disable READA

There's a bug in bread() which can cause it to misinterpret a failed READA request as an IO error on SMP.

If a filesytem block is subject to READA and __make_request() decides that the request would block then __make_request() will (via end_buffer_io_sync) mark the buffer as not uptodate and will unlock it.

But if another CPU attempts to bread() the same buffer at the same time, the following happens:

1: ll_rw_block() sees the buffer is locked and does nothing at all

2: bread() does a wait_on_buffer()

3: the other CPU (the one doing READA) unlocks the non uptodate buffer

4: bread() sees the buffer come unlocked. It's not uptodate, so
bread() bogusly reports an IO error.

I haven't seen it in the wild, because it is rare to get that much read I/O in flight. reiserfs and ext3 (at least) use READA.

An appropriate fix for 2.4 is to disable READA.

 ll_rw_blk.c |    2 ++
 1 files changed, 2 insertions

drivers/block/ll_rw_blk.c renamed from drivers/block/ll_rw_blk.c~no-readahead +2 -0

@@ -841,7 +841,9 @@ static int __make_request(request_queue_
 	rw_ahead = 0;	/* normal case; gets changed below for READA */
 	switch (rw) {
 		case READA:
+#if 0	/* bread() misinterprets failed READA attempts as IO errors on SMP */
 			rw_ahead = 1;
+#endif
 			rw = READ;	/* drop into READ */
 		case READ:
 		case WRITE:

.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Who is using READA this days ?

Quoted 42 lines · Andrew Morton

There's a bug in bread() which can cause it to misinterpret a failed READA request as an IO error on SMP.

If a filesytem block is subject to READA and __make_request() decides that the request would block then __make_request() will (via end_buffer_io_sync) mark the buffer as not uptodate and will unlock it.

But if another CPU attempts to bread() the same buffer at the same time, the following happens:

1: ll_rw_block() sees the buffer is locked and does nothing at all

2: bread() does a wait_on_buffer()

3: the other CPU (the one doing READA) unlocks the non uptodate buffer

4: bread() sees the buffer come unlocked. It's not uptodate, so
bread() bogusly reports an IO error.

I haven't seen it in the wild, because it is rare to get that much read I/O in flight. reiserfs and ext3 (at least) use READA.

An appropriate fix for 2.4 is to disable READA.

 ll_rw_blk.c |    2 ++
 1 files changed, 2 insertions

drivers/block/ll_rw_blk.c renamed from drivers/block/ll_rw_blk.c~no-readahead +2 -0

@@ -841,7 +841,9 @@ static int __make_request(request_queue_
 	rw_ahead = 0;	/* normal case; gets changed below for READA */
 	switch (rw) {
 		case READA:
+#if 0	/* bread() misinterprets failed READA attempts as IO errors on SMP */
 			rw_ahead = 1;
+#endif
 			rw = READ;	/* drop into READ */
 		case READ:
 		case WRITE: