[PATCH i-g-t] tests/intel/xe_exec_compute_mode: Stress test to validate long running workload on compute mode

[Zbigniew Kempczyński]

On Fri, Feb 23, 2024 at 12:04:30PM +0530, sai.gowtham.ch at intel.com wrote:
> From: Sai Gowtham Ch <sai.gowtham.ch at intel.com>
> 
> Test validates long running compute mode by submitting a spinner for 30s
> long.
> 
> Cc: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
> Cc: Janga Rahul Kumar <janga.rahul.kumar at intel.com>
> Signed-off-by: Sai Gowtham Ch <sai.gowtham.ch at intel.com>
> ---
>  tests/intel/xe_exec_compute_mode.c | 73 ++++++++++++++++++++++++++++++
>  1 file changed, 73 insertions(+)
> 
> diff --git a/tests/intel/xe_exec_compute_mode.c b/tests/intel/xe_exec_compute_mode.c
> index 7dad71509..0a71bfdc5 100644
> --- a/tests/intel/xe_exec_compute_mode.c
> +++ b/tests/intel/xe_exec_compute_mode.c
> @@ -400,6 +400,76 @@ static void non_block(int fd, int expect)
>  	xe_vm_destroy(fd, vm);
>  }
>  
> +/**
> + * SUBTEST: lr-stress
> + * Description: Stress LR mode workload for 30s.
> + * Test category: functionality test
> + */
> +static void lr_stress(int fd)
> +{
> +	uint64_t addr = 0x1a0000;
> +	struct drm_xe_sync sync[1] = {
> +		{.type = DRM_XE_SYNC_TYPE_USER_FENCE,
> +		 .flags = DRM_XE_SYNC_FLAG_SIGNAL,
> +		 .timeline_value = USER_FENCE_VALUE},
> +	};
> +	struct drm_xe_exec exec = {
> +		.num_batch_buffer = 1,
> +		.num_syncs = 1,
> +		.syncs = to_user_pointer(&sync),
> +	};
> +	struct  {
> +		struct xe_spin spin;
> +		uint64_t vm_sync;
> +		uint32_t data;
> +		uint64_t exec_sync;
> +	} *data;
> +	struct xe_spin_opts spin_opts = { .preempt = false };
> +	struct drm_xe_engine *engine;
> +	size_t bo_size;
> +	uint32_t vm;
> +	uint32_t exec_queue;
> +	uint32_t bo;
> +
> +	vm = xe_vm_create(fd, DRM_XE_VM_CREATE_FLAG_LR_MODE, 0);
> +	bo_size = sizeof(*data);
> +	bo_size = xe_bb_size(fd, bo_size);
> +
> +	engine = xe_engine(fd, 1);
> +	bo = xe_bo_create(fd, vm, bo_size, vram_if_possible(fd, engine->instance.gt_id), 0);
> +
> +	data = xe_bo_map(fd, bo, bo_size);
> +	memset(data, 0, bo_size);
> +	exec_queue = xe_exec_queue_create(fd, vm, &engine->instance, 0);
> +
> +	sync[0].addr = to_user_pointer(&data[0].vm_sync);
> +	xe_vm_bind_async(fd, vm, engine->instance.gt_id, bo, 0,
> +				addr, bo_size, sync, 1);
> +	xe_wait_ufence(fd, &data[0].vm_sync, USER_FENCE_VALUE, 0, ONE_SEC);
> +	data[0].vm_sync = 0;
> +
> +	spin_opts.addr = addr + (char *)&data[0].spin - (char *)data;
> +	xe_spin_init(&data[0].spin, &spin_opts);
> +	sync[0].addr = addr + (char *)&data[0].exec_sync - (char *)data;
> +	exec.exec_queue_id = exec_queue;
> +	exec.address = spin_opts.addr;
> +	xe_exec(fd, &exec);
> +	xe_spin_wait_started(&data[0].spin);
> +	sleep(30);

May you prove spinner is still spinning at this point and job is not
killed?

--
Zbigniew

> +
> +	xe_spin_end(&data[0].spin);
> +	xe_wait_ufence(fd, &data[0].exec_sync, USER_FENCE_VALUE, 0, ONE_SEC);
> +
> +	sync[0].addr = to_user_pointer(&data[0].vm_sync);
> +	xe_vm_unbind_async(fd, vm, 0, 0, addr, bo_size, sync, 1);
> +	xe_wait_ufence(fd, &data[0].vm_sync, USER_FENCE_VALUE, 0, ONE_SEC);
> +	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;
> @@ -460,6 +530,9 @@ igt_main
>  	igt_subtest("non-blocking")
>  		non_block(fd, EWOULDBLOCK);
>  
> +	igt_subtest("lr-stress")
> +		lr_stress(fd);
> +
>  
>  	igt_fixture
>  		drm_close_driver(fd);
> -- 
> 2.39.1
>