Re: [patch 13/13] signalfd/timerfd/asyncfd v5 - KAIO asyncfd support (example/maybe-broken) ...

[Davide Libenzi <davidel@xmailserver.org>]

On Wed, 14 Mar 2007, Benjamin LaHaise wrote:

> On Wed, Mar 14, 2007 at 04:41:58PM -0700, Davide Libenzi wrote:
> > Yeah, of course. I do not plan revolutions. Just asking if it's a possible 
> > thing to do. I can mlock the userspace ring, if imposing that burden over 
> > aio_complete() is seen as too heavy.
> 
> I'm not sure I follow what you're doing -- why isn't asyncfd merely calling 
> io_getevents() instead of reinventing everything the ringbuffer does?  The 
> aio ringbuffer is already locked in memory.  Fwiw, the aio ringbuffer was 
> originally wired up to a file descriptor, but that gave way to the actual 
> syscall in order to enforce proper typechecking and typical usage scenarios 
> with timeouts.

The purpose of asyncfd is to provide a pollable (by the mean of 
f_op->poll) device that can be hosted inside a standard select/poll/epoll 
wait subsystem, and that, at the same time, provide a zero-copy way for 
kernel code (KAIO and syslets/threadlets were my thought) to deliver 
results to userspace.



> Also, there have been patches floating around for aio_poll and a way to get 
> epoll wakeups into the aio event queue.  They deserve serious consideration 
> if this asyncfd seems necessary.

I don't want to talk about the AIO poll code, because last time I saw it, 
it did not look shiny.
But I think we can agree that ppl needs to have a way to wait for both 
block I/O (covered by either KAIO or syslets/threadlets) and all the other 
world (covered by epoll). This has been pretty clear for me, looking at 
the continuous request I got to provide block I/O completions through 
epoll, and looking at the hackage that ppl has currently to do in 
userspace to achieve that.
Now that I'm seeing I can wait for both block and net I/O, I got excited ;)



- Davide