On Fri, Jul 05, 2019 at 09:12:37AM +0200, Hannes Reinecke wrote:
> On 7/4/19 3:19 PM, Paolo Bonzini wrote:
> > On 19/06/19 12:31, Paolo Bonzini wrote:
> >>> I'm a bit unsure if 'bd->last' is always set; it's quite obvious that
> >>> it's present if set, but what about requests with 'bd->last == false' ?
> >>> Is there a guarantee that they will _always_ be followed with a request
> >>> with bd->last == true?
> >>> And if so, is there a guarantee that this request is part of the same batch?
> >> It's complicated.  A request with bd->last == false _will_ always be
> >> followed by a request with bd->last == true in the same batch.  However,
> >> due to e.g. errors it may be possible that the last request is not sent.
> >>  In that case, the block layer sends commit_rqs, as documented in the
> >> comment above, to flush the requests that have been sent already.
> >>
> >> So, a driver that obeys bd->last (or SCMD_LAST) but does not implement
> >> commit_rqs is bound to have bugs, which is why this patch was not split
> >> further.
> >>
> >> Makes sense?
> > 
> > Hannes, can you provide your Reviewed-by?
> > 
> Well ... since you asked for it:
> 
> Where is the 'commit_rqs' callback actually used?
> I seem to be going blind, but I can't find it; should be somewhere in
> the first patch, no?
> As per description:
> 
>  * The commit_rqs function is used to trigger a hardware
>  * doorbell after some requests have been queued with
>  * queuecommand, when an error is encountered before sending
>  * the request with SCMD_LAST set.
> 
> So it should be somewhere in the error path, probably scsi_error or
> something. But I don't seem to be able to find it ...

The block layer calls scsi_mq_ops->commit_rqs() from
blk_mq_dispatch_rq_list() and blk_mq_try_issue_list_directly().

Stefan