F_RDLCK lock to FreeBSD NFS server fails to R/O target file (nasty little bug)

[Rick Macklem]

John wrote:
> After pondering the best way to allow the VOP_ACCESS() call to
> only query for the permissions really needed, I've come up with
> a patch that minimally adds one parameter to the nlm_get_vfs_state()
> function call with the lock type from the original argp.
> 
> http://people.freebsd.org/~jwd/nlm_prot_impl.c.accmode.patch
> 
> I'd appreciate a review and seeing what might be required to commit
> this prior to 9 release.
> 
Doing a little testing with your patch, I came across this nasty little
bug.
I have a little program that does the following:
- opens a file read/write
- waits for user input
- does read and write locks (non overlapping byte ranges)
- does an unlock of all bytes covered by the above locks

I ran it as follows:
- had a file on the NFSv3 mount point (which is using the NLM)
  not owned by the user running the above program, with file
  mode 0666
- the program opens the file read/write and waits for user input
--> on the server, I did a "chmod 622 <file>"
- gave it user input, so it then did the above locks
  (with your patch, the read lock fails although the read lock
   works for the unpatched NLM and my variant of the patch)
- it also does the unlock and doesn't complain of a failure.

---> The ugly bug is that, for your patch, the unlock has failed
     silently.
If you subsequently try and write lock a byte range overlapping
the write lock byte range above, the attempt fails, since it is
still locked by the client.
---> I don't know of an easy way to get rid of that lock!!!
     (It persists after the client unmounts the NFS file system.)

Cute, eh:-)

Btw, I suspect you can get the unpatched code to fail the same way
if you did a "chmod 600 <file>" after the write lock, but before
the unlock.

I'm almost thinking that an unlock shouldn't do a VOP_ACCESS() of
any kind, but since that isn't current behaviour???

rick

> Thanks,
> John
> 
> ----- John's Original Message -----
> > Hi Fellow NFS'ers,
> >
> >    I believe I have found the problem we've been having with read
> >    locks
> > while attaching to a FreeBSD NFS server.
> >
> >    In sys/nlm/nlm_prot_impl.c, function nlm_get_vfs_state(), there
> >    is a call
> > to VOP_ACCESS() as follows:
> >
> >         /*
> >          * Check cred.
> >          */
> >         NLM_DEBUG(3, "nlm_get_vfs_state(): Calling
> >         VOP_ACCESS(VWRITE) with cred->cr_uid=%d\n",cred->cr_uid);
> >         error = VOP_ACCESS(vs->vs_vp, VWRITE, cred, curthread);
> >         if (error) {
> >                 NLM_DEBUG(3, "nlm_get_vfs_state(): caller_name = %s
> >                 VOP_ACCESS() returns %d\n",
> >                 host->nh_caller_name, error);
> >                 goto out;
> >         }
> >
> >   The file being accessed is read only to the user, and open()ed
> >   with
> > O_RDONLY. The lock being requested is for a read.
> >
> > fd = open(filename, O_RDONLY, 0);
> > ...
> >
> > lblk.l_type = F_RDLCK;
> > lblk.l_start = 0;
> > lblk.l_whence= SEEK_SET;
> > lblk.l_len = 0;
> > lblk.l_pid = 0;
> > rc = fcntl(fd, F_SETLK, &lblk);
> >
> >    Running the above from a remote system, the lock call fails with
> > errno set to ENOLCK. Given cred->cr_uid comes in as 227 which is
> > my uid on the remote system. Since the file is R/O to me, and the
> > VOP_ACCESS() is asking for VWRITE, it fails with errno 13, EACCES,
> > Permission denied.
> >
> >    The above operations work correctly to some of our other
> > favorite big-name nfs vendors :-)
> >
> >    Opinions on the "correct" way to fix this?
> >
> > 1. Since we're only asking for a read lock, why do we need to ask
> >    for VWRITE? I may not understand an underlying requirement for
> >    the VWRITE so please feel free to educate me if needed.
> >
> >    Something like: request == F_RDLCK ? VREAD : VWRITE
> >    (need to figure out where to get the request from in this
> >    context).
> >
> > 2. Attempt VWRITE, fallback to VREAD... seems off to me though.
> >
> > 3. Other?
> >
> >    I appreciate any thoughts on this.
> >
> > Thanks,
> > John
> >
> >    While they might not follow style(9) completely, I've uploaded
> > my patch to nlm_prot.impl.c with the NLM_DEBUG() calls i've added.
> > I'd appreciate it if someone would consider committing them so
> > who ever debugs this file next will have them available.
> >
> > http://people.freebsd.org/~jwd/nlm_prot_impl.c.patch
> >
> _______________________________________________
> freebsd-fs at freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-fs
> To unsubscribe, send any mail to "freebsd-fs-unsubscribe at freebsd.org"