error returns from ->queue_rq

error returns from ->queue_rq

[Christoph Hellwig]

Hi Jens,

seems like with the SCSI work I introduced the first
BLK_MQ_RQ_QUEUE_ERROR error return in the tree, and immediately ran into
the first pitfall.  The code as-is expects rq->errors set to an error
value, which otherwise is an  internal field used by the block layer and
some drivers, but not part of the communication protocol between the
two.

We can either make it part of the protocol for blk-mq, which would
require documenting and praying driver writers get it right, or
alternatively we could map BLK_MQ_RQ_QUEUE_ERROR to -EIO and if
nessecary introduce other return values if we need to return other
errors.  The third option would be to remove BLK_MQ_RQ_QUEUE_ERROR
entirely and require drivers to call blk_mq_end_io themselves from
->queue_rq, mirroring the ->queuecommand error handling.  I'm undecided
between options 2 and 3, but I'd rather avoid the current pitfall.

Re: error returns from ->queue_rq

[Jens Axboe]

On Tue, Feb 11 2014, Christoph Hellwig wrote:
> Hi Jens,
> 
> seems like with the SCSI work I introduced the first
> BLK_MQ_RQ_QUEUE_ERROR error return in the tree, and immediately ran into
> the first pitfall.  The code as-is expects rq->errors set to an error
> value, which otherwise is an  internal field used by the block layer and
> some drivers, but not part of the communication protocol between the
> two.
> 
> We can either make it part of the protocol for blk-mq, which would
> require documenting and praying driver writers get it right, or
> alternatively we could map BLK_MQ_RQ_QUEUE_ERROR to -EIO and if
> nessecary introduce other return values if we need to return other
> errors.  The third option would be to remove BLK_MQ_RQ_QUEUE_ERROR
> entirely and require drivers to call blk_mq_end_io themselves from
> ->queue_rq, mirroring the ->queuecommand error handling.  I'm undecided
> between options 2 and 3, but I'd rather avoid the current pitfall.

Lets go for option #2. I agree that we should not make this a new
depedency, it's a lot more robust to just have it be -EIO and add other
error returns as needed.

-- 
Jens Axboe