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