[Intel-gfx] [PATCH v3 2/2] drm/i915: add syncobj timeline support

[Lionel Landwerlin]

On 01/08/2019 18:16, Chris Wilson wrote:
> Quoting Lionel Landwerlin (2019-08-01 15:29:34)
>> On 31/07/2019 23:03, Chris Wilson wrote:
>>> Quoting Lionel Landwerlin (2019-07-31 15:07:33)
...
>>> I think I have convinced myself that with the split between wait before,
>>> signal after combined with the rule that seqno point along the syncobj
>>> are monotonic, you should not be able to generate an AB-BA deadlock
>>> between concurrent clients.
>>
>> Can you come up with an example that would deadlock?
> Timeline holds 2,1; a wait on 2 will fail with -EINVAL to userspace.
> (Though possibly perfectly valid behaviour on the part of the user.)
>
> Timeline holds 2, with 1 being submitted. A wait on 1 waits on 2
> instead. If 1 gains a dependency on A (e.g. bad userspace or an


You mean : 1 gains a dependency on 2?


> implicit fence, it's a concurrent wait/submit so expect the worst, i.e.
> userspace has to be racing with itself to get into this mess), you
> now have a deadlock.


Not quite sure I see why...

2 doesn't have any dependency on 1, it just happens to have a number higher.


-Lionel


>   (The assumption being that the syncpt along the
> timeline are themselves not strictly ordered, and considering they are
> external syncobj, that seems like a reasonable generalisation.)
> -Chris
>