The i915_gem_init_hw() function calls a bunch of smaller initialisation
functions. Multiple of which have generic sections and per ring sections. This
means multiple passes are done over the rings. Each pass writes data to the ring
which floats around in that ring's OLR until some random point in the future
when an add_request() is done by some random other piece of code.

This patch breaks i915_ppgtt_init_hw() in two with the per ring initialisation
now being done in i915_ppgtt_init_ring(). The ring looping is now done at the
top level in i915_gem_init_hw().

v2: Fix dumb loop variable re-use.

For: VIZ-5115
Signed-off-by: NJohn Harrison <John.C.Harrison@Intel.com>
Reviewed-by: Tomas Elf <tomas.elf@intel.com> (v1)
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Added explicit request creation and submission to the GPU idle code path.

For: VIZ-5115
Signed-off-by: NJohn Harrison <John.C.Harrison@Intel.com>
Reviewed-by: NTomas Elf <tomas.elf@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

In order to explcitly track all GPU work (and completely remove the outstanding
lazy request), it is necessary to add extra i915_add_request() calls to various
places. Some of these do not need the implicit cache flush done as part of the
standard batch buffer submission process.

This patch adds a flag to _add_request() to specify whether the flush is
required or not.

For: VIZ-5115
Signed-off-by: NJohn Harrison <John.C.Harrison@Intel.com>
Reviewed-by: NTomas Elf <tomas.elf@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

The alloc_request() function does not actually return the newly allocated
request. Instead, it must be pulled from ring->outstanding_lazy_request. This
patch fixes this so that code can create a request and start using it knowing
exactly which request it actually owns.

v2: Updated for new i915_gem_request_alloc() scheme.

For: VIZ-5115
Signed-off-by: NJohn Harrison <John.C.Harrison@Intel.com>
Reviewed-by: NTomas Elf <tomas.elf@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

In execlist mode, the context object pointer is written in to the request
structure (and reference counted) at the point of request creation. In legacy
mode, this only happens inside i915_add_request().

This patch updates the legacy code path to match the execlist version. This
allows all the intermediate code between request creation and request submission
to get at the context object given only a request structure. Thus negating the
need to pass context pointers here, there and everywhere.

v2: Moved the context reference so it does not need to be undone if the
get_seqno() fails.

v3: Fixed execlist mode always hitting a warning about invalid last_contexts
(which don't exist in execlist mode).

v4: Updated for new i915_gem_request_alloc() scheme.

For: VIZ-5115
Signed-off-by: NJohn Harrison <John.C.Harrison@Intel.com>
Reviewed-by: NTomas Elf <tomas.elf@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

The i915_add_request() function is called to keep track of work that has been
written to the ring buffer. It adds epilogue commands to track progress (seqno
updates and such), moves the request structure onto the right list and other
such house keeping tasks. However, the work itself has already been written to
the ring and will get executed whether or not the add request call succeeds. So
no matter what goes wrong, there isn't a whole lot of point in failing the call.

At the moment, this is fine(ish). If the add request does bail early on and not
do the housekeeping, the request will still float around in the
ring->outstanding_lazy_request field and be picked up next time. It means
multiple pieces of work will be tagged as the same request and driver can't
actually wait for the first piece of work until something else has been
submitted. But it all sort of hangs together.

This patch series is all about removing the OLR and guaranteeing that each piece
of work gets its own personal request. That means that there is no more
'hoovering up of forgotten requests'. If the request does not get tracked then
it will be leaked. Thus the add request call _must_ not fail. The previous patch
should have already ensured that it _will_ not fail by removing the potential
for running out of ring space. This patch enforces the rule by actually removing
the early exit paths and the return code.

Note that if something does manage to fail and the epilogue commands don't get
written to the ring, the driver will still hang together. The request will be
added to the tracking lists. And as in the old case, any subsequent work will
generate a new seqno which will suffice for marking the old one as complete.

v2: Improved WARNings (Tomas Elf review request).

For: VIZ-5115
Signed-off-by: NJohn Harrison <John.C.Harrison@Intel.com>
Reviewed-by: NTomas Elf <tomas.elf@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

It is a bad idea for i915_add_request() to fail. The work will already have been
send to the ring and will be processed, but there will not be any tracking or
management of that work.

The only way the add request call can fail is if it can't write its epilogue
commands to the ring (cache flushing, seqno updates, interrupt signalling). The
reasons for that are mostly down to running out of ring buffer space and the
problems associated with trying to get some more. This patch prevents that
situation from happening in the first place.

When a request is created, it marks sufficient space as reserved for the
epilogue commands. Thus guaranteeing that by the time the epilogue is written,
there will be plenty of space for it. Note that a ring_begin() call is required
to actually reserve the space (and do any potential waiting). However, that is
not currently done at request creation time. This is because the ring_begin()
code can allocate a request. Hence calling begin() from the request allocation
code would lead to infinite recursion! Later patches in this series remove the
need for begin() to do the allocate. At that point, it becomes safe for the
allocate to call begin() and really reserve the space.

Until then, there is a potential for insufficient space to be available at the
point of calling i915_add_request(). However, that would only be in the case
where the request was created and immediately submitted without ever calling
ring_begin() and adding any work to that request. Which should never happen. And
even if it does, and if that request happens to fall down the tiny window of
opportunity for failing due to being out of ring space then does it really
matter because the request wasn't doing anything in the first place?

v2: Updated the 'reserved space too small' warning to include the offending
sizes. Added a 'cancel' operation to clean up when a request is abandoned. Added
re-initialisation of tracking state after a buffer wrap to keep the sanity
checks accurate.

v3: Incremented the reserved size to accommodate Ironlake (after finally
managing to run on an ILK system). Also fixed missing wrap code in LRC mode.

v4: Added extra comment and removed duplicate WARN (feedback from Tomas).

For: VIZ-5115
CC: Tomas Elf <tomas.elf@intel.com>
Signed-off-by: NJohn Harrison <John.C.Harrison@Intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

This patch doesn't have any functional change, but organize fruntbuffer
invalidate and busy by removing unecesarry signature argument for ring.

It was unsed on mark_fb_busy and only used on fb_obj_invalidate for the
same ORIGIN_CS usage. So let's clean it a bit

Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: NRodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Apparently we can have requests even if though the active list is empty,
so do the request retirement regardless of whether there's anything
on the active list.

The way it happened here is that during suspend intel_ring_idle()
notices the olr hanging around and then proceeds to get rid of it by
adding a request. However since there was nothing on the active lists
i915_gem_retire_requests() didn't clean those up, and so the idle work
never runs, and we leave the GPU "busy" during suspend resulting in a
WARN later.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NJani Nikula <jani.nikula@intel.com>

In commit 1854d5ca
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Tue Apr 7 16:20:32 2015 +0100

    drm/i915: Deminish contribution of wait-boosting from clients

we removed an atomic timer based check for allowing waitboosting and
moved it below the mutex taken during RPS. However, that mutex can be
held for long periods of time on Vallyview/Cherryview as communication
with the PCU is slow. As clients may frequently wait for results (e.g.
such as tranform feedback) we introduced contention between the client
and the RPS worker. We can take advantage of the RPS worker, by
switching the wait boost decision to use spin locks and defer the
actual reclocking to the worker.

Fixes a regression of up to 45% on Baytrail and Baswell!

v2 (Daniel):
- Use max_freq_softlimit instead of the not-yet-merged boost
  frequency.
- Don't inject a fake irq into the boost work, instead treat
  client_boost as just another legit waker.

v3: Drop the now unused mask (Chris).

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=90112
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> (v1)
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

As Daniel commented on

commit b7ffe1362c5f468b853223acc9268804aa92afc8
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Mon Apr 27 13:41:24 2015 +0100

    drm/i915: Free RPS boosts for all laggards

it is better to be explicit when sharing hardcoded values such as
throttle/boost timeouts. Make it so!
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

After allocating from the slab cache, we then need to free the request
back into the slab cache upon error (and not call kfree as that leads
to eventual memory corruption).

Fixes regression from
commit efab6d8d
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Tue Apr 7 16:20:57 2015 +0100

    drm/i915: Use a separate slab for requests
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

If the client stalls on a congested request, chosen to be 20ms old to
match throttling, allow the client a free RPS boost.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
[danvet: s/rq/req/]
[danvet: s/0/NULL/ reported by 0-day build]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Now that we have internal clients, rather than faking a whole
drm_i915_file_private just for tracking RPS boosts, create a new struct
intel_rps_client and pass it along when waiting.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
[danvet: s/rq/req/]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Ring switches can occur many times per frame, and are often out of
control, causing frequent RPS boosting for no practical benefit. Treat
the sw semaphore synchronisation as a separate client and only allow it
to boost once per busy/idle cycle.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
[danvet: s/rq/req/]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Currently, we only track the last request globally across all engines.
This prevents us from issuing concurrent read requests on e.g. the RCS
and BCS engines (or more likely the render and media engines). Without
semaphores, we incur costly stalls as we synchronise between rings -
greatly impacting the current performance of Broadwell versus Haswell in
certain workloads (like video decode). With the introduction of
reference counted requests, it is much easier to track the last request
per ring, as well as the last global write request so that we can
optimise inter-engine read read requests (as well as better optimise
certain CPU waits).

v2: Fix inverted readonly condition for nonblocking waits.
v3: Handle non-continguous engine array after waits
v4: Rebase, tidy, rewrite ring list debugging
v5: Use obj->active as a bitfield, it looks cool
v6: Micro-optimise, mostly involving moving code around
v7: Fix retire-requests-upto for execlists (and multiple rq->ringbuf)
v8: Rebase
v9: Refactor i915_gem_object_sync() to allow the compiler to better
optimise it.

Benchmark: igt/gem_read_read_speed
hsw:gt3e (with semaphores):
Before: Time to read-read 1024k:		275.794µs
After:  Time to read-read 1024k:		123.260µs

hsw:gt3e (w/o semaphores):
Before: Time to read-read 1024k:		230.433µs
After:  Time to read-read 1024k:		124.593µs

bdw-u (w/o semaphores):             Before          After
Time to read-read 1x1:            26.274µs       10.350µs
Time to read-read 128x128:        40.097µs       21.366µs
Time to read-read 256x256:        77.087µs       42.608µs
Time to read-read 512x512:       281.999µs      181.155µs
Time to read-read 1024x1024:    1196.141µs     1118.223µs
Time to read-read 2048x2048:    5639.072µs     5225.837µs
Time to read-read 4096x4096:   22401.662µs    21137.067µs
Time to read-read 8192x8192:   89617.735µs    85637.681µs

Testcase: igt/gem_concurrent_blit (read-read and friends)
Cc: Lionel Landwerlin <lionel.g.landwerlin@linux.intel.com>
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> [v8]
[danvet: s/\<rq\>/req/g]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

The merged seqno->request conversion from John called request
variables req, but some (not all) of Chris' recent patches changed
those to just rq. We've had a lenghty (and inconclusive) discussion on
irc which is the more meaningful name with maybe at most a slight bias
towards req.

Given that the "don't change names without good reason to avoid
conflicts" rule applies, so lets go back to a req everywhere for
consistency. I'll sed any patches for which this will cause conflicts
before applying.

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: John Harrison <John.C.Harrison@Intel.com>
[danvet: s/origina/merged/ as pointed out by Chris - the first
mass-conversion patch was from Chris, the merged one from John.]
Signed-off-by: NDaniel Vetter <daniel.vetter@intel.com>

We no longer interpolate domains in the same manner, and even if we did,
we should trust setting either of the other write domains would trigger
an invalidation rather than force it. Remove the tweaking of the
read_domains since it serves no purpose and use
i915_gem_object_wait_rendering() directly.

Note that this goes back to

commit a8198eea
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Wed Apr 13 22:04:09 2011 +0100

    drm/i915: Introduce i915_gem_object_finish_gpu()

and gpu domain tracking died in

commit cc889e0f
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date:   Wed Jun 13 20:45:19 2012 +0200

    drm/i915: disable flushing_list/gpu_write_list

which is more than 1 year older.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
[danvet: Add notes with information dug out of git history.]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Lost in

commit c5ad54cf
Author: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Date:   Wed May 6 14:36:09 2015 +0300

    drm/i915: Use partial view in mmap fault handler

Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@intel.com>

We do not yet support tiled objects bigger than the mappable
aperture size so reject them.
Reported-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NJoonas Lahtinen <joonas.lahtinen@linux.intel.com>
[danvet: Rework the check a bit to avoid warnings.]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Use partial view for huge BOs (bigger than half the mappable aperture)
in fault handler so that they can be accessed withough trying to make
room for them by evicting other objects.

v2:
- Only use partial views in the case where early rejection was
  previously done.
- Account variable type changes from previous reroll.
v3:
- Add a comment about overwriting existing page entries.
  (Tvrtko Ursulin)
- Whitespace fixes.

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Signed-off-by: NJoonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: NTvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Do not skip special GGTT views when considering whether an object
is pinned or not.

Wrong behaviour was introduced in;

commit ec7adb6e
Author: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Date:   Mon Mar 16 14:11:13 2015 +0200

    drm/i915: Do not use ggtt_view with (aliasing) PPGTT

Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Signed-off-by: NJoonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: NTvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

GGTT VMA sizes might be smaller than the whole object size due to
different GGTT views.

v2:
- Separate GGTT view constraint calculations from normal view
  constraint calculations (Chris Wilson)
v3:
- Do not bother with debug wording. (Tvrtko Ursulin)
v4:
- Clearer logic for calculating map_and_fenceable (Tvrtko Ursulin)

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: NJoonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: NTvrtko Ursulin <tvrtko.ursulin@intel.com>
[danvet: Drop BUG_ON, it's redudant.]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Reading a single value from the object, the locking only provides futile
protection against userspace races. The locking is useless so remove it.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NMika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Unbinding doesn't always lead to unconditional destruction
of vma. This destruction avoidance happens if vma is part of
execbuffer relocation list or if vma is being considered for
eviction in i915_gem_evict_something().

For those other users, mark the vma unbound so that
the correct state of this vma is preserved.
Reported-by: NChris Wilson <chris@chris-wilson.co.ok>
Cc: Chris Wilson <chris@chris-wilson.co.ok>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: NMika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

WaIdleLiteRestore is an execlists-only workaround, and requires the driver
to ensure that any context always has HEAD!=TAIL when attempting lite
restore.

Add two extra MI_NOOP instructions at the end of each request, but keep
the requests tail pointing before the MI_NOOPs. We may not need to
executed them, and this is why request->tail is sampled before adding
these extra instructions.

If we submit a context to the ELSP which has previously been submitted,
move the tail pointer past the MI_NOOPs. This ensures HEAD!=TAIL.

v2: Move overallocation to gen8_emit_request, and added note about
sampling request->tail in commit message (Chris).

v3: Remove redundant request->tail assignment in __i915_add_request, in
lrc mode this is already set in execlists_context_queue.
Do not add wa implementation details inside gem (Chris).

v4: Apply the wa whenever the req has been resubmitted and update
comment (Chris).

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NThomas Daniel <thomas.daniel@intel.com>
Signed-off-by: NMichel Thierry <michel.thierry@intel.com>
Signed-off-by: NJani Nikula <jani.nikula@intel.com>

Currently we have the problem that the decision whether ptes need to
be (re)written is splattered all over the codebase. Move all that into
i915_vma_bind. This needs a few changes:
- Just reuse the PIN_* flags for i915_vma_bind and do the conversion
  to vma->bound in there to avoid duplicating the conversion code all
  over.
- We need to make binding for EXECBUF (i.e. pick aliasing ppgtt if
  around) explicit, add PIN_USER for that.
- Two callers want to update ptes, give them a PIN_UPDATE for that.

Of course we still want to avoid double-binding, but that should be
taken care of:
- A ppgtt vma will only ever see PIN_USER, so no issue with
  double-binding.
- A ggtt vma with aliasing ppgtt needs both types of binding, and we
  track that properly now.
- A ggtt vma without aliasing ppgtt could be bound twice. In the
  lower-level ->bind_vma functions hence unconditionally set
  GLOBAL_BIND when writing the ggtt ptes.

There's still a bit room for cleanup, but that's for follow-up
patches.

v2: Fixup fumbles.

v3: s/PIN_EXECBUF/PIN_USER/ for clearer meaning, suggested by Chris.

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NMika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@intel.com>

It's true that we might need to context switch, but both the signalling
and implementation of the same are a few source files away. Remove it.
Reviewed-by: NMika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

They change with the address space and not with each vma, so move them
into the right pile of vfuncs. Save 2 pointers per vma and clarifies
the code.
Reviewed-by: NMika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@intel.com>

Purpose of this tracking is to know when to flush the cache between
the CPU and the non-coherent display engine. Prior to:

   commit 121920fa
   Author: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
   Date:   Mon Mar 23 11:10:37 2015 +0000

       drm/i915/skl: Query display address through a wrapper

This worked by a mix of direct flag manipulation and checking for
existence of a pinned GGTT VMA.

With the introduction of rotated display mappings this approach is
no longer correct.

New simpler approach is to just keep this count over calls which pin
and unpin objects to and from display, at the slight cost of extra
space in every bo.

(Inspired and extracted code from a larger rework by Chris Wilson.)

v2: Remove the limit since it is not well defined. (Chris Wilson, Ville Syrjälä)
v3: Commit message corrections. (Chris Wilson)
Signed-off-by: NTvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NJoonas Lahtinen <joonas.lahtinen@linux.intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

One month passed between posting a patch and it getting merged, and
unfortunately even though it still applies, it needs fixing to account
for changes in function parameters since:

   commit d385612e15b8b6eb3db328d83f1872ef8a381788
   Author: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
   Date:   Tue Mar 17 14:45:29 2015 +0000

       drm/i915: Log view type when printing warnings
Signed-off-by: NTvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
[danvet: Squash in fixup from Tvrtko to fix the rebase conflict.]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

The obj->pin_mappable flag only exists for debug purposes and is a
hindrance that is mistreated with rotated GGTT views. For debug
purposes, it suffices to mark objects with pin_display as being of note.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NTvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

This provides a nice boost to mesa in swap bound scenarios (as mesa
throttles itself to the previous frame and given the scenario that will
complete shortly). It will also provide a good boost to systems running
with semaphores disabled and so frequently waiting on the GPU as it
switches rings. In the most favourable of microbenchmarks, this can
increase performance by around 15% - though in practice improvements
will be marginal and rarely noticeable.

v2: Account for user timeouts
v3: Limit the spinning to a single jiffie (~1us) at most. On an
otherwise idle system, there is no scheduler contention and so without a
limit we would spin until the GPU is ready.
v4: Drop forcewake - the lazy coherent access doesn't require it, and we
have no reason to believe that the forcewake itself improves seqno
coherency - it only adds delay.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Cc: Eero Tamminen <eero.t.tamminen@intel.com>
Cc: "Rantala, Valtteri" <valtteri.rantala@intel.com>
Reviewed-by: NTvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Move to i915_vma_bind as it is part of the binding.
Suggested-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Michel Thierry <michel.thierry@intel.com>
Signed-off-by: NMika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

We already assign a unique identifier to every request: seqno. That
someone felt like adding a second one without even mentioning why and
tweaking ABI smells very fishy.

Fixes regression from
commit b3a38998
Author: Nick Hoath <nicholas.hoath@intel.com>
Date:   Thu Feb 19 16:30:47 2015 +0000

    drm/i915: Fix a use after free, and unbalanced refcounting

v2: Rebase
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Nick Hoath <nicholas.hoath@intel.com>
Cc: Thomas Daniel <thomas.daniel@intel.com>
Cc: Daniel Vetter <daniel@ffwll.ch>
Cc: Jani Nikula <jani.nikula@intel.com>
[danvet: Fixup because different merge order.]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

vma are more frequently allocated than objects and so should equally
benefit from having a dedicated slab.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

requests are even more frequently allocated than objects and equally
benefit from having a dedicated slab.

v2: Rebase
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

When we submit a request to the GPU, we first take the rpm wakelock, and
only release it once the GPU has been idle for a small period of time
after all requests have been complete. This means that we are sure no
new interrupt can arrive whilst we do not hold the rpm wakelock and so
can drop the individual get/put around every single request inside
execlists.

Note: to close one potential issue we should mark the GPU as busy
earlier in __i915_add_request.

To elaborate: The issue is that we emit the irq signalling sequence
before we grab the rpm reference, which means we could miss the
resulting interrupt (since that's not set up when suspended). The only
bad side effect is a missed interrupt, gt mmio writes automatically
wake up the hw itself. But otoh we have an umbrella rpm reference for
the entirety of execbuf, as long as that's there we're covered.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
[danvet: Explain a bit more about the add_request issue, which after
some irc chatting with Chris turns out to not be an issue really.]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Now with the trimmed memcpy before the command parser, we try to
allocate many different sizes of batches, predominantly one or two
pages. We can therefore speed up searching for a good sized batch by
keeping the objects of buckets of roughly the same size.

v2: Add a comment about bucket sizes
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>