Ringbuffers are now being written to either through LLC or WC paths, so
treating them as simply iomem is no longer adequate. However, for the
older !llc hardware, the hardware is documentated as treating the TAIL
register update as serialising, so we can relax the barriers when filling
the rings (but even if it were not, it is still an uncached register write
and so serialising anyway.).
For simplicity, let's ignore the iomem annotation.
v2: Remove iomem from ringbuffer->virtual_address
v3: And for good measure add iomem elsewhere to keep sparse happy
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com> #v2
Link: http://patchwork.freedesktop.org/patch/msgid/1469005202-9659-8-git-send-email-chris@chris-wilson.co.uk
Link: http://patchwork.freedesktop.org/patch/msgid/1469017917-15134-7-git-send-email-chris@chris-wilson.co.uk
dma-buf provides a generic fence class for interoperation between
drivers. Internally we use the request structure as a fence, and so with
only a little bit of interfacing we can rebase those requests on top of
dma-buf fences. This will allow us, in the future, to pass those fences
back to userspace or between drivers.
v2: The fence_context needs to be globally unique, not just unique to
this device.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1469002875-2335-4-git-send-email-chris@chris-wilson.co.uk
Common code deserves to be put in a separate file from legacy and
execlists implementation for clarity and ease of maintenance.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Chris Wilson <chris-wilson.co.uk>
With the unified common engine setup done, and the execlist engine
initialization loop clearly split into two phases, we can eliminate
the separate legacy engine initialization code.
v2: Fix cleanup path for legacy.
v3: Rename constructors. (Chris Wilson)
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Chris Wilson <chris-wilson.co.uk>
intel_lrc.c has a table of "logical rings" (meaning engines), while
intel_ringbuffer.c has separately open-coded initialisation for each
engine. We can deduplicate this somewhat by using the same first-stage
engine-setup function for both modes.
So here we expose the function that transfers information from the
static table of (all) known engines to the dev_priv->engine array of
engines available on this device (adjusting the names along the way)
and then embed calls to it in both the LRC and the legacy-mode setup.
Signed-off-by: Dave Gordon <david.s.gordon@intel.com>
Reviewed-by: Chris Wilson <chris-wilson.co.uk>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
As we inspect both the tasklet (to check for an active bottom-half) and
set the irq-posted flag at the same time (both in the interrupt handler
and then in the bottom-halt), group those two together into the same
cacheline. (Not having total control over placement of the struct means
we can't guarantee the cacheline boundary, we need to align the kmalloc
and then each struct, but the grouping should help.)
v2: Try a couple of different names for the state touched by the user
interrupt handler.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1467805142-22219-3-git-send-email-chris@chris-wilson.co.uk
Now that we have (near) universal GPU recovery code, we can inject a
real hang from userspace and not need any fakery. Not only does this
mean that the testing is far more realistic, but we can simplify the
kernel in the process.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Arun Siluvery <arun.siluvery@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1467616119-4093-7-git-send-email-chris@chris-wilson.co.uk
The retire worker is a low frequency task that makes sure we retire
outstanding requests if userspace is being lax. We only need to start it
once as it remains active until the GPU is idle, so do a cheap test
before the more expensive queue_work(). A consequence of this is that we
need correct locking in the worker to make the hot path of request
submission cheap. To keep the symmetry and keep hangcheck strictly bound
by the GPU's wakelock, we move the cancel_sync(hangcheck) to the idle
worker before dropping the wakelock.
v2: Guard against RCU fouling the breadcrumbs bottom-half whilst we kick
the waiter.
v3: Remove the wakeref assertion squelching (now we hold a wakeref for
the hangcheck, any rpm error there is genuine).
v4: To prevent excess work when retiring requests, we split the busy
flag into two, a boolean to denote whether we hold the wakeref and a
bitmask of active engines.
v5: Reorder cancelling hangcheck upon idling to avoid a race where we
might cancel a hangcheck after being preempted by a new task
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
References: https://bugs.freedesktop.org/show_bug.cgi?id=88437
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1467616119-4093-1-git-send-email-chris@chris-wilson.co.uk
With only a single callsite for intel_engine_cs->irq_get and ->irq_put,
we can reduce the code size by moving the common preamble into the
caller, and we can also eliminate the reference counting.
For completeness, as we are no longer doing reference counting on irq,
rename the get/put vfunctions to enable/disable respectively and are
able to review the use of posting reads. We only require the
serialisation with hardware when enabling the interrupt (i.e. so we
cannot miss an interrupt by going to sleep before the hardware truly
enables it).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-18-git-send-email-chris@chris-wilson.co.uk
Under the assumption that enabling signaling will be a frequent
operation, lets preallocate our attachments for signaling inside the
(rather large) request struct (and so benefiting from the slab cache).
v2: Convert from void * to more meaningful names and types.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-17-git-send-email-chris@chris-wilson.co.uk
If we convert the tracing over from direct use of ring->irq_get() and
over to the breadcrumb infrastructure, we only have a single user of the
ring->irq_get and so we will be able to simplify the driver routines
(eliminating the redundant validation and irq refcounting).
Process context is preferred over softirq (or even hardirq) for a couple
of reasons:
- we already utilize process context to have fast wakeup of a single
client (i.e. the client waiting for the GPU inspects the seqno for
itself following an interrupt to avoid the overhead of a context
switch before it returns to userspace)
- engine->irq_seqno() is not suitable for use from an softirq/hardirq
context as we may require long waits (100-250us) to ensure the seqno
write is posted before we read it from the CPU
A signaling framework is a requirement for enabling dma-fences.
v2: Move to a signaling framework based upon the waiter.
v3: Track the first-signal to avoid having to walk the rbtree everytime.
v4: Mark the signaler thread as RT priority to reduce latency in the
indirect wakeups.
v5: Make failure to allocate the thread fatal.
v6: Rename kthreads to i915/signal:%u
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-16-git-send-email-chris@chris-wilson.co.uk
If we flag the seqno as potentially stale upon receiving an interrupt,
we can use that information to reduce the frequency that we apply the
heavyweight coherent seqno read (i.e. if we wake up a chain of waiters).
v2: Use cmpxchg to replace READ_ONCE/WRITE_ONCE for more explicit
control of the ordering wrt to interrupt generation and interrupt
checking in the bottom-half.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-14-git-send-email-chris@chris-wilson.co.uk
After the elimination of using the scratch page for Ironlake's
breadcrumb, we no longer need to kmap the object. We therefore can move
it into the high unmappable space and do not need to force the object to
be coherent (i.e. snooped on !llc platforms).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-9-git-send-email-chris@chris-wilson.co.uk
By using the same address for storing the HWS on every platform, we can
remove the platform specific vfuncs and reduce the get-seqno routine to
a single read of a cached memory location.
v2: Fix semaphore_passed() to look at the signaling engine (not the
waiter's)
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-8-git-send-email-chris@chris-wilson.co.uk
One particularly stressful scenario consists of many independent tasks
all competing for GPU time and waiting upon the results (e.g. realtime
transcoding of many, many streams). One bottleneck in particular is that
each client waits on its own results, but every client is woken up after
every batchbuffer - hence the thunder of hooves as then every client must
do its heavyweight dance to read a coherent seqno to see if it is the
lucky one.
Ideally, we only want one client to wake up after the interrupt and
check its request for completion. Since the requests must retire in
order, we can select the first client on the oldest request to be woken.
Once that client has completed his wait, we can then wake up the
next client and so on. However, all clients then incur latency as every
process in the chain may be delayed for scheduling - this may also then
cause some priority inversion. To reduce the latency, when a client
is added or removed from the list, we scan the tree for completed
seqno and wake up all the completed waiters in parallel.
Using igt/benchmarks/gem_latency, we can demonstrate this effect. The
benchmark measures the number of GPU cycles between completion of a
batch and the client waking up from a call to wait-ioctl. With many
concurrent waiters, with each on a different request, we observe that
the wakeup latency before the patch scales nearly linearly with the
number of waiters (before external factors kick in making the scaling much
worse). After applying the patch, we can see that only the single waiter
for the request is being woken up, providing a constant wakeup latency
for every operation. However, the situation is not quite as rosy for
many waiters on the same request, though to the best of my knowledge this
is much less likely in practice. Here, we can observe that the
concurrent waiters incur extra latency from being woken up by the
solitary bottom-half, rather than directly by the interrupt. This
appears to be scheduler induced (having discounted adverse effects from
having a rbtree walk/erase in the wakeup path), each additional
wake_up_process() costs approximately 1us on big core. Another effect of
performing the secondary wakeups from the first bottom-half is the
incurred delay this imposes on high priority threads - rather than
immediately returning to userspace and leaving the interrupt handler to
wake the others.
To offset the delay incurred with additional waiters on a request, we
could use a hybrid scheme that did a quick read in the interrupt handler
and dequeued all the completed waiters (incurring the overhead in the
interrupt handler, not the best plan either as we then incur GPU
submission latency) but we would still have to wake up the bottom-half
every time to do the heavyweight slow read. Or we could only kick the
waiters on the seqno with the same priority as the current task (i.e. in
the realtime waiter scenario, only it is woken up immediately by the
interrupt and simply queues the next waiter before returning to userspace,
minimising its delay at the expense of the chain, and also reducing
contention on its scheduler runqueue). This is effective at avoid long
pauses in the interrupt handler and at avoiding the extra latency in
realtime/high-priority waiters.
v2: Convert from a kworker per engine into a dedicated kthread for the
bottom-half.
v3: Rename request members and tweak comments.
v4: Use a per-engine spinlock in the breadcrumbs bottom-half.
v5: Fix race in locklessly checking waiter status and kicking the task on
adding a new waiter.
v6: Fix deciding when to force the timer to hide missing interrupts.
v7: Move the bottom-half from the kthread to the first client process.
v8: Reword a few comments
v9: Break the busy loop when the interrupt is unmasked or has fired.
v10: Comments, unnecessary churn, better debugging from Tvrtko
v11: Wake all completed waiters on removing the current bottom-half to
reduce the latency of waking up a herd of clients all waiting on the
same request.
v12: Rearrange missed-interrupt fault injection so that it works with
igt/drv_missed_irq_hang
v13: Rename intel_breadcrumb and friends to intel_wait in preparation
for signal handling.
v14: RCU commentary, assert_spin_locked
v15: Hide BUG_ON behind the compiler; report on gem_latency findings.
v16: Sort seqno-groups by priority so that first-waiter has the highest
task priority (and so avoid priority inversion).
v17: Add waiters to post-mortem GPU hang state.
v18: Return early for a completed wait after acquiring the spinlock.
Avoids adding ourselves to the tree if the is already complete, and
skips the awkward question of why we don't do completion wakeups for
waits earlier than or equal to ourselves.
v19: Prepare for init_breadcrumbs to fail. Later patches may want to
allocate during init, so be prepared to propagate back the error code.
Testcase: igt/gem_concurrent_blit
Testcase: igt/benchmarks/gem_latency
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com>
Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Cc: Dave Gordon <david.s.gordon@intel.com>
Cc: "Goel, Akash" <akash.goel@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18
Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
Replace the macro initializer with a programatic loop which
results in smaller code and hopefully just as clear.
v2: Rebase.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Our goal is to rename the anonymous per-engine struct beneath the
current intel_context. However, after a lively debate resolving around
the confusion between intel_context_engine and intel_engine_context, the
realisation is that the two structs target different users. The outer
struct is API / user facing, and so carries the higher level GEM
information. The inner struct is hw facing. Thus we want to name the
inner struct intel_context and the outer one i915_gem_context. As the
first step, we need to rename the current struct:
s/struct intel_context/struct i915_gem_context/
which fits much better with its constructors already conveying the
i915_gem_context prefix!
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Dave Gordon <david.s.gordon@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1464098023-3294-1-git-send-email-chris@chris-wilson.co.uk
text data bss dec hex filename
6309351 3578714 696320 10584385 a18141 vmlinux
6308391 3578714 696320 10583425 a17d81 vmlinux
Almost 1KiB of code reduction.
v2: More s/INTEL_INFO()->gen/INTEL_GEN()/ and IS_GENx() conversions
text data bss dec hex filename
6304579 3578778 696320 10579677 a16edd vmlinux
6303427 3578778 696320 10578525 a16a5d vmlinux
Now over 1KiB!
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1462545621-30125-3-git-send-email-chris@chris-wilson.co.uk
With the introduction of a distinct engine->id vs the hardware id, we need
to fix up the value we use for selecting the target engine when signaling
a semaphore. Note that these values can be merged with engine->guc_id.
Fixes: de1add3605
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1461932305-14637-3-git-send-email-chris@chris-wilson.co.uk
For legacy ringbuffer mode, we need the new ordered breadcrumb emission
tried and tested on execlists in order to avoid the dreaded "missed
interrupt" syndrome. A secondary advantage of the execlists method is
that it writes to an arbitrary address, useful if one wants to write a
breadcrumb elsewhere.
This fix is taken from commit 7c17d37737 (drm/i915: Use ordered seqno
write interrupt generation on gen8+ execlists).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1461932305-14637-1-git-send-email-chris@chris-wilson.co.uk
With 5 rings and a flush, we need 192 bytes of space to emit the
breadcrumb and semaphores. However, we need some spare room the size of
the single largest packet (36 dwords, 144 bytes) to accommodate
wraparound giving a grand total of 336 bytes
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1461917226-9132-1-git-send-email-chris@chris-wilson.co.uk
With the previous patch having extended the pinned lifetime of
contexts by referencing the previous context from the current
request until the latter is retired (completed by the GPU),
we can now remove usage of execlist retired queue entirely.
This is because the above now guarantees that all execlist
object access requirements are satisfied by this new tracking,
and we can stop taking additional references and stop keeping
request on the execlists retired queue.
The latter was a source of significant scalability issues in
the driver causing performance hits on some tests. Most
dramatical of which was igt/gem_close_race which had run time
in tens of minutes which is now reduced to tens of seconds.
Signed-off-by: Tvrtko Ursulin <tvrtko@ursulin.net>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/1461833819-3991-24-git-send-email-chris@chris-wilson.co.uk
Now that we share intel_ring_begin(), reserving space for the tail of
the request is identical between legacy/execlists and so the tautology
can be removed. In the process, we move the reserved space tracking
from the ringbuffer on to the request. This is to enable us to reorder
the reserved space allocation in the next patch.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1461833819-3991-13-git-send-email-chris@chris-wilson.co.uk
Combine the near identical implementations of intel_logical_ring_begin()
and intel_ring_begin() - the only difference is that the logical wait
has to check for a matching ring (which is assumed by legacy).
In the process some debug messages are culled as there were following a
WARN if we hit an actual error.
v2: Updated commentary
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1461833819-3991-12-git-send-email-chris@chris-wilson.co.uk
Rather than blindly waking up all forcewake domains on command
submission, we can teach each engine what is (or are) the correct
one to take.
On platforms with multiple forcewake domains like VLV, CHV, SKL
and BXT, this has the potential of lowering the GPU and CPU
power use and submission latency.
To implement it we add a function named
intel_uncore_forcewake_for_reg whose purpose is to query which
forcewake domains need to be taken to read or write a specific
register with raw mmio accessors.
These enables the execlists engine setup to query which
forcewake domains are relevant per engine on the currently
running platform.
v2:
* Kerneldoc.
* Split from intel_uncore.c macro extraction, WARN_ON,
no warns on old platforms. (Chris Wilson)
v3:
* Single domain per engine, mention all registers,
bi-directional function and a new name, fix handling
of gen6 and gen7 writes. (Chris Wilson)
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/1460468251-14069-1-git-send-email-tvrtko.ursulin@linux.intel.com
When reading from the HWS page, we use barrier() to prevent the compiler
optimising away the read from the volatile (may be updated by the GPU)
memory address. This is more suited to READ_ONCE(); make it so.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1460195877-20520-5-git-send-email-chris@chris-wilson.co.uk
Rather than call a function to compute the matching cachelines and
clflush them, just call the clflush *instruction* directly. We also know
that we can use the unpatched plain clflush rather than the clflushopt
alternative.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1460195877-20520-4-git-send-email-chris@chris-wilson.co.uk
Only declare a missed interrupt if we find that the GPU is idle with
waiters and a hangcheck interval has passed in which no new user
interrupts have been raised.
v2: Clear the stuck interrupt marker between successful batches
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1460195877-20520-3-git-send-email-chris@chris-wilson.co.uk
In order to simplify future patches, extract the
lazy_coherency optimisation our of the engine->get_seqno() vfunc into
its own callback.
v2: Rename the barrier to engine->irq_seqno_barrier to try and better
reflect that the barrier is only required after the user interrupt before
reading the seqno (to ensure that the seqno update lands in time as we
do not have strict seqno-irq ordering on all platforms).
Reviewed-by: Dave Gordon <david.s.gordon@intel.com> [#v2]
v3: Comments for hangcheck paranoia. Mika wanted to keep the extra
barrier inside the hangcheck, just in case. I can argue that it doesn't
provide a barrier against anything, but the side-effects of applying the
barrier may prevent a false declaration of a hung GPU.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Cc: Dave Gordon <david.s.gordon@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1460195877-20520-2-git-send-email-chris@chris-wilson.co.uk
Doing a lot of work in the interrupt handler introduces huge
latencies to the system as a whole.
Most dramatic effect can be seen by running an all engine
stress test like igt/gem_exec_nop/all where, when the kernel
config is lean enough, the whole system can be brought into
multi-second periods of complete non-interactivty. That can
look for example like this:
NMI watchdog: BUG: soft lockup - CPU#0 stuck for 23s! [kworker/u8:3:143]
Modules linked in: [redacted for brevity]
CPU: 0 PID: 143 Comm: kworker/u8:3 Tainted: G U L 4.5.0-160321+ #183
Hardware name: Intel Corporation Broadwell Client platform/WhiteTip Mountain 1
Workqueue: i915 gen6_pm_rps_work [i915]
task: ffff8800aae88000 ti: ffff8800aae90000 task.ti: ffff8800aae90000
RIP: 0010:[<ffffffff8104a3c2>] [<ffffffff8104a3c2>] __do_softirq+0x72/0x1d0
RSP: 0000:ffff88014f403f38 EFLAGS: 00000206
RAX: ffff8800aae94000 RBX: 0000000000000000 RCX: 00000000000006e0
RDX: 0000000000000020 RSI: 0000000004208060 RDI: 0000000000215d80
RBP: ffff88014f403f80 R08: 0000000b1b42c180 R09: 0000000000000022
R10: 0000000000000004 R11: 00000000ffffffff R12: 000000000000a030
R13: 0000000000000082 R14: ffff8800aa4d0080 R15: 0000000000000082
FS: 0000000000000000(0000) GS:ffff88014f400000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007fa53b90c000 CR3: 0000000001a0a000 CR4: 00000000001406f0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Stack:
042080601b33869f ffff8800aae94000 00000000fffc2678 ffff88010000000a
0000000000000000 000000000000a030 0000000000005302 ffff8800aa4d0080
0000000000000206 ffff88014f403f90 ffffffff8104a716 ffff88014f403fa8
Call Trace:
<IRQ>
[<ffffffff8104a716>] irq_exit+0x86/0x90
[<ffffffff81031e7d>] smp_apic_timer_interrupt+0x3d/0x50
[<ffffffff814f3eac>] apic_timer_interrupt+0x7c/0x90
<EOI>
[<ffffffffa01c5b40>] ? gen8_write64+0x1a0/0x1a0 [i915]
[<ffffffff814f2b39>] ? _raw_spin_unlock_irqrestore+0x9/0x20
[<ffffffffa01c5c44>] gen8_write32+0x104/0x1a0 [i915]
[<ffffffff8132c6a2>] ? n_tty_receive_buf_common+0x372/0xae0
[<ffffffffa017cc9e>] gen6_set_rps_thresholds+0x1be/0x330 [i915]
[<ffffffffa017eaf0>] gen6_set_rps+0x70/0x200 [i915]
[<ffffffffa0185375>] intel_set_rps+0x25/0x30 [i915]
[<ffffffffa01768fd>] gen6_pm_rps_work+0x10d/0x2e0 [i915]
[<ffffffff81063852>] ? finish_task_switch+0x72/0x1c0
[<ffffffff8105ab29>] process_one_work+0x139/0x350
[<ffffffff8105b186>] worker_thread+0x126/0x490
[<ffffffff8105b060>] ? rescuer_thread+0x320/0x320
[<ffffffff8105fa64>] kthread+0xc4/0xe0
[<ffffffff8105f9a0>] ? kthread_create_on_node+0x170/0x170
[<ffffffff814f351f>] ret_from_fork+0x3f/0x70
[<ffffffff8105f9a0>] ? kthread_create_on_node+0x170/0x170
I could not explain, or find a code path, which would explain
a +20 second lockup, but from some instrumentation it was
apparent the interrupts off proportion of time was between
10-25% under heavy load which is quite bad.
When a interrupt "cliff" is reached, which was >~320k irq/s on
my machine, the whole system goes into a terrible state of the
above described multi-second lockups.
By moving the GT interrupt handling to a tasklet in a most
simple way, the problem above disappears completely.
Testing the effect on sytem-wide latencies using
igt/gem_syslatency shows the following before this patch:
gem_syslatency: cycles=1532739, latency mean=416531.829us max=2499237us
gem_syslatency: cycles=1839434, latency mean=1458099.157us max=4998944us
gem_syslatency: cycles=1432570, latency mean=2688.451us max=1201185us
gem_syslatency: cycles=1533543, latency mean=416520.499us max=2498886us
This shows that the unrelated process is experiencing huge
delays in its wake-up latency. After the patch the results
look like this:
gem_syslatency: cycles=808907, latency mean=53.133us max=1640us
gem_syslatency: cycles=862154, latency mean=62.778us max=2117us
gem_syslatency: cycles=856039, latency mean=58.079us max=2123us
gem_syslatency: cycles=841683, latency mean=56.914us max=1667us
Showing a huge improvement in the unrelated process wake-up
latency. It also shows an approximate halving in the number
of total empty batches submitted during the test. This may
not be worrying since the test puts the driver under
a very unrealistic load with ncpu threads doing empty batch
submission to all GPU engines each.
Another benefit compared to the hard-irq handling is that now
work on all engines can be dispatched in parallel since we can
have up to number of CPUs active tasklets. (While previously
a single hard-irq would serially dispatch on one engine after
another.)
More interesting scenario with regards to throughput is
"gem_latency -n 100" which shows 25% better throughput and
CPU usage, and 14% better dispatch latencies.
I did not find any gains or regressions with Synmark2 or
GLbench under light testing. More benchmarking is certainly
required.
v2:
* execlists_lock should be taken as spin_lock_bh when
queuing work from userspace now. (Chris Wilson)
* uncore.lock must be taken with spin_lock_irq when
submitting requests since that now runs from either
softirq or process context.
v3:
* Expanded commit message with more testing data;
* converted missed locking sites to _bh;
* added execlist_lock comment. (Chris Wilson)
v4:
* Mention dispatch parallelism in commit. (Chris Wilson)
* Do not hold uncore.lock over MMIO reads since the block
is already serialised per-engine via the tasklet itself.
(Chris Wilson)
* intel_lrc_irq_handler should be static. (Chris Wilson)
* Cancel/sync the tasklet on GPU reset. (Chris Wilson)
* Document and WARN that tasklet cannot be active/pending
on engine cleanup. (Chris Wilson/Imre Deak)
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Imre Deak <imre.deak@intel.com>
Testcase: igt/gem_exec_nop/all
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94350
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/1459768316-6670-1-git-send-email-tvrtko.ursulin@linux.intel.com
For Haswell, we will want another table of registers while retaining
the large common table of whitelisted registers shared by all gen7
devices.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
[danvet: Pipe patch through sed -e 's/\<ring\>/engine/g' to make it
apply.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Some trivial ones, first pass done with Coccinelle:
@@
@@
(
- I915_NUM_RINGS
+ I915_NUM_ENGINES
|
- intel_ring_flag
+ intel_engine_flag
|
- for_each_ring
+ for_each_engine
|
- i915_gem_request_get_ring
+ i915_gem_request_get_engine
|
- intel_ring_idle
+ intel_engine_idle
|
- i915_gem_reset_ring_status
+ i915_gem_reset_engine_status
|
- i915_gem_reset_ring_cleanup
+ i915_gem_reset_engine_cleanup
|
- init_ring_lists
+ init_engine_lists
)
But that didn't fully work so I cleaned it up with:
for f in *.[hc]; do sed -i -e s/I915_NUM_RINGS/I915_NUM_ENGINES/ $f; done
for f in *.[hc]; do sed -i -e s/i915_gem_request_get_ring/i915_gem_request_get_engine/ $f; done
for f in *.[hc]; do sed -i -e s/intel_ring_flag/intel_engine_flag/ $f; done
for f in *.[hc]; do sed -i -e s/intel_ring_idle/intel_engine_idle/ $f; done
for f in *.[hc]; do sed -i -e s/init_ring_lists/init_engine_lists/ $f; done
for f in *.[hc]; do sed -i -e s/i915_gem_reset_ring_cleanup/i915_gem_reset_engine_cleanup/ $f; done
for f in *.[hc]; do sed -i -e s/i915_gem_reset_ring_status/i915_gem_reset_engine_status/ $f; done
v2: Rebase.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
With full-ppgtt, it takes the GPU an eon to traverse the entire 256PiB
address space, causing a loop to be detected. Under the current scheme,
if ACTHD walks off the end of a batch buffer and into an empty
address space, we "never" detect the hang. If we always increment the
score as the ACTHD is progressing then we will eventually timeout (after
~46.5s (31 * 1.5s) without advancing onto a new batch). To counter act
this, increase the amount we reduce the score for good batches, so that
only a series of almost-bad batches trigger a full reset. DoS detection
suffers slightly but series of long running shader tests will benefit.
Based on a patch from Chris Wilson.
Testcase: igt/drv_hangman/hangcheck-unterminated
Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/1456930109-21532-1-git-send-email-mika.kuoppala@intel.com
Assorted changes in the areas of code cleanup, reduction of
invariant conditional in the interrupt handler and lock
contention and MMIO access optimisation.
* Remove needless initialization.
* Improve cache locality by reorganizing code and/or using
branch hints to keep unexpected or error conditions out
of line.
* Favor busy submit path vs. empty queue.
* Less branching in hot-paths.
v2:
* Avoid mmio reads when possible. (Chris Wilson)
* Use natural integer size for csb indices.
* Remove useless return value from execlists_update_context.
* Extract 32-bit ppgtt PDPs update so it is out of line and
shared with two callers.
* Grab forcewake across all mmio operations to ease the
load on uncore lock and use chepear mmio ops.
v3:
* Removed some more pointless u8 data types.
* Removed unused return from execlists_context_queue.
* Commit message updates.
v4:
* Unclumsify the unqueue if statement. (Chris Wilson)
* Hide forcewake from the queuing function. (Chris Wilson)
Version 3 now makes the irq handling code path ~20% smaller on
48-bit PPGTT hardware, and a little bit less elsewhere. Hot
paths are mostly in-line now and hammering on the uncore
spinlock is greatly reduced together with mmio traffic to an
extent.
Benchmarking with "gem_latency -n 100" (keep submitting
batches with 100 nop instruction) shows approximately 4% higher
throughput, 2% less CPU time and 22% smaller latencies. This was
on a big-core while small-cores could benefit even more.
Most likely reason for the improvements are the MMIO
optimization and uncore lock traffic reduction.
One odd result is with "gem_latency -n 0" (dispatching empty
batches) which shows 5% more throughput, 8% less CPU time,
25% better producer and consumer latencies, but 15% higher
dispatch latency which is yet unexplained.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/1456505912-22286-1-git-send-email-tvrtko.ursulin@linux.intel.com
Previously GuC uses ring id as engine id because of same definition.
But this is not true since this commit:
commit de1add3605
Author: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Date: Fri Jan 15 15:12:50 2016 +0000
drm/i915: Decouple execbuf uAPI from internal implementation
Added GuC engine id into GuC interface to decouple it from ring id used
by driver.
v2: Keep ring name print out in debugfs; using for_each_ring() where
possible to keep driver consistent. (Chris W.)
Signed-off-by: Alex Dai <yu.dai@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1453579094-29860-1-git-send-email-yu.dai@intel.com
Tvrtko was looking through the execbuffer-ioctl and noticed that the
uABI was tightly coupled to our internal engine identifiers. Close
inspection also revealed that we leak those internal engine identifiers
through the busy-ioctl, and those internal identifiers already do not
match the user identifiers. Fortuitiously, there is only one user of the
set of busy rings from the busy-ioctl, and they only wish to choose
between the RENDER and the BLT engines.
Let's fix the userspace ABI while we still can.
v2: Update the uAPI documentation to explain the identifiers.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Testcase: igt/gem_busy
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1452876706-21620-1-git-send-email-chris@chris-wilson.co.uk
At the moment execbuf ring selection is fully coupled to
internal ring ids which is not a good thing on its own.
This dependency is also spread between two source files and
not spelled out at either side which makes it hidden and
fragile.
This patch decouples this dependency by introducing an explicit
translation table of execbuf uAPI to ring id close to the only
call site (i915_gem_do_execbuffer).
This way we are free to change driver internal implementation
details without breaking userspace. All state relating to the
uAPI is now contained in, or next to, i915_gem_do_execbuffer.
As a side benefit, this patch decreases the compiled size
of i915_gem_do_execbuffer.
v2: Extract ring selection into eb_select_ring. (Chris Wilson)
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1452870770-13981-1-git-send-email-tvrtko.ursulin@linux.intel.com
Broadwell and later currently use the same unordered command sequence to
update the seqno in the HWS status page and then assert the user
interrupt. We should apply the w/a from legacy (where we do an mmio
read to delay the seqno read after the interrupt), but this is not
enough to enforce coherent seqno visibilty on Skylake. Rather than
search for the proper post-interrupt seqno barrier, use a strongly
ordered command sequence to write the seqno, then assert the user
interrupt from the ring.
v2: Move around the wa tail dwords to avoid adding duplicate code.
v3: Add references, comments on workarounds and bit5 check.
References: https://bugs.freedesktop.org/show_bug.cgi?id=93693
Testcase: igt/gem_ring_sync_loop #skl
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1453297415-17793-1-git-send-email-mika.kuoppala@intel.com
Now that we've eliminated a lot of uses of ring->default_context,
we can eliminate the pointer itself.
All the engines share the same default intel_context, so we can just
keep a single reference to it in the dev_priv structure rather than one
in each of the engine[] elements. This make refcounting more sensible
too, as we now have a refcount of one for the one pointer, rather than
a refcount of one but multiple pointers.
From an idea by Chris Wilson.
v2: transform an extra instance of ring->default_context introduced by
42f1cae8c drm/i915: Restore inhibiting the load of the default context
That patch's commentary includes:
v2: Mark the global default context as uninitialized on GPU reset so
that the context-local workarounds are reloaded upon re-enabling
The code implementing that now also benefits from the replacement of
the multiple (per-ring) pointers to the default context with a single
pointer to the unique kernel context.
v4: Rebased, remove underused local (Nick Hoath)
Signed-off-by: Dave Gordon <david.s.gordon@intel.com>
Reviewed-by: Nick Hoath <nicholas.hoath@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/1453230175-19330-3-git-send-email-david.s.gordon@intel.com
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Purpose is to avoid calling i915_gem_obj_ggtt_offset from the
interrupt context without the big lock held.
v2: Renamed gtt_start to gtt_offset. (Daniel Vetter)
v3: Cache the VMA instead of address. (Chris Wilson)
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/1452870629-13830-2-git-send-email-tvrtko.ursulin@linux.intel.com
Chris Wilson