Hi Christian,
Your explanation matches my code study and test results. Well, I will remove original read. Shirish and I will re-test it. I will submit it after test done.
As for other related code, yes, I assumed there might be similar issues as well. My plan is to create internal ticket and write test code to check if any problem can be hit. Then we can do fix. I have the
current patch focus on the current Chrome issue.
Thanks for your review and information.
BR,
Louis
From: Christian König <ckoenig.leichtzumerken@xxxxxxxxx>
Sent: Tuesday, May 28, 2019 3:24 PM
To: Li, Ching-shih (Louis) <Ching-shih.Li@xxxxxxx>; Liu, Leo <Leo.Liu@xxxxxxx>; S, Shirish <Shirish.S@xxxxxxx>; Grodzovsky, Andrey <Andrey.Grodzovsky@xxxxxxx>; Zhang, Jerry <Jerry.Zhang@xxxxxxx>; Deng, Emily <Emily.Deng@xxxxxxx>; Deucher, Alexander <Alexander.Deucher@xxxxxxx>
Cc: amd-gfx@xxxxxxxxxxxxxxxxxxxxx
Subject: Re: [PATCH] drm/amdgpu: fix ring test failure issue during s3 in vce 3.0
[CAUTION: External Email]
Wow, really good catch!
The underlying problem is most likely that VCE block is either power or clock gated and because of this the readptr read always returns zero.
Now amdgpu_ring_alloc() informs the power management code that the block is about to be used and so the gating is turned off.
Mhm, that is probably wrong at a hole bunch of other places, at least the UVD and VCN code comes to mind.
I agree with Leo that you should remove the original read (so to not read twice) and it would be realy nice if you could double check the other code (UVD/VCN) for similar problems as well.
Regards,
Christian.
Am 27.05.19 um 19:20 schrieb Li, Ching-shih (Louis):
I don’t mean to read it twice. The solution is to make first read later. I didn’t modify the original code to make code difference less and simple. I guess it should work to remove the original read there.
int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
{
struct amdgpu_device *adev = ring->adev;
uint32_t rptr = amdgpu_ring_get_rptr(ring);
unsigned i;
int r, timeout = adev->usec_timeout;
/* skip ring test for sriov*/
if (amdgpu_sriov_vf(adev))
return 0;
r = amdgpu_ring_alloc(ring, 16);
if (r)
return r;
amdgpu_ring_write(ring, VCE_CMD_END);
amdgpu_ring_commit(ring);
Above is original code, rptr is updated when called, and below is your patch, my question is why do you need to get rptr twice?
@@ -1084,6 +1084,8 @@ int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
if (r)
return r;
+ rptr = amdgpu_ring_get_rptr(ring);
+
amdgpu_ring_write(ring, VCE_CMD_END);
amdgpu_ring_commit(ring);
On 5/27/19 12:22 PM, Li, Ching-shih (Louis) wrote:
Hi Leo,
Yes, I confirm it is the root cause
for the Chrome S3 issue. Whenever system is resumed, the original instruction always gets zero. However, I have no idea why it fails, and didn’t verify this problem on CRB or any other Linux platform yet.
Although I think the ideal solution is an indicator, e.g. a register, for driver to check if related firmware and hardware are ready to work. So driver can make sure it is ok to read rptr. Without any reference
document, I can only try to solve the problem by modifying driver. Debug traces reveal that only first rptr read fails, but the read in check loop is ok. Therefore, a solution comes to mind: to update rptr later for initial rptr value. Tests prove it working
in Chrome platforms. Fyi~
BR,
Louis
On 5/27/19 3:42 AM, S, Shirish wrote:
From: Louis Li <Ching-shih.Li@xxxxxxx>
[What]
vce ring test fails consistently during resume in s3 cycle, due to
mismatch read & write pointers.
On debug/analysis its found that rptr to be compared is not being
correctly updated/read, which leads to this failure.
Below is the failure signature:
[drm:amdgpu_vce_ring_test_ring] *ERROR* amdgpu: ring 12 test failed
[drm:amdgpu_device_ip_resume_phase2] *ERROR* resume of IP block <vce_v3_0> failed -110
[drm:amdgpu_device_resume] *ERROR* amdgpu_device_ip_resume failed (-110).
[How]
fetch rptr appropriately, meaning move its read location further down
in the code flow.
With this patch applied the s3 failure is no more seen for >5k s3 cycles,
which otherwise is pretty consistent.
Signed-off-by: Louis Li <Ching-shih.Li@xxxxxxx>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c
index c021b11..92f9d46 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c
@@ -1084,6 +1084,8 @@ int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
if (r)
return r;
+ rptr = amdgpu_ring_get_rptr(ring);
+
The rptr update is there:
uint32_t rptr = amdgpu_ring_get_rptr(ring);
Are you sure this is the root cause?
Regards,
Leo
amdgpu_ring_write(ring, VCE_CMD_END);
amdgpu_ring_commit(ring);
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
|
