[PATCH 24/31] net/tcp: Allow asynchronous delete for TCP-AO keys (MKTs)

[Dmitry Safonov <dima@arista.com>]

Delete becomes very, very fast - almost free, but after setsockopt()
syscall returns, the key is still alive until next RCU grace period.
Which is fine for listen sockets as userspace needs to be aware of
setsockopt(TCP_AO) and accept() race and resolve it with verification
by getsockopt() after TCP connection was accepted.

The benchmark results (on non-loaded box, worse with more RCU work pending):
> ok 33    Worst case delete    16384 keys: min=5ms max=10ms mean=6.93904ms stddev=0.263421
> ok 34        Add a new key    16384 keys: min=1ms max=4ms mean=2.17751ms stddev=0.147564
> ok 35 Remove random-search    16384 keys: min=5ms max=10ms mean=6.50243ms stddev=0.254999
> ok 36         Remove async    16384 keys: min=0ms max=0ms mean=0.0296107ms stddev=0.0172078

Co-developed-by: Francesco Ruggeri <fruggeri@arista.com>
Signed-off-by: Francesco Ruggeri <fruggeri@arista.com>
Co-developed-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Salam Noureddine <noureddine@arista.com>
Signed-off-by: Dmitry Safonov <dima@arista.com>
---
 include/uapi/linux/tcp.h |  3 +++
 net/ipv4/tcp_ao.c        | 17 ++++++++++++++++-
 2 files changed, 19 insertions(+), 1 deletion(-)

diff --git a/include/uapi/linux/tcp.h b/include/uapi/linux/tcp.h
index 453187d21da8..42850ae6e99d 100644
--- a/include/uapi/linux/tcp.h
+++ b/include/uapi/linux/tcp.h
@@ -353,6 +353,9 @@ struct tcp_diag_md5sig {
 #define TCP_AO_CMDF_CURR	(1 << 0)	/* Only checks field sndid */
 #define TCP_AO_CMDF_NEXT	(1 << 1)	/* Only checks field rcvid */
 #define TCP_AO_CMDF_ACCEPT_ICMP	(1 << 2)	/* Accept incoming ICMPs */
+#define TCP_AO_CMDF_DEL_ASYNC	(1 << 3)	/* Asynchronious delete, valid
+						 * only for listen sockets
+						 */
 
 #define TCP_AO_GET_CURR		TCP_AO_CMDF_CURR
 #define TCP_AO_GET_NEXT		TCP_AO_CMDF_NEXT
diff --git a/net/ipv4/tcp_ao.c b/net/ipv4/tcp_ao.c
index 5ab16b857c29..8e75432c0cc8 100644
--- a/net/ipv4/tcp_ao.c
+++ b/net/ipv4/tcp_ao.c
@@ -1422,7 +1422,7 @@ static bool tcp_ao_mkt_overlap_v6(struct tcp_ao *cmd,
 #define TCP_AO_CMDF_ADDMOD_VALID					\
 	(TCP_AO_CMDF_CURR | TCP_AO_CMDF_NEXT | TCP_AO_CMDF_ACCEPT_ICMP)
 #define TCP_AO_CMDF_DEL_VALID						\
-	(TCP_AO_CMDF_CURR | TCP_AO_CMDF_NEXT)
+	(TCP_AO_CMDF_CURR | TCP_AO_CMDF_NEXT | TCP_AO_CMDF_DEL_ASYNC)
 #define TCP_AO_GETF_VALID						\
 	(TCP_AO_GET_ALL | TCP_AO_GET_CURR | TCP_AO_GET_NEXT)
 
@@ -1547,11 +1547,26 @@ static int tcp_ao_delete_key(struct sock *sk, struct tcp_ao_key *key,
 
 	hlist_del_rcu(&key->node);
 
+	/* Support for async delete on listening sockets: as they don't
+	 * need current_key/rnext_key maintaining, we don't need to check
+	 * them and we can just free all resources in RCU fashion.
+	 */
+	if (cmd->tcpa_flags & TCP_AO_CMDF_DEL_ASYNC) {
+		if (sk->sk_state != TCP_LISTEN)
+			return -EINVAL;
+		atomic_sub(tcp_ao_sizeof_key(key), &sk->sk_omem_alloc);
+		call_rcu(&key->rcu, tcp_ao_key_free_rcu);
+		return 0;
+	}
+
 	/* At this moment another CPU could have looked this key up
 	 * while it was unlinked from the list. Wait for RCU grace period,
 	 * after which the key is off-list and can't be looked up again;
 	 * the rx path [just before RCU came] might have used it and set it
 	 * as current_key (very unlikely).
+	 * Free the key with next RCU grace period (in case it was
+	 * current_key before tcp_ao_current_rnext() might have
+	 * changed it in forced-delete).
 	 */
 	synchronize_rcu();
 	err = tcp_ao_current_rnext(sk, cmd->tcpa_flags,
-- 
2.37.2