This exercises both the wc mmappings and the extended get_tiling ioctl.
Userspace cannot handle bit17 swizzling through wc mmaps (because bit17
requires swizzling based on the actual physical address of the page -
which is unknown to userspace) and so we need an extended get_tiling
ioctl to report the actual as well as the logical swizzling on an
object. We then check that the contents of the object are tiled and
swizzled correctly when viewed through a wc mmap.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
We already allocate enough objects to thrash the ppGTT VMs, so allow us
to reuse the batch buffers for some efficiency gains and through the
contention more towards the ctx->vm.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Play nice, especially with the subtest, and wait for the children to
exit before finishing the test. If we don't we end up with a fork bomb
for some unknown reason...
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Now that we are no longer busy-spinning inside random(), we can spend
more time exercising i915.ko
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Just to make sure this keeps working since a patch to WARN_ON using
dumb buffers in execbuf was accidentally merged.
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
The android makefiles were passing the -std=c99 flag to the
compiler which disables the typeof keyword. This causes a
build fail for a recent addition to igt_aux.h.
Change this to -std=gnu99, which is the flag used in the
linux build
Signed-off-by: Tim Gore <tim.gore@intel.com>
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
Add a function to lock memory into RAM and use it in the
gem_tiled_swapping test to reduce the amount of allocated memory
required to force swapping. This also reduces the amount of time
required for the test to complete, since the data set is smaller.
The following durations were recorded with gem_tiled_swapping on a
haswell system before the change:
Subtest non-threaded: SUCCESS (55.889s)
Subtest threaded: SUCCESS (810.532s)
and after:
Subtest non-threaded: SUCCESS (11.804s)
Subtest threaded: SUCCESS (268.336s)
v2: add various assertions and requirements and make sure
gem_tiled_swapping works on systems with less RAM (Daniel Vetter)
v3: fix allocation size calculation
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
Use images with a smaller file size, created by saving the left and
right stereo output of glxgears.
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
The valid pipe/connector combo check was lost in
commit 57259d714d3fe1170cf931af72648219856a9918
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date: Mon Nov 24 16:08:32 2014 +0100
lib/igt_debugfs: Don't setup crc in _new
Restore it to make the test again useful on CHV.
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Report SKIP only if none of the pipe/connector combos worked, instead of
trying to report for each pipe separately.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Aim the dummy load to the current back buffer instead if the front
buffer. Assuming the idea is to get the next flip to be stuck behind
the dummy load?
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Try to tune the dummy load to ~1 second. The calibration happens the
first time dummy load is generated.
v2: Actually do the number of ops intended and
calibrate to 1 second and not 2
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Make the dummy load independent of the display resolution by using a
two fixed size dummy bos to generate the load. As a final step do
another copy from one of the dummy bos to the fb to make sure there's
a dependency between the dummy load and any subsequent operation on
the fb.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
When setting the calculated middle frequency value the test assumes that
the HW/kernel rounds this value according to a 50MHz step value. This is
not so at least on VLV/CHV, on my B0 BYT-M for example this step value
is 22MHz, so there the test will fail.
To fix this get the nearest supported value by setting the target
frequency as a min or max frequency and read it back. The kernel will
round the returned value to the nearest supported.
v2:
- remove the 50MHz rounding that was done for non-VLV platforms, the new
way of rounding should provide the correct value for all platforms
(Ville)
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Imre Deak <imre.deak@intel.com>
When changing the sysfs GT min/max frequencies, the kernel won't
explicitly change the current frequency, unless it becomes out of bound
based on the new min/max values. The test happens to work on non-VLV
platforms because on those the kernel resets the current frequency
unconditionally (to adjust the RPS interrupt mask as a side-effect) and
that will lead to an RPS interrupt setting the minimum frequency.
To fix this load the GPU after decreasing the min frequency and before
checking the current frequency. This should set the current frequency to
the minimum.
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Imre Deak <imre.deak@intel.com>
At least on VLV when forcing a new GT frequency by writing to the
min/max freq sysfs entries the kernel doesn't wait until the new
frequency settles, so the subsequent readback check might fail. To fix
this wait until the current frequency is between the min/max values
using a 10ms timeout.
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Imre Deak <imre.deak@intel.com>
As we require a pipe enabled to generate vblanks, the first step is to
then to check that pipe 0 is active or else skip the test.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Check that the more obvious userspace error conditions are handled by
the kernel, ideally without loss of data. These include nonblocking
waits, passing invalid buffers and passing buffers of the incorrect
length.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Here is a cheap way for this test to give consistent results. This
doesn't change the usefulness of this test, hopefully.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=85270
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
gem_gpgpu_fill was added in commit 4ec8479 (tests: Add gem_gpgpu_fill),
but wasn't added to .gitignore.
Cc: Zhenyu Wang <zhenyuw@linux.intel.com>
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
This is simply a copy of gem_media_fill but using new
GPGPU fill operation.
v2: Use general fill func pointer.
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
This test will not run on Android as the coreu service
remains running even after the android system is stopped.
Coreu is a client of drm and when the test finds this it
fails an assert.
Coreu is started by the init process and there is no
tidy, non invasive way to stop it (init just restarts it).
Coreu isn't doing anything and would not be expected to
interfere with this test. In addition, all the other
igt tests just rely on the user/test script to ensure
that there are no other drm clients, so this test can
do the same. On Android we must rely on coreu being
dormant when this test runs.
Signed-off-by: Tim Gore <tim.gore@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
kms_pwrite_crc was recently added and requires cairo, so
add this to the list of tests to exclude if cairo is not
avaiable
Signed-off-by: Tim Gore <tim.gore@intel.com>
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
If we wait for one vblank, we may end up returning almost immediately,
so trying to assert anything but >0 about the minimum duration is
bogus.
Instead wait for two vblanks and then we can assert that we should have
be blocked for at least one frame. And move the upper bound to a little
over two frames to match.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=79050
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
gem_ring_sync_copy uses a lot of memory and gets OOM killed on smaller
systems (eg android devices). Most of the allocation is for "busy work"
to keep the render rings busy and for this we can just re-use the same
few buffers over and over. This enables the test to be run on low end
devices.
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Tim Gore <tim.gore@intel.com>
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
The problem is that this causes writes to registers, and if the pipe
is off they might go nowhere (e.g. when runtime pm is enabled).
Furthermore we can only really check once the modeset setup is done,
but again most tests set up the CRC structure before calling
igt_commit and friends. We could add crc restore support to the
kernel's rpm code, but that will end up being rather invasive and
fragile hard-to-test code.
Now originally this was needed back when CRC support wasn't available
everywhere. But that's fixed now.
So given all this just drop that sanity check and make sure that we
only touch the debugfs file (and so the hw state) when we know the
pipe is running in the desired configuration.
A complementary kernel patch will try to catch offenders by returning
-EIO if the pipe is off.
v2: Forgot to git add one hunk.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86092
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
drm_open_any keeps a buffer handle around for the cleanup sync work,
so we can only grab the buffer count after the latst drm_open_any
call. Otherwise we'll detect a fake leak.
This broke in
commit 2f2c491cf3167befe7c79e4b17afb4f6284dfc84
Author: Mika Kuoppala <mika.kuoppala@intel.com>
Date: Fri Mar 28 10:52:46 2014 +0200
lib/drmtest: don't dup quiescent fd
since that additional open drm fd keeps a gem object for the default
context around. Hence why this also only blows up on gen6+ - earlier
platforms don't have hw context support.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=79821
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
This has the upside that we'll never forget to add it to thrashing
tests. But we'll also never miss to move it when adding basic
functionality tests to existing binaries. Chris already started this
refining work in e.g.
commit d77eda6614a1955717f224be023dedf74eb7735d
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Fri Nov 14 07:45:40 2014 +0000
igt/gem_linear_blits: Require that we do the full test
by moving igt_skip_on_simulation into subtests.
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
More in line with the usual igt pattern and simplifies the code -
every called just wrapped it in igt_require.
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
This triggers a device suspended WARN in the kernel in
gen6_ggtt_insert_entries() while calling the GEM pwrite ioctl.
The sequence is suggested by Daniel.
Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Add a test to verify that pwriting to a future scanout buffer works
correctly. The specific problem occurs when the buffer is already
UC/WT before the pwrite, not the current scanout buffer, and not
currently in the CPU write domain. With the buggy kernel no clflush
will be performed after the pwrite, and hence we end up with cache
dirt on the display.
The problem only affects LLC platforms (non-LLC would clflush anyway
after pwrite), but we can let the test run on all platforms.
v2: Fix typos in commit message and add to .gitignore
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
kms_flip/nonblocking-read will block indefinitely if it fails, so
introduce a timeout to indicate test failure.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=85718
Signed-off-by: Thomas Wood <thomas.wood@intel.com>
