[Intel-gfx] [Linux v4.10.0-rc1+] Still call-traces after suspend-resume (pm? i915? cpu/hotplug?)

[Rafael J. Wysocki]

On 1/31/2017 11:58 AM, Imre Deak wrote:
> Hi Rafael,

Hi,

> On Mon, Jan 30, 2017 at 11:44:37PM +0100, Rafael J. Wysocki wrote:
>> On 1/24/2017 2:33 AM, Sedat Dilek wrote:
>>> On Fri, Dec 30, 2016 at 3:02 PM, Rafael J. Wysocki <rafael at kernel.org> wrote:
>>>> On Fri, Dec 30, 2016 at 12:40 PM, Sedat Dilek <sedat.dilek at gmail.com> 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).
> Not sure what's the root cause, but thought to clarify the above:
>
> Yes, i915_gem_do_execbuffer() does take an RPM reference to optimize
> things, so the RPM get in aliasing_gtt_bind_vma() won't need to resume
> the device on this path. This isn't a guarantee though,
> aliasing_gtt_bind_vma() could be called from other places without an RPM
> reference. So preemption being disabled at that point is not
> intentional. I also can't see where on the above path it would get
> disabled due to a bug or otherwise.

The i915 code is correct AFAICS, but the core complains about a nested 
invocation of pm_runtime_get_sync() with disabled interrupts (which 
looks OK though), so the complaint is a false positive.

As I said, the way to go here appears to be to tweak the core, which I'm 
going to do.

Thanks,
Rafael