[PATCH net-next 0/3] net: make moving a device to another netns cheaper

[PATCH net-next 0/3] 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 about 75 per second in a 16
vCPU guest no matter how 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. Routes through nexthop objects are not
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 make the IPv4 rt_genid and the IPv6 fib6_sernum unique
across netns, which fixes that leak for nexthop objects and leaves
nothing that relies on the barrier. Patch 3 removes it. Patches 1 and 2
fix bugs that exist today and can go to net on their own if preferred.

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

                   before      after
  1 thread        67 /s      440 /s
  4 threads       75 /s      409 /s
  16 threads      76 /s      283 /s

Signed-off-by: Qiliang Yuan <odys.yuan@gmail.com>
---
Qiliang Yuan (3):
      ipv4: draw rt_genid from a sequence shared by all netns
      ipv6: take fib6_sernum from a sequence shared by all netns
      net: remove rcu_barrier() from __dev_change_net_namespace()

 include/net/net_namespace.h |  5 +----
 net/core/dev.c              |  1 -
 net/ipv4/route.c            | 16 +++++++++++++++-
 net/ipv6/af_inet6.c         |  1 -
 net/ipv6/ip6_fib.c          | 14 ++++++++++++--
 5 files changed, 28 insertions(+), 9 deletions(-)
---
base-commit: 6b3e73774815910a39d8ca75b9ad7df963f61d50
change-id: 20261016-bug-net-netns-move-rcu-barrier-d998c6395830

Best regards,
-- 
Qiliang Yuan <odys.yuan@gmail.com>

[PATCH net-next 1/3] 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, including the initial one, from one global
sequence, so that a cached route never matches the rt_genid of another
netns.

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          2         0
  datagrams sent after move       2353         0

Fixes: 4c7e8084fd46 ("ipv4: Plumb support for nexthop object in a fib_info")
Signed-off-by: Qiliang Yuan <odys.yuan@gmail.com>
---
 include/net/net_namespace.h |  5 +----
 net/ipv4/route.c            | 16 +++++++++++++++-
 2 files changed, 16 insertions(+), 5 deletions(-)

diff --git a/include/net/net_namespace.h b/include/net/net_namespace.h
index 46b4c67e2966f..9c78f1d90fb62 100644
--- a/include/net/net_namespace.h
+++ b/include/net/net_namespace.h
@@ -556,10 +556,7 @@ 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 595a3b7f17830..bd52d14476f6f 100644
--- a/net/ipv4/route.c
+++ b/net/ipv4/route.c
@@ -393,6 +393,20 @@ static inline int ip_rt_proc_init(void)
 }
 #endif /* CONFIG_PROC_FS */
 
+/*
+ * Draw every rt_genid 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. Distinct values make that check fail once the device
+ * has moved to another netns.
+ */
+static atomic_t rt_genid_seq;
+
+void rt_genid_bump_ipv4(struct net *net)
+{
+	atomic_set(&net->ipv4.rt_genid, atomic_inc_return(&rt_genid_seq));
+}
+EXPORT_SYMBOL_GPL(rt_genid_bump_ipv4);
+
 static inline bool rt_is_expired(const struct rtable *rth)
 {
 	bool res;
@@ -3716,7 +3730,7 @@ static struct pernet_operations __net_initdata ip_rt_ops = {
 
 static __net_init int rt_genid_init(struct net *net)
 {
-	atomic_set(&net->ipv4.rt_genid, 0);
+	rt_genid_bump_ipv4(net);
 	atomic_set(&net->fnhe_genid, 0);
 	atomic_set(&net->ipv4.dev_addr_genid, get_random_u32());
 	return 0;

-- 
2.43.0

Re: [PATCH net-next 1/3] ipv4: draw rt_genid from a sequence shared by all netns

[kuniyu]

On Fri, 16 Oct 2026 14:09:53 +0800, Qiliang Yuan wrote:
> 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, including the initial one, from one global
> sequence, so that a cached route never matches the rt_genid of another
> netns.
> 
> 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          2         0
>   datagrams sent after move       2353         0
> 
> Fixes: 4c7e8084fd46 ("ipv4: Plumb support for nexthop object in a fib_info")
> Signed-off-by: Qiliang Yuan <odys.yuan@gmail.com>

Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>

[PATCH net-next 2/3] 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, including the initial one, from one global
sequence, so that a cached route never matches the fib6_sernum of
another netns.

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          3         0
  datagrams sent after move       8187         0

Fixes: 8f34e53b60b3 ("ipv6: Use global sernum for dst validation with nexthop objects")
Signed-off-by: Qiliang Yuan <odys.yuan@gmail.com>
---
 net/ipv6/af_inet6.c |  1 -
 net/ipv6/ip6_fib.c  | 14 ++++++++++++--
 2 files changed, 12 insertions(+), 3 deletions(-)

diff --git a/net/ipv6/af_inet6.c b/net/ipv6/af_inet6.c
index a9b8b4b42a1af..db4cccfdc565c 100644
--- a/net/ipv6/af_inet6.c
+++ b/net/ipv6/af_inet6.c
@@ -945,7 +945,6 @@ static int __net_init inet6_net_init(struct net *net)
 	net->ipv6.sysctl.max_dst_opts_len = IP6_DEFAULT_MAX_DST_OPTS_LEN;
 	net->ipv6.sysctl.max_hbh_opts_len = IP6_DEFAULT_MAX_HBH_OPTS_LEN;
 	net->ipv6.sysctl.fib_notify_on_flag_change = 0;
-	atomic_set(&net->ipv6.fib6_sernum, 1);
 
 	net->ipv6.sysctl.ioam6_id = IOAM6_DEFAULT_ID;
 	net->ipv6.sysctl.ioam6_id_wide = IOAM6_DEFAULT_ID_WIDE;
diff --git a/net/ipv6/ip6_fib.c b/net/ipv6/ip6_fib.c
index 0dc6ebee6e568..c66466be64451 100644
--- a/net/ipv6/ip6_fib.c
+++ b/net/ipv6/ip6_fib.c
@@ -89,13 +89,21 @@ static void fib6_walker_unlink(struct net *net, struct fib6_walker *w)
 	write_unlock_bh(&net->ipv6.fib6_walker_lock);
 }
 
+/*
+ * Draw every fib6_sernum from one sequence shared by all netns, so that a
+ * route cached against one netns never matches the sernum of another one,
+ * e.g. after its device moved.
+ */
+static atomic_t fib6_sernum_seq;
+
 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));
+	} while (!atomic_try_cmpxchg(&fib6_sernum_seq, &old, new));
+	atomic_set(&net->ipv6.fib6_sernum, new);
 
 	return new;
 }
@@ -2466,6 +2474,8 @@ static int __net_init fib6_net_init(struct net *net)
 	size_t size = sizeof(struct hlist_head) * FIB6_TABLE_HASHSZ;
 	int err;
 
+	fib6_new_sernum(net);
+
 	err = fib6_notifier_init(net);
 	if (err)
 		return err;

-- 
2.43.0

Re: [PATCH net-next 2/3] ipv6: take fib6_sernum from a sequence shared by all netns

[kuniyu]

On Fri, 16 Oct 2026 14:09:54 +0800, Qiliang Yuan wrote:
> 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, including the initial one, from one global
> sequence, so that a cached route never matches the fib6_sernum of
> another netns.
> 
> 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          3         0
>   datagrams sent after move       8187         0
> 
> Fixes: 8f34e53b60b3 ("ipv6: Use global sernum for dst validation with nexthop objects")
> Signed-off-by: Qiliang Yuan <odys.yuan@gmail.com>

Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>

Re: [PATCH net-next 2/3] ipv6: take fib6_sernum from a sequence shared by all netns

[Eric Dumazet]

Le ven. 16 oct. 2026 à 08:10, Qiliang Yuan <odys.yuan@gmail.com> a écrit :
>
> 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, including the initial one, from one global
> sequence, so that a cached route never matches the fib6_sernum of
> another netns.
>
> 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          3         0
>   datagrams sent after move       8187         0
>
> Fixes: 8f34e53b60b3 ("ipv6: Use global sernum for dst validation with nexthop objects")
> Signed-off-by: Qiliang Yuan <odys.yuan@gmail.com>
> ---
>  net/ipv6/af_inet6.c |  1 -
>  net/ipv6/ip6_fib.c  | 14 ++++++++++++--
>  2 files changed, 12 insertions(+), 3 deletions(-)
>
> diff --git a/net/ipv6/af_inet6.c b/net/ipv6/af_inet6.c
> index a9b8b4b42a1af..db4cccfdc565c 100644
> --- a/net/ipv6/af_inet6.c
> +++ b/net/ipv6/af_inet6.c
> @@ -945,7 +945,6 @@ static int __net_init inet6_net_init(struct net *net)
>         net->ipv6.sysctl.max_dst_opts_len = IP6_DEFAULT_MAX_DST_OPTS_LEN;
>         net->ipv6.sysctl.max_hbh_opts_len = IP6_DEFAULT_MAX_HBH_OPTS_LEN;
>         net->ipv6.sysctl.fib_notify_on_flag_change = 0;
> -       atomic_set(&net->ipv6.fib6_sernum, 1);
>
>         net->ipv6.sysctl.ioam6_id = IOAM6_DEFAULT_ID;
>         net->ipv6.sysctl.ioam6_id_wide = IOAM6_DEFAULT_ID_WIDE;
> diff --git a/net/ipv6/ip6_fib.c b/net/ipv6/ip6_fib.c
> index 0dc6ebee6e568..c66466be64451 100644
> --- a/net/ipv6/ip6_fib.c
> +++ b/net/ipv6/ip6_fib.c
> @@ -89,13 +89,21 @@ static void fib6_walker_unlink(struct net *net, struct fib6_walker *w)
>         write_unlock_bh(&net->ipv6.fib6_walker_lock);
>  }
>
> +/*
> + * Draw every fib6_sernum from one sequence shared by all netns, so that a
> + * route cached against one netns never matches the sernum of another one,
> + * e.g. after its device moved.
> + */
> +static atomic_t fib6_sernum_seq;
> +
>  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));
> +       } while (!atomic_try_cmpxchg(&fib6_sernum_seq, &old, new));
> +       atomic_set(&net->ipv6.fib6_sernum, new);
>

The atomic_set() of net->ipv6.fib6_sernum after the cmpxchg()
 is not ordered with other writers. fib6_new_sernum() callers only
 hold the per-table tb6_lock, so two tables of the same netns can
 race and leave the netns with the older of two values. Since values
 are never reused, I think this only causes extra invalidations, but
 the changelog should say so: fib6_sernum is no longer monotonic per
 netns.

>         return new;
>  }
> @@ -2466,6 +2474,8 @@ static int __net_init fib6_net_init(struct net *net)
>         size_t size = sizeof(struct hlist_head) * FIB6_TABLE_HASHSZ;
>         int err;
>
> +       fib6_new_sernum(net);

There is no need to draw a value from the shared sequence at netns
creation. Netns creation is exactly what you are trying to make
cheaper, and the shared cacheline does not need to be written there.

You can instead keep per-netns initial values that the shared
sequence never hands out:

keep atomic_set(&net->ipv6.fib6_sernum, 1) in
inet6_net_init() (0 is not usable, rt->sernum == 0 and
FIB6_NO_SERNUM_CHANGE have their own meaning), and make the
sequence never return 1 (restart at 2)

A netns that never bumped can then not match a value from the
sequence. Also, the moved device is down after the move, and
NETDEV_UP in the new netns bumps it (rt_cache_flush() for IPv4,
the route inserted by addrconf_add_dev() for IPv6) before the
device can transmit.

[PATCH net-next 3/3] 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. Cached IPv4 and IPv6 routes through the moved device are
rejected once its netns changes, since rt_genid and fib6_sernum are now
unique across netns.

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 16 vCPU
guest, median of 5 runs:

                   before      after
  1 thread        68 /s      440 /s
  4 threads       74 /s      409 /s
  16 threads      77 /s      283 /s

Signed-off-by: Qiliang Yuan <odys.yuan@gmail.com>
---
 net/core/dev.c | 1 -
 1 file changed, 1 deletion(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 8eb9bf1ff9eb6..ef172905109ee 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -12994,7 +12994,6 @@ 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

Re: [PATCH net-next 3/3] net: remove rcu_barrier() from __dev_change_net_namespace()

[Eric Dumazet]

Le ven. 16 oct. 2026 à 08:10, Qiliang Yuan <odys.yuan@gmail.com> a écrit :
>
> 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. Cached IPv4 and IPv6 routes through the moved device are
> rejected once its netns changes, since rt_genid and fib6_sernum are now
> unique across netns.
>
> 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 16 vCPU
> guest, median of 5 runs:
>
>                    before      after
>   1 thread        68 /s      440 /s
>   4 threads       74 /s      409 /s
>   16 threads      77 /s      283 /s
>
> Signed-off-by: Qiliang Yuan <odys.yuan@gmail.com>
> ---

I think this is mostly fine. I was working on this months ago but got
distracted.

I sent two patches fixing pre-existing issues that would be widened by
your patch.

https://patchwork.kernel.org/project/netdevbpf/patch/20261016084802.2725187-1-edumazet@kernel.org/
https://patchwork.kernel.org/project/netdevbpf/patch/20261016091556.2750902-1-edumazet@kernel.org/

Reviewed-by: Eric Dumazet <edumazet@kernel.org>

Thanks.

Re: [PATCH net-next 3/3] net: remove rcu_barrier() from __dev_change_net_namespace()

[kuniyu]

On Fri, 16 Oct 2026 14:09:55 +0800, Qiliang Yuan wrote:
> 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. Cached IPv4 and IPv6 routes through the moved device are
> rejected once its netns changes, since rt_genid and fib6_sernum are now
> unique across netns.
> 
> 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 16 vCPU
> guest, median of 5 runs:
> 
>                    before      after
>   1 thread        68 /s      440 /s
>   4 threads       74 /s      409 /s
>   16 threads      77 /s      283 /s
> 
> Signed-off-by: Qiliang Yuan <odys.yuan@gmail.com>

Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>

Thanks !