xref: /freebsd/sys/kern/uipc_mbuf.c (revision e430d1ed78d021db4e9760d9800393b627156745)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1988, 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *	@(#)uipc_mbuf.c	8.2 (Berkeley) 1/4/94
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include "opt_param.h"
38 #include "opt_mbuf_stress_test.h"
39 #include "opt_mbuf_profiling.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/limits.h>
45 #include <sys/lock.h>
46 #include <sys/malloc.h>
47 #include <sys/mbuf.h>
48 #include <sys/sysctl.h>
49 #include <sys/domain.h>
50 #include <sys/protosw.h>
51 #include <sys/uio.h>
52 #include <sys/sdt.h>
53 
54 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__init,
55     "struct mbuf *", "mbufinfo_t *",
56     "uint32_t", "uint32_t",
57     "uint16_t", "uint16_t",
58     "uint32_t", "uint32_t",
59     "uint32_t", "uint32_t");
60 
61 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__gethdr,
62     "uint32_t", "uint32_t",
63     "uint16_t", "uint16_t",
64     "struct mbuf *", "mbufinfo_t *");
65 
66 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__get,
67     "uint32_t", "uint32_t",
68     "uint16_t", "uint16_t",
69     "struct mbuf *", "mbufinfo_t *");
70 
71 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__getcl,
72     "uint32_t", "uint32_t",
73     "uint16_t", "uint16_t",
74     "uint32_t", "uint32_t",
75     "struct mbuf *", "mbufinfo_t *");
76 
77 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__clget,
78     "struct mbuf *", "mbufinfo_t *",
79     "uint32_t", "uint32_t",
80     "uint32_t", "uint32_t");
81 
82 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__cljget,
83     "struct mbuf *", "mbufinfo_t *",
84     "uint32_t", "uint32_t",
85     "uint32_t", "uint32_t",
86     "void*", "void*");
87 
88 SDT_PROBE_DEFINE(sdt, , , m__cljset);
89 
90 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__free,
91         "struct mbuf *", "mbufinfo_t *");
92 
93 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__freem,
94     "struct mbuf *", "mbufinfo_t *");
95 
96 #include <security/mac/mac_framework.h>
97 
98 int	max_linkhdr;
99 int	max_protohdr;
100 int	max_hdr;
101 int	max_datalen;
102 #ifdef MBUF_STRESS_TEST
103 int	m_defragpackets;
104 int	m_defragbytes;
105 int	m_defraguseless;
106 int	m_defragfailure;
107 int	m_defragrandomfailures;
108 #endif
109 
110 /*
111  * sysctl(8) exported objects
112  */
113 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RD,
114 	   &max_linkhdr, 0, "Size of largest link layer header");
115 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RD,
116 	   &max_protohdr, 0, "Size of largest protocol layer header");
117 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RD,
118 	   &max_hdr, 0, "Size of largest link plus protocol header");
119 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RD,
120 	   &max_datalen, 0, "Minimum space left in mbuf after max_hdr");
121 #ifdef MBUF_STRESS_TEST
122 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
123 	   &m_defragpackets, 0, "");
124 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
125 	   &m_defragbytes, 0, "");
126 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
127 	   &m_defraguseless, 0, "");
128 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
129 	   &m_defragfailure, 0, "");
130 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
131 	   &m_defragrandomfailures, 0, "");
132 #endif
133 
134 /*
135  * Ensure the correct size of various mbuf parameters.  It could be off due
136  * to compiler-induced padding and alignment artifacts.
137  */
138 CTASSERT(MSIZE - offsetof(struct mbuf, m_dat) == MLEN);
139 CTASSERT(MSIZE - offsetof(struct mbuf, m_pktdat) == MHLEN);
140 
141 /*
142  * mbuf data storage should be 64-bit aligned regardless of architectural
143  * pointer size; check this is the case with and without a packet header.
144  */
145 CTASSERT(offsetof(struct mbuf, m_dat) % 8 == 0);
146 CTASSERT(offsetof(struct mbuf, m_pktdat) % 8 == 0);
147 
148 /*
149  * While the specific values here don't matter too much (i.e., +/- a few
150  * words), we do want to ensure that changes to these values are carefully
151  * reasoned about and properly documented.  This is especially the case as
152  * network-protocol and device-driver modules encode these layouts, and must
153  * be recompiled if the structures change.  Check these values at compile time
154  * against the ones documented in comments in mbuf.h.
155  *
156  * NB: Possibly they should be documented there via #define's and not just
157  * comments.
158  */
159 #if defined(__LP64__)
160 CTASSERT(offsetof(struct mbuf, m_dat) == 32);
161 CTASSERT(sizeof(struct pkthdr) == 56);
162 CTASSERT(sizeof(struct m_ext) == 48);
163 #else
164 CTASSERT(offsetof(struct mbuf, m_dat) == 24);
165 CTASSERT(sizeof(struct pkthdr) == 48);
166 CTASSERT(sizeof(struct m_ext) == 28);
167 #endif
168 
169 /*
170  * Assert that the queue(3) macros produce code of the same size as an old
171  * plain pointer does.
172  */
173 #ifdef INVARIANTS
174 static struct mbuf __used m_assertbuf;
175 CTASSERT(sizeof(m_assertbuf.m_slist) == sizeof(m_assertbuf.m_next));
176 CTASSERT(sizeof(m_assertbuf.m_stailq) == sizeof(m_assertbuf.m_next));
177 CTASSERT(sizeof(m_assertbuf.m_slistpkt) == sizeof(m_assertbuf.m_nextpkt));
178 CTASSERT(sizeof(m_assertbuf.m_stailqpkt) == sizeof(m_assertbuf.m_nextpkt));
179 #endif
180 
181 /*
182  * Attach the cluster from *m to *n, set up m_ext in *n
183  * and bump the refcount of the cluster.
184  */
185 void
186 mb_dupcl(struct mbuf *n, struct mbuf *m)
187 {
188 	volatile u_int *refcnt;
189 
190 	KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
191 	KASSERT(!(n->m_flags & M_EXT), ("%s: M_EXT set on %p", __func__, n));
192 
193 	/*
194 	 * Cache access optimization.  For most kinds of external
195 	 * storage we don't need full copy of m_ext, since the
196 	 * holder of the 'ext_count' is responsible to carry the
197 	 * free routine and its arguments.  Exclusion is EXT_EXTREF,
198 	 * where 'ext_cnt' doesn't point into mbuf at all.
199 	 */
200 	if (m->m_ext.ext_type == EXT_EXTREF)
201 		bcopy(&m->m_ext, &n->m_ext, sizeof(struct m_ext));
202 	else
203 		bcopy(&m->m_ext, &n->m_ext, m_ext_copylen);
204 	n->m_flags |= M_EXT;
205 	n->m_flags |= m->m_flags & M_RDONLY;
206 
207 	/* See if this is the mbuf that holds the embedded refcount. */
208 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
209 		refcnt = n->m_ext.ext_cnt = &m->m_ext.ext_count;
210 		n->m_ext.ext_flags &= ~EXT_FLAG_EMBREF;
211 	} else {
212 		KASSERT(m->m_ext.ext_cnt != NULL,
213 		    ("%s: no refcounting pointer on %p", __func__, m));
214 		refcnt = m->m_ext.ext_cnt;
215 	}
216 
217 	if (*refcnt == 1)
218 		*refcnt += 1;
219 	else
220 		atomic_add_int(refcnt, 1);
221 }
222 
223 void
224 m_demote_pkthdr(struct mbuf *m)
225 {
226 
227 	M_ASSERTPKTHDR(m);
228 
229 	m_tag_delete_chain(m, NULL);
230 	m->m_flags &= ~M_PKTHDR;
231 	bzero(&m->m_pkthdr, sizeof(struct pkthdr));
232 }
233 
234 /*
235  * Clean up mbuf (chain) from any tags and packet headers.
236  * If "all" is set then the first mbuf in the chain will be
237  * cleaned too.
238  */
239 void
240 m_demote(struct mbuf *m0, int all, int flags)
241 {
242 	struct mbuf *m;
243 
244 	for (m = all ? m0 : m0->m_next; m != NULL; m = m->m_next) {
245 		KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt in m %p, m0 %p",
246 		    __func__, m, m0));
247 		if (m->m_flags & M_PKTHDR)
248 			m_demote_pkthdr(m);
249 		m->m_flags = m->m_flags & (M_EXT | M_RDONLY | M_NOFREE | flags);
250 	}
251 }
252 
253 /*
254  * Sanity checks on mbuf (chain) for use in KASSERT() and general
255  * debugging.
256  * Returns 0 or panics when bad and 1 on all tests passed.
257  * Sanitize, 0 to run M_SANITY_ACTION, 1 to garble things so they
258  * blow up later.
259  */
260 int
261 m_sanity(struct mbuf *m0, int sanitize)
262 {
263 	struct mbuf *m;
264 	caddr_t a, b;
265 	int pktlen = 0;
266 
267 #ifdef INVARIANTS
268 #define	M_SANITY_ACTION(s)	panic("mbuf %p: " s, m)
269 #else
270 #define	M_SANITY_ACTION(s)	printf("mbuf %p: " s, m)
271 #endif
272 
273 	for (m = m0; m != NULL; m = m->m_next) {
274 		/*
275 		 * Basic pointer checks.  If any of these fails then some
276 		 * unrelated kernel memory before or after us is trashed.
277 		 * No way to recover from that.
278 		 */
279 		a = M_START(m);
280 		b = a + M_SIZE(m);
281 		if ((caddr_t)m->m_data < a)
282 			M_SANITY_ACTION("m_data outside mbuf data range left");
283 		if ((caddr_t)m->m_data > b)
284 			M_SANITY_ACTION("m_data outside mbuf data range right");
285 		if ((caddr_t)m->m_data + m->m_len > b)
286 			M_SANITY_ACTION("m_data + m_len exeeds mbuf space");
287 
288 		/* m->m_nextpkt may only be set on first mbuf in chain. */
289 		if (m != m0 && m->m_nextpkt != NULL) {
290 			if (sanitize) {
291 				m_freem(m->m_nextpkt);
292 				m->m_nextpkt = (struct mbuf *)0xDEADC0DE;
293 			} else
294 				M_SANITY_ACTION("m->m_nextpkt on in-chain mbuf");
295 		}
296 
297 		/* packet length (not mbuf length!) calculation */
298 		if (m0->m_flags & M_PKTHDR)
299 			pktlen += m->m_len;
300 
301 		/* m_tags may only be attached to first mbuf in chain. */
302 		if (m != m0 && m->m_flags & M_PKTHDR &&
303 		    !SLIST_EMPTY(&m->m_pkthdr.tags)) {
304 			if (sanitize) {
305 				m_tag_delete_chain(m, NULL);
306 				/* put in 0xDEADC0DE perhaps? */
307 			} else
308 				M_SANITY_ACTION("m_tags on in-chain mbuf");
309 		}
310 
311 		/* M_PKTHDR may only be set on first mbuf in chain */
312 		if (m != m0 && m->m_flags & M_PKTHDR) {
313 			if (sanitize) {
314 				bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
315 				m->m_flags &= ~M_PKTHDR;
316 				/* put in 0xDEADCODE and leave hdr flag in */
317 			} else
318 				M_SANITY_ACTION("M_PKTHDR on in-chain mbuf");
319 		}
320 	}
321 	m = m0;
322 	if (pktlen && pktlen != m->m_pkthdr.len) {
323 		if (sanitize)
324 			m->m_pkthdr.len = 0;
325 		else
326 			M_SANITY_ACTION("m_pkthdr.len != mbuf chain length");
327 	}
328 	return 1;
329 
330 #undef	M_SANITY_ACTION
331 }
332 
333 /*
334  * Non-inlined part of m_init().
335  */
336 int
337 m_pkthdr_init(struct mbuf *m, int how)
338 {
339 #ifdef MAC
340 	int error;
341 #endif
342 	m->m_data = m->m_pktdat;
343 	bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
344 #ifdef NUMA
345 	m->m_pkthdr.numa_domain = M_NODOM;
346 #endif
347 #ifdef MAC
348 	/* If the label init fails, fail the alloc */
349 	error = mac_mbuf_init(m, how);
350 	if (error)
351 		return (error);
352 #endif
353 
354 	return (0);
355 }
356 
357 /*
358  * "Move" mbuf pkthdr from "from" to "to".
359  * "from" must have M_PKTHDR set, and "to" must be empty.
360  */
361 void
362 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
363 {
364 
365 #if 0
366 	/* see below for why these are not enabled */
367 	M_ASSERTPKTHDR(to);
368 	/* Note: with MAC, this may not be a good assertion. */
369 	KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags),
370 	    ("m_move_pkthdr: to has tags"));
371 #endif
372 #ifdef MAC
373 	/*
374 	 * XXXMAC: It could be this should also occur for non-MAC?
375 	 */
376 	if (to->m_flags & M_PKTHDR)
377 		m_tag_delete_chain(to, NULL);
378 #endif
379 	to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT);
380 	if ((to->m_flags & M_EXT) == 0)
381 		to->m_data = to->m_pktdat;
382 	to->m_pkthdr = from->m_pkthdr;		/* especially tags */
383 	SLIST_INIT(&from->m_pkthdr.tags);	/* purge tags from src */
384 	from->m_flags &= ~M_PKTHDR;
385 }
386 
387 /*
388  * Duplicate "from"'s mbuf pkthdr in "to".
389  * "from" must have M_PKTHDR set, and "to" must be empty.
390  * In particular, this does a deep copy of the packet tags.
391  */
392 int
393 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
394 {
395 
396 #if 0
397 	/*
398 	 * The mbuf allocator only initializes the pkthdr
399 	 * when the mbuf is allocated with m_gethdr(). Many users
400 	 * (e.g. m_copy*, m_prepend) use m_get() and then
401 	 * smash the pkthdr as needed causing these
402 	 * assertions to trip.  For now just disable them.
403 	 */
404 	M_ASSERTPKTHDR(to);
405 	/* Note: with MAC, this may not be a good assertion. */
406 	KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags), ("m_dup_pkthdr: to has tags"));
407 #endif
408 	MBUF_CHECKSLEEP(how);
409 #ifdef MAC
410 	if (to->m_flags & M_PKTHDR)
411 		m_tag_delete_chain(to, NULL);
412 #endif
413 	to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT);
414 	if ((to->m_flags & M_EXT) == 0)
415 		to->m_data = to->m_pktdat;
416 	to->m_pkthdr = from->m_pkthdr;
417 	SLIST_INIT(&to->m_pkthdr.tags);
418 	return (m_tag_copy_chain(to, from, how));
419 }
420 
421 /*
422  * Lesser-used path for M_PREPEND:
423  * allocate new mbuf to prepend to chain,
424  * copy junk along.
425  */
426 struct mbuf *
427 m_prepend(struct mbuf *m, int len, int how)
428 {
429 	struct mbuf *mn;
430 
431 	if (m->m_flags & M_PKTHDR)
432 		mn = m_gethdr(how, m->m_type);
433 	else
434 		mn = m_get(how, m->m_type);
435 	if (mn == NULL) {
436 		m_freem(m);
437 		return (NULL);
438 	}
439 	if (m->m_flags & M_PKTHDR)
440 		m_move_pkthdr(mn, m);
441 	mn->m_next = m;
442 	m = mn;
443 	if (len < M_SIZE(m))
444 		M_ALIGN(m, len);
445 	m->m_len = len;
446 	return (m);
447 }
448 
449 /*
450  * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
451  * continuing for "len" bytes.  If len is M_COPYALL, copy to end of mbuf.
452  * The wait parameter is a choice of M_WAITOK/M_NOWAIT from caller.
453  * Note that the copy is read-only, because clusters are not copied,
454  * only their reference counts are incremented.
455  */
456 struct mbuf *
457 m_copym(struct mbuf *m, int off0, int len, int wait)
458 {
459 	struct mbuf *n, **np;
460 	int off = off0;
461 	struct mbuf *top;
462 	int copyhdr = 0;
463 
464 	KASSERT(off >= 0, ("m_copym, negative off %d", off));
465 	KASSERT(len >= 0, ("m_copym, negative len %d", len));
466 	MBUF_CHECKSLEEP(wait);
467 	if (off == 0 && m->m_flags & M_PKTHDR)
468 		copyhdr = 1;
469 	while (off > 0) {
470 		KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
471 		if (off < m->m_len)
472 			break;
473 		off -= m->m_len;
474 		m = m->m_next;
475 	}
476 	np = &top;
477 	top = NULL;
478 	while (len > 0) {
479 		if (m == NULL) {
480 			KASSERT(len == M_COPYALL,
481 			    ("m_copym, length > size of mbuf chain"));
482 			break;
483 		}
484 		if (copyhdr)
485 			n = m_gethdr(wait, m->m_type);
486 		else
487 			n = m_get(wait, m->m_type);
488 		*np = n;
489 		if (n == NULL)
490 			goto nospace;
491 		if (copyhdr) {
492 			if (!m_dup_pkthdr(n, m, wait))
493 				goto nospace;
494 			if (len == M_COPYALL)
495 				n->m_pkthdr.len -= off0;
496 			else
497 				n->m_pkthdr.len = len;
498 			copyhdr = 0;
499 		}
500 		n->m_len = min(len, m->m_len - off);
501 		if (m->m_flags & M_EXT) {
502 			n->m_data = m->m_data + off;
503 			mb_dupcl(n, m);
504 		} else
505 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
506 			    (u_int)n->m_len);
507 		if (len != M_COPYALL)
508 			len -= n->m_len;
509 		off = 0;
510 		m = m->m_next;
511 		np = &n->m_next;
512 	}
513 
514 	return (top);
515 nospace:
516 	m_freem(top);
517 	return (NULL);
518 }
519 
520 /*
521  * Copy an entire packet, including header (which must be present).
522  * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
523  * Note that the copy is read-only, because clusters are not copied,
524  * only their reference counts are incremented.
525  * Preserve alignment of the first mbuf so if the creator has left
526  * some room at the beginning (e.g. for inserting protocol headers)
527  * the copies still have the room available.
528  */
529 struct mbuf *
530 m_copypacket(struct mbuf *m, int how)
531 {
532 	struct mbuf *top, *n, *o;
533 
534 	MBUF_CHECKSLEEP(how);
535 	n = m_get(how, m->m_type);
536 	top = n;
537 	if (n == NULL)
538 		goto nospace;
539 
540 	if (!m_dup_pkthdr(n, m, how))
541 		goto nospace;
542 	n->m_len = m->m_len;
543 	if (m->m_flags & M_EXT) {
544 		n->m_data = m->m_data;
545 		mb_dupcl(n, m);
546 	} else {
547 		n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
548 		bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
549 	}
550 
551 	m = m->m_next;
552 	while (m) {
553 		o = m_get(how, m->m_type);
554 		if (o == NULL)
555 			goto nospace;
556 
557 		n->m_next = o;
558 		n = n->m_next;
559 
560 		n->m_len = m->m_len;
561 		if (m->m_flags & M_EXT) {
562 			n->m_data = m->m_data;
563 			mb_dupcl(n, m);
564 		} else {
565 			bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
566 		}
567 
568 		m = m->m_next;
569 	}
570 	return top;
571 nospace:
572 	m_freem(top);
573 	return (NULL);
574 }
575 
576 /*
577  * Copy data from an mbuf chain starting "off" bytes from the beginning,
578  * continuing for "len" bytes, into the indicated buffer.
579  */
580 void
581 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
582 {
583 	u_int count;
584 
585 	KASSERT(off >= 0, ("m_copydata, negative off %d", off));
586 	KASSERT(len >= 0, ("m_copydata, negative len %d", len));
587 	while (off > 0) {
588 		KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
589 		if (off < m->m_len)
590 			break;
591 		off -= m->m_len;
592 		m = m->m_next;
593 	}
594 	while (len > 0) {
595 		KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
596 		count = min(m->m_len - off, len);
597 		bcopy(mtod(m, caddr_t) + off, cp, count);
598 		len -= count;
599 		cp += count;
600 		off = 0;
601 		m = m->m_next;
602 	}
603 }
604 
605 /*
606  * Copy a packet header mbuf chain into a completely new chain, including
607  * copying any mbuf clusters.  Use this instead of m_copypacket() when
608  * you need a writable copy of an mbuf chain.
609  */
610 struct mbuf *
611 m_dup(const struct mbuf *m, int how)
612 {
613 	struct mbuf **p, *top = NULL;
614 	int remain, moff, nsize;
615 
616 	MBUF_CHECKSLEEP(how);
617 	/* Sanity check */
618 	if (m == NULL)
619 		return (NULL);
620 	M_ASSERTPKTHDR(m);
621 
622 	/* While there's more data, get a new mbuf, tack it on, and fill it */
623 	remain = m->m_pkthdr.len;
624 	moff = 0;
625 	p = &top;
626 	while (remain > 0 || top == NULL) {	/* allow m->m_pkthdr.len == 0 */
627 		struct mbuf *n;
628 
629 		/* Get the next new mbuf */
630 		if (remain >= MINCLSIZE) {
631 			n = m_getcl(how, m->m_type, 0);
632 			nsize = MCLBYTES;
633 		} else {
634 			n = m_get(how, m->m_type);
635 			nsize = MLEN;
636 		}
637 		if (n == NULL)
638 			goto nospace;
639 
640 		if (top == NULL) {		/* First one, must be PKTHDR */
641 			if (!m_dup_pkthdr(n, m, how)) {
642 				m_free(n);
643 				goto nospace;
644 			}
645 			if ((n->m_flags & M_EXT) == 0)
646 				nsize = MHLEN;
647 			n->m_flags &= ~M_RDONLY;
648 		}
649 		n->m_len = 0;
650 
651 		/* Link it into the new chain */
652 		*p = n;
653 		p = &n->m_next;
654 
655 		/* Copy data from original mbuf(s) into new mbuf */
656 		while (n->m_len < nsize && m != NULL) {
657 			int chunk = min(nsize - n->m_len, m->m_len - moff);
658 
659 			bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
660 			moff += chunk;
661 			n->m_len += chunk;
662 			remain -= chunk;
663 			if (moff == m->m_len) {
664 				m = m->m_next;
665 				moff = 0;
666 			}
667 		}
668 
669 		/* Check correct total mbuf length */
670 		KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
671 		    	("%s: bogus m_pkthdr.len", __func__));
672 	}
673 	return (top);
674 
675 nospace:
676 	m_freem(top);
677 	return (NULL);
678 }
679 
680 /*
681  * Concatenate mbuf chain n to m.
682  * Both chains must be of the same type (e.g. MT_DATA).
683  * Any m_pkthdr is not updated.
684  */
685 void
686 m_cat(struct mbuf *m, struct mbuf *n)
687 {
688 	while (m->m_next)
689 		m = m->m_next;
690 	while (n) {
691 		if (!M_WRITABLE(m) ||
692 		    M_TRAILINGSPACE(m) < n->m_len) {
693 			/* just join the two chains */
694 			m->m_next = n;
695 			return;
696 		}
697 		/* splat the data from one into the other */
698 		bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
699 		    (u_int)n->m_len);
700 		m->m_len += n->m_len;
701 		n = m_free(n);
702 	}
703 }
704 
705 /*
706  * Concatenate two pkthdr mbuf chains.
707  */
708 void
709 m_catpkt(struct mbuf *m, struct mbuf *n)
710 {
711 
712 	M_ASSERTPKTHDR(m);
713 	M_ASSERTPKTHDR(n);
714 
715 	m->m_pkthdr.len += n->m_pkthdr.len;
716 	m_demote(n, 1, 0);
717 
718 	m_cat(m, n);
719 }
720 
721 void
722 m_adj(struct mbuf *mp, int req_len)
723 {
724 	int len = req_len;
725 	struct mbuf *m;
726 	int count;
727 
728 	if ((m = mp) == NULL)
729 		return;
730 	if (len >= 0) {
731 		/*
732 		 * Trim from head.
733 		 */
734 		while (m != NULL && len > 0) {
735 			if (m->m_len <= len) {
736 				len -= m->m_len;
737 				m->m_len = 0;
738 				m = m->m_next;
739 			} else {
740 				m->m_len -= len;
741 				m->m_data += len;
742 				len = 0;
743 			}
744 		}
745 		if (mp->m_flags & M_PKTHDR)
746 			mp->m_pkthdr.len -= (req_len - len);
747 	} else {
748 		/*
749 		 * Trim from tail.  Scan the mbuf chain,
750 		 * calculating its length and finding the last mbuf.
751 		 * If the adjustment only affects this mbuf, then just
752 		 * adjust and return.  Otherwise, rescan and truncate
753 		 * after the remaining size.
754 		 */
755 		len = -len;
756 		count = 0;
757 		for (;;) {
758 			count += m->m_len;
759 			if (m->m_next == (struct mbuf *)0)
760 				break;
761 			m = m->m_next;
762 		}
763 		if (m->m_len >= len) {
764 			m->m_len -= len;
765 			if (mp->m_flags & M_PKTHDR)
766 				mp->m_pkthdr.len -= len;
767 			return;
768 		}
769 		count -= len;
770 		if (count < 0)
771 			count = 0;
772 		/*
773 		 * Correct length for chain is "count".
774 		 * Find the mbuf with last data, adjust its length,
775 		 * and toss data from remaining mbufs on chain.
776 		 */
777 		m = mp;
778 		if (m->m_flags & M_PKTHDR)
779 			m->m_pkthdr.len = count;
780 		for (; m; m = m->m_next) {
781 			if (m->m_len >= count) {
782 				m->m_len = count;
783 				if (m->m_next != NULL) {
784 					m_freem(m->m_next);
785 					m->m_next = NULL;
786 				}
787 				break;
788 			}
789 			count -= m->m_len;
790 		}
791 	}
792 }
793 
794 /*
795  * Rearange an mbuf chain so that len bytes are contiguous
796  * and in the data area of an mbuf (so that mtod will work
797  * for a structure of size len).  Returns the resulting
798  * mbuf chain on success, frees it and returns null on failure.
799  * If there is room, it will add up to max_protohdr-len extra bytes to the
800  * contiguous region in an attempt to avoid being called next time.
801  */
802 struct mbuf *
803 m_pullup(struct mbuf *n, int len)
804 {
805 	struct mbuf *m;
806 	int count;
807 	int space;
808 
809 	/*
810 	 * If first mbuf has no cluster, and has room for len bytes
811 	 * without shifting current data, pullup into it,
812 	 * otherwise allocate a new mbuf to prepend to the chain.
813 	 */
814 	if ((n->m_flags & M_EXT) == 0 &&
815 	    n->m_data + len < &n->m_dat[MLEN] && n->m_next) {
816 		if (n->m_len >= len)
817 			return (n);
818 		m = n;
819 		n = n->m_next;
820 		len -= m->m_len;
821 	} else {
822 		if (len > MHLEN)
823 			goto bad;
824 		m = m_get(M_NOWAIT, n->m_type);
825 		if (m == NULL)
826 			goto bad;
827 		if (n->m_flags & M_PKTHDR)
828 			m_move_pkthdr(m, n);
829 	}
830 	space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
831 	do {
832 		count = min(min(max(len, max_protohdr), space), n->m_len);
833 		bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
834 		  (u_int)count);
835 		len -= count;
836 		m->m_len += count;
837 		n->m_len -= count;
838 		space -= count;
839 		if (n->m_len)
840 			n->m_data += count;
841 		else
842 			n = m_free(n);
843 	} while (len > 0 && n);
844 	if (len > 0) {
845 		(void) m_free(m);
846 		goto bad;
847 	}
848 	m->m_next = n;
849 	return (m);
850 bad:
851 	m_freem(n);
852 	return (NULL);
853 }
854 
855 /*
856  * Like m_pullup(), except a new mbuf is always allocated, and we allow
857  * the amount of empty space before the data in the new mbuf to be specified
858  * (in the event that the caller expects to prepend later).
859  */
860 struct mbuf *
861 m_copyup(struct mbuf *n, int len, int dstoff)
862 {
863 	struct mbuf *m;
864 	int count, space;
865 
866 	if (len > (MHLEN - dstoff))
867 		goto bad;
868 	m = m_get(M_NOWAIT, n->m_type);
869 	if (m == NULL)
870 		goto bad;
871 	if (n->m_flags & M_PKTHDR)
872 		m_move_pkthdr(m, n);
873 	m->m_data += dstoff;
874 	space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
875 	do {
876 		count = min(min(max(len, max_protohdr), space), n->m_len);
877 		memcpy(mtod(m, caddr_t) + m->m_len, mtod(n, caddr_t),
878 		    (unsigned)count);
879 		len -= count;
880 		m->m_len += count;
881 		n->m_len -= count;
882 		space -= count;
883 		if (n->m_len)
884 			n->m_data += count;
885 		else
886 			n = m_free(n);
887 	} while (len > 0 && n);
888 	if (len > 0) {
889 		(void) m_free(m);
890 		goto bad;
891 	}
892 	m->m_next = n;
893 	return (m);
894  bad:
895 	m_freem(n);
896 	return (NULL);
897 }
898 
899 /*
900  * Partition an mbuf chain in two pieces, returning the tail --
901  * all but the first len0 bytes.  In case of failure, it returns NULL and
902  * attempts to restore the chain to its original state.
903  *
904  * Note that the resulting mbufs might be read-only, because the new
905  * mbuf can end up sharing an mbuf cluster with the original mbuf if
906  * the "breaking point" happens to lie within a cluster mbuf. Use the
907  * M_WRITABLE() macro to check for this case.
908  */
909 struct mbuf *
910 m_split(struct mbuf *m0, int len0, int wait)
911 {
912 	struct mbuf *m, *n;
913 	u_int len = len0, remain;
914 
915 	MBUF_CHECKSLEEP(wait);
916 	for (m = m0; m && len > m->m_len; m = m->m_next)
917 		len -= m->m_len;
918 	if (m == NULL)
919 		return (NULL);
920 	remain = m->m_len - len;
921 	if (m0->m_flags & M_PKTHDR && remain == 0) {
922 		n = m_gethdr(wait, m0->m_type);
923 		if (n == NULL)
924 			return (NULL);
925 		n->m_next = m->m_next;
926 		m->m_next = NULL;
927 		n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
928 		n->m_pkthdr.len = m0->m_pkthdr.len - len0;
929 		m0->m_pkthdr.len = len0;
930 		return (n);
931 	} else if (m0->m_flags & M_PKTHDR) {
932 		n = m_gethdr(wait, m0->m_type);
933 		if (n == NULL)
934 			return (NULL);
935 		n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
936 		n->m_pkthdr.len = m0->m_pkthdr.len - len0;
937 		m0->m_pkthdr.len = len0;
938 		if (m->m_flags & M_EXT)
939 			goto extpacket;
940 		if (remain > MHLEN) {
941 			/* m can't be the lead packet */
942 			M_ALIGN(n, 0);
943 			n->m_next = m_split(m, len, wait);
944 			if (n->m_next == NULL) {
945 				(void) m_free(n);
946 				return (NULL);
947 			} else {
948 				n->m_len = 0;
949 				return (n);
950 			}
951 		} else
952 			M_ALIGN(n, remain);
953 	} else if (remain == 0) {
954 		n = m->m_next;
955 		m->m_next = NULL;
956 		return (n);
957 	} else {
958 		n = m_get(wait, m->m_type);
959 		if (n == NULL)
960 			return (NULL);
961 		M_ALIGN(n, remain);
962 	}
963 extpacket:
964 	if (m->m_flags & M_EXT) {
965 		n->m_data = m->m_data + len;
966 		mb_dupcl(n, m);
967 	} else {
968 		bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
969 	}
970 	n->m_len = remain;
971 	m->m_len = len;
972 	n->m_next = m->m_next;
973 	m->m_next = NULL;
974 	return (n);
975 }
976 /*
977  * Routine to copy from device local memory into mbufs.
978  * Note that `off' argument is offset into first mbuf of target chain from
979  * which to begin copying the data to.
980  */
981 struct mbuf *
982 m_devget(char *buf, int totlen, int off, struct ifnet *ifp,
983     void (*copy)(char *from, caddr_t to, u_int len))
984 {
985 	struct mbuf *m;
986 	struct mbuf *top = NULL, **mp = &top;
987 	int len;
988 
989 	if (off < 0 || off > MHLEN)
990 		return (NULL);
991 
992 	while (totlen > 0) {
993 		if (top == NULL) {	/* First one, must be PKTHDR */
994 			if (totlen + off >= MINCLSIZE) {
995 				m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
996 				len = MCLBYTES;
997 			} else {
998 				m = m_gethdr(M_NOWAIT, MT_DATA);
999 				len = MHLEN;
1000 
1001 				/* Place initial small packet/header at end of mbuf */
1002 				if (m && totlen + off + max_linkhdr <= MHLEN) {
1003 					m->m_data += max_linkhdr;
1004 					len -= max_linkhdr;
1005 				}
1006 			}
1007 			if (m == NULL)
1008 				return NULL;
1009 			m->m_pkthdr.rcvif = ifp;
1010 			m->m_pkthdr.len = totlen;
1011 		} else {
1012 			if (totlen + off >= MINCLSIZE) {
1013 				m = m_getcl(M_NOWAIT, MT_DATA, 0);
1014 				len = MCLBYTES;
1015 			} else {
1016 				m = m_get(M_NOWAIT, MT_DATA);
1017 				len = MLEN;
1018 			}
1019 			if (m == NULL) {
1020 				m_freem(top);
1021 				return NULL;
1022 			}
1023 		}
1024 		if (off) {
1025 			m->m_data += off;
1026 			len -= off;
1027 			off = 0;
1028 		}
1029 		m->m_len = len = min(totlen, len);
1030 		if (copy)
1031 			copy(buf, mtod(m, caddr_t), (u_int)len);
1032 		else
1033 			bcopy(buf, mtod(m, caddr_t), (u_int)len);
1034 		buf += len;
1035 		*mp = m;
1036 		mp = &m->m_next;
1037 		totlen -= len;
1038 	}
1039 	return (top);
1040 }
1041 
1042 /*
1043  * Copy data from a buffer back into the indicated mbuf chain,
1044  * starting "off" bytes from the beginning, extending the mbuf
1045  * chain if necessary.
1046  */
1047 void
1048 m_copyback(struct mbuf *m0, int off, int len, c_caddr_t cp)
1049 {
1050 	int mlen;
1051 	struct mbuf *m = m0, *n;
1052 	int totlen = 0;
1053 
1054 	if (m0 == NULL)
1055 		return;
1056 	while (off > (mlen = m->m_len)) {
1057 		off -= mlen;
1058 		totlen += mlen;
1059 		if (m->m_next == NULL) {
1060 			n = m_get(M_NOWAIT, m->m_type);
1061 			if (n == NULL)
1062 				goto out;
1063 			bzero(mtod(n, caddr_t), MLEN);
1064 			n->m_len = min(MLEN, len + off);
1065 			m->m_next = n;
1066 		}
1067 		m = m->m_next;
1068 	}
1069 	while (len > 0) {
1070 		if (m->m_next == NULL && (len > m->m_len - off)) {
1071 			m->m_len += min(len - (m->m_len - off),
1072 			    M_TRAILINGSPACE(m));
1073 		}
1074 		mlen = min (m->m_len - off, len);
1075 		bcopy(cp, off + mtod(m, caddr_t), (u_int)mlen);
1076 		cp += mlen;
1077 		len -= mlen;
1078 		mlen += off;
1079 		off = 0;
1080 		totlen += mlen;
1081 		if (len == 0)
1082 			break;
1083 		if (m->m_next == NULL) {
1084 			n = m_get(M_NOWAIT, m->m_type);
1085 			if (n == NULL)
1086 				break;
1087 			n->m_len = min(MLEN, len);
1088 			m->m_next = n;
1089 		}
1090 		m = m->m_next;
1091 	}
1092 out:	if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1093 		m->m_pkthdr.len = totlen;
1094 }
1095 
1096 /*
1097  * Append the specified data to the indicated mbuf chain,
1098  * Extend the mbuf chain if the new data does not fit in
1099  * existing space.
1100  *
1101  * Return 1 if able to complete the job; otherwise 0.
1102  */
1103 int
1104 m_append(struct mbuf *m0, int len, c_caddr_t cp)
1105 {
1106 	struct mbuf *m, *n;
1107 	int remainder, space;
1108 
1109 	for (m = m0; m->m_next != NULL; m = m->m_next)
1110 		;
1111 	remainder = len;
1112 	space = M_TRAILINGSPACE(m);
1113 	if (space > 0) {
1114 		/*
1115 		 * Copy into available space.
1116 		 */
1117 		if (space > remainder)
1118 			space = remainder;
1119 		bcopy(cp, mtod(m, caddr_t) + m->m_len, space);
1120 		m->m_len += space;
1121 		cp += space, remainder -= space;
1122 	}
1123 	while (remainder > 0) {
1124 		/*
1125 		 * Allocate a new mbuf; could check space
1126 		 * and allocate a cluster instead.
1127 		 */
1128 		n = m_get(M_NOWAIT, m->m_type);
1129 		if (n == NULL)
1130 			break;
1131 		n->m_len = min(MLEN, remainder);
1132 		bcopy(cp, mtod(n, caddr_t), n->m_len);
1133 		cp += n->m_len, remainder -= n->m_len;
1134 		m->m_next = n;
1135 		m = n;
1136 	}
1137 	if (m0->m_flags & M_PKTHDR)
1138 		m0->m_pkthdr.len += len - remainder;
1139 	return (remainder == 0);
1140 }
1141 
1142 /*
1143  * Apply function f to the data in an mbuf chain starting "off" bytes from
1144  * the beginning, continuing for "len" bytes.
1145  */
1146 int
1147 m_apply(struct mbuf *m, int off, int len,
1148     int (*f)(void *, void *, u_int), void *arg)
1149 {
1150 	u_int count;
1151 	int rval;
1152 
1153 	KASSERT(off >= 0, ("m_apply, negative off %d", off));
1154 	KASSERT(len >= 0, ("m_apply, negative len %d", len));
1155 	while (off > 0) {
1156 		KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1157 		if (off < m->m_len)
1158 			break;
1159 		off -= m->m_len;
1160 		m = m->m_next;
1161 	}
1162 	while (len > 0) {
1163 		KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1164 		count = min(m->m_len - off, len);
1165 		rval = (*f)(arg, mtod(m, caddr_t) + off, count);
1166 		if (rval)
1167 			return (rval);
1168 		len -= count;
1169 		off = 0;
1170 		m = m->m_next;
1171 	}
1172 	return (0);
1173 }
1174 
1175 /*
1176  * Return a pointer to mbuf/offset of location in mbuf chain.
1177  */
1178 struct mbuf *
1179 m_getptr(struct mbuf *m, int loc, int *off)
1180 {
1181 
1182 	while (loc >= 0) {
1183 		/* Normal end of search. */
1184 		if (m->m_len > loc) {
1185 			*off = loc;
1186 			return (m);
1187 		} else {
1188 			loc -= m->m_len;
1189 			if (m->m_next == NULL) {
1190 				if (loc == 0) {
1191 					/* Point at the end of valid data. */
1192 					*off = m->m_len;
1193 					return (m);
1194 				}
1195 				return (NULL);
1196 			}
1197 			m = m->m_next;
1198 		}
1199 	}
1200 	return (NULL);
1201 }
1202 
1203 void
1204 m_print(const struct mbuf *m, int maxlen)
1205 {
1206 	int len;
1207 	int pdata;
1208 	const struct mbuf *m2;
1209 
1210 	if (m == NULL) {
1211 		printf("mbuf: %p\n", m);
1212 		return;
1213 	}
1214 
1215 	if (m->m_flags & M_PKTHDR)
1216 		len = m->m_pkthdr.len;
1217 	else
1218 		len = -1;
1219 	m2 = m;
1220 	while (m2 != NULL && (len == -1 || len)) {
1221 		pdata = m2->m_len;
1222 		if (maxlen != -1 && pdata > maxlen)
1223 			pdata = maxlen;
1224 		printf("mbuf: %p len: %d, next: %p, %b%s", m2, m2->m_len,
1225 		    m2->m_next, m2->m_flags, "\20\20freelist\17skipfw"
1226 		    "\11proto5\10proto4\7proto3\6proto2\5proto1\4rdonly"
1227 		    "\3eor\2pkthdr\1ext", pdata ? "" : "\n");
1228 		if (pdata)
1229 			printf(", %*D\n", pdata, (u_char *)m2->m_data, "-");
1230 		if (len != -1)
1231 			len -= m2->m_len;
1232 		m2 = m2->m_next;
1233 	}
1234 	if (len > 0)
1235 		printf("%d bytes unaccounted for.\n", len);
1236 	return;
1237 }
1238 
1239 u_int
1240 m_fixhdr(struct mbuf *m0)
1241 {
1242 	u_int len;
1243 
1244 	len = m_length(m0, NULL);
1245 	m0->m_pkthdr.len = len;
1246 	return (len);
1247 }
1248 
1249 u_int
1250 m_length(struct mbuf *m0, struct mbuf **last)
1251 {
1252 	struct mbuf *m;
1253 	u_int len;
1254 
1255 	len = 0;
1256 	for (m = m0; m != NULL; m = m->m_next) {
1257 		len += m->m_len;
1258 		if (m->m_next == NULL)
1259 			break;
1260 	}
1261 	if (last != NULL)
1262 		*last = m;
1263 	return (len);
1264 }
1265 
1266 /*
1267  * Defragment a mbuf chain, returning the shortest possible
1268  * chain of mbufs and clusters.  If allocation fails and
1269  * this cannot be completed, NULL will be returned, but
1270  * the passed in chain will be unchanged.  Upon success,
1271  * the original chain will be freed, and the new chain
1272  * will be returned.
1273  *
1274  * If a non-packet header is passed in, the original
1275  * mbuf (chain?) will be returned unharmed.
1276  */
1277 struct mbuf *
1278 m_defrag(struct mbuf *m0, int how)
1279 {
1280 	struct mbuf *m_new = NULL, *m_final = NULL;
1281 	int progress = 0, length;
1282 
1283 	MBUF_CHECKSLEEP(how);
1284 	if (!(m0->m_flags & M_PKTHDR))
1285 		return (m0);
1286 
1287 	m_fixhdr(m0); /* Needed sanity check */
1288 
1289 #ifdef MBUF_STRESS_TEST
1290 	if (m_defragrandomfailures) {
1291 		int temp = arc4random() & 0xff;
1292 		if (temp == 0xba)
1293 			goto nospace;
1294 	}
1295 #endif
1296 
1297 	if (m0->m_pkthdr.len > MHLEN)
1298 		m_final = m_getcl(how, MT_DATA, M_PKTHDR);
1299 	else
1300 		m_final = m_gethdr(how, MT_DATA);
1301 
1302 	if (m_final == NULL)
1303 		goto nospace;
1304 
1305 	if (m_dup_pkthdr(m_final, m0, how) == 0)
1306 		goto nospace;
1307 
1308 	m_new = m_final;
1309 
1310 	while (progress < m0->m_pkthdr.len) {
1311 		length = m0->m_pkthdr.len - progress;
1312 		if (length > MCLBYTES)
1313 			length = MCLBYTES;
1314 
1315 		if (m_new == NULL) {
1316 			if (length > MLEN)
1317 				m_new = m_getcl(how, MT_DATA, 0);
1318 			else
1319 				m_new = m_get(how, MT_DATA);
1320 			if (m_new == NULL)
1321 				goto nospace;
1322 		}
1323 
1324 		m_copydata(m0, progress, length, mtod(m_new, caddr_t));
1325 		progress += length;
1326 		m_new->m_len = length;
1327 		if (m_new != m_final)
1328 			m_cat(m_final, m_new);
1329 		m_new = NULL;
1330 	}
1331 #ifdef MBUF_STRESS_TEST
1332 	if (m0->m_next == NULL)
1333 		m_defraguseless++;
1334 #endif
1335 	m_freem(m0);
1336 	m0 = m_final;
1337 #ifdef MBUF_STRESS_TEST
1338 	m_defragpackets++;
1339 	m_defragbytes += m0->m_pkthdr.len;
1340 #endif
1341 	return (m0);
1342 nospace:
1343 #ifdef MBUF_STRESS_TEST
1344 	m_defragfailure++;
1345 #endif
1346 	if (m_final)
1347 		m_freem(m_final);
1348 	return (NULL);
1349 }
1350 
1351 /*
1352  * Defragment an mbuf chain, returning at most maxfrags separate
1353  * mbufs+clusters.  If this is not possible NULL is returned and
1354  * the original mbuf chain is left in its present (potentially
1355  * modified) state.  We use two techniques: collapsing consecutive
1356  * mbufs and replacing consecutive mbufs by a cluster.
1357  *
1358  * NB: this should really be named m_defrag but that name is taken
1359  */
1360 struct mbuf *
1361 m_collapse(struct mbuf *m0, int how, int maxfrags)
1362 {
1363 	struct mbuf *m, *n, *n2, **prev;
1364 	u_int curfrags;
1365 
1366 	/*
1367 	 * Calculate the current number of frags.
1368 	 */
1369 	curfrags = 0;
1370 	for (m = m0; m != NULL; m = m->m_next)
1371 		curfrags++;
1372 	/*
1373 	 * First, try to collapse mbufs.  Note that we always collapse
1374 	 * towards the front so we don't need to deal with moving the
1375 	 * pkthdr.  This may be suboptimal if the first mbuf has much
1376 	 * less data than the following.
1377 	 */
1378 	m = m0;
1379 again:
1380 	for (;;) {
1381 		n = m->m_next;
1382 		if (n == NULL)
1383 			break;
1384 		if (M_WRITABLE(m) &&
1385 		    n->m_len < M_TRAILINGSPACE(m)) {
1386 			bcopy(mtod(n, void *), mtod(m, char *) + m->m_len,
1387 				n->m_len);
1388 			m->m_len += n->m_len;
1389 			m->m_next = n->m_next;
1390 			m_free(n);
1391 			if (--curfrags <= maxfrags)
1392 				return m0;
1393 		} else
1394 			m = n;
1395 	}
1396 	KASSERT(maxfrags > 1,
1397 		("maxfrags %u, but normal collapse failed", maxfrags));
1398 	/*
1399 	 * Collapse consecutive mbufs to a cluster.
1400 	 */
1401 	prev = &m0->m_next;		/* NB: not the first mbuf */
1402 	while ((n = *prev) != NULL) {
1403 		if ((n2 = n->m_next) != NULL &&
1404 		    n->m_len + n2->m_len < MCLBYTES) {
1405 			m = m_getcl(how, MT_DATA, 0);
1406 			if (m == NULL)
1407 				goto bad;
1408 			bcopy(mtod(n, void *), mtod(m, void *), n->m_len);
1409 			bcopy(mtod(n2, void *), mtod(m, char *) + n->m_len,
1410 				n2->m_len);
1411 			m->m_len = n->m_len + n2->m_len;
1412 			m->m_next = n2->m_next;
1413 			*prev = m;
1414 			m_free(n);
1415 			m_free(n2);
1416 			if (--curfrags <= maxfrags)	/* +1 cl -2 mbufs */
1417 				return m0;
1418 			/*
1419 			 * Still not there, try the normal collapse
1420 			 * again before we allocate another cluster.
1421 			 */
1422 			goto again;
1423 		}
1424 		prev = &n->m_next;
1425 	}
1426 	/*
1427 	 * No place where we can collapse to a cluster; punt.
1428 	 * This can occur if, for example, you request 2 frags
1429 	 * but the packet requires that both be clusters (we
1430 	 * never reallocate the first mbuf to avoid moving the
1431 	 * packet header).
1432 	 */
1433 bad:
1434 	return NULL;
1435 }
1436 
1437 #ifdef MBUF_STRESS_TEST
1438 
1439 /*
1440  * Fragment an mbuf chain.  There's no reason you'd ever want to do
1441  * this in normal usage, but it's great for stress testing various
1442  * mbuf consumers.
1443  *
1444  * If fragmentation is not possible, the original chain will be
1445  * returned.
1446  *
1447  * Possible length values:
1448  * 0	 no fragmentation will occur
1449  * > 0	each fragment will be of the specified length
1450  * -1	each fragment will be the same random value in length
1451  * -2	each fragment's length will be entirely random
1452  * (Random values range from 1 to 256)
1453  */
1454 struct mbuf *
1455 m_fragment(struct mbuf *m0, int how, int length)
1456 {
1457 	struct mbuf *m_first, *m_last;
1458 	int divisor = 255, progress = 0, fraglen;
1459 
1460 	if (!(m0->m_flags & M_PKTHDR))
1461 		return (m0);
1462 
1463 	if (length == 0 || length < -2)
1464 		return (m0);
1465 	if (length > MCLBYTES)
1466 		length = MCLBYTES;
1467 	if (length < 0 && divisor > MCLBYTES)
1468 		divisor = MCLBYTES;
1469 	if (length == -1)
1470 		length = 1 + (arc4random() % divisor);
1471 	if (length > 0)
1472 		fraglen = length;
1473 
1474 	m_fixhdr(m0); /* Needed sanity check */
1475 
1476 	m_first = m_getcl(how, MT_DATA, M_PKTHDR);
1477 	if (m_first == NULL)
1478 		goto nospace;
1479 
1480 	if (m_dup_pkthdr(m_first, m0, how) == 0)
1481 		goto nospace;
1482 
1483 	m_last = m_first;
1484 
1485 	while (progress < m0->m_pkthdr.len) {
1486 		if (length == -2)
1487 			fraglen = 1 + (arc4random() % divisor);
1488 		if (fraglen > m0->m_pkthdr.len - progress)
1489 			fraglen = m0->m_pkthdr.len - progress;
1490 
1491 		if (progress != 0) {
1492 			struct mbuf *m_new = m_getcl(how, MT_DATA, 0);
1493 			if (m_new == NULL)
1494 				goto nospace;
1495 
1496 			m_last->m_next = m_new;
1497 			m_last = m_new;
1498 		}
1499 
1500 		m_copydata(m0, progress, fraglen, mtod(m_last, caddr_t));
1501 		progress += fraglen;
1502 		m_last->m_len = fraglen;
1503 	}
1504 	m_freem(m0);
1505 	m0 = m_first;
1506 	return (m0);
1507 nospace:
1508 	if (m_first)
1509 		m_freem(m_first);
1510 	/* Return the original chain on failure */
1511 	return (m0);
1512 }
1513 
1514 #endif
1515 
1516 /*
1517  * Copy the contents of uio into a properly sized mbuf chain.
1518  */
1519 struct mbuf *
1520 m_uiotombuf(struct uio *uio, int how, int len, int align, int flags)
1521 {
1522 	struct mbuf *m, *mb;
1523 	int error, length;
1524 	ssize_t total;
1525 	int progress = 0;
1526 
1527 	/*
1528 	 * len can be zero or an arbitrary large value bound by
1529 	 * the total data supplied by the uio.
1530 	 */
1531 	if (len > 0)
1532 		total = (uio->uio_resid < len) ? uio->uio_resid : len;
1533 	else
1534 		total = uio->uio_resid;
1535 
1536 	/*
1537 	 * The smallest unit returned by m_getm2() is a single mbuf
1538 	 * with pkthdr.  We can't align past it.
1539 	 */
1540 	if (align >= MHLEN)
1541 		return (NULL);
1542 
1543 	/*
1544 	 * Give us the full allocation or nothing.
1545 	 * If len is zero return the smallest empty mbuf.
1546 	 */
1547 	m = m_getm2(NULL, max(total + align, 1), how, MT_DATA, flags);
1548 	if (m == NULL)
1549 		return (NULL);
1550 	m->m_data += align;
1551 
1552 	/* Fill all mbufs with uio data and update header information. */
1553 	for (mb = m; mb != NULL; mb = mb->m_next) {
1554 		length = min(M_TRAILINGSPACE(mb), total - progress);
1555 
1556 		error = uiomove(mtod(mb, void *), length, uio);
1557 		if (error) {
1558 			m_freem(m);
1559 			return (NULL);
1560 		}
1561 
1562 		mb->m_len = length;
1563 		progress += length;
1564 		if (flags & M_PKTHDR)
1565 			m->m_pkthdr.len += length;
1566 	}
1567 	KASSERT(progress == total, ("%s: progress != total", __func__));
1568 
1569 	return (m);
1570 }
1571 
1572 /*
1573  * Copy an mbuf chain into a uio limited by len if set.
1574  */
1575 int
1576 m_mbuftouio(struct uio *uio, const struct mbuf *m, int len)
1577 {
1578 	int error, length, total;
1579 	int progress = 0;
1580 
1581 	if (len > 0)
1582 		total = min(uio->uio_resid, len);
1583 	else
1584 		total = uio->uio_resid;
1585 
1586 	/* Fill the uio with data from the mbufs. */
1587 	for (; m != NULL; m = m->m_next) {
1588 		length = min(m->m_len, total - progress);
1589 
1590 		error = uiomove(mtod(m, void *), length, uio);
1591 		if (error)
1592 			return (error);
1593 
1594 		progress += length;
1595 	}
1596 
1597 	return (0);
1598 }
1599 
1600 /*
1601  * Create a writable copy of the mbuf chain.  While doing this
1602  * we compact the chain with a goal of producing a chain with
1603  * at most two mbufs.  The second mbuf in this chain is likely
1604  * to be a cluster.  The primary purpose of this work is to create
1605  * a writable packet for encryption, compression, etc.  The
1606  * secondary goal is to linearize the data so the data can be
1607  * passed to crypto hardware in the most efficient manner possible.
1608  */
1609 struct mbuf *
1610 m_unshare(struct mbuf *m0, int how)
1611 {
1612 	struct mbuf *m, *mprev;
1613 	struct mbuf *n, *mfirst, *mlast;
1614 	int len, off;
1615 
1616 	mprev = NULL;
1617 	for (m = m0; m != NULL; m = mprev->m_next) {
1618 		/*
1619 		 * Regular mbufs are ignored unless there's a cluster
1620 		 * in front of it that we can use to coalesce.  We do
1621 		 * the latter mainly so later clusters can be coalesced
1622 		 * also w/o having to handle them specially (i.e. convert
1623 		 * mbuf+cluster -> cluster).  This optimization is heavily
1624 		 * influenced by the assumption that we're running over
1625 		 * Ethernet where MCLBYTES is large enough that the max
1626 		 * packet size will permit lots of coalescing into a
1627 		 * single cluster.  This in turn permits efficient
1628 		 * crypto operations, especially when using hardware.
1629 		 */
1630 		if ((m->m_flags & M_EXT) == 0) {
1631 			if (mprev && (mprev->m_flags & M_EXT) &&
1632 			    m->m_len <= M_TRAILINGSPACE(mprev)) {
1633 				/* XXX: this ignores mbuf types */
1634 				memcpy(mtod(mprev, caddr_t) + mprev->m_len,
1635 				    mtod(m, caddr_t), m->m_len);
1636 				mprev->m_len += m->m_len;
1637 				mprev->m_next = m->m_next;	/* unlink from chain */
1638 				m_free(m);			/* reclaim mbuf */
1639 			} else {
1640 				mprev = m;
1641 			}
1642 			continue;
1643 		}
1644 		/*
1645 		 * Writable mbufs are left alone (for now).
1646 		 */
1647 		if (M_WRITABLE(m)) {
1648 			mprev = m;
1649 			continue;
1650 		}
1651 
1652 		/*
1653 		 * Not writable, replace with a copy or coalesce with
1654 		 * the previous mbuf if possible (since we have to copy
1655 		 * it anyway, we try to reduce the number of mbufs and
1656 		 * clusters so that future work is easier).
1657 		 */
1658 		KASSERT(m->m_flags & M_EXT, ("m_flags 0x%x", m->m_flags));
1659 		/* NB: we only coalesce into a cluster or larger */
1660 		if (mprev != NULL && (mprev->m_flags & M_EXT) &&
1661 		    m->m_len <= M_TRAILINGSPACE(mprev)) {
1662 			/* XXX: this ignores mbuf types */
1663 			memcpy(mtod(mprev, caddr_t) + mprev->m_len,
1664 			    mtod(m, caddr_t), m->m_len);
1665 			mprev->m_len += m->m_len;
1666 			mprev->m_next = m->m_next;	/* unlink from chain */
1667 			m_free(m);			/* reclaim mbuf */
1668 			continue;
1669 		}
1670 
1671 		/*
1672 		 * Allocate new space to hold the copy and copy the data.
1673 		 * We deal with jumbo mbufs (i.e. m_len > MCLBYTES) by
1674 		 * splitting them into clusters.  We could just malloc a
1675 		 * buffer and make it external but too many device drivers
1676 		 * don't know how to break up the non-contiguous memory when
1677 		 * doing DMA.
1678 		 */
1679 		n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS);
1680 		if (n == NULL) {
1681 			m_freem(m0);
1682 			return (NULL);
1683 		}
1684 		if (m->m_flags & M_PKTHDR) {
1685 			KASSERT(mprev == NULL, ("%s: m0 %p, m %p has M_PKTHDR",
1686 			    __func__, m0, m));
1687 			m_move_pkthdr(n, m);
1688 		}
1689 		len = m->m_len;
1690 		off = 0;
1691 		mfirst = n;
1692 		mlast = NULL;
1693 		for (;;) {
1694 			int cc = min(len, MCLBYTES);
1695 			memcpy(mtod(n, caddr_t), mtod(m, caddr_t) + off, cc);
1696 			n->m_len = cc;
1697 			if (mlast != NULL)
1698 				mlast->m_next = n;
1699 			mlast = n;
1700 #if 0
1701 			newipsecstat.ips_clcopied++;
1702 #endif
1703 
1704 			len -= cc;
1705 			if (len <= 0)
1706 				break;
1707 			off += cc;
1708 
1709 			n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS);
1710 			if (n == NULL) {
1711 				m_freem(mfirst);
1712 				m_freem(m0);
1713 				return (NULL);
1714 			}
1715 		}
1716 		n->m_next = m->m_next;
1717 		if (mprev == NULL)
1718 			m0 = mfirst;		/* new head of chain */
1719 		else
1720 			mprev->m_next = mfirst;	/* replace old mbuf */
1721 		m_free(m);			/* release old mbuf */
1722 		mprev = mfirst;
1723 	}
1724 	return (m0);
1725 }
1726 
1727 #ifdef MBUF_PROFILING
1728 
1729 #define MP_BUCKETS 32 /* don't just change this as things may overflow.*/
1730 struct mbufprofile {
1731 	uintmax_t wasted[MP_BUCKETS];
1732 	uintmax_t used[MP_BUCKETS];
1733 	uintmax_t segments[MP_BUCKETS];
1734 } mbprof;
1735 
1736 #define MP_MAXDIGITS 21	/* strlen("16,000,000,000,000,000,000") == 21 */
1737 #define MP_NUMLINES 6
1738 #define MP_NUMSPERLINE 16
1739 #define MP_EXTRABYTES 64	/* > strlen("used:\nwasted:\nsegments:\n") */
1740 /* work out max space needed and add a bit of spare space too */
1741 #define MP_MAXLINE ((MP_MAXDIGITS+1) * MP_NUMSPERLINE)
1742 #define MP_BUFSIZE ((MP_MAXLINE * MP_NUMLINES) + 1 + MP_EXTRABYTES)
1743 
1744 char mbprofbuf[MP_BUFSIZE];
1745 
1746 void
1747 m_profile(struct mbuf *m)
1748 {
1749 	int segments = 0;
1750 	int used = 0;
1751 	int wasted = 0;
1752 
1753 	while (m) {
1754 		segments++;
1755 		used += m->m_len;
1756 		if (m->m_flags & M_EXT) {
1757 			wasted += MHLEN - sizeof(m->m_ext) +
1758 			    m->m_ext.ext_size - m->m_len;
1759 		} else {
1760 			if (m->m_flags & M_PKTHDR)
1761 				wasted += MHLEN - m->m_len;
1762 			else
1763 				wasted += MLEN - m->m_len;
1764 		}
1765 		m = m->m_next;
1766 	}
1767 	/* be paranoid.. it helps */
1768 	if (segments > MP_BUCKETS - 1)
1769 		segments = MP_BUCKETS - 1;
1770 	if (used > 100000)
1771 		used = 100000;
1772 	if (wasted > 100000)
1773 		wasted = 100000;
1774 	/* store in the appropriate bucket */
1775 	/* don't bother locking. if it's slightly off, so what? */
1776 	mbprof.segments[segments]++;
1777 	mbprof.used[fls(used)]++;
1778 	mbprof.wasted[fls(wasted)]++;
1779 }
1780 
1781 static void
1782 mbprof_textify(void)
1783 {
1784 	int offset;
1785 	char *c;
1786 	uint64_t *p;
1787 
1788 	p = &mbprof.wasted[0];
1789 	c = mbprofbuf;
1790 	offset = snprintf(c, MP_MAXLINE + 10,
1791 	    "wasted:\n"
1792 	    "%ju %ju %ju %ju %ju %ju %ju %ju "
1793 	    "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1794 	    p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1795 	    p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1796 #ifdef BIG_ARRAY
1797 	p = &mbprof.wasted[16];
1798 	c += offset;
1799 	offset = snprintf(c, MP_MAXLINE,
1800 	    "%ju %ju %ju %ju %ju %ju %ju %ju "
1801 	    "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1802 	    p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1803 	    p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1804 #endif
1805 	p = &mbprof.used[0];
1806 	c += offset;
1807 	offset = snprintf(c, MP_MAXLINE + 10,
1808 	    "used:\n"
1809 	    "%ju %ju %ju %ju %ju %ju %ju %ju "
1810 	    "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1811 	    p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1812 	    p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1813 #ifdef BIG_ARRAY
1814 	p = &mbprof.used[16];
1815 	c += offset;
1816 	offset = snprintf(c, MP_MAXLINE,
1817 	    "%ju %ju %ju %ju %ju %ju %ju %ju "
1818 	    "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1819 	    p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1820 	    p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1821 #endif
1822 	p = &mbprof.segments[0];
1823 	c += offset;
1824 	offset = snprintf(c, MP_MAXLINE + 10,
1825 	    "segments:\n"
1826 	    "%ju %ju %ju %ju %ju %ju %ju %ju "
1827 	    "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1828 	    p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1829 	    p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1830 #ifdef BIG_ARRAY
1831 	p = &mbprof.segments[16];
1832 	c += offset;
1833 	offset = snprintf(c, MP_MAXLINE,
1834 	    "%ju %ju %ju %ju %ju %ju %ju %ju "
1835 	    "%ju %ju %ju %ju %ju %ju %ju %jju",
1836 	    p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1837 	    p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1838 #endif
1839 }
1840 
1841 static int
1842 mbprof_handler(SYSCTL_HANDLER_ARGS)
1843 {
1844 	int error;
1845 
1846 	mbprof_textify();
1847 	error = SYSCTL_OUT(req, mbprofbuf, strlen(mbprofbuf) + 1);
1848 	return (error);
1849 }
1850 
1851 static int
1852 mbprof_clr_handler(SYSCTL_HANDLER_ARGS)
1853 {
1854 	int clear, error;
1855 
1856 	clear = 0;
1857 	error = sysctl_handle_int(oidp, &clear, 0, req);
1858 	if (error || !req->newptr)
1859 		return (error);
1860 
1861 	if (clear) {
1862 		bzero(&mbprof, sizeof(mbprof));
1863 	}
1864 
1865 	return (error);
1866 }
1867 
1868 
1869 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofile, CTLTYPE_STRING|CTLFLAG_RD,
1870 	    NULL, 0, mbprof_handler, "A", "mbuf profiling statistics");
1871 
1872 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofileclr, CTLTYPE_INT|CTLFLAG_RW,
1873 	    NULL, 0, mbprof_clr_handler, "I", "clear mbuf profiling statistics");
1874 #endif
1875 
1876