Re: [RFC] ACPI PM during kernel poweroff/reboot

[Furquan Shaikh <furquan@google.com>]

On Wed, Nov 25, 2020 at 10:29 AM Furquan Shaikh <furquan@google.com> wrote:
>
> On Wed, Nov 25, 2020 at 9:51 AM Rafael J. Wysocki <rafael@kernel.org> wrote:
> >
> > On Wed, Nov 25, 2020 at 6:43 PM Furquan Shaikh <furquan@google.com> wrote:
> > >
> > > On Wed, Nov 25, 2020 at 8:39 AM Rafael J. Wysocki <rafael@kernel.org> wrote:
> > > >
> > > > On Thu, Nov 12, 2020 at 8:19 PM Furquan Shaikh <furquan@google.com> wrote:
> > > > >
> > > > > On x86 Chromebooks, we have observed this issue for a long time now -
> > > > > when the system is powered off or rebooted, ACPI PM is not invoked and
> > > > > this results in PowerResource _OFF methods not being invoked for any
> > > > > of the devices. The _OFF methods are invoked correctly in case of
> > > > > suspend-to-idle (S0ix) and suspend-to-memory(S3). However, they do not
> > > > > get invoked when `poweroff` or `reboot` are triggered.
> > > > >
> > > > > One of the differences between suspend, hibernate and shutdown paths
> > > > > in Linux kernel is that the shutdown path does not use the typical
> > > > > device PM phases (prepare, freeze/suspend, poweroff) as used by
> > > > > suspend/hibernate. Instead the shutdown path makes use of
> > > > > .shutdown_pre() and .shutdown() callbacks.
> > > > >
> > > > > If I understand correctly, .shutdown() has been around for a long time
> > > > > and existed even before the PM callbacks were added. Thus,
> > > > > pm->poweroff() and .shutdown() are supposed to be analogous and
> > > > > consistent in the behavior.
> > > >
> > > > Well, not quite.
> > > >
> > > > ->shutdown() is expected to be a lightweight operation also suitable
> > > > for kexec() and similar situations where ->poweroff() may not work.
> > > >
> > > > > This is why runtime PM is disallowed by
> > > > > device_shutdown() before it calls .shutdown() (i.e. to keep behavior
> > > > > consistent for both paths). However, in practice, there are
> > > > > differences in behavior for the pm->poweroff() and .shutdown() paths
> > > > > since the shutdown path does not execute any PM domain operations.
> > > >
> > > > That's correct.
> > > >
> > > > > Because of this difference in behavior, shutdown path never invokes
> > > > > ACPI PM and thus the ACPI PowerResources are not turned off when the
> > > > > system is rebooted or powered off (sleep S5). On Chromebooks, it is
> > > > > critical to run the _OFF methods for poweroff/reboot in order to
> > > > > ensure that the device power off sequencing requirements are met.
> > > > > Currently, these requirements are violated which impact the
> > > > > reliability of devices over the lifetime of the platform.
> > > > >
> > > > > There are a few ways in which this can be addressed:
> > > > >
> > > > > 1. Similar to the case of hibernation, a new
> > > > > PMSG_POWEROFF/PM_EVENT_POWEROFF can be introduced to invoke device
> > > > > power management phases using `dpm_suspend_start(PMSG_POWEROFF)` and
> > > > > `dpm_suspend_end(PMSG_POWEROFF)`. However, as the shutdown path uses
> > > > > the class/bus/driver .shutdown() callbacks, adding dpm phases for
> > > > > poweroff complicates the order of operations. If the dpm phases are
> > > > > run before .shutdown() callbacks, then it will result in the callbacks
> > > > > accessing devices after they are powered off. If the .shutdown()
> > > > > callbacks are run before dpm phases, then the pm->poweroff() calls are
> > > > > made after the device shutdown is done. Since .shutdown() and
> > > > > pm->poweroff() are supposed to be analogous, having both calls in the
> > > > > shutdown path is not only redundant but also results in incorrect
> > > > > behavior.
> > > > >
> > > > > 2. Another option is to update device_shutdown() to make
> > > > > pm_domain.poweroff calls after the class/bus/driver .shutdown() is
> > > > > done. However, this suffers from the same problem as #1 above i.e. it
> > > > > is redundant and creates conflicting order of operations.
> > > > >
> > > > > 3. Third possible solution is to detach the device from the PM domain
> > > > > after it is shutdown. Currently, device drivers perform a detach
> > > > > operation only when the device is removed. However, in case of
> > > > > poweroff/reboot as the device is already shutdown, detaching PM domain
> > > > > will give it the opportunity to ensure that any power resources are
> > > > > correctly turned off before the system shuts down.
> > > >
> > > > 4. Make Chromebooks call something like hibernation_platform_enter()
> > > > on S5 entries (including reboot).
> > >
> > > Actually, Chromebooks do not support S4 and hence CONFIG_HIBERNATION.
> >
> > This doesn't matter.  The ->poweroff callbacks can still be used by
> > them (of course, that part of the current hibernation support code
> > needs to be put under a more general Kconfig option for that, but this
> > is a technical detail).
>
> Ah I see what you are saying. Just to be sure I understand this
> correctly. Is this what you are thinking:
> 1. Extract hibernation_platform_enter() and any other helpers required
> to trigger the PM phases for shutdown into a separate unit controlled
> by a more general Kconfig.
> 2. Add a new Kconfig that enables support for performing PM phases
> during the poweroff/reboot phases.
> 3. Based on this new Kconfig selection, LINUX_REBOOT_CMD_RESTART,
> LINUX_REBOOT_CMD_HALT, LINUX_REBOOT_CMD_POWER_OFF will be updated to
> use the new paths instead of the current lightweight calls.

I am currently exploring this approach to see how the components need
to be organized to make use of hibernation_platform_enter by more than
just the hibernation path. Please let me know if the above summary
doesn't align with your suggestion.

Meanwhile, I have also sent out a formal patch for detaching the PM
domain: https://lore.kernel.org/linux-acpi/20201201213019.1558738-1-furquan@google.com/T/#u
to ensure that this addresses the issue with ACPI PM domain.

I will continue working on the above suggestion as well, but it might
take some time for me to get a good understanding of the current paths
and to cleanly implement the support for PM phases during
poweroff/reboot cases.

Thanks,
Furquan

>
> >
> > > This is done for a number of reasons including security. Hence, I
> > > don't think using hibernation_platform_enter() would be an option.
> >
> > Yes, it is an option.
> >
> > Having "hibernation" in the name need not mean that the given piece of
> > code is really hibernation-specific ...
>
> Sorry, I had misunderstood the suggestion before. I have attempted to
> outline your proposal with some more details above.
>
> >
> > > >
> > > > > Out of these, I think #3 makes the most sense as it does not introduce
> > > > > any conflicting operations. I verified that the following diff results
> > > > > in _OFF methods getting invoked in both poweroff and reboot cases:
> > > >
> > > > This won't work for PCI devices though, only for devices in the ACPI
> > > > PM domain, so it is not sufficient in general.
> > >
> > > That is true. The proposed solution only handles detaching of PM
> > > domains. I understand your point about this not working for any
> > > devices not part of the PM domain. The issues that we have observed in
> > > shutdown/reboot paths have been specific to ACPI power resources
> > > controlling the sequencing to external devices.
> >
> > PCI devices PM can use power resources too.  For instance, this has
> > been quite common for discrete GPUs in laptops IIRC.
>
> Sorry about my naive question: Is the power resource not described
> using ACPI in this case? (I haven't run into a situation with PCI
> devices using non-ACPI power resources, so curious to understand the
> scenario).