xref: /dflybsd-src/sys/kern/uipc_mbuf.c (revision 67bf99c4e3c62e257027c8f0d3b312f44cfe622f)
1 /*
2  * (MPSAFE)
3  *
4  * Copyright (c) 2004 Jeffrey M. Hsu.  All rights reserved.
5  * Copyright (c) 2004 The DragonFly Project.  All rights reserved.
6  *
7  * This code is derived from software contributed to The DragonFly Project
8  * by Jeffrey M. Hsu.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of The DragonFly Project nor the names of its
19  *    contributors may be used to endorse or promote products derived
20  *    from this software without specific, prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  */
35 
36 /*
37  * Copyright (c) 1982, 1986, 1988, 1991, 1993
38  *	The Regents of the University of California.  All rights reserved.
39  *
40  * Redistribution and use in source and binary forms, with or without
41  * modification, are permitted provided that the following conditions
42  * are met:
43  * 1. Redistributions of source code must retain the above copyright
44  *    notice, this list of conditions and the following disclaimer.
45  * 2. Redistributions in binary form must reproduce the above copyright
46  *    notice, this list of conditions and the following disclaimer in the
47  *    documentation and/or other materials provided with the distribution.
48  * 3. All advertising materials mentioning features or use of this software
49  *    must display the following acknowledgement:
50  *	This product includes software developed by the University of
51  *	California, Berkeley and its contributors.
52  * 4. Neither the name of the University nor the names of its contributors
53  *    may be used to endorse or promote products derived from this software
54  *    without specific prior written permission.
55  *
56  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
57  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
58  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
59  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
60  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
61  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
62  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
63  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
64  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
65  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
66  * SUCH DAMAGE.
67  *
68  * @(#)uipc_mbuf.c	8.2 (Berkeley) 1/4/94
69  * $FreeBSD: src/sys/kern/uipc_mbuf.c,v 1.51.2.24 2003/04/15 06:59:29 silby Exp $
70  * $DragonFly: src/sys/kern/uipc_mbuf.c,v 1.70 2008/11/20 14:21:01 sephe Exp $
71  */
72 
73 #include "opt_param.h"
74 #include "opt_mbuf_stress_test.h"
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/malloc.h>
78 #include <sys/mbuf.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
81 #include <sys/domain.h>
82 #include <sys/objcache.h>
83 #include <sys/tree.h>
84 #include <sys/protosw.h>
85 #include <sys/uio.h>
86 #include <sys/thread.h>
87 #include <sys/globaldata.h>
88 
89 #include <sys/thread2.h>
90 #include <sys/spinlock2.h>
91 
92 #include <machine/atomic.h>
93 #include <machine/limits.h>
94 
95 #include <vm/vm.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_extern.h>
98 
99 #ifdef INVARIANTS
100 #include <machine/cpu.h>
101 #endif
102 
103 /*
104  * mbuf cluster meta-data
105  */
106 struct mbcluster {
107 	int32_t	mcl_refs;
108 	void	*mcl_data;
109 };
110 
111 /*
112  * mbuf tracking for debugging purposes
113  */
114 #ifdef MBUF_DEBUG
115 
116 static MALLOC_DEFINE(M_MTRACK, "mtrack", "mtrack");
117 
118 struct mbctrack;
119 RB_HEAD(mbuf_rb_tree, mbtrack);
120 RB_PROTOTYPE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *);
121 
122 struct mbtrack {
123 	RB_ENTRY(mbtrack) rb_node;
124 	int trackid;
125 	struct mbuf *m;
126 };
127 
128 static int
129 mbtrack_cmp(struct mbtrack *mb1, struct mbtrack *mb2)
130 {
131 	if (mb1->m < mb2->m)
132 		return(-1);
133 	if (mb1->m > mb2->m)
134 		return(1);
135 	return(0);
136 }
137 
138 RB_GENERATE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *, m);
139 
140 struct mbuf_rb_tree	mbuf_track_root;
141 static struct spinlock	mbuf_track_spin = SPINLOCK_INITIALIZER(mbuf_track_spin);
142 
143 static void
144 mbuftrack(struct mbuf *m)
145 {
146 	struct mbtrack *mbt;
147 
148 	mbt = kmalloc(sizeof(*mbt), M_MTRACK, M_INTWAIT|M_ZERO);
149 	spin_lock(&mbuf_track_spin);
150 	mbt->m = m;
151 	if (mbuf_rb_tree_RB_INSERT(&mbuf_track_root, mbt)) {
152 		spin_unlock(&mbuf_track_spin);
153 		panic("mbuftrack: mbuf %p already being tracked\n", m);
154 	}
155 	spin_unlock(&mbuf_track_spin);
156 }
157 
158 static void
159 mbufuntrack(struct mbuf *m)
160 {
161 	struct mbtrack *mbt;
162 
163 	spin_lock(&mbuf_track_spin);
164 	mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m);
165 	if (mbt == NULL) {
166 		spin_unlock(&mbuf_track_spin);
167 		panic("mbufuntrack: mbuf %p was not tracked\n", m);
168 	} else {
169 		mbuf_rb_tree_RB_REMOVE(&mbuf_track_root, mbt);
170 		spin_unlock(&mbuf_track_spin);
171 		kfree(mbt, M_MTRACK);
172 	}
173 }
174 
175 void
176 mbuftrackid(struct mbuf *m, int trackid)
177 {
178 	struct mbtrack *mbt;
179 	struct mbuf *n;
180 
181 	spin_lock(&mbuf_track_spin);
182 	while (m) {
183 		n = m->m_nextpkt;
184 		while (m) {
185 			mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m);
186 			if (mbt == NULL) {
187 				spin_unlock(&mbuf_track_spin);
188 				panic("mbuftrackid: mbuf %p not tracked", m);
189 			}
190 			mbt->trackid = trackid;
191 			m = m->m_next;
192 		}
193 		m = n;
194 	}
195 	spin_unlock(&mbuf_track_spin);
196 }
197 
198 static int
199 mbuftrack_callback(struct mbtrack *mbt, void *arg)
200 {
201 	struct sysctl_req *req = arg;
202 	char buf[64];
203 	int error;
204 
205 	ksnprintf(buf, sizeof(buf), "mbuf %p track %d\n", mbt->m, mbt->trackid);
206 
207 	spin_unlock(&mbuf_track_spin);
208 	error = SYSCTL_OUT(req, buf, strlen(buf));
209 	spin_lock(&mbuf_track_spin);
210 	if (error)
211 		return(-error);
212 	return(0);
213 }
214 
215 static int
216 mbuftrack_show(SYSCTL_HANDLER_ARGS)
217 {
218 	int error;
219 
220 	spin_lock(&mbuf_track_spin);
221 	error = mbuf_rb_tree_RB_SCAN(&mbuf_track_root, NULL,
222 				     mbuftrack_callback, req);
223 	spin_unlock(&mbuf_track_spin);
224 	return (-error);
225 }
226 SYSCTL_PROC(_kern_ipc, OID_AUTO, showmbufs, CTLFLAG_RD|CTLTYPE_STRING,
227 	    0, 0, mbuftrack_show, "A", "Show all in-use mbufs");
228 
229 #else
230 
231 #define mbuftrack(m)
232 #define mbufuntrack(m)
233 
234 #endif
235 
236 static void mbinit(void *);
237 SYSINIT(mbuf, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, mbinit, NULL)
238 
239 static u_long	mbtypes[SMP_MAXCPU][MT_NTYPES];
240 
241 static struct mbstat mbstat[SMP_MAXCPU];
242 int	max_linkhdr;
243 int	max_protohdr;
244 int	max_hdr;
245 int	max_datalen;
246 int	m_defragpackets;
247 int	m_defragbytes;
248 int	m_defraguseless;
249 int	m_defragfailure;
250 #ifdef MBUF_STRESS_TEST
251 int	m_defragrandomfailures;
252 #endif
253 
254 struct objcache *mbuf_cache, *mbufphdr_cache;
255 struct objcache *mclmeta_cache;
256 struct objcache *mbufcluster_cache, *mbufphdrcluster_cache;
257 
258 int	nmbclusters;
259 int	nmbufs;
260 
261 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW,
262 	   &max_linkhdr, 0, "");
263 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW,
264 	   &max_protohdr, 0, "");
265 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0, "");
266 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW,
267 	   &max_datalen, 0, "");
268 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW,
269 	   &mbuf_wait, 0, "");
270 static int do_mbstat(SYSCTL_HANDLER_ARGS);
271 
272 SYSCTL_PROC(_kern_ipc, KIPC_MBSTAT, mbstat, CTLTYPE_STRUCT|CTLFLAG_RD,
273 	0, 0, do_mbstat, "S,mbstat", "");
274 
275 static int do_mbtypes(SYSCTL_HANDLER_ARGS);
276 
277 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbtypes, CTLTYPE_ULONG|CTLFLAG_RD,
278 	0, 0, do_mbtypes, "LU", "");
279 
280 static int
281 do_mbstat(SYSCTL_HANDLER_ARGS)
282 {
283 	struct mbstat mbstat_total;
284 	struct mbstat *mbstat_totalp;
285 	int i;
286 
287 	bzero(&mbstat_total, sizeof(mbstat_total));
288 	mbstat_totalp = &mbstat_total;
289 
290 	for (i = 0; i < ncpus; i++)
291 	{
292 		mbstat_total.m_mbufs += mbstat[i].m_mbufs;
293 		mbstat_total.m_clusters += mbstat[i].m_clusters;
294 		mbstat_total.m_spare += mbstat[i].m_spare;
295 		mbstat_total.m_clfree += mbstat[i].m_clfree;
296 		mbstat_total.m_drops += mbstat[i].m_drops;
297 		mbstat_total.m_wait += mbstat[i].m_wait;
298 		mbstat_total.m_drain += mbstat[i].m_drain;
299 		mbstat_total.m_mcfail += mbstat[i].m_mcfail;
300 		mbstat_total.m_mpfail += mbstat[i].m_mpfail;
301 
302 	}
303 	/*
304 	 * The following fields are not cumulative fields so just
305 	 * get their values once.
306 	 */
307 	mbstat_total.m_msize = mbstat[0].m_msize;
308 	mbstat_total.m_mclbytes = mbstat[0].m_mclbytes;
309 	mbstat_total.m_minclsize = mbstat[0].m_minclsize;
310 	mbstat_total.m_mlen = mbstat[0].m_mlen;
311 	mbstat_total.m_mhlen = mbstat[0].m_mhlen;
312 
313 	return(sysctl_handle_opaque(oidp, mbstat_totalp, sizeof(mbstat_total), req));
314 }
315 
316 static int
317 do_mbtypes(SYSCTL_HANDLER_ARGS)
318 {
319 	u_long totals[MT_NTYPES];
320 	int i, j;
321 
322 	for (i = 0; i < MT_NTYPES; i++)
323 		totals[i] = 0;
324 
325 	for (i = 0; i < ncpus; i++)
326 	{
327 		for (j = 0; j < MT_NTYPES; j++)
328 			totals[j] += mbtypes[i][j];
329 	}
330 
331 	return(sysctl_handle_opaque(oidp, totals, sizeof(totals), req));
332 }
333 
334 /*
335  * These are read-only because we do not currently have any code
336  * to adjust the objcache limits after the fact.  The variables
337  * may only be set as boot-time tunables.
338  */
339 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RD,
340 	   &nmbclusters, 0, "Maximum number of mbuf clusters available");
341 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RD, &nmbufs, 0,
342 	   "Maximum number of mbufs available");
343 
344 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
345 	   &m_defragpackets, 0, "");
346 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
347 	   &m_defragbytes, 0, "");
348 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
349 	   &m_defraguseless, 0, "");
350 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
351 	   &m_defragfailure, 0, "");
352 #ifdef MBUF_STRESS_TEST
353 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
354 	   &m_defragrandomfailures, 0, "");
355 #endif
356 
357 static MALLOC_DEFINE(M_MBUF, "mbuf", "mbuf");
358 static MALLOC_DEFINE(M_MBUFCL, "mbufcl", "mbufcl");
359 static MALLOC_DEFINE(M_MCLMETA, "mclmeta", "mclmeta");
360 
361 static void m_reclaim (void);
362 static void m_mclref(void *arg);
363 static void m_mclfree(void *arg);
364 
365 #ifndef NMBCLUSTERS
366 #define NMBCLUSTERS	(512 + maxusers * 16)
367 #endif
368 #ifndef NMBUFS
369 #define NMBUFS		(nmbclusters * 2)
370 #endif
371 
372 /*
373  * Perform sanity checks of tunables declared above.
374  */
375 static void
376 tunable_mbinit(void *dummy)
377 {
378 	/*
379 	 * This has to be done before VM init.
380 	 */
381 	nmbclusters = NMBCLUSTERS;
382 	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
383 	nmbufs = NMBUFS;
384 	TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
385 	/* Sanity checks */
386 	if (nmbufs < nmbclusters * 2)
387 		nmbufs = nmbclusters * 2;
388 }
389 SYSINIT(tunable_mbinit, SI_BOOT1_TUNABLES, SI_ORDER_ANY,
390 	tunable_mbinit, NULL);
391 
392 /* "number of clusters of pages" */
393 #define NCL_INIT	1
394 
395 #define NMB_INIT	16
396 
397 /*
398  * The mbuf object cache only guarantees that m_next and m_nextpkt are
399  * NULL and that m_data points to the beginning of the data area.  In
400  * particular, m_len and m_pkthdr.len are uninitialized.  It is the
401  * responsibility of the caller to initialize those fields before use.
402  */
403 
404 static boolean_t __inline
405 mbuf_ctor(void *obj, void *private, int ocflags)
406 {
407 	struct mbuf *m = obj;
408 
409 	m->m_next = NULL;
410 	m->m_nextpkt = NULL;
411 	m->m_data = m->m_dat;
412 	m->m_flags = 0;
413 
414 	return (TRUE);
415 }
416 
417 /*
418  * Initialize the mbuf and the packet header fields.
419  */
420 static boolean_t
421 mbufphdr_ctor(void *obj, void *private, int ocflags)
422 {
423 	struct mbuf *m = obj;
424 
425 	m->m_next = NULL;
426 	m->m_nextpkt = NULL;
427 	m->m_data = m->m_pktdat;
428 	m->m_flags = M_PKTHDR | M_PHCACHE;
429 
430 	m->m_pkthdr.rcvif = NULL;	/* eliminate XXX JH */
431 	SLIST_INIT(&m->m_pkthdr.tags);
432 	m->m_pkthdr.csum_flags = 0;	/* eliminate XXX JH */
433 	m->m_pkthdr.fw_flags = 0;	/* eliminate XXX JH */
434 
435 	return (TRUE);
436 }
437 
438 /*
439  * A mbcluster object consists of 2K (MCLBYTES) cluster and a refcount.
440  */
441 static boolean_t
442 mclmeta_ctor(void *obj, void *private, int ocflags)
443 {
444 	struct mbcluster *cl = obj;
445 	void *buf;
446 
447 	if (ocflags & M_NOWAIT)
448 		buf = kmalloc(MCLBYTES, M_MBUFCL, M_NOWAIT | M_ZERO);
449 	else
450 		buf = kmalloc(MCLBYTES, M_MBUFCL, M_INTWAIT | M_ZERO);
451 	if (buf == NULL)
452 		return (FALSE);
453 	cl->mcl_refs = 0;
454 	cl->mcl_data = buf;
455 	return (TRUE);
456 }
457 
458 static void
459 mclmeta_dtor(void *obj, void *private)
460 {
461 	struct mbcluster *mcl = obj;
462 
463 	KKASSERT(mcl->mcl_refs == 0);
464 	kfree(mcl->mcl_data, M_MBUFCL);
465 }
466 
467 static void
468 linkcluster(struct mbuf *m, struct mbcluster *cl)
469 {
470 	/*
471 	 * Add the cluster to the mbuf.  The caller will detect that the
472 	 * mbuf now has an attached cluster.
473 	 */
474 	m->m_ext.ext_arg = cl;
475 	m->m_ext.ext_buf = cl->mcl_data;
476 	m->m_ext.ext_ref = m_mclref;
477 	m->m_ext.ext_free = m_mclfree;
478 	m->m_ext.ext_size = MCLBYTES;
479 	atomic_add_int(&cl->mcl_refs, 1);
480 
481 	m->m_data = m->m_ext.ext_buf;
482 	m->m_flags |= M_EXT | M_EXT_CLUSTER;
483 }
484 
485 static boolean_t
486 mbufphdrcluster_ctor(void *obj, void *private, int ocflags)
487 {
488 	struct mbuf *m = obj;
489 	struct mbcluster *cl;
490 
491 	mbufphdr_ctor(obj, private, ocflags);
492 	cl = objcache_get(mclmeta_cache, ocflags);
493 	if (cl == NULL) {
494 		++mbstat[mycpu->gd_cpuid].m_drops;
495 		return (FALSE);
496 	}
497 	m->m_flags |= M_CLCACHE;
498 	linkcluster(m, cl);
499 	return (TRUE);
500 }
501 
502 static boolean_t
503 mbufcluster_ctor(void *obj, void *private, int ocflags)
504 {
505 	struct mbuf *m = obj;
506 	struct mbcluster *cl;
507 
508 	mbuf_ctor(obj, private, ocflags);
509 	cl = objcache_get(mclmeta_cache, ocflags);
510 	if (cl == NULL) {
511 		++mbstat[mycpu->gd_cpuid].m_drops;
512 		return (FALSE);
513 	}
514 	m->m_flags |= M_CLCACHE;
515 	linkcluster(m, cl);
516 	return (TRUE);
517 }
518 
519 /*
520  * Used for both the cluster and cluster PHDR caches.
521  *
522  * The mbuf may have lost its cluster due to sharing, deal
523  * with the situation by checking M_EXT.
524  */
525 static void
526 mbufcluster_dtor(void *obj, void *private)
527 {
528 	struct mbuf *m = obj;
529 	struct mbcluster *mcl;
530 
531 	if (m->m_flags & M_EXT) {
532 		KKASSERT((m->m_flags & M_EXT_CLUSTER) != 0);
533 		mcl = m->m_ext.ext_arg;
534 		KKASSERT(mcl->mcl_refs == 1);
535 		mcl->mcl_refs = 0;
536 		objcache_put(mclmeta_cache, mcl);
537 	}
538 }
539 
540 struct objcache_malloc_args mbuf_malloc_args = { MSIZE, M_MBUF };
541 struct objcache_malloc_args mclmeta_malloc_args =
542 	{ sizeof(struct mbcluster), M_MCLMETA };
543 
544 /* ARGSUSED*/
545 static void
546 mbinit(void *dummy)
547 {
548 	int mb_limit, cl_limit;
549 	int limit;
550 	int i;
551 
552 	/*
553 	 * Initialize statistics
554 	 */
555 	for (i = 0; i < ncpus; i++) {
556 		atomic_set_long_nonlocked(&mbstat[i].m_msize, MSIZE);
557 		atomic_set_long_nonlocked(&mbstat[i].m_mclbytes, MCLBYTES);
558 		atomic_set_long_nonlocked(&mbstat[i].m_minclsize, MINCLSIZE);
559 		atomic_set_long_nonlocked(&mbstat[i].m_mlen, MLEN);
560 		atomic_set_long_nonlocked(&mbstat[i].m_mhlen, MHLEN);
561 	}
562 
563 	/*
564 	 * Create objtect caches and save cluster limits, which will
565 	 * be used to adjust backing kmalloc pools' limit later.
566 	 */
567 
568 	mb_limit = cl_limit = 0;
569 
570 	limit = nmbufs;
571 	mbuf_cache = objcache_create("mbuf", &limit, 0,
572 	    mbuf_ctor, NULL, NULL,
573 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
574 	mb_limit += limit;
575 
576 	limit = nmbufs;
577 	mbufphdr_cache = objcache_create("mbuf pkt hdr", &limit, 64,
578 	    mbufphdr_ctor, NULL, NULL,
579 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
580 	mb_limit += limit;
581 
582 	cl_limit = nmbclusters;
583 	mclmeta_cache = objcache_create("cluster mbuf", &cl_limit, 0,
584 	    mclmeta_ctor, mclmeta_dtor, NULL,
585 	    objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args);
586 
587 	limit = nmbclusters;
588 	mbufcluster_cache = objcache_create("mbuf + cluster", &limit, 0,
589 	    mbufcluster_ctor, mbufcluster_dtor, NULL,
590 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
591 	mb_limit += limit;
592 
593 	limit = nmbclusters;
594 	mbufphdrcluster_cache = objcache_create("mbuf pkt hdr + cluster",
595 	    &limit, 64, mbufphdrcluster_ctor, mbufcluster_dtor, NULL,
596 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
597 	mb_limit += limit;
598 
599 	/*
600 	 * Adjust backing kmalloc pools' limit
601 	 *
602 	 * NOTE: We raise the limit by another 1/8 to take the effect
603 	 * of loosememuse into account.
604 	 */
605 	cl_limit += cl_limit / 8;
606 	kmalloc_raise_limit(mclmeta_malloc_args.mtype,
607 			    mclmeta_malloc_args.objsize * cl_limit);
608 	kmalloc_raise_limit(M_MBUFCL, MCLBYTES * cl_limit);
609 
610 	mb_limit += mb_limit / 8;
611 	kmalloc_raise_limit(mbuf_malloc_args.mtype,
612 			    mbuf_malloc_args.objsize * mb_limit);
613 }
614 
615 /*
616  * Return the number of references to this mbuf's data.  0 is returned
617  * if the mbuf is not M_EXT, a reference count is returned if it is
618  * M_EXT | M_EXT_CLUSTER, and 99 is returned if it is a special M_EXT.
619  */
620 int
621 m_sharecount(struct mbuf *m)
622 {
623 	switch (m->m_flags & (M_EXT | M_EXT_CLUSTER)) {
624 	case 0:
625 		return (0);
626 	case M_EXT:
627 		return (99);
628 	case M_EXT | M_EXT_CLUSTER:
629 		return (((struct mbcluster *)m->m_ext.ext_arg)->mcl_refs);
630 	}
631 	/* NOTREACHED */
632 	return (0);		/* to shut up compiler */
633 }
634 
635 /*
636  * change mbuf to new type
637  */
638 void
639 m_chtype(struct mbuf *m, int type)
640 {
641 	struct globaldata *gd = mycpu;
642 
643 	atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1);
644 	atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1);
645 	atomic_set_short_nonlocked(&m->m_type, type);
646 }
647 
648 static void
649 m_reclaim(void)
650 {
651 	struct domain *dp;
652 	struct protosw *pr;
653 
654 	kprintf("Debug: m_reclaim() called\n");
655 
656 	SLIST_FOREACH(dp, &domains, dom_next) {
657 		for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) {
658 			if (pr->pr_drain)
659 				(*pr->pr_drain)();
660 		}
661 	}
662 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_drain, 1);
663 }
664 
665 static void __inline
666 updatestats(struct mbuf *m, int type)
667 {
668 	struct globaldata *gd = mycpu;
669 
670 	m->m_type = type;
671 	mbuftrack(m);
672 	KKASSERT(m->m_next == NULL);
673 	KKASSERT(m->m_nextpkt == NULL);
674 
675 	atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1);
676 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1);
677 
678 }
679 
680 /*
681  * Allocate an mbuf.
682  */
683 struct mbuf *
684 m_get(int how, int type)
685 {
686 	struct mbuf *m;
687 	int ntries = 0;
688 	int ocf = MBTOM(how);
689 
690 retryonce:
691 
692 	m = objcache_get(mbuf_cache, ocf);
693 
694 	if (m == NULL) {
695 		if ((how & MB_TRYWAIT) && ntries++ == 0) {
696 			struct objcache *reclaimlist[] = {
697 				mbufphdr_cache,
698 				mbufcluster_cache,
699 				mbufphdrcluster_cache
700 			};
701 			const int nreclaims = __arysize(reclaimlist);
702 
703 			if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
704 				m_reclaim();
705 			goto retryonce;
706 		}
707 		++mbstat[mycpu->gd_cpuid].m_drops;
708 		return (NULL);
709 	}
710 
711 	updatestats(m, type);
712 	return (m);
713 }
714 
715 struct mbuf *
716 m_gethdr(int how, int type)
717 {
718 	struct mbuf *m;
719 	int ocf = MBTOM(how);
720 	int ntries = 0;
721 
722 retryonce:
723 
724 	m = objcache_get(mbufphdr_cache, ocf);
725 
726 	if (m == NULL) {
727 		if ((how & MB_TRYWAIT) && ntries++ == 0) {
728 			struct objcache *reclaimlist[] = {
729 				mbuf_cache,
730 				mbufcluster_cache, mbufphdrcluster_cache
731 			};
732 			const int nreclaims = __arysize(reclaimlist);
733 
734 			if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
735 				m_reclaim();
736 			goto retryonce;
737 		}
738 		++mbstat[mycpu->gd_cpuid].m_drops;
739 		return (NULL);
740 	}
741 
742 	updatestats(m, type);
743 	return (m);
744 }
745 
746 /*
747  * Get a mbuf (not a mbuf cluster!) and zero it.
748  * Deprecated.
749  */
750 struct mbuf *
751 m_getclr(int how, int type)
752 {
753 	struct mbuf *m;
754 
755 	m = m_get(how, type);
756 	if (m != NULL)
757 		bzero(m->m_data, MLEN);
758 	return (m);
759 }
760 
761 /*
762  * Returns an mbuf with an attached cluster.
763  * Because many network drivers use this kind of buffers a lot, it is
764  * convenient to keep a small pool of free buffers of this kind.
765  * Even a small size such as 10 gives about 10% improvement in the
766  * forwarding rate in a bridge or router.
767  */
768 struct mbuf *
769 m_getcl(int how, short type, int flags)
770 {
771 	struct mbuf *m;
772 	int ocflags = MBTOM(how);
773 	int ntries = 0;
774 
775 retryonce:
776 
777 	if (flags & M_PKTHDR)
778 		m = objcache_get(mbufphdrcluster_cache, ocflags);
779 	else
780 		m = objcache_get(mbufcluster_cache, ocflags);
781 
782 	if (m == NULL) {
783 		if ((how & MB_TRYWAIT) && ntries++ == 0) {
784 			struct objcache *reclaimlist[1];
785 
786 			if (flags & M_PKTHDR)
787 				reclaimlist[0] = mbufcluster_cache;
788 			else
789 				reclaimlist[0] = mbufphdrcluster_cache;
790 			if (!objcache_reclaimlist(reclaimlist, 1, ocflags))
791 				m_reclaim();
792 			goto retryonce;
793 		}
794 		++mbstat[mycpu->gd_cpuid].m_drops;
795 		return (NULL);
796 	}
797 
798 	m->m_type = type;
799 
800 	mbuftrack(m);
801 
802 	atomic_add_long_nonlocked(&mbtypes[mycpu->gd_cpuid][type], 1);
803 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
804 	return (m);
805 }
806 
807 /*
808  * Allocate chain of requested length.
809  */
810 struct mbuf *
811 m_getc(int len, int how, int type)
812 {
813 	struct mbuf *n, *nfirst = NULL, **ntail = &nfirst;
814 	int nsize;
815 
816 	while (len > 0) {
817 		n = m_getl(len, how, type, 0, &nsize);
818 		if (n == NULL)
819 			goto failed;
820 		n->m_len = 0;
821 		*ntail = n;
822 		ntail = &n->m_next;
823 		len -= nsize;
824 	}
825 	return (nfirst);
826 
827 failed:
828 	m_freem(nfirst);
829 	return (NULL);
830 }
831 
832 /*
833  * Allocate len-worth of mbufs and/or mbuf clusters (whatever fits best)
834  * and return a pointer to the head of the allocated chain. If m0 is
835  * non-null, then we assume that it is a single mbuf or an mbuf chain to
836  * which we want len bytes worth of mbufs and/or clusters attached, and so
837  * if we succeed in allocating it, we will just return a pointer to m0.
838  *
839  * If we happen to fail at any point during the allocation, we will free
840  * up everything we have already allocated and return NULL.
841  *
842  * Deprecated.  Use m_getc() and m_cat() instead.
843  */
844 struct mbuf *
845 m_getm(struct mbuf *m0, int len, int type, int how)
846 {
847 	struct mbuf *nfirst;
848 
849 	nfirst = m_getc(len, how, type);
850 
851 	if (m0 != NULL) {
852 		m_last(m0)->m_next = nfirst;
853 		return (m0);
854 	}
855 
856 	return (nfirst);
857 }
858 
859 /*
860  * Adds a cluster to a normal mbuf, M_EXT is set on success.
861  * Deprecated.  Use m_getcl() instead.
862  */
863 void
864 m_mclget(struct mbuf *m, int how)
865 {
866 	struct mbcluster *mcl;
867 
868 	KKASSERT((m->m_flags & M_EXT) == 0);
869 	mcl = objcache_get(mclmeta_cache, MBTOM(how));
870 	if (mcl != NULL) {
871 		linkcluster(m, mcl);
872 		atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters,
873 					  1);
874 	} else {
875 		++mbstat[mycpu->gd_cpuid].m_drops;
876 	}
877 }
878 
879 /*
880  * Updates to mbcluster must be MPSAFE.  Only an entity which already has
881  * a reference to the cluster can ref it, so we are in no danger of
882  * racing an add with a subtract.  But the operation must still be atomic
883  * since multiple entities may have a reference on the cluster.
884  *
885  * m_mclfree() is almost the same but it must contend with two entities
886  * freeing the cluster at the same time.
887  */
888 static void
889 m_mclref(void *arg)
890 {
891 	struct mbcluster *mcl = arg;
892 
893 	atomic_add_int(&mcl->mcl_refs, 1);
894 }
895 
896 /*
897  * When dereferencing a cluster we have to deal with a N->0 race, where
898  * N entities free their references simultaniously.  To do this we use
899  * atomic_fetchadd_int().
900  */
901 static void
902 m_mclfree(void *arg)
903 {
904 	struct mbcluster *mcl = arg;
905 
906 	if (atomic_fetchadd_int(&mcl->mcl_refs, -1) == 1)
907 		objcache_put(mclmeta_cache, mcl);
908 }
909 
910 /*
911  * Free a single mbuf and any associated external storage.  The successor,
912  * if any, is returned.
913  *
914  * We do need to check non-first mbuf for m_aux, since some of existing
915  * code does not call M_PREPEND properly.
916  * (example: call to bpf_mtap from drivers)
917  */
918 struct mbuf *
919 m_free(struct mbuf *m)
920 {
921 	struct mbuf *n;
922 	struct globaldata *gd = mycpu;
923 
924 	KASSERT(m->m_type != MT_FREE, ("freeing free mbuf %p", m));
925 	KASSERT(M_TRAILINGSPACE(m) >= 0, ("overflowed mbuf %p", m));
926 	atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1);
927 
928 	n = m->m_next;
929 
930 	/*
931 	 * Make sure the mbuf is in constructed state before returning it
932 	 * to the objcache.
933 	 */
934 	m->m_next = NULL;
935 	mbufuntrack(m);
936 #ifdef notyet
937 	KKASSERT(m->m_nextpkt == NULL);
938 #else
939 	if (m->m_nextpkt != NULL) {
940 		static int afewtimes = 10;
941 
942 		if (afewtimes-- > 0) {
943 			kprintf("mfree: m->m_nextpkt != NULL\n");
944 			print_backtrace(-1);
945 		}
946 		m->m_nextpkt = NULL;
947 	}
948 #endif
949 	if (m->m_flags & M_PKTHDR) {
950 		m_tag_delete_chain(m);		/* eliminate XXX JH */
951 	}
952 
953 	m->m_flags &= (M_EXT | M_EXT_CLUSTER | M_CLCACHE | M_PHCACHE);
954 
955 	/*
956 	 * Clean the M_PKTHDR state so we can return the mbuf to its original
957 	 * cache.  This is based on the PHCACHE flag which tells us whether
958 	 * the mbuf was originally allocated out of a packet-header cache
959 	 * or a non-packet-header cache.
960 	 */
961 	if (m->m_flags & M_PHCACHE) {
962 		m->m_flags |= M_PKTHDR;
963 		m->m_pkthdr.rcvif = NULL;	/* eliminate XXX JH */
964 		m->m_pkthdr.csum_flags = 0;	/* eliminate XXX JH */
965 		m->m_pkthdr.fw_flags = 0;	/* eliminate XXX JH */
966 		SLIST_INIT(&m->m_pkthdr.tags);
967 	}
968 
969 	/*
970 	 * Handle remaining flags combinations.  M_CLCACHE tells us whether
971 	 * the mbuf was originally allocated from a cluster cache or not,
972 	 * and is totally separate from whether the mbuf is currently
973 	 * associated with a cluster.
974 	 */
975 	switch(m->m_flags & (M_CLCACHE | M_EXT | M_EXT_CLUSTER)) {
976 	case M_CLCACHE | M_EXT | M_EXT_CLUSTER:
977 		/*
978 		 * mbuf+cluster cache case.  The mbuf was allocated from the
979 		 * combined mbuf_cluster cache and can be returned to the
980 		 * cache if the cluster hasn't been shared.
981 		 */
982 		if (m_sharecount(m) == 1) {
983 			/*
984 			 * The cluster has not been shared, we can just
985 			 * reset the data pointer and return the mbuf
986 			 * to the cluster cache.  Note that the reference
987 			 * count is left intact (it is still associated with
988 			 * an mbuf).
989 			 */
990 			m->m_data = m->m_ext.ext_buf;
991 			if (m->m_flags & M_PHCACHE)
992 				objcache_put(mbufphdrcluster_cache, m);
993 			else
994 				objcache_put(mbufcluster_cache, m);
995 			atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
996 		} else {
997 			/*
998 			 * Hell.  Someone else has a ref on this cluster,
999 			 * we have to disconnect it which means we can't
1000 			 * put it back into the mbufcluster_cache, we
1001 			 * have to destroy the mbuf.
1002 			 *
1003 			 * Other mbuf references to the cluster will typically
1004 			 * be M_EXT | M_EXT_CLUSTER but without M_CLCACHE.
1005 			 *
1006 			 * XXX we could try to connect another cluster to
1007 			 * it.
1008 			 */
1009 			m->m_ext.ext_free(m->m_ext.ext_arg);
1010 			m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
1011 			if (m->m_flags & M_PHCACHE)
1012 				objcache_dtor(mbufphdrcluster_cache, m);
1013 			else
1014 				objcache_dtor(mbufcluster_cache, m);
1015 		}
1016 		break;
1017 	case M_EXT | M_EXT_CLUSTER:
1018 		/*
1019 		 * Normal cluster associated with an mbuf that was allocated
1020 		 * from the normal mbuf pool rather then the cluster pool.
1021 		 * The cluster has to be independantly disassociated from the
1022 		 * mbuf.
1023 		 */
1024 		if (m_sharecount(m) == 1)
1025 			atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
1026 		/* fall through */
1027 	case M_EXT:
1028 		/*
1029 		 * Normal cluster association case, disconnect the cluster from
1030 		 * the mbuf.  The cluster may or may not be custom.
1031 		 */
1032 		m->m_ext.ext_free(m->m_ext.ext_arg);
1033 		m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
1034 		/* fall through */
1035 	case 0:
1036 		/*
1037 		 * return the mbuf to the mbuf cache.
1038 		 */
1039 		if (m->m_flags & M_PHCACHE) {
1040 			m->m_data = m->m_pktdat;
1041 			objcache_put(mbufphdr_cache, m);
1042 		} else {
1043 			m->m_data = m->m_dat;
1044 			objcache_put(mbuf_cache, m);
1045 		}
1046 		atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1);
1047 		break;
1048 	default:
1049 		if (!panicstr)
1050 			panic("bad mbuf flags %p %08x\n", m, m->m_flags);
1051 		break;
1052 	}
1053 	return (n);
1054 }
1055 
1056 void
1057 m_freem(struct mbuf *m)
1058 {
1059 	while (m)
1060 		m = m_free(m);
1061 }
1062 
1063 /*
1064  * mbuf utility routines
1065  */
1066 
1067 /*
1068  * Lesser-used path for M_PREPEND: allocate new mbuf to prepend to chain and
1069  * copy junk along.
1070  */
1071 struct mbuf *
1072 m_prepend(struct mbuf *m, int len, int how)
1073 {
1074 	struct mbuf *mn;
1075 
1076 	if (m->m_flags & M_PKTHDR)
1077 	    mn = m_gethdr(how, m->m_type);
1078 	else
1079 	    mn = m_get(how, m->m_type);
1080 	if (mn == NULL) {
1081 		m_freem(m);
1082 		return (NULL);
1083 	}
1084 	if (m->m_flags & M_PKTHDR)
1085 		M_MOVE_PKTHDR(mn, m);
1086 	mn->m_next = m;
1087 	m = mn;
1088 	if (len < MHLEN)
1089 		MH_ALIGN(m, len);
1090 	m->m_len = len;
1091 	return (m);
1092 }
1093 
1094 /*
1095  * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
1096  * continuing for "len" bytes.  If len is M_COPYALL, copy to end of mbuf.
1097  * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller.
1098  * Note that the copy is read-only, because clusters are not copied,
1099  * only their reference counts are incremented.
1100  */
1101 struct mbuf *
1102 m_copym(const struct mbuf *m, int off0, int len, int wait)
1103 {
1104 	struct mbuf *n, **np;
1105 	int off = off0;
1106 	struct mbuf *top;
1107 	int copyhdr = 0;
1108 
1109 	KASSERT(off >= 0, ("m_copym, negative off %d", off));
1110 	KASSERT(len >= 0, ("m_copym, negative len %d", len));
1111 	if (off == 0 && (m->m_flags & M_PKTHDR))
1112 		copyhdr = 1;
1113 	while (off > 0) {
1114 		KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
1115 		if (off < m->m_len)
1116 			break;
1117 		off -= m->m_len;
1118 		m = m->m_next;
1119 	}
1120 	np = &top;
1121 	top = NULL;
1122 	while (len > 0) {
1123 		if (m == NULL) {
1124 			KASSERT(len == M_COPYALL,
1125 			    ("m_copym, length > size of mbuf chain"));
1126 			break;
1127 		}
1128 		/*
1129 		 * Because we are sharing any cluster attachment below,
1130 		 * be sure to get an mbuf that does not have a cluster
1131 		 * associated with it.
1132 		 */
1133 		if (copyhdr)
1134 			n = m_gethdr(wait, m->m_type);
1135 		else
1136 			n = m_get(wait, m->m_type);
1137 		*np = n;
1138 		if (n == NULL)
1139 			goto nospace;
1140 		if (copyhdr) {
1141 			if (!m_dup_pkthdr(n, m, wait))
1142 				goto nospace;
1143 			if (len == M_COPYALL)
1144 				n->m_pkthdr.len -= off0;
1145 			else
1146 				n->m_pkthdr.len = len;
1147 			copyhdr = 0;
1148 		}
1149 		n->m_len = min(len, m->m_len - off);
1150 		if (m->m_flags & M_EXT) {
1151 			KKASSERT((n->m_flags & M_EXT) == 0);
1152 			n->m_data = m->m_data + off;
1153 			m->m_ext.ext_ref(m->m_ext.ext_arg);
1154 			n->m_ext = m->m_ext;
1155 			n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1156 		} else {
1157 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1158 			    (unsigned)n->m_len);
1159 		}
1160 		if (len != M_COPYALL)
1161 			len -= n->m_len;
1162 		off = 0;
1163 		m = m->m_next;
1164 		np = &n->m_next;
1165 	}
1166 	if (top == NULL)
1167 		atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1168 	return (top);
1169 nospace:
1170 	m_freem(top);
1171 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1172 	return (NULL);
1173 }
1174 
1175 /*
1176  * Copy an entire packet, including header (which must be present).
1177  * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
1178  * Note that the copy is read-only, because clusters are not copied,
1179  * only their reference counts are incremented.
1180  * Preserve alignment of the first mbuf so if the creator has left
1181  * some room at the beginning (e.g. for inserting protocol headers)
1182  * the copies also have the room available.
1183  */
1184 struct mbuf *
1185 m_copypacket(struct mbuf *m, int how)
1186 {
1187 	struct mbuf *top, *n, *o;
1188 
1189 	n = m_gethdr(how, m->m_type);
1190 	top = n;
1191 	if (!n)
1192 		goto nospace;
1193 
1194 	if (!m_dup_pkthdr(n, m, how))
1195 		goto nospace;
1196 	n->m_len = m->m_len;
1197 	if (m->m_flags & M_EXT) {
1198 		KKASSERT((n->m_flags & M_EXT) == 0);
1199 		n->m_data = m->m_data;
1200 		m->m_ext.ext_ref(m->m_ext.ext_arg);
1201 		n->m_ext = m->m_ext;
1202 		n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1203 	} else {
1204 		n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
1205 		bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1206 	}
1207 
1208 	m = m->m_next;
1209 	while (m) {
1210 		o = m_get(how, m->m_type);
1211 		if (!o)
1212 			goto nospace;
1213 
1214 		n->m_next = o;
1215 		n = n->m_next;
1216 
1217 		n->m_len = m->m_len;
1218 		if (m->m_flags & M_EXT) {
1219 			KKASSERT((n->m_flags & M_EXT) == 0);
1220 			n->m_data = m->m_data;
1221 			m->m_ext.ext_ref(m->m_ext.ext_arg);
1222 			n->m_ext = m->m_ext;
1223 			n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1224 		} else {
1225 			bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1226 		}
1227 
1228 		m = m->m_next;
1229 	}
1230 	return top;
1231 nospace:
1232 	m_freem(top);
1233 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1234 	return (NULL);
1235 }
1236 
1237 /*
1238  * Copy data from an mbuf chain starting "off" bytes from the beginning,
1239  * continuing for "len" bytes, into the indicated buffer.
1240  */
1241 void
1242 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
1243 {
1244 	unsigned count;
1245 
1246 	KASSERT(off >= 0, ("m_copydata, negative off %d", off));
1247 	KASSERT(len >= 0, ("m_copydata, negative len %d", len));
1248 	while (off > 0) {
1249 		KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
1250 		if (off < m->m_len)
1251 			break;
1252 		off -= m->m_len;
1253 		m = m->m_next;
1254 	}
1255 	while (len > 0) {
1256 		KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
1257 		count = min(m->m_len - off, len);
1258 		bcopy(mtod(m, caddr_t) + off, cp, count);
1259 		len -= count;
1260 		cp += count;
1261 		off = 0;
1262 		m = m->m_next;
1263 	}
1264 }
1265 
1266 /*
1267  * Copy a packet header mbuf chain into a completely new chain, including
1268  * copying any mbuf clusters.  Use this instead of m_copypacket() when
1269  * you need a writable copy of an mbuf chain.
1270  */
1271 struct mbuf *
1272 m_dup(struct mbuf *m, int how)
1273 {
1274 	struct mbuf **p, *top = NULL;
1275 	int remain, moff, nsize;
1276 
1277 	/* Sanity check */
1278 	if (m == NULL)
1279 		return (NULL);
1280 	KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __func__));
1281 
1282 	/* While there's more data, get a new mbuf, tack it on, and fill it */
1283 	remain = m->m_pkthdr.len;
1284 	moff = 0;
1285 	p = &top;
1286 	while (remain > 0 || top == NULL) {	/* allow m->m_pkthdr.len == 0 */
1287 		struct mbuf *n;
1288 
1289 		/* Get the next new mbuf */
1290 		n = m_getl(remain, how, m->m_type, top == NULL ? M_PKTHDR : 0,
1291 			   &nsize);
1292 		if (n == NULL)
1293 			goto nospace;
1294 		if (top == NULL)
1295 			if (!m_dup_pkthdr(n, m, how))
1296 				goto nospace0;
1297 
1298 		/* Link it into the new chain */
1299 		*p = n;
1300 		p = &n->m_next;
1301 
1302 		/* Copy data from original mbuf(s) into new mbuf */
1303 		n->m_len = 0;
1304 		while (n->m_len < nsize && m != NULL) {
1305 			int chunk = min(nsize - n->m_len, m->m_len - moff);
1306 
1307 			bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1308 			moff += chunk;
1309 			n->m_len += chunk;
1310 			remain -= chunk;
1311 			if (moff == m->m_len) {
1312 				m = m->m_next;
1313 				moff = 0;
1314 			}
1315 		}
1316 
1317 		/* Check correct total mbuf length */
1318 		KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
1319 			("%s: bogus m_pkthdr.len", __func__));
1320 	}
1321 	return (top);
1322 
1323 nospace:
1324 	m_freem(top);
1325 nospace0:
1326 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1327 	return (NULL);
1328 }
1329 
1330 /*
1331  * Copy the non-packet mbuf data chain into a new set of mbufs, including
1332  * copying any mbuf clusters.  This is typically used to realign a data
1333  * chain by nfs_realign().
1334  *
1335  * The original chain is left intact.  how should be MB_WAIT or MB_DONTWAIT
1336  * and NULL can be returned if MB_DONTWAIT is passed.
1337  *
1338  * Be careful to use cluster mbufs, a large mbuf chain converted to non
1339  * cluster mbufs can exhaust our supply of mbufs.
1340  */
1341 struct mbuf *
1342 m_dup_data(struct mbuf *m, int how)
1343 {
1344 	struct mbuf **p, *n, *top = NULL;
1345 	int mlen, moff, chunk, gsize, nsize;
1346 
1347 	/*
1348 	 * Degenerate case
1349 	 */
1350 	if (m == NULL)
1351 		return (NULL);
1352 
1353 	/*
1354 	 * Optimize the mbuf allocation but do not get too carried away.
1355 	 */
1356 	if (m->m_next || m->m_len > MLEN)
1357 		gsize = MCLBYTES;
1358 	else
1359 		gsize = MLEN;
1360 
1361 	/* Chain control */
1362 	p = &top;
1363 	n = NULL;
1364 	nsize = 0;
1365 
1366 	/*
1367 	 * Scan the mbuf chain until nothing is left, the new mbuf chain
1368 	 * will be allocated on the fly as needed.
1369 	 */
1370 	while (m) {
1371 		mlen = m->m_len;
1372 		moff = 0;
1373 
1374 		while (mlen) {
1375 			KKASSERT(m->m_type == MT_DATA);
1376 			if (n == NULL) {
1377 				n = m_getl(gsize, how, MT_DATA, 0, &nsize);
1378 				n->m_len = 0;
1379 				if (n == NULL)
1380 					goto nospace;
1381 				*p = n;
1382 				p = &n->m_next;
1383 			}
1384 			chunk = imin(mlen, nsize);
1385 			bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1386 			mlen -= chunk;
1387 			moff += chunk;
1388 			n->m_len += chunk;
1389 			nsize -= chunk;
1390 			if (nsize == 0)
1391 				n = NULL;
1392 		}
1393 		m = m->m_next;
1394 	}
1395 	*p = NULL;
1396 	return(top);
1397 nospace:
1398 	*p = NULL;
1399 	m_freem(top);
1400 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1401 	return (NULL);
1402 }
1403 
1404 /*
1405  * Concatenate mbuf chain n to m.
1406  * Both chains must be of the same type (e.g. MT_DATA).
1407  * Any m_pkthdr is not updated.
1408  */
1409 void
1410 m_cat(struct mbuf *m, struct mbuf *n)
1411 {
1412 	m = m_last(m);
1413 	while (n) {
1414 		if (m->m_flags & M_EXT ||
1415 		    m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) {
1416 			/* just join the two chains */
1417 			m->m_next = n;
1418 			return;
1419 		}
1420 		/* splat the data from one into the other */
1421 		bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1422 		    (u_int)n->m_len);
1423 		m->m_len += n->m_len;
1424 		n = m_free(n);
1425 	}
1426 }
1427 
1428 void
1429 m_adj(struct mbuf *mp, int req_len)
1430 {
1431 	int len = req_len;
1432 	struct mbuf *m;
1433 	int count;
1434 
1435 	if ((m = mp) == NULL)
1436 		return;
1437 	if (len >= 0) {
1438 		/*
1439 		 * Trim from head.
1440 		 */
1441 		while (m != NULL && len > 0) {
1442 			if (m->m_len <= len) {
1443 				len -= m->m_len;
1444 				m->m_len = 0;
1445 				m = m->m_next;
1446 			} else {
1447 				m->m_len -= len;
1448 				m->m_data += len;
1449 				len = 0;
1450 			}
1451 		}
1452 		m = mp;
1453 		if (mp->m_flags & M_PKTHDR)
1454 			m->m_pkthdr.len -= (req_len - len);
1455 	} else {
1456 		/*
1457 		 * Trim from tail.  Scan the mbuf chain,
1458 		 * calculating its length and finding the last mbuf.
1459 		 * If the adjustment only affects this mbuf, then just
1460 		 * adjust and return.  Otherwise, rescan and truncate
1461 		 * after the remaining size.
1462 		 */
1463 		len = -len;
1464 		count = 0;
1465 		for (;;) {
1466 			count += m->m_len;
1467 			if (m->m_next == NULL)
1468 				break;
1469 			m = m->m_next;
1470 		}
1471 		if (m->m_len >= len) {
1472 			m->m_len -= len;
1473 			if (mp->m_flags & M_PKTHDR)
1474 				mp->m_pkthdr.len -= len;
1475 			return;
1476 		}
1477 		count -= len;
1478 		if (count < 0)
1479 			count = 0;
1480 		/*
1481 		 * Correct length for chain is "count".
1482 		 * Find the mbuf with last data, adjust its length,
1483 		 * and toss data from remaining mbufs on chain.
1484 		 */
1485 		m = mp;
1486 		if (m->m_flags & M_PKTHDR)
1487 			m->m_pkthdr.len = count;
1488 		for (; m; m = m->m_next) {
1489 			if (m->m_len >= count) {
1490 				m->m_len = count;
1491 				break;
1492 			}
1493 			count -= m->m_len;
1494 		}
1495 		while (m->m_next)
1496 			(m = m->m_next) ->m_len = 0;
1497 	}
1498 }
1499 
1500 /*
1501  * Set the m_data pointer of a newly-allocated mbuf
1502  * to place an object of the specified size at the
1503  * end of the mbuf, longword aligned.
1504  */
1505 void
1506 m_align(struct mbuf *m, int len)
1507 {
1508 	int adjust;
1509 
1510 	if (m->m_flags & M_EXT)
1511 		adjust = m->m_ext.ext_size - len;
1512 	else if (m->m_flags & M_PKTHDR)
1513 		adjust = MHLEN - len;
1514 	else
1515 		adjust = MLEN - len;
1516 	m->m_data += adjust &~ (sizeof(long)-1);
1517 }
1518 
1519 /*
1520  * Rearrange an mbuf chain so that len bytes are contiguous
1521  * and in the data area of an mbuf (so that mtod will work for a structure
1522  * of size len).  Returns the resulting mbuf chain on success, frees it and
1523  * returns null on failure.  If there is room, it will add up to
1524  * max_protohdr-len extra bytes to the contiguous region in an attempt to
1525  * avoid being called next time.
1526  */
1527 struct mbuf *
1528 m_pullup(struct mbuf *n, int len)
1529 {
1530 	struct mbuf *m;
1531 	int count;
1532 	int space;
1533 
1534 	/*
1535 	 * If first mbuf has no cluster, and has room for len bytes
1536 	 * without shifting current data, pullup into it,
1537 	 * otherwise allocate a new mbuf to prepend to the chain.
1538 	 */
1539 	if (!(n->m_flags & M_EXT) &&
1540 	    n->m_data + len < &n->m_dat[MLEN] &&
1541 	    n->m_next) {
1542 		if (n->m_len >= len)
1543 			return (n);
1544 		m = n;
1545 		n = n->m_next;
1546 		len -= m->m_len;
1547 	} else {
1548 		if (len > MHLEN)
1549 			goto bad;
1550 		if (n->m_flags & M_PKTHDR)
1551 			m = m_gethdr(MB_DONTWAIT, n->m_type);
1552 		else
1553 			m = m_get(MB_DONTWAIT, n->m_type);
1554 		if (m == NULL)
1555 			goto bad;
1556 		m->m_len = 0;
1557 		if (n->m_flags & M_PKTHDR)
1558 			M_MOVE_PKTHDR(m, n);
1559 	}
1560 	space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
1561 	do {
1562 		count = min(min(max(len, max_protohdr), space), n->m_len);
1563 		bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1564 		  (unsigned)count);
1565 		len -= count;
1566 		m->m_len += count;
1567 		n->m_len -= count;
1568 		space -= count;
1569 		if (n->m_len)
1570 			n->m_data += count;
1571 		else
1572 			n = m_free(n);
1573 	} while (len > 0 && n);
1574 	if (len > 0) {
1575 		m_free(m);
1576 		goto bad;
1577 	}
1578 	m->m_next = n;
1579 	return (m);
1580 bad:
1581 	m_freem(n);
1582 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1583 	return (NULL);
1584 }
1585 
1586 /*
1587  * Partition an mbuf chain in two pieces, returning the tail --
1588  * all but the first len0 bytes.  In case of failure, it returns NULL and
1589  * attempts to restore the chain to its original state.
1590  *
1591  * Note that the resulting mbufs might be read-only, because the new
1592  * mbuf can end up sharing an mbuf cluster with the original mbuf if
1593  * the "breaking point" happens to lie within a cluster mbuf. Use the
1594  * M_WRITABLE() macro to check for this case.
1595  */
1596 struct mbuf *
1597 m_split(struct mbuf *m0, int len0, int wait)
1598 {
1599 	struct mbuf *m, *n;
1600 	unsigned len = len0, remain;
1601 
1602 	for (m = m0; m && len > m->m_len; m = m->m_next)
1603 		len -= m->m_len;
1604 	if (m == NULL)
1605 		return (NULL);
1606 	remain = m->m_len - len;
1607 	if (m0->m_flags & M_PKTHDR) {
1608 		n = m_gethdr(wait, m0->m_type);
1609 		if (n == NULL)
1610 			return (NULL);
1611 		n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1612 		n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1613 		m0->m_pkthdr.len = len0;
1614 		if (m->m_flags & M_EXT)
1615 			goto extpacket;
1616 		if (remain > MHLEN) {
1617 			/* m can't be the lead packet */
1618 			MH_ALIGN(n, 0);
1619 			n->m_next = m_split(m, len, wait);
1620 			if (n->m_next == NULL) {
1621 				m_free(n);
1622 				return (NULL);
1623 			} else {
1624 				n->m_len = 0;
1625 				return (n);
1626 			}
1627 		} else
1628 			MH_ALIGN(n, remain);
1629 	} else if (remain == 0) {
1630 		n = m->m_next;
1631 		m->m_next = 0;
1632 		return (n);
1633 	} else {
1634 		n = m_get(wait, m->m_type);
1635 		if (n == NULL)
1636 			return (NULL);
1637 		M_ALIGN(n, remain);
1638 	}
1639 extpacket:
1640 	if (m->m_flags & M_EXT) {
1641 		KKASSERT((n->m_flags & M_EXT) == 0);
1642 		n->m_data = m->m_data + len;
1643 		m->m_ext.ext_ref(m->m_ext.ext_arg);
1644 		n->m_ext = m->m_ext;
1645 		n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1646 	} else {
1647 		bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1648 	}
1649 	n->m_len = remain;
1650 	m->m_len = len;
1651 	n->m_next = m->m_next;
1652 	m->m_next = 0;
1653 	return (n);
1654 }
1655 
1656 /*
1657  * Routine to copy from device local memory into mbufs.
1658  * Note: "offset" is ill-defined and always called as 0, so ignore it.
1659  */
1660 struct mbuf *
1661 m_devget(char *buf, int len, int offset, struct ifnet *ifp,
1662     void (*copy)(volatile const void *from, volatile void *to, size_t length))
1663 {
1664 	struct mbuf *m, *mfirst = NULL, **mtail;
1665 	int nsize, flags;
1666 
1667 	if (copy == NULL)
1668 		copy = bcopy;
1669 	mtail = &mfirst;
1670 	flags = M_PKTHDR;
1671 
1672 	while (len > 0) {
1673 		m = m_getl(len, MB_DONTWAIT, MT_DATA, flags, &nsize);
1674 		if (m == NULL) {
1675 			m_freem(mfirst);
1676 			return (NULL);
1677 		}
1678 		m->m_len = min(len, nsize);
1679 
1680 		if (flags & M_PKTHDR) {
1681 			if (len + max_linkhdr <= nsize)
1682 				m->m_data += max_linkhdr;
1683 			m->m_pkthdr.rcvif = ifp;
1684 			m->m_pkthdr.len = len;
1685 			flags = 0;
1686 		}
1687 
1688 		copy(buf, m->m_data, (unsigned)m->m_len);
1689 		buf += m->m_len;
1690 		len -= m->m_len;
1691 		*mtail = m;
1692 		mtail = &m->m_next;
1693 	}
1694 
1695 	return (mfirst);
1696 }
1697 
1698 /*
1699  * Routine to pad mbuf to the specified length 'padto'.
1700  */
1701 int
1702 m_devpad(struct mbuf *m, int padto)
1703 {
1704 	struct mbuf *last = NULL;
1705 	int padlen;
1706 
1707 	if (padto <= m->m_pkthdr.len)
1708 		return 0;
1709 
1710 	padlen = padto - m->m_pkthdr.len;
1711 
1712 	/* if there's only the packet-header and we can pad there, use it. */
1713 	if (m->m_pkthdr.len == m->m_len && M_TRAILINGSPACE(m) >= padlen) {
1714 		last = m;
1715 	} else {
1716 		/*
1717 		 * Walk packet chain to find last mbuf. We will either
1718 		 * pad there, or append a new mbuf and pad it
1719 		 */
1720 		for (last = m; last->m_next != NULL; last = last->m_next)
1721 			; /* EMPTY */
1722 
1723 		/* `last' now points to last in chain. */
1724 		if (M_TRAILINGSPACE(last) < padlen) {
1725 			struct mbuf *n;
1726 
1727 			/* Allocate new empty mbuf, pad it.  Compact later. */
1728 			MGET(n, MB_DONTWAIT, MT_DATA);
1729 			if (n == NULL)
1730 				return ENOBUFS;
1731 			n->m_len = 0;
1732 			last->m_next = n;
1733 			last = n;
1734 		}
1735 	}
1736 	KKASSERT(M_TRAILINGSPACE(last) >= padlen);
1737 	KKASSERT(M_WRITABLE(last));
1738 
1739 	/* Now zero the pad area */
1740 	bzero(mtod(last, char *) + last->m_len, padlen);
1741 	last->m_len += padlen;
1742 	m->m_pkthdr.len += padlen;
1743 	return 0;
1744 }
1745 
1746 /*
1747  * Copy data from a buffer back into the indicated mbuf chain,
1748  * starting "off" bytes from the beginning, extending the mbuf
1749  * chain if necessary.
1750  */
1751 void
1752 m_copyback(struct mbuf *m0, int off, int len, caddr_t cp)
1753 {
1754 	int mlen;
1755 	struct mbuf *m = m0, *n;
1756 	int totlen = 0;
1757 
1758 	if (m0 == NULL)
1759 		return;
1760 	while (off > (mlen = m->m_len)) {
1761 		off -= mlen;
1762 		totlen += mlen;
1763 		if (m->m_next == NULL) {
1764 			n = m_getclr(MB_DONTWAIT, m->m_type);
1765 			if (n == NULL)
1766 				goto out;
1767 			n->m_len = min(MLEN, len + off);
1768 			m->m_next = n;
1769 		}
1770 		m = m->m_next;
1771 	}
1772 	while (len > 0) {
1773 		mlen = min (m->m_len - off, len);
1774 		bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen);
1775 		cp += mlen;
1776 		len -= mlen;
1777 		mlen += off;
1778 		off = 0;
1779 		totlen += mlen;
1780 		if (len == 0)
1781 			break;
1782 		if (m->m_next == NULL) {
1783 			n = m_get(MB_DONTWAIT, m->m_type);
1784 			if (n == NULL)
1785 				break;
1786 			n->m_len = min(MLEN, len);
1787 			m->m_next = n;
1788 		}
1789 		m = m->m_next;
1790 	}
1791 out:	if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1792 		m->m_pkthdr.len = totlen;
1793 }
1794 
1795 /*
1796  * Append the specified data to the indicated mbuf chain,
1797  * Extend the mbuf chain if the new data does not fit in
1798  * existing space.
1799  *
1800  * Return 1 if able to complete the job; otherwise 0.
1801  */
1802 int
1803 m_append(struct mbuf *m0, int len, c_caddr_t cp)
1804 {
1805 	struct mbuf *m, *n;
1806 	int remainder, space;
1807 
1808 	for (m = m0; m->m_next != NULL; m = m->m_next)
1809 		;
1810 	remainder = len;
1811 	space = M_TRAILINGSPACE(m);
1812 	if (space > 0) {
1813 		/*
1814 		 * Copy into available space.
1815 		 */
1816 		if (space > remainder)
1817 			space = remainder;
1818 		bcopy(cp, mtod(m, caddr_t) + m->m_len, space);
1819 		m->m_len += space;
1820 		cp += space, remainder -= space;
1821 	}
1822 	while (remainder > 0) {
1823 		/*
1824 		 * Allocate a new mbuf; could check space
1825 		 * and allocate a cluster instead.
1826 		 */
1827 		n = m_get(MB_DONTWAIT, m->m_type);
1828 		if (n == NULL)
1829 			break;
1830 		n->m_len = min(MLEN, remainder);
1831 		bcopy(cp, mtod(n, caddr_t), n->m_len);
1832 		cp += n->m_len, remainder -= n->m_len;
1833 		m->m_next = n;
1834 		m = n;
1835 	}
1836 	if (m0->m_flags & M_PKTHDR)
1837 		m0->m_pkthdr.len += len - remainder;
1838 	return (remainder == 0);
1839 }
1840 
1841 /*
1842  * Apply function f to the data in an mbuf chain starting "off" bytes from
1843  * the beginning, continuing for "len" bytes.
1844  */
1845 int
1846 m_apply(struct mbuf *m, int off, int len,
1847     int (*f)(void *, void *, u_int), void *arg)
1848 {
1849 	u_int count;
1850 	int rval;
1851 
1852 	KASSERT(off >= 0, ("m_apply, negative off %d", off));
1853 	KASSERT(len >= 0, ("m_apply, negative len %d", len));
1854 	while (off > 0) {
1855 		KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1856 		if (off < m->m_len)
1857 			break;
1858 		off -= m->m_len;
1859 		m = m->m_next;
1860 	}
1861 	while (len > 0) {
1862 		KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1863 		count = min(m->m_len - off, len);
1864 		rval = (*f)(arg, mtod(m, caddr_t) + off, count);
1865 		if (rval)
1866 			return (rval);
1867 		len -= count;
1868 		off = 0;
1869 		m = m->m_next;
1870 	}
1871 	return (0);
1872 }
1873 
1874 /*
1875  * Return a pointer to mbuf/offset of location in mbuf chain.
1876  */
1877 struct mbuf *
1878 m_getptr(struct mbuf *m, int loc, int *off)
1879 {
1880 
1881 	while (loc >= 0) {
1882 		/* Normal end of search. */
1883 		if (m->m_len > loc) {
1884 			*off = loc;
1885 			return (m);
1886 		} else {
1887 			loc -= m->m_len;
1888 			if (m->m_next == NULL) {
1889 				if (loc == 0) {
1890 					/* Point at the end of valid data. */
1891 					*off = m->m_len;
1892 					return (m);
1893 				}
1894 				return (NULL);
1895 			}
1896 			m = m->m_next;
1897 		}
1898 	}
1899 	return (NULL);
1900 }
1901 
1902 void
1903 m_print(const struct mbuf *m)
1904 {
1905 	int len;
1906 	const struct mbuf *m2;
1907 
1908 	len = m->m_pkthdr.len;
1909 	m2 = m;
1910 	while (len) {
1911 		kprintf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-");
1912 		len -= m2->m_len;
1913 		m2 = m2->m_next;
1914 	}
1915 	return;
1916 }
1917 
1918 /*
1919  * "Move" mbuf pkthdr from "from" to "to".
1920  * "from" must have M_PKTHDR set, and "to" must be empty.
1921  */
1922 void
1923 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
1924 {
1925 	KASSERT((to->m_flags & M_PKTHDR), ("m_move_pkthdr: not packet header"));
1926 
1927 	to->m_flags |= from->m_flags & M_COPYFLAGS;
1928 	to->m_pkthdr = from->m_pkthdr;		/* especially tags */
1929 	SLIST_INIT(&from->m_pkthdr.tags);	/* purge tags from src */
1930 }
1931 
1932 /*
1933  * Duplicate "from"'s mbuf pkthdr in "to".
1934  * "from" must have M_PKTHDR set, and "to" must be empty.
1935  * In particular, this does a deep copy of the packet tags.
1936  */
1937 int
1938 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
1939 {
1940 	KASSERT((to->m_flags & M_PKTHDR), ("m_dup_pkthdr: not packet header"));
1941 
1942 	to->m_flags = (from->m_flags & M_COPYFLAGS) |
1943 		      (to->m_flags & ~M_COPYFLAGS);
1944 	to->m_pkthdr = from->m_pkthdr;
1945 	SLIST_INIT(&to->m_pkthdr.tags);
1946 	return (m_tag_copy_chain(to, from, how));
1947 }
1948 
1949 /*
1950  * Defragment a mbuf chain, returning the shortest possible
1951  * chain of mbufs and clusters.  If allocation fails and
1952  * this cannot be completed, NULL will be returned, but
1953  * the passed in chain will be unchanged.  Upon success,
1954  * the original chain will be freed, and the new chain
1955  * will be returned.
1956  *
1957  * If a non-packet header is passed in, the original
1958  * mbuf (chain?) will be returned unharmed.
1959  *
1960  * m_defrag_nofree doesn't free the passed in mbuf.
1961  */
1962 struct mbuf *
1963 m_defrag(struct mbuf *m0, int how)
1964 {
1965 	struct mbuf *m_new;
1966 
1967 	if ((m_new = m_defrag_nofree(m0, how)) == NULL)
1968 		return (NULL);
1969 	if (m_new != m0)
1970 		m_freem(m0);
1971 	return (m_new);
1972 }
1973 
1974 struct mbuf *
1975 m_defrag_nofree(struct mbuf *m0, int how)
1976 {
1977 	struct mbuf	*m_new = NULL, *m_final = NULL;
1978 	int		progress = 0, length, nsize;
1979 
1980 	if (!(m0->m_flags & M_PKTHDR))
1981 		return (m0);
1982 
1983 #ifdef MBUF_STRESS_TEST
1984 	if (m_defragrandomfailures) {
1985 		int temp = karc4random() & 0xff;
1986 		if (temp == 0xba)
1987 			goto nospace;
1988 	}
1989 #endif
1990 
1991 	m_final = m_getl(m0->m_pkthdr.len, how, MT_DATA, M_PKTHDR, &nsize);
1992 	if (m_final == NULL)
1993 		goto nospace;
1994 	m_final->m_len = 0;	/* in case m0->m_pkthdr.len is zero */
1995 
1996 	if (m_dup_pkthdr(m_final, m0, how) == 0)
1997 		goto nospace;
1998 
1999 	m_new = m_final;
2000 
2001 	while (progress < m0->m_pkthdr.len) {
2002 		length = m0->m_pkthdr.len - progress;
2003 		if (length > MCLBYTES)
2004 			length = MCLBYTES;
2005 
2006 		if (m_new == NULL) {
2007 			m_new = m_getl(length, how, MT_DATA, 0, &nsize);
2008 			if (m_new == NULL)
2009 				goto nospace;
2010 		}
2011 
2012 		m_copydata(m0, progress, length, mtod(m_new, caddr_t));
2013 		progress += length;
2014 		m_new->m_len = length;
2015 		if (m_new != m_final)
2016 			m_cat(m_final, m_new);
2017 		m_new = NULL;
2018 	}
2019 	if (m0->m_next == NULL)
2020 		m_defraguseless++;
2021 	m_defragpackets++;
2022 	m_defragbytes += m_final->m_pkthdr.len;
2023 	return (m_final);
2024 nospace:
2025 	m_defragfailure++;
2026 	if (m_new)
2027 		m_free(m_new);
2028 	m_freem(m_final);
2029 	return (NULL);
2030 }
2031 
2032 /*
2033  * Move data from uio into mbufs.
2034  */
2035 struct mbuf *
2036 m_uiomove(struct uio *uio)
2037 {
2038 	struct mbuf *m;			/* current working mbuf */
2039 	struct mbuf *head = NULL;	/* result mbuf chain */
2040 	struct mbuf **mp = &head;
2041 	int flags = M_PKTHDR;
2042 	int nsize;
2043 	int error;
2044 	int resid;
2045 
2046 	do {
2047 		if (uio->uio_resid > INT_MAX)
2048 			resid = INT_MAX;
2049 		else
2050 			resid = (int)uio->uio_resid;
2051 		m = m_getl(resid, MB_WAIT, MT_DATA, flags, &nsize);
2052 		if (flags) {
2053 			m->m_pkthdr.len = 0;
2054 			/* Leave room for protocol headers. */
2055 			if (resid < MHLEN)
2056 				MH_ALIGN(m, resid);
2057 			flags = 0;
2058 		}
2059 		m->m_len = imin(nsize, resid);
2060 		error = uiomove(mtod(m, caddr_t), m->m_len, uio);
2061 		if (error) {
2062 			m_free(m);
2063 			goto failed;
2064 		}
2065 		*mp = m;
2066 		mp = &m->m_next;
2067 		head->m_pkthdr.len += m->m_len;
2068 	} while (uio->uio_resid > 0);
2069 
2070 	return (head);
2071 
2072 failed:
2073 	m_freem(head);
2074 	return (NULL);
2075 }
2076 
2077 struct mbuf *
2078 m_last(struct mbuf *m)
2079 {
2080 	while (m->m_next)
2081 		m = m->m_next;
2082 	return (m);
2083 }
2084 
2085 /*
2086  * Return the number of bytes in an mbuf chain.
2087  * If lastm is not NULL, also return the last mbuf.
2088  */
2089 u_int
2090 m_lengthm(struct mbuf *m, struct mbuf **lastm)
2091 {
2092 	u_int len = 0;
2093 	struct mbuf *prev = m;
2094 
2095 	while (m) {
2096 		len += m->m_len;
2097 		prev = m;
2098 		m = m->m_next;
2099 	}
2100 	if (lastm != NULL)
2101 		*lastm = prev;
2102 	return (len);
2103 }
2104 
2105 /*
2106  * Like m_lengthm(), except also keep track of mbuf usage.
2107  */
2108 u_int
2109 m_countm(struct mbuf *m, struct mbuf **lastm, u_int *pmbcnt)
2110 {
2111 	u_int len = 0, mbcnt = 0;
2112 	struct mbuf *prev = m;
2113 
2114 	while (m) {
2115 		len += m->m_len;
2116 		mbcnt += MSIZE;
2117 		if (m->m_flags & M_EXT)
2118 			mbcnt += m->m_ext.ext_size;
2119 		prev = m;
2120 		m = m->m_next;
2121 	}
2122 	if (lastm != NULL)
2123 		*lastm = prev;
2124 	*pmbcnt = mbcnt;
2125 	return (len);
2126 }
2127