svn commit: r359016 - head/sys/netinet

[Gleb Smirnoff]

  Andrew,

On Mon, Mar 16, 2020 at 02:03:27PM +0000, Andrew Gallatin wrote:
A> Log:
A>   Avoid a cache miss accessing an mbuf ext_pgs pointer when doing SW kTLS.
A>   
A>   For a Netflix 90Gb/s 100% TLS software kTLS workload, this reduces
A>   the CPI of tcp_m_copym() from ~3.5 to ~2.5 as reported by vtune.
A>   
A>   Reviewed by:	jtl, rrs
A>   Sponsored by:	Netflix
A>   Differential Revision:	https://reviews.freebsd.org/D23998
A> 
A> Modified:
A>   head/sys/netinet/tcp_output.c
A> 
A> Modified: head/sys/netinet/tcp_output.c
A> ==============================================================================
A> --- head/sys/netinet/tcp_output.c	Mon Mar 16 13:53:29 2020	(r359015)
A> +++ head/sys/netinet/tcp_output.c	Mon Mar 16 14:03:27 2020	(r359016)
A> @@ -1907,7 +1907,7 @@ tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *ple
A>  	top = NULL;
A>  	pkthdrlen = NULL;
A>  #ifdef KERN_TLS
A> -	if (m->m_flags & M_NOMAP)
A> +	if (hw_tls && (m->m_flags & M_NOMAP))
A>  		tls = m->m_ext.ext_pgs->tls;
A>  	else
A>  		tls = NULL;

IMHO, such changes must always be accompanied by a comment. Otherwise, I
can easily imagine someone in couple of years "optimizing" it back with
commit message "Remove extraneous check. m->m_ext.ext_pgs->tls is NULL when tls is off"

-- 
Gleb Smirnoff