Increate the GPU load slightly, but mitigate the CPU overhead from
clflushing by keeping the object alive using flink. This has also the
side-effect of magnifying the desired busy-close race.
References: https://bugs.freedesktop.org/show_bug.cgi?id=71029
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
The challenge here is to race GPU activity versus the gem_close. Using a
large objects makes the window of GPU activity larger - except on !llc
systems we then incur massive overhead from clflush, likely destroying
any race (due to mutex contention). And that mutex contention is
amplified by the number of cores - bad news for Baytrail.
Give up and make the objects smaller and hope that the test is run
frequently enough to catch the race.
References: https://bugs.freedesktop.org/show_bug.cgi?id=71029
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
With Full PPGTT, each new fd creates a new context and thus a new
PPGTT, so we have to reduce the number of simultaneous fds or face
OOM problems. For every new PPGTT, its PDEs are stored in the GGTT
which imposes a limit of 1024 new contexts. We want to leave at
least 1/4 of the GGTT available for "important" stuff like scanout
buffers, so never open more than 768 fds.
v2: Unconditionally use the lower limit, as suggested by Daniel
Vetter
Signed-off-by: Oscar Mateo <oscar.mateo@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
In the past new testcases with subtest often forgot to add the call to
igt_exit at the end of their main() function. That is now caught with
a bit more obnoxious asserts, but it's still a nuissance.
This little igt_main macro takes care of that (and also of calling the
subtest machinery initialization code correctly).
If no one objects I'll roll this out for all the simple cases (i.e.
those tests that don't have additional argv parsing on top of the
subtest machinery).
v2: Roll it out across the board.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
