If you create a subvolume in btrfs and access it (by name) without
mounting it, then the subvolume looks like a separate mount to some
extent, returning a different st_dev to stat(), but it doesn't look like
a separate mount in that it isn't listed in /proc/mounts. This
inconsistency can confuse tools.
This patch causes these subvolumes to become separate mounts by using
the VFS' automount functionality, much like NFS uses automount when it
discovered mountpoints on the server.
The VFS currently makes it impossible to auto-mount a directory on to itself
(i.e. a bind mount). For NFS this isn't a problem as a new superblock
is created for the child filesystem so there are two separate dentries
(and inodes) for the one directory: one in the parent filesystem, one in
the child (note that the two superblocks share a common connection to
the server so there is still a lot of commonality).
BTRFS has chosen instead to use a single superblock for all subvolumes.
This results in a single dentry for the subvol-root. A dentry which
must be auto-mounted on itself.
This creates 2 problems.
Firstly, we want to be selective about when the dentry triggers an
automount. When the dentry is not the root of a vfsmount, we do want to
trigger an automount. When it is the root, we don't.
I have modified follow_managed() to understand that if ->d_manage()
returns 1, then an automount should be suppressed when
path->mnt->mnt_root == path->dentry.
Secondly, finish_automount() explicitly tests for the case of mounting a
dentry on top of itself and fails with -ELOOP. I've changed this to
only fail if the mounted-on directory is the root in its own vfsmount.
So now mounting a dentry on itself can be OK, but mounting the same
dentry on top of that will cause -ELOOP.
With those two VFS changes, it is straight forward to enable automounts for
btrfs subvolumes and to use clone_private_mount to create the required
vfsmount.
I have not added infrastructure to time-out the submounts after a period
of inactivity in the way that NFS does. If that is desirable (I'm not
sure of the benefits) it can easily be added as a subsequent patch.
If this approach is acceptable, I will resubmit addressing any issues
raised and with proper updates for the automount documentation.
Thanks for any comments,
NeilBrown
diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index 611b66d73e80..e96c53590f72 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -5621,6 +5621,23 @@ static void btrfs_dentry_release(struct dentry *dentry)
kfree(dentry->d_fsdata);
}
+static int btrfs_dentry_manage(struct dentry *dentry, bool in_rcu)
+{
+ /* This is a 'rebind automount'. So only trigger automount
+ * when the dentry isn't the root of a mountpoint.
+ */
+ return 1;
+}
+
+static struct vfsmount *btrfs_dentry_automount(struct path *path)
+{
+ struct vfsmount *mnt;
+ mnt = clone_private_mount(path);
+ if (!IS_ERR(mnt))
+ mntget(mnt);
+ return mnt;
+}
+
static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
unsigned int flags)
{
@@ -5633,6 +5650,8 @@ static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
else
return ERR_CAST(inode);
}
+ if (inode && inode->i_ino == BTRFS_FIRST_FREE_OBJECTID)
+ dentry->d_flags |= DCACHE_MANAGE_TRANSIT | DCACHE_NEED_AUTOMOUNT;
return d_splice_alias(inode, dentry);
}
@@ -9990,4 +10009,6 @@ static const struct inode_operations btrfs_symlink_inode_operations = {
const struct dentry_operations btrfs_dentry_operations = {
.d_delete = btrfs_dentry_delete,
.d_release = btrfs_dentry_release,
+ .d_manage = btrfs_dentry_manage,
+ .d_automount = btrfs_dentry_automount,
};
diff --git a/fs/namei.c b/fs/namei.c
index 33e9495a3129..07e4bbbadae1 100644
--- a/fs/namei.c
+++ b/fs/namei.c
@@ -1178,6 +1178,7 @@ static int follow_managed(struct path *path, struct nameidata *nd)
unlikely(managed != 0)) {
/* Allow the filesystem to manage the transit without i_mutex
* being held. */
+ ret = 0;
if (managed & DCACHE_MANAGE_TRANSIT) {
BUG_ON(!path->dentry->d_op);
BUG_ON(!path->dentry->d_op->d_manage);
@@ -1207,7 +1208,12 @@ static int follow_managed(struct path *path, struct nameidata *nd)
/* Handle an automount point */
if (managed & DCACHE_NEED_AUTOMOUNT) {
- ret = follow_automount(path, nd, &need_mntput);
+ if (ret == 1 && path->mnt->mnt_root == path->dentry) {
+ /* only automount when not the root */
+ ret = 0;
+ break;
+ } else
+ ret = follow_automount(path, nd, &need_mntput);
if (ret < 0)
break;
continue;
@@ -1219,7 +1225,7 @@ static int follow_managed(struct path *path, struct nameidata *nd)
if (need_mntput && path->mnt == mnt)
mntput(path->mnt);
- if (ret == -EISDIR)
+ if (ret == -EISDIR || ret > 0)
ret = 0;
if (need_mntput)
nd->flags |= LOOKUP_JUMPED;
diff --git a/fs/namespace.c b/fs/namespace.c
index 0570729c87fd..adfcb125bef0 100644
--- a/fs/namespace.c
+++ b/fs/namespace.c
@@ -2431,7 +2431,8 @@ int finish_automount(struct vfsmount *m, struct path *path)
BUG_ON(mnt_get_count(mnt) < 2);
if (m->mnt_sb == path->mnt->mnt_sb &&
- m->mnt_root == path->dentry) {
+ m->mnt_root == path->dentry &&
+ path->mnt->mnt_root == path->dentry) {
err = -ELOOP;
goto fail;
}
Attachment:
signature.asc
Description: PGP signature
