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/export.h> 18 #include <linux/init.h> 19 #include <linux/net.h> 20 #include <linux/netdevice.h> 21 #include <linux/poll.h> 22 #include <linux/rculist.h> 23 #include <linux/skbuff.h> 24 #include <linux/socket.h> 25 #include <linux/uaccess.h> 26 #include <linux/workqueue.h> 27 #include <net/strparser.h> 28 #include <net/netns/generic.h> 29 #include <net/sock.h> 30 31 static struct workqueue_struct *strp_wq; 32 33 struct _strp_msg { 34 /* Internal cb structure. struct strp_msg must be first for passing 35 * to upper layer. 36 */ 37 struct strp_msg strp; 38 int accum_len; 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 if (unlikely(orig_offset)) { 119 /* Getting data with a non-zero offset when a message is 120 * in progress is not expected. If it does happen, we 121 * need to clone and pull since we can't deal with 122 * offsets in the skbs for a message expect in the head. 123 */ 124 orig_skb = skb_clone(orig_skb, GFP_ATOMIC); 125 if (!orig_skb) { 126 STRP_STATS_INCR(strp->stats.mem_fail); 127 desc->error = -ENOMEM; 128 return 0; 129 } 130 if (!pskb_pull(orig_skb, orig_offset)) { 131 STRP_STATS_INCR(strp->stats.mem_fail); 132 kfree_skb(orig_skb); 133 desc->error = -ENOMEM; 134 return 0; 135 } 136 cloned_orig = true; 137 orig_offset = 0; 138 } 139 140 if (!strp->skb_nextp) { 141 /* We are going to append to the frags_list of head. 142 * Need to unshare the frag_list. 143 */ 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 if (unlikely(skb_shinfo(head)->frag_list)) { 152 /* We can't append to an sk_buff that already 153 * has a frag_list. We create a new head, point 154 * the frag_list of that to the old head, and 155 * then are able to use the old head->next for 156 * appending to the message. 157 */ 158 if (WARN_ON(head->next)) { 159 desc->error = -EINVAL; 160 return 0; 161 } 162 163 skb = alloc_skb_for_msg(head); 164 if (!skb) { 165 STRP_STATS_INCR(strp->stats.mem_fail); 166 desc->error = -ENOMEM; 167 return 0; 168 } 169 170 strp->skb_nextp = &head->next; 171 strp->skb_head = skb; 172 head = skb; 173 } else { 174 strp->skb_nextp = 175 &skb_shinfo(head)->frag_list; 176 } 177 } 178 } 179 180 while (eaten < orig_len) { 181 /* Always clone since we will consume something */ 182 skb = skb_clone(orig_skb, GFP_ATOMIC); 183 if (!skb) { 184 STRP_STATS_INCR(strp->stats.mem_fail); 185 desc->error = -ENOMEM; 186 break; 187 } 188 189 cand_len = orig_len - eaten; 190 191 head = strp->skb_head; 192 if (!head) { 193 head = skb; 194 strp->skb_head = head; 195 /* Will set skb_nextp on next packet if needed */ 196 strp->skb_nextp = NULL; 197 stm = _strp_msg(head); 198 memset(stm, 0, sizeof(*stm)); 199 stm->strp.offset = orig_offset + eaten; 200 } else { 201 /* Unclone if we are appending to an skb that we 202 * already share a frag_list with. 203 */ 204 if (skb_has_frag_list(skb)) { 205 err = skb_unclone(skb, GFP_ATOMIC); 206 if (err) { 207 STRP_STATS_INCR(strp->stats.mem_fail); 208 desc->error = err; 209 break; 210 } 211 } 212 213 stm = _strp_msg(head); 214 *strp->skb_nextp = skb; 215 strp->skb_nextp = &skb->next; 216 head->data_len += skb->len; 217 head->len += skb->len; 218 head->truesize += skb->truesize; 219 } 220 221 if (!stm->strp.full_len) { 222 ssize_t len; 223 224 len = (*strp->cb.parse_msg)(strp, head); 225 226 if (!len) { 227 /* Need more header to determine length */ 228 if (!stm->accum_len) { 229 /* Start RX timer for new message */ 230 strp_start_timer(strp, timeo); 231 } 232 stm->accum_len += cand_len; 233 eaten += cand_len; 234 STRP_STATS_INCR(strp->stats.need_more_hdr); 235 WARN_ON(eaten != orig_len); 236 break; 237 } else if (len < 0) { 238 if (len == -ESTRPIPE && stm->accum_len) { 239 len = -ENODATA; 240 strp->unrecov_intr = 1; 241 } else { 242 strp->interrupted = 1; 243 } 244 strp_parser_err(strp, len, desc); 245 break; 246 } else if (len > max_msg_size) { 247 /* Message length exceeds maximum allowed */ 248 STRP_STATS_INCR(strp->stats.msg_too_big); 249 strp_parser_err(strp, -EMSGSIZE, desc); 250 break; 251 } else if (len <= (ssize_t)head->len - 252 skb->len - stm->strp.offset) { 253 /* Length must be into new skb (and also 254 * greater than zero) 255 */ 256 STRP_STATS_INCR(strp->stats.bad_hdr_len); 257 strp_parser_err(strp, -EPROTO, desc); 258 break; 259 } 260 261 stm->strp.full_len = len; 262 } 263 264 extra = (ssize_t)(stm->accum_len + cand_len) - 265 stm->strp.full_len; 266 267 if (extra < 0) { 268 /* Message not complete yet. */ 269 if (stm->strp.full_len - stm->accum_len > 270 strp_peek_len(strp)) { 271 /* Don't have the whole message in the socket 272 * buffer. Set strp->need_bytes to wait for 273 * the rest of the message. Also, set "early 274 * eaten" since we've already buffered the skb 275 * but don't consume yet per strp_read_sock. 276 */ 277 278 if (!stm->accum_len) { 279 /* Start RX timer for new message */ 280 strp_start_timer(strp, timeo); 281 } 282 283 stm->accum_len += cand_len; 284 eaten += cand_len; 285 strp->need_bytes = stm->strp.full_len - 286 stm->accum_len; 287 STRP_STATS_ADD(strp->stats.bytes, cand_len); 288 desc->count = 0; /* Stop reading socket */ 289 break; 290 } 291 stm->accum_len += cand_len; 292 eaten += cand_len; 293 WARN_ON(eaten != orig_len); 294 break; 295 } 296 297 /* Positive extra indicates more bytes than needed for the 298 * message 299 */ 300 301 WARN_ON(extra > cand_len); 302 303 eaten += (cand_len - extra); 304 305 /* Hurray, we have a new message! */ 306 cancel_delayed_work(&strp->msg_timer_work); 307 strp->skb_head = NULL; 308 strp->need_bytes = 0; 309 STRP_STATS_INCR(strp->stats.msgs); 310 311 /* Give skb to upper layer */ 312 strp->cb.rcv_msg(strp, head); 313 314 if (unlikely(strp->paused)) { 315 /* Upper layer paused strp */ 316 break; 317 } 318 } 319 320 if (cloned_orig) 321 kfree_skb(orig_skb); 322 323 STRP_STATS_ADD(strp->stats.bytes, eaten); 324 325 return eaten; 326 } 327 328 int strp_process(struct strparser *strp, struct sk_buff *orig_skb, 329 unsigned int orig_offset, size_t orig_len, 330 size_t max_msg_size, long timeo) 331 { 332 read_descriptor_t desc; /* Dummy arg to strp_recv */ 333 334 desc.arg.data = strp; 335 336 return __strp_recv(&desc, orig_skb, orig_offset, orig_len, 337 max_msg_size, timeo); 338 } 339 EXPORT_SYMBOL_GPL(strp_process); 340 341 static int strp_recv(read_descriptor_t *desc, struct sk_buff *orig_skb, 342 unsigned int orig_offset, size_t orig_len) 343 { 344 struct strparser *strp = (struct strparser *)desc->arg.data; 345 346 return __strp_recv(desc, orig_skb, orig_offset, orig_len, 347 strp->sk->sk_rcvbuf, strp->sk->sk_rcvtimeo); 348 } 349 350 static int default_read_sock_done(struct strparser *strp, int err) 351 { 352 return err; 353 } 354 355 /* Called with lock held on lower socket */ 356 static int strp_read_sock(struct strparser *strp) 357 { 358 struct socket *sock = strp->sk->sk_socket; 359 read_descriptor_t desc; 360 361 if (unlikely(!sock || !sock->ops || !sock->ops->read_sock)) 362 return -EBUSY; 363 364 desc.arg.data = strp; 365 desc.error = 0; 366 desc.count = 1; /* give more than one skb per call */ 367 368 /* sk should be locked here, so okay to do read_sock */ 369 sock->ops->read_sock(strp->sk, &desc, strp_recv); 370 371 desc.error = strp->cb.read_sock_done(strp, desc.error); 372 373 return desc.error; 374 } 375 376 /* Lower sock lock held */ 377 void strp_data_ready(struct strparser *strp) 378 { 379 if (unlikely(strp->stopped) || strp->paused) 380 return; 381 382 /* This check is needed to synchronize with do_strp_work. 383 * do_strp_work acquires a process lock (lock_sock) whereas 384 * the lock held here is bh_lock_sock. The two locks can be 385 * held by different threads at the same time, but bh_lock_sock 386 * allows a thread in BH context to safely check if the process 387 * lock is held. In this case, if the lock is held, queue work. 388 */ 389 if (sock_owned_by_user_nocheck(strp->sk)) { 390 queue_work(strp_wq, &strp->work); 391 return; 392 } 393 394 if (strp->need_bytes) { 395 if (strp_peek_len(strp) < strp->need_bytes) 396 return; 397 } 398 399 if (strp_read_sock(strp) == -ENOMEM) 400 queue_work(strp_wq, &strp->work); 401 } 402 EXPORT_SYMBOL_GPL(strp_data_ready); 403 404 static void do_strp_work(struct strparser *strp) 405 { 406 /* We need the read lock to synchronize with strp_data_ready. We 407 * need the socket lock for calling strp_read_sock. 408 */ 409 strp->cb.lock(strp); 410 411 if (unlikely(strp->stopped)) 412 goto out; 413 414 if (strp->paused) 415 goto out; 416 417 if (strp_read_sock(strp) == -ENOMEM) 418 queue_work(strp_wq, &strp->work); 419 420 out: 421 strp->cb.unlock(strp); 422 } 423 424 static void strp_work(struct work_struct *w) 425 { 426 do_strp_work(container_of(w, struct strparser, work)); 427 } 428 429 static void strp_msg_timeout(struct work_struct *w) 430 { 431 struct strparser *strp = container_of(w, struct strparser, 432 msg_timer_work.work); 433 434 /* Message assembly timed out */ 435 STRP_STATS_INCR(strp->stats.msg_timeouts); 436 strp->cb.lock(strp); 437 strp->cb.abort_parser(strp, -ETIMEDOUT); 438 strp->cb.unlock(strp); 439 } 440 441 static void strp_sock_lock(struct strparser *strp) 442 { 443 lock_sock(strp->sk); 444 } 445 446 static void strp_sock_unlock(struct strparser *strp) 447 { 448 release_sock(strp->sk); 449 } 450 451 int strp_init(struct strparser *strp, struct sock *sk, 452 const struct strp_callbacks *cb) 453 { 454 455 if (!cb || !cb->rcv_msg || !cb->parse_msg) 456 return -EINVAL; 457 458 /* The sk (sock) arg determines the mode of the stream parser. 459 * 460 * If the sock is set then the strparser is in receive callback mode. 461 * The upper layer calls strp_data_ready to kick receive processing 462 * and strparser calls the read_sock function on the socket to 463 * get packets. 464 * 465 * If the sock is not set then the strparser is in general mode. 466 * The upper layer calls strp_process for each skb to be parsed. 467 */ 468 469 if (!sk) { 470 if (!cb->lock || !cb->unlock) 471 return -EINVAL; 472 } 473 474 memset(strp, 0, sizeof(*strp)); 475 476 strp->sk = sk; 477 478 strp->cb.lock = cb->lock ? : strp_sock_lock; 479 strp->cb.unlock = cb->unlock ? : strp_sock_unlock; 480 strp->cb.rcv_msg = cb->rcv_msg; 481 strp->cb.parse_msg = cb->parse_msg; 482 strp->cb.read_sock_done = cb->read_sock_done ? : default_read_sock_done; 483 strp->cb.abort_parser = cb->abort_parser ? : strp_abort_strp; 484 485 INIT_DELAYED_WORK(&strp->msg_timer_work, strp_msg_timeout); 486 INIT_WORK(&strp->work, strp_work); 487 488 return 0; 489 } 490 EXPORT_SYMBOL_GPL(strp_init); 491 492 /* Sock process lock held (lock_sock) */ 493 void __strp_unpause(struct strparser *strp) 494 { 495 strp->paused = 0; 496 497 if (strp->need_bytes) { 498 if (strp_peek_len(strp) < strp->need_bytes) 499 return; 500 } 501 strp_read_sock(strp); 502 } 503 EXPORT_SYMBOL_GPL(__strp_unpause); 504 505 void strp_unpause(struct strparser *strp) 506 { 507 strp->paused = 0; 508 509 /* Sync setting paused with RX work */ 510 smp_mb(); 511 512 queue_work(strp_wq, &strp->work); 513 } 514 EXPORT_SYMBOL_GPL(strp_unpause); 515 516 /* strp must already be stopped so that strp_recv will no longer be called. 517 * Note that strp_done is not called with the lower socket held. 518 */ 519 void strp_done(struct strparser *strp) 520 { 521 WARN_ON(!strp->stopped); 522 523 cancel_delayed_work_sync(&strp->msg_timer_work); 524 cancel_work_sync(&strp->work); 525 526 if (strp->skb_head) { 527 kfree_skb(strp->skb_head); 528 strp->skb_head = NULL; 529 } 530 } 531 EXPORT_SYMBOL_GPL(strp_done); 532 533 void strp_stop(struct strparser *strp) 534 { 535 strp->stopped = 1; 536 } 537 EXPORT_SYMBOL_GPL(strp_stop); 538 539 void strp_check_rcv(struct strparser *strp) 540 { 541 queue_work(strp_wq, &strp->work); 542 } 543 EXPORT_SYMBOL_GPL(strp_check_rcv); 544 545 static int __init strp_dev_init(void) 546 { 547 strp_wq = create_singlethread_workqueue("kstrp"); 548 if (unlikely(!strp_wq)) 549 return -ENOMEM; 550 551 return 0; 552 } 553 device_initcall(strp_dev_init); 554