Quoting Tvrtko Ursulin (2018-04-24 09:55:20)
>
> On 23/04/2018 18:08, Chris Wilson wrote:
> > Quoting Tvrtko Ursulin (2018-04-23 17:52:54)
> >>
> >> On 23/04/2018 14:43, Chris Wilson wrote:
> >>> In the existing ABI, each engine operates its own timeline
> >>> (fence.context) and so should execute independently of any other. If we
> >>> install a blocker on all other engines, that should not affect execution
> >>> on the local engine.
> >>>
> >>> Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> >>> Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
> >>> ---
> >>> tests/gem_exec_schedule.c | 90 +++++++++++++++++++++++++++++++++++----
> >>> 1 file changed, 82 insertions(+), 8 deletions(-)
> >>>
> >>> diff --git a/tests/gem_exec_schedule.c b/tests/gem_exec_schedule.c
> >>> index 5d0f215b2..471275169 100644
> >>> --- a/tests/gem_exec_schedule.c
> >>> +++ b/tests/gem_exec_schedule.c
> >>> @@ -49,9 +49,9 @@
> >>>
> >>> IGT_TEST_DESCRIPTION("Check that we can control the order of execution");
> >>>
> >>> -static void store_dword(int fd, uint32_t ctx, unsigned ring,
> >>> - uint32_t target, uint32_t offset, uint32_t value,
> >>> - uint32_t cork, unsigned write_domain)
> >>> +static uint32_t __store_dword(int fd, uint32_t ctx, unsigned ring,
> >>> + uint32_t target, uint32_t offset, uint32_t value,
> >>> + uint32_t cork, unsigned write_domain)
> >>> {
> >>> const int gen = intel_gen(intel_get_drm_devid(fd));
> >>> struct drm_i915_gem_exec_object2 obj[3];
> >>> @@ -100,7 +100,17 @@ static void store_dword(int fd, uint32_t ctx, unsigned ring,
> >>> batch[++i] = MI_BATCH_BUFFER_END;
> >>> gem_write(fd, obj[2].handle, 0, batch, sizeof(batch));
> >>> gem_execbuf(fd, &execbuf);
> >>> - gem_close(fd, obj[2].handle);
> >>> +
> >>> + return obj[2].handle;
> >>> +}
> >>> +
> >>> +static void store_dword(int fd, uint32_t ctx, unsigned ring,
> >>> + uint32_t target, uint32_t offset, uint32_t value,
> >>> + uint32_t cork, unsigned write_domain)
> >>> +{
> >>> + gem_close(fd, __store_dword(fd, ctx, ring,
> >>> + target, offset, value,
> >>> + cork, write_domain));
> >>> }
> >>>
> >>> static uint32_t create_highest_priority(int fd)
> >>> @@ -161,6 +171,64 @@ static void fifo(int fd, unsigned ring)
> >>> munmap(ptr, 4096);
> >>> }
> >>>
> >>> +static void independent(int fd, unsigned int engine)
> >>> +{
> >>> + IGT_CORK_HANDLE(cork);
> >>> + uint32_t scratch, plug, batch;
> >>> + igt_spin_t *spin = NULL;
> >>> + unsigned int other;
> >>> + uint32_t *ptr;
> >>> +
> >>> + igt_require(engine != 0);
> >>> +
> >>> + scratch = gem_create(fd, 4096);
> >>> + plug = igt_cork_plug(&cork, fd);
> >>> +
> >>> + /* Check that we can submit to engine while all others are blocked */
> >>> + for_each_physical_engine(fd, other) {
> >>> + if (other == engine)
> >>> + continue;
> >>> +
> >>> + if (spin == NULL) {
> >>> + spin = __igt_spin_batch_new(fd, 0, other, 0);
> >>> + } else {
> >>> + struct drm_i915_gem_exec_object2 obj = {
> >>> + .handle = spin->handle,
> >>> + };
> >>> + struct drm_i915_gem_execbuffer2 eb = {
> >>> + .buffer_count = 1,
> >>> + .buffers_ptr = to_user_pointer(&obj),
> >>> + .flags = other,
> >>> + };
> >>> + gem_execbuf(fd, &eb);
> >>> + }
> >>> +
> >>> + store_dword(fd, 0, other, scratch, 0, other, plug, 0);
> >>> + }
> >>> + igt_require(spin);
> >>> +
> >>> + /* Same priority, but different timeline (as different engine) */
> >>> + batch = __store_dword(fd, 0, engine, scratch, 0, engine, plug, 0);
> >>> +
> >>> + unplug_show_queue(fd, &cork, engine);
> >>> + gem_close(fd, plug);
> >>> +
> >>> + gem_sync(fd, batch);
> >>> + gem_close(fd, batch);
> >>
> >> Strictly speaking I think you need to use the poll-able spinner and wait
> >> on it here, before the busy assert. It's unlikely, but spinners on
> >> 'other' engines are getting submitted async to the store dword batch on
> >> 'engine'.
> >
> > We've waited for its completion, so we know batch is idle and the others
> > are still busy. We then check its seqno is written to the scratch; so
> > using pollable here is redundant. And then we check that the others are
> > run after.
>
> Yeah I was confused, thinking busy check on spinner could return false
> if the respective tasklet on those engines hadn't ran yet - but of
> course busy is true immediately after execbuf so as I said, total confusion.
So having checked the logic should produce N-1 blocked engines and
ourselves free to complete; care for the r-b so we can add it to the mix
of tests?
-Chris
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
