ipv6: avoid high order allocations
authorEric Dumazet <eric.dumazet@gmail.com>
Sat, 5 Jun 2010 10:03:30 +0000 (03:03 -0700)
committerDavid S. Miller <davem@davemloft.net>
Sat, 5 Jun 2010 10:03:30 +0000 (03:03 -0700)
With mtu=9000, mld_newpack() use order-2 GFP_ATOMIC allocations, that
are very unreliable, on machines where PAGE_SIZE=4K

Limit allocated skbs to be at most one page. (order-0 allocations)

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv6/mcast.c

index 59f1881968c70aad15b0296cf6400c6b0b3e49cb..ab1622d7d409f0a853a956ee06db49aad930050a 100644 (file)
@@ -1356,7 +1356,10 @@ static struct sk_buff *mld_newpack(struct net_device *dev, int size)
                     IPV6_TLV_PADN, 0 };
 
        /* we assume size > sizeof(ra) here */
-       skb = sock_alloc_send_skb(sk, size + LL_ALLOCATED_SPACE(dev), 1, &err);
+       size += LL_ALLOCATED_SPACE(dev);
+       /* limit our allocations to order-0 page */
+       size = min_t(int, size, SKB_MAX_ORDER(0, 0));
+       skb = sock_alloc_send_skb(sk, size, 1, &err);
 
        if (!skb)
                return NULL;