AIO work in librt

AIO work in librt

[Mark Brown]

Given this conversation from last year:

Andreas Jaeger wrote:
> Ulrich Drepper <drepper at redhat dot com> writes:
> > Andreas Jaeger wrote:
> >> AFAIK Ben's kernel level implementation is in current 2.5 kernels.
> >> Can we base now a glibc implementation on this?  Anybody already
> >> working on this?
> >
> > I doN't think the kernel side is complete enough.
> 
> Ben, what is still missing on the kernel side?  Looking at fs/aio.c I
> thought it was complete.

We are thinking about making this work in librt (and submitting it, of 
course).
Andreas, did you learn anything from Ben? Anyone else know of a
factor that keeps this from being able to work?


-------------------
Mark S. Brown 
IBM Systems Group, Linux Technology Center
bmark@us.ibm.com

Re: AIO work in librt

[Ulrich Drepper]

Mark Brown wrote:

> We are thinking about making this work in librt (and submitting it, of 
> course).

Making what work?  RHEL ships with some support for kernel aio already.
 This code (or more specifically: the kernel side) is not yet in
sufficient shape to be accepted.  That's what's keeping it back.  The
librt side should be done for the current kernel aio approach which is
everything but sane.

So the task is to design an implement a workable aio implementation in
the kernel and I haven't seen anything like this lately.

-- 
➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖

Re: AIO work in librt

[Jakub Jelinek]

On Fri, Apr 23, 2004 at 11:17:09AM -0500, Mark Brown wrote:
> Given this conversation from last year:
> 
> Andreas Jaeger wrote:
> > Ulrich Drepper <drepper at redhat dot com> writes:
> > > Andreas Jaeger wrote:
> > >> AFAIK Ben's kernel level implementation is in current 2.5 kernels.
> > >> Can we base now a glibc implementation on this?  Anybody already
> > >> working on this?
> > >
> > > I doN't think the kernel side is complete enough.
> > 
> > Ben, what is still missing on the kernel side?  Looking at fs/aio.c I
> > thought it was complete.
> 
> We are thinking about making this work in librt (and submitting it, of 
> course).
> Andreas, did you learn anything from Ben? Anyone else know of a
> factor that keeps this from being able to work?

RHEL glibc already includes the rtkaio glibc add-on which is librt using
the kernel AIO interface.
Unfortunately, the kernel side is very limited and doesn't have good
support for the things userland POSIX AIO needs (e.g. in the area of
notification and cancellation, or that many things are serialized in
the kernel rather than truely asynchronous; and from what I've heard
the 2.6 aio is even in worse shape than what we used to have in 2.4.x RHEL
kernels).

	Jakub

Re: AIO work in librt

[Mark Brown]

Jakub Jelinekl writes (as did Ulrich in another note):
> RHEL glibc already includes the rtkaio glibc add-on which is librt using
> the kernel AIO interface.

We were looking for something that could go into the main tree, but given 
the
status below...

> Unfortunately, the kernel side is very limited and doesn't have good
> support for the things userland POSIX AIO needs (e.g. in the area of
> notification and cancellation, or that many things are serialized in
> the kernel rather than truely asynchronous; and from what I've heard
> the 2.6 aio is even in worse shape than what we used to have in 2.4.x 
RHEL
> kernels).

Not being an aio expert myself, I didn't know things were in this shape.
I'll pass your opinions here to the team that's proposing the work.


-------------------
Mark S. Brown 
bmark@us.ibm.com