xref: /openbmc/linux/net/strparser/strparser.c (revision 0c7beb2d)
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 };
39 
40 static inline struct _strp_msg *_strp_msg(struct sk_buff *skb)
41 {
42 	return (struct _strp_msg *)((void *)skb->cb +
43 		offsetof(struct qdisc_skb_cb, data));
44 }
45 
46 /* Lower lock held */
47 static void strp_abort_strp(struct strparser *strp, int err)
48 {
49 	/* Unrecoverable error in receive */
50 
51 	cancel_delayed_work(&strp->msg_timer_work);
52 
53 	if (strp->stopped)
54 		return;
55 
56 	strp->stopped = 1;
57 
58 	if (strp->sk) {
59 		struct sock *sk = strp->sk;
60 
61 		/* Report an error on the lower socket */
62 		sk->sk_err = -err;
63 		sk->sk_error_report(sk);
64 	}
65 }
66 
67 static void strp_start_timer(struct strparser *strp, long timeo)
68 {
69 	if (timeo && timeo != LONG_MAX)
70 		mod_delayed_work(strp_wq, &strp->msg_timer_work, timeo);
71 }
72 
73 /* Lower lock held */
74 static void strp_parser_err(struct strparser *strp, int err,
75 			    read_descriptor_t *desc)
76 {
77 	desc->error = err;
78 	kfree_skb(strp->skb_head);
79 	strp->skb_head = NULL;
80 	strp->cb.abort_parser(strp, err);
81 }
82 
83 static inline int strp_peek_len(struct strparser *strp)
84 {
85 	if (strp->sk) {
86 		struct socket *sock = strp->sk->sk_socket;
87 
88 		return sock->ops->peek_len(sock);
89 	}
90 
91 	/* If we don't have an associated socket there's nothing to peek.
92 	 * Return int max to avoid stopping the strparser.
93 	 */
94 
95 	return INT_MAX;
96 }
97 
98 /* Lower socket lock held */
99 static int __strp_recv(read_descriptor_t *desc, struct sk_buff *orig_skb,
100 		       unsigned int orig_offset, size_t orig_len,
101 		       size_t max_msg_size, long timeo)
102 {
103 	struct strparser *strp = (struct strparser *)desc->arg.data;
104 	struct _strp_msg *stm;
105 	struct sk_buff *head, *skb;
106 	size_t eaten = 0, cand_len;
107 	ssize_t extra;
108 	int err;
109 	bool cloned_orig = false;
110 
111 	if (strp->paused)
112 		return 0;
113 
114 	head = strp->skb_head;
115 	if (head) {
116 		/* Message already in progress */
117 		if (unlikely(orig_offset)) {
118 			/* Getting data with a non-zero offset when a message is
119 			 * in progress is not expected. If it does happen, we
120 			 * need to clone and pull since we can't deal with
121 			 * offsets in the skbs for a message expect in the head.
122 			 */
123 			orig_skb = skb_clone(orig_skb, GFP_ATOMIC);
124 			if (!orig_skb) {
125 				STRP_STATS_INCR(strp->stats.mem_fail);
126 				desc->error = -ENOMEM;
127 				return 0;
128 			}
129 			if (!pskb_pull(orig_skb, orig_offset)) {
130 				STRP_STATS_INCR(strp->stats.mem_fail);
131 				kfree_skb(orig_skb);
132 				desc->error = -ENOMEM;
133 				return 0;
134 			}
135 			cloned_orig = true;
136 			orig_offset = 0;
137 		}
138 
139 		if (!strp->skb_nextp) {
140 			/* We are going to append to the frags_list of head.
141 			 * Need to unshare the frag_list.
142 			 */
143 			err = skb_unclone(head, GFP_ATOMIC);
144 			if (err) {
145 				STRP_STATS_INCR(strp->stats.mem_fail);
146 				desc->error = err;
147 				return 0;
148 			}
149 
150 			if (unlikely(skb_shinfo(head)->frag_list)) {
151 				/* We can't append to an sk_buff that already
152 				 * has a frag_list. We create a new head, point
153 				 * the frag_list of that to the old head, and
154 				 * then are able to use the old head->next for
155 				 * appending to the message.
156 				 */
157 				if (WARN_ON(head->next)) {
158 					desc->error = -EINVAL;
159 					return 0;
160 				}
161 
162 				skb = alloc_skb(0, GFP_ATOMIC);
163 				if (!skb) {
164 					STRP_STATS_INCR(strp->stats.mem_fail);
165 					desc->error = -ENOMEM;
166 					return 0;
167 				}
168 				skb->len = head->len;
169 				skb->data_len = head->len;
170 				skb->truesize = head->truesize;
171 				*_strp_msg(skb) = *_strp_msg(head);
172 				strp->skb_nextp = &head->next;
173 				skb_shinfo(skb)->frag_list = head;
174 				strp->skb_head = skb;
175 				head = skb;
176 			} else {
177 				strp->skb_nextp =
178 				    &skb_shinfo(head)->frag_list;
179 			}
180 		}
181 	}
182 
183 	while (eaten < orig_len) {
184 		/* Always clone since we will consume something */
185 		skb = skb_clone(orig_skb, GFP_ATOMIC);
186 		if (!skb) {
187 			STRP_STATS_INCR(strp->stats.mem_fail);
188 			desc->error = -ENOMEM;
189 			break;
190 		}
191 
192 		cand_len = orig_len - eaten;
193 
194 		head = strp->skb_head;
195 		if (!head) {
196 			head = skb;
197 			strp->skb_head = head;
198 			/* Will set skb_nextp on next packet if needed */
199 			strp->skb_nextp = NULL;
200 			stm = _strp_msg(head);
201 			memset(stm, 0, sizeof(*stm));
202 			stm->strp.offset = orig_offset + eaten;
203 		} else {
204 			/* Unclone if we are appending to an skb that we
205 			 * already share a frag_list with.
206 			 */
207 			if (skb_has_frag_list(skb)) {
208 				err = skb_unclone(skb, GFP_ATOMIC);
209 				if (err) {
210 					STRP_STATS_INCR(strp->stats.mem_fail);
211 					desc->error = err;
212 					break;
213 				}
214 			}
215 
216 			stm = _strp_msg(head);
217 			*strp->skb_nextp = skb;
218 			strp->skb_nextp = &skb->next;
219 			head->data_len += skb->len;
220 			head->len += skb->len;
221 			head->truesize += skb->truesize;
222 		}
223 
224 		if (!stm->strp.full_len) {
225 			ssize_t len;
226 
227 			len = (*strp->cb.parse_msg)(strp, head);
228 
229 			if (!len) {
230 				/* Need more header to determine length */
231 				if (!stm->accum_len) {
232 					/* Start RX timer for new message */
233 					strp_start_timer(strp, timeo);
234 				}
235 				stm->accum_len += cand_len;
236 				eaten += cand_len;
237 				STRP_STATS_INCR(strp->stats.need_more_hdr);
238 				WARN_ON(eaten != orig_len);
239 				break;
240 			} else if (len < 0) {
241 				if (len == -ESTRPIPE && stm->accum_len) {
242 					len = -ENODATA;
243 					strp->unrecov_intr = 1;
244 				} else {
245 					strp->interrupted = 1;
246 				}
247 				strp_parser_err(strp, len, desc);
248 				break;
249 			} else if (len > max_msg_size) {
250 				/* Message length exceeds maximum allowed */
251 				STRP_STATS_INCR(strp->stats.msg_too_big);
252 				strp_parser_err(strp, -EMSGSIZE, desc);
253 				break;
254 			} else if (len <= (ssize_t)head->len -
255 					  skb->len - stm->strp.offset) {
256 				/* Length must be into new skb (and also
257 				 * greater than zero)
258 				 */
259 				STRP_STATS_INCR(strp->stats.bad_hdr_len);
260 				strp_parser_err(strp, -EPROTO, desc);
261 				break;
262 			}
263 
264 			stm->strp.full_len = len;
265 		}
266 
267 		extra = (ssize_t)(stm->accum_len + cand_len) -
268 			stm->strp.full_len;
269 
270 		if (extra < 0) {
271 			/* Message not complete yet. */
272 			if (stm->strp.full_len - stm->accum_len >
273 			    strp_peek_len(strp)) {
274 				/* Don't have the whole message in the socket
275 				 * buffer. Set strp->need_bytes to wait for
276 				 * the rest of the message. Also, set "early
277 				 * eaten" since we've already buffered the skb
278 				 * but don't consume yet per strp_read_sock.
279 				 */
280 
281 				if (!stm->accum_len) {
282 					/* Start RX timer for new message */
283 					strp_start_timer(strp, timeo);
284 				}
285 
286 				stm->accum_len += cand_len;
287 				eaten += cand_len;
288 				strp->need_bytes = stm->strp.full_len -
289 						       stm->accum_len;
290 				STRP_STATS_ADD(strp->stats.bytes, cand_len);
291 				desc->count = 0; /* Stop reading socket */
292 				break;
293 			}
294 			stm->accum_len += cand_len;
295 			eaten += cand_len;
296 			WARN_ON(eaten != orig_len);
297 			break;
298 		}
299 
300 		/* Positive extra indicates more bytes than needed for the
301 		 * message
302 		 */
303 
304 		WARN_ON(extra > cand_len);
305 
306 		eaten += (cand_len - extra);
307 
308 		/* Hurray, we have a new message! */
309 		cancel_delayed_work(&strp->msg_timer_work);
310 		strp->skb_head = NULL;
311 		strp->need_bytes = 0;
312 		STRP_STATS_INCR(strp->stats.msgs);
313 
314 		/* Give skb to upper layer */
315 		strp->cb.rcv_msg(strp, head);
316 
317 		if (unlikely(strp->paused)) {
318 			/* Upper layer paused strp */
319 			break;
320 		}
321 	}
322 
323 	if (cloned_orig)
324 		kfree_skb(orig_skb);
325 
326 	STRP_STATS_ADD(strp->stats.bytes, eaten);
327 
328 	return eaten;
329 }
330 
331 int strp_process(struct strparser *strp, struct sk_buff *orig_skb,
332 		 unsigned int orig_offset, size_t orig_len,
333 		 size_t max_msg_size, long timeo)
334 {
335 	read_descriptor_t desc; /* Dummy arg to strp_recv */
336 
337 	desc.arg.data = strp;
338 
339 	return __strp_recv(&desc, orig_skb, orig_offset, orig_len,
340 			   max_msg_size, timeo);
341 }
342 EXPORT_SYMBOL_GPL(strp_process);
343 
344 static int strp_recv(read_descriptor_t *desc, struct sk_buff *orig_skb,
345 		     unsigned int orig_offset, size_t orig_len)
346 {
347 	struct strparser *strp = (struct strparser *)desc->arg.data;
348 
349 	return __strp_recv(desc, orig_skb, orig_offset, orig_len,
350 			   strp->sk->sk_rcvbuf, strp->sk->sk_rcvtimeo);
351 }
352 
353 static int default_read_sock_done(struct strparser *strp, int err)
354 {
355 	return err;
356 }
357 
358 /* Called with lock held on lower socket */
359 static int strp_read_sock(struct strparser *strp)
360 {
361 	struct socket *sock = strp->sk->sk_socket;
362 	read_descriptor_t desc;
363 
364 	if (unlikely(!sock || !sock->ops || !sock->ops->read_sock))
365 		return -EBUSY;
366 
367 	desc.arg.data = strp;
368 	desc.error = 0;
369 	desc.count = 1; /* give more than one skb per call */
370 
371 	/* sk should be locked here, so okay to do read_sock */
372 	sock->ops->read_sock(strp->sk, &desc, strp_recv);
373 
374 	desc.error = strp->cb.read_sock_done(strp, desc.error);
375 
376 	return desc.error;
377 }
378 
379 /* Lower sock lock held */
380 void strp_data_ready(struct strparser *strp)
381 {
382 	if (unlikely(strp->stopped) || strp->paused)
383 		return;
384 
385 	/* This check is needed to synchronize with do_strp_work.
386 	 * do_strp_work acquires a process lock (lock_sock) whereas
387 	 * the lock held here is bh_lock_sock. The two locks can be
388 	 * held by different threads at the same time, but bh_lock_sock
389 	 * allows a thread in BH context to safely check if the process
390 	 * lock is held. In this case, if the lock is held, queue work.
391 	 */
392 	if (sock_owned_by_user_nocheck(strp->sk)) {
393 		queue_work(strp_wq, &strp->work);
394 		return;
395 	}
396 
397 	if (strp->need_bytes) {
398 		if (strp_peek_len(strp) < strp->need_bytes)
399 			return;
400 	}
401 
402 	if (strp_read_sock(strp) == -ENOMEM)
403 		queue_work(strp_wq, &strp->work);
404 }
405 EXPORT_SYMBOL_GPL(strp_data_ready);
406 
407 static void do_strp_work(struct strparser *strp)
408 {
409 	/* We need the read lock to synchronize with strp_data_ready. We
410 	 * need the socket lock for calling strp_read_sock.
411 	 */
412 	strp->cb.lock(strp);
413 
414 	if (unlikely(strp->stopped))
415 		goto out;
416 
417 	if (strp->paused)
418 		goto out;
419 
420 	if (strp_read_sock(strp) == -ENOMEM)
421 		queue_work(strp_wq, &strp->work);
422 
423 out:
424 	strp->cb.unlock(strp);
425 }
426 
427 static void strp_work(struct work_struct *w)
428 {
429 	do_strp_work(container_of(w, struct strparser, work));
430 }
431 
432 static void strp_msg_timeout(struct work_struct *w)
433 {
434 	struct strparser *strp = container_of(w, struct strparser,
435 					      msg_timer_work.work);
436 
437 	/* Message assembly timed out */
438 	STRP_STATS_INCR(strp->stats.msg_timeouts);
439 	strp->cb.lock(strp);
440 	strp->cb.abort_parser(strp, -ETIMEDOUT);
441 	strp->cb.unlock(strp);
442 }
443 
444 static void strp_sock_lock(struct strparser *strp)
445 {
446 	lock_sock(strp->sk);
447 }
448 
449 static void strp_sock_unlock(struct strparser *strp)
450 {
451 	release_sock(strp->sk);
452 }
453 
454 int strp_init(struct strparser *strp, struct sock *sk,
455 	      const struct strp_callbacks *cb)
456 {
457 
458 	if (!cb || !cb->rcv_msg || !cb->parse_msg)
459 		return -EINVAL;
460 
461 	/* The sk (sock) arg determines the mode of the stream parser.
462 	 *
463 	 * If the sock is set then the strparser is in receive callback mode.
464 	 * The upper layer calls strp_data_ready to kick receive processing
465 	 * and strparser calls the read_sock function on the socket to
466 	 * get packets.
467 	 *
468 	 * If the sock is not set then the strparser is in general mode.
469 	 * The upper layer calls strp_process for each skb to be parsed.
470 	 */
471 
472 	if (!sk) {
473 		if (!cb->lock || !cb->unlock)
474 			return -EINVAL;
475 	}
476 
477 	memset(strp, 0, sizeof(*strp));
478 
479 	strp->sk = sk;
480 
481 	strp->cb.lock = cb->lock ? : strp_sock_lock;
482 	strp->cb.unlock = cb->unlock ? : strp_sock_unlock;
483 	strp->cb.rcv_msg = cb->rcv_msg;
484 	strp->cb.parse_msg = cb->parse_msg;
485 	strp->cb.read_sock_done = cb->read_sock_done ? : default_read_sock_done;
486 	strp->cb.abort_parser = cb->abort_parser ? : strp_abort_strp;
487 
488 	INIT_DELAYED_WORK(&strp->msg_timer_work, strp_msg_timeout);
489 	INIT_WORK(&strp->work, strp_work);
490 
491 	return 0;
492 }
493 EXPORT_SYMBOL_GPL(strp_init);
494 
495 /* Sock process lock held (lock_sock) */
496 void __strp_unpause(struct strparser *strp)
497 {
498 	strp->paused = 0;
499 
500 	if (strp->need_bytes) {
501 		if (strp_peek_len(strp) < strp->need_bytes)
502 			return;
503 	}
504 	strp_read_sock(strp);
505 }
506 EXPORT_SYMBOL_GPL(__strp_unpause);
507 
508 void strp_unpause(struct strparser *strp)
509 {
510 	strp->paused = 0;
511 
512 	/* Sync setting paused with RX work */
513 	smp_mb();
514 
515 	queue_work(strp_wq, &strp->work);
516 }
517 EXPORT_SYMBOL_GPL(strp_unpause);
518 
519 /* strp must already be stopped so that strp_recv will no longer be called.
520  * Note that strp_done is not called with the lower socket held.
521  */
522 void strp_done(struct strparser *strp)
523 {
524 	WARN_ON(!strp->stopped);
525 
526 	cancel_delayed_work_sync(&strp->msg_timer_work);
527 	cancel_work_sync(&strp->work);
528 
529 	if (strp->skb_head) {
530 		kfree_skb(strp->skb_head);
531 		strp->skb_head = NULL;
532 	}
533 }
534 EXPORT_SYMBOL_GPL(strp_done);
535 
536 void strp_stop(struct strparser *strp)
537 {
538 	strp->stopped = 1;
539 }
540 EXPORT_SYMBOL_GPL(strp_stop);
541 
542 void strp_check_rcv(struct strparser *strp)
543 {
544 	queue_work(strp_wq, &strp->work);
545 }
546 EXPORT_SYMBOL_GPL(strp_check_rcv);
547 
548 static int __init strp_mod_init(void)
549 {
550 	strp_wq = create_singlethread_workqueue("kstrp");
551 	if (unlikely(!strp_wq))
552 		return -ENOMEM;
553 
554 	return 0;
555 }
556 
557 static void __exit strp_mod_exit(void)
558 {
559 	destroy_workqueue(strp_wq);
560 }
561 module_init(strp_mod_init);
562 module_exit(strp_mod_exit);
563 MODULE_LICENSE("GPL");
564