Re: [PATCH 3/6] netfilter: ctnetlink: use GFP_ATOMIC in all allocations

On Tue, Mar 06, 2012 at 04:50:21AM -0800, Eric Dumazet wrote:

On Tue, 2012-03-06 at 12:22 +0100, pablo@netfilter.org wrote:

quoted

From: Pablo Neira Ayuso <pablo@netfilter.org>

All ctnetlink operations are invoked inside rcu_read_lock
(see net/netfilter/nfnetlink.c).

Allocations have to be atomic, as RCU requires.

Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
---
 net/netfilter/nf_conntrack_netlink.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/net/netfilter/nf_conntrack_netlink.c b/net/netfilter/nf_conntrack_netlink.c
index 1068769..867843f 100644
--- a/net/netfilter/nf_conntrack_netlink.c
+++ b/net/netfilter/nf_conntrack_netlink.c

@@ -1002,7 +1002,7 @@ ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
 	ct = nf_ct_tuplehash_to_ctrack(h);
 
 	err = -ENOMEM;
-	skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
+	skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
 	if (skb2 == NULL) {
 		nf_ct_put(ct);
 		return -ENOMEM;

@@ -1865,7 +1865,7 @@ ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
 	}
 
 	err = -ENOMEM;
-	skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
+	skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
 	if (skb2 == NULL) {
 		nf_ct_expect_put(exp);
 		goto out;

This cant be right.

Really this must be kept as GFP_KERNEL allocations.

Only if .call_rcu member is used in place of .call rcu_read_lock() is
held instead of nfnl_lock().

I thought we couldn't sleep while holding rcu_read_lock.

You should take a look at all GFP_ATOMIC uses in
net/netfilter/nf_conntrack_netlink.c and check if they can be GFP_KERNEL
instead.

David, can you take all patches except this one?

I'll have to rebase my tree after this, sorry.