Date:   Mon, 20 Dec 2021 15:03:43 -0800
From:   Jakub Kicinski <kuba@kernel.org>
To:     Herbert Xu <herbert@gondor.apana.org.au>,
        Ard Biesheuvel <ardb@kernel.org>
Cc:     linux-crypto@vger.kernel.org, Sabrina Dubroca <sd@queasysnail.net>
Subject: Re: x86 AES crypto alignment
Message-ID: <20211220150343.4e12a4d2@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
In-Reply-To: <20211207212907.6e91821b@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
References: <20211207113252.162701ed@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
        <20211208044037.GA11399@gondor.apana.org.au>
        <20211207212907.6e91821b@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Precedence: bulk

On Tue, 7 Dec 2021 21:29:07 -0800 Jakub Kicinski wrote:
> On Wed, 8 Dec 2021 15:40:37 +1100 Herbert Xu wrote:
> > On Tue, Dec 07, 2021 at 11:32:52AM -0800, Jakub Kicinski wrote:  
> > > Hi!
> > > 
> > > The x86 AES crypto (gcm(aes)) requires 16B alignment which is hard to
> > > achieve in networking. Is there any reason for this? On any moderately 
> > > recent Intel platform aligned and unaligned vmovdq should have the same
> > > performance (reportedly).    
> > 
> > There is no such thing as an alignment requirement.  If an algorithm
> > specifies an alignment and you pass it a request which is unaligned,
> > the Crypto API will automatically align the data for you.
> > 
> > So what is the actual problem here?  
> 
> By align you mean copy right? I'm trying to avoid the copy.

Hm, I'm benchmarking things now and it appears to be a regression
introduced somewhere around 5.11 / 5.12. I don't see the memcpy 
eating 20% of performance on 5.10. Bisection time.