cairo-trace already depended upon HAVE_FUNLOCKFILE for its
thread-safety.
[This is a candidate for 1.10.]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
All cairo_xcb_picture_t now get their own, private Picture instead of possibly
sharing it with a cairo_xcb_surface_t.
This solves a cyclic dependency which caused cairo_xcb_picture_t to be leaked
until their device was finished.
When an xcb surface was used as a source, a cairo_xcb_picture_t was created
which kept a reference to the surface. This picture was then added as a snapshot
to the surface which caused the surface to own a reference to the picture.
As a result, the Picture and possibly its associated Pixmap were not freed on
the X11 server which could thus run out of memory.
This change causes more Pictures to be created which could possibly slow down
cairo-xcb. There seems to be no measurable difference with cairo-perf-trace. For
the micro benchmarks, this has the most impact on paint_similar-rgba_source.512:
min(ticks) min(ms) median(ms) stddev. iterations overhead
before 158732 0.159 0.159 0.11% 4: 1651.49
after 162579 0.163 0.163 0.18% 4: 1612.41
Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=34912
Signed-off-by: Uli Schlachter <psychon@znc.in>
When running cairo-test-suite with "-a", backends are also tested with a
non-zero device offset. However, for "xcb-window&" the boilerplate incorrectly
overwrote the device offset with a zero offset again.
This caused all test results to be offset by 25 pixels which obviously makes all
tests fail.
Just removing the call to cairo_surface_set_device_offset solves the problem.
Signed-off-by: Uli Schlachter <psychon@znc.in>
This flag was only ever tested, but never set. Also, there is no X11 server out
there which actually supports this.
AFAIK it was just an experiment which turned into dead code.
Signed-off-by: Uli Schlachter <psychon@znc.in>
RENDER 0.11 added PDF operators and the cairo-xcb backend is already prepared
for handling them. However, it doesn't actually handle them.
If you used such an operator with RENDER 0.11, _operator_is_supported() would
return TRUE, but later on an ASSERT_NOT_REACHED in _render_operator() is hit.
Signed-off-by: Uli Schlachter <psychon@znc.in>
All of these defines where not used anymore and this '#if 0'd code was so for
ages and most likely would stay commented out for more ages, too.
Signed-off-by: Uli Schlachter <psychon@znc.in>
We never bound any other value to the sampler uniforms, so why bother
resetting them all the time?
[ 0] before firefox-talos-gfx 29.972 30.493 0.89% 3/3
[ 0] after firefox-talos-gfx 29.397 29.599 1.27% 3/3
Signed-off-by: Eric Anholt <eric@anholt.net>
Because of GLX's unfortunate requirement that only one context have a
thread current at a time, we had to unbind the context, triggering a
flush, and eating of all the CPU. With a small tweak to the GLX spec
by GLX_MESA_multithread_makecurrent, and a small tweak to Mesa that
consisted mostly of deleting the "is this thread already bound
elsewhere?" check, we can bind our context to any thread and use it
safely as long as our usage of it is mutexed, which cairo-gl should
already be doing.
[ 0] before firefox-talos-gfx 50.260 50.525 0.25% 3/3
[ 0] after firefox-talos-gfx 32.091 32.422 0.65% 3/3
Since its inception, cairo-gl has been plagued by the "how am I really
supposed to use it?" problem. This lays down my expectations for how
cairo-gl will interact with other usage of the GL API.
If PIXMAN_HAS_ATOMIC_OPS, ix and iy will be changed if the source
surface is of subsurface type. They need to be reset to 0 before being
passed to _cairo_matrix_to_pixman_matrix_offset() or the acquired
surface will have an incorrect offset .
The pixman_image_ref() and pixman_image_unref() functions
aren't threadsafe in current pixman, so we can't share
pixman images across contexts where they may end up
moving between threads. The main use of these functions
is in keeping a global cache of pixman_image_ts for
solid colours and reusing the same pixman_image_t for
patterns pointing to the same image data where possible.
This patch removes those uses from cairo by wrapping
them in a PIXMAN_HAS_ATOMIC_OPS conditional. There is
no automatic test this, so we will have to manually
turn this code back on when pixman does gain atomic
ops.
The only two remaining calls to pixman_image_ref() are
done on locally created images so they are safe.
Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=34177
In order to be able to make an egl context current when using a
pbuffer surface, that surface must have been created with a config
that is compatible with the context config. For Mesa, this means that
the configs must be the same.
Test case taken from the WebKit test suite, failure originally reported
by Zan Dobersek <zandobersek@gmail.com>. WebKit test is
LayoutTests/canvas/philip/tests/2d.path.rect.selfintersect.html
It is not correct to rely on the version defined in render.h. The
Xrender.h header is independent and might not define some functions
available in RENDER 0.10.
Their availability must be detected at configure time and the stubs
must be defined only if the functions are not available.
Fixes https://bugs.freedesktop.org/show_bug.cgi?id=31906
Draw/readbuffers are a property of the framebuffer, not of the
context. So we can leave them in place across bindings. I left the
window drawbuffer setting in place until we decide how to handle
interoperating of cairo-gl with normal GL usage.
[ 0] before firefox-talos-gfx 67.552 67.561 0.22% 3/3
[ 0] after firefox-talos-gfx 66.689 66.913 0.41% 3/3
This test creates a similar surface using cairo_push_group () and
assumes its size will be w x h. This is not true when testing similar
surfaces with "cairo-test-suite -a", because they are bigger than the
actual ROI, which corresponds to the rectangle (0,0) - (w,h).
Clipping to the ROI guarantees that the surface created by
cairo_push_group () always has the same size.
Fixes surface-pattern-scale-down-extend.
The test was incorrectly translating the "target" of the context
instead of the "group target" (i.e. the current destination).
"cairo-test-suite -a" runs the tests on similar surfaces created using
cairo_push_group (), thus without this change the device-offset did
not affect the real destination surface.
Fixes clip-device-offset.
This test called cairo_save () without restoring again. Normally, this doesn't
cause a problem. However, when cairo-test-suite is called with "-a", it also
executes the tests on a similar surface. This then caused
cairo_pop_group_to_source to be called on a cairo_save state which causes a
CAIRO_STATUS_INVALID_RESTORE since commit 5d95ae924e.
Signed-off-by: Uli Schlachter <psychon@znc.in>
Whenever a surface is used as a source, cairo-xcb creates a Picture for it and
adds it as a snapshot to the source. Additionally, there is a per-screen cache
that makes sure old Pictures are destroyed when more than 16 MiB of image data
is used.
There are two problems here:
- Normally, the Picture would be destroyed once it is no longer a snapshot for
the original source surface. However, the cache kept it alive for longer which
means it was using memory but couldn't actually be used any more. (The
Pictures are found via _cairo_surface_has_snapshot)
- When a xcb surface is used as a source, we create a new Picture for it, but
when that surface is cairo_surface_finish'd, we don't FreePicture this
Picture. When the drawable was a window, this may later on cause a BadPicture
error when cairo tries to explicitly free a picture that was already
implicitly freed when its window was destroyed.
The second point caused lots of tests in the test suite to fail on an assert due
to the above-mentioned BadPicture error, but just removing the cache makes both
these problems go away.
The only downside is that the cache enforced a limit of 16 MiB of pixel data
that was used for source surfaces on the server. After this commit the Picture
will be kept alive until the snapshot is detached. If this becomes a problem
memory-wise, a new solution will have to be invented...
Signed-off-by: Uli Schlachter <psychon@znc.in>
The xcb backend uses caches for linear, radial and surface patterns so that it
doesn't have to create a new cairo_xcb_picture_t on every use. Whenever any
picture is finished, the surface pattern cache is checked on whether it still
contains that picture.
This means that we cannot destroy the surface_pattern_cache until after the
linear and radial_pattern_cache are destroyed or else we'll be querying the
surface pattern cache after it was finished.
This wasn't noticed before because _cairo_xcb_picture_finish couldn't acquire
the cairo_xcb_connection_t and thus failed earlier. The previous commit changed
this and now lots of tests in the test suite trigger the crash.
Signed-off-by: Uli Schlachter <psychon@znc.in>
The problem is that the finish callback might still have to work with the
device, so it can't be made unusable yet. This is in contrast to what
cairo_surface_finish() does, but when finishing a surface it's quite unlikely
that its cairo_surface_t still has to be passed around to functions outside of
the backend.
This "fixes" some problems with the xcb-surface-source test where
_cairo_surface_snapshot()-style snapshots break when the underlying
cairo_device_t is finished. (In the following backtrace, cairo_device_acquire
fails because the device's ->finished member is already TRUE)
#0 cairo_device_acquire (device=0x690f70) at cairo-device.c:414
#1 0x00007ffff7b884fb in _cairo_xcb_connection_acquire (surface=0x69c610, use_shm=<value optimized out>, image_out=0x7fffffffcc58) at cairo-xcb-private.h:246
#2 _get_image (surface=0x69c610, use_shm=<value optimized out>, image_out=0x7fffffffcc58) at cairo-xcb-surface.c:425
#3 0x00007ffff7b893cf in _cairo_xcb_surface_acquire_source_image (abstract_surface=0x69c610, image_out=0x7fffffffcce8, image_extra=0x7fffffffcce0)
at cairo-xcb-surface.c:561
#4 0x00007ffff7b601d4 in _cairo_surface_acquire_source_image (surface=0x690f70, image_out=0x7fffffffcce8, image_extra=0x0) at cairo-surface.c:1458
#5 0x00007ffff7b65dcf in _cairo_surface_snapshot_copy_on_write (surface=<value optimized out>) at cairo-surface-snapshot.c:125
#6 0x00007ffff7b609b1 in _cairo_surface_detach_snapshot (snapshot=0x690f70) at cairo-surface.c:330
#7 0x00007ffff7b606dc in _cairo_surface_detach_snapshots (surface=0x69c610) at cairo-surface.c:315
#8 cairo_surface_flush (surface=0x69c610) at cairo-surface.c:1126
#9 0x00007ffff7b6074d in cairo_surface_finish (surface=0x690f70) at cairo-surface.c:718
#10 0x00007ffff7b881ff in _cairo_xcb_screen_finish (screen=0x691920) at cairo-xcb-screen.c:61
#11 0x00007ffff7b8629c in _device_finish (device=0x690f70) at cairo-xcb-connection.c:546
#12 0x0000000000446f7c in cleanup (data=0x69c370) at xcb-surface-source.c:47
#13 0x00007ffff7b1e56a in _cairo_user_data_array_fini (array=0x693628) at cairo-array.c:390
#14 0x00007ffff7b60868 in cairo_surface_destroy (surface=0x693600) at cairo-surface.c:651
#15 0x0000000000447599 in draw (cr=0x7ffff7ddc288, width=<value optimized out>, height=<value optimized out>) at surface-source.c:149
#16 0x000000000040c1c3 in cairo_test_for_target (ctx=0x7fffffffe640, target=0x669668, dev_offset=<value optimized out>, similar=<value optimized out>)
at cairo-test.c:984
#17 0x000000000040d997 in _cairo_test_context_run_for_target (ctx=<value optimized out>, target=<value optimized out>, similar=<value optimized out>,
dev_offset=<value optimized out>) at cairo-test.c:1617
#18 0x000000000040ee09 in _cairo_test_runner_draw (argc=<value optimized out>, argv=<value optimized out>) at cairo-test-runner.c:269
#19 main (argc=<value optimized out>, argv=<value optimized out>) at cairo-test-runner.c:924
Signed-off-by: Uli Schlachter <psychon@znc.in>
