[Intel-gfx] [PATCH 3/3] [v3] intel_error_decode: Fix ACTHD/HEAD mess with libdrm

Chris Wilson chris at chris-wilson.co.uk
Tue Apr 23 10:30:33 CEST 2013


On Mon, Apr 22, 2013 at 06:11:12PM -0700, Ben Widawsky wrote:
> This patch is an enormous mess, and I'd be fine if people didn't want
> it. However I have made the code do what we want at least on the two
> error dumps I've tried.
> 
> The way that it works is it attempts to identify which ACTHD belongs to
> the ring, or batch, and add the appropriate offset as necessary so the
> libdrm decoder can do the right thing.
> 
> What I do is put each ACTHD in a fixed part of the array, and assume the
> error dump will dump each ring in gtt ascending order (ie. RCS offset <
> VCS offset < BCS offset). I know, its hacky.
> 
> Signed-off-by: Ben Widawsky <ben at bwidawsk.net>
> ---

> +static void emit_acthd(struct drm_intel_decode *decode_ctx,
> +		uint64_t gtt_offset, uint64_t size, int is_batch)
> +{
> +	/* XXX: This depends on always being in ascending gtt space order */
> +	static int which_acthd = 0;

You have ring_name available to sort on.
I would prefer the use of RING_HEAD when decoding the ringbuffers, so
that the current HEAD is always printed there - acting like a frame
pointer in a stacktrace. Making that distinction between rings and
batches should make this code simplier.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre



More information about the Intel-gfx mailing list