I recognize that wall from the house that I grew up in. Never noticed the dude sitting there though.

I dont know what a system cell is, but that's a cool myspace photo icon you're rocking there!

The real question is: are you getting ready to unleash something wicked on LJ fools on Saturday?

It's finally going in? Nice!

So long as I can specify a non-length to mean "until EOF".

nanothreads accomplish a similar task even more efficiently dont they? difference being file support I suppose...

Huh?

splice lets you do fd->fd copies all in-kernel without copying to/from userspace. What is a nanothread and how does that help?

your userpic is adorable.

nanothreads reduce/eliminate the overhead of copy to/from user/kernel. it is mainly useful in multithreaded applications and for multi-core or hyper threaded kernels.. but i have seen some other interesting uses of it.

anyway, here is one of the many papers on it:
http://www.hpclab.ceid.upatras.gr/publications/hpclab-tr-021298.pdf.gz

Oh look, kernel support for CopyFile() :)

It's been a long time coming, yes.

Interesting. Maybe I'm missing something, but it looks like another device-specific version of sendfile(). Why don't they implement a generic sendfilev() that works with any type of descriptor?

splice is the generic version. They're talking about implementing sendfile now in terms of splice.

Thought it was weird you hadn't commented on this, but I missed the post the first time around :)

splice()! Woot!

I'm confused on how to buffer socket data and replay using splice though... Looks like socket -> pipe buffer, pull up to 16 pages (was that limit ever removed?) into the pipe buffer, then -> splice() onto the remote socket. Along with cork/userspace tricks for adding headers.

You can't splice from a socket fd to a pipe buffer yet. Eventually they'll have socket to socket, with an implicit pipe buffer. That's when things get really fun.

Ahh okay. I had read about the socket to socket coming soon, but no details. Thanks :)