cvs commit: src/sys/nfsclient nfs_bio.c nfs_vfsops.c nfsargs.h nfsmount.h src/sys/sys buf.h bufobj.h src/sys/kern vfs_bio.c

[Brian Fundakowski Feldman]

On Sun, Jun 12, 2005 at 06:30:49AM -0400, Daniel Eischen wrote:
> On Sun, 12 Jun 2005, Alfred Perlstein wrote:
> 
> > * Brian Fundakowski Feldman <green at freebsd.org> [050612 01:26] wrote:
> > > On Sun, Jun 12, 2005 at 01:08:33AM -0700, Alfred Perlstein wrote:
> > > >
> > > > Seriously, have you tested what happens to a libc_r app that
> > > > opens an nfs file F_SYNC?  My guess is that it's not pretty.
> > >
> > > This code path is related to O_NONBLOCK, not O_FSYNC.  O_FSYNC is
> > > synonymous with the slow fallback path that large transactional block
> > > now takes, rather than deadlocking.  O_NONBLOCK really means that
> > > whatever they do, they are required to check for EAGAIN.
> >
> > To make it perfectly clear.
> >
> > If an application linked against libc_r opens a file with O_FSYNC.
> > Libc_r will set O_NONBLOCK (it does so for each open(2))
> > A write on that descriptor will return EAGAIN (to libc_r)
> > Libc_r will then attempt to select(2) on this decriptor, which
> >   will return "ready" (as do all select(2)'s on disk files)
> >
> > The question is:
> >
> > Will Libc_r then busy spin?
> 
> Yes, for the most part.
> 
> > If so, how many other apps might get screwed just sometimes (over
> > nfs) because only _half_ of this "solution" is implemented?
> >
> > Or is my thinking on this wrong?
> 
> I think I agree with Alfred.
> 
> If select() returns ready, then you should be able to write
> some part of your buffer.  Also, anyone using a file descriptor
> in non-blocking mode should expect short writes and loop until
> the entire buffer has been written.
> 
> >From my understanding, disk I/O has always returned ready, then
> blocked in the kernel if necessary.  I think that if we are going
> to start honoring non-blocking mode for disk I/O (or NFS, whatever),
> it should be done fully.  That means you allow short writes when
> select() returns ready, and select() doesn't return ready if no
> data can be written.

 That's a good point.  I wasn't more than vaguely aware of libc_r's
continued existance and usage of such things.  Fixing select(3) to
match up would be easiest, right?  Short writes are definitely not
allowed for a non-socket, though.

-- 
Brian Fundakowski Feldman                           \'[ FreeBSD ]''''''''''\
  <> green at FreeBSD.org                               \  The Power to Serve! \
 Opinions expressed are my own.                       \,,,,,,,,,,,,,,,,,,,,,,\