[patch] nmount ro, rw and negated option handling

[Craig Rodrigues]

Hi,

I disagree with your patch and do not approve it.

I prefer something simpler:


Index: vfs_mount.c
===================================================================
--- vfs_mount.c	(revision 216607)
+++ vfs_mount.c	(working copy)
@@ -522,10 +522,9 @@
 	struct vfsopt *opt, *noro_opt, *tmp_opt;
 	char *fstype, *fspath, *errmsg;
 	int error, fstypelen, fspathlen, errmsg_len, errmsg_pos;
-	int has_rw, has_noro;
 
 	errmsg = fspath = NULL;
-	errmsg_len = has_noro = has_rw = fspathlen = 0;
+	errmsg_len = fspathlen = 0;
 	errmsg_pos = -1;
 
 	error = vfs_buildopts(fsoptions, &optlist);
@@ -624,7 +623,8 @@
 		}
 		else if (strcmp(opt->name, "rw") == 0) {
 			fsflags &= ~MNT_RDONLY;
-			has_rw = 1;
+			free(opt->name, M_MOUNT);
+			opt->name = strdup("noro", M_MOUNT);
 		}
 		else if (strcmp(opt->name, "ro") == 0)
 			fsflags |= MNT_RDONLY;
@@ -642,22 +642,6 @@
 	}
 
 	/*
-	 * If "rw" was specified as a mount option, and we
-	 * are trying to update a mount-point from "ro" to "rw",
-	 * we need a mount option "noro", since in vfs_mergeopts(),
-	 * "noro" will cancel "ro", but "rw" will not do anything.
-	 */
-	if (has_rw && !has_noro) {
-		noro_opt = malloc(sizeof(struct vfsopt), M_MOUNT, M_WAITOK);
-		noro_opt->name = strdup("noro", M_MOUNT);
-		noro_opt->value = NULL;
-		noro_opt->len = 0;
-		noro_opt->pos = -1;
-		noro_opt->seen = 1;
-		TAILQ_INSERT_TAIL(optlist, noro_opt, link);
-	}
-
-	/*
 	 * Be ultra-paranoid about making sure the type and fspath
 	 * variables will fit in our mp buffers, including the
 	 * terminating NUL.


ZFS can be changed to check for "rw" or "noro".

-- 
Craig Rodrigues
rodrigc at crodrigues.org



On Fri, Jan 14, 2011 at 02:24:54PM +0200, Jaakko Heinonen wrote:
> 
> Hi,
> 
> Currently nmount(2) allows a mount point to have "ro", "rw", and "noro"
> string options concurrently active. This can cause erratic behavior
> demonstrated by this example:
> 
> 1. Have mountd(8) running.
> 2. # mdconfig -a -t vnode -f ufsimg
> 3. # mount -o ro,rw /dev/md0 /mnt
> 
> After these steps the mount point has string options "ro", "rw" and
> "noro" active but the MNT_RDONLY flag is not set. Eventually this will
> lead to "ffs_sync: rofs mod" (or similar) panic because the ffs code
> marks the file system read-only due to the "ro" string option.
> (MNT_RDONLY flag is used in most places for read-only check.)
> 
> I wrote a patch to do following changes:
> 
> - vfs_equalopts() now recognizes "ro" and "rw" as equal options
> - vfs_mergeopts() uses vfs_sanitizeopts() to merge options. This ensures
>   that if the same option shows up several times (negated or not), only
>   the last one is taken in account. There is still a problem when for
>   example option "foo" and "nofoo" are merged: the "nofoo" option will
>   become an active option. This is not a regression however and
>   currently I don't know an easy way to solve this because the list of
>   valid options is not available in vfs_mergeopts().
> - vfs_donmount() always converts "norw"/"rdonly" to "ro" and "noro" to
>   "rw". Thus the mount point will always have either "rw" or "ro"
>   option. I haven't seen any in-tree file system to test for "noro" but
>   at least ZFS tests for "rw". That's why I chose "rw" instead or
>   "noro".
> 
> The patch is available here:
> 
> 	http://people.freebsd.org/~jh/patches/nmount-ro-rw.diff
> 
> Reviews and testing would be appreciated.
> 
> Here are some references to bug reports which the patch attempts to
> resolve:
> 
> http://lists.freebsd.org/pipermail/freebsd-current/2009-September/thread.html#11385
> http://lists.freebsd.org/pipermail/freebsd-questions/2009-August/204124.html
> http://www.freebsd.org/cgi/query-pr.cgi?pr=kern/133614
> http://www.freebsd.org/cgi/query-pr.cgi?pr=kern/150206
> 
> -- 
> Jaakko