Skip to contentSkip to conversation

drm/i915: Handle sync_seqno correctly when seqno has wrapped.

PATCH In review Mika Kuoppala 4 participants 9 messages Nov 13, 2012 – Nov 29, 2012

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/i915: Handle sync_seqno correctly when seqno has wrapped.

If seqno has wrapped, normal compare operation will give wrong results. i915_seqno_passed can handle the wrap so use it instead.

---
 drivers/gpu/drm/i915/i915_gem.c |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +2 -2
index cdcf19d..35d5db8 100644
@@ -2138,7 +2138,7 @@ i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
 	seqno = ring->get_seqno(ring, true);
 
 	for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++)
-		if (seqno >= ring->sync_seqno[i])
+		if (i915_seqno_passed(seqno, ring->sync_seqno[i]))
 			ring->sync_seqno[i] = 0;
 
 	while (!list_empty(&ring->request_list)) {
@@ -2376,7 +2376,7 @@ i915_gem_object_sync(struct drm_i915_gem_object *obj,
 	idx = intel_ring_sync_index(from, to);
 
 	seqno = obj->last_read_seqno;
-	if (seqno <= from->sync_seqno[idx])
+	if (i915_seqno_passed(from->sync_seqno[idx], seqno))
 		return 0;
 
 	ret = i915_gem_check_olr(obj->ring, seqno);
Signature · 1 line
1.7.9.5
#1 lore raw
Quoted 2 lines · Mika Kuoppala

If seqno has wrapped, normal compare operation will give wrong results. i915_seqno_passed can handle the wrap so use it instead.

I'm still a little wary of this patch, as it means that we are emitting requests that are not associated with an execbuffer and so the wraparound detection is lost. Perhaps if you were to make the seqno wraparound handling explicit we could all sleep more soundly...

And then I'd accept this patch for making seqno handling consistent at least.
-Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#2 lore raw
In reply to Mika Kuoppala
Quoted 22 lines · On Tue, 13 Nov 2012 15:51:36 +0200 Mika Kuoppala…

If seqno has wrapped, normal compare operation will give wrong results. i915_seqno_passed can handle the wrap so use it instead.

---
 drivers/gpu/drm/i915/i915_gem.c |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +1 -1
index cdcf19d..35d5db8 100644
@@ -2138,7 +2138,7 @@ i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
 	seqno = ring->get_seqno(ring, true);
 
 	for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++)
-		if (seqno >= ring->sync_seqno[i])
+		if (i915_seqno_passed(seqno, ring->sync_seqno[i]))
 			ring->sync_seqno[i] = 0;
 
 	while (!list_empty(&ring->request_list)) {

drivers/gpu/drm/i915/i915_gem.c i915_gem_object_sync(struct drm_i915_gem_object *obj, · line 2382

Quoted 8 lines
drivers/gpu/drm/i915/i915_gem.c +1 -1
@@ -2376,7 +2376,7 @@ i915_gem_object_sync(struct drm_i915_gem_object *obj,
 	idx = intel_ring_sync_index(from, to);
 
 	seqno = obj->last_read_seqno;
-	if (seqno <= from->sync_seqno[idx])
+	if (i915_seqno_passed(from->sync_seqno[idx], seqno))
 		return 0;
 
 	ret = i915_gem_check_olr(obj->ring, seqno);

Can you add another patch on top of this to have the starting seqno be a near pre-wrap value so we can catch bugs even without igt.

Also as an overall comment, I want the patches to guarantee to catch the bug you found, which I think with the randomness of gem_stress - isn't. Specifically, we want the waiting ring to be waiting on a pre-wrapped value. Maybe I missed that guarantee, but if there is a quick/dirty way to make that happen, that would better than running an arbitrary number of gem_stress tests.

Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center
#3 lore raw
Quoted 7 lines · Ben Widawsky

Also as an overall comment, I want the patches to guarantee to catch the bug you found, which I think with the randomness of gem_stress - isn't. Specifically, we want the waiting ring to be waiting on a pre-wrapped value. Maybe I missed that guarantee, but if there is a quick/dirty way to make that happen, that would better than running an arbitrary number of gem_stress tests.

I think running just gem_stress is ok - as long as the test has a reasonable good chance of blowing up. On future platforms something else than semaphores might blow up, or we might simply botch a seqno comparison. So imo having a test that just beats a bit on the systems+the wrap-around after each boot/resume should give us excellent coverage, and trying to engineer a perfect test for the single failure mode we now have in front of us might actually reduce coverage.
-Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#4 lore raw
Quoted 9 lines · On Tue, 13 Nov 2012 17:45:14 +0100 Daniel Vetter <[email protected]> wrote:
Quoted 7 lines · Ben Widawsky

Also as an overall comment, I want the patches to guarantee to catch the bug you found, which I think with the randomness of gem_stress - isn't. Specifically, we want the waiting ring to be waiting on a pre-wrapped value. Maybe I missed that guarantee, but if there is a quick/dirty way to make that happen, that would better than running an arbitrary number of gem_stress tests.

Quoted 9 lines

I think running just gem_stress is ok - as long as the test has a reasonable good chance of blowing up. On future platforms something else than semaphores might blow up, or we might simply botch a seqno comparison. So imo having a test that just beats a bit on the systems+the wrap-around after each boot/resume should give us excellent coverage, and trying to engineer a perfect test for the single failure mode we now have in front of us might actually reduce coverage.
-Daniel

I didn't say don't run gem_stress...

Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center
#5 lore raw
In reply to Daniel Vetter
Quoted 3 lines · Daniel Vetter

trying to engineer a perfect test for the
single failure mode we now have in front of us might actually reduce coverage.

This is a critical point worth remembering when writing tests: Having a test case that detects a past bug is great, but a test case that detects a future bug is better.
-Chris.

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#6 lore raw
In reply to Ben Widawsky
Quoted 2 lines · Ben Widawsky

Can you add another patch on top of this to have the starting seqno be a near pre-wrap value so we can catch bugs even without igt.

I tried that but failed. I suspect that there is something in hw/sw that doesn't like if first seqno is >0x80000000. Something in hw needs to initialized explicitly if seqno is not starting from 1?

#7 lore raw
Quoted 9 lines · On Mon, 19 Nov 2012 12:55:22 +0200 Mika Kuoppala…
Quoted 2 lines · Ben Widawsky

Can you add another patch on top of this to have the starting seqno be a near pre-wrap value so we can catch bugs even without igt.

I tried that but failed. I suspect that there is something in hw/sw that doesn't like if first seqno is >0x80000000. Something in hw needs to initialized explicitly if seqno is not starting from 1?

Nothing that I am aware of. If you have the first be 0, and then add a huge jump right after that, does it work?

Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center
#8 lore raw
Quoted 8 lines · Ben Widawsky
Quoted 5 lines · On Mon, 19 Nov 2012 12:55:22 +0200 Mika Kuoppala…
Quoted 2 lines · Ben Widawsky

Can you add another patch on top of this to have the starting seqno be a near pre-wrap value so we can catch bugs even without igt.

Quoted 7 lines
Quoted 4 lines

I tried that but failed. I suspect that there is something in hw/sw that doesn't like if first seqno is >0x80000000. Something in hw needs to initialized explicitly if seqno is not starting from 1?

Nothing that I am aware of. If you have the first be 0, and then add a huge jump right after that, does it work?

We can't jump more than 0x80000000-1 as i915_seqno_passed() breaks. But with the patches for preallocate seqnos now in, I was able to get things working when first seqno was 0xFFFF0000.

-Mika