Re: [igt-dev] [PATCH i-g-t 2/3] i915/gem_ctx_shared: Make gem_ctx_shared understand static priority mapping

[Matthew Brost <matthew.brost@intel.com>]

On Wed, Oct 06, 2021 at 09:12:41AM +0100, Tvrtko Ursulin wrote:
> 
> On 05/10/2021 17:44, Matthew Brost wrote:
> > On Tue, Oct 05, 2021 at 11:44:02AM +0100, Tvrtko Ursulin wrote:
> > > 
> > > On 05/10/2021 00:26, Daniele Ceraolo Spurio wrote:
> > > > 
> > > > 
> > > > On 9/16/2021 11:03 AM, Matthew Brost wrote:
> > > > > The i915 currently has 2k visible priority levels which are currently
> > > > > unique. This is changing to statically map these 2k levels into 3
> > > > > buckets:
> > > > > 
> > > > > low: < 0
> > > > > mid: 0
> > > > > high: > 0
> > > > > 
> > > > > Update gem_ctx_shared to understand this. This entails updating
> > > > > promotion test to use 3 levels that will map into different buckets and
> > > > > also add bit of delay after releasing a cork beforing completing the
> > > > > spinners to give time to the i915 schedule to process the fence and
> > > > > release and queue the requests.
> > > > > 
> > > > > v2: Add a delay between starting releasing spinner and cork in
> > > > > promotion
> > > > > v3:
> > > > >    (Daniele)
> > > > >     - Always add delay, update commit message
> > > > > 
> > > > > Signed-off-by: Matthew Brost <matthew.brost@intel.com>
> > > > > ---
> > > > >    tests/i915/gem_ctx_shared.c | 5 +++--
> > > > >    1 file changed, 3 insertions(+), 2 deletions(-)
> > > > > 
> > > > > diff --git a/tests/i915/gem_ctx_shared.c b/tests/i915/gem_ctx_shared.c
> > > > > index ea1b5dd1b..7f88871b8 100644
> > > > > --- a/tests/i915/gem_ctx_shared.c
> > > > > +++ b/tests/i915/gem_ctx_shared.c
> > > > > @@ -622,6 +622,7 @@ static void unplug_show_queue(int i915, struct
> > > > > igt_cork *c, uint64_t ahnd,
> > > > >        igt_cork_unplug(c); /* batches will now be queued on the engine */
> > > > >        igt_debugfs_dump(i915, "i915_engine_info");
> > > > > +    usleep(250000);
> > > > 
> > > > Same as previous patch, with a comment added:
> > > > 
> > > > Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
> > > 
> > > Come on guys, it is a bit bad and lazy for good taste no? 250ms more test
> > 
> > Yea, this is way too long of a sleep. 25ms seems just fine.
> 
> Until you get 1/1000 failures in CI on some platforms due phase of the moon.
> Adding sleeps should be avoided where possible.
> 

Yea, that is always a risk. Looking at this test there is already 1
other sleep in the test and in gem_exec_schedule there are 5 other
sleeps. I'm not breaking precedent here. 

> > 
> > On TGL /w the updated sleep:
> > gem_ctx_shared --r *promotion*
> > IGT-Version: 1.26-g7e489b053 (x86_64) (Linux: 5.15.0-rc3-guc+ x86_64)
> > Starting subtest: Q-promotion
> > Starting dynamic subtest: rcs0
> > Dynamic subtest rcs0: SUCCESS (0.059s)
> > Starting dynamic subtest: bcs0
> > Dynamic subtest bcs0: SUCCESS (0.059s)
> > Starting dynamic subtest: vcs0
> > Dynamic subtest vcs0: SUCCESS (0.060s)
> > Starting dynamic subtest: vecs0
> > Dynamic subtest vecs0: SUCCESS (0.061s)
> > Subtest Q-promotion: SUCCESS (0.239s)
> > 
> > > runtime, multiplied by number of tests and subtests (engines) will add up
> > > and over shadows the current test time. For instance current state is:
> > > 
> > > # tests/gem_ctx_shared --r *promotion*
> > > IGT-Version: 1.26-NO-GIT (x86_64) (Linux: 5.15.0-rc4 x86_64)
> > > Starting subtest: Q-promotion
> > > Starting dynamic subtest: rcs0
> > > Dynamic subtest rcs0: SUCCESS (0.174s)
> > > Starting dynamic subtest: bcs0
> > > Dynamic subtest bcs0: SUCCESS (0.224s)
> > > Starting dynamic subtest: vcs0
> > > Dynamic subtest vcs0: SUCCESS (0.153s)
> > > Starting dynamic subtest: vecs0
> > > Dynamic subtest vecs0: SUCCESS (0.153s)
> > > Subtest Q-promotion: SUCCESS (0.708s)
> > > 
> > > This patch instantly makes this 1.7 seconds instead. Add the in/out order
> > > subtests, other tests, platforms with more engines, in a world where CI time
> > > budget is scarce - can't we do better than this and not settle on sprinkling
> > > of usleep all over the place?
> > > 
> > > Like maybe add out fences to the two requests, merge them, and wait on that
> > > (since there is no implicit write declared so set domain does not suffice)?
> > > 
> > 
> > Not sure I follow. Anyways reposting now with a smaller timeout value.
> > Unless we hear something we plan on merging this tomorrow.
> 
> The in/out-order test as example. It needs to wait until two dword writes
> have completed, right? I am saying pass output fences out from spinners and
> wait on them before checking content of the shared buffer.
> 
> I also think sleep after uncorking is also conceptually misleading because
> that's not where the problem lies. Problem is set domain does not actually
> wait for sdw completion (existing comment even hints so "no write hazard
> lies!"). If I am right sporadic fail can in *theory* happen with execlists
> as well.
> 

Still not following what you are saying. The sleep just adds period of
time for all the uncorked requests to make it into the GuC scheduler,
without it is possible for the lower priority request (submitted first)
completes before the higher priority request (submitted after) makes it
into the GuC. The sleep ensures alls the requests are in the GuC
scheduler still stuck behind spinning requests on the hardware, then
after the spinners complete the requests are scheduled in order of
priority. Yes, this possible with execlists too but I think the timing
is different enough it doesn't happen.

Again adding a sleep here seems legit to me, the same problem would
exist if we used fences as you suggest (e.g. the lower priority request
would be submitted first + could complete before the higher priority
request is submitted). Let's not over think this one.

Matt

> Regards,
> 
> Tvrtko
> 
> > 
> > Matt
> > 
> > > Unless I am missing something it would be strictly correct for execlists
> > > backend as well since it now works just because it is fast enough.
> > > 
> > > Regards,
> > > 
> > > Tvrtko
> > > 
> > > > 
> > > > Daniele
> > > > 
> > > > >        for (int n = 0; n < ARRAY_SIZE(spin); n++) {
> > > > >            ahnd = spin[n]->ahnd;
> > > > > @@ -831,10 +832,10 @@ static void promotion(int i915, const
> > > > > intel_ctx_cfg_t *cfg, unsigned ring)
> > > > >        gem_context_set_priority(i915, ctx[LO]->id, MIN_PRIO);
> > > > >        ctx[HI] = intel_ctx_create(i915, &q_cfg);
> > > > > -    gem_context_set_priority(i915, ctx[HI]->id, 0);
> > > > > +    gem_context_set_priority(i915, ctx[HI]->id, MAX_PRIO);
> > > > >        ctx[NOISE] = intel_ctx_create(i915, &q_cfg);
> > > > > -    gem_context_set_priority(i915, ctx[NOISE]->id, MIN_PRIO/2);
> > > > > +    gem_context_set_priority(i915, ctx[NOISE]->id, 0);
> > > > >        result = gem_create(i915, 4096);
> > > > >        result_offset = get_offset(ahnd, result, 4096, 0);
> > > >