Re: [PATCH] drm/atomic: swap_state should stall on cleanup_done

[Maarten Lankhorst]

Op 01-11-2019 om 15:59 schreef Rob Clark:
> On Fri, Nov 1, 2019 at 7:47 AM Maarten Lankhorst
> <maarten.lankhorst@xxxxxxxxxxxxxxx> wrote:
>> Op 31-10-2019 om 23:36 schreef Rob Clark:
>>> From: Rob Clark <robdclark@xxxxxxxxxxxx>
>>>
>>> Stalling on cleanup_done ensures that any atomic state related to a
>>> nonblock commit no longer has dangling references to per-object state
>>> that can be freed.
>>>
>>> Otherwise, if a !nonblock commit completes after a nonblock commit has
>>> swapped state (ie. the synchronous part of the nonblock commit comes
>>> before the !nonblock commit), but before the asynchronous part of the
>>> nonblock commit completes, what was the new per-object state in the
>>> nonblock commit can be freed.
>>>
>>> This shows up with the new self-refresh helper, as _update_avg_times()
>>> dereferences the original old and new crtc_state.
>>>
>>> Fixes: d4da4e33341c ("drm: Measure Self Refresh Entry/Exit times to avoid thrashing")
>>> Cc: Sean Paul <seanpaul@xxxxxxxxxxxx>
>>> Signed-off-by: Rob Clark <robdclark@xxxxxxxxxxxx>
>>> ---
>>> Other possibilities:
>>> 1) maybe block later before freeing atomic state?
>>> 2) refcount individual per-object state
>>>
>>>  drivers/gpu/drm/drm_atomic_helper.c | 6 +++---
>>>  1 file changed, 3 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/drm_atomic_helper.c b/drivers/gpu/drm/drm_atomic_helper.c
>>> index 3ef2ac52ce94..a5d95429f91b 100644
>>> --- a/drivers/gpu/drm/drm_atomic_helper.c
>>> +++ b/drivers/gpu/drm/drm_atomic_helper.c
>>> @@ -2711,7 +2711,7 @@ int drm_atomic_helper_swap_state(struct drm_atomic_state *state,
>>>                       if (!commit)
>>>                               continue;
>>>
>>> -                     ret = wait_for_completion_interruptible(&commit->hw_done);
>>> +                     ret = wait_for_completion_interruptible(&commit->cleanup_done);
>>>                       if (ret)
>>>                               return ret;
>>>               }
>>> @@ -2722,7 +2722,7 @@ int drm_atomic_helper_swap_state(struct drm_atomic_state *state,
>>>                       if (!commit)
>>>                               continue;
>>>
>>> -                     ret = wait_for_completion_interruptible(&commit->hw_done);
>>> +                     ret = wait_for_completion_interruptible(&commit->cleanup_done);
>>>                       if (ret)
>>>                               return ret;
>>>               }
>>> @@ -2733,7 +2733,7 @@ int drm_atomic_helper_swap_state(struct drm_atomic_state *state,
>>>                       if (!commit)
>>>                               continue;
>>>
>>> -                     ret = wait_for_completion_interruptible(&commit->hw_done);
>>> +                     ret = wait_for_completion_interruptible(&commit->cleanup_done);
>>>                       if (ret)
>>>                               return ret;
>>>               }
>> Nack, hw_done means all new_crtc_state (from the old commit pov) dereferences are done.
>>
> hmm, it would be nice if the for_each_blah_in_state() iterators would
> splat on incorrect usage, then..  it tool a while to track down what
> was going wrong.  And Sean claimed the self refresh helpers worked for
> him on rockchip/i915 (although I'm starting to suspect maybe he just
> didn't have enough debug options enabled to poison freed memory?)
Could do a memset on the new arrays after hw_done?
>> Self refresh helpers should be fixed. :)
> Looks like what they need out of crtc_state is pretty minimal, maybe
> they could extract out crtc_state->self_refresh_active earlier..

Yeah, something like that would work. :)