The 'clear' subtest used to exercise correctness of object memory clearing
on passing a batch with the object to GPU for processing. However, commit
e25913a1a79d ("i915/gem_mmap_offset: Ignore ENOSPC error for making
residency execbuf"), while resolving an issue of unnecessary failures on
ENOSPC errors, introduced an alternative method of clearing the object
memory, with random selection of one of those methods on each iteration.
The new method expects the memory to be cleared in pagefault handler path.
Since those two methods may give different results, mixing them in one
exercise may make the picture unclear, especially if something goes wrong.
The exercise is executed in several parallel threads, one per CPU. Each
thread repeats the exercise in a time only limited loop, with no delay
between consecutive iterations. In case of passing system memory objects
to execbuf, that happens to exhaust all available physical memory, which
is neither the matter nor the goal nor requirement of the exercise.
Move the pagefault method, free from the issue, to a new separate subtest,
and make sure sufficient amount of physical memory is available before
calling another execbuf.
v2: Limit the scope of the fix to SMEM exercise.
Janusz Krzysztofik (2):
tests/gem_mmap_offset: Split 'clear' subtest
tests/gem_mmap_offset: Fix OOM hits
tests/intel/gem_mmap_offset.c | 22 ++++++++++++++++++----
1 file changed, 18 insertions(+), 4 deletions(-)
--
2.47.0
