scsi: core: Fix a race between scsi_done() and scsi_timeout()

If there is a race between scsi_done() and scsi_timeout() and if
scsi_timeout() loses the race, scsi_timeout() should not reset the request
timer. Hence change the return value for this case from BLK_EH_RESET_TIMER
into BLK_EH_DONE.

Although the block layer holds a reference on a request (req->ref) while
calling a timeout handler, restarting the timer (blk_add_timer()) while a
request is being completed is racy.

Reviewed-by: Mike Christie <[e-mail]>
Cc: Keith Busch <[e-mail]>
Cc: Christoph Hellwig <[e-mail]>
Cc: Ming Lei <[e-mail]>
Cc: John Garry <[e-mail]>
Cc: Hannes Reinecke <[e-mail]>
Reported-by: Adrian Hunter <[e-mail]>
Fixes: 15f73f5b3e59 ("blk-mq: move failure injection out of blk_mq_complete_request")
Signed-off-by: Bart Van Assche <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]
Signed-off-by: Martin K. Petersen <[e-mail]>

drivers/scsi/scsi_error.c

···
 
 	if (rtn == BLK_EH_DONE) {
 		/*
-		 * Set the command to complete first in order to prevent a real
-		 * completion from releasing the command while error handling
-		 * is using it. If the command was already completed, then the
-		 * lower level driver beat the timeout handler, and it is safe
-		 * to return without escalating error recovery.
-		 *
-		 * If timeout handling lost the race to a real completion, the
-		 * block layer may ignore that due to a fake timeout injection,
-		 * so return RESET_TIMER to allow error handling another shot
-		 * at this command.
+		 * If scsi_done() has already set SCMD_STATE_COMPLETE, do not
+		 * modify *scmd.
 		 */
 		if (test_and_set_bit(SCMD_STATE_COMPLETE, &scmd->state))
-			return BLK_EH_RESET_TIMER;
+			return BLK_EH_DONE;
 		if (scsi_abort_command(scmd) != SUCCESS) {
 			set_host_byte(scmd, DID_TIME_OUT);
 			scsi_eh_scmd_add(scmd);