| b.liu | e958203 | 2025-04-17 19:18:16 +0800 | [diff] [blame] | 1 | From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001 | 
|  | 2 | From: "Jason A. Donenfeld" <Jason@zx2c4.com> | 
|  | 3 | Date: Mon, 22 Feb 2021 17:25:48 +0100 | 
|  | 4 | Subject: [PATCH] wireguard: queueing: get rid of per-peer ring buffers | 
|  | 5 | MIME-Version: 1.0 | 
|  | 6 | Content-Type: text/plain; charset=UTF-8 | 
|  | 7 | Content-Transfer-Encoding: 8bit | 
|  | 8 |  | 
|  | 9 | commit 8b5553ace83cced775eefd0f3f18b5c6214ccf7a upstream. | 
|  | 10 |  | 
|  | 11 | Having two ring buffers per-peer means that every peer results in two | 
|  | 12 | massive ring allocations. On an 8-core x86_64 machine, this commit | 
|  | 13 | reduces the per-peer allocation from 18,688 bytes to 1,856 bytes, which | 
|  | 14 | is an 90% reduction. Ninety percent! With some single-machine | 
|  | 15 | deployments approaching 500,000 peers, we're talking about a reduction | 
|  | 16 | from 7 gigs of memory down to 700 megs of memory. | 
|  | 17 |  | 
|  | 18 | In order to get rid of these per-peer allocations, this commit switches | 
|  | 19 | to using a list-based queueing approach. Currently GSO fragments are | 
|  | 20 | chained together using the skb->next pointer (the skb_list_* singly | 
|  | 21 | linked list approach), so we form the per-peer queue around the unused | 
|  | 22 | skb->prev pointer (which sort of makes sense because the links are | 
|  | 23 | pointing backwards). Use of skb_queue_* is not possible here, because | 
|  | 24 | that is based on doubly linked lists and spinlocks. Multiple cores can | 
|  | 25 | write into the queue at any given time, because its writes occur in the | 
|  | 26 | start_xmit path or in the udp_recv path. But reads happen in a single | 
|  | 27 | workqueue item per-peer, amounting to a multi-producer, single-consumer | 
|  | 28 | paradigm. | 
|  | 29 |  | 
|  | 30 | The MPSC queue is implemented locklessly and never blocks. However, it | 
|  | 31 | is not linearizable (though it is serializable), with a very tight and | 
|  | 32 | unlikely race on writes, which, when hit (some tiny fraction of the | 
|  | 33 | 0.15% of partial adds on a fully loaded 16-core x86_64 system), causes | 
|  | 34 | the queue reader to terminate early. However, because every packet sent | 
|  | 35 | queues up the same workqueue item after it is fully added, the worker | 
|  | 36 | resumes again, and stopping early isn't actually a problem, since at | 
|  | 37 | that point the packet wouldn't have yet been added to the encryption | 
|  | 38 | queue. These properties allow us to avoid disabling interrupts or | 
|  | 39 | spinning. The design is based on Dmitry Vyukov's algorithm [1]. | 
|  | 40 |  | 
|  | 41 | Performance-wise, ordinarily list-based queues aren't preferable to | 
|  | 42 | ringbuffers, because of cache misses when following pointers around. | 
|  | 43 | However, we *already* have to follow the adjacent pointers when working | 
|  | 44 | through fragments, so there shouldn't actually be any change there. A | 
|  | 45 | potential downside is that dequeueing is a bit more complicated, but the | 
|  | 46 | ptr_ring structure used prior had a spinlock when dequeueing, so all and | 
|  | 47 | all the difference appears to be a wash. | 
|  | 48 |  | 
|  | 49 | Actually, from profiling, the biggest performance hit, by far, of this | 
|  | 50 | commit winds up being atomic_add_unless(count, 1, max) and atomic_ | 
|  | 51 | dec(count), which account for the majority of CPU time, according to | 
|  | 52 | perf. In that sense, the previous ring buffer was superior in that it | 
|  | 53 | could check if it was full by head==tail, which the list-based approach | 
|  | 54 | cannot do. | 
|  | 55 |  | 
|  | 56 | But all and all, this enables us to get massive memory savings, allowing | 
|  | 57 | WireGuard to scale for real world deployments, without taking much of a | 
|  | 58 | performance hit. | 
|  | 59 |  | 
|  | 60 | [1] http://www.1024cores.net/home/lock-free-algorithms/queues/intrusive-mpsc-node-based-queue | 
|  | 61 |  | 
|  | 62 | Reviewed-by: Dmitry Vyukov <dvyukov@google.com> | 
|  | 63 | Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com> | 
|  | 64 | Fixes: e7096c131e51 ("net: WireGuard secure network tunnel") | 
|  | 65 | Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com> | 
|  | 66 | Signed-off-by: Jakub Kicinski <kuba@kernel.org> | 
|  | 67 | Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com> | 
|  | 68 | --- | 
|  | 69 | drivers/net/wireguard/device.c   | 12 ++--- | 
|  | 70 | drivers/net/wireguard/device.h   | 15 +++--- | 
|  | 71 | drivers/net/wireguard/peer.c     | 28 ++++------- | 
|  | 72 | drivers/net/wireguard/peer.h     |  4 +- | 
|  | 73 | drivers/net/wireguard/queueing.c | 86 +++++++++++++++++++++++++------- | 
|  | 74 | drivers/net/wireguard/queueing.h | 45 ++++++++++++----- | 
|  | 75 | drivers/net/wireguard/receive.c  | 16 +++--- | 
|  | 76 | drivers/net/wireguard/send.c     | 31 ++++-------- | 
|  | 77 | 8 files changed, 144 insertions(+), 93 deletions(-) | 
|  | 78 |  | 
|  | 79 | --- a/drivers/net/wireguard/device.c | 
|  | 80 | +++ b/drivers/net/wireguard/device.c | 
|  | 81 | @@ -235,8 +235,8 @@ static void wg_destruct(struct net_devic | 
|  | 82 | destroy_workqueue(wg->handshake_receive_wq); | 
|  | 83 | destroy_workqueue(wg->handshake_send_wq); | 
|  | 84 | destroy_workqueue(wg->packet_crypt_wq); | 
|  | 85 | -	wg_packet_queue_free(&wg->decrypt_queue, true); | 
|  | 86 | -	wg_packet_queue_free(&wg->encrypt_queue, true); | 
|  | 87 | +	wg_packet_queue_free(&wg->decrypt_queue); | 
|  | 88 | +	wg_packet_queue_free(&wg->encrypt_queue); | 
|  | 89 | rcu_barrier(); /* Wait for all the peers to be actually freed. */ | 
|  | 90 | wg_ratelimiter_uninit(); | 
|  | 91 | memzero_explicit(&wg->static_identity, sizeof(wg->static_identity)); | 
|  | 92 | @@ -338,12 +338,12 @@ static int wg_newlink(struct net *src_ne | 
|  | 93 | goto err_destroy_handshake_send; | 
|  | 94 |  | 
|  | 95 | ret = wg_packet_queue_init(&wg->encrypt_queue, wg_packet_encrypt_worker, | 
|  | 96 | -				   true, MAX_QUEUED_PACKETS); | 
|  | 97 | +				   MAX_QUEUED_PACKETS); | 
|  | 98 | if (ret < 0) | 
|  | 99 | goto err_destroy_packet_crypt; | 
|  | 100 |  | 
|  | 101 | ret = wg_packet_queue_init(&wg->decrypt_queue, wg_packet_decrypt_worker, | 
|  | 102 | -				   true, MAX_QUEUED_PACKETS); | 
|  | 103 | +				   MAX_QUEUED_PACKETS); | 
|  | 104 | if (ret < 0) | 
|  | 105 | goto err_free_encrypt_queue; | 
|  | 106 |  | 
|  | 107 | @@ -368,9 +368,9 @@ static int wg_newlink(struct net *src_ne | 
|  | 108 | err_uninit_ratelimiter: | 
|  | 109 | wg_ratelimiter_uninit(); | 
|  | 110 | err_free_decrypt_queue: | 
|  | 111 | -	wg_packet_queue_free(&wg->decrypt_queue, true); | 
|  | 112 | +	wg_packet_queue_free(&wg->decrypt_queue); | 
|  | 113 | err_free_encrypt_queue: | 
|  | 114 | -	wg_packet_queue_free(&wg->encrypt_queue, true); | 
|  | 115 | +	wg_packet_queue_free(&wg->encrypt_queue); | 
|  | 116 | err_destroy_packet_crypt: | 
|  | 117 | destroy_workqueue(wg->packet_crypt_wq); | 
|  | 118 | err_destroy_handshake_send: | 
|  | 119 | --- a/drivers/net/wireguard/device.h | 
|  | 120 | +++ b/drivers/net/wireguard/device.h | 
|  | 121 | @@ -27,13 +27,14 @@ struct multicore_worker { | 
|  | 122 |  | 
|  | 123 | struct crypt_queue { | 
|  | 124 | struct ptr_ring ring; | 
|  | 125 | -	union { | 
|  | 126 | -		struct { | 
|  | 127 | -			struct multicore_worker __percpu *worker; | 
|  | 128 | -			int last_cpu; | 
|  | 129 | -		}; | 
|  | 130 | -		struct work_struct work; | 
|  | 131 | -	}; | 
|  | 132 | +	struct multicore_worker __percpu *worker; | 
|  | 133 | +	int last_cpu; | 
|  | 134 | +}; | 
|  | 135 | + | 
|  | 136 | +struct prev_queue { | 
|  | 137 | +	struct sk_buff *head, *tail, *peeked; | 
|  | 138 | +	struct { struct sk_buff *next, *prev; } empty; // Match first 2 members of struct sk_buff. | 
|  | 139 | +	atomic_t count; | 
|  | 140 | }; | 
|  | 141 |  | 
|  | 142 | struct wg_device { | 
|  | 143 | --- a/drivers/net/wireguard/peer.c | 
|  | 144 | +++ b/drivers/net/wireguard/peer.c | 
|  | 145 | @@ -32,27 +32,22 @@ struct wg_peer *wg_peer_create(struct wg | 
|  | 146 | peer = kzalloc(sizeof(*peer), GFP_KERNEL); | 
|  | 147 | if (unlikely(!peer)) | 
|  | 148 | return ERR_PTR(ret); | 
|  | 149 | -	peer->device = wg; | 
|  | 150 | +	if (dst_cache_init(&peer->endpoint_cache, GFP_KERNEL)) | 
|  | 151 | +		goto err; | 
|  | 152 |  | 
|  | 153 | +	peer->device = wg; | 
|  | 154 | wg_noise_handshake_init(&peer->handshake, &wg->static_identity, | 
|  | 155 | public_key, preshared_key, peer); | 
|  | 156 | -	if (dst_cache_init(&peer->endpoint_cache, GFP_KERNEL)) | 
|  | 157 | -		goto err_1; | 
|  | 158 | -	if (wg_packet_queue_init(&peer->tx_queue, wg_packet_tx_worker, false, | 
|  | 159 | -				 MAX_QUEUED_PACKETS)) | 
|  | 160 | -		goto err_2; | 
|  | 161 | -	if (wg_packet_queue_init(&peer->rx_queue, NULL, false, | 
|  | 162 | -				 MAX_QUEUED_PACKETS)) | 
|  | 163 | -		goto err_3; | 
|  | 164 | - | 
|  | 165 | peer->internal_id = atomic64_inc_return(&peer_counter); | 
|  | 166 | peer->serial_work_cpu = nr_cpumask_bits; | 
|  | 167 | wg_cookie_init(&peer->latest_cookie); | 
|  | 168 | wg_timers_init(peer); | 
|  | 169 | wg_cookie_checker_precompute_peer_keys(peer); | 
|  | 170 | spin_lock_init(&peer->keypairs.keypair_update_lock); | 
|  | 171 | -	INIT_WORK(&peer->transmit_handshake_work, | 
|  | 172 | -		  wg_packet_handshake_send_worker); | 
|  | 173 | +	INIT_WORK(&peer->transmit_handshake_work, wg_packet_handshake_send_worker); | 
|  | 174 | +	INIT_WORK(&peer->transmit_packet_work, wg_packet_tx_worker); | 
|  | 175 | +	wg_prev_queue_init(&peer->tx_queue); | 
|  | 176 | +	wg_prev_queue_init(&peer->rx_queue); | 
|  | 177 | rwlock_init(&peer->endpoint_lock); | 
|  | 178 | kref_init(&peer->refcount); | 
|  | 179 | skb_queue_head_init(&peer->staged_packet_queue); | 
|  | 180 | @@ -68,11 +63,7 @@ struct wg_peer *wg_peer_create(struct wg | 
|  | 181 | pr_debug("%s: Peer %llu created\n", wg->dev->name, peer->internal_id); | 
|  | 182 | return peer; | 
|  | 183 |  | 
|  | 184 | -err_3: | 
|  | 185 | -	wg_packet_queue_free(&peer->tx_queue, false); | 
|  | 186 | -err_2: | 
|  | 187 | -	dst_cache_destroy(&peer->endpoint_cache); | 
|  | 188 | -err_1: | 
|  | 189 | +err: | 
|  | 190 | kfree(peer); | 
|  | 191 | return ERR_PTR(ret); | 
|  | 192 | } | 
|  | 193 | @@ -197,8 +188,7 @@ static void rcu_release(struct rcu_head | 
|  | 194 | struct wg_peer *peer = container_of(rcu, struct wg_peer, rcu); | 
|  | 195 |  | 
|  | 196 | dst_cache_destroy(&peer->endpoint_cache); | 
|  | 197 | -	wg_packet_queue_free(&peer->rx_queue, false); | 
|  | 198 | -	wg_packet_queue_free(&peer->tx_queue, false); | 
|  | 199 | +	WARN_ON(wg_prev_queue_peek(&peer->tx_queue) || wg_prev_queue_peek(&peer->rx_queue)); | 
|  | 200 |  | 
|  | 201 | /* The final zeroing takes care of clearing any remaining handshake key | 
|  | 202 | * material and other potentially sensitive information. | 
|  | 203 | --- a/drivers/net/wireguard/peer.h | 
|  | 204 | +++ b/drivers/net/wireguard/peer.h | 
|  | 205 | @@ -36,7 +36,7 @@ struct endpoint { | 
|  | 206 |  | 
|  | 207 | struct wg_peer { | 
|  | 208 | struct wg_device *device; | 
|  | 209 | -	struct crypt_queue tx_queue, rx_queue; | 
|  | 210 | +	struct prev_queue tx_queue, rx_queue; | 
|  | 211 | struct sk_buff_head staged_packet_queue; | 
|  | 212 | int serial_work_cpu; | 
|  | 213 | bool is_dead; | 
|  | 214 | @@ -46,7 +46,7 @@ struct wg_peer { | 
|  | 215 | rwlock_t endpoint_lock; | 
|  | 216 | struct noise_handshake handshake; | 
|  | 217 | atomic64_t last_sent_handshake; | 
|  | 218 | -	struct work_struct transmit_handshake_work, clear_peer_work; | 
|  | 219 | +	struct work_struct transmit_handshake_work, clear_peer_work, transmit_packet_work; | 
|  | 220 | struct cookie latest_cookie; | 
|  | 221 | struct hlist_node pubkey_hash; | 
|  | 222 | u64 rx_bytes, tx_bytes; | 
|  | 223 | --- a/drivers/net/wireguard/queueing.c | 
|  | 224 | +++ b/drivers/net/wireguard/queueing.c | 
|  | 225 | @@ -9,8 +9,7 @@ struct multicore_worker __percpu * | 
|  | 226 | wg_packet_percpu_multicore_worker_alloc(work_func_t function, void *ptr) | 
|  | 227 | { | 
|  | 228 | int cpu; | 
|  | 229 | -	struct multicore_worker __percpu *worker = | 
|  | 230 | -		alloc_percpu(struct multicore_worker); | 
|  | 231 | +	struct multicore_worker __percpu *worker = alloc_percpu(struct multicore_worker); | 
|  | 232 |  | 
|  | 233 | if (!worker) | 
|  | 234 | return NULL; | 
|  | 235 | @@ -23,7 +22,7 @@ wg_packet_percpu_multicore_worker_alloc( | 
|  | 236 | } | 
|  | 237 |  | 
|  | 238 | int wg_packet_queue_init(struct crypt_queue *queue, work_func_t function, | 
|  | 239 | -			 bool multicore, unsigned int len) | 
|  | 240 | +			 unsigned int len) | 
|  | 241 | { | 
|  | 242 | int ret; | 
|  | 243 |  | 
|  | 244 | @@ -31,25 +30,78 @@ int wg_packet_queue_init(struct crypt_qu | 
|  | 245 | ret = ptr_ring_init(&queue->ring, len, GFP_KERNEL); | 
|  | 246 | if (ret) | 
|  | 247 | return ret; | 
|  | 248 | -	if (function) { | 
|  | 249 | -		if (multicore) { | 
|  | 250 | -			queue->worker = wg_packet_percpu_multicore_worker_alloc( | 
|  | 251 | -				function, queue); | 
|  | 252 | -			if (!queue->worker) { | 
|  | 253 | -				ptr_ring_cleanup(&queue->ring, NULL); | 
|  | 254 | -				return -ENOMEM; | 
|  | 255 | -			} | 
|  | 256 | -		} else { | 
|  | 257 | -			INIT_WORK(&queue->work, function); | 
|  | 258 | -		} | 
|  | 259 | +	queue->worker = wg_packet_percpu_multicore_worker_alloc(function, queue); | 
|  | 260 | +	if (!queue->worker) { | 
|  | 261 | +		ptr_ring_cleanup(&queue->ring, NULL); | 
|  | 262 | +		return -ENOMEM; | 
|  | 263 | } | 
|  | 264 | return 0; | 
|  | 265 | } | 
|  | 266 |  | 
|  | 267 | -void wg_packet_queue_free(struct crypt_queue *queue, bool multicore) | 
|  | 268 | +void wg_packet_queue_free(struct crypt_queue *queue) | 
|  | 269 | { | 
|  | 270 | -	if (multicore) | 
|  | 271 | -		free_percpu(queue->worker); | 
|  | 272 | +	free_percpu(queue->worker); | 
|  | 273 | WARN_ON(!__ptr_ring_empty(&queue->ring)); | 
|  | 274 | ptr_ring_cleanup(&queue->ring, NULL); | 
|  | 275 | } | 
|  | 276 | + | 
|  | 277 | +#define NEXT(skb) ((skb)->prev) | 
|  | 278 | +#define STUB(queue) ((struct sk_buff *)&queue->empty) | 
|  | 279 | + | 
|  | 280 | +void wg_prev_queue_init(struct prev_queue *queue) | 
|  | 281 | +{ | 
|  | 282 | +	NEXT(STUB(queue)) = NULL; | 
|  | 283 | +	queue->head = queue->tail = STUB(queue); | 
|  | 284 | +	queue->peeked = NULL; | 
|  | 285 | +	atomic_set(&queue->count, 0); | 
|  | 286 | +	BUILD_BUG_ON( | 
|  | 287 | +		offsetof(struct sk_buff, next) != offsetof(struct prev_queue, empty.next) - | 
|  | 288 | +							offsetof(struct prev_queue, empty) || | 
|  | 289 | +		offsetof(struct sk_buff, prev) != offsetof(struct prev_queue, empty.prev) - | 
|  | 290 | +							 offsetof(struct prev_queue, empty)); | 
|  | 291 | +} | 
|  | 292 | + | 
|  | 293 | +static void __wg_prev_queue_enqueue(struct prev_queue *queue, struct sk_buff *skb) | 
|  | 294 | +{ | 
|  | 295 | +	WRITE_ONCE(NEXT(skb), NULL); | 
|  | 296 | +	WRITE_ONCE(NEXT(xchg_release(&queue->head, skb)), skb); | 
|  | 297 | +} | 
|  | 298 | + | 
|  | 299 | +bool wg_prev_queue_enqueue(struct prev_queue *queue, struct sk_buff *skb) | 
|  | 300 | +{ | 
|  | 301 | +	if (!atomic_add_unless(&queue->count, 1, MAX_QUEUED_PACKETS)) | 
|  | 302 | +		return false; | 
|  | 303 | +	__wg_prev_queue_enqueue(queue, skb); | 
|  | 304 | +	return true; | 
|  | 305 | +} | 
|  | 306 | + | 
|  | 307 | +struct sk_buff *wg_prev_queue_dequeue(struct prev_queue *queue) | 
|  | 308 | +{ | 
|  | 309 | +	struct sk_buff *tail = queue->tail, *next = smp_load_acquire(&NEXT(tail)); | 
|  | 310 | + | 
|  | 311 | +	if (tail == STUB(queue)) { | 
|  | 312 | +		if (!next) | 
|  | 313 | +			return NULL; | 
|  | 314 | +		queue->tail = next; | 
|  | 315 | +		tail = next; | 
|  | 316 | +		next = smp_load_acquire(&NEXT(next)); | 
|  | 317 | +	} | 
|  | 318 | +	if (next) { | 
|  | 319 | +		queue->tail = next; | 
|  | 320 | +		atomic_dec(&queue->count); | 
|  | 321 | +		return tail; | 
|  | 322 | +	} | 
|  | 323 | +	if (tail != READ_ONCE(queue->head)) | 
|  | 324 | +		return NULL; | 
|  | 325 | +	__wg_prev_queue_enqueue(queue, STUB(queue)); | 
|  | 326 | +	next = smp_load_acquire(&NEXT(tail)); | 
|  | 327 | +	if (next) { | 
|  | 328 | +		queue->tail = next; | 
|  | 329 | +		atomic_dec(&queue->count); | 
|  | 330 | +		return tail; | 
|  | 331 | +	} | 
|  | 332 | +	return NULL; | 
|  | 333 | +} | 
|  | 334 | + | 
|  | 335 | +#undef NEXT | 
|  | 336 | +#undef STUB | 
|  | 337 | --- a/drivers/net/wireguard/queueing.h | 
|  | 338 | +++ b/drivers/net/wireguard/queueing.h | 
|  | 339 | @@ -17,12 +17,13 @@ struct wg_device; | 
|  | 340 | struct wg_peer; | 
|  | 341 | struct multicore_worker; | 
|  | 342 | struct crypt_queue; | 
|  | 343 | +struct prev_queue; | 
|  | 344 | struct sk_buff; | 
|  | 345 |  | 
|  | 346 | /* queueing.c APIs: */ | 
|  | 347 | int wg_packet_queue_init(struct crypt_queue *queue, work_func_t function, | 
|  | 348 | -			 bool multicore, unsigned int len); | 
|  | 349 | -void wg_packet_queue_free(struct crypt_queue *queue, bool multicore); | 
|  | 350 | +			 unsigned int len); | 
|  | 351 | +void wg_packet_queue_free(struct crypt_queue *queue); | 
|  | 352 | struct multicore_worker __percpu * | 
|  | 353 | wg_packet_percpu_multicore_worker_alloc(work_func_t function, void *ptr); | 
|  | 354 |  | 
|  | 355 | @@ -135,8 +136,31 @@ static inline int wg_cpumask_next_online | 
|  | 356 | return cpu; | 
|  | 357 | } | 
|  | 358 |  | 
|  | 359 | +void wg_prev_queue_init(struct prev_queue *queue); | 
|  | 360 | + | 
|  | 361 | +/* Multi producer */ | 
|  | 362 | +bool wg_prev_queue_enqueue(struct prev_queue *queue, struct sk_buff *skb); | 
|  | 363 | + | 
|  | 364 | +/* Single consumer */ | 
|  | 365 | +struct sk_buff *wg_prev_queue_dequeue(struct prev_queue *queue); | 
|  | 366 | + | 
|  | 367 | +/* Single consumer */ | 
|  | 368 | +static inline struct sk_buff *wg_prev_queue_peek(struct prev_queue *queue) | 
|  | 369 | +{ | 
|  | 370 | +	if (queue->peeked) | 
|  | 371 | +		return queue->peeked; | 
|  | 372 | +	queue->peeked = wg_prev_queue_dequeue(queue); | 
|  | 373 | +	return queue->peeked; | 
|  | 374 | +} | 
|  | 375 | + | 
|  | 376 | +/* Single consumer */ | 
|  | 377 | +static inline void wg_prev_queue_drop_peeked(struct prev_queue *queue) | 
|  | 378 | +{ | 
|  | 379 | +	queue->peeked = NULL; | 
|  | 380 | +} | 
|  | 381 | + | 
|  | 382 | static inline int wg_queue_enqueue_per_device_and_peer( | 
|  | 383 | -	struct crypt_queue *device_queue, struct crypt_queue *peer_queue, | 
|  | 384 | +	struct crypt_queue *device_queue, struct prev_queue *peer_queue, | 
|  | 385 | struct sk_buff *skb, struct workqueue_struct *wq, int *next_cpu) | 
|  | 386 | { | 
|  | 387 | int cpu; | 
|  | 388 | @@ -145,8 +169,9 @@ static inline int wg_queue_enqueue_per_d | 
|  | 389 | /* We first queue this up for the peer ingestion, but the consumer | 
|  | 390 | * will wait for the state to change to CRYPTED or DEAD before. | 
|  | 391 | */ | 
|  | 392 | -	if (unlikely(ptr_ring_produce_bh(&peer_queue->ring, skb))) | 
|  | 393 | +	if (unlikely(!wg_prev_queue_enqueue(peer_queue, skb))) | 
|  | 394 | return -ENOSPC; | 
|  | 395 | + | 
|  | 396 | /* Then we queue it up in the device queue, which consumes the | 
|  | 397 | * packet as soon as it can. | 
|  | 398 | */ | 
|  | 399 | @@ -157,9 +182,7 @@ static inline int wg_queue_enqueue_per_d | 
|  | 400 | return 0; | 
|  | 401 | } | 
|  | 402 |  | 
|  | 403 | -static inline void wg_queue_enqueue_per_peer(struct crypt_queue *queue, | 
|  | 404 | -					     struct sk_buff *skb, | 
|  | 405 | -					     enum packet_state state) | 
|  | 406 | +static inline void wg_queue_enqueue_per_peer_tx(struct sk_buff *skb, enum packet_state state) | 
|  | 407 | { | 
|  | 408 | /* We take a reference, because as soon as we call atomic_set, the | 
|  | 409 | * peer can be freed from below us. | 
|  | 410 | @@ -167,14 +190,12 @@ static inline void wg_queue_enqueue_per_ | 
|  | 411 | struct wg_peer *peer = wg_peer_get(PACKET_PEER(skb)); | 
|  | 412 |  | 
|  | 413 | atomic_set_release(&PACKET_CB(skb)->state, state); | 
|  | 414 | -	queue_work_on(wg_cpumask_choose_online(&peer->serial_work_cpu, | 
|  | 415 | -					       peer->internal_id), | 
|  | 416 | -		      peer->device->packet_crypt_wq, &queue->work); | 
|  | 417 | +	queue_work_on(wg_cpumask_choose_online(&peer->serial_work_cpu, peer->internal_id), | 
|  | 418 | +		      peer->device->packet_crypt_wq, &peer->transmit_packet_work); | 
|  | 419 | wg_peer_put(peer); | 
|  | 420 | } | 
|  | 421 |  | 
|  | 422 | -static inline void wg_queue_enqueue_per_peer_napi(struct sk_buff *skb, | 
|  | 423 | -						  enum packet_state state) | 
|  | 424 | +static inline void wg_queue_enqueue_per_peer_rx(struct sk_buff *skb, enum packet_state state) | 
|  | 425 | { | 
|  | 426 | /* We take a reference, because as soon as we call atomic_set, the | 
|  | 427 | * peer can be freed from below us. | 
|  | 428 | --- a/drivers/net/wireguard/receive.c | 
|  | 429 | +++ b/drivers/net/wireguard/receive.c | 
|  | 430 | @@ -444,7 +444,6 @@ packet_processed: | 
|  | 431 | int wg_packet_rx_poll(struct napi_struct *napi, int budget) | 
|  | 432 | { | 
|  | 433 | struct wg_peer *peer = container_of(napi, struct wg_peer, napi); | 
|  | 434 | -	struct crypt_queue *queue = &peer->rx_queue; | 
|  | 435 | struct noise_keypair *keypair; | 
|  | 436 | struct endpoint endpoint; | 
|  | 437 | enum packet_state state; | 
|  | 438 | @@ -455,11 +454,10 @@ int wg_packet_rx_poll(struct napi_struct | 
|  | 439 | if (unlikely(budget <= 0)) | 
|  | 440 | return 0; | 
|  | 441 |  | 
|  | 442 | -	while ((skb = __ptr_ring_peek(&queue->ring)) != NULL && | 
|  | 443 | +	while ((skb = wg_prev_queue_peek(&peer->rx_queue)) != NULL && | 
|  | 444 | (state = atomic_read_acquire(&PACKET_CB(skb)->state)) != | 
|  | 445 | PACKET_STATE_UNCRYPTED) { | 
|  | 446 | -		__ptr_ring_discard_one(&queue->ring); | 
|  | 447 | -		peer = PACKET_PEER(skb); | 
|  | 448 | +		wg_prev_queue_drop_peeked(&peer->rx_queue); | 
|  | 449 | keypair = PACKET_CB(skb)->keypair; | 
|  | 450 | free = true; | 
|  | 451 |  | 
|  | 452 | @@ -508,7 +506,7 @@ void wg_packet_decrypt_worker(struct wor | 
|  | 453 | enum packet_state state = | 
|  | 454 | likely(decrypt_packet(skb, PACKET_CB(skb)->keypair)) ? | 
|  | 455 | PACKET_STATE_CRYPTED : PACKET_STATE_DEAD; | 
|  | 456 | -		wg_queue_enqueue_per_peer_napi(skb, state); | 
|  | 457 | +		wg_queue_enqueue_per_peer_rx(skb, state); | 
|  | 458 | if (need_resched()) | 
|  | 459 | cond_resched(); | 
|  | 460 | } | 
|  | 461 | @@ -531,12 +529,10 @@ static void wg_packet_consume_data(struc | 
|  | 462 | if (unlikely(READ_ONCE(peer->is_dead))) | 
|  | 463 | goto err; | 
|  | 464 |  | 
|  | 465 | -	ret = wg_queue_enqueue_per_device_and_peer(&wg->decrypt_queue, | 
|  | 466 | -						   &peer->rx_queue, skb, | 
|  | 467 | -						   wg->packet_crypt_wq, | 
|  | 468 | -						   &wg->decrypt_queue.last_cpu); | 
|  | 469 | +	ret = wg_queue_enqueue_per_device_and_peer(&wg->decrypt_queue, &peer->rx_queue, skb, | 
|  | 470 | +						   wg->packet_crypt_wq, &wg->decrypt_queue.last_cpu); | 
|  | 471 | if (unlikely(ret == -EPIPE)) | 
|  | 472 | -		wg_queue_enqueue_per_peer_napi(skb, PACKET_STATE_DEAD); | 
|  | 473 | +		wg_queue_enqueue_per_peer_rx(skb, PACKET_STATE_DEAD); | 
|  | 474 | if (likely(!ret || ret == -EPIPE)) { | 
|  | 475 | rcu_read_unlock_bh(); | 
|  | 476 | return; | 
|  | 477 | --- a/drivers/net/wireguard/send.c | 
|  | 478 | +++ b/drivers/net/wireguard/send.c | 
|  | 479 | @@ -239,8 +239,7 @@ void wg_packet_send_keepalive(struct wg_ | 
|  | 480 | wg_packet_send_staged_packets(peer); | 
|  | 481 | } | 
|  | 482 |  | 
|  | 483 | -static void wg_packet_create_data_done(struct sk_buff *first, | 
|  | 484 | -				       struct wg_peer *peer) | 
|  | 485 | +static void wg_packet_create_data_done(struct wg_peer *peer, struct sk_buff *first) | 
|  | 486 | { | 
|  | 487 | struct sk_buff *skb, *next; | 
|  | 488 | bool is_keepalive, data_sent = false; | 
|  | 489 | @@ -262,22 +261,19 @@ static void wg_packet_create_data_done(s | 
|  | 490 |  | 
|  | 491 | void wg_packet_tx_worker(struct work_struct *work) | 
|  | 492 | { | 
|  | 493 | -	struct crypt_queue *queue = container_of(work, struct crypt_queue, | 
|  | 494 | -						 work); | 
|  | 495 | +	struct wg_peer *peer = container_of(work, struct wg_peer, transmit_packet_work); | 
|  | 496 | struct noise_keypair *keypair; | 
|  | 497 | enum packet_state state; | 
|  | 498 | struct sk_buff *first; | 
|  | 499 | -	struct wg_peer *peer; | 
|  | 500 |  | 
|  | 501 | -	while ((first = __ptr_ring_peek(&queue->ring)) != NULL && | 
|  | 502 | +	while ((first = wg_prev_queue_peek(&peer->tx_queue)) != NULL && | 
|  | 503 | (state = atomic_read_acquire(&PACKET_CB(first)->state)) != | 
|  | 504 | PACKET_STATE_UNCRYPTED) { | 
|  | 505 | -		__ptr_ring_discard_one(&queue->ring); | 
|  | 506 | -		peer = PACKET_PEER(first); | 
|  | 507 | +		wg_prev_queue_drop_peeked(&peer->tx_queue); | 
|  | 508 | keypair = PACKET_CB(first)->keypair; | 
|  | 509 |  | 
|  | 510 | if (likely(state == PACKET_STATE_CRYPTED)) | 
|  | 511 | -			wg_packet_create_data_done(first, peer); | 
|  | 512 | +			wg_packet_create_data_done(peer, first); | 
|  | 513 | else | 
|  | 514 | kfree_skb_list(first); | 
|  | 515 |  | 
|  | 516 | @@ -306,16 +302,14 @@ void wg_packet_encrypt_worker(struct wor | 
|  | 517 | break; | 
|  | 518 | } | 
|  | 519 | } | 
|  | 520 | -		wg_queue_enqueue_per_peer(&PACKET_PEER(first)->tx_queue, first, | 
|  | 521 | -					  state); | 
|  | 522 | +		wg_queue_enqueue_per_peer_tx(first, state); | 
|  | 523 | if (need_resched()) | 
|  | 524 | cond_resched(); | 
|  | 525 | } | 
|  | 526 | } | 
|  | 527 |  | 
|  | 528 | -static void wg_packet_create_data(struct sk_buff *first) | 
|  | 529 | +static void wg_packet_create_data(struct wg_peer *peer, struct sk_buff *first) | 
|  | 530 | { | 
|  | 531 | -	struct wg_peer *peer = PACKET_PEER(first); | 
|  | 532 | struct wg_device *wg = peer->device; | 
|  | 533 | int ret = -EINVAL; | 
|  | 534 |  | 
|  | 535 | @@ -323,13 +317,10 @@ static void wg_packet_create_data(struct | 
|  | 536 | if (unlikely(READ_ONCE(peer->is_dead))) | 
|  | 537 | goto err; | 
|  | 538 |  | 
|  | 539 | -	ret = wg_queue_enqueue_per_device_and_peer(&wg->encrypt_queue, | 
|  | 540 | -						   &peer->tx_queue, first, | 
|  | 541 | -						   wg->packet_crypt_wq, | 
|  | 542 | -						   &wg->encrypt_queue.last_cpu); | 
|  | 543 | +	ret = wg_queue_enqueue_per_device_and_peer(&wg->encrypt_queue, &peer->tx_queue, first, | 
|  | 544 | +						   wg->packet_crypt_wq, &wg->encrypt_queue.last_cpu); | 
|  | 545 | if (unlikely(ret == -EPIPE)) | 
|  | 546 | -		wg_queue_enqueue_per_peer(&peer->tx_queue, first, | 
|  | 547 | -					  PACKET_STATE_DEAD); | 
|  | 548 | +		wg_queue_enqueue_per_peer_tx(first, PACKET_STATE_DEAD); | 
|  | 549 | err: | 
|  | 550 | rcu_read_unlock_bh(); | 
|  | 551 | if (likely(!ret || ret == -EPIPE)) | 
|  | 552 | @@ -393,7 +384,7 @@ void wg_packet_send_staged_packets(struc | 
|  | 553 | packets.prev->next = NULL; | 
|  | 554 | wg_peer_get(keypair->entry.peer); | 
|  | 555 | PACKET_CB(packets.next)->keypair = keypair; | 
|  | 556 | -	wg_packet_create_data(packets.next); | 
|  | 557 | +	wg_packet_create_data(peer, packets.next); | 
|  | 558 | return; | 
|  | 559 |  | 
|  | 560 | out_invalid: |