On 8/5/24 10:47 AM, Damien Le Moal wrote:
On 2024/08/03 13:42, Mike Christie wrote:
Do you just want to retry UAs or all internal passthrough command
errors that go through here?
I think that we definitely do *not* want to retry all internal commands. E.g. I
do not see the point in retrying INQUIRY or VPD page accesses failing on device
scan.
For just the specific UA or NOT_READY/0x04/0x01 case for an example.
Does every scsi core passthrough caller want that retried? If so, then
I can see where you are coming from where you feel it's a bug. I would
agree we would normally want to retry that in general. Maybe others know
about some specific old case though.
I think the command Bart had a problem with is START_STOP_UNIT. If that is the
only case causing problems, then we should probably improve
sd_start_stop_device() to allow retries. But if there are more commands that
needs the same, then the patch that Bart is proposing makes sense I think, but
it will require an audit of all REQ_OP_DRV_XXX internal users to make sure that
they all use that command with -> allowed set to 0. But hopefully, most cases go
through scsi_execute_cmnd(), which should simplify this audit.
Thanks Mike and Damien for having shared your thoughts. I only need the
START STOP UNIT command to be retried. So the lowest risk approach is to
introduce a mechanism that only causes that command to be retried. I
plan to look into this and prepare a patch.
Bart.
