Hello!
> i believe it might still make some limited sense for normal sendmsg()
> and higher MTUs (or 8k NFS) - we could copy & checksum stuff into the
> ->tcp_page if SG is possible and thus the SG capability improves the VM.
> (because we can allocate at PAGE_SIZE granularity.)
There is no problem in doing this for udp and af_unix sockets.
But I suspect it is not very useful. NFS allocations do not stress
VM too much and high order allocations are still cheaper than allocations
of vector of pages. Seems, this requires some caching to allocate
vectors of pages.
For tcp it is also not difficult, though requires combined segmentizer
and pager. Easy to do, but not useful because interfaces with large
mtu not doing checksummming in hardware... do they exist if to forget
about buggy SK?
Real advantage is that this scheme allows to get rid of too early
fragmentation, but any efforts in this direction are worthful only
after netfilter and all the popular devices (f.e. including ppp)
start to understand non-linear skbs.
Alexey
|