[Intel-gfx] [PATCH 4/7] drm/i915/perf: lock powergating configuration to default when active

[Tvrtko Ursulin]

On 11/09/2018 21:11, Lionel Landwerlin wrote:
> On 10/09/2018 14:44, Tvrtko Ursulin wrote:
>>
>> On 07/09/2018 10:55, Lionel Landwerlin wrote:
>>> On 07/09/2018 10:39, Tvrtko Ursulin wrote:
>>>>
>>>> On 07/09/2018 10:23, Lionel Landwerlin wrote:
>>>>> On 07/09/2018 09:26, Tvrtko Ursulin wrote:
>>>>>>
>>>>>> On 06/09/2018 11:36, Lionel Landwerlin wrote:
>>>>>>> On 06/09/2018 11:22, Chris Wilson wrote:
>>>>>>>> Quoting Lionel Landwerlin (2018-09-06 11:18:01)
>>>>>>>>> On 06/09/2018 11:10, Chris Wilson wrote:
>>>>>>>>>> Quoting Lionel Landwerlin (2018-09-06 10:57:47)
>>>>>>>>>>> On 05/09/2018 15:22, Tvrtko Ursulin wrote:
>>>>>>>>>>>> From: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
>>>>>>>>>>>>
>>>>>>>>>>>> If some of the contexts submitting workloads to the GPU have 
>>>>>>>>>>>> been
>>>>>>>>>>>> configured to shutdown slices/subslices, we might loose the NOA
>>>>>>>>>>>> configurations written in the NOA muxes.
>>>>>>>>>>>>
>>>>>>>>>>>> One possible solution to this problem is to reprogram the 
>>>>>>>>>>>> NOA muxes
>>>>>>>>>>>> when we switch to a new context. We initially tried this in the
>>>>>>>>>>>> workaround batchbuffer but some concerns where raised about 
>>>>>>>>>>>> the cost
>>>>>>>>>>>> of reprogramming at every context switch. This solution is 
>>>>>>>>>>>> also not
>>>>>>>>>>>> without consequences from the userspace point of view. 
>>>>>>>>>>>> Reprogramming
>>>>>>>>>>>> of the muxes can only happen once the powergating 
>>>>>>>>>>>> configuration has
>>>>>>>>>>>> changed (which happens after context switch). This means for 
>>>>>>>>>>>> a window
>>>>>>>>>>>> of time during the recording, counters recorded by the OA 
>>>>>>>>>>>> unit might
>>>>>>>>>>>> be invalid. This requires userspace dealing with OA reports 
>>>>>>>>>>>> to discard
>>>>>>>>>>>> the invalid values.
>>>>>>>>>>>>
>>>>>>>>>>>> Minimizing the reprogramming could be implemented by 
>>>>>>>>>>>> tracking of the
>>>>>>>>>>>> last programmed configuration somewhere in GGTT and use 
>>>>>>>>>>>> MI_PREDICATE
>>>>>>>>>>>> to discard some of the programming commands, but the command 
>>>>>>>>>>>> streamer
>>>>>>>>>>>> would still have to parse all the MI_LRI instructions in the
>>>>>>>>>>>> workaround batchbuffer.
>>>>>>>>>>>>
>>>>>>>>>>>> Another solution, which this change implements, is to simply 
>>>>>>>>>>>> disregard
>>>>>>>>>>>> the user requested configuration for the period of time when 
>>>>>>>>>>>> i915/perf
>>>>>>>>>>>> is active. There is no known issue with this apart from a 
>>>>>>>>>>>> performance
>>>>>>>>>>>> penality for some media workloads that benefit from running 
>>>>>>>>>>>> on a
>>>>>>>>>>>> partially powergated GPU. We already prevent RC6 from 
>>>>>>>>>>>> affecting the
>>>>>>>>>>>> programming so it doesn't sound completely unreasonable to 
>>>>>>>>>>>> hold on
>>>>>>>>>>>> powergating for the same reason.
>>>>>>>>>>>>
>>>>>>>>>>>> v2: Leave RPCS programming in intel_lrc.c (Lionel)
>>>>>>>>>>>>
>>>>>>>>>>>> v3: Update for s/union intel_sseu/struct intel_sseu/ (Lionel)
>>>>>>>>>>>>        More to_intel_context() (Tvrtko)
>>>>>>>>>>>>        s/dev_priv/i915/ (Tvrtko)
>>>>>>>>>>>>
>>>>>>>>>>>> Tvrtko Ursulin:
>>>>>>>>>>>>
>>>>>>>>>>>> v4:
>>>>>>>>>>>>     * Rebase for make_rpcs changes.
>>>>>>>>>>>>
>>>>>>>>>>>> v5:
>>>>>>>>>>>>     * Apply OA restriction from make_rpcs directly.
>>>>>>>>>>>>
>>>>>>>>>>>> v6:
>>>>>>>>>>>>     * Rebase for context image setup changes.
>>>>>>>>>>>>
>>>>>>>>>>>> Signed-off-by: Lionel Landwerlin 
>>>>>>>>>>>> <lionel.g.landwerlin at intel.com>
>>>>>>>>>>>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>>>>>>>>> ---
>>>>>>>>>>>>     drivers/gpu/drm/i915/i915_perf.c |  5 +++++
>>>>>>>>>>>>     drivers/gpu/drm/i915/intel_lrc.c | 30 
>>>>>>>>>>>> ++++++++++++++++++++----------
>>>>>>>>>>>>     drivers/gpu/drm/i915/intel_lrc.h |  3 +++
>>>>>>>>>>>>     3 files changed, 28 insertions(+), 10 deletions(-)
>>>>>>>>>>>>
>>>>>>>>>>>> diff --git a/drivers/gpu/drm/i915/i915_perf.c 
>>>>>>>>>>>> b/drivers/gpu/drm/i915/i915_perf.c
>>>>>>>>>>>> index ccb20230df2c..dd65b72bddd4 100644
>>>>>>>>>>>> --- a/drivers/gpu/drm/i915/i915_perf.c
>>>>>>>>>>>> +++ b/drivers/gpu/drm/i915/i915_perf.c
>>>>>>>>>>>> @@ -1677,6 +1677,11 @@ static void 
>>>>>>>>>>>> gen8_update_reg_state_unlocked(struct i915_gem_context *ctx,
>>>>>>>>>>>>                 CTX_REG(reg_state, state_offset, 
>>>>>>>>>>>> flex_regs[i], value);
>>>>>>>>>>>>         }
>>>>>>>>>>>> +
>>>>>>>>>>>> +     CTX_REG(reg_state, CTX_R_PWR_CLK_STATE, 
>>>>>>>>>>>> GEN8_R_PWR_CLK_STATE,
>>>>>>>>>>>> +             gen8_make_rpcs(dev_priv,
>>>>>>>>>>>> + &to_intel_context(ctx,
>>>>>>>>>>>> + dev_priv->engine[RCS])->sseu));
>>>>>>>>>>> I think there is one issue I missed on the previous 
>>>>>>>>>>> iterations of this
>>>>>>>>>>> patch.
>>>>>>>>>>>
>>>>>>>>>>> This gen8_update_reg_state_unlocked() is called when the GPU 
>>>>>>>>>>> is parked
>>>>>>>>>>> on the kernel context.
>>>>>>>>>>>
>>>>>>>>>>> It's supposed to update all contexts, but I think we might 
>>>>>>>>>>> not be able
>>>>>>>>>>> to update the kernel context image while the GPU is using it.
>>>>>>>>>> The kernel context is only ever taken in extremis (you are either
>>>>>>>>>> parking or stalling userspace) so I don't care.
>>>>>>>>>
>>>>>>>>> The patch exposing the RPCS configuration to userspace will 
>>>>>>>>> make use of
>>>>>>>>> the kernel context while OA/perf is enabled. Even if it 
>>>>>>>>> reprograms the
>>>>>>>>> locked value that will break the power configuration stability 
>>>>>>>>> on Gen11
>>>>>>>>> (because the locked configuration will be different from the 
>>>>>>>>> kernel
>>>>>>>>> context configuration).
>>>>>>>> Sure, but as you point out that's only on changing configuration.
>>>>>>>>
>>>>>>>> What's missing in the patch is that we only bail early if the 
>>>>>>>> new sseu
>>>>>>>> matches the ce->sseu, but that doesn't necessarily match whats 
>>>>>>>> in the
>>>>>>>> context due to OA. (Or maybe I missed the conversion to rpcs 
>>>>>>>> value and
>>>>>>>> checking.)
>>>>>>>> -Chris
>>>>>>>>
>>>>>>>
>>>>>>> Yep, because the gen8_make_rpcs() post processes the values store 
>>>>>>> at the gem context level, we risk rerunning the kernel context to 
>>>>>>> write the exiting value.
>>>>>>> Sorry this is all so messy :(
>>>>>>
>>>>>> Lets see if I managed to follow here.
>>>>>>
>>>>>> The current code indeed bails out at the set ctx param level if 
>>>>>> the requested state matches the ce->state. My thinking was that 
>>>>>> ce->state is the master state and whatever happens in "post 
>>>>>> processing" via gen8_make_rpcs should be hidden from it since the 
>>>>>> design is that the i915_perf.c will re-configure all contexts when 
>>>>>> the OA active status changes (to either direction).
>>>>>>
>>>>>> So I don't see a problem in those two interactions.
>>>>>
>>>>>
>>>>> Let's say you have contextA with sseu(slice,subslice)=(0x1/0xff) 
>>>>> for ICL.
>>>>>
>>>>> You then enable OA which locks the configuration at (0x1,0xf).
>>>>>
>>>>> The kernel context has retained its (0x1/0xff) configuration.
>>>>>
>>>>>
>>>>> And after you change the config of contextA to (0x1,0x7).
>>>>>
>>>>>
>>>>> This would lead to the kernel context scheduled with (0x1,0xff) 
>>>>> while OA is active.
>>>>
>>>> Okay that's a problem discussed in the paragraph below - that the 
>>>> kernel context is not updated at all. But is it a problem for OA? 
>>>> Will it mess up some counters even if kernel context isn't executing 
>>>> anything interacting with them? Or is it?
>>>
>>>
>>> What the HW is going to do to the NOA logic in power configuration 
>>> changes is not really documented.
>>> Experimentally on SKL GT4, it seems a change in power configuration 
>>> will trigger a power off of everything before applying the power at 
>>> the new configuration.
>>> So that would imply loosing the NOA programming when we switch to the 
>>> kernel context which means invalid values in the counters.
>>>
>>>
>>>>
>>>>>
>>>>>>
>>>>>> Apart from one, get_param_sseu will lie a bit - we can discuss 
>>>>>> about this one more. At one point I suggested we have two sets of 
>>>>>> masks in the uAPI, requested and active in a way. So userspace 
>>>>>> could query what it set and what is actually active.
>>>>>>
>>>>>> Now second issue is if i915_perf.c is able to reprogram the kernel 
>>>>>> config.
>>>>>>
>>>>>> Here its true, it will write to the context image and that will 
>>>>>> get overwritten by context save.
>>>>>>
>>>>>> If that is a problem for OA, I was initially if a throw-away 
>>>>>> second "kernel" context could be use to re-program the real one, 
>>>>>> but perhaps even simpler - what about a mmio write to program the 
>>>>>> RPCS while kernel context is active?
>>>>>
>>>>>
>>>>> Documentation says : "This register must not be programmed directly 
>>>>> through CPU MMIO cycle."
>>>>>
>>>>>
>>>>> Sorry :(
>>>>
>>>> Ugh.. okay, help me understand if kernel context absolutely needs to 
>>>> follow the "lock" for OA to work and then we'll see what to do.
>>>
>>> I think so.
>>>
>>> Your idea of a throw away context to reprogramming every seems sound.
>>
>> I was in the middle of refactoring the series to implement this when I 
>> started suspecting the need for it.
>>
>> Premise of the problem statement was the kernel context doesn't get 
>> updated by the OA code, but, there is a loop in 
>> gen8_configure_all_contexts which goes through exactly all of them 
>> after it has idled the GPU.
>>
>> AFAICS that means it is able to map the kernel context state and edit 
>> it so everything seems fine from this angle. (Kernel context is 
>> perma-pinned in software, but after idling the GPU we know it is not 
>> actually on the GPU so it safe to edit it's image.)
>>
>> Am I missing some hole here, and if not, why does the code needs to 
>> have a primary update via a request in 
>> gen8_switch_to_updated_kernel_context? Context image after idling 
>> seems again sufficient.
> 
> 
> My understanding is that the current code puts the GPU idle on the 
> kernel context.
> When idling stops, the GPU will save the last values from the HW 
> register into the kernel context image.
> We currently don't see any issue because we also run a set of commands 
> on the kernel context prior to idle that mirror the contexts image edition.

We actually retire everything, and the kernel context, wait for any 
pending interrupts (context save) and check that CS reports idle. So I 
think it is actually simpler than the OA code currently does it. I just 
wanted a way to express it as a test.

> That won't be the case for the RPCS register because we can't load it 
> from the command streamer.
> 
>>
>> Would it be possible to exercise the hypothetical loss of NOA 
>> configuration from an IGT, if kernel context wasn't correctly updated?
> 
> 
> The test configs saved in the kernel use NOA and their counters are 
> progressing at different ratios from the "core clock" (RING_TIMESTAMP 
> register).
> In IGT tests/perf.c, the gen8_sanity_check_test_oa_reports() verify that 
> the counters progress as expected.
> 
> But I can't tell from what part of the GT the test configs are sourcing 
> signals from. If they do from unslice, then that won't work :(

So not testable? Shall we still apply the doctrine of "if in doubt rip 
it out"? Or too risky? Shouldn't be.. looks pretty obvious we check 
properly for idle. If we simplify the worst that can happen is that OA 
becomes glitchy paired with Gen11 media workloads which is not so 
critical, but I don't think it will break to start with.

Regards,

Tvrtko