2 * Copyright (c) 2004 Jeffrey M. Hsu. All rights reserved.
3 * Copyright (c) 1982, 1986, 1988, 1991, 1993
4 * The Regents of the University of California. All rights reserved.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 * 3. All advertising materials mentioning features or use of this software
15 * must display the following acknowledgement:
16 * This product includes software developed by the University of
17 * California, Berkeley and its contributors.
18 * 4. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * @(#)uipc_mbuf.c 8.2 (Berkeley) 1/4/94
35 * $FreeBSD: src/sys/kern/uipc_mbuf.c,v 1.51.2.24 2003/04/15 06:59:29 silby Exp $
36 * $DragonFly: src/sys/kern/uipc_mbuf.c,v 1.20 2004/06/07 02:18:13 dillon Exp $
39 #include "opt_param.h"
40 #include "opt_mbuf_stress_test.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/malloc.h>
45 #include <sys/kernel.h>
46 #include <sys/sysctl.h>
47 #include <sys/domain.h>
48 #include <sys/protosw.h>
50 #include <sys/thread.h>
51 #include <sys/globaldata.h>
54 #include <vm/vm_kern.h>
55 #include <vm/vm_extern.h>
58 #include <machine/cpu.h>
61 static void mbinit (void *);
62 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbinit, NULL)
68 u_long mbtypes[MT_NTYPES];
70 union mcluster *mclfree;
79 #ifdef MBUF_STRESS_TEST
80 int m_defragrandomfailures;
85 u_int m_mballoc_wid = 0;
86 u_int m_clalloc_wid = 0;
88 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW,
90 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW,
91 &max_protohdr, 0, "");
92 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0, "");
93 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW,
95 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW,
97 SYSCTL_STRUCT(_kern_ipc, KIPC_MBSTAT, mbstat, CTLFLAG_RW, &mbstat, mbstat, "");
98 SYSCTL_OPAQUE(_kern_ipc, OID_AUTO, mbtypes, CTLFLAG_RD, mbtypes,
99 sizeof(mbtypes), "LU", "");
100 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RD,
101 &nmbclusters, 0, "Maximum number of mbuf clusters available");
102 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RD, &nmbufs, 0,
103 "Maximum number of mbufs available");
104 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
105 &m_defragpackets, 0, "");
106 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
107 &m_defragbytes, 0, "");
108 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
109 &m_defraguseless, 0, "");
110 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
111 &m_defragfailure, 0, "");
112 #ifdef MBUF_STRESS_TEST
113 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
114 &m_defragrandomfailures, 0, "");
117 static void m_reclaim (void);
120 #define NMBCLUSTERS (512 + maxusers * 16)
123 #define NMBUFS (nmbclusters * 4)
127 * Perform sanity checks of tunables declared above.
130 tunable_mbinit(void *dummy)
134 * This has to be done before VM init.
136 nmbclusters = NMBCLUSTERS;
137 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
139 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
141 if (nmbufs < nmbclusters * 2)
142 nmbufs = nmbclusters * 2;
146 SYSINIT(tunable_mbinit, SI_SUB_TUNABLES, SI_ORDER_ANY, tunable_mbinit, NULL);
148 /* "number of clusters of pages" */
159 mmbfree = NULL; mclfree = NULL;
160 mbstat.m_msize = MSIZE;
161 mbstat.m_mclbytes = MCLBYTES;
162 mbstat.m_minclsize = MINCLSIZE;
163 mbstat.m_mlen = MLEN;
164 mbstat.m_mhlen = MHLEN;
167 if (m_mballoc(NMB_INIT, MB_DONTWAIT) == 0)
169 #if MCLBYTES <= PAGE_SIZE
170 if (m_clalloc(NCL_INIT, MB_DONTWAIT) == 0)
173 /* It's OK to call contigmalloc in this context. */
174 if (m_clalloc(16, MB_WAIT) == 0)
184 * Allocate at least nmb mbufs and place on mbuf free list.
185 * Must be called at splimp.
189 m_mballoc(int nmb, int how)
196 * If we've hit the mbuf limit, stop allocating from mb_map,
197 * (or trying to) in order to avoid dipping into the section of
198 * mb_map which we've "reserved" for clusters.
200 if ((nmb + mbstat.m_mbufs) > nmbufs)
204 * Once we run out of map space, it will be impossible to get
205 * any more (nothing is ever freed back to the map)
206 * -- however you are not dead as m_reclaim might
207 * still be able to free a substantial amount of space.
209 * XXX Furthermore, we can also work with "recycled" mbufs (when
210 * we're calling with MB_WAIT the sleep procedure will be woken
211 * up when an mbuf is freed. See m_mballoc_wait()).
216 nbytes = round_page(nmb * MSIZE);
217 p = (caddr_t)kmem_malloc(mb_map, nbytes, M_NOWAIT);
218 if (p == 0 && how == MB_WAIT) {
220 p = (caddr_t)kmem_malloc(mb_map, nbytes, M_WAITOK);
224 * Either the map is now full, or `how' is M_NOWAIT and there
230 nmb = nbytes / MSIZE;
231 for (i = 0; i < nmb; i++) {
232 ((struct mbuf *)p)->m_next = mmbfree;
233 mmbfree = (struct mbuf *)p;
236 mbstat.m_mbufs += nmb;
237 mbtypes[MT_FREE] += nmb;
242 * Once the mb_map has been exhausted and if the call to the allocation macros
243 * (or, in some cases, functions) is with MB_WAIT, then it is necessary to rely
244 * solely on reclaimed mbufs. Here we wait for an mbuf to be freed for a
245 * designated (mbuf_wait) time.
248 m_mballoc_wait(int caller, int type)
255 if ((tsleep(&m_mballoc_wid, 0, "mballc", mbuf_wait)) == EWOULDBLOCK)
260 * Now that we (think) that we've got something, we will redo an
261 * MGET, but avoid getting into another instance of m_mballoc_wait()
262 * XXX: We retry to fetch _even_ if the sleep timed out. This is left
263 * this way, purposely, in the [unlikely] case that an mbuf was
264 * freed but the sleep was not awakened in time.
269 MGET(p, MB_DONTWAIT, type);
272 MGETHDR(p, MB_DONTWAIT, type);
275 panic("m_mballoc_wait: invalid caller (%d)", caller);
279 if (p != NULL) { /* We waited and got something... */
281 /* Wake up another if we have more free. */
289 #if MCLBYTES > PAGE_SIZE
290 static int i_want_my_mcl;
298 tsleep(&i_want_my_mcl, 0, "mclalloc", 0);
300 for (; i_want_my_mcl; i_want_my_mcl--) {
301 if (m_clalloc(1, MB_WAIT) == 0)
302 printf("m_clalloc failed even in process context!\n");
307 static struct thread *mclallocthread;
308 static struct kproc_desc mclalloc_kp = {
313 SYSINIT(mclallocthread, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start,
318 * Allocate some number of mbuf clusters
319 * and place on cluster free list.
320 * Must be called at splimp.
324 m_clalloc(int ncl, int how)
331 * If we've hit the mcluster number limit, stop allocating from
332 * mb_map, (or trying to) in order to avoid dipping into the section
333 * of mb_map which we've "reserved" for mbufs.
335 if ((ncl + mbstat.m_clusters) > nmbclusters)
339 * Once we run out of map space, it will be impossible
340 * to get any more (nothing is ever freed back to the
341 * map). From this point on, we solely rely on freed
347 #if MCLBYTES > PAGE_SIZE
348 if (how != MB_WAIT) {
349 i_want_my_mcl += ncl;
350 wakeup(&i_want_my_mcl);
354 p = contigmalloc_map(MCLBYTES * ncl, M_DEVBUF, M_WAITOK, 0ul,
355 ~0ul, PAGE_SIZE, 0, mb_map);
359 p = (caddr_t)kmem_malloc(mb_map, ctob(npg),
360 how != MB_WAIT ? M_NOWAIT : M_WAITOK);
361 ncl = ncl * PAGE_SIZE / MCLBYTES;
364 * Either the map is now full, or `how' is M_NOWAIT and there
368 static int last_report ; /* when we did that (in ticks) */
371 if (ticks < last_report || (ticks - last_report) >= hz) {
373 printf("All mbuf clusters exhausted, please see tuning(7).\n");
378 for (i = 0; i < ncl; i++) {
379 ((union mcluster *)p)->mcl_next = mclfree;
380 mclfree = (union mcluster *)p;
384 mbstat.m_clusters += ncl;
389 * Once the mb_map submap has been exhausted and the allocation is called with
390 * MB_WAIT, we rely on the mclfree union pointers. If nothing is free, we will
391 * sleep for a designated amount of time (mbuf_wait) or until we're woken up
392 * due to sudden mcluster availability.
400 /* If in interrupt context, and INVARIANTS, maintain sanity and die. */
401 KASSERT(mycpu->gd_intr_nesting_level == 0, ("CLALLOC: CANNOT WAIT IN INTERRUPT"));
403 /* Sleep until something's available or until we expire. */
405 if ((tsleep(&m_clalloc_wid, 0, "mclalc", mbuf_wait)) == EWOULDBLOCK)
409 * Now that we (think) that we've got something, we will redo and
410 * MGET, but avoid getting into another instance of m_clalloc_wait()
412 p = m_mclalloc(MB_DONTWAIT);
415 if (p != NULL) { /* We waited and got something... */
417 /* Wake up another if we have more free. */
427 * When MGET fails, ask protocols to free space when short of memory,
428 * then re-attempt to allocate an mbuf.
431 m_retry(int i, int t)
437 * Must only do the reclaim if not in an interrupt context.
440 KASSERT(mycpu->gd_intr_nesting_level == 0,
441 ("MBALLOC: CANNOT WAIT IN INTERRUPT"));
447 (void)m_mballoc(1, i);
456 m->m_data = m->m_dat;
461 static int last_report ; /* when we did that (in ticks) */
465 if (ticks < last_report || (ticks - last_report) >= hz) {
467 printf("All mbufs exhausted, please see tuning(7).\n");
475 * As above; retry an MGETHDR.
478 m_retryhdr(int i, int t)
484 * Must only do the reclaim if not in an interrupt context.
487 KASSERT(mycpu->gd_intr_nesting_level == 0,
488 ("MBALLOC: CANNOT WAIT IN INTERRUPT"));
494 (void)m_mballoc(1, i);
503 m->m_data = m->m_pktdat;
504 m->m_flags = M_PKTHDR;
505 m->m_pkthdr.rcvif = NULL;
506 SLIST_INIT(&m->m_pkthdr.tags);
507 m->m_pkthdr.csum_flags = 0;
511 static int last_report ; /* when we did that (in ticks) */
515 if (ticks < last_report || (ticks - last_report) >= hz) {
517 printf("All mbufs exhausted, please see tuning(7).\n");
531 for (dp = domains; dp; dp = dp->dom_next) {
532 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) {
542 * Space allocation routines.
543 * These are also available as macros
544 * for critical paths.
547 m_get(int how, int type)
554 (void)m_mballoc(1, how);
563 m->m_data = m->m_dat;
568 m = m_retry(how, type);
569 if (m == NULL && how == MB_WAIT)
570 m = m_mballoc_wait(MGET_C, type);
576 m_gethdr(int how, int type)
583 (void)m_mballoc(1, how);
592 m->m_data = m->m_pktdat;
593 m->m_flags = M_PKTHDR;
594 m->m_pkthdr.rcvif = NULL;
595 SLIST_INIT(&m->m_pkthdr.tags);
596 m->m_pkthdr.csum_flags = 0;
600 m = m_retryhdr(how, type);
601 if (m == NULL && how == MB_WAIT)
602 m = m_mballoc_wait(MGETHDR_C, type);
608 m_getclr(int how, int type)
615 bzero(mtod(m, caddr_t), MLEN);
620 * m_getcl() returns an mbuf with an attached cluster.
621 * Because many network drivers use this kind of buffers a lot, it is
622 * convenient to keep a small pool of free buffers of this kind.
623 * Even a small size such as 10 gives about 10% improvement in the
624 * forwarding rate in a bridge or router.
625 * The size of this free list is controlled by the sysctl variable
626 * mcl_pool_max. The list is populated on m_freem(), and used in
627 * m_getcl() if elements are available.
629 static struct mbuf *mcl_pool;
630 static int mcl_pool_now;
631 static int mcl_pool_max = 10;
633 SYSCTL_INT(_kern_ipc, OID_AUTO, mcl_pool_max, CTLFLAG_RW, &mcl_pool_max, 0,
634 "Maximum number of mbufs+cluster in free list");
635 SYSCTL_INT(_kern_ipc, OID_AUTO, mcl_pool_now, CTLFLAG_RD, &mcl_pool_now, 0,
636 "Current number of mbufs+cluster in free list");
639 m_getcl(int how, short type, int flags)
644 if (flags & M_PKTHDR) {
645 if (type == MT_DATA && mcl_pool) {
647 mcl_pool = mp->m_nextpkt;
650 mp->m_nextpkt = NULL;
651 mp->m_data = mp->m_ext.ext_buf;
652 mp->m_flags = M_PKTHDR|M_EXT;
653 mp->m_pkthdr.rcvif = NULL;
654 mp->m_pkthdr.csum_flags = 0;
657 MGETHDR(mp, how, type);
662 if ( (mp->m_flags & M_EXT) == 0) {
673 * m_getm(m, len, how, type)
675 * This will allocate len-worth of mbufs and/or mbuf clusters (whatever fits
676 * best) and return a pointer to the top of the allocated chain. If m is
677 * non-null, then we assume that it is a single mbuf or an mbuf chain to
678 * which we want len bytes worth of mbufs and/or clusters attached, and so
679 * if we succeed in allocating it, we will just return a pointer to m.
681 * If we happen to fail at any point during the allocation, we will free
682 * up everything we have already allocated and return NULL.
686 m_getm(struct mbuf *m, int len, int how, int type)
688 struct mbuf *top, *tail, *mp, *mtail = NULL;
690 KASSERT(len >= 0, ("len is < 0 in m_getm"));
695 else if (len > MINCLSIZE) {
697 if ((mp->m_flags & M_EXT) == 0) {
703 len -= M_TRAILINGSPACE(mp);
706 for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next);
718 if (len > MINCLSIZE) {
720 if ((mp->m_flags & M_EXT) == 0)
725 len -= M_TRAILINGSPACE(mp);
738 * m_mclalloc() - Allocates an mbuf cluster.
750 mp = (caddr_t)mclfree;
752 KKASSERT((struct mbuf *)mp >= mbutl &&
753 (struct mbuf *)mp < mbute);
754 mclrefcnt[mtocl(mp)]++;
756 mclfree = ((union mcluster *)mp)->mcl_next;
762 return(m_clalloc_wait());
767 * m_mclget() - Adds a cluster to a normal mbuf, M_EXT is set on success.
770 m_mclget(struct mbuf *m, int how)
772 m->m_ext.ext_buf = m_mclalloc(how);
773 if (m->m_ext.ext_buf != NULL) {
774 m->m_data = m->m_ext.ext_buf;
776 m->m_ext.ext_free = NULL;
777 m->m_ext.ext_ref = NULL;
778 m->m_ext.ext_size = MCLBYTES;
783 _m_mclfree(caddr_t data)
785 union mcluster *mp = (union mcluster *)data;
787 KASSERT(mclrefcnt[mtocl(mp)] > 0, ("freeing free cluster"));
788 KKASSERT((struct mbuf *)mp >= mbutl &&
789 (struct mbuf *)mp < mbute);
790 if (--mclrefcnt[mtocl(mp)] == 0) {
791 mp->mcl_next = mclfree;
799 m_mclfree(caddr_t mp)
809 * Free a single mbuf and any associated external storage. The successor,
810 * if any, is returned.
812 * We do need to check non-first mbuf for m_aux, since some of existing
813 * code does not call M_PREPEND properly.
814 * (example: call to bpf_mtap from drivers)
817 m_free(struct mbuf *m)
823 KASSERT(m->m_type != MT_FREE, ("freeing free mbuf"));
824 mbtypes[m->m_type]--;
825 if ((m->m_flags & M_PKTHDR) != 0)
826 m_tag_delete_chain(m, NULL);
827 if (m->m_flags & M_EXT) {
828 if (m->m_ext.ext_free != NULL) {
829 m->m_ext.ext_free(m->m_ext.ext_buf, m->m_ext.ext_size);
831 _m_mclfree(m->m_ext.ext_buf); /* inlined */
846 m_freem(struct mbuf *m)
851 * Try to keep a small pool of mbuf+cluster for quick use in
852 * device drivers. A good candidate is a M_PKTHDR buffer with
853 * only one cluster attached. Other mbufs, or those exceeding
854 * the pool size, are just m_free'd in the usual way.
855 * The following code makes sure that m_next, m_type,
856 * m_pkthdr.aux and m_ext.* are properly initialized.
857 * Other fields in the mbuf are initialized in m_getcl()
860 if (mcl_pool_now < mcl_pool_max && m && m->m_next == NULL &&
861 (m->m_flags & (M_PKTHDR|M_EXT)) == (M_PKTHDR|M_EXT) &&
862 m->m_type == MT_DATA && M_EXT_WRITABLE(m) ) {
863 m_tag_delete_chain(m, NULL);
864 m->m_nextpkt = mcl_pool;
875 * Mbuffer utility routines.
879 * Lesser-used path for M_PREPEND:
880 * allocate new mbuf to prepend to chain,
884 m_prepend(struct mbuf *m, int len, int how)
888 MGET(mn, how, m->m_type);
889 if (mn == (struct mbuf *)NULL) {
891 return ((struct mbuf *)NULL);
893 if (m->m_flags & M_PKTHDR)
894 M_MOVE_PKTHDR(mn, m);
904 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
905 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
906 * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller.
907 * Note that the copy is read-only, because clusters are not copied,
908 * only their reference counts are incremented.
910 #define MCFail (mbstat.m_mcfail)
913 m_copym(const struct mbuf *m, int off0, int len, int wait)
915 struct mbuf *n, **np;
920 KASSERT(off >= 0, ("m_copym, negative off %d", off));
921 KASSERT(len >= 0, ("m_copym, negative len %d", len));
922 if (off == 0 && m->m_flags & M_PKTHDR)
925 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
935 KASSERT(len == M_COPYALL,
936 ("m_copym, length > size of mbuf chain"));
939 MGET(n, wait, m->m_type);
944 if (!m_dup_pkthdr(n, m, wait))
946 if (len == M_COPYALL)
947 n->m_pkthdr.len -= off0;
949 n->m_pkthdr.len = len;
952 n->m_len = min(len, m->m_len - off);
953 if (m->m_flags & M_EXT) {
954 n->m_data = m->m_data + off;
955 if (m->m_ext.ext_ref == NULL) {
957 &mclrefcnt[mtocl(m->m_ext.ext_buf)], 1);
961 (*m->m_ext.ext_ref)(m->m_ext.ext_buf,
968 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
970 if (len != M_COPYALL)
986 * Copy an entire packet, including header (which must be present).
987 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
988 * Note that the copy is read-only, because clusters are not copied,
989 * only their reference counts are incremented.
990 * Preserve alignment of the first mbuf so if the creator has left
991 * some room at the beginning (e.g. for inserting protocol headers)
992 * the copies also have the room available.
995 m_copypacket(struct mbuf *m, int how)
997 struct mbuf *top, *n, *o;
999 MGET(n, how, m->m_type);
1004 if (!m_dup_pkthdr(n, m, how))
1006 n->m_len = m->m_len;
1007 if (m->m_flags & M_EXT) {
1008 n->m_data = m->m_data;
1009 if (m->m_ext.ext_ref == NULL)
1010 atomic_add_char(&mclrefcnt[mtocl(m->m_ext.ext_buf)], 1);
1014 (*m->m_ext.ext_ref)(m->m_ext.ext_buf,
1018 n->m_ext = m->m_ext;
1019 n->m_flags |= M_EXT;
1021 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
1022 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1027 MGET(o, how, m->m_type);
1034 n->m_len = m->m_len;
1035 if (m->m_flags & M_EXT) {
1036 n->m_data = m->m_data;
1037 if (m->m_ext.ext_ref == NULL) {
1039 &mclrefcnt[mtocl(m->m_ext.ext_buf)], 1);
1043 (*m->m_ext.ext_ref)(m->m_ext.ext_buf,
1047 n->m_ext = m->m_ext;
1048 n->m_flags |= M_EXT;
1050 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1063 * Copy data from an mbuf chain starting "off" bytes from the beginning,
1064 * continuing for "len" bytes, into the indicated buffer.
1067 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
1071 KASSERT(off >= 0, ("m_copydata, negative off %d", off));
1072 KASSERT(len >= 0, ("m_copydata, negative len %d", len));
1074 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
1081 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
1082 count = min(m->m_len - off, len);
1083 bcopy(mtod(m, caddr_t) + off, cp, count);
1092 * Copy a packet header mbuf chain into a completely new chain, including
1093 * copying any mbuf clusters. Use this instead of m_copypacket() when
1094 * you need a writable copy of an mbuf chain.
1097 m_dup(struct mbuf *m, int how)
1099 struct mbuf **p, *top = NULL;
1100 int remain, moff, nsize;
1105 KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __FUNCTION__));
1107 /* While there's more data, get a new mbuf, tack it on, and fill it */
1108 remain = m->m_pkthdr.len;
1111 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */
1114 /* Get the next new mbuf */
1115 MGET(n, how, m->m_type);
1118 if (top == NULL) { /* first one, must be PKTHDR */
1119 if (!m_dup_pkthdr(n, m, how))
1122 } else /* not the first one */
1124 if (remain >= MINCLSIZE) {
1126 if ((n->m_flags & M_EXT) == 0) {
1134 /* Link it into the new chain */
1138 /* Copy data from original mbuf(s) into new mbuf */
1139 while (n->m_len < nsize && m != NULL) {
1140 int chunk = min(nsize - n->m_len, m->m_len - moff);
1142 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1146 if (moff == m->m_len) {
1152 /* Check correct total mbuf length */
1153 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
1154 ("%s: bogus m_pkthdr.len", __FUNCTION__));
1165 * Concatenate mbuf chain n to m.
1166 * Both chains must be of the same type (e.g. MT_DATA).
1167 * Any m_pkthdr is not updated.
1170 m_cat(struct mbuf *m, struct mbuf *n)
1175 if (m->m_flags & M_EXT ||
1176 m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) {
1177 /* just join the two chains */
1181 /* splat the data from one into the other */
1182 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1184 m->m_len += n->m_len;
1190 m_adj(struct mbuf *mp, int req_len)
1196 if ((m = mp) == NULL)
1202 while (m != NULL && len > 0) {
1203 if (m->m_len <= len) {
1214 if (mp->m_flags & M_PKTHDR)
1215 m->m_pkthdr.len -= (req_len - len);
1218 * Trim from tail. Scan the mbuf chain,
1219 * calculating its length and finding the last mbuf.
1220 * If the adjustment only affects this mbuf, then just
1221 * adjust and return. Otherwise, rescan and truncate
1222 * after the remaining size.
1228 if (m->m_next == (struct mbuf *)0)
1232 if (m->m_len >= len) {
1234 if (mp->m_flags & M_PKTHDR)
1235 mp->m_pkthdr.len -= len;
1242 * Correct length for chain is "count".
1243 * Find the mbuf with last data, adjust its length,
1244 * and toss data from remaining mbufs on chain.
1247 if (m->m_flags & M_PKTHDR)
1248 m->m_pkthdr.len = count;
1249 for (; m; m = m->m_next) {
1250 if (m->m_len >= count) {
1257 (m = m->m_next) ->m_len = 0;
1262 * Rearange an mbuf chain so that len bytes are contiguous
1263 * and in the data area of an mbuf (so that mtod will work for a structure
1264 * of size len). Returns the resulting mbuf chain on success, frees it and
1265 * returns null on failure. If there is room, it will add up to
1266 * max_protohdr-len extra bytes to the contiguous region in an attempt to
1267 * avoid being called next time.
1269 #define MPFail (mbstat.m_mpfail)
1272 m_pullup(struct mbuf *n, int len)
1279 * If first mbuf has no cluster, and has room for len bytes
1280 * without shifting current data, pullup into it,
1281 * otherwise allocate a new mbuf to prepend to the chain.
1283 if ((n->m_flags & M_EXT) == 0 &&
1284 n->m_data + len < &n->m_dat[MLEN] && n->m_next) {
1285 if (n->m_len >= len)
1293 MGET(m, MB_DONTWAIT, n->m_type);
1297 if (n->m_flags & M_PKTHDR)
1298 M_MOVE_PKTHDR(m, n);
1300 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
1302 count = min(min(max(len, max_protohdr), space), n->m_len);
1303 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1313 } while (len > 0 && n);
1327 * Partition an mbuf chain in two pieces, returning the tail --
1328 * all but the first len0 bytes. In case of failure, it returns NULL and
1329 * attempts to restore the chain to its original state.
1331 * Note that the resulting mbufs might be read-only, because the new
1332 * mbuf can end up sharing an mbuf cluster with the original mbuf if
1333 * the "breaking point" happens to lie within a cluster mbuf. Use the
1334 * M_WRITABLE() macro to check for this case.
1337 m_split(struct mbuf *m0, int len0, int wait)
1340 unsigned len = len0, remain;
1342 for (m = m0; m && len > m->m_len; m = m->m_next)
1346 remain = m->m_len - len;
1347 if (m0->m_flags & M_PKTHDR) {
1348 MGETHDR(n, wait, m0->m_type);
1351 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1352 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1353 m0->m_pkthdr.len = len0;
1354 if (m->m_flags & M_EXT)
1356 if (remain > MHLEN) {
1357 /* m can't be the lead packet */
1359 n->m_next = m_split(m, len, wait);
1360 if (n->m_next == 0) {
1368 MH_ALIGN(n, remain);
1369 } else if (remain == 0) {
1374 MGET(n, wait, m->m_type);
1380 if (m->m_flags & M_EXT) {
1381 n->m_flags |= M_EXT;
1382 n->m_ext = m->m_ext;
1383 if (m->m_ext.ext_ref == NULL)
1384 atomic_add_char(&mclrefcnt[mtocl(m->m_ext.ext_buf)], 1);
1388 (*m->m_ext.ext_ref)(m->m_ext.ext_buf,
1392 n->m_data = m->m_data + len;
1394 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1398 n->m_next = m->m_next;
1403 * Routine to copy from device local memory into mbufs.
1406 m_devget(char *buf, int totlen, int off0, struct ifnet *ifp,
1407 void (*copy) (char *from, caddr_t to, u_int len))
1410 struct mbuf *top = 0, **mp = ⊤
1411 int off = off0, len;
1418 cp += off + 2 * sizeof(u_short);
1419 totlen -= 2 * sizeof(u_short);
1421 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1424 m->m_pkthdr.rcvif = ifp;
1425 m->m_pkthdr.len = totlen;
1428 while (totlen > 0) {
1430 MGET(m, MB_DONTWAIT, MT_DATA);
1437 len = min(totlen, epkt - cp);
1438 if (len >= MINCLSIZE) {
1439 MCLGET(m, MB_DONTWAIT);
1440 if (m->m_flags & M_EXT)
1441 m->m_len = len = min(len, MCLBYTES);
1446 * Place initial small packet/header at end of mbuf.
1448 if (len < m->m_len) {
1449 if (top == 0 && len + max_linkhdr <= m->m_len)
1450 m->m_data += max_linkhdr;
1456 copy(cp, mtod(m, caddr_t), (unsigned)len);
1458 bcopy(cp, mtod(m, caddr_t), (unsigned)len);
1470 * Copy data from a buffer back into the indicated mbuf chain,
1471 * starting "off" bytes from the beginning, extending the mbuf
1472 * chain if necessary.
1475 m_copyback(struct mbuf *m0, int off, int len, caddr_t cp)
1478 struct mbuf *m = m0, *n;
1483 while (off > (mlen = m->m_len)) {
1486 if (m->m_next == 0) {
1487 n = m_getclr(MB_DONTWAIT, m->m_type);
1490 n->m_len = min(MLEN, len + off);
1496 mlen = min (m->m_len - off, len);
1497 bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen);
1505 if (m->m_next == 0) {
1506 n = m_get(MB_DONTWAIT, m->m_type);
1509 n->m_len = min(MLEN, len);
1514 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1515 m->m_pkthdr.len = totlen;
1519 m_print(const struct mbuf *m)
1522 const struct mbuf *m2;
1524 len = m->m_pkthdr.len;
1527 printf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-");
1535 * "Move" mbuf pkthdr from "from" to "to".
1536 * "from" must have M_PKTHDR set, and "to" must be empty.
1539 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
1541 KASSERT((to->m_flags & M_EXT) == 0, ("m_move_pkthdr: to has cluster"));
1543 to->m_flags = from->m_flags & M_COPYFLAGS;
1544 to->m_data = to->m_pktdat;
1545 to->m_pkthdr = from->m_pkthdr; /* especially tags */
1546 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */
1547 from->m_flags &= ~M_PKTHDR;
1551 * Duplicate "from"'s mbuf pkthdr in "to".
1552 * "from" must have M_PKTHDR set, and "to" must be empty.
1553 * In particular, this does a deep copy of the packet tags.
1556 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
1558 to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT);
1559 if ((to->m_flags & M_EXT) == 0)
1560 to->m_data = to->m_pktdat;
1561 to->m_pkthdr = from->m_pkthdr;
1562 SLIST_INIT(&to->m_pkthdr.tags);
1563 return (m_tag_copy_chain(to, from, how));
1567 * Defragment a mbuf chain, returning the shortest possible
1568 * chain of mbufs and clusters. If allocation fails and
1569 * this cannot be completed, NULL will be returned, but
1570 * the passed in chain will be unchanged. Upon success,
1571 * the original chain will be freed, and the new chain
1574 * If a non-packet header is passed in, the original
1575 * mbuf (chain?) will be returned unharmed.
1578 m_defrag(struct mbuf *m0, int how)
1580 struct mbuf *m_new = NULL, *m_final = NULL;
1581 int progress = 0, length;
1583 if (!(m0->m_flags & M_PKTHDR))
1586 #ifdef MBUF_STRESS_TEST
1587 if (m_defragrandomfailures) {
1588 int temp = arc4random() & 0xff;
1594 if (m0->m_pkthdr.len > MHLEN)
1595 m_final = m_getcl(how, MT_DATA, M_PKTHDR);
1597 m_final = m_gethdr(how, MT_DATA);
1599 if (m_final == NULL)
1602 if (m_dup_pkthdr(m_final, m0, how) == NULL)
1607 while (progress < m0->m_pkthdr.len) {
1608 length = m0->m_pkthdr.len - progress;
1609 if (length > MCLBYTES)
1612 if (m_new == NULL) {
1614 m_new = m_getcl(how, MT_DATA, 0);
1616 m_new = m_get(how, MT_DATA);
1621 m_copydata(m0, progress, length, mtod(m_new, caddr_t));
1623 m_new->m_len = length;
1624 if (m_new != m_final)
1625 m_cat(m_final, m_new);
1628 if (m0->m_next == NULL)
1633 m_defragbytes += m0->m_pkthdr.len;
1645 * Move data from uio into mbufs.
1646 * A length of zero means copy the whole uio.
1649 m_uiomove(struct uio *uio, int wait, int len0)
1651 struct mbuf *head; /* result mbuf chain */
1652 struct mbuf *m; /* current working mbuf */
1654 int resid, datalen, error;
1656 resid = (len0 == 0) ? uio->uio_resid : min(len0, uio->uio_resid);
1661 if (resid > MHLEN) {
1662 m = m_getcl(wait, MT_DATA, head == NULL ? M_PKTHDR : 0);
1665 if (m->m_flags & M_PKTHDR)
1666 m->m_pkthdr.len = 0;
1669 MGETHDR(m, wait, MT_DATA);
1672 m->m_pkthdr.len = 0;
1673 /* Leave room for protocol headers. */
1677 MGET(m, wait, MT_DATA);
1682 datalen = min(MCLBYTES, resid);
1683 error = uiomove(mtod(m, caddr_t), datalen, uio);
1691 head->m_pkthdr.len += datalen;
1693 } while (resid > 0);