[Intel-gfx] [PATCH 2/2] drm/i915/gt: Do not schedule normal requests immediately along virtual

[Tvrtko Ursulin]

On 27/05/2020 08:03, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2020-05-27 07:51:44)
>>
>> On 26/05/2020 10:07, Chris Wilson wrote:
>>> When we push a virtual request onto the HW, we update the rq->engine to
>>> point to the physical engine. A request that is then submitted by the
>>> user that waits upon the virtual engine, but along the physical engine
>>> in use, will then see that it is due to be submitted to the same engine
>>> and take a shortcut (and be queued without waiting for the completion
>>> fence). However, the virtual request may be preempted (either by higher
>>> priority users, or by timeslicing) and removed from the physical engine
>>> to be migrated over to one of its siblings. The dependent normal request
>>> however is oblivious to the removal of the virtual request and remains
>>> queued to execute on HW, believing that once it reaches the head of its
>>> queue all of its predecessors will have completed executing!
>>>
>>> v2: Beware restriction of signal->execution_mask prior to submission.
>>>
>>> Fixes: 6d06779e8672 ("drm/i915: Load balancing across a virtual engine")
>>> Testcase: igt/gem_exec_balancer/sliced
>>> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
>>> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>> Cc: <stable at vger.kernel.org> # v5.3+
>>> ---
>>>    drivers/gpu/drm/i915/i915_request.c | 25 +++++++++++++++++++++----
>>>    1 file changed, 21 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
>>> index 33bbad623e02..0b07ccc7e9bc 100644
>>> --- a/drivers/gpu/drm/i915/i915_request.c
>>> +++ b/drivers/gpu/drm/i915/i915_request.c
>>> @@ -1237,6 +1237,25 @@ i915_request_await_execution(struct i915_request *rq,
>>>        return 0;
>>>    }
>>>    
>>> +static int
>>> +await_request_submit(struct i915_request *to, struct i915_request *from)
>>> +{
>>> +     /*
>>> +      * If we are waiting on a virtual engine, then it may be
>>> +      * constrained to execute on a single engine *prior* to submission.
>>> +      * When it is submitted, it will be first submitted to the virtual
>>> +      * engine and then passed to the physical engine. We cannot allow
>>> +      * the waiter to be submitted immediately to the physical engine
>>> +      * as it may then bypass the virtual request.
>>> +      */
>>> +     if (to->engine == READ_ONCE(from->engine))
>>> +             return i915_sw_fence_await_sw_fence_gfp(&to->submit,
>>> +                                                     &from->submit,
>>> +                                                     I915_FENCE_GFP);
>>> +     else
>>> +             return __i915_request_await_execution(to, from, NULL);
>>
>> If I am following correctly this branch will be the virtual <-> physical
>> or virtual <-> virtual dependency on the same physical engine. Why is
>> await_execution sufficient in this case? Is there something preventing
>> timeslicing between the two (not wanted right!) once from start
>> execution (not finishes).
> 
> Timeslicing is only between independent requests. When we expire a
> request, we also expire all of its waiters along the same engine, so
> that the execution order remains intact.

Via scheduler dependencies - they are enough?

Regards,

Tvrtko

> 
> await_execution is a more restrictive form of the
> await_sw_fence(submit), in that 'to' can only be submitted once 'from'
> reaches HW and not simply once 'from' reaches its engine submission
> queue.