1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3  * Copyright (C) 2015-2019 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved.
4  */
5 
6 #ifndef _WG_QUEUEING_H
7 #define _WG_QUEUEING_H
8 
9 #include "peer.h"
10 #include <linux/types.h>
11 #include <linux/skbuff.h>
12 #include <linux/ip.h>
13 #include <linux/ipv6.h>
14 #include <net/ip_tunnels.h>
15 
16 struct wg_device;
17 struct wg_peer;
18 struct multicore_worker;
19 struct crypt_queue;
20 struct prev_queue;
21 struct sk_buff;
22 
23 /* queueing.c APIs: */
24 int wg_packet_queue_init(struct crypt_queue *queue, work_func_t function,
25 			 unsigned int len);
26 void wg_packet_queue_free(struct crypt_queue *queue, bool purge);
27 struct multicore_worker __percpu *
28 wg_packet_percpu_multicore_worker_alloc(work_func_t function, void *ptr);
29 
30 /* receive.c APIs: */
31 void wg_packet_receive(struct wg_device *wg, struct sk_buff *skb);
32 void wg_packet_handshake_receive_worker(struct work_struct *work);
33 /* NAPI poll function: */
34 int wg_packet_rx_poll(struct napi_struct *napi, int budget);
35 /* Workqueue worker: */
36 void wg_packet_decrypt_worker(struct work_struct *work);
37 
38 /* send.c APIs: */
39 void wg_packet_send_queued_handshake_initiation(struct wg_peer *peer,
40 						bool is_retry);
41 void wg_packet_send_handshake_response(struct wg_peer *peer);
42 void wg_packet_send_handshake_cookie(struct wg_device *wg,
43 				     struct sk_buff *initiating_skb,
44 				     __le32 sender_index);
45 void wg_packet_send_keepalive(struct wg_peer *peer);
46 void wg_packet_purge_staged_packets(struct wg_peer *peer);
47 void wg_packet_send_staged_packets(struct wg_peer *peer);
48 /* Workqueue workers: */
49 void wg_packet_handshake_send_worker(struct work_struct *work);
50 void wg_packet_tx_worker(struct work_struct *work);
51 void wg_packet_encrypt_worker(struct work_struct *work);
52 
53 enum packet_state {
54 	PACKET_STATE_UNCRYPTED,
55 	PACKET_STATE_CRYPTED,
56 	PACKET_STATE_DEAD
57 };
58 
59 struct packet_cb {
60 	u64 nonce;
61 	struct noise_keypair *keypair;
62 	atomic_t state;
63 	u32 mtu;
64 	u8 ds;
65 };
66 
67 #define PACKET_CB(skb) ((struct packet_cb *)((skb)->cb))
68 #define PACKET_PEER(skb) (PACKET_CB(skb)->keypair->entry.peer)
69 
wg_check_packet_protocol(struct sk_buff * skb)70 static inline bool wg_check_packet_protocol(struct sk_buff *skb)
71 {
72 	__be16 real_protocol = ip_tunnel_parse_protocol(skb);
73 	return real_protocol && skb->protocol == real_protocol;
74 }
75 
wg_reset_packet(struct sk_buff * skb,bool encapsulating)76 static inline void wg_reset_packet(struct sk_buff *skb, bool encapsulating)
77 {
78 	u8 l4_hash = skb->l4_hash;
79 	u8 sw_hash = skb->sw_hash;
80 	u32 hash = skb->hash;
81 	skb_scrub_packet(skb, true);
82 	memset(&skb->headers_start, 0,
83 	       offsetof(struct sk_buff, headers_end) -
84 		       offsetof(struct sk_buff, headers_start));
85 	if (encapsulating) {
86 		skb->l4_hash = l4_hash;
87 		skb->sw_hash = sw_hash;
88 		skb->hash = hash;
89 	}
90 	skb->queue_mapping = 0;
91 	skb->nohdr = 0;
92 	skb->peeked = 0;
93 	skb->mac_len = 0;
94 	skb->dev = NULL;
95 #ifdef CONFIG_NET_SCHED
96 	skb->tc_index = 0;
97 #endif
98 	skb_reset_redirect(skb);
99 	skb->hdr_len = skb_headroom(skb);
100 	skb_reset_mac_header(skb);
101 	skb_reset_network_header(skb);
102 	skb_reset_transport_header(skb);
103 	skb_probe_transport_header(skb);
104 	skb_reset_inner_headers(skb);
105 }
106 
wg_cpumask_choose_online(int * stored_cpu,unsigned int id)107 static inline int wg_cpumask_choose_online(int *stored_cpu, unsigned int id)
108 {
109 	unsigned int cpu = *stored_cpu, cpu_index, i;
110 
111 	if (unlikely(cpu == nr_cpumask_bits ||
112 		     !cpumask_test_cpu(cpu, cpu_online_mask))) {
113 		cpu_index = id % cpumask_weight(cpu_online_mask);
114 		cpu = cpumask_first(cpu_online_mask);
115 		for (i = 0; i < cpu_index; ++i)
116 			cpu = cpumask_next(cpu, cpu_online_mask);
117 		*stored_cpu = cpu;
118 	}
119 	return cpu;
120 }
121 
122 /* This function is racy, in the sense that next is unlocked, so it could return
123  * the same CPU twice. A race-free version of this would be to instead store an
124  * atomic sequence number, do an increment-and-return, and then iterate through
125  * every possible CPU until we get to that index -- choose_cpu. However that's
126  * a bit slower, and it doesn't seem like this potential race actually
127  * introduces any performance loss, so we live with it.
128  */
wg_cpumask_next_online(int * next)129 static inline int wg_cpumask_next_online(int *next)
130 {
131 	int cpu = *next;
132 
133 	while (unlikely(!cpumask_test_cpu(cpu, cpu_online_mask)))
134 		cpu = cpumask_next(cpu, cpu_online_mask) % nr_cpumask_bits;
135 	*next = cpumask_next(cpu, cpu_online_mask) % nr_cpumask_bits;
136 	return cpu;
137 }
138 
139 void wg_prev_queue_init(struct prev_queue *queue);
140 
141 /* Multi producer */
142 bool wg_prev_queue_enqueue(struct prev_queue *queue, struct sk_buff *skb);
143 
144 /* Single consumer */
145 struct sk_buff *wg_prev_queue_dequeue(struct prev_queue *queue);
146 
147 /* Single consumer */
wg_prev_queue_peek(struct prev_queue * queue)148 static inline struct sk_buff *wg_prev_queue_peek(struct prev_queue *queue)
149 {
150 	if (queue->peeked)
151 		return queue->peeked;
152 	queue->peeked = wg_prev_queue_dequeue(queue);
153 	return queue->peeked;
154 }
155 
156 /* Single consumer */
wg_prev_queue_drop_peeked(struct prev_queue * queue)157 static inline void wg_prev_queue_drop_peeked(struct prev_queue *queue)
158 {
159 	queue->peeked = NULL;
160 }
161 
wg_queue_enqueue_per_device_and_peer(struct crypt_queue * device_queue,struct prev_queue * peer_queue,struct sk_buff * skb,struct workqueue_struct * wq,int * next_cpu)162 static inline int wg_queue_enqueue_per_device_and_peer(
163 	struct crypt_queue *device_queue, struct prev_queue *peer_queue,
164 	struct sk_buff *skb, struct workqueue_struct *wq, int *next_cpu)
165 {
166 	int cpu;
167 
168 	atomic_set_release(&PACKET_CB(skb)->state, PACKET_STATE_UNCRYPTED);
169 	/* We first queue this up for the peer ingestion, but the consumer
170 	 * will wait for the state to change to CRYPTED or DEAD before.
171 	 */
172 	if (unlikely(!wg_prev_queue_enqueue(peer_queue, skb)))
173 		return -ENOSPC;
174 
175 	/* Then we queue it up in the device queue, which consumes the
176 	 * packet as soon as it can.
177 	 */
178 	cpu = wg_cpumask_next_online(next_cpu);
179 	if (unlikely(ptr_ring_produce_bh(&device_queue->ring, skb)))
180 		return -EPIPE;
181 	queue_work_on(cpu, wq, &per_cpu_ptr(device_queue->worker, cpu)->work);
182 	return 0;
183 }
184 
wg_queue_enqueue_per_peer_tx(struct sk_buff * skb,enum packet_state state)185 static inline void wg_queue_enqueue_per_peer_tx(struct sk_buff *skb, enum packet_state state)
186 {
187 	/* We take a reference, because as soon as we call atomic_set, the
188 	 * peer can be freed from below us.
189 	 */
190 	struct wg_peer *peer = wg_peer_get(PACKET_PEER(skb));
191 
192 	atomic_set_release(&PACKET_CB(skb)->state, state);
193 	queue_work_on(wg_cpumask_choose_online(&peer->serial_work_cpu, peer->internal_id),
194 		      peer->device->packet_crypt_wq, &peer->transmit_packet_work);
195 	wg_peer_put(peer);
196 }
197 
wg_queue_enqueue_per_peer_rx(struct sk_buff * skb,enum packet_state state)198 static inline void wg_queue_enqueue_per_peer_rx(struct sk_buff *skb, enum packet_state state)
199 {
200 	/* We take a reference, because as soon as we call atomic_set, the
201 	 * peer can be freed from below us.
202 	 */
203 	struct wg_peer *peer = wg_peer_get(PACKET_PEER(skb));
204 
205 	atomic_set_release(&PACKET_CB(skb)->state, state);
206 	napi_schedule(&peer->napi);
207 	wg_peer_put(peer);
208 }
209 
210 #ifdef DEBUG
211 bool wg_packet_counter_selftest(void);
212 #endif
213 
214 #endif /* _WG_QUEUEING_H */
215