Avoid deadlocks during insertion into SP-GiST indexes.

SP-GiST's original scheme for avoiding deadlocks during concurrent index
insertions doesn't work, as per report from Hailong Li, and there isn't any
evident way to make it work completely.  We could possibly lock individual
inner tuples instead of their whole pages, but preliminary experimentation
suggests that the performance penalty would be huge.  Instead, if we fail
to get a buffer lock while descending the tree, just restart the tree
descent altogether.  We keep the old tuple positioning rules, though, in
hopes of reducing the number of cases where this can happen.

Teodor Sigaev, somewhat edited by Tom Lane

src/backend/access/spgist/README

@@ -201,17 +201,31 @@ space utilization, but doesn't change the basis of the algorithm.
 CONCURRENCY
 
 While descending the tree, the insertion algorithm holds exclusive lock on
-two tree levels at a time, ie both parent and child pages (parent and child
-pages can be the same, see notes above). There is a possibility of deadlock
-between two insertions if there are cross-referenced pages in different
-branches.  That is, if inner tuple on page M has a child on page N while
-an inner tuple from another branch is on page N and has a child on page M,
-then two insertions descending the two branches could deadlock.  To prevent
-deadlocks we introduce a concept of "triple parity" of pages: if inner tuple
-is on page with BlockNumber N, then its child tuples should be placed on the
-same page, or else on a page with BlockNumber M where (N+1) mod 3 == M mod 3.
-This rule guarantees that tuples on page M will have no children on page N,
-since (M+1) mod 3 != N mod 3.
+two tree levels at a time, ie both parent and child pages (but parent and
+child pages can be the same, see notes above).  There is a possibility of
+deadlock between two insertions if there are cross-referenced pages in
+different branches.  That is, if inner tuple on page M has a child on page N
+while an inner tuple from another branch is on page N and has a child on
+page M, then two insertions descending the two branches could deadlock,
+since they will each hold their parent page's lock while trying to get the
+child page's lock.
+
+Currently, we deal with this by conditionally locking buffers as we descend
+the tree.  If we fail to get lock on a buffer, we release both buffers and
+restart the insertion process.  This is potentially inefficient, but the
+locking costs of a more deterministic approach seem very high.
+
+To reduce the number of cases where that happens, we introduce a concept of
+"triple parity" of pages: if inner tuple is on page with BlockNumber N, then
+its child tuples should be placed on the same page, or else on a page with
+BlockNumber M where (N+1) mod 3 == M mod 3.  This rule ensures that tuples
+on page M will have no children on page N, since (M+1) mod 3 != N mod 3.
+That makes it unlikely that two insertion processes will conflict against
+each other while descending the tree.  It's not perfect though: in the first
+place, we could still get a deadlock among three or more insertion processes,
+and in the second place, it's impractical to preserve this invariant in every
+case when we expand or split an inner tuple.  So we still have to allow for
+deadlocks.
 
 Insertion may also need to take locks on an additional inner and/or leaf page
 to add tuples of the right type(s), when there's not enough room on the pages

src/backend/access/spgist/spgdoinsert.c

@@ -1836,9 +1836,13 @@ spgSplitNodeAction(Relation index, SpGistState *state,
 }
 
 /*
- * Insert one item into the index
+ * Insert one item into the index.
+ *
+ * Returns true on success, false if we failed to complete the insertion
+ * because of conflict with a concurrent insert.  In the latter case,
+ * caller should re-call spgdoinsert() with the same args.
  */
-void
+bool
 spgdoinsert(Relation index, SpGistState *state,
 			ItemPointer heapPtr, Datum datum, bool isnull)
 {
@@ -1927,12 +1931,32 @@ spgdoinsert(Relation index, SpGistState *state,
 								&isNew);
 			current.blkno = BufferGetBlockNumber(current.buffer);
 		}
-		else if (parent.buffer == InvalidBuffer ||
-				 current.blkno != parent.blkno)
+		else if (parent.buffer == InvalidBuffer)
 		{
+			/* we hold no parent-page lock, so no deadlock is possible */
 			current.buffer = ReadBuffer(index, current.blkno);
 			LockBuffer(current.buffer, BUFFER_LOCK_EXCLUSIVE);
 		}
+		else if (current.blkno != parent.blkno)
+		{
+			/* descend to a new child page */
+			current.buffer = ReadBuffer(index, current.blkno);
+
+			/*
+			 * Attempt to acquire lock on child page.  We must beware of
+			 * deadlock against another insertion process descending from that
+			 * page to our parent page (see README).  If we fail to get lock,
+			 * abandon the insertion and tell our caller to start over.  XXX
+			 * this could be improved; perhaps it'd be worth sleeping a bit
+			 * before giving up?
+			 */
+			if (!ConditionalLockBuffer(current.buffer))
+			{
+				ReleaseBuffer(current.buffer);
+				UnlockReleaseBuffer(parent.buffer);
+				return false;
+			}
+		}
 		else
 		{
 			/* inner tuple can be stored on the same page as parent one */
@@ -2131,4 +2155,6 @@ spgdoinsert(Relation index, SpGistState *state,
 		SpGistSetLastUsedPage(index, parent.buffer);
 		UnlockReleaseBuffer(parent.buffer);
 	}
+
+	return true;
 }

src/backend/access/spgist/spginsert.c

@@ -45,7 +45,10 @@ spgistBuildCallback(Relation index, HeapTuple htup, Datum *values,
 	/* Work in temp context, and reset it after each tuple */
 	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
 
-	spgdoinsert(index, &buildstate->spgstate, &htup->t_self, *values, *isnull);
+	/* No concurrent insertions can be happening, so failure is unexpected */
+	if (!spgdoinsert(index, &buildstate->spgstate, &htup->t_self,
+					 *values, *isnull))
+		elog(ERROR, "unexpected spgdoinsert() failure");
 
 	MemoryContextSwitchTo(oldCtx);
 	MemoryContextReset(buildstate->tmpCtx);
@@ -219,7 +222,17 @@ spginsert(PG_FUNCTION_ARGS)
 
 	initSpGistState(&spgstate, index);
 
-	spgdoinsert(index, &spgstate, ht_ctid, *values, *isnull);
+	/*
+	 * We might have to repeat spgdoinsert() multiple times, if conflicts
+	 * occur with concurrent insertions.  If so, reset the insertCtx each time
+	 * to avoid cumulative memory consumption.	That means we also have to
+	 * redo initSpGistState(), but it's cheap enough not to matter.
+	 */
+	while (!spgdoinsert(index, &spgstate, ht_ctid, *values, *isnull))
+	{
+		MemoryContextReset(insertCtx);
+		initSpGistState(&spgstate, index);
+	}
 
 	SpGistUpdateMetaPage(index);
 

src/include/access/spgist_private.h

@@ -652,7 +652,7 @@ extern void spgPageIndexMultiDelete(SpGistState *state, Page page,
 						OffsetNumber *itemnos, int nitems,
 						int firststate, int reststate,
 						BlockNumber blkno, OffsetNumber offnum);
-extern void spgdoinsert(Relation index, SpGistState *state,
+extern bool spgdoinsert(Relation index, SpGistState *state,
 			ItemPointer heapPtr, Datum datum, bool isnull);
 
 #endif   /* SPGIST_PRIVATE_H */