[Intel-gfx] [PATCH 1/2] drm/i915/pmu: Fix KMD and GuC race on accessing busyness

[Tvrtko Ursulin]

On 28/01/2022 21:53, John Harrison wrote:
> On 1/28/2022 01:34, Tvrtko Ursulin wrote:
>>
>> John,
>>
>> What CI results were used to merge this particular single patch? 
>> Unless I am not seeing it, it was always set in pair with something else.
>>
>> First with "drm/i915/pmu: Use PM timestamp instead of RING TIMESTAMP 
>> for reference", which was merged earlier in the week and it had a 
>> standalone CI results so that is fine.
> 
> It seemed plausible to say that it had a green CI run when included with 
> a dependent patch given that the dependent patch has already been merged.

Given it was later sent in a series with a different patch that was 
quite confusing.

> We don't normally require a brand new CI run any time the tree has moved 
> on since the previous CI was run. You would never get anything merged if 
> that was necessary!

Not when the tree has moved, but when series gain or lose patches.

Here we had these CI runs:

1. Patch A + Patch B
2. Patch A
3. Patch B + Patch C

Patch A got merged based on 2 one day. Patch B merged alone, based on 
what you say, the next day...

> So to me this just counts as pushing a tested patch set to a slightly 
> newer tree where the rebase happened to mean one patch vanished (because 
> exactly the same patch was already merged).

.. I say please do a CI run for B standalone. That way the Link: in the 
merged patch points to a series where maintainer can click on it and see 
it was green and all is good. These situations are so rare that there 
are no concerns about CI time.

If the link leads me somewhere else, which does not make sense, then it 
is extra work to untangle the story. I don't want to hunt on the mailing 
list between multiple postings of the patch with the same name and check 
whether it is the same patch, or just the same title where someone 
forgot to add a change log or something. Keeping it simple is the 
cheapest option here.

Regards,

Tvrtko

> 
> John.
> 
> 
>>
>> Other postings I can see it was in tandem with "drm/i915/pmu: Use 
>> existing uncore helper to read gpm_timestamp", which wasn't reviewed 
>> or merged.
>>
>> So it looks to me, again unless I am not seeing anything, that it got 
>> merged without ever having a standalone CI results. This is therefore 
>> a reminder that BKM should be to always have CI results for the exact 
>> series/patch being merged.
>>
>> If there is a situation where a subset of a series is conceptually 
>> ready before the rest, in the past what we used to do is send the 
>> reviewed portion as "--subject-prefix=CI" and then "--suppress-cc=all" 
>> and then merge when CI gives all green.
>>
>> Regards,
>>
>> Tvrtko
>>
>> On 27/01/2022 02:00, Umesh Nerlige Ramappa wrote:
>>> GuC updates shared memory and KMD reads it. Since this is not
>>> synchronized, we run into a race where the value read is inconsistent.
>>> Sometimes the inconsistency is in reading the upper MSB bytes of the
>>> last_switch_in value. 2 types of cases are seen - upper 8 bits are zero
>>> and upper 24 bits are zero. Since these are non-zero values, it is
>>> not trivial to determine validity of these values. Instead we read the
>>> values multiple times until they are consistent. In test runs, 3
>>> attempts results in consistent values. The upper bound is set to 6
>>> attempts and may need to be tuned as per any new occurences.
>>>
>>> Since the duration that gt is parked can vary, the patch also updates
>>> the gt timestamp on unpark before starting the worker.
>>>
>>> v2:
>>> - Initialize i
>>> - Use READ_ONCE to access engine record
>>>
>>> Fixes: 77cdd054dd2c ("drm/i915/pmu: Connect engine busyness stats 
>>> from GuC to pmu")
>>> Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa at intel.com>
>>> Reviewed-by: Alan Previn <alan.previn.teres.alexis at intel.com>
>>> ---
>>>   .../gpu/drm/i915/gt/uc/intel_guc_submission.c | 58 +++++++++++++++++--
>>>   1 file changed, 54 insertions(+), 4 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 db9615dcb0ec..4e9154cacc58 100644
>>> --- a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
>>> +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
>>> @@ -1114,6 +1114,19 @@ __extend_last_switch(struct intel_guc *guc, 
>>> u64 *prev_start, u32 new_start)
>>>       if (new_start == lower_32_bits(*prev_start))
>>>           return;
>>>   +    /*
>>> +     * When gt is unparked, we update the gt timestamp and start the 
>>> ping
>>> +     * worker that updates the gt_stamp every POLL_TIME_CLKS. As 
>>> long as gt
>>> +     * is unparked, all switched in contexts will have a start time 
>>> that is
>>> +     * within +/- POLL_TIME_CLKS of the most recent gt_stamp.
>>> +     *
>>> +     * If neither gt_stamp nor new_start has rolled over, then the
>>> +     * gt_stamp_hi does not need to be adjusted, however if one of 
>>> them has
>>> +     * rolled over, we need to adjust gt_stamp_hi accordingly.
>>> +     *
>>> +     * The below conditions address the cases of new_start rollover and
>>> +     * gt_stamp_last rollover respectively.
>>> +     */
>>>       if (new_start < gt_stamp_last &&
>>>           (new_start - gt_stamp_last) <= POLL_TIME_CLKS)
>>>           gt_stamp_hi++;
>>> @@ -1125,17 +1138,45 @@ __extend_last_switch(struct intel_guc *guc, 
>>> u64 *prev_start, u32 new_start)
>>>       *prev_start = ((u64)gt_stamp_hi << 32) | new_start;
>>>   }
>>>   -static void guc_update_engine_gt_clks(struct intel_engine_cs *engine)
>>> +/*
>>> + * GuC updates shared memory and KMD reads it. Since this is not 
>>> synchronized,
>>> + * we run into a race where the value read is inconsistent. 
>>> Sometimes the
>>> + * inconsistency is in reading the upper MSB bytes of the last_in 
>>> value when
>>> + * this race occurs. 2 types of cases are seen - upper 8 bits are 
>>> zero and upper
>>> + * 24 bits are zero. Since these are non-zero values, it is 
>>> non-trivial to
>>> + * determine validity of these values. Instead we read the values 
>>> multiple times
>>> + * until they are consistent. In test runs, 3 attempts results in 
>>> consistent
>>> + * values. The upper bound is set to 6 attempts and may need to be 
>>> tuned as per
>>> + * any new occurences.
>>> + */
>>> +static void __get_engine_usage_record(struct intel_engine_cs *engine,
>>> +                      u32 *last_in, u32 *id, u32 *total)
>>>   {
>>>       struct guc_engine_usage_record *rec = 
>>> intel_guc_engine_usage(engine);
>>> +    int i = 0;
>>> +
>>> +    do {
>>> +        *last_in = READ_ONCE(rec->last_switch_in_stamp);
>>> +        *id = READ_ONCE(rec->current_context_index);
>>> +        *total = READ_ONCE(rec->total_runtime);
>>> +
>>> +        if (READ_ONCE(rec->last_switch_in_stamp) == *last_in &&
>>> +            READ_ONCE(rec->current_context_index) == *id &&
>>> +            READ_ONCE(rec->total_runtime) == *total)
>>> +            break;
>>> +    } while (++i < 6);
>>> +}
>>> +
>>> +static void guc_update_engine_gt_clks(struct intel_engine_cs *engine)
>>> +{
>>>       struct intel_engine_guc_stats *stats = &engine->stats.guc;
>>>       struct intel_guc *guc = &engine->gt->uc.guc;
>>> -    u32 last_switch = rec->last_switch_in_stamp;
>>> -    u32 ctx_id = rec->current_context_index;
>>> -    u32 total = rec->total_runtime;
>>> +    u32 last_switch, ctx_id, total;
>>>         lockdep_assert_held(&guc->timestamp.lock);
>>>   +    __get_engine_usage_record(engine, &last_switch, &ctx_id, &total);
>>> +
>>>       stats->running = ctx_id != ~0U && last_switch;
>>>       if (stats->running)
>>>           __extend_last_switch(guc, &stats->start_gt_clk, last_switch);
>>> @@ -1237,6 +1278,10 @@ static ktime_t guc_engine_busyness(struct 
>>> intel_engine_cs *engine, ktime_t *now)
>>>       if (!in_reset && intel_gt_pm_get_if_awake(gt)) {
>>>           stats_saved = *stats;
>>>           gt_stamp_saved = guc->timestamp.gt_stamp;
>>> +        /*
>>> +         * Update gt_clks, then gt timestamp to simplify the 
>>> 'gt_stamp -
>>> +         * start_gt_clk' calculation below for active engines.
>>> +         */
>>>           guc_update_engine_gt_clks(engine);
>>>           guc_update_pm_timestamp(guc, now);
>>>           intel_gt_pm_put_async(gt);
>>> @@ -1365,10 +1410,15 @@ void intel_guc_busyness_park(struct intel_gt 
>>> *gt)
>>>   void intel_guc_busyness_unpark(struct intel_gt *gt)
>>>   {
>>>       struct intel_guc *guc = &gt->uc.guc;
>>> +    unsigned long flags;
>>> +    ktime_t unused;
>>>         if (!guc_submission_initialized(guc))
>>>           return;
>>>   +    spin_lock_irqsave(&guc->timestamp.lock, flags);
>>> +    guc_update_pm_timestamp(guc, &unused);
>>> +    spin_unlock_irqrestore(&guc->timestamp.lock, flags);
>>>       mod_delayed_work(system_highpri_wq, &guc->timestamp.work,
>>>                guc->timestamp.ping_delay);
>>>   }
>>>
>