Re: [PATCH RFC v2 02/24] scsi: allocate separate queue for reserved commands

[Hannes Reinecke]

On 4/7/20 4:35 PM, John Garry wrote:
> On 07/04/2020 15:00, Hannes Reinecke wrote:
> > On 4/7/20 1:54 PM, John Garry wrote:
> > > On 06/04/2020 10:05, Hannes Reinecke wrote:
[ .. ]
> > > > This would be okay if 'this_id' would have been defined by the driver;
> > > > sadly, most drivers which are affected here do set 'this_id' to -1.
> > > > So we wouldn't have a nice target ID to allocate the device from, let
> > > > alone the problem that we would have to emulate a complete scsi device
> > > > with all required minimal command support etc.
> > > > And I'm not quite sure how well that would play with the exising SCSI
> > > > host template; the device we'll be allocating would have basically
> > > > nothing in common with the 'normal' SCSI devices.
> > > >
> > > > What we could do, though, is to try it the other way round:
> > > > Lift the request queue from scsi_get_host_dev() into the scsi host
> > > > itself, so that scsi_get_host_dev() can use that queue, but we also
> > > > would be able to use it without a SCSI device attached.
> > >
> > > wouldn't that limit 1x scsi device per host, not that I know if any 
> > > more would ever be required? But it does still seem better to use the 
> > > request queue in the scsi device.
> > My concern is this:
> >
> > struct scsi_device *scsi_get_host_dev(struct Scsi_Host *shost)
> > {
> > ÂÂÂÂÂ[ .. ]
> > ÂÂÂÂÂstarget = scsi_alloc_target(&shost->shost_gendev, 0, 
> > shost->this_id);
> > ÂÂÂÂÂ[ .. ]
> >
> > and we have typically:
> >
> > drivers/scsi/hisi_sas/hisi_sas_v3_hw.c: .this_idÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ = -1,
> >
> > It's _very_ uncommon to have a negative number as the SCSI target 
> > device; in fact, it _is_ an unsigned int already.
>
> FWIW, the only other driver (gdth) which I see uses this API has this_id 
> = -1 in the scsi host template.
>
> > But alright, I'll give it a go; let's see what I'll end up with.
>
> note: If we want a fixed scsi_device per host, calling 
> scsi_mq_setup_tags() -> scsi_get_host_dev() will fail as shost state is 
> not running. Maybe we need to juggle some things there to provide a 
> generic solution.
It might even get worse, as during device setup things like 
'slave_alloc' etc is getting called, which has a fair chance of getting 
confused for non-existing devices.
Cf qla2xxx:qla2xx_slave_alloc() is calling starget_to_rport(), which 
will get us a nice oops when accessing a target which is _not_ the child 
of a fc remote port.
And this is why I'm not utterly keen on this approach; auditing all 
these callbacks is _not_ fun.

Cheers,

Hannes
--
Dr. Hannes Reinecke            Teamlead Storage & Networking
hare@xxxxxxx                               +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 NÃrnberg
HRB 36809 (AG NÃrnberg), GeschÃftsfÃhrer: Felix ImendÃrffer