projects/linux·blob v4.18 · refs

563 lines · 13.2 KiB

1/*
2 * Stream Parser
3 *
4 * Copyright (c) 2016 Tom Herbert <[email protected]>
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
30static struct workqueue_struct *strp_wq;
31
32struct _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
40static 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 */
47static 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
67static 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 */
74static 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
83static 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 */
99static 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 since we may be appending to an skb that we
205 * already share a frag_list with.
206 */
207 err = skb_unclone(skb, GFP_ATOMIC);
208 if (err) {
209 STRP_STATS_INCR(strp->stats.mem_fail);
210 desc->error = err;
211 break;
212 }
213
214 stm = _strp_msg(head);
215 *strp->skb_nextp = skb;
216 strp->skb_nextp = &skb->next;
217 head->data_len += skb->len;
218 head->len += skb->len;
219 head->truesize += skb->truesize;
220 }
221
222 if (!stm->strp.full_len) {
223 ssize_t len;
224
225 len = (*strp->cb.parse_msg)(strp, head);
226
227 if (!len) {
228 /* Need more header to determine length */
229 if (!stm->accum_len) {
230 /* Start RX timer for new message */
231 strp_start_timer(strp, timeo);
232 }
233 stm->accum_len += cand_len;
234 eaten += cand_len;
235 STRP_STATS_INCR(strp->stats.need_more_hdr);
236 WARN_ON(eaten != orig_len);
237 break;
238 } else if (len < 0) {
239 if (len == -ESTRPIPE && stm->accum_len) {
240 len = -ENODATA;
241 strp->unrecov_intr = 1;
242 } else {
243 strp->interrupted = 1;
244 }
245 strp_parser_err(strp, len, desc);
246 break;
247 } else if (len > max_msg_size) {
248 /* Message length exceeds maximum allowed */
249 STRP_STATS_INCR(strp->stats.msg_too_big);
250 strp_parser_err(strp, -EMSGSIZE, desc);
251 break;
252 } else if (len <= (ssize_t)head->len -
253 skb->len - stm->strp.offset) {
254 /* Length must be into new skb (and also
255 * greater than zero)
256 */
257 STRP_STATS_INCR(strp->stats.bad_hdr_len);
258 strp_parser_err(strp, -EPROTO, desc);
259 break;
260 }
261
262 stm->strp.full_len = len;
263 }
264
265 extra = (ssize_t)(stm->accum_len + cand_len) -
266 stm->strp.full_len;
267
268 if (extra < 0) {
269 /* Message not complete yet. */
270 if (stm->strp.full_len - stm->accum_len >
271 strp_peek_len(strp)) {
272 /* Don't have the whole message in the socket
273 * buffer. Set strp->need_bytes to wait for
274 * the rest of the message. Also, set "early
275 * eaten" since we've already buffered the skb
276 * but don't consume yet per strp_read_sock.
277 */
278
279 if (!stm->accum_len) {
280 /* Start RX timer for new message */
281 strp_start_timer(strp, timeo);
282 }
283
284 stm->accum_len += cand_len;
285 eaten += cand_len;
286 strp->need_bytes = stm->strp.full_len -
287 stm->accum_len;
288 STRP_STATS_ADD(strp->stats.bytes, cand_len);
289 desc->count = 0; /* Stop reading socket */
290 break;
291 }
292 stm->accum_len += cand_len;
293 eaten += cand_len;
294 WARN_ON(eaten != orig_len);
295 break;
296 }
297
298 /* Positive extra indicates ore bytes than needed for the
299 * message
300 */
301
302 WARN_ON(extra > cand_len);
303
304 eaten += (cand_len - extra);
305
306 /* Hurray, we have a new message! */
307 cancel_delayed_work(&strp->msg_timer_work);
308 strp->skb_head = NULL;
309 strp->need_bytes = 0;
310 STRP_STATS_INCR(strp->stats.msgs);
311
312 /* Give skb to upper layer */
313 strp->cb.rcv_msg(strp, head);
314
315 if (unlikely(strp->paused)) {
316 /* Upper layer paused strp */
317 break;
318 }
319 }
320
321 if (cloned_orig)
322 kfree_skb(orig_skb);
323
324 STRP_STATS_ADD(strp->stats.bytes, eaten);
325
326 return eaten;
327}
328
329int strp_process(struct strparser *strp, struct sk_buff *orig_skb,
330 unsigned int orig_offset, size_t orig_len,
331 size_t max_msg_size, long timeo)
332{
333 read_descriptor_t desc; /* Dummy arg to strp_recv */
334
335 desc.arg.data = strp;
336
337 return __strp_recv(&desc, orig_skb, orig_offset, orig_len,
338 max_msg_size, timeo);
339}
340EXPORT_SYMBOL_GPL(strp_process);
341
342static int strp_recv(read_descriptor_t *desc, struct sk_buff *orig_skb,
343 unsigned int orig_offset, size_t orig_len)
344{
345 struct strparser *strp = (struct strparser *)desc->arg.data;
346
347 return __strp_recv(desc, orig_skb, orig_offset, orig_len,
348 strp->sk->sk_rcvbuf, strp->sk->sk_rcvtimeo);
349}
350
351static int default_read_sock_done(struct strparser *strp, int err)
352{
353 return err;
354}
355
356/* Called with lock held on lower socket */
357static int strp_read_sock(struct strparser *strp)
358{
359 struct socket *sock = strp->sk->sk_socket;
360 read_descriptor_t desc;
361
362 if (unlikely(!sock || !sock->ops || !sock->ops->read_sock))
363 return -EBUSY;
364
365 desc.arg.data = strp;
366 desc.error = 0;
367 desc.count = 1; /* give more than one skb per call */
368
369 /* sk should be locked here, so okay to do read_sock */
370 sock->ops->read_sock(strp->sk, &desc, strp_recv);
371
372 desc.error = strp->cb.read_sock_done(strp, desc.error);
373
374 return desc.error;
375}
376
377/* Lower sock lock held */
378void strp_data_ready(struct strparser *strp)
379{
380 if (unlikely(strp->stopped) || strp->paused)
381 return;
382
383 /* This check is needed to synchronize with do_strp_work.
384 * do_strp_work acquires a process lock (lock_sock) whereas
385 * the lock held here is bh_lock_sock. The two locks can be
386 * held by different threads at the same time, but bh_lock_sock
387 * allows a thread in BH context to safely check if the process
388 * lock is held. In this case, if the lock is held, queue work.
389 */
390 if (sock_owned_by_user_nocheck(strp->sk)) {
391 queue_work(strp_wq, &strp->work);
392 return;
393 }
394
395 if (strp->need_bytes) {
396 if (strp_peek_len(strp) < strp->need_bytes)
397 return;
398 }
399
400 if (strp_read_sock(strp) == -ENOMEM)
401 queue_work(strp_wq, &strp->work);
402}
403EXPORT_SYMBOL_GPL(strp_data_ready);
404
405static void do_strp_work(struct strparser *strp)
406{
407 read_descriptor_t rd_desc;
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 rd_desc.arg.data = strp;
421
422 if (strp_read_sock(strp) == -ENOMEM)
423 queue_work(strp_wq, &strp->work);
424
425out:
426 strp->cb.unlock(strp);
427}
428
429static void strp_work(struct work_struct *w)
430{
431 do_strp_work(container_of(w, struct strparser, work));
432}
433
434static 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
446static void strp_sock_lock(struct strparser *strp)
447{
448 lock_sock(strp->sk);
449}
450
451static void strp_sock_unlock(struct strparser *strp)
452{
453 release_sock(strp->sk);
454}
455
456int 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}
495EXPORT_SYMBOL_GPL(strp_init);
496
497/* Sock process lock held (lock_sock) */
498void __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}
508EXPORT_SYMBOL_GPL(__strp_unpause);
509
510void 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}
519EXPORT_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 */
524void 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}
536EXPORT_SYMBOL_GPL(strp_done);
537
538void strp_stop(struct strparser *strp)
539{
540 strp->stopped = 1;
541}
542EXPORT_SYMBOL_GPL(strp_stop);
543
544void strp_check_rcv(struct strparser *strp)
545{
546 queue_work(strp_wq, &strp->work);
547}
548EXPORT_SYMBOL_GPL(strp_check_rcv);
549
550static int __init strp_mod_init(void)
551{
552 strp_wq = create_singlethread_workqueue("kstrp");
553
554 return 0;
555}
556
557static void __exit strp_mod_exit(void)
558{
559 destroy_workqueue(strp_wq);
560}
561module_init(strp_mod_init);
562module_exit(strp_mod_exit);
563MODULE_LICENSE("GPL");