xref: /openbmc/linux/net/ipv4/inet_fragment.c (revision f71a261a)
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 /* 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.
30  *
31  * Invariants:
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.
34  */
35 struct ipfrag_skb_cb {
36 	union {
37 		struct inet_skb_parm	h4;
38 		struct inet6_skb_parm	h6;
39 	};
40 	struct sk_buff		*next_frag;
41 	int			frag_run_len;
42 };
43 
44 #define FRAG_CB(skb)		((struct ipfrag_skb_cb *)((skb)->cb))
45 
46 static void fragcb_clear(struct sk_buff *skb)
47 {
48 	RB_CLEAR_NODE(&skb->rbnode);
49 	FRAG_CB(skb)->next_frag = NULL;
50 	FRAG_CB(skb)->frag_run_len = skb->len;
51 }
52 
53 /* Append skb to the last "run". */
54 static void fragrun_append_to_last(struct inet_frag_queue *q,
55 				   struct sk_buff *skb)
56 {
57 	fragcb_clear(skb);
58 
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;
62 }
63 
64 /* Create a new "run" with the skb. */
65 static void fragrun_create(struct inet_frag_queue *q, struct sk_buff *skb)
66 {
67 	BUILD_BUG_ON(sizeof(struct ipfrag_skb_cb) > sizeof(skb->cb));
68 	fragcb_clear(skb);
69 
70 	if (q->last_run_head)
71 		rb_link_node(&skb->rbnode, &q->last_run_head->rbnode,
72 			     &q->last_run_head->rbnode.rb_right);
73 	else
74 		rb_link_node(&skb->rbnode, NULL, &q->rb_fragments.rb_node);
75 	rb_insert_color(&skb->rbnode, &q->rb_fragments);
76 
77 	q->fragments_tail = skb;
78 	q->last_run_head = skb;
79 }
80 
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
84  */
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,
90 
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,
99 };
100 EXPORT_SYMBOL(ip_frag_ecn_table);
101 
102 int inet_frags_init(struct inet_frags *f)
103 {
104 	f->frags_cachep = kmem_cache_create(f->frags_cache_name, f->qsize, 0, 0,
105 					    NULL);
106 	if (!f->frags_cachep)
107 		return -ENOMEM;
108 
109 	refcount_set(&f->refcnt, 1);
110 	init_completion(&f->completion);
111 	return 0;
112 }
113 EXPORT_SYMBOL(inet_frags_init);
114 
115 void inet_frags_fini(struct inet_frags *f)
116 {
117 	if (refcount_dec_and_test(&f->refcnt))
118 		complete(&f->completion);
119 
120 	wait_for_completion(&f->completion);
121 
122 	kmem_cache_destroy(f->frags_cachep);
123 	f->frags_cachep = NULL;
124 }
125 EXPORT_SYMBOL(inet_frags_fini);
126 
127 /* called from rhashtable_free_and_destroy() at netns_frags dismantle */
128 static void inet_frags_free_cb(void *ptr, void *arg)
129 {
130 	struct inet_frag_queue *fq = ptr;
131 	int count;
132 
133 	count = del_timer_sync(&fq->timer) ? 1 : 0;
134 
135 	spin_lock_bh(&fq->lock);
136 	if (!(fq->flags & INET_FRAG_COMPLETE)) {
137 		fq->flags |= INET_FRAG_COMPLETE;
138 		count++;
139 	} else if (fq->flags & INET_FRAG_HASH_DEAD) {
140 		count++;
141 	}
142 	spin_unlock_bh(&fq->lock);
143 
144 	if (refcount_sub_and_test(count, &fq->refcnt))
145 		inet_frag_destroy(fq);
146 }
147 
148 static LLIST_HEAD(fqdir_free_list);
149 
150 static void fqdir_free_fn(struct work_struct *work)
151 {
152 	struct llist_node *kill_list;
153 	struct fqdir *fqdir, *tmp;
154 	struct inet_frags *f;
155 
156 	/* Atomically snapshot the list of fqdirs to free */
157 	kill_list = llist_del_all(&fqdir_free_list);
158 
159 	/* We need to make sure all ongoing call_rcu(..., inet_frag_destroy_rcu)
160 	 * have completed, since they need to dereference fqdir.
161 	 * Would it not be nice to have kfree_rcu_barrier() ? :)
162 	 */
163 	rcu_barrier();
164 
165 	llist_for_each_entry_safe(fqdir, tmp, kill_list, free_list) {
166 		f = fqdir->f;
167 		if (refcount_dec_and_test(&f->refcnt))
168 			complete(&f->completion);
169 
170 		kfree(fqdir);
171 	}
172 }
173 
174 static DECLARE_WORK(fqdir_free_work, fqdir_free_fn);
175 
176 static void fqdir_work_fn(struct work_struct *work)
177 {
178 	struct fqdir *fqdir = container_of(work, struct fqdir, destroy_work);
179 
180 	rhashtable_free_and_destroy(&fqdir->rhashtable, inet_frags_free_cb, NULL);
181 
182 	if (llist_add(&fqdir->free_list, &fqdir_free_list))
183 		queue_work(system_wq, &fqdir_free_work);
184 }
185 
186 int fqdir_init(struct fqdir **fqdirp, struct inet_frags *f, struct net *net)
187 {
188 	struct fqdir *fqdir = kzalloc(sizeof(*fqdir), GFP_KERNEL);
189 	int res;
190 
191 	if (!fqdir)
192 		return -ENOMEM;
193 	fqdir->f = f;
194 	fqdir->net = net;
195 	res = rhashtable_init(&fqdir->rhashtable, &fqdir->f->rhash_params);
196 	if (res < 0) {
197 		kfree(fqdir);
198 		return res;
199 	}
200 	refcount_inc(&f->refcnt);
201 	*fqdirp = fqdir;
202 	return 0;
203 }
204 EXPORT_SYMBOL(fqdir_init);
205 
206 static struct workqueue_struct *inet_frag_wq;
207 
208 static int __init inet_frag_wq_init(void)
209 {
210 	inet_frag_wq = create_workqueue("inet_frag_wq");
211 	if (!inet_frag_wq)
212 		panic("Could not create inet frag workq");
213 	return 0;
214 }
215 
216 pure_initcall(inet_frag_wq_init);
217 
218 void fqdir_exit(struct fqdir *fqdir)
219 {
220 	INIT_WORK(&fqdir->destroy_work, fqdir_work_fn);
221 	queue_work(inet_frag_wq, &fqdir->destroy_work);
222 }
223 EXPORT_SYMBOL(fqdir_exit);
224 
225 void inet_frag_kill(struct inet_frag_queue *fq)
226 {
227 	if (del_timer(&fq->timer))
228 		refcount_dec(&fq->refcnt);
229 
230 	if (!(fq->flags & INET_FRAG_COMPLETE)) {
231 		struct fqdir *fqdir = fq->fqdir;
232 
233 		fq->flags |= INET_FRAG_COMPLETE;
234 		rcu_read_lock();
235 		/* The RCU read lock provides a memory barrier
236 		 * guaranteeing that if fqdir->dead is false then
237 		 * the hash table destruction will not start until
238 		 * after we unlock.  Paired with fqdir_pre_exit().
239 		 */
240 		if (!READ_ONCE(fqdir->dead)) {
241 			rhashtable_remove_fast(&fqdir->rhashtable, &fq->node,
242 					       fqdir->f->rhash_params);
243 			refcount_dec(&fq->refcnt);
244 		} else {
245 			fq->flags |= INET_FRAG_HASH_DEAD;
246 		}
247 		rcu_read_unlock();
248 	}
249 }
250 EXPORT_SYMBOL(inet_frag_kill);
251 
252 static void inet_frag_destroy_rcu(struct rcu_head *head)
253 {
254 	struct inet_frag_queue *q = container_of(head, struct inet_frag_queue,
255 						 rcu);
256 	struct inet_frags *f = q->fqdir->f;
257 
258 	if (f->destructor)
259 		f->destructor(q);
260 	kmem_cache_free(f->frags_cachep, q);
261 }
262 
263 unsigned int inet_frag_rbtree_purge(struct rb_root *root)
264 {
265 	struct rb_node *p = rb_first(root);
266 	unsigned int sum = 0;
267 
268 	while (p) {
269 		struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode);
270 
271 		p = rb_next(p);
272 		rb_erase(&skb->rbnode, root);
273 		while (skb) {
274 			struct sk_buff *next = FRAG_CB(skb)->next_frag;
275 
276 			sum += skb->truesize;
277 			kfree_skb(skb);
278 			skb = next;
279 		}
280 	}
281 	return sum;
282 }
283 EXPORT_SYMBOL(inet_frag_rbtree_purge);
284 
285 void inet_frag_destroy(struct inet_frag_queue *q)
286 {
287 	struct fqdir *fqdir;
288 	unsigned int sum, sum_truesize = 0;
289 	struct inet_frags *f;
290 
291 	WARN_ON(!(q->flags & INET_FRAG_COMPLETE));
292 	WARN_ON(del_timer(&q->timer) != 0);
293 
294 	/* Release all fragment data. */
295 	fqdir = q->fqdir;
296 	f = fqdir->f;
297 	sum_truesize = inet_frag_rbtree_purge(&q->rb_fragments);
298 	sum = sum_truesize + f->qsize;
299 
300 	call_rcu(&q->rcu, inet_frag_destroy_rcu);
301 
302 	sub_frag_mem_limit(fqdir, sum);
303 }
304 EXPORT_SYMBOL(inet_frag_destroy);
305 
306 static struct inet_frag_queue *inet_frag_alloc(struct fqdir *fqdir,
307 					       struct inet_frags *f,
308 					       void *arg)
309 {
310 	struct inet_frag_queue *q;
311 
312 	q = kmem_cache_zalloc(f->frags_cachep, GFP_ATOMIC);
313 	if (!q)
314 		return NULL;
315 
316 	q->fqdir = fqdir;
317 	f->constructor(q, arg);
318 	add_frag_mem_limit(fqdir, f->qsize);
319 
320 	timer_setup(&q->timer, f->frag_expire, 0);
321 	spin_lock_init(&q->lock);
322 	refcount_set(&q->refcnt, 3);
323 
324 	return q;
325 }
326 
327 static struct inet_frag_queue *inet_frag_create(struct fqdir *fqdir,
328 						void *arg,
329 						struct inet_frag_queue **prev)
330 {
331 	struct inet_frags *f = fqdir->f;
332 	struct inet_frag_queue *q;
333 
334 	q = inet_frag_alloc(fqdir, f, arg);
335 	if (!q) {
336 		*prev = ERR_PTR(-ENOMEM);
337 		return NULL;
338 	}
339 	mod_timer(&q->timer, jiffies + fqdir->timeout);
340 
341 	*prev = rhashtable_lookup_get_insert_key(&fqdir->rhashtable, &q->key,
342 						 &q->node, f->rhash_params);
343 	if (*prev) {
344 		q->flags |= INET_FRAG_COMPLETE;
345 		inet_frag_kill(q);
346 		inet_frag_destroy(q);
347 		return NULL;
348 	}
349 	return q;
350 }
351 
352 /* TODO : call from rcu_read_lock() and no longer use refcount_inc_not_zero() */
353 struct inet_frag_queue *inet_frag_find(struct fqdir *fqdir, void *key)
354 {
355 	/* This pairs with WRITE_ONCE() in fqdir_pre_exit(). */
356 	long high_thresh = READ_ONCE(fqdir->high_thresh);
357 	struct inet_frag_queue *fq = NULL, *prev;
358 
359 	if (!high_thresh || frag_mem_limit(fqdir) > high_thresh)
360 		return NULL;
361 
362 	rcu_read_lock();
363 
364 	prev = rhashtable_lookup(&fqdir->rhashtable, key, fqdir->f->rhash_params);
365 	if (!prev)
366 		fq = inet_frag_create(fqdir, key, &prev);
367 	if (!IS_ERR_OR_NULL(prev)) {
368 		fq = prev;
369 		if (!refcount_inc_not_zero(&fq->refcnt))
370 			fq = NULL;
371 	}
372 	rcu_read_unlock();
373 	return fq;
374 }
375 EXPORT_SYMBOL(inet_frag_find);
376 
377 int inet_frag_queue_insert(struct inet_frag_queue *q, struct sk_buff *skb,
378 			   int offset, int end)
379 {
380 	struct sk_buff *last = q->fragments_tail;
381 
382 	/* RFC5722, Section 4, amended by Errata ID : 3089
383 	 *                          When reassembling an IPv6 datagram, if
384 	 *   one or more its constituent fragments is determined to be an
385 	 *   overlapping fragment, the entire datagram (and any constituent
386 	 *   fragments) MUST be silently discarded.
387 	 *
388 	 * Duplicates, however, should be ignored (i.e. skb dropped, but the
389 	 * queue/fragments kept for later reassembly).
390 	 */
391 	if (!last)
392 		fragrun_create(q, skb);  /* First fragment. */
393 	else if (last->ip_defrag_offset + last->len < end) {
394 		/* This is the common case: skb goes to the end. */
395 		/* Detect and discard overlaps. */
396 		if (offset < last->ip_defrag_offset + last->len)
397 			return IPFRAG_OVERLAP;
398 		if (offset == last->ip_defrag_offset + last->len)
399 			fragrun_append_to_last(q, skb);
400 		else
401 			fragrun_create(q, skb);
402 	} else {
403 		/* Binary search. Note that skb can become the first fragment,
404 		 * but not the last (covered above).
405 		 */
406 		struct rb_node **rbn, *parent;
407 
408 		rbn = &q->rb_fragments.rb_node;
409 		do {
410 			struct sk_buff *curr;
411 			int curr_run_end;
412 
413 			parent = *rbn;
414 			curr = rb_to_skb(parent);
415 			curr_run_end = curr->ip_defrag_offset +
416 					FRAG_CB(curr)->frag_run_len;
417 			if (end <= curr->ip_defrag_offset)
418 				rbn = &parent->rb_left;
419 			else if (offset >= curr_run_end)
420 				rbn = &parent->rb_right;
421 			else if (offset >= curr->ip_defrag_offset &&
422 				 end <= curr_run_end)
423 				return IPFRAG_DUP;
424 			else
425 				return IPFRAG_OVERLAP;
426 		} while (*rbn);
427 		/* Here we have parent properly set, and rbn pointing to
428 		 * one of its NULL left/right children. Insert skb.
429 		 */
430 		fragcb_clear(skb);
431 		rb_link_node(&skb->rbnode, parent, rbn);
432 		rb_insert_color(&skb->rbnode, &q->rb_fragments);
433 	}
434 
435 	skb->ip_defrag_offset = offset;
436 
437 	return IPFRAG_OK;
438 }
439 EXPORT_SYMBOL(inet_frag_queue_insert);
440 
441 void *inet_frag_reasm_prepare(struct inet_frag_queue *q, struct sk_buff *skb,
442 			      struct sk_buff *parent)
443 {
444 	struct sk_buff *fp, *head = skb_rb_first(&q->rb_fragments);
445 	struct sk_buff **nextp;
446 	int delta;
447 
448 	if (head != skb) {
449 		fp = skb_clone(skb, GFP_ATOMIC);
450 		if (!fp)
451 			return NULL;
452 		FRAG_CB(fp)->next_frag = FRAG_CB(skb)->next_frag;
453 		if (RB_EMPTY_NODE(&skb->rbnode))
454 			FRAG_CB(parent)->next_frag = fp;
455 		else
456 			rb_replace_node(&skb->rbnode, &fp->rbnode,
457 					&q->rb_fragments);
458 		if (q->fragments_tail == skb)
459 			q->fragments_tail = fp;
460 		skb_morph(skb, head);
461 		FRAG_CB(skb)->next_frag = FRAG_CB(head)->next_frag;
462 		rb_replace_node(&head->rbnode, &skb->rbnode,
463 				&q->rb_fragments);
464 		consume_skb(head);
465 		head = skb;
466 	}
467 	WARN_ON(head->ip_defrag_offset != 0);
468 
469 	delta = -head->truesize;
470 
471 	/* Head of list must not be cloned. */
472 	if (skb_unclone(head, GFP_ATOMIC))
473 		return NULL;
474 
475 	delta += head->truesize;
476 	if (delta)
477 		add_frag_mem_limit(q->fqdir, delta);
478 
479 	/* If the first fragment is fragmented itself, we split
480 	 * it to two chunks: the first with data and paged part
481 	 * and the second, holding only fragments.
482 	 */
483 	if (skb_has_frag_list(head)) {
484 		struct sk_buff *clone;
485 		int i, plen = 0;
486 
487 		clone = alloc_skb(0, GFP_ATOMIC);
488 		if (!clone)
489 			return NULL;
490 		skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
491 		skb_frag_list_init(head);
492 		for (i = 0; i < skb_shinfo(head)->nr_frags; i++)
493 			plen += skb_frag_size(&skb_shinfo(head)->frags[i]);
494 		clone->data_len = head->data_len - plen;
495 		clone->len = clone->data_len;
496 		head->truesize += clone->truesize;
497 		clone->csum = 0;
498 		clone->ip_summed = head->ip_summed;
499 		add_frag_mem_limit(q->fqdir, clone->truesize);
500 		skb_shinfo(head)->frag_list = clone;
501 		nextp = &clone->next;
502 	} else {
503 		nextp = &skb_shinfo(head)->frag_list;
504 	}
505 
506 	return nextp;
507 }
508 EXPORT_SYMBOL(inet_frag_reasm_prepare);
509 
510 void inet_frag_reasm_finish(struct inet_frag_queue *q, struct sk_buff *head,
511 			    void *reasm_data, bool try_coalesce)
512 {
513 	struct sk_buff **nextp = reasm_data;
514 	struct rb_node *rbn;
515 	struct sk_buff *fp;
516 	int sum_truesize;
517 
518 	skb_push(head, head->data - skb_network_header(head));
519 
520 	/* Traverse the tree in order, to build frag_list. */
521 	fp = FRAG_CB(head)->next_frag;
522 	rbn = rb_next(&head->rbnode);
523 	rb_erase(&head->rbnode, &q->rb_fragments);
524 
525 	sum_truesize = head->truesize;
526 	while (rbn || fp) {
527 		/* fp points to the next sk_buff in the current run;
528 		 * rbn points to the next run.
529 		 */
530 		/* Go through the current run. */
531 		while (fp) {
532 			struct sk_buff *next_frag = FRAG_CB(fp)->next_frag;
533 			bool stolen;
534 			int delta;
535 
536 			sum_truesize += fp->truesize;
537 			if (head->ip_summed != fp->ip_summed)
538 				head->ip_summed = CHECKSUM_NONE;
539 			else if (head->ip_summed == CHECKSUM_COMPLETE)
540 				head->csum = csum_add(head->csum, fp->csum);
541 
542 			if (try_coalesce && skb_try_coalesce(head, fp, &stolen,
543 							     &delta)) {
544 				kfree_skb_partial(fp, stolen);
545 			} else {
546 				fp->prev = NULL;
547 				memset(&fp->rbnode, 0, sizeof(fp->rbnode));
548 				fp->sk = NULL;
549 
550 				head->data_len += fp->len;
551 				head->len += fp->len;
552 				head->truesize += fp->truesize;
553 
554 				*nextp = fp;
555 				nextp = &fp->next;
556 			}
557 
558 			fp = next_frag;
559 		}
560 		/* Move to the next run. */
561 		if (rbn) {
562 			struct rb_node *rbnext = rb_next(rbn);
563 
564 			fp = rb_to_skb(rbn);
565 			rb_erase(rbn, &q->rb_fragments);
566 			rbn = rbnext;
567 		}
568 	}
569 	sub_frag_mem_limit(q->fqdir, sum_truesize);
570 
571 	*nextp = NULL;
572 	skb_mark_not_on_list(head);
573 	head->prev = NULL;
574 	head->tstamp = q->stamp;
575 	head->mono_delivery_time = q->mono_delivery_time;
576 }
577 EXPORT_SYMBOL(inet_frag_reasm_finish);
578 
579 struct sk_buff *inet_frag_pull_head(struct inet_frag_queue *q)
580 {
581 	struct sk_buff *head, *skb;
582 
583 	head = skb_rb_first(&q->rb_fragments);
584 	if (!head)
585 		return NULL;
586 	skb = FRAG_CB(head)->next_frag;
587 	if (skb)
588 		rb_replace_node(&head->rbnode, &skb->rbnode,
589 				&q->rb_fragments);
590 	else
591 		rb_erase(&head->rbnode, &q->rb_fragments);
592 	memset(&head->rbnode, 0, sizeof(head->rbnode));
593 	barrier();
594 
595 	if (head == q->fragments_tail)
596 		q->fragments_tail = NULL;
597 
598 	sub_frag_mem_limit(q->fqdir, head->truesize);
599 
600 	return head;
601 }
602 EXPORT_SYMBOL(inet_frag_pull_head);
603