Skip to contentSkip to conversation

drm/amdgpu/sriov: give 8s for recover vram under RUNTIME

PATCH In review Emily Deng 3 participants 6 messages Aug 8, 2018 – Aug 9, 2018 4 messages not in the archive

Subsystems (2)
lore
Share & tools
lore

Emily Deng

author
#0 lore raw

PATCH 1/1 drm/amdgpu/sriov: give 8s for recover vram under RUNTIME

Modify the commit message

Extend the timeout for recovering vram bos from shadows on sr-iov to cover the worst case scenario for timeslices and VFs

Under runtime, the wait fence time could be quite long when other VFs are in exclusive mode. For example, for 4 VF, every VF's exclusive timeout time is set to 3s, then the worst case is 9s. If the VF number is more than 4,then the worst case time will be longer.
The 8s is the test data, with setting to 8s, it will pass the TDR test for 1000 times.

SWDEV-161490

Change-Id Ifc32d56ca7fde01b1f4fe2b0db6959b51909008a
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -1
index 1d933db..ef82ad1 100644
@@ -3124,7 +3124,7 @@ static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
 	long tmo;
 
 	if (amdgpu_sriov_runtime(adev))
-		tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
+		tmo = msecs_to_jiffies(8000);
 	else
 		tmo = msecs_to_jiffies(100);
 
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1533694400-21433-1-git-send-email-Emily.Deng-5C7GfCeVMHo@public.gmane.org

Reviewed-by
#2 lore raw

________________________________
From: amd-gfx <amd-gfx-bounces-PD4FTy7X32lNgt0PjOBp9y5qC8QIuHrW@public.gmane.org> on behalf of Emily Deng <[email protected]>
Sent: Tuesday, August 7, 2018 10:13 PM
To: [email protected]
Cc: Deng, Emily; Liu, Monk
Subject: [PATCH] drm/amdgpu/sriov: give 8s for recover vram under RUNTIME

Modify the commit message

Extend the timeout for recovering vram bos from shadows on sr-iov to cover the worst case scenario for timeslices and VFs

Under runtime, the wait fence time could be quite long when other VFs are in exclusive mode. For example, for 4 VF, every VF's exclusive timeout time is set to 3s, then the worst case is 9s. If the VF number is more than 4,then the worst case time will be longer.
The 8s is the test data, with setting to 8s, it will pass the TDR test for 1000 times.

SWDEV-161490

Change-Id Ifc32d56ca7fde01b1f4fe2b0db6959b51909008a
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -1
index 1d933db..ef82ad1 100644
@@ -3124,7 +3124,7 @@ static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
         long tmo;
 
         if (amdgpu_sriov_runtime(adev))
-               tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
+               tmo = msecs_to_jiffies(8000);
         else
                 tmo = msecs_to_jiffies(100);
 

--
2.7.4

_______________________________________________
amd-gfx mailing list [email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#3 lore raw
In reply to a message not in the archive

Dear Deng,

Quoted 1 line · Emily Deng

Modify the commit message

I guess the line above is a leftover from some template, and can be removed?

Quoted 8 lines

Extend the timeout for recovering vram bos from shadows on sr-iov to cover the worst case scenario for timeslices and VFs

Under runtime, the wait fence time could be quite long when other VFs are in exclusive mode. For example, for 4 VF, every VF's exclusive timeout time is set to 3s, then the worst case is 9s. If the VF number is more than 4,then the worst case time will be longer.

Nit: Missing space after the comma.

How did you get to nine seconds? Isn’t it four times three, which is twelve?

Quoted 17 lines

The 8s is the test data, with setting to 8s, it will pass the TDR test for 1000 times.

SWDEV-161490

Change-Id Ifc32d56ca7fde01b1f4fe2b0db6959b51909008a
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index 1d933db..ef82ad1 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev) · line 3127

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -1
@@ -3124,7 +3124,7 @@ static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
 	long tmo;
 
 	if (amdgpu_sriov_runtime(adev))
-		tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
+		tmo = msecs_to_jiffies(8000);

Actually, I do not understand the change at all. Isn’t that a module parameter?

```
$ git grep amdgpu_lockup_timeout
drivers/gpu/drm/amd/amdgpu/amdgpu.h:extern int amdgpu_lockup_timeout;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:     if (amdgpu_lockup_timeout == 0) {
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:             amdgpu_lockup_timeout = 10000;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:             tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c:int amdgpu_lockup_timeout = 10000;
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c:module_param_named(lockup_timeout, amdgpu_lockup_timeout, int, 0444);
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c:                      timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c:  if (amdgpu_lockup_timeout == 0)
```

So, if it’s not set, the time-out is set to 10(!) seconds anyway, isn’t it? What am I missing?

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev) · line 3130

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
 	else
 		tmo = msecs_to_jiffies(100);
 

In my opinion, such time-outs need to have a big FIXME added to it, and VF’s exclusive time-outs should be drastically decreased.

Kind regards,

Paul

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 7f631492-3d79-97e7-b46a-5e928090e2c9-KUpvgZVWgV9o1qOY/[email protected]

#5 lore raw

Hi Paul,
Answers as below:

-----Original Message-----
From: Paul Menzel <[email protected]>
Sent: Wednesday, August 8, 2018 10:29 PM
To: Deng, Emily <[email protected]>; [email protected]
Cc: Liu, Monk <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/sriov: give 8s for recover vram under RUNTIME

Dear Deng,

Quoted 1 line · Emily Deng

Modify the commit message

I guess the line above is a leftover from some template, and can be removed?
[Emily] It is removed

Quoted 8 lines

Extend the timeout for recovering vram bos from shadows on sr-iov to cover the worst case scenario for timeslices and VFs

Under runtime, the wait fence time could be quite long when other VFs are in exclusive mode. For example, for 4 VF, every VF's exclusive timeout time is set to 3s, then the worst case is 9s. If the VF number is more than 4,then the worst case time will be longer.

Nit: Missing space after the comma.

How did you get to nine seconds? Isn’t it four times three, which is
twelve?
[Emily] It is 3 times three, not 4, as the VF only need to wait the other three VF's exclusive timeout.

Quoted 17 lines

The 8s is the test data, with setting to 8s, it will pass the TDR test for 1000 times.

SWDEV-161490

Change-Id Ifc32d56ca7fde01b1f4fe2b0db6959b51909008a
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index 1d933db..ef82ad1 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev) · line 3127

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -1
@@ -3124,7 +3124,7 @@ static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
 	long tmo;
 
 	if (amdgpu_sriov_runtime(adev))
-		tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
+		tmo = msecs_to_jiffies(8000);

Actually, I do not understand the change at all. Isn’t that a module parameter?

```
$ git grep amdgpu_lockup_timeout
drivers/gpu/drm/amd/amdgpu/amdgpu.h:extern int amdgpu_lockup_timeout;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:     if (amdgpu_lockup_timeout == 0) {
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:             amdgpu_lockup_timeout = 10000;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:             tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c:int amdgpu_lockup_timeout = 10000;
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c:module_param_named(lockup_timeout, amdgpu_lockup_timeout, int, 0444);
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c:                      timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c:  if (amdgpu_lockup_timeout == 0)
```

So, if it’s not set, the time-out is set to 10(!) seconds anyway, isn’t it? What am I missing?

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev) · line 3130

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
 	else
 		tmo = msecs_to_jiffies(100);
 

In my opinion, such time-outs need to have a big FIXME added to it, and VF’s exclusive time-outs should be drastically decreased.

Kind regards,

Paul

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): CY4PR12MB1125A65D7368C804EEB1A3288F250-rpdhrqHFk07v2MZdTKcfDgdYzm3356FpvxpqHgZTriW3zl9H0oFU5g@public.gmane.org

#7 lore raw

Dear Emily,

Thank you for your reply.

Am 09.08.2018 um 04:11 schrieb Deng, Emily:

Quoted 2 lines

Hi Paul,
Answers as below:

It’d be great if you had an mail user agent (email program) that supports quoting. Maybe you could configure Outlook(?) that way? That would tremendously useful for the mailing list archives too.

Quoted 8 lines

-----Original Message-----
From: Paul Menzel <[email protected]>
Sent: Wednesday, August 8, 2018 10:29 PM
To: Deng, Emily <[email protected]>; [email protected]
Cc: Liu, Monk <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/sriov: give 8s for recover vram under RUNTIME

On 08/08/18 04:13, Emily Deng wrote:

[…]

Quoted 9 lines
Quoted 8 lines

Extend the timeout for recovering vram bos from shadows on sr-iov to cover the worst case scenario for timeslices and VFs

Under runtime, the wait fence time could be quite long when other VFs are in exclusive mode. For example, for 4 VF, every VF's exclusive timeout time is set to 3s, then the worst case is 9s. If the VF number is more than 4,then the worst case time will be longer.

Quoted 6 lines

Nit: Missing space after the comma.

How did you get to nine seconds? Isn’t it four times three, which is twelve?
[Emily] It is 3 times three, not 4, as the VF only need to wait the other three VF's exclusive timeout.

Thank you, I didn’t know that.

(It looks like you missed my inline comments for the change itself below.)

Quoted 41 lines
Quoted 17 lines

The 8s is the test data, with setting to 8s, it will pass the TDR test for 1000 times.

SWDEV-161490

Change-Id Ifc32d56ca7fde01b1f4fe2b0db6959b51909008a
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index 1d933db..ef82ad1 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev) · line 3127

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -1
@@ -3124,7 +3124,7 @@ static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
  	long tmo;
 
  	if (amdgpu_sriov_runtime(adev))
-		tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
+		tmo = msecs_to_jiffies(8000);

Actually, I do not understand the change at all. Isn’t that a module parameter?

```
$ git grep amdgpu_lockup_timeout
drivers/gpu/drm/amd/amdgpu/amdgpu.h:extern int amdgpu_lockup_timeout;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:     if (amdgpu_lockup_timeout == 0) {
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:             amdgpu_lockup_timeout = 10000;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:             tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c:int amdgpu_lockup_timeout = 10000;
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c:module_param_named(lockup_timeout, amdgpu_lockup_timeout, int, 0444);
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c:                      timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c:  if (amdgpu_lockup_timeout == 0)
```

So, if it’s not set, the time-out is set to 10(!) seconds anyway, isn’t it? What am I missing?

Quoted 7 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev) · line 3130

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
  	else
  		tmo = msecs_to_jiffies(100);
 

In my opinion, such time-outs need to have a big FIXME added to it, and VF’s exclusive time-outs should be drastically decreased.

Kind regards,

Paul
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 5d61f338-cef6-fce3-6c29-ff4c503855aa-KUpvgZVWgV9o1qOY/[email protected]

#9 lore raw

Hi Menzel,
Thanks for your advice, will have a try.

Best wishes
Emily Deng

-----Original Message-----
From: Paul Menzel <[email protected]>
Sent: Thursday, August 9, 2018 2:47 PM
To: Deng, Emily <[email protected]>
Cc: [email protected]; Liu, Monk <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/sriov: give 8s for recover vram under RUNTIME

Dear Emily,

Thank you for your reply.

Am 09.08.2018 um 04:11 schrieb Deng, Emily:

Quoted 2 lines

Hi Paul,
Answers as below:

It’d be great if you had an mail user agent (email program) that supports quoting. Maybe you could configure Outlook(?) that way? That would tremendously useful for the mailing list archives too.

Quoted 9 lines

-----Original Message-----
From: Paul Menzel <[email protected]>
Sent: Wednesday, August 8, 2018 10:29 PM
To: Deng, Emily <[email protected]>; [email protected]
Cc: Liu, Monk <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/sriov: give 8s for recover vram under RUNTIME

On 08/08/18 04:13, Emily Deng wrote:

[…]

Quoted 8 lines
Quoted 7 lines

Extend the timeout for recovering vram bos from shadows on sr-iov to cover the worst case scenario for timeslices and VFs

Under runtime, the wait fence time could be quite long when other VFs are in exclusive mode. For example, for 4 VF, every VF's exclusive timeout time is set to 3s, then the worst case is 9s. If the VF number is more than 4,then the worst case time will be longer.

Quoted 6 lines

Nit: Missing space after the comma.

How did you get to nine seconds? Isn’t it four times three, which is twelve?
[Emily] It is 3 times three, not 4, as the VF only need to wait the other three VF's exclusive timeout.

Thank you, I didn’t know that.

(It looks like you missed my inline comments for the change itself below.)

Quoted 42 lines
Quoted 18 lines

The 8s is the test data, with setting to 8s, it will pass the TDR test for 1000 times.

SWDEV-161490

Change-Id Ifc32d56ca7fde01b1f4fe2b0db6959b51909008a
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index 1d933db..ef82ad1 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev) · line 3127

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -1
@@ -3124,7 +3124,7 @@ static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
  	long tmo;
 
  	if (amdgpu_sriov_runtime(adev))
-		tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
+		tmo = msecs_to_jiffies(8000);

Actually, I do not understand the change at all. Isn’t that a module parameter?

```

$ git grep amdgpu_lockup_timeout
drivers/gpu/drm/amd/amdgpu/amdgpu.h:extern int amdgpu_lockup_timeout;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:     if (amdgpu_lockup_timeout == 0) {
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:             amdgpu_lockup_timeout = 10000;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c:             tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c:int amdgpu_lockup_timeout = 
10000; drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c:module_param_named(lockup_timeout, amdgpu_lockup_timeout, int, 0444);
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c:                      timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c:  if (amdgpu_lockup_timeout == 
0) ```

So, if it’s not set, the time-out is set to 10(!) seconds anyway, isn’t it? What am I missing?

Quoted 7 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev) · line 3130

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
  	else
  		tmo = msecs_to_jiffies(100);
 

In my opinion, such time-outs need to have a big FIXME added to it, and VF’s exclusive time-outs should be drastically decreased.

Kind regards,

Paul
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx