With the global kmem_cache shrink infrastructure gone there's nothing
special and we can convert them over.
I'm doing this split up into each patch because there's quite a bit of
noise with removing the static global.slab_dependencies|priorities to just a
slab_dependencies|priorities.
v2: Make slab static (Jason, 0day)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210727121037.2041102-8-daniel.vetter@ffwll.ch
Implement a simple static mapping algorithm of the i915 priority levels
(int, -1k to 1k exposed to user) to the 4 GuC levels. Mapping is as
follows:
i915 level < 0 -> GuC low level (3)
i915 level == 0 -> GuC normal level (2)
i915 level < INT_MAX -> GuC high level (1)
i915 level == INT_MAX -> GuC highest level (0)
We believe this mapping should cover the UMD use cases (3 distinct user
levels + 1 kernel level).
In addition to static mapping, a simple counter system is attached to
each context tracking the number of requests inflight on the context at
each level. This is needed as the GuC levels are per context while in
the i915 levels are per request.
v2:
(Daniele)
- Add BUILD_BUG_ON to enforce ordering of priority levels
- Add missing lockdep to guc_prio_fini
- Check for return before setting context registered flag
- Map DISPLAY priority or higher to highest guc prio
- Update comment for guc_prio
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210727002348.97202-33-matthew.brost@intel.com
If submission is disabled by the backend for any reason, reset the GPU
immediately in the heartbeat code as the backend can't be reenabled
until the GPU is reset.
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210727002348.97202-10-matthew.brost@intel.com
This essentially reverts
commit 84a1074920
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Wed Jan 24 11:36:08 2018 +0000
drm/i915: Shrink the GEM kmem_caches upon idling
mm/vmscan.c:do_shrink_slab() is a thing, if there's an issue with it
then we need to fix that there, not hand-roll our own slab shrinking
code in i915.
Also when this was added there was only one other caller of
kmem_cache_shrink (added 2005 to the acpi code). Now there's a 2nd one
outside of i915 code in a kunit test, which seems legit since that
wants to very carefully control what's in the kmem_cache. This out of
a total of over 500 calls to kmem_cache_create. This alone should have
been warning sign enough that we're doing something silly.
Noticed while reviewing a patch set from Jason to fix up some issues
in our i915_init() and i915_exit() module load/cleanup code. Now that
i915_globals.c isn't any different than normal init/exit functions, we
should convert them over to one unified table and remove
i915_globals.[hc] entirely.
v2: Improve commit message (Jason)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: David Airlie <airlied@linux.ie>
Cc: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210721183229.4136488-1-daniel.vetter@ffwll.ch
The submission tasklet operates on i915_sched_engine, thus it is the
correct place for it.
v3:
(Jason Ekstrand)
Change sched_engine->engine to a void* private data pointer
Add kernel doc
v4:
(Daniele)
Update private_data comment
Set queue_priority_hint in kick_execlists
v5:
(CI)
Rebase and fix build error
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210618010638.98941-9-matthew.brost@intel.com
Rather passing around an intel_engine_cs in the scheduling code, pass
around a i915_sched_engine.
v3:
(Jason Ekstrand)
Add READ_ONCE around rq->engine in lock_sched_engine
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210618010638.98941-8-matthew.brost@intel.com
Not all back-ends require a kick after a scheduling update, so make the
kick a call-back function that the back-end can opt-in to. Also move
the current kick function from the scheduler to the execlists file as it
is specific to that back-end.
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210618010638.98941-7-matthew.brost@intel.com
Move active request tracking and its lock to i915_sched_engine. This
lock is also the submission lock so having it in the i915_sched_engine
is the correct place.
v3:
(Jason Ekstrand)
Add kernel doc
v6:
Rebase
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.comk>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210618010638.98941-5-matthew.brost@intel.com
Introduce i915_sched_engine object which is lower level data structure
that i915_scheduler / generic code can operate on without touching
execlist specific structures. This allows additional submission backends
to be added without breaking the layering. Currently the execlists
backend uses 1 of these object per each engine (physical or virtual) but
future backends like the GuC will point to less instances utilizing the
reference counting.
This is a bit of detour to integrating the i915 with the DRM scheduler
but this object will still exist when the DRM scheduler lands in the
i915. It will however look a bit different. It will encapsulate the
drm_gpu_scheduler object plus and common variables (to the backends)
related to scheduling. Regardless this is a step in the right direction.
This patch starts the aforementioned transition by moving the priolist
into the i915_sched_engine object.
v3:
(Jason Ekstrand)
Update comment next to intel_engine_cs.virtual
Add kernel doc
(Checkpatch)
Fix double the in commit message
v4:
(Daniele)
Update comment message.
Add comment about subclass field
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210618010638.98941-2-matthew.brost@intel.com
Since we are not using any internal priority levels, and in the next few
patches will introduce a new index for which the optimisation is not so
lear cut, discard the small table within the priolist.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Andi Shyti <andi.shyti@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210120121439.17600-1-chris@chris-wilson.co.uk
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
When we know that we are inside the timeline mutex, or inside the
submission flow (under active.lock or the holder's rcu lock), we know
that the rq->hwsp is stable and we can use the simpler direct version.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Andi Shyti <andi.shyti@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210114135612.13210-1-chris@chris-wilson.co.uk
Rather than having special case code for opportunistically calling
process_csb() and performing a direct submit while holding the engine
spinlock for submitting the request, simply call the tasklet directly.
This allows us to retain the direct submission path, including the CS
draining to allow fast/immediate submissions, without requiring any
duplicated code paths, and most importantly greatly simplifying the
control flow by removing reentrancy. This will enable us to close a few
races in the virtual engines in the next few patches.
The trickiest part here is to ensure that paired operations (such as
schedule_in/schedule_out) remain under consistent locking domains,
e.g. when pulled outside of the engine->active.lock
v2: Use bh kicking, see commit 3c53776e29 ("Mark HI and TASKLET
softirq synchronous").
v3: Update engine-reset to be tasklet aware
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20201224135544.1713-1-chris@chris-wilson.co.uk
Include the signalers each request in the timeline is waiting on, as a
means to try and identify the cause of a stall. This can be quite
verbose, even as for now we only show each request in the timeline and
its immediate antecedents.
This generates output like:
Timeline 886: { count 1, ready: 0, inflight: 0, seqno: { current: 664, last: 666 }, engine: rcs0 }
U 886:29a- prio=0 @ 134ms: gem_exec_parall<4621>
U bc1:27a- prio=0 @ 134ms: gem_exec_parall[4917]
Timeline 825: { count 1, ready: 0, inflight: 0, seqno: { current: 802, last: 804 }, engine: vcs0 }
U 825:324 prio=0 @ 107ms: gem_exec_parall<4518>
U b75:140- prio=0 @ 110ms: gem_exec_parall<5486>
Timeline b46: { count 1, ready: 0, inflight: 0, seqno: { current: 782, last: 784 }, engine: vcs0 }
U b46:310- prio=0 @ 70ms: gem_exec_parall<5428>
U c11:170- prio=0 @ 70ms: gem_exec_parall[5501]
Timeline 96b: { count 1, ready: 0, inflight: 0, seqno: { current: 632, last: 634 }, engine: vcs0 }
U 96b:27a- prio=0 @ 67ms: gem_exec_parall<4878>
U b75:19e- prio=0 @ 67ms: gem_exec_parall<5486>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20201119165616.10834-6-chris@chris-wilson.co.uk
We recorded the execlists->queue_priority_hint update for the inflight
request without kicking the tasklet. The next submitted request then
failed to be scheduled as it had a lower priority than the hint, leaving
the HW running with only the inflight request.
Fixes: 6cebcf746f ("drm/i915: Tweak scheduler's kick_submission()")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200519063123.20673-1-chris@chris-wilson.co.uk
(cherry picked from commit b86fc6e5e8)
Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Now that we have fast timeslicing on semaphores, we no longer need to
prioritise none-semaphore work as we will yield any work blocked on a
semaphore to the next in the queue. Previously with no timeslicing,
blocking on the semaphore caused extremely bad scheduling with multiple
clients utilising multiple rings. Now, there is no impact and we can
remove the complication.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200513173504.28322-1-chris@chris-wilson.co.uk
Upon waiting a request (when asked), we gave that request a small
priority boost, not enough for it to cause preemption, but enough for it
to be scheduled next before all equals. We also used that bit to give
new clients a small priority boost, similar to FQ_CODEL, such that we
favoured short interactive tasks ahead of long running streams.
However, this is causing lots of complications with timeslicing where we
both want to honour the boost and yet ignore it. Those complications
cause unexpected user behaviour (tasks not being timesliced and run
concurrently as epxected), and the easiest way to resolve that is to
remove the boost. Hopefully, we can find a compromise again if we need
to, but in theory timeslicing itself and future more advanced schedulers
should give us the interactivity boost we seek.
Testcase: igt/gem_exec_schedule/lateslice
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200507152338.7452-3-chris@chris-wilson.co.uk
We recorded the dependencies for WAIT_FOR_SUBMIT in order that we could
correctly perform priority inheritance from the parallel branches to the
common trunk. However, for the purpose of timeslicing and reset
handling, the dependency is weak -- as we the pair of requests are
allowed to run in parallel and not in strict succession.
The real significance though is that this allows us to rearrange
groups of WAIT_FOR_SUBMIT linked requests along the single engine, and
so can resolve user level inter-batch scheduling dependencies from user
semaphores.
Fixes: c81471f5e9 ("drm/i915: Copy across scheduler behaviour flags across submit fences")
Testcase: igt/gem_exec_fence/submit
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: <stable@vger.kernel.org> # v5.6+
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200507155109.8892-1-chris@chris-wilson.co.uk
Since we may kick more than engine, and may kick each one a couple of
times, coalesce the tasklet execution to the end. This also ensures that
we have the chance to run the tasklet immediately after priority
bumping.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200331114852.11583-1-chris@chris-wilson.co.uk
We only call i915_schedule() when we know we have changed the priority
on a request and so require to propagate any change in priority to its
signalers (for PI). By unconditionally checking all of our signalers, we
avoid skipping changes made prior to construction of the request (as the
request may be waited upon before submission when used in parallel).
References: https://gitlab.freedesktop.org/drm/intel/issues/1318
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200306071614.2846708-2-chris@chris-wilson.co.uk
In preparation for making GEM execbuf parallel, we need to be prepared
to handle very early declaration of dependencies -- even before our
signaler has itself been submitted.
References: a79ca656b6 ("drm/i915: Push the wakeref->count deferral to the backend")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200220123608.1666271-1-chris@chris-wilson.co.uk
While we know that the waiters cannot disappear as we walk our list
(only that they might be added), the same cannot be said for our
signalers as they may be completed by the HW and retired as we process
this request. Ergo we need to use rcu to protect the list iteration and
remember to mark up the list_del_rcu.
v2: Mark the deps as safe-for-rcu
Fixes: 793c226173 ("drm/i915/gt: Protect execlists_hold/unhold from new waiters")
Fixes: 32ff621fd7 ("drm/i915/gt: Allow temporary suspension of inflight requests")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200220075025.1539375-1-chris@chris-wilson.co.uk
As we may add new waiters to a request as it is being run, we need to
mark the list iteration as being safe for concurrent addition.
v2: Mika spotted that we used the same trick for signalers_list, so warn
the compiler about the lockless walk there as well.
Fixes: 32ff621fd7 ("drm/i915/gt: Allow temporary suspension of inflight requests")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200207110213.2734386-1-chris@chris-wilson.co.uk
If we keep track of when the i915_request.sched.link is on the HW
runlist, or in the priority queue we can simplify our interactions with
the request (such as during rescheduling). This also simplifies the next
patch where we introduce a new in-between list, for requests that are
ready but neither on the run list or in the queue.
v2: Update i915_sched_node.link explanation for current usage where it
is a link on both the queue and on the runlists.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200116184754.2860848-1-chris@chris-wilson.co.uk
Keep the intel_context as being the primary state for i915_request, with
the GEM context a backpointer from the low level state for the rarer
cases we need client information. Our goal is to remove such references
to clients from the backend, and leave the HW submission agnostic to
client interfaces and self-contained.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Andi Shyti <andi.shyti@intel.com>
Reviewed-by: Andi Shyti <andi.shyti@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191220101230.256839-1-chris@chris-wilson.co.uk
We want the bonded request to have the same scheduler properties as its
master so that it is placed at the same depth in the queue. For example,
consider we have requests A, B and B', where B & B' are a bonded pair to
run in parallel on two engines.
A -> B
\- B'
B will run after A and so may be scheduled on an idle engine and wait on
A using a semaphore. B' sees B being executed and so enters the queue on
the same engine as A. As B' did not inherit the semaphore-chain from B,
it may have higher precedence than A and so preempts execution. However,
B' then sits on a semaphore waiting for B, who is waiting for A, who is
blocked by B.
Ergo B' needs to inherit the scheduler properties from B (i.e. the
semaphore chain) so that it is scheduled with the same priority as B and
will not be executed ahead of Bs dependencies.
Furthermore, to prevent the priorities changing via the expose fence on
B', we need to couple in the dependencies for PI. This requires us to
relax our sanity-checks that dependencies are strictly in order.
v2: Synchronise (B, B') execution on all platforms, regardless of using
a scheduler, any no-op syncs should be elided.
Fixes: ee1136908e ("drm/i915/execlists: Virtual engine bonding")
Closes: https://gitlab.freedesktop.org/drm/intel/issues/464
Testcase: igt/gem_exec_balancer/bonded-chain
Testcase: igt/gem_exec_balancer/bonded-semaphore
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191210151332.3902215-1-chris@chris-wilson.co.uk
As we start peeking into requests for longer and longer, e.g.
incorporating use of spinlocks when only protected by an
rcu_read_lock(), we need to be careful in how we reset the request when
recycling and need to preserve any barriers that may still be in use as
the request is reset for reuse.
Quoting Linus Torvalds:
> If there is refcounting going on then why use SLAB_TYPESAFE_BY_RCU?
.. because the object can be accessed (by RCU) after the refcount has
gone down to zero, and the thing has been released.
That's the whole and only point of SLAB_TYPESAFE_BY_RCU.
That flag basically says:
"I may end up accessing this object *after* it has been free'd,
because there may be RCU lookups in flight"
This has nothing to do with constructors. It's ok if the object gets
reused as an object of the same type and does *not* get
re-initialized, because we're perfectly fine seeing old stale data.
What it guarantees is that the slab isn't shared with any other kind
of object, _and_ that the underlying pages are free'd after an RCU
quiescent period (so the pages aren't shared with another kind of
object either during an RCU walk).
And it doesn't necessarily have to have a constructor, because the
thing that a RCU walk will care about is
(a) guaranteed to be an object that *has* been on some RCU list (so
it's not a "new" object)
(b) the RCU walk needs to have logic to verify that it's still the
*same* object and hasn't been re-used as something else.
In contrast, a SLAB_TYPESAFE_BY_RCU memory gets free'd and re-used
immediately, but because it gets reused as the same kind of object,
the RCU walker can "know" what parts have meaning for re-use, in a way
it couidn't if the re-use was random.
That said, it *is* subtle, and people should be careful.
> So the re-use might initialize the fields lazily, not necessarily using a ctor.
If you have a well-defined refcount, and use "atomic_inc_not_zero()"
to guard the speculative RCU access section, and use
"atomic_dec_and_test()" in the freeing section, then you should be
safe wrt new allocations.
If you have a completely new allocation that has "random stale
content", you know that it cannot be on the RCU list, so there is no
speculative access that can ever see that random content.
So the only case you need to worry about is a re-use allocation, and
you know that the refcount will start out as zero even if you don't
have a constructor.
So you can think of the refcount itself as always having a zero
constructor, *BUT* you need to be careful with ordering.
In particular, whoever does the allocation needs to then set the
refcount to a non-zero value *after* it has initialized all the other
fields. And in particular, it needs to make sure that it uses the
proper memory ordering to do so.
NOTE! One thing to be very worried about is that re-initializing
whatever RCU lists means that now the RCU walker may be walking on the
wrong list so the walker may do the right thing for this particular
entry, but it may miss walking *other* entries. So then you can get
spurious lookup failures, because the RCU walker never walked all the
way to the end of the right list. That ends up being a much more
subtle bug.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191122094924.629690-1-chris@chris-wilson.co.uk
Since the execlists_active() is no longer protected by the
engine->active.lock, we need to protect the request pointer with RCU to
prevent it being freed as we evaluate whether or not we need to preempt.
Fixes: df40306902 ("drm/i915/execlists: Lift process_csb() out of the irq-off spinlock")
Fixes: 13ed13a4dc ("drm/i915: Don't set queue_priority_hint if we don't kick the submission")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191104090158.2959-2-chris@chris-wilson.co.uk
If we change the priority of the active context, then it has no impact
on the decision of whether to preempt the active context -- we don't
preempt the context with itself. In this situation, we elide the tasklet
rescheduling and should *not* be marking up the queue_priority_hint as
that may mask a later submission where we decide we don't have to kick
the tasklet as a higher priority submission is pending (spoiler alert,
it was not).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191021080226.537-1-chris@chris-wilson.co.uk
Normally, we try and skip submission if ELSP[1] is filled. However, we
may desire to enable timeslicing due to the queue priority, even if
ELSP[1] itself does not require timeslicing. That is the queue is equal
priority to ELSP[0] and higher priority then ELSP[1]. Previously, we
would wait until the context switch to preempt the current ELSP[1], but
with timeslicing, we want to preempt ELSP[0] and replace it with the
queue.
In writing the test case, it become quickly apparent that we were also
suppressing the tasklet during promotion and so failing to notice when
the queue started requiring timeslicing.
Fixes: 2229adc813 ("drm/i915/execlist: Trim immediate timeslice expiry")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191018072027.31948-1-chris@chris-wilson.co.uk
If we have multiple contexts of equal priority pending execution,
activate a timer to demote the currently executing context in favour of
the next in the queue when that timeslice expires. This enforces
fairness between contexts (so long as they allow preemption -- forced
preemption, in the future, will kick those who do not obey) and allows
us to avoid userspace blocking forward progress with e.g. unbounded
MI_SEMAPHORE_WAIT.
For the starting point here, we use the jiffie as our timeslice so that
we should be reasonably efficient wrt frequent CPU wakeups.
Testcase: igt/gem_exec_scheduler/semaphore-resolve
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190620142052.19311-2-chris@chris-wilson.co.uk
When using a global seqno, we required a precise stop-the-workd event to
handle preemption and unwind the global seqno counter. To accomplish
this, we would preempt to a special out-of-band context and wait for the
machine to report that it was idle. Given an idle machine, we could very
precisely see which requests had completed and which we needed to feed
back into the run queue.
However, now that we have scrapped the global seqno, we no longer need
to precisely unwind the global counter and only track requests by their
per-context seqno. This allows us to loosely unwind inflight requests
while scheduling a preemption, with the enormous caveat that the
requests we put back on the run queue are still _inflight_ (until the
preemption request is complete). This makes request tracking much more
messy, as at any point then we can see a completed request that we
believe is not currently scheduled for execution. We also have to be
careful not to rewind RING_TAIL past RING_HEAD on preempting to the
running context, and for this we use a semaphore to prevent completion
of the request before continuing.
To accomplish this feat, we change how we track requests scheduled to
the HW. Instead of appending our requests onto a single list as we
submit, we track each submission to ELSP as its own block. Then upon
receiving the CS preemption event, we promote the pending block to the
inflight block (discarding what was previously being tracked). As normal
CS completion events arrive, we then remove stale entries from the
inflight tracker.
v2: Be a tinge paranoid and ensure we flush the write into the HWS page
for the GPU semaphore to pick in a timely fashion.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190620142052.19311-1-chris@chris-wilson.co.uk
To continue the onslaught of removing the assumption of a global
execution ordering, another casualty is the engine->timeline. Without an
actual timeline to track, it is overkill and we can replace it with a
much less grand plain list. We still need a list of requests inflight,
for the simple purpose of finding inflight requests (for retiring,
resetting, preemption etc).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190614164606.15633-3-chris@chris-wilson.co.uk
Having allowed the user to define a set of engines that they will want
to only use, we go one step further and allow them to bind those engines
into a single virtual instance. Submitting a batch to the virtual engine
will then forward it to any one of the set in a manner as best to
distribute load. The virtual engine has a single timeline across all
engines (it operates as a single queue), so it is not able to concurrently
run batches across multiple engines by itself; that is left up to the user
to submit multiple concurrent batches to multiple queues. Multiple users
will be load balanced across the system.
The mechanism used for load balancing in this patch is a late greedy
balancer. When a request is ready for execution, it is added to each
engine's queue, and when an engine is ready for its next request it
claims it from the virtual engine. The first engine to do so, wins, i.e.
the request is executed at the earliest opportunity (idle moment) in the
system.
As not all HW is created equal, the user is still able to skip the
virtual engine and execute the batch on a specific engine, all within the
same queue. It will then be executed in order on the correct engine,
with execution on other virtual engines being moved away due to the load
detection.
A couple of areas for potential improvement left!
- The virtual engine always take priority over equal-priority tasks.
Mostly broken up by applying FQ_CODEL rules for prioritising new clients,
and hopefully the virtual and real engines are not then congested (i.e.
all work is via virtual engines, or all work is to the real engine).
- We require the breadcrumb irq around every virtual engine request. For
normal engines, we eliminate the need for the slow round trip via
interrupt by using the submit fence and queueing in order. For virtual
engines, we have to allow any job to transfer to a new ring, and cannot
coalesce the submissions, so require the completion fence instead,
forcing the persistent use of interrupts.
- We only drip feed single requests through each virtual engine and onto
the physical engines, even if there was enough work to fill all ELSP,
leaving small stalls with an idle CS event at the end of every request.
Could we be greedy and fill both slots? Being lazy is virtuous for load
distribution on less-than-full workloads though.
Other areas of improvement are more general, such as reducing lock
contention, reducing dispatch overhead, looking at direct submission
rather than bouncing around tasklets etc.
sseu: Lift the restriction to allow sseu to be reconfigured on virtual
engines composed of RENDER_CLASS (rcs).
v2: macroize check_user_mbz()
v3: Cancel virtual engines on wedging
v4: Commence commenting
v5: Replace 64b sibling_mask with a list of class:instance
v6: Drop the one-element array in the uabi
v7: Assert it is an virtual engine in to_virtual_engine()
v8: Skip over holes in [class][inst] so we can selftest with (vcs0, vcs2)
Link: https://github.com/intel/media-driver/pull/283
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190521211134.16117-6-chris@chris-wilson.co.uk
The handling of the no-preemption priority level imposes the restriction
that we need to maintain the implied ordering even though preemption is
disabled. Otherwise we may end up with an AB-BA deadlock across multiple
engine due to a real preemption event reordering the no-preemption
WAITs. To resolve this issue we currently promote all requests to WAIT
on unsubmission, however this interferes with the timeslicing
requirement that we do not apply any implicit promotion that will defeat
the round-robin timeslice list. (If we automatically promote the active
request it will go back to the head of the queue and not the tail!)
So we need implicit promotion to prevent reordering around semaphores
where we are not allowed to preempt, and we must avoid implicit
promotion on unsubmission. So instead of at unsubmit, if we apply that
implicit promotion on adding the dependency, we avoid the semaphore
deadlock and we also reduce the gains made by the promotion for user
space waiting. Furthermore, by keeping the earlier dependencies at a
higher level, we reduce the search space for timeslicing without
altering runtime scheduling too badly (no dependencies at all will be
assigned a higher priority for rrul).
v2: Limit the bump to external edges (as originally intended) i.e.
between contexts and out to the user.
Testcase: igt/gem_concurrent_blit
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190515130052.4475-3-chris@chris-wilson.co.uk
In all likelihood, the priority and node are already in the CPU cache
and by checking them first, we can avoid having to chase the
*request->hwsp for the current breadcrumb.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190513120102.29660-3-chris@chris-wilson.co.uk
To simplify the next patch, update bump_priority and schedule to accept
the internal i915_sched_ndoe directly and not expect a request pointer.
add/remove: 0/0 grow/shrink: 2/1 up/down: 8/-15 (-7)
Function old new delta
i915_schedule_bump_priority 109 113 +4
i915_schedule 50 54 +4
__i915_schedule 922 907 -15
v2: Adopt node for the old rq local, since it no longer is a request but
the origin node.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190513120102.29660-2-chris@chris-wilson.co.uk
If we couple the scheduler more tightly with the execlists policy, we
can apply the preemption policy to the question of whether we need to
kick the tasklet at all for this priority bump.
v2: Rephrase it as a core i915 policy and not an execlists foible.
v3: Pull the kick together.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190507122544.12698-1-chris@chris-wilson.co.uk
Consider two tasks that are running in parallel on a pair of engines
(vcs0, vcs1), but then must complete on a shared engine (rcs0). To
maximise throughput, we want to run the first ready task on rcs0 (i.e.
the first task that completes on either of vcs0 or vcs1). When using
semaphores, however, we will instead queue onto rcs in submission order.
To resolve this incorrect ordering, we want to re-evaluate the priority
queue when each of the request is ready. Normally this happens because
we only insert into the priority queue requests that are ready, but with
semaphores we are inserting ahead of their readiness and to compensate
we penalize those tasks with reduced priority (so that tasks that do not
need to busywait should naturally be run first). However, given a series
of tasks that each use semaphores, the queue degrades into submission
fifo rather than readiness fifo, and so to counter this we give a small
boost to semaphore users as their dependent tasks are completed (and so
we no longer require any busywait prior to running the user task as they
are then ready themselves).
v2: Fixup irqsave for schedule_lock (Tvrtko)
Testcase: igt/gem_exec_schedule/semaphore-codependency
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>
Cc: Dmitry Ermilov <dmitry.ermilov@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190409152922.23894-1-chris@chris-wilson.co.uk
Ideally we only need one semaphore per ring to accommodate waiting on
multiple engines in parallel. However, since we do not know which fences
we will finally be waiting on, we emit a semaphore for every fence. It
turns out to be quite easy to trick ourselves into exhausting our
ringbuffer causing an error, just by feeding in a batch that depends on
several thousand contexts.
Since we never can be waiting on more than one semaphore in parallel
(other than perhaps the desire to busywait on multiple engines), just
pick the first fence for our semaphore. If we pick the wrong fence to
busywait on, we just miss an opportunity to reduce latency.
An adaption might be to use sched.flags as either a semaphore counter,
or to track the first busywait on each engine, converting it back to a
single use bit prior to closing the request.
v2: Track first semaphore used per-engine (this caters for our basic
igt that semaphores are working).
Reported-by: Mika Kuoppala <mika.kuoppala@intel.com>
Testcase: igt/gem_exec_fence/long-history
Fixes: e886196469 ("drm/i915: Use HW semaphores for inter-engine synchronisation on gen8+")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190401162641.10963-3-chris@chris-wilson.co.uk
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Rather than manually add every new global into each hook, use
i915_global_register() function and keep a list of registered globals to
invoke instead.
However, I haven't found a way for random drivers to add an .init table
to avoid having to manually add ourselves to i915_globals_init() each
time.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20190305213830.18094-1-chris@chris-wilson.co.uk
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
We don't want to busywait on the GPU if we have other work to do. If we
give non-busywaiting workloads higher (initial) priority than workloads
that require a busywait, we will prioritise work that is ready to run
immediately. We then also have to be careful that we don't give earlier
semaphores an accidental boost because later work doesn't wait on other
rings, hence we keep a history of semaphore usage of the dependency chain.
v2: Stop rolling the bits into a chain and just use a flag in case this
request or any of our dependencies use a semaphore. The rolling around
was contagious as Tvrtko was heard to fall off his chair.
Testcase: igt/gem_exec_schedule/semaphore
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190301170901.8340-4-chris@chris-wilson.co.uk
WAIT is occasionally suppressed by virtue of preempted requests being
promoted to NEWCLIENT if they have not all ready received that boost.
Make this consistent for all WAIT boosts that they are not allowed to
preempt executing contexts and are merely granted the right to be at the
front of the queue for the next execution slot. This is in keeping with
the desire that the WAIT boost be a minor tweak that does not give
excessive promotion to its user and open ourselves to trivial abuse.
The problem with the inconsistent WAIT preemption becomes more apparent
as the preemption is propagated across the engines, where one engine may
preempt and the other not, and we be relying on the exact execution
order being consistent across engines (e.g. using HW semaphores to
coordinate parallel execution).
v2: Also protect GuC submission from false preemption loops.
v3: Build bug safeguards and better debug messages for st.
v4: Do the priority bumping in unsubmit (i.e. on preemption/reset
unwind), applying it earlier during submit causes out-of-order execution
combined with execute fences.
v5: Call sw_fence_fini for our dummy request (Matthew)
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190228220639.3173-1-chris@chris-wilson.co.uk
