[Intel-gfx] [PATCH] drm/i915: Avoid GPU stalls from kswapd

[Chris Wilson]

On Mon, Jun 22, 2015 at 05:35:11PM +0200, Daniel Vetter wrote:
> On Mon, Jun 22, 2015 at 03:18:12PM +0100, Chris Wilson wrote:
> > On Mon, Jun 22, 2015 at 04:11:01PM +0200, Daniel Vetter wrote:
> > > On Fri, Jun 19, 2015 at 02:04:16PM +0100, Chris Wilson wrote:
> > > > Exclude active GPU pages from the purview of the background shrinker
> > > > (kswapd), as these cause uncontrollable GPU stalls. Given that the
> > > > shrinker is rerun until the freelists are satisfied, we should have
> > > > opportunity in subsequent passes to recover the pages once idle. If the
> > > > machine does run out of memory entirely, we have the forced idling in the
> > > > oom-notifier as a means of releasing all the pages we can before an oom
> > > > is prematurely executed.
> > > > 
> > > > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > > > ---
> > > >  drivers/gpu/drm/i915/i915_drv.h          | 1 +
> > > >  drivers/gpu/drm/i915/i915_gem_shrinker.c | 8 ++++++--
> > > >  2 files changed, 7 insertions(+), 2 deletions(-)
> > > > 
> > > > diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> > > > index 71f4ca5088e2..e0dcd018379f 100644
> > > > --- a/drivers/gpu/drm/i915/i915_drv.h
> > > > +++ b/drivers/gpu/drm/i915/i915_drv.h
> > > > @@ -3185,6 +3185,7 @@ unsigned long i915_gem_shrink(struct drm_i915_private *dev_priv,
> > > >  #define I915_SHRINK_PURGEABLE 0x1
> > > >  #define I915_SHRINK_UNBOUND 0x2
> > > >  #define I915_SHRINK_BOUND 0x4
> > > > +#define I915_SHRINK_ACTIVE 0x8
> > > >  unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv);
> > > >  void i915_gem_shrinker_init(struct drm_i915_private *dev_priv);
> > > >  
> > > > diff --git a/drivers/gpu/drm/i915/i915_gem_shrinker.c b/drivers/gpu/drm/i915/i915_gem_shrinker.c
> > > > index bd1cf921aead..8d25ec8a6559 100644
> > > > --- a/drivers/gpu/drm/i915/i915_gem_shrinker.c
> > > > +++ b/drivers/gpu/drm/i915/i915_gem_shrinker.c
> > > > @@ -123,6 +123,10 @@ i915_gem_shrink(struct drm_i915_private *dev_priv,
> > > >  			    obj->madv != I915_MADV_DONTNEED)
> > > >  				continue;
> > > >  
> > > > +			if ((flags & I915_SHRINK_ACTIVE) == 0 &&
> > > 
> > > Isn't there a "I'm kswapd" process flag we could use to make the shrinker
> > > a bit more clever between synchronous reclaim and kswap reclaim? Or has
> > > synchronous reclaim died as a thing?
> > 
> > If we cared we can use the lock-stealer as a flag for when waiting may
> > be acceptable. But the better heuristic imo is to delay the stall until
> > all other sources of cache memory have been exhausted. Then either
> > kswapd or direct reclaim will flush the GPU in preference to killing
> > processes.
> 
> Well I'm more concerned about fairness since atm the shrinker stall gives
> us a very crude throttling of gpu processes. It makes sense not to stall
> kswapd though since that would just punish the system overall for no
> useful purpose at all. But punishing active processes would imo be a
> feature. Since we hold the BKL while shrinking that'll also automatically
> stall and gem users.

But you are not punishing individual GPU processes, but all GPU
processes, so it doesn't seem that fair to me. In terms of mm,
throttling is performed on processes when they dirty pages. If we want
to hook into the system sense of fairness, we should look to similarly
charge every process as it accesses the bo for the first time (since
put_pages at least).
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre