The legacy context switch for ringbuffer submission is multistaged,
where each of those stages may fail. However, we were updating global
state after some stages, and so we had to force the incomplete request
to be submitted because we could not unwind. Save the global state
before performing the switches, and so enable us to unwind back to the
previous global state should any phase fail. We then must cancel the
request instead of submitting it should the construction fail.
v2: s/saved_ctx/from_ctx/; s/ctx/to_ctx/ etc.
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/20171123152631.31385-1-chris@chris-wilson.co.uk
We can use engine busy stats instead of the sampling timer for
better accuracy.
By doing this we replace the stohastic sampling with busyness
metric derived directly from engine activity. This is context
switch interrupt driven, so as accurate as we can get from
software tracking.
As a secondary benefit, we can also not run the sampling timer
in cases only busyness metric is enabled.
v2: Rebase.
v3:
* Rebase, comments.
* Leave engine busyness controls out of workers.
v4: Checkpatch cleanup.
v5: Added comment to pmu_needs_timer change.
v6:
* Rebase.
* Fix style of some comments. (Chris Wilson)
v7: Rebase and commit message update. (Chris Wilson)
v8: Add delayed stats disabling to improve accuracy in face of
CPU hotplug events.
v9: Rebase.
v10: Rebase - i915_modparams.enable_execlists removal.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20171121181852.16128-6-tvrtko.ursulin@linux.intel.com
Track total time requests have been executing on the hardware.
We add new kernel API to allow software tracking of time GPU
engines are spending executing requests.
Both per-engine and global API is added with the latter also
being exported for use by external users.
v2:
* Squashed with the internal API.
* Dropped static key.
* Made per-engine.
* Store time in monotonic ktime.
v3: Moved stats clearing to disable.
v4:
* Comments.
* Don't export the API just yet.
v5: Whitespace cleanup.
v6:
* Rename ref to active.
* Drop engine aggregate stats for now.
* Account initial busy period after enabling stats.
v7:
* Rebase.
v8:
* Move context in notification after the notifier. (Chris Wilson)
v9:
In cases where stats tracking is getting disabled while there is
an active context on an engine, add up the current value to the
total. This also implies we don't clear the total when tracking
is disabled any longer. There is no real need to do so because
we define the stats as relative while enabled, meaning
comparison between two samples while tracking is enabled is the
valid usage. However, when busy stats will later be plugged into
the perf PMU API, it is beneficial to not reset the total, since
the PMU core likes to do some counter disable/enable cycles on
startup, and while doing so during a single long context
executing on an engine we would lose some accuracy and so make
unit testing more difficult than needs to be.
v10:
* Fix accounting for preemption.
v11:
* Rebase for i915_modparams.enable_execlists removal.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20171121181852.16128-5-tvrtko.ursulin@linux.intel.com
If only a subset of events is enabled we can afford to suspend
the sampling timer when the GPU is idle and so save some cycles
and power.
v2: Rebase and limit timer even more.
v3: Rebase.
v4: Rebase.
v5: Skip action if perf PMU failed to register.
v6: Checkpatch cleanup.
v7:
* Add a common helper to start the timer if needed. (Chris Wilson)
* Add comment explaining bitwise logic in pmu_needs_timer.
v8: Fix some comments styles. (Chris Wilson)
v9: Rebase.
v10: Move function declarations to i915_pmu.h.
v11: Rename functions to i915_pmu_gt_(un)parked. (Chris Wilson)
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20171121181852.16128-3-tvrtko.ursulin@linux.intel.com
From: Chris Wilson <chris@chris-wilson.co.uk>
From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
From: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>
The first goal is to be able to measure GPU (and invidual ring) busyness
without having to poll registers from userspace. (Which not only incurs
holding the forcewake lock indefinitely, perturbing the system, but also
runs the risk of hanging the machine.) As an alternative we can use the
perf event counter interface to sample the ring registers periodically
and send those results to userspace.
Functionality we are exporting to userspace is via the existing perf PMU
API and can be exercised via the existing tools. For example:
perf stat -a -e i915/rcs0-busy/ -I 1000
Will print the render engine busynnes once per second. All the performance
counters can be enumerated (perf list) and have their unit of measure
correctly reported in sysfs.
v1-v2 (Chris Wilson):
v2: Use a common timer for the ring sampling.
v3: (Tvrtko Ursulin)
* Decouple uAPI from i915 engine ids.
* Complete uAPI defines.
* Refactor some code to helpers for clarity.
* Skip sampling disabled engines.
* Expose counters in sysfs.
* Pass in fake regs to avoid null ptr deref in perf core.
* Convert to class/instance uAPI.
* Use shared driver code for rc6 residency, power and frequency.
v4: (Dmitry Rogozhkin)
* Register PMU with .task_ctx_nr=perf_invalid_context
* Expose cpumask for the PMU with the single CPU in the mask
* Properly support pmu->stop(): it should call pmu->read()
* Properly support pmu->del(): it should call stop(event, PERF_EF_UPDATE)
* Introduce refcounting of event subscriptions.
* Make pmu.busy_stats a refcounter to avoid busy stats going away
with some deleted event.
* Expose cpumask for i915 PMU to avoid multiple events creation of
the same type followed by counter aggregation by perf-stat.
* Track CPUs getting online/offline to migrate perf context. If (likely)
cpumask will initially set CPU0, CONFIG_BOOTPARAM_HOTPLUG_CPU0 will be
needed to see effect of CPU status tracking.
* End result is that only global events are supported and perf stat
works correctly.
* Deny perf driver level sampling - it is prohibited for uncore PMU.
v5: (Tvrtko Ursulin)
* Don't hardcode number of engine samplers.
* Rewrite event ref-counting for correctness and simplicity.
* Store initial counter value when starting already enabled events
to correctly report values to all listeners.
* Fix RC6 residency readout.
* Comments, GPL header.
v6:
* Add missing entry to v4 changelog.
* Fix accounting in CPU hotplug case by copying the approach from
arch/x86/events/intel/cstate.c. (Dmitry Rogozhkin)
v7:
* Log failure message only on failure.
* Remove CPU hotplug notification state on unregister.
v8:
* Fix error unwind on failed registration.
* Checkpatch cleanup.
v9:
* Drop the energy metric, it is available via intel_rapl_perf.
(Ville Syrjälä)
* Use HAS_RC6(p). (Chris Wilson)
* Handle unsupported non-engine events. (Dmitry Rogozhkin)
* Rebase for intel_rc6_residency_ns needing caller managed
runtime pm.
* Drop HAS_RC6 checks from the read callback since creating those
events will be rejected at init time already.
* Add counter units to sysfs so perf stat output is nicer.
* Cleanup the attribute tables for brevity and readability.
v10:
* Fixed queued accounting.
v11:
* Move intel_engine_lookup_user to intel_engine_cs.c
* Commit update. (Joonas Lahtinen)
v12:
* More accurate sampling. (Chris Wilson)
* Store and report frequency in MHz for better usability from
perf stat.
* Removed metrics: queued, interrupts, rc6 counters.
* Sample engine busyness based on seqno difference only
for less MMIO (and forcewake) on all platforms. (Chris Wilson)
v13:
* Comment spelling, use mul_u32_u32 to work around potential GCC
issue and somne code alignment changes. (Chris Wilson)
v14:
* Rebase.
v15:
* Rebase for RPS refactoring.
v16:
* Use the dynamic slot in the CPU hotplug state machine so that we are
free to setup our state as multi-instance. Previously we were re-using
the CPUHP_AP_PERF_X86_UNCORE_ONLINE slot which is neither used as
multi-instance, nor owned by our driver to start with.
* Register the CPU hotplug handlers after the PMU, otherwise the callback
will get called before the PMU is initialized which can end up in
perf_pmu_migrate_context with an un-initialized base.
* Added workaround for a probable bug in cpuhp core.
v17:
* Remove workaround for the cpuhp bug.
v18:
* Rebase for drm_i915_gem_engine_class getting upstream before us.
v19:
* Rebase. (trivial)
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171121181852.16128-2-tvrtko.ursulin@linux.intel.com
Rename enum plane to enum i9xx_plane_id to make it clear that it only
applies to pre-SKL platforms.
enum i9xx_plane_id is a global identifier, whereas enum plane_id is
per-pipe. We need the old global identifier to index the primary plane
(and the pre-g4x sprite C if we ever expose it) registers on pre-SKL
platforms.
v2: Reorder patches
v3: s/old_plane_id/i9xx_plane_id/ (Daniel)
Pimp the commit message a bit
Note that i9xx_plane_id doesn't apply to SKL+
v4: Rebase due to power domain handling in plane readout
v5: Rebase due to crtc->dspaddr_offset removal
v6: s/plane/i9xx_plane/ etc. (James)
Cc: James Ausmus <james.ausmus@intel.com>
Cc: Daniel Vetter <daniel@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20171117191917.11506-4-ville.syrjala@linux.intel.com
Reviewed-by: James Ausmus <james.ausmus@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Add a .get_hw_state() method for planes, returning true or false
depending on whether the plane is enabled. Use it to rewrite the
plane enabled/disabled asserts in platform agnostic fashion.
We do lose the pre-gen4 plane<->pipe mapping checks, but since we're
supposed sanitize that anyway it doesn't really matter.
v2: Reoder patches to not depend on enum old_plane_id
Just call assert_plane_disabled() from assert_planes_disabled()
v3: Deal with disabled power wells in .get_hw_state()
v4: Rebase due skl primary plane code removal
Cc: Thierry Reding <thierry.reding@gmail.com>
Cc: Alex Villacís Lasso <alexvillacislasso@hotmail.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> #v2
Tested-by: Thierry Reding <thierry.reding@gmail.com> #v2
Link: https://patchwork.freedesktop.org/patch/msgid/20171117191917.11506-2-ville.syrjala@linux.intel.com
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
I should have admitted defeat long ago as there has been a rare but
persistent error on Sandybridge where semaphore signaling did not
propagate to the waiter, leading to a GPU hang.
With the work on fence signaling for v4.9, the impact of using CPU driven
signaling was greatly reduced wrt to the latency of GPU semaphores,
though without logical rings support, the benefit of reordering work to
avoid bubbles is not realised (i.e. as it stands fence signaling is just
a slower, more costly version of HW semaphores; but works more
consistently). As a rough indicator of the difference,
with semaphores:
Sequential (3 engines, 1 processes): average 5.470us per cycle [expected 4.988us]
w/o semaphores:
Sequential (3 engines, 1 processes): average 15.771us per cycle [expected 4.923us]
In comparison, v3.4:
with semaphores:
Sequential (3 engines, 1 processes): average 16.066us per cycle [expected 11.842us]
w/o semaphores:
Sequential (3 engines, 1 processes): average 23.460us per cycle [expected 11.839us]
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=54226 #and 100+ dupes
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Acked-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171120205504.21892-3-chris@chris-wilson.co.uk
Since removing the module parameter to force selection of ringbuffer
emission for gen8, the code is defunct. Remove it.
To put the difference into perspective, a couple of microbenchmarks
(bdw i7-5557u, 20170324):
ring execlists
exec continuous nops on all rings: 1.491us 2.223us
exec sequential nops on each ring: 12.508us 53.682us
single nop + sync: 9.272us 30.291us
vblank_mode=0 glxgears: ~11000fps ~9000fps
Since the earlier submission, gen8 ringbuffer submission has fallen
further and further behind in features. So while ringbuffer may hold the
throughput crown, in terms of interactive latency, execlists is much
better. Alas, we have no convenient metrics for such, other than
demonstrating things we can do with execlists but can not using
legacy ringbuffer submission.
We have made a few improvements to lowlevel execlists throughput,
and ringbuffer currently panics on boot! (bdw i7-5557u, 20171026):
ring execlists
exec continuous nops on all rings: n/a 1.921us
exec sequential nops on each ring: n/a 44.621us
single nop + sync: n/a 21.953us
vblank_mode=0 glxgears: n/a ~18500fps
References: https://bugs.freedesktop.org/show_bug.cgi?id=87725
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Once-upon-a-time-Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171120205504.21892-2-chris@chris-wilson.co.uk
The hardware needs some time to process the information received in the
ExecList Submission Port, and expects us to not write anything more until
it has 'acknowledged' this new submission by sending an IDLE_ACTIVE or
PREEMPTED CSB event.
If we do not follow this, the driver could write new data into the ELSP
before HW had finishing fetching the previous one, putting us in
'undefined behaviour' space.
This seems to be the problem causing the spurious PREEMPTED & COMPLETE
events after a COMPLETE like the one below:
[] vcs0: sw rd pointer = 2, hw wr pointer = 0, current 'head' = 3.
[] vcs0: Execlist CSB[0]: 0x00000018 _ 0x00000007
[] vcs0: Execlist CSB[1]: 0x00000001 _ 0x00000000
[] vcs0: Execlist CSB[2]: 0x00000018 _ 0x00000007 <<< COMPLETE
[] vcs0: Execlist CSB[3]: 0x00000012 _ 0x00000007 <<< PREEMPTED & COMPLETE
[] vcs0: Execlist CSB[4]: 0x00008002 _ 0x00000006
[] vcs0: Execlist CSB[5]: 0x00000014 _ 0x00000006
The ELSP writes that lead to this CSB sequence show that the HW hadn't
started executing the previous execlist (the one with only ctx 0x6) by the
time the new one was submitted; this is a bit more clear in the data
show in the EXECLIST_STATUS register at the time of the ELSP write.
[] vcs0: ELSP[0] = 0x0_0 [execlist1] - status_reg = 0x0_302
[] vcs0: ELSP[1] = 0x6_fedb2119 [execlist0] - status_reg = 0x0_8302
[] vcs0: ELSP[2] = 0x7_fedaf119 [execlist1] - status_reg = 0x0_8308
[] vcs0: ELSP[3] = 0x6_fedb2119 [execlist0] - status_reg = 0x7_8308
Note that having to wait for this ack does not disable lite-restores,
although it may reduce their numbers.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=102035
Signed-off-by: Michel Thierry <michel.thierry@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/<20171118003038.7935-1-michel.thierry@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171120123458.23242-4-chris@chris-wilson.co.uk
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Tested-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Now that we have this stored in the device info, we can drop it from perf
part of the driver.
Note that this requires to init perf after we've computed the frequency,
hence why we move i915_perf_init() from i915_driver_init_early() to after
intel_device_info_runtime_init().
v2: Use div_u64 (Chris)
v3: Drop u64 divs by switching to kHz (Chris/Ville)
Move i915_perf_fini to i915_driver_cleanup_hw (Matthew)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171113181902.12411-2-lionel.g.landwerlin@intel.com
As the request will, in the following patch, implicitly invoke a
context-switch on construction, we should precede that with a GPU TLB
invalidation. Also, even before using GGTT, we always want to invalidate
the TLBs for any updates (as well as the ppgtt invalidates that are
unconditionally applied by execbuf). Since we almost always require the
TLB invalidate, do it unconditionally on request allocation and so we can
remove it from all other paths.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171120102002.22254-1-chris@chris-wilson.co.uk
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
