From: David Gwynne <david@gwynne.id.au>
Subject: Re: ix(4): fix rss hashing on big endian archs
To: tech@openbsd.org
Date: Thu, 23 Apr 2026 07:40:14 +1000

On Wed, Apr 22, 2026 at 10:42:31AM +0200, Claudio Jeker wrote:
> On Wed, Apr 22, 2026 at 01:08:54PM +1000, David Gwynne wrote:
> > i noticed that packets seem to be hashed differently by ix(4) and the
> > network stack on sparc64. the diff below seems to fix it, and it still
> > works as expected on arm64.
> 
> How did you notice this?

i was doing a single fast tcp transfer while looking at stuff in systat,
and noticed that ix had interrupts firing on two sets of rings instead
of one like i expect. txq and rxq kstats confirmed the tx and rx
packets were going to different ring indexes.

if you want a clearer indicator, i could add a counters to the tcp stats
that says whether the flowid in rx packets matches the flowid on the tcp
pcb.

> > ok?
> 
> Now this screams for bus_space_write_raw_4() but I see that the macro zoo
> does not make this a good option.
> 
> I would expect this to be a le32toh() so we end up with htole32(le32toh())
> combo but for all archs we care htole32(x) == le32toh(x) so it really is a
> bikeshed.

i can use do this.

> OK claudio@
> 
> Are ixl and ice affected as well?

i havent moved on to looking at ixl yet and i dont have any ice.

>  
> > Index: if_ix.c
> > ===================================================================
> > RCS file: /cvs/src/sys/dev/pci/if_ix.c,v
> > diff -u -p -r1.223 if_ix.c
> > --- if_ix.c	26 Feb 2026 23:38:10 -0000	1.223
> > +++ if_ix.c	22 Apr 2026 03:04:16 -0000
> > @@ -3061,8 +3059,13 @@ ixgbe_initialize_rss_mapping(struct ix_s
> >  	}
> >  
> >  	/* Now fill our hash function seeds */
> > -	for (i = 0; i < 10; i++)
> > -		IXGBE_WRITE_REG(hw, IXGBE_RSSRK(i), rss_key[i]);
> > +	for (i = 0; i < nitems(rss_key); i++) {
> > +		/*
> > +		 * twist the words so the rss key will be handled
> > +		 * as bytes regardless of the hosts endianness.
> > +		 */
> > +		IXGBE_WRITE_REG(hw, IXGBE_RSSRK(i), htole32(rss_key[i]));
> > +	}
> >  
> >  	/*
> >  	 * Disable UDP - IP fragments aren't currently being handled
> > 
> 
> -- 
> :wq Claudio