We may try to preempt the currently executing request, only to find that
after unravelling all the dependencies that the original executing
context is still the earliest in the topological sort and re-submitted
back to HW (if we do detect some change in the ELSP that requires
re-submission). However, due to the way we check for wrap-around during
the unravelling, we mark any context that has been submitted just once
(i.e. with the rq->wa_tail set, but the ring->tail earlier) as
potentially wrapping and requiring a forced restore on resubmission.
This was expected to be not a problem, as it was anticipated that most
unwinding for preemption would result in a context switch and the few
that did not would be lost in the noise. It did not take long for
someone to find one particular workload where the cost of those extra
context restores was measurable.
However, since we know the wa_tail is of fixed size, and we know that a
request must be larger than the wa_tail itself, we can safely maintain
the check for request wrapping and check against a slightly future point
in the ring that includes an expected wa_tail. (That is if the
ring->tail is already set to rq->wa_tail, including another 8 bytes in
the check does not invalidate the incremental wrap detection.)
Fixes: 8ab3a3812a ("drm/i915/gt: Incrementally check for rewinding")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Bruce Chang <yu.bruce.chang@intel.com>
Cc: Ramalingam C <ramalingam.c@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: <stable@vger.kernel.org> # v5.4+
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20201002083425.4605-1-chris@chris-wilson.co.uk
When running gem_exec_nop, it floods the system with many requests (with
the goal of userspace submitting faster than the HW can process a single
empty batch). This causes the driver to continually resubmit new
requests onto the end of an active context, a flood of lite-restore
preemptions. If we time this just right, Tigerlake hangs.
Inserting a small delay between the processing of CS events and
submitting the next context, prevents the hang. Naturally it does not
occur with debugging enabled. The suspicion then is that this is related
to the issues with the CS event buffer, and inserting an mmio read of
the CS pointer status appears to be very successful in preventing the
hang. Other registers, or uncached reads, or plain mb, do not prevent
the hang, suggesting that register is key -- but that the hang can be
prevented by a simple udelay, suggests it is just a timing issue like
that encountered by commit 233c1ae3c8 ("drm/i915/gt: Wait for CSB
entries on Tigerlake"). Also note that the hang is not prevented by
applying CTX_DESC_FORCE_RESTORE, or by inserting a delay on the GPU
between requests.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Bruce Chang <yu.bruce.chang@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: stable@vger.kernel.org
Acked-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20201015195023.32346-1-chris@chris-wilson.co.uk
Since SKL the eLLC has been sitting on the far side of the system
agent, meaning the display engine can utilize it. Let's enable that.
I chose WB for the caching mode, because my numbers are indicating
that WT might actually be WB and WC might actually be UC. I'm not
100% sure that is indeed the case but at least my simple rendercopy
based benchmark didn't see any difference in performance.
Also if I configure things to do LLCeLLC+WT I still get cache dirt
on my screen, suggesting that is in fact operating in WB mode
anyway. This is also the reason I had to fix the MOCS target cache
to really say PTE rather than LLC+eLLC.
Since SKL the eLLC has been sitting on the far side of the system agent,
meaning the display engine can utilize it. Let's enable that.
Eero's earlier benchmarks numbers:
"* Results in GfxBench and Unigine (Valley/Heaven) tests were within daily
variation on the tested SKL machines
* SKL GT4e (128MB eLLC) / Wayland / Weston:
+15-20% SynMark TexMem512 (512MB of textures)
+4-6% SynMark TerrainFly*, CSCloth, ShMapVsm
-5-10% SynMark TexMem128 (128MB of textures)
* SKL GT3e (64MB eLLC) / Xorg / Unity:
+4-8% GpuTest Triangle fullscreen (FullHD)
-5-10% GpuTest Triangle windowed (1/2 screen)
* SKL GT2 (no eLLC) / Xorg / Unity:
* Some of the higher FPS SynMark pixel and vertex shader tests
are few percent higher, more than daily variance
=> Do you see any reason why this machine would be impacted
although it doesn't eLLC?"
Caveats:
- Still haven't tested with a prime setup
- Still not entirely sure this a good idea, but I've been
using it on my cfl anyway :)
v2: Split the MOCS PTE change out
Cc: Eero Tamminen <eero.t.tamminen@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20201007120329.17076-3-ville.syrjala@linux.intel.com
Link: https://patchwork.freedesktop.org/patch/msgid/20201015122138.30161-3-chris@chris-wilson.co.uk
Since we track the idle_pulse for flushing the barriers and avoid
re-emitting the pulse upon idling if no futher action is required, this
also impacts the heartbeat. Before emitting a fresh heartbeat, we look
at the engine idle status and assume that if the pulse was the last
request emitted along the heartbeat, the engine is idling and a
heartbeat pulse not required. This assumption fails, but we can reuse
the idle pulse as the heartbeat if we are yet to emit one, and so track
the status of that pulse for our engine health check.
This impacts tgl/rcs0 as we rely on the heartbeat for our healthcheck for
the normal preemption detection mechanism is disabled by default.
Testcase: igt/gem_exec_schedule/preempt-hang/rcs0 #tgl
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20201006094653.7558-1-chris@chris-wilson.co.uk
As the previous patch fixed the places where we walk the whole scatterlist
for DMA addresses, this patch fixes the random lookup functionality.
To achieve this we have to add a second lookup iterator and add a
i915_gem_object_get_sg_dma helper, to be used analoguous to existing
i915_gem_object_get_sg_dma. Therefore two lookup caches are maintained per
object and they are flushed at the same point for simplicity. (Strictly
speaking the DMA cache should be flushed from i915_gem_gtt_finish_pages,
but today this conincides with unsetting of the pages in general.)
Partial VMA view is then fixed to use the new DMA lookup and properly
query sg length.
v2:
* Checkpatch.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Matthew Auld <matthew.auld@intel.com>
Cc: Lu Baolu <baolu.lu@linux.intel.com>
Cc: Tom Murphy <murphyt7@tcd.ie>
Cc: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20201006092508.1064287-2-tvrtko.ursulin@linux.intel.com
Currently we do a final scrub of the HW state upon release. However,
when rebinding the device, this is too late as the device may either
have been partially rebound or the device is no longer accessible. If
the device has been removed before release, the reset goes astray
leaving the device in an inconsistent state, unlikely to work without a
full PCI reset. Furthermore, if the device is partially rebound before
the HW scrubbing, there may be leftover HW state that should have been
scrubbed. Either way, we need to push the scrubbing earlier before the
removal, so into unregister. The danger is that on older machines,
resetting the GPU also impact the display engine and so the reset should
be after modesetting is disabled (and before reuse we need to recover
modesetting).
Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/2508
Testcase: igt/core_hotunplug
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/20200929112639.24223-1-chris@chris-wilson.co.uk
If we manage to hit the intel_gt_set_wedged_on_fini() while active, i.e.
module unload during a stress test, we may cancel the requests but not
clean up. This leads to a very slow module unload as we wait for
something or other to trigger the retirement flushing, or timeout and
unload with a bunch of warnings. Instead if we explicitly cancel then
cleanup on an active unload, it should be instant and quiet.
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/20200930163253.2789-3-chris@chris-wilson.co.uk
We only allow persistent requests to remain on the GPU past the closure
of their containing context (and process) so long as they are continuously
checked for hangs or allow other requests to preempt them, as we need to
ensure forward progress of the system. If we allow persistent contexts
to remain on the system after the the hangcheck mechanism is disabled,
the system may grind to a halt. On disabling the mechanism, we sent a
pulse along the engine to remove all executing contexts from the engine
which would check for hung contexts -- but we did not prevent those
contexts from being resubmitted if they survived the final hangcheck.
Fixes: 9a40bddd47 ("drm/i915/gt: Expose heartbeat interval via sysfs")
Testcase: igt/gem_ctx_persistence/heartbeat-stop
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: <stable@vger.kernel.org> # v5.7+
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Acked-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200928221510.26044-1-chris@chris-wilson.co.uk
Shrink the hold time for the error capture mutex to just around the
acquire/release of the PTE used for reading back the object via the
Global GTT. For platforms that do not need the GGTT read back, we can
skip the mutex entirely and allow concurrent error capture. Where we do
use the GGTT, by restricting the hold time around the slow readback and
compression, we are more resilient against softlockups (khungtaskd) as
the heartbeat may well also trigger an error while the first is on
going, and this allows the heartbeat reset to skip past the capture and
not be stalled.
Testcase: igt/gem_exec_capture/many-*
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200916090059.3189-3-chris@chris-wilson.co.uk
On Tigerlake, we are seeing a repeat of commit d8f5053117 ("drm/i915/icl:
Forcibly evict stale csb entries") where, presumably, due to a missing
Global Observation Point synchronisation, the write pointer of the CSB
ringbuffer is updated _prior_ to the contents of the ringbuffer. That is
we see the GPU report more context-switch entries for us to parse, but
those entries have not been written, leading us to process stale events,
and eventually report a hung GPU.
However, this effect appears to be much more severe than we previously
saw on Icelake (though it might be best if we try the same approach
there as well and measure), and Bruce suggested the good idea of resetting
the CSB entry after use so that we can detect when it has been updated by
the GPU. By instrumenting how long that may be, we can set a reliable
upper bound for how long we should wait for:
513 late, avg of 61 retries (590 ns), max of 1061 retries (10099 ns)
Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/2045
References: d8f5053117 ("drm/i915/icl: Forcibly evict stale csb entries")
References: HSDES#22011327657, HSDES#1508287568
Suggested-by: Bruce Chang <yu.bruce.chang@intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Bruce Chang <yu.bruce.chang@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: stable@vger.kernel.org # v5.4
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200915134923.30088-2-chris@chris-wilson.co.uk
If we create a new node, it is possible for the slab allocator to return
us a recently freed node. If that node was just retired, it will retain
the current jiffy as its node->age. There is then a miniscule window,
where as that node is retired, it will appear on the free list with an
incorrect age and be eligible for reuse by one thread, and then by a
second thread as the correct node->age is written.
Fixes: 06b73c2d0b ("drm/i915/gt: Delay taking the spinlock for grabbing from the buffer pool")
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/20200915091417.4086-3-chris@chris-wilson.co.uk
On 32b, highmem using a finite set of indirect PTE (i.e. vmap) to provide
virtual mappings of the high pages. As these are finite, map_new_virtual()
must wait for some other kmap() to finish when it runs out. If we map a
large number of objects, there is no method for it to tell us to release
the mappings, and we deadlock.
However, if we make an explicit vmap of the page, that uses a larger
vmalloc arena, and also has the ability to tell us to release unwanted
mappings. Most importantly, it will fail and propagate an error instead
of waiting forever.
Fixes: fb8621d3be ("drm/i915: Avoid allocating a vmap arena for a single page") #x86-32
References: e87666b52f ("drm/i915/shrinker: Hook up vmap allocation failure notifier")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Cc: <stable@vger.kernel.org> # v4.7+
Link: https://patchwork.freedesktop.org/patch/msgid/20200915091417.4086-1-chris@chris-wilson.co.uk
The hwsp_gtt object is used for sub-allocation and could therefore
be shared by many contexts causing unnecessary contention during
concurrent context pinning.
However since we're currently locking it only for pinning, it remains
resident until we unpin it, and therefore it's safe to drop the
lock early, allowing for concurrent thread access.
Signed-off-by: Thomas Hellström <thomas.hellstrom@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
To implement preempt-to-busy (and so efficient timeslicing and best utilization
of the hardware submission ports) we let the GPU run asynchronously in respect
to the ELSP submission queue. This created challenges in keeping and accessing
the driver state mirroring the asynchronous GPU execution.
Previous fix 1d9221e9d3 ("drm/i915: Skip signaling a signaled request")
however did not correctly serialize request retirement with the execution
callbacks.
We were using the i915_request.lock to serialise adding an execution callback
with __i915_request_submit. However, if we use an atomic llist_add to serialise
multiple waiters and then check to see if the request is already executing, we
can remove the irq-spinlock and fix serialization between retirement and
execution callbacks in one go.
v2: Avoid using the irq_work when outside of the irq-spinlocks, where we
can execute the callbacks immediately.
v3: Pay close attention to the order of setting ACTIVE on retirement, we
need to ensure the request is signaled and breadcrumbs detached before
we finish removing the request from the engine.
v4: Expanded commit message.
Fixes: 1d9221e9d3 ("drm/i915: Skip signaling a signaled 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/20200716142207.13003-2-chris@chris-wilson.co.uk
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
[Joonas: Rebased and reordered into drm-intel-gt-next branch]
[Joonas: Added expanded commit message from Tvrtko and Chris]
Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
To implement preempt-to-busy (and so efficient timeslicing and best utilization
of the hardware submission ports) we let the GPU run asynchronously in respect
to the ELSP submission queue. This created challenges in keeping and accessing
the driver state mirroring the asynchronous GPU execution.
The latest occurence of this was spotted by KCSAN:
[ 1413.563200] BUG: KCSAN: data-race in __await_execution+0x217/0x370 [i915]
[ 1413.563221]
[ 1413.563236] race at unknown origin, with read to 0xffff88885bb6c478 of 8 bytes by task 9654 on cpu 1:
[ 1413.563548] __await_execution+0x217/0x370 [i915]
[ 1413.563891] i915_request_await_dma_fence+0x4eb/0x6a0 [i915]
[ 1413.564235] i915_request_await_object+0x421/0x490 [i915]
[ 1413.564577] i915_gem_do_execbuffer+0x29b7/0x3c40 [i915]
[ 1413.564967] i915_gem_execbuffer2_ioctl+0x22f/0x5c0 [i915]
[ 1413.564998] drm_ioctl_kernel+0x156/0x1b0
[ 1413.565022] drm_ioctl+0x2ff/0x480
[ 1413.565046] __x64_sys_ioctl+0x87/0xd0
[ 1413.565069] do_syscall_64+0x4d/0x80
[ 1413.565094] entry_SYSCALL_64_after_hwframe+0x44/0xa9
To complicate matters, we have to both avoid the read tearing of *active and
avoid any write tearing as perform the pending[] -> inflight[] promotion of the
execlists.
This is because we cannot rely on the memcpy doing u64 aligned copies on all
kernels/platforms and so we opt to open-code it with explicit WRITE_ONCE
annotations to satisfy KCSAN.
v2: When in doubt, write the same comment again.
v3: Expanded commit message.
Fixes: b55230e5e8 ("drm/i915: Check for awaits on still currently executing requests")
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/20200716142207.13003-1-chris@chris-wilson.co.uk
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
[Joonas: Rebased and reordered into drm-intel-gt-next branch]
[Joonas: Added expanded commit message from Tvrtko and Chris]
Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
