As the engine->kernel_context is used within the engine-pm barrier, we
have to be careful when emitting requests outside of the barrier, as the
strict timeline locking rules do not apply. Instead, we must ensure the
engine_park() cannot be entered as we build the request, which is
simplest by taking an explicit engine-pm wakeref around the request
construction.
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/20191125105858.1718307-1-chris@chris-wilson.co.uk
Bonded request submission is designed to allow requests to execute in
parallel as laid out by the user. If the master request is already
finished before its bonded pair is submitted, the pair were not destined
to run in parallel and we lose the information about the master engine
to dictate selection of the secondary. If the second request was
required to be run on a particular engine in a virtual set, that should
have been specified, rather than left to the whims of a random
unconnected requests!
In the selftest, I made the mistake of not ensuring the master would
overlap with its bonded pairs, meaning that it could indeed complete
before we submitted the bonds. Those bonds were then free to select any
available engine in their virtual set, and not the one expected by the
test.
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/20191122112152.660743-1-chris@chris-wilson.co.uk
Verify that we can execute a long chain of dependent requests from
userspace, each one slightly more important than the last.
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/20191114225736.616885-4-chris@chris-wilson.co.uk
An interesting observation made with our parallel selftests was that on
our small/single cpu systems we would call kthread_stop() before the
kthreads were spawned. If this happens, the kthread is never run at all;
completely bypassing the test.
A simple yield() from the parent will ensure that all children have the
opportunity to start before we reap them.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191101084940.31838-1-chris@chris-wilson.co.uk
Execlists uses a scheduling quantum (a timeslice) to alternate execution
between ready-to-run contexts of equal priority. This ensures that all
users (though only if they of equal importance) have the opportunity to
run and prevents livelocks where contexts may have implicit ordering due
to userspace semaphores. However, not all workloads necessarily benefit
from timeslicing and in the extreme some sysadmin may want to disable or
reduce the timeslicing granularity.
The timeslicing mechanism can be compiled out^W^W disabled (but should
DCE!) with
./scripts/config --set-val DRM_I915_TIMESLICE_DURATION 0
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191029091632.26281-1-chris@chris-wilson.co.uk
As we use hard coded offsets for a few locations within the context
image, include those in the selftests to assert that they are valid.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191028121803.29408-1-chris@chris-wilson.co.uk
The location of RING_MI_MODE (used to stop the ring across resets) moved
for Tigerlake. Fixup the new location and include a selftest to verify
the location in the default context image.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Acked-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191026082220.32632-1-chris@chris-wilson.co.uk
On schedule-out (CS completion) of a banned context, scrub the context
image so that we do not replay the active payload. The intent is that we
skip banned payloads on request submission so that the timeline
advancement continues on in the background. However, if we are returning
to a preempted request, i915_request_skip() is ineffective and instead we
need to patch up the context image so that it continues from the start
of the next request.
v2: Fixup cancellation so that we only scrub the payload of the active
request and do not short-circuit the breadcrumbs (which might cause
other contexts to execute out of order).
v3: Grammar pass
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/20191023133108.21401-3-chris@chris-wilson.co.uk
If the preempted context takes too long to relinquish control, e.g. it
is stuck inside a shader with arbitration disabled, evict that context
with an engine reset. This ensures that preemptions are reasonably
responsive, providing a tighter QoS for the more important context at
the cost of flagging unresponsive contexts more frequently (i.e. instead
of using an ~10s hangcheck, we now evict at ~100ms). The challenge of
lies in picking a timeout that can be reasonably serviced by HW for
typical workloads, balancing the existing clients against the needs for
responsiveness.
Note that coupled with timeslicing, this will lead to rapid GPU "hang"
detection with multiple active contexts vying for GPU time.
The forced preemption mechanism can be compiled out with
./scripts/config --set-val DRM_I915_PREEMPT_TIMEOUT 0
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/20191023133108.21401-2-chris@chris-wilson.co.uk
The actual conditions are that we know the GPU is not accessing the
context, and we hold a pin on the context image to allow CPU access. We
used a fake lock on ce->pin_mutex so that we could try and use lockdep
to assert that access is serialised, but the various different
hardirq/softirq contexts where we need to *fake* holding the pin_mutex
are causing more trouble.
Still it would be nice if we did have a way to reassure ourselves that
the direct update to the context image is serialised with GPU execution.
In the meantime, stop lockdep complaining about false irq inversions.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111923
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022122845.25038-1-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
Where the function, or code segment, operates on intel_gt, we need to
start passing it instead of i915 to for_each_engine(_masked).
This is another partial step in migration of i915->engines[] to
gt->engines[].
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
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/20191017094500.21831-2-tvrtko.ursulin@linux.intel.com
We want the general purpose registers to be clear in all new contexts so
that we can be confident that no information is leaked from one to the
next.
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/20191014090757.32111-7-chris@chris-wilson.co.uk
Check the logical ring context by asserting that the registers hold
expected start during execution. (It's a bit chicken-and-egg for how
could we manage to execute our request if the registers were not being
updated. Still, it's nice to verify that the HW is working as expected.)
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/20191014090757.32111-6-chris@chris-wilson.co.uk
Keep track of the GEM contexts underneath i915->gem.contexts and assign
them their own lock for the purposes of list management.
v2: Focus on lock tracking; ctx->vm is protected by ctx->mutex
v3: Correct split with removal of logical HW ID
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/20191004134015.13204-15-chris@chris-wilson.co.uk
We don't need to hold struct_mutex now for retiring requests, so drop it
from i915_retire_requests() and i915_gem_wait_for_idle(), finally
removing I915_WAIT_LOCKED for good.
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/20191004134015.13204-8-chris@chris-wilson.co.uk
Forgo the struct_mutex serialisation for i915_active, and interpose its
own mutex handling for active/retire.
This is a multi-layered sleight-of-hand. First, we had to ensure that no
active/retire callbacks accidentally inverted the mutex ordering rules,
nor assumed that they were themselves serialised by struct_mutex. More
challenging though, is the rule over updating elements of the active
rbtree. Instead of the whole i915_active now being serialised by
struct_mutex, allocations/rotations of the tree are serialised by the
i915_active.mutex and individual nodes are serialised by the caller
using the i915_timeline.mutex (we need to use nested spinlocks to
interact with the dma_fence callback lists).
The pain point here is that instead of a single mutex around execbuf, we
now have to take a mutex for active tracker (one for each vma, context,
etc) and a couple of spinlocks for each fence update. The improvement in
fine grained locking allowing for multiple concurrent clients
(eventually!) should be worth it in typical loads.
v2: Add some comments that barely elucidate anything :(
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/20191004134015.13204-6-chris@chris-wilson.co.uk
If execlists's lite-restore is based on the common GEM context tag
rather than the per-intel_context LRCA, then a context switch between
two intel_contexts on the same engine derived from the same GEM context
will perform a lite-restore instead of a full context switch. We can
exploit this by poisoning the ringbuffer of the first context and trying
to trick a simple RING_TAIL update (i.e. lite-restore)
v2: Also check what happens if preempt ce[0] with ce[1] (both instances
on the same engine from the same parent context) [Tvrtko]
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/20191002183459.26614-1-chris@chris-wilson.co.uk
As we execute GPU resets on a gt/ basis, and use the intel_gt as the
primary for all other reset functions, also use it for the has-reset?
predicates. Gradually simplifying the churn of pointers.
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/20190927211749.2181-1-chris@chris-wilson.co.uk
Before we submit the first context to HW, we need to construct a valid
image of the register state. This layout is defined by the HW and should
match the layout generated by HW when it saves the context image.
Asserting that this should be equivalent should help avoid any undefined
behaviour and verify that we haven't missed anything important!
Of course, having insisted that the initial register state within the
LRC should match that returned by HW, we need to ensure that it does.
v2: Drop the RELATIVE_MMIO flag from gen11, we ignore it for
constructing the lrc image.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190924145950.3011-1-chris@chris-wilson.co.uk
The request->timeline is only valid until the request is retired (i.e.
before it is completed). Upon retiring the request, the context may be
unpinned and freed, and along with it the timeline may be freed. We
therefore need to be very careful when chasing rq->timeline that the
pointer does not disappear beneath us. The vast majority of users are in
a protected context, either during request construction or retirement,
where the timeline->mutex is held and the timeline cannot disappear. It
is those few off the beaten path (where we access a second timeline) that
need extra scrutiny -- to be added in the next patch after first adding
the warnings about dangerous access.
One complication, where we cannot use the timeline->mutex itself, is
during request submission onto hardware (under spinlocks). Here, we want
to check on the timeline to finalize the breadcrumb, and so we need to
impose a second rule to ensure that the request->timeline is indeed
valid. As we are submitting the request, it's context and timeline must
be pinned, as it will be used by the hardware. Since it is pinned, we
know the request->timeline must still be valid, and we cannot submit the
idle barrier until after we release the engine->active.lock, ergo while
submitting and holding that spinlock, a second thread cannot release the
timeline.
v2: Don't be lazy inside selftests; hold the timeline->mutex for as long
as we need it, and tidy up acquiring the timeline with a bit of
refactoring (i915_active_add_request)
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/20190919111912.21631-1-chris@chris-wilson.co.uk
Make sure that when submitting requests, we always serialize against
potential vma moves and clflushes.
Time for a i915_request_await_vma() interface!
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190819112033.30638-1-chris@chris-wilson.co.uk
The order in which we store the engines inside default_engines() for the
legacy ctx->engines[] has to match the legacy I915_EXEC_RING selector
mapping in execbuf::user_map. If we present VCS2 as being the second
instance of the video engine, legacy userspace calls that I915_EXEC_BSD2
and so we need to insert it into the second video slot.
v2: Record the legacy mapping (hopefully we can remove this need in the
future)
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111328
Fixes: 2edda80db3 ("drm/i915: Rename engines to match their user interface")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com> #v1
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190808110612.23539-2-chris@chris-wilson.co.uk
To maintain a fast lookup from a GT centric irq handler, we want the
engine lookup tables on the intel_gt. To avoid having multiple copies of
the same multi-dimension lookup table, move the generic user engine
lookup into an rbtree (for fast and flexible indexing).
v2: Split uabi_instance cf uabi_class
v3: Set uabi_class/uabi_instance after collating all engines to provide a
stable uabi across parallel unordered construction.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> #v2
Link: https://patchwork.freedesktop.org/patch/msgid/20190806124300.24945-2-chris@chris-wilson.co.uk
Teach igt_spinner to only use our internal structs, decoupling the
interface from the GEM contexts. This makes it easier to avoid
requiring ce->gem_context back references for kernel_context that may
have them in future.
v2: Lift engine lock to verify_wa() caller.
v3: Less than v2, but more so
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/20190731081126.9139-1-chris@chris-wilson.co.uk
Preempt-to-busy uses a GPU semaphore to enforce an idle-barrier across
preemption, but mediated gvt does not fully support semaphores.
v2: Fiddle around with the flags and settle on using has-semaphores for
the core bits so that we retain the ability to preempt our own
semaphores.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Zhenyu Wang <zhenyuw@linux.intel.com>
Cc: Xiaolin Zhang <xiaolin.zhang@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Acked-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190709091233.8573-1-chris@chris-wilson.co.uk
GVT forces single port submission of individual requests. We do not
enjoy the context amalgamation that the test depends upon for setting up
the test (where port 0 has a large number of requests with a priority
change somewhere in the middle). Under single request submission of gvt
it is quite able for the preemption event to occur while another context
is active and so there be a real need to act upon that preemption.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111108
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Zhenyu Wang <zhenyuw@linux.intel.com>
Acked-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190712082549.25053-1-chris@chris-wilson.co.uk
Having taken the first step in encapsulating the functionality by moving
the related files under gt/, the next step is to start encapsulating by
passing around the relevant structs rather than the global
drm_i915_private. In this step, we pass intel_gt to intel_reset.c
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190712192953.9187-1-chris@chris-wilson.co.uk
We want to set this flag in the next commit on requests containing
perf queries so that the result of the perf query can just be a delta
of global counters, rather than doing post processing of the OA
buffer.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
[ickle: add basic selftest for nopreempt]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190709164227.25859-1-chris@chris-wilson.co.uk
We frequently, but not frequently enough!, remember to flush residual
operations and objects at the end of a live subtest. The purpose is to
cleanup after every subtest, leaving a clean slate for the next subtest,
and perform early detection of leaky state. As this should ideally be
common for all live subtests, pull the task into a common teardown
routine.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190703091726.11690-1-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
Since commit eb8d0f5af4 ("drm/i915: Remove GPU reset dependence on
struct_mutex"), the I915_WAIT_LOCKED flags passed to i915_request_wait()
has been defunct. Now go ahead and remove it from all callers.
References: eb8d0f5af4 ("drm/i915: Remove GPU reset dependence on struct_mutex")
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/20190618074153.16055-3-chris@chris-wilson.co.uk
The functions where internally already only using the structure, so we
need to just flip the interface.
v2: rebase
Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20190613232156.34940-7-daniele.ceraolospurio@intel.com
We should check "request[n]" instead of just "request".
Fixes: 78e41ddd21 ("drm/i915: Apply an execution_mask to the virtual_engine")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.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/20190529110355.GA19119@mwanda
Use the per-object local lock to control the cache domain of the
individual GEM objects, not struct_mutex. This is a huge leap forward
for us in terms of object-level synchronisation; execbuffers are
coordinated using the ww_mutex and pread/pwrite is finally fully
serialised again.
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/20190528092956.14910-10-chris@chris-wilson.co.uk
Some users require that when a master batch is executed on one particular
engine, a companion batch is run simultaneously on a specific slave
engine. For this purpose, we introduce virtual engine bonding, allowing
maps of master:slaves to be constructed to constrain which physical
engines a virtual engine may select given a fence on a master engine.
For the moment, we continue to ignore the issue of preemption deferring
the master request for later. Ideally, we would like to then also remove
the slave and run something else rather than have it stall the pipeline.
With load balancing, we should be able to move workload around it, but
there is a similar stall on the master pipeline while it may wait for
the slave to be executed. At the cost of more latency for the bonded
request, it may be interesting to launch both on their engines in
lockstep. (Bubbles abound.)
Opens: Also what about bonding an engine as its own master? It doesn't
break anything internally, so allow the silliness.
v2: Emancipate the bonds
v3: Couple in delayed scheduling for the selftests
v4: Handle invalid mutually exclusive bonding
v5: Mention what the uapi does
v6: s/nbond/num_bonds/
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/20190521211134.16117-9-chris@chris-wilson.co.uk
Allow the user to direct which physical engines of the virtual engine
they wish to execute one, as sometimes it is necessary to override the
load balancing algorithm.
v2: Only kick the virtual engines on context-out if required
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/20190521211134.16117-7-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
