * because it might not count the prepared transaction as running.
950 : */
951 GIC 361 : ShmemVariableCache->xactCompletionCount++;
952 ECB :
953 : /* Clear the subtransaction-XID cache too */
954 GIC 361 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
955 ECB : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
956 GIC 361 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
957 ECB : {
958 GIC 149 : ProcGlobal->subxidStates[pgxactoff].count = 0;
959 CBC 149 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
960 149 : proc->subxidStatus.count = 0;
961 149 : proc->subxidStatus.overflowed = false;
962 ECB : }
963 :
964 GIC 361 : LWLockRelease(ProcArrayLock);
965 CBC 361 : }
966 ECB :
967 : /*
968 : * Update ShmemVariableCache->latestCompletedXid to point to latestXid if
969 : * currently older.
970 : */
971 : static void
972 GIC 298679 : MaintainLatestCompletedXid(TransactionId latestXid)
973 ECB : {
974 GIC 298679 : FullTransactionId cur_latest = ShmemVariableCache->latestCompletedXid;
975 ECB :
976 GIC 298679 : Assert(FullTransactionIdIsValid(cur_latest));
977 CBC 298679 : Assert(!RecoveryInProgress());
978 298679 : Assert(LWLockHeldByMe(ProcArrayLock));
979 ECB :
980 GIC 298679 : if (TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
981 ECB : {
982 GIC 288173 : ShmemVariableCache->latestCompletedXid =
983 CBC 288173 : FullXidRelativeTo(cur_latest, latestXid);
984 ECB : }
985 :
986 GIC 298679 : Assert(IsBootstrapProcessingMode() ||
987 ECB : FullTransactionIdIsNormal(ShmemVariableCache->latestCompletedXid));
988 GIC 298679 : }
989 ECB :
990 : /*
991 : * Same as MaintainLatestCompletedXid, except for use during WAL replay.
992 : */
993 : static void
994 GIC 18250 : MaintainLatestCompletedXidRecovery(TransactionId latestXid)
995 ECB : {
996 GIC 18250 : FullTransactionId cur_latest = ShmemVariableCache->latestCompletedXid;
997 ECB : FullTransactionId rel;
998 :
999 GIC 18250 : Assert(AmStartupProcess() || !IsUnderPostmaster);
1000 CBC 18250 : Assert(LWLockHeldByMe(ProcArrayLock));
1001 ECB :
1002 : /*
1003 : * Need a FullTransactionId to compare latestXid with. Can't rely on
1004 : * latestCompletedXid to be initialized in recovery. But in recovery it's
1005 : * safe to access nextXid without a lock for the startup process.
1006 : */
1007 GIC 18250 : rel = ShmemVariableCache->nextXid;
1008 CBC 18250 : Assert(FullTransactionIdIsValid(ShmemVariableCache->nextXid));
1009 ECB :
1010 GIC 36429 : if (!FullTransactionIdIsValid(cur_latest) ||
1011 CBC 18179 : TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
1012 ECB : {
1013 GIC 14716 : ShmemVariableCache->latestCompletedXid =
1014 CBC 14716 : FullXidRelativeTo(rel, latestXid);
1015 ECB : }
1016 :
1017 GIC 18250 : Assert(FullTransactionIdIsNormal(ShmemVariableCache->latestCompletedXid));
1018 CBC 18250 : }
1019 ECB :
1020 : /*
1021 : * ProcArrayInitRecovery -- initialize recovery xid mgmt environment
1022 : *
1023 : * Remember up to where the startup process initialized the CLOG and subtrans
1024 : * so we can ensure it's initialized gaplessly up to the point where necessary
1025 : * while in recovery.
1026 : */
1027 : void
1028 GIC 71 : ProcArrayInitRecovery(TransactionId initializedUptoXID)
1029 ECB : {
1030 GIC 71 : Assert(standbyState == STANDBY_INITIALIZED);
1031 CBC 71 : Assert(TransactionIdIsNormal(initializedUptoXID));
1032 ECB :
1033 : /*
1034 : * we set latestObservedXid to the xid SUBTRANS has been initialized up
1035 : * to, so we can extend it from that point onwards in
1036 : * RecordKnownAssignedTransactionIds, and when we get consistent in
1037 : * ProcArrayApplyRecoveryInfo().
1038 : */
1039 GIC 71 : latestObservedXid = initializedUptoXID;
1040 CBC 71 : TransactionIdRetreat(latestObservedXid);
1041 71 : }
1042 ECB :
1043 : /*
1044 : * ProcArrayApplyRecoveryInfo -- apply recovery info about xids
1045 : *
1046 : * Takes us through 3 states: Initialized, Pending and Ready.
1047 : * Normal case is to go all the way to Ready straight away, though there
1048 : * are atypical cases where we need to take it in steps.
1049 : *
1050 : * Use the data about running transactions on the primary to create the initial
1051 : * state of KnownAssignedXids. We also use these records to regularly prune
1052 : * KnownAssignedXids because we know it is possible that some transactions
1053 : * with FATAL errors fail to write abort records, which could cause eventual
1054 : * overflow.
1055 : *
1056 : * See comments for LogStandbySnapshot().
1057 : */
1058 : void
1059 GIC 195 : ProcArrayApplyRecoveryInfo(RunningTransactions running)
1060 ECB : {
1061 : TransactionId *xids;
1062 : int nxids;
1063 : int i;
1064 :
1065 GIC 195 : Assert(standbyState >= STANDBY_INITIALIZED);
1066 CBC 195 : Assert(TransactionIdIsValid(running->nextXid));
1067 195 : Assert(TransactionIdIsValid(running->oldestRunningXid));
1068 195 : Assert(TransactionIdIsNormal(running->latestCompletedXid));
1069 ECB :
1070 : /*
1071 : * Remove stale transactions, if any.
1072 : */
1073 GIC 195 : ExpireOldKnownAssignedTransactionIds(running->oldestRunningXid);
1074 ECB :
1075 : /*
1076 : * Remove stale locks, if any.
1077 : */
1078 GIC 195 : StandbyReleaseOldLocks(running->oldestRunningXid);
1079 ECB :
1080 : /*
1081 : * If our snapshot is already valid, nothing else to do...
1082 : */
1083 GIC 195 : if (standbyState == STANDBY_SNAPSHOT_READY)
1084 CBC 124 : return;
1085 ECB :
1086 : /*
1087 : * If our initial RunningTransactionsData had an overflowed snapshot then
1088 : * we knew we were missing some subxids from our snapshot. If we continue
1089 : * to see overflowed snapshots then we might never be able to start up, so
1090 : * we make another test to see if our snapshot is now valid. We know that
1091 : * the missing subxids are equal to or earlier than nextXid. After we
1092 : * initialise we continue to apply changes during recovery, so once the
1093 : * oldestRunningXid is later than the nextXid from the initial snapshot we
1094 : * know that we no longer have missing information and can mark the
1095 : * snapshot as valid.
1096 : */
1097 GIC 71 : if (standbyState == STANDBY_SNAPSHOT_PENDING)
1098 ECB : {
1099 : /*
1100 : * If the snapshot isn't overflowed or if its empty we can reset our
1101 : * pending state and use this snapshot instead.
1102 : */
1103 UIC 0 : if (!running->subxid_overflow || running->xcnt == 0)
1104 EUB : {
1105 : /*
1106 : * If we have already collected known assigned xids, we need to
1107 : * throw them away before we apply the recovery snapshot.
1108 : */
1109 UIC 0 : KnownAssignedXidsReset();
1110 UBC 0 : standbyState = STANDBY_INITIALIZED;
1111 EUB : }
1112 : else
1113 : {
1114 UIC 0 : if (TransactionIdPrecedes(standbySnapshotPendingXmin,
1115 EUB : running->oldestRunningXid))
1116 : {
1117 UIC 0 : standbyState = STANDBY_SNAPSHOT_READY;
1118 UBC 0 : elog(trace_recovery(DEBUG1),
1119 EUB : "recovery snapshots are now enabled");
1120 : }
1121 : else
1122 UIC 0 : elog(trace_recovery(DEBUG1),
1123 EUB : "recovery snapshot waiting for non-overflowed snapshot or "
1124 : "until oldest active xid on standby is at least %u (now %u)",
1125 : standbySnapshotPendingXmin,
1126 : running->oldestRunningXid);
1127 UIC 0 : return;
1128 EUB : }
1129 : }
1130 :
1131 GIC 71 : Assert(standbyState == STANDBY_INITIALIZED);
1132 ECB :
1133 : /*
1134 : * NB: this can be reached at least twice, so make sure new code can deal
1135 : * with that.
1136 : */
1137 :
1138 : /*
1139 : * Nobody else is running yet, but take locks anyhow
1140 : */
1141 GIC 71 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1142 ECB :
1143 : /*
1144 : * KnownAssignedXids is sorted so we cannot just add the xids, we have to
1145 : * sort them first.
1146 : *
1147 : * Some of the new xids are top-level xids and some are subtransactions.
1148 : * We don't call SubTransSetParent because it doesn't matter yet. If we
1149 : * aren't overflowed then all xids will fit in snapshot and so we don't
1150 : * need subtrans. If we later overflow, an xid assignment record will add
1151 : * xids to subtrans. If RunningTransactionsData is overflowed then we
1152 : * don't have enough information to correctly update subtrans anyway.
1153 : */
1154 :
1155 : /*
1156 : * Allocate a temporary array to avoid modifying the array passed as
1157 : * argument.
1158 : */
1159 GIC 71 : xids = palloc(sizeof(TransactionId) * (running->xcnt + running->subxcnt));
1160 ECB :
1161 : /*
1162 : * Add to the temp array any xids which have not already completed.
1163 : */
1164 GIC 71 : nxids = 0;
1165 CBC 73 : for (i = 0; i < running->xcnt + running->subxcnt; i++)
1166 ECB : {
1167 GIC 2 : TransactionId xid = running->xids[i];
1168 ECB :
1169 : /*
1170 : * The running-xacts snapshot can contain xids that were still visible
1171 : * in the procarray when the snapshot was taken, but were already
1172 : * WAL-logged as completed. They're not running anymore, so ignore
1173 : * them.
1174 : */
1175 GIC 2 : if (TransactionIdDidCommit(xid) || TransactionIdDidAbort(xid))
1176 LBC 0 : continue;
1177 EUB :
1178 GIC 2 : xids[nxids++] = xid;
1179 ECB : }
1180 :
1181 GIC 71 : if (nxids > 0)
1182 ECB : {
1183 GIC 2 : if (procArray->numKnownAssignedXids != 0)
1184 ECB : {
1185 UIC 0 : LWLockRelease(ProcArrayLock);
1186 UBC 0 : elog(ERROR, "KnownAssignedXids is not empty");
1187 EUB : }
1188 :
1189 : /*
1190 : * Sort the array so that we can add them safely into
1191 : * KnownAssignedXids.
1192 : *
1193 : * We have to sort them logically, because in KnownAssignedXidsAdd we
1194 : * call TransactionIdFollowsOrEquals and so on. But we know these XIDs
1195 : * come from RUNNING_XACTS, which means there are only normal XIDs
1196 : * from the same epoch, so this is safe.
1197 : */
1198 GIC 2 : qsort(xids, nxids, sizeof(TransactionId), xidLogicalComparator);
1199 ECB :
1200 : /*
1201 : * Add the sorted snapshot into KnownAssignedXids. The running-xacts
1202 : * snapshot may include duplicated xids because of prepared
1203 : * transactions, so ignore them.
1204 : */
1205 GIC 4 : for (i = 0; i < nxids; i++)
1206 ECB : {
1207 GIC 2 : if (i > 0 && TransactionIdEquals(xids[i - 1], xids[i]))
1208 ECB : {
1209 UIC 0 : elog(DEBUG1,
1210 EUB : "found duplicated transaction %u for KnownAssignedXids insertion",
1211 : xids[i]);
1212 UIC 0 : continue;
1213 EUB : }
1214 GIC 2 : KnownAssignedXidsAdd(xids[i], xids[i], true);
1215 ECB : }
1216 :
1217 GIC 2 : KnownAssignedXidsDisplay(trace_recovery(DEBUG3));
1218 ECB : }
1219 :
1220 GIC 71 : pfree(xids);
1221 ECB :
1222 : /*
1223 : * latestObservedXid is at least set to the point where SUBTRANS was
1224 : * started up to (cf. ProcArrayInitRecovery()) or to the biggest xid
1225 : * RecordKnownAssignedTransactionIds() was called for. Initialize
1226 : * subtrans from thereon, up to nextXid - 1.
1227 : *
1228 : * We need to duplicate parts of RecordKnownAssignedTransactionId() here,
1229 : * because we've just added xids to the known assigned xids machinery that
1230 : * haven't gone through RecordKnownAssignedTransactionId().
1231 : */
1232 GIC 71 : Assert(TransactionIdIsNormal(latestObservedXid));
1233 CBC 71 : TransactionIdAdvance(latestObservedXid);
1234 142 : while (TransactionIdPrecedes(latestObservedXid, running->nextXid))
1235 ECB : {
1236 UIC 0 : ExtendSUBTRANS(latestObservedXid);
1237 UBC 0 : TransactionIdAdvance(latestObservedXid);
1238 EUB : }
1239 GIC 71 : TransactionIdRetreat(latestObservedXid); /* = running->nextXid - 1 */
1240 ECB :
1241 : /* ----------
1242 : * Now we've got the running xids we need to set the global values that
1243 : * are used to track snapshots as they evolve further.
1244 : *
1245 : * - latestCompletedXid which will be the xmax for snapshots
1246 : * - lastOverflowedXid which shows whether snapshots overflow
1247 : * - nextXid
1248 : *
1249 : * If the snapshot overflowed, then we still initialise with what we know,
1250 : * but the recovery snapshot isn't fully valid yet because we know there
1251 : * are some subxids missing. We don't know the specific subxids that are
1252 : * missing, so conservatively assume the last one is latestObservedXid.
1253 : * ----------
1254 : */
1255 GIC 71 : if (running->subxid_overflow)
1256 ECB : {
1257 UIC 0 : standbyState = STANDBY_SNAPSHOT_PENDING;
1258 EUB :
1259 UIC 0 : standbySnapshotPendingXmin = latestObservedXid;
1260 UBC 0 : procArray->lastOverflowedXid = latestObservedXid;
1261 EUB : }
1262 : else
1263 : {
1264 GIC 71 : standbyState = STANDBY_SNAPSHOT_READY;
1265 ECB :
1266 GIC 71 : standbySnapshotPendingXmin = InvalidTransactionId;
1267 ECB : }
1268 :
1269 : /*
1270 : * If a transaction wrote a commit record in the gap between taking and
1271 : * logging the snapshot then latestCompletedXid may already be higher than
1272 : * the value from the snapshot, so check before we use the incoming value.
1273 : * It also might not yet be set at all.
1274 : */
1275 GIC 71 : MaintainLatestCompletedXidRecovery(running->latestCompletedXid);
1276 ECB :
1277 : /*
1278 : * NB: No need to increment ShmemVariableCache->xactCompletionCount here,
1279 : * nobody can see it yet.
1280 : */
1281 :
1282 GIC 71 : LWLockRelease(ProcArrayLock);
1283 ECB :
1284 : /* ShmemVariableCache->nextXid must be beyond any observed xid. */
1285 GIC 71 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
1286 ECB :
1287 GIC 71 : Assert(FullTransactionIdIsValid(ShmemVariableCache->nextXid));
1288 ECB :
1289 GIC 71 : KnownAssignedXidsDisplay(trace_recovery(DEBUG3));
1290 CBC 71 : if (standbyState == STANDBY_SNAPSHOT_READY)
1291 71 : elog(trace_recovery(DEBUG1), "recovery snapshots are now enabled");
1292 ECB : else
1293 UIC 0 : elog(trace_recovery(DEBUG1),
1294 EUB : "recovery snapshot waiting for non-overflowed snapshot or "
1295 : "until oldest active xid on standby is at least %u (now %u)",
1296 : standbySnapshotPendingXmin,
1297 : running->oldestRunningXid);
1298 : }
1299 :
1300 : /*
1301 : * ProcArrayApplyXidAssignment
1302 : * Process an XLOG_XACT_ASSIGNMENT WAL record
1303 : */
1304 : void
1305 GIC 21 : ProcArrayApplyXidAssignment(TransactionId topxid,
1306 ECB : int nsubxids, TransactionId *subxids)
1307 : {
1308 : TransactionId max_xid;
1309 : int i;
1310 :
1311 GIC 21 : Assert(standbyState >= STANDBY_INITIALIZED);
1312 ECB :
1313 GIC 21 : max_xid = TransactionIdLatest(topxid, nsubxids, subxids);
1314 ECB :
1315 : /*
1316 : * Mark all the subtransactions as observed.
1317 : *
1318 : * NOTE: This will fail if the subxid contains too many previously
1319 : * unobserved xids to fit into known-assigned-xids. That shouldn't happen
1320 : * as the code stands, because xid-assignment records should never contain
1321 : * more than PGPROC_MAX_CACHED_SUBXIDS entries.
1322 : */
1323 GIC 21 : RecordKnownAssignedTransactionIds(max_xid);
1324 ECB :
1325 : /*
1326 : * Notice that we update pg_subtrans with the top-level xid, rather than
1327 : * the parent xid. This is a difference between normal processing and
1328 : * recovery, yet is still correct in all cases. The reason is that
1329 : * subtransaction commit is not marked in clog until commit processing, so
1330 : * all aborted subtransactions have already been clearly marked in clog.
1331 : * As a result we are able to refer directly to the top-level
1332 : * transaction's state rather than skipping through all the intermediate
1333 : * states in the subtransaction tree. This should be the first time we
1334 : * have attempted to SubTransSetParent().
1335 : */
1336 GIC 1365 : for (i = 0; i < nsubxids; i++)
1337 CBC 1344 : SubTransSetParent(subxids[i], topxid);
1338 ECB :
1339 : /* KnownAssignedXids isn't maintained yet, so we're done for now */
1340 GIC 21 : if (standbyState == STANDBY_INITIALIZED)
1341 LBC 0 : return;
1342 EUB :
1343 : /*
1344 : * Uses same locking as transaction commit
1345 : */
1346 GIC 21 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1347 ECB :
1348 : /*
1349 : * Remove subxids from known-assigned-xacts.
1350 : */
1351 GIC 21 : KnownAssignedXidsRemoveTree(InvalidTransactionId, nsubxids, subxids);
1352 ECB :
1353 : /*
1354 : * Advance lastOverflowedXid to be at least the last of these subxids.
1355 : */
1356 GIC 21 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, max_xid))
1357 CBC 21 : procArray->lastOverflowedXid = max_xid;
1358 ECB :
1359 GIC 21 : LWLockRelease(ProcArrayLock);
1360 ECB : }
1361 :
1362 : /*
1363 : * TransactionIdIsInProgress -- is given transaction running in some backend
1364 : *
1365 : * Aside from some shortcuts such as checking RecentXmin and our own Xid,
1366 : * there are four possibilities for finding a running transaction:
1367 : *
1368 : * 1. The given Xid is a main transaction Id. We will find this out cheaply
1369 : * by looking at ProcGlobal->xids.
1370 : *
1371 : * 2. The given Xid is one of the cached subxact Xids in the PGPROC array.
1372 : * We can find this out cheaply too.
1373 : *
1374 : * 3. In Hot Standby mode, we must search the KnownAssignedXids list to see
1375 : * if the Xid is running on the primary.
1376 : *
1377 : * 4. Search the SubTrans tree to find the Xid's topmost parent, and then see
1378 : * if that is running according to ProcGlobal->xids[] or KnownAssignedXids.
1379 : * This is the slowest way, but sadly it has to be done always if the others
1380 : * failed, unless we see that the cached subxact sets are complete (none have
1381 : * overflowed).
1382 : *
1383 : * ProcArrayLock has to be held while we do 1, 2, 3. If we save the top Xids
1384 : * while doing 1 and 3, we can release the ProcArrayLock while we do 4.
1385 : * This buys back some concurrency (and we can't retrieve the main Xids from
1386 : * ProcGlobal->xids[] again anyway; see GetNewTransactionId).
1387 : */
1388 : bool
1389 GIC 7228314 : TransactionIdIsInProgress(TransactionId xid)
1390 ECB : {
1391 : static TransactionId *xids = NULL;
1392 : static TransactionId *other_xids;
1393 : XidCacheStatus *other_subxidstates;
1394 GIC 7228314 : int nxids = 0;
1395 CBC 7228314 : ProcArrayStruct *arrayP = procArray;
1396 ECB : TransactionId topxid;
1397 : TransactionId latestCompletedXid;
1398 : int mypgxactoff;
1399 : int numProcs;
1400 : int j;
1401 :
1402 : /*
1403 : * Don't bother checking a transaction older than RecentXmin; it could not
1404 : * possibly still be running. (Note: in particular, this guarantees that
1405 : * we reject InvalidTransactionId, FrozenTransactionId, etc as not
1406 : * running.)
1407 : */
1408 GIC 7228314 : if (TransactionIdPrecedes(xid, RecentXmin))
1409 ECB : {
1410 : xc_by_recent_xmin_inc();
1411 GIC 5968169 : return false;
1412 ECB : }
1413 :
1414 : /*
1415 : * We may have just checked the status of this transaction, so if it is
1416 : * already known to be completed, we can fall out without any access to
1417 : * shared memory.
1418 : */
1419 GIC 1260145 : if (TransactionIdEquals(cachedXidIsNotInProgress, xid))
1420 ECB : {
1421 : xc_by_known_xact_inc();
1422 GIC 953515 : return false;
1423 ECB : }
1424 :
1425 : /*
1426 : * Also, we can handle our own transaction (and subtransactions) without
1427 : * any access to shared memory.
1428 : */
1429 GIC 306630 : if (TransactionIdIsCurrentTransactionId(xid))
1430 ECB : {
1431 : xc_by_my_xact_inc();
1432 GIC 287949 : return true;
1433 ECB : }
1434 :
1435 : /*
1436 : * If first time through, get workspace to remember main XIDs in. We
1437 : * malloc it permanently to avoid repeated palloc/pfree overhead.
1438 : */
1439 GIC 18681 : if (xids == NULL)
1440 ECB : {
1441 : /*
1442 : * In hot standby mode, reserve enough space to hold all xids in the
1443 : * known-assigned list. If we later finish recovery, we no longer need
1444 : * the bigger array, but we don't bother to shrink it.
1445 : */
1446 GIC 600 : int maxxids = RecoveryInProgress() ? TOTAL_MAX_CACHED_SUBXIDS : arrayP->maxProcs;
1447 ECB :
1448 GIC 600 : xids = (TransactionId *) malloc(maxxids * sizeof(TransactionId));
1449 CBC 600 : if (xids == NULL)
1450 LBC 0 : ereport(ERROR,
1451 EUB : (errcode(ERRCODE_OUT_OF_MEMORY),
1452 : errmsg("out of memory")));
1453 : }
1454 :
1455 GIC 18681 : other_xids = ProcGlobal->xids;
1456 CBC 18681 : other_subxidstates = ProcGlobal->subxidStates;
1457 ECB :
1458 GIC 18681 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1459 ECB :
1460 : /*
1461 : * Now that we have the lock, we can check latestCompletedXid; if the
1462 : * target Xid is after that, it's surely still running.
1463 : */
1464 GIC 18681 : latestCompletedXid =
1465 CBC 18681 : XidFromFullTransactionId(ShmemVariableCache->latestCompletedXid);
1466 18681 : if (TransactionIdPrecedes(latestCompletedXid, xid))
1467 ECB : {
1468 GIC 4590 : LWLockRelease(ProcArrayLock);
1469 ECB : xc_by_latest_xid_inc();
1470 GIC 4590 : return true;
1471 ECB : }
1472 :
1473 : /* No shortcuts, gotta grovel through the array */
1474 GIC 14091 : mypgxactoff = MyProc->pgxactoff;
1475 CBC 14091 : numProcs = arrayP->numProcs;
1476 130983 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
1477 ECB : {
1478 : int pgprocno;
1479 : PGPROC *proc;
1480 : TransactionId pxid;
1481 : int pxids;
1482 :
1483 : /* Ignore ourselves --- dealt with it above */
1484 GIC 119817 : if (pgxactoff == mypgxactoff)
1485 CBC 12122 : continue;
1486 ECB :
1487 : /* Fetch xid just once - see GetNewTransactionId */
1488 GIC 107695 : pxid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
1489 ECB :
1490 GIC 107695 : if (!TransactionIdIsValid(pxid))
1491 CBC 75489 : continue;
1492 ECB :
1493 : /*
1494 : * Step 1: check the main Xid
1495 : */
1496 GIC 32206 : if (TransactionIdEquals(pxid, xid))
1497 ECB : {
1498 GIC 2387 : LWLockRelease(ProcArrayLock);
1499 ECB : xc_by_main_xid_inc();
1500 GIC 2387 : return true;
1501 ECB : }
1502 :
1503 : /*
1504 : * We can ignore main Xids that are younger than the target Xid, since
1505 : * the target could not possibly be their child.
1506 : */
1507 GIC 29819 : if (TransactionIdPrecedes(xid, pxid))
1508 CBC 13120 : continue;
1509 ECB :
1510 : /*
1511 : * Step 2: check the cached child-Xids arrays
1512 : */
1513 GIC 16699 : pxids = other_subxidstates[pgxactoff].count;
1514 CBC 16699 : pg_read_barrier(); /* pairs with barrier in GetNewTransactionId() */
1515 16699 : pgprocno = arrayP->pgprocnos[pgxactoff];
1516 16699 : proc = &allProcs[pgprocno];
1517 52181 : for (j = pxids - 1; j >= 0; j--)
1518 ECB : {
1519 : /* Fetch xid just once - see GetNewTransactionId */
1520 GIC 36020 : TransactionId cxid = UINT32_ACCESS_ONCE(proc->subxids.xids[j]);
1521 ECB :
1522 GIC 36020 : if (TransactionIdEquals(cxid, xid))
1523 ECB : {
1524 GIC 538 : LWLockRelease(ProcArrayLock);
1525 ECB : xc_by_child_xid_inc();
1526 GIC 538 : return true;
1527 ECB : }
1528 : }
1529 :
1530 : /*
1531 : * Save the main Xid for step 4. We only need to remember main Xids
1532 : * that have uncached children. (Note: there is no race condition
1533 : * here because the overflowed flag cannot be cleared, only set, while
1534 : * we hold ProcArrayLock. So we can't miss an Xid that we need to
1535 : * worry about.)
1536 : */
1537 GIC 16161 : if (other_subxidstates[pgxactoff].overflowed)
1538 CBC 157 : xids[nxids++] = pxid;
1539 ECB : }
1540 :
1541 : /*
1542 : * Step 3: in hot standby mode, check the known-assigned-xids list. XIDs
1543 : * in the list must be treated as running.
1544 : */
1545 GIC 11166 : if (RecoveryInProgress())
1546 ECB : {
1547 : /* none of the PGPROC entries should have XIDs in hot standby mode */
1548 UIC 0 : Assert(nxids == 0);
1549 EUB :
1550 UIC 0 : if (KnownAssignedXidExists(xid))
1551 EUB : {
1552 UIC 0 : LWLockRelease(ProcArrayLock);
1553 EUB : xc_by_known_assigned_inc();
1554 UIC 0 : return true;
1555 EUB : }
1556 :
1557 : /*
1558 : * If the KnownAssignedXids overflowed, we have to check pg_subtrans
1559 : * too. Fetch all xids from KnownAssignedXids that are lower than
1560 : * xid, since if xid is a subtransaction its parent will always have a
1561 : * lower value. Note we will collect both main and subXIDs here, but
1562 : * there's no help for it.
1563 : */
1564 UIC 0 : if (TransactionIdPrecedesOrEquals(xid, procArray->lastOverflowedXid))
1565 UBC 0 : nxids = KnownAssignedXidsGet(xids, xid);
1566 EUB : }
1567 :
1568 GIC 11166 : LWLockRelease(ProcArrayLock);
1569 ECB :
1570 : /*
1571 : * If none of the relevant caches overflowed, we know the Xid is not
1572 : * running without even looking at pg_subtrans.
1573 : */
1574 GIC 11166 : if (nxids == 0)
1575 ECB : {
1576 : xc_no_overflow_inc();
1577 GIC 11009 : cachedXidIsNotInProgress = xid;
1578 CBC 11009 : return false;
1579 ECB : }
1580 :
1581 : /*
1582 : * Step 4: have to check pg_subtrans.
1583 : *
1584 : * At this point, we know it's either a subtransaction of one of the Xids
1585 : * in xids[], or it's not running. If it's an already-failed
1586 : * subtransaction, we want to say "not running" even though its parent may
1587 : * still be running. So first, check pg_xact to see if it's been aborted.
1588 : */
1589 : xc_slow_answer_inc();
1590 :
1591 GIC 157 : if (TransactionIdDidAbort(xid))
1592 ECB : {
1593 UIC 0 : cachedXidIsNotInProgress = xid;
1594 UBC 0 : return false;
1595 EUB : }
1596 :
1597 : /*
1598 : * It isn't aborted, so check whether the transaction tree it belongs to
1599 : * is still running (or, more precisely, whether it was running when we
1600 : * held ProcArrayLock).
1601 : */
1602 GIC 157 : topxid = SubTransGetTopmostTransaction(xid);
1603 CBC 157 : Assert(TransactionIdIsValid(topxid));
1604 GNC 314 : if (!TransactionIdEquals(topxid, xid) &&
1605 157 : pg_lfind32(topxid, xids, nxids))
1606 157 : return true;
1607 :
1608 UIC 0 : cachedXidIsNotInProgress = xid;
1609 0 : return false;
1610 : }
1611 :
1612 : /*
1613 : * TransactionIdIsActive -- is xid the top-level XID of an active backend?
1614 : *
1615 : * This differs from TransactionIdIsInProgress in that it ignores prepared
1616 : * transactions, as well as transactions running on the primary if we're in
1617 EUB : * hot standby. Also, we ignore subtransactions since that's not needed
1618 : * for current uses.
1619 : */
1620 : bool
1621 UBC 0 : TransactionIdIsActive(TransactionId xid)
1622 : {
1623 UIC 0 : bool result = false;
1624 0 : ProcArrayStruct *arrayP = procArray;
1625 0 : TransactionId *other_xids = ProcGlobal->xids;
1626 : int i;
1627 :
1628 EUB : /*
1629 : * Don't bother checking a transaction older than RecentXmin; it could not
1630 : * possibly still be running.
1631 : */
1632 UIC 0 : if (TransactionIdPrecedes(xid, RecentXmin))
1633 UBC 0 : return false;
1634 :
1635 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1636 EUB :
1637 UIC 0 : for (i = 0; i < arrayP->numProcs; i++)
1638 : {
1639 0 : int pgprocno = arrayP->pgprocnos[i];
1640 UBC 0 : PGPROC *proc = &allProcs[pgprocno];
1641 : TransactionId pxid;
1642 EUB :
1643 : /* Fetch xid just once - see GetNewTransactionId */
1644 UIC 0 : pxid = UINT32_ACCESS_ONCE(other_xids[i]);
1645 EUB :
1646 UBC 0 : if (!TransactionIdIsValid(pxid))
1647 UIC 0 : continue;
1648 EUB :
1649 UIC 0 : if (proc->pid == 0)
1650 UBC 0 : continue; /* ignore prepared transactions */
1651 EUB :
1652 UIC 0 : if (TransactionIdEquals(pxid, xid))
1653 : {
1654 0 : result = true;
1655 UBC 0 : break;
1656 : }
1657 EUB : }
1658 :
1659 UIC 0 : LWLockRelease(ProcArrayLock);
1660 :
1661 0 : return result;
1662 : }
1663 :
1664 :
1665 : /*
1666 : * Determine XID horizons.
1667 : *
1668 : * This is used by wrapper functions like GetOldestNonRemovableTransactionId()
1669 : * (for VACUUM), GetReplicationHorizons() (for hot_standby_feedback), etc as
1670 : * well as "internally" by GlobalVisUpdate() (see comment above struct
1671 : * GlobalVisState).
1672 : *
1673 : * See the definition of ComputeXidHorizonsResult for the various computed
1674 : * horizons.
1675 : *
1676 : * For VACUUM separate horizons (used to decide which deleted tuples must
1677 : * be preserved), for shared and non-shared tables are computed. For shared
1678 : * relations backends in all databases must be considered, but for non-shared
1679 : * relations that's not required, since only backends in my own database could
1680 : * ever see the tuples in them. Also, we can ignore concurrently running lazy
1681 : * VACUUMs because (a) they must be working on other tables, and (b) they
1682 : * don't need to do snapshot-based lookups.
1683 : *
1684 : * This also computes a horizon used to truncate pg_subtrans. For that
1685 : * backends in all databases have to be considered, and concurrently running
1686 : * lazy VACUUMs cannot be ignored, as they still may perform pg_subtrans
1687 : * accesses.
1688 : *
1689 : * Note: we include all currently running xids in the set of considered xids.
1690 : * This ensures that if a just-started xact has not yet set its snapshot,
1691 : * when it does set the snapshot it cannot set xmin less than what we compute.
1692 : * See notes in src/backend/access/transam/README.
1693 : *
1694 : * Note: despite the above, it's possible for the calculated values to move
1695 : * backwards on repeated calls. The calculated values are conservative, so
1696 : * that anything older is definitely not considered as running by anyone
1697 : * anymore, but the exact values calculated depend on a number of things. For
1698 : * example, if there are no transactions running in the current database, the
1699 : * horizon for normal tables will be latestCompletedXid. If a transaction
1700 : * begins after that, its xmin will include in-progress transactions in other
1701 : * databases that started earlier, so another call will return a lower value.
1702 : * Nonetheless it is safe to vacuum a table in the current database with the
1703 : * first result. There are also replication-related effects: a walsender
1704 : * process can set its xmin based on transactions that are no longer running
1705 : * on the primary but are still being replayed on the standby, thus possibly
1706 : * making the values go backwards. In this case there is a possibility that
1707 : * we lose data that the standby would like to have, but unless the standby
1708 : * uses a replication slot to make its xmin persistent there is little we can
1709 : * do about that --- data is only protected if the walsender runs continuously
1710 : * while queries are executed on the standby. (The Hot Standby code deals
1711 : * with such cases by failing standby queries that needed to access
1712 : * already-removed data, so there's no integrity bug.) The computed values
1713 : * are also adjusted with vacuum_defer_cleanup_age, so increasing that setting
1714 : * on the fly is another easy way to make horizons move backwards, with no
1715 : * consequences for data integrity.
1716 : *
1717 : * Note: the approximate horizons (see definition of GlobalVisState) are
1718 : * updated by the computations done here. That's currently required for
1719 : * correctness and a small optimization. Without doing so it's possible that
1720 : * heap vacuum's call to heap_page_prune() uses a more conservative horizon
1721 ECB : * than later when deciding which tuples can be removed - which the code
1722 : * doesn't expect (breaking HOT).
1723 : */
1724 : static void
1725 CBC 204530 : ComputeXidHorizons(ComputeXidHorizonsResult *h)
1726 ECB : {
1727 GIC 204530 : ProcArrayStruct *arrayP = procArray;
1728 : TransactionId kaxmin;
1729 CBC 204530 : bool in_recovery = RecoveryInProgress();
1730 GIC 204530 : TransactionId *other_xids = ProcGlobal->xids;
1731 ECB :
1732 : /* inferred after ProcArrayLock is released */
1733 CBC 204530 : h->catalog_oldest_nonremovable = InvalidTransactionId;
1734 :
1735 GIC 204530 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1736 :
1737 204530 : h->latest_completed = ShmemVariableCache->latestCompletedXid;
1738 :
1739 : /*
1740 : * We initialize the MIN() calculation with latestCompletedXid + 1. This
1741 : * is a lower bound for the XIDs that might appear in the ProcArray later,
1742 : * and so protects us against overestimating the result due to future
1743 : * additions.
1744 ECB : */
1745 : {
1746 : TransactionId initial;
1747 :
1748 CBC 204530 : initial = XidFromFullTransactionId(h->latest_completed);
1749 204530 : Assert(TransactionIdIsValid(initial));
1750 204530 : TransactionIdAdvance(initial);
1751 :
1752 GIC 204530 : h->oldest_considered_running = initial;
1753 204530 : h->shared_oldest_nonremovable = initial;
1754 204530 : h->data_oldest_nonremovable = initial;
1755 :
1756 : /*
1757 : * Only modifications made by this backend affect the horizon for
1758 : * temporary relations. Instead of a check in each iteration of the
1759 : * loop over all PGPROCs it is cheaper to just initialize to the
1760 : * current top-level xid any.
1761 : *
1762 : * Without an assigned xid we could use a horizon as aggressive as
1763 : * ReadNewTransactionid(), but we can get away with the much cheaper
1764 ECB : * latestCompletedXid + 1: If this backend has no xid there, by
1765 : * definition, can't be any newer changes in the temp table than
1766 : * latestCompletedXid.
1767 : */
1768 GIC 204530 : if (TransactionIdIsValid(MyProc->xid))
1769 53112 : h->temp_oldest_nonremovable = MyProc->xid;
1770 : else
1771 151418 : h->temp_oldest_nonremovable = initial;
1772 : }
1773 :
1774 : /*
1775 ECB : * Fetch slot horizons while ProcArrayLock is held - the
1776 : * LWLockAcquire/LWLockRelease are a barrier, ensuring this happens inside
1777 : * the lock.
1778 : */
1779 GIC 204530 : h->slot_xmin = procArray->replication_slot_xmin;
1780 CBC 204530 : h->slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
1781 ECB :
1782 CBC 712499 : for (int index = 0; index < arrayP->numProcs; index++)
1783 : {
1784 GIC 507969 : int pgprocno = arrayP->pgprocnos[index];
1785 507969 : PGPROC *proc = &allProcs[pgprocno];
1786 507969 : int8 statusFlags = ProcGlobal->statusFlags[index];
1787 ECB : TransactionId xid;
1788 : TransactionId xmin;
1789 :
1790 : /* Fetch xid just once - see GetNewTransactionId */
1791 GIC 507969 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
1792 507969 : xmin = UINT32_ACCESS_ONCE(proc->xmin);
1793 :
1794 : /*
1795 : * Consider both the transaction's Xmin, and its Xid.
1796 : *
1797 ECB : * We must check both because a transaction might have an Xmin but not
1798 : * (yet) an Xid; conversely, if it has an Xid, that could determine
1799 : * some not-yet-set Xmin.
1800 : */
1801 CBC 507969 : xmin = TransactionIdOlder(xmin, xid);
1802 :
1803 : /* if neither is set, this proc doesn't influence the horizon */
1804 GIC 507969 : if (!TransactionIdIsValid(xmin))
1805 205291 : continue;
1806 :
1807 : /*
1808 : * Don't ignore any procs when determining which transactions might be
1809 ECB : * considered running. While slots should ensure logical decoding
1810 : * backends are protected even without this check, it can't hurt to
1811 : * include them here as well..
1812 : */
1813 GIC 302678 : h->oldest_considered_running =
1814 302678 : TransactionIdOlder(h->oldest_considered_running, xmin);
1815 :
1816 : /*
1817 ECB : * Skip over backends either vacuuming (which is ok with rows being
1818 : * removed, as long as pg_subtrans is not truncated) or doing logical
1819 : * decoding (which manages xmin separately, check below).
1820 : */
1821 CBC 302678 : if (statusFlags & (PROC_IN_VACUUM | PROC_IN_LOGICAL_DECODING))
1822 37947 : continue;
1823 :
1824 : /* shared tables need to take backends in all databases into account */
1825 GIC 264731 : h->shared_oldest_nonremovable =
1826 264731 : TransactionIdOlder(h->shared_oldest_nonremovable, xmin);
1827 :
1828 : /*
1829 : * Normally sessions in other databases are ignored for anything but
1830 : * the shared horizon.
1831 : *
1832 : * However, include them when MyDatabaseId is not (yet) set. A
1833 : * backend in the process of starting up must not compute a "too
1834 : * aggressive" horizon, otherwise we could end up using it to prune
1835 : * still-needed data away. If the current backend never connects to a
1836 : * database this is harmless, because data_oldest_nonremovable will
1837 : * never be utilized.
1838 : *
1839 : * Also, sessions marked with PROC_AFFECTS_ALL_HORIZONS should always
1840 : * be included. (This flag is used for hot standby feedback, which
1841 : * can't be tied to a specific database.)
1842 : *
1843 ECB : * Also, while in recovery we cannot compute an accurate per-database
1844 : * horizon, as all xids are managed via the KnownAssignedXids
1845 : * machinery.
1846 : */
1847 GIC 264731 : if (proc->databaseId == MyDatabaseId ||
1848 CBC 6263 : MyDatabaseId == InvalidOid ||
1849 300 : (statusFlags & PROC_AFFECTS_ALL_HORIZONS) ||
1850 : in_recovery)
1851 : {
1852 GIC 264431 : h->data_oldest_nonremovable =
1853 264431 : TransactionIdOlder(h->data_oldest_nonremovable, xmin);
1854 : }
1855 : }
1856 :
1857 ECB : /*
1858 : * If in recovery fetch oldest xid in KnownAssignedXids, will be applied
1859 : * after lock is released.
1860 : */
1861 GIC 204530 : if (in_recovery)
1862 116 : kaxmin = KnownAssignedXidsGetOldestXmin();
1863 :
1864 ECB : /*
1865 : * No other information from shared state is needed, release the lock
1866 : * immediately. The rest of the computations can be done without a lock.
1867 : */
1868 CBC 204530 : LWLockRelease(ProcArrayLock);
1869 ECB :
1870 CBC 204530 : if (in_recovery)
1871 ECB : {
1872 CBC 116 : h->oldest_considered_running =
1873 116 : TransactionIdOlder(h->oldest_considered_running, kaxmin);
1874 GIC 116 : h->shared_oldest_nonremovable =
1875 116 : TransactionIdOlder(h->shared_oldest_nonremovable, kaxmin);
1876 116 : h->data_oldest_nonremovable =
1877 116 : TransactionIdOlder(h->data_oldest_nonremovable, kaxmin);
1878 : /* temp relations cannot be accessed in recovery */
1879 : }
1880 : else
1881 : {
1882 : /*
1883 : * Compute the cutoff XID by subtracting vacuum_defer_cleanup_age.
1884 : *
1885 : * vacuum_defer_cleanup_age provides some additional "slop" for the
1886 : * benefit of hot standby queries on standby servers. This is quick
1887 : * and dirty, and perhaps not all that useful unless the primary has a
1888 : * predictable transaction rate, but it offers some protection when
1889 : * there's no walsender connection. Note that we are assuming
1890 : * vacuum_defer_cleanup_age isn't large enough to cause wraparound ---
1891 : * so guc.c should limit it to no more than the xidStopLimit threshold
1892 : * in varsup.c. Also note that we intentionally don't apply
1893 : * vacuum_defer_cleanup_age on standby servers.
1894 : *
1895 ECB : * Need to use TransactionIdRetreatSafely() instead of open-coding the
1896 : * subtraction, to prevent creating an xid before
1897 : * FirstNormalTransactionId.
1898 : */
1899 GIC 204414 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1900 ECB : h->shared_oldest_nonremovable));
1901 GIC 204414 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1902 EUB : h->data_oldest_nonremovable));
1903 :
1904 GIC 204414 : if (vacuum_defer_cleanup_age > 0)
1905 EUB : {
1906 UIC 0 : TransactionIdRetreatSafely(&h->oldest_considered_running,
1907 : vacuum_defer_cleanup_age,
1908 EUB : h->latest_completed);
1909 UIC 0 : TransactionIdRetreatSafely(&h->shared_oldest_nonremovable,
1910 : vacuum_defer_cleanup_age,
1911 : h->latest_completed);
1912 0 : TransactionIdRetreatSafely(&h->data_oldest_nonremovable,
1913 : vacuum_defer_cleanup_age,
1914 EUB : h->latest_completed);
1915 : /* defer doesn't apply to temp relations */
1916 :
1917 :
1918 UIC 0 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1919 : h->shared_oldest_nonremovable));
1920 0 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1921 : h->data_oldest_nonremovable));
1922 : }
1923 : }
1924 ECB :
1925 : /*
1926 : * Check whether there are replication slots requiring an older xmin.
1927 : */
1928 GIC 204530 : h->shared_oldest_nonremovable =
1929 204530 : TransactionIdOlder(h->shared_oldest_nonremovable, h->slot_xmin);
1930 204530 : h->data_oldest_nonremovable =
1931 204530 : TransactionIdOlder(h->data_oldest_nonremovable, h->slot_xmin);
1932 :
1933 : /*
1934 : * The only difference between catalog / data horizons is that the slot's
1935 : * catalog xmin is applied to the catalog one (so catalogs can be accessed
1936 ECB : * for logical decoding). Initialize with data horizon, and then back up
1937 : * further if necessary. Have to back up the shared horizon as well, since
1938 : * that also can contain catalogs.
1939 : */
1940 CBC 204530 : h->shared_oldest_nonremovable_raw = h->shared_oldest_nonremovable;
1941 204530 : h->shared_oldest_nonremovable =
1942 204530 : TransactionIdOlder(h->shared_oldest_nonremovable,
1943 : h->slot_catalog_xmin);
1944 GIC 204530 : h->catalog_oldest_nonremovable = h->data_oldest_nonremovable;
1945 204530 : h->catalog_oldest_nonremovable =
1946 204530 : TransactionIdOlder(h->catalog_oldest_nonremovable,
1947 : h->slot_catalog_xmin);
1948 :
1949 ECB : /*
1950 : * It's possible that slots / vacuum_defer_cleanup_age backed up the
1951 : * horizons further than oldest_considered_running. Fix.
1952 : */
1953 CBC 204530 : h->oldest_considered_running =
1954 GIC 204530 : TransactionIdOlder(h->oldest_considered_running,
1955 ECB : h->shared_oldest_nonremovable);
1956 CBC 204530 : h->oldest_considered_running =
1957 GIC 204530 : TransactionIdOlder(h->oldest_considered_running,
1958 : h->catalog_oldest_nonremovable);
1959 204530 : h->oldest_considered_running =
1960 204530 : TransactionIdOlder(h->oldest_considered_running,
1961 : h->data_oldest_nonremovable);
1962 :
1963 ECB : /*
1964 : * shared horizons have to be at least as old as the oldest visible in
1965 : * current db
1966 : */
1967 GIC 204530 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1968 : h->data_oldest_nonremovable));
1969 204530 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1970 : h->catalog_oldest_nonremovable));
1971 :
1972 ECB : /*
1973 : * Horizons need to ensure that pg_subtrans access is still possible for
1974 : * the relevant backends.
1975 : */
1976 CBC 204530 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1977 : h->shared_oldest_nonremovable));
1978 204530 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1979 : h->catalog_oldest_nonremovable));
1980 204530 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1981 : h->data_oldest_nonremovable));
1982 GIC 204530 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1983 ECB : h->temp_oldest_nonremovable));
1984 GIC 204530 : Assert(!TransactionIdIsValid(h->slot_xmin) ||
1985 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1986 : h->slot_xmin));
1987 204530 : Assert(!TransactionIdIsValid(h->slot_catalog_xmin) ||
1988 ECB : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1989 : h->slot_catalog_xmin));
1990 :
1991 : /* update approximate horizons with the computed horizons */
1992 GIC 204530 : GlobalVisUpdateApply(h);
1993 204530 : }
1994 :
1995 : /*
1996 ECB : * Determine what kind of visibility horizon needs to be used for a
1997 : * relation. If rel is NULL, the most conservative horizon is used.
1998 : */
1999 : static inline GlobalVisHorizonKind
2000 GIC 14117213 : GlobalVisHorizonKindForRel(Relation rel)
2001 : {
2002 ECB : /*
2003 : * Other relkkinds currently don't contain xids, nor always the necessary
2004 : * logical decoding markers.
2005 : */
2006 GIC 14117213 : Assert(!rel ||
2007 ECB : rel->rd_rel->relkind == RELKIND_RELATION ||
2008 : rel->rd_rel->relkind == RELKIND_MATVIEW ||
2009 : rel->rd_rel->relkind == RELKIND_TOASTVALUE);
2010 :
2011 CBC 14117213 : if (rel == NULL || rel->rd_rel->relisshared || RecoveryInProgress())
2012 68019 : return VISHORIZON_SHARED;
2013 14049194 : else if (IsCatalogRelation(rel) ||
2014 GIC 11358891 : RelationIsAccessibleInLogicalDecoding(rel))
2015 CBC 2690307 : return VISHORIZON_CATALOG;
2016 GIC 11358887 : else if (!RELATION_IS_LOCAL(rel))
2017 11305763 : return VISHORIZON_DATA;
2018 : else
2019 53124 : return VISHORIZON_TEMP;
2020 : }
2021 :
2022 : /*
2023 : * Return the oldest XID for which deleted tuples must be preserved in the
2024 : * passed table.
2025 : *
2026 : * If rel is not NULL the horizon may be considerably more recent than
2027 : * otherwise (i.e. fewer tuples will be removable). In the NULL case a horizon
2028 : * that is correct (but not optimal) for all relations will be returned.
2029 : *
2030 ECB : * This is used by VACUUM to decide which deleted tuples must be preserved in
2031 : * the passed in table.
2032 : */
2033 : TransactionId
2034 CBC 80228 : GetOldestNonRemovableTransactionId(Relation rel)
2035 : {
2036 ECB : ComputeXidHorizonsResult horizons;
2037 :
2038 CBC 80228 : ComputeXidHorizons(&horizons);
2039 ECB :
2040 CBC 80228 : switch (GlobalVisHorizonKindForRel(rel))
2041 ECB : {
2042 CBC 12107 : case VISHORIZON_SHARED:
2043 12107 : return horizons.shared_oldest_nonremovable;
2044 43630 : case VISHORIZON_CATALOG:
2045 43630 : return horizons.catalog_oldest_nonremovable;
2046 GIC 13134 : case VISHORIZON_DATA:
2047 13134 : return horizons.data_oldest_nonremovable;
2048 11357 : case VISHORIZON_TEMP:
2049 GBC 11357 : return horizons.temp_oldest_nonremovable;
2050 : }
2051 :
2052 : /* just to prevent compiler warnings */
2053 UIC 0 : return InvalidTransactionId;
2054 : }
2055 :
2056 : /*
2057 : * Return the oldest transaction id any currently running backend might still
2058 : * consider running. This should not be used for visibility / pruning
2059 ECB : * determinations (see GetOldestNonRemovableTransactionId()), but for
2060 : * decisions like up to where pg_subtrans can be truncated.
2061 : */
2062 : TransactionId
2063 CBC 2336 : GetOldestTransactionIdConsideredRunning(void)
2064 : {
2065 ECB : ComputeXidHorizonsResult horizons;
2066 :
2067 GIC 2336 : ComputeXidHorizons(&horizons);
2068 :
2069 2336 : return horizons.oldest_considered_running;
2070 : }
2071 :
2072 ECB : /*
2073 : * Return the visibility horizons for a hot standby feedback message.
2074 : */
2075 : void
2076 CBC 21 : GetReplicationHorizons(TransactionId *xmin, TransactionId *catalog_xmin)
2077 : {
2078 : ComputeXidHorizonsResult horizons;
2079 :
2080 GIC 21 : ComputeXidHorizons(&horizons);
2081 :
2082 : /*
2083 : * Don't want to use shared_oldest_nonremovable here, as that contains the
2084 ECB : * effect of replication slot's catalog_xmin. We want to send a separate
2085 : * feedback for the catalog horizon, so the primary can remove data table
2086 : * contents more aggressively.
2087 : */
2088 GIC 21 : *xmin = horizons.shared_oldest_nonremovable_raw;
2089 21 : *catalog_xmin = horizons.slot_catalog_xmin;
2090 21 : }
2091 :
2092 : /*
2093 : * GetMaxSnapshotXidCount -- get max size for snapshot XID array
2094 ECB : *
2095 : * We have to export this for use by snapmgr.c.
2096 : */
2097 : int
2098 GIC 25076 : GetMaxSnapshotXidCount(void)
2099 : {
2100 25076 : return procArray->maxProcs;
2101 : }
2102 :
2103 : /*
2104 : * GetMaxSnapshotSubxidCount -- get max size for snapshot sub-XID array
2105 ECB : *
2106 : * We have to export this for use by snapmgr.c.
2107 : */
2108 : int
2109 GIC 24930 : GetMaxSnapshotSubxidCount(void)
2110 : {
2111 24930 : return TOTAL_MAX_CACHED_SUBXIDS;
2112 : }
2113 :
2114 ECB : /*
2115 : * Initialize old_snapshot_threshold specific parts of a newly build snapshot.
2116 : */
2117 : static void
2118 GIC 2597176 : GetSnapshotDataInitOldSnapshot(Snapshot snapshot)
2119 : {
2120 2597176 : if (!OldSnapshotThresholdActive())
2121 : {
2122 ECB : /*
2123 : * If not using "snapshot too old" feature, fill related fields with
2124 : * dummy values that don't require any locking.
2125 : */
2126 GIC 2593616 : snapshot->lsn = InvalidXLogRecPtr;
2127 2593616 : snapshot->whenTaken = 0;
2128 : }
2129 : else
2130 : {
2131 : /*
2132 ECB : * Capture the current time and WAL stream location in case this
2133 : * snapshot becomes old enough to need to fall back on the special
2134 : * "old snapshot" logic.
2135 : */
2136 CBC 3560 : snapshot->lsn = GetXLogInsertRecPtr();
2137 GIC 3560 : snapshot->whenTaken = GetSnapshotCurrentTimestamp();
2138 3560 : MaintainOldSnapshotTimeMapping(snapshot->whenTaken, snapshot->xmin);
2139 : }
2140 2597176 : }
2141 :
2142 : /*
2143 : * Helper function for GetSnapshotData() that checks if the bulk of the
2144 : * visibility information in the snapshot is still valid. If so, it updates
2145 : * the fields that need to change and returns true. Otherwise it returns
2146 : * false.
2147 : *
2148 ECB : * This very likely can be evolved to not need ProcArrayLock held (at very
2149 : * least in the case we already hold a snapshot), but that's for another day.
2150 : */
2151 : static bool
2152 CBC 2597176 : GetSnapshotDataReuse(Snapshot snapshot)
2153 : {
2154 ECB : uint64 curXactCompletionCount;
2155 :
2156 GIC 2597176 : Assert(LWLockHeldByMe(ProcArrayLock));
2157 ECB :
2158 CBC 2597176 : if (unlikely(snapshot->snapXactCompletionCount == 0))
2159 23627 : return false;
2160 :
2161 GIC 2573549 : curXactCompletionCount = ShmemVariableCache->xactCompletionCount;
2162 2573549 : if (curXactCompletionCount != snapshot->snapXactCompletionCount)
2163 672660 : return false;
2164 :
2165 : /*
2166 : * If the current xactCompletionCount is still the same as it was at the
2167 : * time the snapshot was built, we can be sure that rebuilding the
2168 : * contents of the snapshot the hard way would result in the same snapshot
2169 : * contents:
2170 : *
2171 : * As explained in transam/README, the set of xids considered running by
2172 : * GetSnapshotData() cannot change while ProcArrayLock is held. Snapshot
2173 : * contents only depend on transactions with xids and xactCompletionCount
2174 : * is incremented whenever a transaction with an xid finishes (while
2175 : * holding ProcArrayLock) exclusively). Thus the xactCompletionCount check
2176 : * ensures we would detect if the snapshot would have changed.
2177 : *
2178 : * As the snapshot contents are the same as it was before, it is safe to
2179 : * re-enter the snapshot's xmin into the PGPROC array. None of the rows
2180 : * visible under the snapshot could already have been removed (that'd
2181 ECB : * require the set of running transactions to change) and it fulfills the
2182 : * requirement that concurrent GetSnapshotData() calls yield the same
2183 : * xmin.
2184 : */
2185 CBC 1900889 : if (!TransactionIdIsValid(MyProc->xmin))
2186 GIC 581099 : MyProc->xmin = TransactionXmin = snapshot->xmin;
2187 ECB :
2188 CBC 1900889 : RecentXmin = snapshot->xmin;
2189 1900889 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2190 ECB :
2191 GIC 1900889 : snapshot->curcid = GetCurrentCommandId(false);
2192 CBC 1900889 : snapshot->active_count = 0;
2193 GIC 1900889 : snapshot->regd_count = 0;
2194 CBC 1900889 : snapshot->copied = false;
2195 :
2196 GIC 1900889 : GetSnapshotDataInitOldSnapshot(snapshot);
2197 :
2198 1900889 : return true;
2199 : }
2200 :
2201 : /*
2202 : * GetSnapshotData -- returns information about running transactions.
2203 : *
2204 : * The returned snapshot includes xmin (lowest still-running xact ID),
2205 : * xmax (highest completed xact ID + 1), and a list of running xact IDs
2206 : * in the range xmin <= xid < xmax. It is used as follows:
2207 : * All xact IDs < xmin are considered finished.
2208 : * All xact IDs >= xmax are considered still running.
2209 : * For an xact ID xmin <= xid < xmax, consult list to see whether
2210 : * it is considered running or not.
2211 : * This ensures that the set of transactions seen as "running" by the
2212 : * current xact will not change after it takes the snapshot.
2213 : *
2214 : * All running top-level XIDs are included in the snapshot, except for lazy
2215 : * VACUUM processes. We also try to include running subtransaction XIDs,
2216 : * but since PGPROC has only a limited cache area for subxact XIDs, full
2217 : * information may not be available. If we find any overflowed subxid arrays,
2218 : * we have to mark the snapshot's subxid data as overflowed, and extra work
2219 : * *may* need to be done to determine what's running (see XidInMVCCSnapshot()
2220 : * in heapam_visibility.c).
2221 : *
2222 : * We also update the following backend-global variables:
2223 : * TransactionXmin: the oldest xmin of any snapshot in use in the
2224 : * current transaction (this is the same as MyProc->xmin).
2225 : * RecentXmin: the xmin computed for the most recent snapshot. XIDs
2226 : * older than this are known not running any more.
2227 : *
2228 : * And try to advance the bounds of GlobalVis{Shared,Catalog,Data,Temp}Rels
2229 : * for the benefit of the GlobalVisTest* family of functions.
2230 : *
2231 ECB : * Note: this function should probably not be called with an argument that's
2232 : * not statically allocated (see xip allocation below).
2233 : */
2234 : Snapshot
2235 GIC 2597176 : GetSnapshotData(Snapshot snapshot)
2236 : {
2237 CBC 2597176 : ProcArrayStruct *arrayP = procArray;
2238 2597176 : TransactionId *other_xids = ProcGlobal->xids;
2239 ECB : TransactionId xmin;
2240 : TransactionId xmax;
2241 GIC 2597176 : int count = 0;
2242 2597176 : int subcount = 0;
2243 2597176 : bool suboverflowed = false;
2244 : FullTransactionId latest_completed;
2245 : TransactionId oldestxid;
2246 ECB : int mypgxactoff;
2247 : TransactionId myxid;
2248 : uint64 curXactCompletionCount;
2249 :
2250 GIC 2597176 : TransactionId replication_slot_xmin = InvalidTransactionId;
2251 2597176 : TransactionId replication_slot_catalog_xmin = InvalidTransactionId;
2252 :
2253 2597176 : Assert(snapshot != NULL);
2254 :
2255 : /*
2256 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2257 : * be sufficient. But it seems better to do the malloc while not holding
2258 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2259 : * more subxip storage than is probably needed.
2260 : *
2261 : * This does open a possibility for avoiding repeated malloc/free: since
2262 ECB : * maxProcs does not change at runtime, we can simply reuse the previous
2263 : * xip arrays if any. (This relies on the fact that all callers pass
2264 : * static SnapshotData structs.)
2265 : */
2266 GIC 2597176 : if (snapshot->xip == NULL)
2267 : {
2268 ECB : /*
2269 : * First call for this snapshot. Snapshot is same size whether or not
2270 : * we are in recovery, see later comments.
2271 EUB : */
2272 GIC 23432 : snapshot->xip = (TransactionId *)
2273 23432 : malloc(GetMaxSnapshotXidCount() * sizeof(TransactionId));
2274 CBC 23432 : if (snapshot->xip == NULL)
2275 LBC 0 : ereport(ERROR,
2276 ECB : (errcode(ERRCODE_OUT_OF_MEMORY),
2277 : errmsg("out of memory")));
2278 GBC 23432 : Assert(snapshot->subxip == NULL);
2279 GIC 23432 : snapshot->subxip = (TransactionId *)
2280 23432 : malloc(GetMaxSnapshotSubxidCount() * sizeof(TransactionId));
2281 23432 : if (snapshot->subxip == NULL)
2282 UIC 0 : ereport(ERROR,
2283 : (errcode(ERRCODE_OUT_OF_MEMORY),
2284 : errmsg("out of memory")));
2285 : }
2286 :
2287 ECB : /*
2288 : * It is sufficient to get shared lock on ProcArrayLock, even if we are
2289 : * going to set MyProc->xmin.
2290 : */
2291 CBC 2597176 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2292 ECB :
2293 GIC 2597176 : if (GetSnapshotDataReuse(snapshot))
2294 : {
2295 CBC 1900889 : LWLockRelease(ProcArrayLock);
2296 1900889 : return snapshot;
2297 ECB : }
2298 :
2299 GIC 696287 : latest_completed = ShmemVariableCache->latestCompletedXid;
2300 CBC 696287 : mypgxactoff = MyProc->pgxactoff;
2301 696287 : myxid = other_xids[mypgxactoff];
2302 GIC 696287 : Assert(myxid == MyProc->xid);
2303 :
2304 CBC 696287 : oldestxid = ShmemVariableCache->oldestXid;
2305 696287 : curXactCompletionCount = ShmemVariableCache->xactCompletionCount;
2306 ECB :
2307 : /* xmax is always latestCompletedXid + 1 */
2308 GIC 696287 : xmax = XidFromFullTransactionId(latest_completed);
2309 CBC 696287 : TransactionIdAdvance(xmax);
2310 GIC 696287 : Assert(TransactionIdIsNormal(xmax));
2311 :
2312 ECB : /* initialize xmin calculation with xmax */
2313 CBC 696287 : xmin = xmax;
2314 :
2315 ECB : /* take own xid into account, saves a check inside the loop */
2316 GIC 696287 : if (TransactionIdIsNormal(myxid) && NormalTransactionIdPrecedes(myxid, xmin))
2317 CBC 22367 : xmin = myxid;
2318 :
2319 696287 : snapshot->takenDuringRecovery = RecoveryInProgress();
2320 ECB :
2321 CBC 696287 : if (!snapshot->takenDuringRecovery)
2322 ECB : {
2323 CBC 695461 : int numProcs = arrayP->numProcs;
2324 GIC 695461 : TransactionId *xip = snapshot->xip;
2325 695461 : int *pgprocnos = arrayP->pgprocnos;
2326 695461 : XidCacheStatus *subxidStates = ProcGlobal->subxidStates;
2327 695461 : uint8 *allStatusFlags = ProcGlobal->statusFlags;
2328 :
2329 ECB : /*
2330 : * First collect set of pgxactoff/xids that need to be included in the
2331 : * snapshot.
2332 : */
2333 GIC 3559582 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
2334 : {
2335 ECB : /* Fetch xid just once - see GetNewTransactionId */
2336 GIC 2864121 : TransactionId xid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
2337 : uint8 statusFlags;
2338 :
2339 2864121 : Assert(allProcs[arrayP->pgprocnos[pgxactoff]].pgxactoff == pgxactoff);
2340 :
2341 ECB : /*
2342 : * If the transaction has no XID assigned, we can skip it; it
2343 : * won't have sub-XIDs either.
2344 : */
2345 GIC 2864121 : if (likely(xid == InvalidTransactionId))
2346 2324866 : continue;
2347 :
2348 : /*
2349 ECB : * We don't include our own XIDs (if any) in the snapshot. It
2350 : * needs to be included in the xmin computation, but we did so
2351 : * outside the loop.
2352 : */
2353 GIC 539255 : if (pgxactoff == mypgxactoff)
2354 70899 : continue;
2355 :
2356 : /*
2357 : * The only way we are able to get here with a non-normal xid is
2358 ECB : * during bootstrap - with this backend using
2359 : * BootstrapTransactionId. But the above test should filter that
2360 : * out.
2361 : */
2362 GIC 468356 : Assert(TransactionIdIsNormal(xid));
2363 :
2364 : /*
2365 ECB : * If the XID is >= xmax, we can skip it; such transactions will
2366 : * be treated as running anyway (and any sub-XIDs will also be >=
2367 : * xmax).
2368 : */
2369 GIC 468356 : if (!NormalTransactionIdPrecedes(xid, xmax))
2370 93952 : continue;
2371 :
2372 ECB : /*
2373 : * Skip over backends doing logical decoding which manages xmin
2374 EUB : * separately (check below) and ones running LAZY VACUUM.
2375 : */
2376 CBC 374404 : statusFlags = allStatusFlags[pgxactoff];
2377 374404 : if (statusFlags & (PROC_IN_LOGICAL_DECODING | PROC_IN_VACUUM))
2378 UIC 0 : continue;
2379 :
2380 CBC 374404 : if (NormalTransactionIdPrecedes(xid, xmin))
2381 GIC 216005 : xmin = xid;
2382 :
2383 : /* Add XID to snapshot. */
2384 374404 : xip[count++] = xid;
2385 :
2386 : /*
2387 : * Save subtransaction XIDs if possible (if we've already
2388 : * overflowed, there's no point). Note that the subxact XIDs must
2389 : * be later than their parent, so no need to check them against
2390 : * xmin. We could filter against xmax, but it seems better not to
2391 : * do that much work while holding the ProcArrayLock.
2392 : *
2393 : * The other backend can add more subxids concurrently, but cannot
2394 : * remove any. Hence it's important to fetch nxids just once.
2395 : * Should be safe to use memcpy, though. (We needn't worry about
2396 : * missing any xids added concurrently, because they must postdate
2397 ECB : * xmax.)
2398 : *
2399 : * Again, our own XIDs are not included in the snapshot.
2400 : */
2401 CBC 374404 : if (!suboverflowed)
2402 : {
2403 :
2404 374404 : if (subxidStates[pgxactoff].overflowed)
2405 GIC 28 : suboverflowed = true;
2406 ECB : else
2407 : {
2408 CBC 374376 : int nsubxids = subxidStates[pgxactoff].count;
2409 ECB :
2410 GIC 374376 : if (nsubxids > 0)
2411 ECB : {
2412 GIC 4762 : int pgprocno = pgprocnos[pgxactoff];
2413 CBC 4762 : PGPROC *proc = &allProcs[pgprocno];
2414 ECB :
2415 GIC 4762 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2416 ECB :
2417 GIC 4762 : memcpy(snapshot->subxip + subcount,
2418 GNC 4762 : proc->subxids.xids,
2419 : nsubxids * sizeof(TransactionId));
2420 GIC 4762 : subcount += nsubxids;
2421 : }
2422 : }
2423 : }
2424 : }
2425 : }
2426 : else
2427 : {
2428 : /*
2429 : * We're in hot standby, so get XIDs from KnownAssignedXids.
2430 : *
2431 : * We store all xids directly into subxip[]. Here's why:
2432 : *
2433 : * In recovery we don't know which xids are top-level and which are
2434 : * subxacts, a design choice that greatly simplifies xid processing.
2435 : *
2436 : * It seems like we would want to try to put xids into xip[] only, but
2437 : * that is fairly small. We would either need to make that bigger or
2438 : * to increase the rate at which we WAL-log xid assignment; neither is
2439 : * an appealing choice.
2440 : *
2441 : * We could try to store xids into xip[] first and then into subxip[]
2442 : * if there are too many xids. That only works if the snapshot doesn't
2443 : * overflow because we do not search subxip[] in that case. A simpler
2444 : * way is to just store all xids in the subxip array because this is
2445 : * by far the bigger array. We just leave the xip array empty.
2446 : *
2447 : * Either way we need to change the way XidInMVCCSnapshot() works
2448 : * depending upon when the snapshot was taken, or change normal
2449 : * snapshot processing so it matches.
2450 : *
2451 : * Note: It is possible for recovery to end before we finish taking
2452 : * the snapshot, and for newly assigned transaction ids to be added to
2453 ECB : * the ProcArray. xmax cannot change while we hold ProcArrayLock, so
2454 : * those newly added transaction ids would be filtered away, so we
2455 : * need not be concerned about them.
2456 : */
2457 CBC 826 : subcount = KnownAssignedXidsGetAndSetXmin(snapshot->subxip, &xmin,
2458 : xmax);
2459 :
2460 GIC 826 : if (TransactionIdPrecedesOrEquals(xmin, procArray->lastOverflowedXid))
2461 4 : suboverflowed = true;
2462 : }
2463 :
2464 :
2465 : /*
2466 ECB : * Fetch into local variable while ProcArrayLock is held - the
2467 : * LWLockRelease below is a barrier, ensuring this happens inside the
2468 : * lock.
2469 : */
2470 CBC 696287 : replication_slot_xmin = procArray->replication_slot_xmin;
2471 GIC 696287 : replication_slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
2472 ECB :
2473 GIC 696287 : if (!TransactionIdIsValid(MyProc->xmin))
2474 349709 : MyProc->xmin = TransactionXmin = xmin;
2475 :
2476 696287 : LWLockRelease(ProcArrayLock);
2477 :
2478 : /* maintain state for GlobalVis* */
2479 : {
2480 : TransactionId def_vis_xid;
2481 : TransactionId def_vis_xid_data;
2482 : FullTransactionId def_vis_fxid;
2483 : FullTransactionId def_vis_fxid_data;
2484 : FullTransactionId oldestfxid;
2485 :
2486 : /*
2487 ECB : * Converting oldestXid is only safe when xid horizon cannot advance,
2488 : * i.e. holding locks. While we don't hold the lock anymore, all the
2489 : * necessary data has been gathered with lock held.
2490 : */
2491 CBC 696287 : oldestfxid = FullXidRelativeTo(latest_completed, oldestxid);
2492 :
2493 : /* apply vacuum_defer_cleanup_age */
2494 GIC 696287 : def_vis_xid_data = xmin;
2495 696287 : TransactionIdRetreatSafely(&def_vis_xid_data,
2496 ECB : vacuum_defer_cleanup_age,
2497 : oldestfxid);
2498 :
2499 : /* Check whether there's a replication slot requiring an older xmin. */
2500 GIC 696287 : def_vis_xid_data =
2501 696287 : TransactionIdOlder(def_vis_xid_data, replication_slot_xmin);
2502 :
2503 ECB : /*
2504 : * Rows in non-shared, non-catalog tables possibly could be vacuumed
2505 : * if older than this xid.
2506 : */
2507 GIC 696287 : def_vis_xid = def_vis_xid_data;
2508 :
2509 : /*
2510 ECB : * Check whether there's a replication slot requiring an older catalog
2511 : * xmin.
2512 : */
2513 : def_vis_xid =
2514 GIC 696287 : TransactionIdOlder(replication_slot_catalog_xmin, def_vis_xid);
2515 :
2516 696287 : def_vis_fxid = FullXidRelativeTo(latest_completed, def_vis_xid);
2517 696287 : def_vis_fxid_data = FullXidRelativeTo(latest_completed, def_vis_xid_data);
2518 :
2519 : /*
2520 : * Check if we can increase upper bound. As a previous
2521 ECB : * GlobalVisUpdate() might have computed more aggressive values, don't
2522 : * overwrite them if so.
2523 : */
2524 : GlobalVisSharedRels.definitely_needed =
2525 GIC 696287 : FullTransactionIdNewer(def_vis_fxid,
2526 : GlobalVisSharedRels.definitely_needed);
2527 ECB : GlobalVisCatalogRels.definitely_needed =
2528 GIC 696287 : FullTransactionIdNewer(def_vis_fxid,
2529 : GlobalVisCatalogRels.definitely_needed);
2530 ECB : GlobalVisDataRels.definitely_needed =
2531 GIC 696287 : FullTransactionIdNewer(def_vis_fxid_data,
2532 ECB : GlobalVisDataRels.definitely_needed);
2533 : /* See temp_oldest_nonremovable computation in ComputeXidHorizons() */
2534 GIC 696287 : if (TransactionIdIsNormal(myxid))
2535 ECB : GlobalVisTempRels.definitely_needed =
2536 CBC 70289 : FullXidRelativeTo(latest_completed, myxid);
2537 : else
2538 : {
2539 GIC 625998 : GlobalVisTempRels.definitely_needed = latest_completed;
2540 625998 : FullTransactionIdAdvance(&GlobalVisTempRels.definitely_needed);
2541 : }
2542 :
2543 : /*
2544 : * Check if we know that we can initialize or increase the lower
2545 : * bound. Currently the only cheap way to do so is to use
2546 : * ShmemVariableCache->oldestXid as input.
2547 : *
2548 ECB : * We should definitely be able to do better. We could e.g. put a
2549 : * global lower bound value into ShmemVariableCache.
2550 : */
2551 : GlobalVisSharedRels.maybe_needed =
2552 GIC 696287 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
2553 : oldestfxid);
2554 ECB : GlobalVisCatalogRels.maybe_needed =
2555 GIC 696287 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
2556 : oldestfxid);
2557 ECB : GlobalVisDataRels.maybe_needed =
2558 GIC 696287 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
2559 : oldestfxid);
2560 ECB : /* accurate value known */
2561 CBC 696287 : GlobalVisTempRels.maybe_needed = GlobalVisTempRels.definitely_needed;
2562 : }
2563 ECB :
2564 CBC 696287 : RecentXmin = xmin;
2565 696287 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2566 ECB :
2567 CBC 696287 : snapshot->xmin = xmin;
2568 696287 : snapshot->xmax = xmax;
2569 GIC 696287 : snapshot->xcnt = count;
2570 CBC 696287 : snapshot->subxcnt = subcount;
2571 GIC 696287 : snapshot->suboverflowed = suboverflowed;
2572 696287 : snapshot->snapXactCompletionCount = curXactCompletionCount;
2573 :
2574 696287 : snapshot->curcid = GetCurrentCommandId(false);
2575 :
2576 ECB : /*
2577 : * This is a new snapshot, so set both refcounts are zero, and mark it as
2578 : * not copied in persistent memory.
2579 : */
2580 CBC 696287 : snapshot->active_count = 0;
2581 GIC 696287 : snapshot->regd_count = 0;
2582 CBC 696287 : snapshot->copied = false;
2583 :
2584 GIC 696287 : GetSnapshotDataInitOldSnapshot(snapshot);
2585 :
2586 696287 : return snapshot;
2587 : }
2588 :
2589 : /*
2590 : * ProcArrayInstallImportedXmin -- install imported xmin into MyProc->xmin
2591 : *
2592 : * This is called when installing a snapshot imported from another
2593 : * transaction. To ensure that OldestXmin doesn't go backwards, we must
2594 : * check that the source transaction is still running, and we'd better do
2595 : * that atomically with installing the new xmin.
2596 ECB : *
2597 : * Returns true if successful, false if source xact is no longer running.
2598 : */
2599 : bool
2600 CBC 18 : ProcArrayInstallImportedXmin(TransactionId xmin,
2601 : VirtualTransactionId *sourcevxid)
2602 : {
2603 18 : bool result = false;
2604 18 : ProcArrayStruct *arrayP = procArray;
2605 EUB : int index;
2606 :
2607 GIC 18 : Assert(TransactionIdIsNormal(xmin));
2608 CBC 18 : if (!sourcevxid)
2609 UIC 0 : return false;
2610 ECB :
2611 : /* Get lock so source xact can't end while we're doing this */
2612 CBC 18 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2613 ECB :
2614 CBC 53 : for (index = 0; index < arrayP->numProcs; index++)
2615 : {
2616 GIC 53 : int pgprocno = arrayP->pgprocnos[index];
2617 53 : PGPROC *proc = &allProcs[pgprocno];
2618 CBC 53 : int statusFlags = ProcGlobal->statusFlags[index];
2619 EUB : TransactionId xid;
2620 :
2621 : /* Ignore procs running LAZY VACUUM */
2622 CBC 53 : if (statusFlags & PROC_IN_VACUUM)
2623 LBC 0 : continue;
2624 ECB :
2625 EUB : /* We are only interested in the specific virtual transaction. */
2626 GIC 53 : if (proc->backendId != sourcevxid->backendId)
2627 35 : continue;
2628 18 : if (proc->lxid != sourcevxid->localTransactionId)
2629 UIC 0 : continue;
2630 :
2631 : /*
2632 : * We check the transaction's database ID for paranoia's sake: if it's
2633 ECB : * in another DB then its xmin does not cover us. Caller should have
2634 EUB : * detected this already, so we just treat any funny cases as
2635 : * "transaction not found".
2636 : */
2637 GIC 18 : if (proc->databaseId != MyDatabaseId)
2638 UIC 0 : continue;
2639 ECB :
2640 : /*
2641 : * Likewise, let's just make real sure its xmin does cover us.
2642 EUB : */
2643 GIC 18 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2644 18 : if (!TransactionIdIsNormal(xid) ||
2645 18 : !TransactionIdPrecedesOrEquals(xid, xmin))
2646 UIC 0 : continue;
2647 :
2648 : /*
2649 : * We're good. Install the new xmin. As in GetSnapshotData, set
2650 ECB : * TransactionXmin too. (Note that because snapmgr.c called
2651 : * GetSnapshotData first, we'll be overwriting a valid xmin here, so
2652 : * we don't check that.)
2653 : */
2654 GIC 18 : MyProc->xmin = TransactionXmin = xmin;
2655 :
2656 CBC 18 : result = true;
2657 GIC 18 : break;
2658 ECB : }
2659 :
2660 GIC 18 : LWLockRelease(ProcArrayLock);
2661 :
2662 18 : return result;
2663 : }
2664 :
2665 : /*
2666 : * ProcArrayInstallRestoredXmin -- install restored xmin into MyProc->xmin
2667 : *
2668 : * This is like ProcArrayInstallImportedXmin, but we have a pointer to the
2669 : * PGPROC of the transaction from which we imported the snapshot, rather than
2670 : * an XID.
2671 : *
2672 : * Note that this function also copies statusFlags from the source `proc` in
2673 : * order to avoid the case where MyProc's xmin needs to be skipped for
2674 : * computing xid horizon.
2675 ECB : *
2676 : * Returns true if successful, false if source xact is no longer running.
2677 : */
2678 : bool
2679 GIC 1453 : ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc)
2680 ECB : {
2681 CBC 1453 : bool result = false;
2682 : TransactionId xid;
2683 :
2684 GIC 1453 : Assert(TransactionIdIsNormal(xmin));
2685 1453 : Assert(proc != NULL);
2686 ECB :
2687 : /*
2688 : * Get an exclusive lock so that we can copy statusFlags from source proc.
2689 : */
2690 GIC 1453 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2691 :
2692 : /*
2693 : * Be certain that the referenced PGPROC has an advertised xmin which is
2694 ECB : * no later than the one we're installing, so that the system-wide xmin
2695 : * can't go backwards. Also, make sure it's running in the same database,
2696 : * so that the per-database xmin cannot go backwards.
2697 : */
2698 GIC 1453 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2699 1453 : if (proc->databaseId == MyDatabaseId &&
2700 1453 : TransactionIdIsNormal(xid) &&
2701 1453 : TransactionIdPrecedesOrEquals(xid, xmin))
2702 : {
2703 ECB : /*
2704 : * Install xmin and propagate the statusFlags that affect how the
2705 : * value is interpreted by vacuum.
2706 : */
2707 GIC 1453 : MyProc->xmin = TransactionXmin = xmin;
2708 CBC 1453 : MyProc->statusFlags = (MyProc->statusFlags & ~PROC_XMIN_FLAGS) |
2709 GIC 1453 : (proc->statusFlags & PROC_XMIN_FLAGS);
2710 1453 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2711 ECB :
2712 GIC 1453 : result = true;
2713 ECB : }
2714 :
2715 GIC 1453 : LWLockRelease(ProcArrayLock);
2716 :
2717 1453 : return result;
2718 : }
2719 :
2720 : /*
2721 : * GetRunningTransactionData -- returns information about running transactions.
2722 : *
2723 : * Similar to GetSnapshotData but returns more information. We include
2724 : * all PGPROCs with an assigned TransactionId, even VACUUM processes and
2725 : * prepared transactions.
2726 : *
2727 : * We acquire XidGenLock and ProcArrayLock, but the caller is responsible for
2728 : * releasing them. Acquiring XidGenLock ensures that no new XIDs enter the proc
2729 : * array until the caller has WAL-logged this snapshot, and releases the
2730 : * lock. Acquiring ProcArrayLock ensures that no transactions commit until the
2731 : * lock is released.
2732 : *
2733 : * The returned data structure is statically allocated; caller should not
2734 : * modify it, and must not assume it is valid past the next call.
2735 : *
2736 : * This is never executed during recovery so there is no need to look at
2737 : * KnownAssignedXids.
2738 : *
2739 : * Dummy PGPROCs from prepared transaction are included, meaning that this
2740 : * may return entries with duplicated TransactionId values coming from
2741 : * transaction finishing to prepare. Nothing is done about duplicated
2742 : * entries here to not hold on ProcArrayLock more than necessary.
2743 : *
2744 : * We don't worry about updating other counters, we want to keep this as
2745 : * simple as possible and leave GetSnapshotData() as the primary code for
2746 : * that bookkeeping.
2747 : *
2748 ECB : * Note that if any transaction has overflowed its cached subtransactions
2749 : * then there is no real need include any subtransactions.
2750 : */
2751 : RunningTransactions
2752 GIC 1749 : GetRunningTransactionData(void)
2753 ECB : {
2754 : /* result workspace */
2755 : static RunningTransactionsData CurrentRunningXactsData;
2756 :
2757 GIC 1749 : ProcArrayStruct *arrayP = procArray;
2758 1749 : TransactionId *other_xids = ProcGlobal->xids;
2759 1749 : RunningTransactions CurrentRunningXacts = &CurrentRunningXactsData;
2760 : TransactionId latestCompletedXid;
2761 : TransactionId oldestRunningXid;
2762 : TransactionId *xids;
2763 : int index;
2764 ECB : int count;
2765 : int subcount;
2766 : bool suboverflowed;
2767 :
2768 GIC 1749 : Assert(!RecoveryInProgress());
2769 :
2770 : /*
2771 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2772 : * be sufficient. But it seems better to do the malloc while not holding
2773 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2774 : * more subxip storage than is probably needed.
2775 ECB : *
2776 : * Should only be allocated in bgwriter, since only ever executed during
2777 : * checkpoints.
2778 : */
2779 GIC 1749 : if (CurrentRunningXacts->xids == NULL)
2780 ECB : {
2781 : /*
2782 : * First call
2783 EUB : */
2784 GIC 667 : CurrentRunningXacts->xids = (TransactionId *)
2785 667 : malloc(TOTAL_MAX_CACHED_SUBXIDS * sizeof(TransactionId));
2786 667 : if (CurrentRunningXacts->xids == NULL)
2787 UIC 0 : ereport(ERROR,
2788 ECB : (errcode(ERRCODE_OUT_OF_MEMORY),
2789 : errmsg("out of memory")));
2790 : }
2791 :
2792 GIC 1749 : xids = CurrentRunningXacts->xids;
2793 :
2794 1749 : count = subcount = 0;
2795 1749 : suboverflowed = false;
2796 :
2797 ECB : /*
2798 : * Ensure that no xids enter or leave the procarray while we obtain
2799 : * snapshot.
2800 : */
2801 CBC 1749 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2802 1749 : LWLockAcquire(XidGenLock, LW_SHARED);
2803 ECB :
2804 GIC 1749 : latestCompletedXid =
2805 1749 : XidFromFullTransactionId(ShmemVariableCache->latestCompletedXid);
2806 1749 : oldestRunningXid =
2807 1749 : XidFromFullTransactionId(ShmemVariableCache->nextXid);
2808 ECB :
2809 : /*
2810 : * Spin over procArray collecting all xids
2811 : */
2812 GIC 5129 : for (index = 0; index < arrayP->numProcs; index++)
2813 ECB : {
2814 : TransactionId xid;
2815 :
2816 : /* Fetch xid just once - see GetNewTransactionId */
2817 GIC 3380 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2818 :
2819 ECB : /*
2820 : * We don't need to store transactions that don't have a TransactionId
2821 : * yet because they will not show as running on a standby server.
2822 : */
2823 GIC 3380 : if (!TransactionIdIsValid(xid))
2824 2075 : continue;
2825 :
2826 : /*
2827 ECB : * Be careful not to exclude any xids before calculating the values of
2828 : * oldestRunningXid and suboverflowed, since these are used to clean
2829 : * up transaction information held on standbys.
2830 : */
2831 CBC 1305 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2832 GIC 1295 : oldestRunningXid = xid;
2833 :
2834 1305 : if (ProcGlobal->subxidStates[index].overflowed)
2835 1 : suboverflowed = true;
2836 :
2837 : /*
2838 : * If we wished to exclude xids this would be the right place for it.
2839 : * Procs with the PROC_IN_VACUUM flag set don't usually assign xids,
2840 : * but they do during truncation at the end when they get the lock and
2841 ECB : * truncate, so it is not much of a problem to include them if they
2842 : * are seen and it is cleaner to include them.
2843 : */
2844 :
2845 GIC 1305 : xids[count++] = xid;
2846 : }
2847 :
2848 ECB : /*
2849 : * Spin over procArray collecting all subxids, but only if there hasn't
2850 : * been a suboverflow.
2851 : */
2852 CBC 1749 : if (!suboverflowed)
2853 : {
2854 1748 : XidCacheStatus *other_subxidstates = ProcGlobal->subxidStates;
2855 ECB :
2856 GIC 5126 : for (index = 0; index < arrayP->numProcs; index++)
2857 : {
2858 3378 : int pgprocno = arrayP->pgprocnos[index];
2859 3378 : PGPROC *proc = &allProcs[pgprocno];
2860 : int nsubxids;
2861 :
2862 ECB : /*
2863 : * Save subtransaction XIDs. Other backends can't add or remove
2864 : * entries while we're holding XidGenLock.
2865 : */
2866 CBC 3378 : nsubxids = other_subxidstates[index].count;
2867 GIC 3378 : if (nsubxids > 0)
2868 ECB : {
2869 : /* barrier not really required, as XidGenLock is held, but ... */
2870 CBC 7 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2871 ECB :
2872 GNC 7 : memcpy(&xids[count], proc->subxids.xids,
2873 : nsubxids * sizeof(TransactionId));
2874 GIC 7 : count += nsubxids;
2875 7 : subcount += nsubxids;
2876 :
2877 : /*
2878 : * Top-level XID of a transaction is always less than any of
2879 : * its subxids, so we don't need to check if any of the
2880 : * subxids are smaller than oldestRunningXid
2881 : */
2882 : }
2883 : }
2884 : }
2885 :
2886 : /*
2887 : * It's important *not* to include the limits set by slots here because
2888 : * snapbuild.c uses oldestRunningXid to manage its xmin horizon. If those
2889 : * were to be included here the initial value could never increase because
2890 : * of a circular dependency where slots only increase their limits when
2891 ECB : * running xacts increases oldestRunningXid and running xacts only
2892 : * increases if slots do.
2893 : */
2894 :
2895 CBC 1749 : CurrentRunningXacts->xcnt = count - subcount;
2896 1749 : CurrentRunningXacts->subxcnt = subcount;
2897 GIC 1749 : CurrentRunningXacts->subxid_overflow = suboverflowed;
2898 CBC 1749 : CurrentRunningXacts->nextXid = XidFromFullTransactionId(ShmemVariableCache->nextXid);
2899 1749 : CurrentRunningXacts->oldestRunningXid = oldestRunningXid;
2900 1749 : CurrentRunningXacts->latestCompletedXid = latestCompletedXid;
2901 :
2902 GIC 1749 : Assert(TransactionIdIsValid(CurrentRunningXacts->nextXid));
2903 1749 : Assert(TransactionIdIsValid(CurrentRunningXacts->oldestRunningXid));
2904 CBC 1749 : Assert(TransactionIdIsNormal(CurrentRunningXacts->latestCompletedXid));
2905 :
2906 : /* We don't release the locks here, the caller is responsible for that */
2907 :
2908 GIC 1749 : return CurrentRunningXacts;
2909 : }
2910 :
2911 : /*
2912 : * GetOldestActiveTransactionId()
2913 : *
2914 : * Similar to GetSnapshotData but returns just oldestActiveXid. We include
2915 : * all PGPROCs with an assigned TransactionId, even VACUUM processes.
2916 : * We look at all databases, though there is no need to include WALSender
2917 : * since this has no effect on hot standby conflicts.
2918 : *
2919 : * This is never executed during recovery so there is no need to look at
2920 : * KnownAssignedXids.
2921 : *
2922 : * We don't worry about updating other counters, we want to keep this as
2923 ECB : * simple as possible and leave GetSnapshotData() as the primary code for
2924 : * that bookkeeping.
2925 : */
2926 : TransactionId
2927 GIC 1353 : GetOldestActiveTransactionId(void)
2928 : {
2929 1353 : ProcArrayStruct *arrayP = procArray;
2930 CBC 1353 : TransactionId *other_xids = ProcGlobal->xids;
2931 : TransactionId oldestRunningXid;
2932 : int index;
2933 :
2934 GIC 1353 : Assert(!RecoveryInProgress());
2935 :
2936 : /*
2937 : * Read nextXid, as the upper bound of what's still active.
2938 : *
2939 ECB : * Reading a TransactionId is atomic, but we must grab the lock to make
2940 : * sure that all XIDs < nextXid are already present in the proc array (or
2941 : * have already completed), when we spin over it.
2942 : */
2943 GIC 1353 : LWLockAcquire(XidGenLock, LW_SHARED);
2944 1353 : oldestRunningXid = XidFromFullTransactionId(ShmemVariableCache->nextXid);
2945 1353 : LWLockRelease(XidGenLock);
2946 ECB :
2947 : /*
2948 : * Spin over procArray collecting all xids and subxids.
2949 : */
2950 GIC 1353 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2951 3088 : for (index = 0; index < arrayP->numProcs; index++)
2952 ECB : {
2953 : TransactionId xid;
2954 :
2955 : /* Fetch xid just once - see GetNewTransactionId */
2956 GIC 1735 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2957 ECB :
2958 CBC 1735 : if (!TransactionIdIsNormal(xid))
2959 GIC 460 : continue;
2960 :
2961 1275 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2962 1257 : oldestRunningXid = xid;
2963 :
2964 : /*
2965 : * Top-level XID of a transaction is always less than any of its
2966 ECB : * subxids, so we don't need to check if any of the subxids are
2967 : * smaller than oldestRunningXid
2968 : */
2969 : }
2970 GIC 1353 : LWLockRelease(ProcArrayLock);
2971 :
2972 1353 : return oldestRunningXid;
2973 : }
2974 :
2975 : /*
2976 : * GetOldestSafeDecodingTransactionId -- lowest xid not affected by vacuum
2977 : *
2978 : * Returns the oldest xid that we can guarantee not to have been affected by
2979 : * vacuum, i.e. no rows >= that xid have been vacuumed away unless the
2980 : * transaction aborted. Note that the value can (and most of the time will) be
2981 : * much more conservative than what really has been affected by vacuum, but we
2982 : * currently don't have better data available.
2983 : *
2984 : * This is useful to initialize the cutoff xid after which a new changeset
2985 : * extraction replication slot can start decoding changes.
2986 : *
2987 : * Must be called with ProcArrayLock held either shared or exclusively,
2988 ECB : * although most callers will want to use exclusive mode since it is expected
2989 : * that the caller will immediately use the xid to peg the xmin horizon.
2990 : */
2991 : TransactionId
2992 GIC 497 : GetOldestSafeDecodingTransactionId(bool catalogOnly)
2993 ECB : {
2994 GIC 497 : ProcArrayStruct *arrayP = procArray;
2995 ECB : TransactionId oldestSafeXid;
2996 : int index;
2997 GIC 497 : bool recovery_in_progress = RecoveryInProgress();
2998 :
2999 497 : Assert(LWLockHeldByMe(ProcArrayLock));
3000 :
3001 : /*
3002 : * Acquire XidGenLock, so no transactions can acquire an xid while we're
3003 : * running. If no transaction with xid were running concurrently a new xid
3004 : * could influence the RecentXmin et al.
3005 ECB : *
3006 : * We initialize the computation to nextXid since that's guaranteed to be
3007 : * a safe, albeit pessimal, value.
3008 : */
3009 GIC 497 : LWLockAcquire(XidGenLock, LW_SHARED);
3010 497 : oldestSafeXid = XidFromFullTransactionId(ShmemVariableCache->nextXid);
3011 :
3012 : /*
3013 : * If there's already a slot pegging the xmin horizon, we can start with
3014 : * that value, it's guaranteed to be safe since it's computed by this
3015 ECB : * routine initially and has been enforced since. We can always use the
3016 : * slot's general xmin horizon, but the catalog horizon is only usable
3017 : * when only catalog data is going to be looked at.
3018 : */
3019 GIC 656 : if (TransactionIdIsValid(procArray->replication_slot_xmin) &&
3020 CBC 159 : TransactionIdPrecedes(procArray->replication_slot_xmin,
3021 ECB : oldestSafeXid))
3022 CBC 2 : oldestSafeXid = procArray->replication_slot_xmin;
3023 :
3024 497 : if (catalogOnly &&
3025 GIC 233 : TransactionIdIsValid(procArray->replication_slot_catalog_xmin) &&
3026 46 : TransactionIdPrecedes(procArray->replication_slot_catalog_xmin,
3027 : oldestSafeXid))
3028 11 : oldestSafeXid = procArray->replication_slot_catalog_xmin;
3029 :
3030 : /*
3031 : * If we're not in recovery, we walk over the procarray and collect the
3032 : * lowest xid. Since we're called with ProcArrayLock held and have
3033 : * acquired XidGenLock, no entries can vanish concurrently, since
3034 : * ProcGlobal->xids[i] is only set with XidGenLock held and only cleared
3035 : * with ProcArrayLock held.
3036 : *
3037 : * In recovery we can't lower the safe value besides what we've computed
3038 ECB : * above, so we'll have to wait a bit longer there. We unfortunately can
3039 : * *not* use KnownAssignedXidsGetOldestXmin() since the KnownAssignedXids
3040 : * machinery can miss values and return an older value than is safe.
3041 : */
3042 GIC 497 : if (!recovery_in_progress)
3043 : {
3044 476 : TransactionId *other_xids = ProcGlobal->xids;
3045 ECB :
3046 : /*
3047 : * Spin over procArray collecting min(ProcGlobal->xids[i])
3048 : */
3049 GIC 2478 : for (index = 0; index < arrayP->numProcs; index++)
3050 ECB : {
3051 : TransactionId xid;
3052 :
3053 : /* Fetch xid just once - see GetNewTransactionId */
3054 GIC 2002 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
3055 ECB :
3056 CBC 2002 : if (!TransactionIdIsNormal(xid))
3057 GIC 1994 : continue;
3058 :
3059 8 : if (TransactionIdPrecedes(xid, oldestSafeXid))
3060 CBC 8 : oldestSafeXid = xid;
3061 : }
3062 ECB : }
3063 :
3064 GIC 497 : LWLockRelease(XidGenLock);
3065 :
3066 497 : return oldestSafeXid;
3067 : }
3068 :
3069 : /*
3070 : * GetVirtualXIDsDelayingChkpt -- Get the VXIDs of transactions that are
3071 : * delaying checkpoint because they have critical actions in progress.
3072 : *
3073 : * Constructs an array of VXIDs of transactions that are currently in commit
3074 : * critical sections, as shown by having specified delayChkptFlags bits set
3075 : * in their PGPROC.
3076 : *
3077 : * Returns a palloc'd array that should be freed by the caller.
3078 : * *nvxids is the number of valid entries.
3079 : *
3080 : * Note that because backends set or clear delayChkptFlags without holding any
3081 : * lock, the result is somewhat indeterminate, but we don't really care. Even
3082 : * in a multiprocessor with delayed writes to shared memory, it should be
3083 : * certain that setting of delayChkptFlags will propagate to shared memory
3084 : * when the backend takes a lock, so we cannot fail to see a virtual xact as
3085 : * delayChkptFlags if it's already inserted its commit record. Whether it
3086 ECB : * takes a little while for clearing of delayChkptFlags to propagate is
3087 : * unimportant for correctness.
3088 : */
3089 : VirtualTransactionId *
3090 CBC 4670 : GetVirtualXIDsDelayingChkpt(int *nvxids, int type)
3091 : {
3092 : VirtualTransactionId *vxids;
3093 4670 : ProcArrayStruct *arrayP = procArray;
3094 GIC 4670 : int count = 0;
3095 : int index;
3096 :
3097 CBC 4670 : Assert(type != 0);
3098 :
3099 ECB : /* allocate what's certainly enough result space */
3100 : vxids = (VirtualTransactionId *)
3101 CBC 4670 : palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
3102 :
3103 4670 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3104 ECB :
3105 GIC 9695 : for (index = 0; index < arrayP->numProcs; index++)
3106 ECB : {
3107 GIC 5025 : int pgprocno = arrayP->pgprocnos[index];
3108 5025 : PGPROC *proc = &allProcs[pgprocno];
3109 :
3110 CBC 5025 : if ((proc->delayChkptFlags & type) != 0)
3111 ECB : {
3112 : VirtualTransactionId vxid;
3113 :
3114 GIC 8 : GET_VXID_FROM_PGPROC(vxid, *proc);
3115 8 : if (VirtualTransactionIdIsValid(vxid))
3116 CBC 8 : vxids[count++] = vxid;
3117 : }
3118 ECB : }
3119 :
3120 GIC 4670 : LWLockRelease(ProcArrayLock);
3121 :
3122 4670 : *nvxids = count;
3123 4670 : return vxids;
3124 : }
3125 :
3126 : /*
3127 : * HaveVirtualXIDsDelayingChkpt -- Are any of the specified VXIDs delaying?
3128 : *
3129 : * This is used with the results of GetVirtualXIDsDelayingChkpt to see if any
3130 : * of the specified VXIDs are still in critical sections of code.
3131 : *
3132 ECB : * Note: this is O(N^2) in the number of vxacts that are/were delaying, but
3133 : * those numbers should be small enough for it not to be a problem.
3134 : */
3135 : bool
3136 GIC 9 : HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids, int type)
3137 : {
3138 CBC 9 : bool result = false;
3139 GIC 9 : ProcArrayStruct *arrayP = procArray;
3140 ECB : int index;
3141 :
3142 CBC 9 : Assert(type != 0);
3143 :
3144 9 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3145 ECB :
3146 GIC 105 : for (index = 0; index < arrayP->numProcs; index++)
3147 : {
3148 CBC 98 : int pgprocno = arrayP->pgprocnos[index];
3149 GIC 98 : PGPROC *proc = &allProcs[pgprocno];
3150 ECB : VirtualTransactionId vxid;
3151 :
3152 GIC 98 : GET_VXID_FROM_PGPROC(vxid, *proc);
3153 :
3154 98 : if ((proc->delayChkptFlags & type) != 0 &&
3155 CBC 5 : VirtualTransactionIdIsValid(vxid))
3156 : {
3157 ECB : int i;
3158 :
3159 CBC 10 : for (i = 0; i < nvxids; i++)
3160 ECB : {
3161 GIC 7 : if (VirtualTransactionIdEquals(vxid, vxids[i]))
3162 : {
3163 CBC 2 : result = true;
3164 2 : break;
3165 : }
3166 : }
3167 GIC 5 : if (result)
3168 CBC 2 : break;
3169 : }
3170 ECB : }
3171 :
3172 GIC 9 : LWLockRelease(ProcArrayLock);
3173 :
3174 9 : return result;
3175 : }
3176 :
3177 : /*
3178 : * BackendPidGetProc -- get a backend's PGPROC given its PID
3179 : *
3180 : * Returns NULL if not found. Note that it is up to the caller to be
3181 ECB : * sure that the question remains meaningful for long enough for the
3182 : * answer to be used ...
3183 : */
3184 : PGPROC *
3185 CBC 4163 : BackendPidGetProc(int pid)
3186 EUB : {
3187 : PGPROC *result;
3188 ECB :
3189 GIC 4163 : if (pid == 0) /* never match dummy PGPROCs */
3190 LBC 0 : return NULL;
3191 :
3192 CBC 4163 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3193 :
3194 4163 : result = BackendPidGetProcWithLock(pid);
3195 :
3196 GIC 4163 : LWLockRelease(ProcArrayLock);
3197 :
3198 4163 : return result;
3199 : }
3200 :
3201 : /*
3202 : * BackendPidGetProcWithLock -- get a backend's PGPROC given its PID
3203 : *
3204 ECB : * Same as above, except caller must be holding ProcArrayLock. The found
3205 : * entry, if any, can be assumed to be valid as long as the lock remains held.
3206 : */
3207 : PGPROC *
3208 GIC 7355 : BackendPidGetProcWithLock(int pid)
3209 : {
3210 CBC 7355 : PGPROC *result = NULL;
3211 GBC 7355 : ProcArrayStruct *arrayP = procArray;
3212 : int index;
3213 ECB :
3214 GIC 7355 : if (pid == 0) /* never match dummy PGPROCs */
3215 LBC 0 : return NULL;
3216 :
3217 CBC 22070 : for (index = 0; index < arrayP->numProcs; index++)
3218 : {
3219 20370 : PGPROC *proc = &allProcs[arrayP->pgprocnos[index]];
3220 ECB :
3221 GIC 20370 : if (proc->pid == pid)
3222 : {
3223 5655 : result = proc;
3224 CBC 5655 : break;
3225 : }
3226 : }
3227 :
3228 GIC 7355 : return result;
3229 : }
3230 :
3231 : /*
3232 : * BackendXidGetPid -- get a backend's pid given its XID
3233 : *
3234 : * Returns 0 if not found or it's a prepared transaction. Note that
3235 : * it is up to the caller to be sure that the question remains
3236 : * meaningful for long enough for the answer to be used ...
3237 : *
3238 : * Only main transaction Ids are considered. This function is mainly
3239 : * useful for determining what backend owns a lock.
3240 : *
3241 ECB : * Beware that not every xact has an XID assigned. However, as long as you
3242 : * only call this using an XID found on disk, you're safe.
3243 : */
3244 : int
3245 CBC 30 : BackendXidGetPid(TransactionId xid)
3246 : {
3247 GIC 30 : int result = 0;
3248 CBC 30 : ProcArrayStruct *arrayP = procArray;
3249 GBC 30 : TransactionId *other_xids = ProcGlobal->xids;
3250 : int index;
3251 ECB :
3252 GIC 30 : if (xid == InvalidTransactionId) /* never match invalid xid */
3253 LBC 0 : return 0;
3254 :
3255 CBC 30 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3256 ECB :
3257 GIC 92 : for (index = 0; index < arrayP->numProcs; index++)
3258 ECB : {
3259 GIC 84 : int pgprocno = arrayP->pgprocnos[index];
3260 CBC 84 : PGPROC *proc = &allProcs[pgprocno];
3261 ECB :
3262 GIC 84 : if (other_xids[index] == xid)
3263 : {
3264 22 : result = proc->pid;
3265 CBC 22 : break;
3266 : }
3267 ECB : }
3268 :
3269 GIC 30 : LWLockRelease(ProcArrayLock);
3270 :
3271 30 : return result;
3272 : }
3273 :
3274 : /*
3275 : * IsBackendPid -- is a given pid a running backend
3276 ECB : *
3277 : * This is not called by the backend, but is called by external modules.
3278 : */
3279 : bool
3280 GIC 2 : IsBackendPid(int pid)
3281 : {
3282 2 : return (BackendPidGetProc(pid) != NULL);
3283 : }
3284 :
3285 :
3286 : /*
3287 : * GetCurrentVirtualXIDs -- returns an array of currently active VXIDs.
3288 : *
3289 : * The array is palloc'd. The number of valid entries is returned into *nvxids.
3290 : *
3291 : * The arguments allow filtering the set of VXIDs returned. Our own process
3292 : * is always skipped. In addition:
3293 : * If limitXmin is not InvalidTransactionId, skip processes with
3294 : * xmin > limitXmin.
3295 : * If excludeXmin0 is true, skip processes with xmin = 0.
3296 : * If allDbs is false, skip processes attached to other databases.
3297 : * If excludeVacuum isn't zero, skip processes for which
3298 : * (statusFlags & excludeVacuum) is not zero.
3299 : *
3300 : * Note: the purpose of the limitXmin and excludeXmin0 parameters is to
3301 : * allow skipping backends whose oldest live snapshot is no older than
3302 : * some snapshot we have. Since we examine the procarray with only shared
3303 : * lock, there are race conditions: a backend could set its xmin just after
3304 : * we look. Indeed, on multiprocessors with weak memory ordering, the
3305 : * other backend could have set its xmin *before* we look. We know however
3306 : * that such a backend must have held shared ProcArrayLock overlapping our
3307 : * own hold of ProcArrayLock, else we would see its xmin update. Therefore,
3308 : * any snapshot the other backend is taking concurrently with our scan cannot
3309 ECB : * consider any transactions as still running that we think are committed
3310 : * (since backends must hold ProcArrayLock exclusive to commit).
3311 : */
3312 : VirtualTransactionId *
3313 GIC 321 : GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0,
3314 ECB : bool allDbs, int excludeVacuum,
3315 : int *nvxids)
3316 : {
3317 : VirtualTransactionId *vxids;
3318 GIC 321 : ProcArrayStruct *arrayP = procArray;
3319 321 : int count = 0;
3320 ECB : int index;
3321 :
3322 : /* allocate what's certainly enough result space */
3323 : vxids = (VirtualTransactionId *)
3324 CBC 321 : palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
3325 :
3326 321 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3327 ECB :
3328 CBC 2041 : for (index = 0; index < arrayP->numProcs; index++)
3329 : {
3330 1720 : int pgprocno = arrayP->pgprocnos[index];
3331 1720 : PGPROC *proc = &allProcs[pgprocno];
3332 GIC 1720 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3333 ECB :
3334 CBC 1720 : if (proc == MyProc)
3335 GIC 321 : continue;
3336 ECB :
3337 GIC 1399 : if (excludeVacuum & statusFlags)
3338 12 : continue;
3339 ECB :
3340 GIC 1387 : if (allDbs || proc->databaseId == MyDatabaseId)
3341 ECB : {
3342 : /* Fetch xmin just once - might change on us */
3343 GIC 685 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3344 :
3345 685 : if (excludeXmin0 && !TransactionIdIsValid(pxmin))
3346 434 : continue;
3347 :
3348 ECB : /*
3349 : * InvalidTransactionId precedes all other XIDs, so a proc that
3350 : * hasn't set xmin yet will not be rejected by this test.
3351 : */
3352 GIC 502 : if (!TransactionIdIsValid(limitXmin) ||
3353 CBC 251 : TransactionIdPrecedesOrEquals(pxmin, limitXmin))
3354 ECB : {
3355 : VirtualTransactionId vxid;
3356 :
3357 GIC 239 : GET_VXID_FROM_PGPROC(vxid, *proc);
3358 239 : if (VirtualTransactionIdIsValid(vxid))
3359 239 : vxids[count++] = vxid;
3360 ECB : }
3361 : }
3362 : }
3363 :
3364 GIC 321 : LWLockRelease(ProcArrayLock);
3365 :
3366 321 : *nvxids = count;
3367 321 : return vxids;
3368 : }
3369 :
3370 : /*
3371 : * GetConflictingVirtualXIDs -- returns an array of currently active VXIDs.
3372 : *
3373 : * Usage is limited to conflict resolution during recovery on standby servers.
3374 : * limitXmin is supplied as either a cutoff with snapshotConflictHorizon
3375 : * semantics, or InvalidTransactionId in cases where caller cannot accurately
3376 : * determine a safe snapshotConflictHorizon value.
3377 : *
3378 : * If limitXmin is InvalidTransactionId then we want to kill everybody,
3379 : * so we're not worried if they have a snapshot or not, nor does it really
3380 : * matter what type of lock we hold. Caller must avoid calling here with
3381 : * snapshotConflictHorizon style cutoffs that were set to InvalidTransactionId
3382 : * during original execution, since that actually indicates that there is
3383 : * definitely no need for a recovery conflict (the snapshotConflictHorizon
3384 : * convention for InvalidTransactionId values is the opposite of our own!).
3385 : *
3386 : * All callers that are checking xmins always now supply a valid and useful
3387 : * value for limitXmin. The limitXmin is always lower than the lowest
3388 : * numbered KnownAssignedXid that is not already a FATAL error. This is
3389 : * because we only care about cleanup records that are cleaning up tuple
3390 : * versions from committed transactions. In that case they will only occur
3391 : * at the point where the record is less than the lowest running xid. That
3392 : * allows us to say that if any backend takes a snapshot concurrently with
3393 : * us then the conflict assessment made here would never include the snapshot
3394 : * that is being derived. So we take LW_SHARED on the ProcArray and allow
3395 : * concurrent snapshots when limitXmin is valid. We might think about adding
3396 : * Assert(limitXmin < lowest(KnownAssignedXids))
3397 : * but that would not be true in the case of FATAL errors lagging in array,
3398 : * but we already know those are bogus anyway, so we skip that test.
3399 : *
3400 : * If dbOid is valid we skip backends attached to other databases.
3401 : *
3402 : * Be careful to *not* pfree the result from this function. We reuse
3403 : * this array sufficiently often that we use malloc for the result.
3404 : */
3405 : VirtualTransactionId *
3406 8625 : GetConflictingVirtualXIDs(TransactionId limitXmin, Oid dbOid)
3407 ECB : {
3408 : static VirtualTransactionId *vxids;
3409 GIC 8625 : ProcArrayStruct *arrayP = procArray;
3410 CBC 8625 : int count = 0;
3411 ECB : int index;
3412 :
3413 : /*
3414 : * If first time through, get workspace to remember main XIDs in. We
3415 : * malloc it permanently to avoid repeated palloc/pfree overhead. Allow
3416 : * result space, remembering room for a terminator.
3417 : */
3418 GIC 8625 : if (vxids == NULL)
3419 ECB : {
3420 GIC 20 : vxids = (VirtualTransactionId *)
3421 CBC 20 : malloc(sizeof(VirtualTransactionId) * (arrayP->maxProcs + 1));
3422 20 : if (vxids == NULL)
3423 LBC 0 : ereport(ERROR,
3424 EUB : (errcode(ERRCODE_OUT_OF_MEMORY),
3425 : errmsg("out of memory")));
3426 : }
3427 :
3428 GIC 8625 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3429 ECB :
3430 GIC 8778 : for (index = 0; index < arrayP->numProcs; index++)
3431 ECB : {
3432 GIC 153 : int pgprocno = arrayP->pgprocnos[index];
3433 CBC 153 : PGPROC *proc = &allProcs[pgprocno];
3434 ECB :
3435 : /* Exclude prepared transactions */
3436 GIC 153 : if (proc->pid == 0)
3437 LBC 0 : continue;
3438 EUB :
3439 GIC 153 : if (!OidIsValid(dbOid) ||
3440 CBC 146 : proc->databaseId == dbOid)
3441 ECB : {
3442 : /* Fetch xmin just once - can't change on us, but good coding */
3443 GIC 15 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3444 ECB :
3445 : /*
3446 : * We ignore an invalid pxmin because this means that backend has
3447 : * no snapshot currently. We hold a Share lock to avoid contention
3448 : * with users taking snapshots. That is not a problem because the
3449 : * current xmin is always at least one higher than the latest
3450 : * removed xid, so any new snapshot would never conflict with the
3451 : * test here.
3452 : */
3453 GIC 15 : if (!TransactionIdIsValid(limitXmin) ||
3454 CBC 1 : (TransactionIdIsValid(pxmin) && !TransactionIdFollows(pxmin, limitXmin)))
3455 ECB : {
3456 : VirtualTransactionId vxid;
3457 :
3458 GIC 2 : GET_VXID_FROM_PGPROC(vxid, *proc);
3459 CBC 2 : if (VirtualTransactionIdIsValid(vxid))
3460 2 : vxids[count++] = vxid;
3461 ECB : }
3462 : }
3463 : }
3464 :
3465 GIC 8625 : LWLockRelease(ProcArrayLock);
3466 ECB :
3467 : /* add the terminator */
3468 GIC 8625 : vxids[count].backendId = InvalidBackendId;
3469 CBC 8625 : vxids[count].localTransactionId = InvalidLocalTransactionId;
3470 ECB :
3471 GIC 8625 : return vxids;
3472 ECB : }
3473 :
3474 : /*
3475 : * CancelVirtualTransaction - used in recovery conflict processing
3476 : *
3477 : * Returns pid of the process signaled, or 0 if not found.
3478 : */
3479 : pid_t
3480 GIC 3 : CancelVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode)
3481 ECB : {
3482 GIC 3 : return SignalVirtualTransaction(vxid, sigmode, true);
3483 ECB : }
3484 :
3485 : pid_t
3486 GIC 5 : SignalVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode,
3487 ECB : bool conflictPending)
3488 : {
3489 GIC 5 : ProcArrayStruct *arrayP = procArray;
3490 ECB : int index;
3491 GIC 5 : pid_t pid = 0;
3492 ECB :
3493 GIC 5 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3494 ECB :
3495 GIC 5 : for (index = 0; index < arrayP->numProcs; index++)
3496 ECB : {
3497 GIC 5 : int pgprocno = arrayP->pgprocnos[index];
3498 CBC 5 : PGPROC *proc = &allProcs[pgprocno];
3499 ECB : VirtualTransactionId procvxid;
3500 :
3501 GIC 5 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3502 ECB :
3503 GIC 5 : if (procvxid.backendId == vxid.backendId &&
3504 CBC 5 : procvxid.localTransactionId == vxid.localTransactionId)
3505 ECB : {
3506 GIC 5 : proc->recoveryConflictPending = conflictPending;
3507 CBC 5 : pid = proc->pid;
3508 5 : if (pid != 0)
3509 ECB : {
3510 : /*
3511 : * Kill the pid if it's still here. If not, that's what we
3512 : * wanted so ignore any errors.
3513 : */
3514 GIC 5 : (void) SendProcSignal(pid, sigmode, vxid.backendId);
3515 ECB : }
3516 GIC 5 : break;
3517 ECB : }
3518 : }
3519 :
3520 GIC 5 : LWLockRelease(ProcArrayLock);
3521 ECB :
3522 GIC 5 : return pid;
3523 ECB : }
3524 :
3525 : /*
3526 : * MinimumActiveBackends --- count backends (other than myself) that are
3527 : * in active transactions. Return true if the count exceeds the
3528 : * minimum threshold passed. This is used as a heuristic to decide if
3529 : * a pre-XLOG-flush delay is worthwhile during commit.
3530 : *
3531 : * Do not count backends that are blocked waiting for locks, since they are
3532 : * not going to get to run until someone else commits.
3533 : */
3534 : bool
3535 UIC 0 : MinimumActiveBackends(int min)
3536 EUB : {
3537 UIC 0 : ProcArrayStruct *arrayP = procArray;
3538 UBC 0 : int count = 0;
3539 EUB : int index;
3540 :
3541 : /* Quick short-circuit if no minimum is specified */
3542 UIC 0 : if (min == 0)
3543 UBC 0 : return true;
3544 EUB :
3545 : /*
3546 : * Note: for speed, we don't acquire ProcArrayLock. This is a little bit
3547 : * bogus, but since we are only testing fields for zero or nonzero, it
3548 : * should be OK. The result is only used for heuristic purposes anyway...
3549 : */
3550 UIC 0 : for (index = 0; index < arrayP->numProcs; index++)
3551 EUB : {
3552 UIC 0 : int pgprocno = arrayP->pgprocnos[index];
3553 UBC 0 : PGPROC *proc = &allProcs[pgprocno];
3554 EUB :
3555 : /*
3556 : * Since we're not holding a lock, need to be prepared to deal with
3557 : * garbage, as someone could have incremented numProcs but not yet
3558 : * filled the structure.
3559 : *
3560 : * If someone just decremented numProcs, 'proc' could also point to a
3561 : * PGPROC entry that's no longer in the array. It still points to a
3562 : * PGPROC struct, though, because freed PGPROC entries just go to the
3563 : * free list and are recycled. Its contents are nonsense in that case,
3564 : * but that's acceptable for this function.
3565 : */
3566 UIC 0 : if (pgprocno == -1)
3567 UBC 0 : continue; /* do not count deleted entries */
3568 0 : if (proc == MyProc)
3569 0 : continue; /* do not count myself */
3570 0 : if (proc->xid == InvalidTransactionId)
3571 0 : continue; /* do not count if no XID assigned */
3572 0 : if (proc->pid == 0)
3573 0 : continue; /* do not count prepared xacts */
3574 0 : if (proc->waitLock != NULL)
3575 0 : continue; /* do not count if blocked on a lock */
3576 0 : count++;
3577 0 : if (count >= min)
3578 0 : break;
3579 EUB : }
3580 :
3581 UIC 0 : return count >= min;
3582 EUB : }
3583 :
3584 : /*
3585 : * CountDBBackends --- count backends that are using specified database
3586 : */
3587 : int
3588 GIC 11 : CountDBBackends(Oid databaseid)
3589 ECB : {
3590 GIC 11 : ProcArrayStruct *arrayP = procArray;
3591 CBC 11 : int count = 0;
3592 ECB : int index;
3593 :
3594 GIC 11 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3595 ECB :
3596 GIC 20 : for (index = 0; index < arrayP->numProcs; index++)
3597 ECB : {
3598 GIC 9 : int pgprocno = arrayP->pgprocnos[index];
3599 CBC 9 : PGPROC *proc = &allProcs[pgprocno];
3600 ECB :
3601 GIC 9 : if (proc->pid == 0)
3602 LBC 0 : continue; /* do not count prepared xacts */
3603 GBC 9 : if (!OidIsValid(databaseid) ||
3604 CBC 9 : proc->databaseId == databaseid)
3605 2 : count++;
3606 ECB : }
3607 :
3608 GIC 11 : LWLockRelease(ProcArrayLock);
3609 ECB :
3610 GIC 11 : return count;
3611 ECB : }
3612 :
3613 : /*
3614 : * CountDBConnections --- counts database backends ignoring any background
3615 : * worker processes
3616 : */
3617 : int
3618 UIC 0 : CountDBConnections(Oid databaseid)
3619 EUB : {
3620 UIC 0 : ProcArrayStruct *arrayP = procArray;
3621 UBC 0 : int count = 0;
3622 EUB : int index;
3623 :
3624 UIC 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3625 EUB :
3626 UIC 0 : for (index = 0; index < arrayP->numProcs; index++)
3627 EUB : {
3628 UIC 0 : int pgprocno = arrayP->pgprocnos[index];
3629 UBC 0 : PGPROC *proc = &allProcs[pgprocno];
3630 EUB :
3631 UIC 0 : if (proc->pid == 0)
3632 UBC 0 : continue; /* do not count prepared xacts */
3633 0 : if (proc->isBackgroundWorker)
3634 0 : continue; /* do not count background workers */
3635 0 : if (!OidIsValid(databaseid) ||
3636 0 : proc->databaseId == databaseid)
3637 0 : count++;
3638 EUB : }
3639 :
3640 UIC 0 : LWLockRelease(ProcArrayLock);
3641 EUB :
3642 UIC 0 : return count;
3643 EUB : }
3644 :
3645 : /*
3646 : * CancelDBBackends --- cancel backends that are using specified database
3647 : */
3648 : void
3649 GIC 10 : CancelDBBackends(Oid databaseid, ProcSignalReason sigmode, bool conflictPending)
3650 ECB : {
3651 GIC 10 : ProcArrayStruct *arrayP = procArray;
3652 ECB : int index;
3653 :
3654 : /* tell all backends to die */
3655 GIC 10 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3656 ECB :
3657 GIC 21 : for (index = 0; index < arrayP->numProcs; index++)
3658 ECB : {
3659 GIC 11 : int pgprocno = arrayP->pgprocnos[index];
3660 CBC 11 : PGPROC *proc = &allProcs[pgprocno];
3661 ECB :
3662 GIC 11 : if (databaseid == InvalidOid || proc->databaseId == databaseid)
3663 ECB : {
3664 : VirtualTransactionId procvxid;
3665 : pid_t pid;
3666 :
3667 GIC 10 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3668 ECB :
3669 GIC 10 : proc->recoveryConflictPending = conflictPending;
3670 CBC 10 : pid = proc->pid;
3671 10 : if (pid != 0)
3672 ECB : {
3673 : /*
3674 : * Kill the pid if it's still here. If not, that's what we
3675 : * wanted so ignore any errors.
3676 : */
3677 GIC 10 : (void) SendProcSignal(pid, sigmode, procvxid.backendId);
3678 ECB : }
3679 : }
3680 : }
3681 :
3682 GIC 10 : LWLockRelease(ProcArrayLock);
3683 CBC 10 : }
3684 ECB :
3685 : /*
3686 : * CountUserBackends --- count backends that are used by specified user
3687 : */
3688 : int
3689 UIC 0 : CountUserBackends(Oid roleid)
3690 EUB : {
3691 UIC 0 : ProcArrayStruct *arrayP = procArray;
3692 UBC 0 : int count = 0;
3693 EUB : int index;
3694 :
3695 UIC 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3696 EUB :
3697 UIC 0 : for (index = 0; index < arrayP->numProcs; index++)
3698 EUB : {
3699 UIC 0 : int pgprocno = arrayP->pgprocnos[index];
3700 UBC 0 : PGPROC *proc = &allProcs[pgprocno];
3701 EUB :
3702 UIC 0 : if (proc->pid == 0)
3703 UBC 0 : continue; /* do not count prepared xacts */
3704 0 : if (proc->isBackgroundWorker)
3705 0 : continue; /* do not count background workers */
3706 0 : if (proc->roleId == roleid)
3707 0 : count++;
3708 EUB : }
3709 :
3710 UIC 0 : LWLockRelease(ProcArrayLock);
3711 EUB :
3712 UIC 0 : return count;
3713 EUB : }
3714 :
3715 : /*
3716 : * CountOtherDBBackends -- check for other backends running in the given DB
3717 : *
3718 : * If there are other backends in the DB, we will wait a maximum of 5 seconds
3719 : * for them to exit. Autovacuum backends are encouraged to exit early by
3720 : * sending them SIGTERM, but normal user backends are just waited for.
3721 : *
3722 : * The current backend is always ignored; it is caller's responsibility to
3723 : * check whether the current backend uses the given DB, if it's important.
3724 : *
3725 : * Returns true if there are (still) other backends in the DB, false if not.
3726 : * Also, *nbackends and *nprepared are set to the number of other backends
3727 : * and prepared transactions in the DB, respectively.
3728 : *
3729 : * This function is used to interlock DROP DATABASE and related commands
3730 : * against there being any active backends in the target DB --- dropping the
3731 : * DB while active backends remain would be a Bad Thing. Note that we cannot
3732 : * detect here the possibility of a newly-started backend that is trying to
3733 : * connect to the doomed database, so additional interlocking is needed during
3734 : * backend startup. The caller should normally hold an exclusive lock on the
3735 : * target DB before calling this, which is one reason we mustn't wait
3736 : * indefinitely.
3737 : */
3738 : bool
3739 GIC 822 : CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared)
3740 ECB : {
3741 GIC 822 : ProcArrayStruct *arrayP = procArray;
3742 ECB :
3743 : #define MAXAUTOVACPIDS 10 /* max autovacs to SIGTERM per iteration */
3744 : int autovac_pids[MAXAUTOVACPIDS];
3745 : int tries;
3746 :
3747 : /* 50 tries with 100ms sleep between tries makes 5 sec total wait */
3748 GIC 822 : for (tries = 0; tries < 50; tries++)
3749 ECB : {
3750 GIC 822 : int nautovacs = 0;
3751 CBC 822 : bool found = false;
3752 ECB : int index;
3753 :
3754 GIC 822 : CHECK_FOR_INTERRUPTS();
3755 ECB :
3756 GIC 822 : *nbackends = *nprepared = 0;
3757 ECB :
3758 GIC 822 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3759 ECB :
3760 GIC 2107 : for (index = 0; index < arrayP->numProcs; index++)
3761 ECB : {
3762 GIC 1285 : int pgprocno = arrayP->pgprocnos[index];
3763 CBC 1285 : PGPROC *proc = &allProcs[pgprocno];
3764 1285 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3765 ECB :
3766 GIC 1285 : if (proc->databaseId != databaseId)
3767 CBC 679 : continue;
3768 606 : if (proc == MyProc)
3769 606 : continue;
3770 ECB :
3771 UIC 0 : found = true;
3772 EUB :
3773 UIC 0 : if (proc->pid == 0)
3774 UBC 0 : (*nprepared)++;
3775 EUB : else
3776 : {
3777 UIC 0 : (*nbackends)++;
3778 UBC 0 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
3779 EUB : nautovacs < MAXAUTOVACPIDS)
3780 UIC 0 : autovac_pids[nautovacs++] = proc->pid;
3781 EUB : }
3782 : }
3783 :
3784 GIC 822 : LWLockRelease(ProcArrayLock);
3785 ECB :
3786 GIC 822 : if (!found)
3787 CBC 822 : return false; /* no conflicting backends, so done */
3788 ECB :
3789 : /*
3790 : * Send SIGTERM to any conflicting autovacuums before sleeping. We
3791 : * postpone this step until after the loop because we don't want to
3792 : * hold ProcArrayLock while issuing kill(). We have no idea what might
3793 : * block kill() inside the kernel...
3794 : */
3795 UIC 0 : for (index = 0; index < nautovacs; index++)
3796 UBC 0 : (void) kill(autovac_pids[index], SIGTERM); /* ignore any error */
3797 EUB :
3798 : /* sleep, then try again */
3799 UIC 0 : pg_usleep(100 * 1000L); /* 100ms */
3800 EUB : }
3801 :
3802 UIC 0 : return true; /* timed out, still conflicts */
3803 EUB : }
3804 :
3805 : /*
3806 : * Terminate existing connections to the specified database. This routine
3807 : * is used by the DROP DATABASE command when user has asked to forcefully
3808 : * drop the database.
3809 : *
3810 : * The current backend is always ignored; it is caller's responsibility to
3811 : * check whether the current backend uses the given DB, if it's important.
3812 : *
3813 : * It doesn't allow to terminate the connections even if there is a one
3814 : * backend with the prepared transaction in the target database.
3815 : */
3816 : void
3817 GIC 1 : TerminateOtherDBBackends(Oid databaseId)
3818 ECB : {
3819 GIC 1 : ProcArrayStruct *arrayP = procArray;
3820 CBC 1 : List *pids = NIL;
3821 1 : int nprepared = 0;
3822 ECB : int i;
3823 :
3824 GIC 1 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3825 ECB :
3826 GIC 4 : for (i = 0; i < procArray->numProcs; i++)
3827 ECB : {
3828 GIC 3 : int pgprocno = arrayP->pgprocnos[i];
3829 CBC 3 : PGPROC *proc = &allProcs[pgprocno];
3830 ECB :
3831 GIC 3 : if (proc->databaseId != databaseId)
3832 CBC 3 : continue;
3833 LBC 0 : if (proc == MyProc)
3834 UBC 0 : continue;
3835 EUB :
3836 UIC 0 : if (proc->pid != 0)
3837 UBC 0 : pids = lappend_int(pids, proc->pid);
3838 EUB : else
3839 UIC 0 : nprepared++;
3840 EUB : }
3841 :
3842 GIC 1 : LWLockRelease(ProcArrayLock);
3843 ECB :
3844 GIC 1 : if (nprepared > 0)
3845 LBC 0 : ereport(ERROR,
3846 EUB : (errcode(ERRCODE_OBJECT_IN_USE),
3847 : errmsg("database \"%s\" is being used by prepared transactions",
3848 : get_database_name(databaseId)),
3849 : errdetail_plural("There is %d prepared transaction using the database.",
3850 : "There are %d prepared transactions using the database.",
3851 : nprepared,
3852 : nprepared)));
3853 :
3854 GIC 1 : if (pids)
3855 ECB : {
3856 : ListCell *lc;
3857 :
3858 : /*
3859 : * Check whether we have the necessary rights to terminate other
3860 : * sessions. We don't terminate any session until we ensure that we
3861 : * have rights on all the sessions to be terminated. These checks are
3862 : * the same as we do in pg_terminate_backend.
3863 : *
3864 : * In this case we don't raise some warnings - like "PID %d is not a
3865 : * PostgreSQL server process", because for us already finished session
3866 : * is not a problem.
3867 : */
3868 UIC 0 : foreach(lc, pids)
3869 EUB : {
3870 UIC 0 : int pid = lfirst_int(lc);
3871 UBC 0 : PGPROC *proc = BackendPidGetProc(pid);
3872 EUB :
3873 UIC 0 : if (proc != NULL)
3874 EUB : {
3875 : /* Only allow superusers to signal superuser-owned backends. */
3876 UIC 0 : if (superuser_arg(proc->roleId) && !superuser())
3877 UBC 0 : ereport(ERROR,
3878 EUB : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3879 : errmsg("must be a superuser to terminate superuser process")));
3880 :
3881 : /* Users can signal backends they have role membership in. */
3882 UIC 0 : if (!has_privs_of_role(GetUserId(), proc->roleId) &&
3883 UBC 0 : !has_privs_of_role(GetUserId(), ROLE_PG_SIGNAL_BACKEND))
3884 0 : ereport(ERROR,
3885 EUB : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3886 : errmsg("permission denied to terminate process"),
3887 : errdetail("Only roles with privileges of the role whose process is being terminated or with privileges of the \"%s\" role may terminate this process.",
3888 : "pg_signal_backend")));
3889 : }
3890 : }
3891 :
3892 : /*
3893 : * There's a race condition here: once we release the ProcArrayLock,
3894 : * it's possible for the session to exit before we issue kill. That
3895 : * race condition possibility seems too unlikely to worry about. See
3896 : * pg_signal_backend.
3897 : */
3898 UIC 0 : foreach(lc, pids)
3899 : {
3900 0 : int pid = lfirst_int(lc);
3901 UBC 0 : PGPROC *proc = BackendPidGetProc(pid);
3902 :
3903 0 : if (proc != NULL)
3904 EUB : {
3905 : /*
3906 : * If we have setsid(), signal the backend's whole process
3907 : * group
3908 : */
3909 : #ifdef HAVE_SETSID
3910 UIC 0 : (void) kill(-pid, SIGTERM);
3911 : #else
3912 : (void) kill(pid, SIGTERM);
3913 EUB : #endif
3914 : }
3915 : }
3916 : }
3917 GIC 1 : }
3918 :
3919 : /*
3920 ECB : * ProcArraySetReplicationSlotXmin
3921 : *
3922 : * Install limits to future computations of the xmin horizon to prevent vacuum
3923 : * and HOT pruning from removing affected rows still needed by clients with
3924 : * replication slots.
3925 : */
3926 : void
3927 GIC 2189 : ProcArraySetReplicationSlotXmin(TransactionId xmin, TransactionId catalog_xmin,
3928 : bool already_locked)
3929 : {
3930 CBC 2189 : Assert(!already_locked || LWLockHeldByMe(ProcArrayLock));
3931 :
3932 GIC 2189 : if (!already_locked)
3933 CBC 1847 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3934 :
3935 2189 : procArray->replication_slot_xmin = xmin;
3936 2189 : procArray->replication_slot_catalog_xmin = catalog_xmin;
3937 :
3938 2189 : if (!already_locked)
3939 1847 : LWLockRelease(ProcArrayLock);
3940 :
3941 2189 : elog(DEBUG1, "xmin required by slots: data %u, catalog %u",
3942 ECB : xmin, catalog_xmin);
3943 GIC 2189 : }
3944 ECB :
3945 : /*
3946 : * ProcArrayGetReplicationSlotXmin
3947 : *
3948 : * Return the current slot xmin limits. That's useful to be able to remove
3949 : * data that's older than those limits.
3950 : */
3951 : void
3952 GIC 21 : ProcArrayGetReplicationSlotXmin(TransactionId *xmin,
3953 : TransactionId *catalog_xmin)
3954 : {
3955 CBC 21 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3956 :
3957 GIC 21 : if (xmin != NULL)
3958 LBC 0 : *xmin = procArray->replication_slot_xmin;
3959 :
3960 CBC 21 : if (catalog_xmin != NULL)
3961 GBC 21 : *catalog_xmin = procArray->replication_slot_catalog_xmin;
3962 :
3963 CBC 21 : LWLockRelease(ProcArrayLock);
3964 21 : }
3965 :
3966 ECB : /*
3967 : * XidCacheRemoveRunningXids
3968 : *
3969 : * Remove a bunch of TransactionIds from the list of known-running
3970 : * subtransactions for my backend. Both the specified xid and those in
3971 : * the xids[] array (of length nxids) are removed from the subxids cache.
3972 : * latestXid must be the latest XID among the group.
3973 : */
3974 : void
3975 GIC 610 : XidCacheRemoveRunningXids(TransactionId xid,
3976 : int nxids, const TransactionId *xids,
3977 : TransactionId latestXid)
3978 ECB : {
3979 : int i,
3980 : j;
3981 : XidCacheStatus *mysubxidstat;
3982 :
3983 GIC 610 : Assert(TransactionIdIsValid(xid));
3984 :
3985 : /*
3986 ECB : * We must hold ProcArrayLock exclusively in order to remove transactions
3987 : * from the PGPROC array. (See src/backend/access/transam/README.) It's
3988 : * possible this could be relaxed since we know this routine is only used
3989 : * to abort subtransactions, but pending closer analysis we'd best be
3990 : * conservative.
3991 : *
3992 : * Note that we do not have to be careful about memory ordering of our own
3993 : * reads wrt. GetNewTransactionId() here - only this process can modify
3994 : * relevant fields of MyProc/ProcGlobal->xids[]. But we do have to be
3995 : * careful about our own writes being well ordered.
3996 : */
3997 GIC 610 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3998 :
3999 610 : mysubxidstat = &ProcGlobal->subxidStates[MyProc->pgxactoff];
4000 ECB :
4001 : /*
4002 : * Under normal circumstances xid and xids[] will be in increasing order,
4003 : * as will be the entries in subxids. Scan backwards to avoid O(N^2)
4004 : * behavior when removing a lot of xids.
4005 : */
4006 GIC 640 : for (i = nxids - 1; i >= 0; i--)
4007 : {
4008 30 : TransactionId anxid = xids[i];
4009 ECB :
4010 GIC 30 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4011 ECB : {
4012 GIC 30 : if (TransactionIdEquals(MyProc->subxids.xids[j], anxid))
4013 ECB : {
4014 GIC 30 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4015 CBC 30 : pg_write_barrier();
4016 GIC 30 : mysubxidstat->count--;
4017 CBC 30 : MyProc->subxidStatus.count--;
4018 30 : break;
4019 ECB : }
4020 : }
4021 :
4022 : /*
4023 : * Ordinarily we should have found it, unless the cache has
4024 : * overflowed. However it's also possible for this routine to be
4025 : * invoked multiple times for the same subtransaction, in case of an
4026 : * error during AbortSubTransaction. So instead of Assert, emit a
4027 : * debug warning.
4028 : */
4029 GIC 30 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4030 UIC 0 : elog(WARNING, "did not find subXID %u in MyProc", anxid);
4031 : }
4032 ECB :
4033 GBC 610 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4034 : {
4035 GIC 610 : if (TransactionIdEquals(MyProc->subxids.xids[j], xid))
4036 ECB : {
4037 GIC 610 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4038 CBC 610 : pg_write_barrier();
4039 GIC 610 : mysubxidstat->count--;
4040 CBC 610 : MyProc->subxidStatus.count--;
4041 610 : break;
4042 ECB : }
4043 : }
4044 : /* Ordinarily we should have found it, unless the cache has overflowed */
4045 GIC 610 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4046 UIC 0 : elog(WARNING, "did not find subXID %u in MyProc", xid);
4047 :
4048 ECB : /* Also advance global latestCompletedXid while holding the lock */
4049 GBC 610 : MaintainLatestCompletedXid(latestXid);
4050 :
4051 : /* ... and xactCompletionCount */
4052 CBC 610 : ShmemVariableCache->xactCompletionCount++;
4053 :
4054 GIC 610 : LWLockRelease(ProcArrayLock);
4055 CBC 610 : }
4056 :
4057 ECB : #ifdef XIDCACHE_DEBUG
4058 :
4059 : /*
4060 : * Print stats about effectiveness of XID cache
4061 : */
4062 : static void
4063 : DisplayXidCache(void)
4064 : {
4065 : fprintf(stderr,
4066 : "XidCache: xmin: %ld, known: %ld, myxact: %ld, latest: %ld, mainxid: %ld, childxid: %ld, knownassigned: %ld, nooflo: %ld, slow: %ld\n",
4067 : xc_by_recent_xmin,
4068 : xc_by_known_xact,
4069 : xc_by_my_xact,
4070 : xc_by_latest_xid,
4071 : xc_by_main_xid,
4072 : xc_by_child_xid,
4073 : xc_by_known_assigned,
4074 : xc_no_overflow,
4075 : xc_slow_answer);
4076 : }
4077 : #endif /* XIDCACHE_DEBUG */
4078 :
4079 : /*
4080 : * If rel != NULL, return test state appropriate for relation, otherwise
4081 : * return state usable for all relations. The latter may consider XIDs as
4082 : * not-yet-visible-to-everyone that a state for a specific relation would
4083 : * already consider visible-to-everyone.
4084 : *
4085 : * This needs to be called while a snapshot is active or registered, otherwise
4086 : * there are wraparound and other dangers.
4087 : *
4088 : * See comment for GlobalVisState for details.
4089 : */
4090 : GlobalVisState *
4091 GIC 14036985 : GlobalVisTestFor(Relation rel)
4092 : {
4093 14036985 : GlobalVisState *state = NULL;
4094 ECB :
4095 : /* XXX: we should assert that a snapshot is pushed or registered */
4096 CBC 14036985 : Assert(RecentXmin);
4097 :
4098 GIC 14036985 : switch (GlobalVisHorizonKindForRel(rel))
4099 ECB : {
4100 GIC 55912 : case VISHORIZON_SHARED:
4101 CBC 55912 : state = &GlobalVisSharedRels;
4102 GIC 55912 : break;
4103 CBC 2646677 : case VISHORIZON_CATALOG:
4104 2646677 : state = &GlobalVisCatalogRels;
4105 2646677 : break;
4106 11292629 : case VISHORIZON_DATA:
4107 11292629 : state = &GlobalVisDataRels;
4108 11292629 : break;
4109 41767 : case VISHORIZON_TEMP:
4110 41767 : state = &GlobalVisTempRels;
4111 41767 : break;
4112 ECB : }
4113 :
4114 CBC 14036985 : Assert(FullTransactionIdIsValid(state->definitely_needed) &&
4115 : FullTransactionIdIsValid(state->maybe_needed));
4116 :
4117 14036985 : return state;
4118 : }
4119 :
4120 ECB : /*
4121 : * Return true if it's worth updating the accurate maybe_needed boundary.
4122 : *
4123 : * As it is somewhat expensive to determine xmin horizons, we don't want to
4124 : * repeatedly do so when there is a low likelihood of it being beneficial.
4125 : *
4126 : * The current heuristic is that we update only if RecentXmin has changed
4127 : * since the last update. If the oldest currently running transaction has not
4128 : * finished, it is unlikely that recomputing the horizon would be useful.
4129 : */
4130 : static bool
4131 GIC 426533 : GlobalVisTestShouldUpdate(GlobalVisState *state)
4132 : {
4133 : /* hasn't been updated yet */
4134 CBC 426533 : if (!TransactionIdIsValid(ComputeXidHorizonsResultLastXmin))
4135 GIC 6917 : return true;
4136 :
4137 ECB : /*
4138 : * If the maybe_needed/definitely_needed boundaries are the same, it's
4139 : * unlikely to be beneficial to refresh boundaries.
4140 : */
4141 GIC 419616 : if (FullTransactionIdFollowsOrEquals(state->maybe_needed,
4142 : state->definitely_needed))
4143 15 : return false;
4144 ECB :
4145 : /* does the last snapshot built have a different xmin? */
4146 CBC 419601 : return RecentXmin != ComputeXidHorizonsResultLastXmin;
4147 : }
4148 :
4149 ECB : static void
4150 GIC 204530 : GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons)
4151 : {
4152 : GlobalVisSharedRels.maybe_needed =
4153 CBC 204530 : FullXidRelativeTo(horizons->latest_completed,
4154 : horizons->shared_oldest_nonremovable);
4155 : GlobalVisCatalogRels.maybe_needed =
4156 204530 : FullXidRelativeTo(horizons->latest_completed,
4157 : horizons->catalog_oldest_nonremovable);
4158 : GlobalVisDataRels.maybe_needed =
4159 204530 : FullXidRelativeTo(horizons->latest_completed,
4160 : horizons->data_oldest_nonremovable);
4161 : GlobalVisTempRels.maybe_needed =
4162 204530 : FullXidRelativeTo(horizons->latest_completed,
4163 : horizons->temp_oldest_nonremovable);
4164 :
4165 ECB : /*
4166 : * In longer running transactions it's possible that transactions we
4167 : * previously needed to treat as running aren't around anymore. So update
4168 : * definitely_needed to not be earlier than maybe_needed.
4169 : */
4170 : GlobalVisSharedRels.definitely_needed =
4171 GIC 204530 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
4172 : GlobalVisSharedRels.definitely_needed);
4173 : GlobalVisCatalogRels.definitely_needed =
4174 CBC 204530 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
4175 : GlobalVisCatalogRels.definitely_needed);
4176 : GlobalVisDataRels.definitely_needed =
4177 204530 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
4178 : GlobalVisDataRels.definitely_needed);
4179 GIC 204530 : GlobalVisTempRels.definitely_needed = GlobalVisTempRels.maybe_needed;
4180 ECB :
4181 GIC 204530 : ComputeXidHorizonsResultLastXmin = RecentXmin;
4182 CBC 204530 : }
4183 :
4184 ECB : /*
4185 : * Update boundaries in GlobalVis{Shared,Catalog, Data}Rels
4186 : * using ComputeXidHorizons().
4187 : */
4188 : static void
4189 GIC 121945 : GlobalVisUpdate(void)
4190 : {
4191 : ComputeXidHorizonsResult horizons;
4192 ECB :
4193 : /* updates the horizons as a side-effect */
4194 GIC 121945 : ComputeXidHorizons(&horizons);
4195 121945 : }
4196 :
4197 ECB : /*
4198 : * Return true if no snapshot still considers fxid to be running.
4199 : *
4200 : * The state passed needs to have been initialized for the relation fxid is
4201 : * from (NULL is also OK), otherwise the result may not be correct.
4202 : *
4203 : * See comment for GlobalVisState for details.
4204 : */
4205 : bool
4206 GIC 9950611 : GlobalVisTestIsRemovableFullXid(GlobalVisState *state,
4207 : FullTransactionId fxid)
4208 : {
4209 ECB : /*
4210 : * If fxid is older than maybe_needed bound, it definitely is visible to
4211 : * everyone.
4212 : */
4213 GIC 9950611 : if (FullTransactionIdPrecedes(fxid, state->maybe_needed))
4214 2832131 : return true;
4215 :
4216 ECB : /*
4217 : * If fxid is >= definitely_needed bound, it is very likely to still be
4218 : * considered running.
4219 : */
4220 GIC 7118480 : if (FullTransactionIdFollowsOrEquals(fxid, state->definitely_needed))
4221 6691968 : return false;
4222 :
4223 ECB : /*
4224 : * fxid is between maybe_needed and definitely_needed, i.e. there might or
4225 : * might not exist a snapshot considering fxid running. If it makes sense,
4226 : * update boundaries and recheck.
4227 : */
4228 GIC 426512 : if (GlobalVisTestShouldUpdate(state))
4229 : {
4230 121939 : GlobalVisUpdate();
4231 ECB :
4232 GIC 121939 : Assert(FullTransactionIdPrecedes(fxid, state->definitely_needed));
4233 ECB :
4234 GIC 121939 : return FullTransactionIdPrecedes(fxid, state->maybe_needed);
4235 ECB : }
4236 : else
4237 CBC 304573 : return false;
4238 : }
4239 :
4240 ECB : /*
4241 : * Wrapper around GlobalVisTestIsRemovableFullXid() for 32bit xids.
4242 : *
4243 : * It is crucial that this only gets called for xids from a source that
4244 : * protects against xid wraparounds (e.g. from a table and thus protected by
4245 : * relfrozenxid).
4246 : */
4247 : bool
4248 GIC 9949850 : GlobalVisTestIsRemovableXid(GlobalVisState *state, TransactionId xid)
4249 : {
4250 : FullTransactionId fxid;
4251 ECB :
4252 : /*
4253 : * Convert 32 bit argument to FullTransactionId. We can do so safely
4254 : * because we know the xid has to, at the very least, be between
4255 : * [oldestXid, nextXid), i.e. within 2 billion of xid. To avoid taking a
4256 : * lock to determine either, we can just compare with
4257 : * state->definitely_needed, which was based on those value at the time
4258 : * the current snapshot was built.
4259 : */
4260 GIC 9949850 : fxid = FullXidRelativeTo(state->definitely_needed, xid);
4261 :
4262 9949850 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4263 ECB : }
4264 :
4265 : /*
4266 : * Return FullTransactionId below which all transactions are not considered
4267 : * running anymore.
4268 : *
4269 : * Note: This is less efficient than testing with
4270 : * GlobalVisTestIsRemovableFullXid as it likely requires building an accurate
4271 : * cutoff, even in the case all the XIDs compared with the cutoff are outside
4272 : * [maybe_needed, definitely_needed).
4273 : */
4274 : FullTransactionId
4275 GIC 21 : GlobalVisTestNonRemovableFullHorizon(GlobalVisState *state)
4276 : {
4277 : /* acquire accurate horizon if not already done */
4278 CBC 21 : if (GlobalVisTestShouldUpdate(state))
4279 GIC 6 : GlobalVisUpdate();
4280 :
4281 CBC 21 : return state->maybe_needed;
4282 ECB : }
4283 :
4284 : /* Convenience wrapper around GlobalVisTestNonRemovableFullHorizon */
4285 : TransactionId
4286 GIC 21 : GlobalVisTestNonRemovableHorizon(GlobalVisState *state)
4287 : {
4288 : FullTransactionId cutoff;
4289 ECB :
4290 GIC 21 : cutoff = GlobalVisTestNonRemovableFullHorizon(state);
4291 :
4292 21 : return XidFromFullTransactionId(cutoff);
4293 ECB : }
4294 :
4295 : /*
4296 : * Convenience wrapper around GlobalVisTestFor() and
4297 : * GlobalVisTestIsRemovableFullXid(), see their comments.
4298 : */
4299 : bool
4300 GIC 761 : GlobalVisCheckRemovableFullXid(Relation rel, FullTransactionId fxid)
4301 : {
4302 : GlobalVisState *state;
4303 ECB :
4304 GIC 761 : state = GlobalVisTestFor(rel);
4305 :
4306 761 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4307 ECB : }
4308 :
4309 : /*
4310 : * Convenience wrapper around GlobalVisTestFor() and
4311 : * GlobalVisTestIsRemovableXid(), see their comments.
4312 : */
4313 : bool
4314 GIC 6 : GlobalVisCheckRemovableXid(Relation rel, TransactionId xid)
4315 : {
4316 : GlobalVisState *state;
4317 ECB :
4318 GIC 6 : state = GlobalVisTestFor(rel);
4319 :
4320 6 : return GlobalVisTestIsRemovableXid(state, xid);
4321 ECB : }
4322 :
4323 : /*
4324 : * Safely retract *xid by retreat_by, store the result in *xid.
4325 : *
4326 : * Need to be careful to prevent *xid from retreating below
4327 : * FirstNormalTransactionId during epoch 0. This is important to prevent
4328 : * generating xids that cannot be converted to a FullTransactionId without
4329 : * wrapping around.
4330 : *
4331 : * If retreat_by would lead to a too old xid, FirstNormalTransactionId is
4332 : * returned instead.
4333 : */
4334 : static void
4335 GIC 696287 : TransactionIdRetreatSafely(TransactionId *xid, int retreat_by, FullTransactionId rel)
4336 : {
4337 696287 : TransactionId original_xid = *xid;
4338 ECB : FullTransactionId fxid;
4339 : uint64 fxid_i;
4340 :
4341 GIC 696287 : Assert(TransactionIdIsNormal(original_xid));
4342 696287 : Assert(retreat_by >= 0); /* relevant GUCs are stored as ints */
4343 696287 : AssertTransactionIdInAllowableRange(original_xid);
4344 ECB :
4345 CBC 696287 : if (retreat_by == 0)
4346 696287 : return;
4347 :
4348 LBC 0 : fxid = FullXidRelativeTo(rel, original_xid);
4349 0 : fxid_i = U64FromFullTransactionId(fxid);
4350 :
4351 UBC 0 : if ((fxid_i - FirstNormalTransactionId) <= retreat_by)
4352 0 : *xid = FirstNormalTransactionId;
4353 : else
4354 EUB : {
4355 UBC 0 : *xid = TransactionIdRetreatedBy(original_xid, retreat_by);
4356 UIC 0 : Assert(TransactionIdIsNormal(*xid));
4357 0 : Assert(NormalTransactionIdPrecedes(*xid, original_xid));
4358 EUB : }
4359 : }
4360 :
4361 : /*
4362 : * Convert a 32 bit transaction id into 64 bit transaction id, by assuming it
4363 : * is within MaxTransactionId / 2 of XidFromFullTransactionId(rel).
4364 : *
4365 : * Be very careful about when to use this function. It can only safely be used
4366 : * when there is a guarantee that xid is within MaxTransactionId / 2 xids of
4367 : * rel. That e.g. can be guaranteed if the caller assures a snapshot is
4368 : * held by the backend and xid is from a table (where vacuum/freezing ensures
4369 : * the xid has to be within that range), or if xid is from the procarray and
4370 : * prevents xid wraparound that way.
4371 : */
4372 : static inline FullTransactionId
4373 GIC 13230009 : FullXidRelativeTo(FullTransactionId rel, TransactionId xid)
4374 : {
4375 13230009 : TransactionId rel_xid = XidFromFullTransactionId(rel);
4376 ECB :
4377 GIC 13230009 : Assert(TransactionIdIsValid(xid));
4378 CBC 13230009 : Assert(TransactionIdIsValid(rel_xid));
4379 :
4380 ECB : /* not guaranteed to find issues, but likely to catch mistakes */
4381 CBC 13230009 : AssertTransactionIdInAllowableRange(xid);
4382 :
4383 GIC 26460018 : return FullTransactionIdFromU64(U64FromFullTransactionId(rel)
4384 CBC 13230009 : + (int32) (xid - rel_xid));
4385 : }
4386 ECB :
4387 :
4388 : /* ----------------------------------------------
4389 : * KnownAssignedTransactionIds sub-module
4390 : * ----------------------------------------------
4391 : */
4392 :
4393 : /*
4394 : * In Hot Standby mode, we maintain a list of transactions that are (or were)
4395 : * running on the primary at the current point in WAL. These XIDs must be
4396 : * treated as running by standby transactions, even though they are not in
4397 : * the standby server's PGPROC array.
4398 : *
4399 : * We record all XIDs that we know have been assigned. That includes all the
4400 : * XIDs seen in WAL records, plus all unobserved XIDs that we can deduce have
4401 : * been assigned. We can deduce the existence of unobserved XIDs because we
4402 : * know XIDs are assigned in sequence, with no gaps. The KnownAssignedXids
4403 : * list expands as new XIDs are observed or inferred, and contracts when
4404 : * transaction completion records arrive.
4405 : *
4406 : * During hot standby we do not fret too much about the distinction between
4407 : * top-level XIDs and subtransaction XIDs. We store both together in the
4408 : * KnownAssignedXids list. In backends, this is copied into snapshots in
4409 : * GetSnapshotData(), taking advantage of the fact that XidInMVCCSnapshot()
4410 : * doesn't care about the distinction either. Subtransaction XIDs are
4411 : * effectively treated as top-level XIDs and in the typical case pg_subtrans
4412 : * links are *not* maintained (which does not affect visibility).
4413 : *
4414 : * We have room in KnownAssignedXids and in snapshots to hold maxProcs *
4415 : * (1 + PGPROC_MAX_CACHED_SUBXIDS) XIDs, so every primary transaction must
4416 : * report its subtransaction XIDs in a WAL XLOG_XACT_ASSIGNMENT record at
4417 : * least every PGPROC_MAX_CACHED_SUBXIDS. When we receive one of these
4418 : * records, we mark the subXIDs as children of the top XID in pg_subtrans,
4419 : * and then remove them from KnownAssignedXids. This prevents overflow of
4420 : * KnownAssignedXids and snapshots, at the cost that status checks for these
4421 : * subXIDs will take a slower path through TransactionIdIsInProgress().
4422 : * This means that KnownAssignedXids is not necessarily complete for subXIDs,
4423 : * though it should be complete for top-level XIDs; this is the same situation
4424 : * that holds with respect to the PGPROC entries in normal running.
4425 : *
4426 : * When we throw away subXIDs from KnownAssignedXids, we need to keep track of
4427 : * that, similarly to tracking overflow of a PGPROC's subxids array. We do
4428 : * that by remembering the lastOverflowedXid, ie the last thrown-away subXID.
4429 : * As long as that is within the range of interesting XIDs, we have to assume
4430 : * that subXIDs are missing from snapshots. (Note that subXID overflow occurs
4431 : * on primary when 65th subXID arrives, whereas on standby it occurs when 64th
4432 : * subXID arrives - that is not an error.)
4433 : *
4434 : * Should a backend on primary somehow disappear before it can write an abort
4435 : * record, then we just leave those XIDs in KnownAssignedXids. They actually
4436 : * aborted but we think they were running; the distinction is irrelevant
4437 : * because either way any changes done by the transaction are not visible to
4438 : * backends in the standby. We prune KnownAssignedXids when
4439 : * XLOG_RUNNING_XACTS arrives, to forestall possible overflow of the
4440 : * array due to such dead XIDs.
4441 : */
4442 :
4443 : /*
4444 : * RecordKnownAssignedTransactionIds
4445 : * Record the given XID in KnownAssignedXids, as well as any preceding
4446 : * unobserved XIDs.
4447 : *
4448 : * RecordKnownAssignedTransactionIds() should be run for *every* WAL record
4449 : * associated with a transaction. Must be called for each record after we
4450 : * have executed StartupCLOG() et al, since we must ExtendCLOG() etc..
4451 : *
4452 : * Called during recovery in analogy with and in place of GetNewTransactionId()
4453 : */
4454 : void
4455 GIC 2260738 : RecordKnownAssignedTransactionIds(TransactionId xid)
4456 : {
4457 2260738 : Assert(standbyState >= STANDBY_INITIALIZED);
4458 CBC 2260738 : Assert(TransactionIdIsValid(xid));
4459 GIC 2260738 : Assert(TransactionIdIsValid(latestObservedXid));
4460 ECB :
4461 CBC 2260738 : elog(trace_recovery(DEBUG4), "record known xact %u latestObservedXid %u",
4462 ECB : xid, latestObservedXid);
4463 :
4464 : /*
4465 : * When a newly observed xid arrives, it is frequently the case that it is
4466 : * *not* the next xid in sequence. When this occurs, we must treat the
4467 : * intervening xids as running also.
4468 : */
4469 GIC 2260738 : if (TransactionIdFollows(xid, latestObservedXid))
4470 : {
4471 : TransactionId next_expected_xid;
4472 ECB :
4473 : /*
4474 : * Extend subtrans like we do in GetNewTransactionId() during normal
4475 : * operation using individual extend steps. Note that we do not need
4476 : * to extend clog since its extensions are WAL logged.
4477 : *
4478 : * This part has to be done regardless of standbyState since we
4479 : * immediately start assigning subtransactions to their toplevel
4480 : * transactions.
4481 : */
4482 GIC 19416 : next_expected_xid = latestObservedXid;
4483 39120 : while (TransactionIdPrecedes(next_expected_xid, xid))
4484 : {
4485 CBC 19704 : TransactionIdAdvance(next_expected_xid);
4486 19704 : ExtendSUBTRANS(next_expected_xid);
4487 : }
4488 19416 : Assert(next_expected_xid == xid);
4489 ECB :
4490 : /*
4491 : * If the KnownAssignedXids machinery isn't up yet, there's nothing
4492 : * more to do since we don't track assigned xids yet.
4493 : */
4494 GIC 19416 : if (standbyState <= STANDBY_INITIALIZED)
4495 : {
4496 4 : latestObservedXid = xid;
4497 CBC 4 : return;
4498 : }
4499 ECB :
4500 : /*
4501 : * Add (latestObservedXid, xid] onto the KnownAssignedXids array.
4502 : */
4503 GIC 19412 : next_expected_xid = latestObservedXid;
4504 19412 : TransactionIdAdvance(next_expected_xid);
4505 19412 : KnownAssignedXidsAdd(next_expected_xid, xid, false);
4506 ECB :
4507 : /*
4508 : * Now we can advance latestObservedXid
4509 : */
4510 GIC 19412 : latestObservedXid = xid;
4511 :
4512 : /* ShmemVariableCache->nextXid must be beyond any observed xid */
4513 CBC 19412 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
4514 : }
4515 : }
4516 ECB :
4517 : /*
4518 : * ExpireTreeKnownAssignedTransactionIds
4519 : * Remove the given XIDs from KnownAssignedXids.
4520 : *
4521 : * Called during recovery in analogy with and in place of ProcArrayEndTransaction()
4522 : */
4523 : void
4524 GIC 18179 : ExpireTreeKnownAssignedTransactionIds(TransactionId xid, int nsubxids,
4525 : TransactionId *subxids, TransactionId max_xid)
4526 : {
4527 CBC 18179 : Assert(standbyState >= STANDBY_INITIALIZED);
4528 :
4529 : /*
4530 ECB : * Uses same locking as transaction commit
4531 : */
4532 GIC 18179 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4533 :
4534 18179 : KnownAssignedXidsRemoveTree(xid, nsubxids, subxids);
4535 ECB :
4536 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4537 CBC 18179 : MaintainLatestCompletedXidRecovery(max_xid);
4538 :
4539 : /* ... and xactCompletionCount */
4540 18179 : ShmemVariableCache->xactCompletionCount++;
4541 :
4542 GIC 18179 : LWLockRelease(ProcArrayLock);
4543 CBC 18179 : }
4544 :
4545 ECB : /*
4546 : * ExpireAllKnownAssignedTransactionIds
4547 : * Remove all entries in KnownAssignedXids and reset lastOverflowedXid.
4548 : */
4549 : void
4550 GIC 71 : ExpireAllKnownAssignedTransactionIds(void)
4551 : {
4552 71 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4553 CBC 71 : KnownAssignedXidsRemovePreceding(InvalidTransactionId);
4554 :
4555 ECB : /*
4556 : * Reset lastOverflowedXid. Currently, lastOverflowedXid has no use after
4557 : * the call of this function. But do this for unification with what
4558 : * ExpireOldKnownAssignedTransactionIds() do.
4559 : */
4560 GIC 71 : procArray->lastOverflowedXid = InvalidTransactionId;
4561 71 : LWLockRelease(ProcArrayLock);
4562 71 : }
4563 ECB :
4564 : /*
4565 : * ExpireOldKnownAssignedTransactionIds
4566 : * Remove KnownAssignedXids entries preceding the given XID and
4567 : * potentially reset lastOverflowedXid.
4568 : */
4569 : void
4570 GIC 195 : ExpireOldKnownAssignedTransactionIds(TransactionId xid)
4571 : {
4572 195 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4573 ECB :
4574 : /*
4575 : * Reset lastOverflowedXid if we know all transactions that have been
4576 : * possibly running are being gone. Not doing so could cause an incorrect
4577 : * lastOverflowedXid value, which makes extra snapshots be marked as
4578 : * suboverflowed.
4579 : */
4580 GIC 195 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, xid))
4581 192 : procArray->lastOverflowedXid = InvalidTransactionId;
4582 195 : KnownAssignedXidsRemovePreceding(xid);
4583 CBC 195 : LWLockRelease(ProcArrayLock);
4584 195 : }
4585 ECB :
4586 : /*
4587 : * KnownAssignedTransactionIdsIdleMaintenance
4588 : * Opportunistically do maintenance work when the startup process
4589 : * is about to go idle.
4590 : */
4591 : void
4592 GIC 12995 : KnownAssignedTransactionIdsIdleMaintenance(void)
4593 : {
4594 12995 : KnownAssignedXidsCompress(KAX_STARTUP_PROCESS_IDLE, false);
4595 CBC 12995 : }
4596 :
4597 ECB :
4598 : /*
4599 : * Private module functions to manipulate KnownAssignedXids
4600 : *
4601 : * There are 5 main uses of the KnownAssignedXids data structure:
4602 : *
4603 : * * backends taking snapshots - all valid XIDs need to be copied out
4604 : * * backends seeking to determine presence of a specific XID
4605 : * * startup process adding new known-assigned XIDs
4606 : * * startup process removing specific XIDs as transactions end
4607 : * * startup process pruning array when special WAL records arrive
4608 : *
4609 : * This data structure is known to be a hot spot during Hot Standby, so we
4610 : * go to some lengths to make these operations as efficient and as concurrent
4611 : * as possible.
4612 : *
4613 : * The XIDs are stored in an array in sorted order --- TransactionIdPrecedes
4614 : * order, to be exact --- to allow binary search for specific XIDs. Note:
4615 : * in general TransactionIdPrecedes would not provide a total order, but
4616 : * we know that the entries present at any instant should not extend across
4617 : * a large enough fraction of XID space to wrap around (the primary would
4618 : * shut down for fear of XID wrap long before that happens). So it's OK to
4619 : * use TransactionIdPrecedes as a binary-search comparator.
4620 : *
4621 : * It's cheap to maintain the sortedness during insertions, since new known
4622 : * XIDs are always reported in XID order; we just append them at the right.
4623 : *
4624 : * To keep individual deletions cheap, we need to allow gaps in the array.
4625 : * This is implemented by marking array elements as valid or invalid using
4626 : * the parallel boolean array KnownAssignedXidsValid[]. A deletion is done
4627 : * by setting KnownAssignedXidsValid[i] to false, *without* clearing the
4628 : * XID entry itself. This preserves the property that the XID entries are
4629 : * sorted, so we can do binary searches easily. Periodically we compress
4630 : * out the unused entries; that's much cheaper than having to compress the
4631 : * array immediately on every deletion.
4632 : *
4633 : * The actually valid items in KnownAssignedXids[] and KnownAssignedXidsValid[]
4634 : * are those with indexes tail <= i < head; items outside this subscript range
4635 : * have unspecified contents. When head reaches the end of the array, we
4636 : * force compression of unused entries rather than wrapping around, since
4637 : * allowing wraparound would greatly complicate the search logic. We maintain
4638 : * an explicit tail pointer so that pruning of old XIDs can be done without
4639 : * immediately moving the array contents. In most cases only a small fraction
4640 : * of the array contains valid entries at any instant.
4641 : *
4642 : * Although only the startup process can ever change the KnownAssignedXids
4643 : * data structure, we still need interlocking so that standby backends will
4644 : * not observe invalid intermediate states. The convention is that backends
4645 : * must hold shared ProcArrayLock to examine the array. To remove XIDs from
4646 : * the array, the startup process must hold ProcArrayLock exclusively, for
4647 : * the usual transactional reasons (compare commit/abort of a transaction
4648 : * during normal running). Compressing unused entries out of the array
4649 : * likewise requires exclusive lock. To add XIDs to the array, we just insert
4650 : * them into slots to the right of the head pointer and then advance the head
4651 : * pointer. This wouldn't require any lock at all, except that on machines
4652 : * with weak memory ordering we need to be careful that other processors
4653 : * see the array element changes before they see the head pointer change.
4654 : * We handle this by using a spinlock to protect reads and writes of the
4655 : * head/tail pointers. (We could dispense with the spinlock if we were to
4656 : * create suitable memory access barrier primitives and use those instead.)
4657 : * The spinlock must be taken to read or write the head/tail pointers unless
4658 : * the caller holds ProcArrayLock exclusively.
4659 : *
4660 : * Algorithmic analysis:
4661 : *
4662 : * If we have a maximum of M slots, with N XIDs currently spread across
4663 : * S elements then we have N <= S <= M always.
4664 : *
4665 : * * Adding a new XID is O(1) and needs little locking (unless compression
4666 : * must happen)
4667 : * * Compressing the array is O(S) and requires exclusive lock
4668 : * * Removing an XID is O(logS) and requires exclusive lock
4669 : * * Taking a snapshot is O(S) and requires shared lock
4670 : * * Checking for an XID is O(logS) and requires shared lock
4671 : *
4672 : * In comparison, using a hash table for KnownAssignedXids would mean that
4673 : * taking snapshots would be O(M). If we can maintain S << M then the
4674 : * sorted array technique will deliver significantly faster snapshots.
4675 : * If we try to keep S too small then we will spend too much time compressing,
4676 : * so there is an optimal point for any workload mix. We use a heuristic to
4677 : * decide when to compress the array, though trimming also helps reduce
4678 : * frequency of compressing. The heuristic requires us to track the number of
4679 : * currently valid XIDs in the array (N). Except in special cases, we'll
4680 : * compress when S >= 2N. Bounding S at 2N in turn bounds the time for
4681 : * taking a snapshot to be O(N), which it would have to be anyway.
4682 : */
4683 :
4684 :
4685 : /*
4686 : * Compress KnownAssignedXids by shifting valid data down to the start of the
4687 : * array, removing any gaps.
4688 : *
4689 : * A compression step is forced if "reason" is KAX_NO_SPACE, otherwise
4690 : * we do it only if a heuristic indicates it's a good time to do it.
4691 : *
4692 : * Compression requires holding ProcArrayLock in exclusive mode.
4693 : * Caller must pass haveLock = true if it already holds the lock.
4694 : */
4695 : static void
4696 GIC 31390 : KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock)
4697 : {
4698 31390 : ProcArrayStruct *pArray = procArray;
4699 ECB : int head,
4700 : tail,
4701 : nelements;
4702 : int compress_index;
4703 : int i;
4704 :
4705 : /* Counters for compression heuristics */
4706 : static unsigned int transactionEndsCounter;
4707 : static TimestampTz lastCompressTs;
4708 :
4709 : /* Tuning constants */
4710 : #define KAX_COMPRESS_FREQUENCY 128 /* in transactions */
4711 : #define KAX_COMPRESS_IDLE_INTERVAL 1000 /* in ms */
4712 :
4713 : /*
4714 : * Since only the startup process modifies the head/tail pointers, we
4715 : * don't need a lock to read them here.
4716 : */
4717 GIC 31390 : head = pArray->headKnownAssignedXids;
4718 31390 : tail = pArray->tailKnownAssignedXids;
4719 31390 : nelements = head - tail;
4720 ECB :
4721 : /*
4722 : * If we can choose whether to compress, use a heuristic to avoid
4723 : * compressing too often or not often enough. "Compress" here simply
4724 : * means moving the values to the beginning of the array, so it is not as
4725 : * complex or costly as typical data compression algorithms.
4726 : */
4727 GIC 31390 : if (nelements == pArray->numKnownAssignedXids)
4728 : {
4729 : /*
4730 ECB : * When there are no gaps between head and tail, don't bother to
4731 : * compress, except in the KAX_NO_SPACE case where we must compress to
4732 : * create some space after the head.
4733 : */
4734 GIC 15008 : if (reason != KAX_NO_SPACE)
4735 15008 : return;
4736 : }
4737 CBC 16382 : else if (reason == KAX_TRANSACTION_END)
4738 ECB : {
4739 : /*
4740 : * Consider compressing only once every so many commits. Frequency
4741 : * determined by benchmarks.
4742 : */
4743 GIC 10929 : if ((transactionEndsCounter++) % KAX_COMPRESS_FREQUENCY != 0)
4744 10835 : return;
4745 :
4746 ECB : /*
4747 : * Furthermore, compress only if the used part of the array is less
4748 : * than 50% full (see comments above).
4749 : */
4750 GIC 94 : if (nelements < 2 * pArray->numKnownAssignedXids)
4751 3 : return;
4752 : }
4753 CBC 5453 : else if (reason == KAX_STARTUP_PROCESS_IDLE)
4754 ECB : {
4755 : /*
4756 : * We're about to go idle for lack of new WAL, so we might as well
4757 : * compress. But not too often, to avoid ProcArray lock contention
4758 : * with readers.
4759 : */
4760 GIC 5447 : if (lastCompressTs != 0)
4761 : {
4762 : TimestampTz compress_after;
4763 ECB :
4764 GIC 5447 : compress_after = TimestampTzPlusMilliseconds(lastCompressTs,
4765 : KAX_COMPRESS_IDLE_INTERVAL);
4766 5447 : if (GetCurrentTimestamp() < compress_after)
4767 CBC 5430 : return;
4768 : }
4769 ECB : }
4770 :
4771 : /* Need to compress, so get the lock if we don't have it. */
4772 GIC 114 : if (!haveLock)
4773 17 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4774 :
4775 ECB : /*
4776 : * We compress the array by reading the valid values from tail to head,
4777 : * re-aligning data to 0th element.
4778 : */
4779 GIC 114 : compress_index = 0;
4780 6592 : for (i = tail; i < head; i++)
4781 : {
4782 CBC 6478 : if (KnownAssignedXidsValid[i])
4783 ECB : {
4784 GIC 341 : KnownAssignedXids[compress_index] = KnownAssignedXids[i];
4785 CBC 341 : KnownAssignedXidsValid[compress_index] = true;
4786 GIC 341 : compress_index++;
4787 ECB : }
4788 : }
4789 CBC 114 : Assert(compress_index == pArray->numKnownAssignedXids);
4790 :
4791 GIC 114 : pArray->tailKnownAssignedXids = 0;
4792 CBC 114 : pArray->headKnownAssignedXids = compress_index;
4793 :
4794 114 : if (!haveLock)
4795 17 : LWLockRelease(ProcArrayLock);
4796 :
4797 ECB : /* Update timestamp for maintenance. No need to hold lock for this. */
4798 CBC 114 : lastCompressTs = GetCurrentTimestamp();
4799 : }
4800 :
4801 ECB : /*
4802 : * Add xids into KnownAssignedXids at the head of the array.
4803 : *
4804 : * xids from from_xid to to_xid, inclusive, are added to the array.
4805 : *
4806 : * If exclusive_lock is true then caller already holds ProcArrayLock in
4807 : * exclusive mode, so we need no extra locking here. Else caller holds no
4808 : * lock, so we need to be sure we maintain sufficient interlocks against
4809 : * concurrent readers. (Only the startup process ever calls this, so no need
4810 : * to worry about concurrent writers.)
4811 : */
4812 : static void
4813 GIC 19414 : KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
4814 : bool exclusive_lock)
4815 : {
4816 CBC 19414 : ProcArrayStruct *pArray = procArray;
4817 : TransactionId next_xid;
4818 : int head,
4819 ECB : tail;
4820 : int nxids;
4821 : int i;
4822 :
4823 GIC 19414 : Assert(TransactionIdPrecedesOrEquals(from_xid, to_xid));
4824 :
4825 : /*
4826 ECB : * Calculate how many array slots we'll need. Normally this is cheap; in
4827 : * the unusual case where the XIDs cross the wrap point, we do it the hard
4828 : * way.
4829 : */
4830 GIC 19414 : if (to_xid >= from_xid)
4831 19414 : nxids = to_xid - from_xid + 1;
4832 : else
4833 ECB : {
4834 LBC 0 : nxids = 1;
4835 UIC 0 : next_xid = from_xid;
4836 0 : while (TransactionIdPrecedes(next_xid, to_xid))
4837 EUB : {
4838 UBC 0 : nxids++;
4839 0 : TransactionIdAdvance(next_xid);
4840 : }
4841 EUB : }
4842 :
4843 : /*
4844 : * Since only the startup process modifies the head/tail pointers, we
4845 : * don't need a lock to read them here.
4846 : */
4847 GIC 19414 : head = pArray->headKnownAssignedXids;
4848 19414 : tail = pArray->tailKnownAssignedXids;
4849 :
4850 CBC 19414 : Assert(head >= 0 && head <= pArray->maxKnownAssignedXids);
4851 19414 : Assert(tail >= 0 && tail < pArray->maxKnownAssignedXids);
4852 :
4853 ECB : /*
4854 : * Verify that insertions occur in TransactionId sequence. Note that even
4855 : * if the last existing element is marked invalid, it must still have a
4856 : * correctly sequenced XID value.
4857 : */
4858 GIC 32320 : if (head > tail &&
4859 12906 : TransactionIdFollowsOrEquals(KnownAssignedXids[head - 1], from_xid))
4860 : {
4861 LBC 0 : KnownAssignedXidsDisplay(LOG);
4862 0 : elog(ERROR, "out-of-order XID insertion in KnownAssignedXids");
4863 : }
4864 EUB :
4865 : /*
4866 : * If our xids won't fit in the remaining space, compress out free space
4867 : */
4868 GIC 19414 : if (head + nxids > pArray->maxKnownAssignedXids)
4869 : {
4870 UIC 0 : KnownAssignedXidsCompress(KAX_NO_SPACE, exclusive_lock);
4871 ECB :
4872 UIC 0 : head = pArray->headKnownAssignedXids;
4873 EUB : /* note: we no longer care about the tail pointer */
4874 :
4875 : /*
4876 : * If it still won't fit then we're out of memory
4877 : */
4878 UIC 0 : if (head + nxids > pArray->maxKnownAssignedXids)
4879 0 : elog(ERROR, "too many KnownAssignedXids");
4880 : }
4881 EUB :
4882 : /* Now we can insert the xids into the space starting at head */
4883 GIC 19414 : next_xid = from_xid;
4884 39116 : for (i = 0; i < nxids; i++)
4885 : {
4886 CBC 19702 : KnownAssignedXids[head] = next_xid;
4887 19702 : KnownAssignedXidsValid[head] = true;
4888 GIC 19702 : TransactionIdAdvance(next_xid);
4889 CBC 19702 : head++;
4890 ECB : }
4891 :
4892 : /* Adjust count of number of valid entries */
4893 GIC 19414 : pArray->numKnownAssignedXids += nxids;
4894 :
4895 : /*
4896 ECB : * Now update the head pointer. We use a spinlock to protect this
4897 : * pointer, not because the update is likely to be non-atomic, but to
4898 : * ensure that other processors see the above array updates before they
4899 : * see the head pointer change.
4900 : *
4901 : * If we're holding ProcArrayLock exclusively, there's no need to take the
4902 : * spinlock.
4903 : */
4904 GIC 19414 : if (exclusive_lock)
4905 2 : pArray->headKnownAssignedXids = head;
4906 : else
4907 ECB : {
4908 CBC 19412 : SpinLockAcquire(&pArray->known_assigned_xids_lck);
4909 GIC 19412 : pArray->headKnownAssignedXids = head;
4910 19412 : SpinLockRelease(&pArray->known_assigned_xids_lck);
4911 ECB : }
4912 CBC 19414 : }
4913 ECB :
4914 : /*
4915 : * KnownAssignedXidsSearch
4916 : *
4917 : * Searches KnownAssignedXids for a specific xid and optionally removes it.
4918 : * Returns true if it was found, false if not.
4919 : *
4920 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4921 : * Exclusive lock must be held for remove = true.
4922 : */
4923 : static bool
4924 GIC 20841 : KnownAssignedXidsSearch(TransactionId xid, bool remove)
4925 : {
4926 20841 : ProcArrayStruct *pArray = procArray;
4927 ECB : int first,
4928 : last;
4929 : int head;
4930 : int tail;
4931 GIC 20841 : int result_index = -1;
4932 :
4933 20841 : if (remove)
4934 ECB : {
4935 : /* we hold ProcArrayLock exclusively, so no need for spinlock */
4936 CBC 20841 : tail = pArray->tailKnownAssignedXids;
4937 GIC 20841 : head = pArray->headKnownAssignedXids;
4938 : }
4939 ECB : else
4940 : {
4941 : /* take spinlock to ensure we see up-to-date array contents */
4942 UIC 0 : SpinLockAcquire(&pArray->known_assigned_xids_lck);
4943 0 : tail = pArray->tailKnownAssignedXids;
4944 0 : head = pArray->headKnownAssignedXids;
4945 UBC 0 : SpinLockRelease(&pArray->known_assigned_xids_lck);
4946 EUB : }
4947 :
4948 : /*
4949 : * Standard binary search. Note we can ignore the KnownAssignedXidsValid
4950 : * array here, since even invalid entries will contain sorted XIDs.
4951 : */
4952 GIC 20841 : first = tail;
4953 20841 : last = head - 1;
4954 75440 : while (first <= last)
4955 ECB : {
4956 : int mid_index;
4957 : TransactionId mid_xid;
4958 :
4959 GIC 74249 : mid_index = (first + last) / 2;
4960 74249 : mid_xid = KnownAssignedXids[mid_index];
4961 :
4962 CBC 74249 : if (xid == mid_xid)
4963 ECB : {
4964 GIC 19650 : result_index = mid_index;
4965 CBC 19650 : break;
4966 : }
4967 54599 : else if (TransactionIdPrecedes(xid, mid_xid))
4968 11139 : last = mid_index - 1;
4969 : else
4970 43460 : first = mid_index + 1;
4971 ECB : }
4972 :
4973 CBC 20841 : if (result_index < 0)
4974 GIC 1191 : return false; /* not in array */
4975 :
4976 CBC 19650 : if (!KnownAssignedXidsValid[result_index])
4977 30 : return false; /* in array, but invalid */
4978 :
4979 19620 : if (remove)
4980 ECB : {
4981 GIC 19620 : KnownAssignedXidsValid[result_index] = false;
4982 ECB :
4983 GIC 19620 : pArray->numKnownAssignedXids--;
4984 CBC 19620 : Assert(pArray->numKnownAssignedXids >= 0);
4985 :
4986 ECB : /*
4987 : * If we're removing the tail element then advance tail pointer over
4988 : * any invalid elements. This will speed future searches.
4989 : */
4990 GIC 19620 : if (result_index == tail)
4991 : {
4992 7694 : tail++;
4993 CBC 13483 : while (tail < head && !KnownAssignedXidsValid[tail])
4994 GIC 5789 : tail++;
4995 CBC 7694 : if (tail >= head)
4996 ECB : {
4997 : /* Array is empty, so we can reset both pointers */
4998 CBC 6501 : pArray->headKnownAssignedXids = 0;
4999 GIC 6501 : pArray->tailKnownAssignedXids = 0;
5000 : }
5001 ECB : else
5002 : {
5003 GIC 1193 : pArray->tailKnownAssignedXids = tail;
5004 : }
5005 : }
5006 ECB : }
5007 :
5008 GIC 19620 : return true;
5009 : }
5010 :
5011 ECB : /*
5012 : * Is the specified XID present in KnownAssignedXids[]?
5013 : *
5014 : * Caller must hold ProcArrayLock in shared or exclusive mode.
5015 : */
5016 : static bool
5017 UIC 0 : KnownAssignedXidExists(TransactionId xid)
5018 : {
5019 0 : Assert(TransactionIdIsValid(xid));
5020 EUB :
5021 UIC 0 : return KnownAssignedXidsSearch(xid, false);
5022 EUB : }
5023 :
5024 : /*
5025 : * Remove the specified XID from KnownAssignedXids[].
5026 : *
5027 : * Caller must hold ProcArrayLock in exclusive mode.
5028 : */
5029 : static void
5030 GIC 20841 : KnownAssignedXidsRemove(TransactionId xid)
5031 : {
5032 20841 : Assert(TransactionIdIsValid(xid));
5033 ECB :
5034 GIC 20841 : elog(trace_recovery(DEBUG4), "remove KnownAssignedXid %u", xid);
5035 ECB :
5036 : /*
5037 : * Note: we cannot consider it an error to remove an XID that's not
5038 : * present. We intentionally remove subxact IDs while processing
5039 : * XLOG_XACT_ASSIGNMENT, to avoid array overflow. Then those XIDs will be
5040 : * removed again when the top-level xact commits or aborts.
5041 : *
5042 : * It might be possible to track such XIDs to distinguish this case from
5043 : * actual errors, but it would be complicated and probably not worth it.
5044 : * So, just ignore the search result.
5045 : */
5046 GIC 20841 : (void) KnownAssignedXidsSearch(xid, true);
5047 20841 : }
5048 :
5049 ECB : /*
5050 : * KnownAssignedXidsRemoveTree
5051 : * Remove xid (if it's not InvalidTransactionId) and all the subxids.
5052 : *
5053 : * Caller must hold ProcArrayLock in exclusive mode.
5054 : */
5055 : static void
5056 GIC 18200 : KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
5057 : TransactionId *subxids)
5058 : {
5059 ECB : int i;
5060 :
5061 GIC 18200 : if (TransactionIdIsValid(xid))
5062 18179 : KnownAssignedXidsRemove(xid);
5063 :
5064 CBC 20862 : for (i = 0; i < nsubxids; i++)
5065 2662 : KnownAssignedXidsRemove(subxids[i]);
5066 :
5067 ECB : /* Opportunistically compress the array */
5068 CBC 18200 : KnownAssignedXidsCompress(KAX_TRANSACTION_END, true);
5069 GIC 18200 : }
5070 :
5071 ECB : /*
5072 : * Prune KnownAssignedXids up to, but *not* including xid. If xid is invalid
5073 : * then clear the whole table.
5074 : *
5075 : * Caller must hold ProcArrayLock in exclusive mode.
5076 : */
5077 : static void
5078 GIC 266 : KnownAssignedXidsRemovePreceding(TransactionId removeXid)
5079 : {
5080 266 : ProcArrayStruct *pArray = procArray;
5081 CBC 266 : int count = 0;
5082 : int head,
5083 ECB : tail,
5084 : i;
5085 :
5086 GIC 266 : if (!TransactionIdIsValid(removeXid))
5087 : {
5088 71 : elog(trace_recovery(DEBUG4), "removing all KnownAssignedXids");
5089 CBC 71 : pArray->numKnownAssignedXids = 0;
5090 GIC 71 : pArray->headKnownAssignedXids = pArray->tailKnownAssignedXids = 0;
5091 CBC 71 : return;
5092 ECB : }
5093 :
5094 CBC 195 : elog(trace_recovery(DEBUG4), "prune KnownAssignedXids to %u", removeXid);
5095 :
5096 : /*
5097 ECB : * Mark entries invalid starting at the tail. Since array is sorted, we
5098 : * can stop as soon as we reach an entry >= removeXid.
5099 : */
5100 GIC 195 : tail = pArray->tailKnownAssignedXids;
5101 195 : head = pArray->headKnownAssignedXids;
5102 :
5103 CBC 195 : for (i = tail; i < head; i++)
5104 ECB : {
5105 GIC 35 : if (KnownAssignedXidsValid[i])
5106 ECB : {
5107 GIC 35 : TransactionId knownXid = KnownAssignedXids[i];
5108 ECB :
5109 GIC 35 : if (TransactionIdFollowsOrEquals(knownXid, removeXid))
5110 CBC 35 : break;
5111 :
5112 LBC 0 : if (!StandbyTransactionIdIsPrepared(knownXid))
5113 ECB : {
5114 UIC 0 : KnownAssignedXidsValid[i] = false;
5115 UBC 0 : count++;
5116 : }
5117 EUB : }
5118 : }
5119 :
5120 GIC 195 : pArray->numKnownAssignedXids -= count;
5121 195 : Assert(pArray->numKnownAssignedXids >= 0);
5122 :
5123 ECB : /*
5124 : * Advance the tail pointer if we've marked the tail item invalid.
5125 : */
5126 GIC 195 : for (i = tail; i < head; i++)
5127 : {
5128 35 : if (KnownAssignedXidsValid[i])
5129 CBC 35 : break;
5130 : }
5131 195 : if (i >= head)
5132 ECB : {
5133 : /* Array is empty, so we can reset both pointers */
5134 CBC 160 : pArray->headKnownAssignedXids = 0;
5135 GIC 160 : pArray->tailKnownAssignedXids = 0;
5136 : }
5137 ECB : else
5138 : {
5139 GIC 35 : pArray->tailKnownAssignedXids = i;
5140 : }
5141 :
5142 ECB : /* Opportunistically compress the array */
5143 GIC 195 : KnownAssignedXidsCompress(KAX_PRUNE, true);
5144 : }
5145 :
5146 ECB : /*
5147 : * KnownAssignedXidsGet - Get an array of xids by scanning KnownAssignedXids.
5148 : * We filter out anything >= xmax.
5149 : *
5150 : * Returns the number of XIDs stored into xarray[]. Caller is responsible
5151 : * that array is large enough.
5152 : *
5153 : * Caller must hold ProcArrayLock in (at least) shared mode.
5154 : */
5155 : static int
5156 UIC 0 : KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax)
5157 : {
5158 0 : TransactionId xtmp = InvalidTransactionId;
5159 EUB :
5160 UIC 0 : return KnownAssignedXidsGetAndSetXmin(xarray, &xtmp, xmax);
5161 EUB : }
5162 :
5163 : /*
5164 : * KnownAssignedXidsGetAndSetXmin - as KnownAssignedXidsGet, plus
5165 : * we reduce *xmin to the lowest xid value seen if not already lower.
5166 : *
5167 : * Caller must hold ProcArrayLock in (at least) shared mode.
5168 : */
5169 : static int
5170 GIC 826 : KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, TransactionId *xmin,
5171 : TransactionId xmax)
5172 : {
5173 CBC 826 : int count = 0;
5174 : int head,
5175 : tail;
5176 ECB : int i;
5177 :
5178 : /*
5179 : * Fetch head just once, since it may change while we loop. We can stop
5180 : * once we reach the initially seen head, since we are certain that an xid
5181 : * cannot enter and then leave the array while we hold ProcArrayLock. We
5182 : * might miss newly-added xids, but they should be >= xmax so irrelevant
5183 : * anyway.
5184 : *
5185 : * Must take spinlock to ensure we see up-to-date array contents.
5186 : */
5187 GIC 826 : SpinLockAcquire(&procArray->known_assigned_xids_lck);
5188 826 : tail = procArray->tailKnownAssignedXids;
5189 826 : head = procArray->headKnownAssignedXids;
5190 CBC 826 : SpinLockRelease(&procArray->known_assigned_xids_lck);
5191 ECB :
5192 CBC 845 : for (i = tail; i < head; i++)
5193 ECB : {
5194 : /* Skip any gaps in the array */
5195 CBC 58 : if (KnownAssignedXidsValid[i])
5196 : {
5197 GIC 52 : TransactionId knownXid = KnownAssignedXids[i];
5198 ECB :
5199 : /*
5200 : * Update xmin if required. Only the first XID need be checked,
5201 : * since the array is sorted.
5202 : */
5203 GIC 104 : if (count == 0 &&
5204 52 : TransactionIdPrecedes(knownXid, *xmin))
5205 13 : *xmin = knownXid;
5206 ECB :
5207 : /*
5208 : * Filter out anything >= xmax, again relying on sorted property
5209 : * of array.
5210 : */
5211 GIC 104 : if (TransactionIdIsValid(xmax) &&
5212 52 : TransactionIdFollowsOrEquals(knownXid, xmax))
5213 39 : break;
5214 ECB :
5215 : /* Add knownXid into output array */
5216 CBC 13 : xarray[count++] = knownXid;
5217 : }
5218 : }
5219 ECB :
5220 GIC 826 : return count;
5221 : }
5222 :
5223 ECB : /*
5224 : * Get oldest XID in the KnownAssignedXids array, or InvalidTransactionId
5225 : * if nothing there.
5226 : */
5227 : static TransactionId
5228 GIC 116 : KnownAssignedXidsGetOldestXmin(void)
5229 : {
5230 : int head,
5231 ECB : tail;
5232 : int i;
5233 :
5234 : /*
5235 : * Fetch head just once, since it may change while we loop.
5236 : */
5237 GIC 116 : SpinLockAcquire(&procArray->known_assigned_xids_lck);
5238 116 : tail = procArray->tailKnownAssignedXids;
5239 116 : head = procArray->headKnownAssignedXids;
5240 CBC 116 : SpinLockRelease(&procArray->known_assigned_xids_lck);
5241 ECB :
5242 CBC 116 : for (i = tail; i < head; i++)
5243 ECB : {
5244 : /* Skip any gaps in the array */
5245 CBC 5 : if (KnownAssignedXidsValid[i])
5246 GIC 5 : return KnownAssignedXids[i];
5247 : }
5248 ECB :
5249 CBC 111 : return InvalidTransactionId;
5250 : }
5251 :
5252 ECB : /*
5253 : * Display KnownAssignedXids to provide debug trail
5254 : *
5255 : * Currently this is only called within startup process, so we need no
5256 : * special locking.
5257 : *
5258 : * Note this is pretty expensive, and much of the expense will be incurred
5259 : * even if the elog message will get discarded. It's not currently called
5260 : * in any performance-critical places, however, so no need to be tenser.
5261 : */
5262 : static void
5263 GIC 73 : KnownAssignedXidsDisplay(int trace_level)
5264 : {
5265 73 : ProcArrayStruct *pArray = procArray;
5266 ECB : StringInfoData buf;
5267 : int head,
5268 : tail,
5269 : i;
5270 GIC 73 : int nxids = 0;
5271 :
5272 73 : tail = pArray->tailKnownAssignedXids;
5273 CBC 73 : head = pArray->headKnownAssignedXids;
5274 :
5275 73 : initStringInfo(&buf);
5276 ECB :
5277 GIC 77 : for (i = tail; i < head; i++)
5278 ECB : {
5279 GIC 4 : if (KnownAssignedXidsValid[i])
5280 ECB : {
5281 GIC 4 : nxids++;
5282 CBC 4 : appendStringInfo(&buf, "[%d]=%u ", i, KnownAssignedXids[i]);
5283 : }
5284 ECB : }
5285 :
5286 GIC 73 : elog(trace_level, "%d KnownAssignedXids (num=%d tail=%d head=%d) %s",
5287 : nxids,
5288 : pArray->numKnownAssignedXids,
5289 ECB : pArray->tailKnownAssignedXids,
5290 : pArray->headKnownAssignedXids,
5291 : buf.data);
5292 :
5293 GIC 73 : pfree(buf.data);
5294 73 : }
5295 :
5296 ECB : /*
5297 : * KnownAssignedXidsReset
5298 : * Resets KnownAssignedXids to be empty
5299 : */
5300 : static void
5301 UIC 0 : KnownAssignedXidsReset(void)
5302 : {
5303 0 : ProcArrayStruct *pArray = procArray;
5304 EUB :
5305 UIC 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
5306 EUB :
5307 UIC 0 : pArray->numKnownAssignedXids = 0;
5308 UBC 0 : pArray->tailKnownAssignedXids = 0;
5309 UIC 0 : pArray->headKnownAssignedXids = 0;
5310 EUB :
5311 UBC 0 : LWLockRelease(ProcArrayLock);
5312 0 : }
|
