This prevents the case of unbinding the object in order to process the
relocations through the GTT and then rebinding it only to then proceed
to use cpu relocations as the object is now in the CPU write domain. By
choosing to use cpu relocations up front, we can therefore avoid the
rebind penalty.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Avoid stalling and waiting for the GPU by checking to see if there is
sufficient inactive space in the aperture for us to bind the buffer
prior to writing through the GTT. If there is inadequate space we will
have to stall waiting for the GPU, and incur overheads moving objects
about. Instead, only incur the clflush overhead on the target object by
writing through shmem.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
When dealing with a working set larger than the GATT, or even the
mappable aperture when touching through the GTT, we end up with evicting
objects only to rebind them at a new offset again later. Moving an
object into and out of the GTT requires clflushing the pages, thus
causing a double-clflush penalty for rebinding.
To avoid having to clflush on rebinding, we can track the pages as they
are evicted from the GTT and only relinquish those pages on memory
pressure.
As usual, if it were not for the handling of out-of-memory condition and
having to manually shrink our own bo caches, it would be a net reduction
of code. Alas.
Note: The patch also contains a few changes to the last-hope
evict_everything logic in i916_gem_execbuffer.c - we no longer try to
only evict the purgeable stuff in a first try (since that's superflous
and only helps in OOM corner-cases, not fragmented-gtt trashing
situations).
Also, the extraction of the get_pages retry loop from bind_to_gtt (and
other callsites) to get_pages should imo have been a separate patch.
v2: Ditch the newly added put_pages (for unbound objects only) in
i915_gem_reset. A quick irc discussion hasn't revealed any important
reason for this, so if we need this, I'd like to have a git blame'able
explanation for it.
v3: Undo the s/drm_malloc_ab/kmalloc/ in get_pages that Chris noticed.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Split out code movements and rant a bit in the commit message
with a few Notes. Done v2]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Backmerge Linux 3.6-rc2 to resolve a few funny conflicts before we put
even more madness on top:
- drivers/gpu/drm/i915/i915_irq.c: Just a spurious WARN removed in
-fixes, that has been changed in a variable-rename in -next, too.
- drivers/gpu/drm/i915/intel_ringbuffer.c: -next remove scratch_addr
(since all their users have been extracted in another fucntion),
-fixes added another user for a hw workaroudn.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
If a buffer that was the target of a PIPE_CONTROL from userland was a
reused one that hadn't been evicted which had not previously had this
workaround applied, then the early return for a correct
presumed_offset in this function meant we would not bind it into the
GTT and the write would land somewhere else.
Fixes reproducible failures with GL_EXT_timer_query usage in apitrace,
and I also expect it to fix the intermittent OQ issues on snb that
danvet's been working on.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=48019
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=52932
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Carl Worth <cworth@cworth.org>
Tested-by: Carl Worth <cworth@cworth.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
As suggested by Daniel, rip out the independent timers for device and
crtc busyness and integrate the manual powermanagement of the display
engine into the GEM core and its request tracking. The benefits are that
the code is a lot smaller, fewer moving parts and should fit more neatly
into the overall activity tracking of the driver.
v2: Complete overhaul and removal of the racy timers and workers.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
By moving the function to intel_ringbuffer and currying the appropriate
parameter, hopefully we make the callsites easier to read and
understand.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Otherwise once we use the buffer with a BLT command on gen2/3, we will
always regard future command submissions as continuing the fenced
access. However, now that we flush/invalidate between every batch we can
drop this pessimism.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Now that we unconditionally flush and invalidate between every batch
buffer, we no longer need the complex logic to decide which domains
require flushing. Remove it and rejoice.
v2 (danvet): Keep around the flip waiting logic. It's gross and
broken, I know, but we can't just kill that thing ... even if we just
keep it around as a reminder that things are broken.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
This is now handled by a global flag to ensure we emit a flush before
the next serialisation point (if we failed to queue one previously).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
As we always flush the GPU cache prior to emitting the breadcrumb, we no
longer have to worry about the deferred flush causing the
pending_gpu_write to be delayed. So we can instead utilize the known
last_write_seqno to hopefully minimise the wait times.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Request preallocation was added to i915_add_request() in order to
support the overlay. However, not all users care and can quite happily
ignore the failure to allocate the request as they will simply repeat
the request in the future.
By pushing the allocation down into i915_add_request(), we can then
remove some rather ugly error handling in the callers.
v2: Nullify request->file_priv otherwise we chase a garbage pointer
when retiring requests.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Fixes failures in transform feedback on gen7 because our SOL_RESET
flag was setting the transform feedback offsets in the old context
(occasionally happened to be ours) instead of the new context.
Signed-off-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
If we drop the breadcrumb request after a batch due to a signal for
example we aim to fix it up at the next opportunity. In this case we
emit a second batchbuffer with no waits upon the first and so no
opportunity to insert the missing request, so we need to emit the
missing flush for coherency. (Note that that invalidating the render
cache is the same as flushing it, so there should have been no
observable corruption.)
Note that beside simply adding the missing flush, avoiding potential
render corruption, this will also fix at least parts of the problem
introduced by some funny interaction of these two commits:
commit de2b998552
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date: Wed Jul 4 22:52:50 2012 +0200
drm/i915: don't return a spurious -EIO from intel_ring_begin
which allowed intel_ring_begin to return -ERESTARTSYS and
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
which essentially disabled the flushing list.
The issue happens when we submit a batch & emit it, but get
interrupted (thanks to the first patch) while trying to emit the
flush. On the next batch we still assume that the full gpu domain
handling is in effect and hence compute the invalidate&flushing
domains. But thanks to the 2nd patch we totally ignore these and only
invalidate all gpu domains, presuming that any required flushes have
been issued already. Which is wrong and eventually results in us
updating the new write_domain values with the computed
pending_write_domain values, which leaves an object with write_domain
== 0 on the gpu_write_list.
As soon as we try to unbind that object, things blow up.
Fix this by emitting the missing flush according to the new
ring->gpu_caches_dirty flag.
Note that this does _not_ fix all the current cases where we end up
with an object on the flushing_list that can't be flushed.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=52040
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Add bug explanation to commit message.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
This is just the minimal patch to disable all this code so that we can
do decent amounts of QA before we rip it all out.
The complicating thing is that we need to flush the gpu caches after
the batchbuffer is emitted. Which is past the point of no return where
execbuffer can't fail any more (otherwise we risk submitting the same
batch multiple times).
Hence we need to add a flag to track whether any caches associated
with that ring are dirty. And emit the flush in add_request if that's
the case.
Note that this has a quite a few behaviour changes:
- Caches get flushed/invalidated unconditionally.
- Invalidation now happens after potential inter-ring sync.
I've bantered around a bit with Chris on irc whether this fixes
anything, and it might or might not. The only thing clear is that with
these changes it's much easier to reason about correctness.
Also rip out a lone get_next_request_seqno in the execbuffer
retire_commands function. I've dug around and I couldn't figure out
why that is still there, with the outstanding lazy request stuff it
shouldn't be necessary.
v2: Chris Wilson complained that I also invalidate the read caches
when flushing after a batchbuffer. Now optimized.
v3: Added some comments to explain the new flushing behaviour.
Cc: Eric Anholt <eric@anholt.net>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Use the rsvd1 field in execbuf2 to specify the context ID associated
with the workload. This will allow the driver to do the proper context
switch when/if needed.
v2: Add checks for context switches on rings not supporting contexts.
Before the code would silently ignore such requests.
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Rather than use the magic feature tests HAS_BLT/HAS_BSD just check
whether the ring we are about to dispatch the execbuffer on is
initialised.
v2: Use intel_ring_initialized()
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The principle of intel_mark_busy() is that we want to spot the
transition of when the display engine is being used in order to bump
powersaving modes and increase display clocks. As such it is only
important when the display is changing, i.e. when rendering to the
scanout or other sprite/plane, and these are characterised by being
pinned.
v2: Mark the whole device as busy on execbuffer and pageflips as well
and rebase against dinq for the minor bug fix to be immediately
applicable.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: fix compile fail.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Backmerge of drm-next to resolve a few ugly conflicts and to get a few
fixes from 3.4-rc6 (which drm-next has already merged). Note that this
merge also restricts the stencil cache lra evict policy workaround to
snb (as it should) - I had to frob the code anyway because the
CM0_MASK_SHIFT define died in the masked bit cleanups.
We need the backmerge to get Paulo Zanoni's infoframe regression fix
for gm45 - further bugfixes from him touch the same area and would
needlessly conflict.
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Conflicts:
drivers/gpu/drm/i915/intel_display.c
Ok, this is a fun story of git totally messing things up. There
/shouldn't/ be any conflict in here, because the fixes in -rc6 do only
touch functions that have not been changed in -next.
The offending commits in drm-next are 14415745b2..1fa611065 which
simply move a few functions from intel_display.c to intel_pm.c. The
problem seems to be that git diff gets completely confused:
$ git diff 14415745b2..1fa611065
is a nice mess in intel_display.c, and the diff leaks into totally
unrelated functions, whereas
$git diff --minimal 14415745b2..1fa611065
is exactly what we want.
Unfortunately there seems to be no way to teach similar smarts to the
merge diff and conflict generation code, because with the minimal diff
there really shouldn't be any conflicts. For added hilarity, every
time something in that area changes the + and - lines in the diff move
around like crazy, again resulting in new conflicts. So I fear this
mess will stay with us for a little longer (and might result in
another backmerge down the road).
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Unfortunately there has been dri1 userspace that used gem to manage
the gtt and hence also needed cliprects in the execbuf ioctl. So
we can't ever remove that code without breaking the ioctl abi.
But at least we can disable it on gen5+, because these horrible
versions of mesa have not supported these chips.
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
This originates from a hack by me to quickly fix a bug in an earlier
patch where we needed control over whether or not waiting on a seqno
actually did any retire list processing. Since the two operations aren't
clearly related, we should pull the parameter out of the wait function,
and make the caller responsible for retiring if the action is desired.
The only function call site which did not get an explicit retire_request call
(on purpose) is i915_gem_inactive_shrink(). That code was already calling
retire_request a second time.
v2: don't modify any behavior excepit i915_gem_inactive_shrink(Daniel)
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
On 32-bit systems, a large args->num_cliprects from userspace via ioctl
may overflow the allocation size, leading to out-of-bounds access.
This vulnerability was introduced in commit 432e58ed ("drm/i915: Avoid
allocation for execbuffer object list").
Signed-off-by: Xi Wang <xi.wang@gmail.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: stable@vger.kernel.org
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
On 32-bit systems, a large args->buffer_count from userspace via ioctl
may overflow the allocation size, leading to out-of-bounds access.
This vulnerability was introduced in commit 8408c282 ("drm/i915:
First try a normal large kmalloc for the temporary exec buffers").
Signed-off-by: Xi Wang <xi.wang@gmail.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: stable@vger.kernel.org
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
We never succeeded in getting pipelined fencing to work (unresolved
spurious GPU hangs), so begin the process of dismantling and removal
the broken code.
Step 1 is the removal of the pipeline parameter to get_fence().
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Chris Wilson
Daniel Vetter
Eric Anholt
Ben Widawsky
Xi Wang