xref: /openbmc/linux/net/strparser/strparser.c (revision 96de2506)
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 			if (skb_has_frag_list(head)) {
144 				err = skb_unclone(head, GFP_ATOMIC);
145 				if (err) {
146 					STRP_STATS_INCR(strp->stats.mem_fail);
147 					desc->error = err;
148 					return 0;
149 				}
150 			}
151 
152 			if (unlikely(skb_shinfo(head)->frag_list)) {
153 				/* We can't append to an sk_buff that already
154 				 * has a frag_list. We create a new head, point
155 				 * the frag_list of that to the old head, and
156 				 * then are able to use the old head->next for
157 				 * appending to the message.
158 				 */
159 				if (WARN_ON(head->next)) {
160 					desc->error = -EINVAL;
161 					return 0;
162 				}
163 
164 				skb = alloc_skb(0, GFP_ATOMIC);
165 				if (!skb) {
166 					STRP_STATS_INCR(strp->stats.mem_fail);
167 					desc->error = -ENOMEM;
168 					return 0;
169 				}
170 				skb->len = head->len;
171 				skb->data_len = head->len;
172 				skb->truesize = head->truesize;
173 				*_strp_msg(skb) = *_strp_msg(head);
174 				strp->skb_nextp = &head->next;
175 				skb_shinfo(skb)->frag_list = head;
176 				strp->skb_head = skb;
177 				head = skb;
178 			} else {
179 				strp->skb_nextp =
180 				    &skb_shinfo(head)->frag_list;
181 			}
182 		}
183 	}
184 
185 	while (eaten < orig_len) {
186 		/* Always clone since we will consume something */
187 		skb = skb_clone(orig_skb, GFP_ATOMIC);
188 		if (!skb) {
189 			STRP_STATS_INCR(strp->stats.mem_fail);
190 			desc->error = -ENOMEM;
191 			break;
192 		}
193 
194 		cand_len = orig_len - eaten;
195 
196 		head = strp->skb_head;
197 		if (!head) {
198 			head = skb;
199 			strp->skb_head = head;
200 			/* Will set skb_nextp on next packet if needed */
201 			strp->skb_nextp = NULL;
202 			stm = _strp_msg(head);
203 			memset(stm, 0, sizeof(*stm));
204 			stm->strp.offset = orig_offset + eaten;
205 		} else {
206 			/* Unclone if we are appending to an skb that we
207 			 * already share a frag_list with.
208 			 */
209 			if (skb_has_frag_list(skb)) {
210 				err = skb_unclone(skb, GFP_ATOMIC);
211 				if (err) {
212 					STRP_STATS_INCR(strp->stats.mem_fail);
213 					desc->error = err;
214 					break;
215 				}
216 			}
217 
218 			stm = _strp_msg(head);
219 			*strp->skb_nextp = skb;
220 			strp->skb_nextp = &skb->next;
221 			head->data_len += skb->len;
222 			head->len += skb->len;
223 			head->truesize += skb->truesize;
224 		}
225 
226 		if (!stm->strp.full_len) {
227 			ssize_t len;
228 
229 			len = (*strp->cb.parse_msg)(strp, head);
230 
231 			if (!len) {
232 				/* Need more header to determine length */
233 				if (!stm->accum_len) {
234 					/* Start RX timer for new message */
235 					strp_start_timer(strp, timeo);
236 				}
237 				stm->accum_len += cand_len;
238 				eaten += cand_len;
239 				STRP_STATS_INCR(strp->stats.need_more_hdr);
240 				WARN_ON(eaten != orig_len);
241 				break;
242 			} else if (len < 0) {
243 				if (len == -ESTRPIPE && stm->accum_len) {
244 					len = -ENODATA;
245 					strp->unrecov_intr = 1;
246 				} else {
247 					strp->interrupted = 1;
248 				}
249 				strp_parser_err(strp, len, desc);
250 				break;
251 			} else if (len > max_msg_size) {
252 				/* Message length exceeds maximum allowed */
253 				STRP_STATS_INCR(strp->stats.msg_too_big);
254 				strp_parser_err(strp, -EMSGSIZE, desc);
255 				break;
256 			} else if (len <= (ssize_t)head->len -
257 					  skb->len - stm->strp.offset) {
258 				/* Length must be into new skb (and also
259 				 * greater than zero)
260 				 */
261 				STRP_STATS_INCR(strp->stats.bad_hdr_len);
262 				strp_parser_err(strp, -EPROTO, desc);
263 				break;
264 			}
265 
266 			stm->strp.full_len = len;
267 		}
268 
269 		extra = (ssize_t)(stm->accum_len + cand_len) -
270 			stm->strp.full_len;
271 
272 		if (extra < 0) {
273 			/* Message not complete yet. */
274 			if (stm->strp.full_len - stm->accum_len >
275 			    strp_peek_len(strp)) {
276 				/* Don't have the whole message in the socket
277 				 * buffer. Set strp->need_bytes to wait for
278 				 * the rest of the message. Also, set "early
279 				 * eaten" since we've already buffered the skb
280 				 * but don't consume yet per strp_read_sock.
281 				 */
282 
283 				if (!stm->accum_len) {
284 					/* Start RX timer for new message */
285 					strp_start_timer(strp, timeo);
286 				}
287 
288 				stm->accum_len += cand_len;
289 				eaten += cand_len;
290 				strp->need_bytes = stm->strp.full_len -
291 						       stm->accum_len;
292 				STRP_STATS_ADD(strp->stats.bytes, cand_len);
293 				desc->count = 0; /* Stop reading socket */
294 				break;
295 			}
296 			stm->accum_len += cand_len;
297 			eaten += cand_len;
298 			WARN_ON(eaten != orig_len);
299 			break;
300 		}
301 
302 		/* Positive extra indicates ore bytes than needed for the
303 		 * message
304 		 */
305 
306 		WARN_ON(extra > cand_len);
307 
308 		eaten += (cand_len - extra);
309 
310 		/* Hurray, we have a new message! */
311 		cancel_delayed_work(&strp->msg_timer_work);
312 		strp->skb_head = NULL;
313 		strp->need_bytes = 0;
314 		STRP_STATS_INCR(strp->stats.msgs);
315 
316 		/* Give skb to upper layer */
317 		strp->cb.rcv_msg(strp, head);
318 
319 		if (unlikely(strp->paused)) {
320 			/* Upper layer paused strp */
321 			break;
322 		}
323 	}
324 
325 	if (cloned_orig)
326 		kfree_skb(orig_skb);
327 
328 	STRP_STATS_ADD(strp->stats.bytes, eaten);
329 
330 	return eaten;
331 }
332 
333 int strp_process(struct strparser *strp, struct sk_buff *orig_skb,
334 		 unsigned int orig_offset, size_t orig_len,
335 		 size_t max_msg_size, long timeo)
336 {
337 	read_descriptor_t desc; /* Dummy arg to strp_recv */
338 
339 	desc.arg.data = strp;
340 
341 	return __strp_recv(&desc, orig_skb, orig_offset, orig_len,
342 			   max_msg_size, timeo);
343 }
344 EXPORT_SYMBOL_GPL(strp_process);
345 
346 static int strp_recv(read_descriptor_t *desc, struct sk_buff *orig_skb,
347 		     unsigned int orig_offset, size_t orig_len)
348 {
349 	struct strparser *strp = (struct strparser *)desc->arg.data;
350 
351 	return __strp_recv(desc, orig_skb, orig_offset, orig_len,
352 			   strp->sk->sk_rcvbuf, strp->sk->sk_rcvtimeo);
353 }
354 
355 static int default_read_sock_done(struct strparser *strp, int err)
356 {
357 	return err;
358 }
359 
360 /* Called with lock held on lower socket */
361 static int strp_read_sock(struct strparser *strp)
362 {
363 	struct socket *sock = strp->sk->sk_socket;
364 	read_descriptor_t desc;
365 
366 	if (unlikely(!sock || !sock->ops || !sock->ops->read_sock))
367 		return -EBUSY;
368 
369 	desc.arg.data = strp;
370 	desc.error = 0;
371 	desc.count = 1; /* give more than one skb per call */
372 
373 	/* sk should be locked here, so okay to do read_sock */
374 	sock->ops->read_sock(strp->sk, &desc, strp_recv);
375 
376 	desc.error = strp->cb.read_sock_done(strp, desc.error);
377 
378 	return desc.error;
379 }
380 
381 /* Lower sock lock held */
382 void strp_data_ready(struct strparser *strp)
383 {
384 	if (unlikely(strp->stopped) || strp->paused)
385 		return;
386 
387 	/* This check is needed to synchronize with do_strp_work.
388 	 * do_strp_work acquires a process lock (lock_sock) whereas
389 	 * the lock held here is bh_lock_sock. The two locks can be
390 	 * held by different threads at the same time, but bh_lock_sock
391 	 * allows a thread in BH context to safely check if the process
392 	 * lock is held. In this case, if the lock is held, queue work.
393 	 */
394 	if (sock_owned_by_user_nocheck(strp->sk)) {
395 		queue_work(strp_wq, &strp->work);
396 		return;
397 	}
398 
399 	if (strp->need_bytes) {
400 		if (strp_peek_len(strp) < strp->need_bytes)
401 			return;
402 	}
403 
404 	if (strp_read_sock(strp) == -ENOMEM)
405 		queue_work(strp_wq, &strp->work);
406 }
407 EXPORT_SYMBOL_GPL(strp_data_ready);
408 
409 static void do_strp_work(struct strparser *strp)
410 {
411 	/* We need the read lock to synchronize with strp_data_ready. We
412 	 * need the socket lock for calling strp_read_sock.
413 	 */
414 	strp->cb.lock(strp);
415 
416 	if (unlikely(strp->stopped))
417 		goto out;
418 
419 	if (strp->paused)
420 		goto out;
421 
422 	if (strp_read_sock(strp) == -ENOMEM)
423 		queue_work(strp_wq, &strp->work);
424 
425 out:
426 	strp->cb.unlock(strp);
427 }
428 
429 static void strp_work(struct work_struct *w)
430 {
431 	do_strp_work(container_of(w, struct strparser, work));
432 }
433 
434 static void strp_msg_timeout(struct work_struct *w)
435 {
436 	struct strparser *strp = container_of(w, struct strparser,
437 					      msg_timer_work.work);
438 
439 	/* Message assembly timed out */
440 	STRP_STATS_INCR(strp->stats.msg_timeouts);
441 	strp->cb.lock(strp);
442 	strp->cb.abort_parser(strp, -ETIMEDOUT);
443 	strp->cb.unlock(strp);
444 }
445 
446 static void strp_sock_lock(struct strparser *strp)
447 {
448 	lock_sock(strp->sk);
449 }
450 
451 static void strp_sock_unlock(struct strparser *strp)
452 {
453 	release_sock(strp->sk);
454 }
455 
456 int strp_init(struct strparser *strp, struct sock *sk,
457 	      const struct strp_callbacks *cb)
458 {
459 
460 	if (!cb || !cb->rcv_msg || !cb->parse_msg)
461 		return -EINVAL;
462 
463 	/* The sk (sock) arg determines the mode of the stream parser.
464 	 *
465 	 * If the sock is set then the strparser is in receive callback mode.
466 	 * The upper layer calls strp_data_ready to kick receive processing
467 	 * and strparser calls the read_sock function on the socket to
468 	 * get packets.
469 	 *
470 	 * If the sock is not set then the strparser is in general mode.
471 	 * The upper layer calls strp_process for each skb to be parsed.
472 	 */
473 
474 	if (!sk) {
475 		if (!cb->lock || !cb->unlock)
476 			return -EINVAL;
477 	}
478 
479 	memset(strp, 0, sizeof(*strp));
480 
481 	strp->sk = sk;
482 
483 	strp->cb.lock = cb->lock ? : strp_sock_lock;
484 	strp->cb.unlock = cb->unlock ? : strp_sock_unlock;
485 	strp->cb.rcv_msg = cb->rcv_msg;
486 	strp->cb.parse_msg = cb->parse_msg;
487 	strp->cb.read_sock_done = cb->read_sock_done ? : default_read_sock_done;
488 	strp->cb.abort_parser = cb->abort_parser ? : strp_abort_strp;
489 
490 	INIT_DELAYED_WORK(&strp->msg_timer_work, strp_msg_timeout);
491 	INIT_WORK(&strp->work, strp_work);
492 
493 	return 0;
494 }
495 EXPORT_SYMBOL_GPL(strp_init);
496 
497 /* Sock process lock held (lock_sock) */
498 void __strp_unpause(struct strparser *strp)
499 {
500 	strp->paused = 0;
501 
502 	if (strp->need_bytes) {
503 		if (strp_peek_len(strp) < strp->need_bytes)
504 			return;
505 	}
506 	strp_read_sock(strp);
507 }
508 EXPORT_SYMBOL_GPL(__strp_unpause);
509 
510 void strp_unpause(struct strparser *strp)
511 {
512 	strp->paused = 0;
513 
514 	/* Sync setting paused with RX work */
515 	smp_mb();
516 
517 	queue_work(strp_wq, &strp->work);
518 }
519 EXPORT_SYMBOL_GPL(strp_unpause);
520 
521 /* strp must already be stopped so that strp_recv will no longer be called.
522  * Note that strp_done is not called with the lower socket held.
523  */
524 void strp_done(struct strparser *strp)
525 {
526 	WARN_ON(!strp->stopped);
527 
528 	cancel_delayed_work_sync(&strp->msg_timer_work);
529 	cancel_work_sync(&strp->work);
530 
531 	if (strp->skb_head) {
532 		kfree_skb(strp->skb_head);
533 		strp->skb_head = NULL;
534 	}
535 }
536 EXPORT_SYMBOL_GPL(strp_done);
537 
538 void strp_stop(struct strparser *strp)
539 {
540 	strp->stopped = 1;
541 }
542 EXPORT_SYMBOL_GPL(strp_stop);
543 
544 void strp_check_rcv(struct strparser *strp)
545 {
546 	queue_work(strp_wq, &strp->work);
547 }
548 EXPORT_SYMBOL_GPL(strp_check_rcv);
549 
550 static int __init strp_mod_init(void)
551 {
552 	strp_wq = create_singlethread_workqueue("kstrp");
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