[PATCH v2] drm/xe/ufence: Flush xe ordered_wq in case of ufence timeout

[Nirmoy Das]

On 10/25/2024 8:34 PM, Matthew Brost wrote:
> On Fri, Oct 25, 2024 at 11:27:55AM -0700, John Harrison wrote:
>> On 10/25/2024 09:03, Nirmoy Das wrote:
>>> On 10/24/2024 6:32 PM, Jani Nikula wrote:
>>>> On Thu, 24 Oct 2024, Nirmoy Das <nirmoy.das at intel.com> wrote:
>>>>> Flush xe ordered_wq in case of ufence timeout which is observed
>>>>> on LNL and that points to the recent scheduling issue with E-cores.
>>>>>
>>>>> This is similar to the recent fix:
>>>>> commit e51527233804 ("drm/xe/guc/ct: Flush g2h worker in case of g2h
>>>>> response timeout") and should be removed once there is E core
>>>>> scheduling fix.
>>>>>
>>>>> v2: Add platform check(Himal)
>>>>>      s/__flush_workqueue/flush_workqueue(Jani)
>>>>>
>>>>> Cc: Badal Nilawar <badal.nilawar at intel.com>
>>>>> Cc: Jani Nikula <jani.nikula at intel.com>
>>>>> Cc: Matthew Auld <matthew.auld at intel.com>
>>>>> Cc: John Harrison <John.C.Harrison at Intel.com>
>>>>> Cc: Himal Prasad Ghimiray <himal.prasad.ghimiray at intel.com>
>>>>> Cc: Lucas De Marchi <lucas.demarchi at intel.com>
>>>>> Cc: <stable at vger.kernel.org> # v6.11+
>>>>> Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/2754
>>>>> Suggested-by: Matthew Brost <matthew.brost at intel.com>
>>>>> Signed-off-by: Nirmoy Das <nirmoy.das at intel.com>
>>>>> Reviewed-by: Matthew Brost <matthew.brost at intel.com>
>>>>> ---
>>>>>   drivers/gpu/drm/xe/xe_wait_user_fence.c | 14 ++++++++++++++
>>>>>   1 file changed, 14 insertions(+)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/xe/xe_wait_user_fence.c b/drivers/gpu/drm/xe/xe_wait_user_fence.c
>>>>> index f5deb81eba01..78a0ad3c78fe 100644
>>>>> --- a/drivers/gpu/drm/xe/xe_wait_user_fence.c
>>>>> +++ b/drivers/gpu/drm/xe/xe_wait_user_fence.c
>>>>> @@ -13,6 +13,7 @@
>>>>>   #include "xe_device.h"
>>>>>   #include "xe_gt.h"
>>>>>   #include "xe_macros.h"
>>>>> +#include "compat-i915-headers/i915_drv.h"
>>>> Sorry, you just can't use this in xe core. At all. Not even a little
>>>> bit. It's purely for i915 display compat code.
>>>>
>>>> If you need it for the LNL platform check, you need to use:
>>>>
>>>> 	xe->info.platform == XE_LUNARLAKE
>>> Will do that. That macro looked odd but I didn't know a better way.
>>>
>>>> Although platform checks in xe code are generally discouraged.
>>> This issue unfortunately depending on platform instead of graphics IP.
>> But isn't this issue dependent upon the CPU platform not the graphics
>> platform? As in, a DG2 card plugged in to a LNL host will also have this
>> issue. So testing any graphics related value is technically incorrect.


Haven't thought about. LNL only has x8 PCIe lanes shared between NVME and other IOs but thunderbolt based eGPU should be easily doable.

I think I could do "if (boot_cpu_data.x86_vfm == INTEL_LUNARLAKE_M)" instead.

>>
> This is a good point, maybe for now we blindly do this regardless of
> platform. It is basically harmless to do this after a timeout... Also a
> warning message if we can detect this fixed the timeout for CI purposes.

I am open to this as well. Please let me know which one should be a better solution here.


Regards,

Nirmoy

>
> Matt
>
>> John.
>>
>>>
>>> Thanks,
>>>
>>> Nirmoy
>>>
>>>> BR,
>>>> Jani.
>>>>
>>>>
>>>>
>>>>>   #include "xe_exec_queue.h"
>>>>>   static int do_compare(u64 addr, u64 value, u64 mask, u16 op)
>>>>> @@ -155,6 +156,19 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev, void *data,
>>>>>   		}
>>>>>   		if (!timeout) {
>>>>> +			if (IS_LUNARLAKE(xe)) {
>>>>> +				/*
>>>>> +				 * This is analogous to e51527233804 ("drm/xe/guc/ct: Flush g2h
>>>>> +				 * worker in case of g2h response timeout")
>>>>> +				 *
>>>>> +				 * TODO: Drop this change once workqueue scheduling delay issue is
>>>>> +				 * fixed on LNL Hybrid CPU.
>>>>> +				 */
>>>>> +				flush_workqueue(xe->ordered_wq);
>>>>> +				err = do_compare(addr, args->value, args->mask, args->op);
>>>>> +				if (err <= 0)
>>>>> +					break;
>>>>> +			}
>>>>>   			err = -ETIME;
>>>>>   			break;
>>>>>   		}