* [PATCH net-next] net/packet: wait for a single RCU grace period in packet_release()
@ 2026-11-18 21:14 Eric Dumazet
2026-11-19 17:59 ` Willem de Bruijn
0 siblings, 1 reply; 3+ messages in thread
From: Eric Dumazet @ 2026-11-18 21:14 UTC (permalink / raw)
To: David S . Miller, Jakub Kicinski, Paolo Abeni, Willem de Bruijn
Cc: Simon Horman, netdev, Eric Dumazet
Closing an AF_PACKET socket waits for one RCU grace period per ring
in packet_set_ring(), then for another one at the end of
packet_release(). A socket with an rx ring (libpcap, tcpdump) pays for
two grace periods, a socket with both rx and tx rings for three.
All the RCU unpublishing is done at the beginning of packet_release(),
when unregister_prot_hook() removes the socket from the ptype lists or
from its fanout group. Nothing is unpublished after that point:
- When closing, packet_set_ring() finds the socket not running and
does not touch the hooks.
- fanout_release() only clears po->fanout and unlinks the group from
fanout_list, both under fanout_mutex. If it drops the last reference
on the group, all members have already been unlinked, so the fanout
hook was removed by __fanout_unlink().
Call fanout_release() and synchronize_net() right after the socket has
been detached, before the rings are torn down, and skip the wait in
packet_set_ring() when closing. That single grace period covers the
rings, po->rollover and the fanout group.
The setsockopt() path of packet_set_ring() is unchanged.
Average close() duration over 20 sockets bound to lo, in a 4 vCPU
virtme-ng guest:
before after
no ring 7.86 ms 7.76 ms
rx ring 8.01 ms 4.01 ms
tx ring 8.00 ms 4.00 ms
rx + tx rings 11.96 ms 4.00 ms
rx + tx rings, fanout 12.02 ms 4.01 ms
Tested with tools/testing/selftests/net/run_afpackettests
(psock_fanout, psock_tpacket, txring_overwrite) on a kernel without
KASAN or lockdep.
Signed-off-by: Eric Dumazet <edumazet@kernel.org>
---
net/packet/af_packet.c | 17 ++++++++++++-----
1 file changed, 12 insertions(+), 5 deletions(-)
diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
index 6ace2a5b2d177e3de89361a3092536a8f6c5e00a..96c5216a9d6b47700da682d864368eead8572297 100644
--- a/net/packet/af_packet.c
+++ b/net/packet/af_packet.c
@@ -3224,6 +3224,15 @@ static int packet_release(struct socket *sock)
packet_flush_mclist(sk);
+ f = fanout_release(sk);
+
+ /* The socket is no longer reachable from the receive path:
+ * unregister_prot_hook() removed it from the ptype lists and
+ * from its fanout group, if any. Wait once for packets currently
+ * being received, this covers the rings, po->rollover and @f.
+ */
+ synchronize_net();
+
lock_sock(sk);
if (po->rx_ring.pg_vec) {
memset(&req_u, 0, sizeof(req_u));
@@ -3236,10 +3245,6 @@ static int packet_release(struct socket *sock)
}
release_sock(sk);
- f = fanout_release(sk);
-
- synchronize_net();
-
kfree(po->rollover);
if (f) {
fanout_release_data(f);
@@ -4674,7 +4679,9 @@ static int packet_set_ring(struct sock *sk, union tpacket_req_u *req_u,
spin_unlock(&po->bind_lock);
- synchronize_net();
+ /* packet_release() already detached the socket and waited. */
+ if (!closing)
+ synchronize_net();
err = -EBUSY;
mutex_lock(&po->pg_vec_lock);
--
2.53.0
^ permalink raw reply related [flat|nested] 3+ messages in thread* Re: [PATCH net-next] net/packet: wait for a single RCU grace period in packet_release()
2026-11-18 21:14 [PATCH net-next] net/packet: wait for a single RCU grace period in packet_release() Eric Dumazet
@ 2026-11-19 17:59 ` Willem de Bruijn
2026-11-19 18:25 ` Eric Dumazet
0 siblings, 1 reply; 3+ messages in thread
From: Willem de Bruijn @ 2026-11-19 17:59 UTC (permalink / raw)
To: Eric Dumazet, David S . Miller, Jakub Kicinski, Paolo Abeni,
Willem de Bruijn
Cc: Simon Horman, netdev, Eric Dumazet
Eric Dumazet wrote:
> Closing an AF_PACKET socket waits for one RCU grace period per ring
> in packet_set_ring(), then for another one at the end of
> packet_release(). A socket with an rx ring (libpcap, tcpdump) pays for
> two grace periods, a socket with both rx and tx rings for three.
>
> All the RCU unpublishing is done at the beginning of packet_release(),
> when unregister_prot_hook() removes the socket from the ptype lists or
> from its fanout group. Nothing is unpublished after that point:
>
> - When closing, packet_set_ring() finds the socket not running and
> does not touch the hooks.
>
> - fanout_release() only clears po->fanout and unlinks the group from
> fanout_list, both under fanout_mutex. If it drops the last reference
> on the group, all members have already been unlinked, so the fanout
> hook was removed by __fanout_unlink().
>
> Call fanout_release() and synchronize_net() right after the socket has
> been detached, before the rings are torn down, and skip the wait in
> packet_set_ring() when closing. That single grace period covers the
> rings, po->rollover and the fanout group.
>
> The setsockopt() path of packet_set_ring() is unchanged.
>
> Average close() duration over 20 sockets bound to lo, in a 4 vCPU
> virtme-ng guest:
>
> before after
> no ring 7.86 ms 7.76 ms
> rx ring 8.01 ms 4.01 ms
> tx ring 8.00 ms 4.00 ms
> rx + tx rings 11.96 ms 4.00 ms
> rx + tx rings, fanout 12.02 ms 4.01 ms
Just curious, any idea why no ring seems to take twice as long?
Don't spend time on it. It's just not obvious to me.
>
> Tested with tools/testing/selftests/net/run_afpackettests
> (psock_fanout, psock_tpacket, txring_overwrite) on a kernel without
> KASAN or lockdep.
>
> Signed-off-by: Eric Dumazet <edumazet@kernel.org>
Reviewed-by: Willem de Bruijn <willemb@google.com>
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH net-next] net/packet: wait for a single RCU grace period in packet_release()
2026-11-19 17:59 ` Willem de Bruijn
@ 2026-11-19 18:25 ` Eric Dumazet
0 siblings, 0 replies; 3+ messages in thread
From: Eric Dumazet @ 2026-11-19 18:25 UTC (permalink / raw)
To: Willem de Bruijn
Cc: David S . Miller, Jakub Kicinski, Paolo Abeni, Willem de Bruijn,
Simon Horman, netdev
Le jeu. 19 nov. 2026 à 18:59, Willem de Bruijn
<willemdebruijn.kernel@gmail.com> a écrit :
>
> Eric Dumazet wrote:
> > Closing an AF_PACKET socket waits for one RCU grace period per ring
> > in packet_set_ring(), then for another one at the end of
> > packet_release(). A socket with an rx ring (libpcap, tcpdump) pays for
> > two grace periods, a socket with both rx and tx rings for three.
> >
> > All the RCU unpublishing is done at the beginning of packet_release(),
> > when unregister_prot_hook() removes the socket from the ptype lists or
> > from its fanout group. Nothing is unpublished after that point:
> >
> > - When closing, packet_set_ring() finds the socket not running and
> > does not touch the hooks.
> >
> > - fanout_release() only clears po->fanout and unlinks the group from
> > fanout_list, both under fanout_mutex. If it drops the last reference
> > on the group, all members have already been unlinked, so the fanout
> > hook was removed by __fanout_unlink().
> >
> > Call fanout_release() and synchronize_net() right after the socket has
> > been detached, before the rings are torn down, and skip the wait in
> > packet_set_ring() when closing. That single grace period covers the
> > rings, po->rollover and the fanout group.
> >
> > The setsockopt() path of packet_set_ring() is unchanged.
> >
> > Average close() duration over 20 sockets bound to lo, in a 4 vCPU
> > virtme-ng guest:
> >
> > before after
> > no ring 7.86 ms 7.76 ms
> > rx ring 8.01 ms 4.01 ms
> > tx ring 8.00 ms 4.00 ms
> > rx + tx rings 11.96 ms 4.00 ms
> > rx + tx rings, fanout 12.02 ms 4.01 ms
>
> Just curious, any idea why no ring seems to take twice as long?
Benchmark artefact: the test closes sockets in a tight loop, so
close() calls synchronize_net() while another grace period is already
in progress, and waits for the rest of that one plus a full one.
Adding a 20ms sleep before close() brings the no ring case to ~4 ms,
same as the others. In the ring cases, the PACKET_{RX,TX}_RING
setsockopt() (which also calls synchronize_net()) sits between bind()
and close(), which presumably absorbs it.
> Don't spend time on it. It's just not obvious to me.
>
> >
> > Tested with tools/testing/selftests/net/run_afpackettests
> > (psock_fanout, psock_tpacket, txring_overwrite) on a kernel without
> > KASAN or lockdep.
> >
> > Signed-off-by: Eric Dumazet <edumazet@kernel.org>
>
> Reviewed-by: Willem de Bruijn <willemb@google.com>
Thanks!
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2026-11-19 18:25 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-18 21:14 [PATCH net-next] net/packet: wait for a single RCU grace period in packet_release() Eric Dumazet
2026-11-19 17:59 ` Willem de Bruijn
2026-11-19 18:25 ` Eric Dumazet
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox