When IPV6_MROUTE_MULTIPLE_TABLES is enabled, multicast routing tables
must be read under RCU or RTNL lock.

Fixes: d1db275dd3f6 ("ipv6: ip6mr: support multiple tables")
Signed-off-by: Stefan Wiehler <stefan.wiehler@nokia.com>
---
 net/ipv6/ip6mr.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/net/ipv6/ip6mr.c b/net/ipv6/ip6mr.c
index b169b27de7e1..39aac81a30f1 100644
--- a/net/ipv6/ip6mr.c
+++ b/net/ipv6/ip6mr.c
@@ -2633,27 +2633,31 @@ static int ip6mr_rtm_getroute(struct sk_buff *in_skb, struct nlmsghdr *nlh,
 		grp = nla_get_in6_addr(tb[RTA_DST]);
 	tableid = tb[RTA_TABLE] ? nla_get_u32(tb[RTA_TABLE]) : 0;
 
+	rcu_read_lock();
 	mrt = ip6mr_get_table(net, tableid ?: RT_TABLE_DEFAULT);
 	if (!mrt) {
+		rcu_read_unlock();
 		NL_SET_ERR_MSG_MOD(extack, "MR table does not exist");
 		return -ENOENT;
 	}
 
 	/* entries are added/deleted only under RTNL */
-	rcu_read_lock();
 	cache = ip6mr_cache_find(mrt, &src, &grp);
-	rcu_read_unlock();
 	if (!cache) {
+		rcu_read_unlock();
 		NL_SET_ERR_MSG_MOD(extack, "MR cache entry not found");
 		return -ENOENT;
 	}
 
 	skb = nlmsg_new(mr6_msgsize(false, mrt->maxvif), GFP_KERNEL);
-	if (!skb)
+	if (!skb) {
+		rcu_read_unlock();
 		return -ENOBUFS;
+	}
 
 	err = ip6mr_fill_mroute(mrt, skb, NETLINK_CB(in_skb).portid,
 				nlh->nlmsg_seq, cache, RTM_NEWROUTE, 0);
+	rcu_read_unlock();
 	if (err < 0) {
 		kfree_skb(skb);
 		return err;
-- 
2.42.0

Stefan Wiehler <stefan.wiehler@nokia.com> wrote:
> When IPV6_MROUTE_MULTIPLE_TABLES is enabled, multicast routing tables
> must be read under RCU or RTNL lock.
> 
> Fixes: d1db275dd3f6 ("ipv6: ip6mr: support multiple tables")
> Signed-off-by: Stefan Wiehler <stefan.wiehler@nokia.com>
> ---
>  net/ipv6/ip6mr.c | 10 +++++++---
>  1 file changed, 7 insertions(+), 3 deletions(-)
> 
> diff --git a/net/ipv6/ip6mr.c b/net/ipv6/ip6mr.c
> index b169b27de7e1..39aac81a30f1 100644
> --- a/net/ipv6/ip6mr.c
> +++ b/net/ipv6/ip6mr.c
> @@ -2633,27 +2633,31 @@ static int ip6mr_rtm_getroute(struct sk_buff *in_skb, struct nlmsghdr *nlh,
>  		grp = nla_get_in6_addr(tb[RTA_DST]);
>  	tableid = tb[RTA_TABLE] ? nla_get_u32(tb[RTA_TABLE]) : 0;
>  
> +	rcu_read_lock();

AFAICS ip6mr_rtm_getroute() runs with RTNL held, so I don't see
why this patch is needed.

>> When IPV6_MROUTE_MULTIPLE_TABLES is enabled, multicast routing tables
>> must be read under RCU or RTNL lock.
>>
>> Fixes: d1db275dd3f6 ("ipv6: ip6mr: support multiple tables")
>> Signed-off-by: Stefan Wiehler <stefan.wiehler@nokia.com>
>> ---
>>  net/ipv6/ip6mr.c | 10 +++++++---
>>  1 file changed, 7 insertions(+), 3 deletions(-)
>>
>> diff --git a/net/ipv6/ip6mr.c b/net/ipv6/ip6mr.c
>> index b169b27de7e1..39aac81a30f1 100644
>> --- a/net/ipv6/ip6mr.c
>> +++ b/net/ipv6/ip6mr.c
>> @@ -2633,27 +2633,31 @@ static int ip6mr_rtm_getroute(struct sk_buff *in_skb, struct nlmsghdr *nlh,
>>               grp = nla_get_in6_addr(tb[RTA_DST]);
>>       tableid = tb[RTA_TABLE] ? nla_get_u32(tb[RTA_TABLE]) : 0;
>>
>> +     rcu_read_lock();
> 
> AFAICS ip6mr_rtm_getroute() runs with RTNL held, so I don't see
> why this patch is needed.

That's true, but it's called neither with RCU nor RTNL lock when
RTNL_FLAG_DOIT_UNLOCKED is set in rtnetlink_rcv_msg():

> if (flags & RTNL_FLAG_DOIT_UNLOCKED) {
> 	doit = link->doit;
> 	rcu_read_unlock();
> 	if (doit)
> 		err = doit(skb, nlh, extack);
> 	module_put(owner);
> 	return err;
> }
> rcu_read_unlock();

I realized now that I completely misunderstood how ip6mr_rtm_dumproute() gets
called - it should be still safe though because mpls_netconf_dump_devconf() and
getaddr_dumpit() hold the RCU lock while mpls_dump_routes() asserts that the
RTNL lock is held. Is that observation correct?

Stefan Wiehler <stefan.wiehler@nokia.com> wrote:
> >> When IPV6_MROUTE_MULTIPLE_TABLES is enabled, multicast routing tables
> >> must be read under RCU or RTNL lock.
> >>
> >> Fixes: d1db275dd3f6 ("ipv6: ip6mr: support multiple tables")
> >> Signed-off-by: Stefan Wiehler <stefan.wiehler@nokia.com>
> >> ---
> >>  net/ipv6/ip6mr.c | 10 +++++++---
> >>  1 file changed, 7 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/net/ipv6/ip6mr.c b/net/ipv6/ip6mr.c
> >> index b169b27de7e1..39aac81a30f1 100644
> >> --- a/net/ipv6/ip6mr.c
> >> +++ b/net/ipv6/ip6mr.c
> >> @@ -2633,27 +2633,31 @@ static int ip6mr_rtm_getroute(struct sk_buff *in_skb, struct nlmsghdr *nlh,
> >>               grp = nla_get_in6_addr(tb[RTA_DST]);
> >>       tableid = tb[RTA_TABLE] ? nla_get_u32(tb[RTA_TABLE]) : 0;
> >>
> >> +     rcu_read_lock();
> > 
> > AFAICS ip6mr_rtm_getroute() runs with RTNL held, so I don't see
> > why this patch is needed.
> 
> That's true, but it's called neither with RCU nor RTNL lock when
> RTNL_FLAG_DOIT_UNLOCKED is set in rtnetlink_rcv_msg():

Sure, but RTNL_FLAG_DOIT_UNLOCKED is not set for this function:

        err = rtnl_register_module(THIS_MODULE, RTNL_FAMILY_IP6MR, RTM_GETROUTE,
                                   ip6mr_rtm_getroute, ip6mr_rtm_dumproute, 0);

(0 == flag field).  So RNTL is held.  Would of course be nice to convert it to RCU
eventually but thats an enhancement, not a bug fix, so this must be in
separate changesets, targetting net and net-next, respectively.

> I realized now that I completely misunderstood how ip6mr_rtm_dumproute() gets
> called - it should be still safe though because mpls_netconf_dump_devconf() and
> getaddr_dumpit() hold the RCU lock while mpls_dump_routes() asserts that the
> RTNL lock is held. Is that observation correct?

{THIS_MODULE, PF_PHONET, RTM_GETADDR, NULL, getaddr_dumpit, 0},
{THIS_MODULE, PF_MPLS, RTM_GETROUTE, mpls_getroute, mpls_dump_routes, 0},

Both get called with RTNL mutex held, but not within an RCU read side
section.

but:
{THIS_MODULE, PF_MPLS, RTM_GETNETCONF, [..] mpls_netconf_dump_devconf,
       	RTNL_FLAG_DUMP_UNLOCKED}, // == no RTNL held

Means: dump callback is invoked without RTNL mutex.  Those functions
are also called without and RCU read-side section.

Both the get (doit) and the dumper function callbacks need to
explicitly opt-in for RTNL-less invocation at register time.