Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Thu, 10 Oct 2019 23:35:12 +0200
From:   Thibaut Sautereau <thibaut@sautereau.fr>
To:     linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org
Cc:     Alexander Viro <viro@zeniv.linux.org.uk>,
        Christian Brauner <christian@brauner.io>
Subject: NULL pointer deref in put_fs_context with unprivileged LXC
Message-ID: <20191010213512.GA875@gandi.net>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
User-Agent: Mutt/1.12.2 (2019-09-21)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Since v5.1 and as of v5.3.5, I get the following oops every single time
I start an *unprivileged* LXC container:

	BUG: kernel NULL pointer dereference, address: 0000000000000043
	#PF: supervisor read access in kernel mode
	#PF: error_code(0x0000) - not-present page
	PGD 0 P4D 0 
	Oops: 0000 [#1] SMP PTI
	CPU: 3 PID: 3789 Comm: systemd Tainted: G                T 5.3.5 #5
	RIP: 0010:put_fs_context+0x13/0x180
	Code: e4 31 c9 eb c8 e8 1d d6 dc ff 66 66 2e 0f 1f 84 00 00 00 00 00 66 90 41 54 55 48 89 fd 53 48 8b b>
	RSP: 0018:ffffc90000777e10 EFLAGS: 00010286
	RAX: 00000000fffffff3 RBX: 0000000000000000 RCX: ffffc90000777d6c
	RDX: 0000000000000000 RSI: ffff8884062331e8 RDI: fffffffffffffff3
	RBP: ffff8883e772dc00 R08: ffff88840d6bc680 R09: 0000000000000001
	R10: 0000000000000000 R11: 0000000000000001 R12: fffffffffffffff3
	R13: ffff888405ad2860 R14: ffff8883e772dc00 R15: 0000000000000027
	FS:  00007998d1444980(0000) GS:ffff88840f980000(0000) knlGS:0000000000000000
	CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
	CR2: 0000000000000043 CR3: 000000040d236003 CR4: 00000000001606e0
	Call Trace:
	 do_mount+0x2f6/0xab0
	 ksys_mount+0x79/0xc0
	 __x64_sys_mount+0x1d/0x30
	 do_syscall_64+0x68/0x666
	 entry_SYSCALL_64_after_hwframe+0x49/0xbe
	RIP: 0033:0x7998d23aafea
	Code: 48 8b 0d a9 0e 0c 00 f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 4>
	RSP: 002b:00007ffd4b0c8bc8 EFLAGS: 00000246 ORIG_RAX: 00000000000000a5
	RAX: ffffffffffffffda RBX: 00005ae352a55a30 RCX: 00007998d23aafea
	RDX: 00005ae3529fe0b3 RSI: 00005ae3529fe0d5 RDI: 00005ae3529fe0b3
	RBP: 0000000000000007 R08: 00005ae3529fe0ca R09: 00005ae35433fb20
	R10: 000000000000000e R11: 0000000000000246 R12: 00000000fffffffe
	R13: 0000000000000000 R14: 0000000000000000 R15: 0000000000000001
	Modules linked in:
	CR2: 0000000000000043
	---[ end trace 66de701522a6be46 ]---
	RIP: 0010:put_fs_context+0x13/0x180
	Code: e4 31 c9 eb c8 e8 1d d6 dc ff 66 66 2e 0f 1f 84 00 00 00 00 00 66 90 41 54 55 48 89 fd 53 48 8b b>
	RSP: 0018:ffffc90000777e10 EFLAGS: 00010286
	RAX: 00000000fffffff3 RBX: 0000000000000000 RCX: ffffc90000777d6c
	RDX: 0000000000000000 RSI: ffff8884062331e8 RDI: fffffffffffffff3
	RBP: ffff8883e772dc00 R08: ffff88840d6bc680 R09: 0000000000000001
	R10: 0000000000000000 R11: 0000000000000001 R12: fffffffffffffff3
	R13: ffff888405ad2860 R14: ffff8883e772dc00 R15: 0000000000000027
	FS:  00007998d1444980(0000) GS:ffff88840f980000(0000) knlGS:0000000000000000
	CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
	CR2: 0000000000000043 CR3: 000000040d236003 CR4: 00000000001606e0

According to GDB:
	$ gdb fs/fs_context.o
	(gdb) l *put_fs_context+0x13
	0xa53 is in put_fs_context (fs/fs_context.c:494).
	489	void put_fs_context(struct fs_context *fc)
	490	{
	491		struct super_block *sb;
	492
	493		if (fc->root) {
	494			sb = fc->root->d_sb;
	495			dput(fc->root);
	496			fc->root = NULL;
	497			deactivate_super(sb);
	498		}

	$ gdb fs/namespace.o
	(gdb) l *do_mount+0x2f6
	0x5506 is in do_mount (fs/namespace.c:2796).
	2791			err = vfs_get_tree(fc);
	2792		if (!err)
	2793			err = do_new_mount_fc(fc, path, mnt_flags);
	2794
	2795		put_fs_context(fc);
	2796		return err;
	2797	}
	2798
	2799	int finish_automount(struct vfsmount *m, struct path *path)
	2800	{


I don't face this issue when starting the same container as a
privileged one. I tried to strace the container when launched in
foreground and the following snippet may be related to the problem:

	[pid 35813] openat(AT_FDCWD, "/sys/fs", O_RDONLY|O_CLOEXEC|O_PATH|O_DIRECTORY) = 4
	[pid 35813] name_to_handle_at(4, "cgroup", {handle_bytes=128}, 0x7ffcdf6ebac4, AT_SYMLINK_FOLLOW) = -1 EOPNOTSUPP (Operation not supported)
	[pid 35813] name_to_handle_at(4, "", {handle_bytes=128}, 0x7ffcdf6ebac4, AT_EMPTY_PATH) = -1 EOPNOTSUPP (Operation not supported)
	[pid 35813] openat(4, "cgroup", O_RDONLY|O_CLOEXEC|O_PATH) = 5
	[pid 35813] openat(AT_FDCWD, "/proc/self/fdinfo/5", O_RDONLY|O_CLOEXEC) = 6
	[pid 35813] fstat(6, {st_mode=S_IFREG|0400, st_size=0, ...}) = 0
	[pid 35813] fstat(6, {st_mode=S_IFREG|0400, st_size=0, ...}) = 0
	[pid 35813] read(6, "pos:\t0\nflags:\t012000000\nmnt_id:\t"..., 2048) = 36
	[pid 35813] read(6, "", 1024)           = 0
	[pid 35813] close(6)                    = 0
	[pid 35813] close(5)                    = 0
	[pid 35813] openat(AT_FDCWD, "/proc/self/fdinfo/4", O_RDONLY|O_CLOEXEC) = 5
	[pid 35813] fstat(5, {st_mode=S_IFREG|0400, st_size=0, ...}) = 0
	[pid 35813] fstat(5, {st_mode=S_IFREG|0400, st_size=0, ...}) = 0
	[pid 35813] read(5, "pos:\t0\nflags:\t012200000\nmnt_id:\t"..., 2048) = 36
	[pid 35813] read(5, "", 1024)           = 0
	[pid 35813] close(5)                    = 0
	[pid 35813] newfstatat(4, "cgroup", {st_mode=S_IFDIR|0555, st_size=0, ...}, 0) = 0
	[pid 35813] newfstatat(4, "", {st_mode=S_IFDIR|0755, st_size=0, ...}, AT_EMPTY_PATH) = 0
	[pid 35813] close(4)                    = 0
	[pid 35813] stat("/sys/fs", {st_mode=S_IFDIR|0755, st_size=0, ...}) = 0
	[pid 35813] mkdir("/sys/fs/cgroup", 0755) = -1 EEXIST (File exists)
	[pid 35813] stat("/sys/fs/cgroup", {st_mode=S_IFDIR|0555, st_size=0, ...}) = 0
	[pid 35813] mount("tmpfs", "/sys/fs/cgroup", "tmpfs", MS_NOSUID|MS_NODEV|MS_NOEXEC|MS_STRICTATIME, "mode=755") = 0
	[pid 35813] statfs("/sys/fs/cgroup/", {f_type=TMPFS_MAGIC, f_bsize=4096, f_blocks=2032290, f_bfree=2032290, f_bavail=2032290, f_files=2032290, f_ffree=2032289, f_fsid={val=[0, 0]}, f_namelen=255, f_frsize=4096, f_flags=ST_VALID|ST_NOSUID|ST_NODEV|ST_NOEXEC}) = 0
	[pid 35813] statfs("/sys/fs/cgroup/unified/", 0x7ffcdf6ebc10) = -1 ENOENT (No such file or directory)
	[pid 35813] statfs("/sys/fs/cgroup/systemd/", 0x7ffcdf6ebc10) = -1 ENOENT (No such file or directory)
	[pid 35813] openat(AT_FDCWD, "/proc/1/cmdline", O_RDONLY|O_CLOEXEC) = 4
	[pid 35813] prlimit64(0, RLIMIT_STACK, NULL, {rlim_cur=8192*1024, rlim_max=RLIM64_INFINITY}) = 0
	[pid 35813] mmap(NULL, 2101248, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7999ced5b000
	[pid 35813] fstat(4, {st_mode=S_IFREG|0444, st_size=0, ...}) = 0
	[pid 35813] read(4, "/sbin/init\0", 1024) = 11
	[pid 35813] read(4, "", 1024)           = 0
	[pid 35813] mremap(0x7999ced5b000, 2101248, 4096, MREMAP_MAYMOVE) = 0x7999ced5b000
	[pid 35813] close(4)                    = 0
	[pid 35813] munmap(0x7999ced5b000, 4096) = 0
	[pid 35813] openat(AT_FDCWD, "/", O_RDONLY|O_NOFOLLOW|O_CLOEXEC|O_PATH) = 4
	[pid 35813] openat(4, "sys", O_RDONLY|O_NOFOLLOW|O_CLOEXEC|O_PATH) = 5
	[pid 35813] fstat(5, {st_mode=S_IFDIR|0555, st_size=0, ...}) = 0
	[pid 35813] close(4)                    = 0
	[pid 35813] openat(5, "fs", O_RDONLY|O_NOFOLLOW|O_CLOEXEC|O_PATH) = 4
	[pid 35813] fstat(4, {st_mode=S_IFDIR|0755, st_size=0, ...}) = 0
	[pid 35813] close(5)                    = 0
	[pid 35813] openat(4, "cgroup", O_RDONLY|O_NOFOLLOW|O_CLOEXEC|O_PATH) = 5
	[pid 35813] fstat(5, {st_mode=S_IFDIR|0755, st_size=40, ...}) = 0
	[pid 35813] close(4)                    = 0
	[pid 35813] openat(5, "unified", O_RDONLY|O_NOFOLLOW|O_CLOEXEC|O_PATH) = -1 ENOENT (No such file or directory)
	[pid 35813] close(5)                    = 0
	[pid 35813] stat("/sys/fs/cgroup", {st_mode=S_IFDIR|0755, st_size=40, ...}) = 0
	[pid 35813] mkdir("/sys/fs/cgroup/unified", 0755) = 0
	[pid 35813] mount("cgroup2", "/sys/fs/cgroup/unified", "cgroup2", MS_NOSUID|MS_NODEV|MS_NOEXEC, "nsdelegate") = ?
	[pid 35813] +++ killed by SIGKILL +++

I've been trying to reproduce by playing with user namespaces and
cgroup2 mounts but I didn't succeed. Only an lxc-start of an
unprivileged container causes an oops (every single time). I wanted to
dive into the code but I hadn't looked at this part of the kernel since
the recent rework of file system mounting internals, thus I've been
postponing that for weeks now and I thought it was time to report the
bug anyway. Sorry for the lack of more detailed info :/

-- 
Thibaut