Quoting Chris Wilson (2018-02-14 19:29:29)
> Quoting Daniele Ceraolo Spurio (2018-02-14 19:18:26)
> > When an out-fence is returned we expect that the in-fence is not
> > overwritten. Add a test to check for that.
> >
> > Cc: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> > Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@xxxxxxxxx>
> > ---
> > tests/gem_exec_fence.c | 39 +++++++++++++++++++++++++++++++++++++++
> > 1 file changed, 39 insertions(+)
> >
> > diff --git a/tests/gem_exec_fence.c b/tests/gem_exec_fence.c
> > index bd7b1263..f7478c55 100644
> > --- a/tests/gem_exec_fence.c
> > +++ b/tests/gem_exec_fence.c
> > @@ -586,6 +586,42 @@ static void test_parallel(int fd, unsigned int master)
> > gem_close(fd, handle[0]);
> > }
> >
> > +/* check that a fence in doesn't get clobbered when a fence out is returned */
> > +static void test_keep_fence_in(int fd, unsigned int engine)
> > +{
> > + struct drm_i915_gem_execbuffer2 execbuf;
> > + struct drm_i915_gem_exec_object2 obj;
> > + const uint32_t bbe = MI_BATCH_BUFFER_END;
> > + int fence;
> > +
> > + memset(&obj, 0, sizeof(obj));
> > + memset(&execbuf, 0, sizeof(execbuf));
> > +
> > + obj.handle = gem_create(fd, 4096);
> > + gem_write(fd, obj.handle, 0, &bbe, sizeof(bbe));
> > +
> > + execbuf.buffers_ptr = to_user_pointer(&obj);
> > + execbuf.buffer_count = 1;
> > + execbuf.flags = engine | LOCAL_EXEC_FENCE_OUT;
> > +
> > + gem_execbuf_wr(fd, &execbuf);
> > + fence = execbuf.rsvd2 >> 32;
> > +
> > + gem_close(fd, obj.handle);
> > + obj.handle = gem_create(fd, 4096);
> > + gem_write(fd, obj.handle, 0, &bbe, sizeof(bbe));
>
> You don't need to recreate the batch here; just reusing the same handle
> will do (and avoid the extra complication).
>
> > +
> > + execbuf.flags |= LOCAL_EXEC_FENCE_IN;
> > + execbuf.rsvd2 = fence;
> > +
> > + gem_execbuf_wr(fd, &execbuf);
> > + igt_assert_eq(fence, execbuf.rsvd2 & 0xFFFFFFFF);
>
> This would be a good one to throw to the interruptible wolves.
Of course to actually interrupt it, we need to hit a wait. Hmm. I'd use
a variant of the measure ring size approach, scrap the second test, and
then repeatedly re-submit until it hit an EINTR. (Obviously requires a
timer, and updating/checking the fences on each pass).
-Chris
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
