xref: /linux-6.15/include/linux/skmsg.h (revision 1ddbddd7) 
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 	unsigned long			copy;
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 static_assert(BITS_PER_LONG >= NR_MSG_FRAG_IDS);
42 
43 /* UAPI in filter.c depends on struct sk_msg_sg being first element. */
44 struct sk_msg {
45 	struct sk_msg_sg		sg;
46 	void				*data;
47 	void				*data_end;
48 	u32				apply_bytes;
49 	u32				cork_bytes;
50 	u32				flags;
51 	struct sk_buff			*skb;
52 	struct sock			*sk_redir;
53 	struct sock			*sk;
54 	struct list_head		list;
55 };
56 
57 struct sk_psock_progs {
58 	struct bpf_prog			*msg_parser;
59 	struct bpf_prog			*stream_parser;
60 	struct bpf_prog			*stream_verdict;
61 	struct bpf_prog			*skb_verdict;
62 };
63 
64 enum sk_psock_state_bits {
65 	SK_PSOCK_TX_ENABLED,
66 };
67 
68 struct sk_psock_link {
69 	struct list_head		list;
70 	struct bpf_map			*map;
71 	void				*link_raw;
72 };
73 
74 struct sk_psock_work_state {
75 	struct sk_buff			*skb;
76 	u32				len;
77 	u32				off;
78 };
79 
80 struct sk_psock {
81 	struct sock			*sk;
82 	struct sock			*sk_redir;
83 	u32				apply_bytes;
84 	u32				cork_bytes;
85 	u32				eval;
86 	struct sk_msg			*cork;
87 	struct sk_psock_progs		progs;
88 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
89 	struct strparser		strp;
90 #endif
91 	struct sk_buff_head		ingress_skb;
92 	struct list_head		ingress_msg;
93 	spinlock_t			ingress_lock;
94 	unsigned long			state;
95 	struct list_head		link;
96 	spinlock_t			link_lock;
97 	refcount_t			refcnt;
98 	void (*saved_unhash)(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 drop_sk_msg(struct sk_psock *psock, struct sk_msg *msg)
309 {
310 	if (msg->skb)
311 		sock_drop(psock->sk, msg->skb);
312 	kfree(msg);
313 }
314 
315 static inline void sk_psock_queue_msg(struct sk_psock *psock,
316 				      struct sk_msg *msg)
317 {
318 	spin_lock_bh(&psock->ingress_lock);
319 	if (sk_psock_test_state(psock, SK_PSOCK_TX_ENABLED))
320 		list_add_tail(&msg->list, &psock->ingress_msg);
321 	else
322 		drop_sk_msg(psock, msg);
323 	spin_unlock_bh(&psock->ingress_lock);
324 }
325 
326 static inline struct sk_msg *sk_psock_dequeue_msg(struct sk_psock *psock)
327 {
328 	struct sk_msg *msg;
329 
330 	spin_lock_bh(&psock->ingress_lock);
331 	msg = list_first_entry_or_null(&psock->ingress_msg, struct sk_msg, list);
332 	if (msg)
333 		list_del(&msg->list);
334 	spin_unlock_bh(&psock->ingress_lock);
335 	return msg;
336 }
337 
338 static inline struct sk_msg *sk_psock_peek_msg(struct sk_psock *psock)
339 {
340 	struct sk_msg *msg;
341 
342 	spin_lock_bh(&psock->ingress_lock);
343 	msg = list_first_entry_or_null(&psock->ingress_msg, struct sk_msg, list);
344 	spin_unlock_bh(&psock->ingress_lock);
345 	return msg;
346 }
347 
348 static inline struct sk_msg *sk_psock_next_msg(struct sk_psock *psock,
349 					       struct sk_msg *msg)
350 {
351 	struct sk_msg *ret;
352 
353 	spin_lock_bh(&psock->ingress_lock);
354 	if (list_is_last(&msg->list, &psock->ingress_msg))
355 		ret = NULL;
356 	else
357 		ret = list_next_entry(msg, list);
358 	spin_unlock_bh(&psock->ingress_lock);
359 	return ret;
360 }
361 
362 static inline bool sk_psock_queue_empty(const struct sk_psock *psock)
363 {
364 	return psock ? list_empty(&psock->ingress_msg) : true;
365 }
366 
367 static inline void kfree_sk_msg(struct sk_msg *msg)
368 {
369 	if (msg->skb)
370 		consume_skb(msg->skb);
371 	kfree(msg);
372 }
373 
374 static inline void sk_psock_report_error(struct sk_psock *psock, int err)
375 {
376 	struct sock *sk = psock->sk;
377 
378 	sk->sk_err = err;
379 	sk_error_report(sk);
380 }
381 
382 struct sk_psock *sk_psock_init(struct sock *sk, int node);
383 void sk_psock_stop(struct sk_psock *psock, bool wait);
384 
385 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
386 int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock);
387 void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock);
388 void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock);
389 #else
390 static inline int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock)
391 {
392 	return -EOPNOTSUPP;
393 }
394 
395 static inline void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock)
396 {
397 }
398 
399 static inline void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock)
400 {
401 }
402 #endif
403 
404 void sk_psock_start_verdict(struct sock *sk, struct sk_psock *psock);
405 void sk_psock_stop_verdict(struct sock *sk, struct sk_psock *psock);
406 
407 int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock,
408 			 struct sk_msg *msg);
409 
410 static inline struct sk_psock_link *sk_psock_init_link(void)
411 {
412 	return kzalloc(sizeof(struct sk_psock_link),
413 		       GFP_ATOMIC | __GFP_NOWARN);
414 }
415 
416 static inline void sk_psock_free_link(struct sk_psock_link *link)
417 {
418 	kfree(link);
419 }
420 
421 struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock);
422 
423 static inline void sk_psock_cork_free(struct sk_psock *psock)
424 {
425 	if (psock->cork) {
426 		sk_msg_free(psock->sk, psock->cork);
427 		kfree(psock->cork);
428 		psock->cork = NULL;
429 	}
430 }
431 
432 static inline void sk_psock_restore_proto(struct sock *sk,
433 					  struct sk_psock *psock)
434 {
435 	if (psock->psock_update_sk_prot)
436 		psock->psock_update_sk_prot(sk, psock, true);
437 }
438 
439 static inline struct sk_psock *sk_psock_get(struct sock *sk)
440 {
441 	struct sk_psock *psock;
442 
443 	rcu_read_lock();
444 	psock = sk_psock(sk);
445 	if (psock && !refcount_inc_not_zero(&psock->refcnt))
446 		psock = NULL;
447 	rcu_read_unlock();
448 	return psock;
449 }
450 
451 void sk_psock_drop(struct sock *sk, struct sk_psock *psock);
452 
453 static inline void sk_psock_put(struct sock *sk, struct sk_psock *psock)
454 {
455 	if (refcount_dec_and_test(&psock->refcnt))
456 		sk_psock_drop(sk, psock);
457 }
458 
459 static inline void sk_psock_data_ready(struct sock *sk, struct sk_psock *psock)
460 {
461 	if (psock->saved_data_ready)
462 		psock->saved_data_ready(sk);
463 	else
464 		sk->sk_data_ready(sk);
465 }
466 
467 static inline void psock_set_prog(struct bpf_prog **pprog,
468 				  struct bpf_prog *prog)
469 {
470 	prog = xchg(pprog, prog);
471 	if (prog)
472 		bpf_prog_put(prog);
473 }
474 
475 static inline int psock_replace_prog(struct bpf_prog **pprog,
476 				     struct bpf_prog *prog,
477 				     struct bpf_prog *old)
478 {
479 	if (cmpxchg(pprog, old, prog) != old)
480 		return -ENOENT;
481 
482 	if (old)
483 		bpf_prog_put(old);
484 
485 	return 0;
486 }
487 
488 static inline void psock_progs_drop(struct sk_psock_progs *progs)
489 {
490 	psock_set_prog(&progs->msg_parser, NULL);
491 	psock_set_prog(&progs->stream_parser, NULL);
492 	psock_set_prog(&progs->stream_verdict, NULL);
493 	psock_set_prog(&progs->skb_verdict, NULL);
494 }
495 
496 int sk_psock_tls_strp_read(struct sk_psock *psock, struct sk_buff *skb);
497 
498 static inline bool sk_psock_strp_enabled(struct sk_psock *psock)
499 {
500 	if (!psock)
501 		return false;
502 	return !!psock->saved_data_ready;
503 }
504 
505 static inline bool sk_is_udp(const struct sock *sk)
506 {
507 	return sk->sk_type == SOCK_DGRAM &&
508 	       sk->sk_protocol == IPPROTO_UDP;
509 }
510 
511 #if IS_ENABLED(CONFIG_NET_SOCK_MSG)
512 
513 #define BPF_F_STRPARSER	(1UL << 1)
514 
515 /* We only have two bits so far. */
516 #define BPF_F_PTR_MASK ~(BPF_F_INGRESS | BPF_F_STRPARSER)
517 
518 static inline bool skb_bpf_strparser(const struct sk_buff *skb)
519 {
520 	unsigned long sk_redir = skb->_sk_redir;
521 
522 	return sk_redir & BPF_F_STRPARSER;
523 }
524 
525 static inline void skb_bpf_set_strparser(struct sk_buff *skb)
526 {
527 	skb->_sk_redir |= BPF_F_STRPARSER;
528 }
529 
530 static inline bool skb_bpf_ingress(const struct sk_buff *skb)
531 {
532 	unsigned long sk_redir = skb->_sk_redir;
533 
534 	return sk_redir & BPF_F_INGRESS;
535 }
536 
537 static inline void skb_bpf_set_ingress(struct sk_buff *skb)
538 {
539 	skb->_sk_redir |= BPF_F_INGRESS;
540 }
541 
542 static inline void skb_bpf_set_redir(struct sk_buff *skb, struct sock *sk_redir,
543 				     bool ingress)
544 {
545 	skb->_sk_redir = (unsigned long)sk_redir;
546 	if (ingress)
547 		skb->_sk_redir |= BPF_F_INGRESS;
548 }
549 
550 static inline struct sock *skb_bpf_redirect_fetch(const struct sk_buff *skb)
551 {
552 	unsigned long sk_redir = skb->_sk_redir;
553 
554 	return (struct sock *)(sk_redir & BPF_F_PTR_MASK);
555 }
556 
557 static inline void skb_bpf_redirect_clear(struct sk_buff *skb)
558 {
559 	skb->_sk_redir = 0;
560 }
561 #endif /* CONFIG_NET_SOCK_MSG */
562 #endif /* _LINUX_SKMSG_H */
563