xref: /openbmc/linux/net/strparser/strparser.c (revision ba61bb17)
1 /*
2  * Stream Parser
3  *
4  * Copyright (c) 2016 Tom Herbert <tom@herbertland.com>
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2
8  * as published by the Free Software Foundation.
9  */
10 
11 #include <linux/bpf.h>
12 #include <linux/errno.h>
13 #include <linux/errqueue.h>
14 #include <linux/file.h>
15 #include <linux/in.h>
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/net.h>
19 #include <linux/netdevice.h>
20 #include <linux/poll.h>
21 #include <linux/rculist.h>
22 #include <linux/skbuff.h>
23 #include <linux/socket.h>
24 #include <linux/uaccess.h>
25 #include <linux/workqueue.h>
26 #include <net/strparser.h>
27 #include <net/netns/generic.h>
28 #include <net/sock.h>
29 
30 static struct workqueue_struct *strp_wq;
31 
32 struct _strp_msg {
33 	/* Internal cb structure. struct strp_msg must be first for passing
34 	 * to upper layer.
35 	 */
36 	struct strp_msg strp;
37 	int accum_len;
38 	int early_eaten;
39 };
40 
41 static inline struct _strp_msg *_strp_msg(struct sk_buff *skb)
42 {
43 	return (struct _strp_msg *)((void *)skb->cb +
44 		offsetof(struct qdisc_skb_cb, data));
45 }
46 
47 /* Lower lock held */
48 static void strp_abort_strp(struct strparser *strp, int err)
49 {
50 	/* Unrecoverable error in receive */
51 
52 	cancel_delayed_work(&strp->msg_timer_work);
53 
54 	if (strp->stopped)
55 		return;
56 
57 	strp->stopped = 1;
58 
59 	if (strp->sk) {
60 		struct sock *sk = strp->sk;
61 
62 		/* Report an error on the lower socket */
63 		sk->sk_err = -err;
64 		sk->sk_error_report(sk);
65 	}
66 }
67 
68 static void strp_start_timer(struct strparser *strp, long timeo)
69 {
70 	if (timeo && timeo != LONG_MAX)
71 		mod_delayed_work(strp_wq, &strp->msg_timer_work, timeo);
72 }
73 
74 /* Lower lock held */
75 static void strp_parser_err(struct strparser *strp, int err,
76 			    read_descriptor_t *desc)
77 {
78 	desc->error = err;
79 	kfree_skb(strp->skb_head);
80 	strp->skb_head = NULL;
81 	strp->cb.abort_parser(strp, err);
82 }
83 
84 static inline int strp_peek_len(struct strparser *strp)
85 {
86 	if (strp->sk) {
87 		struct socket *sock = strp->sk->sk_socket;
88 
89 		return sock->ops->peek_len(sock);
90 	}
91 
92 	/* If we don't have an associated socket there's nothing to peek.
93 	 * Return int max to avoid stopping the strparser.
94 	 */
95 
96 	return INT_MAX;
97 }
98 
99 /* Lower socket lock held */
100 static int __strp_recv(read_descriptor_t *desc, struct sk_buff *orig_skb,
101 		       unsigned int orig_offset, size_t orig_len,
102 		       size_t max_msg_size, long timeo)
103 {
104 	struct strparser *strp = (struct strparser *)desc->arg.data;
105 	struct _strp_msg *stm;
106 	struct sk_buff *head, *skb;
107 	size_t eaten = 0, cand_len;
108 	ssize_t extra;
109 	int err;
110 	bool cloned_orig = false;
111 
112 	if (strp->paused)
113 		return 0;
114 
115 	head = strp->skb_head;
116 	if (head) {
117 		/* Message already in progress */
118 
119 		stm = _strp_msg(head);
120 		if (unlikely(stm->early_eaten)) {
121 			/* Already some number of bytes on the receive sock
122 			 * data saved in skb_head, just indicate they
123 			 * are consumed.
124 			 */
125 			eaten = orig_len <= stm->early_eaten ?
126 				orig_len : stm->early_eaten;
127 			stm->early_eaten -= eaten;
128 
129 			return eaten;
130 		}
131 
132 		if (unlikely(orig_offset)) {
133 			/* Getting data with a non-zero offset when a message is
134 			 * in progress is not expected. If it does happen, we
135 			 * need to clone and pull since we can't deal with
136 			 * offsets in the skbs for a message expect in the head.
137 			 */
138 			orig_skb = skb_clone(orig_skb, GFP_ATOMIC);
139 			if (!orig_skb) {
140 				STRP_STATS_INCR(strp->stats.mem_fail);
141 				desc->error = -ENOMEM;
142 				return 0;
143 			}
144 			if (!pskb_pull(orig_skb, orig_offset)) {
145 				STRP_STATS_INCR(strp->stats.mem_fail);
146 				kfree_skb(orig_skb);
147 				desc->error = -ENOMEM;
148 				return 0;
149 			}
150 			cloned_orig = true;
151 			orig_offset = 0;
152 		}
153 
154 		if (!strp->skb_nextp) {
155 			/* We are going to append to the frags_list of head.
156 			 * Need to unshare the frag_list.
157 			 */
158 			err = skb_unclone(head, GFP_ATOMIC);
159 			if (err) {
160 				STRP_STATS_INCR(strp->stats.mem_fail);
161 				desc->error = err;
162 				return 0;
163 			}
164 
165 			if (unlikely(skb_shinfo(head)->frag_list)) {
166 				/* We can't append to an sk_buff that already
167 				 * has a frag_list. We create a new head, point
168 				 * the frag_list of that to the old head, and
169 				 * then are able to use the old head->next for
170 				 * appending to the message.
171 				 */
172 				if (WARN_ON(head->next)) {
173 					desc->error = -EINVAL;
174 					return 0;
175 				}
176 
177 				skb = alloc_skb(0, GFP_ATOMIC);
178 				if (!skb) {
179 					STRP_STATS_INCR(strp->stats.mem_fail);
180 					desc->error = -ENOMEM;
181 					return 0;
182 				}
183 				skb->len = head->len;
184 				skb->data_len = head->len;
185 				skb->truesize = head->truesize;
186 				*_strp_msg(skb) = *_strp_msg(head);
187 				strp->skb_nextp = &head->next;
188 				skb_shinfo(skb)->frag_list = head;
189 				strp->skb_head = skb;
190 				head = skb;
191 			} else {
192 				strp->skb_nextp =
193 				    &skb_shinfo(head)->frag_list;
194 			}
195 		}
196 	}
197 
198 	while (eaten < orig_len) {
199 		/* Always clone since we will consume something */
200 		skb = skb_clone(orig_skb, GFP_ATOMIC);
201 		if (!skb) {
202 			STRP_STATS_INCR(strp->stats.mem_fail);
203 			desc->error = -ENOMEM;
204 			break;
205 		}
206 
207 		cand_len = orig_len - eaten;
208 
209 		head = strp->skb_head;
210 		if (!head) {
211 			head = skb;
212 			strp->skb_head = head;
213 			/* Will set skb_nextp on next packet if needed */
214 			strp->skb_nextp = NULL;
215 			stm = _strp_msg(head);
216 			memset(stm, 0, sizeof(*stm));
217 			stm->strp.offset = orig_offset + eaten;
218 		} else {
219 			/* Unclone since we may be appending to an skb that we
220 			 * already share a frag_list with.
221 			 */
222 			err = skb_unclone(skb, GFP_ATOMIC);
223 			if (err) {
224 				STRP_STATS_INCR(strp->stats.mem_fail);
225 				desc->error = err;
226 				break;
227 			}
228 
229 			stm = _strp_msg(head);
230 			*strp->skb_nextp = skb;
231 			strp->skb_nextp = &skb->next;
232 			head->data_len += skb->len;
233 			head->len += skb->len;
234 			head->truesize += skb->truesize;
235 		}
236 
237 		if (!stm->strp.full_len) {
238 			ssize_t len;
239 
240 			len = (*strp->cb.parse_msg)(strp, head);
241 
242 			if (!len) {
243 				/* Need more header to determine length */
244 				if (!stm->accum_len) {
245 					/* Start RX timer for new message */
246 					strp_start_timer(strp, timeo);
247 				}
248 				stm->accum_len += cand_len;
249 				eaten += cand_len;
250 				STRP_STATS_INCR(strp->stats.need_more_hdr);
251 				WARN_ON(eaten != orig_len);
252 				break;
253 			} else if (len < 0) {
254 				if (len == -ESTRPIPE && stm->accum_len) {
255 					len = -ENODATA;
256 					strp->unrecov_intr = 1;
257 				} else {
258 					strp->interrupted = 1;
259 				}
260 				strp_parser_err(strp, len, desc);
261 				break;
262 			} else if (len > max_msg_size) {
263 				/* Message length exceeds maximum allowed */
264 				STRP_STATS_INCR(strp->stats.msg_too_big);
265 				strp_parser_err(strp, -EMSGSIZE, desc);
266 				break;
267 			} else if (len <= (ssize_t)head->len -
268 					  skb->len - stm->strp.offset) {
269 				/* Length must be into new skb (and also
270 				 * greater than zero)
271 				 */
272 				STRP_STATS_INCR(strp->stats.bad_hdr_len);
273 				strp_parser_err(strp, -EPROTO, desc);
274 				break;
275 			}
276 
277 			stm->strp.full_len = len;
278 		}
279 
280 		extra = (ssize_t)(stm->accum_len + cand_len) -
281 			stm->strp.full_len;
282 
283 		if (extra < 0) {
284 			/* Message not complete yet. */
285 			if (stm->strp.full_len - stm->accum_len >
286 			    strp_peek_len(strp)) {
287 				/* Don't have the whole message in the socket
288 				 * buffer. Set strp->need_bytes to wait for
289 				 * the rest of the message. Also, set "early
290 				 * eaten" since we've already buffered the skb
291 				 * but don't consume yet per strp_read_sock.
292 				 */
293 
294 				if (!stm->accum_len) {
295 					/* Start RX timer for new message */
296 					strp_start_timer(strp, timeo);
297 				}
298 
299 				stm->accum_len += cand_len;
300 				strp->need_bytes = stm->strp.full_len -
301 						       stm->accum_len;
302 				stm->early_eaten = cand_len;
303 				STRP_STATS_ADD(strp->stats.bytes, cand_len);
304 				desc->count = 0; /* Stop reading socket */
305 				break;
306 			}
307 			stm->accum_len += cand_len;
308 			eaten += cand_len;
309 			WARN_ON(eaten != orig_len);
310 			break;
311 		}
312 
313 		/* Positive extra indicates ore bytes than needed for the
314 		 * message
315 		 */
316 
317 		WARN_ON(extra > cand_len);
318 
319 		eaten += (cand_len - extra);
320 
321 		/* Hurray, we have a new message! */
322 		cancel_delayed_work(&strp->msg_timer_work);
323 		strp->skb_head = NULL;
324 		strp->need_bytes = 0;
325 		STRP_STATS_INCR(strp->stats.msgs);
326 
327 		/* Give skb to upper layer */
328 		strp->cb.rcv_msg(strp, head);
329 
330 		if (unlikely(strp->paused)) {
331 			/* Upper layer paused strp */
332 			break;
333 		}
334 	}
335 
336 	if (cloned_orig)
337 		kfree_skb(orig_skb);
338 
339 	STRP_STATS_ADD(strp->stats.bytes, eaten);
340 
341 	return eaten;
342 }
343 
344 int strp_process(struct strparser *strp, struct sk_buff *orig_skb,
345 		 unsigned int orig_offset, size_t orig_len,
346 		 size_t max_msg_size, long timeo)
347 {
348 	read_descriptor_t desc; /* Dummy arg to strp_recv */
349 
350 	desc.arg.data = strp;
351 
352 	return __strp_recv(&desc, orig_skb, orig_offset, orig_len,
353 			   max_msg_size, timeo);
354 }
355 EXPORT_SYMBOL_GPL(strp_process);
356 
357 static int strp_recv(read_descriptor_t *desc, struct sk_buff *orig_skb,
358 		     unsigned int orig_offset, size_t orig_len)
359 {
360 	struct strparser *strp = (struct strparser *)desc->arg.data;
361 
362 	return __strp_recv(desc, orig_skb, orig_offset, orig_len,
363 			   strp->sk->sk_rcvbuf, strp->sk->sk_rcvtimeo);
364 }
365 
366 static int default_read_sock_done(struct strparser *strp, int err)
367 {
368 	return err;
369 }
370 
371 /* Called with lock held on lower socket */
372 static int strp_read_sock(struct strparser *strp)
373 {
374 	struct socket *sock = strp->sk->sk_socket;
375 	read_descriptor_t desc;
376 
377 	if (unlikely(!sock || !sock->ops || !sock->ops->read_sock))
378 		return -EBUSY;
379 
380 	desc.arg.data = strp;
381 	desc.error = 0;
382 	desc.count = 1; /* give more than one skb per call */
383 
384 	/* sk should be locked here, so okay to do read_sock */
385 	sock->ops->read_sock(strp->sk, &desc, strp_recv);
386 
387 	desc.error = strp->cb.read_sock_done(strp, desc.error);
388 
389 	return desc.error;
390 }
391 
392 /* Lower sock lock held */
393 void strp_data_ready(struct strparser *strp)
394 {
395 	if (unlikely(strp->stopped) || strp->paused)
396 		return;
397 
398 	/* This check is needed to synchronize with do_strp_work.
399 	 * do_strp_work acquires a process lock (lock_sock) whereas
400 	 * the lock held here is bh_lock_sock. The two locks can be
401 	 * held by different threads at the same time, but bh_lock_sock
402 	 * allows a thread in BH context to safely check if the process
403 	 * lock is held. In this case, if the lock is held, queue work.
404 	 */
405 	if (sock_owned_by_user_nocheck(strp->sk)) {
406 		queue_work(strp_wq, &strp->work);
407 		return;
408 	}
409 
410 	if (strp->need_bytes) {
411 		if (strp_peek_len(strp) < strp->need_bytes)
412 			return;
413 	}
414 
415 	if (strp_read_sock(strp) == -ENOMEM)
416 		queue_work(strp_wq, &strp->work);
417 }
418 EXPORT_SYMBOL_GPL(strp_data_ready);
419 
420 static void do_strp_work(struct strparser *strp)
421 {
422 	read_descriptor_t rd_desc;
423 
424 	/* We need the read lock to synchronize with strp_data_ready. We
425 	 * need the socket lock for calling strp_read_sock.
426 	 */
427 	strp->cb.lock(strp);
428 
429 	if (unlikely(strp->stopped))
430 		goto out;
431 
432 	if (strp->paused)
433 		goto out;
434 
435 	rd_desc.arg.data = strp;
436 
437 	if (strp_read_sock(strp) == -ENOMEM)
438 		queue_work(strp_wq, &strp->work);
439 
440 out:
441 	strp->cb.unlock(strp);
442 }
443 
444 static void strp_work(struct work_struct *w)
445 {
446 	do_strp_work(container_of(w, struct strparser, work));
447 }
448 
449 static void strp_msg_timeout(struct work_struct *w)
450 {
451 	struct strparser *strp = container_of(w, struct strparser,
452 					      msg_timer_work.work);
453 
454 	/* Message assembly timed out */
455 	STRP_STATS_INCR(strp->stats.msg_timeouts);
456 	strp->cb.lock(strp);
457 	strp->cb.abort_parser(strp, -ETIMEDOUT);
458 	strp->cb.unlock(strp);
459 }
460 
461 static void strp_sock_lock(struct strparser *strp)
462 {
463 	lock_sock(strp->sk);
464 }
465 
466 static void strp_sock_unlock(struct strparser *strp)
467 {
468 	release_sock(strp->sk);
469 }
470 
471 int strp_init(struct strparser *strp, struct sock *sk,
472 	      const struct strp_callbacks *cb)
473 {
474 
475 	if (!cb || !cb->rcv_msg || !cb->parse_msg)
476 		return -EINVAL;
477 
478 	/* The sk (sock) arg determines the mode of the stream parser.
479 	 *
480 	 * If the sock is set then the strparser is in receive callback mode.
481 	 * The upper layer calls strp_data_ready to kick receive processing
482 	 * and strparser calls the read_sock function on the socket to
483 	 * get packets.
484 	 *
485 	 * If the sock is not set then the strparser is in general mode.
486 	 * The upper layer calls strp_process for each skb to be parsed.
487 	 */
488 
489 	if (!sk) {
490 		if (!cb->lock || !cb->unlock)
491 			return -EINVAL;
492 	}
493 
494 	memset(strp, 0, sizeof(*strp));
495 
496 	strp->sk = sk;
497 
498 	strp->cb.lock = cb->lock ? : strp_sock_lock;
499 	strp->cb.unlock = cb->unlock ? : strp_sock_unlock;
500 	strp->cb.rcv_msg = cb->rcv_msg;
501 	strp->cb.parse_msg = cb->parse_msg;
502 	strp->cb.read_sock_done = cb->read_sock_done ? : default_read_sock_done;
503 	strp->cb.abort_parser = cb->abort_parser ? : strp_abort_strp;
504 
505 	INIT_DELAYED_WORK(&strp->msg_timer_work, strp_msg_timeout);
506 	INIT_WORK(&strp->work, strp_work);
507 
508 	return 0;
509 }
510 EXPORT_SYMBOL_GPL(strp_init);
511 
512 /* Sock process lock held (lock_sock) */
513 void __strp_unpause(struct strparser *strp)
514 {
515 	strp->paused = 0;
516 
517 	if (strp->need_bytes) {
518 		if (strp_peek_len(strp) < strp->need_bytes)
519 			return;
520 	}
521 	strp_read_sock(strp);
522 }
523 EXPORT_SYMBOL_GPL(__strp_unpause);
524 
525 void strp_unpause(struct strparser *strp)
526 {
527 	strp->paused = 0;
528 
529 	/* Sync setting paused with RX work */
530 	smp_mb();
531 
532 	queue_work(strp_wq, &strp->work);
533 }
534 EXPORT_SYMBOL_GPL(strp_unpause);
535 
536 /* strp must already be stopped so that strp_recv will no longer be called.
537  * Note that strp_done is not called with the lower socket held.
538  */
539 void strp_done(struct strparser *strp)
540 {
541 	WARN_ON(!strp->stopped);
542 
543 	cancel_delayed_work_sync(&strp->msg_timer_work);
544 	cancel_work_sync(&strp->work);
545 
546 	if (strp->skb_head) {
547 		kfree_skb(strp->skb_head);
548 		strp->skb_head = NULL;
549 	}
550 }
551 EXPORT_SYMBOL_GPL(strp_done);
552 
553 void strp_stop(struct strparser *strp)
554 {
555 	strp->stopped = 1;
556 }
557 EXPORT_SYMBOL_GPL(strp_stop);
558 
559 void strp_check_rcv(struct strparser *strp)
560 {
561 	queue_work(strp_wq, &strp->work);
562 }
563 EXPORT_SYMBOL_GPL(strp_check_rcv);
564 
565 static int __init strp_mod_init(void)
566 {
567 	strp_wq = create_singlethread_workqueue("kstrp");
568 
569 	return 0;
570 }
571 
572 static void __exit strp_mod_exit(void)
573 {
574 	destroy_workqueue(strp_wq);
575 }
576 module_init(strp_mod_init);
577 module_exit(strp_mod_exit);
578 MODULE_LICENSE("GPL");
579