RFC: Retrying SCSI pass-through commands

[Bart Van Assche <bvanassche@acm.org>]

Hi,

Recently I noticed that a particular UFS-based device does not resume
correctly. The logs of the device show that sd_start_stop_device() does
not retry the START STOP UNIT command if the device reports a unit
attention. I think that's a bug in the SCSI core. The following hack
makes resume work again. I think this confirms my understanding of this
issue (sd_start_stop_device() sets RQF_PM):

diff --git a/drivers/scsi/scsi_error.c b/drivers/scsi/scsi_error.c
index da7dac77f8cd..e21becc5bcf9 100644
--- a/drivers/scsi/scsi_error.c
+++ b/drivers/scsi/scsi_error.c
@@ -1816,6 +1816,8 @@ bool scsi_noretry_cmd(struct scsi_cmnd *scmd)
          * assume caller has checked sense and determined
          * the check condition was retryable.
          */
+       if (req->rq_flags & RQF_PM)
+               return false;
         if (req->cmd_flags & REQ_FAILFAST_DEV || blk_rq_is_passthrough(req))
                 return true;

My understanding is that SCSI pass-through commands submitted from
user space must not be retried. Are there any objections against
modifying the behavior of the SCSI core such that it retries
REQ_OP_DRV_* operations submitted by the SCSI core, as illustrated
by the pseudo-code below?

diff --git a/drivers/scsi/scsi_error.c b/drivers/scsi/scsi_error.c
index da7dac77f8cd..e21becc5bcf9 100644
--- a/drivers/scsi/scsi_error.c
+++ b/drivers/scsi/scsi_error.c
@@ -1816,6 +1816,12 @@ bool scsi_noretry_cmd(struct scsi_cmnd *scmd)
          * assume caller has checked sense and determined
          * the check condition was retryable.
          */
-       if (req->cmd_flags & REQ_FAILFAST_DEV || blk_rq_is_passthrough(req))
-               return true;
+       if (req->cmd_flags & REQ_FAILFAST_DEV)
+               return true;
+       if (/* submitted by the SCSI core */)
+               return false;
+       if (blk_rq_is_passthrough(req))
+               return true;

Thanks,

Bart.