We use the fence as a backchannel to commuicated the tiling mode to the
DRI2 client, so in such cases we cannot tolerate a failure to set the
tiling mode.
References: https://bugs.freedesktop.org/show_bug.cgi?id=70461#c114
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
I chopped off one too many characters when creating the path in the
presence of a colon-delimited list.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If we are doing synchronous but immediate swaps, we may notice that we
have a vblank placeholder from the last swap. If so, we can replace it
with the swap we need to perform on the next vblank.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As a slightly more convenient approach than loading the EDID on the
kernel command line, allow the user to specify the EDID path in the
Device section of xorg.conf, e.g.
Section "Device"
Identifier "igfx"
Option "CustomEDID" "DP1:/path/to/dp1.edid,DP2:/path/to/dp2.edid"
EndSection
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89945
Modesetting can be slow, and we certainly don't want to have to keep
restarting it due to an interrupt from an invisible cursor.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If we fail to CPU mmap an object, mark it as inaccessible and never try
again. The likely scenario is that is it a PRIME object not backed by
our shmemfs storage and so unmappable.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Suppose that the kernel reported a truncated EDID, we will eventually
crash by trying to access invalid memory. Prevent the crash by fixing up
the EDID header to fit the size read from the kernel (though we don't
fixup the checksum, so strict conformance tests will still complain, but
we won't crash.)
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
After the Window resizes, we should never hand back a buffer of the old
size or else we end up rendering garbage - with the possibilty of GPU
hangs or memory corruption.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91036
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
It's possible if the Window is redirected whilst the client render for
its backbuffer to no longer overlap with the visible Window. In this
case, we attempt to copy an empty region, but it debugging is enabled
this throws an assert.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91120
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
sna_dri2_xchg() is also called on the flip path to do an immediate swap
of a hidden buffer, in which case can_xchg() returns false as it
believes that the swap requires a flip.
Reported-by: Jiri Slaby <jirislaby@gmail.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=70461#c109
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When we record the next scanout for display after the following vblank,
note the size of the allocation as well as its name and bo et al.
References: https://bugs.freedesktop.org/show_bug.cgi?id=70461#c105
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Apply s/cacheing/caching/ because I am alone in my belief of how to
spell cacheing correctly :)
Suggested-by: Sedat Dilek <sedat.dilek@gmail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If we need to handle a FillRectangles request that is larger than the 3D
pipeline can handle, we need to split that operation up into smaller
chunks. This is achieved by redirecting the rendering through a view of
the surface (offset so that coordinates are within range), or else we
have to fallback to creating temporary surfaces.
If we fail to do so, we should hit an assertion that the render target
fits within the 3D pipeline inside gen4_emit_drawing_rectangle()
Reported-by: Zdenek Kabelac <zdenek.kabelac@gmail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Looks like amalgamating the per-CRTC xchg with the blit caused a few
regressions with TearFree single monitor flipping.
References: https://bugs.freedesktop.org/show_bug.cgi?id=91066
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
One assert(active_scanout) was firing when feed in an old flip bo, which
in hindsight was quite expected.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If we cannot cache the current backbuffer, due to either it not fitting
or because a reference to it still exists, delete the cache entry we
were holding for it - or else we slowly leak the cache list and
gradually get slower and slower walking along the growing chain.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
We can fail to submit the batch if the GPU is wedged (or the driver
upset), in which case the bo will be NULL and we should not blindly
dereference it.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Since the /sys/device/backlight never turned up we face an issue with
disambiguating the backlight on multi-GPU devices. Both intel_backlight
and nv_backlight are presented are raw interfaces, and on modern systems
the ACPI interface is defunct, so we need a way to distinguish them. So,
we fallback to our priority table of known interfaces and rank them
accordingly, first by class (platform, native, raw) and then by priority.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Since there were two paths performing similar tasks for swapping a
windowed buffer, they would get easily confused and queue themselves in
conflict with each other as they behaved slightly differently wrt to rapid
swaps. Almagamate the two paths such that the behaviour is the same -
with the complication being in tracking the spare active buffer.
Fixes (some at least!) flickering from
commit 72d208a7f2
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Thu Mar 5 12:26:15 2015 +0000
sna/dri2: Perform swap elision on windows for swap-interval==0
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If we fail to turn off an output via DPMS, disable the entire CRTC in
order to blank the output and save the screeen/power.
Reported-by: Lukas Hejtmanek <xhejtman@fi.muni.cz>
References: https://bugs.freedesktop.org/show_bug.cgi?id=90179
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Fixes copy'n'paste error in
commit d46c793663
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Fri Jun 12 13:50:35 2015 +0100
sna/dri2: Remove the active_scanout flag when deleting the event
Reported-by: Andreas Reis <andreas.reis@gmail.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=90968#c8
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If there is already a pending operation on the drawable, the swap is
queued. It is queued with SWAP_THROTTLE which classes with the
vblank_mode=0 chaining, so the choice is to either review all other uses
of SWAP_THROTTLE and duplicate the chain swap handling, or to choose when
use to the alternate copy or back buffer. This patch opts for the
latter.
Fixes regression from
commit dcb4d323ca
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Thu Jun 11 13:54:49 2015 +0100
sna/dri2: Mark the pending backbuffer copy as active
Reported-by: Christoph Haag <haagch@frickel.club>
Reported-by: Andreas Reis <andreas.reis@gmail.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=90968
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Keep the swap cache around for 100ms, so only if the app drops below
10fps do we start having to recreate buffers on the fly.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As the backend may submit the batch as it completes the copy, we need to
look at what request the dst_bo is in rather than assuming it is the
next_request.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Although we are ordered wrt to the client, there is no point allowing
the client to continue submitting rendering and eventually blocking
whilst we are still waiting for the previous frame to be shown.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Not only is the pixmap->flush flag shared, but so is the lower level
gpu_bo->flush flag, so further to
commit 19d1e4ee19
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Thu Jun 11 13:43:10 2015 +0100
sna/dri2: Be wary of interactions with DRI3 and sna_pixmap->flush
we need to be more careful when asserting the state of gpu_bo->flush
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
We still only keep it alive whilst the event chain is in progress, but
moving the cache onto the window allows us to reuse it easily for triple
buffer window swaps as well as full screen flips.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When the trapezoid is rendering to a surface larger than the 3D pipe can
handle, we split it into tiles. However, the code to do so insisted on
passing along the wrong pointer and consequently crashed.
Based on the patch by Carl Michal.
Reported-by: Carl Michal <michal@physics.ubc.ca>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=90940
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
In order to perform swap elision for windows, we defer the blit until
the next vblank and then do whatever was the last buffer to be presented
by the client. Whilst that copy is pending, in the same manner as a
pending flip, we cannot hand that buffer back to the client for use (or
else they will render over top of it before we copy from it, or even
worse we copy from it in the middle of rendering).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Since both DRI2 and DRI3 manipulate the sna_pixmap->flush flag, we have
to relaxation assertions that it is wholly owned by DRI2.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
We know Present is sending garbage msc into the wait_vblank, so just
warn about the error so that we do not prevent usage of --enable-debug
due to somebody else's bug.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
For PRIME bo, we need to use uncached render targets so that any writes
are flushed out to main memory where they can be immediately read by a
PCI device. For simplicity, we just request that PRIME bo be also
SCANOUTs as that ensures that they will be created with the right
attributes for coherent main memory.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As the slave may use the ProcessPending damage callback to do its own
copying, we need to flush before.
Reported-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If the source is only being exported for reading, we can skip adding it
to the flush list only to perform a no-op.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
In the case of an exported pixmap, e.g. with DRI3, it is possible for
the client to render into the pixmap whilst we are unaware. To serialise
the xserver and the client, we flush all operations on exported pixmaps
before talking to the client. In the case of COW however, we did not
flush the copy-on-write when transferring control to the client, and
thereby we could capture the modified contents.
Bugzilla: https://bugs.kde.org/show_bug.cgi?id=340202
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=90836
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
commit d1bf75f155
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Mon Jun 1 08:56:43 2015 +0000
sna: Compilation fixes for stable distros
caused an apparent regression by using the wrong ring for detecting
whether we could program the BCS tiling registers.
Reported-by: Sedat Dilek <sedat.dilek@gmail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Chris Wilson
Martin Peres
Damien Lespiau
Sedat Dilek