Re: [v2] crypto: zstd - convert to acomp

["Cabiddu, Giovanni" <giovanni.cabiddu@intel.com>]

On Wed, May 21, 2025 at 04:03:31PM +0800, Herbert Xu wrote:
> On Wed, May 21, 2025 at 07:41:11AM +0100, Suman Kumar Chakraborty wrote:
> >
> > +	if ((sg_nents(req->src) == 1) && (sg_nents(req->dst) == 1)) {
> > +		ret = zstd_compress_one(req, ctx, &total_out);
> > +		goto out;
> 
> This isn't right.  Even if the SG list is a singleton, it could
> consist of multiple pages in highmem.
> 
> What's worse, if src and dst are in fact linear, then you're
> dereferencing bogus pointers.
> 
> Just get rid of this optimisation unless you've got numbers
> proving that it makes a difference.
Without this optimization the performance drop is significant, compared
to the original implementation based on scomp. About -40% using
compressible data and 4KB flat buffers.

Is it an option to check if the SG list consists of multiple pages in
highmem like it was in scomp [1]?

Thanks,

[1] https://elixir.bootlin.com/linux/v6.14.7/source/crypto/scompress.c#L138

-- 
Giovanni