xref: /dflybsd-src/sys/kern/uipc_mbuf.c (revision ac2e3f5effc58aa364c7e5c199f35ebbae7cda81)
1 /*
2  * Copyright (c) 1982, 1986, 1988, 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. All advertising materials mentioning features or use of this software
14  *    must display the following acknowledgement:
15  *	This product includes software developed by the University of
16  *	California, Berkeley and its contributors.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	@(#)uipc_mbuf.c	8.2 (Berkeley) 1/4/94
34  * $FreeBSD: src/sys/kern/uipc_mbuf.c,v 1.51.2.24 2003/04/15 06:59:29 silby Exp $
35  * $DragonFly: src/sys/kern/uipc_mbuf.c,v 1.10 2003/07/26 19:42:11 rob Exp $
36  */
37 
38 #include "opt_param.h"
39 #include "opt_mbuf_stress_test.h"
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/malloc.h>
43 #include <sys/mbuf.h>
44 #include <sys/kernel.h>
45 #include <sys/sysctl.h>
46 #include <sys/domain.h>
47 #include <sys/protosw.h>
48 #include <sys/thread.h>
49 #include <sys/globaldata.h>
50 
51 #include <vm/vm.h>
52 #include <vm/vm_kern.h>
53 #include <vm/vm_extern.h>
54 
55 #ifdef INVARIANTS
56 #include <machine/cpu.h>
57 #endif
58 
59 static void mbinit __P((void *));
60 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbinit, NULL)
61 
62 struct mbuf *mbutl;
63 char	*mclrefcnt;
64 struct mbstat mbstat;
65 u_long	mbtypes[MT_NTYPES];
66 struct mbuf *mmbfree;
67 union mcluster *mclfree;
68 int	max_linkhdr;
69 int	max_protohdr;
70 int	max_hdr;
71 int	max_datalen;
72 int	m_defragpackets;
73 int	m_defragbytes;
74 int	m_defraguseless;
75 int	m_defragfailure;
76 #ifdef MBUF_STRESS_TEST
77 int	m_defragrandomfailures;
78 #endif
79 
80 int	nmbclusters;
81 int	nmbufs;
82 u_int	m_mballoc_wid = 0;
83 u_int	m_clalloc_wid = 0;
84 
85 SYSCTL_DECL(_kern_ipc);
86 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW,
87 	   &max_linkhdr, 0, "");
88 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW,
89 	   &max_protohdr, 0, "");
90 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0, "");
91 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW,
92 	   &max_datalen, 0, "");
93 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW,
94 	   &mbuf_wait, 0, "");
95 SYSCTL_STRUCT(_kern_ipc, KIPC_MBSTAT, mbstat, CTLFLAG_RW, &mbstat, mbstat, "");
96 SYSCTL_OPAQUE(_kern_ipc, OID_AUTO, mbtypes, CTLFLAG_RD, mbtypes,
97 	   sizeof(mbtypes), "LU", "");
98 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RD,
99 	   &nmbclusters, 0, "Maximum number of mbuf clusters available");
100 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RD, &nmbufs, 0,
101 	   "Maximum number of mbufs available");
102 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
103 	   &m_defragpackets, 0, "");
104 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
105 	   &m_defragbytes, 0, "");
106 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
107 	   &m_defraguseless, 0, "");
108 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
109 	   &m_defragfailure, 0, "");
110 #ifdef MBUF_STRESS_TEST
111 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
112 	   &m_defragrandomfailures, 0, "");
113 #endif
114 
115 static void	m_reclaim __P((void));
116 
117 #ifndef NMBCLUSTERS
118 #define NMBCLUSTERS	(512 + maxusers * 16)
119 #endif
120 #ifndef NMBUFS
121 #define NMBUFS		(nmbclusters * 4)
122 #endif
123 
124 /*
125  * Perform sanity checks of tunables declared above.
126  */
127 static void
128 tunable_mbinit(void *dummy)
129 {
130 
131 	/*
132 	 * This has to be done before VM init.
133 	 */
134 	nmbclusters = NMBCLUSTERS;
135 	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
136 	nmbufs = NMBUFS;
137 	TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
138 	/* Sanity checks */
139 	if (nmbufs < nmbclusters * 2)
140 		nmbufs = nmbclusters * 2;
141 
142 	return;
143 }
144 SYSINIT(tunable_mbinit, SI_SUB_TUNABLES, SI_ORDER_ANY, tunable_mbinit, NULL);
145 
146 /* "number of clusters of pages" */
147 #define NCL_INIT	1
148 
149 #define NMB_INIT	16
150 
151 /* ARGSUSED*/
152 static void
153 mbinit(dummy)
154 	void *dummy;
155 {
156 	int s;
157 
158 	mmbfree = NULL; mclfree = NULL;
159 	mbstat.m_msize = MSIZE;
160 	mbstat.m_mclbytes = MCLBYTES;
161 	mbstat.m_minclsize = MINCLSIZE;
162 	mbstat.m_mlen = MLEN;
163 	mbstat.m_mhlen = MHLEN;
164 
165 	s = splimp();
166 	if (m_mballoc(NMB_INIT, M_DONTWAIT) == 0)
167 		goto bad;
168 #if MCLBYTES <= PAGE_SIZE
169 	if (m_clalloc(NCL_INIT, M_DONTWAIT) == 0)
170 		goto bad;
171 #else
172 	/* It's OK to call contigmalloc in this context. */
173 	if (m_clalloc(16, M_WAIT) == 0)
174 		goto bad;
175 #endif
176 	splx(s);
177 	return;
178 bad:
179 	panic("mbinit");
180 }
181 
182 /*
183  * Allocate at least nmb mbufs and place on mbuf free list.
184  * Must be called at splimp.
185  */
186 /* ARGSUSED */
187 int
188 m_mballoc(nmb, how)
189 	int nmb;
190 	int how;
191 {
192 	caddr_t p;
193 	int i;
194 	int nbytes;
195 
196 	/*
197 	 * If we've hit the mbuf limit, stop allocating from mb_map,
198 	 * (or trying to) in order to avoid dipping into the section of
199 	 * mb_map which we've "reserved" for clusters.
200 	 */
201 	if ((nmb + mbstat.m_mbufs) > nmbufs)
202 		return (0);
203 
204 	/*
205 	 * Once we run out of map space, it will be impossible to get
206 	 * any more (nothing is ever freed back to the map)
207 	 * -- however you are not dead as m_reclaim might
208 	 * still be able to free a substantial amount of space.
209 	 *
210 	 * XXX Furthermore, we can also work with "recycled" mbufs (when
211 	 * we're calling with M_WAIT the sleep procedure will be woken
212 	 * up when an mbuf is freed. See m_mballoc_wait()).
213 	 */
214 	if (mb_map_full)
215 		return (0);
216 
217 	nbytes = round_page(nmb * MSIZE);
218 	p = (caddr_t)kmem_malloc(mb_map, nbytes, M_NOWAIT);
219 	if (p == 0 && how == M_WAIT) {
220 		mbstat.m_wait++;
221 		p = (caddr_t)kmem_malloc(mb_map, nbytes, M_WAITOK);
222 	}
223 
224 	/*
225 	 * Either the map is now full, or `how' is M_NOWAIT and there
226 	 * are no pages left.
227 	 */
228 	if (p == NULL)
229 		return (0);
230 
231 	nmb = nbytes / MSIZE;
232 	for (i = 0; i < nmb; i++) {
233 		((struct mbuf *)p)->m_next = mmbfree;
234 		mmbfree = (struct mbuf *)p;
235 		p += MSIZE;
236 	}
237 	mbstat.m_mbufs += nmb;
238 	mbtypes[MT_FREE] += nmb;
239 	return (1);
240 }
241 
242 /*
243  * Once the mb_map has been exhausted and if the call to the allocation macros
244  * (or, in some cases, functions) is with M_WAIT, then it is necessary to rely
245  * solely on reclaimed mbufs. Here we wait for an mbuf to be freed for a
246  * designated (mbuf_wait) time.
247  */
248 struct mbuf *
249 m_mballoc_wait(int caller, int type)
250 {
251 	struct mbuf *p;
252 	int s;
253 
254 	s = splimp();
255 	m_mballoc_wid++;
256 	if ((tsleep(&m_mballoc_wid, 0, "mballc", mbuf_wait)) == EWOULDBLOCK)
257 		m_mballoc_wid--;
258 	splx(s);
259 
260 	/*
261 	 * Now that we (think) that we've got something, we will redo an
262 	 * MGET, but avoid getting into another instance of m_mballoc_wait()
263 	 * XXX: We retry to fetch _even_ if the sleep timed out. This is left
264 	 *      this way, purposely, in the [unlikely] case that an mbuf was
265 	 *      freed but the sleep was not awakened in time.
266 	 */
267 	p = NULL;
268 	switch (caller) {
269 	case MGET_C:
270 		MGET(p, M_DONTWAIT, type);
271 		break;
272 	case MGETHDR_C:
273 		MGETHDR(p, M_DONTWAIT, type);
274 		break;
275 	default:
276 		panic("m_mballoc_wait: invalid caller (%d)", caller);
277 	}
278 
279 	s = splimp();
280 	if (p != NULL) {		/* We waited and got something... */
281 		mbstat.m_wait++;
282 		/* Wake up another if we have more free. */
283 		if (mmbfree != NULL)
284 			MMBWAKEUP();
285 	}
286 	splx(s);
287 	return (p);
288 }
289 
290 #if MCLBYTES > PAGE_SIZE
291 static int i_want_my_mcl;
292 
293 static void
294 kproc_mclalloc(void)
295 {
296 	int status;
297 
298 	while (1) {
299 		tsleep(&i_want_my_mcl, 0, "mclalloc", 0);
300 
301 		for (; i_want_my_mcl; i_want_my_mcl--) {
302 			if (m_clalloc(1, M_WAIT) == 0)
303 				printf("m_clalloc failed even in process context!\n");
304 		}
305 	}
306 }
307 
308 static struct thread *mclallocthread;
309 static struct kproc_desc mclalloc_kp = {
310 	"mclalloc",
311 	kproc_mclalloc,
312 	&mclallocthread
313 };
314 SYSINIT(mclallocthread, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start,
315 	   &mclalloc_kp);
316 #endif
317 
318 /*
319  * Allocate some number of mbuf clusters
320  * and place on cluster free list.
321  * Must be called at splimp.
322  */
323 /* ARGSUSED */
324 int
325 m_clalloc(ncl, how)
326 	int ncl;
327 	int how;
328 {
329 	caddr_t p;
330 	int i;
331 	int npg;
332 
333 	/*
334 	 * If we've hit the mcluster number limit, stop allocating from
335 	 * mb_map, (or trying to) in order to avoid dipping into the section
336 	 * of mb_map which we've "reserved" for mbufs.
337 	 */
338 	if ((ncl + mbstat.m_clusters) > nmbclusters)
339 		goto m_clalloc_fail;
340 
341 	/*
342 	 * Once we run out of map space, it will be impossible
343 	 * to get any more (nothing is ever freed back to the
344 	 * map). From this point on, we solely rely on freed
345 	 * mclusters.
346 	 */
347 	if (mb_map_full)
348 		goto m_clalloc_fail;
349 
350 #if MCLBYTES > PAGE_SIZE
351 	if (how != M_WAIT) {
352 		i_want_my_mcl += ncl;
353 		wakeup(&i_want_my_mcl);
354 		mbstat.m_wait++;
355 		p = 0;
356 	} else {
357 		p = contigmalloc1(MCLBYTES * ncl, M_DEVBUF, M_WAITOK, 0ul,
358 				  ~0ul, PAGE_SIZE, 0, mb_map);
359 	}
360 #else
361 	npg = ncl;
362 	p = (caddr_t)kmem_malloc(mb_map, ctob(npg),
363 				 how != M_WAIT ? M_NOWAIT : M_WAITOK);
364 	ncl = ncl * PAGE_SIZE / MCLBYTES;
365 #endif
366 	/*
367 	 * Either the map is now full, or `how' is M_NOWAIT and there
368 	 * are no pages left.
369 	 */
370 	if (p == NULL) {
371 		static int last_report ; /* when we did that (in ticks) */
372 m_clalloc_fail:
373 		mbstat.m_drops++;
374 		if (ticks < last_report || (ticks - last_report) >= hz) {
375 			last_report = ticks;
376 			printf("All mbuf clusters exhausted, please see tuning(7).\n");
377 		}
378 		return (0);
379 	}
380 
381 	for (i = 0; i < ncl; i++) {
382 		((union mcluster *)p)->mcl_next = mclfree;
383 		mclfree = (union mcluster *)p;
384 		p += MCLBYTES;
385 		mbstat.m_clfree++;
386 	}
387 	mbstat.m_clusters += ncl;
388 	return (1);
389 }
390 
391 /*
392  * Once the mb_map submap has been exhausted and the allocation is called with
393  * M_WAIT, we rely on the mclfree union pointers. If nothing is free, we will
394  * sleep for a designated amount of time (mbuf_wait) or until we're woken up
395  * due to sudden mcluster availability.
396  */
397 caddr_t
398 m_clalloc_wait(void)
399 {
400 	caddr_t p;
401 	int s;
402 
403 #ifdef __i386__
404 	/* If in interrupt context, and INVARIANTS, maintain sanity and die. */
405 	KASSERT(mycpu->gd_intr_nesting_level == 0, ("CLALLOC: CANNOT WAIT IN INTERRUPT"));
406 #endif
407 
408 	/* Sleep until something's available or until we expire. */
409 	m_clalloc_wid++;
410 	if ((tsleep(&m_clalloc_wid, 0, "mclalc", mbuf_wait)) == EWOULDBLOCK)
411 		m_clalloc_wid--;
412 
413 	/*
414 	 * Now that we (think) that we've got something, we will redo and
415 	 * MGET, but avoid getting into another instance of m_clalloc_wait()
416 	 */
417 	p = m_mclalloc(M_DONTWAIT);
418 
419 	s = splimp();
420 	if (p != NULL) {	/* We waited and got something... */
421 		mbstat.m_wait++;
422 		/* Wake up another if we have more free. */
423 		if (mclfree != NULL)
424 			MCLWAKEUP();
425 	}
426 
427 	splx(s);
428 	return (p);
429 }
430 
431 /*
432  * When MGET fails, ask protocols to free space when short of memory,
433  * then re-attempt to allocate an mbuf.
434  */
435 struct mbuf *
436 m_retry(i, t)
437 	int i, t;
438 {
439 	struct mbuf *m;
440 	int ms;
441 
442 	/*
443 	 * Must only do the reclaim if not in an interrupt context.
444 	 */
445 	if (i == M_WAIT) {
446 #ifdef __i386__
447 		KASSERT(mycpu->gd_intr_nesting_level == 0,
448 		    ("MBALLOC: CANNOT WAIT IN INTERRUPT"));
449 #endif
450 		m_reclaim();
451 	}
452 
453 	ms = splimp();
454 	if (mmbfree == NULL)
455 		(void)m_mballoc(1, i);
456 	m = mmbfree;
457 	if (m != NULL) {
458 		mmbfree = m->m_next;
459 		mbtypes[MT_FREE]--;
460 		m->m_type = t;
461 		mbtypes[t]++;
462 		m->m_next = NULL;
463 		m->m_nextpkt = NULL;
464 		m->m_data = m->m_dat;
465 		m->m_flags = 0;
466 		splx(ms);
467 		mbstat.m_wait++;
468 	} else {
469 		static int last_report ; /* when we did that (in ticks) */
470 
471 		splx(ms);
472 		mbstat.m_drops++;
473 		if (ticks < last_report || (ticks - last_report) >= hz) {
474 			last_report = ticks;
475 			printf("All mbufs exhausted, please see tuning(7).\n");
476 		}
477 	}
478 
479 	return (m);
480 }
481 
482 /*
483  * As above; retry an MGETHDR.
484  */
485 struct mbuf *
486 m_retryhdr(i, t)
487 	int i, t;
488 {
489 	struct mbuf *m;
490 	int ms;
491 
492 	/*
493 	 * Must only do the reclaim if not in an interrupt context.
494 	 */
495 	if (i == M_WAIT) {
496 #ifdef __i386__
497 		KASSERT(mycpu->gd_intr_nesting_level == 0,
498 		    ("MBALLOC: CANNOT WAIT IN INTERRUPT"));
499 #endif
500 		m_reclaim();
501 	}
502 
503 	ms = splimp();
504 	if (mmbfree == NULL)
505 		(void)m_mballoc(1, i);
506 	m = mmbfree;
507 	if (m != NULL) {
508 		mmbfree = m->m_next;
509 		mbtypes[MT_FREE]--;
510 		m->m_type = t;
511 		mbtypes[t]++;
512 		m->m_next = NULL;
513 		m->m_nextpkt = NULL;
514 		m->m_data = m->m_pktdat;
515 		m->m_flags = M_PKTHDR;
516 		m->m_pkthdr.rcvif = NULL;
517 		SLIST_INIT(&m->m_pkthdr.tags);
518 		m->m_pkthdr.csum_flags = 0;
519 		splx(ms);
520 		mbstat.m_wait++;
521 	} else {
522 		static int last_report ; /* when we did that (in ticks) */
523 
524 		splx(ms);
525 		mbstat.m_drops++;
526 		if (ticks < last_report || (ticks - last_report) >= hz) {
527 			last_report = ticks;
528 			printf("All mbufs exhausted, please see tuning(7).\n");
529 		}
530 	}
531 
532 	return (m);
533 }
534 
535 static void
536 m_reclaim()
537 {
538 	struct domain *dp;
539 	struct protosw *pr;
540 	int s = splimp();
541 
542 	for (dp = domains; dp; dp = dp->dom_next)
543 		for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
544 			if (pr->pr_drain)
545 				(*pr->pr_drain)();
546 	splx(s);
547 	mbstat.m_drain++;
548 }
549 
550 /*
551  * Space allocation routines.
552  * These are also available as macros
553  * for critical paths.
554  */
555 struct mbuf *
556 m_get(how, type)
557 	int how, type;
558 {
559 	struct mbuf *m;
560 	int ms;
561 
562 	ms = splimp();
563 	if (mmbfree == NULL)
564 		(void)m_mballoc(1, how);
565 	m = mmbfree;
566 	if (m != NULL) {
567 		mmbfree = m->m_next;
568 		mbtypes[MT_FREE]--;
569 		m->m_type = type;
570 		mbtypes[type]++;
571 		m->m_next = NULL;
572 		m->m_nextpkt = NULL;
573 		m->m_data = m->m_dat;
574 		m->m_flags = 0;
575 		splx(ms);
576 	} else {
577 		splx(ms);
578 		m = m_retry(how, type);
579 		if (m == NULL && how == M_WAIT)
580 			m = m_mballoc_wait(MGET_C, type);
581 	}
582 	return (m);
583 }
584 
585 struct mbuf *
586 m_gethdr(how, type)
587 	int how, type;
588 {
589 	struct mbuf *m;
590 	int ms;
591 
592 	ms = splimp();
593 	if (mmbfree == NULL)
594 		(void)m_mballoc(1, how);
595 	m = mmbfree;
596 	if (m != NULL) {
597 		mmbfree = m->m_next;
598 		mbtypes[MT_FREE]--;
599 		m->m_type = type;
600 		mbtypes[type]++;
601 		m->m_next = NULL;
602 		m->m_nextpkt = NULL;
603 		m->m_data = m->m_pktdat;
604 		m->m_flags = M_PKTHDR;
605 		m->m_pkthdr.rcvif = NULL;
606 		SLIST_INIT(&m->m_pkthdr.tags);
607 		m->m_pkthdr.csum_flags = 0;
608 		splx(ms);
609 	} else {
610 		splx(ms);
611 		m = m_retryhdr(how, type);
612 		if (m == NULL && how == M_WAIT)
613 			m = m_mballoc_wait(MGETHDR_C, type);
614 	}
615 	return (m);
616 }
617 
618 struct mbuf *
619 m_getclr(how, type)
620 	int how, type;
621 {
622 	struct mbuf *m;
623 
624 	MGET(m, how, type);
625 	if (m == 0)
626 		return (0);
627 	bzero(mtod(m, caddr_t), MLEN);
628 	return (m);
629 }
630 
631 /*
632  * m_getcl() returns an mbuf with an attached cluster.
633  * Because many network drivers use this kind of buffers a lot, it is
634  * convenient to keep a small pool of free buffers of this kind.
635  * Even a small size such as 10 gives about 10% improvement in the
636  * forwarding rate in a bridge or router.
637  * The size of this free list is controlled by the sysctl variable
638  * mcl_pool_max. The list is populated on m_freem(), and used in
639  * m_getcl() if elements are available.
640  */
641 static struct mbuf *mcl_pool;
642 static int mcl_pool_now;
643 static int mcl_pool_max = 0;
644 
645 SYSCTL_INT(_kern_ipc, OID_AUTO, mcl_pool_max, CTLFLAG_RW, &mcl_pool_max, 0,
646            "Maximum number of mbufs+cluster in free list");
647 SYSCTL_INT(_kern_ipc, OID_AUTO, mcl_pool_now, CTLFLAG_RD, &mcl_pool_now, 0,
648            "Current number of mbufs+cluster in free list");
649 
650 struct mbuf *
651 m_getcl(int how, short type, int flags)
652 {
653 	int s = splimp();
654 	struct mbuf *mp;
655 
656 	if (flags & M_PKTHDR) {
657 		if (type == MT_DATA && mcl_pool) {
658 			mp = mcl_pool;
659 			mcl_pool = mp->m_nextpkt;
660 			mcl_pool_now--;
661 			splx(s);
662 			mp->m_nextpkt = NULL;
663 			mp->m_data = mp->m_ext.ext_buf;
664 			mp->m_flags = M_PKTHDR|M_EXT;
665 			mp->m_pkthdr.rcvif = NULL;
666 			mp->m_pkthdr.csum_flags = 0;
667 			return mp;
668 		} else
669 			MGETHDR(mp, how, type);
670 	} else
671 		MGET(mp, how, type);
672 	if (mp) {
673 		MCLGET(mp, how);
674 		if ( (mp->m_flags & M_EXT) == 0) {
675 			m_free(mp);
676 			mp = NULL;
677 		}
678 	}
679 	splx(s);
680 	return mp;
681 }
682 
683 /*
684  * struct mbuf *
685  * m_getm(m, len, how, type)
686  *
687  * This will allocate len-worth of mbufs and/or mbuf clusters (whatever fits
688  * best) and return a pointer to the top of the allocated chain. If m is
689  * non-null, then we assume that it is a single mbuf or an mbuf chain to
690  * which we want len bytes worth of mbufs and/or clusters attached, and so
691  * if we succeed in allocating it, we will just return a pointer to m.
692  *
693  * If we happen to fail at any point during the allocation, we will free
694  * up everything we have already allocated and return NULL.
695  *
696  */
697 struct mbuf *
698 m_getm(struct mbuf *m, int len, int how, int type)
699 {
700 	struct mbuf *top, *tail, *mp, *mtail = NULL;
701 
702 	KASSERT(len >= 0, ("len is < 0 in m_getm"));
703 
704 	MGET(mp, how, type);
705 	if (mp == NULL)
706 		return (NULL);
707 	else if (len > MINCLSIZE) {
708 		MCLGET(mp, how);
709 		if ((mp->m_flags & M_EXT) == 0) {
710 			m_free(mp);
711 			return (NULL);
712 		}
713 	}
714 	mp->m_len = 0;
715 	len -= M_TRAILINGSPACE(mp);
716 
717 	if (m != NULL)
718 		for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next);
719 	else
720 		m = mp;
721 
722 	top = tail = mp;
723 	while (len > 0) {
724 		MGET(mp, how, type);
725 		if (mp == NULL)
726 			goto failed;
727 
728 		tail->m_next = mp;
729 		tail = mp;
730 		if (len > MINCLSIZE) {
731 			MCLGET(mp, how);
732 			if ((mp->m_flags & M_EXT) == 0)
733 				goto failed;
734 		}
735 
736 		mp->m_len = 0;
737 		len -= M_TRAILINGSPACE(mp);
738 	}
739 
740 	if (mtail != NULL)
741 		mtail->m_next = top;
742 	return (m);
743 
744 failed:
745 	m_freem(top);
746 	return (NULL);
747 }
748 
749 /*
750  * m_mclalloc()	- Allocates an mbuf cluster.
751  */
752 caddr_t
753 m_mclalloc(int how)
754 {
755 	caddr_t mp;
756 	int s;
757 
758 	s = splimp();
759 
760 	if (mclfree == NULL)
761 		m_clalloc(1, how);
762 	mp = (caddr_t)mclfree;
763 	if (mp != NULL) {
764 		mclrefcnt[mtocl(mp)]++;
765 		mbstat.m_clfree--;
766 		mclfree = ((union mcluster *)mp)->mcl_next;
767 		splx(s);
768 		return(mp);
769 	}
770 	splx(s);
771 	if (how == M_WAIT)
772 		return(m_clalloc_wait());
773 	return(NULL);
774 }
775 
776 /*
777  *  m_mclget() - Adds a cluster to a normal mbuf, M_EXT is set on success.
778  */
779 void
780 m_mclget(struct mbuf *m, int how)
781 {
782 	m->m_ext.ext_buf = m_mclalloc(how);
783 	if (m->m_ext.ext_buf != NULL) {
784 		m->m_data = m->m_ext.ext_buf;
785 		m->m_flags |= M_EXT;
786 		m->m_ext.ext_free = NULL;
787 		m->m_ext.ext_ref = NULL;
788 		m->m_ext.ext_size = MCLBYTES;
789 	}
790 }
791 
792 static __inline void
793 _m_mclfree(caddr_t data)
794 {
795 	union mcluster *mp = (union mcluster *)data;
796 
797 	KASSERT(mclrefcnt[mtocl(mp)] > 0, ("freeing free cluster"));
798 	if (--mclrefcnt[mtocl(mp)] == 0) {
799 		mp->mcl_next = mclfree;
800 		mclfree = mp;
801 		mbstat.m_clfree++;
802 		MCLWAKEUP();
803 	}
804 }
805 
806 void
807 m_mclfree(caddr_t mp)
808 {
809 	int s = splimp();
810 	_m_mclfree(mp);
811 	splx(s);
812 }
813 
814 /*
815  * m_free()
816  *
817  * Free a single mbuf and any associated external storage.  The successor,
818  * if any, is returned.
819  *
820  * We do need to check non-first mbuf for m_aux, since some of existing
821  * code does not call M_PREPEND properly.
822  * (example: call to bpf_mtap from drivers)
823  */
824 struct mbuf *
825 m_free(struct mbuf *m)
826 {
827 	int s;
828 	struct mbuf *n;
829 
830 	s = splimp();
831 	KASSERT(m->m_type != MT_FREE, ("freeing free mbuf"));
832 	mbtypes[m->m_type]--;
833 	if ((m->m_flags & M_PKTHDR) != 0)
834 		m_tag_delete_chain(m, NULL);
835 	if (m->m_flags & M_EXT) {
836 		if (m->m_ext.ext_free != NULL) {
837 			m->m_ext.ext_free(m->m_ext.ext_buf, m->m_ext.ext_size);
838 		} else {
839 			_m_mclfree(m->m_ext.ext_buf); /* inlined */
840 		}
841 	}
842 	n = m->m_next;
843 	m->m_type = MT_FREE;
844 	mbtypes[MT_FREE]++;
845 	m->m_next = mmbfree;
846 	mmbfree = m;
847 	MMBWAKEUP();
848 	splx(s);
849 
850 	return (n);
851 }
852 
853 void
854 m_freem(struct mbuf *m)
855 {
856 	int s = splimp();
857 
858 	/*
859 	 * Try to keep a small pool of mbuf+cluster for quick use in
860 	 * device drivers. A good candidate is a M_PKTHDR buffer with
861 	 * only one cluster attached. Other mbufs, or those exceeding
862 	 * the pool size, are just m_free'd in the usual way.
863 	 * The following code makes sure that m_next, m_type,
864 	 * m_pkthdr.aux and m_ext.* are properly initialized.
865 	 * Other fields in the mbuf are initialized in m_getcl()
866 	 * upon allocation.
867 	 */
868         if (mcl_pool_now < mcl_pool_max && m && m->m_next == NULL &&
869             (m->m_flags & (M_PKTHDR|M_EXT)) == (M_PKTHDR|M_EXT) &&
870             m->m_type == MT_DATA && M_EXT_WRITABLE(m) ) {
871 		m_tag_delete_chain(m, NULL);
872                 m->m_nextpkt = mcl_pool;
873                 mcl_pool = m;
874                 mcl_pool_now++;
875         } else {
876 		while (m)
877 			m = m_free(m);
878 	}
879 	splx(s);
880 }
881 
882 /*
883  * Mbuffer utility routines.
884  */
885 
886 /*
887  * Lesser-used path for M_PREPEND:
888  * allocate new mbuf to prepend to chain,
889  * copy junk along.
890  */
891 struct mbuf *
892 m_prepend(m, len, how)
893 	struct mbuf *m;
894 	int len, how;
895 {
896 	struct mbuf *mn;
897 
898 	MGET(mn, how, m->m_type);
899 	if (mn == (struct mbuf *)NULL) {
900 		m_freem(m);
901 		return ((struct mbuf *)NULL);
902 	}
903 	if (m->m_flags & M_PKTHDR)
904 		M_MOVE_PKTHDR(mn, m);
905 	mn->m_next = m;
906 	m = mn;
907 	if (len < MHLEN)
908 		MH_ALIGN(m, len);
909 	m->m_len = len;
910 	return (m);
911 }
912 
913 /*
914  * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
915  * continuing for "len" bytes.  If len is M_COPYALL, copy to end of mbuf.
916  * The wait parameter is a choice of M_WAIT/M_DONTWAIT from caller.
917  * Note that the copy is read-only, because clusters are not copied,
918  * only their reference counts are incremented.
919  */
920 #define MCFail (mbstat.m_mcfail)
921 
922 struct mbuf *
923 m_copym(m, off0, len, wait)
924 	const struct mbuf *m;
925 	int off0, wait;
926 	int len;
927 {
928 	struct mbuf *n, **np;
929 	int off = off0;
930 	struct mbuf *top;
931 	int copyhdr = 0;
932 
933 	KASSERT(off >= 0, ("m_copym, negative off %d", off));
934 	KASSERT(len >= 0, ("m_copym, negative len %d", len));
935 	if (off == 0 && m->m_flags & M_PKTHDR)
936 		copyhdr = 1;
937 	while (off > 0) {
938 		KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
939 		if (off < m->m_len)
940 			break;
941 		off -= m->m_len;
942 		m = m->m_next;
943 	}
944 	np = &top;
945 	top = 0;
946 	while (len > 0) {
947 		if (m == 0) {
948 			KASSERT(len == M_COPYALL,
949 			    ("m_copym, length > size of mbuf chain"));
950 			break;
951 		}
952 		MGET(n, wait, m->m_type);
953 		*np = n;
954 		if (n == 0)
955 			goto nospace;
956 		if (copyhdr) {
957 			if (!m_dup_pkthdr(n, m, wait))
958 				goto nospace;
959 			if (len == M_COPYALL)
960 				n->m_pkthdr.len -= off0;
961 			else
962 				n->m_pkthdr.len = len;
963 			copyhdr = 0;
964 		}
965 		n->m_len = min(len, m->m_len - off);
966 		if (m->m_flags & M_EXT) {
967 			n->m_data = m->m_data + off;
968 			if (m->m_ext.ext_ref == NULL) {
969 				atomic_add_char(
970 				    &mclrefcnt[mtocl(m->m_ext.ext_buf)], 1);
971 			} else {
972 				int s = splimp();
973 
974 				(*m->m_ext.ext_ref)(m->m_ext.ext_buf,
975 				    m->m_ext.ext_size);
976 				splx(s);
977 			}
978 			n->m_ext = m->m_ext;
979 			n->m_flags |= M_EXT;
980 		} else
981 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
982 			    (unsigned)n->m_len);
983 		if (len != M_COPYALL)
984 			len -= n->m_len;
985 		off = 0;
986 		m = m->m_next;
987 		np = &n->m_next;
988 	}
989 	if (top == 0)
990 		MCFail++;
991 	return (top);
992 nospace:
993 	m_freem(top);
994 	MCFail++;
995 	return (0);
996 }
997 
998 /*
999  * Copy an entire packet, including header (which must be present).
1000  * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
1001  * Note that the copy is read-only, because clusters are not copied,
1002  * only their reference counts are incremented.
1003  * Preserve alignment of the first mbuf so if the creator has left
1004  * some room at the beginning (e.g. for inserting protocol headers)
1005  * the copies also have the room available.
1006  */
1007 struct mbuf *
1008 m_copypacket(m, how)
1009 	struct mbuf *m;
1010 	int how;
1011 {
1012 	struct mbuf *top, *n, *o;
1013 
1014 	MGET(n, how, m->m_type);
1015 	top = n;
1016 	if (!n)
1017 		goto nospace;
1018 
1019 	if (!m_dup_pkthdr(n, m, how))
1020 		goto nospace;
1021 	n->m_len = m->m_len;
1022 	if (m->m_flags & M_EXT) {
1023 		n->m_data = m->m_data;
1024 		if (m->m_ext.ext_ref == NULL)
1025 			atomic_add_char(&mclrefcnt[mtocl(m->m_ext.ext_buf)], 1);
1026 		else {
1027 			int s = splimp();
1028 
1029 			(*m->m_ext.ext_ref)(m->m_ext.ext_buf,
1030 			    m->m_ext.ext_size);
1031 			splx(s);
1032 		}
1033 		n->m_ext = m->m_ext;
1034 		n->m_flags |= M_EXT;
1035 	} else {
1036 		n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
1037 		bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1038 	}
1039 
1040 	m = m->m_next;
1041 	while (m) {
1042 		MGET(o, how, m->m_type);
1043 		if (!o)
1044 			goto nospace;
1045 
1046 		n->m_next = o;
1047 		n = n->m_next;
1048 
1049 		n->m_len = m->m_len;
1050 		if (m->m_flags & M_EXT) {
1051 			n->m_data = m->m_data;
1052 			if (m->m_ext.ext_ref == NULL) {
1053 				atomic_add_char(
1054 				    &mclrefcnt[mtocl(m->m_ext.ext_buf)], 1);
1055 			} else {
1056 				int s = splimp();
1057 
1058 				(*m->m_ext.ext_ref)(m->m_ext.ext_buf,
1059 				    m->m_ext.ext_size);
1060 				splx(s);
1061 			}
1062 			n->m_ext = m->m_ext;
1063 			n->m_flags |= M_EXT;
1064 		} else {
1065 			bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1066 		}
1067 
1068 		m = m->m_next;
1069 	}
1070 	return top;
1071 nospace:
1072 	m_freem(top);
1073 	MCFail++;
1074 	return 0;
1075 }
1076 
1077 /*
1078  * Copy data from an mbuf chain starting "off" bytes from the beginning,
1079  * continuing for "len" bytes, into the indicated buffer.
1080  */
1081 void
1082 m_copydata(m, off, len, cp)
1083 	const struct mbuf *m;
1084 	int off;
1085 	int len;
1086 	caddr_t cp;
1087 {
1088 	unsigned count;
1089 
1090 	KASSERT(off >= 0, ("m_copydata, negative off %d", off));
1091 	KASSERT(len >= 0, ("m_copydata, negative len %d", len));
1092 	while (off > 0) {
1093 		KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
1094 		if (off < m->m_len)
1095 			break;
1096 		off -= m->m_len;
1097 		m = m->m_next;
1098 	}
1099 	while (len > 0) {
1100 		KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
1101 		count = min(m->m_len - off, len);
1102 		bcopy(mtod(m, caddr_t) + off, cp, count);
1103 		len -= count;
1104 		cp += count;
1105 		off = 0;
1106 		m = m->m_next;
1107 	}
1108 }
1109 
1110 /*
1111  * Copy a packet header mbuf chain into a completely new chain, including
1112  * copying any mbuf clusters.  Use this instead of m_copypacket() when
1113  * you need a writable copy of an mbuf chain.
1114  */
1115 struct mbuf *
1116 m_dup(m, how)
1117 	struct mbuf *m;
1118 	int how;
1119 {
1120 	struct mbuf **p, *top = NULL;
1121 	int remain, moff, nsize;
1122 
1123 	/* Sanity check */
1124 	if (m == NULL)
1125 		return (0);
1126 	KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __FUNCTION__));
1127 
1128 	/* While there's more data, get a new mbuf, tack it on, and fill it */
1129 	remain = m->m_pkthdr.len;
1130 	moff = 0;
1131 	p = &top;
1132 	while (remain > 0 || top == NULL) {	/* allow m->m_pkthdr.len == 0 */
1133 		struct mbuf *n;
1134 
1135 		/* Get the next new mbuf */
1136 		MGET(n, how, m->m_type);
1137 		if (n == NULL)
1138 			goto nospace;
1139 		if (top == NULL) {		/* first one, must be PKTHDR */
1140 			if (!m_dup_pkthdr(n, m, how))
1141 				goto nospace;
1142 			nsize = MHLEN;
1143 		} else				/* not the first one */
1144 			nsize = MLEN;
1145 		if (remain >= MINCLSIZE) {
1146 			MCLGET(n, how);
1147 			if ((n->m_flags & M_EXT) == 0) {
1148 				(void)m_free(n);
1149 				goto nospace;
1150 			}
1151 			nsize = MCLBYTES;
1152 		}
1153 		n->m_len = 0;
1154 
1155 		/* Link it into the new chain */
1156 		*p = n;
1157 		p = &n->m_next;
1158 
1159 		/* Copy data from original mbuf(s) into new mbuf */
1160 		while (n->m_len < nsize && m != NULL) {
1161 			int chunk = min(nsize - n->m_len, m->m_len - moff);
1162 
1163 			bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1164 			moff += chunk;
1165 			n->m_len += chunk;
1166 			remain -= chunk;
1167 			if (moff == m->m_len) {
1168 				m = m->m_next;
1169 				moff = 0;
1170 			}
1171 		}
1172 
1173 		/* Check correct total mbuf length */
1174 		KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
1175 		    	("%s: bogus m_pkthdr.len", __FUNCTION__));
1176 	}
1177 	return (top);
1178 
1179 nospace:
1180 	m_freem(top);
1181 	MCFail++;
1182 	return (0);
1183 }
1184 
1185 /*
1186  * Concatenate mbuf chain n to m.
1187  * Both chains must be of the same type (e.g. MT_DATA).
1188  * Any m_pkthdr is not updated.
1189  */
1190 void
1191 m_cat(m, n)
1192 	struct mbuf *m, *n;
1193 {
1194 	while (m->m_next)
1195 		m = m->m_next;
1196 	while (n) {
1197 		if (m->m_flags & M_EXT ||
1198 		    m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) {
1199 			/* just join the two chains */
1200 			m->m_next = n;
1201 			return;
1202 		}
1203 		/* splat the data from one into the other */
1204 		bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1205 		    (u_int)n->m_len);
1206 		m->m_len += n->m_len;
1207 		n = m_free(n);
1208 	}
1209 }
1210 
1211 void
1212 m_adj(mp, req_len)
1213 	struct mbuf *mp;
1214 	int req_len;
1215 {
1216 	int len = req_len;
1217 	struct mbuf *m;
1218 	int count;
1219 
1220 	if ((m = mp) == NULL)
1221 		return;
1222 	if (len >= 0) {
1223 		/*
1224 		 * Trim from head.
1225 		 */
1226 		while (m != NULL && len > 0) {
1227 			if (m->m_len <= len) {
1228 				len -= m->m_len;
1229 				m->m_len = 0;
1230 				m = m->m_next;
1231 			} else {
1232 				m->m_len -= len;
1233 				m->m_data += len;
1234 				len = 0;
1235 			}
1236 		}
1237 		m = mp;
1238 		if (mp->m_flags & M_PKTHDR)
1239 			m->m_pkthdr.len -= (req_len - len);
1240 	} else {
1241 		/*
1242 		 * Trim from tail.  Scan the mbuf chain,
1243 		 * calculating its length and finding the last mbuf.
1244 		 * If the adjustment only affects this mbuf, then just
1245 		 * adjust and return.  Otherwise, rescan and truncate
1246 		 * after the remaining size.
1247 		 */
1248 		len = -len;
1249 		count = 0;
1250 		for (;;) {
1251 			count += m->m_len;
1252 			if (m->m_next == (struct mbuf *)0)
1253 				break;
1254 			m = m->m_next;
1255 		}
1256 		if (m->m_len >= len) {
1257 			m->m_len -= len;
1258 			if (mp->m_flags & M_PKTHDR)
1259 				mp->m_pkthdr.len -= len;
1260 			return;
1261 		}
1262 		count -= len;
1263 		if (count < 0)
1264 			count = 0;
1265 		/*
1266 		 * Correct length for chain is "count".
1267 		 * Find the mbuf with last data, adjust its length,
1268 		 * and toss data from remaining mbufs on chain.
1269 		 */
1270 		m = mp;
1271 		if (m->m_flags & M_PKTHDR)
1272 			m->m_pkthdr.len = count;
1273 		for (; m; m = m->m_next) {
1274 			if (m->m_len >= count) {
1275 				m->m_len = count;
1276 				break;
1277 			}
1278 			count -= m->m_len;
1279 		}
1280 		while (m->m_next)
1281 			(m = m->m_next) ->m_len = 0;
1282 	}
1283 }
1284 
1285 /*
1286  * Rearange an mbuf chain so that len bytes are contiguous
1287  * and in the data area of an mbuf (so that mtod and dtom
1288  * will work for a structure of size len).  Returns the resulting
1289  * mbuf chain on success, frees it and returns null on failure.
1290  * If there is room, it will add up to max_protohdr-len extra bytes to the
1291  * contiguous region in an attempt to avoid being called next time.
1292  */
1293 #define MPFail (mbstat.m_mpfail)
1294 
1295 struct mbuf *
1296 m_pullup(n, len)
1297 	struct mbuf *n;
1298 	int len;
1299 {
1300 	struct mbuf *m;
1301 	int count;
1302 	int space;
1303 
1304 	/*
1305 	 * If first mbuf has no cluster, and has room for len bytes
1306 	 * without shifting current data, pullup into it,
1307 	 * otherwise allocate a new mbuf to prepend to the chain.
1308 	 */
1309 	if ((n->m_flags & M_EXT) == 0 &&
1310 	    n->m_data + len < &n->m_dat[MLEN] && n->m_next) {
1311 		if (n->m_len >= len)
1312 			return (n);
1313 		m = n;
1314 		n = n->m_next;
1315 		len -= m->m_len;
1316 	} else {
1317 		if (len > MHLEN)
1318 			goto bad;
1319 		MGET(m, M_DONTWAIT, n->m_type);
1320 		if (m == 0)
1321 			goto bad;
1322 		m->m_len = 0;
1323 		if (n->m_flags & M_PKTHDR)
1324 			M_MOVE_PKTHDR(m, n);
1325 	}
1326 	space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
1327 	do {
1328 		count = min(min(max(len, max_protohdr), space), n->m_len);
1329 		bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1330 		  (unsigned)count);
1331 		len -= count;
1332 		m->m_len += count;
1333 		n->m_len -= count;
1334 		space -= count;
1335 		if (n->m_len)
1336 			n->m_data += count;
1337 		else
1338 			n = m_free(n);
1339 	} while (len > 0 && n);
1340 	if (len > 0) {
1341 		(void) m_free(m);
1342 		goto bad;
1343 	}
1344 	m->m_next = n;
1345 	return (m);
1346 bad:
1347 	m_freem(n);
1348 	MPFail++;
1349 	return (0);
1350 }
1351 
1352 /*
1353  * Partition an mbuf chain in two pieces, returning the tail --
1354  * all but the first len0 bytes.  In case of failure, it returns NULL and
1355  * attempts to restore the chain to its original state.
1356  *
1357  * Note that the resulting mbufs might be read-only, because the new
1358  * mbuf can end up sharing an mbuf cluster with the original mbuf if
1359  * the "breaking point" happens to lie within a cluster mbuf. Use the
1360  * M_WRITABLE() macro to check for this case.
1361  */
1362 struct mbuf *
1363 m_split(m0, len0, wait)
1364 	struct mbuf *m0;
1365 	int len0, wait;
1366 {
1367 	struct mbuf *m, *n;
1368 	unsigned len = len0, remain;
1369 
1370 	for (m = m0; m && len > m->m_len; m = m->m_next)
1371 		len -= m->m_len;
1372 	if (m == 0)
1373 		return (0);
1374 	remain = m->m_len - len;
1375 	if (m0->m_flags & M_PKTHDR) {
1376 		MGETHDR(n, wait, m0->m_type);
1377 		if (n == 0)
1378 			return (0);
1379 		n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1380 		n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1381 		m0->m_pkthdr.len = len0;
1382 		if (m->m_flags & M_EXT)
1383 			goto extpacket;
1384 		if (remain > MHLEN) {
1385 			/* m can't be the lead packet */
1386 			MH_ALIGN(n, 0);
1387 			n->m_next = m_split(m, len, wait);
1388 			if (n->m_next == 0) {
1389 				(void) m_free(n);
1390 				return (0);
1391 			} else {
1392 				n->m_len = 0;
1393 				return (n);
1394 			}
1395 		} else
1396 			MH_ALIGN(n, remain);
1397 	} else if (remain == 0) {
1398 		n = m->m_next;
1399 		m->m_next = 0;
1400 		return (n);
1401 	} else {
1402 		MGET(n, wait, m->m_type);
1403 		if (n == 0)
1404 			return (0);
1405 		M_ALIGN(n, remain);
1406 	}
1407 extpacket:
1408 	if (m->m_flags & M_EXT) {
1409 		n->m_flags |= M_EXT;
1410 		n->m_ext = m->m_ext;
1411 		if (m->m_ext.ext_ref == NULL)
1412 			atomic_add_char(&mclrefcnt[mtocl(m->m_ext.ext_buf)], 1);
1413 		else {
1414 			int s = splimp();
1415 
1416 			(*m->m_ext.ext_ref)(m->m_ext.ext_buf,
1417 			    m->m_ext.ext_size);
1418 			splx(s);
1419 		}
1420 		n->m_data = m->m_data + len;
1421 	} else {
1422 		bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1423 	}
1424 	n->m_len = remain;
1425 	m->m_len = len;
1426 	n->m_next = m->m_next;
1427 	m->m_next = 0;
1428 	return (n);
1429 }
1430 /*
1431  * Routine to copy from device local memory into mbufs.
1432  */
1433 struct mbuf *
1434 m_devget(buf, totlen, off0, ifp, copy)
1435 	char *buf;
1436 	int totlen, off0;
1437 	struct ifnet *ifp;
1438 	void (*copy) __P((char *from, caddr_t to, u_int len));
1439 {
1440 	struct mbuf *m;
1441 	struct mbuf *top = 0, **mp = &top;
1442 	int off = off0, len;
1443 	char *cp;
1444 	char *epkt;
1445 
1446 	cp = buf;
1447 	epkt = cp + totlen;
1448 	if (off) {
1449 		cp += off + 2 * sizeof(u_short);
1450 		totlen -= 2 * sizeof(u_short);
1451 	}
1452 	MGETHDR(m, M_DONTWAIT, MT_DATA);
1453 	if (m == 0)
1454 		return (0);
1455 	m->m_pkthdr.rcvif = ifp;
1456 	m->m_pkthdr.len = totlen;
1457 	m->m_len = MHLEN;
1458 
1459 	while (totlen > 0) {
1460 		if (top) {
1461 			MGET(m, M_DONTWAIT, MT_DATA);
1462 			if (m == 0) {
1463 				m_freem(top);
1464 				return (0);
1465 			}
1466 			m->m_len = MLEN;
1467 		}
1468 		len = min(totlen, epkt - cp);
1469 		if (len >= MINCLSIZE) {
1470 			MCLGET(m, M_DONTWAIT);
1471 			if (m->m_flags & M_EXT)
1472 				m->m_len = len = min(len, MCLBYTES);
1473 			else
1474 				len = m->m_len;
1475 		} else {
1476 			/*
1477 			 * Place initial small packet/header at end of mbuf.
1478 			 */
1479 			if (len < m->m_len) {
1480 				if (top == 0 && len + max_linkhdr <= m->m_len)
1481 					m->m_data += max_linkhdr;
1482 				m->m_len = len;
1483 			} else
1484 				len = m->m_len;
1485 		}
1486 		if (copy)
1487 			copy(cp, mtod(m, caddr_t), (unsigned)len);
1488 		else
1489 			bcopy(cp, mtod(m, caddr_t), (unsigned)len);
1490 		cp += len;
1491 		*mp = m;
1492 		mp = &m->m_next;
1493 		totlen -= len;
1494 		if (cp == epkt)
1495 			cp = buf;
1496 	}
1497 	return (top);
1498 }
1499 
1500 /*
1501  * Copy data from a buffer back into the indicated mbuf chain,
1502  * starting "off" bytes from the beginning, extending the mbuf
1503  * chain if necessary.
1504  */
1505 void
1506 m_copyback(m0, off, len, cp)
1507 	struct	mbuf *m0;
1508 	int off;
1509 	int len;
1510 	caddr_t cp;
1511 {
1512 	int mlen;
1513 	struct mbuf *m = m0, *n;
1514 	int totlen = 0;
1515 
1516 	if (m0 == 0)
1517 		return;
1518 	while (off > (mlen = m->m_len)) {
1519 		off -= mlen;
1520 		totlen += mlen;
1521 		if (m->m_next == 0) {
1522 			n = m_getclr(M_DONTWAIT, m->m_type);
1523 			if (n == 0)
1524 				goto out;
1525 			n->m_len = min(MLEN, len + off);
1526 			m->m_next = n;
1527 		}
1528 		m = m->m_next;
1529 	}
1530 	while (len > 0) {
1531 		mlen = min (m->m_len - off, len);
1532 		bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen);
1533 		cp += mlen;
1534 		len -= mlen;
1535 		mlen += off;
1536 		off = 0;
1537 		totlen += mlen;
1538 		if (len == 0)
1539 			break;
1540 		if (m->m_next == 0) {
1541 			n = m_get(M_DONTWAIT, m->m_type);
1542 			if (n == 0)
1543 				break;
1544 			n->m_len = min(MLEN, len);
1545 			m->m_next = n;
1546 		}
1547 		m = m->m_next;
1548 	}
1549 out:	if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1550 		m->m_pkthdr.len = totlen;
1551 }
1552 
1553 void
1554 m_print(const struct mbuf *m)
1555 {
1556 	int len;
1557 	const struct mbuf *m2;
1558 
1559 	len = m->m_pkthdr.len;
1560 	m2 = m;
1561 	while (len) {
1562 		printf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-");
1563 		len -= m2->m_len;
1564 		m2 = m2->m_next;
1565 	}
1566 	return;
1567 }
1568 
1569 /*
1570  * "Move" mbuf pkthdr from "from" to "to".
1571  * "from" must have M_PKTHDR set, and "to" must be empty.
1572  */
1573 void
1574 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
1575 {
1576 	KASSERT((to->m_flags & M_EXT) == 0, ("m_move_pkthdr: to has cluster"));
1577 
1578 	to->m_flags = from->m_flags & M_COPYFLAGS;
1579 	to->m_data = to->m_pktdat;
1580 	to->m_pkthdr = from->m_pkthdr;		/* especially tags */
1581 	SLIST_INIT(&from->m_pkthdr.tags);	/* purge tags from src */
1582 	from->m_flags &= ~M_PKTHDR;
1583 }
1584 
1585 /*
1586  * Duplicate "from"'s mbuf pkthdr in "to".
1587  * "from" must have M_PKTHDR set, and "to" must be empty.
1588  * In particular, this does a deep copy of the packet tags.
1589  */
1590 int
1591 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
1592 {
1593 	to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT);
1594 	if ((to->m_flags & M_EXT) == 0)
1595 		to->m_data = to->m_pktdat;
1596 	to->m_pkthdr = from->m_pkthdr;
1597 	SLIST_INIT(&to->m_pkthdr.tags);
1598 	return (m_tag_copy_chain(to, from, how));
1599 }
1600 
1601 /*
1602  * Defragment a mbuf chain, returning the shortest possible
1603  * chain of mbufs and clusters.  If allocation fails and
1604  * this cannot be completed, NULL will be returned, but
1605  * the passed in chain will be unchanged.  Upon success,
1606  * the original chain will be freed, and the new chain
1607  * will be returned.
1608  *
1609  * If a non-packet header is passed in, the original
1610  * mbuf (chain?) will be returned unharmed.
1611  */
1612 struct mbuf *
1613 m_defrag(struct mbuf *m0, int how)
1614 {
1615 	struct mbuf	*m_new = NULL, *m_final = NULL;
1616 	int		progress = 0, length;
1617 
1618 	if (!(m0->m_flags & M_PKTHDR))
1619 		return (m0);
1620 
1621 #ifdef MBUF_STRESS_TEST
1622 	if (m_defragrandomfailures) {
1623 		int temp = arc4random() & 0xff;
1624 		if (temp == 0xba)
1625 			goto nospace;
1626 	}
1627 #endif
1628 
1629 	if (m0->m_pkthdr.len > MHLEN)
1630 		m_final = m_getcl(how, MT_DATA, M_PKTHDR);
1631 	else
1632 		m_final = m_gethdr(how, MT_DATA);
1633 
1634 	if (m_final == NULL)
1635 		goto nospace;
1636 
1637 	if (m_dup_pkthdr(m_final, m0, how) == NULL)
1638 		goto nospace;
1639 
1640 	m_new = m_final;
1641 
1642 	while (progress < m0->m_pkthdr.len) {
1643 		length = m0->m_pkthdr.len - progress;
1644 		if (length > MCLBYTES)
1645 			length = MCLBYTES;
1646 
1647 		if (m_new == NULL) {
1648 			if (length > MLEN)
1649 				m_new = m_getcl(how, MT_DATA, 0);
1650 			else
1651 				m_new = m_get(how, MT_DATA);
1652 			if (m_new == NULL)
1653 				goto nospace;
1654 		}
1655 
1656 		m_copydata(m0, progress, length, mtod(m_new, caddr_t));
1657 		progress += length;
1658 		m_new->m_len = length;
1659 		if (m_new != m_final)
1660 			m_cat(m_final, m_new);
1661 		m_new = NULL;
1662 	}
1663 	if (m0->m_next == NULL)
1664 		m_defraguseless++;
1665 	m_freem(m0);
1666 	m0 = m_final;
1667 	m_defragpackets++;
1668 	m_defragbytes += m0->m_pkthdr.len;
1669 	return (m0);
1670 nospace:
1671 	m_defragfailure++;
1672 	if (m_new)
1673 		m_free(m_new);
1674 	if (m_final)
1675 		m_freem(m_final);
1676 	return (NULL);
1677 }
1678