xref: /linux-6.15/include/linux/skmsg.h (revision 2d91ecac) 
1 /* SPDX-License-Identifier: GPL-2.0 */
2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
3 
4 #ifndef _LINUX_SKMSG_H
5 #define _LINUX_SKMSG_H
6 
7 #include <linux/bpf.h>
8 #include <linux/filter.h>
9 #include <linux/scatterlist.h>
10 #include <linux/skbuff.h>
11 
12 #include <net/sock.h>
13 #include <net/tcp.h>
14 #include <net/strparser.h>
15 
16 #define MAX_MSG_FRAGS			MAX_SKB_FRAGS
17 #define NR_MSG_FRAG_IDS			(MAX_MSG_FRAGS + 1)
18 
19 enum __sk_action {
20 	__SK_DROP = 0,
21 	__SK_PASS,
22 	__SK_REDIRECT,
23 	__SK_NONE,
24 };
25 
26 struct sk_msg_sg {
27 	u32				start;
28 	u32				curr;
29 	u32				end;
30 	u32				size;
31 	u32				copybreak;
32 	DECLARE_BITMAP(copy, MAX_MSG_FRAGS + 2);
33 	/* The extra two elements:
34 	 * 1) used for chaining the front and sections when the list becomes
35 	 *    partitioned (e.g. end < start). The crypto APIs require the
36 	 *    chaining;
37 	 * 2) to chain tailer SG entries after the message.
38 	 */
39 	struct scatterlist		data[MAX_MSG_FRAGS + 2];
40 };
41 
42 /* UAPI in filter.c depends on struct sk_msg_sg being first element. */
43 struct sk_msg {
44 	struct sk_msg_sg		sg;
45 	void				*data;
46 	void				*data_end;
47 	u32				apply_bytes;
48 	u32				cork_bytes;
49 	u32				flags;
50 	struct sk_buff			*skb;
51 	struct sock			*sk_redir;
52 	struct sock			*sk;
53 	struct list_head		list;
54 };
55 
56 struct sk_psock_progs {
57 	struct bpf_prog			*msg_parser;
58 	struct bpf_prog			*stream_parser;
59 	struct bpf_prog			*stream_verdict;
60 	struct bpf_prog			*skb_verdict;
61 };
62 
63 enum sk_psock_state_bits {
64 	SK_PSOCK_TX_ENABLED,
65 };
66 
67 struct sk_psock_link {
68 	struct list_head		list;
69 	struct bpf_map			*map;
70 	void				*link_raw;
71 };
72 
73 struct sk_psock_work_state {
74 	struct sk_buff			*skb;
75 	u32				len;
76 	u32				off;
77 };
78 
79 struct sk_psock {
80 	struct sock			*sk;
81 	struct sock			*sk_redir;
82 	u32				apply_bytes;
83 	u32				cork_bytes;
84 	u32				eval;
85 	struct sk_msg			*cork;
86 	struct sk_psock_progs		progs;
87 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
88 	struct strparser		strp;
89 #endif
90 	struct sk_buff_head		ingress_skb;
91 	struct list_head		ingress_msg;
92 	spinlock_t			ingress_lock;
93 	unsigned long			state;
94 	struct list_head		link;
95 	spinlock_t			link_lock;
96 	refcount_t			refcnt;
97 	void (*saved_unhash)(struct sock *sk);
98 	void (*saved_destroy)(struct sock *sk);
99 	void (*saved_close)(struct sock *sk, long timeout);
100 	void (*saved_write_space)(struct sock *sk);
101 	void (*saved_data_ready)(struct sock *sk);
102 	int  (*psock_update_sk_prot)(struct sock *sk, struct sk_psock *psock,
103 				     bool restore);
104 	struct proto			*sk_proto;
105 	struct mutex			work_mutex;
106 	struct sk_psock_work_state	work_state;
107 	struct work_struct		work;
108 	struct rcu_work			rwork;
109 };
110 
111 int sk_msg_alloc(struct sock *sk, struct sk_msg *msg, int len,
112 		 int elem_first_coalesce);
113 int sk_msg_clone(struct sock *sk, struct sk_msg *dst, struct sk_msg *src,
114 		 u32 off, u32 len);
115 void sk_msg_trim(struct sock *sk, struct sk_msg *msg, int len);
116 int sk_msg_free(struct sock *sk, struct sk_msg *msg);
117 int sk_msg_free_nocharge(struct sock *sk, struct sk_msg *msg);
118 void sk_msg_free_partial(struct sock *sk, struct sk_msg *msg, u32 bytes);
119 void sk_msg_free_partial_nocharge(struct sock *sk, struct sk_msg *msg,
120 				  u32 bytes);
121 
122 void sk_msg_return(struct sock *sk, struct sk_msg *msg, int bytes);
123 void sk_msg_return_zero(struct sock *sk, struct sk_msg *msg, int bytes);
124 
125 int sk_msg_zerocopy_from_iter(struct sock *sk, struct iov_iter *from,
126 			      struct sk_msg *msg, u32 bytes);
127 int sk_msg_memcopy_from_iter(struct sock *sk, struct iov_iter *from,
128 			     struct sk_msg *msg, u32 bytes);
129 int sk_msg_recvmsg(struct sock *sk, struct sk_psock *psock, struct msghdr *msg,
130 		   int len, int flags);
131 bool sk_msg_is_readable(struct sock *sk);
132 
133 static inline void sk_msg_check_to_free(struct sk_msg *msg, u32 i, u32 bytes)
134 {
135 	WARN_ON(i == msg->sg.end && bytes);
136 }
137 
138 static inline void sk_msg_apply_bytes(struct sk_psock *psock, u32 bytes)
139 {
140 	if (psock->apply_bytes) {
141 		if (psock->apply_bytes < bytes)
142 			psock->apply_bytes = 0;
143 		else
144 			psock->apply_bytes -= bytes;
145 	}
146 }
147 
148 static inline u32 sk_msg_iter_dist(u32 start, u32 end)
149 {
150 	return end >= start ? end - start : end + (NR_MSG_FRAG_IDS - start);
151 }
152 
153 #define sk_msg_iter_var_prev(var)			\
154 	do {						\
155 		if (var == 0)				\
156 			var = NR_MSG_FRAG_IDS - 1;	\
157 		else					\
158 			var--;				\
159 	} while (0)
160 
161 #define sk_msg_iter_var_next(var)			\
162 	do {						\
163 		var++;					\
164 		if (var == NR_MSG_FRAG_IDS)		\
165 			var = 0;			\
166 	} while (0)
167 
168 #define sk_msg_iter_prev(msg, which)			\
169 	sk_msg_iter_var_prev(msg->sg.which)
170 
171 #define sk_msg_iter_next(msg, which)			\
172 	sk_msg_iter_var_next(msg->sg.which)
173 
174 static inline void sk_msg_init(struct sk_msg *msg)
175 {
176 	BUILD_BUG_ON(ARRAY_SIZE(msg->sg.data) - 1 != NR_MSG_FRAG_IDS);
177 	memset(msg, 0, sizeof(*msg));
178 	sg_init_marker(msg->sg.data, NR_MSG_FRAG_IDS);
179 }
180 
181 static inline void sk_msg_xfer(struct sk_msg *dst, struct sk_msg *src,
182 			       int which, u32 size)
183 {
184 	dst->sg.data[which] = src->sg.data[which];
185 	dst->sg.data[which].length  = size;
186 	dst->sg.size		   += size;
187 	src->sg.size		   -= size;
188 	src->sg.data[which].length -= size;
189 	src->sg.data[which].offset += size;
190 }
191 
192 static inline void sk_msg_xfer_full(struct sk_msg *dst, struct sk_msg *src)
193 {
194 	memcpy(dst, src, sizeof(*src));
195 	sk_msg_init(src);
196 }
197 
198 static inline bool sk_msg_full(const struct sk_msg *msg)
199 {
200 	return sk_msg_iter_dist(msg->sg.start, msg->sg.end) == MAX_MSG_FRAGS;
201 }
202 
203 static inline u32 sk_msg_elem_used(const struct sk_msg *msg)
204 {
205 	return sk_msg_iter_dist(msg->sg.start, msg->sg.end);
206 }
207 
208 static inline struct scatterlist *sk_msg_elem(struct sk_msg *msg, int which)
209 {
210 	return &msg->sg.data[which];
211 }
212 
213 static inline struct scatterlist sk_msg_elem_cpy(struct sk_msg *msg, int which)
214 {
215 	return msg->sg.data[which];
216 }
217 
218 static inline struct page *sk_msg_page(struct sk_msg *msg, int which)
219 {
220 	return sg_page(sk_msg_elem(msg, which));
221 }
222 
223 static inline bool sk_msg_to_ingress(const struct sk_msg *msg)
224 {
225 	return msg->flags & BPF_F_INGRESS;
226 }
227 
228 static inline void sk_msg_compute_data_pointers(struct sk_msg *msg)
229 {
230 	struct scatterlist *sge = sk_msg_elem(msg, msg->sg.start);
231 
232 	if (test_bit(msg->sg.start, msg->sg.copy)) {
233 		msg->data = NULL;
234 		msg->data_end = NULL;
235 	} else {
236 		msg->data = sg_virt(sge);
237 		msg->data_end = msg->data + sge->length;
238 	}
239 }
240 
241 static inline void sk_msg_page_add(struct sk_msg *msg, struct page *page,
242 				   u32 len, u32 offset)
243 {
244 	struct scatterlist *sge;
245 
246 	get_page(page);
247 	sge = sk_msg_elem(msg, msg->sg.end);
248 	sg_set_page(sge, page, len, offset);
249 	sg_unmark_end(sge);
250 
251 	__set_bit(msg->sg.end, msg->sg.copy);
252 	msg->sg.size += len;
253 	sk_msg_iter_next(msg, end);
254 }
255 
256 static inline void sk_msg_sg_copy(struct sk_msg *msg, u32 i, bool copy_state)
257 {
258 	do {
259 		if (copy_state)
260 			__set_bit(i, msg->sg.copy);
261 		else
262 			__clear_bit(i, msg->sg.copy);
263 		sk_msg_iter_var_next(i);
264 		if (i == msg->sg.end)
265 			break;
266 	} while (1);
267 }
268 
269 static inline void sk_msg_sg_copy_set(struct sk_msg *msg, u32 start)
270 {
271 	sk_msg_sg_copy(msg, start, true);
272 }
273 
274 static inline void sk_msg_sg_copy_clear(struct sk_msg *msg, u32 start)
275 {
276 	sk_msg_sg_copy(msg, start, false);
277 }
278 
279 static inline struct sk_psock *sk_psock(const struct sock *sk)
280 {
281 	return rcu_dereference_sk_user_data(sk);
282 }
283 
284 static inline void sk_psock_set_state(struct sk_psock *psock,
285 				      enum sk_psock_state_bits bit)
286 {
287 	set_bit(bit, &psock->state);
288 }
289 
290 static inline void sk_psock_clear_state(struct sk_psock *psock,
291 					enum sk_psock_state_bits bit)
292 {
293 	clear_bit(bit, &psock->state);
294 }
295 
296 static inline bool sk_psock_test_state(const struct sk_psock *psock,
297 				       enum sk_psock_state_bits bit)
298 {
299 	return test_bit(bit, &psock->state);
300 }
301 
302 static inline void sock_drop(struct sock *sk, struct sk_buff *skb)
303 {
304 	sk_drops_add(sk, skb);
305 	kfree_skb(skb);
306 }
307 
308 static inline void sk_psock_queue_msg(struct sk_psock *psock,
309 				      struct sk_msg *msg)
310 {
311 	spin_lock_bh(&psock->ingress_lock);
312 	if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED))
313 		list_add_tail(&msg->list, &psock->ingress_msg);
314 	else {
315 		sk_msg_free(psock->sk, msg);
316 		kfree(msg);
317 	}
318 	spin_unlock_bh(&psock->ingress_lock);
319 }
320 
321 static inline struct sk_msg *sk_psock_dequeue_msg(struct sk_psock *psock)
322 {
323 	struct sk_msg *msg;
324 
325 	spin_lock_bh(&psock->ingress_lock);
326 	msg = list_first_entry_or_null(&psock->ingress_msg, struct sk_msg, list);
327 	if (msg)
328 		list_del(&msg->list);
329 	spin_unlock_bh(&psock->ingress_lock);
330 	return msg;
331 }
332 
333 static inline struct sk_msg *sk_psock_peek_msg(struct sk_psock *psock)
334 {
335 	struct sk_msg *msg;
336 
337 	spin_lock_bh(&psock->ingress_lock);
338 	msg = list_first_entry_or_null(&psock->ingress_msg, struct sk_msg, list);
339 	spin_unlock_bh(&psock->ingress_lock);
340 	return msg;
341 }
342 
343 static inline struct sk_msg *sk_psock_next_msg(struct sk_psock *psock,
344 					       struct sk_msg *msg)
345 {
346 	struct sk_msg *ret;
347 
348 	spin_lock_bh(&psock->ingress_lock);
349 	if (list_is_last(&msg->list, &psock->ingress_msg))
350 		ret = NULL;
351 	else
352 		ret = list_next_entry(msg, list);
353 	spin_unlock_bh(&psock->ingress_lock);
354 	return ret;
355 }
356 
357 static inline bool sk_psock_queue_empty(const struct sk_psock *psock)
358 {
359 	return psock ? list_empty(&psock->ingress_msg) : true;
360 }
361 
362 static inline void kfree_sk_msg(struct sk_msg *msg)
363 {
364 	if (msg->skb)
365 		consume_skb(msg->skb);
366 	kfree(msg);
367 }
368 
369 static inline void sk_psock_report_error(struct sk_psock *psock, int err)
370 {
371 	struct sock *sk = psock->sk;
372 
373 	sk->sk_err = err;
374 	sk_error_report(sk);
375 }
376 
377 struct sk_psock *sk_psock_init(struct sock *sk, int node);
378 void sk_psock_stop(struct sk_psock *psock, bool wait);
379 
380 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
381 int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock);
382 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock);
383 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock);
384 #else
385 static inline int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
386 {
387 	return -EOPNOTSUPP;
388 }
389 
390 static inline void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
391 {
392 }
393 
394 static inline void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
395 {
396 }
397 #endif
398 
399 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock);
400 void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock);
401 
402 int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock,
403 			 struct sk_msg *msg);
404 
405 static inline struct sk_psock_link *sk_psock_init_link(void)
406 {
407 	return kzalloc(sizeof(struct sk_psock_link),
408 		       GFP_ATOMIC | __GFP_NOWARN);
409 }
410 
411 static inline void sk_psock_free_link(struct sk_psock_link *link)
412 {
413 	kfree(link);
414 }
415 
416 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock);
417 
418 static inline void sk_psock_cork_free(struct sk_psock *psock)
419 {
420 	if (psock->cork) {
421 		sk_msg_free(psock->sk, psock->cork);
422 		kfree(psock->cork);
423 		psock->cork = NULL;
424 	}
425 }
426 
427 static inline void sk_psock_restore_proto(struct sock *sk,
428 					  struct sk_psock *psock)
429 {
430 	if (psock->psock_update_sk_prot)
431 		psock->psock_update_sk_prot(sk, psock, true);
432 }
433 
434 static inline struct sk_psock *sk_psock_get(struct sock *sk)
435 {
436 	struct sk_psock *psock;
437 
438 	rcu_read_lock();
439 	psock = sk_psock(sk);
440 	if (psock && !refcount_inc_not_zero(&psock->refcnt))
441 		psock = NULL;
442 	rcu_read_unlock();
443 	return psock;
444 }
445 
446 void sk_psock_drop(struct sock *sk, struct sk_psock *psock);
447 
448 static inline void sk_psock_put(struct sock *sk, struct sk_psock *psock)
449 {
450 	if (refcount_dec_and_test(&psock->refcnt))
451 		sk_psock_drop(sk, psock);
452 }
453 
454 static inline void sk_psock_data_ready(struct sock *sk, struct sk_psock *psock)
455 {
456 	if (psock->saved_data_ready)
457 		psock->saved_data_ready(sk);
458 	else
459 		sk->sk_data_ready(sk);
460 }
461 
462 static inline void psock_set_prog(struct bpf_prog **pprog,
463 				  struct bpf_prog *prog)
464 {
465 	prog = xchg(pprog, prog);
466 	if (prog)
467 		bpf_prog_put(prog);
468 }
469 
470 static inline int psock_replace_prog(struct bpf_prog **pprog,
471 				     struct bpf_prog *prog,
472 				     struct bpf_prog *old)
473 {
474 	if (cmpxchg(pprog, old, prog) != old)
475 		return -ENOENT;
476 
477 	if (old)
478 		bpf_prog_put(old);
479 
480 	return 0;
481 }
482 
483 static inline void psock_progs_drop(struct sk_psock_progs *progs)
484 {
485 	psock_set_prog(&progs->msg_parser, NULL);
486 	psock_set_prog(&progs->stream_parser, NULL);
487 	psock_set_prog(&progs->stream_verdict, NULL);
488 	psock_set_prog(&progs->skb_verdict, NULL);
489 }
490 
491 int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb);
492 
493 static inline bool sk_psock_strp_enabled(struct sk_psock *psock)
494 {
495 	if (!psock)
496 		return false;
497 	return !!psock->saved_data_ready;
498 }
499 
500 static inline bool sk_is_udp(const struct sock *sk)
501 {
502 	return sk->sk_type == SOCK_DGRAM &&
503 	       sk->sk_protocol == IPPROTO_UDP;
504 }
505 
506 #if IS_ENABLED(CONFIG_NET_SOCK_MSG)
507 
508 #define BPF_F_STRPARSER	(1UL << 1)
509 
510 /* We only have two bits so far. */
511 #define BPF_F_PTR_MASK ~(BPF_F_INGRESS | BPF_F_STRPARSER)
512 
513 static inline bool skb_bpf_strparser(const struct sk_buff *skb)
514 {
515 	unsigned long sk_redir = skb->_sk_redir;
516 
517 	return sk_redir & BPF_F_STRPARSER;
518 }
519 
520 static inline void skb_bpf_set_strparser(struct sk_buff *skb)
521 {
522 	skb->_sk_redir |= BPF_F_STRPARSER;
523 }
524 
525 static inline bool skb_bpf_ingress(const struct sk_buff *skb)
526 {
527 	unsigned long sk_redir = skb->_sk_redir;
528 
529 	return sk_redir & BPF_F_INGRESS;
530 }
531 
532 static inline void skb_bpf_set_ingress(struct sk_buff *skb)
533 {
534 	skb->_sk_redir |= BPF_F_INGRESS;
535 }
536 
537 static inline void skb_bpf_set_redir(struct sk_buff *skb, struct sock *sk_redir,
538 				     bool ingress)
539 {
540 	skb->_sk_redir = (unsigned long)sk_redir;
541 	if (ingress)
542 		skb->_sk_redir |= BPF_F_INGRESS;
543 }
544 
545 static inline struct sock *skb_bpf_redirect_fetch(const struct sk_buff *skb)
546 {
547 	unsigned long sk_redir = skb->_sk_redir;
548 
549 	return (struct sock *)(sk_redir & BPF_F_PTR_MASK);
550 }
551 
552 static inline void skb_bpf_redirect_clear(struct sk_buff *skb)
553 {
554 	skb->_sk_redir = 0;
555 }
556 #endif /* CONFIG_NET_SOCK_MSG */
557 #endif /* _LINUX_SKMSG_H */
558