[PATCH] audit: ensure userspace is penalized the same as the kernel when under pressure

[cuigaosheng]

在 2021/12/15 0:16, Paul Moore 写道:
> Due to the audit control mutex necessary for serializing audit
> userspace messages we haven't been able to block/penalize userspace
> processes that attempt to send audit records while the system is
> under audit pressure.  The result is that privileged userspace
> applications have a priority boost with respect to audit as they are
> not bound by the same audit queue throttling as the other tasks on
> the system.
>
> This patch attempts to restore some balance to the system when under
> audit pressure by blocking these privileged userspace tasks after
> they have finished their audit processing, and dropped the audit
> control mutex, but before they return to userspace.
>
> Reported-by: Gaosheng Cui <cuigaosheng1 at huawei.com>
> Signed-off-by: Paul Moore <paul at paul-moore.com>
> ---
>   kernel/audit.c |   18 +++++++++++++++++-
>   1 file changed, 17 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/audit.c b/kernel/audit.c
> index 4cebadb5f30d..eab7282668ab 100644
> --- a/kernel/audit.c
> +++ b/kernel/audit.c
> @@ -1540,6 +1540,20 @@ static void audit_receive(struct sk_buff  *skb)
>   		nlh = nlmsg_next(nlh, &len);
>   	}
>   	audit_ctl_unlock();
> +
> +	/* can't block with the ctrl lock, so penalize the sender now */
> +	if (audit_backlog_limit &&
> +	    (skb_queue_len(&audit_queue) > audit_backlog_limit)) {
> +		DECLARE_WAITQUEUE(wait, current);
> +
> +		/* wake kauditd to try and flush the queue */
> +		wake_up_interruptible(&kauditd_wait);
> +
> +		add_wait_queue_exclusive(&audit_backlog_wait, &wait);
> +		set_current_state(TASK_UNINTERRUPTIBLE);
> +		schedule_timeout(audit_backlog_wait_time);
> +		remove_wait_queue(&audit_backlog_wait, &wait);
> +	}
>   }
>   
>   /* Log information about who is connecting to the audit multicast socket */
> @@ -1824,7 +1838,9 @@ struct audit_buffer *audit_log_start(struct audit_context *ctx, gfp_t gfp_mask,
>   	 *    task_tgid_vnr() since auditd_pid is set in audit_receive_msg()
>   	 *    using a PID anchored in the caller's namespace
>   	 * 2. generator holding the audit_cmd_mutex - we don't want to block
> -	 *    while holding the mutex */
> +	 *    while holding the mutex, although we do penalize the sender
> +	 *    later in audit_receive() when it is safe to block
> +	 */
>   	if (!(auditd_test_task(current) || audit_ctl_owner_current())) {
>   		long stime = audit_backlog_wait_time;
>   

Tested-by: Gaosheng Cui<cuigaosheng1 at huawei.com>