[PATCH 3/3] Fix race in ipv6_flowlabel_opt() when inserting two labels

[PATCH 3/3] Fix race in ipv6_flowlabel_opt() when inserting two labels

[Pavel Emelyanov]

In the IPV6_FL_A_GET case the hash is checked for flowlabels
with the given label. If it is not found, the lock, protecting 
the hash, is dropped to be re-get for writing. After this a
newly allocated entry is inserted, but no checks are performed
to catch a classical SMP race, when the conflicting label may 
be inserted on another cpu.

Use the (currently unused) return value from fl_intern() to
return the conflicting entry (if found) and re-check, whether
we can reuse it (IPV6_FL_F_EXCL) or return -EEXISTS.

Also add the comment, about why not re-lookup the current
sock for conflicting flowlabel entry.

Signed-off-by: Pavel Emelyanov <xemul@openvz.org>

---

diff --git a/net/ipv6/ip6_flowlabel.c b/net/ipv6/ip6_flowlabel.c
index f40a086..e55ae1a 100644
--- a/net/ipv6/ip6_flowlabel.c
+++ b/net/ipv6/ip6_flowlabel.c
@@ -154,8 +154,10 @@ static void ip6_fl_gc(unsigned long dummy)
 	write_unlock(&ip6_fl_lock);
 }
 
-static int fl_intern(struct ip6_flowlabel *fl, __be32 label)
+static struct ip6_flowlabel *fl_intern(struct ip6_flowlabel *fl, __be32 label)
 {
+	struct ip6_flowlabel *lfl;
+
 	fl->label = label & IPV6_FLOWLABEL_MASK;
 
 	write_lock_bh(&ip6_fl_lock);
@@ -163,12 +165,26 @@ static int fl_intern(struct ip6_flowlabel *fl, __be32 label)
 		for (;;) {
 			fl->label = htonl(net_random())&IPV6_FLOWLABEL_MASK;
 			if (fl->label) {
-				struct ip6_flowlabel *lfl;
 				lfl = __fl_lookup(fl->label);
 				if (lfl == NULL)
 					break;
 			}
 		}
+	} else {
+		/*
+		 * we dropper the ip6_fl_lock, so this entry could reappear
+		 * and we need to recheck with it.
+		 *
+		 * OTOH no need to search the active socket first, like it is
+		 * done in ipv6_flowlabel_opt - sock is locked, so new entry
+		 * with the same label can only appear on another sock
+		 */
+		lfl = __fl_lookup(fl->label);
+		if (lfl != NULL) {
+			atomic_inc(&lfl->users);
+			write_unlock_bh(&ip6_fl_lock);
+			return lfl;
+		}
 	}
 
 	fl->lastuse = jiffies;
@@ -176,7 +192,7 @@ static int fl_intern(struct ip6_flowlabel *fl, __be32 label)
 	fl_ht[FL_HASH(fl->label)] = fl;
 	atomic_inc(&fl_size);
 	write_unlock_bh(&ip6_fl_lock);
-	return 0;
+	return NULL;
 }
 
 
@@ -429,7 +445,8 @@ int ipv6_flowlabel_opt(struct sock *sk, char __user *optval, int optlen)
 	struct in6_flowlabel_req freq;
 	struct ipv6_fl_socklist *sfl1=NULL;
 	struct ipv6_fl_socklist *sfl, **sflp;
-	struct ip6_flowlabel *fl;
+	struct ip6_flowlabel *fl, *fl1 = NULL;
+
 
 	if (optlen < sizeof(freq))
 		return -EINVAL;
@@ -485,8 +502,6 @@ int ipv6_flowlabel_opt(struct sock *sk, char __user *optval, int optlen)
 		sfl1 = kmalloc(sizeof(*sfl1), GFP_KERNEL);
 
 		if (freq.flr_label) {
-			struct ip6_flowlabel *fl1 = NULL;
-
 			err = -EEXIST;
 			read_lock_bh(&ip6_sk_fl_lock);
 			for (sfl = np->ipv6_fl_list; sfl; sfl = sfl->next) {
@@ -505,6 +520,7 @@ int ipv6_flowlabel_opt(struct sock *sk, char __user *optval, int optlen)
 			if (fl1 == NULL)
 				fl1 = fl_lookup(freq.flr_label);
 			if (fl1) {
+recheck:
 				err = -EEXIST;
 				if (freq.flr_flags&IPV6_FL_F_EXCL)
 					goto release;
@@ -543,9 +559,9 @@ release:
 		if (sfl1 == NULL || (err = mem_check(sk)) != 0)
 			goto done;
 
-		err = fl_intern(fl, freq.flr_label);
-		if (err)
-			goto done;
+		fl1 = fl_intern(fl, freq.flr_label);
+		if (fl1 != NULL)
+			goto recheck;
 
 		if (!freq.flr_label) {
 			if (copy_to_user(&((struct in6_flowlabel_req __user *) optval)->flr_label,

Re: [PATCH 3/3] Fix race in ipv6_flowlabel_opt() when inserting two labels

[David Miller]

From: Pavel Emelyanov <xemul@openvz.org>
Date: Thu, 18 Oct 2007 15:59:14 +0400

> In the IPV6_FL_A_GET case the hash is checked for flowlabels
> with the given label. If it is not found, the lock, protecting 
> the hash, is dropped to be re-get for writing. After this a
> newly allocated entry is inserted, but no checks are performed
> to catch a classical SMP race, when the conflicting label may 
> be inserted on another cpu.
> 
> Use the (currently unused) return value from fl_intern() to
> return the conflicting entry (if found) and re-check, whether
> we can reuse it (IPV6_FL_F_EXCL) or return -EEXISTS.
> 
> Also add the comment, about why not re-lookup the current
> sock for conflicting flowlabel entry.
> 
> Signed-off-by: Pavel Emelyanov <xemul@openvz.org>

Applied.