Re: Thinkpad E595 system deadlock on resume from S3

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue, Oct 3, 2023 at 1:02 PM Petr Tesařík <petr@xxxxxxxxxxx> wrote:
>
> On Tue, 3 Oct 2023 12:15:10 +0200
> "Rafael J. Wysocki" <rafael@xxxxxxxxxx> wrote:
>
> > On Tue, Oct 3, 2023 at 11:31 AM Petr Tesařík <petr@xxxxxxxxxxx> wrote:
> > >
> > > Hi again (adding more recipients),
> > >
> > > On Sat, 30 Sep 2023 12:20:54 +0200
> > > Petr Tesařík <petr@xxxxxxxxxxx> wrote:
> > >
> > > > Hi all,
> > > >
> > > > this time no patch (yet). In short, my Thinkpad running v6.6-rc3 fails
> > > > to resume from S3. It also fails the same way with Tumbleweed v6.5
> > > > kernel. I was able to capture a crash dump of the v6.5 kernel, and
> > > > here's my analysis:
> > > >
> > > > The system never gets to waking up my SATA SSD disk:
> > > >
> > > > [0:0:0:0]    disk    ATA      KINGSTON SEDC600 H5.1  /dev/sda
> > > >
> > > > There is a pending resume work for kworker/u32:12 (PID 11032), but this
> > > > worker is stuck in 'D' state:
> > > >
> > > > >>> prog.stack_trace(11032)
> > > > #0  context_switch (../kernel/sched/core.c:5381:2)
> > > > #1  __schedule (../kernel/sched/core.c:6710:8)
> > > > #2  schedule (../kernel/sched/core.c:6786:3)
> > > > #3  schedule_preempt_disabled (../kernel/sched/core.c:6845:2)
> > > > #4  __mutex_lock_common (../kernel/locking/mutex.c:679:3)
> > > > #5  __mutex_lock (../kernel/locking/mutex.c:747:9)
> > > > #6  acpi_device_hotplug (../drivers/acpi/scan.c:382:2)
> > > > #7  acpi_hotplug_work_fn (../drivers/acpi/osl.c:1162:2)
> > > > #8  process_one_work (../kernel/workqueue.c:2600:2)
> > > > #9  worker_thread (../kernel/workqueue.c:2751:4)
> > > > #10 kthread (../kernel/kthread.c:389:9)
> > > > #11 ret_from_fork (../arch/x86/kernel/process.c:145:3)
> > > > #12 ret_from_fork_asm+0x1b/0x20 (../arch/x86/entry/entry_64.S:304)
> > > >
> > > > acpi_device_hotplug() tries to acquire acpi_scan_lock, which is held by
> > > > systemd-sleep (PID 11002). This task is also in 'D' state:
> > > >
> > > > >>> prog.stack_trace(11002)
> > > > #0  context_switch (../kernel/sched/core.c:5381:2)
> > > > #1  __schedule (../kernel/sched/core.c:6710:8)
> > > > #2  schedule (../kernel/sched/core.c:6786:3)
> > > > #3  schedule_preempt_disabled (../kernel/sched/core.c:6845:2)
> > > > #4  __mutex_lock_common (../kernel/locking/mutex.c:679:3)
> > > > #5  __mutex_lock (../kernel/locking/mutex.c:747:9)
> > > > #6  device_lock (../include/linux/device.h:958:2)
> > > > #7  device_complete (../drivers/base/power/main.c:1063:2)
> > > > #8  dpm_complete (../drivers/base/power/main.c:1121:3)
> > > > #9  suspend_devices_and_enter (../kernel/power/suspend.c:516:2)
> > >
> > > I believe the issue must be somewhere here. The whole suspend and
> > > resume logic in suspend_devices_and_enter() is framed by
> > > platform_suspend_begin() and platform_resume_end().
> > >
> > > My system is an ACPI system, so suspend_ops contains:
> > >
> > >         .begin = acpi_suspend_begin,
> > >         .end = acpi_pm_end,
> > >
> > > Now, acpi_suspend_begin() acquires acpi_scan_lock through
> > > acpi_pm_start(), and the lock is not released until acpi_pm_end().
> > > Since dpm_complete() waits for the completion of a work that tries to
> > > acquire acpi_scan_lock, the system will deadlock.
> >
> > So holding acpi_scan_lock across suspend-resume is basically to
> > prevent the hotplug from taking place then IIRC.
> >
> > > AFAICS either:
> > >
> > > a. the ACPI lock cannot be held while dpm_complete() runs, or
> > > b. ata_scsi_dev_rescan() must not be scheduled before the system is
> > > resumed, or
> > > c. acpi_device_hotplug() must be implemented without taking dev->mutex.
> > >
> > > My gut feeling is that b. is the right answer.
> >
> > It's been a while since I looked at that code last time, but then it
> > has not changed for quite some time too.
> >
> > It looks like the acpi_device_hotplug() path attempts to acquire
> > acpi_scan_lock() while holding dev->mutex which is kind of silly.  I
> > need to check that, though.
>
> Thanks for your willingness. Well, it's not quite what you describe. If
> it was a simple ABBA deadlock, then it would be reported by lockdep.
> No, it's more complicated:
>
> 1. suspend_devices_and_enter() holds acpi_scan_lock,
> 2. an ACPI hotplug work runs, but acpi_device_hotplug() goes to sleep
>    when it gets to acquiring acpi_scan_lock,
> 3. ata_scsi_dev_rescan() submits a SCSI command and waits for its
>    completion while holding dev->mutex,
> 4. the SCSI completion work happens to be put on the same workqueue as
>    the ACPI hotplug work in step 2,
>    ^^^--- THIS is how the two events are serialized!

Which is unexpected.

And quite honestly I'm not sure how this can happen, because
acpi_hotplug_schedule() uses a dedicated workqueue and it is called
from (a) the "eject" sysfs attribute (which cannot happen while system
suspend-resume is in progress) and (b) acpi_bus_notify() which has
nothing to do with SCSI.

Maybe the workqueue used for the SCSI completion is freezable?

> 5. suspend_devices_and_enter() calls dpm_complete() ->
>    device_complete() -> device_lock() and cannot acquire dev->mutex.





[Index of Archives]     [Linux IBM ACPI]     [Linux Power Management]     [Linux Kernel]     [Linux Laptop]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]
  Powered by Linux