The recent addition of lockdep support to reservations and their subsequent
use by TTM showed up a number of potential problems with the way qxl was using
TTM objects.
a) it was allocating objects, and reserving them later without validating
underneath the reservation, which meant in extreme conditions the objects could
be evicted before the reservation ever used them.
b) it was reserving objects straight after allocating them, but with no
ability to back off should the reservations fail. It now allocates the necessary
objects then does a complete reservation pass on them to avoid deadlocks.
c) it had two lists per release tracking objects, unnecessary complicating
the reservation process.
This patch removes the dual object tracking, adds reservations ticket support
to the release and fence object handling. It then ports the internal fb
drawing code and the userspace facing ioctl to use the new interfaces properly,
along with cleanup up the error path handling in some codepaths.
Signed-off-by: Dave Airlie <airlied@redhat.com>
In order to fix an issue with reservations we need to create the releases
as pre-pinned objects, this changes the placement interface and bo creation
interface to allow creating pinned objects to save nested reservations later.
This is just a stepping stone to main fix which follows to actually fix how
qxl deals with reservations.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Due to the nature of qxl hw we cannot queue operations while in an irq
context, so we queue these operations as best we can until atomic allocations
fail, and dequeue them later in a work queue.
Daniel looked over the locking on the list and agrees it should be sufficent.
The atomic allocs use no warn, as the last thing we want if we haven't memory
to allocate space for a printk in an irq context is more printks.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Pull drm updates from Dave Airlie:
"Okay this is the big one, I was stalled on the fbdev pull req as I
stupidly let fbdev guys merge a patch I required to fix a warning with
some patches I had, they ended up merging the patch from the wrong
place, but the warning should be fixed. In future I'll just take the
patch myself!
Outside drm:
There are some snd changes for the HDMI audio interactions on haswell,
they've been acked for inclusion via my tree. This relies on the
wound/wait tree from Ingo which is already merged.
Major changes:
AMD finally released the dynamic power management code for all their
GPUs from r600->present day, this is great, off by default for now but
also a huge amount of code, in fact it is most of this pull request.
Since it landed there has been a lot of community testing and Alex has
sent a lot of fixes for any bugs found so far. I suspect radeon might
now be the biggest kernel driver ever :-P p.s. radeon.dpm=1 to enable
dynamic powermanagement for anyone.
New drivers:
Renesas r-car display unit.
Other highlights:
- core: GEM CMA prime support, use new w/w mutexs for TTM
reservations, cursor hotspot, doc updates
- dvo chips: chrontel 7010B support
- i915: Haswell (fbc, ips, vecs, watermarks, audio powerwell),
Valleyview (enabled by default, rc6), lots of pll reworking, 30bpp
support (this time for sure)
- nouveau: async buffer object deletion, context/register init
updates, kernel vp2 engine support, GF117 support, GK110 accel
support (with external nvidia ucode), context cleanups.
- exynos: memory leak fixes, Add S3C64XX SoC series support, device
tree updates, common clock framework support,
- qxl: cursor hotspot support, multi-monitor support, suspend/resume
support
- mgag200: hw cursor support, g200 mode limiting
- shmobile: prime support
- tegra: fixes mostly
I've been banging on this quite a lot due to the size of it, and it
seems to okay on everything I've tested it on."
* 'drm-next' of git://people.freedesktop.org/~airlied/linux: (811 commits)
drm/radeon/dpm: implement vblank_too_short callback for si
drm/radeon/dpm: implement vblank_too_short callback for cayman
drm/radeon/dpm: implement vblank_too_short callback for btc
drm/radeon/dpm: implement vblank_too_short callback for evergreen
drm/radeon/dpm: implement vblank_too_short callback for 7xx
drm/radeon/dpm: add checks against vblank time
drm/radeon/dpm: add helper to calculate vblank time
drm/radeon: remove stray line in old pm code
drm/radeon/dpm: fix display_gap programming on rv7xx
drm/nvc0/gr: fix gpc firmware regression
drm/nouveau: fix minor thinko causing bo moves to not be async on kepler
drm/radeon/dpm: implement force performance level for TN
drm/radeon/dpm: implement force performance level for ON/LN
drm/radeon/dpm: implement force performance level for SI
drm/radeon/dpm: implement force performance level for cayman
drm/radeon/dpm: implement force performance levels for 7xx/eg/btc
drm/radeon/dpm: add infrastructure to force performance levels
drm/radeon: fix surface setup on r1xx
drm/radeon: add support for 3d perf states on older asics
drm/radeon: set default clocks for SI when DPM is disabled
...
This adds suspend/resume and hibernate support for the KMS driver. it evicts
all the objects, turns off the outputs, and waits for the hw to go idle,
On resume, it resets the memslots, rings, monitors object and forces modeset.
Signed-off-by: Dave Airlie <airlied@redhat.com>
This splits the creation of the monitors config object out so we can
re-use it across suspend/resume later.
Signed-off-by: Dave Airlie <airlied@redhat.com>
This adds support for a default of 4 heads, with a command line
parameter to change the default number.
It also overhauls the modesetting code to handle this case properly,
and send the correct things to the hardware at the right time.
Signed-off-by: Dave Airlie <airlied@redhat.com>
qxl has a feature to allow the userspace driver do arbitrary resizes
when the viewer resizes, this fixes it by removing unnecessary code
from the kernel side.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Now that the code is compatible in semantics, flip the switch.
Use ww_mutex instead of the homegrown implementation.
ww_mutex uses -EDEADLK to signal that the caller has to back off,
and -EALREADY to indicate this buffer is already held by the caller.
ttm used -EAGAIN and -EDEADLK for those, respectively. So some changes
were needed to handle this correctly.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Many of the drivers didn't implement palette/gamma handling, but were forced
to provide stubs for the hooks to avoid drm_fb_helper from oopsing. Now that
the hooks are optional, we can eliminate all the stubs.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@gmail.com>
if the surface is evicted, this validation will happen
to the wrong place, I noticed this with other work I was
doing, haven't seen it go wrong in practice.
Signed-off-by: Dave Airlie <airlied@redhat.com>
This was a bogus way to figure out what the active framebuffer was,
just check if the underlying bo is the primary bo.
Signed-off-by: Dave Airlie <airlied@redhat.com>
So qxl has ioports, but it really really really doesn't want you
to write to them twice, but if you write and get a signal before
the irq arrives to let you know its completed, you have to think
ahead and avoid writing another time.
However this works fine for update area where really multiple
writes aren't the end of the world, however with create primary
surface, you can't ever do multiple writes. So this stop internal
kernel writes from doing interruptible waits, because otherwise
we have no idea if this write is a new one or a continuation of
a previous one.
virtual hw sucks more than real hw.
This fixes lockups and VM crashes when resizing and starting/stopping
X.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Fix build error when CONFIG_DEBUG_FS is disabled:
drivers/gpu/drm/qxl/qxl_debugfs.c: In function 'qxl_debugfs_init':
drivers/gpu/drm/qxl/qxl_debugfs.c:76:2: error: implicit declaration of function 'drm_debugfs_create_files'
drivers/gpu/drm/qxl/qxl_debugfs.c: In function 'qxl_debugfs_takedown':
drivers/gpu/drm/qxl/qxl_debugfs.c:84:2: error: implicit declaration of function 'drm_debugfs_remove_files'
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
/usr/lib/gcc/x86_64-linux-gnu/4.7/include/stddef.h:414:9: sparse: preprocessor token offsetof redefined
include/linux/stddef.h:17:9: this was the original definition
>> drivers/gpu/drm/qxl/qxl_drv.c:49:5: sparse: symbol 'qxl_modeset' was not declared. Should it be static?
Reported-by: kbuild test robot.
Signed-off-by: Dave Airlie <airlied@redhat.com>
QXL is a paravirtual graphics device used by the Spice virtual desktop
interface.
The drivers uses GEM and TTM to manage memory, the qxl hw fencing however
is quite different than normal TTM expects, we have to keep track of a number
of non-linear fence ids per bo that we need to have released by the hardware.
The releases are freed from a workqueue that wakes up and processes the
release ring.
releases are suballocated from a BO, there are 3 release categories, drawables,
surfaces and cursor cmds. The hw also has 3 rings for commands, cursor and release handling.
The hardware also have a surface id tracking mechnaism and the driver encapsulates it completely inside the kernel, userspace never sees the actual hw surface
ids.
This requires a newer version of the QXL userspace driver, so shouldn't be
enabled until that has been placed into your distro of choice.
Authors: Dave Airlie, Alon Levy
v1.1: fixup some issues in the ioctl interface with padding
v1.2: add module device table
v1.3: fix nomodeset, fbcon leak, dumb bo create, release ring irq,
don't try flush release ring (broken hw), fix -modesetting.
v1.4: fbcon cpu usage reduction + suitable accel flags.
Signed-off-by: Alon Levy <alevy@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
