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 #include "../core/sock_destructor.h"
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.
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.
37 struct ipfrag_skb_cb {
39 struct inet_skb_parm h4;
40 struct inet6_skb_parm h6;
42 struct sk_buff *next_frag;
47 #define FRAG_CB(skb) ((struct ipfrag_skb_cb *)((skb)->cb))
49 static void fragcb_clear(struct sk_buff *skb)
51 RB_CLEAR_NODE(&skb->rbnode);
52 FRAG_CB(skb)->next_frag = NULL;
53 FRAG_CB(skb)->frag_run_len = skb->len;
56 /* Append skb to the last "run". */
57 static void fragrun_append_to_last(struct inet_frag_queue *q,
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;
67 /* Create a new "run" with the skb. */
68 static void fragrun_create(struct inet_frag_queue *q, struct sk_buff *skb)
70 BUILD_BUG_ON(sizeof(struct ipfrag_skb_cb) > sizeof(skb->cb));
74 rb_link_node(&skb->rbnode, &q->last_run_head->rbnode,
75 &q->last_run_head->rbnode.rb_right);
77 rb_link_node(&skb->rbnode, NULL, &q->rb_fragments.rb_node);
78 rb_insert_color(&skb->rbnode, &q->rb_fragments);
80 q->fragments_tail = skb;
81 q->last_run_head = skb;
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
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,
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,
103 EXPORT_SYMBOL(ip_frag_ecn_table);
105 int inet_frags_init(struct inet_frags *f)
107 f->frags_cachep = kmem_cache_create(f->frags_cache_name, f->qsize, 0, 0,
109 if (!f->frags_cachep)
112 refcount_set(&f->refcnt, 1);
113 init_completion(&f->completion);
116 EXPORT_SYMBOL(inet_frags_init);
118 void inet_frags_fini(struct inet_frags *f)
120 if (refcount_dec_and_test(&f->refcnt))
121 complete(&f->completion);
123 wait_for_completion(&f->completion);
125 kmem_cache_destroy(f->frags_cachep);
126 f->frags_cachep = NULL;
128 EXPORT_SYMBOL(inet_frags_fini);
130 /* called from rhashtable_free_and_destroy() at netns_frags dismantle */
131 static void inet_frags_free_cb(void *ptr, void *arg)
133 struct inet_frag_queue *fq = ptr;
136 count = del_timer_sync(&fq->timer) ? 1 : 0;
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;
143 } else if (fq->flags & INET_FRAG_HASH_DEAD) {
146 spin_unlock_bh(&fq->lock);
148 if (refcount_sub_and_test(count, &fq->refcnt))
149 inet_frag_destroy(fq);
152 static LLIST_HEAD(fqdir_free_list);
154 static void fqdir_free_fn(struct work_struct *work)
156 struct llist_node *kill_list;
157 struct fqdir *fqdir, *tmp;
158 struct inet_frags *f;
160 /* Atomically snapshot the list of fqdirs to free */
161 kill_list = llist_del_all(&fqdir_free_list);
163 /* We need to make sure all ongoing call_rcu(..., inet_frag_destroy_rcu)
164 * have completed, since they need to dereference fqdir.
165 * Would it not be nice to have kfree_rcu_barrier() ? :)
169 llist_for_each_entry_safe(fqdir, tmp, kill_list, free_list) {
171 if (refcount_dec_and_test(&f->refcnt))
172 complete(&f->completion);
178 static DECLARE_WORK(fqdir_free_work, fqdir_free_fn);
180 static void fqdir_work_fn(struct work_struct *work)
182 struct fqdir *fqdir = container_of(work, struct fqdir, destroy_work);
184 rhashtable_free_and_destroy(&fqdir->rhashtable, inet_frags_free_cb, NULL);
186 if (llist_add(&fqdir->free_list, &fqdir_free_list))
187 queue_work(system_wq, &fqdir_free_work);
190 int fqdir_init(struct fqdir **fqdirp, struct inet_frags *f, struct net *net)
192 struct fqdir *fqdir = kzalloc(sizeof(*fqdir), GFP_KERNEL);
199 res = rhashtable_init(&fqdir->rhashtable, &fqdir->f->rhash_params);
204 refcount_inc(&f->refcnt);
208 EXPORT_SYMBOL(fqdir_init);
210 static struct workqueue_struct *inet_frag_wq;
212 static int __init inet_frag_wq_init(void)
214 inet_frag_wq = create_workqueue("inet_frag_wq");
216 panic("Could not create inet frag workq");
220 pure_initcall(inet_frag_wq_init);
222 void fqdir_exit(struct fqdir *fqdir)
224 INIT_WORK(&fqdir->destroy_work, fqdir_work_fn);
225 queue_work(inet_frag_wq, &fqdir->destroy_work);
227 EXPORT_SYMBOL(fqdir_exit);
229 void inet_frag_kill(struct inet_frag_queue *fq)
231 if (del_timer(&fq->timer))
232 refcount_dec(&fq->refcnt);
234 if (!(fq->flags & INET_FRAG_COMPLETE)) {
235 struct fqdir *fqdir = fq->fqdir;
237 fq->flags |= INET_FRAG_COMPLETE;
239 /* The RCU read lock provides a memory barrier
240 * guaranteeing that if fqdir->dead is false then
241 * the hash table destruction will not start until
242 * after we unlock. Paired with fqdir_pre_exit().
244 if (!READ_ONCE(fqdir->dead)) {
245 rhashtable_remove_fast(&fqdir->rhashtable, &fq->node,
246 fqdir->f->rhash_params);
247 refcount_dec(&fq->refcnt);
249 fq->flags |= INET_FRAG_HASH_DEAD;
254 EXPORT_SYMBOL(inet_frag_kill);
256 static void inet_frag_destroy_rcu(struct rcu_head *head)
258 struct inet_frag_queue *q = container_of(head, struct inet_frag_queue,
260 struct inet_frags *f = q->fqdir->f;
264 kmem_cache_free(f->frags_cachep, q);
267 unsigned int inet_frag_rbtree_purge(struct rb_root *root,
268 enum skb_drop_reason reason)
270 struct rb_node *p = rb_first(root);
271 unsigned int sum = 0;
274 struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode);
277 rb_erase(&skb->rbnode, root);
279 struct sk_buff *next = FRAG_CB(skb)->next_frag;
281 sum += skb->truesize;
282 kfree_skb_reason(skb, reason);
288 EXPORT_SYMBOL(inet_frag_rbtree_purge);
290 void inet_frag_destroy(struct inet_frag_queue *q)
292 unsigned int sum, sum_truesize = 0;
293 enum skb_drop_reason reason;
294 struct inet_frags *f;
297 WARN_ON(!(q->flags & INET_FRAG_COMPLETE));
298 reason = (q->flags & INET_FRAG_DROP) ?
299 SKB_DROP_REASON_FRAG_REASM_TIMEOUT :
301 WARN_ON(del_timer(&q->timer) != 0);
303 /* Release all fragment data. */
306 sum_truesize = inet_frag_rbtree_purge(&q->rb_fragments, reason);
307 sum = sum_truesize + f->qsize;
309 call_rcu(&q->rcu, inet_frag_destroy_rcu);
311 sub_frag_mem_limit(fqdir, sum);
313 EXPORT_SYMBOL(inet_frag_destroy);
315 static struct inet_frag_queue *inet_frag_alloc(struct fqdir *fqdir,
316 struct inet_frags *f,
319 struct inet_frag_queue *q;
321 q = kmem_cache_zalloc(f->frags_cachep, GFP_ATOMIC);
326 f->constructor(q, arg);
327 add_frag_mem_limit(fqdir, f->qsize);
329 timer_setup(&q->timer, f->frag_expire, 0);
330 spin_lock_init(&q->lock);
331 refcount_set(&q->refcnt, 3);
336 static struct inet_frag_queue *inet_frag_create(struct fqdir *fqdir,
338 struct inet_frag_queue **prev)
340 struct inet_frags *f = fqdir->f;
341 struct inet_frag_queue *q;
343 q = inet_frag_alloc(fqdir, f, arg);
345 *prev = ERR_PTR(-ENOMEM);
348 mod_timer(&q->timer, jiffies + fqdir->timeout);
350 *prev = rhashtable_lookup_get_insert_key(&fqdir->rhashtable, &q->key,
351 &q->node, f->rhash_params);
353 q->flags |= INET_FRAG_COMPLETE;
355 inet_frag_destroy(q);
361 /* TODO : call from rcu_read_lock() and no longer use refcount_inc_not_zero() */
362 struct inet_frag_queue *inet_frag_find(struct fqdir *fqdir, void *key)
364 /* This pairs with WRITE_ONCE() in fqdir_pre_exit(). */
365 long high_thresh = READ_ONCE(fqdir->high_thresh);
366 struct inet_frag_queue *fq = NULL, *prev;
368 if (!high_thresh || frag_mem_limit(fqdir) > high_thresh)
373 prev = rhashtable_lookup(&fqdir->rhashtable, key, fqdir->f->rhash_params);
375 fq = inet_frag_create(fqdir, key, &prev);
376 if (!IS_ERR_OR_NULL(prev)) {
378 if (!refcount_inc_not_zero(&fq->refcnt))
384 EXPORT_SYMBOL(inet_frag_find);
386 int inet_frag_queue_insert(struct inet_frag_queue *q, struct sk_buff *skb,
389 struct sk_buff *last = q->fragments_tail;
391 /* RFC5722, Section 4, amended by Errata ID : 3089
392 * When reassembling an IPv6 datagram, if
393 * one or more its constituent fragments is determined to be an
394 * overlapping fragment, the entire datagram (and any constituent
395 * fragments) MUST be silently discarded.
397 * Duplicates, however, should be ignored (i.e. skb dropped, but the
398 * queue/fragments kept for later reassembly).
401 fragrun_create(q, skb); /* First fragment. */
402 else if (FRAG_CB(last)->ip_defrag_offset + last->len < end) {
403 /* This is the common case: skb goes to the end. */
404 /* Detect and discard overlaps. */
405 if (offset < FRAG_CB(last)->ip_defrag_offset + last->len)
406 return IPFRAG_OVERLAP;
407 if (offset == FRAG_CB(last)->ip_defrag_offset + last->len)
408 fragrun_append_to_last(q, skb);
410 fragrun_create(q, skb);
412 /* Binary search. Note that skb can become the first fragment,
413 * but not the last (covered above).
415 struct rb_node **rbn, *parent;
417 rbn = &q->rb_fragments.rb_node;
419 struct sk_buff *curr;
423 curr = rb_to_skb(parent);
424 curr_run_end = FRAG_CB(curr)->ip_defrag_offset +
425 FRAG_CB(curr)->frag_run_len;
426 if (end <= FRAG_CB(curr)->ip_defrag_offset)
427 rbn = &parent->rb_left;
428 else if (offset >= curr_run_end)
429 rbn = &parent->rb_right;
430 else if (offset >= FRAG_CB(curr)->ip_defrag_offset &&
434 return IPFRAG_OVERLAP;
436 /* Here we have parent properly set, and rbn pointing to
437 * one of its NULL left/right children. Insert skb.
440 rb_link_node(&skb->rbnode, parent, rbn);
441 rb_insert_color(&skb->rbnode, &q->rb_fragments);
444 FRAG_CB(skb)->ip_defrag_offset = offset;
448 EXPORT_SYMBOL(inet_frag_queue_insert);
450 void *inet_frag_reasm_prepare(struct inet_frag_queue *q, struct sk_buff *skb,
451 struct sk_buff *parent)
453 struct sk_buff *fp, *head = skb_rb_first(&q->rb_fragments);
454 void (*destructor)(struct sk_buff *);
455 unsigned int orig_truesize = 0;
456 struct sk_buff **nextp = NULL;
457 struct sock *sk = skb->sk;
460 if (sk && is_skb_wmem(skb)) {
461 /* TX: skb->sk might have been passed as argument to
462 * dst->output and must remain valid until tx completes.
464 * Move sk to reassembled skb and fix up wmem accounting.
466 orig_truesize = skb->truesize;
467 destructor = skb->destructor;
471 fp = skb_clone(skb, GFP_ATOMIC);
476 FRAG_CB(fp)->next_frag = FRAG_CB(skb)->next_frag;
477 if (RB_EMPTY_NODE(&skb->rbnode))
478 FRAG_CB(parent)->next_frag = fp;
480 rb_replace_node(&skb->rbnode, &fp->rbnode,
482 if (q->fragments_tail == skb)
483 q->fragments_tail = fp;
486 /* prevent skb_morph from releasing sk */
488 skb->destructor = NULL;
490 skb_morph(skb, head);
491 FRAG_CB(skb)->next_frag = FRAG_CB(head)->next_frag;
492 rb_replace_node(&head->rbnode, &skb->rbnode,
497 WARN_ON(FRAG_CB(head)->ip_defrag_offset != 0);
499 delta = -head->truesize;
501 /* Head of list must not be cloned. */
502 if (skb_unclone(head, GFP_ATOMIC))
505 delta += head->truesize;
507 add_frag_mem_limit(q->fqdir, delta);
509 /* If the first fragment is fragmented itself, we split
510 * it to two chunks: the first with data and paged part
511 * and the second, holding only fragments.
513 if (skb_has_frag_list(head)) {
514 struct sk_buff *clone;
517 clone = alloc_skb(0, GFP_ATOMIC);
520 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
521 skb_frag_list_init(head);
522 for (i = 0; i < skb_shinfo(head)->nr_frags; i++)
523 plen += skb_frag_size(&skb_shinfo(head)->frags[i]);
524 clone->data_len = head->data_len - plen;
525 clone->len = clone->data_len;
526 head->truesize += clone->truesize;
528 clone->ip_summed = head->ip_summed;
529 add_frag_mem_limit(q->fqdir, clone->truesize);
530 skb_shinfo(head)->frag_list = clone;
531 nextp = &clone->next;
533 nextp = &skb_shinfo(head)->frag_list;
538 int ts_delta = head->truesize - orig_truesize;
540 /* if this reassembled skb is fragmented later,
541 * fraglist skbs will get skb->sk assigned from head->sk,
542 * and each frag skb will be released via sock_wfree.
544 * Update sk_wmem_alloc.
547 head->destructor = destructor;
548 refcount_add(ts_delta, &sk->sk_wmem_alloc);
553 EXPORT_SYMBOL(inet_frag_reasm_prepare);
555 void inet_frag_reasm_finish(struct inet_frag_queue *q, struct sk_buff *head,
556 void *reasm_data, bool try_coalesce)
558 struct sock *sk = is_skb_wmem(head) ? head->sk : NULL;
559 const unsigned int head_truesize = head->truesize;
560 struct sk_buff **nextp = reasm_data;
565 skb_push(head, head->data - skb_network_header(head));
567 /* Traverse the tree in order, to build frag_list. */
568 fp = FRAG_CB(head)->next_frag;
569 rbn = rb_next(&head->rbnode);
570 rb_erase(&head->rbnode, &q->rb_fragments);
572 sum_truesize = head->truesize;
574 /* fp points to the next sk_buff in the current run;
575 * rbn points to the next run.
577 /* Go through the current run. */
579 struct sk_buff *next_frag = FRAG_CB(fp)->next_frag;
583 sum_truesize += fp->truesize;
584 if (head->ip_summed != fp->ip_summed)
585 head->ip_summed = CHECKSUM_NONE;
586 else if (head->ip_summed == CHECKSUM_COMPLETE)
587 head->csum = csum_add(head->csum, fp->csum);
589 if (try_coalesce && skb_try_coalesce(head, fp, &stolen,
591 kfree_skb_partial(fp, stolen);
594 memset(&fp->rbnode, 0, sizeof(fp->rbnode));
597 head->data_len += fp->len;
598 head->len += fp->len;
599 head->truesize += fp->truesize;
607 /* Move to the next run. */
609 struct rb_node *rbnext = rb_next(rbn);
612 rb_erase(rbn, &q->rb_fragments);
616 sub_frag_mem_limit(q->fqdir, sum_truesize);
619 skb_mark_not_on_list(head);
621 head->tstamp = q->stamp;
622 head->mono_delivery_time = q->mono_delivery_time;
625 refcount_add(sum_truesize - head_truesize, &sk->sk_wmem_alloc);
627 EXPORT_SYMBOL(inet_frag_reasm_finish);
629 struct sk_buff *inet_frag_pull_head(struct inet_frag_queue *q)
631 struct sk_buff *head, *skb;
633 head = skb_rb_first(&q->rb_fragments);
636 skb = FRAG_CB(head)->next_frag;
638 rb_replace_node(&head->rbnode, &skb->rbnode,
641 rb_erase(&head->rbnode, &q->rb_fragments);
642 memset(&head->rbnode, 0, sizeof(head->rbnode));
645 if (head == q->fragments_tail)
646 q->fragments_tail = NULL;
648 sub_frag_mem_limit(q->fqdir, head->truesize);
652 EXPORT_SYMBOL(inet_frag_pull_head);