On Wed, Jun 14, 2023 at 10:48 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Tue, Jun 13, 2023 at 05:25:07PM +0200, Rafael J. Wysocki wrote:
> > From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> >
> > The addition of might_sleep() to down_timeout() caused the latter to
> > enable interrupts unconditionally in some cases, which in turn broke
> > the ACPI S3 wakeup path in acpi_suspend_enter(), where down_timeout()
> > is called by acpi_disable_all_gpes() via acpi_ut_acquire_mutex().
> >
> > Namely, if CONFIG_DEBUG_ATOMIC_SLEEP is set, might_sleep() causes
> > might_resched() to be used and if CONFIG_PREEMPT_VOLUNTARY is set,
> > this triggers __cond_resched() which may call preempt_schedule_common(),
> > so __schedule() gets invoked and it ends up with enabled interrupts (in
> > the prev == next case).
>
> Urgh, so that code was relying on the lack of contention to not trigger
> the schedule path -- with the added might_sleep() it triggers a
> preemption point.
Right.
> > Now, enabling interrupts early in the S3 wakeup path causes the kernel
> > to crash.
> >
> > Address this by modifying acpi_suspend_enter() to disable GPEs without
> > attempting to acquire the sleeping lock which is not needed in that code
> > path anyway.
> >
> > Fixes: 99409b935c9a locking/semaphore: Add might_sleep() to down_*() family
>
> $ git show -s --pretty='format:%h ("%s")' 99409b935c9a
> 99409b935c9a ("locking/semaphore: Add might_sleep() to down_*() family")
Right, thanks!
> > Reported-by: Srinivas Pandruvada <srinivas.pandruvada@xxxxxxxxxxxxxxx>
> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
>
> Acked-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
Thanks!
> > ---
> > drivers/acpi/acpica/achware.h | 2 --
> > drivers/acpi/sleep.c | 16 ++++++++++++----
> > include/acpi/acpixf.h | 1 +
> > 3 files changed, 13 insertions(+), 6 deletions(-)
> >
> > Index: linux-pm/drivers/acpi/acpica/achware.h
> > ===================================================================
> > --- linux-pm.orig/drivers/acpi/acpica/achware.h
> > +++ linux-pm/drivers/acpi/acpica/achware.h
> > @@ -101,8 +101,6 @@ acpi_status
> > acpi_hw_get_gpe_status(struct acpi_gpe_event_info *gpe_event_info,
> > acpi_event_status *event_status);
> >
> > -acpi_status acpi_hw_disable_all_gpes(void);
> > -
> > acpi_status acpi_hw_enable_all_runtime_gpes(void);
> >
> > acpi_status acpi_hw_enable_all_wakeup_gpes(void);
> > Index: linux-pm/include/acpi/acpixf.h
> > ===================================================================
> > --- linux-pm.orig/include/acpi/acpixf.h
> > +++ linux-pm/include/acpi/acpixf.h
> > @@ -761,6 +761,7 @@ ACPI_HW_DEPENDENT_RETURN_STATUS(acpi_sta
> > acpi_event_status
> > *event_status))
> > ACPI_HW_DEPENDENT_RETURN_UINT32(u32 acpi_dispatch_gpe(acpi_handle gpe_device, u32 gpe_number))
> > +ACPI_HW_DEPENDENT_RETURN_STATUS(acpi_status acpi_hw_disable_all_gpes(void))
> > ACPI_HW_DEPENDENT_RETURN_STATUS(acpi_status acpi_disable_all_gpes(void))
> > ACPI_HW_DEPENDENT_RETURN_STATUS(acpi_status acpi_enable_all_runtime_gpes(void))
> > ACPI_HW_DEPENDENT_RETURN_STATUS(acpi_status acpi_enable_all_wakeup_gpes(void))
> > Index: linux-pm/drivers/acpi/sleep.c
> > ===================================================================
> > --- linux-pm.orig/drivers/acpi/sleep.c
> > +++ linux-pm/drivers/acpi/sleep.c
> > @@ -636,11 +636,19 @@ static int acpi_suspend_enter(suspend_st
> > }
> >
> > /*
> > - * Disable and clear GPE status before interrupt is enabled. Some GPEs
> > - * (like wakeup GPE) haven't handler, this can avoid such GPE misfire.
> > - * acpi_leave_sleep_state will reenable specific GPEs later
> > + * Disable all GPE and clear their status bits before interrupts are
> > + * enabled. Some GPEs (like wakeup GPEs) have no handlers and this can
> > + * prevent them from producing spurious interrups.
> > + *
> > + * acpi_leave_sleep_state() will reenable specific GPEs later.
> > + *
> > + * Because this code runs on one CPU with disabled interrupts (all of
> > + * the other CPUs are offline at that time), it need not acquire any
> > + * sleeping locks which maybe harmful due to instrumentation even if
> > + * those locks are not contended, so avoid doing that by using a low-
> > + * level library routine here.
>
> I'm not sure I'd call the implicit preemption point 'instrumentation'
> but yeah, fair enough I suppose.
OK, I'll send a v2 with this clarified.
> > */
> > - acpi_disable_all_gpes();
> > + acpi_hw_disable_all_gpes();
> > /* Allow EC transactions to happen. */
> > acpi_ec_unblock_transactions();
