> On Aug 7, 2025, at 6:35 AM, Gleb Smirnoff <gleb...@freebsd.org> wrote:
> 
>  Michael,
> 
> On Fri, Aug 01, 2025 at 08:17:35AM +0000, Michael Tuexen wrote:
> M>     When the SCTP, TCP, or UDP implementation send a packet, it does not
> M>     compute the corresponding checksum but defers that. The network layer
> M>     will determine whether the network interface selected for the packet
> M>     has the requested capability and computes the checksum in software,
> M>     if the selected network interface doesn't have the requested
> M>     capability.
> M>     Do this not only for packets being sent by the local SCTP, TCP,
> M>     and UDP stack, but also when forwarding packets. Furthermore, when
> M>     such packets are delivered to a local SCTP, TCP, or UDP stack, do not
> M>     compute or validate the checksum, since such packets never have been on
> M>     the wire.
> M>     This allows to support checksum offloading also in the case of local
> M>     virtual machines or jails.
> M>     Support for epair, vtnet, and tap interfaces will be added in
> M>     separate commits.
> 
> Not a request for any action, but a general comment on the topic.
> 
> Imagine we are developing an IP stack from scratch for modern computers.  Most
> NICs do hardware checksumming, large fraction of installations run in
> containers/VMs and communicate via virtual interfaces.  With modern reality we
> would probably do not do any checksumming in the stack at all, just completely
> ignore the checksum field.  It would always be obligation of the NIC driver to
> care about the checksums.  The existing software checksumming code would be
> just a library that drivers for legacy hardware call. Local traffic won't set
> and won't check checksums neither any related flags.

I have ever had the same idea, but that requires lots of modification to in-tree
drivers and I finally abort.

Think it twice, maybe a new driver flag say `IFF_CSUM_OFFLOAD` is needed to
indicate the driver is capable to delay calculating the checksums, and 
eventually
all drivers are converted and tested.

> 
> -- 
> Gleb Smirnoff

Best regards,
Zhenlei


Reply via email to