Movatterモバイル変換

[0]ホーム
URL:

Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

 Sign up
Appearance settings

Commitdb122d4

Browse files
committed
Cope with catcache entries becoming stale during detoasting.
We've long had a policy that any toasted fields in a catalog tupleshould be pulled in-line before entering the tuple in a catalog cache.However, that requires access to the catalog's toast table, and we'lltypically do AcceptInvalidationMessages while opening the toast table.So it's possible that the catalog tuple is outdated by the time wefinish detoasting it. Since no cache entry exists yet, we can'tmark the entry stale during AcceptInvalidationMessages, and insteadwe'll press forward and build an apparently-valid cache entry. Theupshot is that we have a race condition whereby an out-of-date entrycould be made in a backend's catalog cache, and persist thereindefinitely causing indeterminate misbehavior.To fix, use the existing systable_recheck_tuple code to recheckwhether the catalog tuple is still up-to-date after we finishdetoasting it. If not, loop around and restart the process ofsearching the catalog and constructing cache entries from the top.The case is rare enough that this shouldn't create any meaningfulperformance penalty, even in the SearchCatCacheList case wherewe need to tear down and reconstruct the whole list.Indeed, the case is so rare that AFAICT it doesn't occur duringour regression tests, and there doesn't seem to be any easy wayto build a test that would exercise it reliably. To allowtesting of the retry code paths, add logic (in USE_ASSERT_CHECKINGbuilds only) that randomly pretends that the recheck failed aboutone time out of a thousand. This is enough to ensure that we'llpass through the retry paths during most regression test runs.By adding an extra level of looping, this commit creates a needto reindent most of SearchCatCacheMiss and SearchCatCacheList.I'll do that separately, to allow putting those changes in.git-blame-ignore-revs.Patch by me; thanks to Alexander Lakhin for having built a testcase to prove the bug is real, and to Xiaoran Wang for review.Back-patch to all supported branches.Discussion:https://postgr.es/m/1393953.1698353013@sss.pgh.pa.usDiscussion:https://postgr.es/m/CAGjhLkOoBEC9mLsnB42d3CO1vcMx71MLSEuigeABbQ8oRdA6gw@mail.gmail.com
1 parent924daec commitdb122d4

File tree

1 file changed

+123
-40
lines changed

1 file changed

+123
-40
lines changed

‎src/backend/utils/cache/catcache.c

Lines changed: 123 additions & 40 deletions
Original file line numberDiff line numberDiff line change
@@ -89,10 +89,10 @@ static void CatCachePrintStats(int code, Datum arg);
8989
staticvoidCatCacheRemoveCTup(CatCache*cache,CatCTup*ct);
9090
staticvoidCatCacheRemoveCList(CatCache*cache,CatCList*cl);
9191
staticvoidCatalogCacheInitializeCache(CatCache*cache);
92-
staticCatCTup*CatalogCacheCreateEntry(CatCache*cache,HeapTuplentp,
92+
staticCatCTup*CatalogCacheCreateEntry(CatCache*cache,
93+
HeapTuplentp,SysScanDescscandesc,
9394
Datum*arguments,
94-
uint32hashValue,IndexhashIndex,
95-
boolnegative);
95+
uint32hashValue,IndexhashIndex);
9696

9797
staticvoidCatCacheFreeKeys(TupleDesctupdesc,intnkeys,int*attnos,
9898
Datum*keys);
@@ -1324,6 +1324,7 @@ SearchCatCacheMiss(CatCache *cache,
13241324
SysScanDescscandesc;
13251325
HeapTuplentp;
13261326
CatCTup*ct;
1327+
boolstale;
13271328
Datumarguments[CATCACHE_MAXKEYS];
13281329

13291330
/* Initialize local parameter array */
@@ -1332,16 +1333,6 @@ SearchCatCacheMiss(CatCache *cache,
13321333
arguments[2]=v3;
13331334
arguments[3]=v4;
13341335

1335-
/*
1336-
* Ok, need to make a lookup in the relation, copy the scankey and fill
1337-
* out any per-call fields.
1338-
*/
1339-
memcpy(cur_skey,cache->cc_skey,sizeof(ScanKeyData)*nkeys);
1340-
cur_skey[0].sk_argument=v1;
1341-
cur_skey[1].sk_argument=v2;
1342-
cur_skey[2].sk_argument=v3;
1343-
cur_skey[3].sk_argument=v4;
1344-
13451336
/*
13461337
* Tuple was not found in cache, so we have to try to retrieve it directly
13471338
* from the relation. If found, we will add it to the cache; if not
@@ -1356,9 +1347,28 @@ SearchCatCacheMiss(CatCache *cache,
13561347
* will eventually age out of the cache, so there's no functional problem.
13571348
* This case is rare enough that it's not worth expending extra cycles to
13581349
* detect.
1350+
*
1351+
* Another case, which we *must* handle, is that the tuple could become
1352+
* outdated during CatalogCacheCreateEntry's attempt to detoast it (since
1353+
* AcceptInvalidationMessages can run during TOAST table access). We do
1354+
* not want to return already-stale catcache entries, so we loop around
1355+
* and do the table scan again if that happens.
13591356
*/
13601357
relation=table_open(cache->cc_reloid,AccessShareLock);
13611358

1359+
do
1360+
{
1361+
/*
1362+
* Ok, need to make a lookup in the relation, copy the scankey and
1363+
* fill out any per-call fields. (We must re-do this when retrying,
1364+
* because systable_beginscan scribbles on the scankey.)
1365+
*/
1366+
memcpy(cur_skey,cache->cc_skey,sizeof(ScanKeyData)*nkeys);
1367+
cur_skey[0].sk_argument=v1;
1368+
cur_skey[1].sk_argument=v2;
1369+
cur_skey[2].sk_argument=v3;
1370+
cur_skey[3].sk_argument=v4;
1371+
13621372
scandesc=systable_beginscan(relation,
13631373
cache->cc_indexoid,
13641374
IndexScanOK(cache,cur_skey),
@@ -1367,12 +1377,18 @@ SearchCatCacheMiss(CatCache *cache,
13671377
cur_skey);
13681378

13691379
ct=NULL;
1380+
stale= false;
13701381

13711382
while (HeapTupleIsValid(ntp=systable_getnext(scandesc)))
13721383
{
1373-
ct=CatalogCacheCreateEntry(cache,ntp,arguments,
1374-
hashValue,hashIndex,
1375-
false);
1384+
ct=CatalogCacheCreateEntry(cache,ntp,scandesc,NULL,
1385+
hashValue,hashIndex);
1386+
/* upon failure, we must start the scan over */
1387+
if (ct==NULL)
1388+
{
1389+
stale= true;
1390+
break;
1391+
}
13761392
/* immediately set the refcount to 1 */
13771393
ResourceOwnerEnlargeCatCacheRefs(CurrentResourceOwner);
13781394
ct->refcount++;
@@ -1381,6 +1397,7 @@ SearchCatCacheMiss(CatCache *cache,
13811397
}
13821398

13831399
systable_endscan(scandesc);
1400+
}while (stale);
13841401

13851402
table_close(relation,AccessShareLock);
13861403

@@ -1399,9 +1416,11 @@ SearchCatCacheMiss(CatCache *cache,
13991416
if (IsBootstrapProcessingMode())
14001417
returnNULL;
14011418

1402-
ct=CatalogCacheCreateEntry(cache,NULL,arguments,
1403-
hashValue,hashIndex,
1404-
true);
1419+
ct=CatalogCacheCreateEntry(cache,NULL,NULL,arguments,
1420+
hashValue,hashIndex);
1421+
1422+
/* Creating a negative cache entry shouldn't fail */
1423+
Assert(ct!=NULL);
14051424

14061425
CACHE_elog(DEBUG2,"SearchCatCache(%s): Contains %d/%d tuples",
14071426
cache->cc_relname,cache->cc_ntup,CacheHdr->ch_ntup);
@@ -1608,7 +1627,8 @@ SearchCatCacheList(CatCache *cache,
16081627
* We have to bump the member refcounts temporarily to ensure they won't
16091628
* get dropped from the cache while loading other members. We use a PG_TRY
16101629
* block to ensure we can undo those refcounts if we get an error before
1611-
* we finish constructing the CatCList.
1630+
* we finish constructing the CatCList. ctlist must be valid throughout
1631+
* the PG_TRY block.
16121632
*/
16131633
ResourceOwnerEnlargeCatCacheListRefs(CurrentResourceOwner);
16141634

@@ -1619,19 +1639,23 @@ SearchCatCacheList(CatCache *cache,
16191639
ScanKeyDatacur_skey[CATCACHE_MAXKEYS];
16201640
Relationrelation;
16211641
SysScanDescscandesc;
1622-
1623-
/*
1624-
* Ok, need to make a lookup in the relation, copy the scankey and
1625-
* fill out any per-call fields.
1626-
*/
1627-
memcpy(cur_skey,cache->cc_skey,sizeof(ScanKeyData)*cache->cc_nkeys);
1628-
cur_skey[0].sk_argument=v1;
1629-
cur_skey[1].sk_argument=v2;
1630-
cur_skey[2].sk_argument=v3;
1631-
cur_skey[3].sk_argument=v4;
1642+
boolstale;
16321643

16331644
relation=table_open(cache->cc_reloid,AccessShareLock);
16341645

1646+
do
1647+
{
1648+
/*
1649+
* Ok, need to make a lookup in the relation, copy the scankey and
1650+
* fill out any per-call fields. (We must re-do this when
1651+
* retrying, because systable_beginscan scribbles on the scankey.)
1652+
*/
1653+
memcpy(cur_skey,cache->cc_skey,sizeof(ScanKeyData)*cache->cc_nkeys);
1654+
cur_skey[0].sk_argument=v1;
1655+
cur_skey[1].sk_argument=v2;
1656+
cur_skey[2].sk_argument=v3;
1657+
cur_skey[3].sk_argument=v4;
1658+
16351659
scandesc=systable_beginscan(relation,
16361660
cache->cc_indexoid,
16371661
IndexScanOK(cache,cur_skey),
@@ -1642,6 +1666,8 @@ SearchCatCacheList(CatCache *cache,
16421666
/* The list will be ordered iff we are doing an index scan */
16431667
ordered= (scandesc->irel!=NULL);
16441668

1669+
stale= false;
1670+
16451671
while (HeapTupleIsValid(ntp=systable_getnext(scandesc)))
16461672
{
16471673
uint32hashValue;
@@ -1684,9 +1710,32 @@ SearchCatCacheList(CatCache *cache,
16841710
if (!found)
16851711
{
16861712
/* We didn't find a usable entry, so make a new one */
1687-
ct=CatalogCacheCreateEntry(cache,ntp,arguments,
1688-
hashValue,hashIndex,
1689-
false);
1713+
ct=CatalogCacheCreateEntry(cache,ntp,scandesc,NULL,
1714+
hashValue,hashIndex);
1715+
/* upon failure, we must start the scan over */
1716+
if (ct==NULL)
1717+
{
1718+
/*
1719+
* Release refcounts on any items we already had. We dare
1720+
* not try to free them if they're now unreferenced, since
1721+
* an error while doing that would result in the PG_CATCH
1722+
* below doing extra refcount decrements. Besides, we'll
1723+
* likely re-adopt those items in the next iteration, so
1724+
* it's not worth complicating matters to try to get rid
1725+
* of them.
1726+
*/
1727+
foreach(ctlist_item,ctlist)
1728+
{
1729+
ct= (CatCTup*)lfirst(ctlist_item);
1730+
Assert(ct->c_list==NULL);
1731+
Assert(ct->refcount>0);
1732+
ct->refcount--;
1733+
}
1734+
/* Reset ctlist in preparation for new try */
1735+
ctlist=NIL;
1736+
stale= true;
1737+
break;
1738+
}
16901739
}
16911740

16921741
/* Careful here: add entry to ctlist, then bump its refcount */
@@ -1696,6 +1745,7 @@ SearchCatCacheList(CatCache *cache,
16961745
}
16971746

16981747
systable_endscan(scandesc);
1748+
}while (stale);
16991749

17001750
table_close(relation,AccessShareLock);
17011751

@@ -1803,22 +1853,42 @@ ReleaseCatCacheList(CatCList *list)
18031853
* CatalogCacheCreateEntry
18041854
*Create a new CatCTup entry, copying the given HeapTuple and other
18051855
*supplied data into it. The new entry initially has refcount 0.
1856+
*
1857+
* To create a normal cache entry, ntp must be the HeapTuple just fetched
1858+
* from scandesc, and "arguments" is not used. To create a negative cache
1859+
* entry, pass NULL for ntp and scandesc; then "arguments" is the cache
1860+
* keys to use. In either case, hashValue/hashIndex are the hash values
1861+
* computed from the cache keys.
1862+
*
1863+
* Returns NULL if we attempt to detoast the tuple and observe that it
1864+
* became stale. (This cannot happen for a negative entry.) Caller must
1865+
* retry the tuple lookup in that case.
18061866
*/
18071867
staticCatCTup*
1808-
CatalogCacheCreateEntry(CatCache*cache,HeapTuplentp,Datum*arguments,
1809-
uint32hashValue,IndexhashIndex,
1810-
boolnegative)
1868+
CatalogCacheCreateEntry(CatCache*cache,HeapTuplentp,SysScanDescscandesc,
1869+
Datum*arguments,
1870+
uint32hashValue,IndexhashIndex)
18111871
{
18121872
CatCTup*ct;
18131873
HeapTupledtp;
18141874
MemoryContextoldcxt;
18151875

1816-
/* negative entries have no tuple associated */
18171876
if (ntp)
18181877
{
18191878
inti;
18201879

1821-
Assert(!negative);
1880+
/*
1881+
* The visibility recheck below essentially never fails during our
1882+
* regression tests, and there's no easy way to force it to fail for
1883+
* testing purposes. To ensure we have test coverage for the retry
1884+
* paths in our callers, make debug builds randomly fail about 0.1% of
1885+
* the times through this code path, even when there's no toasted
1886+
* fields.
1887+
*/
1888+
#ifdefUSE_ASSERT_CHECKING
1889+
if (random() <= (MAX_RANDOM_VALUE /1000))
1890+
returnNULL;
1891+
#endif
18221892

18231893
/*
18241894
* If there are any out-of-line toasted fields in the tuple, expand
@@ -1828,7 +1898,20 @@ CatalogCacheCreateEntry(CatCache *cache, HeapTuple ntp, Datum *arguments,
18281898
* something using a slightly stale catcache entry.
18291899
*/
18301900
if (HeapTupleHasExternal(ntp))
1901+
{
18311902
dtp=toast_flatten_tuple(ntp,cache->cc_tupdesc);
1903+
1904+
/*
1905+
* The tuple could become stale while we are doing toast table
1906+
* access (since AcceptInvalidationMessages can run then), so we
1907+
* must recheck its visibility afterwards.
1908+
*/
1909+
if (!systable_recheck_tuple(scandesc,ntp))
1910+
{
1911+
heap_freetuple(dtp);
1912+
returnNULL;
1913+
}
1914+
}
18321915
else
18331916
dtp=ntp;
18341917

@@ -1867,7 +1950,7 @@ CatalogCacheCreateEntry(CatCache *cache, HeapTuple ntp, Datum *arguments,
18671950
}
18681951
else
18691952
{
1870-
Assert(negative);
1953+
/* Set up keys for anegative cache entry */
18711954
oldcxt=MemoryContextSwitchTo(CacheMemoryContext);
18721955
ct= (CatCTup*)palloc(sizeof(CatCTup));
18731956

@@ -1889,7 +1972,7 @@ CatalogCacheCreateEntry(CatCache *cache, HeapTuple ntp, Datum *arguments,
18891972
ct->c_list=NULL;
18901973
ct->refcount=0;/* for the moment */
18911974
ct->dead= false;
1892-
ct->negative=negative;
1975+
ct->negative=(ntp==NULL);
18931976
ct->hash_value=hashValue;
18941977

18951978
dlist_push_head(&cache->cc_bucket[hashIndex],&ct->cache_elem);

0 commit comments

Comments
 (0)
[8]ページ先頭
©2009-2025 Movatter.jp