xref: /dflybsd-src/sys/kern/uipc_mbuf.c (revision 3bcb6e5e80cb05cecaddd866d4b15b428334b9aa)
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  */
71 
72 #include "opt_param.h"
73 #include "opt_mbuf_stress_test.h"
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/file.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, *mjclmeta_cache;
256 struct objcache *mbufcluster_cache, *mbufphdrcluster_cache;
257 struct objcache *mbufjcluster_cache, *mbufphdrjcluster_cache;
258 
259 int	nmbclusters;
260 int	nmbufs;
261 
262 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW,
263 	&max_linkhdr, 0, "Max size of a link-level header");
264 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW,
265 	&max_protohdr, 0, "Max size of a protocol header");
266 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0,
267 	"Max size of link+protocol headers");
268 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW,
269 	&max_datalen, 0, "Max data payload size without headers");
270 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW,
271 	&mbuf_wait, 0, "Time in ticks to sleep after failed mbuf allocations");
272 static int do_mbstat(SYSCTL_HANDLER_ARGS);
273 
274 SYSCTL_PROC(_kern_ipc, KIPC_MBSTAT, mbstat, CTLTYPE_STRUCT|CTLFLAG_RD,
275 	0, 0, do_mbstat, "S,mbstat", "mbuf usage statistics");
276 
277 static int do_mbtypes(SYSCTL_HANDLER_ARGS);
278 
279 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbtypes, CTLTYPE_ULONG|CTLFLAG_RD,
280 	0, 0, do_mbtypes, "LU", "");
281 
282 static int
283 do_mbstat(SYSCTL_HANDLER_ARGS)
284 {
285 	struct mbstat mbstat_total;
286 	struct mbstat *mbstat_totalp;
287 	int i;
288 
289 	bzero(&mbstat_total, sizeof(mbstat_total));
290 	mbstat_totalp = &mbstat_total;
291 
292 	for (i = 0; i < ncpus; i++)
293 	{
294 		mbstat_total.m_mbufs += mbstat[i].m_mbufs;
295 		mbstat_total.m_clusters += mbstat[i].m_clusters;
296 		mbstat_total.m_spare += mbstat[i].m_spare;
297 		mbstat_total.m_clfree += mbstat[i].m_clfree;
298 		mbstat_total.m_drops += mbstat[i].m_drops;
299 		mbstat_total.m_wait += mbstat[i].m_wait;
300 		mbstat_total.m_drain += mbstat[i].m_drain;
301 		mbstat_total.m_mcfail += mbstat[i].m_mcfail;
302 		mbstat_total.m_mpfail += mbstat[i].m_mpfail;
303 
304 	}
305 	/*
306 	 * The following fields are not cumulative fields so just
307 	 * get their values once.
308 	 */
309 	mbstat_total.m_msize = mbstat[0].m_msize;
310 	mbstat_total.m_mclbytes = mbstat[0].m_mclbytes;
311 	mbstat_total.m_minclsize = mbstat[0].m_minclsize;
312 	mbstat_total.m_mlen = mbstat[0].m_mlen;
313 	mbstat_total.m_mhlen = mbstat[0].m_mhlen;
314 
315 	return(sysctl_handle_opaque(oidp, mbstat_totalp, sizeof(mbstat_total), req));
316 }
317 
318 static int
319 do_mbtypes(SYSCTL_HANDLER_ARGS)
320 {
321 	u_long totals[MT_NTYPES];
322 	int i, j;
323 
324 	for (i = 0; i < MT_NTYPES; i++)
325 		totals[i] = 0;
326 
327 	for (i = 0; i < ncpus; i++)
328 	{
329 		for (j = 0; j < MT_NTYPES; j++)
330 			totals[j] += mbtypes[i][j];
331 	}
332 
333 	return(sysctl_handle_opaque(oidp, totals, sizeof(totals), req));
334 }
335 
336 /*
337  * These are read-only because we do not currently have any code
338  * to adjust the objcache limits after the fact.  The variables
339  * may only be set as boot-time tunables.
340  */
341 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RD,
342 	   &nmbclusters, 0, "Maximum number of mbuf clusters available");
343 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RD, &nmbufs, 0,
344 	   "Maximum number of mbufs available");
345 
346 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
347 	   &m_defragpackets, 0, "Number of defragment packets");
348 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
349 	   &m_defragbytes, 0, "Number of defragment bytes");
350 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
351 	   &m_defraguseless, 0, "Number of useless defragment mbuf chain operations");
352 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
353 	   &m_defragfailure, 0, "Number of failed defragment mbuf chain operations");
354 #ifdef MBUF_STRESS_TEST
355 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
356 	   &m_defragrandomfailures, 0, "");
357 #endif
358 
359 static MALLOC_DEFINE(M_MBUF, "mbuf", "mbuf");
360 static MALLOC_DEFINE(M_MBUFCL, "mbufcl", "mbufcl");
361 static MALLOC_DEFINE(M_MJBUFCL, "mbufcl", "mbufcl");
362 static MALLOC_DEFINE(M_MCLMETA, "mclmeta", "mclmeta");
363 static MALLOC_DEFINE(M_MJCLMETA, "mjclmeta", "mjclmeta");
364 
365 static void m_reclaim (void);
366 static void m_mclref(void *arg);
367 static void m_mclfree(void *arg);
368 
369 /*
370  * NOTE: Default NMBUFS must take into account a possible DOS attack
371  *	 using fd passing on unix domain sockets.
372  */
373 #ifndef NMBCLUSTERS
374 #define NMBCLUSTERS	(512 + maxusers * 16)
375 #endif
376 #ifndef NMBUFS
377 #define NMBUFS		(nmbclusters * 2 + maxfiles)
378 #endif
379 
380 /*
381  * Perform sanity checks of tunables declared above.
382  */
383 static void
384 tunable_mbinit(void *dummy)
385 {
386 	/*
387 	 * This has to be done before VM init.
388 	 */
389 	nmbclusters = NMBCLUSTERS;
390 	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
391 	nmbufs = NMBUFS;
392 	TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
393 	/* Sanity checks */
394 	if (nmbufs < nmbclusters * 2)
395 		nmbufs = nmbclusters * 2;
396 }
397 SYSINIT(tunable_mbinit, SI_BOOT1_TUNABLES, SI_ORDER_ANY,
398 	tunable_mbinit, NULL);
399 
400 /* "number of clusters of pages" */
401 #define NCL_INIT	1
402 
403 #define NMB_INIT	16
404 
405 /*
406  * The mbuf object cache only guarantees that m_next and m_nextpkt are
407  * NULL and that m_data points to the beginning of the data area.  In
408  * particular, m_len and m_pkthdr.len are uninitialized.  It is the
409  * responsibility of the caller to initialize those fields before use.
410  */
411 
412 static __inline boolean_t
413 mbuf_ctor(void *obj, void *private, int ocflags)
414 {
415 	struct mbuf *m = obj;
416 
417 	m->m_next = NULL;
418 	m->m_nextpkt = NULL;
419 	m->m_data = m->m_dat;
420 	m->m_flags = 0;
421 
422 	return (TRUE);
423 }
424 
425 /*
426  * Initialize the mbuf and the packet header fields.
427  */
428 static boolean_t
429 mbufphdr_ctor(void *obj, void *private, int ocflags)
430 {
431 	struct mbuf *m = obj;
432 
433 	m->m_next = NULL;
434 	m->m_nextpkt = NULL;
435 	m->m_data = m->m_pktdat;
436 	m->m_flags = M_PKTHDR | M_PHCACHE;
437 
438 	m->m_pkthdr.rcvif = NULL;	/* eliminate XXX JH */
439 	SLIST_INIT(&m->m_pkthdr.tags);
440 	m->m_pkthdr.csum_flags = 0;	/* eliminate XXX JH */
441 	m->m_pkthdr.fw_flags = 0;	/* eliminate XXX JH */
442 
443 	return (TRUE);
444 }
445 
446 /*
447  * A mbcluster object consists of 2K (MCLBYTES) cluster and a refcount.
448  */
449 static boolean_t
450 mclmeta_ctor(void *obj, void *private, int ocflags)
451 {
452 	struct mbcluster *cl = obj;
453 	void *buf;
454 
455 	if (ocflags & M_NOWAIT)
456 		buf = kmalloc(MCLBYTES, M_MBUFCL, M_NOWAIT | M_ZERO);
457 	else
458 		buf = kmalloc(MCLBYTES, M_MBUFCL, M_INTWAIT | M_ZERO);
459 	if (buf == NULL)
460 		return (FALSE);
461 	cl->mcl_refs = 0;
462 	cl->mcl_data = buf;
463 	return (TRUE);
464 }
465 
466 static boolean_t
467 mjclmeta_ctor(void *obj, void *private, int ocflags)
468 {
469 	struct mbcluster *cl = obj;
470 	void *buf;
471 
472 	if (ocflags & M_NOWAIT)
473 		buf = kmalloc(MJUMPAGESIZE, M_MBUFCL, M_NOWAIT | M_ZERO);
474 	else
475 		buf = kmalloc(MJUMPAGESIZE, M_MBUFCL, M_INTWAIT | M_ZERO);
476 	if (buf == NULL)
477 		return (FALSE);
478 	cl->mcl_refs = 0;
479 	cl->mcl_data = buf;
480 	return (TRUE);
481 }
482 
483 static void
484 mclmeta_dtor(void *obj, void *private)
485 {
486 	struct mbcluster *mcl = obj;
487 
488 	KKASSERT(mcl->mcl_refs == 0);
489 	kfree(mcl->mcl_data, M_MBUFCL);
490 }
491 
492 static void
493 linkjcluster(struct mbuf *m, struct mbcluster *cl, uint size)
494 {
495 	/*
496 	 * Add the cluster to the mbuf.  The caller will detect that the
497 	 * mbuf now has an attached cluster.
498 	 */
499 	m->m_ext.ext_arg = cl;
500 	m->m_ext.ext_buf = cl->mcl_data;
501 	m->m_ext.ext_ref = m_mclref;
502 	m->m_ext.ext_free = m_mclfree;
503 	m->m_ext.ext_size = size;
504 	atomic_add_int(&cl->mcl_refs, 1);
505 
506 	m->m_data = m->m_ext.ext_buf;
507 	m->m_flags |= M_EXT | M_EXT_CLUSTER;
508 }
509 
510 static void
511 linkcluster(struct mbuf *m, struct mbcluster *cl)
512 {
513 	linkjcluster(m, cl, MCLBYTES);
514 }
515 
516 static boolean_t
517 mbufphdrcluster_ctor(void *obj, void *private, int ocflags)
518 {
519 	struct mbuf *m = obj;
520 	struct mbcluster *cl;
521 
522 	mbufphdr_ctor(obj, private, ocflags);
523 	cl = objcache_get(mclmeta_cache, ocflags);
524 	if (cl == NULL) {
525 		++mbstat[mycpu->gd_cpuid].m_drops;
526 		return (FALSE);
527 	}
528 	m->m_flags |= M_CLCACHE;
529 	linkcluster(m, cl);
530 	return (TRUE);
531 }
532 
533 static boolean_t
534 mbufphdrjcluster_ctor(void *obj, void *private, int ocflags)
535 {
536 	struct mbuf *m = obj;
537 	struct mbcluster *cl;
538 
539 	mbufphdr_ctor(obj, private, ocflags);
540 	cl = objcache_get(mjclmeta_cache, ocflags);
541 	if (cl == NULL) {
542 		++mbstat[mycpu->gd_cpuid].m_drops;
543 		return (FALSE);
544 	}
545 	m->m_flags |= M_CLCACHE;
546 	linkjcluster(m, cl, MJUMPAGESIZE);
547 	return (TRUE);
548 }
549 
550 static boolean_t
551 mbufcluster_ctor(void *obj, void *private, int ocflags)
552 {
553 	struct mbuf *m = obj;
554 	struct mbcluster *cl;
555 
556 	mbuf_ctor(obj, private, ocflags);
557 	cl = objcache_get(mclmeta_cache, ocflags);
558 	if (cl == NULL) {
559 		++mbstat[mycpu->gd_cpuid].m_drops;
560 		return (FALSE);
561 	}
562 	m->m_flags |= M_CLCACHE;
563 	linkcluster(m, cl);
564 	return (TRUE);
565 }
566 
567 static boolean_t
568 mbufjcluster_ctor(void *obj, void *private, int ocflags)
569 {
570 	struct mbuf *m = obj;
571 	struct mbcluster *cl;
572 
573 	mbuf_ctor(obj, private, ocflags);
574 	cl = objcache_get(mjclmeta_cache, ocflags);
575 	if (cl == NULL) {
576 		++mbstat[mycpu->gd_cpuid].m_drops;
577 		return (FALSE);
578 	}
579 	m->m_flags |= M_CLCACHE;
580 	linkjcluster(m, cl, MJUMPAGESIZE);
581 	return (TRUE);
582 }
583 
584 /*
585  * Used for both the cluster and cluster PHDR caches.
586  *
587  * The mbuf may have lost its cluster due to sharing, deal
588  * with the situation by checking M_EXT.
589  */
590 static void
591 mbufcluster_dtor(void *obj, void *private)
592 {
593 	struct mbuf *m = obj;
594 	struct mbcluster *mcl;
595 
596 	if (m->m_flags & M_EXT) {
597 		KKASSERT((m->m_flags & M_EXT_CLUSTER) != 0);
598 		mcl = m->m_ext.ext_arg;
599 		KKASSERT(mcl->mcl_refs == 1);
600 		mcl->mcl_refs = 0;
601 		if (m->m_flags & M_EXT && m->m_ext.ext_size != MCLBYTES)
602 			objcache_put(mjclmeta_cache, mcl);
603 		else
604 			objcache_put(mclmeta_cache, mcl);
605 	}
606 }
607 
608 struct objcache_malloc_args mbuf_malloc_args = { MSIZE, M_MBUF };
609 struct objcache_malloc_args mclmeta_malloc_args =
610 	{ sizeof(struct mbcluster), M_MCLMETA };
611 
612 /* ARGSUSED*/
613 static void
614 mbinit(void *dummy)
615 {
616 	int mb_limit, cl_limit;
617 	int limit;
618 	int i;
619 
620 	/*
621 	 * Initialize statistics
622 	 */
623 	for (i = 0; i < ncpus; i++) {
624 		atomic_set_long_nonlocked(&mbstat[i].m_msize, MSIZE);
625 		atomic_set_long_nonlocked(&mbstat[i].m_mclbytes, MCLBYTES);
626 		atomic_set_long_nonlocked(&mbstat[i].m_mjumpagesize, MJUMPAGESIZE);
627 		atomic_set_long_nonlocked(&mbstat[i].m_minclsize, MINCLSIZE);
628 		atomic_set_long_nonlocked(&mbstat[i].m_mlen, MLEN);
629 		atomic_set_long_nonlocked(&mbstat[i].m_mhlen, MHLEN);
630 	}
631 
632 	/*
633 	 * Create objtect caches and save cluster limits, which will
634 	 * be used to adjust backing kmalloc pools' limit later.
635 	 */
636 
637 	mb_limit = cl_limit = 0;
638 
639 	limit = nmbufs;
640 	mbuf_cache = objcache_create("mbuf",
641 	    &limit, 0,
642 	    mbuf_ctor, NULL, NULL,
643 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
644 	mb_limit += limit;
645 
646 	limit = nmbufs;
647 	mbufphdr_cache = objcache_create("mbuf pkt hdr",
648 	    &limit, nmbufs / 4,
649 	    mbufphdr_ctor, NULL, NULL,
650 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
651 	mb_limit += limit;
652 
653 	cl_limit = nmbclusters;
654 	mclmeta_cache = objcache_create("cluster mbuf",
655 	    &cl_limit, 0,
656 	    mclmeta_ctor, mclmeta_dtor, NULL,
657 	    objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args);
658 
659 	cl_limit = nmbclusters;
660 	mjclmeta_cache = objcache_create("jcluster mbuf",
661 	    &cl_limit, 0,
662 	    mjclmeta_ctor, mclmeta_dtor, NULL,
663 	    objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args);
664 
665 	limit = nmbclusters;
666 	mbufcluster_cache = objcache_create("mbuf + cluster",
667 	    &limit, 0,
668 	    mbufcluster_ctor, mbufcluster_dtor, NULL,
669 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
670 	mb_limit += limit;
671 
672 	limit = nmbclusters;
673 	mbufphdrcluster_cache = objcache_create("mbuf pkt hdr + cluster",
674 	    &limit, nmbclusters / 16,
675 	    mbufphdrcluster_ctor, mbufcluster_dtor, NULL,
676 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
677 	mb_limit += limit;
678 
679 	limit = nmbclusters;
680 	mbufjcluster_cache = objcache_create("mbuf + jcluster",
681 	    &limit, 0,
682 	    mbufjcluster_ctor, mbufcluster_dtor, NULL,
683 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
684 	mb_limit += limit;
685 
686 	limit = nmbclusters;
687 	mbufphdrjcluster_cache = objcache_create("mbuf pkt hdr + jcluster",
688 	    &limit, nmbclusters / 16,
689 	    mbufphdrjcluster_ctor, mbufcluster_dtor, NULL,
690 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
691 	mb_limit += limit;
692 
693 	/*
694 	 * Adjust backing kmalloc pools' limit
695 	 *
696 	 * NOTE: We raise the limit by another 1/8 to take the effect
697 	 * of loosememuse into account.
698 	 */
699 	cl_limit += cl_limit / 8;
700 	kmalloc_raise_limit(mclmeta_malloc_args.mtype,
701 			    mclmeta_malloc_args.objsize * cl_limit);
702 	kmalloc_raise_limit(M_MBUFCL, MCLBYTES * cl_limit * 3/4 + MJUMPAGESIZE * cl_limit / 4);
703 	/*kmalloc_raise_limit(M_MBUFCL, MCLBYTES * cl_limit);*/
704 
705 	mb_limit += mb_limit / 8;
706 	kmalloc_raise_limit(mbuf_malloc_args.mtype,
707 			    mbuf_malloc_args.objsize * mb_limit);
708 }
709 
710 /*
711  * Return the number of references to this mbuf's data.  0 is returned
712  * if the mbuf is not M_EXT, a reference count is returned if it is
713  * M_EXT | M_EXT_CLUSTER, and 99 is returned if it is a special M_EXT.
714  */
715 int
716 m_sharecount(struct mbuf *m)
717 {
718 	switch (m->m_flags & (M_EXT | M_EXT_CLUSTER)) {
719 	case 0:
720 		return (0);
721 	case M_EXT:
722 		return (99);
723 	case M_EXT | M_EXT_CLUSTER:
724 		return (((struct mbcluster *)m->m_ext.ext_arg)->mcl_refs);
725 	}
726 	/* NOTREACHED */
727 	return (0);		/* to shut up compiler */
728 }
729 
730 /*
731  * change mbuf to new type
732  */
733 void
734 m_chtype(struct mbuf *m, int type)
735 {
736 	struct globaldata *gd = mycpu;
737 
738 	atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1);
739 	atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1);
740 	atomic_set_short_nonlocked(&m->m_type, type);
741 }
742 
743 static void
744 m_reclaim(void)
745 {
746 	struct domain *dp;
747 	struct protosw *pr;
748 
749 	kprintf("Debug: m_reclaim() called\n");
750 
751 	SLIST_FOREACH(dp, &domains, dom_next) {
752 		for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) {
753 			if (pr->pr_drain)
754 				(*pr->pr_drain)();
755 		}
756 	}
757 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_drain, 1);
758 }
759 
760 static __inline void
761 updatestats(struct mbuf *m, int type)
762 {
763 	struct globaldata *gd = mycpu;
764 
765 	m->m_type = type;
766 	mbuftrack(m);
767 #ifdef MBUF_DEBUG
768 	KASSERT(m->m_next == NULL, ("mbuf %p: bad m_next in get", m));
769 	KASSERT(m->m_nextpkt == NULL, ("mbuf %p: bad m_nextpkt in get", m));
770 #endif
771 
772 	atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1);
773 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1);
774 
775 }
776 
777 /*
778  * Allocate an mbuf.
779  */
780 struct mbuf *
781 m_get(int how, int type)
782 {
783 	struct mbuf *m;
784 	int ntries = 0;
785 	int ocf = MBTOM(how);
786 
787 retryonce:
788 
789 	m = objcache_get(mbuf_cache, ocf);
790 
791 	if (m == NULL) {
792 		if ((how & MB_TRYWAIT) && ntries++ == 0) {
793 			struct objcache *reclaimlist[] = {
794 				mbufphdr_cache,
795 				mbufcluster_cache,
796 				mbufphdrcluster_cache,
797 				mbufjcluster_cache,
798 				mbufphdrjcluster_cache
799 			};
800 			const int nreclaims = NELEM(reclaimlist);
801 
802 			if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
803 				m_reclaim();
804 			goto retryonce;
805 		}
806 		++mbstat[mycpu->gd_cpuid].m_drops;
807 		return (NULL);
808 	}
809 #ifdef MBUF_DEBUG
810 	KASSERT(m->m_data == m->m_dat, ("mbuf %p: bad m_data in get", m));
811 #endif
812 	m->m_len = 0;
813 
814 	updatestats(m, type);
815 	return (m);
816 }
817 
818 struct mbuf *
819 m_gethdr(int how, int type)
820 {
821 	struct mbuf *m;
822 	int ocf = MBTOM(how);
823 	int ntries = 0;
824 
825 retryonce:
826 
827 	m = objcache_get(mbufphdr_cache, ocf);
828 
829 	if (m == NULL) {
830 		if ((how & MB_TRYWAIT) && ntries++ == 0) {
831 			struct objcache *reclaimlist[] = {
832 				mbuf_cache,
833 				mbufcluster_cache, mbufphdrcluster_cache,
834 				mbufjcluster_cache, mbufphdrjcluster_cache
835 			};
836 			const int nreclaims = NELEM(reclaimlist);
837 
838 			if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
839 				m_reclaim();
840 			goto retryonce;
841 		}
842 		++mbstat[mycpu->gd_cpuid].m_drops;
843 		return (NULL);
844 	}
845 #ifdef MBUF_DEBUG
846 	KASSERT(m->m_data == m->m_pktdat, ("mbuf %p: bad m_data in get", m));
847 #endif
848 	m->m_len = 0;
849 	m->m_pkthdr.len = 0;
850 
851 	updatestats(m, type);
852 	return (m);
853 }
854 
855 /*
856  * Get a mbuf (not a mbuf cluster!) and zero it.
857  * Deprecated.
858  */
859 struct mbuf *
860 m_getclr(int how, int type)
861 {
862 	struct mbuf *m;
863 
864 	m = m_get(how, type);
865 	if (m != NULL)
866 		bzero(m->m_data, MLEN);
867 	return (m);
868 }
869 
870 struct mbuf *
871 m_getjcl(int how, short type, int flags, size_t size)
872 {
873 	struct mbuf *m = NULL;
874 	struct objcache *mbclc, *mbphclc;
875 	int ocflags = MBTOM(how);
876 	int ntries = 0;
877 
878 	switch (size) {
879 		case MCLBYTES:
880 			mbclc = mbufcluster_cache;
881 			mbphclc = mbufphdrcluster_cache;
882 			break;
883 		default:
884 			mbclc = mbufjcluster_cache;
885 			mbphclc = mbufphdrjcluster_cache;
886 			break;
887 	}
888 
889 retryonce:
890 
891 	if (flags & M_PKTHDR)
892 		m = objcache_get(mbphclc, ocflags);
893 	else
894 		m = objcache_get(mbclc, ocflags);
895 
896 	if (m == NULL) {
897 		if ((how & MB_TRYWAIT) && ntries++ == 0) {
898 			struct objcache *reclaimlist[1];
899 
900 			if (flags & M_PKTHDR)
901 				reclaimlist[0] = mbclc;
902 			else
903 				reclaimlist[0] = mbphclc;
904 			if (!objcache_reclaimlist(reclaimlist, 1, ocflags))
905 				m_reclaim();
906 			goto retryonce;
907 		}
908 		++mbstat[mycpu->gd_cpuid].m_drops;
909 		return (NULL);
910 	}
911 
912 #ifdef MBUF_DEBUG
913 	KASSERT(m->m_data == m->m_ext.ext_buf,
914 		("mbuf %p: bad m_data in get", m));
915 #endif
916 	m->m_type = type;
917 	m->m_len = 0;
918 	m->m_pkthdr.len = 0;	/* just do it unconditonally */
919 
920 	mbuftrack(m);
921 
922 	atomic_add_long_nonlocked(&mbtypes[mycpu->gd_cpuid][type], 1);
923 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
924 	return (m);
925 }
926 
927 /*
928  * Returns an mbuf with an attached cluster.
929  * Because many network drivers use this kind of buffers a lot, it is
930  * convenient to keep a small pool of free buffers of this kind.
931  * Even a small size such as 10 gives about 10% improvement in the
932  * forwarding rate in a bridge or router.
933  */
934 struct mbuf *
935 m_getcl(int how, short type, int flags)
936 {
937 	return (m_getjcl(how, type, flags, MCLBYTES));
938 }
939 
940 /*
941  * Allocate chain of requested length.
942  */
943 struct mbuf *
944 m_getc(int len, int how, int type)
945 {
946 	struct mbuf *n, *nfirst = NULL, **ntail = &nfirst;
947 	int nsize;
948 
949 	while (len > 0) {
950 		n = m_getl(len, how, type, 0, &nsize);
951 		if (n == NULL)
952 			goto failed;
953 		n->m_len = 0;
954 		*ntail = n;
955 		ntail = &n->m_next;
956 		len -= nsize;
957 	}
958 	return (nfirst);
959 
960 failed:
961 	m_freem(nfirst);
962 	return (NULL);
963 }
964 
965 /*
966  * Allocate len-worth of mbufs and/or mbuf clusters (whatever fits best)
967  * and return a pointer to the head of the allocated chain. If m0 is
968  * non-null, then we assume that it is a single mbuf or an mbuf chain to
969  * which we want len bytes worth of mbufs and/or clusters attached, and so
970  * if we succeed in allocating it, we will just return a pointer to m0.
971  *
972  * If we happen to fail at any point during the allocation, we will free
973  * up everything we have already allocated and return NULL.
974  *
975  * Deprecated.  Use m_getc() and m_cat() instead.
976  */
977 struct mbuf *
978 m_getm(struct mbuf *m0, int len, int type, int how)
979 {
980 	struct mbuf *nfirst;
981 
982 	nfirst = m_getc(len, how, type);
983 
984 	if (m0 != NULL) {
985 		m_last(m0)->m_next = nfirst;
986 		return (m0);
987 	}
988 
989 	return (nfirst);
990 }
991 
992 /*
993  * Adds a cluster to a normal mbuf, M_EXT is set on success.
994  * Deprecated.  Use m_getcl() instead.
995  */
996 void
997 m_mclget(struct mbuf *m, int how)
998 {
999 	struct mbcluster *mcl;
1000 
1001 	KKASSERT((m->m_flags & M_EXT) == 0);
1002 	mcl = objcache_get(mclmeta_cache, MBTOM(how));
1003 	if (mcl != NULL) {
1004 		linkcluster(m, mcl);
1005 		atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters,
1006 					  1);
1007 	} else {
1008 		++mbstat[mycpu->gd_cpuid].m_drops;
1009 	}
1010 }
1011 
1012 /*
1013  * Updates to mbcluster must be MPSAFE.  Only an entity which already has
1014  * a reference to the cluster can ref it, so we are in no danger of
1015  * racing an add with a subtract.  But the operation must still be atomic
1016  * since multiple entities may have a reference on the cluster.
1017  *
1018  * m_mclfree() is almost the same but it must contend with two entities
1019  * freeing the cluster at the same time.
1020  */
1021 static void
1022 m_mclref(void *arg)
1023 {
1024 	struct mbcluster *mcl = arg;
1025 
1026 	atomic_add_int(&mcl->mcl_refs, 1);
1027 }
1028 
1029 /*
1030  * When dereferencing a cluster we have to deal with a N->0 race, where
1031  * N entities free their references simultaniously.  To do this we use
1032  * atomic_fetchadd_int().
1033  */
1034 static void
1035 m_mclfree(void *arg)
1036 {
1037 	struct mbcluster *mcl = arg;
1038 
1039 	if (atomic_fetchadd_int(&mcl->mcl_refs, -1) == 1)
1040 		objcache_put(mclmeta_cache, mcl);
1041 }
1042 
1043 /*
1044  * Free a single mbuf and any associated external storage.  The successor,
1045  * if any, is returned.
1046  *
1047  * We do need to check non-first mbuf for m_aux, since some of existing
1048  * code does not call M_PREPEND properly.
1049  * (example: call to bpf_mtap from drivers)
1050  */
1051 
1052 #ifdef MBUF_DEBUG
1053 
1054 struct mbuf  *
1055 _m_free(struct mbuf *m, const char *func)
1056 
1057 #else
1058 
1059 struct mbuf *
1060 m_free(struct mbuf *m)
1061 
1062 #endif
1063 {
1064 	struct mbuf *n;
1065 	struct globaldata *gd = mycpu;
1066 
1067 	KASSERT(m->m_type != MT_FREE, ("freeing free mbuf %p", m));
1068 	KASSERT(M_TRAILINGSPACE(m) >= 0, ("overflowed mbuf %p", m));
1069 	atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1);
1070 
1071 	n = m->m_next;
1072 
1073 	/*
1074 	 * Make sure the mbuf is in constructed state before returning it
1075 	 * to the objcache.
1076 	 */
1077 	m->m_next = NULL;
1078 	mbufuntrack(m);
1079 #ifdef MBUF_DEBUG
1080 	m->m_hdr.mh_lastfunc = func;
1081 #endif
1082 #ifdef notyet
1083 	KKASSERT(m->m_nextpkt == NULL);
1084 #else
1085 	if (m->m_nextpkt != NULL) {
1086 		static int afewtimes = 10;
1087 
1088 		if (afewtimes-- > 0) {
1089 			kprintf("mfree: m->m_nextpkt != NULL\n");
1090 			print_backtrace(-1);
1091 		}
1092 		m->m_nextpkt = NULL;
1093 	}
1094 #endif
1095 	if (m->m_flags & M_PKTHDR) {
1096 		m_tag_delete_chain(m);		/* eliminate XXX JH */
1097 	}
1098 
1099 	m->m_flags &= (M_EXT | M_EXT_CLUSTER | M_CLCACHE | M_PHCACHE);
1100 
1101 	/*
1102 	 * Clean the M_PKTHDR state so we can return the mbuf to its original
1103 	 * cache.  This is based on the PHCACHE flag which tells us whether
1104 	 * the mbuf was originally allocated out of a packet-header cache
1105 	 * or a non-packet-header cache.
1106 	 */
1107 	if (m->m_flags & M_PHCACHE) {
1108 		m->m_flags |= M_PKTHDR;
1109 		m->m_pkthdr.rcvif = NULL;	/* eliminate XXX JH */
1110 		m->m_pkthdr.csum_flags = 0;	/* eliminate XXX JH */
1111 		m->m_pkthdr.fw_flags = 0;	/* eliminate XXX JH */
1112 		SLIST_INIT(&m->m_pkthdr.tags);
1113 	}
1114 
1115 	/*
1116 	 * Handle remaining flags combinations.  M_CLCACHE tells us whether
1117 	 * the mbuf was originally allocated from a cluster cache or not,
1118 	 * and is totally separate from whether the mbuf is currently
1119 	 * associated with a cluster.
1120 	 */
1121 	switch(m->m_flags & (M_CLCACHE | M_EXT | M_EXT_CLUSTER)) {
1122 	case M_CLCACHE | M_EXT | M_EXT_CLUSTER:
1123 		/*
1124 		 * mbuf+cluster cache case.  The mbuf was allocated from the
1125 		 * combined mbuf_cluster cache and can be returned to the
1126 		 * cache if the cluster hasn't been shared.
1127 		 */
1128 		if (m_sharecount(m) == 1) {
1129 			/*
1130 			 * The cluster has not been shared, we can just
1131 			 * reset the data pointer and return the mbuf
1132 			 * to the cluster cache.  Note that the reference
1133 			 * count is left intact (it is still associated with
1134 			 * an mbuf).
1135 			 */
1136 			m->m_data = m->m_ext.ext_buf;
1137 			if (m->m_flags & M_EXT && m->m_ext.ext_size != MCLBYTES) {
1138 				if (m->m_flags & M_PHCACHE)
1139 					objcache_put(mbufphdrjcluster_cache, m);
1140 				else
1141 					objcache_put(mbufjcluster_cache, m);
1142 			} else {
1143 				if (m->m_flags & M_PHCACHE)
1144 					objcache_put(mbufphdrcluster_cache, m);
1145 				else
1146 					objcache_put(mbufcluster_cache, m);
1147 			}
1148 			atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
1149 		} else {
1150 			/*
1151 			 * Hell.  Someone else has a ref on this cluster,
1152 			 * we have to disconnect it which means we can't
1153 			 * put it back into the mbufcluster_cache, we
1154 			 * have to destroy the mbuf.
1155 			 *
1156 			 * Other mbuf references to the cluster will typically
1157 			 * be M_EXT | M_EXT_CLUSTER but without M_CLCACHE.
1158 			 *
1159 			 * XXX we could try to connect another cluster to
1160 			 * it.
1161 			 */
1162 
1163 			m->m_ext.ext_free(m->m_ext.ext_arg);
1164 			m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
1165 			if (m->m_ext.ext_size == MCLBYTES) {
1166 				if (m->m_flags & M_PHCACHE)
1167 					objcache_dtor(mbufphdrcluster_cache, m);
1168 				else
1169 					objcache_dtor(mbufcluster_cache, m);
1170 			} else {
1171 				if (m->m_flags & M_PHCACHE)
1172 					objcache_dtor(mbufphdrjcluster_cache, m);
1173 				else
1174 					objcache_dtor(mbufjcluster_cache, m);
1175 			}
1176 		}
1177 		break;
1178 	case M_EXT | M_EXT_CLUSTER:
1179 		/*
1180 		 * Normal cluster associated with an mbuf that was allocated
1181 		 * from the normal mbuf pool rather then the cluster pool.
1182 		 * The cluster has to be independantly disassociated from the
1183 		 * mbuf.
1184 		 */
1185 		if (m_sharecount(m) == 1)
1186 			atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
1187 		/* fall through */
1188 	case M_EXT:
1189 		/*
1190 		 * Normal cluster association case, disconnect the cluster from
1191 		 * the mbuf.  The cluster may or may not be custom.
1192 		 */
1193 		m->m_ext.ext_free(m->m_ext.ext_arg);
1194 		m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
1195 		/* fall through */
1196 	case 0:
1197 		/*
1198 		 * return the mbuf to the mbuf cache.
1199 		 */
1200 		if (m->m_flags & M_PHCACHE) {
1201 			m->m_data = m->m_pktdat;
1202 			objcache_put(mbufphdr_cache, m);
1203 		} else {
1204 			m->m_data = m->m_dat;
1205 			objcache_put(mbuf_cache, m);
1206 		}
1207 		atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1);
1208 		break;
1209 	default:
1210 		if (!panicstr)
1211 			panic("bad mbuf flags %p %08x\n", m, m->m_flags);
1212 		break;
1213 	}
1214 	return (n);
1215 }
1216 
1217 #ifdef MBUF_DEBUG
1218 
1219 void
1220 _m_freem(struct mbuf *m, const char *func)
1221 {
1222 	while (m)
1223 		m = _m_free(m, func);
1224 }
1225 
1226 #else
1227 
1228 void
1229 m_freem(struct mbuf *m)
1230 {
1231 	while (m)
1232 		m = m_free(m);
1233 }
1234 
1235 #endif
1236 
1237 void
1238 m_extadd(struct mbuf *m, caddr_t buf, u_int size,  void (*reff)(void *),
1239     void (*freef)(void *), void *arg)
1240 {
1241 	m->m_ext.ext_arg = arg;
1242 	m->m_ext.ext_buf = buf;
1243 	m->m_ext.ext_ref = reff;
1244 	m->m_ext.ext_free = freef;
1245 	m->m_ext.ext_size = size;
1246 	reff(arg);
1247 	m->m_data = buf;
1248 	m->m_flags |= M_EXT;
1249 }
1250 
1251 /*
1252  * mbuf utility routines
1253  */
1254 
1255 /*
1256  * Lesser-used path for M_PREPEND: allocate new mbuf to prepend to chain and
1257  * copy junk along.
1258  */
1259 struct mbuf *
1260 m_prepend(struct mbuf *m, int len, int how)
1261 {
1262 	struct mbuf *mn;
1263 
1264 	if (m->m_flags & M_PKTHDR)
1265 	    mn = m_gethdr(how, m->m_type);
1266 	else
1267 	    mn = m_get(how, m->m_type);
1268 	if (mn == NULL) {
1269 		m_freem(m);
1270 		return (NULL);
1271 	}
1272 	if (m->m_flags & M_PKTHDR)
1273 		M_MOVE_PKTHDR(mn, m);
1274 	mn->m_next = m;
1275 	m = mn;
1276 	if (len < MHLEN)
1277 		MH_ALIGN(m, len);
1278 	m->m_len = len;
1279 	return (m);
1280 }
1281 
1282 /*
1283  * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
1284  * continuing for "len" bytes.  If len is M_COPYALL, copy to end of mbuf.
1285  * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller.
1286  * Note that the copy is read-only, because clusters are not copied,
1287  * only their reference counts are incremented.
1288  */
1289 struct mbuf *
1290 m_copym(const struct mbuf *m, int off0, int len, int wait)
1291 {
1292 	struct mbuf *n, **np;
1293 	int off = off0;
1294 	struct mbuf *top;
1295 	int copyhdr = 0;
1296 
1297 	KASSERT(off >= 0, ("m_copym, negative off %d", off));
1298 	KASSERT(len >= 0, ("m_copym, negative len %d", len));
1299 	if (off == 0 && (m->m_flags & M_PKTHDR))
1300 		copyhdr = 1;
1301 	while (off > 0) {
1302 		KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
1303 		if (off < m->m_len)
1304 			break;
1305 		off -= m->m_len;
1306 		m = m->m_next;
1307 	}
1308 	np = &top;
1309 	top = NULL;
1310 	while (len > 0) {
1311 		if (m == NULL) {
1312 			KASSERT(len == M_COPYALL,
1313 			    ("m_copym, length > size of mbuf chain"));
1314 			break;
1315 		}
1316 		/*
1317 		 * Because we are sharing any cluster attachment below,
1318 		 * be sure to get an mbuf that does not have a cluster
1319 		 * associated with it.
1320 		 */
1321 		if (copyhdr)
1322 			n = m_gethdr(wait, m->m_type);
1323 		else
1324 			n = m_get(wait, m->m_type);
1325 		*np = n;
1326 		if (n == NULL)
1327 			goto nospace;
1328 		if (copyhdr) {
1329 			if (!m_dup_pkthdr(n, m, wait))
1330 				goto nospace;
1331 			if (len == M_COPYALL)
1332 				n->m_pkthdr.len -= off0;
1333 			else
1334 				n->m_pkthdr.len = len;
1335 			copyhdr = 0;
1336 		}
1337 		n->m_len = min(len, m->m_len - off);
1338 		if (m->m_flags & M_EXT) {
1339 			KKASSERT((n->m_flags & M_EXT) == 0);
1340 			n->m_data = m->m_data + off;
1341 			m->m_ext.ext_ref(m->m_ext.ext_arg);
1342 			n->m_ext = m->m_ext;
1343 			n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1344 		} else {
1345 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1346 			    (unsigned)n->m_len);
1347 		}
1348 		if (len != M_COPYALL)
1349 			len -= n->m_len;
1350 		off = 0;
1351 		m = m->m_next;
1352 		np = &n->m_next;
1353 	}
1354 	if (top == NULL)
1355 		atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1356 	return (top);
1357 nospace:
1358 	m_freem(top);
1359 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1360 	return (NULL);
1361 }
1362 
1363 /*
1364  * Copy an entire packet, including header (which must be present).
1365  * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
1366  * Note that the copy is read-only, because clusters are not copied,
1367  * only their reference counts are incremented.
1368  * Preserve alignment of the first mbuf so if the creator has left
1369  * some room at the beginning (e.g. for inserting protocol headers)
1370  * the copies also have the room available.
1371  */
1372 struct mbuf *
1373 m_copypacket(struct mbuf *m, int how)
1374 {
1375 	struct mbuf *top, *n, *o;
1376 
1377 	n = m_gethdr(how, m->m_type);
1378 	top = n;
1379 	if (!n)
1380 		goto nospace;
1381 
1382 	if (!m_dup_pkthdr(n, m, how))
1383 		goto nospace;
1384 	n->m_len = m->m_len;
1385 	if (m->m_flags & M_EXT) {
1386 		KKASSERT((n->m_flags & M_EXT) == 0);
1387 		n->m_data = m->m_data;
1388 		m->m_ext.ext_ref(m->m_ext.ext_arg);
1389 		n->m_ext = m->m_ext;
1390 		n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1391 	} else {
1392 		n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
1393 		bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1394 	}
1395 
1396 	m = m->m_next;
1397 	while (m) {
1398 		o = m_get(how, m->m_type);
1399 		if (!o)
1400 			goto nospace;
1401 
1402 		n->m_next = o;
1403 		n = n->m_next;
1404 
1405 		n->m_len = m->m_len;
1406 		if (m->m_flags & M_EXT) {
1407 			KKASSERT((n->m_flags & M_EXT) == 0);
1408 			n->m_data = m->m_data;
1409 			m->m_ext.ext_ref(m->m_ext.ext_arg);
1410 			n->m_ext = m->m_ext;
1411 			n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1412 		} else {
1413 			bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1414 		}
1415 
1416 		m = m->m_next;
1417 	}
1418 	return top;
1419 nospace:
1420 	m_freem(top);
1421 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1422 	return (NULL);
1423 }
1424 
1425 /*
1426  * Copy data from an mbuf chain starting "off" bytes from the beginning,
1427  * continuing for "len" bytes, into the indicated buffer.
1428  */
1429 void
1430 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
1431 {
1432 	unsigned count;
1433 
1434 	KASSERT(off >= 0, ("m_copydata, negative off %d", off));
1435 	KASSERT(len >= 0, ("m_copydata, negative len %d", len));
1436 	while (off > 0) {
1437 		KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
1438 		if (off < m->m_len)
1439 			break;
1440 		off -= m->m_len;
1441 		m = m->m_next;
1442 	}
1443 	while (len > 0) {
1444 		KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
1445 		count = min(m->m_len - off, len);
1446 		bcopy(mtod(m, caddr_t) + off, cp, count);
1447 		len -= count;
1448 		cp += count;
1449 		off = 0;
1450 		m = m->m_next;
1451 	}
1452 }
1453 
1454 /*
1455  * Copy a packet header mbuf chain into a completely new chain, including
1456  * copying any mbuf clusters.  Use this instead of m_copypacket() when
1457  * you need a writable copy of an mbuf chain.
1458  */
1459 struct mbuf *
1460 m_dup(struct mbuf *m, int how)
1461 {
1462 	struct mbuf **p, *top = NULL;
1463 	int remain, moff, nsize;
1464 
1465 	/* Sanity check */
1466 	if (m == NULL)
1467 		return (NULL);
1468 	KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __func__));
1469 
1470 	/* While there's more data, get a new mbuf, tack it on, and fill it */
1471 	remain = m->m_pkthdr.len;
1472 	moff = 0;
1473 	p = &top;
1474 	while (remain > 0 || top == NULL) {	/* allow m->m_pkthdr.len == 0 */
1475 		struct mbuf *n;
1476 
1477 		/* Get the next new mbuf */
1478 		n = m_getl(remain, how, m->m_type, top == NULL ? M_PKTHDR : 0,
1479 			   &nsize);
1480 		if (n == NULL)
1481 			goto nospace;
1482 		if (top == NULL)
1483 			if (!m_dup_pkthdr(n, m, how))
1484 				goto nospace0;
1485 
1486 		/* Link it into the new chain */
1487 		*p = n;
1488 		p = &n->m_next;
1489 
1490 		/* Copy data from original mbuf(s) into new mbuf */
1491 		n->m_len = 0;
1492 		while (n->m_len < nsize && m != NULL) {
1493 			int chunk = min(nsize - n->m_len, m->m_len - moff);
1494 
1495 			bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1496 			moff += chunk;
1497 			n->m_len += chunk;
1498 			remain -= chunk;
1499 			if (moff == m->m_len) {
1500 				m = m->m_next;
1501 				moff = 0;
1502 			}
1503 		}
1504 
1505 		/* Check correct total mbuf length */
1506 		KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
1507 			("%s: bogus m_pkthdr.len", __func__));
1508 	}
1509 	return (top);
1510 
1511 nospace:
1512 	m_freem(top);
1513 nospace0:
1514 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1515 	return (NULL);
1516 }
1517 
1518 /*
1519  * Copy the non-packet mbuf data chain into a new set of mbufs, including
1520  * copying any mbuf clusters.  This is typically used to realign a data
1521  * chain by nfs_realign().
1522  *
1523  * The original chain is left intact.  how should be MB_WAIT or MB_DONTWAIT
1524  * and NULL can be returned if MB_DONTWAIT is passed.
1525  *
1526  * Be careful to use cluster mbufs, a large mbuf chain converted to non
1527  * cluster mbufs can exhaust our supply of mbufs.
1528  */
1529 struct mbuf *
1530 m_dup_data(struct mbuf *m, int how)
1531 {
1532 	struct mbuf **p, *n, *top = NULL;
1533 	int mlen, moff, chunk, gsize, nsize;
1534 
1535 	/*
1536 	 * Degenerate case
1537 	 */
1538 	if (m == NULL)
1539 		return (NULL);
1540 
1541 	/*
1542 	 * Optimize the mbuf allocation but do not get too carried away.
1543 	 */
1544 	if (m->m_next || m->m_len > MLEN)
1545 		if (m->m_flags & M_EXT && m->m_ext.ext_size == MCLBYTES)
1546 			gsize = MCLBYTES;
1547 		else
1548 			gsize = MJUMPAGESIZE;
1549 	else
1550 		gsize = MLEN;
1551 
1552 	/* Chain control */
1553 	p = &top;
1554 	n = NULL;
1555 	nsize = 0;
1556 
1557 	/*
1558 	 * Scan the mbuf chain until nothing is left, the new mbuf chain
1559 	 * will be allocated on the fly as needed.
1560 	 */
1561 	while (m) {
1562 		mlen = m->m_len;
1563 		moff = 0;
1564 
1565 		while (mlen) {
1566 			KKASSERT(m->m_type == MT_DATA);
1567 			if (n == NULL) {
1568 				n = m_getl(gsize, how, MT_DATA, 0, &nsize);
1569 				n->m_len = 0;
1570 				if (n == NULL)
1571 					goto nospace;
1572 				*p = n;
1573 				p = &n->m_next;
1574 			}
1575 			chunk = imin(mlen, nsize);
1576 			bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1577 			mlen -= chunk;
1578 			moff += chunk;
1579 			n->m_len += chunk;
1580 			nsize -= chunk;
1581 			if (nsize == 0)
1582 				n = NULL;
1583 		}
1584 		m = m->m_next;
1585 	}
1586 	*p = NULL;
1587 	return(top);
1588 nospace:
1589 	*p = NULL;
1590 	m_freem(top);
1591 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1592 	return (NULL);
1593 }
1594 
1595 /*
1596  * Concatenate mbuf chain n to m.
1597  * Both chains must be of the same type (e.g. MT_DATA).
1598  * Any m_pkthdr is not updated.
1599  */
1600 void
1601 m_cat(struct mbuf *m, struct mbuf *n)
1602 {
1603 	m = m_last(m);
1604 	while (n) {
1605 		if (m->m_flags & M_EXT ||
1606 		    m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) {
1607 			/* just join the two chains */
1608 			m->m_next = n;
1609 			return;
1610 		}
1611 		/* splat the data from one into the other */
1612 		bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1613 		    (u_int)n->m_len);
1614 		m->m_len += n->m_len;
1615 		n = m_free(n);
1616 	}
1617 }
1618 
1619 void
1620 m_adj(struct mbuf *mp, int req_len)
1621 {
1622 	int len = req_len;
1623 	struct mbuf *m;
1624 	int count;
1625 
1626 	if ((m = mp) == NULL)
1627 		return;
1628 	if (len >= 0) {
1629 		/*
1630 		 * Trim from head.
1631 		 */
1632 		while (m != NULL && len > 0) {
1633 			if (m->m_len <= len) {
1634 				len -= m->m_len;
1635 				m->m_len = 0;
1636 				m = m->m_next;
1637 			} else {
1638 				m->m_len -= len;
1639 				m->m_data += len;
1640 				len = 0;
1641 			}
1642 		}
1643 		m = mp;
1644 		if (mp->m_flags & M_PKTHDR)
1645 			m->m_pkthdr.len -= (req_len - len);
1646 	} else {
1647 		/*
1648 		 * Trim from tail.  Scan the mbuf chain,
1649 		 * calculating its length and finding the last mbuf.
1650 		 * If the adjustment only affects this mbuf, then just
1651 		 * adjust and return.  Otherwise, rescan and truncate
1652 		 * after the remaining size.
1653 		 */
1654 		len = -len;
1655 		count = 0;
1656 		for (;;) {
1657 			count += m->m_len;
1658 			if (m->m_next == NULL)
1659 				break;
1660 			m = m->m_next;
1661 		}
1662 		if (m->m_len >= len) {
1663 			m->m_len -= len;
1664 			if (mp->m_flags & M_PKTHDR)
1665 				mp->m_pkthdr.len -= len;
1666 			return;
1667 		}
1668 		count -= len;
1669 		if (count < 0)
1670 			count = 0;
1671 		/*
1672 		 * Correct length for chain is "count".
1673 		 * Find the mbuf with last data, adjust its length,
1674 		 * and toss data from remaining mbufs on chain.
1675 		 */
1676 		m = mp;
1677 		if (m->m_flags & M_PKTHDR)
1678 			m->m_pkthdr.len = count;
1679 		for (; m; m = m->m_next) {
1680 			if (m->m_len >= count) {
1681 				m->m_len = count;
1682 				break;
1683 			}
1684 			count -= m->m_len;
1685 		}
1686 		while (m->m_next)
1687 			(m = m->m_next) ->m_len = 0;
1688 	}
1689 }
1690 
1691 /*
1692  * Set the m_data pointer of a newly-allocated mbuf
1693  * to place an object of the specified size at the
1694  * end of the mbuf, longword aligned.
1695  */
1696 void
1697 m_align(struct mbuf *m, int len)
1698 {
1699 	int adjust;
1700 
1701 	if (m->m_flags & M_EXT)
1702 		adjust = m->m_ext.ext_size - len;
1703 	else if (m->m_flags & M_PKTHDR)
1704 		adjust = MHLEN - len;
1705 	else
1706 		adjust = MLEN - len;
1707 	m->m_data += adjust &~ (sizeof(long)-1);
1708 }
1709 
1710 /*
1711  * Rearrange an mbuf chain so that len bytes are contiguous
1712  * and in the data area of an mbuf (so that mtod will work for a structure
1713  * of size len).  Returns the resulting mbuf chain on success, frees it and
1714  * returns null on failure.  If there is room, it will add up to
1715  * max_protohdr-len extra bytes to the contiguous region in an attempt to
1716  * avoid being called next time.
1717  */
1718 struct mbuf *
1719 m_pullup(struct mbuf *n, int len)
1720 {
1721 	struct mbuf *m;
1722 	int count;
1723 	int space;
1724 
1725 	/*
1726 	 * If first mbuf has no cluster, and has room for len bytes
1727 	 * without shifting current data, pullup into it,
1728 	 * otherwise allocate a new mbuf to prepend to the chain.
1729 	 */
1730 	if (!(n->m_flags & M_EXT) &&
1731 	    n->m_data + len < &n->m_dat[MLEN] &&
1732 	    n->m_next) {
1733 		if (n->m_len >= len)
1734 			return (n);
1735 		m = n;
1736 		n = n->m_next;
1737 		len -= m->m_len;
1738 	} else {
1739 		if (len > MHLEN)
1740 			goto bad;
1741 		if (n->m_flags & M_PKTHDR)
1742 			m = m_gethdr(MB_DONTWAIT, n->m_type);
1743 		else
1744 			m = m_get(MB_DONTWAIT, n->m_type);
1745 		if (m == NULL)
1746 			goto bad;
1747 		m->m_len = 0;
1748 		if (n->m_flags & M_PKTHDR)
1749 			M_MOVE_PKTHDR(m, n);
1750 	}
1751 	space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
1752 	do {
1753 		count = min(min(max(len, max_protohdr), space), n->m_len);
1754 		bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1755 		  (unsigned)count);
1756 		len -= count;
1757 		m->m_len += count;
1758 		n->m_len -= count;
1759 		space -= count;
1760 		if (n->m_len)
1761 			n->m_data += count;
1762 		else
1763 			n = m_free(n);
1764 	} while (len > 0 && n);
1765 	if (len > 0) {
1766 		m_free(m);
1767 		goto bad;
1768 	}
1769 	m->m_next = n;
1770 	return (m);
1771 bad:
1772 	m_freem(n);
1773 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1774 	return (NULL);
1775 }
1776 
1777 /*
1778  * Partition an mbuf chain in two pieces, returning the tail --
1779  * all but the first len0 bytes.  In case of failure, it returns NULL and
1780  * attempts to restore the chain to its original state.
1781  *
1782  * Note that the resulting mbufs might be read-only, because the new
1783  * mbuf can end up sharing an mbuf cluster with the original mbuf if
1784  * the "breaking point" happens to lie within a cluster mbuf. Use the
1785  * M_WRITABLE() macro to check for this case.
1786  */
1787 struct mbuf *
1788 m_split(struct mbuf *m0, int len0, int wait)
1789 {
1790 	struct mbuf *m, *n;
1791 	unsigned len = len0, remain;
1792 
1793 	for (m = m0; m && len > m->m_len; m = m->m_next)
1794 		len -= m->m_len;
1795 	if (m == NULL)
1796 		return (NULL);
1797 	remain = m->m_len - len;
1798 	if (m0->m_flags & M_PKTHDR) {
1799 		n = m_gethdr(wait, m0->m_type);
1800 		if (n == NULL)
1801 			return (NULL);
1802 		n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1803 		n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1804 		m0->m_pkthdr.len = len0;
1805 		if (m->m_flags & M_EXT)
1806 			goto extpacket;
1807 		if (remain > MHLEN) {
1808 			/* m can't be the lead packet */
1809 			MH_ALIGN(n, 0);
1810 			n->m_next = m_split(m, len, wait);
1811 			if (n->m_next == NULL) {
1812 				m_free(n);
1813 				return (NULL);
1814 			} else {
1815 				n->m_len = 0;
1816 				return (n);
1817 			}
1818 		} else
1819 			MH_ALIGN(n, remain);
1820 	} else if (remain == 0) {
1821 		n = m->m_next;
1822 		m->m_next = 0;
1823 		return (n);
1824 	} else {
1825 		n = m_get(wait, m->m_type);
1826 		if (n == NULL)
1827 			return (NULL);
1828 		M_ALIGN(n, remain);
1829 	}
1830 extpacket:
1831 	if (m->m_flags & M_EXT) {
1832 		KKASSERT((n->m_flags & M_EXT) == 0);
1833 		n->m_data = m->m_data + len;
1834 		m->m_ext.ext_ref(m->m_ext.ext_arg);
1835 		n->m_ext = m->m_ext;
1836 		n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1837 	} else {
1838 		bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1839 	}
1840 	n->m_len = remain;
1841 	m->m_len = len;
1842 	n->m_next = m->m_next;
1843 	m->m_next = 0;
1844 	return (n);
1845 }
1846 
1847 /*
1848  * Routine to copy from device local memory into mbufs.
1849  * Note: "offset" is ill-defined and always called as 0, so ignore it.
1850  */
1851 struct mbuf *
1852 m_devget(char *buf, int len, int offset, struct ifnet *ifp,
1853     void (*copy)(volatile const void *from, volatile void *to, size_t length))
1854 {
1855 	struct mbuf *m, *mfirst = NULL, **mtail;
1856 	int nsize, flags;
1857 
1858 	if (copy == NULL)
1859 		copy = bcopy;
1860 	mtail = &mfirst;
1861 	flags = M_PKTHDR;
1862 
1863 	while (len > 0) {
1864 		m = m_getl(len, MB_DONTWAIT, MT_DATA, flags, &nsize);
1865 		if (m == NULL) {
1866 			m_freem(mfirst);
1867 			return (NULL);
1868 		}
1869 		m->m_len = min(len, nsize);
1870 
1871 		if (flags & M_PKTHDR) {
1872 			if (len + max_linkhdr <= nsize)
1873 				m->m_data += max_linkhdr;
1874 			m->m_pkthdr.rcvif = ifp;
1875 			m->m_pkthdr.len = len;
1876 			flags = 0;
1877 		}
1878 
1879 		copy(buf, m->m_data, (unsigned)m->m_len);
1880 		buf += m->m_len;
1881 		len -= m->m_len;
1882 		*mtail = m;
1883 		mtail = &m->m_next;
1884 	}
1885 
1886 	return (mfirst);
1887 }
1888 
1889 /*
1890  * Routine to pad mbuf to the specified length 'padto'.
1891  */
1892 int
1893 m_devpad(struct mbuf *m, int padto)
1894 {
1895 	struct mbuf *last = NULL;
1896 	int padlen;
1897 
1898 	if (padto <= m->m_pkthdr.len)
1899 		return 0;
1900 
1901 	padlen = padto - m->m_pkthdr.len;
1902 
1903 	/* if there's only the packet-header and we can pad there, use it. */
1904 	if (m->m_pkthdr.len == m->m_len && M_TRAILINGSPACE(m) >= padlen) {
1905 		last = m;
1906 	} else {
1907 		/*
1908 		 * Walk packet chain to find last mbuf. We will either
1909 		 * pad there, or append a new mbuf and pad it
1910 		 */
1911 		for (last = m; last->m_next != NULL; last = last->m_next)
1912 			; /* EMPTY */
1913 
1914 		/* `last' now points to last in chain. */
1915 		if (M_TRAILINGSPACE(last) < padlen) {
1916 			struct mbuf *n;
1917 
1918 			/* Allocate new empty mbuf, pad it.  Compact later. */
1919 			MGET(n, MB_DONTWAIT, MT_DATA);
1920 			if (n == NULL)
1921 				return ENOBUFS;
1922 			n->m_len = 0;
1923 			last->m_next = n;
1924 			last = n;
1925 		}
1926 	}
1927 	KKASSERT(M_TRAILINGSPACE(last) >= padlen);
1928 	KKASSERT(M_WRITABLE(last));
1929 
1930 	/* Now zero the pad area */
1931 	bzero(mtod(last, char *) + last->m_len, padlen);
1932 	last->m_len += padlen;
1933 	m->m_pkthdr.len += padlen;
1934 	return 0;
1935 }
1936 
1937 /*
1938  * Copy data from a buffer back into the indicated mbuf chain,
1939  * starting "off" bytes from the beginning, extending the mbuf
1940  * chain if necessary.
1941  */
1942 void
1943 m_copyback(struct mbuf *m0, int off, int len, caddr_t cp)
1944 {
1945 	int mlen;
1946 	struct mbuf *m = m0, *n;
1947 	int totlen = 0;
1948 
1949 	if (m0 == NULL)
1950 		return;
1951 	while (off > (mlen = m->m_len)) {
1952 		off -= mlen;
1953 		totlen += mlen;
1954 		if (m->m_next == NULL) {
1955 			n = m_getclr(MB_DONTWAIT, m->m_type);
1956 			if (n == NULL)
1957 				goto out;
1958 			n->m_len = min(MLEN, len + off);
1959 			m->m_next = n;
1960 		}
1961 		m = m->m_next;
1962 	}
1963 	while (len > 0) {
1964 		mlen = min (m->m_len - off, len);
1965 		bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen);
1966 		cp += mlen;
1967 		len -= mlen;
1968 		mlen += off;
1969 		off = 0;
1970 		totlen += mlen;
1971 		if (len == 0)
1972 			break;
1973 		if (m->m_next == NULL) {
1974 			n = m_get(MB_DONTWAIT, m->m_type);
1975 			if (n == NULL)
1976 				break;
1977 			n->m_len = min(MLEN, len);
1978 			m->m_next = n;
1979 		}
1980 		m = m->m_next;
1981 	}
1982 out:	if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1983 		m->m_pkthdr.len = totlen;
1984 }
1985 
1986 /*
1987  * Append the specified data to the indicated mbuf chain,
1988  * Extend the mbuf chain if the new data does not fit in
1989  * existing space.
1990  *
1991  * Return 1 if able to complete the job; otherwise 0.
1992  */
1993 int
1994 m_append(struct mbuf *m0, int len, c_caddr_t cp)
1995 {
1996 	struct mbuf *m, *n;
1997 	int remainder, space;
1998 
1999 	for (m = m0; m->m_next != NULL; m = m->m_next)
2000 		;
2001 	remainder = len;
2002 	space = M_TRAILINGSPACE(m);
2003 	if (space > 0) {
2004 		/*
2005 		 * Copy into available space.
2006 		 */
2007 		if (space > remainder)
2008 			space = remainder;
2009 		bcopy(cp, mtod(m, caddr_t) + m->m_len, space);
2010 		m->m_len += space;
2011 		cp += space, remainder -= space;
2012 	}
2013 	while (remainder > 0) {
2014 		/*
2015 		 * Allocate a new mbuf; could check space
2016 		 * and allocate a cluster instead.
2017 		 */
2018 		n = m_get(MB_DONTWAIT, m->m_type);
2019 		if (n == NULL)
2020 			break;
2021 		n->m_len = min(MLEN, remainder);
2022 		bcopy(cp, mtod(n, caddr_t), n->m_len);
2023 		cp += n->m_len, remainder -= n->m_len;
2024 		m->m_next = n;
2025 		m = n;
2026 	}
2027 	if (m0->m_flags & M_PKTHDR)
2028 		m0->m_pkthdr.len += len - remainder;
2029 	return (remainder == 0);
2030 }
2031 
2032 /*
2033  * Apply function f to the data in an mbuf chain starting "off" bytes from
2034  * the beginning, continuing for "len" bytes.
2035  */
2036 int
2037 m_apply(struct mbuf *m, int off, int len,
2038     int (*f)(void *, void *, u_int), void *arg)
2039 {
2040 	u_int count;
2041 	int rval;
2042 
2043 	KASSERT(off >= 0, ("m_apply, negative off %d", off));
2044 	KASSERT(len >= 0, ("m_apply, negative len %d", len));
2045 	while (off > 0) {
2046 		KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
2047 		if (off < m->m_len)
2048 			break;
2049 		off -= m->m_len;
2050 		m = m->m_next;
2051 	}
2052 	while (len > 0) {
2053 		KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
2054 		count = min(m->m_len - off, len);
2055 		rval = (*f)(arg, mtod(m, caddr_t) + off, count);
2056 		if (rval)
2057 			return (rval);
2058 		len -= count;
2059 		off = 0;
2060 		m = m->m_next;
2061 	}
2062 	return (0);
2063 }
2064 
2065 /*
2066  * Return a pointer to mbuf/offset of location in mbuf chain.
2067  */
2068 struct mbuf *
2069 m_getptr(struct mbuf *m, int loc, int *off)
2070 {
2071 
2072 	while (loc >= 0) {
2073 		/* Normal end of search. */
2074 		if (m->m_len > loc) {
2075 			*off = loc;
2076 			return (m);
2077 		} else {
2078 			loc -= m->m_len;
2079 			if (m->m_next == NULL) {
2080 				if (loc == 0) {
2081 					/* Point at the end of valid data. */
2082 					*off = m->m_len;
2083 					return (m);
2084 				}
2085 				return (NULL);
2086 			}
2087 			m = m->m_next;
2088 		}
2089 	}
2090 	return (NULL);
2091 }
2092 
2093 void
2094 m_print(const struct mbuf *m)
2095 {
2096 	int len;
2097 	const struct mbuf *m2;
2098 
2099 	len = m->m_pkthdr.len;
2100 	m2 = m;
2101 	while (len) {
2102 		kprintf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-");
2103 		len -= m2->m_len;
2104 		m2 = m2->m_next;
2105 	}
2106 	return;
2107 }
2108 
2109 /*
2110  * "Move" mbuf pkthdr from "from" to "to".
2111  * "from" must have M_PKTHDR set, and "to" must be empty.
2112  */
2113 void
2114 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
2115 {
2116 	KASSERT((to->m_flags & M_PKTHDR), ("m_move_pkthdr: not packet header"));
2117 
2118 	to->m_flags |= from->m_flags & M_COPYFLAGS;
2119 	to->m_pkthdr = from->m_pkthdr;		/* especially tags */
2120 	SLIST_INIT(&from->m_pkthdr.tags);	/* purge tags from src */
2121 }
2122 
2123 /*
2124  * Duplicate "from"'s mbuf pkthdr in "to".
2125  * "from" must have M_PKTHDR set, and "to" must be empty.
2126  * In particular, this does a deep copy of the packet tags.
2127  */
2128 int
2129 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
2130 {
2131 	KASSERT((to->m_flags & M_PKTHDR), ("m_dup_pkthdr: not packet header"));
2132 
2133 	to->m_flags = (from->m_flags & M_COPYFLAGS) |
2134 		      (to->m_flags & ~M_COPYFLAGS);
2135 	to->m_pkthdr = from->m_pkthdr;
2136 	SLIST_INIT(&to->m_pkthdr.tags);
2137 	return (m_tag_copy_chain(to, from, how));
2138 }
2139 
2140 /*
2141  * Defragment a mbuf chain, returning the shortest possible
2142  * chain of mbufs and clusters.  If allocation fails and
2143  * this cannot be completed, NULL will be returned, but
2144  * the passed in chain will be unchanged.  Upon success,
2145  * the original chain will be freed, and the new chain
2146  * will be returned.
2147  *
2148  * If a non-packet header is passed in, the original
2149  * mbuf (chain?) will be returned unharmed.
2150  *
2151  * m_defrag_nofree doesn't free the passed in mbuf.
2152  */
2153 struct mbuf *
2154 m_defrag(struct mbuf *m0, int how)
2155 {
2156 	struct mbuf *m_new;
2157 
2158 	if ((m_new = m_defrag_nofree(m0, how)) == NULL)
2159 		return (NULL);
2160 	if (m_new != m0)
2161 		m_freem(m0);
2162 	return (m_new);
2163 }
2164 
2165 struct mbuf *
2166 m_defrag_nofree(struct mbuf *m0, int how)
2167 {
2168 	struct mbuf	*m_new = NULL, *m_final = NULL;
2169 	int		progress = 0, length, nsize;
2170 
2171 	if (!(m0->m_flags & M_PKTHDR))
2172 		return (m0);
2173 
2174 #ifdef MBUF_STRESS_TEST
2175 	if (m_defragrandomfailures) {
2176 		int temp = karc4random() & 0xff;
2177 		if (temp == 0xba)
2178 			goto nospace;
2179 	}
2180 #endif
2181 
2182 	m_final = m_getl(m0->m_pkthdr.len, how, MT_DATA, M_PKTHDR, &nsize);
2183 	if (m_final == NULL)
2184 		goto nospace;
2185 	m_final->m_len = 0;	/* in case m0->m_pkthdr.len is zero */
2186 
2187 	if (m_dup_pkthdr(m_final, m0, how) == 0)
2188 		goto nospace;
2189 
2190 	m_new = m_final;
2191 
2192 	while (progress < m0->m_pkthdr.len) {
2193 		length = m0->m_pkthdr.len - progress;
2194 		if (length > MCLBYTES)
2195 			length = MCLBYTES;
2196 
2197 		if (m_new == NULL) {
2198 			m_new = m_getl(length, how, MT_DATA, 0, &nsize);
2199 			if (m_new == NULL)
2200 				goto nospace;
2201 		}
2202 
2203 		m_copydata(m0, progress, length, mtod(m_new, caddr_t));
2204 		progress += length;
2205 		m_new->m_len = length;
2206 		if (m_new != m_final)
2207 			m_cat(m_final, m_new);
2208 		m_new = NULL;
2209 	}
2210 	if (m0->m_next == NULL)
2211 		m_defraguseless++;
2212 	m_defragpackets++;
2213 	m_defragbytes += m_final->m_pkthdr.len;
2214 	return (m_final);
2215 nospace:
2216 	m_defragfailure++;
2217 	if (m_new)
2218 		m_free(m_new);
2219 	m_freem(m_final);
2220 	return (NULL);
2221 }
2222 
2223 /*
2224  * Move data from uio into mbufs.
2225  */
2226 struct mbuf *
2227 m_uiomove(struct uio *uio)
2228 {
2229 	struct mbuf *m;			/* current working mbuf */
2230 	struct mbuf *head = NULL;	/* result mbuf chain */
2231 	struct mbuf **mp = &head;
2232 	int flags = M_PKTHDR;
2233 	int nsize;
2234 	int error;
2235 	int resid;
2236 
2237 	do {
2238 		if (uio->uio_resid > INT_MAX)
2239 			resid = INT_MAX;
2240 		else
2241 			resid = (int)uio->uio_resid;
2242 		m = m_getl(resid, MB_WAIT, MT_DATA, flags, &nsize);
2243 		if (flags) {
2244 			m->m_pkthdr.len = 0;
2245 			/* Leave room for protocol headers. */
2246 			if (resid < MHLEN)
2247 				MH_ALIGN(m, resid);
2248 			flags = 0;
2249 		}
2250 		m->m_len = imin(nsize, resid);
2251 		error = uiomove(mtod(m, caddr_t), m->m_len, uio);
2252 		if (error) {
2253 			m_free(m);
2254 			goto failed;
2255 		}
2256 		*mp = m;
2257 		mp = &m->m_next;
2258 		head->m_pkthdr.len += m->m_len;
2259 	} while (uio->uio_resid > 0);
2260 
2261 	return (head);
2262 
2263 failed:
2264 	m_freem(head);
2265 	return (NULL);
2266 }
2267 
2268 struct mbuf *
2269 m_last(struct mbuf *m)
2270 {
2271 	while (m->m_next)
2272 		m = m->m_next;
2273 	return (m);
2274 }
2275 
2276 /*
2277  * Return the number of bytes in an mbuf chain.
2278  * If lastm is not NULL, also return the last mbuf.
2279  */
2280 u_int
2281 m_lengthm(struct mbuf *m, struct mbuf **lastm)
2282 {
2283 	u_int len = 0;
2284 	struct mbuf *prev = m;
2285 
2286 	while (m) {
2287 		len += m->m_len;
2288 		prev = m;
2289 		m = m->m_next;
2290 	}
2291 	if (lastm != NULL)
2292 		*lastm = prev;
2293 	return (len);
2294 }
2295 
2296 /*
2297  * Like m_lengthm(), except also keep track of mbuf usage.
2298  */
2299 u_int
2300 m_countm(struct mbuf *m, struct mbuf **lastm, u_int *pmbcnt)
2301 {
2302 	u_int len = 0, mbcnt = 0;
2303 	struct mbuf *prev = m;
2304 
2305 	while (m) {
2306 		len += m->m_len;
2307 		mbcnt += MSIZE;
2308 		if (m->m_flags & M_EXT)
2309 			mbcnt += m->m_ext.ext_size;
2310 		prev = m;
2311 		m = m->m_next;
2312 	}
2313 	if (lastm != NULL)
2314 		*lastm = prev;
2315 	*pmbcnt = mbcnt;
2316 	return (len);
2317 }
2318