Currently, the backend scheduling code abuses struct_mutex into order to
have a global lock to manipulate a temporary list (without widespread
allocation) and to protect against list modifications. This is an
extraneous coupling to struct_mutex and further can not extend beyond
the local device.
Pull all the code that needs to be under the one true lock into
i915_scheduler.c, and make it so.
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/20181001144755.7978-2-chris@chris-wilson.co.uk
Taken from an idea used for FQ_CODEL, we give the first request of a
new request flows a small priority boost. These flows are likely to
correspond with short, interactive tasks and so be more latency sensitive
than the longer free running queues. As soon as the client has more than
one request in the queue, further requests are not boosted and it settles
down into ordinary steady state behaviour. Such small kicks dramatically
help combat the starvation issue, by allowing each client the opportunity
to run even when the system is under heavy throughput load (within the
constraints of the user selected priority).
v2: Mark the preempted request as the start of a new flow, to prevent a
single client being continually gazumped by its peers.
Testcase: igt/benchmarks/rrul
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@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/20181001144755.7978-1-chris@chris-wilson.co.uk
If we try and fail to allocate a i915_request, we apply some
backpressure on the clients to throttle the memory allocations coming
from i915.ko. Currently, we wait until completely idle, but this is far
too heavy and leads to some situations where the only escape is to
declare a client hung and reset the GPU. The intent is to only ratelimit
the allocation requests and to allow ourselves to recycle requests and
memory from any long queues built up by a client hog.
Although the system memory is inherently a global resources, we don't
want to overly penalize an unlucky client to pay the price of reaping a
hog. To reduce the influence of one client on another, we can instead of
waiting for the entire GPU to idle, impose a barrier on the local client.
(One end goal for request allocation is for scalability to many
concurrent allocators; simultaneous execbufs.)
To prevent ourselves from getting caught out by long running requests
(requests that may never finish without userspace intervention, whom we
are blocking) we need to impose a finite timeout, ideally shorter than
hangcheck. A long time ago Paul McKenney suggested that RCU users should
ratelimit themselves using judicious use of cond_synchronize_rcu(). This
gives us the opportunity to reduce our indefinite wait for the GPU to
idle to a wait for the RCU grace period of the previous allocation along
this timeline to expire, satisfying both the local and finite properties
we desire for our ratelimiting.
There are still a few global steps (reclaim not least amongst those!)
when we exhaust the immediate slab pool, at least now the wait is itself
decoupled from struct_mutex for our glorious highly parallel future!
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=106680
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180914080017.30308-1-chris@chris-wilson.co.uk
Usually we have no idea about the upper bound we need to wait to catch
up with userspace when idling the device, but in a few situations we
know the system was idle beforehand and can provide a short timeout in
order to very quickly catch a failure, long before hangcheck kicks in.
In the following patches, we will use the timeout to curtain two overly
long waits, where we know we can expect the GPU to complete within a
reasonable time or declare it broken.
In particular, with a broken GPU we expect it to fail during the initial
GPU setup where do a couple of context switches to record the defaults.
This is a task that takes a few milliseconds even on the slowest of
devices, but we may have to wait 60s for hangcheck to give in and
declare the machine inoperable. In this a case where any gpu hang is
unacceptable, both from a timeliness and practical standpoint.
The other improvement is that in selftests, we do not need to arm an
independent timer to inject a wedge, as we can just limit the timeout on
the wait directly.
v2: Include the timeout parameter in the trace.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180709122044.7028-1-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
The discovery with trying to enable full-ppgtt was that we were
completely failing to the load both the mm and context following the
reset. Although we were performing mmio to set the PP_DIR (per-process
GTT) and CCID (context), these were taking no effect (the assumption was
that this would trigger reload of the context and restore the page
tables). It was not until we performed the LRI + MI_SET_CONTEXT in a
following context switch would anything occur.
Since we are then required to reset the context image and PP_DIR using
CS commands, we place those commands into every batch. The hardware
should recognise the no-ops and eliminate the expensive context loads,
but we still have to pay the cost of using cross-powerwell register
writes. In practice, this has no effect on actual context switch times,
and only adds a few hundred nanoseconds to no-op switches. We can improve
the latter by eliminating the w/a around known no-op switches, but there
is an ulterior motive to keeping them.
Always emitting the context switch at the beginning of the request (and
relying on HW to skip unneeded switches) does have one key advantage.
Should we implement request reordering on Haswell, we will not know in
advance what the previous executing context was on the GPU and so we
would not be able to elide the MI_SET_CONTEXT commands ourselves and
always have to emit them. Having our hand forced now actually prepares
us for later.
Now since that context and mm follow the request, we no longer (and not
for a long time since requests took over!) require a trace point to tell
when we write the switch into the ring, since it is always. (This is
even more important when you remember that simply writing into the ring
bears no relation to the current mm.)
v2: Sandybridge has to agree to use LRI as well.
Testcase: igt/drv_selftests/live_hangcheck
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Matthew Auld <matthew.william.auld@gmail.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180611110845.31890-1-chris@chris-wilson.co.uk
We were not very carefully checking to see if an older request on the
engine was an earlier switch-to-kernel-context before deciding to emit a
new switch. The end result would be that we could get into a permanent
loop of trying to emit a new request to perform the switch simply to
flush the existing switch.
What we need is a means of tracking the completion of each timeline
versus the kernel context, that is to detect if a more recent request
has been submitted that would result in a switch away from the kernel
context. To realise this, we need only to look in our syncmap on the
kernel context and check that we have synchronized against all active
rings.
v2: Since all ringbuffer clients currently share the same timeline, we do
have to use the gem_context to distinguish clients.
As a bonus, include all the tracing used to debug the death inside
suspend.
v3: Test, test, test. Construct a selftest to exercise and assert the
expected behaviour that multiple switch-to-contexts do not emit
redundant requests.
Reported-by: Mika Kuoppala <mika.kuoppala@intel.com>
Fixes: a89d1f921c ("drm/i915: Split i915_gem_timeline into individual timelines")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180524081135.15278-1-chris@chris-wilson.co.uk
During request submission, we call the engine->schedule() function so
that we may reorder the active requests as required for inheriting the
new request's priority. This may schedule several tasklets to run on the
local CPU, but we will need to schedule the tasklets again for the new
request. Delay all the local tasklets until the end, so that we only
have to process the queue just once.
v2: Beware PREEMPT_RCU, as then local_bh_disable() is then not a
superset of rcu_read_lock().
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180507135731.10587-2-chris@chris-wilson.co.uk
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
We need to move to a more flexible timeline that doesn't assume one
fence context per engine, and so allow for a single timeline to be used
across a combination of engines. This means that preallocating a fence
context per engine is now a hindrance, and so we want to introduce the
singular timeline. From the code perspective, this has the notable
advantage of clearing up a lot of mirky semantics and some clumsy
pointer chasing.
By splitting the timeline up into a single entity rather than an array
of per-engine timelines, we can realise the goal of the previous patch
of tracking the timeline alongside the ring.
v2: Tweak wait_for_idle to stop the compiling thinking that ret may be
uninitialised.
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/20180502163839.3248-2-chris@chris-wilson.co.uk
In the future, we want to move a request between engines. To achieve
this, we first realise that we have two timelines in effect here. The
first runs through the GTT is required for ordering vma access, which is
tracked currently by engine. The second is implied by sequential
execution of commands inside the ringbuffer. This timeline is one that
maps to userspace's expectations when submitting requests (i.e. given the
same context, batch A is executed before batch B). As the rings's
timelines map to userspace and the GTT timeline an implementation
detail, move the timeline from the GTT into the ring itself (per-context
in logical-ring-contexts/execlists, or a global per-engine timeline for
the shared ringbuffers in legacy submission.
The two timelines are still assumed to be equivalent at the moment (no
migrating requests between engines yet) and so we can simply move from
one to the other without adding extra ordering.
v2: Reinforce that one isn't allowed to mix the engine execution
timeline with the client timeline from userspace (on the ring).
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/20180502163839.3248-1-chris@chris-wilson.co.uk
In the next patch, rings are the central timeline as requests may jump
between engines. Therefore in the future as we retire in order along the
engine timeline, we may retire out-of-order within a ring (as the ring now
occurs along multiple engines), leading to much hilarity in miscomputing
the position of ring->head.
As an added bonus, retiring along the ring reduces the penalty of having
one execlists client do cleanup for another (old legacy submission
shares a ring between all clients). The downside is that slow and
irregular (off the critical path) process of cleaning up stale requests
after userspace becomes a modicum less efficient.
In the long run, it will become apparent that the ordered
ring->request_list matches the ring->timeline, a fun challenge for the
future will be unifying the two lists to avoid duplication!
v2: We need both engine-order and ring-order processing to maintain our
knowledge of where individual rings have completed upto as well as
knowing what was last executing on any engine. And finally by decoupling
retiring the contexts on the engine and the timelines along the rings,
we do have to keep a reference to the context on each request
(previously it was guaranteed by the context being pinned).
v3: Not just a reference to the context, but we need to keep it pinned
as we manipulate the rings; i.e. we need a pin for both the manipulation
of the engine state during its retirements, and a separate pin for the
manipulation of the ring state.
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/20180430131503.5375-3-chris@chris-wilson.co.uk
In commit 9b6586ae9f ("drm/i915: Keep a global seqno per-engine"), we
moved from a global inflight counter to per-engine counters in the
hope that will be easy to run concurrently in future. However, with the
advent of the desire to move requests between engines, we do need a
global counter to preserve the semantics that no engine wraps in the
middle of a submit. (Although this semantic is now only required for gen7
semaphore support, which only supports greater-then comparisons!)
v2: Keep a global counter of all requests ever submitted and force the
reset when it wraps.
References: 9b6586ae9f ("drm/i915: Keep a global seqno per-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/20180430131503.5375-1-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
Both request_submit and request_unsubmit deal with transferring the
request from the client's timeline onto the execution timeline and back
again. As both functions deal with a pair of timeline's, using a
shorthand for just one of them is slightly confusing, especially as the
different functions use the shorthand for the alternate timeline.
Instead, use the full version of each timeline so it should be easier to
keep track of the transfer between the request/client and the engine.
v2: Refactor the common lock+list_move
v3: Be clear we require the other timeline list to be locked as well.
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/20180322131034.6036-1-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
