[Intel-gfx] [PATCH 4/7] drm/i915/guc: Don't hog IRQs when destroying contexts

[Tvrtko Ursulin]

On 17/12/2021 11:06, Tvrtko Ursulin wrote:
> On 14/12/2021 17:04, Matthew Brost wrote:
>> From: John Harrison <John.C.Harrison at Intel.com>
>>
>> While attempting to debug a CT deadlock issue in various CI failures
>> (most easily reproduced with gem_ctx_create/basic-files), I was seeing
>> CPU deadlock errors being reported. This were because the context
>> destroy loop was blocking waiting on H2G space from inside an IRQ
>> spinlock. There no was deadlock as such, it's just that the H2G queue
>> was full of context destroy commands and GuC was taking a long time to
>> process them. However, the kernel was seeing the large amount of time
>> spent inside the IRQ lock as a dead CPU. Various Bad Things(tm) would
>> then happen (heartbeat failures, CT deadlock errors, outstanding H2G
>> WARNs, etc.).
>>
>> Re-working the loop to only acquire the spinlock around the list
>> management (which is all it is meant to protect) rather than the
>> entire destroy operation seems to fix all the above issues.
>>
>> v2:
>>   (John Harrison)
>>    - Fix typo in comment message
>>
>> Signed-off-by: John Harrison <John.C.Harrison at Intel.com>
>> Signed-off-by: Matthew Brost <matthew.brost at intel.com>
>> Reviewed-by: Matthew Brost <matthew.brost at intel.com>
>> ---
>>   .../gpu/drm/i915/gt/uc/intel_guc_submission.c | 45 ++++++++++++-------
>>   1 file changed, 28 insertions(+), 17 deletions(-)
>>
>> 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 36c2965db49b..96fcf869e3ff 100644
>> --- a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
>> +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
>> @@ -2644,7 +2644,6 @@ static inline void guc_lrc_desc_unpin(struct 
>> intel_context *ce)
>>       unsigned long flags;
>>       bool disabled;
>> -    lockdep_assert_held(&guc->submission_state.lock);
>>       GEM_BUG_ON(!intel_gt_pm_is_awake(gt));
>>       GEM_BUG_ON(!lrc_desc_registered(guc, ce->guc_id.id));
>>       GEM_BUG_ON(ce != __get_context(guc, ce->guc_id.id));
>> @@ -2660,7 +2659,7 @@ static inline void guc_lrc_desc_unpin(struct 
>> intel_context *ce)
>>       }
>>       spin_unlock_irqrestore(&ce->guc_state.lock, flags);
>>       if (unlikely(disabled)) {
>> -        __release_guc_id(guc, ce);
>> +        release_guc_id(guc, ce);
>>           __guc_context_destroy(ce);
>>           return;
>>       }
>> @@ -2694,36 +2693,48 @@ static void __guc_context_destroy(struct 
>> intel_context *ce)
>>   static void guc_flush_destroyed_contexts(struct intel_guc *guc)
>>   {
>> -    struct intel_context *ce, *cn;
>> +    struct intel_context *ce;
>>       unsigned long flags;
>>       GEM_BUG_ON(!submission_disabled(guc) &&
>>              guc_submission_initialized(guc));
>> -    spin_lock_irqsave(&guc->submission_state.lock, flags);
>> -    list_for_each_entry_safe(ce, cn,
>> -                 &guc->submission_state.destroyed_contexts,
>> -                 destroyed_link) {
>> -        list_del_init(&ce->destroyed_link);
>> -        __release_guc_id(guc, ce);
>> +    while (!list_empty(&guc->submission_state.destroyed_contexts)) {
> 
> Are lockless false negatives a concern here - I mean this thread not 
> seeing something just got added to the list?
> 
>> +        spin_lock_irqsave(&guc->submission_state.lock, flags);
>> +        ce = 
>> list_first_entry_or_null(&guc->submission_state.destroyed_contexts,
>> +                          struct intel_context,
>> +                          destroyed_link);
>> +        if (ce)
>> +            list_del_init(&ce->destroyed_link);
>> +        spin_unlock_irqrestore(&guc->submission_state.lock, flags);
>> +
>> +        if (!ce)
>> +            break;
>> +
>> +        release_guc_id(guc, ce);
> 
> This looks suboptimal and in conflict with this part of the commit message:
> 
> """
>   Re-working the loop to only acquire the spinlock around the list
>   management (which is all it is meant to protect) rather than the
>   entire destroy operation seems to fix all the above issues.
> """
> 
> Because you end up doing:
> 
> ... loop ...
>    spin_lock_irqsave(&guc->submission_state.lock, flags);
>    list_del_init(&ce->destroyed_link);
>    spin_unlock_irqrestore(&guc->submission_state.lock, flags);
> 
>    release_guc_id, which calls:
>      spin_lock_irqsave(&guc->submission_state.lock, flags);
>      __release_guc_id(guc, ce);
>      spin_unlock_irqrestore(&guc->submission_state.lock, flags);
> 
> So a) the lock seems to be protecting more than just list management, or 
> release_guc_if is wrong, and b) the loop ends up with highly 
> questionable hammering on the lock.
> 
> Is there any point to this part of the patch? Or the only business end 
> of the patch is below:
> 
>>           __guc_context_destroy(ce);
>>       }
>> -    spin_unlock_irqrestore(&guc->submission_state.lock, flags);
>>   }
>>   static void deregister_destroyed_contexts(struct intel_guc *guc)
>>   {
>> -    struct intel_context *ce, *cn;
>> +    struct intel_context *ce;
>>       unsigned long flags;
>> -    spin_lock_irqsave(&guc->submission_state.lock, flags);
>> -    list_for_each_entry_safe(ce, cn,
>> -                 &guc->submission_state.destroyed_contexts,
>> -                 destroyed_link) {
>> -        list_del_init(&ce->destroyed_link);
>> +    while (!list_empty(&guc->submission_state.destroyed_contexts)) {
>> +        spin_lock_irqsave(&guc->submission_state.lock, flags);
>> +        ce = 
>> list_first_entry_or_null(&guc->submission_state.destroyed_contexts,
>> +                          struct intel_context,
>> +                          destroyed_link);
>> +        if (ce)
>> +            list_del_init(&ce->destroyed_link);
>> +        spin_unlock_irqrestore(&guc->submission_state.lock, flags);
>> +
>> +        if (!ce)
>> +            break;
>> +
>>           guc_lrc_desc_unpin(ce);
> 
> Here?
> 
> Not wanting/needing to nest ce->guc_state.lock under 
> guc->submission_state.lock, and call the CPU cycle expensive 
> deregister_context?
> 
> 1)
> Could you unlink en masse, under the assumption destroyed contexts are 
> not reachable from anywhere else at this point, so under a single lock 
> hold?
> 
> 2)
> But then you also end up with guc_lrc_desc_unpin calling 
> __release_guc_id, which when called by release_guc_id does take 
> guc->submission_state.lock and here it does not. Is it then clear which 
> operations inside __release_guc_id need the lock? Bitmap or IDA?

Ah no, with 2nd point I missed you changed guc_lrc_desc_unpin to call 
release_guc_id.

Question on the merit of change in guc_flush_destroyed_contexts remains, 
and also whether at both places you could do group unlink (one lock 
hold), put on a private list, and then unpin/deregister.

Regards,

Tvrtko