Re: [PATCH 01/11] blk-mq: Add blk_mq_init_queue_ops()

[Hannes Reinecke]

On 3/22/22 13:30, John Garry wrote:
> On 22/03/2022 12:16, Hannes Reinecke wrote:
> > On 3/22/22 12:33, John Garry wrote:
> > > On 22/03/2022 11:18, Christoph Hellwig wrote:
> > > > On Tue, Mar 22, 2022 at 06:39:35PM +0800, John Garry wrote:
> > > > > Add an API to allocate a request queue which accepts a custom set of
> > > > > blk_mq_ops for that request queue.
> > > > >
> > > > > The reason which we may want custom ops is for queuing requests 
> > > > > which we
> > > > > don't want to go through the normal queuing path.
> > > >
> > > > Eww.  I really do not think we should do separate ops per queue, as 
> > > > that
> > > > is going to get us into a deep mess eventually.
> > >
> > > Yeah... so far (here) it works out quite nicely, as we don't need to 
> > > change the SCSI blk mq ops nor allocate a scsi_device - everything is 
> > > just separate.
> > >
> > > The other method mentioned previously was to add the request 
> > > "reserved" flag and add new paths in scsi_queue_rq() et al to handle 
> > > this, but that gets messy.
> > >
> > > Any other ideas ...?
> >
> > As outlined in the other mail, I think might be useful is to have a 
> > _third_ type of requests (in addition to the normal and the reserved 
> > ones).
> > That one would be allocated from the normal I/O pool (and hence could 
> > fail if the pool is exhausted), but would be able to carry a different 
> > payload (type) than the normal requests.
>
> As mentioned in the cover letter response, it just seems best to keep 
> the normal scsi_cmnd payload but have other means to add on the internal 
> command data, like using host_scribble or scsi_cmnd priv data.
Well; I found that most drivers I had been looking at the scsi command 
payload isn't used at all; the drivers primarily cared about the 
(driver-provided) payload, and were completely ignoring the scsi command 
payload.

Similar for ATA/libsas: you basically never issue real scsi commands, 
but either 'raw' ATA requests or SCSI TMFs. None of which are scsi 
commands, so providing them is a bit of a waste.

(And causes irritations, too, as a scsi command requires associated 
pointers like ->device etc to be set up. Which makes it tricky to use 
for the initial device setup.)

> > And we could have a separate queue_rq for these requests, as we can 
> > differentiate them in the block layer.
>
> I don't know, let me think about it. Maybe we could add an "internal" 
> blk flag, which uses a separate "internal" queue_rq callback.
Yeah, that's what I had in mind.

Cheers,

Hannes
--
Dr. Hannes Reinecke		           Kernel Storage Architect
hare@xxxxxxx			                  +49 911 74053 688
SUSE Software Solutions Germany GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), GF: Felix Imendörffer