2 * Copyright (c) 1982, 1986, 1988, 1991, 1993
3 * The Regents of the University of California. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 3. All advertising materials mentioning features or use of this software
14 * must display the following acknowledgement:
15 * This product includes software developed by the University of
16 * California, Berkeley and its contributors.
17 * 4. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * @(#)uipc_mbuf.c 8.2 (Berkeley) 1/4/94
34 * $FreeBSD: src/sys/kern/uipc_mbuf.c,v 1.51.2.24 2003/04/15 06:59:29 silby Exp $
35 * $DragonFly: src/sys/kern/uipc_mbuf.c,v 1.13 2003/10/15 16:48:03 hmp Exp $
38 #include "opt_param.h"
39 #include "opt_mbuf_stress_test.h"
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/malloc.h>
44 #include <sys/kernel.h>
45 #include <sys/sysctl.h>
46 #include <sys/domain.h>
47 #include <sys/protosw.h>
48 #include <sys/thread.h>
49 #include <sys/globaldata.h>
52 #include <vm/vm_kern.h>
53 #include <vm/vm_extern.h>
56 #include <machine/cpu.h>
59 static void mbinit (void *);
60 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbinit, NULL)
65 u_long mbtypes[MT_NTYPES];
67 union mcluster *mclfree;
76 #ifdef MBUF_STRESS_TEST
77 int m_defragrandomfailures;
82 u_int m_mballoc_wid = 0;
83 u_int m_clalloc_wid = 0;
85 SYSCTL_DECL(_kern_ipc);
86 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW,
88 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW,
89 &max_protohdr, 0, "");
90 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0, "");
91 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW,
93 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW,
95 SYSCTL_STRUCT(_kern_ipc, KIPC_MBSTAT, mbstat, CTLFLAG_RW, &mbstat, mbstat, "");
96 SYSCTL_OPAQUE(_kern_ipc, OID_AUTO, mbtypes, CTLFLAG_RD, mbtypes,
97 sizeof(mbtypes), "LU", "");
98 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RD,
99 &nmbclusters, 0, "Maximum number of mbuf clusters available");
100 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RD, &nmbufs, 0,
101 "Maximum number of mbufs available");
102 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
103 &m_defragpackets, 0, "");
104 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
105 &m_defragbytes, 0, "");
106 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
107 &m_defraguseless, 0, "");
108 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
109 &m_defragfailure, 0, "");
110 #ifdef MBUF_STRESS_TEST
111 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
112 &m_defragrandomfailures, 0, "");
115 static void m_reclaim (void);
118 #define NMBCLUSTERS (512 + maxusers * 16)
121 #define NMBUFS (nmbclusters * 4)
125 * Perform sanity checks of tunables declared above.
128 tunable_mbinit(void *dummy)
132 * This has to be done before VM init.
134 nmbclusters = NMBCLUSTERS;
135 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
137 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
139 if (nmbufs < nmbclusters * 2)
140 nmbufs = nmbclusters * 2;
144 SYSINIT(tunable_mbinit, SI_SUB_TUNABLES, SI_ORDER_ANY, tunable_mbinit, NULL);
146 /* "number of clusters of pages" */
158 mmbfree = NULL; mclfree = NULL;
159 mbstat.m_msize = MSIZE;
160 mbstat.m_mclbytes = MCLBYTES;
161 mbstat.m_minclsize = MINCLSIZE;
162 mbstat.m_mlen = MLEN;
163 mbstat.m_mhlen = MHLEN;
166 if (m_mballoc(NMB_INIT, M_DONTWAIT) == 0)
168 #if MCLBYTES <= PAGE_SIZE
169 if (m_clalloc(NCL_INIT, M_DONTWAIT) == 0)
172 /* It's OK to call contigmalloc in this context. */
173 if (m_clalloc(16, M_WAIT) == 0)
183 * Allocate at least nmb mbufs and place on mbuf free list.
184 * Must be called at splimp.
197 * If we've hit the mbuf limit, stop allocating from mb_map,
198 * (or trying to) in order to avoid dipping into the section of
199 * mb_map which we've "reserved" for clusters.
201 if ((nmb + mbstat.m_mbufs) > nmbufs)
205 * Once we run out of map space, it will be impossible to get
206 * any more (nothing is ever freed back to the map)
207 * -- however you are not dead as m_reclaim might
208 * still be able to free a substantial amount of space.
210 * XXX Furthermore, we can also work with "recycled" mbufs (when
211 * we're calling with M_WAIT the sleep procedure will be woken
212 * up when an mbuf is freed. See m_mballoc_wait()).
217 nbytes = round_page(nmb * MSIZE);
218 p = (caddr_t)kmem_malloc(mb_map, nbytes, M_NOWAIT);
219 if (p == 0 && how == M_WAIT) {
221 p = (caddr_t)kmem_malloc(mb_map, nbytes, M_WAITOK);
225 * Either the map is now full, or `how' is M_NOWAIT and there
231 nmb = nbytes / MSIZE;
232 for (i = 0; i < nmb; i++) {
233 ((struct mbuf *)p)->m_next = mmbfree;
234 mmbfree = (struct mbuf *)p;
237 mbstat.m_mbufs += nmb;
238 mbtypes[MT_FREE] += nmb;
243 * Once the mb_map has been exhausted and if the call to the allocation macros
244 * (or, in some cases, functions) is with M_WAIT, then it is necessary to rely
245 * solely on reclaimed mbufs. Here we wait for an mbuf to be freed for a
246 * designated (mbuf_wait) time.
249 m_mballoc_wait(int caller, int type)
256 if ((tsleep(&m_mballoc_wid, 0, "mballc", mbuf_wait)) == EWOULDBLOCK)
261 * Now that we (think) that we've got something, we will redo an
262 * MGET, but avoid getting into another instance of m_mballoc_wait()
263 * XXX: We retry to fetch _even_ if the sleep timed out. This is left
264 * this way, purposely, in the [unlikely] case that an mbuf was
265 * freed but the sleep was not awakened in time.
270 MGET(p, M_DONTWAIT, type);
273 MGETHDR(p, M_DONTWAIT, type);
276 panic("m_mballoc_wait: invalid caller (%d)", caller);
280 if (p != NULL) { /* We waited and got something... */
282 /* Wake up another if we have more free. */
290 #if MCLBYTES > PAGE_SIZE
291 static int i_want_my_mcl;
299 tsleep(&i_want_my_mcl, 0, "mclalloc", 0);
301 for (; i_want_my_mcl; i_want_my_mcl--) {
302 if (m_clalloc(1, M_WAIT) == 0)
303 printf("m_clalloc failed even in process context!\n");
308 static struct thread *mclallocthread;
309 static struct kproc_desc mclalloc_kp = {
314 SYSINIT(mclallocthread, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start,
319 * Allocate some number of mbuf clusters
320 * and place on cluster free list.
321 * Must be called at splimp.
334 * If we've hit the mcluster number limit, stop allocating from
335 * mb_map, (or trying to) in order to avoid dipping into the section
336 * of mb_map which we've "reserved" for mbufs.
338 if ((ncl + mbstat.m_clusters) > nmbclusters)
342 * Once we run out of map space, it will be impossible
343 * to get any more (nothing is ever freed back to the
344 * map). From this point on, we solely rely on freed
350 #if MCLBYTES > PAGE_SIZE
352 i_want_my_mcl += ncl;
353 wakeup(&i_want_my_mcl);
357 p = contigmalloc_map(MCLBYTES * ncl, M_DEVBUF, M_WAITOK, 0ul,
358 ~0ul, PAGE_SIZE, 0, mb_map);
362 p = (caddr_t)kmem_malloc(mb_map, ctob(npg),
363 how != M_WAIT ? M_NOWAIT : M_WAITOK);
364 ncl = ncl * PAGE_SIZE / MCLBYTES;
367 * Either the map is now full, or `how' is M_NOWAIT and there
371 static int last_report ; /* when we did that (in ticks) */
374 if (ticks < last_report || (ticks - last_report) >= hz) {
376 printf("All mbuf clusters exhausted, please see tuning(7).\n");
381 for (i = 0; i < ncl; i++) {
382 ((union mcluster *)p)->mcl_next = mclfree;
383 mclfree = (union mcluster *)p;
387 mbstat.m_clusters += ncl;
392 * Once the mb_map submap has been exhausted and the allocation is called with
393 * M_WAIT, we rely on the mclfree union pointers. If nothing is free, we will
394 * sleep for a designated amount of time (mbuf_wait) or until we're woken up
395 * due to sudden mcluster availability.
403 /* If in interrupt context, and INVARIANTS, maintain sanity and die. */
404 KASSERT(mycpu->gd_intr_nesting_level == 0, ("CLALLOC: CANNOT WAIT IN INTERRUPT"));
406 /* Sleep until something's available or until we expire. */
408 if ((tsleep(&m_clalloc_wid, 0, "mclalc", mbuf_wait)) == EWOULDBLOCK)
412 * Now that we (think) that we've got something, we will redo and
413 * MGET, but avoid getting into another instance of m_clalloc_wait()
415 p = m_mclalloc(M_DONTWAIT);
418 if (p != NULL) { /* We waited and got something... */
420 /* Wake up another if we have more free. */
430 * When MGET fails, ask protocols to free space when short of memory,
431 * then re-attempt to allocate an mbuf.
441 * Must only do the reclaim if not in an interrupt context.
444 KASSERT(mycpu->gd_intr_nesting_level == 0,
445 ("MBALLOC: CANNOT WAIT IN INTERRUPT"));
451 (void)m_mballoc(1, i);
460 m->m_data = m->m_dat;
465 static int last_report ; /* when we did that (in ticks) */
469 if (ticks < last_report || (ticks - last_report) >= hz) {
471 printf("All mbufs exhausted, please see tuning(7).\n");
479 * As above; retry an MGETHDR.
489 * Must only do the reclaim if not in an interrupt context.
492 KASSERT(mycpu->gd_intr_nesting_level == 0,
493 ("MBALLOC: CANNOT WAIT IN INTERRUPT"));
499 (void)m_mballoc(1, i);
508 m->m_data = m->m_pktdat;
509 m->m_flags = M_PKTHDR;
510 m->m_pkthdr.rcvif = NULL;
511 SLIST_INIT(&m->m_pkthdr.tags);
512 m->m_pkthdr.csum_flags = 0;
516 static int last_report ; /* when we did that (in ticks) */
520 if (ticks < last_report || (ticks - last_report) >= hz) {
522 printf("All mbufs exhausted, please see tuning(7).\n");
536 for (dp = domains; dp; dp = dp->dom_next)
537 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
545 * Space allocation routines.
546 * These are also available as macros
547 * for critical paths.
558 (void)m_mballoc(1, how);
567 m->m_data = m->m_dat;
572 m = m_retry(how, type);
573 if (m == NULL && how == M_WAIT)
574 m = m_mballoc_wait(MGET_C, type);
588 (void)m_mballoc(1, how);
597 m->m_data = m->m_pktdat;
598 m->m_flags = M_PKTHDR;
599 m->m_pkthdr.rcvif = NULL;
600 SLIST_INIT(&m->m_pkthdr.tags);
601 m->m_pkthdr.csum_flags = 0;
605 m = m_retryhdr(how, type);
606 if (m == NULL && how == M_WAIT)
607 m = m_mballoc_wait(MGETHDR_C, type);
621 bzero(mtod(m, caddr_t), MLEN);
626 * m_getcl() returns an mbuf with an attached cluster.
627 * Because many network drivers use this kind of buffers a lot, it is
628 * convenient to keep a small pool of free buffers of this kind.
629 * Even a small size such as 10 gives about 10% improvement in the
630 * forwarding rate in a bridge or router.
631 * The size of this free list is controlled by the sysctl variable
632 * mcl_pool_max. The list is populated on m_freem(), and used in
633 * m_getcl() if elements are available.
635 static struct mbuf *mcl_pool;
636 static int mcl_pool_now;
637 static int mcl_pool_max = 0;
639 SYSCTL_INT(_kern_ipc, OID_AUTO, mcl_pool_max, CTLFLAG_RW, &mcl_pool_max, 0,
640 "Maximum number of mbufs+cluster in free list");
641 SYSCTL_INT(_kern_ipc, OID_AUTO, mcl_pool_now, CTLFLAG_RD, &mcl_pool_now, 0,
642 "Current number of mbufs+cluster in free list");
645 m_getcl(int how, short type, int flags)
650 if (flags & M_PKTHDR) {
651 if (type == MT_DATA && mcl_pool) {
653 mcl_pool = mp->m_nextpkt;
656 mp->m_nextpkt = NULL;
657 mp->m_data = mp->m_ext.ext_buf;
658 mp->m_flags = M_PKTHDR|M_EXT;
659 mp->m_pkthdr.rcvif = NULL;
660 mp->m_pkthdr.csum_flags = 0;
663 MGETHDR(mp, how, type);
668 if ( (mp->m_flags & M_EXT) == 0) {
679 * m_getm(m, len, how, type)
681 * This will allocate len-worth of mbufs and/or mbuf clusters (whatever fits
682 * best) and return a pointer to the top of the allocated chain. If m is
683 * non-null, then we assume that it is a single mbuf or an mbuf chain to
684 * which we want len bytes worth of mbufs and/or clusters attached, and so
685 * if we succeed in allocating it, we will just return a pointer to m.
687 * If we happen to fail at any point during the allocation, we will free
688 * up everything we have already allocated and return NULL.
692 m_getm(struct mbuf *m, int len, int how, int type)
694 struct mbuf *top, *tail, *mp, *mtail = NULL;
696 KASSERT(len >= 0, ("len is < 0 in m_getm"));
701 else if (len > MINCLSIZE) {
703 if ((mp->m_flags & M_EXT) == 0) {
709 len -= M_TRAILINGSPACE(mp);
712 for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next);
724 if (len > MINCLSIZE) {
726 if ((mp->m_flags & M_EXT) == 0)
731 len -= M_TRAILINGSPACE(mp);
744 * m_mclalloc() - Allocates an mbuf cluster.
756 mp = (caddr_t)mclfree;
758 mclrefcnt[mtocl(mp)]++;
760 mclfree = ((union mcluster *)mp)->mcl_next;
766 return(m_clalloc_wait());
771 * m_mclget() - Adds a cluster to a normal mbuf, M_EXT is set on success.
774 m_mclget(struct mbuf *m, int how)
776 m->m_ext.ext_buf = m_mclalloc(how);
777 if (m->m_ext.ext_buf != NULL) {
778 m->m_data = m->m_ext.ext_buf;
780 m->m_ext.ext_free = NULL;
781 m->m_ext.ext_ref = NULL;
782 m->m_ext.ext_size = MCLBYTES;
787 _m_mclfree(caddr_t data)
789 union mcluster *mp = (union mcluster *)data;
791 KASSERT(mclrefcnt[mtocl(mp)] > 0, ("freeing free cluster"));
792 if (--mclrefcnt[mtocl(mp)] == 0) {
793 mp->mcl_next = mclfree;
801 m_mclfree(caddr_t mp)
811 * Free a single mbuf and any associated external storage. The successor,
812 * if any, is returned.
814 * We do need to check non-first mbuf for m_aux, since some of existing
815 * code does not call M_PREPEND properly.
816 * (example: call to bpf_mtap from drivers)
819 m_free(struct mbuf *m)
825 KASSERT(m->m_type != MT_FREE, ("freeing free mbuf"));
826 mbtypes[m->m_type]--;
827 if ((m->m_flags & M_PKTHDR) != 0)
828 m_tag_delete_chain(m, NULL);
829 if (m->m_flags & M_EXT) {
830 if (m->m_ext.ext_free != NULL) {
831 m->m_ext.ext_free(m->m_ext.ext_buf, m->m_ext.ext_size);
833 _m_mclfree(m->m_ext.ext_buf); /* inlined */
848 m_freem(struct mbuf *m)
853 * Try to keep a small pool of mbuf+cluster for quick use in
854 * device drivers. A good candidate is a M_PKTHDR buffer with
855 * only one cluster attached. Other mbufs, or those exceeding
856 * the pool size, are just m_free'd in the usual way.
857 * The following code makes sure that m_next, m_type,
858 * m_pkthdr.aux and m_ext.* are properly initialized.
859 * Other fields in the mbuf are initialized in m_getcl()
862 if (mcl_pool_now < mcl_pool_max && m && m->m_next == NULL &&
863 (m->m_flags & (M_PKTHDR|M_EXT)) == (M_PKTHDR|M_EXT) &&
864 m->m_type == MT_DATA && M_EXT_WRITABLE(m) ) {
865 m_tag_delete_chain(m, NULL);
866 m->m_nextpkt = mcl_pool;
877 * Mbuffer utility routines.
881 * Lesser-used path for M_PREPEND:
882 * allocate new mbuf to prepend to chain,
886 m_prepend(m, len, how)
892 MGET(mn, how, m->m_type);
893 if (mn == (struct mbuf *)NULL) {
895 return ((struct mbuf *)NULL);
897 if (m->m_flags & M_PKTHDR)
898 M_MOVE_PKTHDR(mn, m);
908 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
909 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
910 * The wait parameter is a choice of M_WAIT/M_DONTWAIT from caller.
911 * Note that the copy is read-only, because clusters are not copied,
912 * only their reference counts are incremented.
914 #define MCFail (mbstat.m_mcfail)
917 m_copym(m, off0, len, wait)
918 const struct mbuf *m;
922 struct mbuf *n, **np;
927 KASSERT(off >= 0, ("m_copym, negative off %d", off));
928 KASSERT(len >= 0, ("m_copym, negative len %d", len));
929 if (off == 0 && m->m_flags & M_PKTHDR)
932 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
942 KASSERT(len == M_COPYALL,
943 ("m_copym, length > size of mbuf chain"));
946 MGET(n, wait, m->m_type);
951 if (!m_dup_pkthdr(n, m, wait))
953 if (len == M_COPYALL)
954 n->m_pkthdr.len -= off0;
956 n->m_pkthdr.len = len;
959 n->m_len = min(len, m->m_len - off);
960 if (m->m_flags & M_EXT) {
961 n->m_data = m->m_data + off;
962 if (m->m_ext.ext_ref == NULL) {
964 &mclrefcnt[mtocl(m->m_ext.ext_buf)], 1);
968 (*m->m_ext.ext_ref)(m->m_ext.ext_buf,
975 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
977 if (len != M_COPYALL)
993 * Copy an entire packet, including header (which must be present).
994 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
995 * Note that the copy is read-only, because clusters are not copied,
996 * only their reference counts are incremented.
997 * Preserve alignment of the first mbuf so if the creator has left
998 * some room at the beginning (e.g. for inserting protocol headers)
999 * the copies also have the room available.
1002 m_copypacket(m, how)
1006 struct mbuf *top, *n, *o;
1008 MGET(n, how, m->m_type);
1013 if (!m_dup_pkthdr(n, m, how))
1015 n->m_len = m->m_len;
1016 if (m->m_flags & M_EXT) {
1017 n->m_data = m->m_data;
1018 if (m->m_ext.ext_ref == NULL)
1019 atomic_add_char(&mclrefcnt[mtocl(m->m_ext.ext_buf)], 1);
1023 (*m->m_ext.ext_ref)(m->m_ext.ext_buf,
1027 n->m_ext = m->m_ext;
1028 n->m_flags |= M_EXT;
1030 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
1031 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1036 MGET(o, how, m->m_type);
1043 n->m_len = m->m_len;
1044 if (m->m_flags & M_EXT) {
1045 n->m_data = m->m_data;
1046 if (m->m_ext.ext_ref == NULL) {
1048 &mclrefcnt[mtocl(m->m_ext.ext_buf)], 1);
1052 (*m->m_ext.ext_ref)(m->m_ext.ext_buf,
1056 n->m_ext = m->m_ext;
1057 n->m_flags |= M_EXT;
1059 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1072 * Copy data from an mbuf chain starting "off" bytes from the beginning,
1073 * continuing for "len" bytes, into the indicated buffer.
1076 m_copydata(m, off, len, cp)
1077 const struct mbuf *m;
1084 KASSERT(off >= 0, ("m_copydata, negative off %d", off));
1085 KASSERT(len >= 0, ("m_copydata, negative len %d", len));
1087 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
1094 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
1095 count = min(m->m_len - off, len);
1096 bcopy(mtod(m, caddr_t) + off, cp, count);
1105 * Copy a packet header mbuf chain into a completely new chain, including
1106 * copying any mbuf clusters. Use this instead of m_copypacket() when
1107 * you need a writable copy of an mbuf chain.
1114 struct mbuf **p, *top = NULL;
1115 int remain, moff, nsize;
1120 KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __FUNCTION__));
1122 /* While there's more data, get a new mbuf, tack it on, and fill it */
1123 remain = m->m_pkthdr.len;
1126 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */
1129 /* Get the next new mbuf */
1130 MGET(n, how, m->m_type);
1133 if (top == NULL) { /* first one, must be PKTHDR */
1134 if (!m_dup_pkthdr(n, m, how))
1137 } else /* not the first one */
1139 if (remain >= MINCLSIZE) {
1141 if ((n->m_flags & M_EXT) == 0) {
1149 /* Link it into the new chain */
1153 /* Copy data from original mbuf(s) into new mbuf */
1154 while (n->m_len < nsize && m != NULL) {
1155 int chunk = min(nsize - n->m_len, m->m_len - moff);
1157 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1161 if (moff == m->m_len) {
1167 /* Check correct total mbuf length */
1168 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
1169 ("%s: bogus m_pkthdr.len", __FUNCTION__));
1180 * Concatenate mbuf chain n to m.
1181 * Both chains must be of the same type (e.g. MT_DATA).
1182 * Any m_pkthdr is not updated.
1191 if (m->m_flags & M_EXT ||
1192 m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) {
1193 /* just join the two chains */
1197 /* splat the data from one into the other */
1198 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1200 m->m_len += n->m_len;
1214 if ((m = mp) == NULL)
1220 while (m != NULL && len > 0) {
1221 if (m->m_len <= len) {
1232 if (mp->m_flags & M_PKTHDR)
1233 m->m_pkthdr.len -= (req_len - len);
1236 * Trim from tail. Scan the mbuf chain,
1237 * calculating its length and finding the last mbuf.
1238 * If the adjustment only affects this mbuf, then just
1239 * adjust and return. Otherwise, rescan and truncate
1240 * after the remaining size.
1246 if (m->m_next == (struct mbuf *)0)
1250 if (m->m_len >= len) {
1252 if (mp->m_flags & M_PKTHDR)
1253 mp->m_pkthdr.len -= len;
1260 * Correct length for chain is "count".
1261 * Find the mbuf with last data, adjust its length,
1262 * and toss data from remaining mbufs on chain.
1265 if (m->m_flags & M_PKTHDR)
1266 m->m_pkthdr.len = count;
1267 for (; m; m = m->m_next) {
1268 if (m->m_len >= count) {
1275 (m = m->m_next) ->m_len = 0;
1280 * Rearange an mbuf chain so that len bytes are contiguous
1281 * and in the data area of an mbuf (so that mtod and dtom
1282 * will work for a structure of size len). Returns the resulting
1283 * mbuf chain on success, frees it and returns null on failure.
1284 * If there is room, it will add up to max_protohdr-len extra bytes to the
1285 * contiguous region in an attempt to avoid being called next time.
1287 #define MPFail (mbstat.m_mpfail)
1299 * If first mbuf has no cluster, and has room for len bytes
1300 * without shifting current data, pullup into it,
1301 * otherwise allocate a new mbuf to prepend to the chain.
1303 if ((n->m_flags & M_EXT) == 0 &&
1304 n->m_data + len < &n->m_dat[MLEN] && n->m_next) {
1305 if (n->m_len >= len)
1313 MGET(m, M_DONTWAIT, n->m_type);
1317 if (n->m_flags & M_PKTHDR)
1318 M_MOVE_PKTHDR(m, n);
1320 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
1322 count = min(min(max(len, max_protohdr), space), n->m_len);
1323 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1333 } while (len > 0 && n);
1347 * Partition an mbuf chain in two pieces, returning the tail --
1348 * all but the first len0 bytes. In case of failure, it returns NULL and
1349 * attempts to restore the chain to its original state.
1351 * Note that the resulting mbufs might be read-only, because the new
1352 * mbuf can end up sharing an mbuf cluster with the original mbuf if
1353 * the "breaking point" happens to lie within a cluster mbuf. Use the
1354 * M_WRITABLE() macro to check for this case.
1357 m_split(m0, len0, wait)
1362 unsigned len = len0, remain;
1364 for (m = m0; m && len > m->m_len; m = m->m_next)
1368 remain = m->m_len - len;
1369 if (m0->m_flags & M_PKTHDR) {
1370 MGETHDR(n, wait, m0->m_type);
1373 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1374 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1375 m0->m_pkthdr.len = len0;
1376 if (m->m_flags & M_EXT)
1378 if (remain > MHLEN) {
1379 /* m can't be the lead packet */
1381 n->m_next = m_split(m, len, wait);
1382 if (n->m_next == 0) {
1390 MH_ALIGN(n, remain);
1391 } else if (remain == 0) {
1396 MGET(n, wait, m->m_type);
1402 if (m->m_flags & M_EXT) {
1403 n->m_flags |= M_EXT;
1404 n->m_ext = m->m_ext;
1405 if (m->m_ext.ext_ref == NULL)
1406 atomic_add_char(&mclrefcnt[mtocl(m->m_ext.ext_buf)], 1);
1410 (*m->m_ext.ext_ref)(m->m_ext.ext_buf,
1414 n->m_data = m->m_data + len;
1416 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1420 n->m_next = m->m_next;
1425 * Routine to copy from device local memory into mbufs.
1428 m_devget(buf, totlen, off0, ifp, copy)
1432 void (*copy) (char *from, caddr_t to, u_int len);
1435 struct mbuf *top = 0, **mp = ⊤
1436 int off = off0, len;
1443 cp += off + 2 * sizeof(u_short);
1444 totlen -= 2 * sizeof(u_short);
1446 MGETHDR(m, M_DONTWAIT, MT_DATA);
1449 m->m_pkthdr.rcvif = ifp;
1450 m->m_pkthdr.len = totlen;
1453 while (totlen > 0) {
1455 MGET(m, M_DONTWAIT, MT_DATA);
1462 len = min(totlen, epkt - cp);
1463 if (len >= MINCLSIZE) {
1464 MCLGET(m, M_DONTWAIT);
1465 if (m->m_flags & M_EXT)
1466 m->m_len = len = min(len, MCLBYTES);
1471 * Place initial small packet/header at end of mbuf.
1473 if (len < m->m_len) {
1474 if (top == 0 && len + max_linkhdr <= m->m_len)
1475 m->m_data += max_linkhdr;
1481 copy(cp, mtod(m, caddr_t), (unsigned)len);
1483 bcopy(cp, mtod(m, caddr_t), (unsigned)len);
1495 * Copy data from a buffer back into the indicated mbuf chain,
1496 * starting "off" bytes from the beginning, extending the mbuf
1497 * chain if necessary.
1500 m_copyback(m0, off, len, cp)
1507 struct mbuf *m = m0, *n;
1512 while (off > (mlen = m->m_len)) {
1515 if (m->m_next == 0) {
1516 n = m_getclr(M_DONTWAIT, m->m_type);
1519 n->m_len = min(MLEN, len + off);
1525 mlen = min (m->m_len - off, len);
1526 bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen);
1534 if (m->m_next == 0) {
1535 n = m_get(M_DONTWAIT, m->m_type);
1538 n->m_len = min(MLEN, len);
1543 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1544 m->m_pkthdr.len = totlen;
1548 m_print(const struct mbuf *m)
1551 const struct mbuf *m2;
1553 len = m->m_pkthdr.len;
1556 printf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-");
1564 * "Move" mbuf pkthdr from "from" to "to".
1565 * "from" must have M_PKTHDR set, and "to" must be empty.
1568 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
1570 KASSERT((to->m_flags & M_EXT) == 0, ("m_move_pkthdr: to has cluster"));
1572 to->m_flags = from->m_flags & M_COPYFLAGS;
1573 to->m_data = to->m_pktdat;
1574 to->m_pkthdr = from->m_pkthdr; /* especially tags */
1575 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */
1576 from->m_flags &= ~M_PKTHDR;
1580 * Duplicate "from"'s mbuf pkthdr in "to".
1581 * "from" must have M_PKTHDR set, and "to" must be empty.
1582 * In particular, this does a deep copy of the packet tags.
1585 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
1587 to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT);
1588 if ((to->m_flags & M_EXT) == 0)
1589 to->m_data = to->m_pktdat;
1590 to->m_pkthdr = from->m_pkthdr;
1591 SLIST_INIT(&to->m_pkthdr.tags);
1592 return (m_tag_copy_chain(to, from, how));
1596 * Defragment a mbuf chain, returning the shortest possible
1597 * chain of mbufs and clusters. If allocation fails and
1598 * this cannot be completed, NULL will be returned, but
1599 * the passed in chain will be unchanged. Upon success,
1600 * the original chain will be freed, and the new chain
1603 * If a non-packet header is passed in, the original
1604 * mbuf (chain?) will be returned unharmed.
1607 m_defrag(struct mbuf *m0, int how)
1609 struct mbuf *m_new = NULL, *m_final = NULL;
1610 int progress = 0, length;
1612 if (!(m0->m_flags & M_PKTHDR))
1615 #ifdef MBUF_STRESS_TEST
1616 if (m_defragrandomfailures) {
1617 int temp = arc4random() & 0xff;
1623 if (m0->m_pkthdr.len > MHLEN)
1624 m_final = m_getcl(how, MT_DATA, M_PKTHDR);
1626 m_final = m_gethdr(how, MT_DATA);
1628 if (m_final == NULL)
1631 if (m_dup_pkthdr(m_final, m0, how) == NULL)
1636 while (progress < m0->m_pkthdr.len) {
1637 length = m0->m_pkthdr.len - progress;
1638 if (length > MCLBYTES)
1641 if (m_new == NULL) {
1643 m_new = m_getcl(how, MT_DATA, 0);
1645 m_new = m_get(how, MT_DATA);
1650 m_copydata(m0, progress, length, mtod(m_new, caddr_t));
1652 m_new->m_len = length;
1653 if (m_new != m_final)
1654 m_cat(m_final, m_new);
1657 if (m0->m_next == NULL)
1662 m_defragbytes += m0->m_pkthdr.len;