[PATCH i-g-t] tests/intel/xe_exec_store: fix sync usage

[Matthew Brost]

On Mon, Apr 08, 2024 at 07:09:05PM +0100, Matthew Auld wrote:
> On 08/04/2024 18:55, Matthew Brost wrote:
> > On Mon, Apr 08, 2024 at 06:41:13PM +0100, Matthew Auld wrote:
> > > If using async binds it looks like an in-fence for the exec is needed to
> > > ensure the exec happens after the out-fence from the binds are complete.
> > > Therefore we need to unset DRM_XE_SYNC_FLAG_SIGNAL after doing the
> > > binds, but before the exec, otherwise the sync is rather treated
> > > as an out-fence and the binds can then happen after the exec, leading to
> > > various failures. In addition it looks like async unbind should be
> > > waited on before tearing down the queue/vm which has the bind engine
> > > attached, since the scheduler timeout is immediately set to zero on
> > > destroy, which might then trigger job timeouts. However it looks like
> > > it's also fine to rather just destroy the object and leave KMD to unbind
> > > everything itself. Update the various subtests here to conform to this.
> > > 
> > > In the case of the persistent subtest it looks simpler to use sync
> > > vm_bind since we don't have another sync for the in-fence at hand, plus
> > > we don't seem to need a dedicated bind engine.
> > > 
> > > Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/1270
> > > Signed-off-by: Matthew Auld <matthew.auld at intel.com>
> > > Cc: Matthew Brost <matthew.brost at intel.com>
> > 
> > Changes LGTM but IMO we jus delete this test as I'm unsure what coverage
> > this test is providing.
> 
> Do you mean just delete the entire xe_exec_store?
> 

Yea, not sure how this test ever got added as basically every test in
the suite does a dword store. Probably along the lines of lets port i915
tests to Xe! Not seeing a ton of value in this one... Probably above my
pay grade to make those types od decessions though.

Matt

> > 
> > Anyways:
> > Reviewed-by: Matthew Brost <matthew.brost at intel.com>
> > 
> > > ---
> > >   tests/intel/xe_exec_store.c | 22 ++++++++++++----------
> > >   1 file changed, 12 insertions(+), 10 deletions(-)
> > > 
> > > diff --git a/tests/intel/xe_exec_store.c b/tests/intel/xe_exec_store.c
> > > index c57bcb852..728ce826b 100644
> > > --- a/tests/intel/xe_exec_store.c
> > > +++ b/tests/intel/xe_exec_store.c
> > > @@ -102,13 +102,13 @@ static void persistance_batch(struct data *data, uint64_t addr)
> > >    */
> > >   static void basic_inst(int fd, int inst_type, struct drm_xe_engine_class_instance *eci)
> > >   {
> > > -	struct drm_xe_sync sync = {
> > > -		.type = DRM_XE_SYNC_TYPE_SYNCOBJ,
> > > -		.flags = DRM_XE_SYNC_FLAG_SIGNAL,
> > > +	struct drm_xe_sync sync[2] = {
> > > +		{ .type = DRM_XE_SYNC_TYPE_SYNCOBJ, .flags = DRM_XE_SYNC_FLAG_SIGNAL, },
> > > +		{ .type = DRM_XE_SYNC_TYPE_SYNCOBJ, .flags = DRM_XE_SYNC_FLAG_SIGNAL, }
> > >   	};
> > >   	struct drm_xe_exec exec = {
> > >   		.num_batch_buffer = 1,
> > > -		.num_syncs = 1,
> > > +		.num_syncs = 2,
> > >   		.syncs = to_user_pointer(&sync),
> > >   	};
> > >   	struct data *data;
> > > @@ -122,7 +122,8 @@ static void basic_inst(int fd, int inst_type, struct drm_xe_engine_class_instanc
> > >   	uint32_t bo = 0;
> > >   	syncobj = syncobj_create(fd, 0);
> > > -	sync.handle = syncobj;
> > > +	sync[0].handle = syncobj_create(fd, 0);
> > > +	sync[1].handle = syncobj;
> > >   	vm = xe_vm_create(fd, 0, 0);
> > >   	bo_size = sizeof(*data);
> > > @@ -134,7 +135,7 @@ static void basic_inst(int fd, int inst_type, struct drm_xe_engine_class_instanc
> > >   	exec_queue = xe_exec_queue_create(fd, vm, eci, 0);
> > >   	bind_engine = xe_bind_exec_queue_create(fd, vm, 0);
> > > -	xe_vm_bind_async(fd, vm, bind_engine, bo, 0, addr, bo_size, &sync, 1);
> > > +	xe_vm_bind_async(fd, vm, bind_engine, bo, 0, addr, bo_size, sync, 1);
> > >   	data = xe_bo_map(fd, bo, bo_size);
> > >   	if (inst_type == STORE)
> > > @@ -149,12 +150,14 @@ static void basic_inst(int fd, int inst_type, struct drm_xe_engine_class_instanc
> > >   	exec.exec_queue_id = exec_queue;
> > >   	exec.address = data->addr;
> > > -	sync.flags &= DRM_XE_SYNC_FLAG_SIGNAL;
> > > +	sync[0].flags &= ~DRM_XE_SYNC_FLAG_SIGNAL;
> > > +	sync[1].flags |= DRM_XE_SYNC_FLAG_SIGNAL;
> > >   	xe_exec(fd, &exec);
> > >   	igt_assert(syncobj_wait(fd, &syncobj, 1, INT64_MAX, 0, NULL));
> > >   	igt_assert_eq(data->data, value);
> > > +	syncobj_destroy(fd, sync[0].handle);
> > >   	syncobj_destroy(fd, syncobj);
> > >   	munmap(data, bo_size);
> > >   	gem_close(fd, bo);
> > > @@ -232,7 +235,7 @@ static void store_cachelines(int fd, struct drm_xe_engine_class_instance *eci,
> > >   		batch_map[b++] = value[n];
> > >   	}
> > >   	batch_map[b++] = MI_BATCH_BUFFER_END;
> > > -	sync[0].flags &= DRM_XE_SYNC_FLAG_SIGNAL;
> > > +	sync[0].flags &= ~DRM_XE_SYNC_FLAG_SIGNAL;
> > >   	sync[1].flags |= DRM_XE_SYNC_FLAG_SIGNAL;
> > >   	sync[1].handle = syncobjs;
> > >   	exec.exec_queue_id = exec_queues;
> > > @@ -250,7 +253,6 @@ static void store_cachelines(int fd, struct drm_xe_engine_class_instance *eci,
> > >   	for (i = 0; i < count; i++) {
> > >   		munmap(bo_map[i], bo_size);
> > > -		xe_vm_unbind_async(fd, vm, 0, 0, dst_offset[i], bo_size, sync, 1);
> > >   		gem_close(fd, bo[i]);
> > >   	}
> > > @@ -300,7 +302,7 @@ static void persistent(int fd)
> > >   			      vram_if_possible(fd, engine->instance.gt_id),
> > >   			      DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM);
> > > -	xe_vm_bind_async(fd, vm, 0, sd_batch, 0, addr, batch_size, &sync, 1);
> > > +	xe_vm_bind_sync(fd, vm, sd_batch, 0, addr, batch_size);
> > >   	sd_data = xe_bo_map(fd, sd_batch, batch_size);
> > >   	prt_data = xe_bo_map(fd, prt_batch, batch_size);
> > > -- 
> > > 2.44.0
> > >