Re: [PATCH RFC v3 16/22] ata: libata-scsi: Allocate sdev early in port probe

[Damien Le Moal]

On 10/27/22 17:11, Hannes Reinecke wrote:
> On 10/27/22 03:34, Damien Le Moal wrote:
>> On 10/25/22 19:18, John Garry wrote:
>>> Currently the per-ata device sdev is allocated as part of the scsi
>>> target
>>> scan, which is after the ata port probe.
>>>
>>> However it is useful to have the sdev available in the port probe. As an
>>> example of an advantage, if the request queue is available in the probe
>>> (which it would be if the sdev is available), then it is possible to use
>>> a SCSI cmnd for ATA internal commands. The benefit of this is then we
>>> can
>>> put the ATA qc structure in the SCSI cmnd private data. It will also be
>>> useful if we want to send ATA internal commands as requests.
>>>
>>> Export scsi_target_reap() so that it can be used to put the extra
>>> reference we get when allocating the sdev.
>>>
>>> Signed-off-by: John Garry <john.garry@xxxxxxxxxx>
>>> ---
>>>   drivers/ata/libata-eh.c   |  1 +
>>>   drivers/ata/libata-scsi.c | 23 +++++++++--------------
>>>   drivers/scsi/scsi_scan.c  |  1 +
>>>   3 files changed, 11 insertions(+), 14 deletions(-)
>>>
>>> diff --git a/drivers/ata/libata-eh.c b/drivers/ata/libata-eh.c
>>> index 08e11bc312c2..1ed5b1b64792 100644
>>> --- a/drivers/ata/libata-eh.c
>>> +++ b/drivers/ata/libata-eh.c
>>> @@ -3446,6 +3446,7 @@ static int ata_eh_schedule_probe(struct
>>> ata_device *dev)
>>>         ata_eh_detach_dev(dev);
>>>       ata_dev_init(dev);
>>> +    ata_scsi_setup_sdev(dev);
>>>       ehc->did_probe_mask |= (1 << dev->devno);
>>>       ehc->i.action |= ATA_EH_RESET;
>>>       ehc->saved_xfer_mode[dev->devno] = 0;
>>> diff --git a/drivers/ata/libata-scsi.c b/drivers/ata/libata-scsi.c
>>> index efdba852e363..476e0ef4bd29 100644
>>> --- a/drivers/ata/libata-scsi.c
>>> +++ b/drivers/ata/libata-scsi.c
>>> @@ -1109,7 +1109,12 @@ int ata_scsi_dev_config(struct scsi_device
>>> *sdev, struct ata_device *dev)
>>>       if (dev->flags & ATA_DFLAG_TRUSTED)
>>>           sdev->security_supported = 1;
>>>   -    dev->sdev = sdev;
>>> +    /*
>>> +     * Put extra reference which we get when allocating the starget
>>> +     * initially
>>> +     */
>>> +    scsi_target_reap(scsi_target(sdev));
>>> +
>>>       return 0;
>>>   }
>>>   @@ -4289,26 +4294,16 @@ void ata_scsi_scan_host(struct ata_port
>>> *ap, int sync)
>>>    repeat:
>>>       ata_for_each_link(link, ap, EDGE) {
>>>           ata_for_each_dev(dev, link, ENABLED) {
>>> -            struct scsi_device *sdev;
>>> +            struct Scsi_Host *shost = ap->scsi_host;
>>>               int channel = 0, id = 0;
>>>   -            if (dev->sdev)
>>> -                continue;
>>> -
>>>               if (ata_is_host_link(link))
>>>                   id = dev->devno;
>>>               else
>>>                   channel = link->pmp;
>>>   -            sdev = __scsi_add_device(ap->scsi_host, channel, id, 0,
>>> -                         NULL);
>>> -            if (!IS_ERR(sdev)) {
>>> -                dev->sdev = sdev;
>>> -                ata_scsi_assign_ofnode(dev, ap);
>>
>> Is there something equivalent to what this function does inside
>> scsi_scan_target() ? I had a quick look but did not see anything...
>>
> Typically, the SCSI layer has two ways of scanning.
> One it the old-style serial scanning (originating in the old SCSI
> parallel model):
> The scanning code will blindly scan _all_ devices up to max_luns, and
> attach every device for which the scanning code returns 'OK'.
> The other one is to issue REPORT_LUNS and scan all LUNs returned there.
> 
> Mapped to libata we would need to figure out a stable SCSI enumeration,
> given that we have PMP and slave devices to content with.
> To my knowledge we have ATA ports, each can have either a 'master' and
> 'slave' device, _or_ it be a PMP port when it can support up to 16
> devices, right?

yes

> Point being, master/slave and PMP are exclusive, right?

Never heard of pmp with ide cable :)

> So we can make the master as LUN 0, and the slave as LUN 1.

Yes, but isn't that a little wrong ? That would assume that the ata port
is the device and the ata devices the luns of that device. But beside
the "link busy" stuff that needs to be dealt with, master and slave are
independent devices, unlike LUNs. No ?

> And for PMP we can use each PMP address as LUN <pmp> + 1, and keeping
> the actual device as LUN 0.
> 
> I think we can figure out whether it's a master/slave device setup
> upfront, so we should be able to set max_luns to '2' for these devices.
> For PMP-capable (or devices which _might_ be PMP capable), we could
> emulate the REPORT LUNS command, mapping on the PMP mechanism to figure
> out which devices are connected.
> 
> Would that work?
> 
> Cheers,
> 
> Hannes

-- 
Damien Le Moal
Western Digital Research