xref: /dflybsd-src/sys/kern/uipc_mbuf.c (revision bfc09ba0a4d805c1860f88e64d6ae9a407d3567d)
1 /*
2  * Copyright (c) 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of The DragonFly Project nor the names of its
17  *    contributors may be used to endorse or promote products derived
18  *    from this software without specific, prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * Copyright (c) 1982, 1986, 1988, 1991, 1993
36  *	The Regents of the University of California.  All rights reserved.
37  *
38  * Redistribution and use in source and binary forms, with or without
39  * modification, are permitted provided that the following conditions
40  * are met:
41  * 1. Redistributions of source code must retain the above copyright
42  *    notice, this list of conditions and the following disclaimer.
43  * 2. Redistributions in binary form must reproduce the above copyright
44  *    notice, this list of conditions and the following disclaimer in the
45  *    documentation and/or other materials provided with the distribution.
46  * 3. All advertising materials mentioning features or use of this software
47  *    must display the following acknowledgement:
48  *	This product includes software developed by the University of
49  *	California, Berkeley and its contributors.
50  * 4. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  * @(#)uipc_mbuf.c	8.2 (Berkeley) 1/4/94
67  * $FreeBSD: src/sys/kern/uipc_mbuf.c,v 1.51.2.24 2003/04/15 06:59:29 silby Exp $
68  * $DragonFly: src/sys/kern/uipc_mbuf.c,v 1.70 2008/11/20 14:21:01 sephe Exp $
69  */
70 
71 #include "opt_param.h"
72 #include "opt_mbuf_stress_test.h"
73 #include <sys/param.h>
74 #include <sys/systm.h>
75 #include <sys/malloc.h>
76 #include <sys/mbuf.h>
77 #include <sys/kernel.h>
78 #include <sys/sysctl.h>
79 #include <sys/domain.h>
80 #include <sys/objcache.h>
81 #include <sys/tree.h>
82 #include <sys/protosw.h>
83 #include <sys/uio.h>
84 #include <sys/thread.h>
85 #include <sys/globaldata.h>
86 #include <sys/thread2.h>
87 
88 #include <machine/atomic.h>
89 #include <machine/limits.h>
90 
91 #include <vm/vm.h>
92 #include <vm/vm_kern.h>
93 #include <vm/vm_extern.h>
94 
95 #ifdef INVARIANTS
96 #include <machine/cpu.h>
97 #endif
98 
99 /*
100  * mbuf cluster meta-data
101  */
102 struct mbcluster {
103 	int32_t	mcl_refs;
104 	void	*mcl_data;
105 };
106 
107 /*
108  * mbuf tracking for debugging purposes
109  */
110 #ifdef MBUF_DEBUG
111 
112 static MALLOC_DEFINE(M_MTRACK, "mtrack", "mtrack");
113 
114 struct mbctrack;
115 RB_HEAD(mbuf_rb_tree, mbtrack);
116 RB_PROTOTYPE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *);
117 
118 struct mbtrack {
119 	RB_ENTRY(mbtrack) rb_node;
120 	int trackid;
121 	struct mbuf *m;
122 };
123 
124 static int
125 mbtrack_cmp(struct mbtrack *mb1, struct mbtrack *mb2)
126 {
127 	if (mb1->m < mb2->m)
128 		return(-1);
129 	if (mb1->m > mb2->m)
130 		return(1);
131 	return(0);
132 }
133 
134 RB_GENERATE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *, m);
135 
136 struct mbuf_rb_tree	mbuf_track_root;
137 
138 static void
139 mbuftrack(struct mbuf *m)
140 {
141 	struct mbtrack *mbt;
142 
143 	crit_enter();
144 	mbt = kmalloc(sizeof(*mbt), M_MTRACK, M_INTWAIT|M_ZERO);
145 	mbt->m = m;
146 	if (mbuf_rb_tree_RB_INSERT(&mbuf_track_root, mbt))
147 		panic("mbuftrack: mbuf %p already being tracked\n", m);
148 	crit_exit();
149 }
150 
151 static void
152 mbufuntrack(struct mbuf *m)
153 {
154 	struct mbtrack *mbt;
155 
156 	crit_enter();
157 	mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m);
158 	if (mbt == NULL) {
159 		kprintf("mbufuntrack: mbuf %p was not tracked\n", m);
160 	} else {
161 		mbuf_rb_tree_RB_REMOVE(&mbuf_track_root, mbt);
162 		kfree(mbt, M_MTRACK);
163 	}
164 	crit_exit();
165 }
166 
167 void
168 mbuftrackid(struct mbuf *m, int trackid)
169 {
170 	struct mbtrack *mbt;
171 	struct mbuf *n;
172 
173 	crit_enter();
174 	while (m) {
175 		n = m->m_nextpkt;
176 		while (m) {
177 			mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m);
178 			if (mbt)
179 				mbt->trackid = trackid;
180 			m = m->m_next;
181 		}
182 		m = n;
183 	}
184 	crit_exit();
185 }
186 
187 static int
188 mbuftrack_callback(struct mbtrack *mbt, void *arg)
189 {
190 	struct sysctl_req *req = arg;
191 	char buf[64];
192 	int error;
193 
194 	ksnprintf(buf, sizeof(buf), "mbuf %p track %d\n", mbt->m, mbt->trackid);
195 
196 	error = SYSCTL_OUT(req, buf, strlen(buf));
197 	if (error)
198 		return(-error);
199 	return(0);
200 }
201 
202 static int
203 mbuftrack_show(SYSCTL_HANDLER_ARGS)
204 {
205 	int error;
206 
207 	crit_enter();
208 	error = mbuf_rb_tree_RB_SCAN(&mbuf_track_root, NULL,
209 				     mbuftrack_callback, req);
210 	crit_exit();
211 	return (-error);
212 }
213 SYSCTL_PROC(_kern_ipc, OID_AUTO, showmbufs, CTLFLAG_RD|CTLTYPE_STRING,
214 	    0, 0, mbuftrack_show, "A", "Show all in-use mbufs");
215 
216 #else
217 
218 #define mbuftrack(m)
219 #define mbufuntrack(m)
220 
221 #endif
222 
223 static void mbinit(void *);
224 SYSINIT(mbuf, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, mbinit, NULL)
225 
226 static u_long	mbtypes[SMP_MAXCPU][MT_NTYPES];
227 
228 static struct mbstat mbstat[SMP_MAXCPU];
229 int	max_linkhdr;
230 int	max_protohdr;
231 int	max_hdr;
232 int	max_datalen;
233 int	m_defragpackets;
234 int	m_defragbytes;
235 int	m_defraguseless;
236 int	m_defragfailure;
237 #ifdef MBUF_STRESS_TEST
238 int	m_defragrandomfailures;
239 #endif
240 
241 struct objcache *mbuf_cache, *mbufphdr_cache;
242 struct objcache *mclmeta_cache;
243 struct objcache *mbufcluster_cache, *mbufphdrcluster_cache;
244 
245 int	nmbclusters;
246 int	nmbufs;
247 
248 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW,
249 	   &max_linkhdr, 0, "");
250 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW,
251 	   &max_protohdr, 0, "");
252 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0, "");
253 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW,
254 	   &max_datalen, 0, "");
255 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW,
256 	   &mbuf_wait, 0, "");
257 static int do_mbstat(SYSCTL_HANDLER_ARGS);
258 
259 SYSCTL_PROC(_kern_ipc, KIPC_MBSTAT, mbstat, CTLTYPE_STRUCT|CTLFLAG_RD,
260 	0, 0, do_mbstat, "S,mbstat", "");
261 
262 static int do_mbtypes(SYSCTL_HANDLER_ARGS);
263 
264 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbtypes, CTLTYPE_ULONG|CTLFLAG_RD,
265 	0, 0, do_mbtypes, "LU", "");
266 
267 static int
268 do_mbstat(SYSCTL_HANDLER_ARGS)
269 {
270 	struct mbstat mbstat_total;
271 	struct mbstat *mbstat_totalp;
272 	int i;
273 
274 	bzero(&mbstat_total, sizeof(mbstat_total));
275 	mbstat_totalp = &mbstat_total;
276 
277 	for (i = 0; i < ncpus; i++)
278 	{
279 		mbstat_total.m_mbufs += mbstat[i].m_mbufs;
280 		mbstat_total.m_clusters += mbstat[i].m_clusters;
281 		mbstat_total.m_spare += mbstat[i].m_spare;
282 		mbstat_total.m_clfree += mbstat[i].m_clfree;
283 		mbstat_total.m_drops += mbstat[i].m_drops;
284 		mbstat_total.m_wait += mbstat[i].m_wait;
285 		mbstat_total.m_drain += mbstat[i].m_drain;
286 		mbstat_total.m_mcfail += mbstat[i].m_mcfail;
287 		mbstat_total.m_mpfail += mbstat[i].m_mpfail;
288 
289 	}
290 	/*
291 	 * The following fields are not cumulative fields so just
292 	 * get their values once.
293 	 */
294 	mbstat_total.m_msize = mbstat[0].m_msize;
295 	mbstat_total.m_mclbytes = mbstat[0].m_mclbytes;
296 	mbstat_total.m_minclsize = mbstat[0].m_minclsize;
297 	mbstat_total.m_mlen = mbstat[0].m_mlen;
298 	mbstat_total.m_mhlen = mbstat[0].m_mhlen;
299 
300 	return(sysctl_handle_opaque(oidp, mbstat_totalp, sizeof(mbstat_total), req));
301 }
302 
303 static int
304 do_mbtypes(SYSCTL_HANDLER_ARGS)
305 {
306 	u_long totals[MT_NTYPES];
307 	int i, j;
308 
309 	for (i = 0; i < MT_NTYPES; i++)
310 		totals[i] = 0;
311 
312 	for (i = 0; i < ncpus; i++)
313 	{
314 		for (j = 0; j < MT_NTYPES; j++)
315 			totals[j] += mbtypes[i][j];
316 	}
317 
318 	return(sysctl_handle_opaque(oidp, totals, sizeof(totals), req));
319 }
320 
321 /*
322  * These are read-only because we do not currently have any code
323  * to adjust the objcache limits after the fact.  The variables
324  * may only be set as boot-time tunables.
325  */
326 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RD,
327 	   &nmbclusters, 0, "Maximum number of mbuf clusters available");
328 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RD, &nmbufs, 0,
329 	   "Maximum number of mbufs available");
330 
331 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
332 	   &m_defragpackets, 0, "");
333 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
334 	   &m_defragbytes, 0, "");
335 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
336 	   &m_defraguseless, 0, "");
337 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
338 	   &m_defragfailure, 0, "");
339 #ifdef MBUF_STRESS_TEST
340 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
341 	   &m_defragrandomfailures, 0, "");
342 #endif
343 
344 static MALLOC_DEFINE(M_MBUF, "mbuf", "mbuf");
345 static MALLOC_DEFINE(M_MBUFCL, "mbufcl", "mbufcl");
346 static MALLOC_DEFINE(M_MCLMETA, "mclmeta", "mclmeta");
347 
348 static void m_reclaim (void);
349 static void m_mclref(void *arg);
350 static void m_mclfree(void *arg);
351 
352 #ifndef NMBCLUSTERS
353 #define NMBCLUSTERS	(512 + maxusers * 16)
354 #endif
355 #ifndef NMBUFS
356 #define NMBUFS		(nmbclusters * 2)
357 #endif
358 
359 /*
360  * Perform sanity checks of tunables declared above.
361  */
362 static void
363 tunable_mbinit(void *dummy)
364 {
365 	/*
366 	 * This has to be done before VM init.
367 	 */
368 	nmbclusters = NMBCLUSTERS;
369 	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
370 	nmbufs = NMBUFS;
371 	TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
372 	/* Sanity checks */
373 	if (nmbufs < nmbclusters * 2)
374 		nmbufs = nmbclusters * 2;
375 }
376 SYSINIT(tunable_mbinit, SI_BOOT1_TUNABLES, SI_ORDER_ANY,
377 	tunable_mbinit, NULL);
378 
379 /* "number of clusters of pages" */
380 #define NCL_INIT	1
381 
382 #define NMB_INIT	16
383 
384 /*
385  * The mbuf object cache only guarantees that m_next and m_nextpkt are
386  * NULL and that m_data points to the beginning of the data area.  In
387  * particular, m_len and m_pkthdr.len are uninitialized.  It is the
388  * responsibility of the caller to initialize those fields before use.
389  */
390 
391 static boolean_t __inline
392 mbuf_ctor(void *obj, void *private, int ocflags)
393 {
394 	struct mbuf *m = obj;
395 
396 	m->m_next = NULL;
397 	m->m_nextpkt = NULL;
398 	m->m_data = m->m_dat;
399 	m->m_flags = 0;
400 
401 	return (TRUE);
402 }
403 
404 /*
405  * Initialize the mbuf and the packet header fields.
406  */
407 static boolean_t
408 mbufphdr_ctor(void *obj, void *private, int ocflags)
409 {
410 	struct mbuf *m = obj;
411 
412 	m->m_next = NULL;
413 	m->m_nextpkt = NULL;
414 	m->m_data = m->m_pktdat;
415 	m->m_flags = M_PKTHDR | M_PHCACHE;
416 
417 	m->m_pkthdr.rcvif = NULL;	/* eliminate XXX JH */
418 	SLIST_INIT(&m->m_pkthdr.tags);
419 	m->m_pkthdr.csum_flags = 0;	/* eliminate XXX JH */
420 	m->m_pkthdr.fw_flags = 0;	/* eliminate XXX JH */
421 
422 	return (TRUE);
423 }
424 
425 /*
426  * A mbcluster object consists of 2K (MCLBYTES) cluster and a refcount.
427  */
428 static boolean_t
429 mclmeta_ctor(void *obj, void *private, int ocflags)
430 {
431 	struct mbcluster *cl = obj;
432 	void *buf;
433 
434 	if (ocflags & M_NOWAIT)
435 		buf = kmalloc(MCLBYTES, M_MBUFCL, M_NOWAIT | M_ZERO);
436 	else
437 		buf = kmalloc(MCLBYTES, M_MBUFCL, M_INTWAIT | M_ZERO);
438 	if (buf == NULL)
439 		return (FALSE);
440 	cl->mcl_refs = 0;
441 	cl->mcl_data = buf;
442 	return (TRUE);
443 }
444 
445 static void
446 mclmeta_dtor(void *obj, void *private)
447 {
448 	struct mbcluster *mcl = obj;
449 
450 	KKASSERT(mcl->mcl_refs == 0);
451 	kfree(mcl->mcl_data, M_MBUFCL);
452 }
453 
454 static void
455 linkcluster(struct mbuf *m, struct mbcluster *cl)
456 {
457 	/*
458 	 * Add the cluster to the mbuf.  The caller will detect that the
459 	 * mbuf now has an attached cluster.
460 	 */
461 	m->m_ext.ext_arg = cl;
462 	m->m_ext.ext_buf = cl->mcl_data;
463 	m->m_ext.ext_ref = m_mclref;
464 	m->m_ext.ext_free = m_mclfree;
465 	m->m_ext.ext_size = MCLBYTES;
466 	atomic_add_int(&cl->mcl_refs, 1);
467 
468 	m->m_data = m->m_ext.ext_buf;
469 	m->m_flags |= M_EXT | M_EXT_CLUSTER;
470 }
471 
472 static boolean_t
473 mbufphdrcluster_ctor(void *obj, void *private, int ocflags)
474 {
475 	struct mbuf *m = obj;
476 	struct mbcluster *cl;
477 
478 	mbufphdr_ctor(obj, private, ocflags);
479 	cl = objcache_get(mclmeta_cache, ocflags);
480 	if (cl == NULL)
481 		return (FALSE);
482 	m->m_flags |= M_CLCACHE;
483 	linkcluster(m, cl);
484 	return (TRUE);
485 }
486 
487 static boolean_t
488 mbufcluster_ctor(void *obj, void *private, int ocflags)
489 {
490 	struct mbuf *m = obj;
491 	struct mbcluster *cl;
492 
493 	mbuf_ctor(obj, private, ocflags);
494 	cl = objcache_get(mclmeta_cache, ocflags);
495 	if (cl == NULL)
496 		return (FALSE);
497 	m->m_flags |= M_CLCACHE;
498 	linkcluster(m, cl);
499 	return (TRUE);
500 }
501 
502 /*
503  * Used for both the cluster and cluster PHDR caches.
504  *
505  * The mbuf may have lost its cluster due to sharing, deal
506  * with the situation by checking M_EXT.
507  */
508 static void
509 mbufcluster_dtor(void *obj, void *private)
510 {
511 	struct mbuf *m = obj;
512 	struct mbcluster *mcl;
513 
514 	if (m->m_flags & M_EXT) {
515 		KKASSERT((m->m_flags & M_EXT_CLUSTER) != 0);
516 		mcl = m->m_ext.ext_arg;
517 		KKASSERT(mcl->mcl_refs == 1);
518 		mcl->mcl_refs = 0;
519 		objcache_put(mclmeta_cache, mcl);
520 	}
521 }
522 
523 struct objcache_malloc_args mbuf_malloc_args = { MSIZE, M_MBUF };
524 struct objcache_malloc_args mclmeta_malloc_args =
525 	{ sizeof(struct mbcluster), M_MCLMETA };
526 
527 /* ARGSUSED*/
528 static void
529 mbinit(void *dummy)
530 {
531 	int mb_limit, cl_limit;
532 	int limit;
533 	int i;
534 
535 	/*
536 	 * Initialize statistics
537 	 */
538 	for (i = 0; i < ncpus; i++) {
539 		atomic_set_long_nonlocked(&mbstat[i].m_msize, MSIZE);
540 		atomic_set_long_nonlocked(&mbstat[i].m_mclbytes, MCLBYTES);
541 		atomic_set_long_nonlocked(&mbstat[i].m_minclsize, MINCLSIZE);
542 		atomic_set_long_nonlocked(&mbstat[i].m_mlen, MLEN);
543 		atomic_set_long_nonlocked(&mbstat[i].m_mhlen, MHLEN);
544 	}
545 
546 	/*
547 	 * Create objtect caches and save cluster limits, which will
548 	 * be used to adjust backing kmalloc pools' limit later.
549 	 */
550 
551 	mb_limit = cl_limit = 0;
552 
553 	limit = nmbufs;
554 	mbuf_cache = objcache_create("mbuf", &limit, 0,
555 	    mbuf_ctor, NULL, NULL,
556 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
557 	mb_limit += limit;
558 
559 	limit = nmbufs;
560 	mbufphdr_cache = objcache_create("mbuf pkt hdr", &limit, 64,
561 	    mbufphdr_ctor, NULL, NULL,
562 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
563 	mb_limit += limit;
564 
565 	cl_limit = nmbclusters;
566 	mclmeta_cache = objcache_create("cluster mbuf", &cl_limit, 0,
567 	    mclmeta_ctor, mclmeta_dtor, NULL,
568 	    objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args);
569 
570 	limit = nmbclusters;
571 	mbufcluster_cache = objcache_create("mbuf + cluster", &limit, 0,
572 	    mbufcluster_ctor, mbufcluster_dtor, NULL,
573 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
574 	mb_limit += limit;
575 
576 	limit = nmbclusters;
577 	mbufphdrcluster_cache = objcache_create("mbuf pkt hdr + cluster",
578 	    &limit, 64, mbufphdrcluster_ctor, mbufcluster_dtor, NULL,
579 	    objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
580 	mb_limit += limit;
581 
582 	/*
583 	 * Adjust backing kmalloc pools' limit
584 	 *
585 	 * NOTE: We raise the limit by another 1/8 to take the effect
586 	 * of loosememuse into account.
587 	 */
588 	cl_limit += cl_limit / 8;
589 	kmalloc_raise_limit(mclmeta_malloc_args.mtype,
590 			    mclmeta_malloc_args.objsize * cl_limit);
591 	kmalloc_raise_limit(M_MBUFCL, MCLBYTES * cl_limit);
592 
593 	mb_limit += mb_limit / 8;
594 	kmalloc_raise_limit(mbuf_malloc_args.mtype,
595 			    mbuf_malloc_args.objsize * mb_limit);
596 }
597 
598 /*
599  * Return the number of references to this mbuf's data.  0 is returned
600  * if the mbuf is not M_EXT, a reference count is returned if it is
601  * M_EXT | M_EXT_CLUSTER, and 99 is returned if it is a special M_EXT.
602  */
603 int
604 m_sharecount(struct mbuf *m)
605 {
606 	switch (m->m_flags & (M_EXT | M_EXT_CLUSTER)) {
607 	case 0:
608 		return (0);
609 	case M_EXT:
610 		return (99);
611 	case M_EXT | M_EXT_CLUSTER:
612 		return (((struct mbcluster *)m->m_ext.ext_arg)->mcl_refs);
613 	}
614 	/* NOTREACHED */
615 	return (0);		/* to shut up compiler */
616 }
617 
618 /*
619  * change mbuf to new type
620  */
621 void
622 m_chtype(struct mbuf *m, int type)
623 {
624 	struct globaldata *gd = mycpu;
625 
626 	atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1);
627 	atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1);
628 	atomic_set_short_nonlocked(&m->m_type, type);
629 }
630 
631 static void
632 m_reclaim(void)
633 {
634 	struct domain *dp;
635 	struct protosw *pr;
636 
637 	crit_enter();
638 	SLIST_FOREACH(dp, &domains, dom_next) {
639 		for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) {
640 			if (pr->pr_drain)
641 				(*pr->pr_drain)();
642 		}
643 	}
644 	crit_exit();
645 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_drain, 1);
646 }
647 
648 static void __inline
649 updatestats(struct mbuf *m, int type)
650 {
651 	struct globaldata *gd = mycpu;
652 	m->m_type = type;
653 
654 	mbuftrack(m);
655 
656 	atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1);
657 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1);
658 
659 }
660 
661 /*
662  * Allocate an mbuf.
663  */
664 struct mbuf *
665 m_get(int how, int type)
666 {
667 	struct mbuf *m;
668 	int ntries = 0;
669 	int ocf = MBTOM(how);
670 
671 retryonce:
672 
673 	m = objcache_get(mbuf_cache, ocf);
674 
675 	if (m == NULL) {
676 		if ((how & MB_TRYWAIT) && ntries++ == 0) {
677 			struct objcache *reclaimlist[] = {
678 				mbufphdr_cache,
679 				mbufcluster_cache, mbufphdrcluster_cache
680 			};
681 			const int nreclaims = __arysize(reclaimlist);
682 
683 			if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
684 				m_reclaim();
685 			goto retryonce;
686 		}
687 		return (NULL);
688 	}
689 
690 	updatestats(m, type);
691 	return (m);
692 }
693 
694 struct mbuf *
695 m_gethdr(int how, int type)
696 {
697 	struct mbuf *m;
698 	int ocf = MBTOM(how);
699 	int ntries = 0;
700 
701 retryonce:
702 
703 	m = objcache_get(mbufphdr_cache, ocf);
704 
705 	if (m == NULL) {
706 		if ((how & MB_TRYWAIT) && ntries++ == 0) {
707 			struct objcache *reclaimlist[] = {
708 				mbuf_cache,
709 				mbufcluster_cache, mbufphdrcluster_cache
710 			};
711 			const int nreclaims = __arysize(reclaimlist);
712 
713 			if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
714 				m_reclaim();
715 			goto retryonce;
716 		}
717 		return (NULL);
718 	}
719 
720 	updatestats(m, type);
721 	return (m);
722 }
723 
724 /*
725  * Get a mbuf (not a mbuf cluster!) and zero it.
726  * Deprecated.
727  */
728 struct mbuf *
729 m_getclr(int how, int type)
730 {
731 	struct mbuf *m;
732 
733 	m = m_get(how, type);
734 	if (m != NULL)
735 		bzero(m->m_data, MLEN);
736 	return (m);
737 }
738 
739 /*
740  * Returns an mbuf with an attached cluster.
741  * Because many network drivers use this kind of buffers a lot, it is
742  * convenient to keep a small pool of free buffers of this kind.
743  * Even a small size such as 10 gives about 10% improvement in the
744  * forwarding rate in a bridge or router.
745  */
746 struct mbuf *
747 m_getcl(int how, short type, int flags)
748 {
749 	struct mbuf *m;
750 	int ocflags = MBTOM(how);
751 	int ntries = 0;
752 
753 retryonce:
754 
755 	if (flags & M_PKTHDR)
756 		m = objcache_get(mbufphdrcluster_cache, ocflags);
757 	else
758 		m = objcache_get(mbufcluster_cache, ocflags);
759 
760 	if (m == NULL) {
761 		if ((how & MB_TRYWAIT) && ntries++ == 0) {
762 			struct objcache *reclaimlist[1];
763 
764 			if (flags & M_PKTHDR)
765 				reclaimlist[0] = mbufcluster_cache;
766 			else
767 				reclaimlist[0] = mbufphdrcluster_cache;
768 			if (!objcache_reclaimlist(reclaimlist, 1, ocflags))
769 				m_reclaim();
770 			goto retryonce;
771 		}
772 		return (NULL);
773 	}
774 
775 	m->m_type = type;
776 
777 	mbuftrack(m);
778 
779 	atomic_add_long_nonlocked(&mbtypes[mycpu->gd_cpuid][type], 1);
780 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
781 	return (m);
782 }
783 
784 /*
785  * Allocate chain of requested length.
786  */
787 struct mbuf *
788 m_getc(int len, int how, int type)
789 {
790 	struct mbuf *n, *nfirst = NULL, **ntail = &nfirst;
791 	int nsize;
792 
793 	while (len > 0) {
794 		n = m_getl(len, how, type, 0, &nsize);
795 		if (n == NULL)
796 			goto failed;
797 		n->m_len = 0;
798 		*ntail = n;
799 		ntail = &n->m_next;
800 		len -= nsize;
801 	}
802 	return (nfirst);
803 
804 failed:
805 	m_freem(nfirst);
806 	return (NULL);
807 }
808 
809 /*
810  * Allocate len-worth of mbufs and/or mbuf clusters (whatever fits best)
811  * and return a pointer to the head of the allocated chain. If m0 is
812  * non-null, then we assume that it is a single mbuf or an mbuf chain to
813  * which we want len bytes worth of mbufs and/or clusters attached, and so
814  * if we succeed in allocating it, we will just return a pointer to m0.
815  *
816  * If we happen to fail at any point during the allocation, we will free
817  * up everything we have already allocated and return NULL.
818  *
819  * Deprecated.  Use m_getc() and m_cat() instead.
820  */
821 struct mbuf *
822 m_getm(struct mbuf *m0, int len, int type, int how)
823 {
824 	struct mbuf *nfirst;
825 
826 	nfirst = m_getc(len, how, type);
827 
828 	if (m0 != NULL) {
829 		m_last(m0)->m_next = nfirst;
830 		return (m0);
831 	}
832 
833 	return (nfirst);
834 }
835 
836 /*
837  * Adds a cluster to a normal mbuf, M_EXT is set on success.
838  * Deprecated.  Use m_getcl() instead.
839  */
840 void
841 m_mclget(struct mbuf *m, int how)
842 {
843 	struct mbcluster *mcl;
844 
845 	KKASSERT((m->m_flags & M_EXT) == 0);
846 	mcl = objcache_get(mclmeta_cache, MBTOM(how));
847 	if (mcl != NULL) {
848 		linkcluster(m, mcl);
849 		atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
850 	}
851 }
852 
853 /*
854  * Updates to mbcluster must be MPSAFE.  Only an entity which already has
855  * a reference to the cluster can ref it, so we are in no danger of
856  * racing an add with a subtract.  But the operation must still be atomic
857  * since multiple entities may have a reference on the cluster.
858  *
859  * m_mclfree() is almost the same but it must contend with two entities
860  * freeing the cluster at the same time.  If there is only one reference
861  * count we are the only entity referencing the cluster and no further
862  * locking is required.  Otherwise we must protect against a race to 0
863  * with the serializer.
864  */
865 static void
866 m_mclref(void *arg)
867 {
868 	struct mbcluster *mcl = arg;
869 
870 	atomic_add_int(&mcl->mcl_refs, 1);
871 }
872 
873 /*
874  * When dereferencing a cluster we have to deal with a N->0 race, where
875  * N entities free their references simultaniously.  To do this we use
876  * atomic_fetchadd_int().
877  */
878 static void
879 m_mclfree(void *arg)
880 {
881 	struct mbcluster *mcl = arg;
882 
883 	if (atomic_fetchadd_int(&mcl->mcl_refs, -1) == 1)
884 		objcache_put(mclmeta_cache, mcl);
885 }
886 
887 /*
888  * Free a single mbuf and any associated external storage.  The successor,
889  * if any, is returned.
890  *
891  * We do need to check non-first mbuf for m_aux, since some of existing
892  * code does not call M_PREPEND properly.
893  * (example: call to bpf_mtap from drivers)
894  */
895 struct mbuf *
896 m_free(struct mbuf *m)
897 {
898 	struct mbuf *n;
899 	struct globaldata *gd = mycpu;
900 
901 	KASSERT(m->m_type != MT_FREE, ("freeing free mbuf %p", m));
902 	atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1);
903 
904 	n = m->m_next;
905 
906 	/*
907 	 * Make sure the mbuf is in constructed state before returning it
908 	 * to the objcache.
909 	 */
910 	m->m_next = NULL;
911 	mbufuntrack(m);
912 #ifdef notyet
913 	KKASSERT(m->m_nextpkt == NULL);
914 #else
915 	if (m->m_nextpkt != NULL) {
916 		static int afewtimes = 10;
917 
918 		if (afewtimes-- > 0) {
919 			kprintf("mfree: m->m_nextpkt != NULL\n");
920 			print_backtrace();
921 		}
922 		m->m_nextpkt = NULL;
923 	}
924 #endif
925 	if (m->m_flags & M_PKTHDR) {
926 		m_tag_delete_chain(m);		/* eliminate XXX JH */
927 	}
928 
929 	m->m_flags &= (M_EXT | M_EXT_CLUSTER | M_CLCACHE | M_PHCACHE);
930 
931 	/*
932 	 * Clean the M_PKTHDR state so we can return the mbuf to its original
933 	 * cache.  This is based on the PHCACHE flag which tells us whether
934 	 * the mbuf was originally allocated out of a packet-header cache
935 	 * or a non-packet-header cache.
936 	 */
937 	if (m->m_flags & M_PHCACHE) {
938 		m->m_flags |= M_PKTHDR;
939 		m->m_pkthdr.rcvif = NULL;	/* eliminate XXX JH */
940 		m->m_pkthdr.csum_flags = 0;	/* eliminate XXX JH */
941 		m->m_pkthdr.fw_flags = 0;	/* eliminate XXX JH */
942 		SLIST_INIT(&m->m_pkthdr.tags);
943 	}
944 
945 	/*
946 	 * Handle remaining flags combinations.  M_CLCACHE tells us whether
947 	 * the mbuf was originally allocated from a cluster cache or not,
948 	 * and is totally separate from whether the mbuf is currently
949 	 * associated with a cluster.
950 	 */
951 	crit_enter();
952 	switch(m->m_flags & (M_CLCACHE | M_EXT | M_EXT_CLUSTER)) {
953 	case M_CLCACHE | M_EXT | M_EXT_CLUSTER:
954 		/*
955 		 * mbuf+cluster cache case.  The mbuf was allocated from the
956 		 * combined mbuf_cluster cache and can be returned to the
957 		 * cache if the cluster hasn't been shared.
958 		 */
959 		if (m_sharecount(m) == 1) {
960 			/*
961 			 * The cluster has not been shared, we can just
962 			 * reset the data pointer and return the mbuf
963 			 * to the cluster cache.  Note that the reference
964 			 * count is left intact (it is still associated with
965 			 * an mbuf).
966 			 */
967 			m->m_data = m->m_ext.ext_buf;
968 			if (m->m_flags & M_PHCACHE)
969 				objcache_put(mbufphdrcluster_cache, m);
970 			else
971 				objcache_put(mbufcluster_cache, m);
972 			atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
973 		} else {
974 			/*
975 			 * Hell.  Someone else has a ref on this cluster,
976 			 * we have to disconnect it which means we can't
977 			 * put it back into the mbufcluster_cache, we
978 			 * have to destroy the mbuf.
979 			 *
980 			 * Other mbuf references to the cluster will typically
981 			 * be M_EXT | M_EXT_CLUSTER but without M_CLCACHE.
982 			 *
983 			 * XXX we could try to connect another cluster to
984 			 * it.
985 			 */
986 			m->m_ext.ext_free(m->m_ext.ext_arg);
987 			m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
988 			if (m->m_flags & M_PHCACHE)
989 				objcache_dtor(mbufphdrcluster_cache, m);
990 			else
991 				objcache_dtor(mbufcluster_cache, m);
992 		}
993 		break;
994 	case M_EXT | M_EXT_CLUSTER:
995 		/*
996 		 * Normal cluster associated with an mbuf that was allocated
997 		 * from the normal mbuf pool rather then the cluster pool.
998 		 * The cluster has to be independantly disassociated from the
999 		 * mbuf.
1000 		 */
1001 		if (m_sharecount(m) == 1)
1002 			atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
1003 		/* fall through */
1004 	case M_EXT:
1005 		/*
1006 		 * Normal cluster association case, disconnect the cluster from
1007 		 * the mbuf.  The cluster may or may not be custom.
1008 		 */
1009 		m->m_ext.ext_free(m->m_ext.ext_arg);
1010 		m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
1011 		/* fall through */
1012 	case 0:
1013 		/*
1014 		 * return the mbuf to the mbuf cache.
1015 		 */
1016 		if (m->m_flags & M_PHCACHE) {
1017 			m->m_data = m->m_pktdat;
1018 			objcache_put(mbufphdr_cache, m);
1019 		} else {
1020 			m->m_data = m->m_dat;
1021 			objcache_put(mbuf_cache, m);
1022 		}
1023 		atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1);
1024 		break;
1025 	default:
1026 		if (!panicstr)
1027 			panic("bad mbuf flags %p %08x\n", m, m->m_flags);
1028 		break;
1029 	}
1030 	crit_exit();
1031 	return (n);
1032 }
1033 
1034 void
1035 m_freem(struct mbuf *m)
1036 {
1037 	crit_enter();
1038 	while (m)
1039 		m = m_free(m);
1040 	crit_exit();
1041 }
1042 
1043 /*
1044  * mbuf utility routines
1045  */
1046 
1047 /*
1048  * Lesser-used path for M_PREPEND: allocate new mbuf to prepend to chain and
1049  * copy junk along.
1050  */
1051 struct mbuf *
1052 m_prepend(struct mbuf *m, int len, int how)
1053 {
1054 	struct mbuf *mn;
1055 
1056 	if (m->m_flags & M_PKTHDR)
1057 	    mn = m_gethdr(how, m->m_type);
1058 	else
1059 	    mn = m_get(how, m->m_type);
1060 	if (mn == NULL) {
1061 		m_freem(m);
1062 		return (NULL);
1063 	}
1064 	if (m->m_flags & M_PKTHDR)
1065 		M_MOVE_PKTHDR(mn, m);
1066 	mn->m_next = m;
1067 	m = mn;
1068 	if (len < MHLEN)
1069 		MH_ALIGN(m, len);
1070 	m->m_len = len;
1071 	return (m);
1072 }
1073 
1074 /*
1075  * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
1076  * continuing for "len" bytes.  If len is M_COPYALL, copy to end of mbuf.
1077  * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller.
1078  * Note that the copy is read-only, because clusters are not copied,
1079  * only their reference counts are incremented.
1080  */
1081 struct mbuf *
1082 m_copym(const struct mbuf *m, int off0, int len, int wait)
1083 {
1084 	struct mbuf *n, **np;
1085 	int off = off0;
1086 	struct mbuf *top;
1087 	int copyhdr = 0;
1088 
1089 	KASSERT(off >= 0, ("m_copym, negative off %d", off));
1090 	KASSERT(len >= 0, ("m_copym, negative len %d", len));
1091 	if (off == 0 && m->m_flags & M_PKTHDR)
1092 		copyhdr = 1;
1093 	while (off > 0) {
1094 		KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
1095 		if (off < m->m_len)
1096 			break;
1097 		off -= m->m_len;
1098 		m = m->m_next;
1099 	}
1100 	np = &top;
1101 	top = 0;
1102 	while (len > 0) {
1103 		if (m == NULL) {
1104 			KASSERT(len == M_COPYALL,
1105 			    ("m_copym, length > size of mbuf chain"));
1106 			break;
1107 		}
1108 		/*
1109 		 * Because we are sharing any cluster attachment below,
1110 		 * be sure to get an mbuf that does not have a cluster
1111 		 * associated with it.
1112 		 */
1113 		if (copyhdr)
1114 			n = m_gethdr(wait, m->m_type);
1115 		else
1116 			n = m_get(wait, m->m_type);
1117 		*np = n;
1118 		if (n == NULL)
1119 			goto nospace;
1120 		if (copyhdr) {
1121 			if (!m_dup_pkthdr(n, m, wait))
1122 				goto nospace;
1123 			if (len == M_COPYALL)
1124 				n->m_pkthdr.len -= off0;
1125 			else
1126 				n->m_pkthdr.len = len;
1127 			copyhdr = 0;
1128 		}
1129 		n->m_len = min(len, m->m_len - off);
1130 		if (m->m_flags & M_EXT) {
1131 			KKASSERT((n->m_flags & M_EXT) == 0);
1132 			n->m_data = m->m_data + off;
1133 			m->m_ext.ext_ref(m->m_ext.ext_arg);
1134 			n->m_ext = m->m_ext;
1135 			n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1136 		} else {
1137 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1138 			    (unsigned)n->m_len);
1139 		}
1140 		if (len != M_COPYALL)
1141 			len -= n->m_len;
1142 		off = 0;
1143 		m = m->m_next;
1144 		np = &n->m_next;
1145 	}
1146 	if (top == NULL)
1147 		atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1148 	return (top);
1149 nospace:
1150 	m_freem(top);
1151 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1152 	return (NULL);
1153 }
1154 
1155 /*
1156  * Copy an entire packet, including header (which must be present).
1157  * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
1158  * Note that the copy is read-only, because clusters are not copied,
1159  * only their reference counts are incremented.
1160  * Preserve alignment of the first mbuf so if the creator has left
1161  * some room at the beginning (e.g. for inserting protocol headers)
1162  * the copies also have the room available.
1163  */
1164 struct mbuf *
1165 m_copypacket(struct mbuf *m, int how)
1166 {
1167 	struct mbuf *top, *n, *o;
1168 
1169 	n = m_gethdr(how, m->m_type);
1170 	top = n;
1171 	if (!n)
1172 		goto nospace;
1173 
1174 	if (!m_dup_pkthdr(n, m, how))
1175 		goto nospace;
1176 	n->m_len = m->m_len;
1177 	if (m->m_flags & M_EXT) {
1178 		KKASSERT((n->m_flags & M_EXT) == 0);
1179 		n->m_data = m->m_data;
1180 		m->m_ext.ext_ref(m->m_ext.ext_arg);
1181 		n->m_ext = m->m_ext;
1182 		n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1183 	} else {
1184 		n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
1185 		bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1186 	}
1187 
1188 	m = m->m_next;
1189 	while (m) {
1190 		o = m_get(how, m->m_type);
1191 		if (!o)
1192 			goto nospace;
1193 
1194 		n->m_next = o;
1195 		n = n->m_next;
1196 
1197 		n->m_len = m->m_len;
1198 		if (m->m_flags & M_EXT) {
1199 			KKASSERT((n->m_flags & M_EXT) == 0);
1200 			n->m_data = m->m_data;
1201 			m->m_ext.ext_ref(m->m_ext.ext_arg);
1202 			n->m_ext = m->m_ext;
1203 			n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1204 		} else {
1205 			bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1206 		}
1207 
1208 		m = m->m_next;
1209 	}
1210 	return top;
1211 nospace:
1212 	m_freem(top);
1213 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1214 	return (NULL);
1215 }
1216 
1217 /*
1218  * Copy data from an mbuf chain starting "off" bytes from the beginning,
1219  * continuing for "len" bytes, into the indicated buffer.
1220  */
1221 void
1222 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
1223 {
1224 	unsigned count;
1225 
1226 	KASSERT(off >= 0, ("m_copydata, negative off %d", off));
1227 	KASSERT(len >= 0, ("m_copydata, negative len %d", len));
1228 	while (off > 0) {
1229 		KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
1230 		if (off < m->m_len)
1231 			break;
1232 		off -= m->m_len;
1233 		m = m->m_next;
1234 	}
1235 	while (len > 0) {
1236 		KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
1237 		count = min(m->m_len - off, len);
1238 		bcopy(mtod(m, caddr_t) + off, cp, count);
1239 		len -= count;
1240 		cp += count;
1241 		off = 0;
1242 		m = m->m_next;
1243 	}
1244 }
1245 
1246 /*
1247  * Copy a packet header mbuf chain into a completely new chain, including
1248  * copying any mbuf clusters.  Use this instead of m_copypacket() when
1249  * you need a writable copy of an mbuf chain.
1250  */
1251 struct mbuf *
1252 m_dup(struct mbuf *m, int how)
1253 {
1254 	struct mbuf **p, *top = NULL;
1255 	int remain, moff, nsize;
1256 
1257 	/* Sanity check */
1258 	if (m == NULL)
1259 		return (NULL);
1260 	KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __func__));
1261 
1262 	/* While there's more data, get a new mbuf, tack it on, and fill it */
1263 	remain = m->m_pkthdr.len;
1264 	moff = 0;
1265 	p = &top;
1266 	while (remain > 0 || top == NULL) {	/* allow m->m_pkthdr.len == 0 */
1267 		struct mbuf *n;
1268 
1269 		/* Get the next new mbuf */
1270 		n = m_getl(remain, how, m->m_type, top == NULL ? M_PKTHDR : 0,
1271 			   &nsize);
1272 		if (n == NULL)
1273 			goto nospace;
1274 		if (top == NULL)
1275 			if (!m_dup_pkthdr(n, m, how))
1276 				goto nospace0;
1277 
1278 		/* Link it into the new chain */
1279 		*p = n;
1280 		p = &n->m_next;
1281 
1282 		/* Copy data from original mbuf(s) into new mbuf */
1283 		n->m_len = 0;
1284 		while (n->m_len < nsize && m != NULL) {
1285 			int chunk = min(nsize - n->m_len, m->m_len - moff);
1286 
1287 			bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1288 			moff += chunk;
1289 			n->m_len += chunk;
1290 			remain -= chunk;
1291 			if (moff == m->m_len) {
1292 				m = m->m_next;
1293 				moff = 0;
1294 			}
1295 		}
1296 
1297 		/* Check correct total mbuf length */
1298 		KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
1299 			("%s: bogus m_pkthdr.len", __func__));
1300 	}
1301 	return (top);
1302 
1303 nospace:
1304 	m_freem(top);
1305 nospace0:
1306 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1307 	return (NULL);
1308 }
1309 
1310 /*
1311  * Concatenate mbuf chain n to m.
1312  * Both chains must be of the same type (e.g. MT_DATA).
1313  * Any m_pkthdr is not updated.
1314  */
1315 void
1316 m_cat(struct mbuf *m, struct mbuf *n)
1317 {
1318 	m = m_last(m);
1319 	while (n) {
1320 		if (m->m_flags & M_EXT ||
1321 		    m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) {
1322 			/* just join the two chains */
1323 			m->m_next = n;
1324 			return;
1325 		}
1326 		/* splat the data from one into the other */
1327 		bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1328 		    (u_int)n->m_len);
1329 		m->m_len += n->m_len;
1330 		n = m_free(n);
1331 	}
1332 }
1333 
1334 void
1335 m_adj(struct mbuf *mp, int req_len)
1336 {
1337 	int len = req_len;
1338 	struct mbuf *m;
1339 	int count;
1340 
1341 	if ((m = mp) == NULL)
1342 		return;
1343 	if (len >= 0) {
1344 		/*
1345 		 * Trim from head.
1346 		 */
1347 		while (m != NULL && len > 0) {
1348 			if (m->m_len <= len) {
1349 				len -= m->m_len;
1350 				m->m_len = 0;
1351 				m = m->m_next;
1352 			} else {
1353 				m->m_len -= len;
1354 				m->m_data += len;
1355 				len = 0;
1356 			}
1357 		}
1358 		m = mp;
1359 		if (mp->m_flags & M_PKTHDR)
1360 			m->m_pkthdr.len -= (req_len - len);
1361 	} else {
1362 		/*
1363 		 * Trim from tail.  Scan the mbuf chain,
1364 		 * calculating its length and finding the last mbuf.
1365 		 * If the adjustment only affects this mbuf, then just
1366 		 * adjust and return.  Otherwise, rescan and truncate
1367 		 * after the remaining size.
1368 		 */
1369 		len = -len;
1370 		count = 0;
1371 		for (;;) {
1372 			count += m->m_len;
1373 			if (m->m_next == NULL)
1374 				break;
1375 			m = m->m_next;
1376 		}
1377 		if (m->m_len >= len) {
1378 			m->m_len -= len;
1379 			if (mp->m_flags & M_PKTHDR)
1380 				mp->m_pkthdr.len -= len;
1381 			return;
1382 		}
1383 		count -= len;
1384 		if (count < 0)
1385 			count = 0;
1386 		/*
1387 		 * Correct length for chain is "count".
1388 		 * Find the mbuf with last data, adjust its length,
1389 		 * and toss data from remaining mbufs on chain.
1390 		 */
1391 		m = mp;
1392 		if (m->m_flags & M_PKTHDR)
1393 			m->m_pkthdr.len = count;
1394 		for (; m; m = m->m_next) {
1395 			if (m->m_len >= count) {
1396 				m->m_len = count;
1397 				break;
1398 			}
1399 			count -= m->m_len;
1400 		}
1401 		while (m->m_next)
1402 			(m = m->m_next) ->m_len = 0;
1403 	}
1404 }
1405 
1406 /*
1407  * Rearrange an mbuf chain so that len bytes are contiguous
1408  * and in the data area of an mbuf (so that mtod will work for a structure
1409  * of size len).  Returns the resulting mbuf chain on success, frees it and
1410  * returns null on failure.  If there is room, it will add up to
1411  * max_protohdr-len extra bytes to the contiguous region in an attempt to
1412  * avoid being called next time.
1413  */
1414 struct mbuf *
1415 m_pullup(struct mbuf *n, int len)
1416 {
1417 	struct mbuf *m;
1418 	int count;
1419 	int space;
1420 
1421 	/*
1422 	 * If first mbuf has no cluster, and has room for len bytes
1423 	 * without shifting current data, pullup into it,
1424 	 * otherwise allocate a new mbuf to prepend to the chain.
1425 	 */
1426 	if (!(n->m_flags & M_EXT) &&
1427 	    n->m_data + len < &n->m_dat[MLEN] &&
1428 	    n->m_next) {
1429 		if (n->m_len >= len)
1430 			return (n);
1431 		m = n;
1432 		n = n->m_next;
1433 		len -= m->m_len;
1434 	} else {
1435 		if (len > MHLEN)
1436 			goto bad;
1437 		if (n->m_flags & M_PKTHDR)
1438 			m = m_gethdr(MB_DONTWAIT, n->m_type);
1439 		else
1440 			m = m_get(MB_DONTWAIT, n->m_type);
1441 		if (m == NULL)
1442 			goto bad;
1443 		m->m_len = 0;
1444 		if (n->m_flags & M_PKTHDR)
1445 			M_MOVE_PKTHDR(m, n);
1446 	}
1447 	space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
1448 	do {
1449 		count = min(min(max(len, max_protohdr), space), n->m_len);
1450 		bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1451 		  (unsigned)count);
1452 		len -= count;
1453 		m->m_len += count;
1454 		n->m_len -= count;
1455 		space -= count;
1456 		if (n->m_len)
1457 			n->m_data += count;
1458 		else
1459 			n = m_free(n);
1460 	} while (len > 0 && n);
1461 	if (len > 0) {
1462 		m_free(m);
1463 		goto bad;
1464 	}
1465 	m->m_next = n;
1466 	return (m);
1467 bad:
1468 	m_freem(n);
1469 	atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1470 	return (NULL);
1471 }
1472 
1473 /*
1474  * Partition an mbuf chain in two pieces, returning the tail --
1475  * all but the first len0 bytes.  In case of failure, it returns NULL and
1476  * attempts to restore the chain to its original state.
1477  *
1478  * Note that the resulting mbufs might be read-only, because the new
1479  * mbuf can end up sharing an mbuf cluster with the original mbuf if
1480  * the "breaking point" happens to lie within a cluster mbuf. Use the
1481  * M_WRITABLE() macro to check for this case.
1482  */
1483 struct mbuf *
1484 m_split(struct mbuf *m0, int len0, int wait)
1485 {
1486 	struct mbuf *m, *n;
1487 	unsigned len = len0, remain;
1488 
1489 	for (m = m0; m && len > m->m_len; m = m->m_next)
1490 		len -= m->m_len;
1491 	if (m == NULL)
1492 		return (NULL);
1493 	remain = m->m_len - len;
1494 	if (m0->m_flags & M_PKTHDR) {
1495 		n = m_gethdr(wait, m0->m_type);
1496 		if (n == NULL)
1497 			return (NULL);
1498 		n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1499 		n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1500 		m0->m_pkthdr.len = len0;
1501 		if (m->m_flags & M_EXT)
1502 			goto extpacket;
1503 		if (remain > MHLEN) {
1504 			/* m can't be the lead packet */
1505 			MH_ALIGN(n, 0);
1506 			n->m_next = m_split(m, len, wait);
1507 			if (n->m_next == NULL) {
1508 				m_free(n);
1509 				return (NULL);
1510 			} else {
1511 				n->m_len = 0;
1512 				return (n);
1513 			}
1514 		} else
1515 			MH_ALIGN(n, remain);
1516 	} else if (remain == 0) {
1517 		n = m->m_next;
1518 		m->m_next = 0;
1519 		return (n);
1520 	} else {
1521 		n = m_get(wait, m->m_type);
1522 		if (n == NULL)
1523 			return (NULL);
1524 		M_ALIGN(n, remain);
1525 	}
1526 extpacket:
1527 	if (m->m_flags & M_EXT) {
1528 		KKASSERT((n->m_flags & M_EXT) == 0);
1529 		n->m_data = m->m_data + len;
1530 		m->m_ext.ext_ref(m->m_ext.ext_arg);
1531 		n->m_ext = m->m_ext;
1532 		n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1533 	} else {
1534 		bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1535 	}
1536 	n->m_len = remain;
1537 	m->m_len = len;
1538 	n->m_next = m->m_next;
1539 	m->m_next = 0;
1540 	return (n);
1541 }
1542 
1543 /*
1544  * Routine to copy from device local memory into mbufs.
1545  * Note: "offset" is ill-defined and always called as 0, so ignore it.
1546  */
1547 struct mbuf *
1548 m_devget(char *buf, int len, int offset, struct ifnet *ifp,
1549     void (*copy)(volatile const void *from, volatile void *to, size_t length))
1550 {
1551 	struct mbuf *m, *mfirst = NULL, **mtail;
1552 	int nsize, flags;
1553 
1554 	if (copy == NULL)
1555 		copy = bcopy;
1556 	mtail = &mfirst;
1557 	flags = M_PKTHDR;
1558 
1559 	while (len > 0) {
1560 		m = m_getl(len, MB_DONTWAIT, MT_DATA, flags, &nsize);
1561 		if (m == NULL) {
1562 			m_freem(mfirst);
1563 			return (NULL);
1564 		}
1565 		m->m_len = min(len, nsize);
1566 
1567 		if (flags & M_PKTHDR) {
1568 			if (len + max_linkhdr <= nsize)
1569 				m->m_data += max_linkhdr;
1570 			m->m_pkthdr.rcvif = ifp;
1571 			m->m_pkthdr.len = len;
1572 			flags = 0;
1573 		}
1574 
1575 		copy(buf, m->m_data, (unsigned)m->m_len);
1576 		buf += m->m_len;
1577 		len -= m->m_len;
1578 		*mtail = m;
1579 		mtail = &m->m_next;
1580 	}
1581 
1582 	return (mfirst);
1583 }
1584 
1585 /*
1586  * Routine to pad mbuf to the specified length 'padto'.
1587  */
1588 int
1589 m_devpad(struct mbuf *m, int padto)
1590 {
1591 	struct mbuf *last = NULL;
1592 	int padlen;
1593 
1594 	if (padto <= m->m_pkthdr.len)
1595 		return 0;
1596 
1597 	padlen = padto - m->m_pkthdr.len;
1598 
1599 	/* if there's only the packet-header and we can pad there, use it. */
1600 	if (m->m_pkthdr.len == m->m_len && M_TRAILINGSPACE(m) >= padlen) {
1601 		last = m;
1602 	} else {
1603 		/*
1604 		 * Walk packet chain to find last mbuf. We will either
1605 		 * pad there, or append a new mbuf and pad it
1606 		 */
1607 		for (last = m; last->m_next != NULL; last = last->m_next)
1608 			; /* EMPTY */
1609 
1610 		/* `last' now points to last in chain. */
1611 		if (M_TRAILINGSPACE(last) < padlen) {
1612 			struct mbuf *n;
1613 
1614 			/* Allocate new empty mbuf, pad it.  Compact later. */
1615 			MGET(n, MB_DONTWAIT, MT_DATA);
1616 			if (n == NULL)
1617 				return ENOBUFS;
1618 			n->m_len = 0;
1619 			last->m_next = n;
1620 			last = n;
1621 		}
1622 	}
1623 	KKASSERT(M_TRAILINGSPACE(last) >= padlen);
1624 	KKASSERT(M_WRITABLE(last));
1625 
1626 	/* Now zero the pad area */
1627 	bzero(mtod(last, char *) + last->m_len, padlen);
1628 	last->m_len += padlen;
1629 	m->m_pkthdr.len += padlen;
1630 	return 0;
1631 }
1632 
1633 /*
1634  * Copy data from a buffer back into the indicated mbuf chain,
1635  * starting "off" bytes from the beginning, extending the mbuf
1636  * chain if necessary.
1637  */
1638 void
1639 m_copyback(struct mbuf *m0, int off, int len, caddr_t cp)
1640 {
1641 	int mlen;
1642 	struct mbuf *m = m0, *n;
1643 	int totlen = 0;
1644 
1645 	if (m0 == NULL)
1646 		return;
1647 	while (off > (mlen = m->m_len)) {
1648 		off -= mlen;
1649 		totlen += mlen;
1650 		if (m->m_next == NULL) {
1651 			n = m_getclr(MB_DONTWAIT, m->m_type);
1652 			if (n == NULL)
1653 				goto out;
1654 			n->m_len = min(MLEN, len + off);
1655 			m->m_next = n;
1656 		}
1657 		m = m->m_next;
1658 	}
1659 	while (len > 0) {
1660 		mlen = min (m->m_len - off, len);
1661 		bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen);
1662 		cp += mlen;
1663 		len -= mlen;
1664 		mlen += off;
1665 		off = 0;
1666 		totlen += mlen;
1667 		if (len == 0)
1668 			break;
1669 		if (m->m_next == NULL) {
1670 			n = m_get(MB_DONTWAIT, m->m_type);
1671 			if (n == NULL)
1672 				break;
1673 			n->m_len = min(MLEN, len);
1674 			m->m_next = n;
1675 		}
1676 		m = m->m_next;
1677 	}
1678 out:	if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1679 		m->m_pkthdr.len = totlen;
1680 }
1681 
1682 void
1683 m_print(const struct mbuf *m)
1684 {
1685 	int len;
1686 	const struct mbuf *m2;
1687 
1688 	len = m->m_pkthdr.len;
1689 	m2 = m;
1690 	while (len) {
1691 		kprintf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-");
1692 		len -= m2->m_len;
1693 		m2 = m2->m_next;
1694 	}
1695 	return;
1696 }
1697 
1698 /*
1699  * "Move" mbuf pkthdr from "from" to "to".
1700  * "from" must have M_PKTHDR set, and "to" must be empty.
1701  */
1702 void
1703 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
1704 {
1705 	KASSERT((to->m_flags & M_PKTHDR), ("m_move_pkthdr: not packet header"));
1706 
1707 	to->m_flags |= from->m_flags & M_COPYFLAGS;
1708 	to->m_pkthdr = from->m_pkthdr;		/* especially tags */
1709 	SLIST_INIT(&from->m_pkthdr.tags);	/* purge tags from src */
1710 }
1711 
1712 /*
1713  * Duplicate "from"'s mbuf pkthdr in "to".
1714  * "from" must have M_PKTHDR set, and "to" must be empty.
1715  * In particular, this does a deep copy of the packet tags.
1716  */
1717 int
1718 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
1719 {
1720 	KASSERT((to->m_flags & M_PKTHDR), ("m_dup_pkthdr: not packet header"));
1721 
1722 	to->m_flags = (from->m_flags & M_COPYFLAGS) |
1723 		      (to->m_flags & ~M_COPYFLAGS);
1724 	to->m_pkthdr = from->m_pkthdr;
1725 	SLIST_INIT(&to->m_pkthdr.tags);
1726 	return (m_tag_copy_chain(to, from, how));
1727 }
1728 
1729 /*
1730  * Defragment a mbuf chain, returning the shortest possible
1731  * chain of mbufs and clusters.  If allocation fails and
1732  * this cannot be completed, NULL will be returned, but
1733  * the passed in chain will be unchanged.  Upon success,
1734  * the original chain will be freed, and the new chain
1735  * will be returned.
1736  *
1737  * If a non-packet header is passed in, the original
1738  * mbuf (chain?) will be returned unharmed.
1739  *
1740  * m_defrag_nofree doesn't free the passed in mbuf.
1741  */
1742 struct mbuf *
1743 m_defrag(struct mbuf *m0, int how)
1744 {
1745 	struct mbuf *m_new;
1746 
1747 	if ((m_new = m_defrag_nofree(m0, how)) == NULL)
1748 		return (NULL);
1749 	if (m_new != m0)
1750 		m_freem(m0);
1751 	return (m_new);
1752 }
1753 
1754 struct mbuf *
1755 m_defrag_nofree(struct mbuf *m0, int how)
1756 {
1757 	struct mbuf	*m_new = NULL, *m_final = NULL;
1758 	int		progress = 0, length, nsize;
1759 
1760 	if (!(m0->m_flags & M_PKTHDR))
1761 		return (m0);
1762 
1763 #ifdef MBUF_STRESS_TEST
1764 	if (m_defragrandomfailures) {
1765 		int temp = karc4random() & 0xff;
1766 		if (temp == 0xba)
1767 			goto nospace;
1768 	}
1769 #endif
1770 
1771 	m_final = m_getl(m0->m_pkthdr.len, how, MT_DATA, M_PKTHDR, &nsize);
1772 	if (m_final == NULL)
1773 		goto nospace;
1774 	m_final->m_len = 0;	/* in case m0->m_pkthdr.len is zero */
1775 
1776 	if (m_dup_pkthdr(m_final, m0, how) == 0)
1777 		goto nospace;
1778 
1779 	m_new = m_final;
1780 
1781 	while (progress < m0->m_pkthdr.len) {
1782 		length = m0->m_pkthdr.len - progress;
1783 		if (length > MCLBYTES)
1784 			length = MCLBYTES;
1785 
1786 		if (m_new == NULL) {
1787 			m_new = m_getl(length, how, MT_DATA, 0, &nsize);
1788 			if (m_new == NULL)
1789 				goto nospace;
1790 		}
1791 
1792 		m_copydata(m0, progress, length, mtod(m_new, caddr_t));
1793 		progress += length;
1794 		m_new->m_len = length;
1795 		if (m_new != m_final)
1796 			m_cat(m_final, m_new);
1797 		m_new = NULL;
1798 	}
1799 	if (m0->m_next == NULL)
1800 		m_defraguseless++;
1801 	m_defragpackets++;
1802 	m_defragbytes += m_final->m_pkthdr.len;
1803 	return (m_final);
1804 nospace:
1805 	m_defragfailure++;
1806 	if (m_new)
1807 		m_free(m_new);
1808 	m_freem(m_final);
1809 	return (NULL);
1810 }
1811 
1812 /*
1813  * Move data from uio into mbufs.
1814  */
1815 struct mbuf *
1816 m_uiomove(struct uio *uio)
1817 {
1818 	struct mbuf *m;			/* current working mbuf */
1819 	struct mbuf *head = NULL;	/* result mbuf chain */
1820 	struct mbuf **mp = &head;
1821 	int flags = M_PKTHDR;
1822 	int nsize;
1823 	int error;
1824 	int resid;
1825 
1826 	do {
1827 		if (uio->uio_resid > INT_MAX)
1828 			resid = INT_MAX;
1829 		else
1830 			resid = (int)uio->uio_resid;
1831 		m = m_getl(resid, MB_WAIT, MT_DATA, flags, &nsize);
1832 		if (flags) {
1833 			m->m_pkthdr.len = 0;
1834 			/* Leave room for protocol headers. */
1835 			if (resid < MHLEN)
1836 				MH_ALIGN(m, resid);
1837 			flags = 0;
1838 		}
1839 		m->m_len = imin(nsize, resid);
1840 		error = uiomove(mtod(m, caddr_t), m->m_len, uio);
1841 		if (error) {
1842 			m_free(m);
1843 			goto failed;
1844 		}
1845 		*mp = m;
1846 		mp = &m->m_next;
1847 		head->m_pkthdr.len += m->m_len;
1848 	} while (uio->uio_resid > 0);
1849 
1850 	return (head);
1851 
1852 failed:
1853 	m_freem(head);
1854 	return (NULL);
1855 }
1856 
1857 struct mbuf *
1858 m_last(struct mbuf *m)
1859 {
1860 	while (m->m_next)
1861 		m = m->m_next;
1862 	return (m);
1863 }
1864 
1865 /*
1866  * Return the number of bytes in an mbuf chain.
1867  * If lastm is not NULL, also return the last mbuf.
1868  */
1869 u_int
1870 m_lengthm(struct mbuf *m, struct mbuf **lastm)
1871 {
1872 	u_int len = 0;
1873 	struct mbuf *prev = m;
1874 
1875 	while (m) {
1876 		len += m->m_len;
1877 		prev = m;
1878 		m = m->m_next;
1879 	}
1880 	if (lastm != NULL)
1881 		*lastm = prev;
1882 	return (len);
1883 }
1884 
1885 /*
1886  * Like m_lengthm(), except also keep track of mbuf usage.
1887  */
1888 u_int
1889 m_countm(struct mbuf *m, struct mbuf **lastm, u_int *pmbcnt)
1890 {
1891 	u_int len = 0, mbcnt = 0;
1892 	struct mbuf *prev = m;
1893 
1894 	while (m) {
1895 		len += m->m_len;
1896 		mbcnt += MSIZE;
1897 		if (m->m_flags & M_EXT)
1898 			mbcnt += m->m_ext.ext_size;
1899 		prev = m;
1900 		m = m->m_next;
1901 	}
1902 	if (lastm != NULL)
1903 		*lastm = prev;
1904 	*pmbcnt = mbcnt;
1905 	return (len);
1906 }
1907