[Intel-gfx] [PATCH 17/32] drm/i915: Remove the lazy_coherency parameter from request-completed?

[Dave Gordon]

On 04/01/16 11:26, Chris Wilson wrote:
> On Mon, Jan 04, 2016 at 11:16:04AM +0000, Dave Gordon wrote:
>> On 14/12/15 15:11, Chris Wilson wrote:
>>> On Mon, Dec 14, 2015 at 02:59:30PM +0000, Tvrtko Ursulin wrote:
>>>>
>>>> Hi,
>>>>
>>>> On 11/12/15 11:33, Chris Wilson wrote:
>>>>> Now that we have split out the seqno-barrier from the
>>>>> engine->get_seqno() callback itself, we can move the users of the
>>>>> seqno-barrier to the required callsites simplifying the common code and
>>>>> making the required workaround handling much more explicit.
>>>>
>>>> What bothers me about this patch, and the one preceding it, is that
>>>> I don't see a tangible improvement for the programmer who still has
>>>> to know when to read the seqno and when to "read it harder, read for
>>>> real".
>>>
>>> In earlier patches, I called it irq_barrier.
>>>
>>> It's not reading it harder. It's just that there is a ordering issue
>>> with receiving an interrupt and the seqno write being visible.
>>>
>>>> Barrier in this sense has a relation to the state of things but
>>>> somehow feels too low level to me when used from the code. But to be
>>>> fair I am not sure how to better define it.
>>>>
>>>> Would ring->get_seqno paired with ring->read_seqno perhaps make
>>>> sense? Implementation for ring->read_seqno would just be a flush
>>>> followed by ring->get_seqno then. Or maybe keep the barrier and add
>>>> ring->read_seqno which would be ring->seqno_barrier +
>>>> ring_get_seqno?
>>>
>>> No.
>>> -Chris
>>
>> We could instead put the knowledge about whether and how to read
>> "for real" inside the read-the-seqno function. For example:
>
> You do appreciate the irony that you are on the reviewer list for patches
> that do that?
>
> http://cgit.freedesktop.org/~ickle/linux-2.6/commit/?h=breadcrumbs&id=34409f2d965001d7d63f21a1c5339b07eed6af34

No, I haven't got as far as that one, since it was posted over a week 
after the message at the head of this thread. Anyway, I still can't see 
in that patch anything equivalent to what I described above.

> There is just one place that we need the extra work, after the
> interrupt. All other places only care about the current value of the
> seqno in the CPU cache.
> -Chris

So, we could instead have a per-engine flag which says, interrupt has 
happened, next reader should expect an update (and try hard to see one)?

Flow would be:
     IRQ->handler
         sets flag, wakes first waiter, clears IRQs
     waiter wakes up
         reads seqno (read fn sees flag => coherent, clear flag if new)
         checks whether (only) its request is completed
             not complete: enable IRQ, sleep
             seqno match: dequeue request, enable IRQ, process completion
             match+: dequeue request, wake next, process completion

where 'process completion' involves updating the request state and 
waking all /additional/ waiters on the same request, whereas 'wake next' 
refers to threads waiting on the /next/ request.

Is that what your patch is trying to achieve? It's a bit hard to tell 
with the seqno-read-optimisation and irq-dispatch changes mixed in with 
the r-b-tree and all the other things in this sequence.

I think it would be easier to understand if some of the more obvious 
improvements (such as 18/32, 20/32, 24/32) were pushed in earlier, so 
that the code is as clear as possible before the patches that actually 
change the way things work are applied. And all the reset-related 
patches could be later, as that's an area with some subtlety to it.

.Dave.