drm/i915: Prevent signals from interrupting memory recovery

Commit Message

Chris Wilson Sept. 4, 2012, 8 p.m. UTC

Just a spark of an idea to see if this is valid.

Given extreme memory pressure where we must zap our own caches and bound
buffers or else we fail to allocate a structure, we have a choice of
propagating that interrupt back to userspace or to quietly suppress it
in the hope of making the allocation. i915_gem_shrink_all() is a likely
candidate for that last ditch effort that currently fails to propagate
the interruption back...
---
 drivers/gpu/drm/i915/i915_gem.c |   10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

Comments

Ben Widawsky Sept. 8, 2012, 6:38 a.m. UTC | #1

On Tue,  4 Sep 2012 21:00:58 +0100
Chris Wilson <[e-mail]> wrote:

> Just a spark of an idea to see if this is valid.
> 
> Given extreme memory pressure where we must zap our own caches and
> bound buffers or else we fail to allocate a structure, we have a
> choice of propagating that interrupt back to userspace or to quietly
> suppress it in the hope of making the allocation.
> i915_gem_shrink_all() is a likely candidate for that last ditch
> effort that currently fails to propagate the interruption back...

What's the BKM for making the memory all go bye-bye (ie. how to test
this patch a bit)?

Chris Wilson Sept. 8, 2012, 8:03 a.m. UTC | #2

On Fri, 7 Sep 2012 23:38:05 -0700, Ben Widawsky <[e-mail]> wrote:
> On Tue,  4 Sep 2012 21:00:58 +0100
> Chris Wilson <[e-mail]> wrote:
> 
> > Just a spark of an idea to see if this is valid.
> > 
> > Given extreme memory pressure where we must zap our own caches and
> > bound buffers or else we fail to allocate a structure, we have a
> > choice of propagating that interrupt back to userspace or to quietly
> > suppress it in the hope of making the allocation.
> > i915_gem_shrink_all() is a likely candidate for that last ditch
> > effort that currently fails to propagate the interruption back...
> 
> What's the BKM for making the memory all go bye-bye (ie. how to test
> this patch a bit)?

There are a few i-g-t that purposesly exercise resource starvation, or
you can just use firefox (which has the added bonus of throwing in
signals for free)!
-Chris

Patch

diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 453335c..353332a 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -1868,11 +1868,19 @@  static void
 i915_gem_shrink_all(struct drm_i915_private *dev_priv)
 {
 	struct drm_i915_gem_object *obj, *next;
+	bool was_interruptible;
 
-	i915_gem_evict_everything(dev_priv->dev);
+	/* As we are called in a last-ditch attempt to recover memory,
+	 * make sure we wait for any unbind attempts to finish.
+	 */
+	was_interruptible = dev_priv->mm.interruptible;
+	dev_priv->mm.interruptible = false;
 
+	i915_gem_evict_everything(dev_priv->dev);
 	list_for_each_entry_safe(obj, next, &dev_priv->mm.unbound_list, gtt_list)
 		i915_gem_object_put_pages(obj);
+
+	dev_priv->mm.interruptible = was_interruptible;
 }
 
 static int