audit: rework the audit queue handling

The audit record backlog queue has always been a bit of a mess, and
the moving the multicast send into kauditd_thread() from
audit_log_end() only makes things worse.  This patch attempts to fix
the backlog queue with a better design that should hold up better
under load and have less of a performance impact at syscall
invocation time.

While it looks like there is a log going on in this patch, the main
change is the move from a single backlog queue to three queues:

* A queue for holding records generated from audit_log_end() that
haven't been consumed by kauditd_thread() (audit_queue).

* A queue for holding records that have been sent via multicast but
had a temporary failure when sending via unicast and need a resend
(audit_retry_queue).

* A queue for holding records that haven't been sent via unicast
because no one is listening (audit_hold_queue).

Special care is taken in this patch to ensure that the proper
record ordering is preserved, e.g. we send everything in the hold
queue first, then the retry queue, and finally the main queue.
Signed-off-by: Paul Moore <paul@paul-moore.com>

kernel/audit.c

@@ -138,11 +138,18 @@ static DEFINE_SPINLOCK(audit_freelist_lock);
 static int	   audit_freelist_count;
 static LIST_HEAD(audit_freelist);
 
+/* queue msgs to send via kauditd_task */
 static struct sk_buff_head audit_queue;
-/* queue of skbs to send to auditd when/if it comes back */
+/* queue msgs due to temporary unicast send problems */
+static struct sk_buff_head audit_retry_queue;
+/* queue msgs waiting for new auditd connection */
 static struct sk_buff_head audit_hold_queue;
+
+/* queue servicing thread */
 static struct task_struct *kauditd_task;
 static DECLARE_WAIT_QUEUE_HEAD(kauditd_wait);
+
+/* waitqueue for callers who are blocked on the audit backlog */
 static DECLARE_WAIT_QUEUE_HEAD(audit_backlog_wait);
 
 static struct audit_features af = {.vers = AUDIT_FEATURE_VERSION,
@@ -364,25 +371,6 @@ static int audit_set_failure(u32 state)
 	return audit_do_config_change("audit_failure", &audit_failure, state);
 }
 
-/*
- * Queue skbs to be sent to auditd when/if it comes back.  These skbs should
- * already have been sent via prink/syslog and so if these messages are dropped
- * it is not a huge concern since we already passed the audit_log_lost()
- * notification and stuff.  This is just nice to get audit messages during
- * boot before auditd is running or messages generated while auditd is stopped.
- * This only holds messages is audit_default is set, aka booting with audit=1
- * or building your kernel that way.
- */
-static void audit_hold_skb(struct sk_buff *skb)
-{
-	if (audit_default &&
-	    (!audit_backlog_limit ||
-	     skb_queue_len(&audit_hold_queue) < audit_backlog_limit))
-		skb_queue_tail(&audit_hold_queue, skb);
-	else
-		kfree_skb(skb);
-}
-
 /*
  * For one reason or another this nlh isn't getting delivered to the userspace
  * audit daemon, just send it to printk.
@@ -398,58 +386,114 @@ static void kauditd_printk_skb(struct sk_buff *skb)
 		else
 			audit_log_lost("printk limit exceeded");
 	}
+}
+
+/**
+ * kauditd_hold_skb - Queue an audit record, waiting for auditd
+ * @skb: audit record
+ *
+ * Description:
+ * Queue the audit record, waiting for an instance of auditd.  When this
+ * function is called we haven't given up yet on sending the record, but things
+ * are not looking good.  The first thing we want to do is try to write the
+ * record via printk and then see if we want to try and hold on to the record
+ * and queue it, if we have room.  If we want to hold on to the record, but we
+ * don't have room, record a record lost message.
+ */
+static void kauditd_hold_skb(struct sk_buff *skb)
+{
+	/* at this point it is uncertain if we will ever send this to auditd so
+	 * try to send the message via printk before we go any further */
+	kauditd_printk_skb(skb);
+
+	/* can we just silently drop the message? */
+	if (!audit_default) {
+		kfree_skb(skb);
+		return;
+	}
+
+	/* if we have room, queue the message */
+	if (!audit_backlog_limit ||
+	    skb_queue_len(&audit_hold_queue) < audit_backlog_limit) {
+		skb_queue_tail(&audit_hold_queue, skb);
+		return;
+	}
 
-	audit_hold_skb(skb);
+	/* we have no other options - drop the message */
+	audit_log_lost("kauditd hold queue overflow");
+	kfree_skb(skb);
 }
 
-static void kauditd_send_unicast_skb(struct sk_buff *skb)
+/**
+ * kauditd_retry_skb - Queue an audit record, attempt to send again to auditd
+ * @skb: audit record
+ *
+ * Description:
+ * Not as serious as kauditd_hold_skb() as we still have a connected auditd,
+ * but for some reason we are having problems sending it audit records so
+ * queue the given record and attempt to resend.
+ */
+static void kauditd_retry_skb(struct sk_buff *skb)
 {
-	int err;
-	int attempts = 0;
-#define AUDITD_RETRIES 5
+	/* NOTE: because records should only live in the retry queue for a
+	 * short period of time, before either being sent or moved to the hold
+	 * queue, we don't currently enforce a limit on this queue */
+	skb_queue_tail(&audit_retry_queue, skb);
+}
 
-restart:
-	/* take a reference in case we can't send it and we want to hold it */
-	skb_get(skb);
-	err = netlink_unicast(audit_sock, skb, audit_nlk_portid, 0);
-	if (err < 0) {
-		pr_err("netlink_unicast sending to audit_pid=%d returned error: %d\n",
-		       audit_pid, err);
-		if (audit_pid) {
-			if (err == -ECONNREFUSED || err == -EPERM
-			    || ++attempts >= AUDITD_RETRIES) {
-				char s[32];
-
-				snprintf(s, sizeof(s), "audit_pid=%d reset", audit_pid);
-				audit_log_lost(s);
+/**
+ * auditd_reset - Disconnect the auditd connection
+ *
+ * Description:
+ * Break the auditd/kauditd connection and move all the records in the retry
+ * queue into the hold queue in case auditd reconnects.
+ */
+static void auditd_reset(void)
+{
+	struct sk_buff *skb;
+
+	/* break the connection */
 	audit_pid = 0;
 	audit_sock = NULL;
-			} else {
-				pr_warn("re-scheduling(#%d) write to audit_pid=%d\n",
-					attempts, audit_pid);
-				set_current_state(TASK_INTERRUPTIBLE);
-				schedule();
-				goto restart;
-			}
-		}
-		/* we might get lucky and get this in the next auditd */
-		audit_hold_skb(skb);
-	} else
-		/* drop the extra reference if sent ok */
+
+	/* flush all of the retry queue to the hold queue */
+	while ((skb = skb_dequeue(&audit_retry_queue)))
+		kauditd_hold_skb(skb);
+}
+
+/**
+ * kauditd_send_unicast_skb - Send a record via unicast to auditd
+ * @skb: audit record
+ */
+static int kauditd_send_unicast_skb(struct sk_buff *skb)
+{
+	int rc;
+
+	/* get an extra skb reference in case we fail to send */
+	skb_get(skb);
+	rc = netlink_unicast(audit_sock, skb, audit_nlk_portid, 0);
+	if (rc >= 0) {
 		consume_skb(skb);
+		rc = 0;
+	}
+
+	return rc;
 }
 
 /*
- * kauditd_send_multicast_skb - send the skb to multicast userspace listeners
+ * kauditd_send_multicast_skb - Send a record to any multicast listeners
+ * @skb: audit record
  *
+ * Description:
  * This function doesn't consume an skb as might be expected since it has to
  * copy it anyways.
  */
-static void kauditd_send_multicast_skb(struct sk_buff *skb, gfp_t gfp_mask)
+static void kauditd_send_multicast_skb(struct sk_buff *skb)
 {
 	struct sk_buff *copy;
 	struct audit_net *aunet = net_generic(&init_net, audit_net_id);
 	struct sock *sock = aunet->nlsk;
+	struct nlmsghdr *nlh;
 
 	if (!netlink_has_listeners(sock, AUDIT_NLGRP_READLOG))
 		return;
@@ -464,94 +508,155 @@ static void kauditd_send_multicast_skb(struct sk_buff *skb, gfp_t gfp_mask)
 	 * no reason for new multicast clients to continue with this
 	 * non-compliance.
 	 */
-	copy = skb_copy(skb, gfp_mask);
+	copy = skb_copy(skb, GFP_KERNEL);
 	if (!copy)
 		return;
+	nlh = nlmsg_hdr(copy);
+	nlh->nlmsg_len = skb->len;
 
-	nlmsg_multicast(sock, copy, 0, AUDIT_NLGRP_READLOG, gfp_mask);
+	nlmsg_multicast(sock, copy, 0, AUDIT_NLGRP_READLOG, GFP_KERNEL);
 }
 
-/*
- * flush_hold_queue - empty the hold queue if auditd appears
- *
- * If auditd just started, drain the queue of messages already
- * sent to syslog/printk.  Remember loss here is ok.  We already
- * called audit_log_lost() if it didn't go out normally.  so the
- * race between the skb_dequeue and the next check for audit_pid
- * doesn't matter.
+/**
+ * kauditd_wake_condition - Return true when it is time to wake kauditd_thread
  *
- * If you ever find kauditd to be too slow we can get a perf win
- * by doing our own locking and keeping better track if there
- * are messages in this queue.  I don't see the need now, but
- * in 5 years when I want to play with this again I'll see this
- * note and still have no friggin idea what i'm thinking today.
+ * Description:
+ * This function is for use by the wait_event_freezable() call in
+ * kauditd_thread().
  */
-static void flush_hold_queue(void)
+static int kauditd_wake_condition(void)
 {
-	struct sk_buff *skb;
-
-	if (!audit_default || !audit_pid)
-		return;
-
-	skb = skb_dequeue(&audit_hold_queue);
-	if (likely(!skb))
-		return;
+	static int pid_last = 0;
+	int rc;
+	int pid = audit_pid;
 
-	while (skb && audit_pid) {
-		kauditd_send_unicast_skb(skb);
-		skb = skb_dequeue(&audit_hold_queue);
-	}
+	/* wake on new messages or a change in the connected auditd */
+	rc = skb_queue_len(&audit_queue) || (pid && pid != pid_last);
+	if (rc)
+		pid_last = pid;
 
-	/*
-	 * if auditd just disappeared but we
-	 * dequeued an skb we need to drop ref
-	 */
-	consume_skb(skb);
+	return rc;
 }
 
 static int kauditd_thread(void *dummy)
 {
+	int rc;
+	int auditd = 0;
+	int reschedule = 0;
 	struct sk_buff *skb;
 	struct nlmsghdr *nlh;
 
+#define UNICAST_RETRIES 5
+#define AUDITD_BAD(x,y) \
+	((x) == -ECONNREFUSED || (x) == -EPERM || ++(y) >= UNICAST_RETRIES)
+
+	/* NOTE: we do invalidate the auditd connection flag on any sending
+	 * errors, but we only "restore" the connection flag at specific places
+	 * in the loop in order to help ensure proper ordering of audit
+	 * records */
+
 	set_freezable();
 	while (!kthread_should_stop()) {
-		flush_hold_queue();
+		/* NOTE: possible area for future improvement is to look at
+		 *       the hold and retry queues, since only this thread
+		 *       has access to these queues we might be able to do
+		 *       our own queuing and skip some/all of the locking */
+
+		/* NOTE: it might be a fun experiment to split the hold and
+		 *       retry queue handling to another thread, but the
+		 *       synchronization issues and other overhead might kill
+		 *       any performance gains */
+
+		/* attempt to flush the hold queue */
+		while (auditd && (skb = skb_dequeue(&audit_hold_queue))) {
+			rc = kauditd_send_unicast_skb(skb);
+			if (rc) {
+				/* requeue to the same spot */
+				skb_queue_head(&audit_hold_queue, skb);
+
+				auditd = 0;
+				if (AUDITD_BAD(rc, reschedule)) {
+					auditd_reset();
+					reschedule = 0;
+				}
+			} else
+				/* we were able to send successfully */
+				reschedule = 0;
+		}
+
+		/* attempt to flush the retry queue */
+		while (auditd && (skb = skb_dequeue(&audit_retry_queue))) {
+			rc = kauditd_send_unicast_skb(skb);
+			if (rc) {
+				auditd = 0;
+				if (AUDITD_BAD(rc, reschedule)) {
+					kauditd_hold_skb(skb);
+					auditd_reset();
+					reschedule = 0;
+				} else
+					/* temporary problem (we hope), queue
+					 * to the same spot and retry */
+					skb_queue_head(&audit_retry_queue, skb);
+			} else
+				/* we were able to send successfully */
+				reschedule = 0;
+		}
 
+		/* standard queue processing, try to be as quick as possible */
+quick_loop:
 		skb = skb_dequeue(&audit_queue);
 		if (skb) {
+			/* setup the netlink header, see the comments in
+			 * kauditd_send_multicast_skb() for length quirks */
 			nlh = nlmsg_hdr(skb);
+			nlh->nlmsg_len = skb->len - NLMSG_HDRLEN;
 
-			/* if nlh->nlmsg_len is zero then we haven't attempted
-			 * to send the message to userspace yet, if nlmsg_len
-			 * is non-zero then we have attempted to send it to
-			 * the multicast listeners as well as auditd; keep
-			 * trying to send to auditd but don't repeat the
-			 * multicast send */
-			if (nlh->nlmsg_len == 0) {
-				nlh->nlmsg_len = skb->len;
-				kauditd_send_multicast_skb(skb, GFP_KERNEL);
+			/* attempt to send to any multicast listeners */
+			kauditd_send_multicast_skb(skb);
 
-				/* see the note in kauditd_send_multicast_skb
-				 * regarding the nlh->nlmsg_len value and why
-				 * it differs between the multicast and unicast
-				 * clients */
-				nlh->nlmsg_len -= NLMSG_HDRLEN;
+			/* attempt to send to auditd, queue on failure */
+			if (auditd) {
+				rc = kauditd_send_unicast_skb(skb);
+				if (rc) {
+					auditd = 0;
+					if (AUDITD_BAD(rc, reschedule)) {
+						auditd_reset();
+						reschedule = 0;
 					}
 
-			if (audit_pid)
-				kauditd_send_unicast_skb(skb);
+					/* move to the retry queue */
+					kauditd_retry_skb(skb);
+				} else
+					/* everything is working so go fast! */
+					goto quick_loop;
+			} else if (reschedule)
+				/* we are currently having problems, move to
+				 * the retry queue */
+				kauditd_retry_skb(skb);
 			else
-				kauditd_printk_skb(skb);
+				/* dump the message via printk and hold it */
+				kauditd_hold_skb(skb);
 		} else {
-			/* we have flushed the backlog so wake everyone up who
-			 * is blocked and go to sleep until we have something
-			 * in the backlog again */
+			/* we have flushed the backlog so wake everyone */
 			wake_up(&audit_backlog_wait);
+
+			/* if everything is okay with auditd (if present), go
+			 * to sleep until there is something new in the queue
+			 * or we have a change in the connected auditd;
+			 * otherwise simply reschedule to give things a chance
+			 * to recover */
+			if (reschedule) {
+				set_current_state(TASK_INTERRUPTIBLE);
+				schedule();
+			} else
 				wait_event_freezable(kauditd_wait,
-					     skb_queue_len(&audit_queue));
+						     kauditd_wake_condition());
+
+			/* update the auditd connection status */
+			auditd = (audit_pid ? 1 : 0);
 		}
 	}
+
 	return 0;
 }
 
@@ -616,6 +721,7 @@ static int audit_send_reply_thread(void *arg)
 	kfree(reply);
 	return 0;
 }
+
 /**
  * audit_send_reply - send an audit reply message via netlink
  * @request_skb: skb of request we are replying to (used to target the reply)
@@ -1171,10 +1277,8 @@ static void __net_exit audit_net_exit(struct net *net)
 {
 	struct audit_net *aunet = net_generic(net, audit_net_id);
 	struct sock *sock = aunet->nlsk;
-	if (sock == audit_sock) {
-		audit_pid = 0;
-		audit_sock = NULL;
-	}
+	if (sock == audit_sock)
+		auditd_reset();
 
 	RCU_INIT_POINTER(aunet->nlsk, NULL);
 	synchronize_net();
@@ -1201,6 +1305,7 @@ static int __init audit_init(void)
 	register_pernet_subsys(&audit_net_ops);
 
 	skb_queue_head_init(&audit_queue);
+	skb_queue_head_init(&audit_retry_queue);
 	skb_queue_head_init(&audit_hold_queue);
 	audit_initialized = AUDIT_INITIALIZED;
 	audit_enabled = audit_default;