Re: [RFC 1/4] dma-fence: Add deadline awareness
From: Rob Clark
Date: Wed Jul 28 2021 - 13:54:47 EST
On Wed, Jul 28, 2021 at 10:23 AM Christian König
<christian.koenig@xxxxxxx> wrote:
>
>
>
> Am 28.07.21 um 17:15 schrieb Rob Clark:
> > On Wed, Jul 28, 2021 at 4:37 AM Christian König
> > <ckoenig.leichtzumerken@xxxxxxxxx> wrote:
> >> Am 28.07.21 um 09:03 schrieb Christian König:
> >>> Am 27.07.21 um 16:25 schrieb Rob Clark:
> >>>> On Tue, Jul 27, 2021 at 12:11 AM Christian König
> >>>> <ckoenig.leichtzumerken@xxxxxxxxx> wrote:
> >>>>> Am 27.07.21 um 01:38 schrieb Rob Clark:
> >>>>>> From: Rob Clark <robdclark@xxxxxxxxxxxx>
> >>>>>>
> >>>>>> Add a way to hint to the fence signaler of an upcoming deadline,
> >>>>>> such as
> >>>>>> vblank, which the fence waiter would prefer not to miss. This is to
> >>>>>> aid
> >>>>>> the fence signaler in making power management decisions, like boosting
> >>>>>> frequency as the deadline approaches and awareness of missing
> >>>>>> deadlines
> >>>>>> so that can be factored in to the frequency scaling.
> >>>>>>
> >>>>>> Signed-off-by: Rob Clark <robdclark@xxxxxxxxxxxx>
> >>>>>> ---
> >>>>>> drivers/dma-buf/dma-fence.c | 39
> >>>>>> +++++++++++++++++++++++++++++++++++++
> >>>>>> include/linux/dma-fence.h | 17 ++++++++++++++++
> >>>>>> 2 files changed, 56 insertions(+)
> >>>>>>
> >>>>>> diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c
> >>>>>> index ce0f5eff575d..2e0d25ab457e 100644
> >>>>>> --- a/drivers/dma-buf/dma-fence.c
> >>>>>> +++ b/drivers/dma-buf/dma-fence.c
> >>>>>> @@ -910,6 +910,45 @@ dma_fence_wait_any_timeout(struct dma_fence
> >>>>>> **fences, uint32_t count,
> >>>>>> }
> >>>>>> EXPORT_SYMBOL(dma_fence_wait_any_timeout);
> >>>>>>
> >>>>>> +
> >>>>>> +/**
> >>>>>> + * dma_fence_set_deadline - set desired fence-wait deadline
> >>>>>> + * @fence: the fence that is to be waited on
> >>>>>> + * @deadline: the time by which the waiter hopes for the fence to be
> >>>>>> + * signaled
> >>>>>> + *
> >>>>>> + * Inform the fence signaler of an upcoming deadline, such as
> >>>>>> vblank, by
> >>>>>> + * which point the waiter would prefer the fence to be signaled
> >>>>>> by. This
> >>>>>> + * is intended to give feedback to the fence signaler to aid in power
> >>>>>> + * management decisions, such as boosting GPU frequency if a periodic
> >>>>>> + * vblank deadline is approaching.
> >>>>>> + */
> >>>>>> +void dma_fence_set_deadline(struct dma_fence *fence, ktime_t
> >>>>>> deadline)
> >>>>>> +{
> >>>>>> + unsigned long flags;
> >>>>>> +
> >>>>>> + if (dma_fence_is_signaled(fence))
> >>>>>> + return;
> >>>>>> +
> >>>>>> + spin_lock_irqsave(fence->lock, flags);
> >>>>>> +
> >>>>>> + /* If we already have an earlier deadline, keep it: */
> >>>>>> + if (test_bit(DMA_FENCE_FLAG_HAS_DEADLINE_BIT, &fence->flags) &&
> >>>>>> + ktime_before(fence->deadline, deadline)) {
> >>>>>> + spin_unlock_irqrestore(fence->lock, flags);
> >>>>>> + return;
> >>>>>> + }
> >>>>>> +
> >>>>>> + fence->deadline = deadline;
> >>>>>> + set_bit(DMA_FENCE_FLAG_HAS_DEADLINE_BIT, &fence->flags);
> >>>>>> +
> >>>>>> + spin_unlock_irqrestore(fence->lock, flags);
> >>>>>> +
> >>>>>> + if (fence->ops->set_deadline)
> >>>>>> + fence->ops->set_deadline(fence, deadline);
> >>>>>> +}
> >>>>>> +EXPORT_SYMBOL(dma_fence_set_deadline);
> >>>>>> +
> >>>>>> /**
> >>>>>> * dma_fence_init - Initialize a custom fence.
> >>>>>> * @fence: the fence to initialize
> >>>>>> diff --git a/include/linux/dma-fence.h b/include/linux/dma-fence.h
> >>>>>> index 6ffb4b2c6371..4e6cfe4e6fbc 100644
> >>>>>> --- a/include/linux/dma-fence.h
> >>>>>> +++ b/include/linux/dma-fence.h
> >>>>>> @@ -88,6 +88,7 @@ struct dma_fence {
> >>>>>> /* @timestamp replaced by @rcu on
> >>>>>> dma_fence_release() */
> >>>>>> struct rcu_head rcu;
> >>>>>> };
> >>>>>> + ktime_t deadline;
> >>>>> Mhm, adding the flag sounds ok to me but I'm a bit hesitating adding
> >>>>> the
> >>>>> deadline as extra field here.
> >>>>>
> >>>>> We tuned the dma_fence structure intentionally so that it is only 64
> >>>>> bytes.
> >>>> Hmm, then I guess you wouldn't be a fan of also adding an hrtimer?
> >>>>
> >>>> We could push the ktime_t (and timer) down into the derived fence
> >>>> class, but I think there is going to need to be some extra storage
> >>>> *somewhere*.. maybe the fence signaler could get away with just
> >>>> storing the nearest upcoming deadline per fence-context instead?
> >>> I would just push that into the driver instead.
> >>>
> >>> You most likely don't want the deadline per fence anyway in complex
> >>> scenarios, but rather per frame. And a frame is usually composed from
> >>> multiple fences.
> > Right, I ended up keeping track of the nearest deadline in patch 5/4
> > which added drm/msm support:
> >
> > https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fpatchwork.freedesktop.org%2Fpatch%2F447138%2F&data=04%7C01%7Cchristian.koenig%40amd.com%7Cce6ace85263d448bbc9f08d951d9f06c%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637630819606427306%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=ameszAOlClaZNeUDlYr37ZdIytVXNgiEUKuctjXLqZ0%3D&reserved=0
> >
> > But if we do have the ktime_t in dma_fence in dma_fence, we can add
> > some checks and avoid calling back to the driver if a later deadline
> > is set on a fence that already has an earlier deadline. OTOH I
> > suppose I can push all that back to the driver to start, and we can
> > revisit once we have more drivers implementing deadline support.
>
> I still think that all of this is rather specific to your use case and
> have strong doubt that anybody else will implement that.
i915 does already have a similar thing in it's hand-rolled atomic
commit path. So I think msm won't be the only one. It should be also
useful to the other mobile GPUs with a gpu vs kms driver split,
although looking at the other gpu devfreq implementations, I don't
think they've yet gotten to this point in the fine tuning..
BR,
-R
> >> Thinking more about it we could probably kill the spinlock pointer and
> >> make the flags 32bit if we absolutely need that here.
> > If we had a 'struct dma_fence_context' we could push the spinlock, ops
> > pointer, and u64 context into that and replace with a single
> > dma_fence_context ptr, fwiw
>
> That won't work. We have a lot of use cases where you can't allocate
> memory, but must allocate a context.
>
> Christian.
>
> >
> > BR,
> > -R
> >
> >> But I still don't see the need for that, especially since most drivers
> >> probably won't implement it.
> >>
> >> Regards,
> >> Christian.
> >>
> >>> Regards,
> >>> Christian.
> >>>
> >>>> BR,
> >>>> -R
> >>>>
> >>>>> Regards,
> >>>>> Christian.
> >>>>>
> >>>>>> u64 context;
> >>>>>> u64 seqno;
> >>>>>> unsigned long flags;
> >>>>>> @@ -99,6 +100,7 @@ enum dma_fence_flag_bits {
> >>>>>> DMA_FENCE_FLAG_SIGNALED_BIT,
> >>>>>> DMA_FENCE_FLAG_TIMESTAMP_BIT,
> >>>>>> DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
> >>>>>> + DMA_FENCE_FLAG_HAS_DEADLINE_BIT,
> >>>>>> DMA_FENCE_FLAG_USER_BITS, /* must always be last member */
> >>>>>> };
> >>>>>>
> >>>>>> @@ -261,6 +263,19 @@ struct dma_fence_ops {
> >>>>>> */
> >>>>>> void (*timeline_value_str)(struct dma_fence *fence,
> >>>>>> char *str, int size);
> >>>>>> +
> >>>>>> + /**
> >>>>>> + * @set_deadline:
> >>>>>> + *
> >>>>>> + * Callback to allow a fence waiter to inform the fence
> >>>>>> signaler of an
> >>>>>> + * upcoming deadline, such as vblank, by which point the
> >>>>>> waiter would
> >>>>>> + * prefer the fence to be signaled by. This is intended to
> >>>>>> give feedback
> >>>>>> + * to the fence signaler to aid in power management
> >>>>>> decisions, such as
> >>>>>> + * boosting GPU frequency.
> >>>>>> + *
> >>>>>> + * This callback is optional.
> >>>>>> + */
> >>>>>> + void (*set_deadline)(struct dma_fence *fence, ktime_t deadline);
> >>>>>> };
> >>>>>>
> >>>>>> void dma_fence_init(struct dma_fence *fence, const struct
> >>>>>> dma_fence_ops *ops,
> >>>>>> @@ -586,6 +601,8 @@ static inline signed long dma_fence_wait(struct
> >>>>>> dma_fence *fence, bool intr)
> >>>>>> return ret < 0 ? ret : 0;
> >>>>>> }
> >>>>>>
> >>>>>> +void dma_fence_set_deadline(struct dma_fence *fence, ktime_t
> >>>>>> deadline);
> >>>>>> +
> >>>>>> struct dma_fence *dma_fence_get_stub(void);
> >>>>>> struct dma_fence *dma_fence_allocate_private_stub(void);
> >>>>>> u64 dma_fence_context_alloc(unsigned num);
>