[PATCH] fuse: propagate SB_POSIXACL to submounts

[Tarun Khandelwal <tkhandelwal32@gmail.com>]

From: Tarun Khandelwal <tarun@supabase.io>

fuse_fill_super_common() always sets SB_POSIXACL on the superblock, but
fuse_fill_super_submount() never does. Submount superblocks are created
from an fs_context with no sb_flags, so the superblock of an automounted
FUSE_ATTR_SUBMOUNT directory (e.g. a virtiofs export containing another
host mount) lacks SB_POSIXACL, even though it shares the fuse_conn, and
thus fc->posix_acl, with its parent.

Since IS_POSIXACL() is false on such a submount, the VFS:

 - fails setting POSIX ACLs with -EOPNOTSUPP and hides existing ACLs,
   without sending any request to the server;
 - applies the umask itself on create, so default ACLs on the server are
   not honoured for new files;
 - skips ACL checks in acl_permission_check(), so with
   default_permissions (implied by FUSE_POSIX_ACL) access is decided by
   the mode bits alone and restrictive ACL entries are not enforced.

Copy SB_POSIXACL from the parent superblock, as is already done for the
other superblock fields inherited from it. It cannot be set in
fuse_sb_defaults(), because fuse_fill_super_common() checks whether the
flag was passed at mount time to decide fc->dont_mask before setting it.

Reproducer, with virtiofsd 1.10.0 started with --posix-acl and an export
containing a nested mount:

  host# mount -t tmpfs none /srv/share
  host# mkdir /srv/share/child
  host# mount -t tmpfs none /srv/share/child
  host# touch /srv/share/rootfile /srv/share/child/childfile
  guest# mount -t virtiofs <tag> /mnt
  guest# setfacl -m u:12345:rw /mnt/rootfile           # works
  guest# setfacl -m u:12345:rw /mnt/child/childfile
  setfacl: /mnt/child/childfile: Operation not supported

With this patch the second setfacl succeeds, ACLs set on the host are
reported by getfacl on the submount, and default ACLs and ACL-based
access checks behave the same as on the root mount. Tested on v7.3-rc5
in a QEMU guest.

Signed-off-by: Tarun Khandelwal <tarun@supabase.io>
---
I wasn't sure whether leaving SB_POSIXACL off submount superblocks was
intentional. I would have opened an issue first, but since the fix is a
one-liner I'm sending the patch directly. If there is a reason for the
current behaviour, I'm happy to rework this or drop it.

 fs/fuse/inode.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/fs/fuse/inode.c b/fs/fuse/inode.c
index e9552be36..ef3c7fac4 100644
--- a/fs/fuse/inode.c
+++ b/fs/fuse/inode.c
@@ -1629,6 +1629,8 @@ static int fuse_fill_super_submount(struct super_block *sb,
 	WARN_ON(sb->s_bdi != &noop_backing_dev_info);
 	sb->s_bdi = bdi_get(parent_sb->s_bdi);
 
+	sb->s_flags |= parent_sb->s_flags & SB_POSIXACL;
+
 	sb->s_xattr = parent_sb->s_xattr;
 	sb->s_export_op = parent_sb->s_export_op;
 	sb->s_time_gran = parent_sb->s_time_gran;
-- 
2.43.0