On Mon, Jun 12, 2023 at 4:33 PM Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote: > > On Mon, Jun 12, 2023 at 04:47:09PM +0900, Damien Le Moal wrote: > > On 6/12/23 16:36, Kai-Heng Feng wrote: > > > On Mon, Jun 12, 2023 at 3:22 PM Damien Le Moal <dlemoal@xxxxxxxxxx> wrote: > > >> > > >> On 6/12/23 15:09, Hannes Reinecke wrote: > > >>> On 6/12/23 05:09, Damien Le Moal wrote: > > >>>> On 6/11/23 00:03, Bart Van Assche wrote: > > >>>>> On 6/10/23 06:27, Bagas Sanjaya wrote: > > >>>>>> On 6/10/23 15:55, Pavel Machek wrote: > > >>>>>>>>> #regzbot introduced: v5.0..v6.4-rc5 https://bugzilla.kernel.org/show_bug.cgi?id=217530 > > >>>>>>>>> #regzbot title: Waking up from resume locks up on SCSI CD/DVD drive > > >>>>>>>>> > > >>>>>>>> The reporter had found the culprit (via bisection), so: > > >>>>>>>> > > >>>>>>>> #regzbot introduced: a19a93e4c6a98c > > >>>>>>> Maybe cc the authors of that commit? > > >>>>>> > > >>>>>> Ah! I forgot to do that! Thanks anyway. > > >>>>> > > >>>>> Hi Damien, > > >>>>> > > >>>>> Why does the ATA code call scsi_rescan_device() before system resume has > > >>>>> finished? Would ATA devices still work with the patch below applied? > > >>>> > > >>>> I do not know the PM code well at all, need to dig into it. But your patch > > >>>> worries me as it seems it would prevent rescan of the device on a resume, which > > >>>> can be an issue if the device has changed. > > >>>> > > >>>> I am not yet 100% clear on the root cause for this, but I think it comes from > > >>>> the fact that ata_port_pm_resume() runs before the sci device resume is done, so > > >>>> with scsi_dev->power.is_suspended still true. And ata_port_pm_resume() calls > > >>>> ata_port_resume_async() which triggers EH (which will do reset + rescan) > > >>>> asynchronously. So it looks like we have scsi device resume and libata EH for > > >>>> rescan fighting each others for the scan mutex and device lock, leading to deadlock. > > >>>> > > >>>> Trying to recreate this issue now to confirm and debug further. But I suspect > > >>>> the solution to this may be best implemented in libata, not in scsi. > > >>>> This looks definitely related to this thread: > > >>>> > > >>>> https://lore.kernel.org/linux-scsi/7b553268-69d3-913a-f9de-28f8d45bdb1e@xxxxxxx/ > > >>>> > > >>>> Similaraly to your comment on that thread, having to look at > > >>>> dev->power.is_suspended is not ideal I think. What we need is to have ata and > > >>>> scsi pm resume be synchronized, but I am not yet 100% clear on the scsi layer side. > > >>>> > > >>> Which is my feeling, too. > > >>> libata runs rescan as part of the device discovery, so really it will > > >>> run after resume. And consequently resume really cannot wait for rescan > > >>> to finish. > > >>> > > >>> What I would be looking at is to decouple resume from libata device > > >>> rescan, and have resume to complete before libata EH runs. > > >> > > >> That is the case now, for the ata port at least, even though that is not super > > >> explicit, and not reliable. See ata_port_pm_resume(): I think that the call to > > >> EH in ata_port_pm_resume() -> ata_port_resume_async() -> ata_port_request_pm() > > >> -> ata_port_schedule_eh() should instead use a sync resume, leading to a sync EH > > >> call. > > >> > > >> That EH execution essentially does ata_eh_handle_port_resume(), which calls into > > >> the adapter resume operation. That in itself does not do much beside some > > >> registers accesses to wakeup the port. There should be no issues doing that > > >> synchronously. > > >> > > >> The problem is that after that is done, ata EH calls ata_std_error_handler() -> > > >> ata_do_eh() -> ata_eh_recover() -> ata_eh_revalidate_and_attach() -> > > >> schedule_work(&(ap->scsi_rescan_task)). And the rescan work calls > > >> scsi_rescan_device() (yet in another context than EH) which causes the problem > > >> when the scsi disk device has not been resumed yet (dev->power_is_suspended > > >> still true). > > >> > > >> So it really looks like the solution should be to have ata_scsi_dev_rescan() > > >> wait for the scsi device to resume first, but not sure how to do that with the > > >> pm API. Digging... > > > > > > Probably use dpm_wait_for_children()? Right now it's an internal PM API. > > > > But I am not sure if there is a relationship between ata_device and its > > scsi_device (dev->sdev)... Need to clarify that. > > > > > > Rafael, > > > What do you think? > > Look into the device_pm_wait_for_dev() API. It's the appropriate thing > to use when you want to wait for another device to complete a system PM > transition. (However, it's not appropriate for runtime PM.) > > Of course, if there is a parent-child relationship between two devices > then waiting is never necessary. The PM core guarantees that a parent > will always be at full power when a child changes its power state, > unless pm_suspend_ignore_children() has been called for the parent > device. There are also device links that go beyond the parent-child handling. Please see Documentation/driver-api/device_link.rst and the kerneldoc description of device_link_add() for more information.
