Repair bufmgr deadlock problem reported by Michael Wildpaner. Must take

share lock on a buffer being written out before releasing BufMgrLock in
the BufferAlloc code path; if we do it later we might block on someone
who's re-pinned the buffer.  I believe this is only an issue for BufferAlloc
and not the other places that call FlushBuffer.  BufferSync must continue
to do it the old way since it may well be trying to write buffers that
other backends have pinned; but it should not be holding any conflicting
locks.  FlushRelationBuffers is okay since it's got exclusive lock at the
relation level.

src/backend/storage/buffer/bufmgr.c

@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *	  $PostgreSQL: pgsql/src/backend/storage/buffer/bufmgr.c,v 1.183 2004/12/31 22:00:49 pgsql Exp $
+ *	  $PostgreSQL: pgsql/src/backend/storage/buffer/bufmgr.c,v 1.184 2005/01/03 18:49:41 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -84,7 +84,7 @@ static Buffer ReadBufferInternal(Relation reln, BlockNumber blockNum,
 				   bool bufferLockHeld);
 static BufferDesc *BufferAlloc(Relation reln, BlockNumber blockNum,
 			bool *foundPtr);
-static void FlushBuffer(BufferDesc *buf, SMgrRelation reln);
+static void FlushBuffer(BufferDesc *buf, SMgrRelation reln, bool earlylock);
 static void write_buffer(Buffer buffer, bool unpin);
 
 
@@ -340,6 +340,10 @@ BufferAlloc(Relation reln,
 		 * allocated -- ours.  If it had a pin it wouldn't have been on
 		 * the free list.  No one else could have pinned it between
 		 * StrategyGetBuffer and here because we have the BufMgrLock.
+		 *
+		 * (We must pin the buffer before releasing BufMgrLock ourselves,
+		 * to ensure StrategyGetBuffer won't give the same buffer to someone
+		 * else.)
 		 */
 		Assert(buf->refcount == 0);
 		buf->refcount = 1;
@@ -367,9 +371,20 @@ BufferAlloc(Relation reln,
 
 			/*
 			 * Write the buffer out, being careful to release BufMgrLock
-			 * while doing the I/O.
+			 * while doing the I/O.  We also tell FlushBuffer to share-lock
+			 * the buffer before releasing BufMgrLock.  This is safe because
+			 * we know no other backend currently has the buffer pinned,
+			 * therefore no one can have it locked either, so we can always
+			 * get the lock without blocking.  It is necessary because if
+			 * we release BufMgrLock first, it's possible for someone else
+			 * to pin and exclusive-lock the buffer before we get to the
+			 * share-lock, causing us to block.  If the someone else then
+			 * blocks on a lock we hold, deadlock ensues.  This has been
+			 * observed to happen when two backends are both trying to split
+			 * btree index pages, and the second one just happens to be
+			 * trying to split the page the first one got from the freelist.
 			 */
-			FlushBuffer(buf, NULL);
+			FlushBuffer(buf, NULL, true);
 
 			/*
 			 * Somebody could have allocated another buffer for the same
@@ -766,7 +781,7 @@ BufferSync(int percent, int maxpages)
 		PinBuffer(bufHdr, true);
 		StartBufferIO(bufHdr, false);
 
-		FlushBuffer(bufHdr, NULL);
+		FlushBuffer(bufHdr, NULL, false);
 
 		TerminateBufferIO(bufHdr, 0);
 		UnpinBuffer(bufHdr, true);
@@ -1018,11 +1033,16 @@ BufferGetFileNode(Buffer buffer)
  * If the caller has an smgr reference for the buffer's relation, pass it
  * as the second parameter.  If not, pass NULL.  (Do not open relation
  * while holding BufMgrLock!)
+ *
+ * When earlylock is TRUE, we grab the per-buffer sharelock before releasing
+ * BufMgrLock, rather than after.  Normally this would be a bad idea since
+ * we might deadlock, but it is safe and necessary when called from
+ * BufferAlloc() --- see comments therein.
  */
 static void
-FlushBuffer(BufferDesc *buf, SMgrRelation reln)
+FlushBuffer(BufferDesc *buf, SMgrRelation reln, bool earlylock)
 {
-	Buffer		buffer;
+	Buffer		buffer = BufferDescriptorGetBuffer(buf);
 	XLogRecPtr	recptr;
 	ErrorContextCallback errcontext;
 
@@ -1033,6 +1053,13 @@ FlushBuffer(BufferDesc *buf, SMgrRelation reln)
 	/* To check if block content changed while flushing. - vadim 01/17/97 */
 	buf->flags &= ~BM_JUST_DIRTIED;
 
+	/*
+	 * If earlylock, grab buffer sharelock before anyone else could re-lock
+	 * the buffer.
+	 */
+	if (earlylock)
+		LockBuffer(buffer, BUFFER_LOCK_SHARE);
+
 	/* Release BufMgrLock while doing xlog work */
 	LWLockRelease(BufMgrLock);
 
@@ -1046,14 +1073,13 @@ FlushBuffer(BufferDesc *buf, SMgrRelation reln)
 	if (reln == NULL)
 		reln = smgropen(buf->tag.rnode);
 
-	buffer = BufferDescriptorGetBuffer(buf);
-
 	/*
 	 * Protect buffer content against concurrent update.  (Note that
 	 * hint-bit updates can still occur while the write is in progress,
 	 * but we assume that that will not invalidate the data written.)
 	 */
-	LockBuffer(buffer, BUFFER_LOCK_SHARE);
+	if (!earlylock)
+		LockBuffer(buffer, BUFFER_LOCK_SHARE);
 
 	/*
 	 * Force XLOG flush for buffer' LSN.  This implements the basic WAL
@@ -1485,7 +1511,7 @@ FlushRelationBuffers(Relation rel, BlockNumber firstDelBlock)
 				{
 					StartBufferIO(bufHdr, false);
 
-					FlushBuffer(bufHdr, rel->rd_smgr);
+					FlushBuffer(bufHdr, rel->rd_smgr, false);
 
 					TerminateBufferIO(bufHdr, 0);
 				}