Re: hardware checksumming

Re: hardware checksumming

[Christian Widmer]

On Thursday 23 August 2001 15:54, you wrote:
> > is hardware support by a NIC for checksum generation / offloading not
> > quite usless? the checksumming enging can only be used when UDP/TCP
> > packets are <= the MTU of the NIC (e.g 1500 bytes).
>
> checksums can be chained.
yes that might be easy fores belonging together come one after the other. so 
the NIC ha transmit, since all chained frams only to trak one 
fragmentet packet. but receiving is will become difficult since different
conections can arrive interleaved. something else: the NIC needs quite
a lot of ram or huge fifios. if not it will have to transfer the data twice 
through the PCI (checksumm are at the beginig of a packet not the
end). nobody did spend a second on thinking to implement all in hard-
ware when the interet protokol was desinged.

do you know any NIC that is capable of chaining? currenty i've a dp83820
on my desc and at the moment im not eaven shure if checksumming can
be chained over multiple descriptions (describing one single ethernet frame).

>
> > i expact that UDP/TCP packets are in general bigger than that or is
> > exactly
>
> TCP packets are certainly never larger than the MTU.

Re: hardware checksumming

[Christian Widmer]

oh its me again. sorry but the message bofore was messed up. so lets try 
again.

On Thursday 23 August 2001 15:54, you wrote:
> > is hardware support by a NIC for checksum generation / offloading not
> > quite usless? the checksumming enging can only be used when UDP/TCP
> > packets are <= the MTU of the NIC (e.g 1500 bytes).
>
> checksums can be chained.
yes that might be easy for the NIC's transmitting enging when all fragments 
belonging together come one after the other in the transmit-ring. 
but receiving is difficult since different connections can arrive 
simultanously and reslut in interleaved framgents.

something else: you need quite a lot of ram on the NIC to buffer fragments.
if not it will have to transfer the data twice through the PCI (checksumm are 
at the beginig of a packet not the end). nobody did spend a second on 
thinking to implement all in hardware when the interet protokol was desinged.

do you know any NIC that is capable of chaining? currenty i've a dp83820
on my desc and at the moment im not eaven shure if checksumming can
be chained over multiple descriptions (describing one single ethernet frame).

>
> > i expact that UDP/TCP packets are in general bigger than that or is
> > exactly
>
> TCP packets are certainly never larger than the MTU.

Re: hardware checksumming

[Alexey Kuznetsov]

Hello!

> but receiving is difficult 

Hmm... on receiving it is enough trivial and we do this.


> something else: you need quite a lot of ram on the NIC to buffer fragments.

A lot? Even ne2000 had 16K of memory, which is enough to do this. 
It is the first.

Card need not hold fragments, it may send fragments in backward order.
It is the second.

Well, and the last but not the least: acenic has exactly this amount
of memory: "lots" :-) It is the third. :-)


> do you know any NIC that is capable of chaining?

acenic. We do not use this, though this is also not difficult.

Alexey

Re: hardware checksumming

[Helge Hafting]

Christian Widmer wrote:

> yes that might be easy for the NIC's transmitting enging when all fragments
> belonging together come one after the other in the transmit-ring.
> but receiving is difficult since different connections can arrive
> simultanously and reslut in interleaved framgents.
> 
> something else: you need quite a lot of ram on the NIC to buffer fragments.
> if not it will have to transfer the data twice through the PCI (checksumm are
> at the beginig of a packet not the end). nobody did spend a second on
> thinking to implement all in hardware when the interet protokol was desinged.

You don't have to store entire fragments in a NIC to do
hw checksums.  All you need to store is the partial checksums,
and add to them whenever yet another fragment comes in for
that packet.

Helge Hafting

hardware checksumming

[Christian Widmer]

is hardware support by a NIC for checksum generation / offloading not quite 
usless? the checksumming enging can only be used when UDP/TCP packets
are <= the MTU of the NIC (e.g 1500 bytes). 
i expact that UDP/TCP packets are in general bigger than that or is exactly 
that wrong and the network protokoll stack splits the data to be transfered 
so that each tcp packet is not bigger than the MTU. but whats with UPD there 
this wont work.