All pointer access into a mmap() arena should be wrapped by sigtrap, in
case the kernel generates a SIGBUS (oom, eio, bugs, etc). Add a couple
more missing annotations.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When we mark the GPU as wedge and disable acceleration, always check if
there is a GPU error state and warn if so - it is increasingly rare to
get a pure EIO from throttling thereby circumventing the current error
state checker.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Anytime we access a mmap() we need to be prepared for the kernel to send
us a SIGBUS, but we were missing a few sigtraps around calls to
pixman_fill and pixman_blt.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
These are primarily to serve to detect when we expect to see corruption
as a result of a badly timed EIO rather than fatal, so downgrade the
assertion to a warn.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
We only need a single copy of the cursor image, from which we can create
all the cloned cursors.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
With a MST topology change, we can find outputs may disappear. We offered
a choice as to whether to simply disable them or completely remove them
from the listing of available outputs. However, clients never expected
that their operation on any output could trigger a BadID XError (or that
there is anyway to prevent race conditions). As such an option was made to
disable by default, but allow testing complete removal. Now the RandR
protocol has been clarified such that XID assigned to outputs are now
permanent, as such the option breaks the spec, so drop it.
See randrproto commit 895ee5264524c7c239ee4ef5e39c4e295323fb51
Author: Dave Airlie <airlied@redhat.com>
Date: Wed Apr 22 10:58:18 2015 +1000
randrproto: clarify output XID lifetimes.
This just makes a note that randr won't make outputs disappear
dynamically.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Dave Airlie <airlied@redhat.com>
Backlight helper PID is set to -1 by default, if for some reason it's
not set, we may end up with waitpid(-1, ...) which will hang forever.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=90230
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
A couple of stray set-purgeable points were causing eventual assertions
that we didn't try and set-purgeable twice.
Minor regression from
commit 73ae4197d6
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Fri Apr 24 14:36:50 2015 +0100
sna: Defer marking cache objects purgeable
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
../../../src/uxa/intel_driver.c: At top level:
../../../src/uxa/intel_driver.c:769:1: error: unknown type name 'bool'
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
gcc generates an error at build time because it fails to inline some
functions:
blt.c: In function 'affine_blt':
blt.c:980:1: error: inlining failed in call to always_inline
'bilinear_weight': optimization level attribute mismatch
bilinear_weight(pixman_fixed_t x)
blt.c:1164:7: error: called from here
fy = bilinear_weight(y1);
^
blt.c:980:1: error: inlining failed in call to always_inline
'bilinear_weight': optimization level attribute mismatch
bilinear_weight(pixman_fixed_t x)
blt.c:1163:7: error: called from here
fx = bilinear_weight(x1);
^
blt.c:989:1: error: inlining failed in call to always_inline
'bilinear_interpolation': optimization level attribute mismatch
bilinear_interpolation(uint32_t tl, uint32_t tr,
^
blt.c:1207:11: error: called from here
b[i] = bilinear_interpolation(tl, tr, bl, br, fx, fy);
^
Do not force inlining of these functions and let the compiler decide to
avoid the compilation failure.
v2: fix up the other two force_inlines whilst we are here
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
In a couple of the copy paths, src/dst parameters were reversed,
inverting the direction of the tiling flags.
Reported-by: Andy Furniss <adf.lists@gmail.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=90138
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Bspec says it is an override for the source format, irrespective of
the tiling bit supplied in the command packet. So we need to apply the
FLUSH+LRI more often.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Somewhere, somehow the first glyph character upload into the Y-tiled
glyph cache is corrupt since
commit 75d8776247
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Fri Apr 17 10:29:28 2015 +0100
sna: Enable blitting with Y-tiled surfaces
I have not spotted whether it is a missing magic bit or a bad GTT path,
but ideally we don't want to switch to BLT along that path anyway (since
we are in the middle of rendering with the other engine).
Reported-by: Andy Furniss <adf.lists@gmail.com>
References: https://bugs.freedesktop.org/show_bug.cgi?id=90138
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If the pixmap's GPU bo is already allocated with Y, expect Mesa to deal
with it rather than convert to X.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Since the first flip may require a fixup, we will only really know with
the second flip whether we can start a flip chain with different tiling.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If we receive an error from the pageflip ioctl, we attempt to set the
framebuffer using SETCRTC instead. This is so that we can transition
between pitches and offsets for a CRTC. However, the kernel may reject a
flip also for tiling changes, or for any reason, which we also want to
transparently handle. Rearrange the error message and flip->blit
fallback handling to only do so if we try to flip twice (i.e. after
resetting the CRTC) and we still fail.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Let the read know how to compile dri3info without having to compile the
whole of xf86-video-intel
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Since gen6, there has been a magic register bit to change the
interpretation of the tiling mode between X and Y for BLT operations.
With the advent of DRI3 and scanouts supporting Y, enabling support at
last appears interesting, perhapss even by default for non-scanouts?
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As we treat it as a name we should capitilize each word, but we forgot
that rule (sometimes!) for Broadwell.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
According to BSpec the max number of VS URB entries for CHV is 640.
Based on a patch by Ville Syrjälä.
v2: Start using a GT info block like earlier generations.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com> [v1]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Use the correct register (Yn_01) with first half of the
Y samples instead of using the register (Yn_23) with the
second half twice when computing the green channel.
Also use the Yn_01 register name instead of Yn for the red
channel as well, just for a bit of extra consistency.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89807
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Tested-by: Chris Wilson <chris@chris-wilson.co.uk>
If at first you don't succeed, try, try again.
The kernel API is lame and requires a new interface for something that
is designed to be supported through the old interface...
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If at first you don't success, try, try again.
The kernel API is lame and requires a new interface for something that
is designed to be supported through the old inteface...
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Since creating/destroying the mmu_notifier for sync userptr is fairly
expensive, leak the test userptr object to keep it around for the
lifetime of our process.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Currently we add fake modes to cater for applications that want to
change the screen size but limit themselves to only those that are
available, and do not wish to either use RandR scaling or their own. For
widescreen panels, these 4:3 aspect ratio modes require panel fitting
and so are less than satisfory. Since we add fake modes, let's add a few
common widescreen modes as well (choosing the right aspect ration for
the panel).
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=37858
References: https://bugs.freedesktop.org/show_bug.cgi?id=29890
References: https://bugs.freedesktop.org/show_bug.cgi?id=89996
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Sometimes we want to render with the GPU when doing a mode switch, e.g.
if we need to initialise an output surface. To do so, we need to prepare
the acceleration layer first.
Lots of thanks to Vasily for tracking this one down.
Reported-by: Vasily Khoruzhick <anarsoul@gmail.com>
Bugzilla: https://bugs.archlinux.org/task/43534
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
An oversight when enabling efficient sharing of the shadow_crtc (i.e.
clients flipping to one monitor).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
When using Xinerama, RandR is automatically disabled, and calling RR
routines will trigger an assert() because the RR keys/resources are
not set, leading to an Xserver abort.
Hotplug makes little sense without RandR, so no need to install a
udev monitor if RandR is not available, as done in sna.
v3: Rebase onto backwards compatible rrPrivKey detection [ickle]
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Since dixPrivateKeyRegistered was only introduced in 1.10, we need an
alternative probe for older Xservers.
Suggested-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Chris Wilson
Olivier Fourdan
Ville Syrjälä