ix(intel) vs mlxen(mellanox) 10Gb performance

[Yonghyeon PYUN]

On Wed, Aug 19, 2015 at 08:13:59AM -0400, Rick Macklem wrote:
> Yonghyeon PYUN wrote:
> > On Wed, Aug 19, 2015 at 09:51:44AM +0200, Hans Petter Selasky wrote:
> > > On 08/19/15 09:42, Yonghyeon PYUN wrote:
> > > >On Wed, Aug 19, 2015 at 09:00:52AM +0200, Hans Petter Selasky wrote:
> > > >>On 08/18/15 23:54, Rick Macklem wrote:
> > > >>>Ouch! Yes, I now see that the code that counts the # of mbufs is before
> > > >>>the
> > > >>>code that adds the tcp/ip header mbuf.
> > > >>>
> > > >>>In my opinion, this should be fixed by setting if_hw_tsomaxsegcount to
> > > >>>whatever
> > > >>>the driver provides - 1. It is not the driver's responsibility to know
> > > >>>if
> > > >>>a tcp/ip
> > > >>>header mbuf will be added and is a lot less confusing that expecting the
> > > >>>driver
> > > >>>author to know to subtract one. (I had mistakenly thought that
> > > >>>tcp_output() had
> > > >>>added the tc/ip header mbuf before the loop that counts mbufs in the
> > > >>>list.
> > > >>>Btw,
> > > >>>this tcp/ip header mbuf also has leading space for the MAC layer
> > > >>>header.)
> > > >>>
> > > >>
> > > >>Hi Rick,
> > > >>
> > > >>Your question is good. With the Mellanox hardware we have separate
> > > >>so-called inline data space for the TCP/IP headers, so if the TCP stack
> > > >>subtracts something, then we would need to add something to the limit,
> > > >>because then the scatter gather list is only used for the data part.
> > > >>
> > > >
> > > >I think all drivers in tree don't subtract 1 for
> > > >if_hw_tsomaxsegcount.  Probably touching Mellanox driver would be
> > > >simpler than fixing all other drivers in tree.
> > > 
> > > Hi,
> > > 
> > > If you change the behaviour don't forget to update and/or add comments
> > > describing it. Maybe the amount of subtraction could be defined by some
> > > macro? Then drivers which inline the headers can subtract it?
> > > 
> > 
> > I'm also ok with your suggestion.
> > 
> > > Your suggestion is fine by me.
> > > 
> > 
> > > The initial TSO limits were tried to be preserved, and I believe that
> > > TSO limits never accounted for IP/TCP/ETHERNET/VLAN headers!
> > > 
> > 
> > I guess FreeBSD used to follow MS LSOv1 specification with minor
> > exception in pseudo checksum computation. If I recall correctly the
> > specification says upper stack can generate up to IP_MAXPACKET sized
> > packet.  Other L2 headers like ethernet/vlan header size is not
> > included in the packet and it's drivers responsibility to allocate
> > additional DMA buffers/segments for L2 headers.
> > 
> Yep. The default for if_hw_tsomax was reduced from IP_MAXPACKET to
>   32 * MCLBYTES - max_ethernet_header_size as a workaround/hack so that
> devices limited to 32 transmit segments would work (ie. the entire packet,
> including MAC header would fit in 32 MCLBYTE clusters).
> This implied that many drivers did end up using m_defrag() to copy the mbuf
> list to one made up of 32 MCLBYTE clusters.
> 
> If a driver sets if_hw_tsomaxsegcount correctly, then it can set if_hw_tsomax
> to whatever it can handle as the largest TSO packet (without MAC header) the
> hardware can handle. If it can handle > IP_MAXPACKET, then it can set it to that.
> 

I thought the upper limit was still IP_MAXPACKET. If driver
increase it (i.e. > IP_MAXPACKET,  the length field in the IP
header would overflow which in turn may break firewalls and other
packet handling in IPv4/IPv6 code path.
If the limit no longer apply to network stack, that's great.  Some
controllers can handle up to 256KB TCP/UDP segmentation and
supporting that feature wouldn't be hard.