Since these require non-upstream patches to other components, we don't
want it enabled by default and randomly breaking builds.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
An inconsistency highlighted by 7c51cabaec revealed that we had a
mismatch between the check in move_to_gpu() and how we created the
pixmap. This mismatch resulted in us creating and uploading tiled
pixmaps for single shot textures, and the increase aperture pressure was
causing a regression in firefox-fishbowl on pnv, for example.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As ErrorF/fprintf is not re-entrant due to its mutex we can not use DBG
from code that could be called by a signal handler. X's SIGIO handler
attempts to move the cursor from within the handler (eek!) and so we
need to be careful not to take any locks, such in as the aforementioned
fprintf, along the cursor paths.
Reported-by: Zdenek Kabelac <zkabelac@redhat.com>
References: https://bugs.freedesktop.org/show_bug.cgi?id=50744
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Even with a 1nm process, I doubt we will see 4+GiB cache sizes ;-)
Reported-by: Zdenek Kabelac <zkabelac@redhat.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
All SourcePictures are now converted into Drawables, which had been
assumed by the driver backend. However, the code still existed to
attempt to pass procedural pictures onwards and so set pSrcPix to NULL
which was being flagged by the static analyser as a potential NULL
dereference.
Reported-by: Zdenek Kabelac <zkabelac@redhat.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If the intel_pixmap was NULL we should have failed to create the DRI2
buffer, so we can safely assert here to keep the analyser quiet.
Reported-by: Zdenek Kabelac <zkabelac@redhat.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If the target drawable is too large for the render pipeline, we need to
create a temporary surface. This may fail, so abort if it does.
Reported-by: Zdenek Kabelac <zkabelac@redhat.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Static analysers are dumb and presume readers are too.
Reported-by: Zdenek Kabelac <zkabelac@redhat.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
In order to prevent a leak of the bo when the chain is eventually
torn-down when the client exits.
Reported-by: Andreas Lampersperger <lampersperger.andreas@heidenhain.de>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=50670
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Time to blt from GTT to LLC 16384 bytes: 125.000µs (snb)
Time to blt from GTT to LLC 16384 bytes: 71.000µs (ivb)
Time to blt from GTT to LLC 1048576 bytes: 1400.000µs (snb)
Time to blt from GTT to LLC 1048576 bytes: 938.000µs (ivb)
Time to copy from GTT to LLC 16384 bytes: 118.000µs (snb)
Time to copy from GTT to LLC 16384 bytes: 134.000µs (ivb)
Time to copy from GTT to LLC 1048576 bytes: 6723.000µs (snb)
Time to copy from GTT to LLC 1048576 bytes: 7424.000µs (ivb)
And conversely,
Time to blt from LLC to GTT 16384 bytes: 10.000µs (snb)
Time to blt from LLC to GTT 16384 bytes: 8.000µs (ivb)
Time to blt from LLC to GTT 1048576 bytes: 217.000µs (snb)
Time to blt from LLC to GTT 1048576 bytes: 135.000µs (ivb)
Time to copy from LLC to GTT 16384 bytes: 4.000µs (snb)
Time to copy from LLC to GTT 16384 bytes: 4.000µs (ivb)
Time to copy from LLC to GTT 1048576 bytes: 270.000µs (snb)
Time to copy from LLC to GTT 1048576 bytes: 179.500µs (ivb)
It seems clear then that even with the extra synchronisation cost
copying from the GTT is much preferable with the GPU than using the
uncached reads by the CPU. Streaming write-combines from the CPU into
the GTT seem about as efficient as we can manage, so continue to use the
mapping unless busy.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Regression from commit 3f3bde4f0c
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Thu May 24 11:58:46 2012 +0100
uxa: Only consider an output valid if the kernel reports it attached
When backporting from SNA, a key difference that UXA does not track DPMS
state in its enabled flag and that a DPMS off CRTC is still bound to the
fb. So we do need to rescan the outputs and check that we have a
connector enabled *and* the pipe is running prior to emitting a scanline
wait.
References: https://bugs.freedesktop.org/show_bug.cgi?id=50668
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
These went away in the new server API, and really if this
made any measurable difference, I'd be impressed.
Signed-off-by: Dave Airlie <airlied@redhat.com>
We may as well continue the operation on the CPU if already bound to try to
reduce needless migrations (where the cost of the upload will outweigh
further use).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This is already checked at the beginning of PutImageTextured, so this
check upon the return value of intel_get_pixmap_bo() should only be
required to keep static analysers happy.
Reported-by: Zdenek Kabelac <zdenek.kabelac@gmail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This was written when the RGBA color value was being passed around and
not the pointer to the xRenderColor. As such, the NULL deref check
doesn't gain much and the check for rgba==0 irrelevant in this scenario.
Reported-by: Zdenek Kabelac <zkabelac@redhat.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
And assert that the allocation succeeds for good measure. A great big
dollop of graceful failure is missing...
Reported-by: Zdenek Kabelac <zkabelac@redhat.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Since the static analyser also doesn't like comparing a uint8_t against
~0.
Reported-by: Zdenek Kabelac <zkabelac@redhat.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Chris Wilson
Dave Airlie