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 15 struct wg_device; 16 struct wg_peer; 17 struct multicore_worker; 18 struct crypt_queue; 19 struct sk_buff; 20 21 /* queueing.c APIs: */ 22 int wg_packet_queue_init(struct crypt_queue *queue, work_func_t function, 23 bool multicore, unsigned int len); 24 void wg_packet_queue_free(struct crypt_queue *queue, bool multicore); 25 struct multicore_worker __percpu * 26 wg_packet_percpu_multicore_worker_alloc(work_func_t function, void *ptr); 27 28 /* receive.c APIs: */ 29 void wg_packet_receive(struct wg_device *wg, struct sk_buff *skb); 30 void wg_packet_handshake_receive_worker(struct work_struct *work); 31 /* NAPI poll function: */ 32 int wg_packet_rx_poll(struct napi_struct *napi, int budget); 33 /* Workqueue worker: */ 34 void wg_packet_decrypt_worker(struct work_struct *work); 35 36 /* send.c APIs: */ 37 void wg_packet_send_queued_handshake_initiation(struct wg_peer *peer, 38 bool is_retry); 39 void wg_packet_send_handshake_response(struct wg_peer *peer); 40 void wg_packet_send_handshake_cookie(struct wg_device *wg, 41 struct sk_buff *initiating_skb, 42 __le32 sender_index); 43 void wg_packet_send_keepalive(struct wg_peer *peer); 44 void wg_packet_purge_staged_packets(struct wg_peer *peer); 45 void wg_packet_send_staged_packets(struct wg_peer *peer); 46 /* Workqueue workers: */ 47 void wg_packet_handshake_send_worker(struct work_struct *work); 48 void wg_packet_tx_worker(struct work_struct *work); 49 void wg_packet_encrypt_worker(struct work_struct *work); 50 51 enum packet_state { 52 PACKET_STATE_UNCRYPTED, 53 PACKET_STATE_CRYPTED, 54 PACKET_STATE_DEAD 55 }; 56 57 struct packet_cb { 58 u64 nonce; 59 struct noise_keypair *keypair; 60 atomic_t state; 61 u32 mtu; 62 u8 ds; 63 }; 64 65 #define PACKET_CB(skb) ((struct packet_cb *)((skb)->cb)) 66 #define PACKET_PEER(skb) (PACKET_CB(skb)->keypair->entry.peer) 67 68 /* Returns either the correct skb->protocol value, or 0 if invalid. */ 69 static inline __be16 wg_skb_examine_untrusted_ip_hdr(struct sk_buff *skb) 70 { 71 if (skb_network_header(skb) >= skb->head && 72 (skb_network_header(skb) + sizeof(struct iphdr)) <= 73 skb_tail_pointer(skb) && 74 ip_hdr(skb)->version == 4) 75 return htons(ETH_P_IP); 76 if (skb_network_header(skb) >= skb->head && 77 (skb_network_header(skb) + sizeof(struct ipv6hdr)) <= 78 skb_tail_pointer(skb) && 79 ipv6_hdr(skb)->version == 6) 80 return htons(ETH_P_IPV6); 81 return 0; 82 } 83 84 static inline void wg_reset_packet(struct sk_buff *skb) 85 { 86 skb_scrub_packet(skb, true); 87 memset(&skb->headers_start, 0, 88 offsetof(struct sk_buff, headers_end) - 89 offsetof(struct sk_buff, headers_start)); 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 skb_reset_tc(skb); 98 #endif 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 } 106 107 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 */ 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 static inline int wg_queue_enqueue_per_device_and_peer( 140 struct crypt_queue *device_queue, struct crypt_queue *peer_queue, 141 struct sk_buff *skb, struct workqueue_struct *wq, int *next_cpu) 142 { 143 int cpu; 144 145 atomic_set_release(&PACKET_CB(skb)->state, PACKET_STATE_UNCRYPTED); 146 /* We first queue this up for the peer ingestion, but the consumer 147 * will wait for the state to change to CRYPTED or DEAD before. 148 */ 149 if (unlikely(ptr_ring_produce_bh(&peer_queue->ring, skb))) 150 return -ENOSPC; 151 /* Then we queue it up in the device queue, which consumes the 152 * packet as soon as it can. 153 */ 154 cpu = wg_cpumask_next_online(next_cpu); 155 if (unlikely(ptr_ring_produce_bh(&device_queue->ring, skb))) 156 return -EPIPE; 157 queue_work_on(cpu, wq, &per_cpu_ptr(device_queue->worker, cpu)->work); 158 return 0; 159 } 160 161 static inline void wg_queue_enqueue_per_peer(struct crypt_queue *queue, 162 struct sk_buff *skb, 163 enum packet_state state) 164 { 165 /* We take a reference, because as soon as we call atomic_set, the 166 * peer can be freed from below us. 167 */ 168 struct wg_peer *peer = wg_peer_get(PACKET_PEER(skb)); 169 170 atomic_set_release(&PACKET_CB(skb)->state, state); 171 queue_work_on(wg_cpumask_choose_online(&peer->serial_work_cpu, 172 peer->internal_id), 173 peer->device->packet_crypt_wq, &queue->work); 174 wg_peer_put(peer); 175 } 176 177 static inline void wg_queue_enqueue_per_peer_napi(struct sk_buff *skb, 178 enum packet_state state) 179 { 180 /* We take a reference, because as soon as we call atomic_set, the 181 * peer can be freed from below us. 182 */ 183 struct wg_peer *peer = wg_peer_get(PACKET_PEER(skb)); 184 185 atomic_set_release(&PACKET_CB(skb)->state, state); 186 napi_schedule(&peer->napi); 187 wg_peer_put(peer); 188 } 189 190 #ifdef DEBUG 191 bool wg_packet_counter_selftest(void); 192 #endif 193 194 #endif /* _WG_QUEUEING_H */ 195