In the next patch, we will want to configure the slave request
depending on which physical engine the master request is executed on.
For this, we introduce a callback from the execute fence to convey this
information.
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/20190521211134.16117-8-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
Commit 1413b2bc07 ("drm/i915: Trim NEWCLIENT boosting") had the
intended consequence of not allowing a sequence of work that merely
crossed into a new engine the privilege to be promoted to NEWCLIENT
status. It also had the unintended consequence of actually making
NEWCLIENT effective on heavily oversubscribed transcode machines and
impacting upon their throughput.
If we consider a client packet composed of (rcsA, rcsB, vcs) and 30 of
those clients, using the NEWCLIENT boost that will be scheduled as
rcsA x 30, (rcsB, vcs) x 30
where as before it would have been
(rcsA, rcsB, vcs) x 30
That is with NEWCLIENT only boosting the first request of each client,
we would execute all rcsA requests prior to running on the vcs engines;
acruing a lot of dead time as compared to the previous case where the
vcs engine would be started in parallel to processing the second client.
The previous patch has the effect of delaying submission until it is
required by a third party (either the user with an explicit wait, or by
another client/engine). We reduce the NEWCLIENT bump to a mere WAIT,
which has the effect of removing its preemptive grant and reducing it to
the same level as any other user interaction -- that it will not be
promoted above the interengine dependencies, and so preventing NEWCLIENTS
from starving other engines. This a large nerf to the rrul properties of
the current NEWCLIENT, but it still does give prioritised submission to
new requests from light workloads.
References: b16c765122 ("drm/i915: Priority boost for new clients")
Fixes: 1413b2bc07 ("drm/i915: Trim NEWCLIENT boosting") # customer impact
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/20190515130052.4475-4-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 commit b7404c7ecb ("drm/i915: Bump ready tasks ahead of
busywaits"), I tried cutting a corner in order to not install a signal
for each of our dependencies, and only listened to requests on which we
were intending to busywait. The compromise that was made was that
instead of then being able to promote the request with a full
NOSEMAPHORE like its non-busywaiting brethren, as we had not ensured we
had cleared the semaphore chain, we settled for only using the NEWCLIENT
boost. With an over saturated system with multiple NEWCLIENTS in flight
at any time, this was found to be an inadequate promotion and left us
with a much poorer scheduling order than prior to using semaphores.
The outcome of this patch, is that all requests have NOSEMAPHORE
priority when they have no dependencies and are ready to run and not
busywait, restoring the pre-semaphore ordering on saturated systems.
We can demonstrate the effect of poor scheduling order by oversaturating
the system using gem_wsim on a system with multiple vcs engines
(i.e running the same workloads across more clients than required for
peak throughput, e.g. media_load_balance_17i7.wsim -c4 -b context):
x v5.1 (normalized)
+ tip
* fix
+------------------------------------------------------------------------+
| x |
| x |
| x |
| x |
| %x |
| %%x |
| %%x |
| %%x |
| %%x |
| %%x |
| %%x |
| %%x |
| %%x |
| %%x |
| %%x |
| %#x |
| %#x |
| %#x |
| %#x |
| %#x |
| + %#xx |
| + %#xx |
| + %%#xx |
| + %%#xx |
| + %%#xx |
| + %%#xx |
| + %%##x |
| +++ %%##x |
| +++ %%##x |
| +++ %%##x |
| ++++ %%##x |
| ++++ %%##x |
| ++++ %%##xx |
| ++++ %###xx |
| ++++ %###xx |
| ++++ %###xx |
| ++++ %###xx |
| ++++ + %#O#xx |
| ++++ + %#O#xx |
| ++++++ + %#O#xx |
| ++++++++++ %OOOxxx|
| ++++++++++ + %#OOO#xx|
| + ++++++++++++ ++ +++++ + ++ @@OOOO#xx|
| |A_| |
||__________M_______A____________________| |
| |A_| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 120 0.99456 1.00628 0.999985 1.0001545 0.0024387139
+ 120 0.873021 1.00037 0.884134 0.90148752 0.039190862
Difference at 99.5% confidence
-0.098667 +/- 0.0110762
-9.86517% +/- 1.10745%
(Student's t, pooled s = 0.0277657)
% 120 0.990207 1.00165 0.9970265 0.99699748 0.0021024
Difference at 99.5% confidence
-0.003157 +/- 0.000908245
-0.315651% +/- 0.0908105%
(Student's t, pooled s = 0.00227678)
Fixes: b7404c7ecb ("drm/i915: Bump ready tasks ahead of busywaits")
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/20190515130052.4475-2-chris@chris-wilson.co.uk
Avoid charging us for the presumed busywait if the request was preempted
after successfully using semaphores to reduce inter-engine latency.
v2: Bump the priority to reflect the lack of semaphores now required.
References: ca6e56f654 ("drm/i915: Disable semaphore busywaits on saturated systems")
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/20190515130052.4475-1-chris@chris-wilson.co.uk
Currently there is an underlying assumption that i915_request_unsubmit()
is synchronous wrt the GPU -- that is the request is no longer in flight
as we remove it. In the near future that may change, and this may upset
our signaling as we can process an interrupt for that request while it
is no longer in flight.
CPU0 CPU1
intel_engine_breadcrumbs_irq
(queue request completion)
i915_request_cancel_signaling
... ...
i915_request_enable_signaling
dma_fence_signal
Hence in the time it took us to drop the lock to signal the request, a
preemption event may have occurred and re-queued the request. In the
process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and
so reused the rq->signal_link that was in use on CPU0, leading to bad
pointer chasing in intel_engine_breadcrumbs_irq.
A related issue was that if someone started listening for a signal on a
completed but no longer in-flight request, we missed the opportunity to
immediately signal that request.
Furthermore, as intel_contexts may be immediately released during
request retirement, in order to be entirely sure that
intel_engine_breadcrumbs_irq may no longer dereference the intel_context
(ce->signals and ce->signal_link), we must wait for irq spinlock.
In order to prevent the race, we use a bit in the fence.flags to signal
the transfer onto the signal list inside intel_engine_breadcrumbs_irq.
For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then
quickly signals to any outside observer that the fence is indeed signaled.
v2: Sketch out potential dma-fence API for manual signaling
v3: And the test_and_set_bit()
Fixes: 52c0fdb25c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking")
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/20190508112452.18942-1-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
Acquiring the signaler's timeline takes an active reference to their
HWSP that we would like to avoid if possible, so take it after
performing all of our allocations required to set up the fencing. The
acquisition also provides the final check that the target has not
already signaled allowing us to avoid the semaphore at the last moment.
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/20190503140239.32668-1-chris@chris-wilson.co.uk
Asking the GPU to busywait on a memory address, perhaps not unexpectedly
in hindsight for a shared system, leads to bus contention that affects
CPU programs trying to concurrently access memory. This can manifest as
a drop in transcode throughput on highly over-saturated workloads.
The only clue offered by perf, is that the bus-cycles (perf stat -e
bus-cycles) jumped by 50% when enabling semaphores. This corresponds
with extra CPU active cycles being attributed to intel_idle's mwait.
This patch introduces a heuristic to try and detect when more than one
client is submitting to the GPU pushing it into an oversaturated state.
As we already keep track of when the semaphores are signaled, we can
inspect their state on submitting the busywait batch and if we planned
to use a semaphore but were too late, conclude that the GPU is
overloaded and not try to use semaphores in future requests. In
practice, this means we optimistically try to use semaphores for the
first frame of a transcode job split over multiple engines, and fail if
there are multiple clients active and continue not to use semaphores for
the subsequent frames in the sequence. Periodically, we try to
optimistically switch semaphores back on whenever the client waits to
catch up with the transcode results.
With 1 client, on Broxton J3455, with the relative fps normalized by %cpu:
x no semaphores
+ drm-tip
* patched
+------------------------------------------------------------------------+
| * |
| *+ |
| **+ |
| **+ x |
| x * +**+ x |
| x x * * +***x xx |
| x x * * *+***x *x |
| x x* + * * *****x *x x |
| + x xx+x* + *** * ********* x * |
| + x xx+x* * *** +** ********* xx * |
| * + ++++* + x*x****+*+* ***+*************+x* * |
|*+ +** *+ + +* + *++****** *xxx**********x***+*****************+*++ *|
| |__________A_____M_____| |
| |_______________A____M_________| |
| |____________A___M________| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 120 2.60475 3.50941 3.31123 3.2143953 0.21117399
+ 120 2.3826 3.57077 3.25101 3.1414161 0.28146407
Difference at 95.0% confidence
-0.0729792 +/- 0.0629585
-2.27039% +/- 1.95864%
(Student's t, pooled s = 0.248814)
* 120 2.35536 3.66713 3.2849 3.2059917 0.24618565
No difference proven at 95.0% confidence
With 10 clients over-saturating the pipeline:
x no semaphores
+ drm-tip
* patched
+------------------------------------------------------------------------+
| ++ ** |
| ++ ** |
| ++ ** |
| ++ ** |
| ++ xx *** |
| ++ xx *** |
| ++ xxx*** |
| ++ xxx*** |
| +++ xxx*** |
| +++ xx**** |
| +++ xx**** |
| +++ xx**** |
| +++ xx**** |
| ++++ xx**** |
| +++++ xx**** |
| +++++ x x****** |
| ++++++ xxx******* |
| ++++++ xxx******* |
| ++++++ xxx******* |
| ++++++ xx******** |
| ++++++ xxxx******** |
| ++++++ xxxx******** |
| ++++++++ xxxxx********* |
|+ + + + ++++++++ xxx*xx**********x* *|
| |__A__| |
| |__AM__| |
| |__A_| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 120 2.47855 2.8972 2.72376 2.7193402 0.074604933
+ 120 1.17367 1.77459 1.71977 1.6966782 0.085850697
Difference at 95.0% confidence
-1.02266 +/- 0.0203502
-37.607% +/- 0.748352%
(Student's t, pooled s = 0.0804246)
* 120 2.57868 3.00821 2.80142 2.7923878 0.058646477
Difference at 95.0% confidence
0.0730476 +/- 0.0169791
2.68622% +/- 0.624383%
(Student's t, pooled s = 0.0671018)
Indicating that we've recovered the regression from enabling semaphores
on this saturated setup, with a hint towards an overall improvement.
Very similar, but of smaller magnitude, results are observed on both
Skylake(gt2) and Kabylake(gt4). This may be due to the reduced impact of
bus-cycles, where we see a 50% hit on Broxton, it is only 10% on the big
core, in this particular test.
One observation to make here is that for a greedy client trying to
maximise its own throughput, using semaphores is the right choice. It is
only the holistic system-wide view that semaphores of one client
impacts another and reduces the overall throughput where we would choose
to disable semaphores.
The most noticeable negactive impact this has is on the no-op
microbenchmarks, which are also very notable for having no cpu bus load.
In particular, this increases the runtime and energy consumption of
gem_exec_whisper.
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>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190504070707.30902-1-chris@chris-wilson.co.uk
Currently we submit the semaphore busywait as soon as the signaler is
submitted to HW. However, we may submit the signaler as the tail of a
batch of requests, and even not as the first context in the HW list,
i.e. the busywait may start spinning far in advance of the signaler even
starting.
If we wait until the request before the signaler is completed before
submitting the busywait, we prevent the busywait from starting too
early, if the signaler is not first in submission port.
To handle the case where the signaler is at the start of the second (or
later) submission port, we will need to delay the execution callback
until we know the context is promoted to port0. A challenge for later.
Fixes: e886196469 ("drm/i915: Use HW semaphores for inter-engine synchroni
sation on gen8+")
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/20190501114541.10077-9-chris@chris-wilson.co.uk
Having transitioned GEM over to using intel_context as its primary means
of tracking the GEM context and engine combined and using
i915_request_create(), we can move the older i915_request_alloc()
helper function into selftests/ where the remaining users are confined.
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/20190426163336.15906-9-chris@chris-wilson.co.uk
We switched to a tree of per-engine HW context to accommodate the
introduction of virtual engines. However, we plan to also support
multiple instances of the same engine within the GEM context, defeating
our use of the engine as a key to looking up the HW context. Just
allocate a logical per-engine instance and always use an index into the
ctx->engines[]. Later on, this ctx->engines[] may be replaced by a user
specified map.
v2: Add for_each_gem_engine() helper to iterator within the engines lock
v3: intel_context_create_request() helper
v4: s/unsigned long/unsigned int/ 4 billion engines is quite enough.
v5: Push iterator locking to caller
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/20190426163336.15906-7-chris@chris-wilson.co.uk
In order to separate the reservation phase of building a request from
its emission phase, we need to pull some of the request alloc activities
from deep inside i915_request to the surface, GEM_EXECBUFFER.
v2: Be frivolous, use a local drm_i915_private.
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/20190425050143.811-1-chris@chris-wilson.co.uk
In the current scheme, on submitting a request we take a single global
GEM wakeref, which trickles down to wake up all GT power domains. This
is undesirable as we would like to be able to localise our power
management to the available power domains and to remove the global GEM
operations from the heart of the driver. (The intent there is to push
global GEM decisions to the boundary as used by the GEM user interface.)
Now during request construction, each request is responsible via its
logical context to acquire a wakeref on each power domain it intends to
utilize. Currently, each request takes a wakeref on the engine(s) and
the engines themselves take a chipset wakeref. This gives us a
transition on each engine which we can extend if we want to insert more
powermangement control (such as soft rc6). The global GEM operations
that currently require a struct_mutex are reduced to listening to pm
events from the chipset GT wakeref. As we reduce the struct_mutex
requirement, these listeners should evaporate.
Perhaps the biggest immediate change is that this removes the
struct_mutex requirement around GT power management, allowing us greater
flexibility in request construction. Another important knock-on effect,
is that by tracking engine usage, we can insert a switch back to the
kernel context on that engine immediately, avoiding any extra delay or
inserting global synchronisation barriers. This makes tracking when an
engine and its associated contexts are idle much easier -- important for
when we forgo our assumed execution ordering and need idle barriers to
unpin used contexts. In the process, it means we remove a large chunk of
code whose only purpose was to switch back to the kernel context.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190424200717.1686-5-chris@chris-wilson.co.uk
Start acquiring the logical intel_context and using that as our primary
means for request allocation. This is the initial step to allow us to
avoid requiring struct_mutex for request allocation along the
perma-pinned kernel context, but it also provides a foundation for
breaking up the complex request allocation to handle different scenarios
inside execbuf.
For the purpose of emitting a request from inside retirement (see the
next patch for engine power management), we also need to lift control
over the timeline mutex to the caller.
v2: Note that the request carries the active reference upon construction.
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/20190424200717.1686-4-chris@chris-wilson.co.uk
We wish to start segregating the power management into different control
domains, both with respect to the hardware and the user interface. The
first step is that at the lowest level flow of requests, we want to
process a context event (and not a global GEM operation). In this patch,
we introduce the context callbacks that in future patches will be
redirected to per-engine interfaces leading to global operations as
required.
The intent is that this will be guarded by the timeline->mutex, except
that retiring has not quite finished transitioning over from being
guarded by struct_mutex. So at the moment it is protected by
struct_mutex with a reminded to switch.
v2: Rename default handlers to intel_context_enter_engine.
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/20190424200717.1686-3-chris@chris-wilson.co.uk
Start partitioning off the code that talks to the hardware (GT) from the
uapi layers and move the device facing code under gt/
One casualty is s/intel_ringbuffer.h/intel_engine.h/ with the plan to
subdivide that header and body further (and split out the submission
code from the ringbuffer and logical context handling). This patch aims
to be simple motion so git can fixup inflight patches with little mess.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Acked-by: Jani Nikula <jani.nikula@intel.com>
Acked-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190424174839.7141-1-chris@chris-wilson.co.uk
An interesting discussion regarding "hybrid interrupt polling" for NVMe
came to the conclusion that the ideal busyspin before sleeping was half
of the expected request latency (and better if it was already halfway
through that request). This suggested that we too should look again at
our tradeoff between spinning and waiting. Currently, our spin simply
tries to hide the cost of enabling the interrupt, which is good to avoid
penalising nop requests (i.e. test throughput) and not much else.
Studying real world workloads suggests that a spin of upto 500us can
dramatically boost performance, but the suggestion is that this is not
from avoiding interrupt latency per-se, but from secondary effects of
sleeping such as allowing the CPU reduce cstate and context switch away.
In a truly hybrid interrupt polling scheme, we would aim to sleep until
just before the request completed and then wake up in advance of the
interrupt and do a quick poll to handle completion. This is tricky for
ourselves at the moment as we are not recording request times, and since
we allow preemption, our requests are not on as a nicely ordered
timeline as IO. However, the idea is interesting, for it will certainly
help us decide when busyspinning is worthwhile.
v2: Expose the spin setting via Kconfig options for easier adjustment
and testing.
v3: Don't get caught sneaking in a change to the busyspin parameters.
v4: Explain more about the "hybrid interrupt polling" scheme that we
want to migrate towards.
Suggested-by: Sagar Kamble <sagar.a.kamble@intel.com>
References: http://events.linuxfoundation.org/sites/events/files/slides/lemoal-nvme-polling-vault-2017-final_0.pdf
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Sagar Kamble <sagar.a.kamble@intel.com>
Cc: Eero Tamminen <eero.t.tamminen@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Ben Widawsky <ben@bwidawsk.net>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Sagar Kamble <sagar.a.kamble@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190419182625.11186-1-chris@chris-wilson.co.uk
As i915_requests are put into an RCU-freelist, they may get reused
before debugobjects notice them as being freed. On cleanup, explicitly
call i915_sw_fence_fini() so that the debugobject is properly tracked.
Reported-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Fixes: b7404c7ecb ("drm/i915: Bump ready tasks ahead of busywaits")
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>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190411122445.20060-1-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
The timeline is strictly ordered, so by inserting the timeline->barrier
request into the timeline->last_request it naturally provides the same
barrier. Consolidate the pair of barriers into one as they serve the
same purpose.
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/20190408091728.20207-4-chris@chris-wilson.co.uk
It used to be handy that we only had a couple of headers, but over time
intel_drv.h has become unwieldy. Extract declarations to a separate
header file corresponding to the implementation module, clarifying the
modularity of the driver.
Ensure the new header is self-contained, and do so with minimal further
includes, using forward declarations as needed. Include the new header
only where needed, and sort the modified include directives while at it
and as needed.
No functional changes.
v2: gen6_rps_reset_ei() is in i915_irq.c not intel_pm.c.
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/adc6463b95eef3440fba9826793f7d1c5f3b0b4a.1554461791.git.jani.nikula@intel.com
During request construction, we take the timeline->mutex to ensure
exclusive access to the ringbuffer (for command emission) and the
timeline itself (for command ordering). The timeline->mutex should not
be dropped by callers until we release it in i915_request_add().
lockdep provides a pin/unpin lock facility to detect accidental unlocks
inside critical sections, so put it to use for request construction.
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/20190403082132.327-1-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>
Previously, our view has been always to run the engines independently
within a context. (Multiple engines happened before we had contexts and
timelines, so they always operated independently and that behaviour
persisted into contexts.) However, at the user level the context often
represents a single timeline (e.g. GL contexts) and userspace must
ensure that the individual engines are serialised to present that
ordering to the client (or forgot about this detail entirely and hope no
one notices - a fair ploy if the client can only directly control one
engine themselves ;)
In the next patch, we will want to construct a set of engines that
operate as one, that have a single timeline interwoven between them, to
present a single virtual engine to the user. (They submit to the virtual
engine, then we decide which engine to execute on based.)
To that end, we want to be able to create contexts which have a single
timeline (fence context) shared between all engines, rather than multiple
timelines.
v2: Move the specialised timeline ordering to its own function.
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/20190322092325.5883-4-chris@chris-wilson.co.uk
The timeline->name is only used for convenience in pretty printing the
i915_request.fence->ops->get_timeline_name() and it is just as
convenient to pull it from the gem_context directly. The few instances
of its use inside GEM_TRACE() has proven more of a nuisance than
helpful, so not worth saving imo.
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/20190321140711.11190-4-chris@chris-wilson.co.uk
As the final request on a ring may hold the reference to this ring (via
retiring the last pinned context), we may find ourselves chasing a
dangling pointer on completion of the list.
A quick solution is to hold a reference to the ring itself as we retire
along it so that we only free it after we stop dereferencing it.
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/20190318095204.9913-4-chris@chris-wilson.co.uk
Currently we assume that we know the order in which requests run and so
can determine if we need to reissue a switch-to-kernel-context prior to
idling. That assumption does not hold for the future, so instead of
tracking which barriers have been used, simply determine if we have ever
switched away from the kernel context by using the engine and before
idling ensure that all engines that have been used since the last idle
are synchronously switched back to the kernel context for safety (and
else of shrinking memory while idle).
v2: Use intel_engine_mask_t and ALL_ENGINES
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/20190308093657.8640-3-chris@chris-wilson.co.uk
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
Having introduced per-context seqno, we now have a means to identity
progress across the system without feel of rollback as befell the
global_seqno. That is we can program a MI_SEMAPHORE_WAIT operation in
advance of submission safe in the knowledge that our target seqno and
address is stable.
However, since we are telling the GPU to busy-spin on the target address
until it matches the signaling seqno, we only want to do so when we are
sure that busy-spin will be completed quickly. To achieve this we only
submit the request to HW once the signaler is itself executing (modulo
preemption causing us to wait longer), and we only do so for default and
above priority requests (so that idle priority tasks never themselves
hog the GPU waiting for others).
As might be reasonably expected, HW semaphores excel in inter-engine
synchronisation microbenchmarks (where the 3x reduced latency / increased
throughput more than offset the power cost of spinning on a second ring)
and have significant improvement (can be up to ~10%, most see no change)
for single clients that utilize multiple engines (typically media players
and transcoders), without regressing multiple clients that can saturate
the system or changing the power envelope dramatically.
v3: Drop the older NEQ branch, now we pin the signaler's HWSP anyway.
v4: Tell the world and include it as part of scheduler caps.
Testcase: igt/gem_exec_whisper
Testcase: igt/benchmarks/gem_wsim
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/20190301170901.8340-3-chris@chris-wilson.co.uk
In preparation for enabling HW semaphores, we need to keep in flight
timeline HWSP alive until its use across entire system has completed,
as any other timeline active on the GPU may still refer back to the
already retired timeline. We both have to delay recycling available
cachelines and unpinning old HWSP until the next idle point.
An easy option would be to simply keep all used HWSP until the system as
a whole was idle, i.e. we could release them all at once on parking.
However, on a busy system, we may never see a global idle point,
essentially meaning the resource will be leaked until we are forced to
do a GC pass. We already employ a fine-grained idle detection mechanism
for vma, which we can reuse here so that each cacheline can be freed
immediately after the last request using it is retired.
v3: Keep track of the activity of each cacheline.
v4: cacheline_free() on canceling the seqno tracking
v5: Finally with a testcase to exercise wraparound
v6: Pack cacheline into empty bits of page-aligned vaddr
v7: Use i915_utils to hide the pointer casting around bit manipulation
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/20190301170901.8340-2-chris@chris-wilson.co.uk
A simple mutex used for guarding the flow of requests in and out of the
timeline. In the short-term, it will be used only to guard the addition
of requests into the timeline, taken on alloc and released on commit so
that only one caller can construct a request into the timeline
(important as the seqno and ring pointers must be serialised). This will
be used by observers to ensure that the seqno/hwsp is stable. Later,
when we have reduced retiring to only operate on a single timeline at a
time, we can then use the mutex as the sole guard required for retiring.
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/20190301110547.14758-2-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
As kmem_caches share the same properties (size, allocation/free behaviour)
for all potential devices, we can use global caches. While this
potential has worse fragmentation behaviour (one can argue that
different devices would have different activity lifetimes, but you can
also argue that activity is temporal across the system) it is the
default behaviour of the system at large to amalgamate matching caches.
The benefit for us is much reduced pointer dancing along the frequent
allocation paths.
v2: Defer shrinking until after a global grace period for futureproofing
multiple consumers of the slab caches, similar to the current strategy
for avoiding shrinking too early.
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/20190228102035.5857-1-chris@chris-wilson.co.uk
Having weaned the interrupt handling off using a single global execution
queue, we no longer need to emit a global_seqno. Note that we still have
a few assumptions about execution order along engine timelines, but this
removes the most obvious artefact!
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/20190226094922.31617-3-chris@chris-wilson.co.uk
At a few points in our uABI, we check to see if the driver is wedged and
report -EIO back to the user in that case. However, as we perform the
check and reset asynchronously (where once before they were both
serialised by the struct_mutex), we may instead see the temporary wedging
used to cancel inflight rendering to avoid a deadlock during reset
(caused by either us timing out in our reset handler,
i915_wedge_on_timeout or with malice aforethought in intel_reset_prepare
for a stuck modeset). If we suspect this is the case, that is we see a
wedged driver *and* reset in progress, then wait until the reset is
resolved before reporting upon the wedged status.
v2: might_sleep() (Mika)
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=109580
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>
Link: https://patchwork.freedesktop.org/patch/msgid/20190220145637.23503-1-chris@chris-wilson.co.uk
Currently, we accumulate each time a context hangs the GPU, offset
against the number of requests it submits, and if that score exceeds a
certain threshold, we ban that context from submitting any more requests
(cancelling any work in flight). In contrast, we use a simple timer on
the file, that if we see more than a 9 hangs faster than 60s apart in
total across all of its contexts, we will ban the client from creating
any more contexts. This leads to a confusing situation where the file
may be banned before the context, so lets use a simple timer scheme for
each.
If the context submits 3 hanging requests within a 120s period, declare
it forbidden to ever send more requests.
This has the advantage of not being easy to repair by simply sending
empty requests, but has the disadvantage that if the context is idle
then it is forgiven. However, if the context is idle, it is not
disrupting the system, but a hog can evade the request counting and
cause much more severe disruption to the system.
Updating ban_score from request retirement is dubious as the retirement
is purposely not in sync with request submission (i.e. we try and batch
retirement to reduce overhead and avoid latency on submission), which
leads to surprising situations where we can forgive a hang immediately
due to a backlog of requests from before the hang being retired
afterwards.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190219122215.8941-2-chris@chris-wilson.co.uk
As we currently do not check on submission whether the context is banned
in a timely manner it is possible for some requests to escape
cancellation after their parent context is banned. By moving the ban
into the request submission under the engine->timeline.lock, we
serialise it with the reset and setting of the context ban.
References: eb8d0f5af4 ("drm/i915: Remove GPU reset dependence on struct_mutex")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190213182737.12695-1-chris@chris-wilson.co.uk
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
As time goes by, usage of generic ioctls such as drm_syncobj and
sync_file are on the increase bypassing i915-specific ioctls like
GEM_WAIT. Currently, we only apply waitboosting to our driver ioctls as
we track the file/client and account the waitboosting to them. However,
since commit 7b92c1bd05 ("drm/i915: Avoid keeping waitboost active for
signaling threads"), we no longer have been applying the client
ratelimiting on waitboosts and so that information has only been used
for debug tracking.
Push the application of waitboosting down to the common
i915_request_wait, and apply it to all foreign fence waits as well.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Eero Tamminen <eero.t.tamminen@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190213092504.25709-1-chris@chris-wilson.co.uk
Looking forward, we need to break the struct_mutex dependency on
i915_gem_active. In the meantime, external use of i915_gem_active is
quite beguiling, little do new users suspect that it implies a barrier
as each request it tracks must be ordered wrt the previous one. As one
of many, it can be used to track activity across multiple timelines, a
shared fence, which fits our unordered request submission much better. We
need to steer external users away from the singular, exclusive fence
imposed by i915_gem_active to i915_active instead. As part of that
process, we move i915_gem_active out of i915_request.c into
i915_active.c to start separating the two concepts, and rename it to
i915_active_request (both to tie it to the concept of tracking just one
request, and to give it a longer, less appealing name).
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/20190205130005.2807-5-chris@chris-wilson.co.uk
Timeline barrier allows serialization between different timelines.
After calling i915_timeline_set_barrier with a request, all following
submissions on this timeline will be set up as depending on this request,
or barrier. Once the barrier has been completed it automatically gets
cleared and things continue as normal.
This facility will be used by the upcoming context SSEU code.
v2:
* Assert barrier has been retired on timeline_fini. (Chris Wilson)
* Fix mock_timeline.
v3:
* Improved comment language. (Chris Wilson)
v4:
* Maintain ordering with previous barriers set on the timeline.
v5:
* Rebase.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Suggested-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20190205095032.22673-3-tvrtko.ursulin@linux.intel.com
Limit the NEWCLIENT boost to only give its small priority boost to fresh
clients only that have no dependencies.
The idea for using NEWCLIENT boosting, commit b16c765122 ("drm/i915:
Priority boost for new clients"), is that short-lived streams are often
interactive and require lower latency -- and that by executing those
ahead of the long running hogs, the short-lived clients do little to
interfere with the system throughput by virtue of their short-lived
nature. However, we were only considering the client's own timeline for
determining whether or not it was a fresh stream. This allowed for
compositors to wake up before their vblank and bump all of its client
streams. However, in testing with media-bench this results in chaining
all cooperating contexts together preventing us from being able to
reorder contexts to reduce bubbles (pipeline stalls), overall increasing
latency, and reducing system throughput. The exact opposite of our
intent. The compromise of applying the NEWCLIENT boost to strictly fresh
clients (that do not wait upon anything else) should maintain the
"real-time response under load" characteristics of FQ_CODEL, without
locking together the long chains of dependencies across the system.
References: b16c765122 ("drm/i915: Priority boost for new clients")
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/20190204150101.30759-1-chris@chris-wilson.co.uk
A few years ago, see commit 688e6c7258 ("drm/i915: Slaughter the
thundering i915_wait_request herd"), the issue of handling multiple
clients waiting in parallel was brought to our attention. The
requirement was that every client should be woken immediately upon its
request being signaled, without incurring any cpu overhead.
To handle certain fragility of our hw meant that we could not do a
simple check inside the irq handler (some generations required almost
unbounded delays before we could be sure of seqno coherency) and so
request completion checking required delegation.
Before commit 688e6c7258, the solution was simple. Every client
waiting on a request would be woken on every interrupt and each would do
a heavyweight check to see if their request was complete. Commit
688e6c7258 introduced an rbtree so that only the earliest waiter on
the global timeline would woken, and would wake the next and so on.
(Along with various complications to handle requests being reordered
along the global timeline, and also a requirement for kthread to provide
a delegate for fence signaling that had no process context.)
The global rbtree depends on knowing the execution timeline (and global
seqno). Without knowing that order, we must instead check all contexts
queued to the HW to see which may have advanced. We trim that list by
only checking queued contexts that are being waited on, but still we
keep a list of all active contexts and their active signalers that we
inspect from inside the irq handler. By moving the waiters onto the fence
signal list, we can combine the client wakeup with the dma_fence
signaling (a dramatic reduction in complexity, but does require the HW
being coherent, the seqno must be visible from the cpu before the
interrupt is raised - we keep a timer backup just in case).
Having previously fixed all the issues with irq-seqno serialisation (by
inserting delays onto the GPU after each request instead of random delays
on the CPU after each interrupt), we can rely on the seqno state to
perfom direct wakeups from the interrupt handler. This allows us to
preserve our single context switch behaviour of the current routine,
with the only downside that we lose the RT priority sorting of wakeups.
In general, direct wakeup latency of multiple clients is about the same
(about 10% better in most cases) with a reduction in total CPU time spent
in the waiter (about 20-50% depending on gen). Average herd behaviour is
improved, but at the cost of not delegating wakeups on task_prio.
v2: Capture fence signaling state for error state and add comments to
warm even the most cold of hearts.
v3: Check if the request is still active before busywaiting
v4: Reduce the amount of pointer misdirection with list_for_each_safe
and using a local i915_request variable inside the loops
v5: Add a missing pluralisation to a purely informative selftest message.
References: 688e6c7258 ("drm/i915: Slaughter the thundering i915_wait_request herd")
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/20190129205230.19056-2-chris@chris-wilson.co.uk
To allow requests to forgo a common execution timeline, one question we
need to be able to answer is "is this request running?". To track
whether a request has started on HW, we can emit a breadcrumb at the
beginning of the request and check its timeline's HWSP to see if the
breadcrumb has advanced past the start of this request. (This is in
contrast to the global timeline where we need only ask if we are on the
global timeline and if the timeline has advanced past the end of the
previous request.)
There is still confusion from a preempted request, which has already
started but relinquished the HW to a high priority request. For the
common case, this discrepancy should be negligible. However, for
identification of hung requests, knowing which one was running at the
time of the hang will be much more important.
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/20190129185452.20989-2-chris@chris-wilson.co.uk
Now that we have allocated ourselves a cacheline to store a breadcrumb,
we can emit a write from the GPU into the timeline's HWSP of the
per-context seqno as we complete each request. This drops the mirroring
of the per-engine HWSP and allows each context to operate independently.
We do not need to unwind the per-context timeline, and so requests are
always consistent with the timeline breadcrumb, greatly simplifying the
completion checks as we no longer need to be concerned about the
global_seqno changing mid check.
One complication though is that we have to be wary that the request may
outlive the HWSP and so avoid touching the potentially danging pointer
after we have retired the fence. We also have to guard our access of the
HWSP with RCU, the release of the obj->mm.pages should already be RCU-safe.
At this point, we are emitting both per-context and global seqno and
still using the single per-engine execution timeline for resolving
interrupts.
v2: s/fake_complete/mark_complete/
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/20190128181812.22804-5-chris@chris-wilson.co.uk
