Re: Marvell if_mvgbe

[Masanobu SAITOH <msaitoh%execsw.org@localhost>]

(2012/10/15 15:28), David Laight wrote:
> On Mon, Oct 15, 2012 at 11:33:14AM +1300, Lloyd Parkes wrote:
> > On 13/10/2012, at 1:42 AM, Robert Swindells <rjs%fdy2.co.uk@localhost> wrote:
> >
> > > I have been using the following patch for a while now on the
> > > SheevaPlug, it should reduce the number of receive interrupts
> > > generated by the ethernet controllers.
> >
> > Great. The interrupt load on some of my older gear was so bad that I had to 
> > retire that gear, so I approve of fewer interrupts.
> >
> > > +static uint32_t
> > > +mvgbe_ipg_rx(struct mvgbec_softc *csc, struct mvgbe_softc *sc)
> > > +{
> > > +
> > > +       if (csc->sc_flags & FLAGS_IPG2)
> > > +               return (((sc->sc_ipg_rx & 0x8000) << 10) |
> > > +                       ((sc->sc_ipg_rx & 0x7fff) << 7));
> > > +       else if (csc->sc_flags & FLAGS_IPG1)
> > > +               return ((sc->sc_ipg_rx & 0x3fff) << 8);
> > > +       else
> > > +               return 0;
> > > +}
> > ...
> > > +       if ((csc->sc_flags & FLAGS_IPG1) || (csc->sc_flags & FLAGS_IPG2))
> > > +               sc->sc_ipg_rx = 768;
> >
> > You need to replace all the numeric constants with #defines though.
> > I can read and write network device drivers (that aren't too complicated),
> > but I don't know what this code does without referring to the Marvell
> > data sheets.
>
> Also, remove the 'else' following the 'return's.
>
>         David

 I've added the sysctl interface to mvgbe (if_mvgbe.c rev. 1.25).


> Modified Files:
>         src/sys/dev/marvell: if_mvgbe.c mvgbereg.h
>
> Log Message:
> Add code to modify interrupt coalescing with sysctl.
>  - both TX side an RX side.
>  - different setting for each port
>  - TX side is hw.mvgbe.mvgbe*.ipginttx
>  - RX side is hw.mvgbe.mvgbe*.ipgintrx
>  - The default value is 768.
>  - The lowest value is 0
>  - For highest value, 0x3777 is used for V1, and 0xffff is used for V2.
>
>
> To generate a diff of this commit:
> cvs rdiff -u -r1.24 -r1.25 src/sys/dev/marvell/if_mvgbe.c
> cvs rdiff -u -r1.5 -r1.6 src/sys/dev/marvell/mvgbereg.h


 Example:

> # sysctl -a |grep mvgbe
> kern.root_device = mvgbe0
> net.interfaces.mvgbe0.sndq.len = 0
> net.interfaces.mvgbe0.sndq.maxlen = 256
> net.interfaces.mvgbe0.sndq.drops = 0
> hw.mvgbe.mvgbe0.ipginttx = 768
> hw.mvgbe.mvgbe0.ipgintrx = 768

five# ping -f 192.168.1.30
PING 192.168.1.30 (192.168.1.30): 52 data bytes
.^C.....

----192.168.1.30 PING Statistics----
11500 packets transmitted, 11500 packets received, 0.0% packet loss
round-trip min/avg/max/stddev = 0.125/0.589/99.985/3.764 ms
  2327.7 packets/sec sent,  2327.8 packets/sec received



> # sysctl -w hw.mvgbe.mvgbe0.ipgintrx=0
> hw.mvgbe.mvgbe0.ipgintrx: 768 -> 0

five# ping -f 192.168.1.30
PING 192.168.1.30 (192.168.1.30): 52 data bytes
.^C......

----192.168.1.30 PING Statistics----
4582 packets transmitted, 4582 packets received, 0.0% packet loss
round-trip min/avg/max/stddev = 0.113/0.583/99.895/3.720 ms
  2391.0 packets/sec sent,  2391.1 packets/sec received


> # sysctl -w hw.mvgbe.mvgbe0.ipgintrx=65535
> hw.mvgbe.mvgbe0.ipgintrx: 0 -> 65535

five# ping -f 192.168.1.30
PING 192.168.1.30 (192.168.1.30): 52 data bytes
.^C.......

----192.168.1.30 PING Statistics----
427 packets transmitted, 427 packets received, 0.0% packet loss
round-trip min/avg/max/stddev = 0.413/17.336/114.079/17.427 ms
  100.3 packets/sec sent,  99.8 packets/sec received

The delay and the standard deviation are good at both 0 and 768,
so I think 768 is reasonable on SheevaPlug.

--
-----------------------------------------------
                SAITOH Masanobu (msaitoh%execsw.org@localhost
                                 msaitoh%netbsd.org@localhost)