scsi_lib: correctly retry failed zero length REQ_TYPE_FS commands

commit a621bac3044ed6f7ec5fa0326491b2d4838bfa93 upstream.

When SCSI was written, all commands coming from the filesystem
(REQ_TYPE_FS commands) had data.  This meant that our signal for needing
to complete the command was the number of bytes completed being equal to
the number of bytes in the request.  Unfortunately, with the advent of
flush barriers, we can now get zero length REQ_TYPE_FS commands, which
confuse this logic because they satisfy the condition every time.  This
means they never get retried even for retryable conditions, like UNIT
ATTENTION because we complete them early assuming they're done.  Fix
this by special casing the early completion condition to recognise zero
length commands with errors and let them drop through to the retry code.

Reported-by: Sebastian Parschauer <s.parschauer@gmx.de>
Signed-off-by: James E.J. Bottomley <jejb@linux.vnet.ibm.com>
Tested-by: Jack Wang <jinpu.wang@profitbricks.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
[bwh: Backported to 3.16: adjust context]
Signed-off-by: Ben Hutchings <ben@decadent.org.uk>

diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
index d7062a06f72ef0fa0e7ce70d8d65b7427d1cc689..7e817b1f95f07dbe7c7a522ff8e482c7b00b1aeb 100644 (file)
--- a/drivers/scsi/scsi_lib.c
+++ b/drivers/scsi/scsi_lib.c
@@ -775,9 +775,12 @@ void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
        }
 
        /*
-        * If we finished all bytes in the request we are done now.
+        * special case: failed zero length commands always need to
+        * drop down into the retry code. Otherwise, if we finished
+        * all bytes in the request we are done now.
         */
-       if (!blk_end_request(req, error, good_bytes))
+       if (!(blk_rq_bytes(req) == 0 && error) &&
+           !blk_end_request(req, error, good_bytes))
                goto next_command;
 
        /*