Re: [PATCH 3/4] scsi: Internally retry scsi_execute commands

[Mike Christie <michael.christie@oracle.com>]

On 8/11/22 4:56 AM, Martin Wilck wrote:
>> It could change how 0x04/0x0a is handled because it uses NEEDS_RETRY.
>> However, scsi_dh_alua uses REQ_FAILFAST_DEV so we do not retry in
>> scsi_noretry_cmd like before.
> 
> Not quite following you here - alua_check_sense() is called for any
> command, not just those submitted from the ALUA code.

Ah, I thought you had mentioned alua above because of your alua_rtpg
example. Above I was saying that there was no behavior change for your
alua_rtpg example because it uses REQ_FAILFAST_DEV.


>> 2. Instead of trying to make it general for all scsi_execute_users,
>> we can
>> add SCMD bits for specific cases like DID_TIME_OUT or a SCMD bit that
>> tells
>> scsi_noretry_cmd to not always fail passthrough commands just because
>> they
>> are passthrough. It would work the opposite of the FASTFAIL bits
>> where instead
>> of failing fast, we retry.
>>
>> I think because the cases scsi_noretry_cmd is used for are really
>> specific cases
>> (scsi_decide_disposition sees NEEDS_RETRY, retries < allowed, and
>> REQ_FAILFAST_DEV
>> is not set) that might not be very useful. 
> 
> I don't think it's _that_ speficic. (retries < allowed) is the default
> case, at least for the first failure. REQ_FAILFAST_DEV has very few
> users except for the device handlers, and NEEDS_RETRY is a rather
> frequently used disposition.
I'm saying it's really specific because we only hit this code
path that is causing issues when scsi_check_sense returns NEEDS_RETRY.
There's 5 in there and one in scsi_dh_alua. 4 of them are UAs.

Compared to all the sense errors that we check for in the
scsi_execute callers and including all the times they do a retry for
all errors the 5 cases in scsi_check_sense seemed really specific.

Let me send a patch for this type of design because in the other mail
Christoph was asking for more details. I originally started going that
route so it won't be too much trouble to do a RFC so we can get an
idea of what it will look like.