1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * inet fragments management
5 * Authors: Pavel Emelyanov <xemul@openvz.org>
6 * Started as consolidation of ipv4/ip_fragment.c,
7 * ipv6/reassembly. and ipv6 nf conntrack reassembly
10 #include <linux/list.h>
11 #include <linux/spinlock.h>
12 #include <linux/module.h>
13 #include <linux/timer.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>
22 #include <net/inet_frag.h>
23 #include <net/inet_ecn.h>
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.
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.
35 struct ipfrag_skb_cb {
37 struct inet_skb_parm h4;
38 struct inet6_skb_parm h6;
40 struct sk_buff *next_frag;
44 #define FRAG_CB(skb) ((struct ipfrag_skb_cb *)((skb)->cb))
46 static void fragcb_clear(struct sk_buff *skb)
48 RB_CLEAR_NODE(&skb->rbnode);
49 FRAG_CB(skb)->next_frag = NULL;
50 FRAG_CB(skb)->frag_run_len = skb->len;
53 /* Append skb to the last "run". */
54 static void fragrun_append_to_last(struct inet_frag_queue *q,
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;
64 /* Create a new "run" with the skb. */
65 static void fragrun_create(struct inet_frag_queue *q, struct sk_buff *skb)
67 BUILD_BUG_ON(sizeof(struct ipfrag_skb_cb) > sizeof(skb->cb));
71 rb_link_node(&skb->rbnode, &q->last_run_head->rbnode,
72 &q->last_run_head->rbnode.rb_right);
74 rb_link_node(&skb->rbnode, NULL, &q->rb_fragments.rb_node);
75 rb_insert_color(&skb->rbnode, &q->rb_fragments);
77 q->fragments_tail = skb;
78 q->last_run_head = skb;
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
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,
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,
100 EXPORT_SYMBOL(ip_frag_ecn_table);
102 int inet_frags_init(struct inet_frags *f)
104 f->frags_cachep = kmem_cache_create(f->frags_cache_name, f->qsize, 0, 0,
106 if (!f->frags_cachep)
109 refcount_set(&f->refcnt, 1);
110 init_completion(&f->completion);
113 EXPORT_SYMBOL(inet_frags_init);
115 void inet_frags_fini(struct inet_frags *f)
117 if (refcount_dec_and_test(&f->refcnt))
118 complete(&f->completion);
120 wait_for_completion(&f->completion);
122 kmem_cache_destroy(f->frags_cachep);
123 f->frags_cachep = NULL;
125 EXPORT_SYMBOL(inet_frags_fini);
127 /* called from rhashtable_free_and_destroy() at netns_frags dismantle */
128 static void inet_frags_free_cb(void *ptr, void *arg)
130 struct inet_frag_queue *fq = ptr;
133 count = del_timer_sync(&fq->timer) ? 1 : 0;
135 spin_lock_bh(&fq->lock);
136 if (!(fq->flags & INET_FRAG_COMPLETE)) {
137 fq->flags |= INET_FRAG_COMPLETE;
139 } else if (fq->flags & INET_FRAG_HASH_DEAD) {
142 spin_unlock_bh(&fq->lock);
144 if (refcount_sub_and_test(count, &fq->refcnt))
145 inet_frag_destroy(fq);
148 static void fqdir_work_fn(struct work_struct *work)
150 struct fqdir *fqdir = container_of(work, struct fqdir, destroy_work);
151 struct inet_frags *f = fqdir->f;
153 rhashtable_free_and_destroy(&fqdir->rhashtable, inet_frags_free_cb, NULL);
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() ? :)
161 if (refcount_dec_and_test(&f->refcnt))
162 complete(&f->completion);
167 int fqdir_init(struct fqdir **fqdirp, struct inet_frags *f, struct net *net)
169 struct fqdir *fqdir = kzalloc(sizeof(*fqdir), GFP_KERNEL);
176 res = rhashtable_init(&fqdir->rhashtable, &fqdir->f->rhash_params);
181 refcount_inc(&f->refcnt);
185 EXPORT_SYMBOL(fqdir_init);
187 void fqdir_exit(struct fqdir *fqdir)
189 INIT_WORK(&fqdir->destroy_work, fqdir_work_fn);
190 queue_work(system_wq, &fqdir->destroy_work);
192 EXPORT_SYMBOL(fqdir_exit);
194 void inet_frag_kill(struct inet_frag_queue *fq)
196 if (del_timer(&fq->timer))
197 refcount_dec(&fq->refcnt);
199 if (!(fq->flags & INET_FRAG_COMPLETE)) {
200 struct fqdir *fqdir = fq->fqdir;
202 fq->flags |= INET_FRAG_COMPLETE;
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 fqdir_pre_exit().
209 if (!READ_ONCE(fqdir->dead)) {
210 rhashtable_remove_fast(&fqdir->rhashtable, &fq->node,
211 fqdir->f->rhash_params);
212 refcount_dec(&fq->refcnt);
214 fq->flags |= INET_FRAG_HASH_DEAD;
219 EXPORT_SYMBOL(inet_frag_kill);
221 static void inet_frag_destroy_rcu(struct rcu_head *head)
223 struct inet_frag_queue *q = container_of(head, struct inet_frag_queue,
225 struct inet_frags *f = q->fqdir->f;
229 kmem_cache_free(f->frags_cachep, q);
232 unsigned int inet_frag_rbtree_purge(struct rb_root *root)
234 struct rb_node *p = rb_first(root);
235 unsigned int sum = 0;
238 struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode);
241 rb_erase(&skb->rbnode, root);
243 struct sk_buff *next = FRAG_CB(skb)->next_frag;
245 sum += skb->truesize;
252 EXPORT_SYMBOL(inet_frag_rbtree_purge);
254 void inet_frag_destroy(struct inet_frag_queue *q)
257 unsigned int sum, sum_truesize = 0;
258 struct inet_frags *f;
260 WARN_ON(!(q->flags & INET_FRAG_COMPLETE));
261 WARN_ON(del_timer(&q->timer) != 0);
263 /* Release all fragment data. */
266 sum_truesize = inet_frag_rbtree_purge(&q->rb_fragments);
267 sum = sum_truesize + f->qsize;
269 call_rcu(&q->rcu, inet_frag_destroy_rcu);
271 sub_frag_mem_limit(fqdir, sum);
273 EXPORT_SYMBOL(inet_frag_destroy);
275 static struct inet_frag_queue *inet_frag_alloc(struct fqdir *fqdir,
276 struct inet_frags *f,
279 struct inet_frag_queue *q;
281 q = kmem_cache_zalloc(f->frags_cachep, GFP_ATOMIC);
286 f->constructor(q, arg);
287 add_frag_mem_limit(fqdir, f->qsize);
289 timer_setup(&q->timer, f->frag_expire, 0);
290 spin_lock_init(&q->lock);
291 refcount_set(&q->refcnt, 3);
296 static struct inet_frag_queue *inet_frag_create(struct fqdir *fqdir,
298 struct inet_frag_queue **prev)
300 struct inet_frags *f = fqdir->f;
301 struct inet_frag_queue *q;
303 q = inet_frag_alloc(fqdir, f, arg);
305 *prev = ERR_PTR(-ENOMEM);
308 mod_timer(&q->timer, jiffies + fqdir->timeout);
310 *prev = rhashtable_lookup_get_insert_key(&fqdir->rhashtable, &q->key,
311 &q->node, f->rhash_params);
313 q->flags |= INET_FRAG_COMPLETE;
315 inet_frag_destroy(q);
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)
324 /* This pairs with WRITE_ONCE() in fqdir_pre_exit(). */
325 long high_thresh = READ_ONCE(fqdir->high_thresh);
326 struct inet_frag_queue *fq = NULL, *prev;
328 if (!high_thresh || frag_mem_limit(fqdir) > high_thresh)
333 prev = rhashtable_lookup(&fqdir->rhashtable, key, fqdir->f->rhash_params);
335 fq = inet_frag_create(fqdir, key, &prev);
336 if (!IS_ERR_OR_NULL(prev)) {
338 if (!refcount_inc_not_zero(&fq->refcnt))
344 EXPORT_SYMBOL(inet_frag_find);
346 int inet_frag_queue_insert(struct inet_frag_queue *q, struct sk_buff *skb,
349 struct sk_buff *last = q->fragments_tail;
351 /* RFC5722, Section 4, amended by Errata ID : 3089
352 * When reassembling an IPv6 datagram, if
353 * one or more its constituent fragments is determined to be an
354 * overlapping fragment, the entire datagram (and any constituent
355 * fragments) MUST be silently discarded.
357 * Duplicates, however, should be ignored (i.e. skb dropped, but the
358 * queue/fragments kept for later reassembly).
361 fragrun_create(q, skb); /* First fragment. */
362 else if (last->ip_defrag_offset + last->len < end) {
363 /* This is the common case: skb goes to the end. */
364 /* Detect and discard overlaps. */
365 if (offset < last->ip_defrag_offset + last->len)
366 return IPFRAG_OVERLAP;
367 if (offset == last->ip_defrag_offset + last->len)
368 fragrun_append_to_last(q, skb);
370 fragrun_create(q, skb);
372 /* Binary search. Note that skb can become the first fragment,
373 * but not the last (covered above).
375 struct rb_node **rbn, *parent;
377 rbn = &q->rb_fragments.rb_node;
379 struct sk_buff *curr;
383 curr = rb_to_skb(parent);
384 curr_run_end = curr->ip_defrag_offset +
385 FRAG_CB(curr)->frag_run_len;
386 if (end <= curr->ip_defrag_offset)
387 rbn = &parent->rb_left;
388 else if (offset >= curr_run_end)
389 rbn = &parent->rb_right;
390 else if (offset >= curr->ip_defrag_offset &&
394 return IPFRAG_OVERLAP;
396 /* Here we have parent properly set, and rbn pointing to
397 * one of its NULL left/right children. Insert skb.
400 rb_link_node(&skb->rbnode, parent, rbn);
401 rb_insert_color(&skb->rbnode, &q->rb_fragments);
404 skb->ip_defrag_offset = offset;
408 EXPORT_SYMBOL(inet_frag_queue_insert);
410 void *inet_frag_reasm_prepare(struct inet_frag_queue *q, struct sk_buff *skb,
411 struct sk_buff *parent)
413 struct sk_buff *fp, *head = skb_rb_first(&q->rb_fragments);
414 struct sk_buff **nextp;
418 fp = skb_clone(skb, GFP_ATOMIC);
421 FRAG_CB(fp)->next_frag = FRAG_CB(skb)->next_frag;
422 if (RB_EMPTY_NODE(&skb->rbnode))
423 FRAG_CB(parent)->next_frag = fp;
425 rb_replace_node(&skb->rbnode, &fp->rbnode,
427 if (q->fragments_tail == skb)
428 q->fragments_tail = fp;
429 skb_morph(skb, head);
430 FRAG_CB(skb)->next_frag = FRAG_CB(head)->next_frag;
431 rb_replace_node(&head->rbnode, &skb->rbnode,
436 WARN_ON(head->ip_defrag_offset != 0);
438 delta = -head->truesize;
440 /* Head of list must not be cloned. */
441 if (skb_unclone(head, GFP_ATOMIC))
444 delta += head->truesize;
446 add_frag_mem_limit(q->fqdir, delta);
448 /* If the first fragment is fragmented itself, we split
449 * it to two chunks: the first with data and paged part
450 * and the second, holding only fragments.
452 if (skb_has_frag_list(head)) {
453 struct sk_buff *clone;
456 clone = alloc_skb(0, GFP_ATOMIC);
459 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
460 skb_frag_list_init(head);
461 for (i = 0; i < skb_shinfo(head)->nr_frags; i++)
462 plen += skb_frag_size(&skb_shinfo(head)->frags[i]);
463 clone->data_len = head->data_len - plen;
464 clone->len = clone->data_len;
465 head->truesize += clone->truesize;
467 clone->ip_summed = head->ip_summed;
468 add_frag_mem_limit(q->fqdir, clone->truesize);
469 skb_shinfo(head)->frag_list = clone;
470 nextp = &clone->next;
472 nextp = &skb_shinfo(head)->frag_list;
477 EXPORT_SYMBOL(inet_frag_reasm_prepare);
479 void inet_frag_reasm_finish(struct inet_frag_queue *q, struct sk_buff *head,
480 void *reasm_data, bool try_coalesce)
482 struct sk_buff **nextp = (struct sk_buff **)reasm_data;
487 skb_push(head, head->data - skb_network_header(head));
489 /* Traverse the tree in order, to build frag_list. */
490 fp = FRAG_CB(head)->next_frag;
491 rbn = rb_next(&head->rbnode);
492 rb_erase(&head->rbnode, &q->rb_fragments);
494 sum_truesize = head->truesize;
496 /* fp points to the next sk_buff in the current run;
497 * rbn points to the next run.
499 /* Go through the current run. */
501 struct sk_buff *next_frag = FRAG_CB(fp)->next_frag;
505 sum_truesize += fp->truesize;
506 if (head->ip_summed != fp->ip_summed)
507 head->ip_summed = CHECKSUM_NONE;
508 else if (head->ip_summed == CHECKSUM_COMPLETE)
509 head->csum = csum_add(head->csum, fp->csum);
511 if (try_coalesce && skb_try_coalesce(head, fp, &stolen,
513 kfree_skb_partial(fp, stolen);
516 memset(&fp->rbnode, 0, sizeof(fp->rbnode));
519 head->data_len += fp->len;
520 head->len += fp->len;
521 head->truesize += fp->truesize;
529 /* Move to the next run. */
531 struct rb_node *rbnext = rb_next(rbn);
534 rb_erase(rbn, &q->rb_fragments);
538 sub_frag_mem_limit(q->fqdir, sum_truesize);
541 skb_mark_not_on_list(head);
543 head->tstamp = q->stamp;
545 EXPORT_SYMBOL(inet_frag_reasm_finish);
547 struct sk_buff *inet_frag_pull_head(struct inet_frag_queue *q)
549 struct sk_buff *head, *skb;
551 head = skb_rb_first(&q->rb_fragments);
554 skb = FRAG_CB(head)->next_frag;
556 rb_replace_node(&head->rbnode, &skb->rbnode,
559 rb_erase(&head->rbnode, &q->rb_fragments);
560 memset(&head->rbnode, 0, sizeof(head->rbnode));
563 if (head == q->fragments_tail)
564 q->fragments_tail = NULL;
566 sub_frag_mem_limit(q->fqdir, head->truesize);
570 EXPORT_SYMBOL(inet_frag_pull_head);