| Message ID | 1245282099.31588.69.camel@johannes.local |
|---|---|
| State | RFC, archived |
| Delegated to: | David Miller |
| Headers | show |
Johannes Berg <johannes@sipsolutions.net> writes: > On Wed, 2009-06-17 at 16:24 -0700, Eric W. Biederman wrote: > >> > So it looks like I can also use rcu_read_lock(), but there's no >> > for_each_net_rcu(), should there be? >> >> I'm not using rcu safe list manipulation. What makes it look like >> rcu_read_lock() is safe? > > Indeed. I was looking at rcu_barrier() only. How about the patch below? > > johannes > > Subject: net: make namespace iteration possible under RCU > > We already call rcu_barrier(), so all we need to take > care of is using proper RCU list add/del primitives. This of course gives you a network namespace that can be found by for_each_net rcu while the per net exit functions are running. I think that opens up to races that I don't want to think about. I still haven't tracked down how I am occasionally getting time wait sockets with an invalid network namespace. Eric -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, 2009-06-17 at 16:54 -0700, Eric W. Biederman wrote: > > Subject: net: make namespace iteration possible under RCU > > > > We already call rcu_barrier(), so all we need to take > > care of is using proper RCU list add/del primitives. > > This of course gives you a network namespace that can be found by for_each_net rcu > while the per net exit functions are running. I think that opens up to races > that I don't want to think about. Indeed. Can we move the rcu_barrier() up? Or we could insert a synchronize_rcu() (which is sufficient for rcu_read_lock) before the exit functions are run? > I still haven't tracked down how I am occasionally getting time wait sockets > with an invalid network namespace. Ouch. johannes
Johannes Berg <johannes@sipsolutions.net> writes: > On Wed, 2009-06-17 at 16:54 -0700, Eric W. Biederman wrote: > >> > Subject: net: make namespace iteration possible under RCU >> > >> > We already call rcu_barrier(), so all we need to take >> > care of is using proper RCU list add/del primitives. >> >> This of course gives you a network namespace that can be found by for_each_net rcu >> while the per net exit functions are running. I think that opens up to races >> that I don't want to think about. > > Indeed. Can we move the rcu_barrier() up? Or we could insert a > synchronize_rcu() (which is sufficient for rcu_read_lock) before the > exit functions are run? I don't think we can move the barrier. But adding an extra synchronize_rcu should be fine. We are talking about the slowest of the slow paths here. It is so slow it even gets it's own kernel thread. Eric -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
--- wireless-testing.orig/include/net/net_namespace.h 2009-06-18 01:36:26.000000000 +0200 +++ wireless-testing/include/net/net_namespace.h 2009-06-18 01:39:35.000000000 +0200 @@ -211,6 +211,9 @@ static inline struct net *read_pnet(stru #define for_each_net(VAR) \ list_for_each_entry(VAR, &net_namespace_list, list) +#define for_each_net_rcu(VAR) \ + list_for_each_entry_rcu(VAR, &net_namespace_list, list) + #ifdef CONFIG_NET_NS #define __net_init #define __net_exit --- wireless-testing.orig/net/core/net_namespace.c 2009-06-18 01:36:39.000000000 +0200 +++ wireless-testing/net/core/net_namespace.c 2009-06-18 01:38:36.000000000 +0200 @@ -6,6 +6,7 @@ #include <linux/delay.h> #include <linux/sched.h> #include <linux/idr.h> +#include <linux/rculist.h> #include <net/net_namespace.h> #include <net/netns/generic.h> @@ -134,7 +135,7 @@ struct net *copy_net_ns(unsigned long fl err = setup_net(new_net); if (!err) { rtnl_lock(); - list_add_tail(&new_net->list, &net_namespace_list); + list_add_tail_rcu(&new_net->list, &net_namespace_list); rtnl_unlock(); } mutex_unlock(&net_mutex); @@ -163,7 +164,7 @@ static void cleanup_net(struct work_stru /* Don't let anyone else find us. */ rtnl_lock(); - list_del(&net->list); + list_del_rcu(&net->list); rtnl_unlock(); /* Run all of the network namespace exit methods */ @@ -227,7 +228,7 @@ static int __init net_ns_init(void) err = setup_net(&init_net); rtnl_lock(); - list_add_tail(&init_net.list, &net_namespace_list); + list_add_tail_rcu(&init_net.list, &net_namespace_list); rtnl_unlock(); mutex_unlock(&net_mutex);