xref: /linux/net/ipv4/ip_fragment.c (revision 32594b09854970d7ba83eb2dc8c69a2edd158c8e)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * INET		An implementation of the TCP/IP protocol suite for the LINUX
4  *		operating system.  INET is implemented using the  BSD Socket
5  *		interface as the means of communication with the user level.
6  *
7  *		The IP fragmentation functionality.
8  *
9  * Authors:	Fred N. van Kempen <waltje@uWalt.NL.Mugnet.ORG>
10  *		Alan Cox <alan@lxorguk.ukuu.org.uk>
11  *
12  * Fixes:
13  *		Alan Cox	:	Split from ip.c , see ip_input.c for history.
14  *		David S. Miller :	Begin massive cleanup...
15  *		Andi Kleen	:	Add sysctls.
16  *		xxxx		:	Overlapfrag bug.
17  *		Ultima          :       ip_expire() kernel panic.
18  *		Bill Hawes	:	Frag accounting and evictor fixes.
19  *		John McDonald	:	0 length frag bug.
20  *		Alexey Kuznetsov:	SMP races, threading, cleanup.
21  *		Patrick McHardy :	LRU queue of frag heads for evictor.
22  */
23 
24 #define pr_fmt(fmt) "IPv4: " fmt
25 
26 #include <linux/compiler.h>
27 #include <linux/module.h>
28 #include <linux/types.h>
29 #include <linux/mm.h>
30 #include <linux/jiffies.h>
31 #include <linux/skbuff.h>
32 #include <linux/list.h>
33 #include <linux/ip.h>
34 #include <linux/icmp.h>
35 #include <linux/netdevice.h>
36 #include <linux/jhash.h>
37 #include <linux/random.h>
38 #include <linux/slab.h>
39 #include <net/route.h>
40 #include <net/dst.h>
41 #include <net/sock.h>
42 #include <net/ip.h>
43 #include <net/icmp.h>
44 #include <net/checksum.h>
45 #include <net/inetpeer.h>
46 #include <net/inet_frag.h>
47 #include <linux/tcp.h>
48 #include <linux/udp.h>
49 #include <linux/inet.h>
50 #include <linux/netfilter_ipv4.h>
51 #include <net/inet_ecn.h>
52 #include <net/l3mdev.h>
53 
54 /* NOTE. Logic of IP defragmentation is parallel to corresponding IPv6
55  * code now. If you change something here, _PLEASE_ update ipv6/reassembly.c
56  * as well. Or notify me, at least. --ANK
57  */
58 static const char ip_frag_cache_name[] = "ip4-frags";
59 
60 /* Describe an entry in the "incomplete datagrams" queue. */
61 struct ipq {
62 	struct inet_frag_queue q;
63 
64 	u8		ecn; /* RFC3168 support */
65 	u16		max_df_size; /* largest frag with DF set seen */
66 	int             iif;
67 	unsigned int    rid;
68 	struct inet_peer *peer;
69 };
70 
71 static u8 ip4_frag_ecn(u8 tos)
72 {
73 	return 1 << (tos & INET_ECN_MASK);
74 }
75 
76 static struct inet_frags ip4_frags;
77 
78 static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
79 			 struct sk_buff *prev_tail, struct net_device *dev,
80 			 int *refs);
81 
82 
83 static void ip4_frag_init(struct inet_frag_queue *q, const void *a)
84 {
85 	struct ipq *qp = container_of(q, struct ipq, q);
86 	const struct frag_v4_compare_key *key = a;
87 	struct net *net = q->fqdir->net;
88 	struct inet_peer *p = NULL;
89 
90 	q->key.v4 = *key;
91 	qp->ecn = 0;
92 	if (q->fqdir->max_dist) {
93 		rcu_read_lock();
94 		p = inet_getpeer_v4(net->ipv4.peers, key->saddr, key->vif);
95 		if (p && !refcount_inc_not_zero(&p->refcnt))
96 			p = NULL;
97 		rcu_read_unlock();
98 	}
99 	qp->peer = p;
100 }
101 
102 static void ip4_frag_free(struct inet_frag_queue *q)
103 {
104 	struct ipq *qp;
105 
106 	qp = container_of(q, struct ipq, q);
107 	if (qp->peer)
108 		inet_putpeer(qp->peer);
109 }
110 
111 static bool frag_expire_skip_icmp(u32 user)
112 {
113 	return user == IP_DEFRAG_AF_PACKET ||
114 	       ip_defrag_user_in_between(user, IP_DEFRAG_CONNTRACK_IN,
115 					 __IP_DEFRAG_CONNTRACK_IN_END) ||
116 	       ip_defrag_user_in_between(user, IP_DEFRAG_CONNTRACK_BRIDGE_IN,
117 					 __IP_DEFRAG_CONNTRACK_BRIDGE_IN);
118 }
119 
120 /*
121  * Oops, a fragment queue timed out.  Kill it and send an ICMP reply.
122  */
123 static void ip_expire(struct timer_list *t)
124 {
125 	enum skb_drop_reason reason = SKB_DROP_REASON_FRAG_REASM_TIMEOUT;
126 	struct inet_frag_queue *frag = timer_container_of(frag, t, timer);
127 	const struct iphdr *iph;
128 	struct sk_buff *head = NULL;
129 	struct net *net;
130 	struct ipq *qp;
131 	int refs = 1;
132 
133 	qp = container_of(frag, struct ipq, q);
134 	net = qp->q.fqdir->net;
135 
136 	rcu_read_lock();
137 	spin_lock(&qp->q.lock);
138 
139 	if (qp->q.flags & INET_FRAG_COMPLETE)
140 		goto out;
141 
142 	qp->q.flags |= INET_FRAG_DROP;
143 	inet_frag_kill(&qp->q, &refs);
144 
145 	/* Paired with WRITE_ONCE() in fqdir_pre_exit(). */
146 	if (READ_ONCE(qp->q.fqdir->dead)) {
147 		inet_frag_queue_flush(&qp->q, 0);
148 		goto out;
149 	}
150 
151 	__IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
152 	__IP_INC_STATS(net, IPSTATS_MIB_REASMTIMEOUT);
153 
154 	if (!(qp->q.flags & INET_FRAG_FIRST_IN))
155 		goto out;
156 
157 	/* sk_buff::dev and sk_buff::rbnode are unionized. So we
158 	 * pull the head out of the tree in order to be able to
159 	 * deal with head->dev.
160 	 */
161 	head = inet_frag_pull_head(&qp->q);
162 	if (!head)
163 		goto out;
164 	head->dev = dev_get_by_index_rcu(net, qp->iif);
165 	if (!head->dev)
166 		goto out;
167 
168 
169 	/* skb has no dst, perform route lookup again */
170 	iph = ip_hdr(head);
171 	reason = ip_route_input_noref(head, iph->daddr, iph->saddr,
172 				      ip4h_dscp(iph), head->dev);
173 	if (reason)
174 		goto out;
175 
176 	/* Only an end host needs to send an ICMP
177 	 * "Fragment Reassembly Timeout" message, per RFC792.
178 	 */
179 	reason = SKB_DROP_REASON_FRAG_REASM_TIMEOUT;
180 	if (frag_expire_skip_icmp(qp->q.key.v4.user) &&
181 	    (skb_rtable(head)->rt_type != RTN_LOCAL))
182 		goto out;
183 
184 	spin_unlock(&qp->q.lock);
185 	icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
186 	goto out_rcu_unlock;
187 
188 out:
189 	spin_unlock(&qp->q.lock);
190 out_rcu_unlock:
191 	rcu_read_unlock();
192 	kfree_skb_reason(head, reason);
193 	inet_frag_putn(&qp->q, refs);
194 }
195 
196 /* Find the correct entry in the "incomplete datagrams" queue for
197  * this IP datagram, and create new one, if nothing is found.
198  */
199 static struct ipq *ip_find(struct net *net, struct iphdr *iph,
200 			   u32 user, int vif)
201 {
202 	struct frag_v4_compare_key key = {
203 		.saddr = iph->saddr,
204 		.daddr = iph->daddr,
205 		.user = user,
206 		.vif = vif,
207 		.id = iph->id,
208 		.protocol = iph->protocol,
209 	};
210 	struct inet_frag_queue *q;
211 
212 	q = inet_frag_find(net->ipv4.fqdir, &key);
213 	if (!q)
214 		return NULL;
215 
216 	return container_of(q, struct ipq, q);
217 }
218 
219 /* Is the fragment too far ahead to be part of ipq? */
220 static int ip_frag_too_far(struct ipq *qp)
221 {
222 	struct inet_peer *peer = qp->peer;
223 	unsigned int max = qp->q.fqdir->max_dist;
224 	unsigned int start, end;
225 
226 	int rc;
227 
228 	if (!peer || !max)
229 		return 0;
230 
231 	start = qp->rid;
232 	end = atomic_inc_return(&peer->rid);
233 	qp->rid = end;
234 
235 	rc = qp->q.fragments_tail && (end - start) > max;
236 
237 	if (rc)
238 		__IP_INC_STATS(qp->q.fqdir->net, IPSTATS_MIB_REASMFAILS);
239 
240 	return rc;
241 }
242 
243 static int ip_frag_reinit(struct ipq *qp)
244 {
245 	if (!mod_timer_pending(&qp->q.timer, jiffies + qp->q.fqdir->timeout))
246 		return -ETIMEDOUT;
247 
248 	inet_frag_queue_flush(&qp->q, SKB_DROP_REASON_FRAG_TOO_FAR);
249 
250 	qp->q.flags = 0;
251 	qp->q.len = 0;
252 	qp->q.meat = 0;
253 	qp->iif = 0;
254 	qp->ecn = 0;
255 
256 	return 0;
257 }
258 
259 /* Add new segment to existing queue. */
260 static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb, int *refs)
261 {
262 	struct net *net = qp->q.fqdir->net;
263 	int ihl, end, flags, offset;
264 	struct sk_buff *prev_tail;
265 	struct net_device *dev;
266 	unsigned int fragsize;
267 	int err = -ENOENT;
268 	SKB_DR(reason);
269 	u8 ecn;
270 
271 	/* If reassembly is already done, @skb must be a duplicate frag. */
272 	if (qp->q.flags & INET_FRAG_COMPLETE) {
273 		SKB_DR_SET(reason, DUP_FRAG);
274 		goto err;
275 	}
276 
277 	if (!(IPCB(skb)->flags & IPSKB_FRAG_COMPLETE) &&
278 	    unlikely(ip_frag_too_far(qp)) &&
279 	    unlikely(err = ip_frag_reinit(qp))) {
280 		inet_frag_kill(&qp->q, refs);
281 		goto err;
282 	}
283 
284 	ecn = ip4_frag_ecn(ip_hdr(skb)->tos);
285 	offset = ntohs(ip_hdr(skb)->frag_off);
286 	flags = offset & ~IP_OFFSET;
287 	offset &= IP_OFFSET;
288 	offset <<= 3;		/* offset is in 8-byte chunks */
289 	ihl = ip_hdrlen(skb);
290 
291 	/* Determine the position of this fragment. */
292 	end = offset + skb->len - skb_network_offset(skb) - ihl;
293 	err = -EINVAL;
294 
295 	/* Is this the final fragment? */
296 	if ((flags & IP_MF) == 0) {
297 		/* If we already have some bits beyond end
298 		 * or have different end, the segment is corrupted.
299 		 */
300 		if (end < qp->q.len ||
301 		    ((qp->q.flags & INET_FRAG_LAST_IN) && end != qp->q.len))
302 			goto discard_qp;
303 		qp->q.flags |= INET_FRAG_LAST_IN;
304 		qp->q.len = end;
305 	} else {
306 		if (end&7) {
307 			end &= ~7;
308 			if (skb->ip_summed != CHECKSUM_UNNECESSARY)
309 				skb->ip_summed = CHECKSUM_NONE;
310 		}
311 		if (end > qp->q.len) {
312 			/* Some bits beyond end -> corruption. */
313 			if (qp->q.flags & INET_FRAG_LAST_IN)
314 				goto discard_qp;
315 			qp->q.len = end;
316 		}
317 	}
318 	if (end == offset)
319 		goto discard_qp;
320 
321 	err = -ENOMEM;
322 	if (!pskb_pull(skb, skb_network_offset(skb) + ihl))
323 		goto discard_qp;
324 
325 	err = pskb_trim_rcsum(skb, end - offset);
326 	if (err)
327 		goto discard_qp;
328 
329 	/* Note : skb->rbnode and skb->dev share the same location. */
330 	dev = skb->dev;
331 	/* Makes sure compiler wont do silly aliasing games */
332 	barrier();
333 
334 	prev_tail = qp->q.fragments_tail;
335 	err = inet_frag_queue_insert(&qp->q, skb, offset, end);
336 	if (err)
337 		goto insert_error;
338 
339 	if (dev)
340 		qp->iif = dev->ifindex;
341 
342 	qp->q.stamp = skb->tstamp;
343 	qp->q.tstamp_type = skb->tstamp_type;
344 	qp->q.meat += skb->len;
345 	qp->ecn |= ecn;
346 	add_frag_mem_limit(qp->q.fqdir, skb->truesize);
347 	if (offset == 0)
348 		qp->q.flags |= INET_FRAG_FIRST_IN;
349 
350 	fragsize = skb->len + ihl;
351 
352 	if (fragsize > qp->q.max_size)
353 		qp->q.max_size = fragsize;
354 
355 	if (ip_hdr(skb)->frag_off & htons(IP_DF) &&
356 	    fragsize > qp->max_df_size)
357 		qp->max_df_size = fragsize;
358 
359 	if (qp->q.flags == (INET_FRAG_FIRST_IN | INET_FRAG_LAST_IN) &&
360 	    qp->q.meat == qp->q.len) {
361 		unsigned long orefdst = skb->_skb_refdst;
362 
363 		skb->_skb_refdst = 0UL;
364 		err = ip_frag_reasm(qp, skb, prev_tail, dev, refs);
365 		skb->_skb_refdst = orefdst;
366 		if (err)
367 			inet_frag_kill(&qp->q, refs);
368 		return err;
369 	}
370 
371 	skb_dst_drop(skb);
372 	skb_orphan(skb);
373 	return -EINPROGRESS;
374 
375 insert_error:
376 	if (err == IPFRAG_DUP) {
377 		SKB_DR_SET(reason, DUP_FRAG);
378 		err = -EINVAL;
379 		goto err;
380 	}
381 	err = -EINVAL;
382 	__IP_INC_STATS(net, IPSTATS_MIB_REASM_OVERLAPS);
383 discard_qp:
384 	inet_frag_kill(&qp->q, refs);
385 	__IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
386 err:
387 	kfree_skb_reason(skb, reason);
388 	return err;
389 }
390 
391 static bool ip_frag_coalesce_ok(const struct ipq *qp)
392 {
393 	return qp->q.key.v4.user == IP_DEFRAG_LOCAL_DELIVER;
394 }
395 
396 /* Build a new IP datagram from all its fragments. */
397 static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
398 			 struct sk_buff *prev_tail, struct net_device *dev,
399 			 int *refs)
400 {
401 	struct net *net = qp->q.fqdir->net;
402 	struct iphdr *iph;
403 	void *reasm_data;
404 	int len, err;
405 	u8 ecn;
406 
407 	inet_frag_kill(&qp->q, refs);
408 
409 	ecn = ip_frag_ecn_table[qp->ecn];
410 	if (unlikely(ecn == 0xff)) {
411 		err = -EINVAL;
412 		goto out_fail;
413 	}
414 
415 	/* Make the one we just received the head. */
416 	reasm_data = inet_frag_reasm_prepare(&qp->q, skb, prev_tail);
417 	if (!reasm_data)
418 		goto out_nomem;
419 
420 	len = ip_hdrlen(skb) + qp->q.len;
421 	err = -E2BIG;
422 	if (len > 65535)
423 		goto out_oversize;
424 
425 	inet_frag_reasm_finish(&qp->q, skb, reasm_data,
426 			       ip_frag_coalesce_ok(qp));
427 
428 	skb->dev = dev;
429 	IPCB(skb)->frag_max_size = max(qp->max_df_size, qp->q.max_size);
430 
431 	iph = ip_hdr(skb);
432 	iph->tot_len = htons(len);
433 	iph->tos |= ecn;
434 
435 	/* When we set IP_DF on a refragmented skb we must also force a
436 	 * call to ip_fragment to avoid forwarding a DF-skb of size s while
437 	 * original sender only sent fragments of size f (where f < s).
438 	 *
439 	 * We only set DF/IPSKB_FRAG_PMTU if such DF fragment was the largest
440 	 * frag seen to avoid sending tiny DF-fragments in case skb was built
441 	 * from one very small df-fragment and one large non-df frag.
442 	 */
443 	if (qp->max_df_size == qp->q.max_size) {
444 		IPCB(skb)->flags |= IPSKB_FRAG_PMTU;
445 		iph->frag_off = htons(IP_DF);
446 	} else {
447 		iph->frag_off = 0;
448 	}
449 
450 	ip_send_check(iph);
451 
452 	__IP_INC_STATS(net, IPSTATS_MIB_REASMOKS);
453 	qp->q.rb_fragments = RB_ROOT;
454 	qp->q.fragments_tail = NULL;
455 	qp->q.last_run_head = NULL;
456 	return 0;
457 
458 out_nomem:
459 	net_dbg_ratelimited("queue_glue: no memory for gluing queue %p\n", qp);
460 	err = -ENOMEM;
461 	goto out_fail;
462 out_oversize:
463 	net_info_ratelimited("Oversized IP packet from %pI4\n", &qp->q.key.v4.saddr);
464 out_fail:
465 	__IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
466 	return err;
467 }
468 
469 /* Process an incoming IP datagram fragment. */
470 int ip_defrag(struct net *net, struct sk_buff *skb, u32 user)
471 {
472 	struct net_device *dev;
473 	struct ipq *qp;
474 	int vif;
475 
476 	__IP_INC_STATS(net, IPSTATS_MIB_REASMREQDS);
477 
478 	/* Lookup (or create) queue header */
479 	rcu_read_lock();
480 	dev = skb->dev ? : skb_dst_dev_rcu(skb);
481 	vif = l3mdev_master_ifindex_rcu(dev);
482 	qp = ip_find(net, ip_hdr(skb), user, vif);
483 	if (qp) {
484 		int ret, refs = 0;
485 
486 		spin_lock(&qp->q.lock);
487 
488 		ret = ip_frag_queue(qp, skb, &refs);
489 
490 		spin_unlock(&qp->q.lock);
491 		rcu_read_unlock();
492 		inet_frag_putn(&qp->q, refs);
493 		return ret;
494 	}
495 	rcu_read_unlock();
496 
497 	__IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
498 	kfree_skb(skb);
499 	return -ENOMEM;
500 }
501 EXPORT_SYMBOL(ip_defrag);
502 
503 struct sk_buff *ip_check_defrag(struct net *net, struct sk_buff *skb, u32 user)
504 {
505 	struct iphdr iph;
506 	int netoff;
507 	u32 len;
508 
509 	if (skb->protocol != htons(ETH_P_IP))
510 		return skb;
511 
512 	netoff = skb_network_offset(skb);
513 
514 	if (skb_copy_bits(skb, netoff, &iph, sizeof(iph)) < 0)
515 		return skb;
516 
517 	if (iph.ihl < 5 || iph.version != 4)
518 		return skb;
519 
520 	len = ntohs(iph.tot_len);
521 	if (skb->len < netoff + len || len < (iph.ihl * 4))
522 		return skb;
523 
524 	if (ip_is_fragment(&iph)) {
525 		skb = skb_share_check(skb, GFP_ATOMIC);
526 		if (skb) {
527 			if (!pskb_may_pull(skb, netoff + iph.ihl * 4)) {
528 				kfree_skb(skb);
529 				return NULL;
530 			}
531 			if (pskb_trim_rcsum(skb, netoff + len)) {
532 				kfree_skb(skb);
533 				return NULL;
534 			}
535 			memset(IPCB(skb), 0, sizeof(struct inet_skb_parm));
536 			if (ip_defrag(net, skb, user))
537 				return NULL;
538 			skb_clear_hash(skb);
539 		}
540 	}
541 	return skb;
542 }
543 EXPORT_SYMBOL(ip_check_defrag);
544 
545 #ifdef CONFIG_SYSCTL
546 static int dist_min;
547 
548 static struct ctl_table ip4_frags_ns_ctl_table[] = {
549 	{
550 		.procname	= "ipfrag_high_thresh",
551 		.maxlen		= sizeof(unsigned long),
552 		.mode		= 0644,
553 		.proc_handler	= proc_doulongvec_minmax,
554 	},
555 	{
556 		.procname	= "ipfrag_low_thresh",
557 		.maxlen		= sizeof(unsigned long),
558 		.mode		= 0644,
559 		.proc_handler	= proc_doulongvec_minmax,
560 	},
561 	{
562 		.procname	= "ipfrag_time",
563 		.maxlen		= sizeof(int),
564 		.mode		= 0644,
565 		.proc_handler	= proc_dointvec_jiffies,
566 	},
567 	{
568 		.procname	= "ipfrag_max_dist",
569 		.maxlen		= sizeof(int),
570 		.mode		= 0644,
571 		.proc_handler	= proc_dointvec_minmax,
572 		.extra1		= &dist_min,
573 	},
574 };
575 
576 /* secret interval has been deprecated */
577 static int ip4_frags_secret_interval_unused;
578 static struct ctl_table ip4_frags_ctl_table[] = {
579 	{
580 		.procname	= "ipfrag_secret_interval",
581 		.data		= &ip4_frags_secret_interval_unused,
582 		.maxlen		= sizeof(int),
583 		.mode		= 0644,
584 		.proc_handler	= proc_dointvec_jiffies,
585 	},
586 };
587 
588 static int __net_init ip4_frags_ns_ctl_register(struct net *net)
589 {
590 	struct ctl_table *table;
591 	struct ctl_table_header *hdr;
592 
593 	table = ip4_frags_ns_ctl_table;
594 	if (!net_eq(net, &init_net)) {
595 		table = kmemdup(table, sizeof(ip4_frags_ns_ctl_table), GFP_KERNEL);
596 		if (!table)
597 			goto err_alloc;
598 
599 	}
600 	table[0].data	= &net->ipv4.fqdir->high_thresh;
601 	table[0].extra1	= &net->ipv4.fqdir->low_thresh;
602 	table[1].data	= &net->ipv4.fqdir->low_thresh;
603 	table[1].extra2	= &net->ipv4.fqdir->high_thresh;
604 	table[2].data	= &net->ipv4.fqdir->timeout;
605 	table[3].data	= &net->ipv4.fqdir->max_dist;
606 
607 	hdr = register_net_sysctl_sz(net, "net/ipv4", table,
608 				     ARRAY_SIZE(ip4_frags_ns_ctl_table));
609 	if (!hdr)
610 		goto err_reg;
611 
612 	net->ipv4.frags_hdr = hdr;
613 	return 0;
614 
615 err_reg:
616 	if (!net_eq(net, &init_net))
617 		kfree(table);
618 err_alloc:
619 	return -ENOMEM;
620 }
621 
622 static void __net_exit ip4_frags_ns_ctl_unregister(struct net *net)
623 {
624 	const struct ctl_table *table;
625 
626 	table = net->ipv4.frags_hdr->ctl_table_arg;
627 	unregister_net_sysctl_table(net->ipv4.frags_hdr);
628 	kfree(table);
629 }
630 
631 static void __init ip4_frags_ctl_register(void)
632 {
633 	register_net_sysctl(&init_net, "net/ipv4", ip4_frags_ctl_table);
634 }
635 #else
636 static int ip4_frags_ns_ctl_register(struct net *net)
637 {
638 	return 0;
639 }
640 
641 static void ip4_frags_ns_ctl_unregister(struct net *net)
642 {
643 }
644 
645 static void __init ip4_frags_ctl_register(void)
646 {
647 }
648 #endif
649 
650 static int __net_init ipv4_frags_init_net(struct net *net)
651 {
652 	int res;
653 
654 	res = fqdir_init(&net->ipv4.fqdir, &ip4_frags, net);
655 	if (res < 0)
656 		return res;
657 	/* Fragment cache limits.
658 	 *
659 	 * The fragment memory accounting code, (tries to) account for
660 	 * the real memory usage, by measuring both the size of frag
661 	 * queue struct (inet_frag_queue (ipv4:ipq/ipv6:frag_queue))
662 	 * and the SKB's truesize.
663 	 *
664 	 * A 64K fragment consumes 129736 bytes (44*2944)+200
665 	 * (1500 truesize == 2944, sizeof(struct ipq) == 200)
666 	 *
667 	 * We will commit 4MB at one time. Should we cross that limit
668 	 * we will prune down to 3MB, making room for approx 8 big 64K
669 	 * fragments 8x128k.
670 	 */
671 	net->ipv4.fqdir->high_thresh = 4 * 1024 * 1024;
672 	net->ipv4.fqdir->low_thresh  = 3 * 1024 * 1024;
673 	/*
674 	 * Important NOTE! Fragment queue must be destroyed before MSL expires.
675 	 * RFC791 is wrong proposing to prolongate timer each fragment arrival
676 	 * by TTL.
677 	 */
678 	net->ipv4.fqdir->timeout = IP_FRAG_TIME;
679 
680 	net->ipv4.fqdir->max_dist = 64;
681 
682 	res = ip4_frags_ns_ctl_register(net);
683 	if (res < 0)
684 		fqdir_exit(net->ipv4.fqdir);
685 	return res;
686 }
687 
688 static void __net_exit ipv4_frags_pre_exit_net(struct net *net)
689 {
690 	fqdir_pre_exit(net->ipv4.fqdir);
691 }
692 
693 static void __net_exit ipv4_frags_exit_net(struct net *net)
694 {
695 	ip4_frags_ns_ctl_unregister(net);
696 	fqdir_exit(net->ipv4.fqdir);
697 }
698 
699 static struct pernet_operations ip4_frags_ops = {
700 	.init		= ipv4_frags_init_net,
701 	.pre_exit	= ipv4_frags_pre_exit_net,
702 	.exit		= ipv4_frags_exit_net,
703 };
704 
705 
706 static u32 ip4_key_hashfn(const void *data, u32 len, u32 seed)
707 {
708 	return jhash2(data,
709 		      sizeof(struct frag_v4_compare_key) / sizeof(u32), seed);
710 }
711 
712 static u32 ip4_obj_hashfn(const void *data, u32 len, u32 seed)
713 {
714 	const struct inet_frag_queue *fq = data;
715 
716 	return jhash2((const u32 *)&fq->key.v4,
717 		      sizeof(struct frag_v4_compare_key) / sizeof(u32), seed);
718 }
719 
720 static int ip4_obj_cmpfn(struct rhashtable_compare_arg *arg, const void *ptr)
721 {
722 	const struct frag_v4_compare_key *key = arg->key;
723 	const struct inet_frag_queue *fq = ptr;
724 
725 	return !!memcmp(&fq->key, key, sizeof(*key));
726 }
727 
728 static const struct rhashtable_params ip4_rhash_params = {
729 	.head_offset		= offsetof(struct inet_frag_queue, node),
730 	.key_offset		= offsetof(struct inet_frag_queue, key),
731 	.key_len		= sizeof(struct frag_v4_compare_key),
732 	.hashfn			= ip4_key_hashfn,
733 	.obj_hashfn		= ip4_obj_hashfn,
734 	.obj_cmpfn		= ip4_obj_cmpfn,
735 	.automatic_shrinking	= true,
736 };
737 
738 void __init ipfrag_init(void)
739 {
740 	ip4_frags.constructor = ip4_frag_init;
741 	ip4_frags.destructor = ip4_frag_free;
742 	ip4_frags.qsize = sizeof(struct ipq);
743 	ip4_frags.frag_expire = ip_expire;
744 	ip4_frags.frags_cache_name = ip_frag_cache_name;
745 	ip4_frags.rhash_params = ip4_rhash_params;
746 	if (inet_frags_init(&ip4_frags))
747 		panic("IP: failed to allocate ip4_frags cache\n");
748 	ip4_frags_ctl_register();
749 	register_pernet_subsys(&ip4_frags_ops);
750 }
751