Patchset to add asynchronous open/close to master

[Jeremy Allison]

On Thu, Jun 21, 2012 at 05:38:42PM -0400, simo wrote:
> > 
> > One more point - the reason for using the
> > syscall(XXX, ...) instead of the setfsuid()
> > calls is because glibc could change
> > in the future to catch the setfsuid() calls
> > to do the SETXID uid signalling and we
> > wouldn't know. The syscall() interface
> > is guarenteed to stay the same on the
> > Linux kernel without any glibc interference.
> 
> As far as I know setfsuid is meant to escape the posix madness, so that
> should not happen as there is software relying on setfsuid exactly to
> avoid having the fsuid changed arbitrarily by glibc.
> However any seteuid will reset also the fsuid in the kernel, so you
> cannot mix seteuid calls with setfsuid ones.

Unfortunately even if I used setfsuid I'd still need
to wrap setgroups to ensure the thread has the correct
suplementary groups list.

I have a patch I'm working on that wraps the
setreuid/setregid/setuid/setgid/setgroups calls
(which are the only ones that we use internally)
and makes them direct syscalls to avoid the glibc
per-process manipulations.

This will put us back to the state we *thought*
we were running in :-), and re-fix the lost wakeup
problem with raw glibc aio calls.

Once that is in the aio create code become much
simpler, as I can remove all the paranoia around
race conditions and just ensure I'm running the
thread as the user that invoked the create call.

Jeremy.