[Intel-gfx] [PATCH igt] lib/dummyload: Pad with a few nops so that we do not completely hog the system

[Mika Kuoppala]

Chris Wilson <chris at chris-wilson.co.uk> writes:

> Part of the attraction of using a recursive batch is that it is
> hard on the system (executing the "function" call is apparently
> quite expensive). However, the GPU may hog the entire system for
> a few minutes, preventing even NMI. Quite why this is so is unclear,
> but presumably it relates to the PM_INTRMSK workaround on gen6/gen7.
> If we give the system a break by having the GPU execute a few nops
> between function calls, that appears enough to keep SNB out of
> trouble.
>

Can we trip into the same pit with igt_hang? It has tight
recurse. Adding few noops makes the hangcheck sampler
to hit more sparsely and slow the verdict of hang.

But we can then improve the heuristics ;)

-Mika



> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Tomi Sarvela <tomi.p.sarvela at intel.com>
> Cc: Daniel Vetter <daniel.vetter at intel.com>
> ---
>  lib/igt_dummyload.c | 18 ++++++++++++++++--
>  1 file changed, 16 insertions(+), 2 deletions(-)
>
> diff --git a/lib/igt_dummyload.c b/lib/igt_dummyload.c
> index 00c6a030..c6e57cff 100644
> --- a/lib/igt_dummyload.c
> +++ b/lib/igt_dummyload.c
> @@ -103,7 +103,7 @@ static void emit_recursive_batch(igt_spin_t *spin,
>  		/* dummy write to dependency */
>  		obj[SCRATCH].handle = dep;
>  		fill_reloc(&relocs[obj[BATCH].relocation_count++],
> -			   dep, 256,
> +			   dep, 1020,
>  			   I915_GEM_DOMAIN_RENDER,
>  			   I915_GEM_DOMAIN_RENDER);
>  		execbuf.buffer_count++;
> @@ -112,9 +112,23 @@ static void emit_recursive_batch(igt_spin_t *spin,
>  	spin->batch = batch;
>  	spin->handle = obj[BATCH].handle;
>  
> +	/* Pad with a few nops so that we do not completely hog the system.
> +	 *
> +	 * Part of the attraction of using a recursive batch is that it is
> +	 * hard on the system (executing the "function" call is apparently
> +	 * quite expensive). However, the GPU may hog the entire system for
> +	 * a few minutes, preventing even NMI. Quite why this is so is unclear,
> +	 * but presumably it relates to the PM_INTRMSK workaround on gen6/gen7.
> +	 * If we give the system a break by having the GPU execute a few nops
> +	 * between function calls, that appears enough to keep SNB out of
> +	 * trouble. See https://
> +	 */
> +	batch += 1000;
> +
>  	/* recurse */
>  	fill_reloc(&relocs[obj[BATCH].relocation_count],
> -		   obj[BATCH].handle, 1, I915_GEM_DOMAIN_COMMAND, 0);
> +		   obj[BATCH].handle, (batch - spin->batch) + 1,
> +		   I915_GEM_DOMAIN_COMMAND, 0);
>  	if (gen >= 8) {
>  		*batch++ = MI_BATCH_BUFFER_START | 1 << 8 | 1;
>  		*batch++ = 0;
> -- 
> 2.14.1
>
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/intel-gfx