[PATCH net-next v3 0/4] net: make moving a device to another netns cheaper

[PATCH net-next v3 0/4] net: make moving a device to another netns cheaper

[Qiliang Yuan]

Moving a device to another netns waits in rcu_barrier() with RTNL held.
Container and sandbox managers move one end of a veth pair for every
new network namespace, so this barrier limits how fast such namespaces
can be created on the whole system, to under 80 per second in a 4
vCPU guest however many threads create them.

The barrier was added for NETDEV_UNREGISTER_FINAL, which is long gone,
but it still keeps cached IPv4 routes through the moved device from
being accepted in the old netns, and it lets the RCU callbacks of the
nexthops removed with the device flush their IPv6 route exceptions
before the move. Routes through nexthop objects are not fully covered by
it: their cached dst entries are released by an RCU callback queued from
another RCU callback, so a socket in the old netns can still send
through a device that now belongs to another netns whenever the route
generation numbers of the two netns happen to match.

Patches 1 and 2 draw the IPv4 rt_genid and the IPv6 fib6_sernum bumps
from sequences shared by all netns, so a cached route no longer matches
another netns. Patch 3 flushes the cached routes of a nexthop group when
one of its members is removed, instead of leaving the IPv6 exceptions
through that member valid until it is freed. With these, nothing relies
on the barrier, and patch 4 removes it. Patches 1 to 3 fix bugs that
exist today and can go to net on their own if preferred.

Without patch 3, removing the barrier lets a socket keep sending through
a moved device that carries no IPv6 address and is a member of a nexthop
group, until its exceptions are flushed. Datagrams a socket in netns A
sent through such a device after it moved to netns B, 10 trials:

                   patches 1, 2, 4     patches 1-4
  datagrams               561-6525               0

Creating 400 netns, each with a veth pair whose host end is moved back
to the initial netns, a tap device, addresses and routes, in a 4 vCPU
guest, median of 5 runs:

                   before      after
  1 thread        46 /s       56 /s
  4 threads       58 /s       84 /s
  16 threads      75 /s      146 /s

Signed-off-by: Qiliang Yuan <qiliang.yuan@linux.dev>
---
V2 -> V3:
- Add patch 3: flush the cached routes of a group when one of its
  nexthops is removed. Without it, patch 4 lets a socket keep sending
  through a moved device that has no IPv6 address (Sashiko review)
- 1/4, 2/4: say that cached routes carry values drawn from the sequence
  because inserting a route bumps it, not because NETDEV_UP does, and
  that this holds until the sequence wraps (Sashiko review)
- 1/4, 2/4: add Kuniyuki's Reviewed-by, the code changes only in comments
- 4/4: commit message only, keep Eric's and Kuniyuki's Reviewed-by
- Rebase onto net-next and remeasure in another guest, where the barrier
  costs about 4 ms per move instead of about 10 ms, so the gains are
  smaller than in v2
- Send from my linux.dev address

V1 -> V2:
- 2/3: keep the initial fib6_sernum of each netns and never hand it out
  from the shared sequence, instead of drawing a value at netns creation
  (Eric Dumazet); 1/3: do the same for rt_genid
- 2/3: note that fib6_sernum is no longer monotonic per netns (Eric
  Dumazet); 1/3: likewise for rt_genid
- Drop Kuniyuki's Reviewed-by from 1/3 and 2/3, which changed
- Rebase onto net-next and remeasure

v2: https://lore.kernel.org/r/20261019-bug-net-netns-move-rcu-barrier-v2-0-b6eb7780c492@gmail.com
v1: https://lore.kernel.org/r/20261016-bug-net-netns-move-rcu-barrier-v1-0-4157d798658f@gmail.com

---
Qiliang Yuan (4):
      ipv4: draw rt_genid from a sequence shared by all netns
      ipv6: take fib6_sernum from a sequence shared by all netns
      nexthop: flush cached routes when a nexthop leaves a group
      net: remove rcu_barrier() from __dev_change_net_namespace()

 include/net/net_namespace.h |  5 +----
 net/core/dev.c              |  1 -
 net/ipv4/nexthop.c          | 10 ++++++++++
 net/ipv4/route.c            | 21 +++++++++++++++++++++
 net/ipv6/ip6_fib.c          | 16 +++++++++++++---
 5 files changed, 45 insertions(+), 8 deletions(-)
---
base-commit: fa5196392691051963726787944db872420dfa75
change-id: 20261016-bug-net-netns-move-rcu-barrier-d998c6395830

Best regards,
-- 
Qiliang Yuan <qiliang.yuan@linux.dev>

[PATCH net-next v3 1/4] ipv4: draw rt_genid from a sequence shared by all netns

[Qiliang Yuan]

An IPv4 route records the rt_genid of the netns it was created in, and
rt_is_expired() compares it with the rt_genid of the netns its device
belongs to when the route is checked. Every netns starts its rt_genid
at 0 and increments it on each cache flush, so different netns often
hold the same value.

When a device moves to another netns, routes cached through it remain
usable until dst_dev_put() points them to blackhole_netdev. For a route
using a nexthop object this happens in nexthop_free_rcu(), which
free_fib_info_rcu() queues from inside an RCU callback, so the
rcu_barrier() in __dev_change_net_namespace() does not wait for it. If
the rt_genid of the new netns matches the cached one in that window,
ipv4_dst_check() accepts the route and a socket in the old netns keeps
sending through a device that now belongs to another netns.

Take every rt_genid bump from one sequence shared by all netns, which
never hands out 0, the value every netns starts with. Inserting a route
bumps rt_genid, so the routes a netns caches carry values drawn from the
sequence, which no other netns holds until it wraps after 2^32 bumps.
Concurrent bumps of one netns can leave the older of two values, so
rt_genid is no longer monotonic per netns. It is only compared for
equality, so this costs at most an extra route lookup. A connected UDP
socket in netns A caches a route through a nexthop object on a veth. The
veth is moved to netns B and brought up with one RTM_NEWLINK, then the
socket sends 200000 datagrams. B's rt_genid is bumped 0 to 20 times
before the move, 63 trials:

                                before     after
  trials sending after move          3         0
  datagrams sent after move       3132         0

Fixes: 4c7e8084fd46 ("ipv4: Plumb support for nexthop object in a fib_info")
Signed-off-by: Qiliang Yuan <qiliang.yuan@linux.dev>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
---
 include/net/net_namespace.h |  5 +----
 net/ipv4/route.c            | 21 +++++++++++++++++++++
 2 files changed, 22 insertions(+), 4 deletions(-)

diff --git a/include/net/net_namespace.h b/include/net/net_namespace.h
index c29a22ca178fa..140858a2fe6d4 100644
--- a/include/net/net_namespace.h
+++ b/include/net/net_namespace.h
@@ -557,10 +557,7 @@ static inline int rt_genid_ipv6(const struct net *net)
 }
 #endif
 
-static inline void rt_genid_bump_ipv4(struct net *net)
-{
-	atomic_inc(&net->ipv4.rt_genid);
-}
+void rt_genid_bump_ipv4(struct net *net);
 
 extern void (*__fib6_flush_trees)(struct net *net);
 static inline void rt_genid_bump_ipv6(struct net *net)
diff --git a/net/ipv4/route.c b/net/ipv4/route.c
index d0a5f6d77094c..16d470ab50142 100644
--- a/net/ipv4/route.c
+++ b/net/ipv4/route.c
@@ -393,6 +393,27 @@ static inline int ip_rt_proc_init(void)
 }
 #endif /* CONFIG_PROC_FS */
 
+/*
+ * Draw rt_genid bumps from one sequence shared by all netns. A route keeps
+ * the genid of the netns it was created in, and is checked against the netns
+ * its device is in now, so distinct values make that check fail once the
+ * device has moved. Every netns starts at 0, which the sequence never hands
+ * out, and inserting a route bumps the genid, so cached routes carry values
+ * drawn from the sequence.
+ */
+static atomic_t rt_genid_seq;
+
+void rt_genid_bump_ipv4(struct net *net)
+{
+	int genid;
+
+	do {
+		genid = atomic_inc_return(&rt_genid_seq);
+	} while (!genid);
+	atomic_set(&net->ipv4.rt_genid, genid);
+}
+EXPORT_SYMBOL_GPL(rt_genid_bump_ipv4);
+
 static inline bool rt_is_expired(const struct rtable *rth)
 {
 	bool res;

-- 
2.43.0

[PATCH net-next v3 2/4] ipv6: take fib6_sernum from a sequence shared by all netns

[Qiliang Yuan]

A per-cpu IPv6 route of a nexthop object records fib6_sernum of the
netns it was created in, and rt6_is_valid() compares it with fib6_sernum
of the netns its device belongs to when the route is checked. Every netns
starts its fib6_sernum at 1 and increments it on each tree change, so
different netns often hold the same value.

When a device moves to another netns, such a route remains usable until
nexthop_free_rcu() releases it, and fib6_info_destroy_rcu() queues that
callback from inside an RCU callback, after the device has moved. If the
fib6_sernum of the new netns matches the cached one in that window,
ip6_dst_check() accepts the route and a socket in the old netns keeps
sending through a device that now belongs to another netns.

Take every fib6_sernum bump from one sequence shared by all netns, which
only hands out values from 2 up: every netns still starts at 1, and 0
means there is no sernum to check. Inserting a route bumps fib6_sernum,
so the routes a netns caches carry values drawn from the sequence, which
no other netns holds until it wraps after 2^31 bumps. fib6_new_sernum()
callers only hold their own table's tb6_lock, so two tables of one netns
can store their values in either order, and fib6_sernum is no longer
monotonic per netns. It is only compared for equality, so this costs at
most an extra invalidation. A connected UDP socket in netns A caches a
route through an IPv6 nexthop object on a veth. The veth is moved to
netns B and brought up with one RTM_NEWLINK, then the socket sends
200000 datagrams. A route is added to and removed from B 0 to 20 times
before the move, 63 trials:

                                before     after
  trials sending after move          1         0
  datagrams sent after move       3429         0

Fixes: 8f34e53b60b3 ("ipv6: Use global sernum for dst validation with nexthop objects")
Signed-off-by: Qiliang Yuan <qiliang.yuan@linux.dev>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
---
 net/ipv6/ip6_fib.c | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)

diff --git a/net/ipv6/ip6_fib.c b/net/ipv6/ip6_fib.c
index e1d49897d6ab0..827051cbf7482 100644
--- a/net/ipv6/ip6_fib.c
+++ b/net/ipv6/ip6_fib.c
@@ -123,13 +123,23 @@ static void fib6_walkers_skip_rt(struct net *net, struct fib6_info *rt)
 	fib6_walkers_replace_leaf(net, rt, next);
 }
 
+/*
+ * Draw fib6_sernum bumps from one sequence shared by all netns, so that a
+ * route cached against one netns does not match the sernum of another one,
+ * e.g. after its device moved, until the sequence wraps. Every netns starts
+ * at 1 and the sequence only hands out values from 2 up: 0 means no sernum
+ * to check.
+ */
+static atomic_t fib6_sernum_seq = ATOMIC_INIT(1);
+
 static int fib6_new_sernum(struct net *net)
 {
-	int new, old = atomic_read(&net->ipv6.fib6_sernum);
+	int new, old = atomic_read(&fib6_sernum_seq);
 
 	do {
-		new = old < INT_MAX ? old + 1 : 1;
-	} while (!atomic_try_cmpxchg(&net->ipv6.fib6_sernum, &old, new));
+		new = old < INT_MAX ? old + 1 : 2;
+	} while (!atomic_try_cmpxchg(&fib6_sernum_seq, &old, new));
+	atomic_set(&net->ipv6.fib6_sernum, new);
 
 	return new;
 }

-- 
2.43.0

[PATCH net-next v3 3/4] nexthop: flush cached routes when a nexthop leaves a group

[Qiliang Yuan]

When a nexthop is removed while other members keep its group alive, only
the group array changes: the routes through the group stay in the FIB,
and the dsts they cached through the removed nexthop are released when
that nexthop is freed, from an RCU callback.

Until then those dsts stay valid. IPv6 route exceptions, such as the
ones created on Packet Too Big, are checked against the cookie of their
FIB node, which does not change, so a socket keeps sending through the
removed nexthop. When the nexthop goes away with its device, removing
the device's IPv6 addresses usually flushes the routing tree first, but
not when the device carries no IPv6 address.

Flush the cached routes of each affected group once all readers see its
new array, as replacing a group already does. A connected UDP socket in
netns A sends to a route through group {nh 1, nh 2}, on an IPv6
exception of nh 1 created by Packet Too Big, while nh 1 is deleted.
Datagrams still sent through nh 1 after the new group is published and
before nh 1 is freed, 10 trials, the one or two left after the change
being sent between synchronize_net() returning and the flush:

                          before      after
  datagrams per trial   395-4423        1-2

Fixes: 430a049190de ("nexthop: Add support for nexthop groups")
Signed-off-by: Qiliang Yuan <qiliang.yuan@linux.dev>
---
 net/ipv4/nexthop.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/net/ipv4/nexthop.c b/net/ipv4/nexthop.c
index 42e55b5a755ee..e0089a6450e75 100644
--- a/net/ipv4/nexthop.c
+++ b/net/ipv4/nexthop.c
@@ -2109,6 +2109,9 @@ remove_nh_grp_entry(struct net *net, struct nh_grp_entry *nhge,
 	return false;
 }
 
+static void nh_rt_cache_flush(struct net *net, struct nexthop *nh,
+			      struct nexthop *replaced_nh);
+
 static bool __must_check
 remove_nexthop_from_groups(struct net *net, struct nexthop *nh,
 			   struct nl_info *nlinfo)
@@ -2130,6 +2133,13 @@ remove_nexthop_from_groups(struct net *net, struct nexthop *nh,
 	/* make sure all see the newly published array before releasing rtnl */
 	synchronize_net();
 
+	/* Routes through the groups can still hold dsts cached through @nh,
+	 * which are only released when @nh is freed. Invalidate them now that
+	 * no reader can pick @nh from the old groups any more.
+	 */
+	list_for_each_entry(nhge, &deferred_free, nh_list)
+		nh_rt_cache_flush(net, nhge->nh_parent, nh);
+
 	/* Now safe to free percpu stats — all RCU readers have finished */
 	list_for_each_entry_safe(nhge, tmp, &deferred_free, nh_list) {
 		list_del(&nhge->nh_list);

-- 
2.43.0

[PATCH net-next v3 4/4] net: remove rcu_barrier() from __dev_change_net_namespace()

[Qiliang Yuan]

Commit 6549dd43c043 ("net: dev: fix the incorrect hold of net
namespace's lo device") added rcu_barrier() before
NETDEV_UNREGISTER_FINAL in dev_change_net_namespace(), so that dst
entries released by RCU callbacks were on the dst garbage list before
dst_ifdown() moved them to the loopback device of the old netns.
Commit 070f2d7e264a ("net: Drop NETDEV_UNREGISTER_FINAL") removed the
event but kept the barrier, and dead dst entries now point to
blackhole_netdev instead of a loopback device.

The barrier waits for the RCU callbacks pending on every CPU while RTNL
is held, which takes a full grace period on each netns move. Container
and sandbox managers move one end of a veth pair for every new network
namespace, so this limits how fast such namespaces can be created on the
whole system, regardless of how many threads create them.

Remove the barrier. The RCU callbacks queued by NETDEV_UNREGISTER
handlers free memory and drop device references, which the move does not
wait for. Routes cached through the moved device are rejected once its
netns changes: IPv4 and per-cpu IPv6 routes no longer match the rt_genid
or fib6_sernum of another netns, and the IPv6 exceptions of a nexthop
that leaves a group were already flushed when it left. Creating 400
netns, each with a veth pair whose host end is moved back to the initial
netns, a tap device, addresses and routes, in a 4 vCPU guest, median of
5 runs:

                   before      after
  1 thread        46 /s       56 /s
  4 threads       58 /s       84 /s
  16 threads      75 /s      146 /s

Signed-off-by: Qiliang Yuan <qiliang.yuan@linux.dev>
Reviewed-by: Eric Dumazet <edumazet@kernel.org>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
---
 net/core/dev.c | 1 -
 1 file changed, 1 deletion(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 3b60b15e2175e..de1fd77440543 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -13044,7 +13044,6 @@ int __dev_change_net_namespace(struct net_device *dev, struct net *net,
 	 * the device is just moving and can keep their slaves up.
 	 */
 	call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
-	rcu_barrier();
 
 	new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL);
 

-- 
2.43.0