[Intel-gfx] [igt-dev] [PATCH i-g-t] tests/core_hotunplug: Take care of closing fences before failing

[Marcin Bernatowicz]

On Wed, 2020-10-14 at 18:55 +0200, Janusz Krzysztofik wrote:
> The test was designed to keep track of open device file descriptors
> for safe driver unbind on recovery from a failed subtest.  In that
> context, fences introduced by commit 1fbd127bd4e1 ("core_hotplug:
> Teach the healthcheck how to check execution status") can affect
> device
> recovery as much as an open device file if not closed before unbind.
> 
> Moreover, forced GPU reset which used to be applied on recovery from
> a
> failed i915 GPU health check is no longer reachable since a GPU hang
> hopefully detected by the new health check algorithm can now break
> the
> whole recovery procedure prematurely.
> 
> Refactor local_i915_healthcheck() so it takes care of closing fences
> and returns a result to its caller instead of long jumping on
> failures
> believed to be recoverable.  While avoiding use of igt_assert() and
> friends, report actual source and error code of failures via
> igt_warn_on_f().
> 
> Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik at linux.intel.com
> >
> Cc: Chris Wilson <chris at chris-wilson.co.uk>
> ---
>  tests/core_hotunplug.c | 46 ++++++++++++++++++++++++++++++++------
> ----
>  1 file changed, 35 insertions(+), 11 deletions(-)
> 
> diff --git a/tests/core_hotunplug.c b/tests/core_hotunplug.c
> index 70669c590..d6db02bad 100644
> --- a/tests/core_hotunplug.c
> +++ b/tests/core_hotunplug.c
> @@ -233,9 +233,9 @@ static int merge_fences(int old, int new)
>  		return new;
>  
>  	merge = sync_fence_merge(old, new);
> -	igt_assert(merge != -1);
> -	close(old);
> -	close(new);
> +	/* Assume fence close errors don't affect device close status
> */
> +	igt_ignore_warn(local_close(old, "old fence close failed"));
> +	igt_ignore_warn(local_close(new, "new fence close failed"));
>  
>  	return merge;
>  }
> @@ -249,29 +249,53 @@ static int local_i915_healthcheck(int i915,
> const char *prefix)
>  		.buffer_count = 1,
>  	};
>  	const struct intel_execution_engine2 *engine;
> -	int fence = -1;
> +	int fence = -1, err = 0, status = 1;
>  
>  	local_debug("%s%s\n", prefix, "running i915 GPU healthcheck");
> -	if (local_i915_is_wedged(i915))
> +	if (igt_warn_on_f(local_i915_is_wedged(i915), "GPU found
> wedged\n"))
>  		return -EIO;
>  
> +	/* Assume gem_create()/gem_write() failures are unrecoverable
> */
>  	obj.handle = gem_create(i915, 4096);
>  	gem_write(i915, obj.handle, 0, &bbe, sizeof(bbe));
>  
> +	/* As soon as a fence is open, don't fail before closing it */
>  	__for_each_physical_engine(i915, engine) {
>  		execbuf.flags = engine->flags | I915_EXEC_FENCE_OUT;
> -		gem_execbuf_wr(i915, &execbuf);
> +		err = __gem_execbuf_wr(i915, &execbuf);
> +		if (igt_warn_on_f(err < 0, "__gem_execbuf_wr() returned
> %d\n",
> +				  err))
> +			break;
>  
>  		fence = merge_fences(fence, execbuf.rsvd2 >> 32);
> +		if (igt_warn_on_f(fence < 0, "merge_fences() returned
> %d\n",
> +				  fence)) {
> +			err = fence;
> +			break;
> +		}
> +	}
> +	if (fence >= 0) {
> +		status = sync_fence_wait(fence, -1);
> +		if (igt_warn_on_f(status < 0, "sync_fence_wait()
> returned %d\n",
> +				  status))
> +			err = status;
> +		if (!err)
> +			status = sync_fence_status(fence);
> +
> +		/* Assume fence close errors don't affect device close
> status */
> +		igt_ignore_warn(local_close(fence, "fence close
> failed"));
>  	}
> -	igt_assert(fence != -1);
> +
> +	/* Assume gem_close() failure is unrecoverable */
>  	gem_close(i915, obj.handle);
>  
> -	igt_assert_eq(sync_fence_wait(fence, -1), 0);
> -	igt_assert_eq(sync_fence_status(fence), 1);
> -	close(fence);
> +	if (err < 0)
> +		return err;
> +	if (igt_warn_on_f(status != 1, "sync_fence_status() returned
> %d\n",
> +			  status))
> +		return -1;
>  
> -	if (local_i915_is_wedged(i915))
> +	if (igt_warn_on_f(local_i915_is_wedged(i915), "GPU turned
> wedged\n"))
>  		return -EIO;
LGTM,
Reviewed-by: Marcin Bernatowicz <marcin.bernatowicz at linux.intel.com>

>  
>  	return 0;