[Intel-gfx] [PATCH 1/3] drm/i915/perf: Reduce cpu overhead for blocking perf OA reads

[Umesh Nerlige Ramappa]

On Mon, Apr 13, 2020 at 11:58:18PM -0700, Dixit, Ashutosh wrote:
>On Mon, 13 Apr 2020 22:08:47 -0700, Dixit, Ashutosh wrote:
>>
>> On Mon, 13 Apr 2020 08:48:20 -0700, Umesh Nerlige Ramappa wrote:
>> >
>> > A condition in wait_event_interruptible seems to be checked twice before
>> > waiting on the event to occur. These checks are redundant when hrtimer
>> > events will call oa_buffer_check_unlocked to update the oa_buffer tail
>> > pointers. The redundant checks add cpu overhead. Simplify the check
>> > to reduce cpu overhead when using blocking io to read oa buffer reports.
>> >
>> > Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa at intel.com>
>> > ---
>> >  drivers/gpu/drm/i915/i915_perf.c | 28 +++++++++++++++++++++++++++-
>> >  1 file changed, 27 insertions(+), 1 deletion(-)
>> >
>> > diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
>> > index 5cde3e4e7be6..e28a3ab83fde 100644
>> > --- a/drivers/gpu/drm/i915/i915_perf.c
>> > +++ b/drivers/gpu/drm/i915/i915_perf.c
>> > @@ -541,6 +541,32 @@ static bool oa_buffer_check_unlocked(struct i915_perf_stream *stream)
>> >	return pollin;
>> >  }
>> >
>> > +/**
>> > + * oa_buffer_check_reports - quick check if reports are available
>> > + * @stream: i915 stream instance
>> > + *
>> > + * The return from this function is used as a condition for
>> > + * wait_event_interruptible in blocking read. This is used to detect
>> > + * available reports.
>> > + *
>> > + * A condition in wait_event_interruptible seems to be checked twice before
>> > + * waiting on an event to occur. These checks are redundant when hrtimer events
>> > + * will call oa_buffer_check_unlocked to update the oa_buffer tail pointers. The
>> > + * redundant checks add cpu overhead. We simplify the check to reduce cpu
>> > + * overhead.
>> > + */
>> > +static bool oa_buffer_check_reports(struct i915_perf_stream *stream)
>> > +{
>> > +	unsigned long flags;
>> > +	bool available;
>> > +
>> > +	spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
>> > +	available = stream->oa_buffer.tail != stream->oa_buffer.head;
>> > +	spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
>> > +
>> > +	return available;
>> > +}
>> > +
>> >  /**
>> >   * append_oa_status - Appends a status record to a userspace read() buffer.
>> >   * @stream: An i915-perf stream opened for OA metrics
>> > @@ -1150,7 +1176,7 @@ static int i915_oa_wait_unlocked(struct i915_perf_stream *stream)
>> >		return -EIO;
>> >
>> >	return wait_event_interruptible(stream->poll_wq,
>> > -					oa_buffer_check_unlocked(stream));
>> > +					oa_buffer_check_reports(stream));
>> >  }
>>
>> I think the problem with this patch is that the original code had the
>> property that the condition for data availability is checked (and the OA
>> tail advanced) /before/ entering the wait. So the tail was advanced and if
>> there was data there was no need to wait at all. This change has lost that
>> property. With this change we will first always enter the wait and then get
>> unblocked after the timer interrupt which will advance the tail and wake us
>> up.
>>
>> I think if we want to do this, it is possible to do without losing the
>> original property. Just call oa_buffer_check_unlocked() first (outside
>> wait_event) and if there is data just return. If not, put yourself on
>> stream->poll_wq from which the timer callback will wake us up. I think this
>> is going to be something like prepare_to_wait() / schedule() /
>> finish_wait() pattern of which there are numerous examples in the
>> kernel. So in this case we won't call wait_event_interruptible() which
>> checks the condition upon waking up. No need to define
>> oa_buffer_check_reports() either.
>
>If on the other hand we say that this should actually be the desired
>behavior of the blocking read, that it should not unblock immediately but
>only after the next timer interrupt (so that an app can call the blocking
>read repeatedly without worrying about it will return a little bit of data
>in each call at a considerable amount of CPU load), then we may be able to
>do something like this, but even then we may have to think about what would
>be the correct way to do that. Though this may be that correct way and we
>may just need to change the commit message, but is that what we want?

I am not quite clear on how the blocking read should behave in terms of 
the API itself.

The change here is solely to reduce cpu overhead from the additional 2 
calls to oa_buffer_check_unlocked before blocking and that would happen 
on every call to the read. I agree that the read would block if the 
timer did not update the tail yet, but that makes sense in a way that we 
don't want read to constantly return data when the OA sampling rates are 
high (say 100 us). With this change the behavior becomes consistent 
irrespective of the OA sampling rate and user has the flexibility to set 
the POLL_OA_PERIOD to whatever value is suitable for the OA sampling 
rate chosen.

Thanks,
Umesh