Re: Hardware checksums on bge(4) interfaces

[buhrow%lothlorien.nfbcal.org@localhost (Brian Buhrow)]

        Hello.  Looking at the Linux tg3 driver from the 2.6-36 kernel, it
appears that the problem is a driver bug in NetBSD.  Only the 5700B series
of chips appear to suffer from the hardware checksumming problem.  And,
yes, it affects udp packets as well.
        If I cook up a patch, are any of you in a position to test?
-thanks
-Brian
On Sep 14, 12:48pm, Brian Buhrow wrote:
} Subject: Re: Hardware checksums on bge(4) interfaces
}       Hello.  That other person was me, I believe.  Since Iwas having the
} trouble with an older revision of the BCM575x chip, I assumed that the
} problem was unique to that revision of the chip and didnt report it.
} However, having read Matthias's bug report, and after having dipped into
} the if_bge.c file for a quick read, I do think there's a chance it could be
} a driver bug.
}       It looks like the problem asserts itself when small tcp packets come
} in.  It also looks like one flag is used to tell the NIC to do checksum
} offloading whether it's udp or tcp.  It looks like for at least some
} revisions of the chip, this  was a known issue.  The driver does appear to
} be conservative in that it assumes that all chips have this bug, except
} that it looks like that assumption may have been commented out at some
} point, and thus all chips are assumed to be good.
}       I also think this bug may affect udp checksums as well, and we just
} haven't yet noticed.  I do see millions of udp packets with checksum errors
} on my name servers, which seems somewhat suspicious.  I also see this
} across different revisions of the chip.  I'll look into this a little
} further and see if I can do some instrumentation to find out more
} information.
}       I think sending packets of different sizes will yield different
} results with respect to the error rate.
}       This bug has existed since at least NetBSD-3.0, if my experience is
} any guide.
}       I'd also say that if it does affect udp packets, that it's a pretty
} serious bug and we should determine if it's a chip problem or a driver
} problem.
} 
} -Brian
} On Sep 14,  1:22pm, Chris Ross wrote:
} } Subject: Re: Hardware checksums on bge(4) interfaces
} } On 09/14/2010 01:04 PM, Matthias Scheler wrote:
} } > On Mon, Sep 13, 2010 at 12:34:59PM -0400, Chris Ross wrote:
} } >    
} } >>    Just a question.  Is there any reason *not* to use the various pieces
} } >> of hardware checksumming that are shown as available on my bge(4)
} } >> interfaces?
} } >>      
} } > Yes, please see PR kern/40018 (http://gnats.netbsd.org/40018).
} } >    
} } 
} }    Thanks for that link.  Someone else mentioned that specific problem.  
} } Looking at that PR, however, it's still not clear if this affects bge(4) 
} } hardware in general (in which case, it should not be offered by the 
} } driver), or only specific revisions of bge(4) hardware.
} } 
} }    Does anyone have any knowledge of how to look into this?  Does 
} } broadcom publish any sort of errata for revisions of their various 
} } ethernet parts?
} } 
} }                                                            - Chris
} } 
} } ps,
} }     and on a related (for me) note, if I want these features on for 
} } devices that are part of an agr(4), should I put them on the devices 
} } before adding them to the agr(4), enable these options on the agr(4), or 
} } both?
} } 
} >-- End of excerpt from Chris Ross
} 
} 
>-- End of excerpt from Brian Buhrow