]> git.baikalelectronics.ru Git - kernel.git/commitdiff
wireguard: receive: use ring buffer for incoming handshakes
authorJason A. Donenfeld <Jason@zx2c4.com>
Mon, 29 Nov 2021 15:39:26 +0000 (10:39 -0500)
committerJakub Kicinski <kuba@kernel.org>
Tue, 30 Nov 2021 03:50:50 +0000 (19:50 -0800)
Apparently the spinlock on incoming_handshake's skb_queue is highly
contended, and a torrent of handshake or cookie packets can bring the
data plane to its knees, simply by virtue of enqueueing the handshake
packets to be processed asynchronously. So, we try switching this to a
ring buffer to hopefully have less lock contention. This alleviates the
problem somewhat, though it still isn't perfect, so future patches will
have to improve this further. However, it at least doesn't completely
diminish the data plane.

Reported-by: Streun Fabio <fstreun@student.ethz.ch>
Reported-by: Joel Wanner <joel.wanner@inf.ethz.ch>
Fixes: 0c73bbc77a76 ("net: WireGuard secure network tunnel")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
drivers/net/wireguard/device.c
drivers/net/wireguard/device.h
drivers/net/wireguard/queueing.c
drivers/net/wireguard/queueing.h
drivers/net/wireguard/receive.c

index 77e64ea6be67c36794522bed6a87801d90f1fa3d..a46067c38bf5def99accd4b885d3c3bd3095ad4b 100644 (file)
@@ -98,6 +98,7 @@ static int wg_stop(struct net_device *dev)
 {
        struct wg_device *wg = netdev_priv(dev);
        struct wg_peer *peer;
+       struct sk_buff *skb;
 
        mutex_lock(&wg->device_update_lock);
        list_for_each_entry(peer, &wg->peer_list, peer_list) {
@@ -108,7 +109,9 @@ static int wg_stop(struct net_device *dev)
                wg_noise_reset_last_sent_handshake(&peer->last_sent_handshake);
        }
        mutex_unlock(&wg->device_update_lock);
-       skb_queue_purge(&wg->incoming_handshakes);
+       while ((skb = ptr_ring_consume(&wg->handshake_queue.ring)) != NULL)
+               kfree_skb(skb);
+       atomic_set(&wg->handshake_queue_len, 0);
        wg_socket_reinit(wg, NULL, NULL);
        return 0;
 }
@@ -235,14 +238,13 @@ static void wg_destruct(struct net_device *dev)
        destroy_workqueue(wg->handshake_receive_wq);
        destroy_workqueue(wg->handshake_send_wq);
        destroy_workqueue(wg->packet_crypt_wq);
-       wg_packet_queue_free(&wg->decrypt_queue);
-       wg_packet_queue_free(&wg->encrypt_queue);
+       wg_packet_queue_free(&wg->handshake_queue, true);
+       wg_packet_queue_free(&wg->decrypt_queue, false);
+       wg_packet_queue_free(&wg->encrypt_queue, false);
        rcu_barrier(); /* Wait for all the peers to be actually freed. */
        wg_ratelimiter_uninit();
        memzero_explicit(&wg->static_identity, sizeof(wg->static_identity));
-       skb_queue_purge(&wg->incoming_handshakes);
        free_percpu(dev->tstats);
-       free_percpu(wg->incoming_handshakes_worker);
        kvfree(wg->index_hashtable);
        kvfree(wg->peer_hashtable);
        mutex_unlock(&wg->device_update_lock);
@@ -298,7 +300,6 @@ static int wg_newlink(struct net *src_net, struct net_device *dev,
        init_rwsem(&wg->static_identity.lock);
        mutex_init(&wg->socket_update_lock);
        mutex_init(&wg->device_update_lock);
-       skb_queue_head_init(&wg->incoming_handshakes);
        wg_allowedips_init(&wg->peer_allowedips);
        wg_cookie_checker_init(&wg->cookie_checker, wg);
        INIT_LIST_HEAD(&wg->peer_list);
@@ -316,16 +317,10 @@ static int wg_newlink(struct net *src_net, struct net_device *dev,
        if (!dev->tstats)
                goto err_free_index_hashtable;
 
-       wg->incoming_handshakes_worker =
-               wg_packet_percpu_multicore_worker_alloc(
-                               wg_packet_handshake_receive_worker, wg);
-       if (!wg->incoming_handshakes_worker)
-               goto err_free_tstats;
-
        wg->handshake_receive_wq = alloc_workqueue("wg-kex-%s",
                        WQ_CPU_INTENSIVE | WQ_FREEZABLE, 0, dev->name);
        if (!wg->handshake_receive_wq)
-               goto err_free_incoming_handshakes;
+               goto err_free_tstats;
 
        wg->handshake_send_wq = alloc_workqueue("wg-kex-%s",
                        WQ_UNBOUND | WQ_FREEZABLE, 0, dev->name);
@@ -347,10 +342,15 @@ static int wg_newlink(struct net *src_net, struct net_device *dev,
        if (ret < 0)
                goto err_free_encrypt_queue;
 
-       ret = wg_ratelimiter_init();
+       ret = wg_packet_queue_init(&wg->handshake_queue, wg_packet_handshake_receive_worker,
+                                  MAX_QUEUED_INCOMING_HANDSHAKES);
        if (ret < 0)
                goto err_free_decrypt_queue;
 
+       ret = wg_ratelimiter_init();
+       if (ret < 0)
+               goto err_free_handshake_queue;
+
        ret = register_netdevice(dev);
        if (ret < 0)
                goto err_uninit_ratelimiter;
@@ -367,18 +367,18 @@ static int wg_newlink(struct net *src_net, struct net_device *dev,
 
 err_uninit_ratelimiter:
        wg_ratelimiter_uninit();
+err_free_handshake_queue:
+       wg_packet_queue_free(&wg->handshake_queue, false);
 err_free_decrypt_queue:
-       wg_packet_queue_free(&wg->decrypt_queue);
+       wg_packet_queue_free(&wg->decrypt_queue, false);
 err_free_encrypt_queue:
-       wg_packet_queue_free(&wg->encrypt_queue);
+       wg_packet_queue_free(&wg->encrypt_queue, false);
 err_destroy_packet_crypt:
        destroy_workqueue(wg->packet_crypt_wq);
 err_destroy_handshake_send:
        destroy_workqueue(wg->handshake_send_wq);
 err_destroy_handshake_receive:
        destroy_workqueue(wg->handshake_receive_wq);
-err_free_incoming_handshakes:
-       free_percpu(wg->incoming_handshakes_worker);
 err_free_tstats:
        free_percpu(dev->tstats);
 err_free_index_hashtable:
index 854bc3d97150e1c1dab3befbe64966add4f65746..43c7cebbf50b08f2a1868f0017d0bee8aee700f8 100644 (file)
@@ -39,21 +39,18 @@ struct prev_queue {
 
 struct wg_device {
        struct net_device *dev;
-       struct crypt_queue encrypt_queue, decrypt_queue;
+       struct crypt_queue encrypt_queue, decrypt_queue, handshake_queue;
        struct sock __rcu *sock4, *sock6;
        struct net __rcu *creating_net;
        struct noise_static_identity static_identity;
-       struct workqueue_struct *handshake_receive_wq, *handshake_send_wq;
-       struct workqueue_struct *packet_crypt_wq;
-       struct sk_buff_head incoming_handshakes;
-       int incoming_handshake_cpu;
-       struct multicore_worker __percpu *incoming_handshakes_worker;
+       struct workqueue_struct *packet_crypt_wq,*handshake_receive_wq, *handshake_send_wq;
        struct cookie_checker cookie_checker;
        struct pubkey_hashtable *peer_hashtable;
        struct index_hashtable *index_hashtable;
        struct allowedips peer_allowedips;
        struct mutex device_update_lock, socket_update_lock;
        struct list_head device_list, peer_list;
+       atomic_t handshake_queue_len;
        unsigned int num_peers, device_update_gen;
        u32 fwmark;
        u16 incoming_port;
index 48e7b982a30736bc147712ce52957517e3e862af..1de413b19e3424a2ace2edcbcf0d0d49c4be6167 100644 (file)
@@ -38,11 +38,11 @@ int wg_packet_queue_init(struct crypt_queue *queue, work_func_t function,
        return 0;
 }
 
-void wg_packet_queue_free(struct crypt_queue *queue)
+void wg_packet_queue_free(struct crypt_queue *queue, bool purge)
 {
        free_percpu(queue->worker);
-       WARN_ON(!__ptr_ring_empty(&queue->ring));
-       ptr_ring_cleanup(&queue->ring, NULL);
+       WARN_ON(!purge && !__ptr_ring_empty(&queue->ring));
+       ptr_ring_cleanup(&queue->ring, purge ? (void(*)(void*))kfree_skb : NULL);
 }
 
 #define NEXT(skb) ((skb)->prev)
index 4ef2944a68bc906ebec5167d1e17e281ea67be61..e2388107f7fdc9c040841adfc164459e0c777d7d 100644 (file)
@@ -23,7 +23,7 @@ struct sk_buff;
 /* queueing.c APIs: */
 int wg_packet_queue_init(struct crypt_queue *queue, work_func_t function,
                         unsigned int len);
-void wg_packet_queue_free(struct crypt_queue *queue);
+void wg_packet_queue_free(struct crypt_queue *queue, bool purge);
 struct multicore_worker __percpu *
 wg_packet_percpu_multicore_worker_alloc(work_func_t function, void *ptr);
 
index 7dc84bcca26139991be00759c0228d3126df2671..f4e537e3e8ec6a10861f4216ad1808e2fc5091a7 100644 (file)
@@ -116,8 +116,8 @@ static void wg_receive_handshake_packet(struct wg_device *wg,
                return;
        }
 
-       under_load = skb_queue_len(&wg->incoming_handshakes) >=
-                    MAX_QUEUED_INCOMING_HANDSHAKES / 8;
+       under_load = atomic_read(&wg->handshake_queue_len) >=
+                       MAX_QUEUED_INCOMING_HANDSHAKES / 8;
        if (under_load) {
                last_under_load = ktime_get_coarse_boottime_ns();
        } else if (last_under_load) {
@@ -212,13 +212,14 @@ static void wg_receive_handshake_packet(struct wg_device *wg,
 
 void wg_packet_handshake_receive_worker(struct work_struct *work)
 {
-       struct wg_device *wg = container_of(work, struct multicore_worker,
-                                           work)->ptr;
+       struct crypt_queue *queue = container_of(work, struct multicore_worker, work)->ptr;
+       struct wg_device *wg = container_of(queue, struct wg_device, handshake_queue);
        struct sk_buff *skb;
 
-       while ((skb = skb_dequeue(&wg->incoming_handshakes)) != NULL) {
+       while ((skb = ptr_ring_consume_bh(&queue->ring)) != NULL) {
                wg_receive_handshake_packet(wg, skb);
                dev_kfree_skb(skb);
+               atomic_dec(&wg->handshake_queue_len);
                cond_resched();
        }
 }
@@ -554,21 +555,17 @@ void wg_packet_receive(struct wg_device *wg, struct sk_buff *skb)
        case cpu_to_le32(MESSAGE_HANDSHAKE_RESPONSE):
        case cpu_to_le32(MESSAGE_HANDSHAKE_COOKIE): {
                int cpu;
-
-               if (skb_queue_len(&wg->incoming_handshakes) >
-                           MAX_QUEUED_INCOMING_HANDSHAKES ||
-                   unlikely(!rng_is_initialized())) {
+               if (unlikely(!rng_is_initialized() ||
+                            ptr_ring_produce_bh(&wg->handshake_queue.ring, skb))) {
                        net_dbg_skb_ratelimited("%s: Dropping handshake packet from %pISpfsc\n",
                                                wg->dev->name, skb);
                        goto err;
                }
-               skb_queue_tail(&wg->incoming_handshakes, skb);
-               /* Queues up a call to packet_process_queued_handshake_
-                * packets(skb):
-                */
-               cpu = wg_cpumask_next_online(&wg->incoming_handshake_cpu);
+               atomic_inc(&wg->handshake_queue_len);
+               cpu = wg_cpumask_next_online(&wg->handshake_queue.last_cpu);
+               /* Queues up a call to packet_process_queued_handshake_packets(skb): */
                queue_work_on(cpu, wg->handshake_receive_wq,
-                       &per_cpu_ptr(wg->incoming_handshakes_worker, cpu)->work);
+                             &per_cpu_ptr(wg->handshake_queue.worker, cpu)->work);
                break;
        }
        case cpu_to_le32(MESSAGE_DATA):