2 * inet fragments management
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Authors: Pavel Emelyanov <xemul@openvz.org>
10 * Started as consolidation of ipv4/ip_fragment.c,
11 * ipv6/reassembly. and ipv6 nf conntrack reassembly
14 #include <linux/list.h>
15 #include <linux/spinlock.h>
16 #include <linux/module.h>
17 #include <linux/timer.h>
19 #include <linux/random.h>
20 #include <linux/skbuff.h>
21 #include <linux/rtnetlink.h>
22 #include <linux/slab.h>
25 #include <net/inet_frag.h>
26 #include <net/inet_ecn.h>
30 /* Use skb->cb to track consecutive/adjacent fragments coming at
31 * the end of the queue. Nodes in the rb-tree queue will
32 * contain "runs" of one or more adjacent fragments.
35 * - next_frag is NULL at the tail of a "run";
36 * - the head of a "run" has the sum of all fragment lengths in frag_run_len.
38 struct ipfrag_skb_cb {
40 struct inet_skb_parm h4;
41 struct inet6_skb_parm h6;
43 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)
114 EXPORT_SYMBOL(inet_frags_init);
116 void inet_frags_fini(struct inet_frags *f)
118 /* We must wait that all inet_frag_destroy_rcu() have completed. */
121 kmem_cache_destroy(f->frags_cachep);
122 f->frags_cachep = NULL;
124 EXPORT_SYMBOL(inet_frags_fini);
126 static void inet_frags_free_cb(void *ptr, void *arg)
128 struct inet_frag_queue *fq = ptr;
130 /* If we can not cancel the timer, it means this frag_queue
131 * is already disappearing, we have nothing to do.
132 * Otherwise, we own a refcount until the end of this function.
134 if (!del_timer(&fq->timer))
137 spin_lock_bh(&fq->lock);
138 if (!(fq->flags & INET_FRAG_COMPLETE)) {
139 fq->flags |= INET_FRAG_COMPLETE;
140 atomic_dec(&fq->refcnt);
142 spin_unlock_bh(&fq->lock);
147 void inet_frags_exit_net(struct netns_frags *nf)
149 nf->high_thresh = 0; /* prevent creation of new frags */
151 rhashtable_free_and_destroy(&nf->rhashtable, inet_frags_free_cb, NULL);
153 EXPORT_SYMBOL(inet_frags_exit_net);
155 void inet_frag_kill(struct inet_frag_queue *fq)
157 if (del_timer(&fq->timer))
158 atomic_dec(&fq->refcnt);
160 if (!(fq->flags & INET_FRAG_COMPLETE)) {
161 struct netns_frags *nf = fq->net;
163 fq->flags |= INET_FRAG_COMPLETE;
164 rhashtable_remove_fast(&nf->rhashtable, &fq->node, nf->f->rhash_params);
165 atomic_dec(&fq->refcnt);
168 EXPORT_SYMBOL(inet_frag_kill);
170 static void inet_frag_destroy_rcu(struct rcu_head *head)
172 struct inet_frag_queue *q = container_of(head, struct inet_frag_queue,
174 struct inet_frags *f = q->net->f;
178 kmem_cache_free(f->frags_cachep, q);
181 unsigned int inet_frag_rbtree_purge(struct rb_root *root)
183 struct rb_node *p = rb_first(root);
184 unsigned int sum = 0;
187 struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode);
190 rb_erase(&skb->rbnode, root);
192 struct sk_buff *next = FRAG_CB(skb)->next_frag;
194 sum += skb->truesize;
201 EXPORT_SYMBOL(inet_frag_rbtree_purge);
203 void inet_frag_destroy(struct inet_frag_queue *q)
206 struct netns_frags *nf;
207 unsigned int sum, sum_truesize = 0;
208 struct inet_frags *f;
210 WARN_ON(!(q->flags & INET_FRAG_COMPLETE));
211 WARN_ON(del_timer(&q->timer) != 0);
213 /* Release all fragment data. */
219 struct sk_buff *xp = fp->next;
221 sum_truesize += fp->truesize;
226 sum_truesize = inet_frag_rbtree_purge(&q->rb_fragments);
228 sum = sum_truesize + f->qsize;
230 call_rcu(&q->rcu, inet_frag_destroy_rcu);
232 sub_frag_mem_limit(nf, sum);
234 EXPORT_SYMBOL(inet_frag_destroy);
236 static struct inet_frag_queue *inet_frag_alloc(struct netns_frags *nf,
237 struct inet_frags *f,
240 struct inet_frag_queue *q;
242 if (!nf->high_thresh || frag_mem_limit(nf) > nf->high_thresh)
245 q = kmem_cache_zalloc(f->frags_cachep, GFP_ATOMIC);
250 f->constructor(q, arg);
251 add_frag_mem_limit(nf, f->qsize);
253 setup_timer(&q->timer, f->frag_expire, (unsigned long)q);
254 spin_lock_init(&q->lock);
255 atomic_set(&q->refcnt, 3);
260 static struct inet_frag_queue *inet_frag_create(struct netns_frags *nf,
262 struct inet_frag_queue **prev)
264 struct inet_frags *f = nf->f;
265 struct inet_frag_queue *q;
267 q = inet_frag_alloc(nf, f, arg);
269 *prev = ERR_PTR(-ENOMEM);
272 mod_timer(&q->timer, jiffies + nf->timeout);
274 *prev = rhashtable_lookup_get_insert_key(&nf->rhashtable, &q->key,
275 &q->node, f->rhash_params);
277 q->flags |= INET_FRAG_COMPLETE;
279 inet_frag_destroy(q);
284 EXPORT_SYMBOL(inet_frag_create);
286 /* TODO : call from rcu_read_lock() and no longer use refcount_inc_not_zero() */
287 struct inet_frag_queue *inet_frag_find(struct netns_frags *nf, void *key)
289 struct inet_frag_queue *fq = NULL, *prev;
292 prev = rhashtable_lookup(&nf->rhashtable, key, nf->f->rhash_params);
294 fq = inet_frag_create(nf, key, &prev);
295 if (prev && !IS_ERR(prev)) {
297 if (!atomic_inc_not_zero(&fq->refcnt))
303 EXPORT_SYMBOL(inet_frag_find);
305 int inet_frag_queue_insert(struct inet_frag_queue *q, struct sk_buff *skb,
308 struct sk_buff *last = q->fragments_tail;
310 /* RFC5722, Section 4, amended by Errata ID : 3089
311 * When reassembling an IPv6 datagram, if
312 * one or more its constituent fragments is determined to be an
313 * overlapping fragment, the entire datagram (and any constituent
314 * fragments) MUST be silently discarded.
316 * Duplicates, however, should be ignored (i.e. skb dropped, but the
317 * queue/fragments kept for later reassembly).
320 fragrun_create(q, skb); /* First fragment. */
321 else if (last->ip_defrag_offset + last->len < end) {
322 /* This is the common case: skb goes to the end. */
323 /* Detect and discard overlaps. */
324 if (offset < last->ip_defrag_offset + last->len)
325 return IPFRAG_OVERLAP;
326 if (offset == last->ip_defrag_offset + last->len)
327 fragrun_append_to_last(q, skb);
329 fragrun_create(q, skb);
331 /* Binary search. Note that skb can become the first fragment,
332 * but not the last (covered above).
334 struct rb_node **rbn, *parent;
336 rbn = &q->rb_fragments.rb_node;
338 struct sk_buff *curr;
342 curr = rb_to_skb(parent);
343 curr_run_end = curr->ip_defrag_offset +
344 FRAG_CB(curr)->frag_run_len;
345 if (end <= curr->ip_defrag_offset)
346 rbn = &parent->rb_left;
347 else if (offset >= curr_run_end)
348 rbn = &parent->rb_right;
349 else if (offset >= curr->ip_defrag_offset &&
353 return IPFRAG_OVERLAP;
355 /* Here we have parent properly set, and rbn pointing to
356 * one of its NULL left/right children. Insert skb.
359 rb_link_node(&skb->rbnode, parent, rbn);
360 rb_insert_color(&skb->rbnode, &q->rb_fragments);
363 skb->ip_defrag_offset = offset;
367 EXPORT_SYMBOL(inet_frag_queue_insert);
369 void *inet_frag_reasm_prepare(struct inet_frag_queue *q, struct sk_buff *skb,
370 struct sk_buff *parent)
372 struct sk_buff *fp, *head = skb_rb_first(&q->rb_fragments);
373 struct sk_buff **nextp;
377 fp = skb_clone(skb, GFP_ATOMIC);
380 FRAG_CB(fp)->next_frag = FRAG_CB(skb)->next_frag;
381 if (RB_EMPTY_NODE(&skb->rbnode))
382 FRAG_CB(parent)->next_frag = fp;
384 rb_replace_node(&skb->rbnode, &fp->rbnode,
386 if (q->fragments_tail == skb)
387 q->fragments_tail = fp;
388 skb_morph(skb, head);
389 FRAG_CB(skb)->next_frag = FRAG_CB(head)->next_frag;
390 rb_replace_node(&head->rbnode, &skb->rbnode,
395 WARN_ON(head->ip_defrag_offset != 0);
397 delta = -head->truesize;
399 /* Head of list must not be cloned. */
400 if (skb_unclone(head, GFP_ATOMIC))
403 delta += head->truesize;
405 add_frag_mem_limit(q->net, delta);
407 /* If the first fragment is fragmented itself, we split
408 * it to two chunks: the first with data and paged part
409 * and the second, holding only fragments.
411 if (skb_has_frag_list(head)) {
412 struct sk_buff *clone;
415 clone = alloc_skb(0, GFP_ATOMIC);
418 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
419 skb_frag_list_init(head);
420 for (i = 0; i < skb_shinfo(head)->nr_frags; i++)
421 plen += skb_frag_size(&skb_shinfo(head)->frags[i]);
422 clone->data_len = head->data_len - plen;
423 clone->len = clone->data_len;
424 head->truesize += clone->truesize;
426 clone->ip_summed = head->ip_summed;
427 add_frag_mem_limit(q->net, clone->truesize);
428 skb_shinfo(head)->frag_list = clone;
429 nextp = &clone->next;
431 nextp = &skb_shinfo(head)->frag_list;
436 EXPORT_SYMBOL(inet_frag_reasm_prepare);
438 void inet_frag_reasm_finish(struct inet_frag_queue *q, struct sk_buff *head,
441 struct sk_buff **nextp = (struct sk_buff **)reasm_data;
445 skb_push(head, head->data - skb_network_header(head));
447 /* Traverse the tree in order, to build frag_list. */
448 fp = FRAG_CB(head)->next_frag;
449 rbn = rb_next(&head->rbnode);
450 rb_erase(&head->rbnode, &q->rb_fragments);
452 /* fp points to the next sk_buff in the current run;
453 * rbn points to the next run.
455 /* Go through the current run. */
460 memset(&fp->rbnode, 0, sizeof(fp->rbnode));
462 head->data_len += fp->len;
463 head->len += fp->len;
464 if (head->ip_summed != fp->ip_summed)
465 head->ip_summed = CHECKSUM_NONE;
466 else if (head->ip_summed == CHECKSUM_COMPLETE)
467 head->csum = csum_add(head->csum, fp->csum);
468 head->truesize += fp->truesize;
469 fp = FRAG_CB(fp)->next_frag;
471 /* Move to the next run. */
473 struct rb_node *rbnext = rb_next(rbn);
476 rb_erase(rbn, &q->rb_fragments);
480 sub_frag_mem_limit(q->net, head->truesize);
485 head->tstamp = q->stamp;
487 EXPORT_SYMBOL(inet_frag_reasm_finish);
489 struct sk_buff *inet_frag_pull_head(struct inet_frag_queue *q)
491 struct sk_buff *head;
495 q->fragments = head->next;
499 head = skb_rb_first(&q->rb_fragments);
502 skb = FRAG_CB(head)->next_frag;
504 rb_replace_node(&head->rbnode, &skb->rbnode,
507 rb_erase(&head->rbnode, &q->rb_fragments);
508 memset(&head->rbnode, 0, sizeof(head->rbnode));
511 if (head == q->fragments_tail)
512 q->fragments_tail = NULL;
514 sub_frag_mem_limit(q->net, head->truesize);
518 EXPORT_SYMBOL(inet_frag_pull_head);