On Tue, Oct 3, 2023 at 5:18 PM Petr Tesařík <petr@xxxxxxxxxxx> wrote:
>
> On Tue, 3 Oct 2023 14:57:46 +0200
> "Rafael J. Wysocki" <rafael@xxxxxxxxxx> wrote:
>
> > On Tue, Oct 3, 2023 at 2:51 PM Petr Tesařík <petr@xxxxxxxxxxx> wrote:
> > >
> > > On Tue, 3 Oct 2023 14:48:13 +0200
> > > "Rafael J. Wysocki" <rafael@xxxxxxxxxx> wrote:
> > >
> > > > On Tue, Oct 3, 2023 at 2:40 PM Petr Tesařík <petr@xxxxxxxxxxx> wrote:
> > > > >
> > > > > On Tue, 3 Oct 2023 14:34:56 +0200
> > > > > "Rafael J. Wysocki" <rafael@xxxxxxxxxx> wrote:
> > > > >
> > > > > > 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.
> > > > >
> > > > > Oh, you're right, and I was too quick. They cannot be on the same
> > > > > queue...
> > > > >
> > > > > > Maybe the workqueue used for the SCSI completion is freezable?
> > > > >
> > > > > Yes, that's it:
> > > > >
> > > > > *(struct workqueue_struct *)0xffff97d240b2fe00 = {
> > > > > /* ... */
> > > > > .flags = (unsigned int)4,
> > > > > /* WQ_FREEZABLE = 1 << 2 */
> > > > >
> > > > > Good. But if this workqueue is frozen, the system still cannot make
> > > > > progress.
> > > >
> > > > The problem seems to be that dev->mutex is held while the work item
> > > > goes to a freezable workqueue and is waited for, which is an almost
> > > > guaranteed deadlock scenario.
> > >
> > > Ah. Thanks for explanation and direction! I'm going to dive into the
> > > block layer and/or SCSI code and bug other people with my findings.
> >
> > Please feel free to CC me on that in case I can help.
>
> And here I am again... The frozen workqueue is in fact pm_wq, and the
> work item that is waited for is pm_runtime_work. The block layer calls
> pm_request_resume() on the device to resume the queue.
If it called pm_runtime_resume() instead, this might work.
> I bet the queue should not be resumed this early. In fact, it seems
> that this is somewhat known to the ATA developers, because
> ata_scsi_dev_rescan() contains this beautiful comment and code:
>
> /*
> * If the rescan work was scheduled because of a resume
> * event, the port is already fully resumed, but the
> * SCSI device may not yet be fully resumed. In such
> * case, executing scsi_rescan_device() may cause a
> * deadlock with the PM code on device_lock(). Prevent
> * this by giving up and retrying rescan after a short
> * delay.
> */
> delay_rescan = sdev->sdev_gendev.power.is_suspended;
> if (delay_rescan) {
> scsi_device_put(sdev);
> break;
> }
>
> It just doesn't seem to work as expected, at least not in my case.
Well, calling pm_request_resume() and waiting for the resume to
actually happen is problematic regardless. It is just better to call
pm_runtime_resume() to synchronously resume the device instead.
