[Intel-gfx] [PATCH] drm/i915/guc: Log engine resets

[Tvrtko Ursulin]

On 20/12/2021 18:34, John Harrison wrote:
> On 12/20/2021 07:00, Tvrtko Ursulin wrote:
>> On 17/12/2021 16:22, Matthew Brost wrote:
>>> On Fri, Dec 17, 2021 at 12:15:53PM +0000, Tvrtko Ursulin wrote:
>>>>
>>>> On 14/12/2021 15:07, Tvrtko Ursulin wrote:
>>>>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>>
>>>>> Log engine resets done by the GuC firmware in the similar way it is 
>>>>> done
>>>>> by the execlists backend.
>>>>>
>>>>> This way we have notion of where the hangs are before the GuC gains
>>>>> support for proper error capture.
>>>>
>>>> Ping - any interest to log this info?
>>>>
>>>> All there currently is a non-descriptive "[drm] GPU HANG: ecode
>>>> 12:0:00000000".
>>>>
>>>
>>> Yea, this could be helpful. One suggestion below.
>>>
>>>> Also, will GuC be reporting the reason for the engine reset at any 
>>>> point?
>>>>
>>>
>>> We are working on the error state capture, presumably the registers will
>>> give a clue what caused the hang.
>>>
>>> As for the GuC providing a reason, that isn't defined in the interface
>>> but that is decent idea to provide a hint in G2H what the issue was. Let
>>> me run that by the i915 GuC developers / GuC firmware team and see what
>>> they think.
>>>
> The GuC does not do any hang analysis. So as far as GuC is concerned, 
> the reason is pretty much always going to be pre-emption timeout. There 
> are a few ways the pre-emption itself could be triggered but basically, 
> if GuC resets an active context then it is because it did not pre-empt 
> quickly enough when requested.
> 
> 
>>>> Regards,
>>>>
>>>> Tvrtko
>>>>
>>>>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>> Cc: Matthew Brost <matthew.brost at intel.com>
>>>>> Cc: John Harrison <John.C.Harrison at Intel.com>
>>>>> ---
>>>>>    drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c | 12 +++++++++++-
>>>>>    1 file changed, 11 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c 
>>>>> b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
>>>>> index 97311119da6f..51512123dc1a 100644
>>>>> --- a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
>>>>> +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
>>>>> @@ -11,6 +11,7 @@
>>>>>    #include "gt/intel_context.h"
>>>>>    #include "gt/intel_engine_pm.h"
>>>>>    #include "gt/intel_engine_heartbeat.h"
>>>>> +#include "gt/intel_engine_user.h"
>>>>>    #include "gt/intel_gpu_commands.h"
>>>>>    #include "gt/intel_gt.h"
>>>>>    #include "gt/intel_gt_clock_utils.h"
>>>>> @@ -3934,9 +3935,18 @@ static void capture_error_state(struct 
>>>>> intel_guc *guc,
>>>>>    {
>>>>>        struct intel_gt *gt = guc_to_gt(guc);
>>>>>        struct drm_i915_private *i915 = gt->i915;
>>>>> -    struct intel_engine_cs *engine = 
>>>>> __context_to_physical_engine(ce);
>>>>> +    struct intel_engine_cs *engine = ce->engine;
>>>>>        intel_wakeref_t wakeref;
>>>>> +    if (intel_engine_is_virtual(engine)) {
>>>>> +        drm_notice(&i915->drm, "%s class, engines 0x%x; GuC engine 
>>>>> reset\n",
>>>>> +               intel_engine_class_repr(engine->class),
>>>>> +               engine->mask);
>>>>> +        engine = guc_virtual_get_sibling(engine, 0);
>>>>> +    } else {
>>>>> +        drm_notice(&i915->drm, "%s GuC engine reset\n", 
>>>>> engine->name);
>>>
>>> Probably include the guc_id of the context too then?
>>
>> Is the guc id stable and useful on its own - who would be the user?
> The GuC id is the only thing that matters when trying to correlate KMD 
> activity with a GuC log. So while it might not be of any use or interest 
> to an end user, it is extremely important and useful to a kernel 
> developer attempting to debug an issue. And that includes bug reports 
> from end users that are hard to repro given that the standard error 
> capture will include the GuC log.

On the topic of GuC log - is there a tool in IGT (or will be) which will 
parse the bit saved in the error capture or how is that supposed to be used?

> Also, note that GuC really resets contexts rather than engines. What it 
> reports back to i915 on a reset is simply the GuC id of the context. It 
> is up to i915 to work back from that to determine engine 
> instances/classes if required. And in the case of a virtual context, it 
> is impossible to extract the actual instance number. So your above print 
> about resetting all instances within the virtual engine mask is 
> incorrect/misleading. The reset would have been applied to one and only 
> one of those engines. If you really need to know exactly which engine 
> was poked, you need to look inside the GuC log.

I think I understood that part. :) It wasn't my intent to imply in the 
message multiple engines have been reset, but in the case of veng, log 
the class and mask and the fact there was an engine reset (singular). 
Clearer message can probably be written.

> However, the follow up point is to ask why you need to report the exact 
> class/instance? The end user doesn't care about which specific engine 
> got reset. They only care that their context was reset. Even a KMD 
> developer doesn't really care unless the concern is about a hardware bug 
> rather than a software bug.

I was simply aligning both backends to log as similar information as 
possible. Information is there, just not logged.

Concerning the wider topic, my thinking is end user is mainly interested 
to know there are any engine resets happening (to tie with the 
experience of UI/video glitching or whatever). Going for deeper analysis 
than that is probably beyond the scope of the kernel log and indeed 
error capture territory.

> My view is that the current message is indeed woefully uninformative. 
> However, it is more important to be reporting context identification 
> than engine instances. So sure, add the engine instance description but 
> also add something specific to the ce as well. Ideally (for me) the GuC 
> id and maybe something else that uniquely identifies the context in KMD 
> land for when not using GuC?

Not sure we need to go that far at this level, but even if we do it 
could be a follow up to add new data to both backends. Not sure yet I 
care enough to drive this. My patch was simply a reaction to noticing 
there is zero information currently logged while debugging some DG2 hangs.

Regards,

Tvrtko