4 * Copyright (c) 2004 Jeffrey M. Hsu. All rights reserved.
5 * Copyright (c) 2004 The DragonFly Project. All rights reserved.
7 * This code is derived from software contributed to The DragonFly Project
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
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.
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
37 * Copyright (c) 1982, 1986, 1988, 1991, 1993
38 * The Regents of the University of California. All rights reserved.
40 * Redistribution and use in source and binary forms, with or without
41 * modification, are permitted provided that the following conditions
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.
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
68 * @(#)uipc_mbuf.c 8.2 (Berkeley) 1/4/94
69 * $FreeBSD: src/sys/kern/uipc_mbuf.c,v 1.51.2.24 2003/04/15 06:59:29 silby Exp $
70 * $DragonFly: src/sys/kern/uipc_mbuf.c,v 1.70 2008/11/20 14:21:01 sephe Exp $
73 #include "opt_param.h"
74 #include "opt_mbuf_stress_test.h"
75 #include <sys/param.h>
76 #include <sys/systm.h>
78 #include <sys/malloc.h>
80 #include <sys/kernel.h>
81 #include <sys/sysctl.h>
82 #include <sys/domain.h>
83 #include <sys/objcache.h>
85 #include <sys/protosw.h>
87 #include <sys/thread.h>
88 #include <sys/globaldata.h>
90 #include <sys/thread2.h>
91 #include <sys/spinlock2.h>
93 #include <machine/atomic.h>
94 #include <machine/limits.h>
97 #include <vm/vm_kern.h>
98 #include <vm/vm_extern.h>
101 #include <machine/cpu.h>
105 * mbuf cluster meta-data
113 * mbuf tracking for debugging purposes
117 static MALLOC_DEFINE(M_MTRACK, "mtrack", "mtrack");
120 RB_HEAD(mbuf_rb_tree, mbtrack);
121 RB_PROTOTYPE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *);
124 RB_ENTRY(mbtrack) rb_node;
130 mbtrack_cmp(struct mbtrack *mb1, struct mbtrack *mb2)
139 RB_GENERATE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *, m);
141 struct mbuf_rb_tree mbuf_track_root;
142 static struct spinlock mbuf_track_spin = SPINLOCK_INITIALIZER(mbuf_track_spin);
145 mbuftrack(struct mbuf *m)
149 mbt = kmalloc(sizeof(*mbt), M_MTRACK, M_INTWAIT|M_ZERO);
150 spin_lock(&mbuf_track_spin);
152 if (mbuf_rb_tree_RB_INSERT(&mbuf_track_root, mbt)) {
153 spin_unlock(&mbuf_track_spin);
154 panic("mbuftrack: mbuf %p already being tracked\n", m);
156 spin_unlock(&mbuf_track_spin);
160 mbufuntrack(struct mbuf *m)
164 spin_lock(&mbuf_track_spin);
165 mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m);
167 spin_unlock(&mbuf_track_spin);
168 panic("mbufuntrack: mbuf %p was not tracked\n", m);
170 mbuf_rb_tree_RB_REMOVE(&mbuf_track_root, mbt);
171 spin_unlock(&mbuf_track_spin);
172 kfree(mbt, M_MTRACK);
177 mbuftrackid(struct mbuf *m, int trackid)
182 spin_lock(&mbuf_track_spin);
186 mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m);
188 spin_unlock(&mbuf_track_spin);
189 panic("mbuftrackid: mbuf %p not tracked", m);
191 mbt->trackid = trackid;
196 spin_unlock(&mbuf_track_spin);
200 mbuftrack_callback(struct mbtrack *mbt, void *arg)
202 struct sysctl_req *req = arg;
206 ksnprintf(buf, sizeof(buf), "mbuf %p track %d\n", mbt->m, mbt->trackid);
208 spin_unlock(&mbuf_track_spin);
209 error = SYSCTL_OUT(req, buf, strlen(buf));
210 spin_lock(&mbuf_track_spin);
217 mbuftrack_show(SYSCTL_HANDLER_ARGS)
221 spin_lock(&mbuf_track_spin);
222 error = mbuf_rb_tree_RB_SCAN(&mbuf_track_root, NULL,
223 mbuftrack_callback, req);
224 spin_unlock(&mbuf_track_spin);
227 SYSCTL_PROC(_kern_ipc, OID_AUTO, showmbufs, CTLFLAG_RD|CTLTYPE_STRING,
228 0, 0, mbuftrack_show, "A", "Show all in-use mbufs");
233 #define mbufuntrack(m)
237 static void mbinit(void *);
238 SYSINIT(mbuf, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, mbinit, NULL)
240 static u_long mbtypes[SMP_MAXCPU][MT_NTYPES];
242 static struct mbstat mbstat[SMP_MAXCPU];
251 #ifdef MBUF_STRESS_TEST
252 int m_defragrandomfailures;
255 struct objcache *mbuf_cache, *mbufphdr_cache;
256 struct objcache *mclmeta_cache, *mjclmeta_cache;
257 struct objcache *mbufcluster_cache, *mbufphdrcluster_cache;
258 struct objcache *mbufjcluster_cache, *mbufphdrjcluster_cache;
263 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW,
264 &max_linkhdr, 0, "Max size of a link-level header");
265 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW,
266 &max_protohdr, 0, "Max size of a protocol header");
267 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0,
268 "Max size of link+protocol headers");
269 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW,
270 &max_datalen, 0, "Max data payload size without headers");
271 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW,
272 &mbuf_wait, 0, "Time in ticks to sleep after failed mbuf allocations");
273 static int do_mbstat(SYSCTL_HANDLER_ARGS);
275 SYSCTL_PROC(_kern_ipc, KIPC_MBSTAT, mbstat, CTLTYPE_STRUCT|CTLFLAG_RD,
276 0, 0, do_mbstat, "S,mbstat", "mbuf usage statistics");
278 static int do_mbtypes(SYSCTL_HANDLER_ARGS);
280 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbtypes, CTLTYPE_ULONG|CTLFLAG_RD,
281 0, 0, do_mbtypes, "LU", "");
284 do_mbstat(SYSCTL_HANDLER_ARGS)
286 struct mbstat mbstat_total;
287 struct mbstat *mbstat_totalp;
290 bzero(&mbstat_total, sizeof(mbstat_total));
291 mbstat_totalp = &mbstat_total;
293 for (i = 0; i < ncpus; i++)
295 mbstat_total.m_mbufs += mbstat[i].m_mbufs;
296 mbstat_total.m_clusters += mbstat[i].m_clusters;
297 mbstat_total.m_spare += mbstat[i].m_spare;
298 mbstat_total.m_clfree += mbstat[i].m_clfree;
299 mbstat_total.m_drops += mbstat[i].m_drops;
300 mbstat_total.m_wait += mbstat[i].m_wait;
301 mbstat_total.m_drain += mbstat[i].m_drain;
302 mbstat_total.m_mcfail += mbstat[i].m_mcfail;
303 mbstat_total.m_mpfail += mbstat[i].m_mpfail;
307 * The following fields are not cumulative fields so just
308 * get their values once.
310 mbstat_total.m_msize = mbstat[0].m_msize;
311 mbstat_total.m_mclbytes = mbstat[0].m_mclbytes;
312 mbstat_total.m_minclsize = mbstat[0].m_minclsize;
313 mbstat_total.m_mlen = mbstat[0].m_mlen;
314 mbstat_total.m_mhlen = mbstat[0].m_mhlen;
316 return(sysctl_handle_opaque(oidp, mbstat_totalp, sizeof(mbstat_total), req));
320 do_mbtypes(SYSCTL_HANDLER_ARGS)
322 u_long totals[MT_NTYPES];
325 for (i = 0; i < MT_NTYPES; i++)
328 for (i = 0; i < ncpus; i++)
330 for (j = 0; j < MT_NTYPES; j++)
331 totals[j] += mbtypes[i][j];
334 return(sysctl_handle_opaque(oidp, totals, sizeof(totals), req));
338 * These are read-only because we do not currently have any code
339 * to adjust the objcache limits after the fact. The variables
340 * may only be set as boot-time tunables.
342 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RD,
343 &nmbclusters, 0, "Maximum number of mbuf clusters available");
344 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RD, &nmbufs, 0,
345 "Maximum number of mbufs available");
347 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
348 &m_defragpackets, 0, "Number of defragment packets");
349 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
350 &m_defragbytes, 0, "Number of defragment bytes");
351 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
352 &m_defraguseless, 0, "Number of useless defragment mbuf chain operations");
353 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
354 &m_defragfailure, 0, "Number of failed defragment mbuf chain operations");
355 #ifdef MBUF_STRESS_TEST
356 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
357 &m_defragrandomfailures, 0, "");
360 static MALLOC_DEFINE(M_MBUF, "mbuf", "mbuf");
361 static MALLOC_DEFINE(M_MBUFCL, "mbufcl", "mbufcl");
362 static MALLOC_DEFINE(M_MJBUFCL, "mbufcl", "mbufcl");
363 static MALLOC_DEFINE(M_MCLMETA, "mclmeta", "mclmeta");
364 static MALLOC_DEFINE(M_MJCLMETA, "mjclmeta", "mjclmeta");
366 static void m_reclaim (void);
367 static void m_mclref(void *arg);
368 static void m_mclfree(void *arg);
371 * NOTE: Default NMBUFS must take into account a possible DOS attack
372 * using fd passing on unix domain sockets.
375 #define NMBCLUSTERS (512 + maxusers * 16)
378 #define NMBUFS (nmbclusters * 2 + maxfiles)
382 * Perform sanity checks of tunables declared above.
385 tunable_mbinit(void *dummy)
388 * This has to be done before VM init.
390 nmbclusters = NMBCLUSTERS;
391 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
393 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
395 if (nmbufs < nmbclusters * 2)
396 nmbufs = nmbclusters * 2;
398 SYSINIT(tunable_mbinit, SI_BOOT1_TUNABLES, SI_ORDER_ANY,
399 tunable_mbinit, NULL);
401 /* "number of clusters of pages" */
407 * The mbuf object cache only guarantees that m_next and m_nextpkt are
408 * NULL and that m_data points to the beginning of the data area. In
409 * particular, m_len and m_pkthdr.len are uninitialized. It is the
410 * responsibility of the caller to initialize those fields before use.
413 static boolean_t __inline
414 mbuf_ctor(void *obj, void *private, int ocflags)
416 struct mbuf *m = obj;
420 m->m_data = m->m_dat;
427 * Initialize the mbuf and the packet header fields.
430 mbufphdr_ctor(void *obj, void *private, int ocflags)
432 struct mbuf *m = obj;
436 m->m_data = m->m_pktdat;
437 m->m_flags = M_PKTHDR | M_PHCACHE;
439 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */
440 SLIST_INIT(&m->m_pkthdr.tags);
441 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */
442 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */
448 * A mbcluster object consists of 2K (MCLBYTES) cluster and a refcount.
451 mclmeta_ctor(void *obj, void *private, int ocflags)
453 struct mbcluster *cl = obj;
456 if (ocflags & M_NOWAIT)
457 buf = kmalloc(MCLBYTES, M_MBUFCL, M_NOWAIT | M_ZERO);
459 buf = kmalloc(MCLBYTES, M_MBUFCL, M_INTWAIT | M_ZERO);
468 mjclmeta_ctor(void *obj, void *private, int ocflags)
470 struct mbcluster *cl = obj;
473 if (ocflags & M_NOWAIT)
474 buf = kmalloc(MJUMPAGESIZE, M_MBUFCL, M_NOWAIT | M_ZERO);
476 buf = kmalloc(MJUMPAGESIZE, M_MBUFCL, M_INTWAIT | M_ZERO);
485 mclmeta_dtor(void *obj, void *private)
487 struct mbcluster *mcl = obj;
489 KKASSERT(mcl->mcl_refs == 0);
490 kfree(mcl->mcl_data, M_MBUFCL);
494 linkjcluster(struct mbuf *m, struct mbcluster *cl, uint size)
497 * Add the cluster to the mbuf. The caller will detect that the
498 * mbuf now has an attached cluster.
500 m->m_ext.ext_arg = cl;
501 m->m_ext.ext_buf = cl->mcl_data;
502 m->m_ext.ext_ref = m_mclref;
503 m->m_ext.ext_free = m_mclfree;
504 m->m_ext.ext_size = size;
505 atomic_add_int(&cl->mcl_refs, 1);
507 m->m_data = m->m_ext.ext_buf;
508 m->m_flags |= M_EXT | M_EXT_CLUSTER;
512 linkcluster(struct mbuf *m, struct mbcluster *cl)
514 linkjcluster(m, cl, MCLBYTES);
518 mbufphdrcluster_ctor(void *obj, void *private, int ocflags)
520 struct mbuf *m = obj;
521 struct mbcluster *cl;
523 mbufphdr_ctor(obj, private, ocflags);
524 cl = objcache_get(mclmeta_cache, ocflags);
526 ++mbstat[mycpu->gd_cpuid].m_drops;
529 m->m_flags |= M_CLCACHE;
535 mbufphdrjcluster_ctor(void *obj, void *private, int ocflags)
537 struct mbuf *m = obj;
538 struct mbcluster *cl;
540 mbufphdr_ctor(obj, private, ocflags);
541 cl = objcache_get(mjclmeta_cache, ocflags);
543 ++mbstat[mycpu->gd_cpuid].m_drops;
546 m->m_flags |= M_CLCACHE;
547 linkjcluster(m, cl, MJUMPAGESIZE);
552 mbufcluster_ctor(void *obj, void *private, int ocflags)
554 struct mbuf *m = obj;
555 struct mbcluster *cl;
557 mbuf_ctor(obj, private, ocflags);
558 cl = objcache_get(mclmeta_cache, ocflags);
560 ++mbstat[mycpu->gd_cpuid].m_drops;
563 m->m_flags |= M_CLCACHE;
569 mbufjcluster_ctor(void *obj, void *private, int ocflags)
571 struct mbuf *m = obj;
572 struct mbcluster *cl;
574 mbuf_ctor(obj, private, ocflags);
575 cl = objcache_get(mjclmeta_cache, ocflags);
577 ++mbstat[mycpu->gd_cpuid].m_drops;
580 m->m_flags |= M_CLCACHE;
581 linkjcluster(m, cl, MJUMPAGESIZE);
586 * Used for both the cluster and cluster PHDR caches.
588 * The mbuf may have lost its cluster due to sharing, deal
589 * with the situation by checking M_EXT.
592 mbufcluster_dtor(void *obj, void *private)
594 struct mbuf *m = obj;
595 struct mbcluster *mcl;
597 if (m->m_flags & M_EXT) {
598 KKASSERT((m->m_flags & M_EXT_CLUSTER) != 0);
599 mcl = m->m_ext.ext_arg;
600 KKASSERT(mcl->mcl_refs == 1);
602 if (m->m_flags & M_EXT && m->m_ext.ext_size != MCLBYTES)
603 objcache_put(mjclmeta_cache, mcl);
605 objcache_put(mclmeta_cache, mcl);
609 struct objcache_malloc_args mbuf_malloc_args = { MSIZE, M_MBUF };
610 struct objcache_malloc_args mclmeta_malloc_args =
611 { sizeof(struct mbcluster), M_MCLMETA };
617 int mb_limit, cl_limit;
622 * Initialize statistics
624 for (i = 0; i < ncpus; i++) {
625 atomic_set_long_nonlocked(&mbstat[i].m_msize, MSIZE);
626 atomic_set_long_nonlocked(&mbstat[i].m_mclbytes, MCLBYTES);
627 atomic_set_long_nonlocked(&mbstat[i].m_mjumpagesize, MJUMPAGESIZE);
628 atomic_set_long_nonlocked(&mbstat[i].m_minclsize, MINCLSIZE);
629 atomic_set_long_nonlocked(&mbstat[i].m_mlen, MLEN);
630 atomic_set_long_nonlocked(&mbstat[i].m_mhlen, MHLEN);
634 * Create objtect caches and save cluster limits, which will
635 * be used to adjust backing kmalloc pools' limit later.
638 mb_limit = cl_limit = 0;
641 mbuf_cache = objcache_create("mbuf", &limit, 0,
642 mbuf_ctor, NULL, NULL,
643 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
647 mbufphdr_cache = objcache_create("mbuf pkt hdr", &limit, 128,
648 mbufphdr_ctor, NULL, NULL,
649 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
652 cl_limit = nmbclusters;
653 mclmeta_cache = objcache_create("cluster mbuf", &cl_limit, 0,
654 mclmeta_ctor, mclmeta_dtor, NULL,
655 objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args);
657 cl_limit = nmbclusters;
658 mjclmeta_cache = objcache_create("jcluster mbuf", &cl_limit, 0,
659 mjclmeta_ctor, mclmeta_dtor, NULL,
660 objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args);
663 mbufcluster_cache = objcache_create("mbuf + cluster", &limit, 0,
664 mbufcluster_ctor, mbufcluster_dtor, NULL,
665 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
669 mbufphdrcluster_cache = objcache_create("mbuf pkt hdr + cluster",
670 &limit, 128, mbufphdrcluster_ctor, mbufcluster_dtor, NULL,
671 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
675 mbufjcluster_cache = objcache_create("mbuf + jcluster", &limit, 0,
676 mbufjcluster_ctor, mbufcluster_dtor, NULL,
677 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
681 mbufphdrjcluster_cache = objcache_create("mbuf pkt hdr + jcluster",
682 &limit, 64, mbufphdrjcluster_ctor, mbufcluster_dtor, NULL,
683 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
687 * Adjust backing kmalloc pools' limit
689 * NOTE: We raise the limit by another 1/8 to take the effect
690 * of loosememuse into account.
692 cl_limit += cl_limit / 8;
693 kmalloc_raise_limit(mclmeta_malloc_args.mtype,
694 mclmeta_malloc_args.objsize * cl_limit);
695 kmalloc_raise_limit(M_MBUFCL, MCLBYTES * cl_limit * 3/4 + MJUMPAGESIZE * cl_limit / 4);
696 /*kmalloc_raise_limit(M_MBUFCL, MCLBYTES * cl_limit);*/
698 mb_limit += mb_limit / 8;
699 kmalloc_raise_limit(mbuf_malloc_args.mtype,
700 mbuf_malloc_args.objsize * mb_limit);
704 * Return the number of references to this mbuf's data. 0 is returned
705 * if the mbuf is not M_EXT, a reference count is returned if it is
706 * M_EXT | M_EXT_CLUSTER, and 99 is returned if it is a special M_EXT.
709 m_sharecount(struct mbuf *m)
711 switch (m->m_flags & (M_EXT | M_EXT_CLUSTER)) {
716 case M_EXT | M_EXT_CLUSTER:
717 return (((struct mbcluster *)m->m_ext.ext_arg)->mcl_refs);
720 return (0); /* to shut up compiler */
724 * change mbuf to new type
727 m_chtype(struct mbuf *m, int type)
729 struct globaldata *gd = mycpu;
731 atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1);
732 atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1);
733 atomic_set_short_nonlocked(&m->m_type, type);
742 kprintf("Debug: m_reclaim() called\n");
744 SLIST_FOREACH(dp, &domains, dom_next) {
745 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) {
750 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_drain, 1);
754 updatestats(struct mbuf *m, int type)
756 struct globaldata *gd = mycpu;
761 KASSERT(m->m_next == NULL, ("mbuf %p: bad m_next in get", m));
762 KASSERT(m->m_nextpkt == NULL, ("mbuf %p: bad m_nextpkt in get", m));
765 atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1);
766 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1);
774 m_get(int how, int type)
778 int ocf = MBTOM(how);
782 m = objcache_get(mbuf_cache, ocf);
785 if ((how & MB_TRYWAIT) && ntries++ == 0) {
786 struct objcache *reclaimlist[] = {
789 mbufphdrcluster_cache,
791 mbufphdrjcluster_cache
793 const int nreclaims = NELEM(reclaimlist);
795 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
799 ++mbstat[mycpu->gd_cpuid].m_drops;
803 KASSERT(m->m_data == m->m_dat, ("mbuf %p: bad m_data in get", m));
807 updatestats(m, type);
812 m_gethdr(int how, int type)
815 int ocf = MBTOM(how);
820 m = objcache_get(mbufphdr_cache, ocf);
823 if ((how & MB_TRYWAIT) && ntries++ == 0) {
824 struct objcache *reclaimlist[] = {
826 mbufcluster_cache, mbufphdrcluster_cache,
827 mbufjcluster_cache, mbufphdrjcluster_cache
829 const int nreclaims = NELEM(reclaimlist);
831 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
835 ++mbstat[mycpu->gd_cpuid].m_drops;
839 KASSERT(m->m_data == m->m_pktdat, ("mbuf %p: bad m_data in get", m));
844 updatestats(m, type);
849 * Get a mbuf (not a mbuf cluster!) and zero it.
853 m_getclr(int how, int type)
857 m = m_get(how, type);
859 bzero(m->m_data, MLEN);
864 m_getjcl(int how, short type, int flags, size_t size)
866 struct mbuf *m = NULL;
867 struct objcache *mbclc, *mbphclc;
868 int ocflags = MBTOM(how);
873 mbclc = mbufcluster_cache;
874 mbphclc = mbufphdrcluster_cache;
877 mbclc = mbufjcluster_cache;
878 mbphclc = mbufphdrjcluster_cache;
884 if (flags & M_PKTHDR)
885 m = objcache_get(mbphclc, ocflags);
887 m = objcache_get(mbclc, ocflags);
890 if ((how & MB_TRYWAIT) && ntries++ == 0) {
891 struct objcache *reclaimlist[1];
893 if (flags & M_PKTHDR)
894 reclaimlist[0] = mbclc;
896 reclaimlist[0] = mbphclc;
897 if (!objcache_reclaimlist(reclaimlist, 1, ocflags))
901 ++mbstat[mycpu->gd_cpuid].m_drops;
906 KASSERT(m->m_data == m->m_ext.ext_buf,
907 ("mbuf %p: bad m_data in get", m));
911 m->m_pkthdr.len = 0; /* just do it unconditonally */
915 atomic_add_long_nonlocked(&mbtypes[mycpu->gd_cpuid][type], 1);
916 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
921 * Returns an mbuf with an attached cluster.
922 * Because many network drivers use this kind of buffers a lot, it is
923 * convenient to keep a small pool of free buffers of this kind.
924 * Even a small size such as 10 gives about 10% improvement in the
925 * forwarding rate in a bridge or router.
928 m_getcl(int how, short type, int flags)
930 return (m_getjcl(how, type, flags, MCLBYTES));
934 * Allocate chain of requested length.
937 m_getc(int len, int how, int type)
939 struct mbuf *n, *nfirst = NULL, **ntail = &nfirst;
943 n = m_getl(len, how, type, 0, &nsize);
959 * Allocate len-worth of mbufs and/or mbuf clusters (whatever fits best)
960 * and return a pointer to the head of the allocated chain. If m0 is
961 * non-null, then we assume that it is a single mbuf or an mbuf chain to
962 * which we want len bytes worth of mbufs and/or clusters attached, and so
963 * if we succeed in allocating it, we will just return a pointer to m0.
965 * If we happen to fail at any point during the allocation, we will free
966 * up everything we have already allocated and return NULL.
968 * Deprecated. Use m_getc() and m_cat() instead.
971 m_getm(struct mbuf *m0, int len, int type, int how)
975 nfirst = m_getc(len, how, type);
978 m_last(m0)->m_next = nfirst;
986 * Adds a cluster to a normal mbuf, M_EXT is set on success.
987 * Deprecated. Use m_getcl() instead.
990 m_mclget(struct mbuf *m, int how)
992 struct mbcluster *mcl;
994 KKASSERT((m->m_flags & M_EXT) == 0);
995 mcl = objcache_get(mclmeta_cache, MBTOM(how));
998 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters,
1001 ++mbstat[mycpu->gd_cpuid].m_drops;
1006 * Updates to mbcluster must be MPSAFE. Only an entity which already has
1007 * a reference to the cluster can ref it, so we are in no danger of
1008 * racing an add with a subtract. But the operation must still be atomic
1009 * since multiple entities may have a reference on the cluster.
1011 * m_mclfree() is almost the same but it must contend with two entities
1012 * freeing the cluster at the same time.
1017 struct mbcluster *mcl = arg;
1019 atomic_add_int(&mcl->mcl_refs, 1);
1023 * When dereferencing a cluster we have to deal with a N->0 race, where
1024 * N entities free their references simultaniously. To do this we use
1025 * atomic_fetchadd_int().
1028 m_mclfree(void *arg)
1030 struct mbcluster *mcl = arg;
1032 if (atomic_fetchadd_int(&mcl->mcl_refs, -1) == 1)
1033 objcache_put(mclmeta_cache, mcl);
1037 * Free a single mbuf and any associated external storage. The successor,
1038 * if any, is returned.
1040 * We do need to check non-first mbuf for m_aux, since some of existing
1041 * code does not call M_PREPEND properly.
1042 * (example: call to bpf_mtap from drivers)
1048 _m_free(struct mbuf *m, const char *func)
1053 m_free(struct mbuf *m)
1058 struct globaldata *gd = mycpu;
1060 KASSERT(m->m_type != MT_FREE, ("freeing free mbuf %p", m));
1061 KASSERT(M_TRAILINGSPACE(m) >= 0, ("overflowed mbuf %p", m));
1062 atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1);
1067 * Make sure the mbuf is in constructed state before returning it
1073 m->m_hdr.mh_lastfunc = func;
1076 KKASSERT(m->m_nextpkt == NULL);
1078 if (m->m_nextpkt != NULL) {
1079 static int afewtimes = 10;
1081 if (afewtimes-- > 0) {
1082 kprintf("mfree: m->m_nextpkt != NULL\n");
1083 print_backtrace(-1);
1085 m->m_nextpkt = NULL;
1088 if (m->m_flags & M_PKTHDR) {
1089 m_tag_delete_chain(m); /* eliminate XXX JH */
1092 m->m_flags &= (M_EXT | M_EXT_CLUSTER | M_CLCACHE | M_PHCACHE);
1095 * Clean the M_PKTHDR state so we can return the mbuf to its original
1096 * cache. This is based on the PHCACHE flag which tells us whether
1097 * the mbuf was originally allocated out of a packet-header cache
1098 * or a non-packet-header cache.
1100 if (m->m_flags & M_PHCACHE) {
1101 m->m_flags |= M_PKTHDR;
1102 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */
1103 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */
1104 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */
1105 SLIST_INIT(&m->m_pkthdr.tags);
1109 * Handle remaining flags combinations. M_CLCACHE tells us whether
1110 * the mbuf was originally allocated from a cluster cache or not,
1111 * and is totally separate from whether the mbuf is currently
1112 * associated with a cluster.
1114 switch(m->m_flags & (M_CLCACHE | M_EXT | M_EXT_CLUSTER)) {
1115 case M_CLCACHE | M_EXT | M_EXT_CLUSTER:
1117 * mbuf+cluster cache case. The mbuf was allocated from the
1118 * combined mbuf_cluster cache and can be returned to the
1119 * cache if the cluster hasn't been shared.
1121 if (m_sharecount(m) == 1) {
1123 * The cluster has not been shared, we can just
1124 * reset the data pointer and return the mbuf
1125 * to the cluster cache. Note that the reference
1126 * count is left intact (it is still associated with
1129 m->m_data = m->m_ext.ext_buf;
1130 if (m->m_flags & M_EXT && m->m_ext.ext_size != MCLBYTES) {
1131 if (m->m_flags & M_PHCACHE)
1132 objcache_put(mbufphdrjcluster_cache, m);
1134 objcache_put(mbufjcluster_cache, m);
1136 if (m->m_flags & M_PHCACHE)
1137 objcache_put(mbufphdrcluster_cache, m);
1139 objcache_put(mbufcluster_cache, m);
1141 atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
1144 * Hell. Someone else has a ref on this cluster,
1145 * we have to disconnect it which means we can't
1146 * put it back into the mbufcluster_cache, we
1147 * have to destroy the mbuf.
1149 * Other mbuf references to the cluster will typically
1150 * be M_EXT | M_EXT_CLUSTER but without M_CLCACHE.
1152 * XXX we could try to connect another cluster to
1156 m->m_ext.ext_free(m->m_ext.ext_arg);
1157 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
1158 if (m->m_ext.ext_size == MCLBYTES) {
1159 if (m->m_flags & M_PHCACHE)
1160 objcache_dtor(mbufphdrcluster_cache, m);
1162 objcache_dtor(mbufcluster_cache, m);
1164 if (m->m_flags & M_PHCACHE)
1165 objcache_dtor(mbufphdrjcluster_cache, m);
1167 objcache_dtor(mbufjcluster_cache, m);
1171 case M_EXT | M_EXT_CLUSTER:
1173 * Normal cluster associated with an mbuf that was allocated
1174 * from the normal mbuf pool rather then the cluster pool.
1175 * The cluster has to be independantly disassociated from the
1178 if (m_sharecount(m) == 1)
1179 atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
1183 * Normal cluster association case, disconnect the cluster from
1184 * the mbuf. The cluster may or may not be custom.
1186 m->m_ext.ext_free(m->m_ext.ext_arg);
1187 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
1191 * return the mbuf to the mbuf cache.
1193 if (m->m_flags & M_PHCACHE) {
1194 m->m_data = m->m_pktdat;
1195 objcache_put(mbufphdr_cache, m);
1197 m->m_data = m->m_dat;
1198 objcache_put(mbuf_cache, m);
1200 atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1);
1204 panic("bad mbuf flags %p %08x\n", m, m->m_flags);
1213 _m_freem(struct mbuf *m, const char *func)
1216 m = _m_free(m, func);
1222 m_freem(struct mbuf *m)
1231 * mbuf utility routines
1235 * Lesser-used path for M_PREPEND: allocate new mbuf to prepend to chain and
1239 m_prepend(struct mbuf *m, int len, int how)
1243 if (m->m_flags & M_PKTHDR)
1244 mn = m_gethdr(how, m->m_type);
1246 mn = m_get(how, m->m_type);
1251 if (m->m_flags & M_PKTHDR)
1252 M_MOVE_PKTHDR(mn, m);
1262 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
1263 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
1264 * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller.
1265 * Note that the copy is read-only, because clusters are not copied,
1266 * only their reference counts are incremented.
1269 m_copym(const struct mbuf *m, int off0, int len, int wait)
1271 struct mbuf *n, **np;
1276 KASSERT(off >= 0, ("m_copym, negative off %d", off));
1277 KASSERT(len >= 0, ("m_copym, negative len %d", len));
1278 if (off == 0 && (m->m_flags & M_PKTHDR))
1281 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
1291 KASSERT(len == M_COPYALL,
1292 ("m_copym, length > size of mbuf chain"));
1296 * Because we are sharing any cluster attachment below,
1297 * be sure to get an mbuf that does not have a cluster
1298 * associated with it.
1301 n = m_gethdr(wait, m->m_type);
1303 n = m_get(wait, m->m_type);
1308 if (!m_dup_pkthdr(n, m, wait))
1310 if (len == M_COPYALL)
1311 n->m_pkthdr.len -= off0;
1313 n->m_pkthdr.len = len;
1316 n->m_len = min(len, m->m_len - off);
1317 if (m->m_flags & M_EXT) {
1318 KKASSERT((n->m_flags & M_EXT) == 0);
1319 n->m_data = m->m_data + off;
1320 m->m_ext.ext_ref(m->m_ext.ext_arg);
1321 n->m_ext = m->m_ext;
1322 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1324 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1325 (unsigned)n->m_len);
1327 if (len != M_COPYALL)
1334 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1338 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1343 * Copy an entire packet, including header (which must be present).
1344 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
1345 * Note that the copy is read-only, because clusters are not copied,
1346 * only their reference counts are incremented.
1347 * Preserve alignment of the first mbuf so if the creator has left
1348 * some room at the beginning (e.g. for inserting protocol headers)
1349 * the copies also have the room available.
1352 m_copypacket(struct mbuf *m, int how)
1354 struct mbuf *top, *n, *o;
1356 n = m_gethdr(how, m->m_type);
1361 if (!m_dup_pkthdr(n, m, how))
1363 n->m_len = m->m_len;
1364 if (m->m_flags & M_EXT) {
1365 KKASSERT((n->m_flags & M_EXT) == 0);
1366 n->m_data = m->m_data;
1367 m->m_ext.ext_ref(m->m_ext.ext_arg);
1368 n->m_ext = m->m_ext;
1369 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1371 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
1372 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1377 o = m_get(how, m->m_type);
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);
1392 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1400 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1405 * Copy data from an mbuf chain starting "off" bytes from the beginning,
1406 * continuing for "len" bytes, into the indicated buffer.
1409 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
1413 KASSERT(off >= 0, ("m_copydata, negative off %d", off));
1414 KASSERT(len >= 0, ("m_copydata, negative len %d", len));
1416 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
1423 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
1424 count = min(m->m_len - off, len);
1425 bcopy(mtod(m, caddr_t) + off, cp, count);
1434 * Copy a packet header mbuf chain into a completely new chain, including
1435 * copying any mbuf clusters. Use this instead of m_copypacket() when
1436 * you need a writable copy of an mbuf chain.
1439 m_dup(struct mbuf *m, int how)
1441 struct mbuf **p, *top = NULL;
1442 int remain, moff, nsize;
1447 KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __func__));
1449 /* While there's more data, get a new mbuf, tack it on, and fill it */
1450 remain = m->m_pkthdr.len;
1453 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */
1456 /* Get the next new mbuf */
1457 n = m_getl(remain, how, m->m_type, top == NULL ? M_PKTHDR : 0,
1462 if (!m_dup_pkthdr(n, m, how))
1465 /* Link it into the new chain */
1469 /* Copy data from original mbuf(s) into new mbuf */
1471 while (n->m_len < nsize && m != NULL) {
1472 int chunk = min(nsize - n->m_len, m->m_len - moff);
1474 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1478 if (moff == m->m_len) {
1484 /* Check correct total mbuf length */
1485 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
1486 ("%s: bogus m_pkthdr.len", __func__));
1493 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1498 * Copy the non-packet mbuf data chain into a new set of mbufs, including
1499 * copying any mbuf clusters. This is typically used to realign a data
1500 * chain by nfs_realign().
1502 * The original chain is left intact. how should be MB_WAIT or MB_DONTWAIT
1503 * and NULL can be returned if MB_DONTWAIT is passed.
1505 * Be careful to use cluster mbufs, a large mbuf chain converted to non
1506 * cluster mbufs can exhaust our supply of mbufs.
1509 m_dup_data(struct mbuf *m, int how)
1511 struct mbuf **p, *n, *top = NULL;
1512 int mlen, moff, chunk, gsize, nsize;
1521 * Optimize the mbuf allocation but do not get too carried away.
1523 if (m->m_next || m->m_len > MLEN)
1524 if (m->m_flags & M_EXT && m->m_ext.ext_size == MCLBYTES)
1527 gsize = MJUMPAGESIZE;
1537 * Scan the mbuf chain until nothing is left, the new mbuf chain
1538 * will be allocated on the fly as needed.
1545 KKASSERT(m->m_type == MT_DATA);
1547 n = m_getl(gsize, how, MT_DATA, 0, &nsize);
1554 chunk = imin(mlen, nsize);
1555 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1570 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1575 * Concatenate mbuf chain n to m.
1576 * Both chains must be of the same type (e.g. MT_DATA).
1577 * Any m_pkthdr is not updated.
1580 m_cat(struct mbuf *m, struct mbuf *n)
1584 if (m->m_flags & M_EXT ||
1585 m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) {
1586 /* just join the two chains */
1590 /* splat the data from one into the other */
1591 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1593 m->m_len += n->m_len;
1599 m_adj(struct mbuf *mp, int req_len)
1605 if ((m = mp) == NULL)
1611 while (m != NULL && len > 0) {
1612 if (m->m_len <= len) {
1623 if (mp->m_flags & M_PKTHDR)
1624 m->m_pkthdr.len -= (req_len - len);
1627 * Trim from tail. Scan the mbuf chain,
1628 * calculating its length and finding the last mbuf.
1629 * If the adjustment only affects this mbuf, then just
1630 * adjust and return. Otherwise, rescan and truncate
1631 * after the remaining size.
1637 if (m->m_next == NULL)
1641 if (m->m_len >= len) {
1643 if (mp->m_flags & M_PKTHDR)
1644 mp->m_pkthdr.len -= len;
1651 * Correct length for chain is "count".
1652 * Find the mbuf with last data, adjust its length,
1653 * and toss data from remaining mbufs on chain.
1656 if (m->m_flags & M_PKTHDR)
1657 m->m_pkthdr.len = count;
1658 for (; m; m = m->m_next) {
1659 if (m->m_len >= count) {
1666 (m = m->m_next) ->m_len = 0;
1671 * Set the m_data pointer of a newly-allocated mbuf
1672 * to place an object of the specified size at the
1673 * end of the mbuf, longword aligned.
1676 m_align(struct mbuf *m, int len)
1680 if (m->m_flags & M_EXT)
1681 adjust = m->m_ext.ext_size - len;
1682 else if (m->m_flags & M_PKTHDR)
1683 adjust = MHLEN - len;
1685 adjust = MLEN - len;
1686 m->m_data += adjust &~ (sizeof(long)-1);
1690 * Rearrange an mbuf chain so that len bytes are contiguous
1691 * and in the data area of an mbuf (so that mtod will work for a structure
1692 * of size len). Returns the resulting mbuf chain on success, frees it and
1693 * returns null on failure. If there is room, it will add up to
1694 * max_protohdr-len extra bytes to the contiguous region in an attempt to
1695 * avoid being called next time.
1698 m_pullup(struct mbuf *n, int len)
1705 * If first mbuf has no cluster, and has room for len bytes
1706 * without shifting current data, pullup into it,
1707 * otherwise allocate a new mbuf to prepend to the chain.
1709 if (!(n->m_flags & M_EXT) &&
1710 n->m_data + len < &n->m_dat[MLEN] &&
1712 if (n->m_len >= len)
1720 if (n->m_flags & M_PKTHDR)
1721 m = m_gethdr(MB_DONTWAIT, n->m_type);
1723 m = m_get(MB_DONTWAIT, n->m_type);
1727 if (n->m_flags & M_PKTHDR)
1728 M_MOVE_PKTHDR(m, n);
1730 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
1732 count = min(min(max(len, max_protohdr), space), n->m_len);
1733 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1743 } while (len > 0 && n);
1752 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1757 * Partition an mbuf chain in two pieces, returning the tail --
1758 * all but the first len0 bytes. In case of failure, it returns NULL and
1759 * attempts to restore the chain to its original state.
1761 * Note that the resulting mbufs might be read-only, because the new
1762 * mbuf can end up sharing an mbuf cluster with the original mbuf if
1763 * the "breaking point" happens to lie within a cluster mbuf. Use the
1764 * M_WRITABLE() macro to check for this case.
1767 m_split(struct mbuf *m0, int len0, int wait)
1770 unsigned len = len0, remain;
1772 for (m = m0; m && len > m->m_len; m = m->m_next)
1776 remain = m->m_len - len;
1777 if (m0->m_flags & M_PKTHDR) {
1778 n = m_gethdr(wait, m0->m_type);
1781 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1782 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1783 m0->m_pkthdr.len = len0;
1784 if (m->m_flags & M_EXT)
1786 if (remain > MHLEN) {
1787 /* m can't be the lead packet */
1789 n->m_next = m_split(m, len, wait);
1790 if (n->m_next == NULL) {
1798 MH_ALIGN(n, remain);
1799 } else if (remain == 0) {
1804 n = m_get(wait, m->m_type);
1810 if (m->m_flags & M_EXT) {
1811 KKASSERT((n->m_flags & M_EXT) == 0);
1812 n->m_data = m->m_data + len;
1813 m->m_ext.ext_ref(m->m_ext.ext_arg);
1814 n->m_ext = m->m_ext;
1815 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1817 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1821 n->m_next = m->m_next;
1827 * Routine to copy from device local memory into mbufs.
1828 * Note: "offset" is ill-defined and always called as 0, so ignore it.
1831 m_devget(char *buf, int len, int offset, struct ifnet *ifp,
1832 void (*copy)(volatile const void *from, volatile void *to, size_t length))
1834 struct mbuf *m, *mfirst = NULL, **mtail;
1843 m = m_getl(len, MB_DONTWAIT, MT_DATA, flags, &nsize);
1848 m->m_len = min(len, nsize);
1850 if (flags & M_PKTHDR) {
1851 if (len + max_linkhdr <= nsize)
1852 m->m_data += max_linkhdr;
1853 m->m_pkthdr.rcvif = ifp;
1854 m->m_pkthdr.len = len;
1858 copy(buf, m->m_data, (unsigned)m->m_len);
1869 * Routine to pad mbuf to the specified length 'padto'.
1872 m_devpad(struct mbuf *m, int padto)
1874 struct mbuf *last = NULL;
1877 if (padto <= m->m_pkthdr.len)
1880 padlen = padto - m->m_pkthdr.len;
1882 /* if there's only the packet-header and we can pad there, use it. */
1883 if (m->m_pkthdr.len == m->m_len && M_TRAILINGSPACE(m) >= padlen) {
1887 * Walk packet chain to find last mbuf. We will either
1888 * pad there, or append a new mbuf and pad it
1890 for (last = m; last->m_next != NULL; last = last->m_next)
1893 /* `last' now points to last in chain. */
1894 if (M_TRAILINGSPACE(last) < padlen) {
1897 /* Allocate new empty mbuf, pad it. Compact later. */
1898 MGET(n, MB_DONTWAIT, MT_DATA);
1906 KKASSERT(M_TRAILINGSPACE(last) >= padlen);
1907 KKASSERT(M_WRITABLE(last));
1909 /* Now zero the pad area */
1910 bzero(mtod(last, char *) + last->m_len, padlen);
1911 last->m_len += padlen;
1912 m->m_pkthdr.len += padlen;
1917 * Copy data from a buffer back into the indicated mbuf chain,
1918 * starting "off" bytes from the beginning, extending the mbuf
1919 * chain if necessary.
1922 m_copyback(struct mbuf *m0, int off, int len, caddr_t cp)
1925 struct mbuf *m = m0, *n;
1930 while (off > (mlen = m->m_len)) {
1933 if (m->m_next == NULL) {
1934 n = m_getclr(MB_DONTWAIT, m->m_type);
1937 n->m_len = min(MLEN, len + off);
1943 mlen = min (m->m_len - off, len);
1944 bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen);
1952 if (m->m_next == NULL) {
1953 n = m_get(MB_DONTWAIT, m->m_type);
1956 n->m_len = min(MLEN, len);
1961 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1962 m->m_pkthdr.len = totlen;
1966 * Append the specified data to the indicated mbuf chain,
1967 * Extend the mbuf chain if the new data does not fit in
1970 * Return 1 if able to complete the job; otherwise 0.
1973 m_append(struct mbuf *m0, int len, c_caddr_t cp)
1976 int remainder, space;
1978 for (m = m0; m->m_next != NULL; m = m->m_next)
1981 space = M_TRAILINGSPACE(m);
1984 * Copy into available space.
1986 if (space > remainder)
1988 bcopy(cp, mtod(m, caddr_t) + m->m_len, space);
1990 cp += space, remainder -= space;
1992 while (remainder > 0) {
1994 * Allocate a new mbuf; could check space
1995 * and allocate a cluster instead.
1997 n = m_get(MB_DONTWAIT, m->m_type);
2000 n->m_len = min(MLEN, remainder);
2001 bcopy(cp, mtod(n, caddr_t), n->m_len);
2002 cp += n->m_len, remainder -= n->m_len;
2006 if (m0->m_flags & M_PKTHDR)
2007 m0->m_pkthdr.len += len - remainder;
2008 return (remainder == 0);
2012 * Apply function f to the data in an mbuf chain starting "off" bytes from
2013 * the beginning, continuing for "len" bytes.
2016 m_apply(struct mbuf *m, int off, int len,
2017 int (*f)(void *, void *, u_int), void *arg)
2022 KASSERT(off >= 0, ("m_apply, negative off %d", off));
2023 KASSERT(len >= 0, ("m_apply, negative len %d", len));
2025 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
2032 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
2033 count = min(m->m_len - off, len);
2034 rval = (*f)(arg, mtod(m, caddr_t) + off, count);
2045 * Return a pointer to mbuf/offset of location in mbuf chain.
2048 m_getptr(struct mbuf *m, int loc, int *off)
2052 /* Normal end of search. */
2053 if (m->m_len > loc) {
2058 if (m->m_next == NULL) {
2060 /* Point at the end of valid data. */
2073 m_print(const struct mbuf *m)
2076 const struct mbuf *m2;
2078 len = m->m_pkthdr.len;
2081 kprintf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-");
2089 * "Move" mbuf pkthdr from "from" to "to".
2090 * "from" must have M_PKTHDR set, and "to" must be empty.
2093 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
2095 KASSERT((to->m_flags & M_PKTHDR), ("m_move_pkthdr: not packet header"));
2097 to->m_flags |= from->m_flags & M_COPYFLAGS;
2098 to->m_pkthdr = from->m_pkthdr; /* especially tags */
2099 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */
2103 * Duplicate "from"'s mbuf pkthdr in "to".
2104 * "from" must have M_PKTHDR set, and "to" must be empty.
2105 * In particular, this does a deep copy of the packet tags.
2108 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
2110 KASSERT((to->m_flags & M_PKTHDR), ("m_dup_pkthdr: not packet header"));
2112 to->m_flags = (from->m_flags & M_COPYFLAGS) |
2113 (to->m_flags & ~M_COPYFLAGS);
2114 to->m_pkthdr = from->m_pkthdr;
2115 SLIST_INIT(&to->m_pkthdr.tags);
2116 return (m_tag_copy_chain(to, from, how));
2120 * Defragment a mbuf chain, returning the shortest possible
2121 * chain of mbufs and clusters. If allocation fails and
2122 * this cannot be completed, NULL will be returned, but
2123 * the passed in chain will be unchanged. Upon success,
2124 * the original chain will be freed, and the new chain
2127 * If a non-packet header is passed in, the original
2128 * mbuf (chain?) will be returned unharmed.
2130 * m_defrag_nofree doesn't free the passed in mbuf.
2133 m_defrag(struct mbuf *m0, int how)
2137 if ((m_new = m_defrag_nofree(m0, how)) == NULL)
2145 m_defrag_nofree(struct mbuf *m0, int how)
2147 struct mbuf *m_new = NULL, *m_final = NULL;
2148 int progress = 0, length, nsize;
2150 if (!(m0->m_flags & M_PKTHDR))
2153 #ifdef MBUF_STRESS_TEST
2154 if (m_defragrandomfailures) {
2155 int temp = karc4random() & 0xff;
2161 m_final = m_getl(m0->m_pkthdr.len, how, MT_DATA, M_PKTHDR, &nsize);
2162 if (m_final == NULL)
2164 m_final->m_len = 0; /* in case m0->m_pkthdr.len is zero */
2166 if (m_dup_pkthdr(m_final, m0, how) == 0)
2171 while (progress < m0->m_pkthdr.len) {
2172 length = m0->m_pkthdr.len - progress;
2173 if (length > MCLBYTES)
2176 if (m_new == NULL) {
2177 m_new = m_getl(length, how, MT_DATA, 0, &nsize);
2182 m_copydata(m0, progress, length, mtod(m_new, caddr_t));
2184 m_new->m_len = length;
2185 if (m_new != m_final)
2186 m_cat(m_final, m_new);
2189 if (m0->m_next == NULL)
2192 m_defragbytes += m_final->m_pkthdr.len;
2203 * Move data from uio into mbufs.
2206 m_uiomove(struct uio *uio)
2208 struct mbuf *m; /* current working mbuf */
2209 struct mbuf *head = NULL; /* result mbuf chain */
2210 struct mbuf **mp = &head;
2211 int flags = M_PKTHDR;
2217 if (uio->uio_resid > INT_MAX)
2220 resid = (int)uio->uio_resid;
2221 m = m_getl(resid, MB_WAIT, MT_DATA, flags, &nsize);
2223 m->m_pkthdr.len = 0;
2224 /* Leave room for protocol headers. */
2229 m->m_len = imin(nsize, resid);
2230 error = uiomove(mtod(m, caddr_t), m->m_len, uio);
2237 head->m_pkthdr.len += m->m_len;
2238 } while (uio->uio_resid > 0);
2248 m_last(struct mbuf *m)
2256 * Return the number of bytes in an mbuf chain.
2257 * If lastm is not NULL, also return the last mbuf.
2260 m_lengthm(struct mbuf *m, struct mbuf **lastm)
2263 struct mbuf *prev = m;
2276 * Like m_lengthm(), except also keep track of mbuf usage.
2279 m_countm(struct mbuf *m, struct mbuf **lastm, u_int *pmbcnt)
2281 u_int len = 0, mbcnt = 0;
2282 struct mbuf *prev = m;
2287 if (m->m_flags & M_EXT)
2288 mbcnt += m->m_ext.ext_size;