Re: [Intel-gfx] [PATCH v4 2/7] drm/i915: Fix up locking around dumping requests lists

[John Harrison <john.c.harrison@intel.com>]

On 1/25/2023 10:12, Tvrtko Ursulin wrote:
> On 25/01/2023 18:00, John Harrison wrote:
>> On 1/24/2023 06:40, Tvrtko Ursulin wrote:
>>> On 20/01/2023 23:28, John.C.Harrison@Intel.com wrote:
>>>> From: John Harrison <John.C.Harrison@Intel.com>
>>>>
>>>> The debugfs dump of requests was confused about what state requires
>>>> the execlist lock versus the GuC lock. There was also a bunch of
>>>> duplicated messy code between it and the error capture code.
>>>>
>>>> So refactor the hung request search into a re-usable function. And
>>>> reduce the span of the execlist state lock to only the execlist
>>>> specific code paths. In order to do that, also move the report of hold
>>>> count (which is an execlist only concept) from the top level dump
>>>> function to the lower level execlist specific function. Also, move the
>>>> execlist specific code into the execlist source file.
>>>>
>>>> v2: Rename some functions and move to more appropriate files 
>>>> (Daniele).
>>>
>>> Continuing from yesterday where you pointed out 2/7 exists, after I 
>>> declared capitulation on 1/7.. I think this refactor makes sense and 
>>> definitely improves things a lot.
>>>
>>> On the high level I am only unsure if the patch split could be 
>>> improved. There seem to be three separate things, correct me if I 
>>> missed something:
>>>
>>> 1) Locking fix in intel_guc_find_hung_context
>> This is the change already it's own patch - #1/7. Can't really split 
>> that one up any further. Changing the internal GuC code requires 
>> changing the external common code to match.
>>
>>> 2) Ref counting change throughout
>>> 3) Locking refactor / helper consolidation
>> These two being the changes in this patch - #2/7, yes?
>>
>> The problem is that the reference counting fixes can only be done 
>> once the code has been refactored/reordered. And the refactor/reorder 
>> can only be done if the reference counting is fixed. I guess there 
>> would be some way to do the re-order first but it would require 
>> making even more of a mess of the spinlock activity to keep it all 
>> correct around that intermediate stage. So I don't think it would 
>> noticeably simplify the patch.
>>
>>>
>>> (Or 2 and 3 swapped around, not sure.)
>>>
>>> That IMO might be a bit easier to read because first patch wouldn't 
>>> have two logical changes in it. Maybe easier to backport too if it 
>>> comes to that?
>> I'm not seeing 'two logical changes' in the first patch. Patch #1 
>> fixes the reference counting of finding the hung request. That 
>> involves adding a reference count internally within the spinlock on 
>> the GuC side and moving the external reference count to within the 
>> spinlock on the execlist side and then doing a put in all cases. That 
>> really is a single change. It can't be split without either a) 
>> introducing a get/put mis-match bug or b) making the code really ugly 
>> as an intermediate (while still leaving one or other side broken).
>
> I was thinking this part is wholy standalone:
>
> @@ -4820,6 +4821,8 @@ void intel_guc_find_hung_context(struct 
> intel_engine_cs *engine)
>
>      xa_lock_irqsave(&guc->context_lookup, flags);
>      xa_for_each(&guc->context_lookup, index, ce) {
> +        bool found;
> +
>          if (!kref_get_unless_zero(&ce->ref))
>              continue;
>
> @@ -4836,10 +4839,18 @@ void intel_guc_find_hung_context(struct 
> intel_engine_cs *engine)
>                  goto next;
>          }
>
> +        found = false;
> +        spin_lock(&ce->guc_state.lock);
>          list_for_each_entry(rq, &ce->guc_state.requests, sched.link) {
>              if (i915_test_request_state(rq) != I915_REQUEST_ACTIVE)
>                  continue;
>
> +            found = true;
> +            break;
> +        }
> +        spin_unlock(&ce->guc_state.lock);
> +
> +        if (found) {
>              intel_engine_set_hung_context(engine, ce);
>
>              /* Can only cope with one hang at a time... */
> @@ -4847,6 +4858,7 @@ void intel_guc_find_hung_context(struct 
> intel_engine_cs *engine)
>              xa_lock(&guc->context_lookup);
>              goto done;
>          }
> +
>  next:
>          intel_context_put(ce);
>          xa_lock(&guc->context_lookup);
>
> Am I missing something?
Doh.

Yes, I guess that part is stand alone. I was getting myself confused and 
thinking that was part of moving a get inside the spinlock. But you are 
right, that part is just about using the correct spinlock for that loop.

So yeah, I can split that chunk out to a separate patch. But that is 
splitting patch #1 into #1a and #1b. It doesn't help with patch #2. 
Which is the one I though you were complaining about being too complex. 
Which it is :(. But I'm really not seeing anyway to simplify it given 
how much of a mess the code is in.

John.


>
> Regards,
>
> Tvrtko
>
>>
>> John.
>>
>>>
>>> On the low level it all looks fine to me - hopefully Daniele can do 
>>> a detailed pass.
>>>
>>> Regards,
>>>
>>> Tvrtko
>>>
>>> P.S. Only that intel_context_find_active_request_get hurts my eyes, 
>>> and inflates the diff. I wouldn't rename it but if you guys insist 
>>> okay.
>>>
>>>> Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
>>>> ---
>>>>   drivers/gpu/drm/i915/gt/intel_engine.h        |  4 +-
>>>>   drivers/gpu/drm/i915/gt/intel_engine_cs.c     | 74 
>>>> +++++++++----------
>>>>   .../drm/i915/gt/intel_execlists_submission.c  | 27 +++++++
>>>>   .../drm/i915/gt/intel_execlists_submission.h  |  4 +
>>>>   drivers/gpu/drm/i915/i915_gpu_error.c         | 26 +------
>>>>   5 files changed, 73 insertions(+), 62 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/i915/gt/intel_engine.h 
>>>> b/drivers/gpu/drm/i915/gt/intel_engine.h
>>>> index 0e24af5efee9c..b58c30ac8ef02 100644
>>>> --- a/drivers/gpu/drm/i915/gt/intel_engine.h
>>>> +++ b/drivers/gpu/drm/i915/gt/intel_engine.h
>>>> @@ -250,8 +250,8 @@ void intel_engine_dump_active_requests(struct 
>>>> list_head *requests,
>>>>   ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine,
>>>>                      ktime_t *now);
>>>>   -struct i915_request *
>>>> -intel_engine_execlist_find_hung_request(struct intel_engine_cs 
>>>> *engine);
>>>> +void intel_engine_get_hung_entity(struct intel_engine_cs *engine,
>>>> +                  struct intel_context **ce, struct i915_request 
>>>> **rq);
>>>>     u32 intel_engine_context_size(struct intel_gt *gt, u8 class);
>>>>   struct intel_context *
>>>> diff --git a/drivers/gpu/drm/i915/gt/intel_engine_cs.c 
>>>> b/drivers/gpu/drm/i915/gt/intel_engine_cs.c
>>>> index fbc0a81617e89..1d77e27801bce 100644
>>>> --- a/drivers/gpu/drm/i915/gt/intel_engine_cs.c
>>>> +++ b/drivers/gpu/drm/i915/gt/intel_engine_cs.c
>>>> @@ -2114,17 +2114,6 @@ static void print_request_ring(struct 
>>>> drm_printer *m, struct i915_request *rq)
>>>>       }
>>>>   }
>>>>   -static unsigned long list_count(struct list_head *list)
>>>> -{
>>>> -    struct list_head *pos;
>>>> -    unsigned long count = 0;
>>>> -
>>>> -    list_for_each(pos, list)
>>>> -        count++;
>>>> -
>>>> -    return count;
>>>> -}
>>>> -
>>>>   static unsigned long read_ul(void *p, size_t x)
>>>>   {
>>>>       return *(unsigned long *)(p + x);
>>>> @@ -2216,11 +2205,11 @@ void 
>>>> intel_engine_dump_active_requests(struct list_head *requests,
>>>>       }
>>>>   }
>>>>   -static void engine_dump_active_requests(struct intel_engine_cs 
>>>> *engine, struct drm_printer *m)
>>>> +static void engine_dump_active_requests(struct intel_engine_cs 
>>>> *engine,
>>>> +                    struct drm_printer *m)
>>>>   {
>>>> +    struct intel_context *hung_ce = NULL;
>>>>       struct i915_request *hung_rq = NULL;
>>>> -    struct intel_context *ce;
>>>> -    bool guc;
>>>>         /*
>>>>        * No need for an engine->irq_seqno_barrier() before the 
>>>> seqno reads.
>>>> @@ -2229,29 +2218,20 @@ static void 
>>>> engine_dump_active_requests(struct intel_engine_cs *engine, struct d
>>>>        * But the intention here is just to report an instantaneous 
>>>> snapshot
>>>>        * so that's fine.
>>>>        */
>>>> - lockdep_assert_held(&engine->sched_engine->lock);
>>>> +    intel_engine_get_hung_entity(engine, &hung_ce, &hung_rq);
>>>>         drm_printf(m, "\tRequests:\n");
>>>>   -    guc = intel_uc_uses_guc_submission(&engine->gt->uc);
>>>> -    if (guc) {
>>>> -        ce = intel_engine_get_hung_context(engine);
>>>> -        if (ce)
>>>> -            hung_rq = intel_context_find_active_request_get(ce);
>>>> -    } else {
>>>> -        hung_rq = intel_engine_execlist_find_hung_request(engine);
>>>> -        if (hung_rq)
>>>> -            hung_rq = i915_request_get_rcu(hung_rq);
>>>> -    }
>>>> -
>>>>       if (hung_rq)
>>>>           engine_dump_request(hung_rq, m, "\t\thung");
>>>> +    else if (hung_ce)
>>>> +        drm_printf(m, "\t\tGot hung ce but no hung rq!\n");
>>>>   -    if (guc)
>>>> +    if (intel_uc_uses_guc_submission(&engine->gt->uc))
>>>>           intel_guc_dump_active_requests(engine, hung_rq, m);
>>>>       else
>>>> - intel_engine_dump_active_requests(&engine->sched_engine->requests,
>>>> -                          hung_rq, m);
>>>> +        intel_execlist_dump_active_requests(engine, hung_rq, m);
>>>> +
>>>>       if (hung_rq)
>>>>           i915_request_put(hung_rq);
>>>>   }
>>>> @@ -2263,7 +2243,6 @@ void intel_engine_dump(struct intel_engine_cs 
>>>> *engine,
>>>>       struct i915_gpu_error * const error = &engine->i915->gpu_error;
>>>>       struct i915_request *rq;
>>>>       intel_wakeref_t wakeref;
>>>> -    unsigned long flags;
>>>>       ktime_t dummy;
>>>>         if (header) {
>>>> @@ -2300,13 +2279,8 @@ void intel_engine_dump(struct 
>>>> intel_engine_cs *engine,
>>>>              i915_reset_count(error));
>>>>       print_properties(engine, m);
>>>>   - spin_lock_irqsave(&engine->sched_engine->lock, flags);
>>>>       engine_dump_active_requests(engine, m);
>>>>   -    drm_printf(m, "\tOn hold?: %lu\n",
>>>> - list_count(&engine->sched_engine->hold));
>>>> - spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
>>>> -
>>>>       drm_printf(m, "\tMMIO base:  0x%08x\n", engine->mmio_base);
>>>>       wakeref = intel_runtime_pm_get_if_in_use(engine->uncore->rpm);
>>>>       if (wakeref) {
>>>> @@ -2352,8 +2326,7 @@ intel_engine_create_virtual(struct 
>>>> intel_engine_cs **siblings,
>>>>       return siblings[0]->cops->create_virtual(siblings, count, 
>>>> flags);
>>>>   }
>>>>   -struct i915_request *
>>>> -intel_engine_execlist_find_hung_request(struct intel_engine_cs 
>>>> *engine)
>>>> +static struct i915_request 
>>>> *engine_execlist_find_hung_request(struct intel_engine_cs *engine)
>>>>   {
>>>>       struct i915_request *request, *active = NULL;
>>>>   @@ -2405,6 +2378,33 @@ 
>>>> intel_engine_execlist_find_hung_request(struct intel_engine_cs 
>>>> *engine)
>>>>       return active;
>>>>   }
>>>>   +void intel_engine_get_hung_entity(struct intel_engine_cs *engine,
>>>> +                  struct intel_context **ce, struct i915_request 
>>>> **rq)
>>>> +{
>>>> +    unsigned long flags;
>>>> +
>>>> +    *ce = intel_engine_get_hung_context(engine);
>>>> +    if (*ce) {
>>>> +        intel_engine_clear_hung_context(engine);
>>>> +
>>>> +        *rq = intel_context_find_active_request_get(*ce);
>>>> +        return;
>>>> +    }
>>>> +
>>>> +    /*
>>>> +     * Getting here with GuC enabled means it is a forced error 
>>>> capture
>>>> +     * with no actual hang. So, no need to attempt the execlist 
>>>> search.
>>>> +     */
>>>> +    if (intel_uc_uses_guc_submission(&engine->gt->uc))
>>>> +        return;
>>>> +
>>>> + spin_lock_irqsave(&engine->sched_engine->lock, flags);
>>>> +    *rq = engine_execlist_find_hung_request(engine);
>>>> +    if (*rq)
>>>> +        *rq = i915_request_get_rcu(*rq);
>>>> + spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
>>>> +}
>>>> +
>>>>   void xehp_enable_ccs_engines(struct intel_engine_cs *engine)
>>>>   {
>>>>       /*
>>>> diff --git a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c 
>>>> b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
>>>> index 18ffe55282e59..05995c8577bef 100644
>>>> --- a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
>>>> +++ b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
>>>> @@ -4150,6 +4150,33 @@ void intel_execlists_show_requests(struct 
>>>> intel_engine_cs *engine,
>>>>       spin_unlock_irqrestore(&sched_engine->lock, flags);
>>>>   }
>>>>   +static unsigned long list_count(struct list_head *list)
>>>> +{
>>>> +    struct list_head *pos;
>>>> +    unsigned long count = 0;
>>>> +
>>>> +    list_for_each(pos, list)
>>>> +        count++;
>>>> +
>>>> +    return count;
>>>> +}
>>>> +
>>>> +void intel_execlist_dump_active_requests(struct intel_engine_cs 
>>>> *engine,
>>>> +                     struct i915_request *hung_rq,
>>>> +                     struct drm_printer *m)
>>>> +{
>>>> +    unsigned long flags;
>>>> +
>>>> + spin_lock_irqsave(&engine->sched_engine->lock, flags);
>>>> +
>>>> + 
>>>> intel_engine_dump_active_requests(&engine->sched_engine->requests, 
>>>> hung_rq, m);
>>>> +
>>>> +    drm_printf(m, "\tOn hold?: %lu\n",
>>>> + list_count(&engine->sched_engine->hold));
>>>> +
>>>> + spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
>>>> +}
>>>> +
>>>>   #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
>>>>   #include "selftest_execlists.c"
>>>>   #endif
>>>> diff --git a/drivers/gpu/drm/i915/gt/intel_execlists_submission.h 
>>>> b/drivers/gpu/drm/i915/gt/intel_execlists_submission.h
>>>> index a1aa92c983a51..cb07488a03764 100644
>>>> --- a/drivers/gpu/drm/i915/gt/intel_execlists_submission.h
>>>> +++ b/drivers/gpu/drm/i915/gt/intel_execlists_submission.h
>>>> @@ -32,6 +32,10 @@ void intel_execlists_show_requests(struct 
>>>> intel_engine_cs *engine,
>>>>                               int indent),
>>>>                      unsigned int max);
>>>>   +void intel_execlist_dump_active_requests(struct intel_engine_cs 
>>>> *engine,
>>>> +                     struct i915_request *hung_rq,
>>>> +                     struct drm_printer *m);
>>>> +
>>>>   bool
>>>>   intel_engine_in_execlists_submission_mode(const struct 
>>>> intel_engine_cs *engine);
>>>>   diff --git a/drivers/gpu/drm/i915/i915_gpu_error.c 
>>>> b/drivers/gpu/drm/i915/i915_gpu_error.c
>>>> index 5c73dfa2fb3f6..b20bd6365615b 100644
>>>> --- a/drivers/gpu/drm/i915/i915_gpu_error.c
>>>> +++ b/drivers/gpu/drm/i915/i915_gpu_error.c
>>>> @@ -1596,35 +1596,15 @@ capture_engine(struct intel_engine_cs *engine,
>>>>   {
>>>>       struct intel_engine_capture_vma *capture = NULL;
>>>>       struct intel_engine_coredump *ee;
>>>> -    struct intel_context *ce;
>>>> +    struct intel_context *ce = NULL;
>>>>       struct i915_request *rq = NULL;
>>>> -    unsigned long flags;
>>>>         ee = intel_engine_coredump_alloc(engine, ALLOW_FAIL, 
>>>> dump_flags);
>>>>       if (!ee)
>>>>           return NULL;
>>>>   -    ce = intel_engine_get_hung_context(engine);
>>>> -    if (ce) {
>>>> -        intel_engine_clear_hung_context(engine);
>>>> -        rq = intel_context_find_active_request_get(ce);
>>>> -        if (!rq || !i915_request_started(rq))
>>>> -            goto no_request_capture;
>>>> -    } else {
>>>> -        /*
>>>> -         * Getting here with GuC enabled means it is a forced 
>>>> error capture
>>>> -         * with no actual hang. So, no need to attempt the 
>>>> execlist search.
>>>> -         */
>>>> -        if (!intel_uc_uses_guc_submission(&engine->gt->uc)) {
>>>> - spin_lock_irqsave(&engine->sched_engine->lock, flags);
>>>> -            rq = intel_engine_execlist_find_hung_request(engine);
>>>> -            if (rq)
>>>> -                rq = i915_request_get_rcu(rq);
>>>> - spin_unlock_irqrestore(&engine->sched_engine->lock,
>>>> -                           flags);
>>>> -        }
>>>> -    }
>>>> -    if (!rq)
>>>> +    intel_engine_get_hung_entity(engine, &ce, &rq);
>>>> +    if (!rq || !i915_request_started(rq))
>>>>           goto no_request_capture;
>>>>         capture = intel_engine_coredump_add_request(ee, rq, 
>>>> ATOMIC_MAYFAIL);
>>