Re: [PATCH v5 2/2] drm-uapi/xe: Don't wait on user_fence during exec queue reset

[Francois Dugast <francois.dugast@intel.com>]

On Tue, Dec 12, 2023 at 10:40:40AM +0530, Bommu Krishnaiah wrote:
> Don't wait till timeout on user fence when exec_queue reset is detected
> and return return  proper error code
> 
> Signed-off-by: Bommu Krishnaiah <krishnaiah.bommu@intel.com>
> Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
> Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
> Cc: Francois Dugast <francois.dugast@intel.com>
> ---
>  tests/intel/xe_waitfence.c | 82 ++++++++++++++++++++++++++++++++++++++
>  1 file changed, 82 insertions(+)
> 
> diff --git a/tests/intel/xe_waitfence.c b/tests/intel/xe_waitfence.c
> index c272d8a83..2eb3a28b2 100644
> --- a/tests/intel/xe_waitfence.c
> +++ b/tests/intel/xe_waitfence.c
> @@ -153,6 +153,9 @@ waitfence(int fd, enum waittype wt)
>   *
>   * SUBTEST: invalid-engine
>   * Description: Check query with invalid engine info returns expected error code
> + *
> + * SUBTEST: exec_queue-reset-wait
> + * Description: Don’t wait till timeout on user fence when exec_queue reset is detected and return return proper error
>   */
>  
>  static void
> @@ -228,6 +231,82 @@ invalid_engine(int fd)
>  	do_ioctl_err(fd, DRM_IOCTL_XE_WAIT_USER_FENCE, &wait, EFAULT);
>  }
>  
> +static void
> +exec_queue_reset_wait(int fd)
> +{
> +	uint32_t bo, b;
> +	uint64_t batch_offset;
> +	uint64_t batch_addr;
> +	uint64_t sdi_offset;
> +	uint64_t sdi_addr;
> +	uint64_t addr = 0x1a0000;
> +
> +	struct {
> +		uint32_t batch[16];
> +		uint64_t pad;
> +		uint64_t vm_sync;
> +		uint64_t exec_sync;
> +		uint32_t data;
> +	} *data;
> +
> +#define USER_FENCE_VALUE        0xdeadbeefdeadbeefull
> +	struct drm_xe_sync sync[1] = {
> +		{ .flags = DRM_XE_SYNC_TYPE_USER_FENCE | 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),
> +	};
> +
> +	uint32_t vm = xe_vm_create(fd, DRM_XE_VM_CREATE_FLAG_ASYNC_DEFAULT, 0);
> +	uint32_t exec_queue = xe_exec_queue_create_class(fd, vm, DRM_XE_ENGINE_CLASS_COPY);
> +	struct drm_xe_wait_user_fence1 wait = {

Hi Bommu,

I assume this is a typo and should be like this, right?
s/drm_xe_wait_user_fence1/drm_xe_wait_user_fence/

Francois

> +		.op = DRM_XE_UFENCE_WAIT_OP_EQ,
> +		.flags = 0,
> +		.value = 0xc0ffee,
> +		.mask = DRM_XE_UFENCE_WAIT_MASK_U64,
> +		.timeout = -1,
> +		.exec_queue_id = exec_queue,
> +	};
> +
> +	bo = xe_bo_create(fd, vm, 0x40000, vram_if_possible(fd, 0), 0);
> +	data = xe_bo_map(fd, bo, 0x40000);
> +
> +	batch_offset = (char *)&data[0].batch - (char *)data;
> +	batch_addr = addr + batch_offset;
> +	sdi_offset = (char *)&data[0].data - (char *)data;
> +	sdi_addr = addr + sdi_offset;
> +
> +	b = 0;
> +	data[0].batch[b++] = MI_STORE_DWORD_IMM_GEN4;
> +	data[0].batch[b++] = sdi_addr;
> +	data[0].batch[b++] = sdi_addr >> 32;
> +	data[0].batch[b++] = 0xc0ffee;
> +	data[0].batch[b++] = MI_BATCH_BUFFER_END;
> +	igt_assert(b <= ARRAY_SIZE(data[0].batch));
> +
> +	wait.addr = to_user_pointer(&data[0].exec_sync);
> +	exec.exec_queue_id = exec_queue;
> +	exec.address = batch_addr;
> +
> +	xe_exec(fd, &exec);
> +
> +	/**
> +	  * Don't do the GPU mapping(vm_bind) for object, so that exec_queue
> +	  * reset will happen and xe_wait_ufence will return EIO not ETIME
> +	  */
> +	do_ioctl_err(fd, DRM_IOCTL_XE_WAIT_USER_FENCE, &wait, EIO);
> +
> +	xe_exec_queue_destroy(fd, exec_queue);
> +
> +	if (bo) {
> +		munmap(data, 0x40000);
> +		gem_close(fd, bo);
> +	}
> +}
>  
>  igt_main
>  {
> @@ -254,6 +333,9 @@ igt_main
>  	igt_subtest("invalid-engine")
>  		invalid_engine(fd);
>  
> +	igt_subtest("exec_queue-reset-wait")
> +		exec_queue_reset_wait(fd);
> +
>  	igt_fixture
>  		drm_close_driver(fd);
>  }
> -- 
> 2.25.1
>