[PATCH] PM / runtime: Avoid false-positive warnings from might_sleep_if()
From: Rafael J. Wysocki
Date: Fri Feb 03 2017 - 19:03:57 EST
On Thursday, February 02, 2017 02:34:42 PM Sedat Dilek wrote:
> On Wed, Feb 1, 2017 at 1:22 AM, Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
> > On Mon, Jan 30, 2017 at 11:44 PM, Rafael J. Wysocki
> > <rafael.j.wysocki@xxxxxxxxx> wrote:
> >> On 1/24/2017 2:33 AM, Sedat Dilek wrote:
> >>>
> >>> On Fri, Dec 30, 2016 at 3:02 PM, Rafael J. Wysocki <rafael@xxxxxxxxxx>
> >>> wrote:
> >>>>
> >>>> On Fri, Dec 30, 2016 at 12:40 PM, Sedat Dilek <sedat.dilek@xxxxxxxxx>
> >>>> wrote:
> >>>>>
> >>>>> Hi,
> >>>>>
> >>>>> I have already reported this issue in [1].
> >>>>> One of the issue was solved.
> >>>>> Unfortunately, it looks like there is still a different problem here
> >>>>> (Ubuntu/precise AMD64).
> >>>>>
> >>>>> I tried v4.10-rc1 and latest Linus tree up to...
> >>>>>
> >>>>> commit 98473f9f3f9bd404873cd1178c8be7d6d619f0d1
> >>>>> "mm/filemap: fix parameters to test_bit()"
> >>>>>
> >>>>> Here we go...
> >>>>>
> >>>>> [ 29.636047] BUG: sleeping function called from invalid context at
> >>>>> drivers/base/power/runtime.c:1032
> >>>>> [ 29.636055] in_atomic(): 1, irqs_disabled(): 0, pid: 1500, name: Xorg
> >>>>> [ 29.636058] 1 lock held by Xorg/1500:
> >>>>> [ 29.636060] #0: (&dev->struct_mutex){+.+.+.}, at:
> >>>>> [<ffffffffa0680c13>] i915_mutex_lock_interruptible+0x43/0x140 [i915]
> >>>>> [ 29.636107] CPU: 0 PID: 1500 Comm: Xorg Not tainted
> >>>>> 4.10.0-rc1-6-iniza-amd64 #1
> >>>>> [ 29.636109] Hardware name: SAMSUNG ELECTRONICS CO., LTD.
> >>>>> 530U3BI/530U4BI/530U4BH/530U3BI/530U4BI/530U4BH, BIOS 13XK 03/28/2013
> >>>>> [ 29.636111] Call Trace:
> >>>>> [ 29.636120] dump_stack+0x85/0xc2
> >>>>> [ 29.636124] ___might_sleep+0x196/0x260
> >>>>> [ 29.636127] __might_sleep+0x53/0xb0
> >>>>> [ 29.636131] __pm_runtime_resume+0x7a/0x90
> >>>>> [ 29.636159] intel_runtime_pm_get+0x25/0x90 [i915]
> >>>>> [ 29.636189] aliasing_gtt_bind_vma+0xaa/0xf0 [i915]
> >>>>> [ 29.636220] i915_vma_bind+0xaf/0x1e0 [i915]
> >>>>> [ 29.636248] i915_gem_execbuffer_relocate_entry+0x513/0x6f0 [i915]
> >>>>> [ 29.636272] i915_gem_execbuffer_relocate_vma.isra.34+0x188/0x250
> >>>>> [i915]
> >>>>> [ 29.636275] ? trace_hardirqs_on+0xd/0x10
> >>>>> [ 29.636294] ? i915_gem_execbuffer_reserve_vma.isra.31+0x152/0x1f0
> >>>>> [i915]
> >>>>> [ 29.636316] ? i915_gem_execbuffer_reserve.isra.32+0x372/0x3a0 [i915]
> >>>>> [ 29.636342] i915_gem_do_execbuffer.isra.38+0xa70/0x1a40 [i915]
> >>>>> [ 29.636347] ? __might_fault+0x4e/0xb0
> >>>>> [ 29.636373] i915_gem_execbuffer2+0xc5/0x260 [i915]
> >>>>> [ 29.636376] ? __might_fault+0x4e/0xb0
> >>>>> [ 29.636395] drm_ioctl+0x206/0x450 [drm]
> >>>>> [ 29.636420] ? i915_gem_execbuffer+0x340/0x340 [i915]
> >>>>> [ 29.636425] ? __fget+0x5/0x200
> >>>>> [ 29.636429] do_vfs_ioctl+0x91/0x6f0
> >>>>> [ 29.636431] ? __fget+0x111/0x200
> >>>>> [ 29.636433] ? __fget+0x5/0x200
> >>>>> [ 29.636436] SyS_ioctl+0x79/0x90
> >>>>> [ 29.636441] entry_SYSCALL_64_fastpath+0x23/0xc6
> >>>>>
> >>>>> On suspend/resume I see the same call trace.
> >>>>> [2] points to the "BUG" line.
> >>>>
> >>>> Well, this appears to be an i915 issue, but not a serious one.
> >>>>
> >>>> Clearly, a function that may sleep (pm_runtime_get_sync() in
> >>>> intel_runtime_pm_get()) is called with disabled interrupts. If I
> >>>> understand the code correctly, though, it actually is not going to
> >>>> sleep in this particular case, because pm_runtime_get_sync() has
> >>>> already been called once for this device in the same code path which
> >>>> means that this particular instance will return immediately, so this
> >>>> is a false-positive (most likely).
> >>>>
> >>>> Let me see if I the might_sleep_if() assertion in
> >>>> __pm_runtime_resume(() can be moved to a better place.
> >>>>
> >>> Hi Rafael,
> >>>
> >>> did you had a chance to look at this?
> >>> The problem still remains in Linux v4.10-rc5.
> >>
> >>
> >> No, I didn't.
> >>
> >> As I said, this is not a serious issue.
> >
> > Something like the attached (untested).
> >
> > Please try it and let me know if it makes the splat go away.
> >
>
> Your patch fixes the issue here.
> I tested against vanilla Linux v4.10-rc5.
>
> Feel free to give appropriate credits.
OK, thanks!
Below is a full version with a changelog & tags.
Thanks,
Rafael
---
From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
Subject: [PATCH] PM / runtime: Avoid false-positive warnings from might_sleep_if()
The might_sleep_if() assertions in __pm_runtime_idle(),
__pm_runtime_suspend() and __pm_runtime_resume() may generate
false-positive warnings in some situations. For example, that
happens if a nested pm_runtime_get_sync()/pm_runtime_put() pair
is executed with disabled interrupts within an outer
pm_runtime_get_sync()/pm_runtime_put() section for the same device.
[Generally, pm_runtime_get_sync() may sleep, so it should not be
called with disabled interrupts, but in this particular case the
previous pm_runtime_get_sync() guarantees that the device will not
be suspended, so the inner pm_runtime_get_sync() will return
immediately after incrementing the device's usage counter.]
That started to happen in the i915 driver in 4.10-rc, leading to
the following splat:
BUG: sleeping function called from invalid context at drivers/base/power/runtime.c:1032
in_atomic(): 1, irqs_disabled(): 0, pid: 1500, name: Xorg
1 lock held by Xorg/1500:
#0: (&dev->struct_mutex){+.+.+.}, at:
[<ffffffffa0680c13>] i915_mutex_lock_interruptible+0x43/0x140 [i915]
CPU: 0 PID: 1500 Comm: Xorg Not tainted
Call Trace:
dump_stack+0x85/0xc2
___might_sleep+0x196/0x260
__might_sleep+0x53/0xb0
__pm_runtime_resume+0x7a/0x90
intel_runtime_pm_get+0x25/0x90 [i915]
aliasing_gtt_bind_vma+0xaa/0xf0 [i915]
i915_vma_bind+0xaf/0x1e0 [i915]
i915_gem_execbuffer_relocate_entry+0x513/0x6f0 [i915]
i915_gem_execbuffer_relocate_vma.isra.34+0x188/0x250 [i915]
? trace_hardirqs_on+0xd/0x10
? i915_gem_execbuffer_reserve_vma.isra.31+0x152/0x1f0 [i915]
? i915_gem_execbuffer_reserve.isra.32+0x372/0x3a0 [i915]
i915_gem_do_execbuffer.isra.38+0xa70/0x1a40 [i915]
? __might_fault+0x4e/0xb0
i915_gem_execbuffer2+0xc5/0x260 [i915]
? __might_fault+0x4e/0xb0
drm_ioctl+0x206/0x450 [drm]
? i915_gem_execbuffer+0x340/0x340 [i915]
? __fget+0x5/0x200
do_vfs_ioctl+0x91/0x6f0
? __fget+0x111/0x200
? __fget+0x5/0x200
SyS_ioctl+0x79/0x90
entry_SYSCALL_64_fastpath+0x23/0xc6
even though the code triggering it is correct.
Unfortunately, the might_sleep_if() assertions in question are
too coarse-grained to cover such cases correctly, so make them
a bit less sensitive in order to avoid the false-positives.
Reported-and-tested-by: Sedat Dilek <sedat.dilek@xxxxxxxxx>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
---
drivers/base/power/runtime.c | 11 ++++++-----
1 file changed, 6 insertions(+), 5 deletions(-)
Index: linux-pm/drivers/base/power/runtime.c
===================================================================
--- linux-pm.orig/drivers/base/power/runtime.c
+++ linux-pm/drivers/base/power/runtime.c
@@ -966,13 +966,13 @@ int __pm_runtime_idle(struct device *dev
unsigned long flags;
int retval;
- might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
-
if (rpmflags & RPM_GET_PUT) {
if (!atomic_dec_and_test(&dev->power.usage_count))
return 0;
}
+ might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
+
spin_lock_irqsave(&dev->power.lock, flags);
retval = rpm_idle(dev, rpmflags);
spin_unlock_irqrestore(&dev->power.lock, flags);
@@ -998,13 +998,13 @@ int __pm_runtime_suspend(struct device *
unsigned long flags;
int retval;
- might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
-
if (rpmflags & RPM_GET_PUT) {
if (!atomic_dec_and_test(&dev->power.usage_count))
return 0;
}
+ might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
+
spin_lock_irqsave(&dev->power.lock, flags);
retval = rpm_suspend(dev, rpmflags);
spin_unlock_irqrestore(&dev->power.lock, flags);
@@ -1029,7 +1029,8 @@ int __pm_runtime_resume(struct device *d
unsigned long flags;
int retval;
- might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
+ might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe &&
+ dev->power.runtime_status != RPM_ACTIVE);
if (rpmflags & RPM_GET_PUT)
atomic_inc(&dev->power.usage_count);