1 /*-------------------------------------------------------------------------
2 *
3 * predicate.c
4 * POSTGRES predicate locking
5 * to support full serializable transaction isolation
6 *
7 *
8 * The approach taken is to implement Serializable Snapshot Isolation (SSI)
9 * as initially described in this paper:
10 *
11 * Michael J. Cahill, Uwe Röhm, and Alan D. Fekete. 2008.
12 * Serializable isolation for snapshot databases.
13 * In SIGMOD '08: Proceedings of the 2008 ACM SIGMOD
14 * international conference on Management of data,
15 * pages 729-738, New York, NY, USA. ACM.
16 * http://doi.acm.org/10.1145/1376616.1376690
17 *
18 * and further elaborated in Cahill's doctoral thesis:
19 *
20 * Michael James Cahill. 2009.
21 * Serializable Isolation for Snapshot Databases.
22 * Sydney Digital Theses.
23 * University of Sydney, School of Information Technologies.
24 * http://hdl.handle.net/2123/5353
25 *
26 *
27 * Predicate locks for Serializable Snapshot Isolation (SSI) are SIREAD
28 * locks, which are so different from normal locks that a distinct set of
29 * structures is required to handle them. They are needed to detect
30 * rw-conflicts when the read happens before the write. (When the write
31 * occurs first, the reading transaction can check for a conflict by
32 * examining the MVCC data.)
33 *
34 * (1) Besides tuples actually read, they must cover ranges of tuples
35 * which would have been read based on the predicate. This will
36 * require modelling the predicates through locks against database
37 * objects such as pages, index ranges, or entire tables.
38 *
39 * (2) They must be kept in RAM for quick access. Because of this, it
40 * isn't possible to always maintain tuple-level granularity -- when
41 * the space allocated to store these approaches exhaustion, a
42 * request for a lock may need to scan for situations where a single
43 * transaction holds many fine-grained locks which can be coalesced
44 * into a single coarser-grained lock.
45 *
46 * (3) They never block anything; they are more like flags than locks
47 * in that regard; although they refer to database objects and are
48 * used to identify rw-conflicts with normal write locks.
49 *
50 * (4) While they are associated with a transaction, they must survive
51 * a successful COMMIT of that transaction, and remain until all
52 * overlapping transactions complete. This even means that they
53 * must survive termination of the transaction's process. If a
54 * top level transaction is rolled back, however, it is immediately
55 * flagged so that it can be ignored, and its SIREAD locks can be
56 * released any time after that.
57 *
58 * (5) The only transactions which create SIREAD locks or check for
59 * conflicts with them are serializable transactions.
60 *
61 * (6) When a write lock for a top level transaction is found to cover
62 * an existing SIREAD lock for the same transaction, the SIREAD lock
63 * can be deleted.
64 *
65 * (7) A write from a serializable transaction must ensure that an xact
66 * record exists for the transaction, with the same lifespan (until
67 * all concurrent transaction complete or the transaction is rolled
68 * back) so that rw-dependencies to that transaction can be
69 * detected.
70 *
71 * We use an optimization for read-only transactions. Under certain
72 * circumstances, a read-only transaction's snapshot can be shown to
73 * never have conflicts with other transactions. This is referred to
74 * as a "safe" snapshot (and one known not to be is "unsafe").
75 * However, it can't be determined whether a snapshot is safe until
76 * all concurrent read/write transactions complete.
77 *
78 * Once a read-only transaction is known to have a safe snapshot, it
79 * can release its predicate locks and exempt itself from further
80 * predicate lock tracking. READ ONLY DEFERRABLE transactions run only
81 * on safe snapshots, waiting as necessary for one to be available.
82 *
83 *
84 * Lightweight locks to manage access to the predicate locking shared
85 * memory objects must be taken in this order, and should be released in
86 * reverse order:
87 *
88 * SerializableFinishedListLock
89 * - Protects the list of transactions which have completed but which
90 * may yet matter because they overlap still-active transactions.
91 *
92 * SerializablePredicateListLock
93 * - Protects the linked list of locks held by a transaction. Note
94 * that the locks themselves are also covered by the partition
95 * locks of their respective lock targets; this lock only affects
96 * the linked list connecting the locks related to a transaction.
97 * - All transactions share this single lock (with no partitioning).
98 * - There is never a need for a process other than the one running
99 * an active transaction to walk the list of locks held by that
100 * transaction, except parallel query workers sharing the leader's
101 * transaction. In the parallel case, an extra per-sxact lock is
102 * taken; see below.
103 * - It is relatively infrequent that another process needs to
104 * modify the list for a transaction, but it does happen for such
105 * things as index page splits for pages with predicate locks and
106 * freeing of predicate locked pages by a vacuum process. When
107 * removing a lock in such cases, the lock itself contains the
108 * pointers needed to remove it from the list. When adding a
109 * lock in such cases, the lock can be added using the anchor in
110 * the transaction structure. Neither requires walking the list.
111 * - Cleaning up the list for a terminated transaction is sometimes
112 * not done on a retail basis, in which case no lock is required.
113 * - Due to the above, a process accessing its active transaction's
114 * list always uses a shared lock, regardless of whether it is
115 * walking or maintaining the list. This improves concurrency
116 * for the common access patterns.
117 * - A process which needs to alter the list of a transaction other
118 * than its own active transaction must acquire an exclusive
119 * lock.
120 *
121 * SERIALIZABLEXACT's member 'perXactPredicateListLock'
122 * - Protects the linked list of predicate locks held by a transaction.
123 * Only needed for parallel mode, where multiple backends share the
124 * same SERIALIZABLEXACT object. Not needed if
125 * SerializablePredicateListLock is held exclusively.
126 *
127 * PredicateLockHashPartitionLock(hashcode)
128 * - The same lock protects a target, all locks on that target, and
129 * the linked list of locks on the target.
130 * - When more than one is needed, acquire in ascending address order.
131 * - When all are needed (rare), acquire in ascending index order with
132 * PredicateLockHashPartitionLockByIndex(index).
133 *
134 * SerializableXactHashLock
135 * - Protects both PredXact and SerializableXidHash.
136 *
137 *
138 * Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
139 * Portions Copyright (c) 1994, Regents of the University of California
140 *
141 *
142 * IDENTIFICATION
143 * src/backend/storage/lmgr/predicate.c
144 *
145 *-------------------------------------------------------------------------
146 */
147 /*
148 * INTERFACE ROUTINES
149 *
150 * housekeeping for setting up shared memory predicate lock structures
151 * InitPredicateLocks(void)
152 * PredicateLockShmemSize(void)
153 *
154 * predicate lock reporting
155 * GetPredicateLockStatusData(void)
156 * PageIsPredicateLocked(Relation relation, BlockNumber blkno)
157 *
158 * predicate lock maintenance
159 * GetSerializableTransactionSnapshot(Snapshot snapshot)
160 * SetSerializableTransactionSnapshot(Snapshot snapshot,
161 * VirtualTransactionId *sourcevxid)
162 * RegisterPredicateLockingXid(void)
163 * PredicateLockRelation(Relation relation, Snapshot snapshot)
164 * PredicateLockPage(Relation relation, BlockNumber blkno,
165 * Snapshot snapshot)
166 * PredicateLockTID(Relation relation, ItemPointer tid, Snapshot snapshot,
167 * TransactionId insert_xid)
168 * PredicateLockPageSplit(Relation relation, BlockNumber oldblkno,
169 * BlockNumber newblkno)
170 * PredicateLockPageCombine(Relation relation, BlockNumber oldblkno,
171 * BlockNumber newblkno)
172 * TransferPredicateLocksToHeapRelation(Relation relation)
173 * ReleasePredicateLocks(bool isCommit, bool isReadOnlySafe)
174 *
175 * conflict detection (may also trigger rollback)
176 * CheckForSerializableConflictOut(Relation relation, TransactionId xid,
177 * Snapshot snapshot)
178 * CheckForSerializableConflictIn(Relation relation, ItemPointer tid,
179 * BlockNumber blkno)
180 * CheckTableForSerializableConflictIn(Relation relation)
181 *
182 * final rollback checking
183 * PreCommit_CheckForSerializationFailure(void)
184 *
185 * two-phase commit support
186 * AtPrepare_PredicateLocks(void);
187 * PostPrepare_PredicateLocks(TransactionId xid);
188 * PredicateLockTwoPhaseFinish(TransactionId xid, bool isCommit);
189 * predicatelock_twophase_recover(TransactionId xid, uint16 info,
190 * void *recdata, uint32 len);
191 */
192
193 #include "postgres.h"
194
195 #include "access/parallel.h"
196 #include "access/slru.h"
197 #include "access/subtrans.h"
198 #include "access/transam.h"
199 #include "access/twophase.h"
200 #include "access/twophase_rmgr.h"
201 #include "access/xact.h"
202 #include "access/xlog.h"
203 #include "miscadmin.h"
204 #include "pgstat.h"
205 #include "storage/bufmgr.h"
206 #include "storage/predicate.h"
207 #include "storage/predicate_internals.h"
208 #include "storage/proc.h"
209 #include "storage/procarray.h"
210 #include "utils/rel.h"
211 #include "utils/snapmgr.h"
212
213 /* Uncomment the next line to test the graceful degradation code. */
214 /* #define TEST_SUMMARIZE_SERIAL */
215
216 /*
217 * Test the most selective fields first, for performance.
218 *
219 * a is covered by b if all of the following hold:
220 * 1) a.database = b.database
221 * 2) a.relation = b.relation
222 * 3) b.offset is invalid (b is page-granularity or higher)
223 * 4) either of the following:
224 * 4a) a.offset is valid (a is tuple-granularity) and a.page = b.page
225 * or 4b) a.offset is invalid and b.page is invalid (a is
226 * page-granularity and b is relation-granularity
227 */
228 #define TargetTagIsCoveredBy(covered_target, covering_target) \
229 ((GET_PREDICATELOCKTARGETTAG_RELATION(covered_target) == /* (2) */ \
230 GET_PREDICATELOCKTARGETTAG_RELATION(covering_target)) \
231 && (GET_PREDICATELOCKTARGETTAG_OFFSET(covering_target) == \
232 InvalidOffsetNumber) /* (3) */ \
233 && (((GET_PREDICATELOCKTARGETTAG_OFFSET(covered_target) != \
234 InvalidOffsetNumber) /* (4a) */ \
235 && (GET_PREDICATELOCKTARGETTAG_PAGE(covering_target) == \
236 GET_PREDICATELOCKTARGETTAG_PAGE(covered_target))) \
237 || ((GET_PREDICATELOCKTARGETTAG_PAGE(covering_target) == \
238 InvalidBlockNumber) /* (4b) */ \
239 && (GET_PREDICATELOCKTARGETTAG_PAGE(covered_target) \
240 != InvalidBlockNumber))) \
241 && (GET_PREDICATELOCKTARGETTAG_DB(covered_target) == /* (1) */ \
242 GET_PREDICATELOCKTARGETTAG_DB(covering_target)))
243
244 /*
245 * The predicate locking target and lock shared hash tables are partitioned to
246 * reduce contention. To determine which partition a given target belongs to,
247 * compute the tag's hash code with PredicateLockTargetTagHashCode(), then
248 * apply one of these macros.
249 * NB: NUM_PREDICATELOCK_PARTITIONS must be a power of 2!
250 */
251 #define PredicateLockHashPartition(hashcode) \
252 ((hashcode) % NUM_PREDICATELOCK_PARTITIONS)
253 #define PredicateLockHashPartitionLock(hashcode) \
254 (&MainLWLockArray[PREDICATELOCK_MANAGER_LWLOCK_OFFSET + \
255 PredicateLockHashPartition(hashcode)].lock)
256 #define PredicateLockHashPartitionLockByIndex(i) \
257 (&MainLWLockArray[PREDICATELOCK_MANAGER_LWLOCK_OFFSET + (i)].lock)
258
259 #define NPREDICATELOCKTARGETENTS() \
260 mul_size(max_predicate_locks_per_xact, add_size(MaxBackends, max_prepared_xacts))
261
262 #define SxactIsOnFinishedList(sxact) (!SHMQueueIsDetached(&((sxact)->finishedLink)))
263
264 /*
265 * Note that a sxact is marked "prepared" once it has passed
266 * PreCommit_CheckForSerializationFailure, even if it isn't using
267 * 2PC. This is the point at which it can no longer be aborted.
268 *
269 * The PREPARED flag remains set after commit, so SxactIsCommitted
270 * implies SxactIsPrepared.
271 */
272 #define SxactIsCommitted(sxact) (((sxact)->flags & SXACT_FLAG_COMMITTED) != 0)
273 #define SxactIsPrepared(sxact) (((sxact)->flags & SXACT_FLAG_PREPARED) != 0)
274 #define SxactIsRolledBack(sxact) (((sxact)->flags & SXACT_FLAG_ROLLED_BACK) != 0)
275 #define SxactIsDoomed(sxact) (((sxact)->flags & SXACT_FLAG_DOOMED) != 0)
276 #define SxactIsReadOnly(sxact) (((sxact)->flags & SXACT_FLAG_READ_ONLY) != 0)
277 #define SxactHasSummaryConflictIn(sxact) (((sxact)->flags & SXACT_FLAG_SUMMARY_CONFLICT_IN) != 0)
278 #define SxactHasSummaryConflictOut(sxact) (((sxact)->flags & SXACT_FLAG_SUMMARY_CONFLICT_OUT) != 0)
279 /*
280 * The following macro actually means that the specified transaction has a
281 * conflict out *to a transaction which committed ahead of it*. It's hard
282 * to get that into a name of a reasonable length.
283 */
284 #define SxactHasConflictOut(sxact) (((sxact)->flags & SXACT_FLAG_CONFLICT_OUT) != 0)
285 #define SxactIsDeferrableWaiting(sxact) (((sxact)->flags & SXACT_FLAG_DEFERRABLE_WAITING) != 0)
286 #define SxactIsROSafe(sxact) (((sxact)->flags & SXACT_FLAG_RO_SAFE) != 0)
287 #define SxactIsROUnsafe(sxact) (((sxact)->flags & SXACT_FLAG_RO_UNSAFE) != 0)
288 #define SxactIsPartiallyReleased(sxact) (((sxact)->flags & SXACT_FLAG_PARTIALLY_RELEASED) != 0)
289
290 /*
291 * Compute the hash code associated with a PREDICATELOCKTARGETTAG.
292 *
293 * To avoid unnecessary recomputations of the hash code, we try to do this
294 * just once per function, and then pass it around as needed. Aside from
295 * passing the hashcode to hash_search_with_hash_value(), we can extract
296 * the lock partition number from the hashcode.
297 */
298 #define PredicateLockTargetTagHashCode(predicatelocktargettag) \
299 get_hash_value(PredicateLockTargetHash, predicatelocktargettag)
300
301 /*
302 * Given a predicate lock tag, and the hash for its target,
303 * compute the lock hash.
304 *
305 * To make the hash code also depend on the transaction, we xor the sxid
306 * struct's address into the hash code, left-shifted so that the
307 * partition-number bits don't change. Since this is only a hash, we
308 * don't care if we lose high-order bits of the address; use an
309 * intermediate variable to suppress cast-pointer-to-int warnings.
310 */
311 #define PredicateLockHashCodeFromTargetHashCode(predicatelocktag, targethash) \
312 ((targethash) ^ ((uint32) PointerGetDatum((predicatelocktag)->myXact)) \
313 << LOG2_NUM_PREDICATELOCK_PARTITIONS)
314
315
316 /*
317 * The SLRU buffer area through which we access the old xids.
318 */
319 static SlruCtlData SerialSlruCtlData;
320
321 #define SerialSlruCtl (&SerialSlruCtlData)
322
323 #define SERIAL_PAGESIZE BLCKSZ
324 #define SERIAL_ENTRYSIZE sizeof(SerCommitSeqNo)
325 #define SERIAL_ENTRIESPERPAGE (SERIAL_PAGESIZE / SERIAL_ENTRYSIZE)
326
327 /*
328 * Set maximum pages based on the number needed to track all transactions.
329 */
330 #define SERIAL_MAX_PAGE (MaxTransactionId / SERIAL_ENTRIESPERPAGE)
331
332 #define SerialNextPage(page) (((page) >= SERIAL_MAX_PAGE) ? 0 : (page) + 1)
333
334 #define SerialValue(slotno, xid) (*((SerCommitSeqNo *) \
335 (SerialSlruCtl->shared->page_buffer[slotno] + \
336 ((((uint32) (xid)) % SERIAL_ENTRIESPERPAGE) * SERIAL_ENTRYSIZE))))
337
338 #define SerialPage(xid) (((uint32) (xid)) / SERIAL_ENTRIESPERPAGE)
339
340 typedef struct SerialControlData
341 {
342 int headPage; /* newest initialized page */
343 TransactionId headXid; /* newest valid Xid in the SLRU */
344 TransactionId tailXid; /* oldest xmin we might be interested in */
345 } SerialControlData;
346
347 typedef struct SerialControlData *SerialControl;
348
349 static SerialControl serialControl;
350
351 /*
352 * When the oldest committed transaction on the "finished" list is moved to
353 * SLRU, its predicate locks will be moved to this "dummy" transaction,
354 * collapsing duplicate targets. When a duplicate is found, the later
355 * commitSeqNo is used.
356 */
357 static SERIALIZABLEXACT *OldCommittedSxact;
358
359
360 /*
361 * These configuration variables are used to set the predicate lock table size
362 * and to control promotion of predicate locks to coarser granularity in an
363 * attempt to degrade performance (mostly as false positive serialization
364 * failure) gracefully in the face of memory pressure.
365 */
366 int max_predicate_locks_per_xact; /* set by guc.c */
367 int max_predicate_locks_per_relation; /* set by guc.c */
368 int max_predicate_locks_per_page; /* set by guc.c */
369
370 /*
371 * This provides a list of objects in order to track transactions
372 * participating in predicate locking. Entries in the list are fixed size,
373 * and reside in shared memory. The memory address of an entry must remain
374 * fixed during its lifetime. The list will be protected from concurrent
375 * update externally; no provision is made in this code to manage that. The
376 * number of entries in the list, and the size allowed for each entry is
377 * fixed upon creation.
378 */
379 static PredXactList PredXact;
380
381 /*
382 * This provides a pool of RWConflict data elements to use in conflict lists
383 * between transactions.
384 */
385 static RWConflictPoolHeader RWConflictPool;
386
387 /*
388 * The predicate locking hash tables are in shared memory.
389 * Each backend keeps pointers to them.
390 */
391 static HTAB *SerializableXidHash;
392 static HTAB *PredicateLockTargetHash;
393 static HTAB *PredicateLockHash;
394 static SHM_QUEUE *FinishedSerializableTransactions;
395
396 /*
397 * Tag for a dummy entry in PredicateLockTargetHash. By temporarily removing
398 * this entry, you can ensure that there's enough scratch space available for
399 * inserting one entry in the hash table. This is an otherwise-invalid tag.
400 */
401 static const PREDICATELOCKTARGETTAG ScratchTargetTag = {0, 0, 0, 0};
402 static uint32 ScratchTargetTagHash;
403 static LWLock *ScratchPartitionLock;
404
405 /*
406 * The local hash table used to determine when to combine multiple fine-
407 * grained locks into a single courser-grained lock.
408 */
409 static HTAB *LocalPredicateLockHash = NULL;
410
411 /*
412 * Keep a pointer to the currently-running serializable transaction (if any)
413 * for quick reference. Also, remember if we have written anything that could
414 * cause a rw-conflict.
415 */
416 static SERIALIZABLEXACT *MySerializableXact = InvalidSerializableXact;
417 static bool MyXactDidWrite = false;
418
419 /*
420 * The SXACT_FLAG_RO_UNSAFE optimization might lead us to release
421 * MySerializableXact early. If that happens in a parallel query, the leader
422 * needs to defer the destruction of the SERIALIZABLEXACT until end of
423 * transaction, because the workers still have a reference to it. In that
424 * case, the leader stores it here.
425 */
426 static SERIALIZABLEXACT *SavedSerializableXact = InvalidSerializableXact;
427
428 /* local functions */
429
430 static SERIALIZABLEXACT *CreatePredXact(void);
431 static void ReleasePredXact(SERIALIZABLEXACT *sxact);
432 static SERIALIZABLEXACT *FirstPredXact(void);
433 static SERIALIZABLEXACT *NextPredXact(SERIALIZABLEXACT *sxact);
434
435 static bool RWConflictExists(const SERIALIZABLEXACT *reader, const SERIALIZABLEXACT *writer);
436 static void SetRWConflict(SERIALIZABLEXACT *reader, SERIALIZABLEXACT *writer);
437 static void SetPossibleUnsafeConflict(SERIALIZABLEXACT *roXact, SERIALIZABLEXACT *activeXact);
438 static void ReleaseRWConflict(RWConflict conflict);
439 static void FlagSxactUnsafe(SERIALIZABLEXACT *sxact);
440
441 static bool SerialPagePrecedesLogically(int page1, int page2);
442 static void SerialInit(void);
443 static void SerialAdd(TransactionId xid, SerCommitSeqNo minConflictCommitSeqNo);
444 static SerCommitSeqNo SerialGetMinConflictCommitSeqNo(TransactionId xid);
445 static void SerialSetActiveSerXmin(TransactionId xid);
446
447 static uint32 predicatelock_hash(const void *key, Size keysize);
448 static void SummarizeOldestCommittedSxact(void);
449 static Snapshot GetSafeSnapshot(Snapshot snapshot);
450 static Snapshot GetSerializableTransactionSnapshotInt(Snapshot snapshot,
451 VirtualTransactionId *sourcevxid,
452 int sourcepid);
453 static bool PredicateLockExists(const PREDICATELOCKTARGETTAG *targettag);
454 static bool GetParentPredicateLockTag(const PREDICATELOCKTARGETTAG *tag,
455 PREDICATELOCKTARGETTAG *parent);
456 static bool CoarserLockCovers(const PREDICATELOCKTARGETTAG *newtargettag);
457 static void RemoveScratchTarget(bool lockheld);
458 static void RestoreScratchTarget(bool lockheld);
459 static void RemoveTargetIfNoLongerUsed(PREDICATELOCKTARGET *target,
460 uint32 targettaghash);
461 static void DeleteChildTargetLocks(const PREDICATELOCKTARGETTAG *newtargettag);
462 static int MaxPredicateChildLocks(const PREDICATELOCKTARGETTAG *tag);
463 static bool CheckAndPromotePredicateLockRequest(const PREDICATELOCKTARGETTAG *reqtag);
464 static void DecrementParentLocks(const PREDICATELOCKTARGETTAG *targettag);
465 static void CreatePredicateLock(const PREDICATELOCKTARGETTAG *targettag,
466 uint32 targettaghash,
467 SERIALIZABLEXACT *sxact);
468 static void DeleteLockTarget(PREDICATELOCKTARGET *target, uint32 targettaghash);
469 static bool TransferPredicateLocksToNewTarget(PREDICATELOCKTARGETTAG oldtargettag,
470 PREDICATELOCKTARGETTAG newtargettag,
471 bool removeOld);
472 static void PredicateLockAcquire(const PREDICATELOCKTARGETTAG *targettag);
473 static void DropAllPredicateLocksFromTable(Relation relation,
474 bool transfer);
475 static void SetNewSxactGlobalXmin(void);
476 static void ClearOldPredicateLocks(void);
477 static void ReleaseOneSerializableXact(SERIALIZABLEXACT *sxact, bool partial,
478 bool summarize);
479 static bool XidIsConcurrent(TransactionId xid);
480 static void CheckTargetForConflictsIn(PREDICATELOCKTARGETTAG *targettag);
481 static void FlagRWConflict(SERIALIZABLEXACT *reader, SERIALIZABLEXACT *writer);
482 static void OnConflict_CheckForSerializationFailure(const SERIALIZABLEXACT *reader,
483 SERIALIZABLEXACT *writer);
484 static void CreateLocalPredicateLockHash(void);
485 static void ReleasePredicateLocksLocal(void);
486
487
488 /*------------------------------------------------------------------------*/
489
490 /*
491 * Does this relation participate in predicate locking? Temporary and system
492 * relations are exempt, as are materialized views.
493 */
494 static inline bool
PredicateLockingNeededForRelation(Relation relation)495 PredicateLockingNeededForRelation(Relation relation)
496 {
497 return !(relation->rd_id < FirstBootstrapObjectId ||
498 RelationUsesLocalBuffers(relation) ||
499 relation->rd_rel->relkind == RELKIND_MATVIEW);
500 }
501
502 /*
503 * When a public interface method is called for a read, this is the test to
504 * see if we should do a quick return.
505 *
506 * Note: this function has side-effects! If this transaction has been flagged
507 * as RO-safe since the last call, we release all predicate locks and reset
508 * MySerializableXact. That makes subsequent calls to return quickly.
509 *
510 * This is marked as 'inline' to eliminate the function call overhead in the
511 * common case that serialization is not needed.
512 */
513 static inline bool
SerializationNeededForRead(Relation relation,Snapshot snapshot)514 SerializationNeededForRead(Relation relation, Snapshot snapshot)
515 {
516 /* Nothing to do if this is not a serializable transaction */
517 if (MySerializableXact == InvalidSerializableXact)
518 return false;
519
520 /*
521 * Don't acquire locks or conflict when scanning with a special snapshot.
522 * This excludes things like CLUSTER and REINDEX. They use the wholesale
523 * functions TransferPredicateLocksToHeapRelation() and
524 * CheckTableForSerializableConflictIn() to participate in serialization,
525 * but the scans involved don't need serialization.
526 */
527 if (!IsMVCCSnapshot(snapshot))
528 return false;
529
530 /*
531 * Check if we have just become "RO-safe". If we have, immediately release
532 * all locks as they're not needed anymore. This also resets
533 * MySerializableXact, so that subsequent calls to this function can exit
534 * quickly.
535 *
536 * A transaction is flagged as RO_SAFE if all concurrent R/W transactions
537 * commit without having conflicts out to an earlier snapshot, thus
538 * ensuring that no conflicts are possible for this transaction.
539 */
540 if (SxactIsROSafe(MySerializableXact))
541 {
542 ReleasePredicateLocks(false, true);
543 return false;
544 }
545
546 /* Check if the relation doesn't participate in predicate locking */
547 if (!PredicateLockingNeededForRelation(relation))
548 return false;
549
550 return true; /* no excuse to skip predicate locking */
551 }
552
553 /*
554 * Like SerializationNeededForRead(), but called on writes.
555 * The logic is the same, but there is no snapshot and we can't be RO-safe.
556 */
557 static inline bool
SerializationNeededForWrite(Relation relation)558 SerializationNeededForWrite(Relation relation)
559 {
560 /* Nothing to do if this is not a serializable transaction */
561 if (MySerializableXact == InvalidSerializableXact)
562 return false;
563
564 /* Check if the relation doesn't participate in predicate locking */
565 if (!PredicateLockingNeededForRelation(relation))
566 return false;
567
568 return true; /* no excuse to skip predicate locking */
569 }
570
571
572 /*------------------------------------------------------------------------*/
573
574 /*
575 * These functions are a simple implementation of a list for this specific
576 * type of struct. If there is ever a generalized shared memory list, we
577 * should probably switch to that.
578 */
579 static SERIALIZABLEXACT *
CreatePredXact(void)580 CreatePredXact(void)
581 {
582 PredXactListElement ptle;
583
584 ptle = (PredXactListElement)
585 SHMQueueNext(&PredXact->availableList,
586 &PredXact->availableList,
587 offsetof(PredXactListElementData, link));
588 if (!ptle)
589 return NULL;
590
591 SHMQueueDelete(&ptle->link);
592 SHMQueueInsertBefore(&PredXact->activeList, &ptle->link);
593 return &ptle->sxact;
594 }
595
596 static void
ReleasePredXact(SERIALIZABLEXACT * sxact)597 ReleasePredXact(SERIALIZABLEXACT *sxact)
598 {
599 PredXactListElement ptle;
600
601 Assert(ShmemAddrIsValid(sxact));
602
603 ptle = (PredXactListElement)
604 (((char *) sxact)
605 - offsetof(PredXactListElementData, sxact)
606 + offsetof(PredXactListElementData, link));
607 SHMQueueDelete(&ptle->link);
608 SHMQueueInsertBefore(&PredXact->availableList, &ptle->link);
609 }
610
611 static SERIALIZABLEXACT *
FirstPredXact(void)612 FirstPredXact(void)
613 {
614 PredXactListElement ptle;
615
616 ptle = (PredXactListElement)
617 SHMQueueNext(&PredXact->activeList,
618 &PredXact->activeList,
619 offsetof(PredXactListElementData, link));
620 if (!ptle)
621 return NULL;
622
623 return &ptle->sxact;
624 }
625
626 static SERIALIZABLEXACT *
NextPredXact(SERIALIZABLEXACT * sxact)627 NextPredXact(SERIALIZABLEXACT *sxact)
628 {
629 PredXactListElement ptle;
630
631 Assert(ShmemAddrIsValid(sxact));
632
633 ptle = (PredXactListElement)
634 (((char *) sxact)
635 - offsetof(PredXactListElementData, sxact)
636 + offsetof(PredXactListElementData, link));
637 ptle = (PredXactListElement)
638 SHMQueueNext(&PredXact->activeList,
639 &ptle->link,
640 offsetof(PredXactListElementData, link));
641 if (!ptle)
642 return NULL;
643
644 return &ptle->sxact;
645 }
646
647 /*------------------------------------------------------------------------*/
648
649 /*
650 * These functions manage primitive access to the RWConflict pool and lists.
651 */
652 static bool
RWConflictExists(const SERIALIZABLEXACT * reader,const SERIALIZABLEXACT * writer)653 RWConflictExists(const SERIALIZABLEXACT *reader, const SERIALIZABLEXACT *writer)
654 {
655 RWConflict conflict;
656
657 Assert(reader != writer);
658
659 /* Check the ends of the purported conflict first. */
660 if (SxactIsDoomed(reader)
661 || SxactIsDoomed(writer)
662 || SHMQueueEmpty(&reader->outConflicts)
663 || SHMQueueEmpty(&writer->inConflicts))
664 return false;
665
666 /* A conflict is possible; walk the list to find out. */
667 conflict = (RWConflict)
668 SHMQueueNext(&reader->outConflicts,
669 &reader->outConflicts,
670 offsetof(RWConflictData, outLink));
671 while (conflict)
672 {
673 if (conflict->sxactIn == writer)
674 return true;
675 conflict = (RWConflict)
676 SHMQueueNext(&reader->outConflicts,
677 &conflict->outLink,
678 offsetof(RWConflictData, outLink));
679 }
680
681 /* No conflict found. */
682 return false;
683 }
684
685 static void
SetRWConflict(SERIALIZABLEXACT * reader,SERIALIZABLEXACT * writer)686 SetRWConflict(SERIALIZABLEXACT *reader, SERIALIZABLEXACT *writer)
687 {
688 RWConflict conflict;
689
690 Assert(reader != writer);
691 Assert(!RWConflictExists(reader, writer));
692
693 conflict = (RWConflict)
694 SHMQueueNext(&RWConflictPool->availableList,
695 &RWConflictPool->availableList,
696 offsetof(RWConflictData, outLink));
697 if (!conflict)
698 ereport(ERROR,
699 (errcode(ERRCODE_OUT_OF_MEMORY),
700 errmsg("not enough elements in RWConflictPool to record a read/write conflict"),
701 errhint("You might need to run fewer transactions at a time or increase max_connections.")));
702
703 SHMQueueDelete(&conflict->outLink);
704
705 conflict->sxactOut = reader;
706 conflict->sxactIn = writer;
707 SHMQueueInsertBefore(&reader->outConflicts, &conflict->outLink);
708 SHMQueueInsertBefore(&writer->inConflicts, &conflict->inLink);
709 }
710
711 static void
SetPossibleUnsafeConflict(SERIALIZABLEXACT * roXact,SERIALIZABLEXACT * activeXact)712 SetPossibleUnsafeConflict(SERIALIZABLEXACT *roXact,
713 SERIALIZABLEXACT *activeXact)
714 {
715 RWConflict conflict;
716
717 Assert(roXact != activeXact);
718 Assert(SxactIsReadOnly(roXact));
719 Assert(!SxactIsReadOnly(activeXact));
720
721 conflict = (RWConflict)
722 SHMQueueNext(&RWConflictPool->availableList,
723 &RWConflictPool->availableList,
724 offsetof(RWConflictData, outLink));
725 if (!conflict)
726 ereport(ERROR,
727 (errcode(ERRCODE_OUT_OF_MEMORY),
728 errmsg("not enough elements in RWConflictPool to record a potential read/write conflict"),
729 errhint("You might need to run fewer transactions at a time or increase max_connections.")));
730
731 SHMQueueDelete(&conflict->outLink);
732
733 conflict->sxactOut = activeXact;
734 conflict->sxactIn = roXact;
735 SHMQueueInsertBefore(&activeXact->possibleUnsafeConflicts,
736 &conflict->outLink);
737 SHMQueueInsertBefore(&roXact->possibleUnsafeConflicts,
738 &conflict->inLink);
739 }
740
741 static void
ReleaseRWConflict(RWConflict conflict)742 ReleaseRWConflict(RWConflict conflict)
743 {
744 SHMQueueDelete(&conflict->inLink);
745 SHMQueueDelete(&conflict->outLink);
746 SHMQueueInsertBefore(&RWConflictPool->availableList, &conflict->outLink);
747 }
748
749 static void
FlagSxactUnsafe(SERIALIZABLEXACT * sxact)750 FlagSxactUnsafe(SERIALIZABLEXACT *sxact)
751 {
752 RWConflict conflict,
753 nextConflict;
754
755 Assert(SxactIsReadOnly(sxact));
756 Assert(!SxactIsROSafe(sxact));
757
758 sxact->flags |= SXACT_FLAG_RO_UNSAFE;
759
760 /*
761 * We know this isn't a safe snapshot, so we can stop looking for other
762 * potential conflicts.
763 */
764 conflict = (RWConflict)
765 SHMQueueNext(&sxact->possibleUnsafeConflicts,
766 &sxact->possibleUnsafeConflicts,
767 offsetof(RWConflictData, inLink));
768 while (conflict)
769 {
770 nextConflict = (RWConflict)
771 SHMQueueNext(&sxact->possibleUnsafeConflicts,
772 &conflict->inLink,
773 offsetof(RWConflictData, inLink));
774
775 Assert(!SxactIsReadOnly(conflict->sxactOut));
776 Assert(sxact == conflict->sxactIn);
777
778 ReleaseRWConflict(conflict);
779
780 conflict = nextConflict;
781 }
782 }
783
784 /*------------------------------------------------------------------------*/
785
786 /*
787 * Decide whether a Serial page number is "older" for truncation purposes.
788 * Analogous to CLOGPagePrecedes().
789 */
790 static bool
SerialPagePrecedesLogically(int page1,int page2)791 SerialPagePrecedesLogically(int page1, int page2)
792 {
793 TransactionId xid1;
794 TransactionId xid2;
795
796 xid1 = ((TransactionId) page1) * SERIAL_ENTRIESPERPAGE;
797 xid1 += FirstNormalTransactionId + 1;
798 xid2 = ((TransactionId) page2) * SERIAL_ENTRIESPERPAGE;
799 xid2 += FirstNormalTransactionId + 1;
800
801 return (TransactionIdPrecedes(xid1, xid2) &&
802 TransactionIdPrecedes(xid1, xid2 + SERIAL_ENTRIESPERPAGE - 1));
803 }
804
805 #ifdef USE_ASSERT_CHECKING
806 static void
SerialPagePrecedesLogicallyUnitTests(void)807 SerialPagePrecedesLogicallyUnitTests(void)
808 {
809 int per_page = SERIAL_ENTRIESPERPAGE,
810 offset = per_page / 2;
811 int newestPage,
812 oldestPage,
813 headPage,
814 targetPage;
815 TransactionId newestXact,
816 oldestXact;
817
818 /* GetNewTransactionId() has assigned the last XID it can safely use. */
819 newestPage = 2 * SLRU_PAGES_PER_SEGMENT - 1; /* nothing special */
820 newestXact = newestPage * per_page + offset;
821 Assert(newestXact / per_page == newestPage);
822 oldestXact = newestXact + 1;
823 oldestXact -= 1U << 31;
824 oldestPage = oldestXact / per_page;
825
826 /*
827 * In this scenario, the SLRU headPage pertains to the last ~1000 XIDs
828 * assigned. oldestXact finishes, ~2B XIDs having elapsed since it
829 * started. Further transactions cause us to summarize oldestXact to
830 * tailPage. Function must return false so SerialAdd() doesn't zero
831 * tailPage (which may contain entries for other old, recently-finished
832 * XIDs) and half the SLRU. Reaching this requires burning ~2B XIDs in
833 * single-user mode, a negligible possibility.
834 */
835 headPage = newestPage;
836 targetPage = oldestPage;
837 Assert(!SerialPagePrecedesLogically(headPage, targetPage));
838
839 /*
840 * In this scenario, the SLRU headPage pertains to oldestXact. We're
841 * summarizing an XID near newestXact. (Assume few other XIDs used
842 * SERIALIZABLE, hence the minimal headPage advancement. Assume
843 * oldestXact was long-running and only recently reached the SLRU.)
844 * Function must return true to make SerialAdd() create targetPage.
845 *
846 * Today's implementation mishandles this case, but it doesn't matter
847 * enough to fix. Verify that the defect affects just one page by
848 * asserting correct treatment of its prior page. Reaching this case
849 * requires burning ~2B XIDs in single-user mode, a negligible
850 * possibility. Moreover, if it does happen, the consequence would be
851 * mild, namely a new transaction failing in SimpleLruReadPage().
852 */
853 headPage = oldestPage;
854 targetPage = newestPage;
855 Assert(SerialPagePrecedesLogically(headPage, targetPage - 1));
856 #if 0
857 Assert(SerialPagePrecedesLogically(headPage, targetPage));
858 #endif
859 }
860 #endif
861
862 /*
863 * Initialize for the tracking of old serializable committed xids.
864 */
865 static void
SerialInit(void)866 SerialInit(void)
867 {
868 bool found;
869
870 /*
871 * Set up SLRU management of the pg_serial data.
872 */
873 SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
874 SimpleLruInit(SerialSlruCtl, "Serial",
875 NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
876 LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
877 #ifdef USE_ASSERT_CHECKING
878 SerialPagePrecedesLogicallyUnitTests();
879 #endif
880 SlruPagePrecedesUnitTests(SerialSlruCtl, SERIAL_ENTRIESPERPAGE);
881
882 /*
883 * Create or attach to the SerialControl structure.
884 */
885 serialControl = (SerialControl)
886 ShmemInitStruct("SerialControlData", sizeof(SerialControlData), &found);
887
888 Assert(found == IsUnderPostmaster);
889 if (!found)
890 {
891 /*
892 * Set control information to reflect empty SLRU.
893 */
894 serialControl->headPage = -1;
895 serialControl->headXid = InvalidTransactionId;
896 serialControl->tailXid = InvalidTransactionId;
897 }
898 }
899
900 /*
901 * Record a committed read write serializable xid and the minimum
902 * commitSeqNo of any transactions to which this xid had a rw-conflict out.
903 * An invalid commitSeqNo means that there were no conflicts out from xid.
904 */
905 static void
SerialAdd(TransactionId xid,SerCommitSeqNo minConflictCommitSeqNo)906 SerialAdd(TransactionId xid, SerCommitSeqNo minConflictCommitSeqNo)
907 {
908 TransactionId tailXid;
909 int targetPage;
910 int slotno;
911 int firstZeroPage;
912 bool isNewPage;
913
914 Assert(TransactionIdIsValid(xid));
915
916 targetPage = SerialPage(xid);
917
918 LWLockAcquire(SerialSLRULock, LW_EXCLUSIVE);
919
920 /*
921 * If no serializable transactions are active, there shouldn't be anything
922 * to push out to the SLRU. Hitting this assert would mean there's
923 * something wrong with the earlier cleanup logic.
924 */
925 tailXid = serialControl->tailXid;
926 Assert(TransactionIdIsValid(tailXid));
927
928 /*
929 * If the SLRU is currently unused, zero out the whole active region from
930 * tailXid to headXid before taking it into use. Otherwise zero out only
931 * any new pages that enter the tailXid-headXid range as we advance
932 * headXid.
933 */
934 if (serialControl->headPage < 0)
935 {
936 firstZeroPage = SerialPage(tailXid);
937 isNewPage = true;
938 }
939 else
940 {
941 firstZeroPage = SerialNextPage(serialControl->headPage);
942 isNewPage = SerialPagePrecedesLogically(serialControl->headPage,
943 targetPage);
944 }
945
946 if (!TransactionIdIsValid(serialControl->headXid)
947 || TransactionIdFollows(xid, serialControl->headXid))
948 serialControl->headXid = xid;
949 if (isNewPage)
950 serialControl->headPage = targetPage;
951
952 if (isNewPage)
953 {
954 /* Initialize intervening pages. */
955 while (firstZeroPage != targetPage)
956 {
957 (void) SimpleLruZeroPage(SerialSlruCtl, firstZeroPage);
958 firstZeroPage = SerialNextPage(firstZeroPage);
959 }
960 slotno = SimpleLruZeroPage(SerialSlruCtl, targetPage);
961 }
962 else
963 slotno = SimpleLruReadPage(SerialSlruCtl, targetPage, true, xid);
964
965 SerialValue(slotno, xid) = minConflictCommitSeqNo;
966 SerialSlruCtl->shared->page_dirty[slotno] = true;
967
968 LWLockRelease(SerialSLRULock);
969 }
970
971 /*
972 * Get the minimum commitSeqNo for any conflict out for the given xid. For
973 * a transaction which exists but has no conflict out, InvalidSerCommitSeqNo
974 * will be returned.
975 */
976 static SerCommitSeqNo
SerialGetMinConflictCommitSeqNo(TransactionId xid)977 SerialGetMinConflictCommitSeqNo(TransactionId xid)
978 {
979 TransactionId headXid;
980 TransactionId tailXid;
981 SerCommitSeqNo val;
982 int slotno;
983
984 Assert(TransactionIdIsValid(xid));
985
986 LWLockAcquire(SerialSLRULock, LW_SHARED);
987 headXid = serialControl->headXid;
988 tailXid = serialControl->tailXid;
989 LWLockRelease(SerialSLRULock);
990
991 if (!TransactionIdIsValid(headXid))
992 return 0;
993
994 Assert(TransactionIdIsValid(tailXid));
995
996 if (TransactionIdPrecedes(xid, tailXid)
997 || TransactionIdFollows(xid, headXid))
998 return 0;
999
1000 /*
1001 * The following function must be called without holding SerialSLRULock,
1002 * but will return with that lock held, which must then be released.
1003 */
1004 slotno = SimpleLruReadPage_ReadOnly(SerialSlruCtl,
1005 SerialPage(xid), xid);
1006 val = SerialValue(slotno, xid);
1007 LWLockRelease(SerialSLRULock);
1008 return val;
1009 }
1010
1011 /*
1012 * Call this whenever there is a new xmin for active serializable
1013 * transactions. We don't need to keep information on transactions which
1014 * precede that. InvalidTransactionId means none active, so everything in
1015 * the SLRU can be discarded.
1016 */
1017 static void
SerialSetActiveSerXmin(TransactionId xid)1018 SerialSetActiveSerXmin(TransactionId xid)
1019 {
1020 LWLockAcquire(SerialSLRULock, LW_EXCLUSIVE);
1021
1022 /*
1023 * When no sxacts are active, nothing overlaps, set the xid values to
1024 * invalid to show that there are no valid entries. Don't clear headPage,
1025 * though. A new xmin might still land on that page, and we don't want to
1026 * repeatedly zero out the same page.
1027 */
1028 if (!TransactionIdIsValid(xid))
1029 {
1030 serialControl->tailXid = InvalidTransactionId;
1031 serialControl->headXid = InvalidTransactionId;
1032 LWLockRelease(SerialSLRULock);
1033 return;
1034 }
1035
1036 /*
1037 * When we're recovering prepared transactions, the global xmin might move
1038 * backwards depending on the order they're recovered. Normally that's not
1039 * OK, but during recovery no serializable transactions will commit, so
1040 * the SLRU is empty and we can get away with it.
1041 */
1042 if (RecoveryInProgress())
1043 {
1044 Assert(serialControl->headPage < 0);
1045 if (!TransactionIdIsValid(serialControl->tailXid)
1046 || TransactionIdPrecedes(xid, serialControl->tailXid))
1047 {
1048 serialControl->tailXid = xid;
1049 }
1050 LWLockRelease(SerialSLRULock);
1051 return;
1052 }
1053
1054 Assert(!TransactionIdIsValid(serialControl->tailXid)
1055 || TransactionIdFollows(xid, serialControl->tailXid));
1056
1057 serialControl->tailXid = xid;
1058
1059 LWLockRelease(SerialSLRULock);
1060 }
1061
1062 /*
1063 * Perform a checkpoint --- either during shutdown, or on-the-fly
1064 *
1065 * We don't have any data that needs to survive a restart, but this is a
1066 * convenient place to truncate the SLRU.
1067 */
1068 void
CheckPointPredicate(void)1069 CheckPointPredicate(void)
1070 {
1071 int tailPage;
1072
1073 LWLockAcquire(SerialSLRULock, LW_EXCLUSIVE);
1074
1075 /* Exit quickly if the SLRU is currently not in use. */
1076 if (serialControl->headPage < 0)
1077 {
1078 LWLockRelease(SerialSLRULock);
1079 return;
1080 }
1081
1082 if (TransactionIdIsValid(serialControl->tailXid))
1083 {
1084 /* We can truncate the SLRU up to the page containing tailXid */
1085 tailPage = SerialPage(serialControl->tailXid);
1086 }
1087 else
1088 {
1089 /*----------
1090 * The SLRU is no longer needed. Truncate to head before we set head
1091 * invalid.
1092 *
1093 * XXX: It's possible that the SLRU is not needed again until XID
1094 * wrap-around has happened, so that the segment containing headPage
1095 * that we leave behind will appear to be new again. In that case it
1096 * won't be removed until XID horizon advances enough to make it
1097 * current again.
1098 *
1099 * XXX: This should happen in vac_truncate_clog(), not in checkpoints.
1100 * Consider this scenario, starting from a system with no in-progress
1101 * transactions and VACUUM FREEZE having maximized oldestXact:
1102 * - Start a SERIALIZABLE transaction.
1103 * - Start, finish, and summarize a SERIALIZABLE transaction, creating
1104 * one SLRU page.
1105 * - Consume XIDs to reach xidStopLimit.
1106 * - Finish all transactions. Due to the long-running SERIALIZABLE
1107 * transaction, earlier checkpoints did not touch headPage. The
1108 * next checkpoint will change it, but that checkpoint happens after
1109 * the end of the scenario.
1110 * - VACUUM to advance XID limits.
1111 * - Consume ~2M XIDs, crossing the former xidWrapLimit.
1112 * - Start, finish, and summarize a SERIALIZABLE transaction.
1113 * SerialAdd() declines to create the targetPage, because headPage
1114 * is not regarded as in the past relative to that targetPage. The
1115 * transaction instigating the summarize fails in
1116 * SimpleLruReadPage().
1117 */
1118 tailPage = serialControl->headPage;
1119 serialControl->headPage = -1;
1120 }
1121
1122 LWLockRelease(SerialSLRULock);
1123
1124 /* Truncate away pages that are no longer required */
1125 SimpleLruTruncate(SerialSlruCtl, tailPage);
1126
1127 /*
1128 * Write dirty SLRU pages to disk
1129 *
1130 * This is not actually necessary from a correctness point of view. We do
1131 * it merely as a debugging aid.
1132 *
1133 * We're doing this after the truncation to avoid writing pages right
1134 * before deleting the file in which they sit, which would be completely
1135 * pointless.
1136 */
1137 SimpleLruWriteAll(SerialSlruCtl, true);
1138 }
1139
1140 /*------------------------------------------------------------------------*/
1141
1142 /*
1143 * InitPredicateLocks -- Initialize the predicate locking data structures.
1144 *
1145 * This is called from CreateSharedMemoryAndSemaphores(), which see for
1146 * more comments. In the normal postmaster case, the shared hash tables
1147 * are created here. Backends inherit the pointers
1148 * to the shared tables via fork(). In the EXEC_BACKEND case, each
1149 * backend re-executes this code to obtain pointers to the already existing
1150 * shared hash tables.
1151 */
1152 void
InitPredicateLocks(void)1153 InitPredicateLocks(void)
1154 {
1155 HASHCTL info;
1156 long max_table_size;
1157 Size requestSize;
1158 bool found;
1159
1160 #ifndef EXEC_BACKEND
1161 Assert(!IsUnderPostmaster);
1162 #endif
1163
1164 /*
1165 * Compute size of predicate lock target hashtable. Note these
1166 * calculations must agree with PredicateLockShmemSize!
1167 */
1168 max_table_size = NPREDICATELOCKTARGETENTS();
1169
1170 /*
1171 * Allocate hash table for PREDICATELOCKTARGET structs. This stores
1172 * per-predicate-lock-target information.
1173 */
1174 info.keysize = sizeof(PREDICATELOCKTARGETTAG);
1175 info.entrysize = sizeof(PREDICATELOCKTARGET);
1176 info.num_partitions = NUM_PREDICATELOCK_PARTITIONS;
1177
1178 PredicateLockTargetHash = ShmemInitHash("PREDICATELOCKTARGET hash",
1179 max_table_size,
1180 max_table_size,
1181 &info,
1182 HASH_ELEM | HASH_BLOBS |
1183 HASH_PARTITION | HASH_FIXED_SIZE);
1184
1185 /*
1186 * Reserve a dummy entry in the hash table; we use it to make sure there's
1187 * always one entry available when we need to split or combine a page,
1188 * because running out of space there could mean aborting a
1189 * non-serializable transaction.
1190 */
1191 if (!IsUnderPostmaster)
1192 {
1193 (void) hash_search(PredicateLockTargetHash, &ScratchTargetTag,
1194 HASH_ENTER, &found);
1195 Assert(!found);
1196 }
1197
1198 /* Pre-calculate the hash and partition lock of the scratch entry */
1199 ScratchTargetTagHash = PredicateLockTargetTagHashCode(&ScratchTargetTag);
1200 ScratchPartitionLock = PredicateLockHashPartitionLock(ScratchTargetTagHash);
1201
1202 /*
1203 * Allocate hash table for PREDICATELOCK structs. This stores per
1204 * xact-lock-of-a-target information.
1205 */
1206 info.keysize = sizeof(PREDICATELOCKTAG);
1207 info.entrysize = sizeof(PREDICATELOCK);
1208 info.hash = predicatelock_hash;
1209 info.num_partitions = NUM_PREDICATELOCK_PARTITIONS;
1210
1211 /* Assume an average of 2 xacts per target */
1212 max_table_size *= 2;
1213
1214 PredicateLockHash = ShmemInitHash("PREDICATELOCK hash",
1215 max_table_size,
1216 max_table_size,
1217 &info,
1218 HASH_ELEM | HASH_FUNCTION |
1219 HASH_PARTITION | HASH_FIXED_SIZE);
1220
1221 /*
1222 * Compute size for serializable transaction hashtable. Note these
1223 * calculations must agree with PredicateLockShmemSize!
1224 */
1225 max_table_size = (MaxBackends + max_prepared_xacts);
1226
1227 /*
1228 * Allocate a list to hold information on transactions participating in
1229 * predicate locking.
1230 *
1231 * Assume an average of 10 predicate locking transactions per backend.
1232 * This allows aggressive cleanup while detail is present before data must
1233 * be summarized for storage in SLRU and the "dummy" transaction.
1234 */
1235 max_table_size *= 10;
1236
1237 PredXact = ShmemInitStruct("PredXactList",
1238 PredXactListDataSize,
1239 &found);
1240 Assert(found == IsUnderPostmaster);
1241 if (!found)
1242 {
1243 int i;
1244
1245 SHMQueueInit(&PredXact->availableList);
1246 SHMQueueInit(&PredXact->activeList);
1247 PredXact->SxactGlobalXmin = InvalidTransactionId;
1248 PredXact->SxactGlobalXminCount = 0;
1249 PredXact->WritableSxactCount = 0;
1250 PredXact->LastSxactCommitSeqNo = FirstNormalSerCommitSeqNo - 1;
1251 PredXact->CanPartialClearThrough = 0;
1252 PredXact->HavePartialClearedThrough = 0;
1253 requestSize = mul_size((Size) max_table_size,
1254 PredXactListElementDataSize);
1255 PredXact->element = ShmemAlloc(requestSize);
1256 /* Add all elements to available list, clean. */
1257 memset(PredXact->element, 0, requestSize);
1258 for (i = 0; i < max_table_size; i++)
1259 {
1260 LWLockInitialize(&PredXact->element[i].sxact.perXactPredicateListLock,
1261 LWTRANCHE_PER_XACT_PREDICATE_LIST);
1262 SHMQueueInsertBefore(&(PredXact->availableList),
1263 &(PredXact->element[i].link));
1264 }
1265 PredXact->OldCommittedSxact = CreatePredXact();
1266 SetInvalidVirtualTransactionId(PredXact->OldCommittedSxact->vxid);
1267 PredXact->OldCommittedSxact->prepareSeqNo = 0;
1268 PredXact->OldCommittedSxact->commitSeqNo = 0;
1269 PredXact->OldCommittedSxact->SeqNo.lastCommitBeforeSnapshot = 0;
1270 SHMQueueInit(&PredXact->OldCommittedSxact->outConflicts);
1271 SHMQueueInit(&PredXact->OldCommittedSxact->inConflicts);
1272 SHMQueueInit(&PredXact->OldCommittedSxact->predicateLocks);
1273 SHMQueueInit(&PredXact->OldCommittedSxact->finishedLink);
1274 SHMQueueInit(&PredXact->OldCommittedSxact->possibleUnsafeConflicts);
1275 PredXact->OldCommittedSxact->topXid = InvalidTransactionId;
1276 PredXact->OldCommittedSxact->finishedBefore = InvalidTransactionId;
1277 PredXact->OldCommittedSxact->xmin = InvalidTransactionId;
1278 PredXact->OldCommittedSxact->flags = SXACT_FLAG_COMMITTED;
1279 PredXact->OldCommittedSxact->pid = 0;
1280 }
1281 /* This never changes, so let's keep a local copy. */
1282 OldCommittedSxact = PredXact->OldCommittedSxact;
1283
1284 /*
1285 * Allocate hash table for SERIALIZABLEXID structs. This stores per-xid
1286 * information for serializable transactions which have accessed data.
1287 */
1288 info.keysize = sizeof(SERIALIZABLEXIDTAG);
1289 info.entrysize = sizeof(SERIALIZABLEXID);
1290
1291 SerializableXidHash = ShmemInitHash("SERIALIZABLEXID hash",
1292 max_table_size,
1293 max_table_size,
1294 &info,
1295 HASH_ELEM | HASH_BLOBS |
1296 HASH_FIXED_SIZE);
1297
1298 /*
1299 * Allocate space for tracking rw-conflicts in lists attached to the
1300 * transactions.
1301 *
1302 * Assume an average of 5 conflicts per transaction. Calculations suggest
1303 * that this will prevent resource exhaustion in even the most pessimal
1304 * loads up to max_connections = 200 with all 200 connections pounding the
1305 * database with serializable transactions. Beyond that, there may be
1306 * occasional transactions canceled when trying to flag conflicts. That's
1307 * probably OK.
1308 */
1309 max_table_size *= 5;
1310
1311 RWConflictPool = ShmemInitStruct("RWConflictPool",
1312 RWConflictPoolHeaderDataSize,
1313 &found);
1314 Assert(found == IsUnderPostmaster);
1315 if (!found)
1316 {
1317 int i;
1318
1319 SHMQueueInit(&RWConflictPool->availableList);
1320 requestSize = mul_size((Size) max_table_size,
1321 RWConflictDataSize);
1322 RWConflictPool->element = ShmemAlloc(requestSize);
1323 /* Add all elements to available list, clean. */
1324 memset(RWConflictPool->element, 0, requestSize);
1325 for (i = 0; i < max_table_size; i++)
1326 {
1327 SHMQueueInsertBefore(&(RWConflictPool->availableList),
1328 &(RWConflictPool->element[i].outLink));
1329 }
1330 }
1331
1332 /*
1333 * Create or attach to the header for the list of finished serializable
1334 * transactions.
1335 */
1336 FinishedSerializableTransactions = (SHM_QUEUE *)
1337 ShmemInitStruct("FinishedSerializableTransactions",
1338 sizeof(SHM_QUEUE),
1339 &found);
1340 Assert(found == IsUnderPostmaster);
1341 if (!found)
1342 SHMQueueInit(FinishedSerializableTransactions);
1343
1344 /*
1345 * Initialize the SLRU storage for old committed serializable
1346 * transactions.
1347 */
1348 SerialInit();
1349 }
1350
1351 /*
1352 * Estimate shared-memory space used for predicate lock table
1353 */
1354 Size
PredicateLockShmemSize(void)1355 PredicateLockShmemSize(void)
1356 {
1357 Size size = 0;
1358 long max_table_size;
1359
1360 /* predicate lock target hash table */
1361 max_table_size = NPREDICATELOCKTARGETENTS();
1362 size = add_size(size, hash_estimate_size(max_table_size,
1363 sizeof(PREDICATELOCKTARGET)));
1364
1365 /* predicate lock hash table */
1366 max_table_size *= 2;
1367 size = add_size(size, hash_estimate_size(max_table_size,
1368 sizeof(PREDICATELOCK)));
1369
1370 /*
1371 * Since NPREDICATELOCKTARGETENTS is only an estimate, add 10% safety
1372 * margin.
1373 */
1374 size = add_size(size, size / 10);
1375
1376 /* transaction list */
1377 max_table_size = MaxBackends + max_prepared_xacts;
1378 max_table_size *= 10;
1379 size = add_size(size, PredXactListDataSize);
1380 size = add_size(size, mul_size((Size) max_table_size,
1381 PredXactListElementDataSize));
1382
1383 /* transaction xid table */
1384 size = add_size(size, hash_estimate_size(max_table_size,
1385 sizeof(SERIALIZABLEXID)));
1386
1387 /* rw-conflict pool */
1388 max_table_size *= 5;
1389 size = add_size(size, RWConflictPoolHeaderDataSize);
1390 size = add_size(size, mul_size((Size) max_table_size,
1391 RWConflictDataSize));
1392
1393 /* Head for list of finished serializable transactions. */
1394 size = add_size(size, sizeof(SHM_QUEUE));
1395
1396 /* Shared memory structures for SLRU tracking of old committed xids. */
1397 size = add_size(size, sizeof(SerialControlData));
1398 size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
1399
1400 return size;
1401 }
1402
1403
1404 /*
1405 * Compute the hash code associated with a PREDICATELOCKTAG.
1406 *
1407 * Because we want to use just one set of partition locks for both the
1408 * PREDICATELOCKTARGET and PREDICATELOCK hash tables, we have to make sure
1409 * that PREDICATELOCKs fall into the same partition number as their
1410 * associated PREDICATELOCKTARGETs. dynahash.c expects the partition number
1411 * to be the low-order bits of the hash code, and therefore a
1412 * PREDICATELOCKTAG's hash code must have the same low-order bits as the
1413 * associated PREDICATELOCKTARGETTAG's hash code. We achieve this with this
1414 * specialized hash function.
1415 */
1416 static uint32
predicatelock_hash(const void * key,Size keysize)1417 predicatelock_hash(const void *key, Size keysize)
1418 {
1419 const PREDICATELOCKTAG *predicatelocktag = (const PREDICATELOCKTAG *) key;
1420 uint32 targethash;
1421
1422 Assert(keysize == sizeof(PREDICATELOCKTAG));
1423
1424 /* Look into the associated target object, and compute its hash code */
1425 targethash = PredicateLockTargetTagHashCode(&predicatelocktag->myTarget->tag);
1426
1427 return PredicateLockHashCodeFromTargetHashCode(predicatelocktag, targethash);
1428 }
1429
1430
1431 /*
1432 * GetPredicateLockStatusData
1433 * Return a table containing the internal state of the predicate
1434 * lock manager for use in pg_lock_status.
1435 *
1436 * Like GetLockStatusData, this function tries to hold the partition LWLocks
1437 * for as short a time as possible by returning two arrays that simply
1438 * contain the PREDICATELOCKTARGETTAG and SERIALIZABLEXACT for each lock
1439 * table entry. Multiple copies of the same PREDICATELOCKTARGETTAG and
1440 * SERIALIZABLEXACT will likely appear.
1441 */
1442 PredicateLockData *
GetPredicateLockStatusData(void)1443 GetPredicateLockStatusData(void)
1444 {
1445 PredicateLockData *data;
1446 int i;
1447 int els,
1448 el;
1449 HASH_SEQ_STATUS seqstat;
1450 PREDICATELOCK *predlock;
1451
1452 data = (PredicateLockData *) palloc(sizeof(PredicateLockData));
1453
1454 /*
1455 * To ensure consistency, take simultaneous locks on all partition locks
1456 * in ascending order, then SerializableXactHashLock.
1457 */
1458 for (i = 0; i < NUM_PREDICATELOCK_PARTITIONS; i++)
1459 LWLockAcquire(PredicateLockHashPartitionLockByIndex(i), LW_SHARED);
1460 LWLockAcquire(SerializableXactHashLock, LW_SHARED);
1461
1462 /* Get number of locks and allocate appropriately-sized arrays. */
1463 els = hash_get_num_entries(PredicateLockHash);
1464 data->nelements = els;
1465 data->locktags = (PREDICATELOCKTARGETTAG *)
1466 palloc(sizeof(PREDICATELOCKTARGETTAG) * els);
1467 data->xacts = (SERIALIZABLEXACT *)
1468 palloc(sizeof(SERIALIZABLEXACT) * els);
1469
1470
1471 /* Scan through PredicateLockHash and copy contents */
1472 hash_seq_init(&seqstat, PredicateLockHash);
1473
1474 el = 0;
1475
1476 while ((predlock = (PREDICATELOCK *) hash_seq_search(&seqstat)))
1477 {
1478 data->locktags[el] = predlock->tag.myTarget->tag;
1479 data->xacts[el] = *predlock->tag.myXact;
1480 el++;
1481 }
1482
1483 Assert(el == els);
1484
1485 /* Release locks in reverse order */
1486 LWLockRelease(SerializableXactHashLock);
1487 for (i = NUM_PREDICATELOCK_PARTITIONS - 1; i >= 0; i--)
1488 LWLockRelease(PredicateLockHashPartitionLockByIndex(i));
1489
1490 return data;
1491 }
1492
1493 /*
1494 * Free up shared memory structures by pushing the oldest sxact (the one at
1495 * the front of the SummarizeOldestCommittedSxact queue) into summary form.
1496 * Each call will free exactly one SERIALIZABLEXACT structure and may also
1497 * free one or more of these structures: SERIALIZABLEXID, PREDICATELOCK,
1498 * PREDICATELOCKTARGET, RWConflictData.
1499 */
1500 static void
SummarizeOldestCommittedSxact(void)1501 SummarizeOldestCommittedSxact(void)
1502 {
1503 SERIALIZABLEXACT *sxact;
1504
1505 LWLockAcquire(SerializableFinishedListLock, LW_EXCLUSIVE);
1506
1507 /*
1508 * This function is only called if there are no sxact slots available.
1509 * Some of them must belong to old, already-finished transactions, so
1510 * there should be something in FinishedSerializableTransactions list that
1511 * we can summarize. However, there's a race condition: while we were not
1512 * holding any locks, a transaction might have ended and cleaned up all
1513 * the finished sxact entries already, freeing up their sxact slots. In
1514 * that case, we have nothing to do here. The caller will find one of the
1515 * slots released by the other backend when it retries.
1516 */
1517 if (SHMQueueEmpty(FinishedSerializableTransactions))
1518 {
1519 LWLockRelease(SerializableFinishedListLock);
1520 return;
1521 }
1522
1523 /*
1524 * Grab the first sxact off the finished list -- this will be the earliest
1525 * commit. Remove it from the list.
1526 */
1527 sxact = (SERIALIZABLEXACT *)
1528 SHMQueueNext(FinishedSerializableTransactions,
1529 FinishedSerializableTransactions,
1530 offsetof(SERIALIZABLEXACT, finishedLink));
1531 SHMQueueDelete(&(sxact->finishedLink));
1532
1533 /* Add to SLRU summary information. */
1534 if (TransactionIdIsValid(sxact->topXid) && !SxactIsReadOnly(sxact))
1535 SerialAdd(sxact->topXid, SxactHasConflictOut(sxact)
1536 ? sxact->SeqNo.earliestOutConflictCommit : InvalidSerCommitSeqNo);
1537
1538 /* Summarize and release the detail. */
1539 ReleaseOneSerializableXact(sxact, false, true);
1540
1541 LWLockRelease(SerializableFinishedListLock);
1542 }
1543
1544 /*
1545 * GetSafeSnapshot
1546 * Obtain and register a snapshot for a READ ONLY DEFERRABLE
1547 * transaction. Ensures that the snapshot is "safe", i.e. a
1548 * read-only transaction running on it can execute serializably
1549 * without further checks. This requires waiting for concurrent
1550 * transactions to complete, and retrying with a new snapshot if
1551 * one of them could possibly create a conflict.
1552 *
1553 * As with GetSerializableTransactionSnapshot (which this is a subroutine
1554 * for), the passed-in Snapshot pointer should reference a static data
1555 * area that can safely be passed to GetSnapshotData.
1556 */
1557 static Snapshot
GetSafeSnapshot(Snapshot origSnapshot)1558 GetSafeSnapshot(Snapshot origSnapshot)
1559 {
1560 Snapshot snapshot;
1561
1562 Assert(XactReadOnly && XactDeferrable);
1563
1564 while (true)
1565 {
1566 /*
1567 * GetSerializableTransactionSnapshotInt is going to call
1568 * GetSnapshotData, so we need to provide it the static snapshot area
1569 * our caller passed to us. The pointer returned is actually the same
1570 * one passed to it, but we avoid assuming that here.
1571 */
1572 snapshot = GetSerializableTransactionSnapshotInt(origSnapshot,
1573 NULL, InvalidPid);
1574
1575 if (MySerializableXact == InvalidSerializableXact)
1576 return snapshot; /* no concurrent r/w xacts; it's safe */
1577
1578 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
1579
1580 /*
1581 * Wait for concurrent transactions to finish. Stop early if one of
1582 * them marked us as conflicted.
1583 */
1584 MySerializableXact->flags |= SXACT_FLAG_DEFERRABLE_WAITING;
1585 while (!(SHMQueueEmpty(&MySerializableXact->possibleUnsafeConflicts) ||
1586 SxactIsROUnsafe(MySerializableXact)))
1587 {
1588 LWLockRelease(SerializableXactHashLock);
1589 ProcWaitForSignal(WAIT_EVENT_SAFE_SNAPSHOT);
1590 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
1591 }
1592 MySerializableXact->flags &= ~SXACT_FLAG_DEFERRABLE_WAITING;
1593
1594 if (!SxactIsROUnsafe(MySerializableXact))
1595 {
1596 LWLockRelease(SerializableXactHashLock);
1597 break; /* success */
1598 }
1599
1600 LWLockRelease(SerializableXactHashLock);
1601
1602 /* else, need to retry... */
1603 ereport(DEBUG2,
1604 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
1605 errmsg_internal("deferrable snapshot was unsafe; trying a new one")));
1606 ReleasePredicateLocks(false, false);
1607 }
1608
1609 /*
1610 * Now we have a safe snapshot, so we don't need to do any further checks.
1611 */
1612 Assert(SxactIsROSafe(MySerializableXact));
1613 ReleasePredicateLocks(false, true);
1614
1615 return snapshot;
1616 }
1617
1618 /*
1619 * GetSafeSnapshotBlockingPids
1620 * If the specified process is currently blocked in GetSafeSnapshot,
1621 * write the process IDs of all processes that it is blocked by
1622 * into the caller-supplied buffer output[]. The list is truncated at
1623 * output_size, and the number of PIDs written into the buffer is
1624 * returned. Returns zero if the given PID is not currently blocked
1625 * in GetSafeSnapshot.
1626 */
1627 int
GetSafeSnapshotBlockingPids(int blocked_pid,int * output,int output_size)1628 GetSafeSnapshotBlockingPids(int blocked_pid, int *output, int output_size)
1629 {
1630 int num_written = 0;
1631 SERIALIZABLEXACT *sxact;
1632
1633 LWLockAcquire(SerializableXactHashLock, LW_SHARED);
1634
1635 /* Find blocked_pid's SERIALIZABLEXACT by linear search. */
1636 for (sxact = FirstPredXact(); sxact != NULL; sxact = NextPredXact(sxact))
1637 {
1638 if (sxact->pid == blocked_pid)
1639 break;
1640 }
1641
1642 /* Did we find it, and is it currently waiting in GetSafeSnapshot? */
1643 if (sxact != NULL && SxactIsDeferrableWaiting(sxact))
1644 {
1645 RWConflict possibleUnsafeConflict;
1646
1647 /* Traverse the list of possible unsafe conflicts collecting PIDs. */
1648 possibleUnsafeConflict = (RWConflict)
1649 SHMQueueNext(&sxact->possibleUnsafeConflicts,
1650 &sxact->possibleUnsafeConflicts,
1651 offsetof(RWConflictData, inLink));
1652
1653 while (possibleUnsafeConflict != NULL && num_written < output_size)
1654 {
1655 output[num_written++] = possibleUnsafeConflict->sxactOut->pid;
1656 possibleUnsafeConflict = (RWConflict)
1657 SHMQueueNext(&sxact->possibleUnsafeConflicts,
1658 &possibleUnsafeConflict->inLink,
1659 offsetof(RWConflictData, inLink));
1660 }
1661 }
1662
1663 LWLockRelease(SerializableXactHashLock);
1664
1665 return num_written;
1666 }
1667
1668 /*
1669 * Acquire a snapshot that can be used for the current transaction.
1670 *
1671 * Make sure we have a SERIALIZABLEXACT reference in MySerializableXact.
1672 * It should be current for this process and be contained in PredXact.
1673 *
1674 * The passed-in Snapshot pointer should reference a static data area that
1675 * can safely be passed to GetSnapshotData. The return value is actually
1676 * always this same pointer; no new snapshot data structure is allocated
1677 * within this function.
1678 */
1679 Snapshot
GetSerializableTransactionSnapshot(Snapshot snapshot)1680 GetSerializableTransactionSnapshot(Snapshot snapshot)
1681 {
1682 Assert(IsolationIsSerializable());
1683
1684 /*
1685 * Can't use serializable mode while recovery is still active, as it is,
1686 * for example, on a hot standby. We could get here despite the check in
1687 * check_XactIsoLevel() if default_transaction_isolation is set to
1688 * serializable, so phrase the hint accordingly.
1689 */
1690 if (RecoveryInProgress())
1691 ereport(ERROR,
1692 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1693 errmsg("cannot use serializable mode in a hot standby"),
1694 errdetail("\"default_transaction_isolation\" is set to \"serializable\"."),
1695 errhint("You can use \"SET default_transaction_isolation = 'repeatable read'\" to change the default.")));
1696
1697 /*
1698 * A special optimization is available for SERIALIZABLE READ ONLY
1699 * DEFERRABLE transactions -- we can wait for a suitable snapshot and
1700 * thereby avoid all SSI overhead once it's running.
1701 */
1702 if (XactReadOnly && XactDeferrable)
1703 return GetSafeSnapshot(snapshot);
1704
1705 return GetSerializableTransactionSnapshotInt(snapshot,
1706 NULL, InvalidPid);
1707 }
1708
1709 /*
1710 * Import a snapshot to be used for the current transaction.
1711 *
1712 * This is nearly the same as GetSerializableTransactionSnapshot, except that
1713 * we don't take a new snapshot, but rather use the data we're handed.
1714 *
1715 * The caller must have verified that the snapshot came from a serializable
1716 * transaction; and if we're read-write, the source transaction must not be
1717 * read-only.
1718 */
1719 void
SetSerializableTransactionSnapshot(Snapshot snapshot,VirtualTransactionId * sourcevxid,int sourcepid)1720 SetSerializableTransactionSnapshot(Snapshot snapshot,
1721 VirtualTransactionId *sourcevxid,
1722 int sourcepid)
1723 {
1724 Assert(IsolationIsSerializable());
1725
1726 /*
1727 * If this is called by parallel.c in a parallel worker, we don't want to
1728 * create a SERIALIZABLEXACT just yet because the leader's
1729 * SERIALIZABLEXACT will be installed with AttachSerializableXact(). We
1730 * also don't want to reject SERIALIZABLE READ ONLY DEFERRABLE in this
1731 * case, because the leader has already determined that the snapshot it
1732 * has passed us is safe. So there is nothing for us to do.
1733 */
1734 if (IsParallelWorker())
1735 return;
1736
1737 /*
1738 * We do not allow SERIALIZABLE READ ONLY DEFERRABLE transactions to
1739 * import snapshots, since there's no way to wait for a safe snapshot when
1740 * we're using the snap we're told to. (XXX instead of throwing an error,
1741 * we could just ignore the XactDeferrable flag?)
1742 */
1743 if (XactReadOnly && XactDeferrable)
1744 ereport(ERROR,
1745 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1746 errmsg("a snapshot-importing transaction must not be READ ONLY DEFERRABLE")));
1747
1748 (void) GetSerializableTransactionSnapshotInt(snapshot, sourcevxid,
1749 sourcepid);
1750 }
1751
1752 /*
1753 * Guts of GetSerializableTransactionSnapshot
1754 *
1755 * If sourcevxid is valid, this is actually an import operation and we should
1756 * skip calling GetSnapshotData, because the snapshot contents are already
1757 * loaded up. HOWEVER: to avoid race conditions, we must check that the
1758 * source xact is still running after we acquire SerializableXactHashLock.
1759 * We do that by calling ProcArrayInstallImportedXmin.
1760 */
1761 static Snapshot
GetSerializableTransactionSnapshotInt(Snapshot snapshot,VirtualTransactionId * sourcevxid,int sourcepid)1762 GetSerializableTransactionSnapshotInt(Snapshot snapshot,
1763 VirtualTransactionId *sourcevxid,
1764 int sourcepid)
1765 {
1766 PGPROC *proc;
1767 VirtualTransactionId vxid;
1768 SERIALIZABLEXACT *sxact,
1769 *othersxact;
1770
1771 /* We only do this for serializable transactions. Once. */
1772 Assert(MySerializableXact == InvalidSerializableXact);
1773
1774 Assert(!RecoveryInProgress());
1775
1776 /*
1777 * Since all parts of a serializable transaction must use the same
1778 * snapshot, it is too late to establish one after a parallel operation
1779 * has begun.
1780 */
1781 if (IsInParallelMode())
1782 elog(ERROR, "cannot establish serializable snapshot during a parallel operation");
1783
1784 proc = MyProc;
1785 Assert(proc != NULL);
1786 GET_VXID_FROM_PGPROC(vxid, *proc);
1787
1788 /*
1789 * First we get the sxact structure, which may involve looping and access
1790 * to the "finished" list to free a structure for use.
1791 *
1792 * We must hold SerializableXactHashLock when taking/checking the snapshot
1793 * to avoid race conditions, for much the same reasons that
1794 * GetSnapshotData takes the ProcArrayLock. Since we might have to
1795 * release SerializableXactHashLock to call SummarizeOldestCommittedSxact,
1796 * this means we have to create the sxact first, which is a bit annoying
1797 * (in particular, an elog(ERROR) in procarray.c would cause us to leak
1798 * the sxact). Consider refactoring to avoid this.
1799 */
1800 #ifdef TEST_SUMMARIZE_SERIAL
1801 SummarizeOldestCommittedSxact();
1802 #endif
1803 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
1804 do
1805 {
1806 sxact = CreatePredXact();
1807 /* If null, push out committed sxact to SLRU summary & retry. */
1808 if (!sxact)
1809 {
1810 LWLockRelease(SerializableXactHashLock);
1811 SummarizeOldestCommittedSxact();
1812 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
1813 }
1814 } while (!sxact);
1815
1816 /* Get the snapshot, or check that it's safe to use */
1817 if (!sourcevxid)
1818 snapshot = GetSnapshotData(snapshot);
1819 else if (!ProcArrayInstallImportedXmin(snapshot->xmin, sourcevxid))
1820 {
1821 ReleasePredXact(sxact);
1822 LWLockRelease(SerializableXactHashLock);
1823 ereport(ERROR,
1824 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1825 errmsg("could not import the requested snapshot"),
1826 errdetail("The source process with PID %d is not running anymore.",
1827 sourcepid)));
1828 }
1829
1830 /*
1831 * If there are no serializable transactions which are not read-only, we
1832 * can "opt out" of predicate locking and conflict checking for a
1833 * read-only transaction.
1834 *
1835 * The reason this is safe is that a read-only transaction can only become
1836 * part of a dangerous structure if it overlaps a writable transaction
1837 * which in turn overlaps a writable transaction which committed before
1838 * the read-only transaction started. A new writable transaction can
1839 * overlap this one, but it can't meet the other condition of overlapping
1840 * a transaction which committed before this one started.
1841 */
1842 if (XactReadOnly && PredXact->WritableSxactCount == 0)
1843 {
1844 ReleasePredXact(sxact);
1845 LWLockRelease(SerializableXactHashLock);
1846 return snapshot;
1847 }
1848
1849 /* Maintain serializable global xmin info. */
1850 if (!TransactionIdIsValid(PredXact->SxactGlobalXmin))
1851 {
1852 Assert(PredXact->SxactGlobalXminCount == 0);
1853 PredXact->SxactGlobalXmin = snapshot->xmin;
1854 PredXact->SxactGlobalXminCount = 1;
1855 SerialSetActiveSerXmin(snapshot->xmin);
1856 }
1857 else if (TransactionIdEquals(snapshot->xmin, PredXact->SxactGlobalXmin))
1858 {
1859 Assert(PredXact->SxactGlobalXminCount > 0);
1860 PredXact->SxactGlobalXminCount++;
1861 }
1862 else
1863 {
1864 Assert(TransactionIdFollows(snapshot->xmin, PredXact->SxactGlobalXmin));
1865 }
1866
1867 /* Initialize the structure. */
1868 sxact->vxid = vxid;
1869 sxact->SeqNo.lastCommitBeforeSnapshot = PredXact->LastSxactCommitSeqNo;
1870 sxact->prepareSeqNo = InvalidSerCommitSeqNo;
1871 sxact->commitSeqNo = InvalidSerCommitSeqNo;
1872 SHMQueueInit(&(sxact->outConflicts));
1873 SHMQueueInit(&(sxact->inConflicts));
1874 SHMQueueInit(&(sxact->possibleUnsafeConflicts));
1875 sxact->topXid = GetTopTransactionIdIfAny();
1876 sxact->finishedBefore = InvalidTransactionId;
1877 sxact->xmin = snapshot->xmin;
1878 sxact->pid = MyProcPid;
1879 SHMQueueInit(&(sxact->predicateLocks));
1880 SHMQueueElemInit(&(sxact->finishedLink));
1881 sxact->flags = 0;
1882 if (XactReadOnly)
1883 {
1884 sxact->flags |= SXACT_FLAG_READ_ONLY;
1885
1886 /*
1887 * Register all concurrent r/w transactions as possible conflicts; if
1888 * all of them commit without any outgoing conflicts to earlier
1889 * transactions then this snapshot can be deemed safe (and we can run
1890 * without tracking predicate locks).
1891 */
1892 for (othersxact = FirstPredXact();
1893 othersxact != NULL;
1894 othersxact = NextPredXact(othersxact))
1895 {
1896 if (!SxactIsCommitted(othersxact)
1897 && !SxactIsDoomed(othersxact)
1898 && !SxactIsReadOnly(othersxact))
1899 {
1900 SetPossibleUnsafeConflict(sxact, othersxact);
1901 }
1902 }
1903 }
1904 else
1905 {
1906 ++(PredXact->WritableSxactCount);
1907 Assert(PredXact->WritableSxactCount <=
1908 (MaxBackends + max_prepared_xacts));
1909 }
1910
1911 MySerializableXact = sxact;
1912 MyXactDidWrite = false; /* haven't written anything yet */
1913
1914 LWLockRelease(SerializableXactHashLock);
1915
1916 CreateLocalPredicateLockHash();
1917
1918 return snapshot;
1919 }
1920
1921 static void
CreateLocalPredicateLockHash(void)1922 CreateLocalPredicateLockHash(void)
1923 {
1924 HASHCTL hash_ctl;
1925
1926 /* Initialize the backend-local hash table of parent locks */
1927 Assert(LocalPredicateLockHash == NULL);
1928 hash_ctl.keysize = sizeof(PREDICATELOCKTARGETTAG);
1929 hash_ctl.entrysize = sizeof(LOCALPREDICATELOCK);
1930 LocalPredicateLockHash = hash_create("Local predicate lock",
1931 max_predicate_locks_per_xact,
1932 &hash_ctl,
1933 HASH_ELEM | HASH_BLOBS);
1934 }
1935
1936 /*
1937 * Register the top level XID in SerializableXidHash.
1938 * Also store it for easy reference in MySerializableXact.
1939 */
1940 void
RegisterPredicateLockingXid(TransactionId xid)1941 RegisterPredicateLockingXid(TransactionId xid)
1942 {
1943 SERIALIZABLEXIDTAG sxidtag;
1944 SERIALIZABLEXID *sxid;
1945 bool found;
1946
1947 /*
1948 * If we're not tracking predicate lock data for this transaction, we
1949 * should ignore the request and return quickly.
1950 */
1951 if (MySerializableXact == InvalidSerializableXact)
1952 return;
1953
1954 /* We should have a valid XID and be at the top level. */
1955 Assert(TransactionIdIsValid(xid));
1956
1957 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
1958
1959 /* This should only be done once per transaction. */
1960 Assert(MySerializableXact->topXid == InvalidTransactionId);
1961
1962 MySerializableXact->topXid = xid;
1963
1964 sxidtag.xid = xid;
1965 sxid = (SERIALIZABLEXID *) hash_search(SerializableXidHash,
1966 &sxidtag,
1967 HASH_ENTER, &found);
1968 Assert(!found);
1969
1970 /* Initialize the structure. */
1971 sxid->myXact = MySerializableXact;
1972 LWLockRelease(SerializableXactHashLock);
1973 }
1974
1975
1976 /*
1977 * Check whether there are any predicate locks held by any transaction
1978 * for the page at the given block number.
1979 *
1980 * Note that the transaction may be completed but not yet subject to
1981 * cleanup due to overlapping serializable transactions. This must
1982 * return valid information regardless of transaction isolation level.
1983 *
1984 * Also note that this doesn't check for a conflicting relation lock,
1985 * just a lock specifically on the given page.
1986 *
1987 * One use is to support proper behavior during GiST index vacuum.
1988 */
1989 bool
PageIsPredicateLocked(Relation relation,BlockNumber blkno)1990 PageIsPredicateLocked(Relation relation, BlockNumber blkno)
1991 {
1992 PREDICATELOCKTARGETTAG targettag;
1993 uint32 targettaghash;
1994 LWLock *partitionLock;
1995 PREDICATELOCKTARGET *target;
1996
1997 SET_PREDICATELOCKTARGETTAG_PAGE(targettag,
1998 relation->rd_node.dbNode,
1999 relation->rd_id,
2000 blkno);
2001
2002 targettaghash = PredicateLockTargetTagHashCode(&targettag);
2003 partitionLock = PredicateLockHashPartitionLock(targettaghash);
2004 LWLockAcquire(partitionLock, LW_SHARED);
2005 target = (PREDICATELOCKTARGET *)
2006 hash_search_with_hash_value(PredicateLockTargetHash,
2007 &targettag, targettaghash,
2008 HASH_FIND, NULL);
2009 LWLockRelease(partitionLock);
2010
2011 return (target != NULL);
2012 }
2013
2014
2015 /*
2016 * Check whether a particular lock is held by this transaction.
2017 *
2018 * Important note: this function may return false even if the lock is
2019 * being held, because it uses the local lock table which is not
2020 * updated if another transaction modifies our lock list (e.g. to
2021 * split an index page). It can also return true when a coarser
2022 * granularity lock that covers this target is being held. Be careful
2023 * to only use this function in circumstances where such errors are
2024 * acceptable!
2025 */
2026 static bool
PredicateLockExists(const PREDICATELOCKTARGETTAG * targettag)2027 PredicateLockExists(const PREDICATELOCKTARGETTAG *targettag)
2028 {
2029 LOCALPREDICATELOCK *lock;
2030
2031 /* check local hash table */
2032 lock = (LOCALPREDICATELOCK *) hash_search(LocalPredicateLockHash,
2033 targettag,
2034 HASH_FIND, NULL);
2035
2036 if (!lock)
2037 return false;
2038
2039 /*
2040 * Found entry in the table, but still need to check whether it's actually
2041 * held -- it could just be a parent of some held lock.
2042 */
2043 return lock->held;
2044 }
2045
2046 /*
2047 * Return the parent lock tag in the lock hierarchy: the next coarser
2048 * lock that covers the provided tag.
2049 *
2050 * Returns true and sets *parent to the parent tag if one exists,
2051 * returns false if none exists.
2052 */
2053 static bool
GetParentPredicateLockTag(const PREDICATELOCKTARGETTAG * tag,PREDICATELOCKTARGETTAG * parent)2054 GetParentPredicateLockTag(const PREDICATELOCKTARGETTAG *tag,
2055 PREDICATELOCKTARGETTAG *parent)
2056 {
2057 switch (GET_PREDICATELOCKTARGETTAG_TYPE(*tag))
2058 {
2059 case PREDLOCKTAG_RELATION:
2060 /* relation locks have no parent lock */
2061 return false;
2062
2063 case PREDLOCKTAG_PAGE:
2064 /* parent lock is relation lock */
2065 SET_PREDICATELOCKTARGETTAG_RELATION(*parent,
2066 GET_PREDICATELOCKTARGETTAG_DB(*tag),
2067 GET_PREDICATELOCKTARGETTAG_RELATION(*tag));
2068
2069 return true;
2070
2071 case PREDLOCKTAG_TUPLE:
2072 /* parent lock is page lock */
2073 SET_PREDICATELOCKTARGETTAG_PAGE(*parent,
2074 GET_PREDICATELOCKTARGETTAG_DB(*tag),
2075 GET_PREDICATELOCKTARGETTAG_RELATION(*tag),
2076 GET_PREDICATELOCKTARGETTAG_PAGE(*tag));
2077 return true;
2078 }
2079
2080 /* not reachable */
2081 Assert(false);
2082 return false;
2083 }
2084
2085 /*
2086 * Check whether the lock we are considering is already covered by a
2087 * coarser lock for our transaction.
2088 *
2089 * Like PredicateLockExists, this function might return a false
2090 * negative, but it will never return a false positive.
2091 */
2092 static bool
CoarserLockCovers(const PREDICATELOCKTARGETTAG * newtargettag)2093 CoarserLockCovers(const PREDICATELOCKTARGETTAG *newtargettag)
2094 {
2095 PREDICATELOCKTARGETTAG targettag,
2096 parenttag;
2097
2098 targettag = *newtargettag;
2099
2100 /* check parents iteratively until no more */
2101 while (GetParentPredicateLockTag(&targettag, &parenttag))
2102 {
2103 targettag = parenttag;
2104 if (PredicateLockExists(&targettag))
2105 return true;
2106 }
2107
2108 /* no more parents to check; lock is not covered */
2109 return false;
2110 }
2111
2112 /*
2113 * Remove the dummy entry from the predicate lock target hash, to free up some
2114 * scratch space. The caller must be holding SerializablePredicateListLock,
2115 * and must restore the entry with RestoreScratchTarget() before releasing the
2116 * lock.
2117 *
2118 * If lockheld is true, the caller is already holding the partition lock
2119 * of the partition containing the scratch entry.
2120 */
2121 static void
RemoveScratchTarget(bool lockheld)2122 RemoveScratchTarget(bool lockheld)
2123 {
2124 bool found;
2125
2126 Assert(LWLockHeldByMe(SerializablePredicateListLock));
2127
2128 if (!lockheld)
2129 LWLockAcquire(ScratchPartitionLock, LW_EXCLUSIVE);
2130 hash_search_with_hash_value(PredicateLockTargetHash,
2131 &ScratchTargetTag,
2132 ScratchTargetTagHash,
2133 HASH_REMOVE, &found);
2134 Assert(found);
2135 if (!lockheld)
2136 LWLockRelease(ScratchPartitionLock);
2137 }
2138
2139 /*
2140 * Re-insert the dummy entry in predicate lock target hash.
2141 */
2142 static void
RestoreScratchTarget(bool lockheld)2143 RestoreScratchTarget(bool lockheld)
2144 {
2145 bool found;
2146
2147 Assert(LWLockHeldByMe(SerializablePredicateListLock));
2148
2149 if (!lockheld)
2150 LWLockAcquire(ScratchPartitionLock, LW_EXCLUSIVE);
2151 hash_search_with_hash_value(PredicateLockTargetHash,
2152 &ScratchTargetTag,
2153 ScratchTargetTagHash,
2154 HASH_ENTER, &found);
2155 Assert(!found);
2156 if (!lockheld)
2157 LWLockRelease(ScratchPartitionLock);
2158 }
2159
2160 /*
2161 * Check whether the list of related predicate locks is empty for a
2162 * predicate lock target, and remove the target if it is.
2163 */
2164 static void
RemoveTargetIfNoLongerUsed(PREDICATELOCKTARGET * target,uint32 targettaghash)2165 RemoveTargetIfNoLongerUsed(PREDICATELOCKTARGET *target, uint32 targettaghash)
2166 {
2167 PREDICATELOCKTARGET *rmtarget PG_USED_FOR_ASSERTS_ONLY;
2168
2169 Assert(LWLockHeldByMe(SerializablePredicateListLock));
2170
2171 /* Can't remove it until no locks at this target. */
2172 if (!SHMQueueEmpty(&target->predicateLocks))
2173 return;
2174
2175 /* Actually remove the target. */
2176 rmtarget = hash_search_with_hash_value(PredicateLockTargetHash,
2177 &target->tag,
2178 targettaghash,
2179 HASH_REMOVE, NULL);
2180 Assert(rmtarget == target);
2181 }
2182
2183 /*
2184 * Delete child target locks owned by this process.
2185 * This implementation is assuming that the usage of each target tag field
2186 * is uniform. No need to make this hard if we don't have to.
2187 *
2188 * We acquire an LWLock in the case of parallel mode, because worker
2189 * backends have access to the leader's SERIALIZABLEXACT. Otherwise,
2190 * we aren't acquiring LWLocks for the predicate lock or lock
2191 * target structures associated with this transaction unless we're going
2192 * to modify them, because no other process is permitted to modify our
2193 * locks.
2194 */
2195 static void
DeleteChildTargetLocks(const PREDICATELOCKTARGETTAG * newtargettag)2196 DeleteChildTargetLocks(const PREDICATELOCKTARGETTAG *newtargettag)
2197 {
2198 SERIALIZABLEXACT *sxact;
2199 PREDICATELOCK *predlock;
2200
2201 LWLockAcquire(SerializablePredicateListLock, LW_SHARED);
2202 sxact = MySerializableXact;
2203 if (IsInParallelMode())
2204 LWLockAcquire(&sxact->perXactPredicateListLock, LW_EXCLUSIVE);
2205 predlock = (PREDICATELOCK *)
2206 SHMQueueNext(&(sxact->predicateLocks),
2207 &(sxact->predicateLocks),
2208 offsetof(PREDICATELOCK, xactLink));
2209 while (predlock)
2210 {
2211 SHM_QUEUE *predlocksxactlink;
2212 PREDICATELOCK *nextpredlock;
2213 PREDICATELOCKTAG oldlocktag;
2214 PREDICATELOCKTARGET *oldtarget;
2215 PREDICATELOCKTARGETTAG oldtargettag;
2216
2217 predlocksxactlink = &(predlock->xactLink);
2218 nextpredlock = (PREDICATELOCK *)
2219 SHMQueueNext(&(sxact->predicateLocks),
2220 predlocksxactlink,
2221 offsetof(PREDICATELOCK, xactLink));
2222
2223 oldlocktag = predlock->tag;
2224 Assert(oldlocktag.myXact == sxact);
2225 oldtarget = oldlocktag.myTarget;
2226 oldtargettag = oldtarget->tag;
2227
2228 if (TargetTagIsCoveredBy(oldtargettag, *newtargettag))
2229 {
2230 uint32 oldtargettaghash;
2231 LWLock *partitionLock;
2232 PREDICATELOCK *rmpredlock PG_USED_FOR_ASSERTS_ONLY;
2233
2234 oldtargettaghash = PredicateLockTargetTagHashCode(&oldtargettag);
2235 partitionLock = PredicateLockHashPartitionLock(oldtargettaghash);
2236
2237 LWLockAcquire(partitionLock, LW_EXCLUSIVE);
2238
2239 SHMQueueDelete(predlocksxactlink);
2240 SHMQueueDelete(&(predlock->targetLink));
2241 rmpredlock = hash_search_with_hash_value
2242 (PredicateLockHash,
2243 &oldlocktag,
2244 PredicateLockHashCodeFromTargetHashCode(&oldlocktag,
2245 oldtargettaghash),
2246 HASH_REMOVE, NULL);
2247 Assert(rmpredlock == predlock);
2248
2249 RemoveTargetIfNoLongerUsed(oldtarget, oldtargettaghash);
2250
2251 LWLockRelease(partitionLock);
2252
2253 DecrementParentLocks(&oldtargettag);
2254 }
2255
2256 predlock = nextpredlock;
2257 }
2258 if (IsInParallelMode())
2259 LWLockRelease(&sxact->perXactPredicateListLock);
2260 LWLockRelease(SerializablePredicateListLock);
2261 }
2262
2263 /*
2264 * Returns the promotion limit for a given predicate lock target. This is the
2265 * max number of descendant locks allowed before promoting to the specified
2266 * tag. Note that the limit includes non-direct descendants (e.g., both tuples
2267 * and pages for a relation lock).
2268 *
2269 * Currently the default limit is 2 for a page lock, and half of the value of
2270 * max_pred_locks_per_transaction - 1 for a relation lock, to match behavior
2271 * of earlier releases when upgrading.
2272 *
2273 * TODO SSI: We should probably add additional GUCs to allow a maximum ratio
2274 * of page and tuple locks based on the pages in a relation, and the maximum
2275 * ratio of tuple locks to tuples in a page. This would provide more
2276 * generally "balanced" allocation of locks to where they are most useful,
2277 * while still allowing the absolute numbers to prevent one relation from
2278 * tying up all predicate lock resources.
2279 */
2280 static int
MaxPredicateChildLocks(const PREDICATELOCKTARGETTAG * tag)2281 MaxPredicateChildLocks(const PREDICATELOCKTARGETTAG *tag)
2282 {
2283 switch (GET_PREDICATELOCKTARGETTAG_TYPE(*tag))
2284 {
2285 case PREDLOCKTAG_RELATION:
2286 return max_predicate_locks_per_relation < 0
2287 ? (max_predicate_locks_per_xact
2288 / (-max_predicate_locks_per_relation)) - 1
2289 : max_predicate_locks_per_relation;
2290
2291 case PREDLOCKTAG_PAGE:
2292 return max_predicate_locks_per_page;
2293
2294 case PREDLOCKTAG_TUPLE:
2295
2296 /*
2297 * not reachable: nothing is finer-granularity than a tuple, so we
2298 * should never try to promote to it.
2299 */
2300 Assert(false);
2301 return 0;
2302 }
2303
2304 /* not reachable */
2305 Assert(false);
2306 return 0;
2307 }
2308
2309 /*
2310 * For all ancestors of a newly-acquired predicate lock, increment
2311 * their child count in the parent hash table. If any of them have
2312 * more descendants than their promotion threshold, acquire the
2313 * coarsest such lock.
2314 *
2315 * Returns true if a parent lock was acquired and false otherwise.
2316 */
2317 static bool
CheckAndPromotePredicateLockRequest(const PREDICATELOCKTARGETTAG * reqtag)2318 CheckAndPromotePredicateLockRequest(const PREDICATELOCKTARGETTAG *reqtag)
2319 {
2320 PREDICATELOCKTARGETTAG targettag,
2321 nexttag,
2322 promotiontag;
2323 LOCALPREDICATELOCK *parentlock;
2324 bool found,
2325 promote;
2326
2327 promote = false;
2328
2329 targettag = *reqtag;
2330
2331 /* check parents iteratively */
2332 while (GetParentPredicateLockTag(&targettag, &nexttag))
2333 {
2334 targettag = nexttag;
2335 parentlock = (LOCALPREDICATELOCK *) hash_search(LocalPredicateLockHash,
2336 &targettag,
2337 HASH_ENTER,
2338 &found);
2339 if (!found)
2340 {
2341 parentlock->held = false;
2342 parentlock->childLocks = 1;
2343 }
2344 else
2345 parentlock->childLocks++;
2346
2347 if (parentlock->childLocks >
2348 MaxPredicateChildLocks(&targettag))
2349 {
2350 /*
2351 * We should promote to this parent lock. Continue to check its
2352 * ancestors, however, both to get their child counts right and to
2353 * check whether we should just go ahead and promote to one of
2354 * them.
2355 */
2356 promotiontag = targettag;
2357 promote = true;
2358 }
2359 }
2360
2361 if (promote)
2362 {
2363 /* acquire coarsest ancestor eligible for promotion */
2364 PredicateLockAcquire(&promotiontag);
2365 return true;
2366 }
2367 else
2368 return false;
2369 }
2370
2371 /*
2372 * When releasing a lock, decrement the child count on all ancestor
2373 * locks.
2374 *
2375 * This is called only when releasing a lock via
2376 * DeleteChildTargetLocks (i.e. when a lock becomes redundant because
2377 * we've acquired its parent, possibly due to promotion) or when a new
2378 * MVCC write lock makes the predicate lock unnecessary. There's no
2379 * point in calling it when locks are released at transaction end, as
2380 * this information is no longer needed.
2381 */
2382 static void
DecrementParentLocks(const PREDICATELOCKTARGETTAG * targettag)2383 DecrementParentLocks(const PREDICATELOCKTARGETTAG *targettag)
2384 {
2385 PREDICATELOCKTARGETTAG parenttag,
2386 nexttag;
2387
2388 parenttag = *targettag;
2389
2390 while (GetParentPredicateLockTag(&parenttag, &nexttag))
2391 {
2392 uint32 targettaghash;
2393 LOCALPREDICATELOCK *parentlock,
2394 *rmlock PG_USED_FOR_ASSERTS_ONLY;
2395
2396 parenttag = nexttag;
2397 targettaghash = PredicateLockTargetTagHashCode(&parenttag);
2398 parentlock = (LOCALPREDICATELOCK *)
2399 hash_search_with_hash_value(LocalPredicateLockHash,
2400 &parenttag, targettaghash,
2401 HASH_FIND, NULL);
2402
2403 /*
2404 * There's a small chance the parent lock doesn't exist in the lock
2405 * table. This can happen if we prematurely removed it because an
2406 * index split caused the child refcount to be off.
2407 */
2408 if (parentlock == NULL)
2409 continue;
2410
2411 parentlock->childLocks--;
2412
2413 /*
2414 * Under similar circumstances the parent lock's refcount might be
2415 * zero. This only happens if we're holding that lock (otherwise we
2416 * would have removed the entry).
2417 */
2418 if (parentlock->childLocks < 0)
2419 {
2420 Assert(parentlock->held);
2421 parentlock->childLocks = 0;
2422 }
2423
2424 if ((parentlock->childLocks == 0) && (!parentlock->held))
2425 {
2426 rmlock = (LOCALPREDICATELOCK *)
2427 hash_search_with_hash_value(LocalPredicateLockHash,
2428 &parenttag, targettaghash,
2429 HASH_REMOVE, NULL);
2430 Assert(rmlock == parentlock);
2431 }
2432 }
2433 }
2434
2435 /*
2436 * Indicate that a predicate lock on the given target is held by the
2437 * specified transaction. Has no effect if the lock is already held.
2438 *
2439 * This updates the lock table and the sxact's lock list, and creates
2440 * the lock target if necessary, but does *not* do anything related to
2441 * granularity promotion or the local lock table. See
2442 * PredicateLockAcquire for that.
2443 */
2444 static void
CreatePredicateLock(const PREDICATELOCKTARGETTAG * targettag,uint32 targettaghash,SERIALIZABLEXACT * sxact)2445 CreatePredicateLock(const PREDICATELOCKTARGETTAG *targettag,
2446 uint32 targettaghash,
2447 SERIALIZABLEXACT *sxact)
2448 {
2449 PREDICATELOCKTARGET *target;
2450 PREDICATELOCKTAG locktag;
2451 PREDICATELOCK *lock;
2452 LWLock *partitionLock;
2453 bool found;
2454
2455 partitionLock = PredicateLockHashPartitionLock(targettaghash);
2456
2457 LWLockAcquire(SerializablePredicateListLock, LW_SHARED);
2458 if (IsInParallelMode())
2459 LWLockAcquire(&sxact->perXactPredicateListLock, LW_EXCLUSIVE);
2460 LWLockAcquire(partitionLock, LW_EXCLUSIVE);
2461
2462 /* Make sure that the target is represented. */
2463 target = (PREDICATELOCKTARGET *)
2464 hash_search_with_hash_value(PredicateLockTargetHash,
2465 targettag, targettaghash,
2466 HASH_ENTER_NULL, &found);
2467 if (!target)
2468 ereport(ERROR,
2469 (errcode(ERRCODE_OUT_OF_MEMORY),
2470 errmsg("out of shared memory"),
2471 errhint("You might need to increase max_pred_locks_per_transaction.")));
2472 if (!found)
2473 SHMQueueInit(&(target->predicateLocks));
2474
2475 /* We've got the sxact and target, make sure they're joined. */
2476 locktag.myTarget = target;
2477 locktag.myXact = sxact;
2478 lock = (PREDICATELOCK *)
2479 hash_search_with_hash_value(PredicateLockHash, &locktag,
2480 PredicateLockHashCodeFromTargetHashCode(&locktag, targettaghash),
2481 HASH_ENTER_NULL, &found);
2482 if (!lock)
2483 ereport(ERROR,
2484 (errcode(ERRCODE_OUT_OF_MEMORY),
2485 errmsg("out of shared memory"),
2486 errhint("You might need to increase max_pred_locks_per_transaction.")));
2487
2488 if (!found)
2489 {
2490 SHMQueueInsertBefore(&(target->predicateLocks), &(lock->targetLink));
2491 SHMQueueInsertBefore(&(sxact->predicateLocks),
2492 &(lock->xactLink));
2493 lock->commitSeqNo = InvalidSerCommitSeqNo;
2494 }
2495
2496 LWLockRelease(partitionLock);
2497 if (IsInParallelMode())
2498 LWLockRelease(&sxact->perXactPredicateListLock);
2499 LWLockRelease(SerializablePredicateListLock);
2500 }
2501
2502 /*
2503 * Acquire a predicate lock on the specified target for the current
2504 * connection if not already held. This updates the local lock table
2505 * and uses it to implement granularity promotion. It will consolidate
2506 * multiple locks into a coarser lock if warranted, and will release
2507 * any finer-grained locks covered by the new one.
2508 */
2509 static void
PredicateLockAcquire(const PREDICATELOCKTARGETTAG * targettag)2510 PredicateLockAcquire(const PREDICATELOCKTARGETTAG *targettag)
2511 {
2512 uint32 targettaghash;
2513 bool found;
2514 LOCALPREDICATELOCK *locallock;
2515
2516 /* Do we have the lock already, or a covering lock? */
2517 if (PredicateLockExists(targettag))
2518 return;
2519
2520 if (CoarserLockCovers(targettag))
2521 return;
2522
2523 /* the same hash and LW lock apply to the lock target and the local lock. */
2524 targettaghash = PredicateLockTargetTagHashCode(targettag);
2525
2526 /* Acquire lock in local table */
2527 locallock = (LOCALPREDICATELOCK *)
2528 hash_search_with_hash_value(LocalPredicateLockHash,
2529 targettag, targettaghash,
2530 HASH_ENTER, &found);
2531 locallock->held = true;
2532 if (!found)
2533 locallock->childLocks = 0;
2534
2535 /* Actually create the lock */
2536 CreatePredicateLock(targettag, targettaghash, MySerializableXact);
2537
2538 /*
2539 * Lock has been acquired. Check whether it should be promoted to a
2540 * coarser granularity, or whether there are finer-granularity locks to
2541 * clean up.
2542 */
2543 if (CheckAndPromotePredicateLockRequest(targettag))
2544 {
2545 /*
2546 * Lock request was promoted to a coarser-granularity lock, and that
2547 * lock was acquired. It will delete this lock and any of its
2548 * children, so we're done.
2549 */
2550 }
2551 else
2552 {
2553 /* Clean up any finer-granularity locks */
2554 if (GET_PREDICATELOCKTARGETTAG_TYPE(*targettag) != PREDLOCKTAG_TUPLE)
2555 DeleteChildTargetLocks(targettag);
2556 }
2557 }
2558
2559
2560 /*
2561 * PredicateLockRelation
2562 *
2563 * Gets a predicate lock at the relation level.
2564 * Skip if not in full serializable transaction isolation level.
2565 * Skip if this is a temporary table.
2566 * Clear any finer-grained predicate locks this session has on the relation.
2567 */
2568 void
PredicateLockRelation(Relation relation,Snapshot snapshot)2569 PredicateLockRelation(Relation relation, Snapshot snapshot)
2570 {
2571 PREDICATELOCKTARGETTAG tag;
2572
2573 if (!SerializationNeededForRead(relation, snapshot))
2574 return;
2575
2576 SET_PREDICATELOCKTARGETTAG_RELATION(tag,
2577 relation->rd_node.dbNode,
2578 relation->rd_id);
2579 PredicateLockAcquire(&tag);
2580 }
2581
2582 /*
2583 * PredicateLockPage
2584 *
2585 * Gets a predicate lock at the page level.
2586 * Skip if not in full serializable transaction isolation level.
2587 * Skip if this is a temporary table.
2588 * Skip if a coarser predicate lock already covers this page.
2589 * Clear any finer-grained predicate locks this session has on the relation.
2590 */
2591 void
PredicateLockPage(Relation relation,BlockNumber blkno,Snapshot snapshot)2592 PredicateLockPage(Relation relation, BlockNumber blkno, Snapshot snapshot)
2593 {
2594 PREDICATELOCKTARGETTAG tag;
2595
2596 if (!SerializationNeededForRead(relation, snapshot))
2597 return;
2598
2599 SET_PREDICATELOCKTARGETTAG_PAGE(tag,
2600 relation->rd_node.dbNode,
2601 relation->rd_id,
2602 blkno);
2603 PredicateLockAcquire(&tag);
2604 }
2605
2606 /*
2607 * PredicateLockTID
2608 *
2609 * Gets a predicate lock at the tuple level.
2610 * Skip if not in full serializable transaction isolation level.
2611 * Skip if this is a temporary table.
2612 */
2613 void
PredicateLockTID(Relation relation,ItemPointer tid,Snapshot snapshot,TransactionId tuple_xid)2614 PredicateLockTID(Relation relation, ItemPointer tid, Snapshot snapshot,
2615 TransactionId tuple_xid)
2616 {
2617 PREDICATELOCKTARGETTAG tag;
2618
2619 if (!SerializationNeededForRead(relation, snapshot))
2620 return;
2621
2622 /*
2623 * Return if this xact wrote it.
2624 */
2625 if (relation->rd_index == NULL)
2626 {
2627 /* If we wrote it; we already have a write lock. */
2628 if (TransactionIdIsCurrentTransactionId(tuple_xid))
2629 return;
2630 }
2631
2632 /*
2633 * Do quick-but-not-definitive test for a relation lock first. This will
2634 * never cause a return when the relation is *not* locked, but will
2635 * occasionally let the check continue when there really *is* a relation
2636 * level lock.
2637 */
2638 SET_PREDICATELOCKTARGETTAG_RELATION(tag,
2639 relation->rd_node.dbNode,
2640 relation->rd_id);
2641 if (PredicateLockExists(&tag))
2642 return;
2643
2644 SET_PREDICATELOCKTARGETTAG_TUPLE(tag,
2645 relation->rd_node.dbNode,
2646 relation->rd_id,
2647 ItemPointerGetBlockNumber(tid),
2648 ItemPointerGetOffsetNumber(tid));
2649 PredicateLockAcquire(&tag);
2650 }
2651
2652
2653 /*
2654 * DeleteLockTarget
2655 *
2656 * Remove a predicate lock target along with any locks held for it.
2657 *
2658 * Caller must hold SerializablePredicateListLock and the
2659 * appropriate hash partition lock for the target.
2660 */
2661 static void
DeleteLockTarget(PREDICATELOCKTARGET * target,uint32 targettaghash)2662 DeleteLockTarget(PREDICATELOCKTARGET *target, uint32 targettaghash)
2663 {
2664 PREDICATELOCK *predlock;
2665 SHM_QUEUE *predlocktargetlink;
2666 PREDICATELOCK *nextpredlock;
2667 bool found;
2668
2669 Assert(LWLockHeldByMeInMode(SerializablePredicateListLock,
2670 LW_EXCLUSIVE));
2671 Assert(LWLockHeldByMe(PredicateLockHashPartitionLock(targettaghash)));
2672
2673 predlock = (PREDICATELOCK *)
2674 SHMQueueNext(&(target->predicateLocks),
2675 &(target->predicateLocks),
2676 offsetof(PREDICATELOCK, targetLink));
2677 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
2678 while (predlock)
2679 {
2680 predlocktargetlink = &(predlock->targetLink);
2681 nextpredlock = (PREDICATELOCK *)
2682 SHMQueueNext(&(target->predicateLocks),
2683 predlocktargetlink,
2684 offsetof(PREDICATELOCK, targetLink));
2685
2686 SHMQueueDelete(&(predlock->xactLink));
2687 SHMQueueDelete(&(predlock->targetLink));
2688
2689 hash_search_with_hash_value
2690 (PredicateLockHash,
2691 &predlock->tag,
2692 PredicateLockHashCodeFromTargetHashCode(&predlock->tag,
2693 targettaghash),
2694 HASH_REMOVE, &found);
2695 Assert(found);
2696
2697 predlock = nextpredlock;
2698 }
2699 LWLockRelease(SerializableXactHashLock);
2700
2701 /* Remove the target itself, if possible. */
2702 RemoveTargetIfNoLongerUsed(target, targettaghash);
2703 }
2704
2705
2706 /*
2707 * TransferPredicateLocksToNewTarget
2708 *
2709 * Move or copy all the predicate locks for a lock target, for use by
2710 * index page splits/combines and other things that create or replace
2711 * lock targets. If 'removeOld' is true, the old locks and the target
2712 * will be removed.
2713 *
2714 * Returns true on success, or false if we ran out of shared memory to
2715 * allocate the new target or locks. Guaranteed to always succeed if
2716 * removeOld is set (by using the scratch entry in PredicateLockTargetHash
2717 * for scratch space).
2718 *
2719 * Warning: the "removeOld" option should be used only with care,
2720 * because this function does not (indeed, can not) update other
2721 * backends' LocalPredicateLockHash. If we are only adding new
2722 * entries, this is not a problem: the local lock table is used only
2723 * as a hint, so missing entries for locks that are held are
2724 * OK. Having entries for locks that are no longer held, as can happen
2725 * when using "removeOld", is not in general OK. We can only use it
2726 * safely when replacing a lock with a coarser-granularity lock that
2727 * covers it, or if we are absolutely certain that no one will need to
2728 * refer to that lock in the future.
2729 *
2730 * Caller must hold SerializablePredicateListLock exclusively.
2731 */
2732 static bool
TransferPredicateLocksToNewTarget(PREDICATELOCKTARGETTAG oldtargettag,PREDICATELOCKTARGETTAG newtargettag,bool removeOld)2733 TransferPredicateLocksToNewTarget(PREDICATELOCKTARGETTAG oldtargettag,
2734 PREDICATELOCKTARGETTAG newtargettag,
2735 bool removeOld)
2736 {
2737 uint32 oldtargettaghash;
2738 LWLock *oldpartitionLock;
2739 PREDICATELOCKTARGET *oldtarget;
2740 uint32 newtargettaghash;
2741 LWLock *newpartitionLock;
2742 bool found;
2743 bool outOfShmem = false;
2744
2745 Assert(LWLockHeldByMeInMode(SerializablePredicateListLock,
2746 LW_EXCLUSIVE));
2747
2748 oldtargettaghash = PredicateLockTargetTagHashCode(&oldtargettag);
2749 newtargettaghash = PredicateLockTargetTagHashCode(&newtargettag);
2750 oldpartitionLock = PredicateLockHashPartitionLock(oldtargettaghash);
2751 newpartitionLock = PredicateLockHashPartitionLock(newtargettaghash);
2752
2753 if (removeOld)
2754 {
2755 /*
2756 * Remove the dummy entry to give us scratch space, so we know we'll
2757 * be able to create the new lock target.
2758 */
2759 RemoveScratchTarget(false);
2760 }
2761
2762 /*
2763 * We must get the partition locks in ascending sequence to avoid
2764 * deadlocks. If old and new partitions are the same, we must request the
2765 * lock only once.
2766 */
2767 if (oldpartitionLock < newpartitionLock)
2768 {
2769 LWLockAcquire(oldpartitionLock,
2770 (removeOld ? LW_EXCLUSIVE : LW_SHARED));
2771 LWLockAcquire(newpartitionLock, LW_EXCLUSIVE);
2772 }
2773 else if (oldpartitionLock > newpartitionLock)
2774 {
2775 LWLockAcquire(newpartitionLock, LW_EXCLUSIVE);
2776 LWLockAcquire(oldpartitionLock,
2777 (removeOld ? LW_EXCLUSIVE : LW_SHARED));
2778 }
2779 else
2780 LWLockAcquire(newpartitionLock, LW_EXCLUSIVE);
2781
2782 /*
2783 * Look for the old target. If not found, that's OK; no predicate locks
2784 * are affected, so we can just clean up and return. If it does exist,
2785 * walk its list of predicate locks and move or copy them to the new
2786 * target.
2787 */
2788 oldtarget = hash_search_with_hash_value(PredicateLockTargetHash,
2789 &oldtargettag,
2790 oldtargettaghash,
2791 HASH_FIND, NULL);
2792
2793 if (oldtarget)
2794 {
2795 PREDICATELOCKTARGET *newtarget;
2796 PREDICATELOCK *oldpredlock;
2797 PREDICATELOCKTAG newpredlocktag;
2798
2799 newtarget = hash_search_with_hash_value(PredicateLockTargetHash,
2800 &newtargettag,
2801 newtargettaghash,
2802 HASH_ENTER_NULL, &found);
2803
2804 if (!newtarget)
2805 {
2806 /* Failed to allocate due to insufficient shmem */
2807 outOfShmem = true;
2808 goto exit;
2809 }
2810
2811 /* If we created a new entry, initialize it */
2812 if (!found)
2813 SHMQueueInit(&(newtarget->predicateLocks));
2814
2815 newpredlocktag.myTarget = newtarget;
2816
2817 /*
2818 * Loop through all the locks on the old target, replacing them with
2819 * locks on the new target.
2820 */
2821 oldpredlock = (PREDICATELOCK *)
2822 SHMQueueNext(&(oldtarget->predicateLocks),
2823 &(oldtarget->predicateLocks),
2824 offsetof(PREDICATELOCK, targetLink));
2825 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
2826 while (oldpredlock)
2827 {
2828 SHM_QUEUE *predlocktargetlink;
2829 PREDICATELOCK *nextpredlock;
2830 PREDICATELOCK *newpredlock;
2831 SerCommitSeqNo oldCommitSeqNo = oldpredlock->commitSeqNo;
2832
2833 predlocktargetlink = &(oldpredlock->targetLink);
2834 nextpredlock = (PREDICATELOCK *)
2835 SHMQueueNext(&(oldtarget->predicateLocks),
2836 predlocktargetlink,
2837 offsetof(PREDICATELOCK, targetLink));
2838 newpredlocktag.myXact = oldpredlock->tag.myXact;
2839
2840 if (removeOld)
2841 {
2842 SHMQueueDelete(&(oldpredlock->xactLink));
2843 SHMQueueDelete(&(oldpredlock->targetLink));
2844
2845 hash_search_with_hash_value
2846 (PredicateLockHash,
2847 &oldpredlock->tag,
2848 PredicateLockHashCodeFromTargetHashCode(&oldpredlock->tag,
2849 oldtargettaghash),
2850 HASH_REMOVE, &found);
2851 Assert(found);
2852 }
2853
2854 newpredlock = (PREDICATELOCK *)
2855 hash_search_with_hash_value(PredicateLockHash,
2856 &newpredlocktag,
2857 PredicateLockHashCodeFromTargetHashCode(&newpredlocktag,
2858 newtargettaghash),
2859 HASH_ENTER_NULL,
2860 &found);
2861 if (!newpredlock)
2862 {
2863 /* Out of shared memory. Undo what we've done so far. */
2864 LWLockRelease(SerializableXactHashLock);
2865 DeleteLockTarget(newtarget, newtargettaghash);
2866 outOfShmem = true;
2867 goto exit;
2868 }
2869 if (!found)
2870 {
2871 SHMQueueInsertBefore(&(newtarget->predicateLocks),
2872 &(newpredlock->targetLink));
2873 SHMQueueInsertBefore(&(newpredlocktag.myXact->predicateLocks),
2874 &(newpredlock->xactLink));
2875 newpredlock->commitSeqNo = oldCommitSeqNo;
2876 }
2877 else
2878 {
2879 if (newpredlock->commitSeqNo < oldCommitSeqNo)
2880 newpredlock->commitSeqNo = oldCommitSeqNo;
2881 }
2882
2883 Assert(newpredlock->commitSeqNo != 0);
2884 Assert((newpredlock->commitSeqNo == InvalidSerCommitSeqNo)
2885 || (newpredlock->tag.myXact == OldCommittedSxact));
2886
2887 oldpredlock = nextpredlock;
2888 }
2889 LWLockRelease(SerializableXactHashLock);
2890
2891 if (removeOld)
2892 {
2893 Assert(SHMQueueEmpty(&oldtarget->predicateLocks));
2894 RemoveTargetIfNoLongerUsed(oldtarget, oldtargettaghash);
2895 }
2896 }
2897
2898
2899 exit:
2900 /* Release partition locks in reverse order of acquisition. */
2901 if (oldpartitionLock < newpartitionLock)
2902 {
2903 LWLockRelease(newpartitionLock);
2904 LWLockRelease(oldpartitionLock);
2905 }
2906 else if (oldpartitionLock > newpartitionLock)
2907 {
2908 LWLockRelease(oldpartitionLock);
2909 LWLockRelease(newpartitionLock);
2910 }
2911 else
2912 LWLockRelease(newpartitionLock);
2913
2914 if (removeOld)
2915 {
2916 /* We shouldn't run out of memory if we're moving locks */
2917 Assert(!outOfShmem);
2918
2919 /* Put the scratch entry back */
2920 RestoreScratchTarget(false);
2921 }
2922
2923 return !outOfShmem;
2924 }
2925
2926 /*
2927 * Drop all predicate locks of any granularity from the specified relation,
2928 * which can be a heap relation or an index relation. If 'transfer' is true,
2929 * acquire a relation lock on the heap for any transactions with any lock(s)
2930 * on the specified relation.
2931 *
2932 * This requires grabbing a lot of LW locks and scanning the entire lock
2933 * target table for matches. That makes this more expensive than most
2934 * predicate lock management functions, but it will only be called for DDL
2935 * type commands that are expensive anyway, and there are fast returns when
2936 * no serializable transactions are active or the relation is temporary.
2937 *
2938 * We don't use the TransferPredicateLocksToNewTarget function because it
2939 * acquires its own locks on the partitions of the two targets involved,
2940 * and we'll already be holding all partition locks.
2941 *
2942 * We can't throw an error from here, because the call could be from a
2943 * transaction which is not serializable.
2944 *
2945 * NOTE: This is currently only called with transfer set to true, but that may
2946 * change. If we decide to clean up the locks from a table on commit of a
2947 * transaction which executed DROP TABLE, the false condition will be useful.
2948 */
2949 static void
DropAllPredicateLocksFromTable(Relation relation,bool transfer)2950 DropAllPredicateLocksFromTable(Relation relation, bool transfer)
2951 {
2952 HASH_SEQ_STATUS seqstat;
2953 PREDICATELOCKTARGET *oldtarget;
2954 PREDICATELOCKTARGET *heaptarget;
2955 Oid dbId;
2956 Oid relId;
2957 Oid heapId;
2958 int i;
2959 bool isIndex;
2960 bool found;
2961 uint32 heaptargettaghash;
2962
2963 /*
2964 * Bail out quickly if there are no serializable transactions running.
2965 * It's safe to check this without taking locks because the caller is
2966 * holding an ACCESS EXCLUSIVE lock on the relation. No new locks which
2967 * would matter here can be acquired while that is held.
2968 */
2969 if (!TransactionIdIsValid(PredXact->SxactGlobalXmin))
2970 return;
2971
2972 if (!PredicateLockingNeededForRelation(relation))
2973 return;
2974
2975 dbId = relation->rd_node.dbNode;
2976 relId = relation->rd_id;
2977 if (relation->rd_index == NULL)
2978 {
2979 isIndex = false;
2980 heapId = relId;
2981 }
2982 else
2983 {
2984 isIndex = true;
2985 heapId = relation->rd_index->indrelid;
2986 }
2987 Assert(heapId != InvalidOid);
2988 Assert(transfer || !isIndex); /* index OID only makes sense with
2989 * transfer */
2990
2991 /* Retrieve first time needed, then keep. */
2992 heaptargettaghash = 0;
2993 heaptarget = NULL;
2994
2995 /* Acquire locks on all lock partitions */
2996 LWLockAcquire(SerializablePredicateListLock, LW_EXCLUSIVE);
2997 for (i = 0; i < NUM_PREDICATELOCK_PARTITIONS; i++)
2998 LWLockAcquire(PredicateLockHashPartitionLockByIndex(i), LW_EXCLUSIVE);
2999 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
3000
3001 /*
3002 * Remove the dummy entry to give us scratch space, so we know we'll be
3003 * able to create the new lock target.
3004 */
3005 if (transfer)
3006 RemoveScratchTarget(true);
3007
3008 /* Scan through target map */
3009 hash_seq_init(&seqstat, PredicateLockTargetHash);
3010
3011 while ((oldtarget = (PREDICATELOCKTARGET *) hash_seq_search(&seqstat)))
3012 {
3013 PREDICATELOCK *oldpredlock;
3014
3015 /*
3016 * Check whether this is a target which needs attention.
3017 */
3018 if (GET_PREDICATELOCKTARGETTAG_RELATION(oldtarget->tag) != relId)
3019 continue; /* wrong relation id */
3020 if (GET_PREDICATELOCKTARGETTAG_DB(oldtarget->tag) != dbId)
3021 continue; /* wrong database id */
3022 if (transfer && !isIndex
3023 && GET_PREDICATELOCKTARGETTAG_TYPE(oldtarget->tag) == PREDLOCKTAG_RELATION)
3024 continue; /* already the right lock */
3025
3026 /*
3027 * If we made it here, we have work to do. We make sure the heap
3028 * relation lock exists, then we walk the list of predicate locks for
3029 * the old target we found, moving all locks to the heap relation lock
3030 * -- unless they already hold that.
3031 */
3032
3033 /*
3034 * First make sure we have the heap relation target. We only need to
3035 * do this once.
3036 */
3037 if (transfer && heaptarget == NULL)
3038 {
3039 PREDICATELOCKTARGETTAG heaptargettag;
3040
3041 SET_PREDICATELOCKTARGETTAG_RELATION(heaptargettag, dbId, heapId);
3042 heaptargettaghash = PredicateLockTargetTagHashCode(&heaptargettag);
3043 heaptarget = hash_search_with_hash_value(PredicateLockTargetHash,
3044 &heaptargettag,
3045 heaptargettaghash,
3046 HASH_ENTER, &found);
3047 if (!found)
3048 SHMQueueInit(&heaptarget->predicateLocks);
3049 }
3050
3051 /*
3052 * Loop through all the locks on the old target, replacing them with
3053 * locks on the new target.
3054 */
3055 oldpredlock = (PREDICATELOCK *)
3056 SHMQueueNext(&(oldtarget->predicateLocks),
3057 &(oldtarget->predicateLocks),
3058 offsetof(PREDICATELOCK, targetLink));
3059 while (oldpredlock)
3060 {
3061 PREDICATELOCK *nextpredlock;
3062 PREDICATELOCK *newpredlock;
3063 SerCommitSeqNo oldCommitSeqNo;
3064 SERIALIZABLEXACT *oldXact;
3065
3066 nextpredlock = (PREDICATELOCK *)
3067 SHMQueueNext(&(oldtarget->predicateLocks),
3068 &(oldpredlock->targetLink),
3069 offsetof(PREDICATELOCK, targetLink));
3070
3071 /*
3072 * Remove the old lock first. This avoids the chance of running
3073 * out of lock structure entries for the hash table.
3074 */
3075 oldCommitSeqNo = oldpredlock->commitSeqNo;
3076 oldXact = oldpredlock->tag.myXact;
3077
3078 SHMQueueDelete(&(oldpredlock->xactLink));
3079
3080 /*
3081 * No need for retail delete from oldtarget list, we're removing
3082 * the whole target anyway.
3083 */
3084 hash_search(PredicateLockHash,
3085 &oldpredlock->tag,
3086 HASH_REMOVE, &found);
3087 Assert(found);
3088
3089 if (transfer)
3090 {
3091 PREDICATELOCKTAG newpredlocktag;
3092
3093 newpredlocktag.myTarget = heaptarget;
3094 newpredlocktag.myXact = oldXact;
3095 newpredlock = (PREDICATELOCK *)
3096 hash_search_with_hash_value(PredicateLockHash,
3097 &newpredlocktag,
3098 PredicateLockHashCodeFromTargetHashCode(&newpredlocktag,
3099 heaptargettaghash),
3100 HASH_ENTER,
3101 &found);
3102 if (!found)
3103 {
3104 SHMQueueInsertBefore(&(heaptarget->predicateLocks),
3105 &(newpredlock->targetLink));
3106 SHMQueueInsertBefore(&(newpredlocktag.myXact->predicateLocks),
3107 &(newpredlock->xactLink));
3108 newpredlock->commitSeqNo = oldCommitSeqNo;
3109 }
3110 else
3111 {
3112 if (newpredlock->commitSeqNo < oldCommitSeqNo)
3113 newpredlock->commitSeqNo = oldCommitSeqNo;
3114 }
3115
3116 Assert(newpredlock->commitSeqNo != 0);
3117 Assert((newpredlock->commitSeqNo == InvalidSerCommitSeqNo)
3118 || (newpredlock->tag.myXact == OldCommittedSxact));
3119 }
3120
3121 oldpredlock = nextpredlock;
3122 }
3123
3124 hash_search(PredicateLockTargetHash, &oldtarget->tag, HASH_REMOVE,
3125 &found);
3126 Assert(found);
3127 }
3128
3129 /* Put the scratch entry back */
3130 if (transfer)
3131 RestoreScratchTarget(true);
3132
3133 /* Release locks in reverse order */
3134 LWLockRelease(SerializableXactHashLock);
3135 for (i = NUM_PREDICATELOCK_PARTITIONS - 1; i >= 0; i--)
3136 LWLockRelease(PredicateLockHashPartitionLockByIndex(i));
3137 LWLockRelease(SerializablePredicateListLock);
3138 }
3139
3140 /*
3141 * TransferPredicateLocksToHeapRelation
3142 * For all transactions, transfer all predicate locks for the given
3143 * relation to a single relation lock on the heap.
3144 */
3145 void
TransferPredicateLocksToHeapRelation(Relation relation)3146 TransferPredicateLocksToHeapRelation(Relation relation)
3147 {
3148 DropAllPredicateLocksFromTable(relation, true);
3149 }
3150
3151
3152 /*
3153 * PredicateLockPageSplit
3154 *
3155 * Copies any predicate locks for the old page to the new page.
3156 * Skip if this is a temporary table or toast table.
3157 *
3158 * NOTE: A page split (or overflow) affects all serializable transactions,
3159 * even if it occurs in the context of another transaction isolation level.
3160 *
3161 * NOTE: This currently leaves the local copy of the locks without
3162 * information on the new lock which is in shared memory. This could cause
3163 * problems if enough page splits occur on locked pages without the processes
3164 * which hold the locks getting in and noticing.
3165 */
3166 void
PredicateLockPageSplit(Relation relation,BlockNumber oldblkno,BlockNumber newblkno)3167 PredicateLockPageSplit(Relation relation, BlockNumber oldblkno,
3168 BlockNumber newblkno)
3169 {
3170 PREDICATELOCKTARGETTAG oldtargettag;
3171 PREDICATELOCKTARGETTAG newtargettag;
3172 bool success;
3173
3174 /*
3175 * Bail out quickly if there are no serializable transactions running.
3176 *
3177 * It's safe to do this check without taking any additional locks. Even if
3178 * a serializable transaction starts concurrently, we know it can't take
3179 * any SIREAD locks on the page being split because the caller is holding
3180 * the associated buffer page lock. Memory reordering isn't an issue; the
3181 * memory barrier in the LWLock acquisition guarantees that this read
3182 * occurs while the buffer page lock is held.
3183 */
3184 if (!TransactionIdIsValid(PredXact->SxactGlobalXmin))
3185 return;
3186
3187 if (!PredicateLockingNeededForRelation(relation))
3188 return;
3189
3190 Assert(oldblkno != newblkno);
3191 Assert(BlockNumberIsValid(oldblkno));
3192 Assert(BlockNumberIsValid(newblkno));
3193
3194 SET_PREDICATELOCKTARGETTAG_PAGE(oldtargettag,
3195 relation->rd_node.dbNode,
3196 relation->rd_id,
3197 oldblkno);
3198 SET_PREDICATELOCKTARGETTAG_PAGE(newtargettag,
3199 relation->rd_node.dbNode,
3200 relation->rd_id,
3201 newblkno);
3202
3203 LWLockAcquire(SerializablePredicateListLock, LW_EXCLUSIVE);
3204
3205 /*
3206 * Try copying the locks over to the new page's tag, creating it if
3207 * necessary.
3208 */
3209 success = TransferPredicateLocksToNewTarget(oldtargettag,
3210 newtargettag,
3211 false);
3212
3213 if (!success)
3214 {
3215 /*
3216 * No more predicate lock entries are available. Failure isn't an
3217 * option here, so promote the page lock to a relation lock.
3218 */
3219
3220 /* Get the parent relation lock's lock tag */
3221 success = GetParentPredicateLockTag(&oldtargettag,
3222 &newtargettag);
3223 Assert(success);
3224
3225 /*
3226 * Move the locks to the parent. This shouldn't fail.
3227 *
3228 * Note that here we are removing locks held by other backends,
3229 * leading to a possible inconsistency in their local lock hash table.
3230 * This is OK because we're replacing it with a lock that covers the
3231 * old one.
3232 */
3233 success = TransferPredicateLocksToNewTarget(oldtargettag,
3234 newtargettag,
3235 true);
3236 Assert(success);
3237 }
3238
3239 LWLockRelease(SerializablePredicateListLock);
3240 }
3241
3242 /*
3243 * PredicateLockPageCombine
3244 *
3245 * Combines predicate locks for two existing pages.
3246 * Skip if this is a temporary table or toast table.
3247 *
3248 * NOTE: A page combine affects all serializable transactions, even if it
3249 * occurs in the context of another transaction isolation level.
3250 */
3251 void
PredicateLockPageCombine(Relation relation,BlockNumber oldblkno,BlockNumber newblkno)3252 PredicateLockPageCombine(Relation relation, BlockNumber oldblkno,
3253 BlockNumber newblkno)
3254 {
3255 /*
3256 * Page combines differ from page splits in that we ought to be able to
3257 * remove the locks on the old page after transferring them to the new
3258 * page, instead of duplicating them. However, because we can't edit other
3259 * backends' local lock tables, removing the old lock would leave them
3260 * with an entry in their LocalPredicateLockHash for a lock they're not
3261 * holding, which isn't acceptable. So we wind up having to do the same
3262 * work as a page split, acquiring a lock on the new page and keeping the
3263 * old page locked too. That can lead to some false positives, but should
3264 * be rare in practice.
3265 */
3266 PredicateLockPageSplit(relation, oldblkno, newblkno);
3267 }
3268
3269 /*
3270 * Walk the list of in-progress serializable transactions and find the new
3271 * xmin.
3272 */
3273 static void
SetNewSxactGlobalXmin(void)3274 SetNewSxactGlobalXmin(void)
3275 {
3276 SERIALIZABLEXACT *sxact;
3277
3278 Assert(LWLockHeldByMe(SerializableXactHashLock));
3279
3280 PredXact->SxactGlobalXmin = InvalidTransactionId;
3281 PredXact->SxactGlobalXminCount = 0;
3282
3283 for (sxact = FirstPredXact(); sxact != NULL; sxact = NextPredXact(sxact))
3284 {
3285 if (!SxactIsRolledBack(sxact)
3286 && !SxactIsCommitted(sxact)
3287 && sxact != OldCommittedSxact)
3288 {
3289 Assert(sxact->xmin != InvalidTransactionId);
3290 if (!TransactionIdIsValid(PredXact->SxactGlobalXmin)
3291 || TransactionIdPrecedes(sxact->xmin,
3292 PredXact->SxactGlobalXmin))
3293 {
3294 PredXact->SxactGlobalXmin = sxact->xmin;
3295 PredXact->SxactGlobalXminCount = 1;
3296 }
3297 else if (TransactionIdEquals(sxact->xmin,
3298 PredXact->SxactGlobalXmin))
3299 PredXact->SxactGlobalXminCount++;
3300 }
3301 }
3302
3303 SerialSetActiveSerXmin(PredXact->SxactGlobalXmin);
3304 }
3305
3306 /*
3307 * ReleasePredicateLocks
3308 *
3309 * Releases predicate locks based on completion of the current transaction,
3310 * whether committed or rolled back. It can also be called for a read only
3311 * transaction when it becomes impossible for the transaction to become
3312 * part of a dangerous structure.
3313 *
3314 * We do nothing unless this is a serializable transaction.
3315 *
3316 * This method must ensure that shared memory hash tables are cleaned
3317 * up in some relatively timely fashion.
3318 *
3319 * If this transaction is committing and is holding any predicate locks,
3320 * it must be added to a list of completed serializable transactions still
3321 * holding locks.
3322 *
3323 * If isReadOnlySafe is true, then predicate locks are being released before
3324 * the end of the transaction because MySerializableXact has been determined
3325 * to be RO_SAFE. In non-parallel mode we can release it completely, but it
3326 * in parallel mode we partially release the SERIALIZABLEXACT and keep it
3327 * around until the end of the transaction, allowing each backend to clear its
3328 * MySerializableXact variable and benefit from the optimization in its own
3329 * time.
3330 */
3331 void
ReleasePredicateLocks(bool isCommit,bool isReadOnlySafe)3332 ReleasePredicateLocks(bool isCommit, bool isReadOnlySafe)
3333 {
3334 bool needToClear;
3335 RWConflict conflict,
3336 nextConflict,
3337 possibleUnsafeConflict;
3338 SERIALIZABLEXACT *roXact;
3339
3340 /*
3341 * We can't trust XactReadOnly here, because a transaction which started
3342 * as READ WRITE can show as READ ONLY later, e.g., within
3343 * subtransactions. We want to flag a transaction as READ ONLY if it
3344 * commits without writing so that de facto READ ONLY transactions get the
3345 * benefit of some RO optimizations, so we will use this local variable to
3346 * get some cleanup logic right which is based on whether the transaction
3347 * was declared READ ONLY at the top level.
3348 */
3349 bool topLevelIsDeclaredReadOnly;
3350
3351 /* We can't be both committing and releasing early due to RO_SAFE. */
3352 Assert(!(isCommit && isReadOnlySafe));
3353
3354 /* Are we at the end of a transaction, that is, a commit or abort? */
3355 if (!isReadOnlySafe)
3356 {
3357 /*
3358 * Parallel workers mustn't release predicate locks at the end of
3359 * their transaction. The leader will do that at the end of its
3360 * transaction.
3361 */
3362 if (IsParallelWorker())
3363 {
3364 ReleasePredicateLocksLocal();
3365 return;
3366 }
3367
3368 /*
3369 * By the time the leader in a parallel query reaches end of
3370 * transaction, it has waited for all workers to exit.
3371 */
3372 Assert(!ParallelContextActive());
3373
3374 /*
3375 * If the leader in a parallel query earlier stashed a partially
3376 * released SERIALIZABLEXACT for final clean-up at end of transaction
3377 * (because workers might still have been accessing it), then it's
3378 * time to restore it.
3379 */
3380 if (SavedSerializableXact != InvalidSerializableXact)
3381 {
3382 Assert(MySerializableXact == InvalidSerializableXact);
3383 MySerializableXact = SavedSerializableXact;
3384 SavedSerializableXact = InvalidSerializableXact;
3385 Assert(SxactIsPartiallyReleased(MySerializableXact));
3386 }
3387 }
3388
3389 if (MySerializableXact == InvalidSerializableXact)
3390 {
3391 Assert(LocalPredicateLockHash == NULL);
3392 return;
3393 }
3394
3395 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
3396
3397 /*
3398 * If the transaction is committing, but it has been partially released
3399 * already, then treat this as a roll back. It was marked as rolled back.
3400 */
3401 if (isCommit && SxactIsPartiallyReleased(MySerializableXact))
3402 isCommit = false;
3403
3404 /*
3405 * If we're called in the middle of a transaction because we discovered
3406 * that the SXACT_FLAG_RO_SAFE flag was set, then we'll partially release
3407 * it (that is, release the predicate locks and conflicts, but not the
3408 * SERIALIZABLEXACT itself) if we're the first backend to have noticed.
3409 */
3410 if (isReadOnlySafe && IsInParallelMode())
3411 {
3412 /*
3413 * The leader needs to stash a pointer to it, so that it can
3414 * completely release it at end-of-transaction.
3415 */
3416 if (!IsParallelWorker())
3417 SavedSerializableXact = MySerializableXact;
3418
3419 /*
3420 * The first backend to reach this condition will partially release
3421 * the SERIALIZABLEXACT. All others will just clear their
3422 * backend-local state so that they stop doing SSI checks for the rest
3423 * of the transaction.
3424 */
3425 if (SxactIsPartiallyReleased(MySerializableXact))
3426 {
3427 LWLockRelease(SerializableXactHashLock);
3428 ReleasePredicateLocksLocal();
3429 return;
3430 }
3431 else
3432 {
3433 MySerializableXact->flags |= SXACT_FLAG_PARTIALLY_RELEASED;
3434 /* ... and proceed to perform the partial release below. */
3435 }
3436 }
3437 Assert(!isCommit || SxactIsPrepared(MySerializableXact));
3438 Assert(!isCommit || !SxactIsDoomed(MySerializableXact));
3439 Assert(!SxactIsCommitted(MySerializableXact));
3440 Assert(SxactIsPartiallyReleased(MySerializableXact)
3441 || !SxactIsRolledBack(MySerializableXact));
3442
3443 /* may not be serializable during COMMIT/ROLLBACK PREPARED */
3444 Assert(MySerializableXact->pid == 0 || IsolationIsSerializable());
3445
3446 /* We'd better not already be on the cleanup list. */
3447 Assert(!SxactIsOnFinishedList(MySerializableXact));
3448
3449 topLevelIsDeclaredReadOnly = SxactIsReadOnly(MySerializableXact);
3450
3451 /*
3452 * We don't hold XidGenLock lock here, assuming that TransactionId is
3453 * atomic!
3454 *
3455 * If this value is changing, we don't care that much whether we get the
3456 * old or new value -- it is just used to determine how far
3457 * SxactGlobalXmin must advance before this transaction can be fully
3458 * cleaned up. The worst that could happen is we wait for one more
3459 * transaction to complete before freeing some RAM; correctness of visible
3460 * behavior is not affected.
3461 */
3462 MySerializableXact->finishedBefore = XidFromFullTransactionId(ShmemVariableCache->nextXid);
3463
3464 /*
3465 * If it's not a commit it's either a rollback or a read-only transaction
3466 * flagged SXACT_FLAG_RO_SAFE, and we can clear our locks immediately.
3467 */
3468 if (isCommit)
3469 {
3470 MySerializableXact->flags |= SXACT_FLAG_COMMITTED;
3471 MySerializableXact->commitSeqNo = ++(PredXact->LastSxactCommitSeqNo);
3472 /* Recognize implicit read-only transaction (commit without write). */
3473 if (!MyXactDidWrite)
3474 MySerializableXact->flags |= SXACT_FLAG_READ_ONLY;
3475 }
3476 else
3477 {
3478 /*
3479 * The DOOMED flag indicates that we intend to roll back this
3480 * transaction and so it should not cause serialization failures for
3481 * other transactions that conflict with it. Note that this flag might
3482 * already be set, if another backend marked this transaction for
3483 * abort.
3484 *
3485 * The ROLLED_BACK flag further indicates that ReleasePredicateLocks
3486 * has been called, and so the SerializableXact is eligible for
3487 * cleanup. This means it should not be considered when calculating
3488 * SxactGlobalXmin.
3489 */
3490 MySerializableXact->flags |= SXACT_FLAG_DOOMED;
3491 MySerializableXact->flags |= SXACT_FLAG_ROLLED_BACK;
3492
3493 /*
3494 * If the transaction was previously prepared, but is now failing due
3495 * to a ROLLBACK PREPARED or (hopefully very rare) error after the
3496 * prepare, clear the prepared flag. This simplifies conflict
3497 * checking.
3498 */
3499 MySerializableXact->flags &= ~SXACT_FLAG_PREPARED;
3500 }
3501
3502 if (!topLevelIsDeclaredReadOnly)
3503 {
3504 Assert(PredXact->WritableSxactCount > 0);
3505 if (--(PredXact->WritableSxactCount) == 0)
3506 {
3507 /*
3508 * Release predicate locks and rw-conflicts in for all committed
3509 * transactions. There are no longer any transactions which might
3510 * conflict with the locks and no chance for new transactions to
3511 * overlap. Similarly, existing conflicts in can't cause pivots,
3512 * and any conflicts in which could have completed a dangerous
3513 * structure would already have caused a rollback, so any
3514 * remaining ones must be benign.
3515 */
3516 PredXact->CanPartialClearThrough = PredXact->LastSxactCommitSeqNo;
3517 }
3518 }
3519 else
3520 {
3521 /*
3522 * Read-only transactions: clear the list of transactions that might
3523 * make us unsafe. Note that we use 'inLink' for the iteration as
3524 * opposed to 'outLink' for the r/w xacts.
3525 */
3526 possibleUnsafeConflict = (RWConflict)
3527 SHMQueueNext(&MySerializableXact->possibleUnsafeConflicts,
3528 &MySerializableXact->possibleUnsafeConflicts,
3529 offsetof(RWConflictData, inLink));
3530 while (possibleUnsafeConflict)
3531 {
3532 nextConflict = (RWConflict)
3533 SHMQueueNext(&MySerializableXact->possibleUnsafeConflicts,
3534 &possibleUnsafeConflict->inLink,
3535 offsetof(RWConflictData, inLink));
3536
3537 Assert(!SxactIsReadOnly(possibleUnsafeConflict->sxactOut));
3538 Assert(MySerializableXact == possibleUnsafeConflict->sxactIn);
3539
3540 ReleaseRWConflict(possibleUnsafeConflict);
3541
3542 possibleUnsafeConflict = nextConflict;
3543 }
3544 }
3545
3546 /* Check for conflict out to old committed transactions. */
3547 if (isCommit
3548 && !SxactIsReadOnly(MySerializableXact)
3549 && SxactHasSummaryConflictOut(MySerializableXact))
3550 {
3551 /*
3552 * we don't know which old committed transaction we conflicted with,
3553 * so be conservative and use FirstNormalSerCommitSeqNo here
3554 */
3555 MySerializableXact->SeqNo.earliestOutConflictCommit =
3556 FirstNormalSerCommitSeqNo;
3557 MySerializableXact->flags |= SXACT_FLAG_CONFLICT_OUT;
3558 }
3559
3560 /*
3561 * Release all outConflicts to committed transactions. If we're rolling
3562 * back clear them all. Set SXACT_FLAG_CONFLICT_OUT if any point to
3563 * previously committed transactions.
3564 */
3565 conflict = (RWConflict)
3566 SHMQueueNext(&MySerializableXact->outConflicts,
3567 &MySerializableXact->outConflicts,
3568 offsetof(RWConflictData, outLink));
3569 while (conflict)
3570 {
3571 nextConflict = (RWConflict)
3572 SHMQueueNext(&MySerializableXact->outConflicts,
3573 &conflict->outLink,
3574 offsetof(RWConflictData, outLink));
3575
3576 if (isCommit
3577 && !SxactIsReadOnly(MySerializableXact)
3578 && SxactIsCommitted(conflict->sxactIn))
3579 {
3580 if ((MySerializableXact->flags & SXACT_FLAG_CONFLICT_OUT) == 0
3581 || conflict->sxactIn->prepareSeqNo < MySerializableXact->SeqNo.earliestOutConflictCommit)
3582 MySerializableXact->SeqNo.earliestOutConflictCommit = conflict->sxactIn->prepareSeqNo;
3583 MySerializableXact->flags |= SXACT_FLAG_CONFLICT_OUT;
3584 }
3585
3586 if (!isCommit
3587 || SxactIsCommitted(conflict->sxactIn)
3588 || (conflict->sxactIn->SeqNo.lastCommitBeforeSnapshot >= PredXact->LastSxactCommitSeqNo))
3589 ReleaseRWConflict(conflict);
3590
3591 conflict = nextConflict;
3592 }
3593
3594 /*
3595 * Release all inConflicts from committed and read-only transactions. If
3596 * we're rolling back, clear them all.
3597 */
3598 conflict = (RWConflict)
3599 SHMQueueNext(&MySerializableXact->inConflicts,
3600 &MySerializableXact->inConflicts,
3601 offsetof(RWConflictData, inLink));
3602 while (conflict)
3603 {
3604 nextConflict = (RWConflict)
3605 SHMQueueNext(&MySerializableXact->inConflicts,
3606 &conflict->inLink,
3607 offsetof(RWConflictData, inLink));
3608
3609 if (!isCommit
3610 || SxactIsCommitted(conflict->sxactOut)
3611 || SxactIsReadOnly(conflict->sxactOut))
3612 ReleaseRWConflict(conflict);
3613
3614 conflict = nextConflict;
3615 }
3616
3617 if (!topLevelIsDeclaredReadOnly)
3618 {
3619 /*
3620 * Remove ourselves from the list of possible conflicts for concurrent
3621 * READ ONLY transactions, flagging them as unsafe if we have a
3622 * conflict out. If any are waiting DEFERRABLE transactions, wake them
3623 * up if they are known safe or known unsafe.
3624 */
3625 possibleUnsafeConflict = (RWConflict)
3626 SHMQueueNext(&MySerializableXact->possibleUnsafeConflicts,
3627 &MySerializableXact->possibleUnsafeConflicts,
3628 offsetof(RWConflictData, outLink));
3629 while (possibleUnsafeConflict)
3630 {
3631 nextConflict = (RWConflict)
3632 SHMQueueNext(&MySerializableXact->possibleUnsafeConflicts,
3633 &possibleUnsafeConflict->outLink,
3634 offsetof(RWConflictData, outLink));
3635
3636 roXact = possibleUnsafeConflict->sxactIn;
3637 Assert(MySerializableXact == possibleUnsafeConflict->sxactOut);
3638 Assert(SxactIsReadOnly(roXact));
3639
3640 /* Mark conflicted if necessary. */
3641 if (isCommit
3642 && MyXactDidWrite
3643 && SxactHasConflictOut(MySerializableXact)
3644 && (MySerializableXact->SeqNo.earliestOutConflictCommit
3645 <= roXact->SeqNo.lastCommitBeforeSnapshot))
3646 {
3647 /*
3648 * This releases possibleUnsafeConflict (as well as all other
3649 * possible conflicts for roXact)
3650 */
3651 FlagSxactUnsafe(roXact);
3652 }
3653 else
3654 {
3655 ReleaseRWConflict(possibleUnsafeConflict);
3656
3657 /*
3658 * If we were the last possible conflict, flag it safe. The
3659 * transaction can now safely release its predicate locks (but
3660 * that transaction's backend has to do that itself).
3661 */
3662 if (SHMQueueEmpty(&roXact->possibleUnsafeConflicts))
3663 roXact->flags |= SXACT_FLAG_RO_SAFE;
3664 }
3665
3666 /*
3667 * Wake up the process for a waiting DEFERRABLE transaction if we
3668 * now know it's either safe or conflicted.
3669 */
3670 if (SxactIsDeferrableWaiting(roXact) &&
3671 (SxactIsROUnsafe(roXact) || SxactIsROSafe(roXact)))
3672 ProcSendSignal(roXact->pid);
3673
3674 possibleUnsafeConflict = nextConflict;
3675 }
3676 }
3677
3678 /*
3679 * Check whether it's time to clean up old transactions. This can only be
3680 * done when the last serializable transaction with the oldest xmin among
3681 * serializable transactions completes. We then find the "new oldest"
3682 * xmin and purge any transactions which finished before this transaction
3683 * was launched.
3684 */
3685 needToClear = false;
3686 if (TransactionIdEquals(MySerializableXact->xmin, PredXact->SxactGlobalXmin))
3687 {
3688 Assert(PredXact->SxactGlobalXminCount > 0);
3689 if (--(PredXact->SxactGlobalXminCount) == 0)
3690 {
3691 SetNewSxactGlobalXmin();
3692 needToClear = true;
3693 }
3694 }
3695
3696 LWLockRelease(SerializableXactHashLock);
3697
3698 LWLockAcquire(SerializableFinishedListLock, LW_EXCLUSIVE);
3699
3700 /* Add this to the list of transactions to check for later cleanup. */
3701 if (isCommit)
3702 SHMQueueInsertBefore(FinishedSerializableTransactions,
3703 &MySerializableXact->finishedLink);
3704
3705 /*
3706 * If we're releasing a RO_SAFE transaction in parallel mode, we'll only
3707 * partially release it. That's necessary because other backends may have
3708 * a reference to it. The leader will release the SERIALIZABLEXACT itself
3709 * at the end of the transaction after workers have stopped running.
3710 */
3711 if (!isCommit)
3712 ReleaseOneSerializableXact(MySerializableXact,
3713 isReadOnlySafe && IsInParallelMode(),
3714 false);
3715
3716 LWLockRelease(SerializableFinishedListLock);
3717
3718 if (needToClear)
3719 ClearOldPredicateLocks();
3720
3721 ReleasePredicateLocksLocal();
3722 }
3723
3724 static void
ReleasePredicateLocksLocal(void)3725 ReleasePredicateLocksLocal(void)
3726 {
3727 MySerializableXact = InvalidSerializableXact;
3728 MyXactDidWrite = false;
3729
3730 /* Delete per-transaction lock table */
3731 if (LocalPredicateLockHash != NULL)
3732 {
3733 hash_destroy(LocalPredicateLockHash);
3734 LocalPredicateLockHash = NULL;
3735 }
3736 }
3737
3738 /*
3739 * Clear old predicate locks, belonging to committed transactions that are no
3740 * longer interesting to any in-progress transaction.
3741 */
3742 static void
ClearOldPredicateLocks(void)3743 ClearOldPredicateLocks(void)
3744 {
3745 SERIALIZABLEXACT *finishedSxact;
3746 PREDICATELOCK *predlock;
3747
3748 /*
3749 * Loop through finished transactions. They are in commit order, so we can
3750 * stop as soon as we find one that's still interesting.
3751 */
3752 LWLockAcquire(SerializableFinishedListLock, LW_EXCLUSIVE);
3753 finishedSxact = (SERIALIZABLEXACT *)
3754 SHMQueueNext(FinishedSerializableTransactions,
3755 FinishedSerializableTransactions,
3756 offsetof(SERIALIZABLEXACT, finishedLink));
3757 LWLockAcquire(SerializableXactHashLock, LW_SHARED);
3758 while (finishedSxact)
3759 {
3760 SERIALIZABLEXACT *nextSxact;
3761
3762 nextSxact = (SERIALIZABLEXACT *)
3763 SHMQueueNext(FinishedSerializableTransactions,
3764 &(finishedSxact->finishedLink),
3765 offsetof(SERIALIZABLEXACT, finishedLink));
3766 if (!TransactionIdIsValid(PredXact->SxactGlobalXmin)
3767 || TransactionIdPrecedesOrEquals(finishedSxact->finishedBefore,
3768 PredXact->SxactGlobalXmin))
3769 {
3770 /*
3771 * This transaction committed before any in-progress transaction
3772 * took its snapshot. It's no longer interesting.
3773 */
3774 LWLockRelease(SerializableXactHashLock);
3775 SHMQueueDelete(&(finishedSxact->finishedLink));
3776 ReleaseOneSerializableXact(finishedSxact, false, false);
3777 LWLockAcquire(SerializableXactHashLock, LW_SHARED);
3778 }
3779 else if (finishedSxact->commitSeqNo > PredXact->HavePartialClearedThrough
3780 && finishedSxact->commitSeqNo <= PredXact->CanPartialClearThrough)
3781 {
3782 /*
3783 * Any active transactions that took their snapshot before this
3784 * transaction committed are read-only, so we can clear part of
3785 * its state.
3786 */
3787 LWLockRelease(SerializableXactHashLock);
3788
3789 if (SxactIsReadOnly(finishedSxact))
3790 {
3791 /* A read-only transaction can be removed entirely */
3792 SHMQueueDelete(&(finishedSxact->finishedLink));
3793 ReleaseOneSerializableXact(finishedSxact, false, false);
3794 }
3795 else
3796 {
3797 /*
3798 * A read-write transaction can only be partially cleared. We
3799 * need to keep the SERIALIZABLEXACT but can release the
3800 * SIREAD locks and conflicts in.
3801 */
3802 ReleaseOneSerializableXact(finishedSxact, true, false);
3803 }
3804
3805 PredXact->HavePartialClearedThrough = finishedSxact->commitSeqNo;
3806 LWLockAcquire(SerializableXactHashLock, LW_SHARED);
3807 }
3808 else
3809 {
3810 /* Still interesting. */
3811 break;
3812 }
3813 finishedSxact = nextSxact;
3814 }
3815 LWLockRelease(SerializableXactHashLock);
3816
3817 /*
3818 * Loop through predicate locks on dummy transaction for summarized data.
3819 */
3820 LWLockAcquire(SerializablePredicateListLock, LW_SHARED);
3821 predlock = (PREDICATELOCK *)
3822 SHMQueueNext(&OldCommittedSxact->predicateLocks,
3823 &OldCommittedSxact->predicateLocks,
3824 offsetof(PREDICATELOCK, xactLink));
3825 while (predlock)
3826 {
3827 PREDICATELOCK *nextpredlock;
3828 bool canDoPartialCleanup;
3829
3830 nextpredlock = (PREDICATELOCK *)
3831 SHMQueueNext(&OldCommittedSxact->predicateLocks,
3832 &predlock->xactLink,
3833 offsetof(PREDICATELOCK, xactLink));
3834
3835 LWLockAcquire(SerializableXactHashLock, LW_SHARED);
3836 Assert(predlock->commitSeqNo != 0);
3837 Assert(predlock->commitSeqNo != InvalidSerCommitSeqNo);
3838 canDoPartialCleanup = (predlock->commitSeqNo <= PredXact->CanPartialClearThrough);
3839 LWLockRelease(SerializableXactHashLock);
3840
3841 /*
3842 * If this lock originally belonged to an old enough transaction, we
3843 * can release it.
3844 */
3845 if (canDoPartialCleanup)
3846 {
3847 PREDICATELOCKTAG tag;
3848 PREDICATELOCKTARGET *target;
3849 PREDICATELOCKTARGETTAG targettag;
3850 uint32 targettaghash;
3851 LWLock *partitionLock;
3852
3853 tag = predlock->tag;
3854 target = tag.myTarget;
3855 targettag = target->tag;
3856 targettaghash = PredicateLockTargetTagHashCode(&targettag);
3857 partitionLock = PredicateLockHashPartitionLock(targettaghash);
3858
3859 LWLockAcquire(partitionLock, LW_EXCLUSIVE);
3860
3861 SHMQueueDelete(&(predlock->targetLink));
3862 SHMQueueDelete(&(predlock->xactLink));
3863
3864 hash_search_with_hash_value(PredicateLockHash, &tag,
3865 PredicateLockHashCodeFromTargetHashCode(&tag,
3866 targettaghash),
3867 HASH_REMOVE, NULL);
3868 RemoveTargetIfNoLongerUsed(target, targettaghash);
3869
3870 LWLockRelease(partitionLock);
3871 }
3872
3873 predlock = nextpredlock;
3874 }
3875
3876 LWLockRelease(SerializablePredicateListLock);
3877 LWLockRelease(SerializableFinishedListLock);
3878 }
3879
3880 /*
3881 * This is the normal way to delete anything from any of the predicate
3882 * locking hash tables. Given a transaction which we know can be deleted:
3883 * delete all predicate locks held by that transaction and any predicate
3884 * lock targets which are now unreferenced by a lock; delete all conflicts
3885 * for the transaction; delete all xid values for the transaction; then
3886 * delete the transaction.
3887 *
3888 * When the partial flag is set, we can release all predicate locks and
3889 * in-conflict information -- we've established that there are no longer
3890 * any overlapping read write transactions for which this transaction could
3891 * matter -- but keep the transaction entry itself and any outConflicts.
3892 *
3893 * When the summarize flag is set, we've run short of room for sxact data
3894 * and must summarize to the SLRU. Predicate locks are transferred to a
3895 * dummy "old" transaction, with duplicate locks on a single target
3896 * collapsing to a single lock with the "latest" commitSeqNo from among
3897 * the conflicting locks..
3898 */
3899 static void
ReleaseOneSerializableXact(SERIALIZABLEXACT * sxact,bool partial,bool summarize)3900 ReleaseOneSerializableXact(SERIALIZABLEXACT *sxact, bool partial,
3901 bool summarize)
3902 {
3903 PREDICATELOCK *predlock;
3904 SERIALIZABLEXIDTAG sxidtag;
3905 RWConflict conflict,
3906 nextConflict;
3907
3908 Assert(sxact != NULL);
3909 Assert(SxactIsRolledBack(sxact) || SxactIsCommitted(sxact));
3910 Assert(partial || !SxactIsOnFinishedList(sxact));
3911 Assert(LWLockHeldByMe(SerializableFinishedListLock));
3912
3913 /*
3914 * First release all the predicate locks held by this xact (or transfer
3915 * them to OldCommittedSxact if summarize is true)
3916 */
3917 LWLockAcquire(SerializablePredicateListLock, LW_SHARED);
3918 if (IsInParallelMode())
3919 LWLockAcquire(&sxact->perXactPredicateListLock, LW_EXCLUSIVE);
3920 predlock = (PREDICATELOCK *)
3921 SHMQueueNext(&(sxact->predicateLocks),
3922 &(sxact->predicateLocks),
3923 offsetof(PREDICATELOCK, xactLink));
3924 while (predlock)
3925 {
3926 PREDICATELOCK *nextpredlock;
3927 PREDICATELOCKTAG tag;
3928 SHM_QUEUE *targetLink;
3929 PREDICATELOCKTARGET *target;
3930 PREDICATELOCKTARGETTAG targettag;
3931 uint32 targettaghash;
3932 LWLock *partitionLock;
3933
3934 nextpredlock = (PREDICATELOCK *)
3935 SHMQueueNext(&(sxact->predicateLocks),
3936 &(predlock->xactLink),
3937 offsetof(PREDICATELOCK, xactLink));
3938
3939 tag = predlock->tag;
3940 targetLink = &(predlock->targetLink);
3941 target = tag.myTarget;
3942 targettag = target->tag;
3943 targettaghash = PredicateLockTargetTagHashCode(&targettag);
3944 partitionLock = PredicateLockHashPartitionLock(targettaghash);
3945
3946 LWLockAcquire(partitionLock, LW_EXCLUSIVE);
3947
3948 SHMQueueDelete(targetLink);
3949
3950 hash_search_with_hash_value(PredicateLockHash, &tag,
3951 PredicateLockHashCodeFromTargetHashCode(&tag,
3952 targettaghash),
3953 HASH_REMOVE, NULL);
3954 if (summarize)
3955 {
3956 bool found;
3957
3958 /* Fold into dummy transaction list. */
3959 tag.myXact = OldCommittedSxact;
3960 predlock = hash_search_with_hash_value(PredicateLockHash, &tag,
3961 PredicateLockHashCodeFromTargetHashCode(&tag,
3962 targettaghash),
3963 HASH_ENTER_NULL, &found);
3964 if (!predlock)
3965 ereport(ERROR,
3966 (errcode(ERRCODE_OUT_OF_MEMORY),
3967 errmsg("out of shared memory"),
3968 errhint("You might need to increase max_pred_locks_per_transaction.")));
3969 if (found)
3970 {
3971 Assert(predlock->commitSeqNo != 0);
3972 Assert(predlock->commitSeqNo != InvalidSerCommitSeqNo);
3973 if (predlock->commitSeqNo < sxact->commitSeqNo)
3974 predlock->commitSeqNo = sxact->commitSeqNo;
3975 }
3976 else
3977 {
3978 SHMQueueInsertBefore(&(target->predicateLocks),
3979 &(predlock->targetLink));
3980 SHMQueueInsertBefore(&(OldCommittedSxact->predicateLocks),
3981 &(predlock->xactLink));
3982 predlock->commitSeqNo = sxact->commitSeqNo;
3983 }
3984 }
3985 else
3986 RemoveTargetIfNoLongerUsed(target, targettaghash);
3987
3988 LWLockRelease(partitionLock);
3989
3990 predlock = nextpredlock;
3991 }
3992
3993 /*
3994 * Rather than retail removal, just re-init the head after we've run
3995 * through the list.
3996 */
3997 SHMQueueInit(&sxact->predicateLocks);
3998
3999 if (IsInParallelMode())
4000 LWLockRelease(&sxact->perXactPredicateListLock);
4001 LWLockRelease(SerializablePredicateListLock);
4002
4003 sxidtag.xid = sxact->topXid;
4004 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
4005
4006 /* Release all outConflicts (unless 'partial' is true) */
4007 if (!partial)
4008 {
4009 conflict = (RWConflict)
4010 SHMQueueNext(&sxact->outConflicts,
4011 &sxact->outConflicts,
4012 offsetof(RWConflictData, outLink));
4013 while (conflict)
4014 {
4015 nextConflict = (RWConflict)
4016 SHMQueueNext(&sxact->outConflicts,
4017 &conflict->outLink,
4018 offsetof(RWConflictData, outLink));
4019 if (summarize)
4020 conflict->sxactIn->flags |= SXACT_FLAG_SUMMARY_CONFLICT_IN;
4021 ReleaseRWConflict(conflict);
4022 conflict = nextConflict;
4023 }
4024 }
4025
4026 /* Release all inConflicts. */
4027 conflict = (RWConflict)
4028 SHMQueueNext(&sxact->inConflicts,
4029 &sxact->inConflicts,
4030 offsetof(RWConflictData, inLink));
4031 while (conflict)
4032 {
4033 nextConflict = (RWConflict)
4034 SHMQueueNext(&sxact->inConflicts,
4035 &conflict->inLink,
4036 offsetof(RWConflictData, inLink));
4037 if (summarize)
4038 conflict->sxactOut->flags |= SXACT_FLAG_SUMMARY_CONFLICT_OUT;
4039 ReleaseRWConflict(conflict);
4040 conflict = nextConflict;
4041 }
4042
4043 /* Finally, get rid of the xid and the record of the transaction itself. */
4044 if (!partial)
4045 {
4046 if (sxidtag.xid != InvalidTransactionId)
4047 hash_search(SerializableXidHash, &sxidtag, HASH_REMOVE, NULL);
4048 ReleasePredXact(sxact);
4049 }
4050
4051 LWLockRelease(SerializableXactHashLock);
4052 }
4053
4054 /*
4055 * Tests whether the given top level transaction is concurrent with
4056 * (overlaps) our current transaction.
4057 *
4058 * We need to identify the top level transaction for SSI, anyway, so pass
4059 * that to this function to save the overhead of checking the snapshot's
4060 * subxip array.
4061 */
4062 static bool
XidIsConcurrent(TransactionId xid)4063 XidIsConcurrent(TransactionId xid)
4064 {
4065 Snapshot snap;
4066 uint32 i;
4067
4068 Assert(TransactionIdIsValid(xid));
4069 Assert(!TransactionIdEquals(xid, GetTopTransactionIdIfAny()));
4070
4071 snap = GetTransactionSnapshot();
4072
4073 if (TransactionIdPrecedes(xid, snap->xmin))
4074 return false;
4075
4076 if (TransactionIdFollowsOrEquals(xid, snap->xmax))
4077 return true;
4078
4079 for (i = 0; i < snap->xcnt; i++)
4080 {
4081 if (xid == snap->xip[i])
4082 return true;
4083 }
4084
4085 return false;
4086 }
4087
4088 bool
CheckForSerializableConflictOutNeeded(Relation relation,Snapshot snapshot)4089 CheckForSerializableConflictOutNeeded(Relation relation, Snapshot snapshot)
4090 {
4091 if (!SerializationNeededForRead(relation, snapshot))
4092 return false;
4093
4094 /* Check if someone else has already decided that we need to die */
4095 if (SxactIsDoomed(MySerializableXact))
4096 {
4097 ereport(ERROR,
4098 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
4099 errmsg("could not serialize access due to read/write dependencies among transactions"),
4100 errdetail_internal("Reason code: Canceled on identification as a pivot, during conflict out checking."),
4101 errhint("The transaction might succeed if retried.")));
4102 }
4103
4104 return true;
4105 }
4106
4107 /*
4108 * CheckForSerializableConflictOut
4109 * A table AM is reading a tuple that has been modified. If it determines
4110 * that the tuple version it is reading is not visible to us, it should
4111 * pass in the top level xid of the transaction that created it.
4112 * Otherwise, if it determines that it is visible to us but it has been
4113 * deleted or there is a newer version available due to an update, it
4114 * should pass in the top level xid of the modifying transaction.
4115 *
4116 * This function will check for overlap with our own transaction. If the given
4117 * xid is also serializable and the transactions overlap (i.e., they cannot see
4118 * each other's writes), then we have a conflict out.
4119 */
4120 void
CheckForSerializableConflictOut(Relation relation,TransactionId xid,Snapshot snapshot)4121 CheckForSerializableConflictOut(Relation relation, TransactionId xid, Snapshot snapshot)
4122 {
4123 SERIALIZABLEXIDTAG sxidtag;
4124 SERIALIZABLEXID *sxid;
4125 SERIALIZABLEXACT *sxact;
4126
4127 if (!SerializationNeededForRead(relation, snapshot))
4128 return;
4129
4130 /* Check if someone else has already decided that we need to die */
4131 if (SxactIsDoomed(MySerializableXact))
4132 {
4133 ereport(ERROR,
4134 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
4135 errmsg("could not serialize access due to read/write dependencies among transactions"),
4136 errdetail_internal("Reason code: Canceled on identification as a pivot, during conflict out checking."),
4137 errhint("The transaction might succeed if retried.")));
4138 }
4139 Assert(TransactionIdIsValid(xid));
4140
4141 if (TransactionIdEquals(xid, GetTopTransactionIdIfAny()))
4142 return;
4143
4144 /*
4145 * Find sxact or summarized info for the top level xid.
4146 */
4147 sxidtag.xid = xid;
4148 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
4149 sxid = (SERIALIZABLEXID *)
4150 hash_search(SerializableXidHash, &sxidtag, HASH_FIND, NULL);
4151 if (!sxid)
4152 {
4153 /*
4154 * Transaction not found in "normal" SSI structures. Check whether it
4155 * got pushed out to SLRU storage for "old committed" transactions.
4156 */
4157 SerCommitSeqNo conflictCommitSeqNo;
4158
4159 conflictCommitSeqNo = SerialGetMinConflictCommitSeqNo(xid);
4160 if (conflictCommitSeqNo != 0)
4161 {
4162 if (conflictCommitSeqNo != InvalidSerCommitSeqNo
4163 && (!SxactIsReadOnly(MySerializableXact)
4164 || conflictCommitSeqNo
4165 <= MySerializableXact->SeqNo.lastCommitBeforeSnapshot))
4166 ereport(ERROR,
4167 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
4168 errmsg("could not serialize access due to read/write dependencies among transactions"),
4169 errdetail_internal("Reason code: Canceled on conflict out to old pivot %u.", xid),
4170 errhint("The transaction might succeed if retried.")));
4171
4172 if (SxactHasSummaryConflictIn(MySerializableXact)
4173 || !SHMQueueEmpty(&MySerializableXact->inConflicts))
4174 ereport(ERROR,
4175 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
4176 errmsg("could not serialize access due to read/write dependencies among transactions"),
4177 errdetail_internal("Reason code: Canceled on identification as a pivot, with conflict out to old committed transaction %u.", xid),
4178 errhint("The transaction might succeed if retried.")));
4179
4180 MySerializableXact->flags |= SXACT_FLAG_SUMMARY_CONFLICT_OUT;
4181 }
4182
4183 /* It's not serializable or otherwise not important. */
4184 LWLockRelease(SerializableXactHashLock);
4185 return;
4186 }
4187 sxact = sxid->myXact;
4188 Assert(TransactionIdEquals(sxact->topXid, xid));
4189 if (sxact == MySerializableXact || SxactIsDoomed(sxact))
4190 {
4191 /* Can't conflict with ourself or a transaction that will roll back. */
4192 LWLockRelease(SerializableXactHashLock);
4193 return;
4194 }
4195
4196 /*
4197 * We have a conflict out to a transaction which has a conflict out to a
4198 * summarized transaction. That summarized transaction must have
4199 * committed first, and we can't tell when it committed in relation to our
4200 * snapshot acquisition, so something needs to be canceled.
4201 */
4202 if (SxactHasSummaryConflictOut(sxact))
4203 {
4204 if (!SxactIsPrepared(sxact))
4205 {
4206 sxact->flags |= SXACT_FLAG_DOOMED;
4207 LWLockRelease(SerializableXactHashLock);
4208 return;
4209 }
4210 else
4211 {
4212 LWLockRelease(SerializableXactHashLock);
4213 ereport(ERROR,
4214 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
4215 errmsg("could not serialize access due to read/write dependencies among transactions"),
4216 errdetail_internal("Reason code: Canceled on conflict out to old pivot."),
4217 errhint("The transaction might succeed if retried.")));
4218 }
4219 }
4220
4221 /*
4222 * If this is a read-only transaction and the writing transaction has
4223 * committed, and it doesn't have a rw-conflict to a transaction which
4224 * committed before it, no conflict.
4225 */
4226 if (SxactIsReadOnly(MySerializableXact)
4227 && SxactIsCommitted(sxact)
4228 && !SxactHasSummaryConflictOut(sxact)
4229 && (!SxactHasConflictOut(sxact)
4230 || MySerializableXact->SeqNo.lastCommitBeforeSnapshot < sxact->SeqNo.earliestOutConflictCommit))
4231 {
4232 /* Read-only transaction will appear to run first. No conflict. */
4233 LWLockRelease(SerializableXactHashLock);
4234 return;
4235 }
4236
4237 if (!XidIsConcurrent(xid))
4238 {
4239 /* This write was already in our snapshot; no conflict. */
4240 LWLockRelease(SerializableXactHashLock);
4241 return;
4242 }
4243
4244 if (RWConflictExists(MySerializableXact, sxact))
4245 {
4246 /* We don't want duplicate conflict records in the list. */
4247 LWLockRelease(SerializableXactHashLock);
4248 return;
4249 }
4250
4251 /*
4252 * Flag the conflict. But first, if this conflict creates a dangerous
4253 * structure, ereport an error.
4254 */
4255 FlagRWConflict(MySerializableXact, sxact);
4256 LWLockRelease(SerializableXactHashLock);
4257 }
4258
4259 /*
4260 * Check a particular target for rw-dependency conflict in. A subroutine of
4261 * CheckForSerializableConflictIn().
4262 */
4263 static void
CheckTargetForConflictsIn(PREDICATELOCKTARGETTAG * targettag)4264 CheckTargetForConflictsIn(PREDICATELOCKTARGETTAG *targettag)
4265 {
4266 uint32 targettaghash;
4267 LWLock *partitionLock;
4268 PREDICATELOCKTARGET *target;
4269 PREDICATELOCK *predlock;
4270 PREDICATELOCK *mypredlock = NULL;
4271 PREDICATELOCKTAG mypredlocktag;
4272
4273 Assert(MySerializableXact != InvalidSerializableXact);
4274
4275 /*
4276 * The same hash and LW lock apply to the lock target and the lock itself.
4277 */
4278 targettaghash = PredicateLockTargetTagHashCode(targettag);
4279 partitionLock = PredicateLockHashPartitionLock(targettaghash);
4280 LWLockAcquire(partitionLock, LW_SHARED);
4281 target = (PREDICATELOCKTARGET *)
4282 hash_search_with_hash_value(PredicateLockTargetHash,
4283 targettag, targettaghash,
4284 HASH_FIND, NULL);
4285 if (!target)
4286 {
4287 /* Nothing has this target locked; we're done here. */
4288 LWLockRelease(partitionLock);
4289 return;
4290 }
4291
4292 /*
4293 * Each lock for an overlapping transaction represents a conflict: a
4294 * rw-dependency in to this transaction.
4295 */
4296 predlock = (PREDICATELOCK *)
4297 SHMQueueNext(&(target->predicateLocks),
4298 &(target->predicateLocks),
4299 offsetof(PREDICATELOCK, targetLink));
4300 LWLockAcquire(SerializableXactHashLock, LW_SHARED);
4301 while (predlock)
4302 {
4303 SHM_QUEUE *predlocktargetlink;
4304 PREDICATELOCK *nextpredlock;
4305 SERIALIZABLEXACT *sxact;
4306
4307 predlocktargetlink = &(predlock->targetLink);
4308 nextpredlock = (PREDICATELOCK *)
4309 SHMQueueNext(&(target->predicateLocks),
4310 predlocktargetlink,
4311 offsetof(PREDICATELOCK, targetLink));
4312
4313 sxact = predlock->tag.myXact;
4314 if (sxact == MySerializableXact)
4315 {
4316 /*
4317 * If we're getting a write lock on a tuple, we don't need a
4318 * predicate (SIREAD) lock on the same tuple. We can safely remove
4319 * our SIREAD lock, but we'll defer doing so until after the loop
4320 * because that requires upgrading to an exclusive partition lock.
4321 *
4322 * We can't use this optimization within a subtransaction because
4323 * the subtransaction could roll back, and we would be left
4324 * without any lock at the top level.
4325 */
4326 if (!IsSubTransaction()
4327 && GET_PREDICATELOCKTARGETTAG_OFFSET(*targettag))
4328 {
4329 mypredlock = predlock;
4330 mypredlocktag = predlock->tag;
4331 }
4332 }
4333 else if (!SxactIsDoomed(sxact)
4334 && (!SxactIsCommitted(sxact)
4335 || TransactionIdPrecedes(GetTransactionSnapshot()->xmin,
4336 sxact->finishedBefore))
4337 && !RWConflictExists(sxact, MySerializableXact))
4338 {
4339 LWLockRelease(SerializableXactHashLock);
4340 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
4341
4342 /*
4343 * Re-check after getting exclusive lock because the other
4344 * transaction may have flagged a conflict.
4345 */
4346 if (!SxactIsDoomed(sxact)
4347 && (!SxactIsCommitted(sxact)
4348 || TransactionIdPrecedes(GetTransactionSnapshot()->xmin,
4349 sxact->finishedBefore))
4350 && !RWConflictExists(sxact, MySerializableXact))
4351 {
4352 FlagRWConflict(sxact, MySerializableXact);
4353 }
4354
4355 LWLockRelease(SerializableXactHashLock);
4356 LWLockAcquire(SerializableXactHashLock, LW_SHARED);
4357 }
4358
4359 predlock = nextpredlock;
4360 }
4361 LWLockRelease(SerializableXactHashLock);
4362 LWLockRelease(partitionLock);
4363
4364 /*
4365 * If we found one of our own SIREAD locks to remove, remove it now.
4366 *
4367 * At this point our transaction already has a RowExclusiveLock on the
4368 * relation, so we are OK to drop the predicate lock on the tuple, if
4369 * found, without fearing that another write against the tuple will occur
4370 * before the MVCC information makes it to the buffer.
4371 */
4372 if (mypredlock != NULL)
4373 {
4374 uint32 predlockhashcode;
4375 PREDICATELOCK *rmpredlock;
4376
4377 LWLockAcquire(SerializablePredicateListLock, LW_SHARED);
4378 if (IsInParallelMode())
4379 LWLockAcquire(&MySerializableXact->perXactPredicateListLock, LW_EXCLUSIVE);
4380 LWLockAcquire(partitionLock, LW_EXCLUSIVE);
4381 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
4382
4383 /*
4384 * Remove the predicate lock from shared memory, if it wasn't removed
4385 * while the locks were released. One way that could happen is from
4386 * autovacuum cleaning up an index.
4387 */
4388 predlockhashcode = PredicateLockHashCodeFromTargetHashCode
4389 (&mypredlocktag, targettaghash);
4390 rmpredlock = (PREDICATELOCK *)
4391 hash_search_with_hash_value(PredicateLockHash,
4392 &mypredlocktag,
4393 predlockhashcode,
4394 HASH_FIND, NULL);
4395 if (rmpredlock != NULL)
4396 {
4397 Assert(rmpredlock == mypredlock);
4398
4399 SHMQueueDelete(&(mypredlock->targetLink));
4400 SHMQueueDelete(&(mypredlock->xactLink));
4401
4402 rmpredlock = (PREDICATELOCK *)
4403 hash_search_with_hash_value(PredicateLockHash,
4404 &mypredlocktag,
4405 predlockhashcode,
4406 HASH_REMOVE, NULL);
4407 Assert(rmpredlock == mypredlock);
4408
4409 RemoveTargetIfNoLongerUsed(target, targettaghash);
4410 }
4411
4412 LWLockRelease(SerializableXactHashLock);
4413 LWLockRelease(partitionLock);
4414 if (IsInParallelMode())
4415 LWLockRelease(&MySerializableXact->perXactPredicateListLock);
4416 LWLockRelease(SerializablePredicateListLock);
4417
4418 if (rmpredlock != NULL)
4419 {
4420 /*
4421 * Remove entry in local lock table if it exists. It's OK if it
4422 * doesn't exist; that means the lock was transferred to a new
4423 * target by a different backend.
4424 */
4425 hash_search_with_hash_value(LocalPredicateLockHash,
4426 targettag, targettaghash,
4427 HASH_REMOVE, NULL);
4428
4429 DecrementParentLocks(targettag);
4430 }
4431 }
4432 }
4433
4434 /*
4435 * CheckForSerializableConflictIn
4436 * We are writing the given tuple. If that indicates a rw-conflict
4437 * in from another serializable transaction, take appropriate action.
4438 *
4439 * Skip checking for any granularity for which a parameter is missing.
4440 *
4441 * A tuple update or delete is in conflict if we have a predicate lock
4442 * against the relation or page in which the tuple exists, or against the
4443 * tuple itself.
4444 */
4445 void
CheckForSerializableConflictIn(Relation relation,ItemPointer tid,BlockNumber blkno)4446 CheckForSerializableConflictIn(Relation relation, ItemPointer tid, BlockNumber blkno)
4447 {
4448 PREDICATELOCKTARGETTAG targettag;
4449
4450 if (!SerializationNeededForWrite(relation))
4451 return;
4452
4453 /* Check if someone else has already decided that we need to die */
4454 if (SxactIsDoomed(MySerializableXact))
4455 ereport(ERROR,
4456 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
4457 errmsg("could not serialize access due to read/write dependencies among transactions"),
4458 errdetail_internal("Reason code: Canceled on identification as a pivot, during conflict in checking."),
4459 errhint("The transaction might succeed if retried.")));
4460
4461 /*
4462 * We're doing a write which might cause rw-conflicts now or later.
4463 * Memorize that fact.
4464 */
4465 MyXactDidWrite = true;
4466
4467 /*
4468 * It is important that we check for locks from the finest granularity to
4469 * the coarsest granularity, so that granularity promotion doesn't cause
4470 * us to miss a lock. The new (coarser) lock will be acquired before the
4471 * old (finer) locks are released.
4472 *
4473 * It is not possible to take and hold a lock across the checks for all
4474 * granularities because each target could be in a separate partition.
4475 */
4476 if (tid != NULL)
4477 {
4478 SET_PREDICATELOCKTARGETTAG_TUPLE(targettag,
4479 relation->rd_node.dbNode,
4480 relation->rd_id,
4481 ItemPointerGetBlockNumber(tid),
4482 ItemPointerGetOffsetNumber(tid));
4483 CheckTargetForConflictsIn(&targettag);
4484 }
4485
4486 if (blkno != InvalidBlockNumber)
4487 {
4488 SET_PREDICATELOCKTARGETTAG_PAGE(targettag,
4489 relation->rd_node.dbNode,
4490 relation->rd_id,
4491 blkno);
4492 CheckTargetForConflictsIn(&targettag);
4493 }
4494
4495 SET_PREDICATELOCKTARGETTAG_RELATION(targettag,
4496 relation->rd_node.dbNode,
4497 relation->rd_id);
4498 CheckTargetForConflictsIn(&targettag);
4499 }
4500
4501 /*
4502 * CheckTableForSerializableConflictIn
4503 * The entire table is going through a DDL-style logical mass delete
4504 * like TRUNCATE or DROP TABLE. If that causes a rw-conflict in from
4505 * another serializable transaction, take appropriate action.
4506 *
4507 * While these operations do not operate entirely within the bounds of
4508 * snapshot isolation, they can occur inside a serializable transaction, and
4509 * will logically occur after any reads which saw rows which were destroyed
4510 * by these operations, so we do what we can to serialize properly under
4511 * SSI.
4512 *
4513 * The relation passed in must be a heap relation. Any predicate lock of any
4514 * granularity on the heap will cause a rw-conflict in to this transaction.
4515 * Predicate locks on indexes do not matter because they only exist to guard
4516 * against conflicting inserts into the index, and this is a mass *delete*.
4517 * When a table is truncated or dropped, the index will also be truncated
4518 * or dropped, and we'll deal with locks on the index when that happens.
4519 *
4520 * Dropping or truncating a table also needs to drop any existing predicate
4521 * locks on heap tuples or pages, because they're about to go away. This
4522 * should be done before altering the predicate locks because the transaction
4523 * could be rolled back because of a conflict, in which case the lock changes
4524 * are not needed. (At the moment, we don't actually bother to drop the
4525 * existing locks on a dropped or truncated table at the moment. That might
4526 * lead to some false positives, but it doesn't seem worth the trouble.)
4527 */
4528 void
CheckTableForSerializableConflictIn(Relation relation)4529 CheckTableForSerializableConflictIn(Relation relation)
4530 {
4531 HASH_SEQ_STATUS seqstat;
4532 PREDICATELOCKTARGET *target;
4533 Oid dbId;
4534 Oid heapId;
4535 int i;
4536
4537 /*
4538 * Bail out quickly if there are no serializable transactions running.
4539 * It's safe to check this without taking locks because the caller is
4540 * holding an ACCESS EXCLUSIVE lock on the relation. No new locks which
4541 * would matter here can be acquired while that is held.
4542 */
4543 if (!TransactionIdIsValid(PredXact->SxactGlobalXmin))
4544 return;
4545
4546 if (!SerializationNeededForWrite(relation))
4547 return;
4548
4549 /*
4550 * We're doing a write which might cause rw-conflicts now or later.
4551 * Memorize that fact.
4552 */
4553 MyXactDidWrite = true;
4554
4555 Assert(relation->rd_index == NULL); /* not an index relation */
4556
4557 dbId = relation->rd_node.dbNode;
4558 heapId = relation->rd_id;
4559
4560 LWLockAcquire(SerializablePredicateListLock, LW_EXCLUSIVE);
4561 for (i = 0; i < NUM_PREDICATELOCK_PARTITIONS; i++)
4562 LWLockAcquire(PredicateLockHashPartitionLockByIndex(i), LW_SHARED);
4563 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
4564
4565 /* Scan through target list */
4566 hash_seq_init(&seqstat, PredicateLockTargetHash);
4567
4568 while ((target = (PREDICATELOCKTARGET *) hash_seq_search(&seqstat)))
4569 {
4570 PREDICATELOCK *predlock;
4571
4572 /*
4573 * Check whether this is a target which needs attention.
4574 */
4575 if (GET_PREDICATELOCKTARGETTAG_RELATION(target->tag) != heapId)
4576 continue; /* wrong relation id */
4577 if (GET_PREDICATELOCKTARGETTAG_DB(target->tag) != dbId)
4578 continue; /* wrong database id */
4579
4580 /*
4581 * Loop through locks for this target and flag conflicts.
4582 */
4583 predlock = (PREDICATELOCK *)
4584 SHMQueueNext(&(target->predicateLocks),
4585 &(target->predicateLocks),
4586 offsetof(PREDICATELOCK, targetLink));
4587 while (predlock)
4588 {
4589 PREDICATELOCK *nextpredlock;
4590
4591 nextpredlock = (PREDICATELOCK *)
4592 SHMQueueNext(&(target->predicateLocks),
4593 &(predlock->targetLink),
4594 offsetof(PREDICATELOCK, targetLink));
4595
4596 if (predlock->tag.myXact != MySerializableXact
4597 && !RWConflictExists(predlock->tag.myXact, MySerializableXact))
4598 {
4599 FlagRWConflict(predlock->tag.myXact, MySerializableXact);
4600 }
4601
4602 predlock = nextpredlock;
4603 }
4604 }
4605
4606 /* Release locks in reverse order */
4607 LWLockRelease(SerializableXactHashLock);
4608 for (i = NUM_PREDICATELOCK_PARTITIONS - 1; i >= 0; i--)
4609 LWLockRelease(PredicateLockHashPartitionLockByIndex(i));
4610 LWLockRelease(SerializablePredicateListLock);
4611 }
4612
4613
4614 /*
4615 * Flag a rw-dependency between two serializable transactions.
4616 *
4617 * The caller is responsible for ensuring that we have a LW lock on
4618 * the transaction hash table.
4619 */
4620 static void
FlagRWConflict(SERIALIZABLEXACT * reader,SERIALIZABLEXACT * writer)4621 FlagRWConflict(SERIALIZABLEXACT *reader, SERIALIZABLEXACT *writer)
4622 {
4623 Assert(reader != writer);
4624
4625 /* First, see if this conflict causes failure. */
4626 OnConflict_CheckForSerializationFailure(reader, writer);
4627
4628 /* Actually do the conflict flagging. */
4629 if (reader == OldCommittedSxact)
4630 writer->flags |= SXACT_FLAG_SUMMARY_CONFLICT_IN;
4631 else if (writer == OldCommittedSxact)
4632 reader->flags |= SXACT_FLAG_SUMMARY_CONFLICT_OUT;
4633 else
4634 SetRWConflict(reader, writer);
4635 }
4636
4637 /*----------------------------------------------------------------------------
4638 * We are about to add a RW-edge to the dependency graph - check that we don't
4639 * introduce a dangerous structure by doing so, and abort one of the
4640 * transactions if so.
4641 *
4642 * A serialization failure can only occur if there is a dangerous structure
4643 * in the dependency graph:
4644 *
4645 * Tin ------> Tpivot ------> Tout
4646 * rw rw
4647 *
4648 * Furthermore, Tout must commit first.
4649 *
4650 * One more optimization is that if Tin is declared READ ONLY (or commits
4651 * without writing), we can only have a problem if Tout committed before Tin
4652 * acquired its snapshot.
4653 *----------------------------------------------------------------------------
4654 */
4655 static void
OnConflict_CheckForSerializationFailure(const SERIALIZABLEXACT * reader,SERIALIZABLEXACT * writer)4656 OnConflict_CheckForSerializationFailure(const SERIALIZABLEXACT *reader,
4657 SERIALIZABLEXACT *writer)
4658 {
4659 bool failure;
4660 RWConflict conflict;
4661
4662 Assert(LWLockHeldByMe(SerializableXactHashLock));
4663
4664 failure = false;
4665
4666 /*------------------------------------------------------------------------
4667 * Check for already-committed writer with rw-conflict out flagged
4668 * (conflict-flag on W means that T2 committed before W):
4669 *
4670 * R ------> W ------> T2
4671 * rw rw
4672 *
4673 * That is a dangerous structure, so we must abort. (Since the writer
4674 * has already committed, we must be the reader)
4675 *------------------------------------------------------------------------
4676 */
4677 if (SxactIsCommitted(writer)
4678 && (SxactHasConflictOut(writer) || SxactHasSummaryConflictOut(writer)))
4679 failure = true;
4680
4681 /*------------------------------------------------------------------------
4682 * Check whether the writer has become a pivot with an out-conflict
4683 * committed transaction (T2), and T2 committed first:
4684 *
4685 * R ------> W ------> T2
4686 * rw rw
4687 *
4688 * Because T2 must've committed first, there is no anomaly if:
4689 * - the reader committed before T2
4690 * - the writer committed before T2
4691 * - the reader is a READ ONLY transaction and the reader was concurrent
4692 * with T2 (= reader acquired its snapshot before T2 committed)
4693 *
4694 * We also handle the case that T2 is prepared but not yet committed
4695 * here. In that case T2 has already checked for conflicts, so if it
4696 * commits first, making the above conflict real, it's too late for it
4697 * to abort.
4698 *------------------------------------------------------------------------
4699 */
4700 if (!failure)
4701 {
4702 if (SxactHasSummaryConflictOut(writer))
4703 {
4704 failure = true;
4705 conflict = NULL;
4706 }
4707 else
4708 conflict = (RWConflict)
4709 SHMQueueNext(&writer->outConflicts,
4710 &writer->outConflicts,
4711 offsetof(RWConflictData, outLink));
4712 while (conflict)
4713 {
4714 SERIALIZABLEXACT *t2 = conflict->sxactIn;
4715
4716 if (SxactIsPrepared(t2)
4717 && (!SxactIsCommitted(reader)
4718 || t2->prepareSeqNo <= reader->commitSeqNo)
4719 && (!SxactIsCommitted(writer)
4720 || t2->prepareSeqNo <= writer->commitSeqNo)
4721 && (!SxactIsReadOnly(reader)
4722 || t2->prepareSeqNo <= reader->SeqNo.lastCommitBeforeSnapshot))
4723 {
4724 failure = true;
4725 break;
4726 }
4727 conflict = (RWConflict)
4728 SHMQueueNext(&writer->outConflicts,
4729 &conflict->outLink,
4730 offsetof(RWConflictData, outLink));
4731 }
4732 }
4733
4734 /*------------------------------------------------------------------------
4735 * Check whether the reader has become a pivot with a writer
4736 * that's committed (or prepared):
4737 *
4738 * T0 ------> R ------> W
4739 * rw rw
4740 *
4741 * Because W must've committed first for an anomaly to occur, there is no
4742 * anomaly if:
4743 * - T0 committed before the writer
4744 * - T0 is READ ONLY, and overlaps the writer
4745 *------------------------------------------------------------------------
4746 */
4747 if (!failure && SxactIsPrepared(writer) && !SxactIsReadOnly(reader))
4748 {
4749 if (SxactHasSummaryConflictIn(reader))
4750 {
4751 failure = true;
4752 conflict = NULL;
4753 }
4754 else
4755 conflict = (RWConflict)
4756 SHMQueueNext(&reader->inConflicts,
4757 &reader->inConflicts,
4758 offsetof(RWConflictData, inLink));
4759 while (conflict)
4760 {
4761 SERIALIZABLEXACT *t0 = conflict->sxactOut;
4762
4763 if (!SxactIsDoomed(t0)
4764 && (!SxactIsCommitted(t0)
4765 || t0->commitSeqNo >= writer->prepareSeqNo)
4766 && (!SxactIsReadOnly(t0)
4767 || t0->SeqNo.lastCommitBeforeSnapshot >= writer->prepareSeqNo))
4768 {
4769 failure = true;
4770 break;
4771 }
4772 conflict = (RWConflict)
4773 SHMQueueNext(&reader->inConflicts,
4774 &conflict->inLink,
4775 offsetof(RWConflictData, inLink));
4776 }
4777 }
4778
4779 if (failure)
4780 {
4781 /*
4782 * We have to kill a transaction to avoid a possible anomaly from
4783 * occurring. If the writer is us, we can just ereport() to cause a
4784 * transaction abort. Otherwise we flag the writer for termination,
4785 * causing it to abort when it tries to commit. However, if the writer
4786 * is a prepared transaction, already prepared, we can't abort it
4787 * anymore, so we have to kill the reader instead.
4788 */
4789 if (MySerializableXact == writer)
4790 {
4791 LWLockRelease(SerializableXactHashLock);
4792 ereport(ERROR,
4793 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
4794 errmsg("could not serialize access due to read/write dependencies among transactions"),
4795 errdetail_internal("Reason code: Canceled on identification as a pivot, during write."),
4796 errhint("The transaction might succeed if retried.")));
4797 }
4798 else if (SxactIsPrepared(writer))
4799 {
4800 LWLockRelease(SerializableXactHashLock);
4801
4802 /* if we're not the writer, we have to be the reader */
4803 Assert(MySerializableXact == reader);
4804 ereport(ERROR,
4805 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
4806 errmsg("could not serialize access due to read/write dependencies among transactions"),
4807 errdetail_internal("Reason code: Canceled on conflict out to pivot %u, during read.", writer->topXid),
4808 errhint("The transaction might succeed if retried.")));
4809 }
4810 writer->flags |= SXACT_FLAG_DOOMED;
4811 }
4812 }
4813
4814 /*
4815 * PreCommit_CheckForSerializationFailure
4816 * Check for dangerous structures in a serializable transaction
4817 * at commit.
4818 *
4819 * We're checking for a dangerous structure as each conflict is recorded.
4820 * The only way we could have a problem at commit is if this is the "out"
4821 * side of a pivot, and neither the "in" side nor the pivot has yet
4822 * committed.
4823 *
4824 * If a dangerous structure is found, the pivot (the near conflict) is
4825 * marked for death, because rolling back another transaction might mean
4826 * that we fail without ever making progress. This transaction is
4827 * committing writes, so letting it commit ensures progress. If we
4828 * canceled the far conflict, it might immediately fail again on retry.
4829 */
4830 void
PreCommit_CheckForSerializationFailure(void)4831 PreCommit_CheckForSerializationFailure(void)
4832 {
4833 RWConflict nearConflict;
4834
4835 if (MySerializableXact == InvalidSerializableXact)
4836 return;
4837
4838 Assert(IsolationIsSerializable());
4839
4840 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
4841
4842 /* Check if someone else has already decided that we need to die */
4843 if (SxactIsDoomed(MySerializableXact))
4844 {
4845 Assert(!SxactIsPartiallyReleased(MySerializableXact));
4846 LWLockRelease(SerializableXactHashLock);
4847 ereport(ERROR,
4848 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
4849 errmsg("could not serialize access due to read/write dependencies among transactions"),
4850 errdetail_internal("Reason code: Canceled on identification as a pivot, during commit attempt."),
4851 errhint("The transaction might succeed if retried.")));
4852 }
4853
4854 nearConflict = (RWConflict)
4855 SHMQueueNext(&MySerializableXact->inConflicts,
4856 &MySerializableXact->inConflicts,
4857 offsetof(RWConflictData, inLink));
4858 while (nearConflict)
4859 {
4860 if (!SxactIsCommitted(nearConflict->sxactOut)
4861 && !SxactIsDoomed(nearConflict->sxactOut))
4862 {
4863 RWConflict farConflict;
4864
4865 farConflict = (RWConflict)
4866 SHMQueueNext(&nearConflict->sxactOut->inConflicts,
4867 &nearConflict->sxactOut->inConflicts,
4868 offsetof(RWConflictData, inLink));
4869 while (farConflict)
4870 {
4871 if (farConflict->sxactOut == MySerializableXact
4872 || (!SxactIsCommitted(farConflict->sxactOut)
4873 && !SxactIsReadOnly(farConflict->sxactOut)
4874 && !SxactIsDoomed(farConflict->sxactOut)))
4875 {
4876 /*
4877 * Normally, we kill the pivot transaction to make sure we
4878 * make progress if the failing transaction is retried.
4879 * However, we can't kill it if it's already prepared, so
4880 * in that case we commit suicide instead.
4881 */
4882 if (SxactIsPrepared(nearConflict->sxactOut))
4883 {
4884 LWLockRelease(SerializableXactHashLock);
4885 ereport(ERROR,
4886 (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
4887 errmsg("could not serialize access due to read/write dependencies among transactions"),
4888 errdetail_internal("Reason code: Canceled on commit attempt with conflict in from prepared pivot."),
4889 errhint("The transaction might succeed if retried.")));
4890 }
4891 nearConflict->sxactOut->flags |= SXACT_FLAG_DOOMED;
4892 break;
4893 }
4894 farConflict = (RWConflict)
4895 SHMQueueNext(&nearConflict->sxactOut->inConflicts,
4896 &farConflict->inLink,
4897 offsetof(RWConflictData, inLink));
4898 }
4899 }
4900
4901 nearConflict = (RWConflict)
4902 SHMQueueNext(&MySerializableXact->inConflicts,
4903 &nearConflict->inLink,
4904 offsetof(RWConflictData, inLink));
4905 }
4906
4907 MySerializableXact->prepareSeqNo = ++(PredXact->LastSxactCommitSeqNo);
4908 MySerializableXact->flags |= SXACT_FLAG_PREPARED;
4909
4910 LWLockRelease(SerializableXactHashLock);
4911 }
4912
4913 /*------------------------------------------------------------------------*/
4914
4915 /*
4916 * Two-phase commit support
4917 */
4918
4919 /*
4920 * AtPrepare_Locks
4921 * Do the preparatory work for a PREPARE: make 2PC state file
4922 * records for all predicate locks currently held.
4923 */
4924 void
AtPrepare_PredicateLocks(void)4925 AtPrepare_PredicateLocks(void)
4926 {
4927 PREDICATELOCK *predlock;
4928 SERIALIZABLEXACT *sxact;
4929 TwoPhasePredicateRecord record;
4930 TwoPhasePredicateXactRecord *xactRecord;
4931 TwoPhasePredicateLockRecord *lockRecord;
4932
4933 sxact = MySerializableXact;
4934 xactRecord = &(record.data.xactRecord);
4935 lockRecord = &(record.data.lockRecord);
4936
4937 if (MySerializableXact == InvalidSerializableXact)
4938 return;
4939
4940 /* Generate an xact record for our SERIALIZABLEXACT */
4941 record.type = TWOPHASEPREDICATERECORD_XACT;
4942 xactRecord->xmin = MySerializableXact->xmin;
4943 xactRecord->flags = MySerializableXact->flags;
4944
4945 /*
4946 * Note that we don't include the list of conflicts in our out in the
4947 * statefile, because new conflicts can be added even after the
4948 * transaction prepares. We'll just make a conservative assumption during
4949 * recovery instead.
4950 */
4951
4952 RegisterTwoPhaseRecord(TWOPHASE_RM_PREDICATELOCK_ID, 0,
4953 &record, sizeof(record));
4954
4955 /*
4956 * Generate a lock record for each lock.
4957 *
4958 * To do this, we need to walk the predicate lock list in our sxact rather
4959 * than using the local predicate lock table because the latter is not
4960 * guaranteed to be accurate.
4961 */
4962 LWLockAcquire(SerializablePredicateListLock, LW_SHARED);
4963
4964 /*
4965 * No need to take sxact->perXactPredicateListLock in parallel mode
4966 * because there cannot be any parallel workers running while we are
4967 * preparing a transaction.
4968 */
4969 Assert(!IsParallelWorker() && !ParallelContextActive());
4970
4971 predlock = (PREDICATELOCK *)
4972 SHMQueueNext(&(sxact->predicateLocks),
4973 &(sxact->predicateLocks),
4974 offsetof(PREDICATELOCK, xactLink));
4975
4976 while (predlock != NULL)
4977 {
4978 record.type = TWOPHASEPREDICATERECORD_LOCK;
4979 lockRecord->target = predlock->tag.myTarget->tag;
4980
4981 RegisterTwoPhaseRecord(TWOPHASE_RM_PREDICATELOCK_ID, 0,
4982 &record, sizeof(record));
4983
4984 predlock = (PREDICATELOCK *)
4985 SHMQueueNext(&(sxact->predicateLocks),
4986 &(predlock->xactLink),
4987 offsetof(PREDICATELOCK, xactLink));
4988 }
4989
4990 LWLockRelease(SerializablePredicateListLock);
4991 }
4992
4993 /*
4994 * PostPrepare_Locks
4995 * Clean up after successful PREPARE. Unlike the non-predicate
4996 * lock manager, we do not need to transfer locks to a dummy
4997 * PGPROC because our SERIALIZABLEXACT will stay around
4998 * anyway. We only need to clean up our local state.
4999 */
5000 void
PostPrepare_PredicateLocks(TransactionId xid)5001 PostPrepare_PredicateLocks(TransactionId xid)
5002 {
5003 if (MySerializableXact == InvalidSerializableXact)
5004 return;
5005
5006 Assert(SxactIsPrepared(MySerializableXact));
5007
5008 MySerializableXact->pid = 0;
5009
5010 hash_destroy(LocalPredicateLockHash);
5011 LocalPredicateLockHash = NULL;
5012
5013 MySerializableXact = InvalidSerializableXact;
5014 MyXactDidWrite = false;
5015 }
5016
5017 /*
5018 * PredicateLockTwoPhaseFinish
5019 * Release a prepared transaction's predicate locks once it
5020 * commits or aborts.
5021 */
5022 void
PredicateLockTwoPhaseFinish(TransactionId xid,bool isCommit)5023 PredicateLockTwoPhaseFinish(TransactionId xid, bool isCommit)
5024 {
5025 SERIALIZABLEXID *sxid;
5026 SERIALIZABLEXIDTAG sxidtag;
5027
5028 sxidtag.xid = xid;
5029
5030 LWLockAcquire(SerializableXactHashLock, LW_SHARED);
5031 sxid = (SERIALIZABLEXID *)
5032 hash_search(SerializableXidHash, &sxidtag, HASH_FIND, NULL);
5033 LWLockRelease(SerializableXactHashLock);
5034
5035 /* xid will not be found if it wasn't a serializable transaction */
5036 if (sxid == NULL)
5037 return;
5038
5039 /* Release its locks */
5040 MySerializableXact = sxid->myXact;
5041 MyXactDidWrite = true; /* conservatively assume that we wrote
5042 * something */
5043 ReleasePredicateLocks(isCommit, false);
5044 }
5045
5046 /*
5047 * Re-acquire a predicate lock belonging to a transaction that was prepared.
5048 */
5049 void
predicatelock_twophase_recover(TransactionId xid,uint16 info,void * recdata,uint32 len)5050 predicatelock_twophase_recover(TransactionId xid, uint16 info,
5051 void *recdata, uint32 len)
5052 {
5053 TwoPhasePredicateRecord *record;
5054
5055 Assert(len == sizeof(TwoPhasePredicateRecord));
5056
5057 record = (TwoPhasePredicateRecord *) recdata;
5058
5059 Assert((record->type == TWOPHASEPREDICATERECORD_XACT) ||
5060 (record->type == TWOPHASEPREDICATERECORD_LOCK));
5061
5062 if (record->type == TWOPHASEPREDICATERECORD_XACT)
5063 {
5064 /* Per-transaction record. Set up a SERIALIZABLEXACT. */
5065 TwoPhasePredicateXactRecord *xactRecord;
5066 SERIALIZABLEXACT *sxact;
5067 SERIALIZABLEXID *sxid;
5068 SERIALIZABLEXIDTAG sxidtag;
5069 bool found;
5070
5071 xactRecord = (TwoPhasePredicateXactRecord *) &record->data.xactRecord;
5072
5073 LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
5074 sxact = CreatePredXact();
5075 if (!sxact)
5076 ereport(ERROR,
5077 (errcode(ERRCODE_OUT_OF_MEMORY),
5078 errmsg("out of shared memory")));
5079
5080 /* vxid for a prepared xact is InvalidBackendId/xid; no pid */
5081 sxact->vxid.backendId = InvalidBackendId;
5082 sxact->vxid.localTransactionId = (LocalTransactionId) xid;
5083 sxact->pid = 0;
5084
5085 /* a prepared xact hasn't committed yet */
5086 sxact->prepareSeqNo = RecoverySerCommitSeqNo;
5087 sxact->commitSeqNo = InvalidSerCommitSeqNo;
5088 sxact->finishedBefore = InvalidTransactionId;
5089
5090 sxact->SeqNo.lastCommitBeforeSnapshot = RecoverySerCommitSeqNo;
5091
5092 /*
5093 * Don't need to track this; no transactions running at the time the
5094 * recovered xact started are still active, except possibly other
5095 * prepared xacts and we don't care whether those are RO_SAFE or not.
5096 */
5097 SHMQueueInit(&(sxact->possibleUnsafeConflicts));
5098
5099 SHMQueueInit(&(sxact->predicateLocks));
5100 SHMQueueElemInit(&(sxact->finishedLink));
5101
5102 sxact->topXid = xid;
5103 sxact->xmin = xactRecord->xmin;
5104 sxact->flags = xactRecord->flags;
5105 Assert(SxactIsPrepared(sxact));
5106 if (!SxactIsReadOnly(sxact))
5107 {
5108 ++(PredXact->WritableSxactCount);
5109 Assert(PredXact->WritableSxactCount <=
5110 (MaxBackends + max_prepared_xacts));
5111 }
5112
5113 /*
5114 * We don't know whether the transaction had any conflicts or not, so
5115 * we'll conservatively assume that it had both a conflict in and a
5116 * conflict out, and represent that with the summary conflict flags.
5117 */
5118 SHMQueueInit(&(sxact->outConflicts));
5119 SHMQueueInit(&(sxact->inConflicts));
5120 sxact->flags |= SXACT_FLAG_SUMMARY_CONFLICT_IN;
5121 sxact->flags |= SXACT_FLAG_SUMMARY_CONFLICT_OUT;
5122
5123 /* Register the transaction's xid */
5124 sxidtag.xid = xid;
5125 sxid = (SERIALIZABLEXID *) hash_search(SerializableXidHash,
5126 &sxidtag,
5127 HASH_ENTER, &found);
5128 Assert(sxid != NULL);
5129 Assert(!found);
5130 sxid->myXact = (SERIALIZABLEXACT *) sxact;
5131
5132 /*
5133 * Update global xmin. Note that this is a special case compared to
5134 * registering a normal transaction, because the global xmin might go
5135 * backwards. That's OK, because until recovery is over we're not
5136 * going to complete any transactions or create any non-prepared
5137 * transactions, so there's no danger of throwing away.
5138 */
5139 if ((!TransactionIdIsValid(PredXact->SxactGlobalXmin)) ||
5140 (TransactionIdFollows(PredXact->SxactGlobalXmin, sxact->xmin)))
5141 {
5142 PredXact->SxactGlobalXmin = sxact->xmin;
5143 PredXact->SxactGlobalXminCount = 1;
5144 SerialSetActiveSerXmin(sxact->xmin);
5145 }
5146 else if (TransactionIdEquals(sxact->xmin, PredXact->SxactGlobalXmin))
5147 {
5148 Assert(PredXact->SxactGlobalXminCount > 0);
5149 PredXact->SxactGlobalXminCount++;
5150 }
5151
5152 LWLockRelease(SerializableXactHashLock);
5153 }
5154 else if (record->type == TWOPHASEPREDICATERECORD_LOCK)
5155 {
5156 /* Lock record. Recreate the PREDICATELOCK */
5157 TwoPhasePredicateLockRecord *lockRecord;
5158 SERIALIZABLEXID *sxid;
5159 SERIALIZABLEXACT *sxact;
5160 SERIALIZABLEXIDTAG sxidtag;
5161 uint32 targettaghash;
5162
5163 lockRecord = (TwoPhasePredicateLockRecord *) &record->data.lockRecord;
5164 targettaghash = PredicateLockTargetTagHashCode(&lockRecord->target);
5165
5166 LWLockAcquire(SerializableXactHashLock, LW_SHARED);
5167 sxidtag.xid = xid;
5168 sxid = (SERIALIZABLEXID *)
5169 hash_search(SerializableXidHash, &sxidtag, HASH_FIND, NULL);
5170 LWLockRelease(SerializableXactHashLock);
5171
5172 Assert(sxid != NULL);
5173 sxact = sxid->myXact;
5174 Assert(sxact != InvalidSerializableXact);
5175
5176 CreatePredicateLock(&lockRecord->target, targettaghash, sxact);
5177 }
5178 }
5179
5180 /*
5181 * Prepare to share the current SERIALIZABLEXACT with parallel workers.
5182 * Return a handle object that can be used by AttachSerializableXact() in a
5183 * parallel worker.
5184 */
5185 SerializableXactHandle
ShareSerializableXact(void)5186 ShareSerializableXact(void)
5187 {
5188 return MySerializableXact;
5189 }
5190
5191 /*
5192 * Allow parallel workers to import the leader's SERIALIZABLEXACT.
5193 */
5194 void
AttachSerializableXact(SerializableXactHandle handle)5195 AttachSerializableXact(SerializableXactHandle handle)
5196 {
5197
5198 Assert(MySerializableXact == InvalidSerializableXact);
5199
5200 MySerializableXact = (SERIALIZABLEXACT *) handle;
5201 if (MySerializableXact != InvalidSerializableXact)
5202 CreateLocalPredicateLockHash();
5203 }
5204