Bug 91298

It's bad synchronisation of the read buffer:

diff --git a/tests/spec/arb_shader_atomic_counters/common.c b/tests/spec/arb_shader_atomic_counters/common.c
index 95d809e..bf8309d 100644
--- a/tests/spec/arb_shader_atomic_counters/common.c
+++ b/tests/spec/arb_shader_atomic_counters/common.c
@@ -35,6 +35,7 @@ atomic_counters_probe_buffer(unsigned base, unsigned count,
         uint32_t *p = glMapBufferRange(
                 GL_ATOMIC_COUNTER_BUFFER, base * sizeof(uint32_t),
                 count * sizeof(uint32_t), GL_MAP_READ_BIT);
+        bool pass = true;
         unsigned i;
 
         if (!p) {
@@ -43,17 +44,18 @@ atomic_counters_probe_buffer(unsigned base, unsigned count,
         }
 
         for (i = 0; i < count; ++i) {
-                if (p[i] != expected[i]) {
+                uint32_t *found = p[i];
+                if (found != expected[i]) {
                         printf("Probe value at (%i)\n", i);
                         printf("  Expected: 0x%08x\n", expected[i]);
-                        printf("  Observed: 0x%08x\n", p[i]);
-                        glUnmapBuffer(GL_ATOMIC_COUNTER_BUFFER);
-                        return false;
+                        printf("  Observed: 0x%08x\n", found);
+                        pass = false;
+                        break;
                 }
         }
 
         glUnmapBuffer(GL_ATOMIC_COUNTER_BUFFER);
-        return true;
+        return pass;
 }
 
 bool

should fix up the error message.

s/*found/found/ ofc

I sent ickle's patch to the mailing list, and fixed the same bug in shader_runner and a compute shader test.

However, this just fixes the test failure message to make sense.

The driver still has a race condition that is causing these intermittent failures, and we need to fix that.

It definitely seems to be some kind of race - adding usleep(1000) in the test inbetween glMapBufferRange and the p[i] access effectively hides the problem.

It also works fine on Haswell - I've run it for a few hours with no failures.

INTEL_DEBUG=sync doesn't help.  always_flush_batch=true seems to make it fail a little more often...it at minimum doesn't help...

Chris, by chance do you have any ideas about this bug?

I'm fairly stumped...

piglit.spec.arb_fragment_layer_viewport.layer-gs-writes-in-range depends on atomic counters and suffers from this bug.

I've sent a test to the mailing list that might be related to this same issue. Adding syncs or sleeps does not seem to help there though.

http://lists.freedesktop.org/archives/piglit/2015-November/018227.html

As another maddening data point... I've never been able to reproduce this on my ThinkPad X250.  I had thought it could be a kernel regression (I was on a quite old kernel), but I just ran the test 100 times in a loop without a single failure on 4.2.6-301.fc23.x86_64.

Perhaps it's only a problem with GBM runs?  I don't use GBM.

I've been trying to reproduce this failure on BDW today without success.  It doesn't seem to make a difference for me whether I use X or GBM to run the test case.

I noticed though that the DRM render ring code is missing a DC flush on batchbuffer submission, what could lead to a non-deterministic failure in this test case.  Can anyone try out the attached kernel patch and see if it helps?  If it does this should be considered a DRM bug.

Created attachment 120435 [details] [review]
gen7_flush_batch_dc_writes.patch

Curro's patch fixes the tests that are intermittent.

However, upgrading to Linux 4.3 (to apply the patch) produces an intermitten failures on piglit.spec.arb_shader_image_load_store.qualifiers.  I'll write that bug up.

(In reply to Mark Janes from comment #11)
> Curro's patch fixes the tests that are intermittent.
> 
> However, upgrading to Linux 4.3 (to apply the patch) produces an intermitten
> failures on piglit.spec.arb_shader_image_load_store.qualifiers.  I'll write
> that bug up.

Thanks, I've added your tested-by and sent the patch for review.

Fix pushed to drm-intel-next-queued as

commit 965fd602a6436f689f4f2fe40a6789582778ccd5
Author: Francisco Jerez <currojerez@riseup.net>
Date:   Wed Jan 13 18:59:39 2016 -0800

    drm/i915: Make sure DC writes are coherent on flush.