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 #include "../core/sock_destructor.h" 28 29 /* Use skb->cb to track consecutive/adjacent fragments coming at 30 * the end of the queue. Nodes in the rb-tree queue will 31 * contain "runs" of one or more adjacent fragments. 32 * 33 * Invariants: 34 * - next_frag is NULL at the tail of a "run"; 35 * - the head of a "run" has the sum of all fragment lengths in frag_run_len. 36 */ 37 struct ipfrag_skb_cb { 38 union { 39 struct inet_skb_parm h4; 40 struct inet6_skb_parm h6; 41 }; 42 struct sk_buff *next_frag; 43 int frag_run_len; 44 int ip_defrag_offset; 45 }; 46 47 #define FRAG_CB(skb) ((struct ipfrag_skb_cb *)((skb)->cb)) 48 49 static void fragcb_clear(struct sk_buff *skb) 50 { 51 RB_CLEAR_NODE(&skb->rbnode); 52 FRAG_CB(skb)->next_frag = NULL; 53 FRAG_CB(skb)->frag_run_len = skb->len; 54 } 55 56 /* Append skb to the last "run". */ 57 static void fragrun_append_to_last(struct inet_frag_queue *q, 58 struct sk_buff *skb) 59 { 60 fragcb_clear(skb); 61 62 FRAG_CB(q->last_run_head)->frag_run_len += skb->len; 63 FRAG_CB(q->fragments_tail)->next_frag = skb; 64 q->fragments_tail = skb; 65 } 66 67 /* Create a new "run" with the skb. */ 68 static void fragrun_create(struct inet_frag_queue *q, struct sk_buff *skb) 69 { 70 BUILD_BUG_ON(sizeof(struct ipfrag_skb_cb) > sizeof(skb->cb)); 71 fragcb_clear(skb); 72 73 if (q->last_run_head) 74 rb_link_node(&skb->rbnode, &q->last_run_head->rbnode, 75 &q->last_run_head->rbnode.rb_right); 76 else 77 rb_link_node(&skb->rbnode, NULL, &q->rb_fragments.rb_node); 78 rb_insert_color(&skb->rbnode, &q->rb_fragments); 79 80 q->fragments_tail = skb; 81 q->last_run_head = skb; 82 } 83 84 /* Given the OR values of all fragments, apply RFC 3168 5.3 requirements 85 * Value : 0xff if frame should be dropped. 86 * 0 or INET_ECN_CE value, to be ORed in to final iph->tos field 87 */ 88 const u8 ip_frag_ecn_table[16] = { 89 /* at least one fragment had CE, and others ECT_0 or ECT_1 */ 90 [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0] = INET_ECN_CE, 91 [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1] = INET_ECN_CE, 92 [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = INET_ECN_CE, 93 94 /* invalid combinations : drop frame */ 95 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE] = 0xff, 96 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0] = 0xff, 97 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_1] = 0xff, 98 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff, 99 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0] = 0xff, 100 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1] = 0xff, 101 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff, 102 }; 103 EXPORT_SYMBOL(ip_frag_ecn_table); 104 105 int inet_frags_init(struct inet_frags *f) 106 { 107 f->frags_cachep = kmem_cache_create(f->frags_cache_name, f->qsize, 0, 0, 108 NULL); 109 if (!f->frags_cachep) 110 return -ENOMEM; 111 112 refcount_set(&f->refcnt, 1); 113 init_completion(&f->completion); 114 return 0; 115 } 116 EXPORT_SYMBOL(inet_frags_init); 117 118 void inet_frags_fini(struct inet_frags *f) 119 { 120 if (refcount_dec_and_test(&f->refcnt)) 121 complete(&f->completion); 122 123 wait_for_completion(&f->completion); 124 125 kmem_cache_destroy(f->frags_cachep); 126 f->frags_cachep = NULL; 127 } 128 EXPORT_SYMBOL(inet_frags_fini); 129 130 /* called from rhashtable_free_and_destroy() at netns_frags dismantle */ 131 static void inet_frags_free_cb(void *ptr, void *arg) 132 { 133 struct inet_frag_queue *fq = ptr; 134 int count; 135 136 count = timer_delete_sync(&fq->timer) ? 1 : 0; 137 138 spin_lock_bh(&fq->lock); 139 fq->flags |= INET_FRAG_DROP; 140 if (!(fq->flags & INET_FRAG_COMPLETE)) { 141 fq->flags |= INET_FRAG_COMPLETE; 142 count++; 143 } else if (fq->flags & INET_FRAG_HASH_DEAD) { 144 count++; 145 } 146 spin_unlock_bh(&fq->lock); 147 148 inet_frag_putn(fq, count); 149 } 150 151 static LLIST_HEAD(fqdir_free_list); 152 153 static void fqdir_free_fn(struct work_struct *work) 154 { 155 struct llist_node *kill_list; 156 struct fqdir *fqdir, *tmp; 157 struct inet_frags *f; 158 159 /* Atomically snapshot the list of fqdirs to free */ 160 kill_list = llist_del_all(&fqdir_free_list); 161 162 /* We need to make sure all ongoing call_rcu(..., inet_frag_destroy_rcu) 163 * have completed, since they need to dereference fqdir. 164 * Would it not be nice to have kfree_rcu_barrier() ? :) 165 */ 166 rcu_barrier(); 167 168 llist_for_each_entry_safe(fqdir, tmp, kill_list, free_list) { 169 f = fqdir->f; 170 if (refcount_dec_and_test(&f->refcnt)) 171 complete(&f->completion); 172 173 kfree(fqdir); 174 } 175 } 176 177 static DECLARE_DELAYED_WORK(fqdir_free_work, fqdir_free_fn); 178 179 static void fqdir_work_fn(struct work_struct *work) 180 { 181 struct fqdir *fqdir = container_of(work, struct fqdir, destroy_work); 182 183 rhashtable_free_and_destroy(&fqdir->rhashtable, inet_frags_free_cb, NULL); 184 185 if (llist_add(&fqdir->free_list, &fqdir_free_list)) 186 queue_delayed_work(system_wq, &fqdir_free_work, HZ); 187 } 188 189 int fqdir_init(struct fqdir **fqdirp, struct inet_frags *f, struct net *net) 190 { 191 struct fqdir *fqdir = kzalloc(sizeof(*fqdir), GFP_KERNEL); 192 int res; 193 194 if (!fqdir) 195 return -ENOMEM; 196 fqdir->f = f; 197 fqdir->net = net; 198 res = rhashtable_init(&fqdir->rhashtable, &fqdir->f->rhash_params); 199 if (res < 0) { 200 kfree(fqdir); 201 return res; 202 } 203 refcount_inc(&f->refcnt); 204 *fqdirp = fqdir; 205 return 0; 206 } 207 EXPORT_SYMBOL(fqdir_init); 208 209 static struct workqueue_struct *inet_frag_wq; 210 211 static int __init inet_frag_wq_init(void) 212 { 213 inet_frag_wq = create_workqueue("inet_frag_wq"); 214 if (!inet_frag_wq) 215 panic("Could not create inet frag workq"); 216 return 0; 217 } 218 219 pure_initcall(inet_frag_wq_init); 220 221 void fqdir_exit(struct fqdir *fqdir) 222 { 223 INIT_WORK(&fqdir->destroy_work, fqdir_work_fn); 224 queue_work(inet_frag_wq, &fqdir->destroy_work); 225 } 226 EXPORT_SYMBOL(fqdir_exit); 227 228 void inet_frag_kill(struct inet_frag_queue *fq, int *refs) 229 { 230 if (timer_delete(&fq->timer)) 231 (*refs)++; 232 233 if (!(fq->flags & INET_FRAG_COMPLETE)) { 234 struct fqdir *fqdir = fq->fqdir; 235 236 fq->flags |= INET_FRAG_COMPLETE; 237 rcu_read_lock(); 238 /* The RCU read lock provides a memory barrier 239 * guaranteeing that if fqdir->dead is false then 240 * the hash table destruction will not start until 241 * after we unlock. Paired with fqdir_pre_exit(). 242 */ 243 if (!READ_ONCE(fqdir->dead)) { 244 rhashtable_remove_fast(&fqdir->rhashtable, &fq->node, 245 fqdir->f->rhash_params); 246 (*refs)++; 247 } else { 248 fq->flags |= INET_FRAG_HASH_DEAD; 249 } 250 rcu_read_unlock(); 251 } 252 } 253 EXPORT_SYMBOL(inet_frag_kill); 254 255 static void inet_frag_destroy_rcu(struct rcu_head *head) 256 { 257 struct inet_frag_queue *q = container_of(head, struct inet_frag_queue, 258 rcu); 259 struct inet_frags *f = q->fqdir->f; 260 261 if (f->destructor) 262 f->destructor(q); 263 kmem_cache_free(f->frags_cachep, q); 264 } 265 266 unsigned int inet_frag_rbtree_purge(struct rb_root *root, 267 enum skb_drop_reason reason) 268 { 269 struct rb_node *p = rb_first(root); 270 unsigned int sum = 0; 271 272 while (p) { 273 struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode); 274 275 p = rb_next(p); 276 rb_erase(&skb->rbnode, root); 277 while (skb) { 278 struct sk_buff *next = FRAG_CB(skb)->next_frag; 279 280 sum += skb->truesize; 281 kfree_skb_reason(skb, reason); 282 skb = next; 283 } 284 } 285 return sum; 286 } 287 EXPORT_SYMBOL(inet_frag_rbtree_purge); 288 289 void inet_frag_destroy(struct inet_frag_queue *q) 290 { 291 unsigned int sum, sum_truesize = 0; 292 enum skb_drop_reason reason; 293 struct inet_frags *f; 294 struct fqdir *fqdir; 295 296 WARN_ON(!(q->flags & INET_FRAG_COMPLETE)); 297 reason = (q->flags & INET_FRAG_DROP) ? 298 SKB_DROP_REASON_FRAG_REASM_TIMEOUT : 299 SKB_CONSUMED; 300 WARN_ON(timer_delete(&q->timer) != 0); 301 302 /* Release all fragment data. */ 303 fqdir = q->fqdir; 304 f = fqdir->f; 305 sum_truesize = inet_frag_rbtree_purge(&q->rb_fragments, reason); 306 sum = sum_truesize + f->qsize; 307 308 call_rcu(&q->rcu, inet_frag_destroy_rcu); 309 310 sub_frag_mem_limit(fqdir, sum); 311 } 312 EXPORT_SYMBOL(inet_frag_destroy); 313 314 static struct inet_frag_queue *inet_frag_alloc(struct fqdir *fqdir, 315 struct inet_frags *f, 316 void *arg) 317 { 318 struct inet_frag_queue *q; 319 320 q = kmem_cache_zalloc(f->frags_cachep, GFP_ATOMIC); 321 if (!q) 322 return NULL; 323 324 q->fqdir = fqdir; 325 f->constructor(q, arg); 326 add_frag_mem_limit(fqdir, f->qsize); 327 328 timer_setup(&q->timer, f->frag_expire, 0); 329 spin_lock_init(&q->lock); 330 /* One reference for the timer, one for the hash table. */ 331 refcount_set(&q->refcnt, 2); 332 333 return q; 334 } 335 336 static struct inet_frag_queue *inet_frag_create(struct fqdir *fqdir, 337 void *arg, 338 struct inet_frag_queue **prev) 339 { 340 struct inet_frags *f = fqdir->f; 341 struct inet_frag_queue *q; 342 343 q = inet_frag_alloc(fqdir, f, arg); 344 if (!q) { 345 *prev = ERR_PTR(-ENOMEM); 346 return NULL; 347 } 348 mod_timer(&q->timer, jiffies + fqdir->timeout); 349 350 *prev = rhashtable_lookup_get_insert_key(&fqdir->rhashtable, &q->key, 351 &q->node, f->rhash_params); 352 if (*prev) { 353 /* We could not insert in the hash table, 354 * we need to cancel what inet_frag_alloc() 355 * anticipated. 356 */ 357 int refs = 1; 358 359 q->flags |= INET_FRAG_COMPLETE; 360 inet_frag_kill(q, &refs); 361 inet_frag_putn(q, refs); 362 return NULL; 363 } 364 return q; 365 } 366 367 struct inet_frag_queue *inet_frag_find(struct fqdir *fqdir, void *key) 368 { 369 /* This pairs with WRITE_ONCE() in fqdir_pre_exit(). */ 370 long high_thresh = READ_ONCE(fqdir->high_thresh); 371 struct inet_frag_queue *fq = NULL, *prev; 372 373 if (!high_thresh || frag_mem_limit(fqdir) > high_thresh) 374 return NULL; 375 376 prev = rhashtable_lookup(&fqdir->rhashtable, key, fqdir->f->rhash_params); 377 if (!prev) 378 fq = inet_frag_create(fqdir, key, &prev); 379 if (!IS_ERR_OR_NULL(prev)) 380 fq = prev; 381 return fq; 382 } 383 EXPORT_SYMBOL(inet_frag_find); 384 385 int inet_frag_queue_insert(struct inet_frag_queue *q, struct sk_buff *skb, 386 int offset, int end) 387 { 388 struct sk_buff *last = q->fragments_tail; 389 390 /* RFC5722, Section 4, amended by Errata ID : 3089 391 * When reassembling an IPv6 datagram, if 392 * one or more its constituent fragments is determined to be an 393 * overlapping fragment, the entire datagram (and any constituent 394 * fragments) MUST be silently discarded. 395 * 396 * Duplicates, however, should be ignored (i.e. skb dropped, but the 397 * queue/fragments kept for later reassembly). 398 */ 399 if (!last) 400 fragrun_create(q, skb); /* First fragment. */ 401 else if (FRAG_CB(last)->ip_defrag_offset + last->len < end) { 402 /* This is the common case: skb goes to the end. */ 403 /* Detect and discard overlaps. */ 404 if (offset < FRAG_CB(last)->ip_defrag_offset + last->len) 405 return IPFRAG_OVERLAP; 406 if (offset == FRAG_CB(last)->ip_defrag_offset + last->len) 407 fragrun_append_to_last(q, skb); 408 else 409 fragrun_create(q, skb); 410 } else { 411 /* Binary search. Note that skb can become the first fragment, 412 * but not the last (covered above). 413 */ 414 struct rb_node **rbn, *parent; 415 416 rbn = &q->rb_fragments.rb_node; 417 do { 418 struct sk_buff *curr; 419 int curr_run_end; 420 421 parent = *rbn; 422 curr = rb_to_skb(parent); 423 curr_run_end = FRAG_CB(curr)->ip_defrag_offset + 424 FRAG_CB(curr)->frag_run_len; 425 if (end <= FRAG_CB(curr)->ip_defrag_offset) 426 rbn = &parent->rb_left; 427 else if (offset >= curr_run_end) 428 rbn = &parent->rb_right; 429 else if (offset >= FRAG_CB(curr)->ip_defrag_offset && 430 end <= curr_run_end) 431 return IPFRAG_DUP; 432 else 433 return IPFRAG_OVERLAP; 434 } while (*rbn); 435 /* Here we have parent properly set, and rbn pointing to 436 * one of its NULL left/right children. Insert skb. 437 */ 438 fragcb_clear(skb); 439 rb_link_node(&skb->rbnode, parent, rbn); 440 rb_insert_color(&skb->rbnode, &q->rb_fragments); 441 } 442 443 FRAG_CB(skb)->ip_defrag_offset = offset; 444 445 return IPFRAG_OK; 446 } 447 EXPORT_SYMBOL(inet_frag_queue_insert); 448 449 void *inet_frag_reasm_prepare(struct inet_frag_queue *q, struct sk_buff *skb, 450 struct sk_buff *parent) 451 { 452 struct sk_buff *fp, *head = skb_rb_first(&q->rb_fragments); 453 void (*destructor)(struct sk_buff *); 454 unsigned int orig_truesize = 0; 455 struct sk_buff **nextp = NULL; 456 struct sock *sk = skb->sk; 457 int delta; 458 459 if (sk && is_skb_wmem(skb)) { 460 /* TX: skb->sk might have been passed as argument to 461 * dst->output and must remain valid until tx completes. 462 * 463 * Move sk to reassembled skb and fix up wmem accounting. 464 */ 465 orig_truesize = skb->truesize; 466 destructor = skb->destructor; 467 } 468 469 if (head != skb) { 470 fp = skb_clone(skb, GFP_ATOMIC); 471 if (!fp) { 472 head = skb; 473 goto out_restore_sk; 474 } 475 FRAG_CB(fp)->next_frag = FRAG_CB(skb)->next_frag; 476 if (RB_EMPTY_NODE(&skb->rbnode)) 477 FRAG_CB(parent)->next_frag = fp; 478 else 479 rb_replace_node(&skb->rbnode, &fp->rbnode, 480 &q->rb_fragments); 481 if (q->fragments_tail == skb) 482 q->fragments_tail = fp; 483 484 if (orig_truesize) { 485 /* prevent skb_morph from releasing sk */ 486 skb->sk = NULL; 487 skb->destructor = NULL; 488 } 489 skb_morph(skb, head); 490 FRAG_CB(skb)->next_frag = FRAG_CB(head)->next_frag; 491 rb_replace_node(&head->rbnode, &skb->rbnode, 492 &q->rb_fragments); 493 consume_skb(head); 494 head = skb; 495 } 496 WARN_ON(FRAG_CB(head)->ip_defrag_offset != 0); 497 498 delta = -head->truesize; 499 500 /* Head of list must not be cloned. */ 501 if (skb_unclone(head, GFP_ATOMIC)) 502 goto out_restore_sk; 503 504 delta += head->truesize; 505 if (delta) 506 add_frag_mem_limit(q->fqdir, delta); 507 508 /* If the first fragment is fragmented itself, we split 509 * it to two chunks: the first with data and paged part 510 * and the second, holding only fragments. 511 */ 512 if (skb_has_frag_list(head)) { 513 struct sk_buff *clone; 514 int i, plen = 0; 515 516 clone = alloc_skb(0, GFP_ATOMIC); 517 if (!clone) 518 goto out_restore_sk; 519 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list; 520 skb_frag_list_init(head); 521 for (i = 0; i < skb_shinfo(head)->nr_frags; i++) 522 plen += skb_frag_size(&skb_shinfo(head)->frags[i]); 523 clone->data_len = head->data_len - plen; 524 clone->len = clone->data_len; 525 head->truesize += clone->truesize; 526 clone->csum = 0; 527 clone->ip_summed = head->ip_summed; 528 add_frag_mem_limit(q->fqdir, clone->truesize); 529 skb_shinfo(head)->frag_list = clone; 530 nextp = &clone->next; 531 } else { 532 nextp = &skb_shinfo(head)->frag_list; 533 } 534 535 out_restore_sk: 536 if (orig_truesize) { 537 int ts_delta = head->truesize - orig_truesize; 538 539 /* if this reassembled skb is fragmented later, 540 * fraglist skbs will get skb->sk assigned from head->sk, 541 * and each frag skb will be released via sock_wfree. 542 * 543 * Update sk_wmem_alloc. 544 */ 545 head->sk = sk; 546 head->destructor = destructor; 547 refcount_add(ts_delta, &sk->sk_wmem_alloc); 548 } 549 550 return nextp; 551 } 552 EXPORT_SYMBOL(inet_frag_reasm_prepare); 553 554 void inet_frag_reasm_finish(struct inet_frag_queue *q, struct sk_buff *head, 555 void *reasm_data, bool try_coalesce) 556 { 557 struct sock *sk = is_skb_wmem(head) ? head->sk : NULL; 558 const unsigned int head_truesize = head->truesize; 559 struct sk_buff **nextp = reasm_data; 560 struct rb_node *rbn; 561 struct sk_buff *fp; 562 int sum_truesize; 563 564 skb_push(head, head->data - skb_network_header(head)); 565 566 /* Traverse the tree in order, to build frag_list. */ 567 fp = FRAG_CB(head)->next_frag; 568 rbn = rb_next(&head->rbnode); 569 rb_erase(&head->rbnode, &q->rb_fragments); 570 571 sum_truesize = head->truesize; 572 while (rbn || fp) { 573 /* fp points to the next sk_buff in the current run; 574 * rbn points to the next run. 575 */ 576 /* Go through the current run. */ 577 while (fp) { 578 struct sk_buff *next_frag = FRAG_CB(fp)->next_frag; 579 bool stolen; 580 int delta; 581 582 sum_truesize += fp->truesize; 583 if (head->ip_summed != fp->ip_summed) 584 head->ip_summed = CHECKSUM_NONE; 585 else if (head->ip_summed == CHECKSUM_COMPLETE) 586 head->csum = csum_add(head->csum, fp->csum); 587 588 if (try_coalesce && skb_try_coalesce(head, fp, &stolen, 589 &delta)) { 590 kfree_skb_partial(fp, stolen); 591 } else { 592 fp->prev = NULL; 593 memset(&fp->rbnode, 0, sizeof(fp->rbnode)); 594 fp->sk = NULL; 595 596 head->data_len += fp->len; 597 head->len += fp->len; 598 head->truesize += fp->truesize; 599 600 *nextp = fp; 601 nextp = &fp->next; 602 } 603 604 fp = next_frag; 605 } 606 /* Move to the next run. */ 607 if (rbn) { 608 struct rb_node *rbnext = rb_next(rbn); 609 610 fp = rb_to_skb(rbn); 611 rb_erase(rbn, &q->rb_fragments); 612 rbn = rbnext; 613 } 614 } 615 sub_frag_mem_limit(q->fqdir, sum_truesize); 616 617 *nextp = NULL; 618 skb_mark_not_on_list(head); 619 head->prev = NULL; 620 head->tstamp = q->stamp; 621 head->tstamp_type = q->tstamp_type; 622 623 if (sk) 624 refcount_add(sum_truesize - head_truesize, &sk->sk_wmem_alloc); 625 } 626 EXPORT_SYMBOL(inet_frag_reasm_finish); 627 628 struct sk_buff *inet_frag_pull_head(struct inet_frag_queue *q) 629 { 630 struct sk_buff *head, *skb; 631 632 head = skb_rb_first(&q->rb_fragments); 633 if (!head) 634 return NULL; 635 skb = FRAG_CB(head)->next_frag; 636 if (skb) 637 rb_replace_node(&head->rbnode, &skb->rbnode, 638 &q->rb_fragments); 639 else 640 rb_erase(&head->rbnode, &q->rb_fragments); 641 memset(&head->rbnode, 0, sizeof(head->rbnode)); 642 barrier(); 643 644 if (head == q->fragments_tail) 645 q->fragments_tail = NULL; 646 647 sub_frag_mem_limit(q->fqdir, head->truesize); 648 649 return head; 650 } 651 EXPORT_SYMBOL(inet_frag_pull_head); 652