Fix locking when fixing an incomplete split of a GIN internal page

ginFinishSplit() expects the caller to hold an exclusive lock on the
buffer, but when finishing an earlier "leftover" incomplete split of
an internal page, the caller held a shared lock. That caused an
assertion failure in MarkBufferDirty(). Without assertions, it could
lead to corruption if two backends tried to complete the split at the
same time.

On master, add a test case using the new injection point facility.

Report and analysis by Fei Changhong. Backpatch the fix to all
supported versions.

Reviewed-by: Fei Changhong, Michael Paquier
Discussion: https://www.postgresql.org/message-id/tencent_A3CE810F59132D8E230475A5F0F7A08C8307@qq.com
This commit is contained in:
Heikki Linnakangas 2024-01-29 13:46:22 +02:00
parent edbd1b41ab
commit b899e00e71
1 changed files with 50 additions and 18 deletions

View File

@ -28,6 +28,8 @@ static bool ginPlaceToPage(GinBtree btree, GinBtreeStack *stack,
Buffer childbuf, GinStatsData *buildStats);
static void ginFinishSplit(GinBtree btree, GinBtreeStack *stack,
bool freestack, GinStatsData *buildStats);
static void ginFinishOldSplit(GinBtree btree, GinBtreeStack *stack,
GinStatsData *buildStats, int access);
/*
* Lock buffer by needed method for search.
@ -109,7 +111,7 @@ ginFindLeafPage(GinBtree btree, bool searchMode,
* encounter on the way.
*/
if (!searchMode && GinPageIsIncompleteSplit(page))
ginFinishSplit(btree, stack, false, NULL);
ginFinishOldSplit(btree, stack, NULL, access);
/*
* ok, page is correctly locked, we should check to move right ..,
@ -130,7 +132,7 @@ ginFindLeafPage(GinBtree btree, bool searchMode,
TestForOldSnapshot(snapshot, btree->index, page);
if (!searchMode && GinPageIsIncompleteSplit(page))
ginFinishSplit(btree, stack, false, NULL);
ginFinishOldSplit(btree, stack, NULL, access);
}
if (GinPageIsLeaf(page)) /* we found, return locked page */
@ -166,8 +168,11 @@ ginFindLeafPage(GinBtree btree, bool searchMode,
* Step right from current page.
*
* The next page is locked first, before releasing the current page. This is
* crucial to protect from concurrent page deletion (see comment in
* ginDeletePage).
* crucial to prevent concurrent VACUUM from deleting a page that we are about
* to step to. (The lock-coupling isn't strictly necessary when we are
* traversing the tree to find an insert location, because page deletion grabs
* a cleanup lock on the root to prevent any concurrent inserts. See Page
* deletion section in the README. But there's no harm in doing it always.)
*/
Buffer
ginStepRight(Buffer buffer, Relation index, int lockmode)
@ -264,7 +269,7 @@ ginFindParents(GinBtree btree, GinBtreeStack *stack)
ptr->parent = root;
ptr->off = InvalidOffsetNumber;
ginFinishSplit(btree, ptr, false, NULL);
ginFinishOldSplit(btree, ptr, NULL, GIN_EXCLUSIVE);
}
leftmostBlkno = btree->getLeftMostChild(btree, page);
@ -293,7 +298,7 @@ ginFindParents(GinBtree btree, GinBtreeStack *stack)
ptr->parent = root;
ptr->off = InvalidOffsetNumber;
ginFinishSplit(btree, ptr, false, NULL);
ginFinishOldSplit(btree, ptr, NULL, GIN_EXCLUSIVE);
}
}
@ -674,15 +679,6 @@ ginFinishSplit(GinBtree btree, GinBtreeStack *stack, bool freestack,
bool done;
bool first = true;
/*
* freestack == false when we encounter an incompletely split page during
* a scan, while freestack == true is used in the normal scenario that a
* split is finished right after the initial insert.
*/
if (!freestack)
elog(DEBUG1, "finishing incomplete split of block %u in gin index \"%s\"",
stack->blkno, RelationGetRelationName(btree->index));
/* this loop crawls up the stack until the insertion is complete */
do
{
@ -703,7 +699,7 @@ ginFinishSplit(GinBtree btree, GinBtreeStack *stack, bool freestack,
* would fail.
*/
if (GinPageIsIncompleteSplit(BufferGetPage(parent->buffer)))
ginFinishSplit(btree, parent, false, buildStats);
ginFinishOldSplit(btree, parent, buildStats, GIN_EXCLUSIVE);
/* move right if it's needed */
page = BufferGetPage(parent->buffer);
@ -727,7 +723,7 @@ ginFinishSplit(GinBtree btree, GinBtreeStack *stack, bool freestack,
page = BufferGetPage(parent->buffer);
if (GinPageIsIncompleteSplit(BufferGetPage(parent->buffer)))
ginFinishSplit(btree, parent, false, buildStats);
ginFinishOldSplit(btree, parent, buildStats, GIN_EXCLUSIVE);
}
/* insert the downlink */
@ -763,6 +759,42 @@ ginFinishSplit(GinBtree btree, GinBtreeStack *stack, bool freestack,
freeGinBtreeStack(stack);
}
/*
* An entry point to ginFinishSplit() that is used when we stumble upon an
* existing incompletely split page in the tree, as opposed to completing a
* split that we just made outselves. The difference is that stack->buffer may
* be merely share-locked on entry, and will be upgraded to exclusive mode.
*
* Note: Upgrading the lock momentarily releases it. Doing that in a scan
* would not be OK, because a concurrent VACUUM might delete the page while
* we're not holding the lock. It's OK in an insert, though, because VACUUM
* has a different mechanism that prevents it from running concurrently with
* inserts. (Namely, it holds a cleanup lock on the root.)
*/
static void
ginFinishOldSplit(GinBtree btree, GinBtreeStack *stack, GinStatsData *buildStats, int access)
{
elog(DEBUG1, "finishing incomplete split of block %u in gin index \"%s\"",
stack->blkno, RelationGetRelationName(btree->index));
if (access == GIN_SHARE)
{
LockBuffer(stack->buffer, GIN_UNLOCK);
LockBuffer(stack->buffer, GIN_EXCLUSIVE);
if (!GinPageIsIncompleteSplit(BufferGetPage(stack->buffer)))
{
/*
* Someone else already completed the split while we were not
* holding the lock.
*/
return;
}
}
ginFinishSplit(btree, stack, false, buildStats);
}
/*
* Insert a value to tree described by stack.
*
@ -783,7 +815,7 @@ ginInsertValue(GinBtree btree, GinBtreeStack *stack, void *insertdata,
/* If the leaf page was incompletely split, finish the split first */
if (GinPageIsIncompleteSplit(BufferGetPage(stack->buffer)))
ginFinishSplit(btree, stack, false, buildStats);
ginFinishOldSplit(btree, stack, buildStats, GIN_EXCLUSIVE);
done = ginPlaceToPage(btree, stack,
insertdata, InvalidBlockNumber,