On Thu, Mar 06, 2008 at 09:49:48AM -0800, David O'Brien wrote:
> On Tue, Aug 29, 2006 at 10:00:12PM +0000, Mohan Srinivasan wrote:
> > mohans      2006-08-29 22:00:12 UTC
> >   FreeBSD src repository
> >   Modified files:
> >     sys/nfsclient        nfs_socket.c 
> >   Log:
> >   Fix for a deadlock triggered by a 'umount -f' causing a NFS request to 
> > never
> >   retransmit (or return). Thanks to John Baldwin for helping nail this one.
> >   Revision  Changes    Path
> >   1.144     +14 -2     src/sys/nfsclient/nfs_socket.c
> > http://cvsweb.freebsd.org/src/sys/nfsclient/nfs_socket.c.diff?r1=1.143&r2=1.144
> 
> How does this look for a RELENG_6 version of this fix?
> 
> Index: nfsclient/nfs_socket.c
> ===================================================================
> RCS file: /home/ncvs/src/sys/nfsclient/nfs_socket.c,v
> retrieving revision 1.125.2.18
> diff -u -p -r1.125.2.18 nfs_socket.c
> --- nfsclient/nfs_socket.c    17 Jan 2008 21:04:51 -0000      1.125.2.18
> +++ nfsclient/nfs_socket.c    25 Feb 2008 10:26:59 -0000
> @@ -1323,6 +1323,18 @@ nfs_timer(void *arg)
>                       continue;
>               if (nfs_sigintr(nmp, rep, rep->r_td))
>                       continue;
> +             else {
> +                     /*
> +                      * Terminate request if force-unmount in progress.
> +                      * Note that NFS could have vfs_busy'ed the mount,
> +                      * causing the unmount to wait for the mnt_lock, making
> +                      * this bit of logic necessary.
> +                      */
> +                     if (rep->r_nmp->nm_mountp->mnt_kern_flag & 
> MNTK_UNMOUNTF) {
> +                             nfs_softterm(rep);
> +                             continue;
> +                     }
> +             }
>               if (nmp->nm_tprintf_initial_delay != 0 &&
>                   (rep->r_rexmit > 2 || (rep->r_flags & R_RESENDERR)) &&
>                   rep->r_lastmsg + nmp->nm_tprintf_delay < now.tv_sec) {
> 
The reason for this change was the rev. 1.180 of the nfs_vfsops.c, that
never makes into RELENG_6. NFS client in RELENG_7 is MPsafe and
sufficiently reworked.

I think that scenario that leads to that deadlock is:

The thread A doing fstatfs() vfs_busy()ed the mountpoint, and then:
- unmount (thread B) waits for the mnt_lock;
- the request sent by the thread A or response to that request is lost and
  never retransmitted due to forced umount being performed.
This ends up in the deadlock.

Attachment: pgp6cAUsZgrf5.pgp
Description: PGP signature

Reply via email to