nfsd: notify_change needs elevated write count
J. Bruce Fields [Mon, 24 Feb 2014 19:59:47 +0000 (14:59 -0500)]
commit 9f67f189939eccaa54f3d2c9cf10788abaf2d584 upstream.

Looks like this bug has been here since these write counts were
introduced, not sure why it was just noticed now.

Thanks also to Jan Kara for pointing out the problem.

Reported-by: Matthew Rahtz <mrahtz@rapitasystems.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

fs/nfsd/vfs.c

index 62fd661..d9b298c 100644 (file)
@@ -406,6 +406,7 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
        umode_t         ftype = 0;
        __be32          err;
        int             host_err;
+       bool            get_write_count;
        int             size_change = 0;
 
        if (iap->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_SIZE))
@@ -413,10 +414,18 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
        if (iap->ia_valid & ATTR_SIZE)
                ftype = S_IFREG;
 
+       /* Callers that do fh_verify should do the fh_want_write: */
+       get_write_count = !fhp->fh_dentry;
+
        /* Get inode */
        err = fh_verify(rqstp, fhp, ftype, accmode);
        if (err)
                goto out;
+       if (get_write_count) {
+               host_err = fh_want_write(fhp);
+               if (host_err)
+                       return nfserrno(host_err);
+       }
 
        dentry = fhp->fh_dentry;
        inode = dentry->d_inode;