RX checksum offloading problem

[Yonghyeon PYUN]

On Wed, May 07, 2014 at 10:07:09AM +0200, Michael Tuexen wrote:
> On 07 May 2014, at 09:56, Yonghyeon PYUN <pyunyh at gmail.com> wrote:
> 
> > On Sat, May 03, 2014 at 11:52:47AM +0200, Michael Tuexen wrote:
> >> On 02 May 2014, at 16:02, Bjoern A. Zeeb <bz at FreeBSD.org> wrote:
> >> 
> >>> 
> >>> On 02 May 2014, at 10:22 , Michael Tuexen <Michael.Tuexen at lurchi.franken.de> wrote:
> >>> 
> >>>> Dear all,
> >>>> 
> >>>> during testing I found that FreeBSD head (on a raspberry pi) accepts SCTP packet
> >>>> with bad checksums. After debugging this I figured out that this is a problem with
> >>>> the csum_flags defined in mbuf.h.
> >>>> 
> >>>> The SCTP code on its input path checks for CSUM_SCTP_VALID, which is defined in mbuf.h:
> >>>> #define CSUM_SCTP_VALID         CSUM_L4_VALID
> >>>> This makes sense: If CSUM_SCTP_VALID is set in csum_flags, the packet is considered
> >>>> to have a correct checksum.
> >>>> 
> >>>> For UDP and TCP some drivers calculate the UDP/TCP checksum and set CSUM_DATA_VALID in
> >>>> csum_flags to indicate that the UDP/TCP should consider csum_data to figure out if
> >>>> the packet has a correct checksum. The problem is that CSUM_DATA_VALID is defined as
> >>>> #define CSUM_DATA_VALID         CSUM_L4_VALID
> >>>> In this case the semantic is not that the packet has a valid checksum, but the csum_data
> >>>> field contains information.
> >>>> 
> >>>> Now the following happens (on the raspberry pi the driver used is
> >>>> dev/usb/net/if_smsc.c
> >>>> 
> >>>> 1. A packet is received and if it is not too short, the checksum computed
> >>>> is stored in csum_data and the flag CSUM_DATA_VALID is set. This happens
> >>>> for all IP packets, not only for UDP and TCP packets.
> >>>> 2. In case of SCTP packets, the SCTP interprets CSUM_DATA_VALID as CSUM_SCTP_VALID
> >>>> and accepts the packet. So no SCTP checksum check ever happened.
> >>>> 
> >>>> Alternatives to fix this:
> >>>> 
> >>>> 1. Change all drivers to set CSUM_DATA_VALID only in case of UDP or TCP packets, since
> >>>> it only makes sense in these cases.
> >>> 
> >>> Wait, or for SCTP in cad the crc32 (I think it was)  was actually checked but not otherwise.   This is how it should be imho.  It seems like a driver bug.
> >> I went through the list of drivers and you are right, it seems to be a bug
> >> in if_smsc.c. Most of the other drivers check for UDP/TCP, a small set I can't tell.
> >> 
> > 
> > I'm not sure how the controller computes TCP/UDP checksum values.
> > It seems the publicly available data sheet was highly sanitized so
> > it was useless to me.  The comment in the driver says that the
> Same for me...
> > controller computes RX checksum after the IPv4 header to the end of
> > ethernet frame. After seeing that comment, three questions popped
> > up:
> > 
> > 1. Is the controller smart enough to skip IP options header in
> >   TCP/UDP checksum offloading?
> > 2. How controller handles UDP checksum value 0x0000(i.e. sender
> >   didn't compute UDP checksum)?
> > 3. How the controller can compute TCP checksum of fragmented
> >   packets?
> > 
> > Since you have the controller I guess it's easy to verify all
> > cases.  For case 3, I believe the controller can't handle
> > fragmented frames so driver should have to explicitly check ip_off
> > field of IPv4 header.  See how gem(4)/sk(4)/hme(4) and fxp(4)
> > handle it.
> Let me check this. Is there a tool to send UDP/TCP with IP level options
> or do I need to write a small test program myself?
> 

I recall I used buggy ipsend of ipfilter package in the past but it
would be more easy to write a simple test program or patch driver
to generate those frames.