When booting an arm64 allmodconfig kernel on linux-next (tag
next-20191115). The following "suspicious RCU usage" warning shows up.
This bug seems to have been introduced by commit f0ad0860d01e ("ipv4:
ipmr: support multiple tables") in 2010, but the warning was added only
in this past year by commit 28875945ba98 ("rcu: Add support for
consolidated-RCU reader checking").
[ 32.496021][ T1] =============================
[ 32.497616][ T1] WARNING: suspicious RCU usage
[ 32.499614][ T1] 5.4.0-rc6-next-20191108-00003-gf74bac957b5c-dirty #2 Not tainted
[ 32.502018][ T1] -----------------------------
[ 32.503976][ T1] net/ipv4/ipmr.c:136 RCU-list traversed in non-reader section!!
[ 32.506746][ T1]
[ 32.506746][ T1] other info that might help us debug this:
[ 32.506746][ T1]
[ 32.509794][ T1]
[ 32.509794][ T1] rcu_scheduler_active = 2, debug_locks = 1
[ 32.512661][ T1] 1 lock held by swapper/0/1:
[ 32.514169][ T1] #0: ffffa000150dd678 (pernet_ops_rwsem){+.+.}, at: register_pernet_subsys+0x24/0x50
[ 32.517621][ T1]
[ 32.517621][ T1] stack backtrace:
[ 32.519930][ T1] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.4.0-rc6-next-20191108-00003-gf74bac957b5c-dirty #2
[ 32.523063][ T1] Hardware name: linux,dummy-virt (DT)
[ 32.524787][ T1] Call trace:
[ 32.525946][ T1] dump_backtrace+0x0/0x2d0
[ 32.527433][ T1] show_stack+0x20/0x30
[ 32.528811][ T1] dump_stack+0x204/0x2ac
[ 32.530258][ T1] lockdep_rcu_suspicious+0xf4/0x108
[ 32.531993][ T1] ipmr_get_table+0xc8/0x170
[ 32.533496][ T1] ipmr_new_table+0x48/0xa0
[ 32.535002][ T1] ipmr_net_init+0xe8/0x258
[ 32.536465][ T1] ops_init+0x280/0x2d8
[ 32.537876][ T1] register_pernet_operations+0x210/0x420
[ 32.539707][ T1] register_pernet_subsys+0x30/0x50
[ 32.541372][ T1] ip_mr_init+0x54/0x180
[ 32.542785][ T1] inet_init+0x25c/0x3e8
[ 32.544186][ T1] do_one_initcall+0x4c0/0xad8
[ 32.545757][ T1] kernel_init_freeable+0x3e0/0x500
[ 32.547443][ T1] kernel_init+0x14/0x1f0
[ 32.548875][ T1] ret_from_fork+0x10/0x18
This commit therefore introduces a lockdep-specific variable that
maintains initialization state. It then passes this variable along with
the return value of lockdep_rtnl_is_held() to list_for_each_entry_rcu()
in order to correctly check for proper RCU/locking/initialization state.
Suggested-by: Paul E. McKenney <[email protected]>
Reviewed-by: Paul E. McKenney <[email protected]>
Signed-off-by: Anders Roxell <[email protected]>
---
net/ipv4/ipmr.c | 13 ++++++++++++-
1 file changed, 12 insertions(+), 1 deletion(-)
diff --git a/net/ipv4/ipmr.c b/net/ipv4/ipmr.c
index 6e68def66822..93007c429dae 100644
--- a/net/ipv4/ipmr.c
+++ b/net/ipv4/ipmr.c
@@ -108,9 +108,18 @@ static void igmpmsg_netlink_event(struct mr_table *mrt, struct sk_buff *pkt);
static void mroute_clean_tables(struct mr_table *mrt, int flags);
static void ipmr_expire_process(struct timer_list *t);
+#ifdef CONFIG_PROVE_LOCKING
+int ip_mr_initialized;
+void ip_mr_now_initialized(void) { ip_mr_initialized = 1; }
+#else
+const int ip_mr_initialized = 1;
+void ip_mr_now_initialized(void) { }
+#endif
+
#ifdef CONFIG_IP_MROUTE_MULTIPLE_TABLES
#define ipmr_for_each_table(mrt, net) \
- list_for_each_entry_rcu(mrt, &net->ipv4.mr_tables, list)
+ list_for_each_entry_rcu(mrt, &net->ipv4.mr_tables, list, \
+ (lockdep_rtnl_is_held() || !ip_mr_initialized))
static struct mr_table *ipmr_mr_table_iter(struct net *net,
struct mr_table *mrt)
@@ -3160,6 +3169,8 @@ int __init ip_mr_init(void)
rtnl_register(RTNL_FAMILY_IPMR, RTM_GETLINK,
NULL, ipmr_rtm_dumplink, 0);
+
+ ip_mr_now_initialized();
return 0;
#ifdef CONFIG_IP_PIMSM_V2
--
2.20.1
From: Anders Roxell <[email protected]>
Date: Mon, 18 Nov 2019 10:09:25 +0100
> @@ -108,9 +108,18 @@ static void igmpmsg_netlink_event(struct mr_table *mrt, struct sk_buff *pkt);
> static void mroute_clean_tables(struct mr_table *mrt, int flags);
> static void ipmr_expire_process(struct timer_list *t);
>
> +#ifdef CONFIG_PROVE_LOCKING
> +int ip_mr_initialized;
> +void ip_mr_now_initialized(void) { ip_mr_initialized = 1; }
> +#else
> +const int ip_mr_initialized = 1;
> +void ip_mr_now_initialized(void) { }
> +#endif
This seems excessive and a bit not so pretty.
> +
> #ifdef CONFIG_IP_MROUTE_MULTIPLE_TABLES
> #define ipmr_for_each_table(mrt, net) \
> - list_for_each_entry_rcu(mrt, &net->ipv4.mr_tables, list)
> + list_for_each_entry_rcu(mrt, &net->ipv4.mr_tables, list, \
> + (lockdep_rtnl_is_held() || !ip_mr_initialized))
>
> static struct mr_table *ipmr_mr_table_iter(struct net *net,
> struct mr_table *mrt)
The problematic code path is ipmr_rules_init() done during ipmr_net_init().
You can just wrap this call around RCU locking or take the RTNL mutex.
That way you don't need to rediculous ip_mr_initialized knob which frankly
doesn't even seem accurate to me. It's a centralized global variable
which is holding state about multiple network namespace objects which makes
absolutely no sense at all, it's wrong.
On Tue, Nov 19, 2019 at 02:50:48PM -0800, David Miller wrote:
> From: Anders Roxell <[email protected]>
> Date: Mon, 18 Nov 2019 10:09:25 +0100
>
> > @@ -108,9 +108,18 @@ static void igmpmsg_netlink_event(struct mr_table *mrt, struct sk_buff *pkt);
> > static void mroute_clean_tables(struct mr_table *mrt, int flags);
> > static void ipmr_expire_process(struct timer_list *t);
> >
> > +#ifdef CONFIG_PROVE_LOCKING
> > +int ip_mr_initialized;
> > +void ip_mr_now_initialized(void) { ip_mr_initialized = 1; }
> > +#else
> > +const int ip_mr_initialized = 1;
> > +void ip_mr_now_initialized(void) { }
> > +#endif
>
> This seems excessive and a bit not so pretty.
>
> > +
> > #ifdef CONFIG_IP_MROUTE_MULTIPLE_TABLES
> > #define ipmr_for_each_table(mrt, net) \
> > - list_for_each_entry_rcu(mrt, &net->ipv4.mr_tables, list)
> > + list_for_each_entry_rcu(mrt, &net->ipv4.mr_tables, list, \
> > + (lockdep_rtnl_is_held() || !ip_mr_initialized))
> >
> > static struct mr_table *ipmr_mr_table_iter(struct net *net,
> > struct mr_table *mrt)
>
> The problematic code path is ipmr_rules_init() done during ipmr_net_init().
>
> You can just wrap this call around RCU locking or take the RTNL mutex.
Agreed, that would work quite well.
Thanx, Paul
> That way you don't need to rediculous ip_mr_initialized knob which frankly
> doesn't even seem accurate to me. It's a centralized global variable
> which is holding state about multiple network namespace objects which makes
> absolutely no sense at all, it's wrong.