[Intel-gfx] [PATCH 5/4] dma-fence: Have dma_fence_signal call signal_locked

[Daniel Vetter]

On Thu, Aug 15, 2019 at 8:46 PM Chris Wilson <chris at chris-wilson.co.uk> wrote:
> Quoting Daniel Vetter (2019-08-14 18:20:53)
> > On Sun, Aug 11, 2019 at 10:15:23AM +0100, Chris Wilson wrote:
> > > Now that dma_fence_signal always takes the spinlock to flush the
> > > cb_list, simply take the spinlock and call dma_fence_signal_locked() to
> > > avoid code repetition.
> > >
> > > Suggested-by: Christian König <christian.koenig at amd.com>
> > > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > > Cc: Christian König <christian.koenig at amd.com>
> >
> > Hm, I think this largely defeats the point of having the lockless signal
> > enabling trickery in dma_fence. Maybe that part isn't needed by anyone,
> > but feels like a thing that needs a notch more thought. And if we need it,
> > maybe a bit more cleanup.
>
> You mean dma_fence_enable_sw_signaling(). The only user appears to be to
> flush fences, which is actually the intent of always notifying the signal
> cb. By always doing the callbacks, we can avoid installing the interrupt
> and completely saturating CPUs with irqs, instead doing a batch in a
> leisurely timer callback if not flushed naturally.

Yeah I'm not against ditching this, but can't we ditch a lot more if
we just always take the spinlock in those paths now anyways? Kinda not
worth having the complexity anymore.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch