[Intel-gfx] [RFC 15/44] drm/i915: Added deferred work handler for scheduler

[Daniel Vetter]

On Thu, Jul 24, 2014 at 04:42:55PM +0100, John Harrison wrote:
> 
> On 23/07/2014 19:50, Daniel Vetter wrote:
> >On Wed, Jul 23, 2014 at 5:37 PM, John Harrison
> ><John.C.Harrison at intel.com> wrote:
> >>>>   diff --git a/drivers/gpu/drm/i915/i915_drv.h
> >>>>b/drivers/gpu/drm/i915/i915_drv.h
> >>>>index 0977653..fbafa68 100644
> >>>>--- a/drivers/gpu/drm/i915/i915_drv.h
> >>>>+++ b/drivers/gpu/drm/i915/i915_drv.h
> >>>>@@ -1075,6 +1075,16 @@ struct i915_gem_mm {
> >>>>         struct delayed_work idle_work;
> >>>>         /**
> >>>>+        * New scheme is to get an interrupt after every work packet
> >>>>+        * in order to allow the low latency scheduling of pending
> >>>>+        * packets. The idea behind adding new packets to a pending
> >>>>+        * queue rather than directly into the hardware ring buffer
> >>>>+        * is to allow high priority packets to over take low priority
> >>>>+        * ones.
> >>>>+        */
> >>>>+       struct work_struct scheduler_work;
> >>>Latency for work items isn't too awesome, and e.g. Oscar's execlist code
> >>>latches the next context right away from the irq handler. Why can't we do
> >>>something similar for the scheduler? Fishing the next item out of a
> >>>priority queue shouldn't be expensive ...
> >>>-Daniel
> >>
> >>The problem is that taking batch buffers from the scheduler's queue and
> >>submitting them to the hardware requires lots of processing that is not IRQ
> >>compatible. It isn't just a simple register write. Half of the code in
> >>'i915_gem_do_execbuffer()' must be executed. Probably/possibly it could be
> >>made IRQ friendly but that would place a lot of restrictions on a lot of
> >>code that currently doesn't expect to be restricted. Instead, the submission
> >>is done via a work handler that acquires the driver mutex lock.
> >>
> >>In order to cover the extra latency, the scheduler operates in a
> >>multi-buffered mode and aims to keep eight batch buffers in flight at all
> >>times. That number being obtained empirically by running lots of benchmarks
> >>on Android with lots of different settings and seeing where the buffer size
> >>stopped making a difference.
> >So I've tried to stitch together that part of the scheduler from the
> >patch series. Afaics you do the actual scheduling under the protection
> >of irqsave spinlocks (well you also hold the dev->struct_mutex). That
> >means you disable local interrupts. Up to the actual submit point I
> >spotted two such critcial sections encompassing pretty much all the
> >code.
> >
> >If we'd run the same code from the interrupt handler then only our own
> >interrupt handler is blocked, all other interrupt processing can
> >continue. So that's actually a lot nicer than what you have. In any
> >case you can't do expensive operations under an irqsave spinlock
> >anyway.
> >
> >So either I've missed something big here, or this justification doesn't hold up.
> 
> The irqsave spinlock is only held while manipulating the internal scheduler
> data structures. It is released immediately prior to calling
> i915_gem_do_execbuffer_final(). So the actual submission code path is done
> with the driver mutex but no spinlocks. I'm sure I got 'scheduling while
> atomic' bug checks the one time I accidentally left the spinlock held.

Ok, missed something ;-)

btw for big patch series please upload them somewhere on a public git
(github or so). Generally I review patches only by reading them because
trying to apply them usually results in some conflicts. But with big patch
series like this here that doesn't work, especially when everything is
tightly coupled (iirc I had to jump around in about 10 different patches
to figure out what the work handler looks like). Or maybe I didn't
understand the patch split flow and read it backwards ;-)
-Daniel
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch