On Mon, May 09, 2022 at 02:54:14PM -0500, Serge E. Hallyn wrote:

On Wed, Apr 20, 2022 at 10:06:08AM -0400, Stefan Berger wrote:

quoted

From: Christian Brauner <brauner@kernel.org>

When securityfs creates a new file or directory via
securityfs_create_dentry() it will take an additional reference on the
newly created dentry after it has attached the new inode to the new
dentry and added it to the hashqueues.
If we contrast this with debugfs which has the same underlying logic as
securityfs. It uses a similar pairing as securityfs. Where securityfs
has the securityfs_create_dentry() and securityfs_remove() pairing,
debugfs has the __debugfs_create_file() and debugfs_remove() pairing.

In contrast to securityfs, debugfs doesn't take an additional reference
on the newly created dentry in __debugfs_create_file() which would need
to be put in debugfs_remove().

The additional dget() isn't a problem per se. In the current
implementation of securityfs each created dentry pins the filesystem via

Is 'via' an extra word here or is there a missing word?

I'll delay the rest of my response as the missing word may answer my
remaining question :)

quoted

until it is removed. Since it is virtually guaranteed that there is at
least one user of securityfs that has created dentries the initial
securityfs mount cannot go away until all dentries have been removed.

Since most of the users of the initial securityfs mount don't go away
until the system is shutdown the initial securityfs won't go away when
unmounted. Instead a mount will usually surface the same superblock as
before. The additional dget() doesn't matter in this scenario since it
is required that all dentries have been cleaned up by the respective
users before the superblock can be destroyed, i.e. superblock shutdown
is tied to the lifetime of the associated dentries.

However, in order to support ima namespaces we need to extend securityfs
to support being mounted outside of the initial user namespace. For
namespaced users the pinning logic doesn't make sense. Whereas in the
initial namespace the securityfs instance and the associated data
structures of its users can't go away for reason explained earlier users
of non-initial securityfs instances do go away when the last users of
the namespace are gone.

So for those users we neither want to duplicate the pinning logic nor
make the global securityfs instance display different information based
on the namespace. Both options would be really messy and hacky.

Instead we will simply give each namespace its own securityfs instance
similar to how each ipc namespace has its own mqueue instance and all
entries in there are cleaned up on umount or when the last user of the
associated namespace is gone.

This means that the superblock's lifetime isn't tied to the dentries.
Instead the last umount, without any fds kept open, will trigger a clean
shutdown. But now the additional dget() gets in the way. Instead of
being able to rely on the generic superblock shutdown logic we would
need to drop the additional dentry reference during superblock shutdown
for all associated users. That would force the use of a generic
coordination mechanism for current and future users of securityfs which
is unnecessary. Simply remove the additional dget() in
securityfs_dentry_create().

In securityfs_remove() we will call dget() to take an additional
reference on the dentry about to be removed. After simple_unlink() or
simple_rmdir() have dropped the dentry refcount we can call d_delete()
which will either turn the dentry into negative dentry if our earlier
dget() is the only reference to the dentry, i.e. it has no other users,
or remove it from the hashqueues in case there are additional users.

All of these changes should not have any effect on the userspace
semantics of the initial securityfs mount.

Signed-off-by: Christian Brauner <brauner@kernel.org>
Cc: John Johansen <john.johansen@canonical.com>
Cc: Matthew Garrett <mjg59@srcf.ucam.org>
Cc: Micah Morton <mortonm@chromium.org>
Cc: Kentaro Takeda <takedakn@nttdata.co.jp>
Cc: James Morris <jmorris@namei.org>
Cc: Jarkko Sakkinen <jarkko@kernel.org>
Signed-off-by: Stefan Berger <stefanb@linux.ibm.com>
Reviewed-by: Mimi Zohar <zohar@linux.ibm.com>
---
 security/inode.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/security/inode.c b/security/inode.c
index 6c326939750d..13e6780c4444 100644
--- a/security/inode.c
+++ b/security/inode.c

@@ -159,7 +159,6 @@ static struct dentry *securityfs_create_dentry(const char *name, umode_t mode,
 		inode->i_fop = fops;
 	}
 	d_instantiate(dentry, inode);
-	dget(dentry);
 	inode_unlock(dir);
 	return dentry;
 

@@ -302,10 +301,12 @@ void securityfs_remove(struct dentry *dentry)
 	dir = d_inode(dentry->d_parent);
 	inode_lock(dir);
 	if (simple_positive(dentry)) {
+		dget(dentry);
 		if (d_is_dir(dentry))
 			simple_rmdir(dir, dentry);

Hm, so I realize your patch isn't introducing this, but is the
fact that we ignore the possible -ENOTEMPTY return value of
simple_rmdir() not a problem?

quoted

 		else
 			simple_unlink(dir, dentry);
+		d_delete(dentry);

I'm mostly trying to convince myself that the fact that there is not
a matching dput being removed (to match the dget being removed above)
is ok.  I do think it is, but that belief seems to dictate that right
now dentries must never be being released.

Otherwise, it seems like there must be cases where the next dput could
be called on a dentry that has been freed.

quoted

 		dput(dentry);
 	}
 	inode_unlock(dir);
-- 
2.34.1