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
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
Supplement the per-engine HWSP with a per-timeline HWSP. That is a
per-request pointer through which we can check a local seqno,
abstracting away the presumption of a global seqno. In this first step,
we point each request back into the engine's HWSP so everything
continues to work with the global timeline.
v2: s/i915_request_hwsp/hwsp_seqno/ to emphasis that this is the current
HW value and that we are accessing it via i915_request merely as a
convenience.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190128181812.22804-1-chris@chris-wilson.co.uk
Latency is in the eye of the beholder. In the case where a client stops
and waits for the gpu, give that request chain a small priority boost
(not so that it overtakes higher priority clients, to preserve the
external ordering) so that ideally the wait completes earlier.
v2: Tvrtko recommends to keep the boost-from-user-stall as small as
possible and to allow new client flows to be preferred for interactivity
over stalls.
Testcase: igt/gem_sync/switch-default
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: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20181001144755.7978-3-chris@chris-wilson.co.uk
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
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
We have a few instances of checking seqno-1 to see if the HW has started
the request. Pull those together under a helper.
v2: Pull the !seqno assertion higher, as given seqno==1 we may indeed
check to see if we have started using seqno==0.
Suggested-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
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/20180806112605.20725-1-chris@chris-wilson.co.uk
In the next patch, we will want to be able to use more flexible request
timelines that can hop between engines. From the vma pov, we can then
not rely on the binding of this request to an engine and so can not
ensure that different requests are ordered through a per-engine
timeline, and so we must track activity of all timelines. (We track
activity on the vma itself to prevent unbinding from HW before the HW
has finished accessing it.)
v2: Switch to a rbtree for 32b safety (since using u64 as a radixtree
index is fraught with aliasing of unsigned longs).
v3: s/lookup_active/active_instance/ because we can never agree on names
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/20180706103947.15919-5-chris@chris-wilson.co.uk
In the next patch, we will want to start skipping requests on failing to
complete their payloads. So export the utility function current used to
make requests inoperable following a failed gpu reset.
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/20180706103947.15919-2-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
During suspend we want to flush out all active contexts and their
rendering. To do so we queue a request from the kernel's context, once
we know that request is done, we know the GPU is completely idle. To
speed up that switch bump the GPU clocks.
Switching to the kernel context prior to idling is also used to enforce
a barrier before changing OA properties, and when evicting active
rendering from the global GTT. All cases where we do want to
race-to-idle.
v2: Limit the boosting to only the switch before suspend.
v3: Limit it to the wait-for-idle on suspend.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: David Weinehall <david.weinehall@linux.intel.com>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Tested-by: David Weinehall <david.weinehall@linux.intel.com> #v1
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180531082246.9763-2-chris@chris-wilson.co.uk
To ease the frequent and ugly pointer dance of
&request->gem_context->engine[request->engine->id] during request
submission, store that pointer as request->hw_context. One major
advantage that we will exploit later is that this decouples the logical
context state from the engine itself.
v2: Set mock_context->ops so we don't crash and burn in selftests.
Cleanups from Tvrtko.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Acked-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180517212633.24934-3-chris@chris-wilson.co.uk
In the next patch, we want to store the intel_context pointer inside
i915_request, as it is frequently access via a convoluted dance when
submitting the request to hw. Having two context pointers inside
i915_request leads to confusion so first rename the existing
i915_gem_context pointer to i915_request.gem_context.
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/20180517212633.24934-1-chris@chris-wilson.co.uk
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
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
Having moved the priotree struct into i915_scheduler.h, identify it as
the scheduling element and rebrand into i915_sched. This becomes more
useful as we start attaching more information we require to propagate
through the scheduler.
v2: Use i915_sched_node for future distinctiveness
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/20180418184052.7129-2-chris@chris-wilson.co.uk
Over time the priotree has grown from a sorted list to a more
complicated structure for propagating constraints along the dependency
chain to try and resolve priority inversion. Start to segregate this
information from the rest of the request/fence tracking.
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/20180418184052.7129-1-chris@chris-wilson.co.uk
The goal here is to try and reduce the latency of signaling additional
requests following the wakeup from interrupt by reducing the list of
to-be-signaled requests from an rbtree to a sorted linked list. The
original choice of using an rbtree was to facilitate random insertions
of request into the signaler while maintaining a sorted list. However,
if we assume that most new requests are added when they are submitted,
we see those new requests in execution order making a insertion sort
fast, and the reduction in overhead of each signaler iteration
significant.
Since commit 56299fb7d9 ("drm/i915: Signal first fence from irq handler
if complete"), we signal most fences directly from notify_ring() in the
interrupt handler greatly reducing the amount of work that actually
needs to be done by the signaler kthread. All the thread is then
required to do is operate as the bottom-half, cleaning up after the
interrupt handler and preparing the next waiter. This includes signaling
all later completed fences in a saturated system, but on a mostly idle
system we only have to rebuild the wait rbtree in time for the next
interrupt. With this de-emphasis of the signaler's role, we want to
rejig it's datastructures to reduce the amount of work we require to
both setup the signal tree and maintain it on every interrupt.
References: 56299fb7d9 ("drm/i915: Signal first fence from irq handler if complete")
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>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180222092545.17216-1-chris@chris-wilson.co.uk
We want to de-emphasize the link between the request (dependency,
execution and fence tracking) from GEM and so rename the struct from
drm_i915_gem_request to i915_request. That is we may implement the GEM
user interface on top of requests, but they are an abstraction for
tracking execution rather than an implementation detail of GEM. (Since
they are not tied to HW, we keep the i915 prefix as opposed to intel.)
In short, the spatch:
@@
@@
- struct drm_i915_gem_request
+ struct i915_request
A corollary to contracting the type name, we also harmonise on using
'rq' shorthand for local variables where space if of the essence and
repetition makes 'request' unwieldy. For globals and struct members,
'request' is still much preferred for its clarity.
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>
Cc: Michał Winiarski <michal.winiarski@intel.com>
Cc: Michal Wajdeczko <michal.wajdeczko@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180221095636.6649-1-chris@chris-wilson.co.uk
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Michał Winiarski <michal.winiarski@intel.com>
Acked-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Chris Wilson