[PATCH] LSM: Delete hooks in reverse order for avoiding race

[James Morris]

Please cc the LSM list with LSM related patches.


On Wed, 8 Jan 2020, Huaisheng Ye wrote:

> From: Huaisheng Ye <yehs1 at lenovo.com>
> 
> There is small possibility as race condition when selinux_disable
> has been triggered. security_delete_hooks deletes all selinux hooks
> from security_hook_heads, but there are some selinux functions which
> are being called at the same time.
> 
> Here is a panic accident scene from 4.18 based kernel,
> 
> [   26.654494] SELinux:  Disabled at runtime.
> [   26.654507] BUG: unable to handle kernel NULL pointer dereference
> at 0000000000000020
> [   26.654508] PGD 0 P4D 0
> [   26.654510] Oops: 0002 [#1] SMP NOPTI
> [   26.654512] CPU: 53 PID: 2614 Comm: systemd-cgroups Tainted: G
>      OE    --------- -  - 4.18.0-80.el8.x86_64 #1
> [   26.654512] Hardware name: Lenovo ThinkSystem SR850P
>  -[7D2H]-/-[7D2H]-, BIOS -[TEE145P-1.10]- 12/06/2019
> [   26.654519] RIP: 0010:selinux_socket_post_create+0x80/0x390
> [   26.654520] Code: e9 95 6a 89 00 bd 16 00 00 00 c7 44 24 04 01
>  00 00 00 45 85 c0 0f 85 f6 00 00 00 8b 56 14 85 d2 0f 84 26 01 00
>  00 89 54 24 04 <66> 41 89 6c 24 20 31 c0 41 89 54 24 1c 41 c6 44
>  24 22 01 49 8b 4d
> [   26.654521] RSP: 0018:ffffbf515cc63e48 EFLAGS: 00010246
> [   26.654522] RAX: 0000000000000000 RBX: 0000000000000001 RCX: 0000000000000019
> [   26.654522] RDX: 0000000000000001 RSI: 0000000000000001 RDI: ffffffffab46f680
> [   26.654523] RBP: 0000000000000019 R08: 0000000000000000 R09: ffffbf515cc63e4c
> [   26.654523] R10: 0000000000000000 R11: 0000000000000000 R12: 0000000000000000
> [   26.654524] R13: ffff97d7bb6cbc80 R14: 0000000000000001 R15: ffff97d7bb6cbc80
> [   26.654525] FS:  00007f5c608ea380(0000) GS:ffff97d7bf140000(0000) knlGS:0000000000000000
> [   26.654525] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [   26.654526] CR2: 0000000000000020 CR3: 0000011ebc934004 CR4: 00000000007606e0
> [   26.654527] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> [   26.654528] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> [   26.654528] PKRU: 55555554
> [   26.654528] Call Trace:
> [   26.654535]  security_socket_post_create+0x42/0x60
> [   26.654537] SELinux:  Unregistering netfilter hooks
> [   26.654542]  __sock_create+0x106/0x1a0
> [   26.654545]  __sys_socket+0x57/0xe0
> [   26.654547]  __x64_sys_socket+0x16/0x20
> [   26.654551]  do_syscall_64+0x5b/0x1b0
> [   26.654554]  entry_SYSCALL_64_after_hwframe+0x65/0xca
> 
> The root cause is that, selinux_inode_alloc_security has been deleted
> firstly from security_hook_heads, so security_inode_alloc directly 
> return 0, that means the value of pointer inode->i_security equalling
> to NULL.
> 
> But selinux_socket_post_create hasn't been deleted at that moment, so
> which would involked by mistake. Inside the function, pointer isec
> needs to point to inode->i_security, then a NULL pointer defect happens.
> 
> For current upstream kernel, because of commit
> afb1cbe37440c7f38b9cf46fc331cc9dfd5cce21
> the inode security has been moved out to LSM infrastructure from
> individual security modules like selinux.
> 
> But this patch still can be applied for solving similar issue when
> security_delete_hooks has been used. Also for stable branch v4.19,
> the inode security still need to be created in individual modules.
> 
> The patch has been verified by Lenovo SR850P server through overnight
> reboot cycles.
> 
> Signed-off-by: Huaisheng Ye <yehs1 at lenovo.com>
> ---
>  include/linux/lsm_hooks.h | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/include/linux/lsm_hooks.h b/include/linux/lsm_hooks.h
> index 20d8cf1..57cb2ac 100644
> --- a/include/linux/lsm_hooks.h
> +++ b/include/linux/lsm_hooks.h
> @@ -2164,7 +2164,7 @@ static inline void security_delete_hooks(struct security_hook_list *hooks,
>  	int i;
>  
>  	for (i = 0; i < count; i++)
> -		hlist_del_rcu(&hooks[i].list);
> +		hlist_del_rcu(&hooks[count - 1 - i].list);
>  }
>  #endif /* CONFIG_SECURITY_SELINUX_DISABLE */
>  
> 

-- 
James Morris
<jmorris at namei.org>