[Intel-gfx] [PATCH 06/10] drm/syncobj: use the timeline point in drm_syncobj_find_fence v3

[Koenig, Christian]

Am 13.12.18 um 13:21 schrieb Chris Wilson:
> Quoting Koenig, Christian (2018-12-13 12:11:10)
>> Am 13.12.18 um 12:37 schrieb Chris Wilson:
>>> Quoting Chunming Zhou (2018-12-11 10:34:45)
>>>> From: Christian König <ckoenig.leichtzumerken at gmail.com>
>>>>
>>>> Implement finding the right timeline point in drm_syncobj_find_fence.
>>>>
>>>> v2: return -EINVAL when the point is not submitted yet.
>>>> v3: fix reference counting bug, add flags handling as well
>>>>
>>>> Signed-off-by: Christian König <christian.koenig at amd.com>
>>>> ---
>>>>    drivers/gpu/drm/drm_syncobj.c | 43 ++++++++++++++++++++++++++++++++---
>>>>    1 file changed, 40 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/drm_syncobj.c b/drivers/gpu/drm/drm_syncobj.c
>>>> index 76ce13dafc4d..d964b348ecba 100644
>>>> --- a/drivers/gpu/drm/drm_syncobj.c
>>>> +++ b/drivers/gpu/drm/drm_syncobj.c
>>>> @@ -231,16 +231,53 @@ int drm_syncobj_find_fence(struct drm_file *file_private,
>>>>                              struct dma_fence **fence)
>>>>    {
>>>>           struct drm_syncobj *syncobj = drm_syncobj_find(file_private, handle);
>>>> -       int ret = 0;
>>>> +       struct syncobj_wait_entry wait;
>>>> +       int ret;
>>>>    
>>>>           if (!syncobj)
>>>>                   return -ENOENT;
>>>>    
>>>>           *fence = drm_syncobj_fence_get(syncobj);
>>>> -       if (!*fence) {
>>>> +       drm_syncobj_put(syncobj);
>>>> +
>>>> +       if (*fence) {
>>>> +               ret = dma_fence_chain_find_seqno(fence, point);
>>>> +               if (!ret)
>>>> +                       return 0;
>>>> +               dma_fence_put(*fence);
>>>> +       } else {
>>>>                   ret = -EINVAL;
>>>>           }
>>>> -       drm_syncobj_put(syncobj);
>>>> +
>>>> +       if (!(flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT))
>>>> +               return ret;
>>>> +
>>>> +       memset(&wait, 0, sizeof(wait));
>>>> +       wait.task = current;
>>>> +       wait.point = point;
>>>> +       drm_syncobj_fence_add_wait(syncobj, &wait);
>>>> +
>>>> +       do {
>>>> +               set_current_state(TASK_INTERRUPTIBLE);
>>>> +               if (wait.fence) {
>>>> +                       ret = 0;
>>>> +                       break;
>>>> +               }
>>>> +
>>>> +               if (signal_pending(current)) {
>>>> +                       ret = -ERESTARTSYS;
>>>> +                       break;
>>>> +               }
>>>> +
>>>> +               schedule();
>>>> +       } while (1);
>>> I've previously used a dma_fence_proxy so that we could do nonblocking
>>> waits on future submits. That would be preferrable (a requirement for
>>> our stupid BKL-driven code).
>> That is exactly what I would definitely NAK.
>>
>> I would rather say we should come up with a wait_multiple_events() macro
>> and completely nuke the custom implementation of this in:
>> 1. dma_fence_default_wait and dma_fence_wait_any_timeout
>> 2. the radeon fence implementation
>> 3. the nouveau fence implementation
>> 4. the syncobj code
>>
>> Cause all of them do exactly the same. The dma_fence implementation
>> unfortunately came up with a custom event handling mechanism instead of
>> extending the core Linux wait_event() system.
> I don't want a blocking wait at all.

Ok I wasn't clear enough :) That is exactly what I would NAK!

The wait must be blocking or otherwise you would allow wait-before-signal.

Christian.

> -Chris