[PATCH 2/2] drm/amd/sched: add schuduling policy
Andres Rodriguez
andresx7 at gmail.com
Thu Mar 16 16:31:59 UTC 2017
On 2017-03-16 12:01 PM, Christian König wrote:
> Am 16.03.2017 um 16:31 schrieb Andres Rodriguez:
>>
>>
>> On 2017-03-16 11:13 AM, Andres Rodriguez wrote:
>>>
>>>
>>> On 2017-03-16 05:15 AM, zhoucm1 wrote:
>>>>
>>>>
>>>> On 2017年03月16日 17:10, Christian König wrote:
>>>>> Am 16.03.2017 um 10:00 schrieb Chunming Zhou:
>>>>>> if high priority rq is full, then process with low priority could be
>>>>>> starve.
>>>>>> Add policy for this problem, the high proiority can ahead of next
>>>>>> priority queue,
>>>>>> the ratio is 2 : 1.
>>>>>>
>>>>>> Change-Id: I58f4a6b9cdce8689b18dd8e83dd6e2cf5f99d5fb
>>>>>> Signed-off-by: Chunming Zhou <David1.Zhou at amd.com>
>>>>>
>>>>> Well, the idea behind the high priority queues is to actually starve
>>>>> the low priority queues to a certain amount.
>>>>>
>>>>> At least for the kernel queue that is really desired.
>>>> Yes, agree, but we certainly don't want low priority queue is totally
>>>> dead, which doesn't have chance to run if high priority queue is always
>>>> full and busy.
>>>> If without Andres changes, it doesn't matter. But after Andres changes
>>>> upstream, we need scheduling policy.
>>>>
>>>> Regards,
>>>> David Zhou
>>>
>>> Hey David,
>>>
>>> The desired behavior is actually starvation. This is why allocating a
>>> high priority context is locked behind root privileges at the moment.
>
> Thanks for the quick response, and yes that's what I was expecting as
> requirement for that feature as well.
>
>>>
>>> High priority work includes many different features intended for the
>>> safety of the user wearing the headset. These include:
>>> - stopping the user from tripping over furniture
>>> - preventing motion sickness
>>>
>>> We cannot compromise these safety features in order to run non-critical
>>> tasks.
>>>
>>> The current approach also has the disadvantage of heavily interleaving
>>> work. This is going to have two undesired side effects. First, there
>>> will be a performance overhead from having the queue context switch so
>>> often.
>>>
>>> Second, this approach improves concurrency of work in a ring with mixed
>>> priority work, but actually decreases overall system concurrency. When a
>>> ring's HW queue has any high priority work committed, the whole ring
>>> will be executing at high priority. So interleaving the work will result
>>> in extending the time a ring spends in high priority mode. This
>>> effectively extends time that a ring might spend with CU's reserved
>>> which will have a performance impact on other rings.
>>>
>>> The best approach here is to make sure we don't map high priority work
>>> and regular priority work to the same ring. This is why the LRU policy
>>> for userspace ring ids to kernel ring ids was introduced. This policy
>>> provides the guarantee as a side effect.
>>
>> Wanted to correct myself here. The LRU policy doesn't actually provide
>> a guarantee. It approximates it.
>
> What David is perfectly right with is that this has the potential for
> undesired side effects.
>
> But I completely agree that deadline or fair queue scheduling is tricky
> to implement even when it is desired.
>
> So letting a submission dominate all other is perfectly ok for me as
> long as it is limited to a certain set of processes somehow.
>
> Christian.
I also wanted to comment that I think policies for the gpu_scheduler are
something interesting to explore. Not just at single ring level, but
also policies that tie all the gpu_schedulers for a single ASIC
together. Currently they all operate independently, but technically they
share the underlying HW so there is some resource aliasing.
There has been a lot of experimentation on this area in the VR world by
other OS vendors. Enhancing the policies provided by the SPI through SW
can have some pretty great benefits. Although because of restrictions
these kind of changes are only activated for certain HMDs.
I think we have a good opportunity here to make the linux version of VR
better for all HMDs, and not just the ones associated with the OS vendor.
Regards,
Andres
>
>>
>> Regards,
>> Andres
>>
>>> But further work there could be
>>> useful to actually check context priorities when doing the ring mapping.
>>>
>>> By placing work on different rings, we let the hardware actually
>>> dispatch work according to wave parameters like VGPR usage, etc. Trying
>>> to deduce all this information in the GPU scheduler will get very
>>> complicated.
>>>
>>> This is NAK'd by me.
>>>
>>> Regards,
>>> Andres
>>>
>>>
>>>>>
>>>>> Christian.
>>>>>
>>>>>> ---
>>>>>> drivers/gpu/drm/amd/scheduler/gpu_scheduler.c | 26
>>>>>> +++++++++++++++++++++++---
>>>>>> drivers/gpu/drm/amd/scheduler/gpu_scheduler.h | 2 ++
>>>>>> 2 files changed, 25 insertions(+), 3 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
>>>>>> b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
>>>>>> index 0f439dd..4637b6f 100644
>>>>>> --- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
>>>>>> +++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
>>>>>> @@ -35,11 +35,16 @@
>>>>>> static void amd_sched_process_job(struct fence *f, struct fence_cb
>>>>>> *cb);
>>>>>> /* Initialize a given run queue struct */
>>>>>> -static void amd_sched_rq_init(struct amd_sched_rq *rq)
>>>>>> +static void amd_sched_rq_init(struct amd_gpu_scheduler *sched, enum
>>>>>> + amd_sched_priority pri)
>>>>>> {
>>>>>> + struct amd_sched_rq *rq = &sched->sched_rq[pri];
>>>>>> +
>>>>>> spin_lock_init(&rq->lock);
>>>>>> INIT_LIST_HEAD(&rq->entities);
>>>>>> rq->current_entity = NULL;
>>>>>> + rq->wait_base = pri * 2;
>>>>>> + rq->wait = rq->wait_base;
>>>>>> }
>>>>>> static void amd_sched_rq_add_entity(struct amd_sched_rq *rq,
>>>>>> @@ -494,17 +499,32 @@ static void amd_sched_wakeup(struct
>>>>>> amd_gpu_scheduler *sched)
>>>>>> {
>>>>>> struct amd_sched_entity *entity;
>>>>>> int i;
>>>>>> + bool skip;
>>>>>> if (!amd_sched_ready(sched))
>>>>>> return NULL;
>>>>>> +retry:
>>>>>> + skip = false;
>>>>>> /* Kernel run queue has higher priority than normal run queue*/
>>>>>> for (i = AMD_SCHED_PRIORITY_MAX - 1; i >=
>>>>>> AMD_SCHED_PRIORITY_MIN; i--) {
>>>>>> + if ((i > AMD_SCHED_PRIORITY_MIN) &&
>>>>>> + (sched->sched_rq[i - 1].wait >=
>>>>>> sched->sched_rq[i].wait_base)) {
>>>>>> + sched->sched_rq[i - 1].wait = sched->sched_rq[i -
>>>>>> 1].wait_base;
>>>>>> + skip = true;
>>>>>> + continue;
>>>>>> + }
>>>>>> entity = amd_sched_rq_select_entity(&sched->sched_rq[i]);
>>>>>> - if (entity)
>>>>>> + if (entity) {
>>>>>> + if (i > AMD_SCHED_PRIORITY_MIN)
>>>>>> + sched->sched_rq[i - 1].wait++;
>>>>>> break;
>>>>>> + }
>>>>>> }
>>>>>> + if (!entity && skip)
>>>>>> + goto retry;
>>>>>> +
>>>>>> return entity;
>>>>>> }
>>>>>> @@ -608,7 +628,7 @@ int amd_sched_init(struct amd_gpu_scheduler
>>>>>> *sched,
>>>>>> sched->name = name;
>>>>>> sched->timeout = timeout;
>>>>>> for (i = AMD_SCHED_PRIORITY_MIN; i < AMD_SCHED_PRIORITY_MAX;
>>>>>> i++)
>>>>>> - amd_sched_rq_init(&sched->sched_rq[i]);
>>>>>> + amd_sched_rq_init(sched, i);
>>>>>> init_waitqueue_head(&sched->wake_up_worker);
>>>>>> init_waitqueue_head(&sched->job_scheduled);
>>>>>> diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
>>>>>> b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
>>>>>> index 99f0240..4caed30 100644
>>>>>> --- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
>>>>>> +++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
>>>>>> @@ -64,6 +64,8 @@ struct amd_sched_rq {
>>>>>> spinlock_t lock;
>>>>>> struct list_head entities;
>>>>>> struct amd_sched_entity *current_entity;
>>>>>> + int wait_base;
>>>>>> + int wait;
>>>>>> };
>>>>>> struct amd_sched_fence {
>>>>>
>>>>>
>>>>
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>
>
More information about the amd-gfx
mailing list