net: dont hold rtnl mutex during netlink dump callbacks

Four years ago, Patrick made a change to hold rtnl mutex during netlink
dump callbacks.

I believe it was a wrong move. This slows down concurrent dumps, making
good old /proc/net/ files faster than rtnetlink in some situations.

This occurred to me because one "ip link show dev ..." was _very_ slow
on a workload adding/removing network devices in background.

All dump callbacks are able to use RCU locking now, so this patch does
roughly a revert of commits :

1c2d670f366 : [RTNETLINK]: Hold rtnl_mutex during netlink dump callbacks
6313c1e0992 : [RTNETLINK]: Remove unnecessary locking in dump callbacks

This let writers fight for rtnl mutex and readers going full speed.

It also takes care of phonet : phonet_route_get() is now called from rcu
read section. I renamed it to phonet_route_get_rcu()

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Patrick McHardy <kaber@trash.net>
Cc: Remi Denis-Courmont <remi.denis-courmont@nokia.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 3 insertions, 1 deletions

diff --git a/net/ipv6/ip6_fib.c b/net/ipv6/ip6_fib.c
index dd88df0a5d7f..4076a0b14b20 100644
--- a/net/ipv6/ip6_fib.c
+++ b/net/ipv6/ip6_fib.c
@@ -394,10 +394,11 @@ static int inet6_dump_fib(struct sk_buff *skb, struct netlink_callback *cb)
         arg.net = net;
         w->args = &arg;
 
+        rcu_read_lock();
         for (h = s_h; h < FIB6_TABLE_HASHSZ; h++, s_e = 0) {
                 e = 0;
                 head = &net->ipv6.fib_table_hash[h];
-                hlist_for_each_entry(tb, node, head, tb6_hlist) {
+                hlist_for_each_entry_rcu(tb, node, head, tb6_hlist) {
                         if (e < s_e)
                                 goto next;
                         res = fib6_dump_table(tb, skb, cb);
@@ -408,6 +409,7 @@ next:
                 }
         }
 out:
+        rcu_read_unlock();
         cb->args[1] = e;
         cb->args[0] = h;