-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hello,

On Jul 25, 2007, at 09:12, Izumi Tsutsui wrote:

> macallan@NetBSD.org wrote:
>
>> The difference is that everything ( besides boot64 ) is built with
>> - -march=mips3 -mtune=r5k - this saves a lot of space and seems to be
>> substantially faster than the official builds which default to
>> - -march=mips1
>
> Are all binaries are built with those flags?

All but boot64.

> It's fine to specify CPUFLAGS into GENERIC32_IP3x,
> but we shouldn't do it for userland binaries by default
> because we also support R3K Indigo (IP12) in this port.

I know, I wasn't suggesting that we should do that.

>> Besides that the kernel contains some changes regarding mace and mec -
>
> Could you also put a diff somewhere?

Should show up in f.n.o/pub/NetBSD/misc/macallan/sgimips later today.

>> On my machine this is stable so far, mec isn't blazingly fast ( only 
>> up
>> to 3MB/s ) but I don't get spontaneous resets on heavy network traffic
>> either.
>
> Current mec(4) driver copies all RX mbufs and most TX mbufs
> from/to DMA buffers so xfer rate is limited by memory bandwidth.

Without those delays in the bus access functions I had mec up to about 
7MB/s - hard to reproduce though, usually the machine resets after a 
few seconds heavy traffic. Funny thing is - receiving works just fine, 
even without delays, transmitting is what kills it so I guess there's a 
race hidden somewhere which is masked by the delays.

have fun
Michael
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (Darwin)

iQEVAwUBRqditMpnzkX8Yg2nAQIbrggArkGI/W1sJt0T1kjXr8CDShAq6VBXtO5S
Yi4GJutJCWBYoyV8BnnbyBptsk8TyFJ39GH3WTMXTPC6h7DfCM2dyWcpAn3/ToJa
GizW7wyfAq8FLXDDD0v2nwE7NtAVBtmqitlEBv7Gkmx7Q/qH0HiTjKsekyOp5kQ3
V+qDGAQgyCr0pGEh/hTd9rxzS56Mlo+zu5MzHju8q78MFWy09fMZvjD6ZbmbpbVG
nGKPEYBvj7xtmSE961xOJS6eB4l2ZI/JqpagU9f+K9EnOMBO+37yVeHZdY1QwGRC
juJLBEmN+5hqWsvoow5ugroZQyLvoYpD7kzNh1j9e54boNyHhA82gw==
=NAAc
-----END PGP SIGNATURE-----