Re: Fwd: Waking up from resume locks up on sr device

[Rafael J. Wysocki]

On Wed, Jun 14, 2023 at 4:26 PM Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On Wed, Jun 14, 2023 at 04:35:50PM +0900, Damien Le Moal wrote:
> > On 6/14/23 15:57, Hannes Reinecke wrote:
> > > On 6/14/23 06:49, Damien Le Moal wrote:
> > >> On 6/11/23 18:05, Joe Breuer wrote:
> > >>> I'm the reporter of this issue.
> > >>>
> > >>> I just tried this patch against 6.3.4, and it completely fixes my
> > >>> suspend/resume issue.
> > >>>
> > >>> The optical drive stays usable after resume, even suspending/resuming
> > >>> during playback of CDDA content works flawlessly and playback resumes
> > >>> seamlessly after system resume.
> > >>>
> > >>> So, from my perspective: Good one!
> > >>
> > >> In place of Bart's fix, could you please try this patch ?
> > >>
> > >> diff --git a/drivers/ata/libata-eh.c b/drivers/ata/libata-eh.c
> > >> index b80e68000dd3..a81eb4f882ab 100644
> > >> --- a/drivers/ata/libata-eh.c
> > >> +++ b/drivers/ata/libata-eh.c
> > >> @@ -4006,9 +4006,32 @@ static void ata_eh_handle_port_resume(struct
> > >> ata_port *ap)
> > >>          /* tell ACPI that we're resuming */
> > >>          ata_acpi_on_resume(ap);
> > >>
> > >> -       /* update the flags */
> > >>          spin_lock_irqsave(ap->lock, flags);
> > >> +
> > >> +       /* Update the flags */
> > >>          ap->pflags &= ~(ATA_PFLAG_PM_PENDING | ATA_PFLAG_SUSPENDED);
> > >> +
> > >> +       /*
> > >> +        * Resuming the port will trigger a rescan of the ATA device(s)
> > >> +        * connected to it. Before scheduling the rescan, make sure that
> > >> +        * the associated scsi device(s) are fully resumed as well.
> > >> +        */
> > >> +       ata_for_each_link(link, ap, HOST_FIRST) {
> > >> +               ata_for_each_dev(dev, link, ENABLED) {
> > >> +                       struct scsi_device *sdev = dev->sdev;
> > >> +
> > >> +                       if (!sdev)
> > >> +                               continue;
> > >> +                       if (scsi_device_get(sdev))
> > >> +                               continue;
> > >> +
> > >> +                       spin_unlock_irqrestore(ap->lock, flags);
> > >> +                       device_pm_wait_for_dev(&ap->tdev,
> > >> +                                              &sdev->sdev_gendev);
> > >> +                       scsi_device_put(sdev);
> > >> +                       spin_lock_irqsave(ap->lock, flags);
> > >> +               }
> > >> +       }
> > >>          spin_unlock_irqrestore(ap->lock, flags);
> > >>   }
> > >>   #endif /* CONFIG_PM */
> > >>
> > >> Thanks !
> > >>
> > > Well; not sure if that'll work out.
> > > The whole reason why we initial a rescan is that we need to check if the
> > > ports are still connected, and whether the devices react.
> > > So we can't iterate the ports here as this is the very thing which gets
> > > checked during EH.
> >
> > Hmmm... Right. So we need to move that loop into ata_scsi_dev_rescan(),
> > which itself already loops over the port devices anyway.
> >
> > > We really should claim resume to be finished as soon as we can talk with
> > > the HBA, and kick off EH asynchronously to let it finish the job after
> > > resume has completed.
> >
> > That is what's done already:
> >
> > static int ata_port_pm_resume(struct device *dev)
> > {
> >       ata_port_resume_async(to_ata_port(dev), PMSG_RESUME);
> >       pm_runtime_disable(dev);
> >       pm_runtime_set_active(dev);
> >       pm_runtime_enable(dev);
> >       return 0;
> > }
> >
> > EH is kicked by ata_port_resume_async() -> ata_port_request_pm() and it
> > is async. There is no synchronization in EH with the PM side though. We
> > probably should have EH check that the port resume is done first, which
> > can be done in ata_eh_handle_port_resume() since that is the first thing
> > done when entering EH.
> >
> > The problem remains though that we *must* wait for the scsi device
> > resume to be done before calling scsi_rescan_device(), which is done
> > asynchronously from EH, as a different work. So that one needs to wait
> > for the scsi side resume to be done.
> >
> > I also thought of trigerring the rescan from the scsi side, but since
> > the resume may be asynchronous, we could endup trigerring it with the
> > ata side not yet resumed... That would only turn the problem around
> > instead of solving it.
>
> The order in which devices get resumed isn't arbitrary.  If the system
> is set up not to use async suspends/resumes then the order is always the
> same as the order in which the devices were originally registered (for
> resume, that is -- suspend obviously takes place in the reverse order).
>
> So if you're trying to perform an action that requires two devices to be
> active, you must not do it in the resume handler for the device that was
> registered first.  I don't know how the ATA and SCSI pieces interact
> here, but regardless, this is a pretty strict requirement.
>
> It should be okay to perform the action in the resume handler for the
> device that was registered second.  But if the two devices aren't in an
> ancestor-descendant relationship then you also have to call
> device_pm_wait_for_dev() (or use device links as Rafael mentioned) to
> handle the async case properly.

Note that the bare device_pm_wait_for_dev() is a bit risky though,
because in the sync case it will deadlock if dpm_list is not ordered
properly.

One of the things taken care of by device_link_add() is to ensure that
the ordering of dpm_list will reflect the dependency represented by
the given new device link.

> > Or... Why the heck scsi_rescan_device() is calling device_lock() ? This
> > is the only place in scsi code I can see that takes this lock. I suspect
> > this is to serialize either rescans, or serialize with resume, or both.
> > For serializing rescans, we can use another lock. For serializing with
> > PM, we should wait for PM transitions...
> > Something is not right here.
>
> Here's what commit e27829dc92e5 ("scsi: serialize ->rescan against
> ->remove", written by Christoph Hellwig) says:
>
>     Lock the device embedded in the scsi_device to protect against
>     concurrent calls to ->remove.
>
> That's the commit which added the device_lock() call.
>
> Alan Stern