ovpn: reduce reference-count contention

PATCH New Ralf Lici 1 participant 3 messages Sep 18

Subsystems (2)

Network drivers Openvpn data channel offload

lore

Share & tools

lore

Ralf Lici

author reviewer

2026-09-18 16:04 UTC #0 lore raw

v2 cover ovpn: reduce reference-count contention

Hi,

Currently ovpn holds peer and key-slot references while packets are being processed and until asynchronous crypto completes. With traffic for one peer spread across CPUs, the shared kref cachelines therefore become contended even though both objects are normally long-lived.

This series converts both references to percpu_ref while preserving their existing RCU-delayed destruction. The individual patch messages describe the memory cost and cache-to-cache measurements.

The test used one peer, one AES-128-GCM key and 32 TCP streams on a 32-logical-CPU Ryzen 9 9950X with a ConnectX-5 hairpin. The encrypted RX flow used one source port and one receive queue, so only TX exercised CPU fan-out. perf c2c showed that each shared reference cacheline disappeared after its respective conversion.

Thanks,

Ralf Lici
Mandelbit Srl

---
Changes since v1 https://lore.kernel.org/openvpn-devel/[email protected]/
- In 1/2, distinguish reference acquisitions which extend existing peer ownership from RCU lookups which must reject dying peers. (Sashiko)
- In 2/2, avoid a nested RCU callback during key-slot release which
could outlive the ovpn workqueue during module unload. (Sashiko)

Ralf Lici (2):
ovpn: use percpu references for peers
ovpn: use percpu references for key slots

 drivers/net/ovpn/crypto.c      | 16 +++++++--------
 drivers/net/ovpn/crypto.h      | 27 ++++++++++++++++++++-----
 drivers/net/ovpn/crypto_aead.c | 17 ++++++++++++----
 drivers/net/ovpn/io.c          |  6 +-----
 drivers/net/ovpn/netlink.c     |  6 +++---
 drivers/net/ovpn/peer.c        | 37 +++++++++++++++++++++-------------
 drivers/net/ovpn/peer.h        | 37 ++++++++++++++++++++++++++++------
 drivers/net/ovpn/tcp.c         | 23 ++++++++++-----------
 8 files changed, 112 insertions(+), 57 deletions(-)

base-commit: b8e9e7d82e7eefd5d2d528469d94ec20e96b38c3

Signature · 8 lines

2.55.0



_______________________________________________
Openvpn-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/openvpn-devel

Ralf Lici

author reviewer

2026-09-18 16:04 UTC #1 lore raw

PATCH v2 1/2 ovpn: use percpu references for peers

Each packet holds its peer across packet processing and asynchronous crypto completion. When traffic for one peer is spread across several CPUs, the shared kref cache line therefore moves between those CPUs for every packet.

Replace the peer kref with percpu_ref and kill the initial reference exactly once after the peer has been removed from its lookup structures and detached from its socket. Keep the existing RCU-delayed destruction and release the percpu_ref storage in the final RCU callback.

Use an unconditional percpu_ref_get where the caller already owns a peer reference or otherwise excludes teardown. Such callers must remain able to extend their existing ownership while the killed reference drains. RCU-protected lookup paths instead use percpu_ref_tryget_live_rcu, so they stop admitting new users as soon as teardown starts. This deliberately makes stale TCP socket lookups fail earlier than kref_get_unless_zero, which continued succeeding until the count reached zero.

percpu_ref_init can fail, so propagate allocation failure from peer creation.

This trades a per-CPU counter allocation for a cheaper live data path on these long-lived, heavily shared objects. Specifically, in terms of memory, this costs: 8 bytes per possible CPU for the counters, a 56-byte control object, and 8 additional bytes in struct ovpn_peer.

In a 32-stream test using one peer and one key, perf c2c placed the peer kref cacheline second among system-wide shared cachelines, with 37 sampled HITM loads attributed to the locked reference update. After this change that shared reference line was no longer sampled, while the untouched key-slot kref addressed by the next commit remained visible. RX used a single receive queue, so only TX exercised CPU fan-out in this test.

Signed-off-by Ralf Lici <[email protected]>

---
Changes since v1 https://lore.kernel.org/openvpn-devel/da742f32b00e7a6872062ce36a3dc4add665f189.1789658051.git.ralf@mandelbit.com/
- Use unconditional percpu_ref_get where the caller already owns the
peer or otherwise excludes teardown and percpu_ref_tryget_live_rcu for
TCP socket lookups under RCU. (Sashiko)
- Make the percpu-ref release callback private to peer.c.

 drivers/net/ovpn/io.c      |  6 +-----
 drivers/net/ovpn/netlink.c |  6 +++---
 drivers/net/ovpn/peer.c    | 37 +++++++++++++++++++++++--------------
 drivers/net/ovpn/peer.h    | 37 +++++++++++++++++++++++++++++++------
 drivers/net/ovpn/tcp.c     | 23 +++++++++++------------
 5 files changed, 69 insertions(+), 40 deletions(-)

drivers/net/ovpn/io.c +1 -5

index 9526f8096da6..15072ce960fa 100644
@@ -321,11 +321,7 @@ static bool ovpn_encrypt_one(struct ovpn_peer *peer, struct sk_buff *skb)
 	/* take a reference to the peer because the crypto code may run async.
 	 * ovpn_encrypt_post() will release it upon completion
 	 */
-	if (unlikely(!ovpn_peer_hold(peer))) {
-		DEBUG_NET_WARN_ON_ONCE(1);
-		ovpn_crypto_key_slot_put(ks);
-		return false;
-	}
+	ovpn_peer_hold(peer);
 
 	memset(ovpn_skb_cb(skb), 0, sizeof(struct ovpn_cb));
 	ovpn_encrypt_post(skb, ovpn_aead_encrypt(peer, ks, skb));

drivers/net/ovpn/netlink.c +3 -3

index 4dad85294198..8cdf46b61142 100644
@@ -463,11 +463,11 @@ int ovpn_nl_peer_new_doit(struct sk_buff *skb, struct genl_info *info)
 	ovpn_socket_release(peer);
 peer_release:
 	/* For UDP, the peer is unreachable until added to the hashtables, so
-	 * dropping the initial reference is enough. For TCP, the peer may be
+	 * killing the initial reference is enough. For TCP, the peer may be
 	 * concurrently reachable via sk_user_data->peer until
-	 * ovpn_socket_release() detaches; rely on the refcount.
+	 * ovpn_socket_release() detaches; rely on the percpu reference.
 	 */
-	ovpn_peer_put(peer);
+	ovpn_peer_kill(peer);
 
 	return ret;
 }

drivers/net/ovpn/peer.c +23 ~~-14~~

index c95656ca7c35..970ca7232ae4 100644
@@ -22,6 +22,8 @@
 #include "peer.h"
 #include "socket.h"
 
+static void ovpn_peer_release_ref(struct percpu_ref *ref);
+
 static void unlock_ovpn(struct ovpn_priv *ovpn,
 			 struct llist_head *release_list)
 	__releases(&ovpn->lock)

@@ -33,7 +35,7 @@ static void unlock_ovpn(struct ovpn_priv *ovpn,
 	llist_for_each_entry_safe(peer, next, release_list->first,
 				  release_entry) {
 		ovpn_socket_release(peer);
-		ovpn_peer_put(peer);
+		ovpn_peer_kill(peer);
 	}
 }

@@ -113,7 +115,6 @@ struct ovpn_peer *ovpn_peer_new(struct ovpn_priv *ovpn, u32 id)
 	RCU_INIT_POINTER(peer->bind, NULL);
 	ovpn_crypto_state_init(&peer->crypto);
 	spin_lock_init(&peer->lock);
-	kref_init(&peer->refcount);
 	ovpn_peer_stats_init(&peer->vpn_stats);
 	ovpn_peer_stats_init(&peer->link_stats);
 	INIT_WORK(&peer->keepalive_work, ovpn_peer_keepalive_send);

@@ -127,6 +128,14 @@ struct ovpn_peer *ovpn_peer_new(struct ovpn_priv *ovpn, u32 id)
 		return ERR_PTR(ret);
 	}
 
+	ret = percpu_ref_init(&peer->refcount, ovpn_peer_release_ref, 0,
+			      GFP_KERNEL);
+	if (ret < 0) {
+		dst_cache_destroy(&peer->dst_cache);
+		kfree(peer);
+		return ERR_PTR(ret);
+	}
+
 	netdev_hold(ovpn->dev, &peer->dev_tracker, GFP_KERNEL);
 
 	return peer;

@@ -348,6 +357,7 @@ static void ovpn_peer_release_rcu(struct rcu_head *head)
 	 * perform it in the RCU callback, when all contexts are done
 	 */
 	dst_cache_destroy(&peer->dst_cache);
+	percpu_ref_exit(&peer->refcount);
 	kfree(peer);
 }

@@ -366,12 +376,12 @@ static void ovpn_peer_release(struct ovpn_peer *peer)
 }
 
 /**
- * ovpn_peer_release_kref - callback for kref_put
- * @kref: the kref object belonging to the peer
+ * ovpn_peer_release_ref - callback for the peer percpu reference
+ * @ref: the percpu_ref object belonging to the peer
  */
-void ovpn_peer_release_kref(struct kref *kref)
+static void ovpn_peer_release_ref(struct percpu_ref *ref)
 {
-	struct ovpn_peer *peer = container_of(kref, struct ovpn_peer, refcount);
+	struct ovpn_peer *peer = container_of(ref, struct ovpn_peer, refcount);
 
 	ovpn_peer_release(peer);
 }

@@ -569,7 +579,7 @@ ovpn_peer_get_by_transp_addr_p2p(struct ovpn_priv *ovpn,
 	rcu_read_lock();
 	tmp = rcu_dereference(ovpn->peer);
 	if (likely(tmp && ovpn_peer_transp_match(tmp, ss) &&
-		   ovpn_peer_hold(tmp)))
+		   ovpn_peer_hold_rcu(tmp)))
 		peer = tmp;
 	rcu_read_unlock();

@@ -610,7 +620,7 @@ struct ovpn_peer *ovpn_peer_get_by_transp_addr(struct ovpn_priv *ovpn,
 		if (!ovpn_peer_transp_match(tmp, &ss))
 			continue;
 
-		if (!ovpn_peer_hold(tmp))
+		if (!ovpn_peer_hold_rcu(tmp))
 			continue;
 
 		peer = tmp;

@@ -641,7 +651,7 @@ static struct ovpn_peer *ovpn_peer_get_by_id_p2p(struct ovpn_priv *ovpn,
 
 	rcu_read_lock();
 	tmp = rcu_dereference(ovpn->peer);
-	if (likely(tmp && tmp->id == peer_id && ovpn_peer_hold(tmp)))
+	if (likely(tmp && tmp->id == peer_id && ovpn_peer_hold_rcu(tmp)))
 		peer = tmp;
 	rcu_read_unlock();

@@ -671,7 +681,7 @@ struct ovpn_peer *ovpn_peer_get_by_id(struct ovpn_priv *ovpn, u32 peer_id)
 		if (tmp->id != peer_id)
 			continue;
 
-		if (!ovpn_peer_hold(tmp))
+		if (!ovpn_peer_hold_rcu(tmp))
 			continue;
 
 		peer = tmp;

@@ -745,7 +755,7 @@ struct ovpn_peer *ovpn_peer_get_by_dst(struct ovpn_priv *ovpn,
 	if (ovpn->mode == OVPN_MODE_P2P) {
 		rcu_read_lock();
 		peer = rcu_dereference(ovpn->peer);
-		if (unlikely(peer && !ovpn_peer_hold(peer)))
+		if (unlikely(peer && !ovpn_peer_hold_rcu(peer)))
 			peer = NULL;
 		rcu_read_unlock();
 		return peer;

@@ -763,7 +773,7 @@ struct ovpn_peer *ovpn_peer_get_by_dst(struct ovpn_priv *ovpn,
 		break;
 	}
 
-	if (unlikely(peer && !ovpn_peer_hold(peer)))
+	if (unlikely(peer && !ovpn_peer_hold_rcu(peer)))
 		peer = NULL;
 	rcu_read_unlock();

@@ -1369,8 +1379,7 @@ static time64_t ovpn_peer_keepalive_work_single(struct ovpn_peer *peer,
 		netdev_dbg(peer->ovpn->dev,
 			   "sending keepalive to peer %u\n",
 			   peer->id);
-		if (WARN_ON(!ovpn_peer_hold(peer)))
-			return 0;
+		ovpn_peer_hold(peer);
 		if (!queue_work(ovpn_wq, &peer->keepalive_work))
 			ovpn_peer_put(peer);
 	}

drivers/net/ovpn/peer.h +31 -6

index dfa5c0037e02..489bf2646684 100644
@@ -10,6 +10,7 @@
 #ifndef _NET_OVPN_OVPNPEER_H_
 #define _NET_OVPN_OVPNPEER_H_
 
+#include <linux/percpu-refcount.h>
 #include <net/dst_cache.h>
 #include <net/strparser.h>

@@ -110,7 +111,7 @@ struct ovpn_peer {
 	struct ovpn_peer_stats link_stats;
 	enum ovpn_del_peer_reason delete_reason;
 	spinlock_t lock; /* protects bind  and keepalive* */
-	struct kref refcount;
+	struct percpu_ref refcount;
 	struct rcu_head rcu;
 	struct llist_node release_entry;
 	struct work_struct keepalive_work;

@@ -120,14 +121,26 @@ struct ovpn_peer {
  * ovpn_peer_hold - increase reference counter
  * @peer: the peer whose counter should be increased
  *
- * Return: true if the counter was increased or false if it was zero already
+ * The caller must already own a peer reference or otherwise guarantee that
+ * teardown cannot complete while the new reference is acquired.
  */
-static inline bool ovpn_peer_hold(struct ovpn_peer *peer)
+static inline void ovpn_peer_hold(struct ovpn_peer *peer)
 {
-	return kref_get_unless_zero(&peer->refcount);
+	percpu_ref_get(&peer->refcount);
 }
 
-void ovpn_peer_release_kref(struct kref *kref);
+/**
+ * ovpn_peer_hold_rcu - acquire a live peer reference under RCU
+ * @peer: peer to acquire
+ *
+ * The caller must hold rcu_read_lock.
+ *
+ * Return: true if the reference was acquired, false if teardown has started
+ */
+static inline bool ovpn_peer_hold_rcu(struct ovpn_peer *peer)
+{
+	return percpu_ref_tryget_live_rcu(&peer->refcount);
+}
 
 /**
  * ovpn_peer_put - decrease reference counter

@@ -135,7 +148,19 @@ void ovpn_peer_release_kref(struct kref *kref);
  */
 static inline void ovpn_peer_put(struct ovpn_peer *peer)
 {
-	kref_put(&peer->refcount, ovpn_peer_release_kref);
+	percpu_ref_put(&peer->refcount);
+}
+
+/**
+ * ovpn_peer_kill - drop the peer's initial reference
+ * @peer: peer which has been unpublished and is being destroyed
+ *
+ * This must be called exactly once, after the peer is no longer reachable
+ * through its owning ovpn instance.
+ */
+static inline void ovpn_peer_kill(struct ovpn_peer *peer)
+{
+	percpu_ref_kill(&peer->refcount);
 }
 
 struct ovpn_peer *ovpn_peer_new(struct ovpn_priv *ovpn, u32 id);

drivers/net/ovpn/tcp.c +11 ~~-12~~

index 8fe8a8e750a4..29981c0279e2 100644
@@ -140,21 +140,18 @@ static void ovpn_tcp_rcv(struct strparser *strp, struct sk_buff *skb)
 	/* hold reference to peer as required by ovpn_recv().
 	 *
 	 * NOTE: in this context we should already be holding a reference to
-	 * this peer, therefore ovpn_peer_hold() is not expected to fail
+	 * this peer
 	 */
-	if (WARN_ON(!ovpn_peer_hold(peer)))
-		goto err_nopeer;
+	ovpn_peer_hold(peer);
 
 	ovpn_recv(peer, skb);
 	return;
 err:
-	/* take reference for deferred peer deletion. should never fail */
-	if (WARN_ON(!ovpn_peer_hold(peer)))
-		goto err_nopeer;
+	/* take reference for deferred peer deletion */
+	ovpn_peer_hold(peer);
 	if (!queue_work(ovpn_wq, &peer->tcp.defer_del_work))
 		ovpn_peer_put(peer);
 	ovpn_dev_dstats_rx_dropped(peer->ovpn->dev);
-err_nopeer:
 	kfree_skb(skb);
 }

@@ -168,7 +165,8 @@ static int ovpn_tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
 
 	rcu_read_lock();
 	sock = rcu_dereference_sk_user_data(sk);
-	if (unlikely(!sock || !sock->peer || !ovpn_peer_hold(sock->peer))) {
+	if (unlikely(!sock || !sock->peer ||
+		     !ovpn_peer_hold_rcu(sock->peer))) {
 		rcu_read_unlock();
 		return -EBADF;
 	}

@@ -385,7 +383,7 @@ static void ovpn_tcp_release(struct sock *sk)
 	/* during initialization this function is called before
 	 * assigning sock->peer
 	 */
-	if (unlikely(!peer || !ovpn_peer_hold(peer))) {
+	if (unlikely(!peer || !ovpn_peer_hold_rcu(peer))) {
 		rcu_read_unlock();
 		return;
 	}

@@ -411,7 +409,8 @@ static int ovpn_tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 	lock_sock(sk);
 	rcu_read_lock();
 	sock = rcu_dereference_sk_user_data(sk);
-	if (unlikely(!sock || !sock->peer || !ovpn_peer_hold(sock->peer))) {
+	if (unlikely(!sock || !sock->peer ||
+		     !ovpn_peer_hold_rcu(sock->peer))) {
 		rcu_read_unlock();
 		release_sock(sk);
 		return -EIO;

@@ -588,7 +587,7 @@ static void ovpn_tcp_close(struct sock *sk, long timeout)
 	}
 
 	peer = sock->peer;
-	if (!peer || !ovpn_peer_hold(peer)) {
+	if (!peer || !ovpn_peer_hold_rcu(peer)) {
 		rcu_read_unlock();
 		return;
 	}

@@ -619,7 +618,7 @@ static __poll_t ovpn_tcp_poll(struct file *file, struct socket *sock,
 		return 0;
 	}
 
-	if (ovpn_peer_hold(ovpn_sock->peer)) {
+	if (ovpn_peer_hold_rcu(ovpn_sock->peer)) {
 		peer = ovpn_sock->peer;
 		queue = &peer->tcp.user_queue;
 	}

Signature · 8 lines

2.55.0



_______________________________________________
Openvpn-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/openvpn-devel

Ralf Lici

author reviewer

2026-09-18 16:04 UTC #2 lore raw

PATCH v2 2/2 ovpn: use percpu references for key slots

In reply to Ralf Lici

Packet processing holds a key slot until asynchronous crypto completion. Parallel traffic using one key consequently updates the same kref cache line for every packet, even though key slots are normally long-lived.

Replace the key-slot kref with percpu_ref. Kill the initial reference after atomically unpublishing a slot, while ordinary packet completions only put their live references.

Key destruction needs process context to release the crypto transforms and an RCU grace period for lockless configuration readers which do not hold a key reference. With percpu_ref, the release callback may run from the internal RCU callback which switches the reference to atomic mode. Queueing rcu_work from there would create a nested callback which the module cleanup rcu_barrier is not guaranteed to wait for.

Record the RCU state after unpublishing the slot and queue ordinary work from the release callback instead. The worker conditionally waits for the recorded grace period before freeing the key. This provides explicit RCU synchronization without depending on percpu_ref implementation details or unconditionally starting another grace period.

Peer teardown kills its keys before dropping the final netdevice reference, so rtnl_link_unregister cannot return before the key percpu-ref callbacks have been registered. The existing rcu_barrier drains those callbacks and destroy_workqueue subsequently drains the ordinary key cleanup work.

Initialize the reference only after the rest of the key slot is ready, allowing the existing destruction path to handle allocation failure.

The conversion adds an 8-byte counter per possible CPU and a 56-byte control allocation for each installed key slot; with at most two long-lived slots per peer, this is a bounded memory tradeoff for removing the shared reference cacheline from the packet path.

After the peer conversion, perf c2c still identified the key-slot kref cacheline, with 15 sampled HITM loads attributed to its locked reference update. This change removed that line from the shared-cacheline profile as well. In a set of interleaved 32-flow iperf3 runs, the combined peer and key conversion moved throughput to 8.812 Gbit/s, against 7.734 Gbit/s for the baseline, a +13.95% improvement. Single-stream control found no reproducible regression.

Signed-off-by Ralf Lici <[email protected]>

---
Changes since v1 https://lore.kernel.org/openvpn-devel/9551c9c842e4ac926089badc81f445becca513e5.1789658051.git.ralf@mandelbit.com/
- Replace the nested rcu_work teardown with ordinary work so module
cleanup cannot destroy the workqueue before the work is queued.
(Sashiko)
- Record the RCU state after unpublishing a key and conditionally wait
for RCU-only readers in the final worker. (Sashiko)

 drivers/net/ovpn/crypto.c      | 16 ++++++++--------
 drivers/net/ovpn/crypto.h      | 27 ++++++++++++++++++++++-----
 drivers/net/ovpn/crypto_aead.c | 17 +++++++++++++----
 3 files changed, 43 insertions(+), 17 deletions(-)

drivers/net/ovpn/crypto.c +8 -8

index 7e545428900a..25c3a2b6da48 100644
@@ -18,12 +18,12 @@
 #include "crypto_aead.h"
 #include "crypto.h"
 
-void ovpn_crypto_key_slot_release(struct kref *kref)
+void ovpn_crypto_key_slot_release(struct percpu_ref *ref)
 {
 	struct ovpn_crypto_key_slot *ks;
 
-	ks = container_of(kref, struct ovpn_crypto_key_slot, refcount);
-	queue_rcu_work(ovpn_wq, &ks->free_work);
+	ks = container_of(ref, struct ovpn_crypto_key_slot, refcount);
+	queue_work(ovpn_wq, &ks->free_work);
 }
 
 /* can only be invoked when all peer references have been dropped (i.e. RCU

@@ -36,13 +36,13 @@ void ovpn_crypto_state_release(struct ovpn_crypto_state *cs)
 	ks = rcu_access_pointer(cs->slots[0]);
 	if (ks) {
 		RCU_INIT_POINTER(cs->slots[0], NULL);
-		ovpn_crypto_key_slot_put(ks);
+		ovpn_crypto_key_slot_kill(ks);
 	}
 
 	ks = rcu_access_pointer(cs->slots[1]);
 	if (ks) {
 		RCU_INIT_POINTER(cs->slots[1], NULL);
-		ovpn_crypto_key_slot_put(ks);
+		ovpn_crypto_key_slot_kill(ks);
 	}
 }

@@ -66,7 +66,7 @@ bool ovpn_crypto_kill_key(struct ovpn_crypto_state *cs, u8 key_id)
 	spin_unlock_bh(&cs->lock);
 
 	if (ks)
-		ovpn_crypto_key_slot_put(ks);
+		ovpn_crypto_key_slot_kill(ks);
 
 	/* let the caller know if a key was actually killed */
 	return ks;

@@ -104,7 +104,7 @@ int ovpn_crypto_state_reset(struct ovpn_crypto_state *cs,
 	spin_unlock_bh(&cs->lock);
 
 	if (old)
-		ovpn_crypto_key_slot_put(old);
+		ovpn_crypto_key_slot_kill(old);
 
 	return 0;
 }

@@ -141,7 +141,7 @@ void ovpn_crypto_key_slot_delete(struct ovpn_crypto_state *cs,
 	}
 
 	pr_debug("deleting key slot %u, key_id=%u\n", slot, ks->key_id);
-	ovpn_crypto_key_slot_put(ks);
+	ovpn_crypto_key_slot_kill(ks);
 }
 
 void ovpn_crypto_key_slots_swap(struct ovpn_crypto_state *cs)

drivers/net/ovpn/crypto.h +22 -5

index e3feb16d5498..a0209b1280de 100644
@@ -10,6 +10,8 @@
 #ifndef _NET_OVPN_OVPNCRYPTO_H_
 #define _NET_OVPN_OVPNCRYPTO_H_
 
+#include <linux/percpu-refcount.h>
+#include <linux/rcupdate.h>
 #include <linux/workqueue.h>
 
 #include "pktid.h"

@@ -47,8 +49,9 @@ struct ovpn_crypto_key_slot {
 
 	struct ovpn_pktid_recv pid_recv ____cacheline_aligned_in_smp;
 	struct ovpn_pktid_xmit pid_xmit ____cacheline_aligned_in_smp;
-	struct rcu_work free_work;
-	struct kref refcount;
+	struct work_struct free_work;
+	unsigned long rcu_state;
+	struct percpu_ref refcount;
 };
 
 struct ovpn_crypto_state {

@@ -61,7 +64,7 @@ struct ovpn_crypto_state {
 
 static inline bool ovpn_crypto_key_slot_hold(struct ovpn_crypto_key_slot *ks)
 {
-	return kref_get_unless_zero(&ks->refcount);
+	return percpu_ref_tryget_live_rcu(&ks->refcount);
 }
 
 static inline void ovpn_crypto_state_init(struct ovpn_crypto_state *cs)

@@ -121,11 +124,25 @@ ovpn_crypto_key_slot_primary(const struct ovpn_crypto_state *cs)
 	return ks;
 }
 
-void ovpn_crypto_key_slot_release(struct kref *kref);
+void ovpn_crypto_key_slot_release(struct percpu_ref *ref);
 
 static inline void ovpn_crypto_key_slot_put(struct ovpn_crypto_key_slot *ks)
 {
-	kref_put(&ks->refcount, ovpn_crypto_key_slot_release);
+	percpu_ref_put(&ks->refcount);
+}
+
+/**
+ * ovpn_crypto_key_slot_kill - stop new users and drop the initial reference
+ * @ks: key slot which has already been unpublished
+ *
+ * percpu_ref does not guarantee an RCU grace period before release. Record
+ * the RCU state after unpublishing the key so the final worker can synchronize
+ * with lockless readers without unconditionally starting another grace period.
+ */
+static inline void ovpn_crypto_key_slot_kill(struct ovpn_crypto_key_slot *ks)
+{
+	ks->rcu_state = get_state_synchronize_rcu();
+	percpu_ref_kill(&ks->refcount);
 }
 
 int ovpn_crypto_state_reset(struct ovpn_crypto_state *cs,

drivers/net/ovpn/crypto_aead.c +13 -4

index 74eaf6fac2f5..77ae958a971a 100644
@@ -391,8 +391,13 @@ static void ovpn_aead_crypto_key_slot_free_work(struct work_struct *work)
 {
 	struct ovpn_crypto_key_slot *ks;
 
-	ks = container_of(to_rcu_work(work), struct ovpn_crypto_key_slot,
-			  free_work);
+	ks = container_of(work, struct ovpn_crypto_key_slot, free_work);
+	/* Reaching this worker means every reference held by packet processing
+	 * and asynchronous crypto has been returned. Separately wait for any
+	 * RCU-only reader which observed the slot before it was unpublished.
+	 */
+	cond_synchronize_rcu(ks->rcu_state);
+	percpu_ref_exit(&ks->refcount);
 	ovpn_aead_crypto_key_slot_free(ks);
 	kfree(ks);
 }

@@ -427,8 +432,7 @@ ovpn_aead_crypto_key_slot_new(const struct ovpn_key_config *kc)
 
 	ks->encrypt = NULL;
 	ks->decrypt = NULL;
-	INIT_RCU_WORK(&ks->free_work, ovpn_aead_crypto_key_slot_free_work);
-	kref_init(&ks->refcount);
+	INIT_WORK(&ks->free_work, ovpn_aead_crypto_key_slot_free_work);
 	ks->key_id = kc->key_id;
 
 	ks->encrypt = ovpn_aead_init("encrypt", alg_name,

@@ -458,6 +462,11 @@ ovpn_aead_crypto_key_slot_new(const struct ovpn_key_config *kc)
 	ovpn_pktid_xmit_init(&ks->pid_xmit);
 	ovpn_pktid_recv_init(&ks->pid_recv);
 
+	ret = percpu_ref_init(&ks->refcount, ovpn_crypto_key_slot_release, 0,
+			      GFP_KERNEL);
+	if (ret < 0)
+		goto destroy_ks;
+
 	return ks;
 
 destroy_ks:

Signature · 8 lines

2.55.0



_______________________________________________
Openvpn-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/openvpn-devel