projects/linux·blob v5.2 · refs

554 lines · 13.0 KiB

1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Stream Parser
4 *
5 * Copyright (c) 2016 Tom Herbert <[email protected]>
6 */
7
8#include <linux/bpf.h>
9#include <linux/errno.h>
10#include <linux/errqueue.h>
11#include <linux/file.h>
12#include <linux/in.h>
13#include <linux/kernel.h>
14#include <linux/export.h>
15#include <linux/init.h>
16#include <linux/net.h>
17#include <linux/netdevice.h>
18#include <linux/poll.h>
19#include <linux/rculist.h>
20#include <linux/skbuff.h>
21#include <linux/socket.h>
22#include <linux/uaccess.h>
23#include <linux/workqueue.h>
24#include <net/strparser.h>
25#include <net/netns/generic.h>
26#include <net/sock.h>
27
28static struct workqueue_struct *strp_wq;
29
30struct _strp_msg {
31 /* Internal cb structure. struct strp_msg must be first for passing
32 * to upper layer.
33 */
34 struct strp_msg strp;
35 int accum_len;
36};
37
38static inline struct _strp_msg *_strp_msg(struct sk_buff *skb)
39{
40 return (struct _strp_msg *)((void *)skb->cb +
41 offsetof(struct qdisc_skb_cb, data));
42}
43
44/* Lower lock held */
45static void strp_abort_strp(struct strparser *strp, int err)
46{
47 /* Unrecoverable error in receive */
48
49 cancel_delayed_work(&strp->msg_timer_work);
50
51 if (strp->stopped)
52 return;
53
54 strp->stopped = 1;
55
56 if (strp->sk) {
57 struct sock *sk = strp->sk;
58
59 /* Report an error on the lower socket */
60 sk->sk_err = -err;
61 sk->sk_error_report(sk);
62 }
63}
64
65static void strp_start_timer(struct strparser *strp, long timeo)
66{
67 if (timeo && timeo != LONG_MAX)
68 mod_delayed_work(strp_wq, &strp->msg_timer_work, timeo);
69}
70
71/* Lower lock held */
72static void strp_parser_err(struct strparser *strp, int err,
73 read_descriptor_t *desc)
74{
75 desc->error = err;
76 kfree_skb(strp->skb_head);
77 strp->skb_head = NULL;
78 strp->cb.abort_parser(strp, err);
79}
80
81static inline int strp_peek_len(struct strparser *strp)
82{
83 if (strp->sk) {
84 struct socket *sock = strp->sk->sk_socket;
85
86 return sock->ops->peek_len(sock);
87 }
88
89 /* If we don't have an associated socket there's nothing to peek.
90 * Return int max to avoid stopping the strparser.
91 */
92
93 return INT_MAX;
94}
95
96/* Lower socket lock held */
97static int __strp_recv(read_descriptor_t *desc, struct sk_buff *orig_skb,
98 unsigned int orig_offset, size_t orig_len,
99 size_t max_msg_size, long timeo)
100{
101 struct strparser *strp = (struct strparser *)desc->arg.data;
102 struct _strp_msg *stm;
103 struct sk_buff *head, *skb;
104 size_t eaten = 0, cand_len;
105 ssize_t extra;
106 int err;
107 bool cloned_orig = false;
108
109 if (strp->paused)
110 return 0;
111
112 head = strp->skb_head;
113 if (head) {
114 /* Message already in progress */
115 if (unlikely(orig_offset)) {
116 /* Getting data with a non-zero offset when a message is
117 * in progress is not expected. If it does happen, we
118 * need to clone and pull since we can't deal with
119 * offsets in the skbs for a message expect in the head.
120 */
121 orig_skb = skb_clone(orig_skb, GFP_ATOMIC);
122 if (!orig_skb) {
123 STRP_STATS_INCR(strp->stats.mem_fail);
124 desc->error = -ENOMEM;
125 return 0;
126 }
127 if (!pskb_pull(orig_skb, orig_offset)) {
128 STRP_STATS_INCR(strp->stats.mem_fail);
129 kfree_skb(orig_skb);
130 desc->error = -ENOMEM;
131 return 0;
132 }
133 cloned_orig = true;
134 orig_offset = 0;
135 }
136
137 if (!strp->skb_nextp) {
138 /* We are going to append to the frags_list of head.
139 * Need to unshare the frag_list.
140 */
141 err = skb_unclone(head, GFP_ATOMIC);
142 if (err) {
143 STRP_STATS_INCR(strp->stats.mem_fail);
144 desc->error = err;
145 return 0;
146 }
147
148 if (unlikely(skb_shinfo(head)->frag_list)) {
149 /* We can't append to an sk_buff that already
150 * has a frag_list. We create a new head, point
151 * the frag_list of that to the old head, and
152 * then are able to use the old head->next for
153 * appending to the message.
154 */
155 if (WARN_ON(head->next)) {
156 desc->error = -EINVAL;
157 return 0;
158 }
159
160 skb = alloc_skb(0, GFP_ATOMIC);
161 if (!skb) {
162 STRP_STATS_INCR(strp->stats.mem_fail);
163 desc->error = -ENOMEM;
164 return 0;
165 }
166 skb->len = head->len;
167 skb->data_len = head->len;
168 skb->truesize = head->truesize;
169 *_strp_msg(skb) = *_strp_msg(head);
170 strp->skb_nextp = &head->next;
171 skb_shinfo(skb)->frag_list = head;
172 strp->skb_head = skb;
173 head = skb;
174 } else {
175 strp->skb_nextp =
176 &skb_shinfo(head)->frag_list;
177 }
178 }
179 }
180
181 while (eaten < orig_len) {
182 /* Always clone since we will consume something */
183 skb = skb_clone(orig_skb, GFP_ATOMIC);
184 if (!skb) {
185 STRP_STATS_INCR(strp->stats.mem_fail);
186 desc->error = -ENOMEM;
187 break;
188 }
189
190 cand_len = orig_len - eaten;
191
192 head = strp->skb_head;
193 if (!head) {
194 head = skb;
195 strp->skb_head = head;
196 /* Will set skb_nextp on next packet if needed */
197 strp->skb_nextp = NULL;
198 stm = _strp_msg(head);
199 memset(stm, 0, sizeof(*stm));
200 stm->strp.offset = orig_offset + eaten;
201 } else {
202 /* Unclone if we are appending to an skb that we
203 * already share a frag_list with.
204 */
205 if (skb_has_frag_list(skb)) {
206 err = skb_unclone(skb, GFP_ATOMIC);
207 if (err) {
208 STRP_STATS_INCR(strp->stats.mem_fail);
209 desc->error = err;
210 break;
211 }
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 more 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 /* We need the read lock to synchronize with strp_data_ready. We
408 * need the socket lock for calling strp_read_sock.
409 */
410 strp->cb.lock(strp);
411
412 if (unlikely(strp->stopped))
413 goto out;
414
415 if (strp->paused)
416 goto out;
417
418 if (strp_read_sock(strp) == -ENOMEM)
419 queue_work(strp_wq, &strp->work);
420
421out:
422 strp->cb.unlock(strp);
423}
424
425static void strp_work(struct work_struct *w)
426{
427 do_strp_work(container_of(w, struct strparser, work));
428}
429
430static void strp_msg_timeout(struct work_struct *w)
431{
432 struct strparser *strp = container_of(w, struct strparser,
433 msg_timer_work.work);
434
435 /* Message assembly timed out */
436 STRP_STATS_INCR(strp->stats.msg_timeouts);
437 strp->cb.lock(strp);
438 strp->cb.abort_parser(strp, -ETIMEDOUT);
439 strp->cb.unlock(strp);
440}
441
442static void strp_sock_lock(struct strparser *strp)
443{
444 lock_sock(strp->sk);
445}
446
447static void strp_sock_unlock(struct strparser *strp)
448{
449 release_sock(strp->sk);
450}
451
452int strp_init(struct strparser *strp, struct sock *sk,
453 const struct strp_callbacks *cb)
454{
455
456 if (!cb || !cb->rcv_msg || !cb->parse_msg)
457 return -EINVAL;
458
459 /* The sk (sock) arg determines the mode of the stream parser.
460 *
461 * If the sock is set then the strparser is in receive callback mode.
462 * The upper layer calls strp_data_ready to kick receive processing
463 * and strparser calls the read_sock function on the socket to
464 * get packets.
465 *
466 * If the sock is not set then the strparser is in general mode.
467 * The upper layer calls strp_process for each skb to be parsed.
468 */
469
470 if (!sk) {
471 if (!cb->lock || !cb->unlock)
472 return -EINVAL;
473 }
474
475 memset(strp, 0, sizeof(*strp));
476
477 strp->sk = sk;
478
479 strp->cb.lock = cb->lock ? : strp_sock_lock;
480 strp->cb.unlock = cb->unlock ? : strp_sock_unlock;
481 strp->cb.rcv_msg = cb->rcv_msg;
482 strp->cb.parse_msg = cb->parse_msg;
483 strp->cb.read_sock_done = cb->read_sock_done ? : default_read_sock_done;
484 strp->cb.abort_parser = cb->abort_parser ? : strp_abort_strp;
485
486 INIT_DELAYED_WORK(&strp->msg_timer_work, strp_msg_timeout);
487 INIT_WORK(&strp->work, strp_work);
488
489 return 0;
490}
491EXPORT_SYMBOL_GPL(strp_init);
492
493/* Sock process lock held (lock_sock) */
494void __strp_unpause(struct strparser *strp)
495{
496 strp->paused = 0;
497
498 if (strp->need_bytes) {
499 if (strp_peek_len(strp) < strp->need_bytes)
500 return;
501 }
502 strp_read_sock(strp);
503}
504EXPORT_SYMBOL_GPL(__strp_unpause);
505
506void strp_unpause(struct strparser *strp)
507{
508 strp->paused = 0;
509
510 /* Sync setting paused with RX work */
511 smp_mb();
512
513 queue_work(strp_wq, &strp->work);
514}
515EXPORT_SYMBOL_GPL(strp_unpause);
516
517/* strp must already be stopped so that strp_recv will no longer be called.
518 * Note that strp_done is not called with the lower socket held.
519 */
520void strp_done(struct strparser *strp)
521{
522 WARN_ON(!strp->stopped);
523
524 cancel_delayed_work_sync(&strp->msg_timer_work);
525 cancel_work_sync(&strp->work);
526
527 if (strp->skb_head) {
528 kfree_skb(strp->skb_head);
529 strp->skb_head = NULL;
530 }
531}
532EXPORT_SYMBOL_GPL(strp_done);
533
534void strp_stop(struct strparser *strp)
535{
536 strp->stopped = 1;
537}
538EXPORT_SYMBOL_GPL(strp_stop);
539
540void strp_check_rcv(struct strparser *strp)
541{
542 queue_work(strp_wq, &strp->work);
543}
544EXPORT_SYMBOL_GPL(strp_check_rcv);
545
546static int __init strp_dev_init(void)
547{
548 strp_wq = create_singlethread_workqueue("kstrp");
549 if (unlikely(!strp_wq))
550 return -ENOMEM;
551
552 return 0;
553}
554device_initcall(strp_dev_init);