Even if the object is no longer in the GTT domain, there may still be
a user space mapping which needs to be released.
Without this fix, render-based text (mostly in firefox) would
occasionally get corrupted when the system was under load.
Signed-off-by: Keith Packard <keithp@keithp.com>
This reverts commit 4a684a4117.
Userland has always been required to set the object's domain to GTT
before using it through a GTT mapping, it's not something that the
kernel is supposed to enforce. (The pagefault support is so that we
can handle multiple mappings without userland having to pin across
them, not so that userland can use GTT after GPU domains without
telling the kernel).
Fixes 19.2% +/- 0.8% (n=6) performance regression in cairo-gl
firefox-talos-gfx on my T420 latop.
Signed-off-by: Keith Packard <keithp@keithp.com>
It seems to me that we are leaking 'user_pages' in
drivers/gpu/drm/i915/i915_gem.c::i915_gem_shmem_pread_slow() if
read_cache_page_gfp() fails.
Signed-off-by: Jesper Juhl <jj@chaosbits.net>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Improves full-screen openarena on my laptop 20.3% +/- 4.0% (n=3)
Improves 800x600 nexuiz on my laptop 12.3% +/- 0.1% (n=3)
We have more room to improve with doing LLC caching for display using
GFDT, and in doing LLC+MLC caching, but this was an easy performance
win and incremental improvement toward those two.
Signed-off-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The simplest and common method for ensuring scanout coherency on all
chipsets is to mark the scanout buffers as uncached (and for
userspace to remember to flush the render cache every so often).
We can improve upon this for later generations by marking scanout
objects as GFDT and only flush those cachelines when required. However,
we start simple.
[v2: Move the set to uncached above the clflush. Otherwise, we'd skip
the clflush and try to scan out data that was still sitting in the
cache.]
Signed-off-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
We need to perform a few operations in order to move the object into the
display plane (where it can be accessed coherently by the display
engine) that are important for future safety to forbid whilst pinned. As a
result, we want to need to perform some of the operations before pinning,
but some are required once we have been bound into the GTT. So combine
the pinning performed by all the callers with set_to_display_plane(), so
this complication is contained within the single function.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
[anholt v2: Don't forget that when going from cached to uncached, we
haven't been tracking the write domain from the CPU perspective, since
we haven't needed it for GPU coherency.]
[ickle v3: We also need to make sure we relinquish any fences on older
chipsets and clear the GTT for sane domain tracking.]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Like its siblings finish_gpu(), this function clears the object from the
GTT domain forcing it to be trigger a domain invalidation should we ever
need to use via the GTT again.
Note that the most important side-effect of finishing the GTT domain
(aside from clearing the tracking read/write domains) is that it imposes
an memory barrier so that all accesses are complete before it returns,
which is important if you intend to be modifying translation tables
shortly afterwards. The second most important side-effect is that it
tears down the GTT mappings forcing a page-fault and invalidation on
next user access to the object.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
... reincarnated from i915_gem_object_flush_gpu(). The semantic
difference is that after calling finish_gpu() the object no longer
resides in any GPU domain, and so will cause the GPU caches to be
invalidated if it is ever used again.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
A tile on gen2 has a size of 2kb, stride of 128 bytes and 16 rows.
Userspace was broken and assumed 8 rows. Chris Wilson noted that the
kernel unfortunately can't reliable check that because libdrm rounds
up the size to the next bucket.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Keith Packard <keithp@keithp.com>
Convert our open coded offset_in_page() to the common macro.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Keith Packard <keithp@keithp.com>
Change each shrinker's API by consolidating the existing parameters into
shrink_control struct. This will simplify any further features added w/o
touching each file of shrinker.
[akpm@linux-foundation.org: fix build]
[akpm@linux-foundation.org: fix warning]
[kosaki.motohiro@jp.fujitsu.com: fix up new shrinker API]
[akpm@linux-foundation.org: fix xfs warning]
[akpm@linux-foundation.org: update gfs2]
Signed-off-by: Ying Han <yinghan@google.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Pavel Emelyanov <xemul@openvz.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The registers are the same as on Sandybridge. Fixes scrambled display
in X when it does software drawing to the GTT, and scans the results
out as tiled.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Keith Packard <keithp@keithp.com>
This is once less place to miss a new INTEL_INFO(dev)->gen update now.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Keith Packard <keithp@keithp.com>
Rely on the GPU snooping into the CPU cache for appropriately bound
objects on MI_FLUSH. Or perhaps one day we will have a cache-coherent
CPU/GPU package...
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Keith Packard <keithp@keithp.com>
... to clarify just how we use it inside the driver and remove the
confusion of the poorly matching agp_type names. We still need to
translate through agp_type for interface into the fake AGP driver.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Keith Packard <keithp@keithp.com>
We always skipped flushing the BLT ring if the request flush did not
include the RENDER domain. However, this neglects that we try to flush
the COMMAND domain after every batch and before the breadcrumb interrupt
(to make sure the batch is indeed completed prior to the interrupt
firing and so insuring CPU coherency). As a result of the missing flush,
incoherency did indeed creep in, most notable when using lots of command
buffers and so potentially rewritting an active command buffer (i.e.
the GPU was still executing from it even though the following interrupt
had already fired and the request/buffer retired).
As all ring->flush routines now have the same preconditions, de-duplicate
and move those checks up into i915_gem_flush_ring().
Fixes gem_linear_blit.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=35284
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Tested-by: mengmeng.meng@intel.com
... even though it was disabled. A mistake in the handling of fence reuse
caused us to skip the vital delay of waiting for the object to finish
rendering before changing the register. This resulted in us changing the
fence register whilst the bo was active and so causing the blits to
complete using the wrong stride or even the wrong tiling. (Visually the
effect is that small blocks of the screen look like they have been
interlaced). The fix is to wait for the GPU to finish using the memory
region pointed to by the fence before changing it.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=34584
Cc: Andy Whitcroft <apw@canonical.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
[Note for 2.6.38-stable, we need to reintroduce the interruptible passing]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Tested-by: Dave Airlie <airlied@linux.ie>
When i915_gem_retire_requests_ring calls i915_gem_request_remove_from_client,
the client_list for that request may already be removed in i915_gem_release.
So we may call twice list_del(&request->client_list), resulting in an
oops like this report:
[126167.230394] BUG: unable to handle kernel paging request at 00100104
[126167.230699] IP: [<f8c2ce44>] i915_gem_retire_requests_ring+0xd4/0x240 [i915]
[126167.231042] *pdpt = 00000000314c1001 *pde = 0000000000000000
[126167.231314] Oops: 0002 [#1] SMP
[126167.231471] last sysfs file: /sys/devices/LNXSYSTM:00/device:00/PNP0C0A:00/power_supply/BAT1/current_now
[126167.231901] Modules linked in: snd_seq_dummy nls_utf8 isofs btrfs zlib_deflate libcrc32c ufs qnx4 hfsplus hfs minix ntfs vfat msdos fat jfs xfs exportfs reiserfs cryptd aes_i586 aes_generic binfmt_misc vboxnetadp vboxnetflt vboxdrv parport_pc ppdev snd_hda_codec_hdmi snd_hda_codec_conexant snd_hda_intel snd_hda_codec snd_hwdep arc4 snd_pcm snd_seq_midi snd_rawmidi snd_seq_midi_event snd_seq uvcvideo videodev snd_timer snd_seq_device joydev iwlagn iwlcore mac80211 snd cfg80211 soundcore i915 drm_kms_helper snd_page_alloc psmouse drm serio_raw i2c_algo_bit video lp parport usbhid hid sky2 sdhci_pci ahci sdhci libahci
[126167.232018]
[126167.232018] Pid: 1101, comm: Xorg Not tainted 2.6.38-6-generic-pae #34-Ubuntu Gateway MC7833U /
[126167.232018] EIP: 0060:[<f8c2ce44>] EFLAGS: 00213246 CPU: 0
[126167.232018] EIP is at i915_gem_retire_requests_ring+0xd4/0x240 [i915]
[126167.232018] EAX: 00200200 EBX: f1ac25b0 ECX: 00000040 EDX: 00100100
[126167.232018] ESI: f1a2801c EDI: e87fc060 EBP: ef4d7dd8 ESP: ef4d7db0
[126167.232018] DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
[126167.232018] Process Xorg (pid: 1101, ti=ef4d6000 task=f1ba6500 task.ti=ef4d6000)
[126167.232018] Stack:
[126167.232018] f1a28000 f1a2809c f1a28094 0058bd97 f1aa2400 f1a2801c 0058bd7b 0058bd85
[126167.232018] f1a2801c f1a28000 ef4d7e38 f8c2e995 ef4d7e30 ef4d7e60 c14d1ebc f6b3a040
[126167.232018] f1522cc0 000000db 00000000 f1ba6500 ffffffa1 00000000 00000001 f1a29214
[126167.232018] Call Trace:
Unfortunately the call trace reported was cut, but looking at debug
symbols the crash is at __list_del, when probably list_del is called
twice on the same request->client_list, as the dereferenced value is
LIST_POISON1 + 4, and by looking more at the debug symbols before
list_del call it should have being called by
i915_gem_request_remove_from_client
And as I can see in the code, it seems we indeed have the possibility
to remove a request->client_list twice, which would cause the above,
because we do list_del(&request->client_list) on both
i915_gem_request_remove_from_client and i915_gem_release
As Chris Wilson pointed out, it's indeed the case:
"(...) I had thought that the actual insertion/deletion was serialised
under the struct mutex and the intention of the spinlock was to protect
the unlocked list traversal during throttling. However, I missed that
i915_gem_release() is also called without struct mutex and so we do need
the double check for i915_gem_request_remove_from_client()."
This change does the required check to avoid the duplicate remove of
request->client_list.
Bugzilla: http://bugs.launchpad.net/bugs/733780
Cc: stable@kernel.org # 2.6.38
Signed-off-by: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
* 'intel/drm-intel-next' of ../drm-next: (755 commits)
drm/i915: Only wait on a pending flip if we intend to write to the buffer
drm/i915/dp: Sanity check eDP existence
drm/i915: Rebind the buffer if its alignment constraints changes with tiling
drm/i915: Disable GPU semaphores by default
drm/i915: Do not overflow the MMADDR write FIFO
Revert "drm/i915: fix corruptions on i8xx due to relaxed fencing"
drm/i915: Don't save/restore hardware status page address register
drm/i915: don't store the reg value for HWS_PGA
drm/i915: fix memory corruption with GM965 and >4GB RAM
Linux 2.6.38-rc7
Revert "TPM: Long default timeout fix"
drm/i915: Re-enable GPU semaphores for SandyBridge mobile
drm/i915: Replace vblank PM QoS with "Interrupt-Based AGPBUSY#"
Revert "drm/i915: Use PM QoS to prevent C-State starvation of gen3 GPU"
drm/i915: Allow relocation deltas outside of target bo
drm/i915: Silence an innocuous compiler warning for an unused variable
fs/block_dev.c: fix new kernel-doc warning
ACPI: Fix build for CONFIG_NET unset
mm: <asm-generic/pgtable.h> must include <linux/mm_types.h>
x86: Use u32 instead of long to set reset vector back to 0
...
Conflicts:
drivers/gpu/drm/i915/i915_gem.c
Early gen3 and gen2 chipset do not have the relaxed per-surface tiling
constraints of the later chipsets, so we need to check that the GTT
alignment is correct for the new tiling. If it is not, we need to
rebind.
Reported-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
The code paths for modesetting are growing in complexity as we may need
to move the buffers around in order to fit the scanout in the aperture.
Therefore we face a choice as to whether to thread the interruptible status
through the entire pinning and unbinding code paths or to add a flag to
the device when we may not be interrupted by a signal. This does the
latter and so fixes a few instances of modesetting failures under stress.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Dave Airlie spotted that we had a potential bug should we ever rearrange
the drm_i915_gem_object so not the base drm_gem_object was not its first
member. He noticed that we often convert the return of
drm_gem_object_lookup() immediately into drm_i915_gem_object and then
check the result for nullity. This is only valid when the base object is
the first member and so the superobject has the same address. Play safe
instead and use the compiler to convert back to the original return
address for sanity testing.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
... so we handle that for i915_gem_fault() in the same manner as
ERESTARTSYS, or we send a SIGBUS to the faulting application.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
A lot of minor tweaks to fix the tracepoints, improve the outputting for
ftrace, and to generally make the tracepoints useful again. It is a start
and enough to begin identifying performance issues and gaps in our
coverage.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
By returning EAGAIN upon a wedged GPU before attempting to wait, we
would hit an infinite loop of repeating operation without ever
progressing. Instead this needs to be EIO so that userspace knows that
the GPU is truly wedged and not in the process of error recovery.
Similarly, we need to handle the error recovery during i915_gem_fault.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This is just an idea that might or might not be a good idea,
it basically adds two ioctls to create a dumb and map a dumb buffer
suitable for scanout. The handle can be passed to the KMS ioctls to create
a framebuffer.
It looks to me like it would be useful in the following cases:
a) in development drivers - we can always provide a shadowfb fallback.
b) libkms users - we can clean up libkms a lot and avoid linking
to libdrm_*.
c) plymouth via libkms is a lot easier.
Userspace bits would be just calls + mmaps. We could probably
mark these handles somehow as not being suitable for acceleartion
so as top stop people who are dumber than dumb.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Instead of reporting EIO upfront in the entrance of an ioctl that may or
may not attempt to use the GPU, defer the actual detection of an invalid
ioctl to when we issue a GPU instruction. This allows us to continue to
use bo in video memory (via pread/pwrite and mmap) after the GPU has hung.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
We can only utilize the stolen portion of the GTT if we are in sole
charge of the hardware. This is only true if using GEM and KMS,
otherwise VESA continues to access stolen memory.
Reported-by: Arnd Bergmann <arnd@arndb.de>
Reported-by: Frederic Weisbecker <fweisbec@gmail.com>
Tested-by: Jiri Olsa <jolsa@redhat.com>
Tested-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
There are I915_NUM_RINGS-1 inter-ring synchronisation counters, but we
were clearing I915_NUM_RINGS of them. Oops.
Reported-by: Jiri Slaby <jirislaby@gmail.com>
Tested-by: Jiri Slaby <jirislaby@gmail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Rather than evicting an object at random, which is unlikely to alleviate
the memory pressure sufficient to allow us to continue, zap the entire
aperture. That should give the system long enough to recover and reap
some pages from the evicted objects, forestalling the allocation error
for the new object.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
In order to retire active buffers whilst no client is active, we need to
insert our own flush requests onto the ring.
This is useful for servers that queue up some rendering and then go to
sleep as it allows us to the complete processing of those requests,
potentially making that memory available again much earlier.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
We need to ensure that writes through the GTT land before any
modification to the MMIO registers and so must impose a mandatory write
barrier when flushing the GTT domain. This was revealed by relaxing the
write ordering by experimentally mapping the registers and the GATT as
write-combining.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
The relative-to-general state default is useless as it means having to
rewrite the streaming kernels for each batch. Relative-to-surface is
more useful, as that stream usually needs to be rewritten for each
batch. And absolute addressing mode, vital if you start streaming
state, is also only available by adjusting the register...
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
In order to enforce the correct memory barriers for irq get/put, we need
to perform the actual counting using atomic operations.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
In order for bos to retire eventually, a request must be sent down the
ring. This is expected, for example, by occlusion queries for which mesa
will wait upon (whilst running glean) before issuing more batches and so
the normal activity upon the ring is suspended and we need to emit a
request to clear the idle ring.
Reported-by: Jinjin, Wang <jinjin.wang@intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=30380
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
I'm still seeing tiling corruption of PutImage and CopyArea (I think)
under mutter on pnv, so obviously the pipelining logic is deeply flawed.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
The bulk of the change is to convert the growing list of rings into an
array so that the relationship between the rings and the semaphore sync
registers can be easily computed.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As the tracepoint is now decoupled from when the actual register is
assigned and was never complemented by detailing when the object lost
its fence, it has outlived its limited usefulness. Profiling the actual
stalls is a far more profitable venture anyway.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As the userspace mappings are torn down on every GPU write, we prefer to
track when the buffer is activated (via a fresh i915_gem_fault). This
makes the LRU conceptually simpler. With coherent mappings, the
remaining use-case for set_domain_ioctl is GPU synchronisation.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
With this change, every batchbuffer can use all available fences (save
pinned and scanout, of course) without ever stalling the gpu!
In theory. Currently the actual pipelined update of the register is
disabled due to some stability issues. However, just the deferred update
is a significant win.
Based on a series of patches by Daniel Vetter.
The premise is that before every access to a buffer through the GTT we
have to declare whether we need a register or not. If the access is by
the GPU, a pipelined update to the register is made via the ringbuffer,
and we track the last seqno of the batches that access it. If by the
CPU we wait for the last GPU access and update the register (either
to clear or to set it for the current buffer).
One advantage of being able to pipeline changes is that we can defer the
actual updating of the fence register until we first need to access the
object through the GTT, i.e. we can eliminate the stall on set_tiling.
This is important as the userspace bo cache does not track the tiling
status of active buffers which generate frequent stalls on gen3 when
enabling tiling for an already bound buffer.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
... so that upon first use after resume we will reacquire the fence reg.
Reported-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
We don't track gpu flush request in any special way. So even with
obj->write_domain == 0, a gpu flush might be outstanding but no
yet executed. Even worse, the latest request might use the object
only for reading. So and unconditional call to object_wait_rendering
is needed for !pipelined.
Hence revert that patch fully and untangle the flushing from the
synchronization again.
Reported-by: Keith Packard <keithp@keithp.com>
Tested-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Besides the minimal improvement in reducing the execbuffer overhead, the
real benefit is clarifying a few routines.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
A number of dragons have been seen lurking within the execbuffer code.
The first step is then to isolate them from the rest and begin to
scrutinise them in depth. Suggested by Daniel Vetter.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Simply remove our accounting of objects inside the aperture, keeping
only track of what is in the aperture and its current usage. This
removes the over-complication of BUGs that were attempting to keep the
accounting correct and also removes the overhead of the accounting on
the hot-paths.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
With KMS, we can simply relinquish the fence when we idle the GPU and
reassign it upon first use.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Avoid evicting buffers that will be used later in the batch in order to
make room for the initial buffers by pinning all bound buffers in a
single pass before binding (and evicting for) fresh buffer.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This used to check the precondition that all fences were to be located
in a mappable area, redundant now as those two parameters are combined
into one.
After pinning, we assert that the buffer is bound into the desired
region.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
The pipe control object is allocated by the device for the sole use of the
render ringbuffer. Move this detail from the general code to the render
ring buffer initialisation.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Combining map_and_fenceable revealed a bug in
i915_gem_object_gtt_size() in that it always computed the appropriate
fence size for the object regardless of tiling state which caused us to
over-allocate linear buffers when binding to the GTT.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This is required to restore gtt mappings on resume when agp is gone.
The right way to do this would be to make sturct drm_mm_node embeddable
and use the allocation list maintained by the drm memory manager. But
that's a bigger project. Getting rid of the per bo agp_mem will save
more memory than this wastes, anyway.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Currently if we hit a pagefault when applying a user relocation for the
execbuffer, we bail and return EFAULT to the application. Instead, we
need to unwind, drop the dev->struct_mutex, copy all the relocation
entries to a vmalloc array (to avoid any potential circular deadlocks
when resolving the pagefault), retake the mutex and then apply the
relocations. Afterwards, we need to again drop the lock and copy the
vmalloc array back to userspace.
v2: Incorporate feedback from Daniel Vetter.
Reported-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Commit 2549d6c2 removed the vmalloc used for temporary storage of the
relocation lists used during execbuffer. However, our use of vmalloc was
being protected by an integer overflow check which we do want to
preserve!
Reported-by: Dan Carpenter <error27@gmail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Linus Torvalds found that it was rather trivial to trigger a system
freeze:
In fact, with lockdep, I don't even need to do the sysrq-d thing: it
shows the bug as it happens. It's the X server taking the same lock
recursively.
Here's the problem:
=============================================
[ INFO: possible recursive locking detected ]
2.6.37-rc2-00012-gbdbd01a #7
---------------------------------------------
Xorg/2816 is trying to acquire lock:
(&dev->struct_mutex){+.+.+.}, at: [<ffffffff812c626c>] i915_gem_fault+0x50/0x17e
but task is already holding lock:
(&dev->struct_mutex){+.+.+.}, at: [<ffffffff812c403b>] i915_mutex_lock_interruptible+0x28/0x4a
other info that might help us debug this:
2 locks held by Xorg/2816:
#0: (&dev->struct_mutex){+.+.+.}, at: [<ffffffff812c403b>] i915_mutex_lock_interruptible+0x28/0x4a
#1: (&mm->mmap_sem){++++++}, at: [<ffffffff81022d4f>] page_fault+0x156/0x37b
This recursion was introduced by rearranging the locking to avoid the
double locking on the fast path (4f27b5d and fbd5a26d) and the
introduction of the prefault to encourage the fast paths (b5e4f2b). In
order to undo the problem, we rearrange the code to perform the access
validation upfront, attempt to prefault and then fight for control of the
mutex. the best case scenario where the mutex is uncontended the
prefaulting is not wasted.
Reported-and-tested-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
g33/pineview doesn't have any alignment constrains for unfenced tiled
buffers. But older chips have. Fix this.
Problem introduced in a00b10c360.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
An old and oft reported bug, is that of the GPU hanging on a
MI_WAIT_FOR_EVENT following a mode switch. The cause is that the GPU is
waiting on a scanline counter on an inactive pipe, and so waits for a
very long time until eventually the user reboots his machine.
We can prevent this either by moving the WAIT into the kernel and
thereby incurring considerable cost on every swapbuffers, or by waiting
for the GPU to retire the last batch that accesses the framebuffer
before installing a new one. As mode switches are much rarer than swap
buffers, this looks like an easy choice.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=28964
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=29252
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: stable@kernel.org
... and so prevent a potential circular reference:
[ INFO: possible circular locking dependency detected ]
2.6.37-rc1-uwe1+ #4
-------------------------------------------------------
Xorg/1401 is trying to acquire lock:
(&mm->mmap_sem){++++++}, at: [<c01e4ddb>] might_fault+0x4b/0xa0
but task is already holding lock:
(&dev->struct_mutex){+.+.+.}, at: [<f869c3ac>]
i915_mutex_lock_interruptible+0x3c/0x60 [i915]
which lock already depends on the new lock.
When the locking around the pwrite ioctl was simplified, I did not spot
that the phys path never took any locks and so we introduced this
potential circular reference.
Reported-by: Uwe Helm <uwe.helm@googlemail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Instead of killing the process, just return no page found and reschedule
the process giving the GPU some time to (hopefully) recover.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
a00b10c360 "Only enforce fence limits inside the GTT" also
added a fenceable/mappable disdinction when binding/pinning buffers.
This only complicates the code with no pratical gain:
- In execbuffer this matters on for g33/pineview, as this is the only
chip that needs fences and has an unmappable gtt area. But fences
are only possible in the mappable part of the gtt, so need_fence
implies need_mappable. And need_mappable is only set independantly
with relocations which implies (for sane userspace) that the buffer
is untiled.
- The overlay code is only really used on i8xx, which doesn't have
unmappable gtt. And it doesn't support tiled buffers, currently.
- For all other buffers it's a bug to pass in a tiled bo.
In short, this disdinction doesn't have any practical gain.
I've also reverted mapping the overlay and context pages as possibly
unmappable. It's not worth being overtly clever here, all the big
gains from unmappable are for execbuf bos.
Also add a comment for a clever optimization that confused me
while reading the original patch by Chris Wilson.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When merging Daniel's full-gtt patches I had a set of tweaks which I
thought I had undone. I was half right...
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=31286
Reported-by: jinjin.wang@intel.com
Reported-by: Alexey Fisher <bug-track@fisher-privat.net>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Take two passes to evict everything whilst searching for sufficient free
space to bind the batchbuffer. After searching for sufficient free space
using LRU eviction, evict everything that is purgeable and try again.
Only then if there is insufficient free space (or the GTT is too badly
fragmented) evict everything from the aperture and try one last time.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Accessing the uninitialised obj->pages instead of the local page lead to
an OOPs.
Reported-by: Xavier Chantry <chantry.xavier@gmail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
So long as we adhere to the fence registers rules for alignment and no
overlaps (including with unfenced accesses to linear memory) and account
for the tiled access in our size allocation, we do not have to allocate
the full fenced region for the object. This allows us to fight the bloat
tiling imposed on pre-i965 chipsets and frees up RAM for real use. [Inside
the GTT we still suffer the additional alignment constraints, so it doesn't
magic allow us to render larger scenes without stalls -- we need the
expanded GTT and fence pipelining to overcome those...]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Also spotted by Dan Carpenter.
obj->pin_count is unsigned so the BUG_ON(obj->pin_count<0) will never
trigger.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
The error code is only expected during the actual pruning and not during
the first measurement (nr_to_scan == 0) pass.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
It is possible for the active list to only contain a read-only buffer so
that the ring->gpu_write_list remains entry. This leads to an
inconsistency between i915_gpu_is_active() and i915_gpu_idle() causing
an infinite spin during the shrinker and an assertion failure that
i915_gpu_idle() does indeed flush all buffers from the active lists.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
In order to force a page-fault on a GTT mapping after we start using it
from the GPU and so enforce correct CPU/GPU synchronisation, we need to
invalidate the mapping.
Pointed out by Owain G. Ainsworth.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
By using read_cache_page() for individual pages during pwrite/pread we
can eliminate an unnecessary large allocation (and immediate free) of
obj->pages. Also this eliminates any potential nesting of get/put pages,
simplifying the code and preparing the path for greater things.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Since we rarely use the mmap_offset and it is easily computable from the
obj->map_list.hash, remove it.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Eliminate the racy device unload by embedding a shrinker into each
device. Smaller, simpler code.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
More precisely: For those that _need_ to be mappable. Also add two
BUG_ONs in fault and pin to check the consistency of the mappable
flag.
Changes in v2:
- Add tracking of gtt mappable space (to notice mappable/unmappable
balancing issues).
- Improve the mappable working set tracking by tracking fault and pin
separately.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This way we can make some more educated guesses as to why exactly
we can't use 2G apertures to their full potential ;)
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
At least the part that's currently enabled by the BIOS.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
In i915_gem_object_pin obviously unbind only if mappable is true.
This is the last part to enable gtt_mappable_end != gtt_size, which
the next patch will do.
v2: Fences on g33/pineview only work in the mappable part of the
gtt.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Like before add a parameter mappable (also to gem_object_pin) and
set it depending upon the context. Only bos that are brought into
the gtt due to an execbuffer call can be put into the unmappable
part of the gtt, everything else (especially pinned objects) need
to be put into the mappable part of the gtt.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Add a mappable parameter to i915_gem_evict_something to distinguish
the two cases (non-restricted vs. mappable gtt allocations). No
functional changes because the mappable limit is set to the end of
the gtt currently.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Preparing the ringbuffer for adding new commands can fail (a timeout
whilst waiting for the GPU to catch up and free some space). So check
for any potential error before overwriting HEAD with new commands, and
propagate that error back to the user where possible.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
The ringbuffer keeps a pointer to the parent device, so we can use that
instead of passing around the pointer on the stack.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
* 'drm-core-next' of git://git.kernel.org/pub/scm/linux/kernel/git/airlied/drm-2.6: (476 commits)
vmwgfx: Implement a proper GMR eviction mechanism
drm/radeon/kms: fix r6xx/7xx 1D tiling CS checker v2
drm/radeon/kms: properly compute group_size on 6xx/7xx
drm/radeon/kms: fix 2D tile height alignment in the r600 CS checker
drm/radeon/kms/evergreen: set the clear state to the blit state
drm/radeon/kms: don't poll dac load detect.
gpu: Add Intel GMA500(Poulsbo) Stub Driver
drm/radeon/kms: MC vram map needs to be >= pci aperture size
drm/radeon/kms: implement display watermark support for evergreen
drm/radeon/kms/evergreen: add some additional safe regs v2
drm/radeon/r600: fix tiling issues in CS checker.
drm/i915: Move gpu_write_list to per-ring
drm/i915: Invalidate the to-ring, flush the old-ring when updating domains
drm/i915/ringbuffer: Write the value passed in to the tail register
agp/intel: Restore valid PTE bit for Sandybridge after bdd3072
drm/i915: Fix flushing regression from 9af90d19f
drm/i915/sdvo: Remove unused encoding member
i915: enable AVI infoframe for intel_hdmi.c [v4]
drm/i915: Fix current fb blocking for page flip
drm/i915: IS_IRONLAKE is synonymous with gen == 5
...
Fix up conflicts in
- drivers/gpu/drm/i915/{i915_gem.c, i915/intel_overlay.c}: due to the
new simplified stack-based kmap_atomic() interface
- drivers/gpu/drm/vmwgfx/vmwgfx_drv.c: added .llseek entry due to BKL
removal cleanups.
Keep the current interface but ignore the KM_type and use a stack based
approach.
The advantage is that we get rid of crappy code like:
#define __KM_PTE \
(in_nmi() ? KM_NMI_PTE : \
in_irq() ? KM_IRQ_PTE : \
KM_PTE0)
and in general can stop worrying about what context we're in and what kmap
slots might be appropriate for that.
The downside is that FRV kmap_atomic() gets more expensive.
For now we use a CPP trick suggested by Andrew:
#define kmap_atomic(page, args...) __kmap_atomic(page)
to avoid having to touch all kmap_atomic() users in a single patch.
[ not compiled on:
- mn10300: the arch doesn't actually build with highmem to begin with ]
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix up drivers/gpu/drm/i915/intel_overlay.c]
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Chris Metcalf <cmetcalf@tilera.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: David Miller <davem@davemloft.net>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Dave Airlie <airlied@linux.ie>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
... to prevent flush processing of an idle (or even absent) ring.
This fixes a regression during suspend from 87acb0a5.
Reported-and-tested-by: Alexey Fisher <bug-track@fisher-privat.net>
Tested-by: Peter Clifton <pcjc2@cam.ac.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When the object has been written to by the gpu it remains on the ring
until its flush has been retired. However, when the object is moving to
the ring and the associated cache needs to be invalidated, we need to
perform the flush on the target ring, not the one it came from (which is
NULL in the reported case and so the flush was entirely absent).
Reported-by: Peter Clifton <pcjc2@cam.ac.uk>
Reported-and-tested-by: Alexey Fisher <bug-track@fisher-privat.net>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Whilst moving the code around in 9af90d19f, I dropped the or'ing in of
new write domains which would zero out the write domain for a render
target if later reused as a source later in the batch. This meant that
we might drop a required flush before reading from the render target.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=31043
Reported-by: xunx.fang@intel.com
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Based on an original patch by Zhenyu Wang, this initializes the BLT ring for
SandyBridge and enables support for user execbuffers.
Cc: Zhenyu Wang <zhenyuw@linux.intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If the userspace driver is using a constant relocation array with a
static buffer, they will pass the same relocation array back to the
kernel. So we *do* need to update the presumed offset value in those
relocations to reflect the current object so that they remain correct
with future batchbuffers and we avoid the necessity of having to suspend
execution and perform redundant relocations.
Fixes the regression introduced by 12f889c for applications using
absolute addressing on trees of buffer (i.e. the current consumers of
libdrm_intel.so).
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=30996
Reported-by: Wang, Jinjin <jinjin.wang@intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
To handle retirements, we need per-ring tracking of active objects.
To handle evictions, we need global tracking of active objects.
As we enable more rings, rebuilding the global list from the individual
per-ring lists quickly grows tiresome and overly complicated. Tracking the
active objects in two lists is the lesser of two evils.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
... by always initialising the empty ringbuffer it is always then safe
to check whether it is active.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
The most frequent relocation within a batchbuffer is a contiguous sequence
of vertex buffer relocations, for which we can virtually eliminate the
drm_gem_object_lookup() overhead by caching the last handle to object
translation.
In doing so we refactor the pin and relocate retry loop out of
do_execbuffer into its own helper function and so improve the error
paths.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
One of the primarily consumers of the i915 driver is X, a large signal
driven application. Frequently when writing into the buffers, there is a
pending signal which causes us not to take the interruptible lock but
then we need to take that same lock around the object unreference. By
rearranging the code to do the interruptible lock as the first check, we
can avoid the frequent additional locking around the unreference.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
... to avoid reacquiring it to drop the object reference count on
exit. Note we have to make sure we now drop (and reacquire) the lock
around acquiring the mm semaphore on the slow paths.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
After allocation a handle for the fresh object, we know that we can
safely drop the refcnt without triggering a free so we do not need the
mutex. Strangely, this mutex acquisition is the one that appears on
driver profiles.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Avoid an early eviction of the batch buffer into the uncached GTT
domain, and so do the relocation fixup in cacheable memory.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
... perform an access validation check up front instead and copy them in
on-demand, during i915_gem_object_pin_and_relocate(). As around 20% of
the CPU overhead may be spent inside vmalloc for the relocation entries
when submitting an execbuffer [for x11perf -aa10text], the savings are
considerable and result in around a 10% throughput increase [for glyphs].
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Currently, if a batch buffer refers to an object with a pending flip,
then we sleep until that pending flip is completed (unpinned and
signalled). This is so that a flip can be queued and the user can
continue rendering to the backbuffer oblivious to whether the buffer is
still pinned as the scan out. (The kernel arbitrating at the last moment
to stall the batch and wait until the buffer is unpinned and replaced as
the front buffer.)
As we only have a queue depth of 1, we can simply wait for the current
pending flip to complete and continue rendering. We can achieve this
with a single WAIT_FOR_EVENT command inserted into the ring buffer prior
to executing the batch, *without* stalling the client.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
* 'drm-intel-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/ickle/drm-intel:
drm/i915: Rephrase pwrite bounds checking to avoid any potential overflow
drm/i915: Sanity check pread/pwrite
drm/i915: Use pipe state to tell when pipe is off
drm/i915: vblank status not valid while training display port
drivers/gpu/drm/i915/i915_gem.c: Add missing error handling code
drm/i915: Fix refleak during eviction.
drm/i915: fix GMCH power reporting
Move the access control up from the fast paths, which are no longer
universally taken first, up into the caller. This then duplicates some
sanity checking along the slow paths, but is much simpler.
Tracked as CVE-2010-2962.
Reported-by: Kees Cook <kees@ubuntu.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: stable@kernel.org
Extend the error handling code with operations found in other nearby error
handling code
A simplified version of the sematic match that finds this problem is as
follows: (http://coccinelle.lip6.fr/)
// <smpl>
@r exists@
@r@
statement S1,S2,S3;
constant C1,C2,C3;
@@
*if (...)
{... S1 return -C1;}
...
*if (...)
{... when != S1
return -C2;}
...
*if (...)
{... S1 return -C3;}
// </smpl>
Signed-off-by: Julia Lawall <julia@diku.dk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: stable@kernel.org
The return from move_to_gtt_domain() may indicate a pending signal which
needs to handled as opposed to an actual error, for instance, so report
the original return value rather than forcing an EINVAL.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
* 'drm-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/airlied/drm-2.6:
vmwgfx: Fix fb VRAM pinning failure due to fragmentation
vmwgfx: Remove initialisation of dev::devname
vmwgfx: Enable use of the vblank system
vmwgfx: vt-switch (master drop) fixes
drm/vmwgfx: Fix breakage introduced by commit "drm: block userspace under allocating buffer and having drivers overwrite it (v2)"
drm: Hold the mutex when dropping the last GEM reference (v2)
drm/gem: handlecount isn't really a kref so don't make it one.
drm: i810/i830: fix locked ioctl variant
drm/radeon/kms: add quirk for MSI K9A2GM motherboard
drm/radeon/kms: fix potential segfault in r600_ioctl_wait_idle
drm: Prune GEM vma entries
drm/radeon/kms: fix up encoder info messages for DFP6
drm/radeon: fix PCI ID 5657 to be an RV410
When the GPU is reset, the fence registers are invalidated, so release
the objects and clear them out.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Only drm/i915 does the bookkeeping that makes the information useful,
and the information maintained is driver specific, so move it out of the
core and into its single user.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Dave Airlie <airlied@redhat.com>
There were lots of places being inconsistent since handle count
looked like a kref but it really wasn't.
Fix this my just making handle count an atomic on the object,
and have it increase the normal object kref.
Now i915/radeon/nouveau drivers can drop the normal reference on
userspace object creation, and have the handle hold it.
This patch fixes a memory leak or corruption on unload, because
the driver had no way of knowing if a handle had been actually
added for this object, and the fbcon object needed to know this
to clean itself up properly.
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Keith Packard