[BUG][PATCH][RFC] audit: hang up in audit_log_start executed on auditd

[Gao feng]

On 10/11/2013 09:36 AM, Toshiyuki Okajima wrote:
> Hi. 
> 
> The following reproducer causes auditd daemon hang up.
> (But the hang up is released after the audit_backlog_wait_time passes.)
>  # auditctl -a exit,always -S all
>  # reboot
> 
> 
> I reproduced the hangup on KVM, and then got a crash dump.
> After I analyzed the dump, I found auditd daemon hung up in audit_log_start. 
> (I have confirmed it on linux-3.12-rc4.)
> 
> Like this:
> crash> bt 1426
> PID: 1426   TASK: ffff88007b63e040  CPU: 1   COMMAND: "auditd"
>  #0 [ffff88007cb93918] __schedule at ffffffff8155d980
>  #1 [ffff88007cb939b0] schedule at ffffffff8155de99
>  #2 [ffff88007cb939c0] schedule_timeout at ffffffff8155b840
>  #3 [ffff88007cb93a60] audit_log_start at ffffffff810d3ce5
>  #4 [ffff88007cb93b20] audit_log_config_change at ffffffff810d3ece
>  #5 [ffff88007cb93b60] audit_receive_msg at ffffffff810d4fd6
>  #6 [ffff88007cb93c00] audit_receive at ffffffff810d5173
>  #7 [ffff88007cb93c30] netlink_unicast at ffffffff814c5269
>  #8 [ffff88007cb93c90] netlink_sendmsg at ffffffff814c6386
>  #9 [ffff88007cb93d20] sock_sendmsg at ffffffff814813c0
> #10 [ffff88007cb93e30] SYSC_sendto at ffffffff81481524
> #11 [ffff88007cb93f70] sys_sendto at ffffffff8148157e
> #12 [ffff88007cb93f80] system_call_fastpath at ffffffff81568052
>     RIP: 00007f5c47f7fba3  RSP: 00007fffcf21a118  RFLAGS: 00010202
>     RAX: 000000000000002c  RBX: ffffffff81568052  RCX: 0000000000000000
>     RDX: 0000000000000030  RSI: 00007fffcf21e7d0  RDI: 0000000000000003
>     RBP: 00007fffcf21e7d0   R8: 00007fffcf21a130   R9: 000000000000000c
>     R10: 0000000000000000  R11: 0000000000000293  R12: ffffffff8148157e
>     R13: ffff88007cb93f78  R14: 0000000000000020  R15: 0000000000000030
>     ORIG_RAX: 000000000000002c  CS: 0033  SS: 002b
> 
> 
> The reason is that auditd daemon itself cannot consume its backlog 
> while audit_log_start is calling schedule_timeout on auditd daemon.  
> So, that is a deadlock!
> 
> Therefore, I think audit_log_start shouldn't handle auditd's backlog
> when auditd daemon executes audit_log_start.
> 
> For example, I made the following fix patch.
> --------------------------------------------------------------
> auditd daemon can execute the audit_log_start, and then it can cause 
> a hang up because only auditd daemon can consume the backlog.
> So, audit_log_start executed by auditd daemon should not handle the backlog 
> in case auditd daemon hangs up (while wait_for_auditd is calling).
> 
> Signed-off-by: Toshiyuki Okajima <toshi.okajima at jp.fujitsu.com>
> ---
>  kernel/audit.c |    3 +++
>  1 files changed, 3 insertions(+), 0 deletions(-)
> 
> diff --git a/kernel/audit.c b/kernel/audit.c
> index 7b0e23a..86c389e 100644
> --- a/kernel/audit.c
> +++ b/kernel/audit.c
> @@ -1098,6 +1098,9 @@ struct audit_buffer *audit_log_start(struct audit_context *ctx, gfp_t gfp_mask,
>  	int reserve;
>  	unsigned long timeout_start = jiffies;
>  
> +	if (audit_pid && (audit_pid == current->pid))
> +		return NULL;
> +

audit_log_start can be called in interrupt context, such as iptables AUDIT module,
we can't use current here.
please try the patch below.

diff --git a/kernel/audit.c b/kernel/audit.c
index 7b0e23a..1f35f3d 100644
--- a/kernel/audit.c
+++ b/kernel/audit.c
@@ -270,9 +270,13 @@ static int audit_log_config_change(char *function_name, int new, int old,
                                   int allow_changes)
 {
        struct audit_buffer *ab;
+       gfp_t gfp_mask = GFP_KERNEL;
        int rc = 0;

-       ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
+       if (audit_pid && audit_pid == current->pid)
+               gfp_mask = GFP_ATOMIC;
+
+       ab = audit_log_start(NULL, gfp_mask, AUDIT_CONFIG_CHANGE);
        if (unlikely(!ab))
                return rc;
        audit_log_format(ab, "%s=%d old=%d", function_name, new, old);


Thanks