[igt-dev] [PATCH i-g-t] intel/xe_exec_compute_mode: Add non-blocking subtest

[Matthew Brost]

On Thu, Dec 07, 2023 at 12:35:03PM +0530, sai.gowtham.ch at intel.com wrote:
> From: Sai Gowtham Ch <sai.gowtham.ch at intel.com>
> 
> Fill the ring with maximum workload and expecte kernel to return
> -EWOULDBLOCK error.
> 
> Signed-off-by: Sai Gowtham Ch <sai.gowtham.ch at intel.com>
> ---
>  tests/intel/xe_exec_compute_mode.c | 95 ++++++++++++++++++++++++++++++
>  1 file changed, 95 insertions(+)
> 
> diff --git a/tests/intel/xe_exec_compute_mode.c b/tests/intel/xe_exec_compute_mode.c
> index 7d3004d65..56bd6cb49 100644
> --- a/tests/intel/xe_exec_compute_mode.c
> +++ b/tests/intel/xe_exec_compute_mode.c
> @@ -290,6 +290,97 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
>  		close(map_fd);
>  }
>  
> +/**
> + * SUBTEST: non-blocking
> + * Description: Fill the ring and check we get expected errors.
> + * Test category: functionality test
> + */
> +static void non_block(int fd, int expect)
> +{
> +	struct drm_xe_sync sync = {
> +		.type = DRM_XE_SYNC_TYPE_SYNCOBJ,

DRM_XE_SYNC_TYPE_SYNCOBJ this type shouldn't work with LR VMs. Does this
test pass? AFAIK this shouldn't work... (e.g. bind and execs should
return -EOPNOTSUPP here).

> +		.flags = DRM_XE_SYNC_FLAG_SIGNAL,
> +	};
> +
> +	struct drm_xe_exec exec = {
> +		.num_batch_buffer = 1,
> +		.num_syncs = 1,
> +		.syncs = to_user_pointer(&sync),
> +	};
> +	struct {
> +		uint32_t batch[16];
> +		uint64_t pad;
> +		uint32_t data;
> +		uint64_t addr;
> +	} *data;
> +	struct drm_xe_engine *engine;
> +	uint32_t vm, exec_queue, syncobj;
> +	size_t bo_size;
> +	int value = 0x123456;
> +	uint64_t addr = 0x100000;
> +	uint32_t bo = 0;
> +	int b, count, intr, err;
> +
> +	syncobj = syncobj_create(fd, 0);
> +	sync.handle = syncobj;
> +
> +	vm = xe_vm_create(fd, DRM_XE_VM_CREATE_FLAG_ASYNC_DEFAULT |
> +			      DRM_XE_VM_CREATE_FLAG_LR_MODE, 0);
> +	bo_size = sizeof(*data);
> +	bo_size = ALIGN(bo_size + xe_cs_prefetch_size(fd), xe_get_default_alignment(fd));
> +
> +	engine = xe_engine(fd, 1);
> +	bo = xe_bo_create(fd, vm, bo_size, vram_if_possible(fd, engine->instance.gt_id),
> +					   DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM);
> +
> +	xe_vm_bind_async(fd, vm, engine->instance.gt_id, bo, 0, addr, bo_size, &sync, 1);
> +	exec_queue = xe_exec_queue_create(fd, vm, &engine->instance, 0);
> +	data = xe_bo_map(fd, bo, bo_size);
> +
> +	count = 0;
> +	intr = 0;

The below loop doesn't make tons of sense. It would make a lot more
sense to issue a spin batch first and then a bunch of execs until
-EWOULDBLOCK is hit. Then release the spin batch, wait that batch to
complete, issue another exec and verify it works.

Matt

> +	do {
> +		uint64_t batch_offset = (char *)&(data->batch) - (char *)data;
> +		uint64_t batch_addr = addr + batch_offset;
> +		uint64_t sdi_offset = (char *) & (data->data) - (char *)data;
> +		uint64_t sdi_addr = addr + sdi_offset;
> +
> +		b = 0;
> +		data->batch[b++] = MI_STORE_DWORD_IMM_GEN4;
> +		data->batch[b++] = sdi_addr;
> +		data->batch[b++] = sdi_addr >> 32;
> +		data->batch[b++] = value;
> +		data->batch[b++] = MI_BATCH_BUFFER_END;
> +		igt_assert(b <= ARRAY_SIZE(data->batch));
> +
> +		exec.exec_queue_id = exec_queue;
> +		exec.address = batch_addr;
> +		sync.flags &= DRM_XE_SYNC_FLAG_SIGNAL;
> +
> +		err = __xe_exec(fd, &exec);
> +		igt_assert(syncobj_wait(fd, &sync.handle, 1,
> +					INT64_MAX, 0, NULL));
> +
> +		if (err == -EWOULDBLOCK) {
> +			if (intr != count)
> +				err = 0;
> +			intr = count + 1;
> +		}
> +		if (err)
> +			break;
> +		count++;
> +	} while (1);
> +
> +	igt_assert_eq(err, expect);
> +
> +	syncobj_destroy(fd, syncobj);
> +	munmap(data, bo_size);
> +	gem_close(fd, bo);
> +
> +	xe_exec_queue_destroy(fd, exec_queue);
> +	xe_vm_destroy(fd, vm);
> +}
> +
>  igt_main
>  {
>  	struct drm_xe_engine_class_instance *hwe;
> @@ -347,6 +438,10 @@ igt_main
>  					  s->flags);
>  	}
>  
> +	igt_subtest("non-blocking")
> +		non_block(fd, -EWOULDBLOCK);
> +
> +
>  	igt_fixture
>  		drm_close_driver(fd);
>  }
> -- 
> 2.39.1
>