The affine_blt walks the destination image sampling from the source, so
we need to walk all the output pixels and not the source size.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
A confusion over the use of marking when we need to compute a
transformed cursor image using pixman version when we need to compute
transformed cursor coordinates meant that we failed to compute the
required transformation after simply rotating the monitor.
Should fix the regression from
commit bbbcc4101c
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Fri Sep 18 12:17:07 2015 +0100
sna: Apply cursor hotspot transformation
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
References: https://bugs.archlinux.org/task/46606
Continuing the udevless saga where we query for topology changes on
calls to RRGetInfo() and so where we cannot call RRGetInfo() ourselves
on discovering the changes (to facilitate hotplug). The next step is
explicitly prevent the recursive call.
References: https://bugs.freedesktop.org/show_bug.cgi?id=91929#c10
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As summarised by Mark Kettenis:
Unfortunately commit 2c08d72393 that
was made to resolve this bug introduced infinite recursion if
polling udev for events fails (which is always the case on systems
without udev). The scenario is as follows:
1. The server calls RRGetInfo().
2. RRGetInfo() calls sna_randr_getinfo()
3. sna_randr_getinfo() calls sna_uevent_poll(), which returns false
4. sna_randr_getinfo() calls sna_mode_discover()
5. sna_mode_discover() calls RRGetInfo()
and we jump straight back to step 2, until we run out of stack
space and crash
Reported-by: Mark Kettenis <kettenis@openbsd.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91929#c8
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If we are forced to use the static_request (reserved for allocation
failure), we dispatch the execbuffer synchronously and expect no buffers
to be present afterwards. Rather than just assert that this is so, we
should clean up the buffer list!
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When performing RRGetInfo() we drain any pending uevents before checking
the CRTC/connector statuses. However, some versions of libudev do a
blocking read and so when we call RRGetInfo() from a Timer callback
(outside of the usual sigtimer interrupts) that blocking read lasts
indefinitely.
Reported-by: Sedat Dilek <sedat.dilek@gmail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
For a read-only mmaping, we may be able to use a read-write (as we know
we don't write) if the memory permits us. We can check this by forcing
the get-user-pages via set-domain(GTT).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
After applying scaling to the cursor image, it was being positioned on
the display incorrectly due to the hotspot transformation remaining
unscaled and so offseting the cursor slightly.
References: https://bugs.archlinux.org/task/45572
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If the kernel doesn't accept the update, then the CRTC remains in its
previous mode i.e. we should not also touch the modesetting serial
stamp.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Force the full reprobe, resetting all the output properties on all
outputs if we detect a status change on hotplug. Rather than simply
translate the kernel mode lists ourselves, use the common functions in
case more settings change.
References: https://bugs.freedesktop.org/show_bug.cgi?id=91929
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Since we only update the mode list when the user queries xrandr or upon
a hotplug event, queue a deferred probe for 2s after X starts in order
to fill in the unknown modes.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Some display managers, for example mutter, only use
XRRScreenResourcesCurrent() which means they only query the existing
modes on an output. We need to therefore update the list of available
modes after a hotplug event.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
During the hotplug event, we query the current connector status and use
that to trigger an output changed event to the clients. However, since
we know the new status, we can set that on the RROutput immediately.
Note the modelist is left unchanged, and will only be queried when the
user requests it (though we may want to provide that in the hotplug
notify as well).
References: https://bugs.freedesktop.org/show_bug.cgi?id=91929#c2
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
In order to catch transitions between SwapInterval 0 and 1 in the middle
of a flip sequence, we need to reset our swap_limit every time.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Just in case we want to debug the native vs software paths and allow
switching without recompiling. For the unfortunate
Section "Device"
Identifier "igfx"
Driver "intel"
Option "HWRotation" "off"
EndSection
default is for us to use HW rotation where available (for speed, reduced
resource usage and reduce power usage).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Once again I tried to add an event to a pixmap, but only Windows carry
the CRTC tracking information. Stop it.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
We can use the same macro on top of the xorg list.h interloper as well
as the private version, so do so.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Even though the kernel reports and error and we skip the queuing, we
have information to avoid the roundtrip through the kernel.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
On a headless or hosted system, we do not have any CRTC on which to
report or queue the vblank. Reporting a BadDrawable is evil (as clients
never expect the error) but is better than crashing the Xserver.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Along the blit swap path, we expect the buffer to not be scanout
compatible and not have the same pitch as the front buffer. The reuse
logic failed to take this into account!
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Fixes a regression from
commit 3f128867d9
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Fri Aug 7 15:19:17 2015 +0100
sna: Skip a no-op copy
that forgot that we can flush damage but still have valid contents to
copy from.
Reported-by: Timo Aaltonen <tjaalton@ubuntu.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As the window may be reparented during its swap lifetime, we may be able
to recreate the backbuffer ready for flipping next swap.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Since the DRI2BufferPtr destruction may be delayed, by the time we do
free it, the state may no longer be accurate. (At this level at least,
there is a later assertion to the same regard just as we think we are
releasing the buffer.)
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When we release the accel backend, we mark the interface as wedged.
However, when we release the modesetting backend we flush the vblank
queue which may want to do rendering. So if we do that first then
disable the accel backend, things we go much smoother.
References: https://bugs.freedesktop.org/show_bug.cgi?id=91658#c20
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This so that we permanently exclude the active front buffer from the set
of triple buffer replacements.
References: https://bugs.freedesktop.org/show_bug.cgi?id=91658
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Down one particular patch (overlapping render copies) the
scratch.pScreen is dereferenced (to create a temporary Pixmap). Ensure
that we populate it before hand.
References: https://bugs.freedesktop.org/show_bug.cgi?id=91658
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
To ease tracking of the next swap, stash it on the event (which is then
reused) rather than the back buffer (which changes frequently).
In addition, add debug flags and assertions to track event stages (such
as making sure we do not decouple/free an event that we have sent a
signal back to the client).
References: https://bugs.freedesktop.org/show_bug.cgi?id=91658
References: https://bugs.freedesktop.org/show_bug.cgi?id=70461
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If we don't control the vblank for the window, we cannot use the vblank
signal as it ends up being queued as well.
References: https://bugs.freedesktop.org/show_bug.cgi?id=91658
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
A side effect of commit 4cea803798
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Sat Aug 8 22:10:20 2015 +0100
sna/dri2: Restore caching of fullscreen triple buffers
was that we discovered that we had been inserting the buffer into the
cache list multiple times.
Reported-by: Dan Doel <dan.doel@gmail.com>
References: https://bugs.freedesktop.org/show_bug.cgi?id=91658#c5
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If the DRI2 copy ends up on the CPU, then we will not have a GPU
request for the operation.
Reported-by: Zdenek Kabelac <zkabelac@redhat.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If a client dies with a pending flip, we have to complete it so that the
scanout matches the frontbuffer again. This requires one last flip
following the client disapparance.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Letting GL clients run amok when offscreen (or when hidden) has
unexpected UX behaviour. Instead we can elect to throttle them to the
nearest CRTC and keep the desktop running smoothly.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Similarly to the fix for reflection into rotation, make sure we only set
one Rotation bit afterwards.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When reducing a xy reflection to a 180 degree rotation, make sure
only one rotation bit is set. Also by rotating the bit left, we
can support cases where xy reflection happens with 90/270 degree
rotation.
Signed-off-by: Bob Paauwe <bob.j.paauwe@intel.com>
[ickle: Interleave comments]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When the client misses a swap, we consider it idle and unlikely to swap
again for a while. We try to take advantage of that and remove the old
back buffers. But it is likely to swap again and so having some of that
cache around would be advantageous.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Chris Wilson
Robert Ancell
Olivier Fourdan
Bob Paauwe