namespace: unhash a dentry before detaching the mounts on it

In the Linux kernel, the following vulnerability has been resolved:

namespace: unhash a dentry before detaching the mounts on it

unlink(), rmdir() and rename() remove the entry from the filesystem,
call detach_mounts() on the entry with the inode locked and then call
d_delete() once the inode is unlocked.

The thing is that between detach_mounts() and d_delete() the dentry is
still hashed and positive. But after detach_mounts() nothing covers the
dentry anymore. A lookup that finds this dentry in the dcache can
uncover what the mounts hid.

That's a problem when unlinking files or directories that are
mountpoints in other mount namespaces. Everybody who had the underlying
entry covered can race the detach_mount() call until d_delete() has run.

The race window isn't all that small. It encompasses namespace_unlock()
with a full synchronize_rcu_expedited() grace period and the
inode_unlock() of the entry.

In my experiments three walkers that kept trying read an overmounted
file in 11938 of the 16140 unlinks that removed its mountpoint from a
bind mount of the filesystem within a minute.

Fun fact, d_invalidate() has the same ordering problem but gets it
right. It unhashes the dentry first and detaches the mounts afterwards.
Let's do the same in __detach_mounts():

- A lookup that hasn't found the dentry yet misses it in the dcache and
waits for the directory lock that the caller holds until the name is
gone for good.

- A lockless lookup that found it already fails the mount_lock check
while the dentry still counts as a mountpoint and the d_seq check once
it doesn't, and retries.

This fixes the lockless path. We still need to fix the reference count
lookup in a follow-up patch.

So d_drop() the dentry. The dentry stays positive and held but can't be
found anymore. Then proceed with the detach and unlink.

Reproducer:

The reads were counted with the program below, three walkers for 60 s
in a VM with 4 CPUs. The race window is stretched with the debug patch
pasted here. fs.detach_race_walk_us sleeps in lookup_fast() once it
found a mountpoint dentry and before its mounts are crossed.
fs.detach_race_unlink_us sleeps in vfs_unlink() between detach_mounts()
and d_delete().

// SPDX-License-Identifier: GPL-2.0
/*
* unlink_covered: unlink a file that is a mountpoint in a detached copy of
* its mount, against lookups that open it through that copy.
*
* A is a tmpfs with the file f1 (content SECRET_F) and the plain file
* MARK_A. A' is an open_tree(OPEN_TREE_CLONE) copy of A with MARK_A bound
* on A'/f1, held through an O_PATH fd on its root once the tree fd is
* closed. Walkers open f1 through that fd while the driver unlinks A/f1,
* where nothing is mounted on it. A walker may read MARK_A or get ENOENT.
* A read of SECRET_F is a hit: the name was found after its mount was
* gone. The lockless walks use openat2(RESOLVE_CACHED).
*
* usage: unlink_covered [-t seconds] [-w walkers]
*/
#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif
#include <errno.h>
#include <fcntl.h>
#include <pthread.h>
#include <stdatomic.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <sys/mount.h>
#include <sys/stat.h>
#include <sys/syscall.h>
#include <linux/openat2.h>

#ifndef __NR_open_tree
#define __NR_open_tree 428
#endif
#ifndef __NR_move_mount
#define __NR_move_mount 429
#endif
#ifndef __NR_openat2
#define __NR_openat2 437
#endif
#ifndef OPEN_TREE_CLONE
#define OPEN_TREE_CLONE 1
#endif
#ifndef OPEN_TREE_CLOEXEC
#define OPEN_TREE_CLOEXEC O_CLOEXEC
#endif
#ifndef MOVE_MOUNT_F_EMPTY_PATH
#define MOVE_MOUNT_F_EMPTY_PATH 0x00000004
#endif

#define WORK "/tmp/uc"
#define ADIR WORK "/A"

static int duration = 60, nwalkers = 3;
static atomic_int stop, writer_waiting;
static pthread_rwlock_t cur_lock = PTHREAD_RWLOCK_INITIALIZER;
static int cur_fd = -1; /* the root of A', -1 wh
---truncated---