Whilst the contents of the context is still protected by the big
struct_mutex, this is not much of an improvement. It is just one tiny
step towards reducing our BKL.
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/20170620110547.15947-3-chris@chris-wilson.co.uk
During execbuf, a mandatory step is that we add this request (this
fence) to each object's reservation_object. Inside execbuf, we track the
vma, and to add the fence to the reservation_object then means having to
first chase the obj, incurring another cache miss. We can reduce the
number of cache misses by stashing a pointer to the reservation_object
in the vma itself.
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/20170616140525.6394-1-chris@chris-wilson.co.uk
If the user requires patching of their batch or auxiliary buffers, we
currently make the alterations on the cpu. If they are active on the GPU
at the time, we wait under the struct_mutex for them to finish executing
before we rewrite the contents. This happens if shared relocation trees
are used between different contexts with separate address space (and the
buffers then have different addresses in each), the 3D state will need
to be adjusted between execution on each context. However, we don't need
to use the CPU to do the relocation patching, as we could queue commands
to the GPU to perform it and use fences to serialise the operation with
the current activity and future - so the operation on the GPU appears
just as atomic as performing it immediately. Performing the relocation
rewrites on the GPU is not free, in terms of pure throughput, the number
of relocations/s is about halved - but more importantly so is the time
under the struct_mutex.
v2: Break out the request/batch allocation for clearer error flow.
v3: A few asserts to ensure rq ordering is maintained
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Currently, the last object in the execlist is the always the batch.
However, when building the batch buffer we often know the batch object
first and if we can use the first slot in the execlist we can emit
relocation instructions relative to it immediately and avoid a separate
pass to adjust the relocations to point to the last execlist slot.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
This simply hides the EAGAIN caused by userptr when userspace causes
resource contention. However, it is quite beneficial with highly
contended userptr users as we avoid repeating the setup costs and
kernel-user context switches.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Michał Winiarski <michal.winiarski@intel.com>
When choosing a slot for an execbuffer, we ideally want to use the same
address as last time (so that we don't have to rebind it) and the same
address as expected by the user (so that we don't have to fixup any
relocations pointing to it). If we first try to bind the incoming
execbuffer->offset from the user, or the currently bound offset that
should hopefully achieve the goal of avoiding the rebind cost and the
relocation penalty. However, if the object is not currently bound there
we don't want to arbitrarily unbind an object in our chosen position and
so choose to rebind/relocate the incoming object instead. After we
report the new position back to the user, on the next pass the
relocations should have settled down.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtien@linux.intel.com>
If we take a reference to the object/vma when it is first used in an
execbuf, we can keep that reference until the object's file-local handle
is closed. Thereby saving a frequent ref/unref pair.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
The major scaling bottleneck in execbuffer is the processing of the
execobjects. Creating an auxiliary list is inefficient when compared to
using the execobject array we already have allocated.
Reservation is then split into phases. As we lookup up the VMA, we
try and bind it back into active location. Only if that fails, do we add
it to the unbound list for phase 2. In phase 2, we try and add all those
objects that could not fit into their previous location, with fallback
to retrying all objects and evicting the VM in case of severe
fragmentation. (This is the same as before, except that phase 1 is now
done inline with looking up the VMA to avoid an iteration over the
execobject array. In the ideal case, we eliminate the separate reservation
phase). During the reservation phase, we only evict from the VM between
passes (rather than currently as we try to fit every new VMA). In
testing with Unreal Engine's Atlantis demo which stresses the eviction
logic on gen7 class hardware, this speed up the framerate by a factor of
2.
The second loop amalgamation is between move_to_gpu and move_to_active.
As we always submit the request, even if incomplete, we can use the
current request to track active VMA as we perform the flushes and
synchronisation required.
The next big advancement is to avoid copying back to the user any
execobjects and relocations that are not changed.
v2: Add a Theory of Operation spiel.
v3: Fall back to slow relocations in preparation for flushing userptrs.
v4: Document struct members, factor out eb_validate_vma(), add a few
more comments to explain some magic and hide other magic behind macros.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
If we write a relocation into the buffer, we require our own implicit
synchronisation added after the start of the execbuf, outside of the
user's control. As we may end up clflushing, or doing the patch itself
on the GPU, asynchronously we need to look at the implicit serialisation
on obj->resv and hence need to disable EXEC_OBJECT_ASYNC for this
object.
If the user does trigger a stall for relocations, we make sure the stall
is complete enough so that the batch is not submitted before we complete
those relocations.
Fixes: 77ae995789 ("drm/i915: Enable userspace to opt-out of implicit fencing")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
We can simplify our tracking of pending writes in an execbuf to the
single bit in the vma->exec_entry->flags, but that requires the
relocation function knowing the object's vma. Pass it along.
Note we have only been using a single bit to track flushing since
commit cc889e0f6c
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date: Wed Jun 13 20:45:19 2012 +0200
drm/i915: disable flushing_list/gpu_write_list
unconditionally flushed all render caches before the breadcrumb and
commit 6ac42f4148
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date: Sat Jul 21 12:25:01 2012 +0200
drm/i915: Replace the complex flushing logic with simple invalidate/flush all
did away with the explicit GPU domain tracking. This was then codified
into the ABI with NO_RELOC in
commit ed5982e6ce
Author: Daniel Vetter <daniel.vetter@ffwll.ch> # Oi! Patch stealer!
Date: Thu Jan 17 22:23:36 2013 +0100
drm/i915: Allow userspace to hint that the relocations were known
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
The advent of full-ppgtt lead to an extra indirection between the object
and its binding. That extra indirection has a noticeable impact on how
fast we can convert from the user handles to our internal vma for
execbuffer. In order to bypass the extra indirection, we use a
resizable hashtable to jump from the object to the per-ctx vma.
rhashtable was considered but we don't need the online resizing feature
and the extra complexity proved to undermine its usefulness. Instead, we
simply reallocate the hastable on demand in a background task and
serialize it before iterating.
In non-full-ppgtt modes, multiple files and multiple contexts can share
the same vma. This leads to having multiple possible handle->vma links,
so we only use the first to establish the fast path. The majority of
buffers are not shared and so we should still be able to realise
speedups with multiple clients.
v2: Prettier names, more magic.
v3: Many style tweaks, most notably hiding the misuse of execobj[].rsvd2
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
For ease of use (i.e. avoiding a few checks and function calls), store
the object's cache coherency next to the cache is dirty bit.
Specifically this patch aims to reduce the frequency of no-op calls to
i915_gem_object_clflush() to counter-act the increase of such calls for
GPU only objects in the previous patch.
v2: Replace cache_dirty & ~cache_coherent with cache_dirty &&
!cache_coherent as gcc generates much better code for the latter
(Tvrtko)
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Dongwon Kim <dongwon.kim@intel.com>
Cc: Matt Roper <matthew.d.roper@intel.com>
Tested-by: Dongwon Kim <dongwon.kim@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170616105455.16977-1-chris@chris-wilson.co.uk
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Currently, we only mark the CPU cache as dirty if we skip a clflush.
This leads to some confusion where we have to ask if the object is in
the write domain or missed a clflush. If we always mark the cache as
dirty, this becomes a much simply question to answer.
The goal remains to do as few clflushes as required and to do them as
late as possible, in the hope of deferring the work to a kthread and not
block the caller (e.g. execbuf, flips).
v2: Always call clflush before GPU execution when the cache_dirty flag
is set. This may cause some extra work on llc systems that migrate dirty
buffers back and forth - but we do try to limit that by only setting
cache_dirty at the end of the gpu sequence.
v3: Always mark the cache as dirty upon a level change, as we need to
invalidate any stale cachelines due to external writes.
Reported-by: Dongwon Kim <dongwon.kim@intel.com>
Fixes: a6a7cc4b7d ("drm/i915: Always flush the dirty CPU cache when pinning the scanout")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Dongwon Kim <dongwon.kim@intel.com>
Cc: Matt Roper <matthew.d.roper@intel.com>
Tested-by: Dongwon Kim <dongwon.kim@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170615123850.26843-1-chris@chris-wilson.co.uk
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Currently the vma has one link member that is used for both holding its
place in the execbuf reservation list, and in any eviction list. This
dual property is quite tricky and error prone.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170615081435.17699-3-chris@chris-wilson.co.uk
This has the benefit of not requiring us to manipulate the
vma->exec_link list when tearing down the execbuffer, and is a
marginally cheaper test to detect the user error.
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/20170615081435.17699-2-chris@chris-wilson.co.uk
More stuff for 4.13:
- skl+ wm fixes from Mahesh Kumar
- some refactor and tests for i915_sw_fence (Chris)
- tune execlist/scheduler code (Chris)
- g4x,g33 gpu reset improvements (Chris, Mika)
- guc code cleanup (Michal Wajdeczko, Michał Winiarski)
- dp aux backlight improvements (Puthikorn Voravootivat)
- buffer based guc/host communication (Michal Wajdeczko)
* tag 'drm-intel-next-2017-05-29' of git://anongit.freedesktop.org/git/drm-intel: (253 commits)
drm/i915: Update DRIVER_DATE to 20170529
drm/i915: Keep the forcewake timer alive for 1ms past the most recent use
drm/i915/guc: capture GuC logs if FW fails to load
drm/i915/guc: Introduce buffer based cmd transport
drm/i915/guc: Disable send function on fini
drm: Add definition for eDP backlight frequency
drm/i915: Drop AUX backlight enable check for backlight control
drm/i915: Consolidate #ifdef CONFIG_INTEL_IOMMU
drm/i915: Only GGTT vma may be pinned and prevent shrinking
drm/i915: Serialize GTT/Aperture accesses on BXT
drm/i915: Convert i915_gem_object_ops->flags values to use BIT()
drm/i915/selftests: Silence compiler warning in igt_ctx_exec
drm/i915/guc: Skip port assign on first iteration of GuC dequeue
drm/i915: Remove misleading comment in request_alloc
drm/i915/g33: Improve reset reliability
Revert "drm/i915: Restore lost "Initialized i915" welcome message"
drm/i915/huc: Update GLK HuC version
drm/i915: Check for allocation failure
drm/i915/guc: Remove action status and statistics from debugfs
drm/i915/g4x: Improve gpu reset reliability
...
Now that drm_[cm]alloc* helpers are simple one line wrappers around
kvmalloc_array and drm_free_large is just kvfree alias we can drop
them and replace by their native forms.
This shouldn't introduce any functional change.
Changes since v1
- fix typo in drivers/gpu//drm/etnaviv/etnaviv_gem.c - noticed by 0day
build robot
Suggested-by: Daniel Vetter <daniel@ffwll.ch>
Signed-off-by: Michal Hocko <mhocko@suse.com>drm: drop drm_[cm]alloc* helpers
[danvet: Fixup vgem which grew another user very recently.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Christian König <christian.koenig@amd.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170517122312.GK18247@dhcp22.suse.cz
Introduce a new execobject.flag (EXEC_OBJECT_CAPTURE) that userspace may
use to indicate that it wants the contents of this buffer preserved in
the error state (/sys/class/drm/cardN/error) following a GPU hang
involving this batch.
Use this at your discretion, the contents of the error state. although
compressed, are allocated with GFP_ATOMIC (i.e. limited) and kept for all
eternity (until the error state is destroyed).
Based on an earlier patch by Ben Widawsky <ben@bwidawsk.net>
Testcase: igt/gem_exec_capture
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ben Widawsky <ben@bwidawsk.net>
Cc: Matt Turner <mattst88@gmail.com>
Acked-by: Ben Widawsky <ben@bwidawsk.net>
Acked-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170415093902.22581-1-chris@chris-wilson.co.uk
Old devices have quite severe restrictions for using fences, and unlike
more recent device (anything from Pineview onwards) we need to enforce
those restrictions even for unfenced tiled access from the render
pipeline.
Fixes: 944397f04f ("drm/i915: Store required fence size/alignment for GGTT vma")
Reported-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: <drm-intel-fixes@lists.freedesktop.org> # v4.11-rc1+
Link: http://patchwork.freedesktop.org/patch/msgid/20170325113243.16438-1-chris@chris-wilson.co.uk
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Tested-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Adding to the tail of the client request list as the only other user is
in the throttle ioctl that iterates forwards over the list. It only
needs protection against deletion of a request as it reads it, it simply
won't see a new request added to the end of the list, or it would be too
early and rejected. We can further reduce the number of spinlocks
required when throttling by removing stale requests from the client_list
as we throttle.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170302122525.19675-1-chris@chris-wilson.co.uk
This patch makes the I915_PARAM_HAS_EXEC_CONSTANTS getparam return 0
(indicating the optional feature is not supported), and makes execbuf
always return -EINVAL if the flags are used.
Apparently, no userspace ever shipped which used this optional feature:
I checked the git history of Mesa, xf86-video-intel, libva, and Beignet,
and there were zero commits showing a use of these flags. Kernel commit
72bfa19c8d apparently introduced the feature prematurely. According
to Chris, the intention was to use this in cairo-drm, but "the use was
broken for gen6", so I don't think it ever happened.
'relative_constants_mode' has always been tracked per-device, but this
has actually been wrong ever since hardware contexts were introduced, as
the INSTPM register is saved (and automatically restored) as part of the
render ring context. The software per-device value could therefore get
out of sync with the hardware per-context value. This meant that using
them is actually unsafe: a client which tried to use them could damage
the state of other clients, causing the GPU to interpret their BO
offsets as absolute pointers, leading to bogus memory reads.
These flags were also never ported to execlist mode, making them no-ops
on Gen9+ (which requires execlists), and Gen8 in the default mode.
On Gen8+, userspace can write these registers directly, achieving the
same effect. On Gen6-7.5, it likely makes sense to extend the command
parser to support them. I don't think anyone wants this on Gen4-5.
Based on a patch by Dave Gordon.
v3: Return -ENODEV for the getparam, as this is what we do for other
obsolete features. Suggested by Chris Wilson.
Cc: stable@vger.kernel.org
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92448
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/20170215093446.21291-1-kenneth@whitecape.org
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Flushing the cachelines for an object is slow, can be as much as 100ms
for a large framebuffer. We currently do this under the struct_mutex BKL
on execution or on pageflip. But now with the ability to add fences to
obj->resv for both flips and execbuf (and we naturally wait on the fence
before CPU access), we can move the clflush operation to a workqueue and
signal a fence for completion, thereby doing the work asynchronously and
not blocking the driver or its clients.
v2: Introduce i915_gem_clflush.h and use a new name, split out some
extras into separate patches.
Suggested-by: Akash Goel <akash.goel@intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170222114049.28456-5-chris@chris-wilson.co.uk
The change_domain tracepoint has been inaccurate for a few years - it
doesn't fully capture the domains, especially with userspace bypassing
them. It is defunct, misleading and time to be removed.
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/20170222114049.28456-1-chris@chris-wilson.co.uk
Rename it to i915_gem_request_queue and fix the logged info
equivalent to the i915_gem_request even class. Also moved it
a bit further apart from the i915_gem_request_add tracepoint
since they otherwise provide similar information too close in
time.
v2: Remove sw fence singalling. We will rely on the soon to
come GuC scheduling backend to enable that. (Chris Wilson)
v3: Log hex with 0x prefix for clarity.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
This removes the usage of intel_ring_emit in favour of
directly writing to the ring buffer.
intel_ring_emit was preventing the compiler for optimising
fetch and increment of the current ring buffer pointer and
therefore generating very verbose code for every write.
It had no useful purpose since all ringbuffer operations
are started and ended with intel_ring_begin and
intel_ring_advance respectively, with no bail out in the
middle possible, so it is fine to increment the tail in
intel_ring_begin and let the code manage the pointer
itself.
Useless instruction removal amounts to approximately
two and half kilobytes of saved text on my build.
Not sure if this has any measurable performance
implications but executing a ton of useless instructions
on fast paths cannot be good.
v2:
* Change return from intel_ring_begin to error pointer by
popular demand.
* Move tail increment to intel_ring_advance to enable some
error checking.
v3:
* Move tail advance back into intel_ring_begin.
* Rebase and tidy.
v4:
* Complete rebase after a few months since v3.
v5:
* Remove unecessary cast and fix !debug compile. (Chris Wilson)
v6:
* Make intel_ring_offset take request as well.
* Fix recording of request postfix plus a sprinkle of asserts.
(Chris Wilson)
v7:
* Use intel_ring_offset to get the postfix. (Chris Wilson)
* Convert GVT code as well.
v8:
* Rename *out++ to *cs++.
v9:
* Fix GVT out to cs conversion in GVT.
v10:
* Rebase for new intel_ring_begin in selftests.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Zhi Wang <zhi.a.wang@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170214113242.29241-1-tvrtko.ursulin@linux.intel.com
Chris Wilson needs the new drm_driver->release callback to make sure
the shiny new dma-buf testcases don't oops the driver on unload.
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
We're using non-canonical addresses in drm_mm, and we're making sure that
userspace is using canonical addressing - both in case of softpin
(verifying incoming offset) and when relocating (converting to canonical
when updating offset returned to userspace).
Unfortunately when considering the need for relocations, we're comparing
offset from userspace (in canonical form) with drm_mm node (in
non-canonical form), and as a result, we end up always relocating if our
offsets are in the "problematic" range.
Let's always convert the offsets to avoid the performance impact of
relocations.
Fixes: a5f0edf63b ("drm/i915: Avoid writing relocs with addresses in non-canonical form")
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Michel Thierry <michel.thierry@intel.com>
Reported-by: Michał Pyrzowski <michal.pyrzowski@intel.com>
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170207195559.18798-1-michal.winiarski@intel.com
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Fences are creted/checked before the pm ref is taken, so if we jump to
pre_mutex_err we will uncorrectly call intel_runtime_pm_put.
v2: Massage unwind error paths
Fixes: fec0445caa (drm/i915: Support explicit fencing for execbuf)
Testcase: igt/gem_exec_params
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1486161930-11764-1-git-send-email-daniele.ceraolospurio@intel.com
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
The drm_mm range manager claimed to support top-down insertion, but it
was neither searching for the top-most hole that could fit the
allocation request nor fitting the request to the hole correctly.
In order to search the range efficiently, we create a secondary index
for the holes using either their size or their address. This index
allows us to find the smallest hole or the hole at the bottom or top of
the range efficiently, whilst keeping the hole stack to rapidly service
evictions.
v2: Search for holes both high and low. Rename flags to mode.
v3: Discover rb_entry_safe() and use it!
v4: Kerneldoc for enum drm_mm_insert_mode.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: "Christian König" <christian.koenig@amd.com>
Cc: David Airlie <airlied@linux.ie>
Cc: Russell King <rmk+kernel@armlinux.org.uk>
Cc: Daniel Vetter <daniel.vetter@intel.com>
Cc: Jani Nikula <jani.nikula@linux.intel.com>
Cc: Sean Paul <seanpaul@chromium.org>
Cc: Lucas Stach <l.stach@pengutronix.de>
Cc: Christian Gmeiner <christian.gmeiner@gmail.com>
Cc: Rob Clark <robdclark@gmail.com>
Cc: Thierry Reding <thierry.reding@gmail.com>
Cc: Stephen Warren <swarren@wwwdotorg.org>
Cc: Alexandre Courbot <gnurou@gmail.com>
Cc: Eric Anholt <eric@anholt.net>
Cc: Sinclair Yeh <syeh@vmware.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com> # vmwgfx
Reviewed-by: Lucas Stach <l.stach@pengutronix.de> #etnaviv
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/20170202210438.28702-1-chris@chris-wilson.co.uk
Now that the user can opt-out of implicit fencing, we need to give them
back control over the fencing. We employ sync_file to wrap our
drm_i915_gem_request and provide an fd that userspace can merge with
other sync_file fds and pass back to the kernel to wait upon before
future execution.
Testcase: igt/gem_exec_fence
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Acked-by: Chad Versace <chadversary@chromium.org>
Link: http://patchwork.freedesktop.org/patch/msgid/20170127094008.27489-2-chris@chris-wilson.co.uk
Userspace is faced with a dilemma. The kernel requires implicit fencing
to manage resource usage (we always must wait for the GPU to finish
before releasing its PTE) and for third parties. However, userspace may
wish to avoid this serialisation if it is either using explicit fencing
between parties and wants more fine-grained access to buffers (e.g. it
may partition the buffer between uses and track fences on ranges rather
than the implicit fences tracking the whole object). It follows that
userspace needs a mechanism to avoid the kernel's serialisation on its
implicit fences before execbuf execution.
The next question is whether this is an object, execbuf or context flag.
Hybrid users (such as using explicit EGL_ANDROID_native_sync fencing on
shared winsys buffers, but implicit fencing on internal surfaces)
require a per-object level flag. Given that this flag need to be only
set once for the lifetime of the object, this reduces the convenience of
having an execbuf or context level flag (and avoids having multiple
pieces of uABI controlling the same feature).
Incorrect use of this flag will result in rendering corruption and GPU
hangs - but will not result in use-after-free or similar resource
tracking issues.
Serious caveat: write ordering is not strictly correct after setting
this flag on a render target on multiple engines. This affects all
subsequent GEM operations (execbuf, set-domain, pread) and shared
dma-buf operations. A fix is possible - but costly (both in terms of
further ABI changes and runtime overhead).
Testcase: igt/gem_exec_async
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Acked-by: Chad Versace <chadversary@chromium.org>
Link: http://patchwork.freedesktop.org/patch/msgid/20170127094008.27489-1-chris@chris-wilson.co.uk
Whilst writing testcases to exercise the VMA API, some oddities came to
light, such as i915_gem_obj_lookup_or_create(). Joonas suggested
i915_vma_instance() as a neat replacement, so rename them, move them to
i915_vma.c and add some kerneldoc as a sugary bonus.
s/i915_gem_obj_to_vma/i915_vma_lookup/
s/i915_gem_obj_lookup_or_create_vma/i915_vma_instance/
Suggested-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170116152131.18089-2-chris@chris-wilson.co.uk
Start converting over from the byte count to its semantic macro, either
we want to allocate the size of a physical page in main memory or we
want the size of a virtual page in the GTT. 4096 could mean either, but
PAGE_SIZE and I915_GTT_PAGE_SIZE are explicit and should help improve
code comprehension and future changes. In the future, we may want to use
variable GTT page sizes and so have the challenge of knowing which
hardcoded values were used to represent a physical page vs the virtual
page.
v2: Look for a few more 4096s to convert, discover IS_ALIGNED().
v3: 4096ul paranoia, make fence alignment a distinct value of 4096, keep
bdw stolen w/a as 4096 until we know better.
v4: Add asserts that i915_vma_insert() start/end are aligned to GTT page
sizes.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/20170110144734.26052-1-chris@chris-wilson.co.uk
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Soft-pinning depends upon being able to check for availabilty of an
interval and evict overlapping object from a drm_mm range manager very
quickly. Currently it uses a linear list, and so performance is dire and
not suitable as a general replacement. Worse, the current code will oops
if it tries to evict an active buffer.
It also helps if the routine reports the correct error codes as expected
by its callers and emits a tracepoint upon use.
For posterity since the wrong patch was pushed (i.e. that missed these
key points and had known bugs), this is the changelog that should have
been on commit 506a8e87d8 ("drm/i915: Add soft-pinning API for
execbuffer"):
Userspace can pass in an offset that it presumes the object is located
at. The kernel will then do its utmost to fit the object into that
location. The assumption is that userspace is handling its own object
locations (for example along with full-ppgtt) and that the kernel will
rarely have to make space for the user's requests.
This extends the DRM_IOCTL_I915_GEM_EXECBUFFER2 to do the following:
* if the user supplies a virtual address via the execobject->offset
*and* sets the EXEC_OBJECT_PINNED flag in execobject->flags, then
that object is placed at that offset in the address space selected
by the context specifier in execbuffer.
* the location must be aligned to the GTT page size, 4096 bytes
* as the object is placed exactly as specified, it may be used by this
execbuffer call without relocations pointing to it
It may fail to do so if:
* EINVAL is returned if the object does not have a 4096 byte aligned
address
* the object conflicts with another pinned object (either pinned by
hardware in that address space, e.g. scanouts in the aliasing ppgtt)
or within the same batch.
EBUSY is returned if the location is pinned by hardware
EINVAL is returned if the location is already in use by the batch
* EINVAL is returned if the object conflicts with its own alignment (as meets
the hardware requirements) or if the placement of the object does not fit
within the address space
All other execbuffer errors apply.
Presence of this execbuf extension may be queried by passing
I915_PARAM_HAS_EXEC_SOFTPIN to DRM_IOCTL_I915_GETPARAM and checking for
a reported value of 1 (or greater).
v2: Combine the hole/adjusted-hole ENOSPC checks
v3: More color, more splitting, more blurb.
Fixes: 506a8e87d8 ("drm/i915: Add soft-pinning API for execbuffer")
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/20161205142941.21965-2-chris@chris-wilson.co.uk
We need to distinguish between full i915_vma structs and simple
drm_mm_nodes when considering eviction (i.e. we must be careful not to
treat a mere drm_mm_node as a much larger i915_vma causing memory
corruption, if we are lucky). To do this, color these not-a-vma with -1
(I915_COLOR_UNEVICTABLE).
v2...v200: New name for -1.
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/20161205142941.21965-1-chris@chris-wilson.co.uk
As i915.enable_cmd_parser is an unsafe option, make it read-only at
runtime. Now that it is constant, we can use the value determined during
initialisation as to whether we need the cmdparser at execbuffer time.
v2: Remove the inline for its single user, it is clear enough (and
shorter) without!
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/20161124125851.6615-1-chris@chris-wilson.co.uk
I tried to avoid having to track the write for every VMA by only
tracking writes to the ggtt. However, for the purposes of frontbuffer
tracking this is insufficient as we need to invalidate around writes not
just to the the ggtt but all aliased ppgtt views of the framebuffer. By
moving the critical section to the object and only doing so for
framebuffer writes we can reduce the tracking even further by only
watching framebuffers and not vma.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Paulo Zanoni <paulo.r.zanoni@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161116190704.5293-1-chris@chris-wilson.co.uk
Tested-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
A small selection of macros which can only accept dev_priv from
now on and a resulting trickle of fixups.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: David Weinehall <david.weinehall@linux.intel.com>
A small selection of macros which can only accept dev_priv from
now on and a resulting trickle of fixups.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: David Weinehall <david.weinehall@linux.intel.com>
On LLC, or even snooped, machines rendering via the GPU ends up in the CPU
cache. This cacheline dirt also needs to be flushed to main memory when
moving to an incoherent domain, such as the display's scanout engine.
Mostly, this happens because either the object is marked as dirty from
its first use or is avoided by setting the object into the display
domain from the start.
v2: Treat WT as not requiring a clflush prior to use on the display
engine as well.
Fixes: 0f71979ab7 ("drm/i915: Performed deferred clflush inside set-cache-level")
References: https://bugs.freedesktop.org/show_bug.cgi?id=95414
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Jani Nikula <jani.nikula@linux.intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: <stable@vger.kernel.org> # v4.0+
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161107165204.7008-1-chris@chris-wilson.co.uk
Move has_64bit_reloc into dev_priv->info. This will make it visible
in the feature listing debug output.
v2:
- Keep the struct member to keep GCC fragile but happy (Chris)
v3:
- More detailed commit message (Chris)
- Include forgotten CHV and BXT (Chris)
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/1478162386-5018-1-git-send-email-joonas.lahtinen@linux.intel.com
In preparation to support many distinct timelines, we need to expand the
activity tracking on the GEM object to handle more than just a request
per engine. We already use the struct reservation_object on the dma-buf
to handle many fence contexts, so integrating that into the GEM object
itself is the preferred solution. (For example, we can now share the same
reservation_object between every consumer/producer using this buffer and
skip the manual import/export via dma-buf.)
v2: Reimplement busy-ioctl (by walking the reservation object), postpone
the ABI change for another day. Similarly use the reservation object to
find the last_write request (if active and from i915) for choosing
display CS flips.
Caveats:
* busy-ioctl: busy-ioctl only reports on the native fences, it will not
warn of stalls (in set-domain-ioctl, pread/pwrite etc) if the object is
being rendered to by external fences. It also will not report the same
busy state as wait-ioctl (or polling on the dma-buf) in the same
circumstances. On the plus side, it does retain reporting of which
*i915* engines are engaged with this object.
* non-blocking atomic modesets take a step backwards as the wait for
render completion blocks the ioctl. This is fixed in a subsequent
patch to use a fence instead for awaiting on the rendering, see
"drm/i915: Restore nonblocking awaits for modesetting"
* dynamic array manipulation for shared-fences in reservation is slower
than the previous lockless static assignment (e.g. gem_exec_lut_handle
runtime on ivb goes from 42s to 66s), mainly due to atomic operations
(maintaining the fence refcounts).
* loss of object-level retirement callbacks, emulated by VMA retirement
tracking.
* minor loss of object-level last activity information from debugfs,
could be replaced with per-vma information if desired
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/20161028125858.23563-21-chris@chris-wilson.co.uk
The plan is to make obtaining the backing storage for the object avoid
struct_mutex (i.e. use its own locking). The first step is to update the
API so that normal users only call pin/unpin whilst working on the
backing storage.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161028125858.23563-12-chris@chris-wilson.co.uk
We only need the active reference to keep the object alive after the
handle has been deleted (so as to prevent a synchronous gem_close). Why
then pay the price of a kref on every execbuf when we can insert that
final active ref just in time for the handle deletion?
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/20161028125858.23563-6-chris@chris-wilson.co.uk
