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
Now that the submission backends are controlled via their own spinlocks,
with a wave of a magic wand we can lift the struct_mutex requirement
around GPU reset. That is we allow the submission frontend (userspace)
to keep on submitting while we process the GPU reset as we can suspend
the backend independently.
The major change is around the backoff/handoff strategy for performing
the reset. With no mutex deadlock, we no longer have to coordinate with
any waiter, and just perform the reset immediately.
Testcase: igt/gem_mmap_gtt/hang # regresses
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/20190125132230.22221-3-chris@chris-wilson.co.uk
The majority of runtime-pm operations are bounded and scoped within a
function; these are easy to verify that the wakeref are handled
correctly. We can employ the compiler to help us, and reduce the number
of wakerefs tracked when debugging, by passing around cookies provided
by the various rpm_get functions to their rpm_put counterpart. This
makes the pairing explicit, and given the required wakeref cookie the
compiler can verify that we pass an initialised value to the rpm_put
(quite handy for double checking error paths).
For regular builds, the compiler should be able to eliminate the unused
local variables and the program growth should be minimal. Fwiw, it came
out as a net improvement as gcc was able to refactor rpm_get and
rpm_get_if_in_use together,
v2: Just s/rpm_put/rpm_put_unchecked/ everywhere, leaving the manual
mark up for smaller more targeted patches.
v3: Mention the cookie in Returns
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Jani Nikula <jani.nikula@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190114142129.24398-2-chris@chris-wilson.co.uk
drm-misc-next for 5.1:
UAPI Changes:
Cross-subsystem Changes:
- Turn dma-buf fence sequence numbers into 64 bit numbers
Core Changes:
- Move to a common helper for the DP MST hotplug for radeon, i915 and
amdgpu
- i2c improvements for drm_dp_mst
- Removal of drm_syncobj_cb
- Introduction of an helper to create and attach the TV margin properties
Driver Changes:
- Improve cache flushes for v3d
- Reflection support for vc4
- HDMI overscan support for vc4
- Add implicit fencing support for rockchip and sun4i
- Switch to generic fbdev emulation for virtio
Signed-off-by: Dave Airlie <airlied@redhat.com>
[airlied: applied amdgpu merge fixup]
From: Maxime Ripard <maxime.ripard@bootlin.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190107180333.amklwycudbsub3s5@flea
Two simple selftests which test that both GT and engine workarounds are
not lost after either a full GPU reset, or after the per-engine ones.
(Including checks that one engine reset is not affecting workarounds not
belonging to itself.)
v2:
* Rebase for series refactoring.
* Add spinner for actual engine reset!
* Add idle reset test as well. (Chris Wilson)
* Share existing global_reset_lock. (Chris Wilson)
v3:
* intel_engine_verify_workarounds can be static.
* API rename. (Chris Wilson)
* Move global reset lock out of the loop. (Chris Wilson)
v4:
* Add missing rpm puts. (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/20181203125014.3219-5-tvrtko.ursulin@linux.intel.com
On older HW, gen2/3, fence registers are used for detiling GPU commands
and as such changing those registers requires serialisation with the
requests on the GPU. Anything running on the GPU is subject to a hang,
and so we must be able to recover cleanly in the middle of a stuck wait
on a fence register.
We can simulate using the fence on the GPU simply by marking the fence
as active on the request for this vma, the interface being common to all
gen, thus broadening the test.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180719194746.19111-2-chris@chris-wilson.co.uk
Handling such a late error in request construction is tricky, but to
accommodate future patches which may allocate here, we potentially could
err. To handle the error after already adjusting global state to track
the new request, we must finish and submit the request. But we don't
want to use the request as not everything is being tracked by it, so we
opt to cancel the commands inside the request.
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/20180706103947.15919-3-chris@chris-wilson.co.uk
For symmetry, simplicity and ensuring the request is always truly idle
upon its completion, always emit the closing flush prior to emitting the
request breadcrumb. Previously, we would only emit the flush if we had
started a user batch, but this just leaves all the other paths open to
speculation (do they affect the GPU caches or not?) With mm switching, a
key requirement is that the GPU is flushed and invalidated before hand,
so for absolute safety, we want that closing flush be mandatory.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180612105135.4459-1-chris@chris-wilson.co.uk
When testing reset, we wait for 1s on the main thread for the hang to
start. Meanwhile, we continue submitting requests on all the background
threads, and we may have more threads than cores and so potentially
starve the waiter from being woken within the timeout. As the hang
timeout and the active timeouts are the same, it is hard to distinguish
which caused the timeout. Bump the active thread timeouts to 5s,
compared to the 1s timeout for the hang, so that we preferentially
report the hang timing out, while hopefully ensuring that we do at least
wake up the hang thread first before declaring the background active
timeout.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20180517142442.16979-1-chris@chris-wilson.co.uk
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Even though we weren't injecting guilty requests to be reset, we could
still fall over the issue of resetting the same request too fast -- where
the GPU refuses to start again. (Although it is interesting to note that
reloading the driver is sufficient, suggesting that we could recover if
we delayed the setup after reset?) Continue to paper over the problem by
adding a small delay by waiting for the engine to idle between tests,
and ensure that the engines are idle before starting the idle tests.
v2: Replace single instance of 50 with a magic macro.
References: 028666793a ("drm/i915/selftests: Avoid repeatedly harming the same innocent context")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Michał Winiarski <michal.winiarski@intel.com>
Cc: Michel Thierry <michel.thierry@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Michał Winiarski <michal.winiarski@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180411120346.27618-1-chris@chris-wilson.co.uk
Today we only want to pass along the priority to engine->schedule(), but
in the future we want to have much more control over the various aspects
of the GPU during a context's execution, for example controlling the
frequency allowed. As we need an ever growing number of parameters for
scheduling, move those into a struct for convenience.
v2: Move the anonymous struct into its own function for legibility and
ye olde gcc.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180418184052.7129-3-chris@chris-wilson.co.uk
Currently, we rely on inspecting the hangcheck state from within the
i915_reset() routines to determine which engines were guilty of the
hang. This is problematic for cases where we want to run
i915_handle_error() and call i915_reset() independently of hangcheck.
Instead of relying on the indirect parameter passing, turn it into an
explicit parameter providing the set of stalled engines which then are
treated as guilty until proven innocent.
While we are removing the implicit stalled parameter, also make the
reason into an explicit parameter to i915_reset(). We still need a
back-channel for i915_handle_error() to hand over the task to the locked
waiter, but let's keep that its own channel rather than incriminate
another.
This leaves stalled/seqno as being private to hangcheck, with no more
nefarious snooping by reset, be it whole-device or per-engine. \o/
The only real issue now is that this makes it crystal clear that we
don't actually do any testing of hangcheck per se in
drv_selftest/live_hangcheck, merely of resets!
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Michel Thierry <michel.thierry@intel.com>
Cc: Jeff McGee <jeff.mcgee@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Michel Thierry <michel.thierry@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180406220354.18911-2-chris@chris-wilson.co.uk
We don't handle resetting the kernel context very well, or presumably any
context executing its breadcrumb commands in the ring as opposed to the
batchbuffer and flush. If we trigger a device reset twice in quick
succession while the kernel context is executing, we may end up skipping
the breadcrumb. This is really only a problem for the selftest as
normally there is a large interlude between resets (hangcheck), or we
focus on resetting just one engine and so avoid repeatedly resetting
innocents.
Something to try would be a preempt-to-idle to quiesce the engine before
reset, so that innocent contexts would be spared the reset.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Michał Winiarski <michal.winiarski@intel.com>
CC: Michel Thierry <michel.thierry@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180330131801.18327-1-chris@chris-wilson.co.uk
Not all callers want the GPU error to handled in the same way, so expose
a control parameter. In the first instance, some callers do not want the
heavyweight error capture so add a bit to request the state to be
captured and saved.
v2: Pass msg down to i915_reset/i915_reset_engine so that we include the
reason for the reset in the dev_notice(), superseding the earlier option
to not print that notice.
v3: Stash the reason inside the i915->gpu_error to handover to the direct
reset from the blocking waiter.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Jeff McGee <jeff.mcgee@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Cc: Michel Thierry <michel.thierry@intel.com>
Reviewed-by: Michel Thierry <michel.thierry@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180320100449.1360-2-chris@chris-wilson.co.uk
When injecting rapid resets, we must be careful to at least wait for the
previous reset to have taken effect and the engine restarted. If we
perform a second reset before that has happened, we will notice that the
engine hasn't recovered and declare it lost, wedging the device and
failing. In practice, since we wait for each hanging batch to start
before injecting the reset, this too-fast-reset condition can only be
triggered when moving onto the next engine in the test, so we need only
wait for the existing reset to complete before switching engines.
v2: Wrap up the wait inside a safety net to bail out in case of angry hw.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Michel Thierry <michel.thierry@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180205152431.12163-1-chris@chris-wilson.co.uk
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>
