[PATCH 3/3] drm/amdgpu: Switch to interrupted wait to recover from ring hang.

[Eric W. Biederman]

Andrey Grodzovsky <Andrey.Grodzovsky at amd.com> writes:

> On 04/24/2018 12:30 PM, Eric W. Biederman wrote:
>> "Panariti, David" <David.Panariti at amd.com> writes:
>>
>>> Andrey Grodzovsky <andrey.grodzovsky at amd.com> writes:
>>>> Kind of dma_fence_wait_killable, except that we don't have such API
>>>> (maybe worth adding ?)
>>> Depends on how many places it would be called, or think it might be called.  Can always factor on the 2nd time it's needed.
>>> Factoring, IMO, rarely hurts.  The factored function can easily be visited using `M-.' ;->
>>>
>>> Also, if the wait could be very long, would a log message, something like "xxx has run for Y seconds."  help?
>>> I personally hate hanging w/no info.
>> Ugh.  This loop appears susceptible to loosing wake ups.  There are
>> races between when a wake-up happens, when we clear the sleeping state,
>> and when we test the stat to see if we should stat awake.  So yes
>> implementing a dma_fence_wait_killable that handles of all that
>> correctly sounds like an very good idea.
>
> I am not clear here - could you be more specific about what races will happen
> here, more bellow
>>
>> Eric
>>
>>
>>>> If the ring is hanging for some reason allow to recover the waiting by sending fatal signal.
>>>>
>>>> Originally-by: David Panariti <David.Panariti at amd.com>
>>>> Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky at amd.com>
>>>> ---
>>>>    drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 14 ++++++++++----
>>>>    1 file changed, 10 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
>>>> index eb80edf..37a36af 100644
>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
>>>> @@ -421,10 +421,16 @@ int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx, unsigned ring_id)
>>>>
>>>>        if (other) {
>>>>                signed long r;
>>>> -             r = dma_fence_wait_timeout(other, false, MAX_SCHEDULE_TIMEOUT);
>>>> -             if (r < 0) {
>>>> -                     DRM_ERROR("Error (%ld) waiting for fence!\n", r);
>>>> -                     return r;
>>>> +
>>>> +             while (true) {
>>>> +                     if ((r = dma_fence_wait_timeout(other, true,
>>>> +                                     MAX_SCHEDULE_TIMEOUT)) >= 0)
>>>> +                             return 0;
>>>> +
>
> Do you mean that by the time I reach here some other thread from my group
> already might dequeued SIGKILL since it's a shared signal and hence
> fatal_signal_pending will return false ? Or are you talking about the
> dma_fence_wait_timeout implementation in dma_fence_default_wait with
> schedule_timeout ?

Given Oleg's earlier comment about the scheduler having special cases
for signals I might be wrong.  But in general there is a pattern:

	for (;;) {
		set_current_state(TASK_UNINTERRUPTIBLE);
		if (loop_is_done())
			break;
		schedule();
	}
        set_current_state(TASK_RUNNING);

If you violate that pattern by testing for a condition without
having first set your task as TASK_UNINTERRUPTIBLE (or whatever your
sleep state is).  Then it is possible to miss a wake-up that
tests the condidtion.

Thus I am quite concerned that there is a subtle corner case where
you can miss a wakeup and not retest fatal_signal_pending().

Given that there is is a timeout the worst case might have you sleep
MAX_SCHEDULE_TIMEOUT instead of indefinitely.

Without a comment why this is safe, or having fatal_signal_pending
check integrated into dma_fence_wait_timeout I am not comfortable
with this loop.

Eric


>>>> +                     if (fatal_signal_pending(current)) {
>>>> +                             DRM_ERROR("Error (%ld) waiting for fence!\n", r);
>>>> +                             return r;
>>>> +                     }
>>>>                }
>>>>        }
>>>>
>>>> --
>>>> 2.7.4
>>>>
>> Eric