[NET]: Handle disabled preemption in gfp_any()
authorPatrick McHardy <kaber@trash.net>
Tue, 27 Feb 2007 17:56:42 +0000 (09:56 -0800)
committerDavid S. Miller <davem@sunset.davemloft.net>
Wed, 28 Feb 2007 17:42:13 +0000 (09:42 -0800)
ctnetlink uses netlink_unicast from an atomic_notifier_chain
(which is called within a RCU read side critical section)
without holding further locks. netlink_unicast calls netlink_trim
with the result of gfp_any() for the gfp flags, which are passed
down to pskb_expand_header. gfp_any() only checks for softirq
context and returns GFP_KERNEL, resulting in this warning:

BUG: sleeping function called from invalid context at mm/slab.c:3032
in_atomic():1, irqs_disabled():0
no locks held by rmmod/7010.

Call Trace:
 [<ffffffff8109467f>] debug_show_held_locks+0x9/0xb
 [<ffffffff8100b0b4>] __might_sleep+0xd9/0xdb
 [<ffffffff810b5082>] __kmalloc+0x68/0x110
 [<ffffffff811ba8f2>] pskb_expand_head+0x4d/0x13b
 [<ffffffff81053147>] netlink_broadcast+0xa5/0x2e0
 [<ffffffff881cd1d7>] :nfnetlink:nfnetlink_send+0x83/0x8a
 [<ffffffff8834f6a6>] :nf_conntrack_netlink:ctnetlink_conntrack_event+0x94c/0x96a
 [<ffffffff810624d6>] notifier_call_chain+0x29/0x3e
 [<ffffffff8106251d>] atomic_notifier_call_chain+0x32/0x60
 [<ffffffff881d266d>] :nf_conntrack:destroy_conntrack+0xa5/0x1d3
 [<ffffffff881d194e>] :nf_conntrack:nf_ct_cleanup+0x8c/0x12c
 [<ffffffff881d4614>] :nf_conntrack:kill_l3proto+0x0/0x13
 [<ffffffff881d482a>] :nf_conntrack:nf_conntrack_l3proto_unregister+0x90/0x94
 [<ffffffff883551b3>] :nf_conntrack_ipv4:nf_conntrack_l3proto_ipv4_fini+0x2b/0x5d
 [<ffffffff8109d44f>] sys_delete_module+0x1b5/0x1e6
 [<ffffffff8105f245>] trace_hardirqs_on_thunk+0x35/0x37
 [<ffffffff8105911e>] system_call+0x7e/0x83

Since netlink_unicast is supposed to be callable from within RCU
read side critical sections, make gfp_any() check for in_atomic()
instead of in_softirq().

Additionally nfnetlink_send needs to use gfp_any() as well for the
call to netlink_broadcast).

Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/sock.h
net/netfilter/nfnetlink.c

index 03684e702d13464cbc0a48b95354dc540341f504..2c7d60ca354858ac75ead88773d0886aedcc185c 100644 (file)
@@ -1278,7 +1278,7 @@ static inline int sock_writeable(const struct sock *sk)
 
 static inline gfp_t gfp_any(void)
 {
-       return in_softirq() ? GFP_ATOMIC : GFP_KERNEL;
+       return in_atomic() ? GFP_ATOMIC : GFP_KERNEL;
 }
 
 static inline long sock_rcvtimeo(const struct sock *sk, int noblock)
index 11d504d0ac72068d8c88bc794fe4cb8c1284e3d9..bf23e489e4cdc88eec4a81a8c3461ac2cd5008ee 100644 (file)
@@ -197,13 +197,12 @@ EXPORT_SYMBOL_GPL(nfnetlink_has_listeners);
 
 int nfnetlink_send(struct sk_buff *skb, u32 pid, unsigned group, int echo)
 {
-       gfp_t allocation = in_interrupt() ? GFP_ATOMIC : GFP_KERNEL;
        int err = 0;
 
        NETLINK_CB(skb).dst_group = group;
        if (echo)
                atomic_inc(&skb->users);
-       netlink_broadcast(nfnl, skb, pid, group, allocation);
+       netlink_broadcast(nfnl, skb, pid, group, gfp_any());
        if (echo)
                err = netlink_unicast(nfnl, skb, pid, MSG_DONTWAIT);