[IPV6]: Fix race in ipv6_flowlabel_opt() when inserting two labels
authorPavel Emelyanov <xemul@openvz.org>
Thu, 18 Oct 2007 12:18:56 +0000 (05:18 -0700)
committerDavid S. Miller <davem@davemloft.net>
Thu, 18 Oct 2007 12:18:56 +0000 (05:18 -0700)
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>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv6/ip6_flowlabel.c

index f40a08669db029c0d012342f07873ae14e7d4ae4..e55ae1a1f56023b12775e49d99cd32aec4b85dee 100644 (file)
@@ -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,