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 fq->flags |= INET_FRAG_DROP;
137 if (!(fq->flags & INET_FRAG_COMPLETE)) {
138 fq->flags |= INET_FRAG_COMPLETE;
140 } else if (fq->flags & INET_FRAG_HASH_DEAD) {
143 spin_unlock_bh(&fq->lock);
145 if (refcount_sub_and_test(count, &fq->refcnt))
146 inet_frag_destroy(fq);
149 static LLIST_HEAD(fqdir_free_list);
151 static void fqdir_free_fn(struct work_struct *work)
153 struct llist_node *kill_list;
154 struct fqdir *fqdir, *tmp;
155 struct inet_frags *f;
157 /* Atomically snapshot the list of fqdirs to free */
158 kill_list = llist_del_all(&fqdir_free_list);
160 /* We need to make sure all ongoing call_rcu(..., inet_frag_destroy_rcu)
161 * have completed, since they need to dereference fqdir.
162 * Would it not be nice to have kfree_rcu_barrier() ? :)
166 llist_for_each_entry_safe(fqdir, tmp, kill_list, free_list) {
168 if (refcount_dec_and_test(&f->refcnt))
169 complete(&f->completion);
175 static DECLARE_WORK(fqdir_free_work, fqdir_free_fn);
177 static void fqdir_work_fn(struct work_struct *work)
179 struct fqdir *fqdir = container_of(work, struct fqdir, destroy_work);
181 rhashtable_free_and_destroy(&fqdir->rhashtable, inet_frags_free_cb, NULL);
183 if (llist_add(&fqdir->free_list, &fqdir_free_list))
184 queue_work(system_wq, &fqdir_free_work);
187 int fqdir_init(struct fqdir **fqdirp, struct inet_frags *f, struct net *net)
189 struct fqdir *fqdir = kzalloc(sizeof(*fqdir), GFP_KERNEL);
196 res = rhashtable_init(&fqdir->rhashtable, &fqdir->f->rhash_params);
201 refcount_inc(&f->refcnt);
205 EXPORT_SYMBOL(fqdir_init);
207 static struct workqueue_struct *inet_frag_wq;
209 static int __init inet_frag_wq_init(void)
211 inet_frag_wq = create_workqueue("inet_frag_wq");
213 panic("Could not create inet frag workq");
217 pure_initcall(inet_frag_wq_init);
219 void fqdir_exit(struct fqdir *fqdir)
221 INIT_WORK(&fqdir->destroy_work, fqdir_work_fn);
222 queue_work(inet_frag_wq, &fqdir->destroy_work);
224 EXPORT_SYMBOL(fqdir_exit);
226 void inet_frag_kill(struct inet_frag_queue *fq)
228 if (del_timer(&fq->timer))
229 refcount_dec(&fq->refcnt);
231 if (!(fq->flags & INET_FRAG_COMPLETE)) {
232 struct fqdir *fqdir = fq->fqdir;
234 fq->flags |= INET_FRAG_COMPLETE;
236 /* The RCU read lock provides a memory barrier
237 * guaranteeing that if fqdir->dead is false then
238 * the hash table destruction will not start until
239 * after we unlock. Paired with fqdir_pre_exit().
241 if (!READ_ONCE(fqdir->dead)) {
242 rhashtable_remove_fast(&fqdir->rhashtable, &fq->node,
243 fqdir->f->rhash_params);
244 refcount_dec(&fq->refcnt);
246 fq->flags |= INET_FRAG_HASH_DEAD;
251 EXPORT_SYMBOL(inet_frag_kill);
253 static void inet_frag_destroy_rcu(struct rcu_head *head)
255 struct inet_frag_queue *q = container_of(head, struct inet_frag_queue,
257 struct inet_frags *f = q->fqdir->f;
261 kmem_cache_free(f->frags_cachep, q);
264 unsigned int inet_frag_rbtree_purge(struct rb_root *root,
265 enum skb_drop_reason reason)
267 struct rb_node *p = rb_first(root);
268 unsigned int sum = 0;
271 struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode);
274 rb_erase(&skb->rbnode, root);
276 struct sk_buff *next = FRAG_CB(skb)->next_frag;
278 sum += skb->truesize;
279 kfree_skb_reason(skb, reason);
285 EXPORT_SYMBOL(inet_frag_rbtree_purge);
287 void inet_frag_destroy(struct inet_frag_queue *q)
289 unsigned int sum, sum_truesize = 0;
290 enum skb_drop_reason reason;
291 struct inet_frags *f;
294 WARN_ON(!(q->flags & INET_FRAG_COMPLETE));
295 reason = (q->flags & INET_FRAG_DROP) ?
296 SKB_DROP_REASON_FRAG_REASM_TIMEOUT :
298 WARN_ON(del_timer(&q->timer) != 0);
300 /* Release all fragment data. */
303 sum_truesize = inet_frag_rbtree_purge(&q->rb_fragments, reason);
304 sum = sum_truesize + f->qsize;
306 call_rcu(&q->rcu, inet_frag_destroy_rcu);
308 sub_frag_mem_limit(fqdir, sum);
310 EXPORT_SYMBOL(inet_frag_destroy);
312 static struct inet_frag_queue *inet_frag_alloc(struct fqdir *fqdir,
313 struct inet_frags *f,
316 struct inet_frag_queue *q;
318 q = kmem_cache_zalloc(f->frags_cachep, GFP_ATOMIC);
323 f->constructor(q, arg);
324 add_frag_mem_limit(fqdir, f->qsize);
326 timer_setup(&q->timer, f->frag_expire, 0);
327 spin_lock_init(&q->lock);
328 refcount_set(&q->refcnt, 3);
333 static struct inet_frag_queue *inet_frag_create(struct fqdir *fqdir,
335 struct inet_frag_queue **prev)
337 struct inet_frags *f = fqdir->f;
338 struct inet_frag_queue *q;
340 q = inet_frag_alloc(fqdir, f, arg);
342 *prev = ERR_PTR(-ENOMEM);
345 mod_timer(&q->timer, jiffies + fqdir->timeout);
347 *prev = rhashtable_lookup_get_insert_key(&fqdir->rhashtable, &q->key,
348 &q->node, f->rhash_params);
350 q->flags |= INET_FRAG_COMPLETE;
352 inet_frag_destroy(q);
358 /* TODO : call from rcu_read_lock() and no longer use refcount_inc_not_zero() */
359 struct inet_frag_queue *inet_frag_find(struct fqdir *fqdir, void *key)
361 /* This pairs with WRITE_ONCE() in fqdir_pre_exit(). */
362 long high_thresh = READ_ONCE(fqdir->high_thresh);
363 struct inet_frag_queue *fq = NULL, *prev;
365 if (!high_thresh || frag_mem_limit(fqdir) > high_thresh)
370 prev = rhashtable_lookup(&fqdir->rhashtable, key, fqdir->f->rhash_params);
372 fq = inet_frag_create(fqdir, key, &prev);
373 if (!IS_ERR_OR_NULL(prev)) {
375 if (!refcount_inc_not_zero(&fq->refcnt))
381 EXPORT_SYMBOL(inet_frag_find);
383 int inet_frag_queue_insert(struct inet_frag_queue *q, struct sk_buff *skb,
386 struct sk_buff *last = q->fragments_tail;
388 /* RFC5722, Section 4, amended by Errata ID : 3089
389 * When reassembling an IPv6 datagram, if
390 * one or more its constituent fragments is determined to be an
391 * overlapping fragment, the entire datagram (and any constituent
392 * fragments) MUST be silently discarded.
394 * Duplicates, however, should be ignored (i.e. skb dropped, but the
395 * queue/fragments kept for later reassembly).
398 fragrun_create(q, skb); /* First fragment. */
399 else if (last->ip_defrag_offset + last->len < end) {
400 /* This is the common case: skb goes to the end. */
401 /* Detect and discard overlaps. */
402 if (offset < last->ip_defrag_offset + last->len)
403 return IPFRAG_OVERLAP;
404 if (offset == last->ip_defrag_offset + last->len)
405 fragrun_append_to_last(q, skb);
407 fragrun_create(q, skb);
409 /* Binary search. Note that skb can become the first fragment,
410 * but not the last (covered above).
412 struct rb_node **rbn, *parent;
414 rbn = &q->rb_fragments.rb_node;
416 struct sk_buff *curr;
420 curr = rb_to_skb(parent);
421 curr_run_end = curr->ip_defrag_offset +
422 FRAG_CB(curr)->frag_run_len;
423 if (end <= curr->ip_defrag_offset)
424 rbn = &parent->rb_left;
425 else if (offset >= curr_run_end)
426 rbn = &parent->rb_right;
427 else if (offset >= curr->ip_defrag_offset &&
431 return IPFRAG_OVERLAP;
433 /* Here we have parent properly set, and rbn pointing to
434 * one of its NULL left/right children. Insert skb.
437 rb_link_node(&skb->rbnode, parent, rbn);
438 rb_insert_color(&skb->rbnode, &q->rb_fragments);
441 skb->ip_defrag_offset = offset;
445 EXPORT_SYMBOL(inet_frag_queue_insert);
447 void *inet_frag_reasm_prepare(struct inet_frag_queue *q, struct sk_buff *skb,
448 struct sk_buff *parent)
450 struct sk_buff *fp, *head = skb_rb_first(&q->rb_fragments);
451 struct sk_buff **nextp;
455 fp = skb_clone(skb, GFP_ATOMIC);
458 FRAG_CB(fp)->next_frag = FRAG_CB(skb)->next_frag;
459 if (RB_EMPTY_NODE(&skb->rbnode))
460 FRAG_CB(parent)->next_frag = fp;
462 rb_replace_node(&skb->rbnode, &fp->rbnode,
464 if (q->fragments_tail == skb)
465 q->fragments_tail = fp;
466 skb_morph(skb, head);
467 FRAG_CB(skb)->next_frag = FRAG_CB(head)->next_frag;
468 rb_replace_node(&head->rbnode, &skb->rbnode,
473 WARN_ON(head->ip_defrag_offset != 0);
475 delta = -head->truesize;
477 /* Head of list must not be cloned. */
478 if (skb_unclone(head, GFP_ATOMIC))
481 delta += head->truesize;
483 add_frag_mem_limit(q->fqdir, delta);
485 /* If the first fragment is fragmented itself, we split
486 * it to two chunks: the first with data and paged part
487 * and the second, holding only fragments.
489 if (skb_has_frag_list(head)) {
490 struct sk_buff *clone;
493 clone = alloc_skb(0, GFP_ATOMIC);
496 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
497 skb_frag_list_init(head);
498 for (i = 0; i < skb_shinfo(head)->nr_frags; i++)
499 plen += skb_frag_size(&skb_shinfo(head)->frags[i]);
500 clone->data_len = head->data_len - plen;
501 clone->len = clone->data_len;
502 head->truesize += clone->truesize;
504 clone->ip_summed = head->ip_summed;
505 add_frag_mem_limit(q->fqdir, clone->truesize);
506 skb_shinfo(head)->frag_list = clone;
507 nextp = &clone->next;
509 nextp = &skb_shinfo(head)->frag_list;
514 EXPORT_SYMBOL(inet_frag_reasm_prepare);
516 void inet_frag_reasm_finish(struct inet_frag_queue *q, struct sk_buff *head,
517 void *reasm_data, bool try_coalesce)
519 struct sk_buff **nextp = reasm_data;
524 skb_push(head, head->data - skb_network_header(head));
526 /* Traverse the tree in order, to build frag_list. */
527 fp = FRAG_CB(head)->next_frag;
528 rbn = rb_next(&head->rbnode);
529 rb_erase(&head->rbnode, &q->rb_fragments);
531 sum_truesize = head->truesize;
533 /* fp points to the next sk_buff in the current run;
534 * rbn points to the next run.
536 /* Go through the current run. */
538 struct sk_buff *next_frag = FRAG_CB(fp)->next_frag;
542 sum_truesize += fp->truesize;
543 if (head->ip_summed != fp->ip_summed)
544 head->ip_summed = CHECKSUM_NONE;
545 else if (head->ip_summed == CHECKSUM_COMPLETE)
546 head->csum = csum_add(head->csum, fp->csum);
548 if (try_coalesce && skb_try_coalesce(head, fp, &stolen,
550 kfree_skb_partial(fp, stolen);
553 memset(&fp->rbnode, 0, sizeof(fp->rbnode));
556 head->data_len += fp->len;
557 head->len += fp->len;
558 head->truesize += fp->truesize;
566 /* Move to the next run. */
568 struct rb_node *rbnext = rb_next(rbn);
571 rb_erase(rbn, &q->rb_fragments);
575 sub_frag_mem_limit(q->fqdir, sum_truesize);
578 skb_mark_not_on_list(head);
580 head->tstamp = q->stamp;
581 head->mono_delivery_time = q->mono_delivery_time;
583 EXPORT_SYMBOL(inet_frag_reasm_finish);
585 struct sk_buff *inet_frag_pull_head(struct inet_frag_queue *q)
587 struct sk_buff *head, *skb;
589 head = skb_rb_first(&q->rb_fragments);
592 skb = FRAG_CB(head)->next_frag;
594 rb_replace_node(&head->rbnode, &skb->rbnode,
597 rb_erase(&head->rbnode, &q->rb_fragments);
598 memset(&head->rbnode, 0, sizeof(head->rbnode));
601 if (head == q->fragments_tail)
602 q->fragments_tail = NULL;
604 sub_frag_mem_limit(q->fqdir, head->truesize);
608 EXPORT_SYMBOL(inet_frag_pull_head);