Jason Thorpe <thorpej@shagadelic.org> wrote:
> I, for one, would like to see AIO that did not require a thread that  
> could block, but rather use of completely asynchronous APIs that  
> already exist in our kernel.
Are you talking about some similar way how Solaris goes? I agree, but do we
have a convenient API to do that? I am not sure how it looks..
There might be some differentiation by fp->f_type, eg. maybe for sockets,
also, as mentioned in TODO - separate handling for VCHR devices.
Talking about threads, it is likely that after interrupts-as-threads work we
will have very light threads (in Linux world known as "threadlets/syslets"),
which will became a real LWPs only in block case. Then many things in the
system could be improved, including AIO - it may be switched to 1:1 (one I/O
operation - 1 light-thread) model.

-- 
Best regards,
Mindaugas
www.NetBSD.org