1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * inet fragments management 4 * 5 * Authors: Pavel Emelyanov <xemul@openvz.org> 6 * Started as consolidation of ipv4/ip_fragment.c, 7 * ipv6/reassembly. and ipv6 nf conntrack reassembly 8 */ 9 10 #include <linux/list.h> 11 #include <linux/spinlock.h> 12 #include <linux/module.h> 13 #include <linux/timer.h> 14 #include <linux/mm.h> 15 #include <linux/random.h> 16 #include <linux/skbuff.h> 17 #include <linux/rtnetlink.h> 18 #include <linux/slab.h> 19 #include <linux/rhashtable.h> 20 21 #include <net/sock.h> 22 #include <net/inet_frag.h> 23 #include <net/inet_ecn.h> 24 #include <net/ip.h> 25 #include <net/ipv6.h> 26 27 /* Use skb->cb to track consecutive/adjacent fragments coming at 28 * the end of the queue. Nodes in the rb-tree queue will 29 * contain "runs" of one or more adjacent fragments. 30 * 31 * Invariants: 32 * - next_frag is NULL at the tail of a "run"; 33 * - the head of a "run" has the sum of all fragment lengths in frag_run_len. 34 */ 35 struct ipfrag_skb_cb { 36 union { 37 struct inet_skb_parm h4; 38 struct inet6_skb_parm h6; 39 }; 40 struct sk_buff *next_frag; 41 int frag_run_len; 42 }; 43 44 #define FRAG_CB(skb) ((struct ipfrag_skb_cb *)((skb)->cb)) 45 46 static void fragcb_clear(struct sk_buff *skb) 47 { 48 RB_CLEAR_NODE(&skb->rbnode); 49 FRAG_CB(skb)->next_frag = NULL; 50 FRAG_CB(skb)->frag_run_len = skb->len; 51 } 52 53 /* Append skb to the last "run". */ 54 static void fragrun_append_to_last(struct inet_frag_queue *q, 55 struct sk_buff *skb) 56 { 57 fragcb_clear(skb); 58 59 FRAG_CB(q->last_run_head)->frag_run_len += skb->len; 60 FRAG_CB(q->fragments_tail)->next_frag = skb; 61 q->fragments_tail = skb; 62 } 63 64 /* Create a new "run" with the skb. */ 65 static void fragrun_create(struct inet_frag_queue *q, struct sk_buff *skb) 66 { 67 BUILD_BUG_ON(sizeof(struct ipfrag_skb_cb) > sizeof(skb->cb)); 68 fragcb_clear(skb); 69 70 if (q->last_run_head) 71 rb_link_node(&skb->rbnode, &q->last_run_head->rbnode, 72 &q->last_run_head->rbnode.rb_right); 73 else 74 rb_link_node(&skb->rbnode, NULL, &q->rb_fragments.rb_node); 75 rb_insert_color(&skb->rbnode, &q->rb_fragments); 76 77 q->fragments_tail = skb; 78 q->last_run_head = skb; 79 } 80 81 /* Given the OR values of all fragments, apply RFC 3168 5.3 requirements 82 * Value : 0xff if frame should be dropped. 83 * 0 or INET_ECN_CE value, to be ORed in to final iph->tos field 84 */ 85 const u8 ip_frag_ecn_table[16] = { 86 /* at least one fragment had CE, and others ECT_0 or ECT_1 */ 87 [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0] = INET_ECN_CE, 88 [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1] = INET_ECN_CE, 89 [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = INET_ECN_CE, 90 91 /* invalid combinations : drop frame */ 92 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE] = 0xff, 93 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0] = 0xff, 94 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_1] = 0xff, 95 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff, 96 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0] = 0xff, 97 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1] = 0xff, 98 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff, 99 }; 100 EXPORT_SYMBOL(ip_frag_ecn_table); 101 102 int inet_frags_init(struct inet_frags *f) 103 { 104 f->frags_cachep = kmem_cache_create(f->frags_cache_name, f->qsize, 0, 0, 105 NULL); 106 if (!f->frags_cachep) 107 return -ENOMEM; 108 109 refcount_set(&f->refcnt, 1); 110 init_completion(&f->completion); 111 return 0; 112 } 113 EXPORT_SYMBOL(inet_frags_init); 114 115 void inet_frags_fini(struct inet_frags *f) 116 { 117 if (refcount_dec_and_test(&f->refcnt)) 118 complete(&f->completion); 119 120 wait_for_completion(&f->completion); 121 122 kmem_cache_destroy(f->frags_cachep); 123 f->frags_cachep = NULL; 124 } 125 EXPORT_SYMBOL(inet_frags_fini); 126 127 /* called from rhashtable_free_and_destroy() at netns_frags dismantle */ 128 static void inet_frags_free_cb(void *ptr, void *arg) 129 { 130 struct inet_frag_queue *fq = ptr; 131 int count; 132 133 count = del_timer_sync(&fq->timer) ? 1 : 0; 134 135 spin_lock_bh(&fq->lock); 136 if (!(fq->flags & INET_FRAG_COMPLETE)) { 137 fq->flags |= INET_FRAG_COMPLETE; 138 count++; 139 } else if (fq->flags & INET_FRAG_HASH_DEAD) { 140 count++; 141 } 142 spin_unlock_bh(&fq->lock); 143 144 if (refcount_sub_and_test(count, &fq->refcnt)) 145 inet_frag_destroy(fq); 146 } 147 148 static void fqdir_work_fn(struct work_struct *work) 149 { 150 struct fqdir *fqdir = container_of(work, struct fqdir, destroy_work); 151 struct inet_frags *f = fqdir->f; 152 153 rhashtable_free_and_destroy(&fqdir->rhashtable, inet_frags_free_cb, NULL); 154 155 /* We need to make sure all ongoing call_rcu(..., inet_frag_destroy_rcu) 156 * have completed, since they need to dereference fqdir. 157 * Would it not be nice to have kfree_rcu_barrier() ? :) 158 */ 159 rcu_barrier(); 160 161 if (refcount_dec_and_test(&f->refcnt)) 162 complete(&f->completion); 163 164 kfree(fqdir); 165 } 166 167 int fqdir_init(struct fqdir **fqdirp, struct inet_frags *f, struct net *net) 168 { 169 struct fqdir *fqdir = kzalloc(sizeof(*fqdir), GFP_KERNEL); 170 int res; 171 172 if (!fqdir) 173 return -ENOMEM; 174 fqdir->f = f; 175 fqdir->net = net; 176 res = rhashtable_init(&fqdir->rhashtable, &fqdir->f->rhash_params); 177 if (res < 0) { 178 kfree(fqdir); 179 return res; 180 } 181 refcount_inc(&f->refcnt); 182 *fqdirp = fqdir; 183 return 0; 184 } 185 EXPORT_SYMBOL(fqdir_init); 186 187 void fqdir_exit(struct fqdir *fqdir) 188 { 189 INIT_WORK(&fqdir->destroy_work, fqdir_work_fn); 190 queue_work(system_wq, &fqdir->destroy_work); 191 } 192 EXPORT_SYMBOL(fqdir_exit); 193 194 void inet_frag_kill(struct inet_frag_queue *fq) 195 { 196 if (del_timer(&fq->timer)) 197 refcount_dec(&fq->refcnt); 198 199 if (!(fq->flags & INET_FRAG_COMPLETE)) { 200 struct fqdir *fqdir = fq->fqdir; 201 202 fq->flags |= INET_FRAG_COMPLETE; 203 rcu_read_lock(); 204 /* The RCU read lock provides a memory barrier 205 * guaranteeing that if fqdir->dead is false then 206 * the hash table destruction will not start until 207 * after we unlock. Paired with inet_frags_exit_net(). 208 */ 209 if (!fqdir->dead) { 210 rhashtable_remove_fast(&fqdir->rhashtable, &fq->node, 211 fqdir->f->rhash_params); 212 refcount_dec(&fq->refcnt); 213 } else { 214 fq->flags |= INET_FRAG_HASH_DEAD; 215 } 216 rcu_read_unlock(); 217 } 218 } 219 EXPORT_SYMBOL(inet_frag_kill); 220 221 static void inet_frag_destroy_rcu(struct rcu_head *head) 222 { 223 struct inet_frag_queue *q = container_of(head, struct inet_frag_queue, 224 rcu); 225 struct inet_frags *f = q->fqdir->f; 226 227 if (f->destructor) 228 f->destructor(q); 229 kmem_cache_free(f->frags_cachep, q); 230 } 231 232 unsigned int inet_frag_rbtree_purge(struct rb_root *root) 233 { 234 struct rb_node *p = rb_first(root); 235 unsigned int sum = 0; 236 237 while (p) { 238 struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode); 239 240 p = rb_next(p); 241 rb_erase(&skb->rbnode, root); 242 while (skb) { 243 struct sk_buff *next = FRAG_CB(skb)->next_frag; 244 245 sum += skb->truesize; 246 kfree_skb(skb); 247 skb = next; 248 } 249 } 250 return sum; 251 } 252 EXPORT_SYMBOL(inet_frag_rbtree_purge); 253 254 void inet_frag_destroy(struct inet_frag_queue *q) 255 { 256 struct fqdir *fqdir; 257 unsigned int sum, sum_truesize = 0; 258 struct inet_frags *f; 259 260 WARN_ON(!(q->flags & INET_FRAG_COMPLETE)); 261 WARN_ON(del_timer(&q->timer) != 0); 262 263 /* Release all fragment data. */ 264 fqdir = q->fqdir; 265 f = fqdir->f; 266 sum_truesize = inet_frag_rbtree_purge(&q->rb_fragments); 267 sum = sum_truesize + f->qsize; 268 269 call_rcu(&q->rcu, inet_frag_destroy_rcu); 270 271 sub_frag_mem_limit(fqdir, sum); 272 } 273 EXPORT_SYMBOL(inet_frag_destroy); 274 275 static struct inet_frag_queue *inet_frag_alloc(struct fqdir *fqdir, 276 struct inet_frags *f, 277 void *arg) 278 { 279 struct inet_frag_queue *q; 280 281 q = kmem_cache_zalloc(f->frags_cachep, GFP_ATOMIC); 282 if (!q) 283 return NULL; 284 285 q->fqdir = fqdir; 286 f->constructor(q, arg); 287 add_frag_mem_limit(fqdir, f->qsize); 288 289 timer_setup(&q->timer, f->frag_expire, 0); 290 spin_lock_init(&q->lock); 291 refcount_set(&q->refcnt, 3); 292 293 return q; 294 } 295 296 static struct inet_frag_queue *inet_frag_create(struct fqdir *fqdir, 297 void *arg, 298 struct inet_frag_queue **prev) 299 { 300 struct inet_frags *f = fqdir->f; 301 struct inet_frag_queue *q; 302 303 q = inet_frag_alloc(fqdir, f, arg); 304 if (!q) { 305 *prev = ERR_PTR(-ENOMEM); 306 return NULL; 307 } 308 mod_timer(&q->timer, jiffies + fqdir->timeout); 309 310 *prev = rhashtable_lookup_get_insert_key(&fqdir->rhashtable, &q->key, 311 &q->node, f->rhash_params); 312 if (*prev) { 313 q->flags |= INET_FRAG_COMPLETE; 314 inet_frag_kill(q); 315 inet_frag_destroy(q); 316 return NULL; 317 } 318 return q; 319 } 320 321 /* TODO : call from rcu_read_lock() and no longer use refcount_inc_not_zero() */ 322 struct inet_frag_queue *inet_frag_find(struct fqdir *fqdir, void *key) 323 { 324 struct inet_frag_queue *fq = NULL, *prev; 325 326 if (!fqdir->high_thresh || frag_mem_limit(fqdir) > fqdir->high_thresh) 327 return NULL; 328 329 rcu_read_lock(); 330 331 prev = rhashtable_lookup(&fqdir->rhashtable, key, fqdir->f->rhash_params); 332 if (!prev) 333 fq = inet_frag_create(fqdir, key, &prev); 334 if (!IS_ERR_OR_NULL(prev)) { 335 fq = prev; 336 if (!refcount_inc_not_zero(&fq->refcnt)) 337 fq = NULL; 338 } 339 rcu_read_unlock(); 340 return fq; 341 } 342 EXPORT_SYMBOL(inet_frag_find); 343 344 int inet_frag_queue_insert(struct inet_frag_queue *q, struct sk_buff *skb, 345 int offset, int end) 346 { 347 struct sk_buff *last = q->fragments_tail; 348 349 /* RFC5722, Section 4, amended by Errata ID : 3089 350 * When reassembling an IPv6 datagram, if 351 * one or more its constituent fragments is determined to be an 352 * overlapping fragment, the entire datagram (and any constituent 353 * fragments) MUST be silently discarded. 354 * 355 * Duplicates, however, should be ignored (i.e. skb dropped, but the 356 * queue/fragments kept for later reassembly). 357 */ 358 if (!last) 359 fragrun_create(q, skb); /* First fragment. */ 360 else if (last->ip_defrag_offset + last->len < end) { 361 /* This is the common case: skb goes to the end. */ 362 /* Detect and discard overlaps. */ 363 if (offset < last->ip_defrag_offset + last->len) 364 return IPFRAG_OVERLAP; 365 if (offset == last->ip_defrag_offset + last->len) 366 fragrun_append_to_last(q, skb); 367 else 368 fragrun_create(q, skb); 369 } else { 370 /* Binary search. Note that skb can become the first fragment, 371 * but not the last (covered above). 372 */ 373 struct rb_node **rbn, *parent; 374 375 rbn = &q->rb_fragments.rb_node; 376 do { 377 struct sk_buff *curr; 378 int curr_run_end; 379 380 parent = *rbn; 381 curr = rb_to_skb(parent); 382 curr_run_end = curr->ip_defrag_offset + 383 FRAG_CB(curr)->frag_run_len; 384 if (end <= curr->ip_defrag_offset) 385 rbn = &parent->rb_left; 386 else if (offset >= curr_run_end) 387 rbn = &parent->rb_right; 388 else if (offset >= curr->ip_defrag_offset && 389 end <= curr_run_end) 390 return IPFRAG_DUP; 391 else 392 return IPFRAG_OVERLAP; 393 } while (*rbn); 394 /* Here we have parent properly set, and rbn pointing to 395 * one of its NULL left/right children. Insert skb. 396 */ 397 fragcb_clear(skb); 398 rb_link_node(&skb->rbnode, parent, rbn); 399 rb_insert_color(&skb->rbnode, &q->rb_fragments); 400 } 401 402 skb->ip_defrag_offset = offset; 403 404 return IPFRAG_OK; 405 } 406 EXPORT_SYMBOL(inet_frag_queue_insert); 407 408 void *inet_frag_reasm_prepare(struct inet_frag_queue *q, struct sk_buff *skb, 409 struct sk_buff *parent) 410 { 411 struct sk_buff *fp, *head = skb_rb_first(&q->rb_fragments); 412 struct sk_buff **nextp; 413 int delta; 414 415 if (head != skb) { 416 fp = skb_clone(skb, GFP_ATOMIC); 417 if (!fp) 418 return NULL; 419 FRAG_CB(fp)->next_frag = FRAG_CB(skb)->next_frag; 420 if (RB_EMPTY_NODE(&skb->rbnode)) 421 FRAG_CB(parent)->next_frag = fp; 422 else 423 rb_replace_node(&skb->rbnode, &fp->rbnode, 424 &q->rb_fragments); 425 if (q->fragments_tail == skb) 426 q->fragments_tail = fp; 427 skb_morph(skb, head); 428 FRAG_CB(skb)->next_frag = FRAG_CB(head)->next_frag; 429 rb_replace_node(&head->rbnode, &skb->rbnode, 430 &q->rb_fragments); 431 consume_skb(head); 432 head = skb; 433 } 434 WARN_ON(head->ip_defrag_offset != 0); 435 436 delta = -head->truesize; 437 438 /* Head of list must not be cloned. */ 439 if (skb_unclone(head, GFP_ATOMIC)) 440 return NULL; 441 442 delta += head->truesize; 443 if (delta) 444 add_frag_mem_limit(q->fqdir, delta); 445 446 /* If the first fragment is fragmented itself, we split 447 * it to two chunks: the first with data and paged part 448 * and the second, holding only fragments. 449 */ 450 if (skb_has_frag_list(head)) { 451 struct sk_buff *clone; 452 int i, plen = 0; 453 454 clone = alloc_skb(0, GFP_ATOMIC); 455 if (!clone) 456 return NULL; 457 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list; 458 skb_frag_list_init(head); 459 for (i = 0; i < skb_shinfo(head)->nr_frags; i++) 460 plen += skb_frag_size(&skb_shinfo(head)->frags[i]); 461 clone->data_len = head->data_len - plen; 462 clone->len = clone->data_len; 463 head->truesize += clone->truesize; 464 clone->csum = 0; 465 clone->ip_summed = head->ip_summed; 466 add_frag_mem_limit(q->fqdir, clone->truesize); 467 skb_shinfo(head)->frag_list = clone; 468 nextp = &clone->next; 469 } else { 470 nextp = &skb_shinfo(head)->frag_list; 471 } 472 473 return nextp; 474 } 475 EXPORT_SYMBOL(inet_frag_reasm_prepare); 476 477 void inet_frag_reasm_finish(struct inet_frag_queue *q, struct sk_buff *head, 478 void *reasm_data, bool try_coalesce) 479 { 480 struct sk_buff **nextp = (struct sk_buff **)reasm_data; 481 struct rb_node *rbn; 482 struct sk_buff *fp; 483 int sum_truesize; 484 485 skb_push(head, head->data - skb_network_header(head)); 486 487 /* Traverse the tree in order, to build frag_list. */ 488 fp = FRAG_CB(head)->next_frag; 489 rbn = rb_next(&head->rbnode); 490 rb_erase(&head->rbnode, &q->rb_fragments); 491 492 sum_truesize = head->truesize; 493 while (rbn || fp) { 494 /* fp points to the next sk_buff in the current run; 495 * rbn points to the next run. 496 */ 497 /* Go through the current run. */ 498 while (fp) { 499 struct sk_buff *next_frag = FRAG_CB(fp)->next_frag; 500 bool stolen; 501 int delta; 502 503 sum_truesize += fp->truesize; 504 if (head->ip_summed != fp->ip_summed) 505 head->ip_summed = CHECKSUM_NONE; 506 else if (head->ip_summed == CHECKSUM_COMPLETE) 507 head->csum = csum_add(head->csum, fp->csum); 508 509 if (try_coalesce && skb_try_coalesce(head, fp, &stolen, 510 &delta)) { 511 kfree_skb_partial(fp, stolen); 512 } else { 513 fp->prev = NULL; 514 memset(&fp->rbnode, 0, sizeof(fp->rbnode)); 515 fp->sk = NULL; 516 517 head->data_len += fp->len; 518 head->len += fp->len; 519 head->truesize += fp->truesize; 520 521 *nextp = fp; 522 nextp = &fp->next; 523 } 524 525 fp = next_frag; 526 } 527 /* Move to the next run. */ 528 if (rbn) { 529 struct rb_node *rbnext = rb_next(rbn); 530 531 fp = rb_to_skb(rbn); 532 rb_erase(rbn, &q->rb_fragments); 533 rbn = rbnext; 534 } 535 } 536 sub_frag_mem_limit(q->fqdir, sum_truesize); 537 538 *nextp = NULL; 539 skb_mark_not_on_list(head); 540 head->prev = NULL; 541 head->tstamp = q->stamp; 542 } 543 EXPORT_SYMBOL(inet_frag_reasm_finish); 544 545 struct sk_buff *inet_frag_pull_head(struct inet_frag_queue *q) 546 { 547 struct sk_buff *head, *skb; 548 549 head = skb_rb_first(&q->rb_fragments); 550 if (!head) 551 return NULL; 552 skb = FRAG_CB(head)->next_frag; 553 if (skb) 554 rb_replace_node(&head->rbnode, &skb->rbnode, 555 &q->rb_fragments); 556 else 557 rb_erase(&head->rbnode, &q->rb_fragments); 558 memset(&head->rbnode, 0, sizeof(head->rbnode)); 559 barrier(); 560 561 if (head == q->fragments_tail) 562 q->fragments_tail = NULL; 563 564 sub_frag_mem_limit(q->fqdir, head->truesize); 565 566 return head; 567 } 568 EXPORT_SYMBOL(inet_frag_pull_head); 569