Re: [PATCH v14 03/22] vfs: Add MAY_DELETE_SELF and MAY_DELETE_CHILD permission flags

On Fri, Nov 6, 2015 at 10:26 PM, Andreas Dilger [off-list ref] wrote:

On Nov 5, 2015, at 4:39 AM, Andreas Gruenbacher [off-list ref] wrote:

quoted

Normally, deleting a file requires MAY_WRITE access to the parent
directory.  With richacls, a file may be deleted with MAY_DELETE_CHILD access
to the parent directory or with MAY_DELETE_SELF access to the file.

To support that, pass the MAY_DELETE_CHILD mask flag to inode_permission()
when checking for delete access inside a directory, and MAY_DELETE_SELF
when checking for delete access to a file itelf.

The MAY_DELETE_SELF permission overrides the sticky directory check.

Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Reviewed-by: J. Bruce Fields <redacted>
---
fs/namei.c         | 21 ++++++++++++---------
include/linux/fs.h |  2 ++
2 files changed, 14 insertions(+), 9 deletions(-)

diff --git a/fs/namei.c b/fs/namei.c
index 0259392..2eab19e 100644
--- a/fs/namei.c
+++ b/fs/namei.c

@@ -453,9 +453,9 @@ static int sb_permission(struct super_block *sb, struct inode *inode, int mask)
 * this, letting us set arbitrary permissions for filesystem access without
 * changing the "normal" UIDs which are used for other things.
 *
- * MAY_WRITE must be set in @mask whenever MAY_APPEND, MAY_CREATE_FILE, or
- * MAY_CREATE_DIR are set.  That way, file systems that don't support these
- * permissions will check for MAY_WRITE instead.
+ * MAY_WRITE must be set in @mask whenever MAY_APPEND, MAY_CREATE_FILE,
+ * MAY_CREATE_DIR, or MAY_DELETE_CHILD are set.  That way, file systems that
+ * don't support these permissions will check for MAY_WRITE instead.
 */

int inode_permission(struct inode *inode, int mask)
{

@@ -2555,7 +2555,7 @@ static int may_delete(struct inode *dir, struct dentry *victim,
                    bool isdir, bool replace)

{
      struct inode *inode = d_backing_inode(victim);
-     int error, mask = MAY_WRITE | MAY_EXEC;
+     int error, mask = MAY_EXEC;

      if (d_is_negative(victim))
              return -ENOENT;

@@ -2565,15 +2565,18 @@ static int may_delete(struct inode *dir, struct dentry *victim,
      audit_inode_child(dir, victim, AUDIT_TYPE_CHILD_DELETE);

      if (replace)
-             mask |= isdir ? MAY_CREATE_DIR : MAY_CREATE_FILE;
-     error = inode_permission(dir, mask);
+             mask |= MAY_WRITE | (isdir ? MAY_CREATE_DIR : MAY_CREATE_FILE);
+     error = inode_permission(dir, mask | MAY_WRITE | MAY_DELETE_CHILD);
+     if (!error && check_sticky(dir, inode))
+             error = -EPERM;
+     if (error && IS_RICHACL(inode) &&
+         inode_permission(inode, MAY_DELETE_SELF) == 0)
+             error = 0;

This looks like a serious bug, as it is overriding other errors returned
from inode_permission() such as -EROFS from sb_permission() or even from
generic_permission->acl_permission_check().  Clearing the error returned
by an earlier check doesn't seem safe, only new errors should be added.

See further below ...

The call to inode_permission(inode) is also duplicating the sb_permission()
check from inode_permission(dir), so at most should be __inode_permission().

This could be optimized, but that would make the code harder to read.
I don't think it really matters: this only affects file with the
MAY_DELETE_SELF permission, and sb_permissions is really cheap anyway.

It looks like this would be correct if you check MAY_DELETE_SELF together
with check_sticky():

        if (!error && check_sticky(dir, inode) &&
            !__inode_permission(inode, MAY_DELETE_SELF))
                error = -EPERM;

Nope, this version would deny users with MAY_DELETE_SELF access to a
file and MAY_EXEC access to the directory the right to delete the file
because @error would already be set then.

You are right that the check for MAY_DELETE_SELF overrides too much
here though; we need an additional inode_permission(dir, ...) check in
there.

I've pushed a fixed version here that I will do some more testing on
before reposting:

  git://git.kernel.org/pub/scm/linux/kernel/git/agruen/linux-richacl.git
richacl-wip

I ommitted the IS_RICHACL() check here, since that should probably be in
__inode_permission() when MAY_DELETE_SELF is passed?

The IS_RICHACL() check is indeed done in inode_permission ->
__inode_permission -> do_inode_permission -> generic_permission ->
acl_permission_check. I've duplicated that check here to avoid slowing
down may_delete in the common case where the filesystem doesn't
support richacls.

Thanks,
Andreas

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs