2 * Copyright (c) 2004 Jeffrey M. Hsu. All rights reserved.
3 * Copyright (c) 2004 The DragonFly Project. All rights reserved.
5 * This code is derived from software contributed to The DragonFly Project
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of The DragonFly Project nor the names of its
17 * contributors may be used to endorse or promote products derived
18 * from this software without specific, prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * Copyright (c) 1982, 1986, 1988, 1991, 1993
36 * The Regents of the University of California. All rights reserved.
38 * Redistribution and use in source and binary forms, with or without
39 * modification, are permitted provided that the following conditions
41 * 1. Redistributions of source code must retain the above copyright
42 * notice, this list of conditions and the following disclaimer.
43 * 2. Redistributions in binary form must reproduce the above copyright
44 * notice, this list of conditions and the following disclaimer in the
45 * documentation and/or other materials provided with the distribution.
46 * 3. All advertising materials mentioning features or use of this software
47 * must display the following acknowledgement:
48 * This product includes software developed by the University of
49 * California, Berkeley and its contributors.
50 * 4. Neither the name of the University nor the names of its contributors
51 * may be used to endorse or promote products derived from this software
52 * without specific prior written permission.
54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
66 * @(#)uipc_mbuf.c 8.2 (Berkeley) 1/4/94
67 * $FreeBSD: src/sys/kern/uipc_mbuf.c,v 1.51.2.24 2003/04/15 06:59:29 silby Exp $
68 * $DragonFly: src/sys/kern/uipc_mbuf.c,v 1.62 2007/05/13 22:56:59 dillon Exp $
71 #include "opt_param.h"
73 #include "opt_mbuf_stress_test.h"
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/malloc.h>
78 #include <sys/kernel.h>
79 #include <sys/sysctl.h>
80 #include <sys/domain.h>
81 #include <sys/objcache.h>
82 #include <sys/protosw.h>
84 #include <sys/thread.h>
85 #include <sys/globaldata.h>
86 #include <sys/serialize.h>
87 #include <sys/thread2.h>
90 #include <vm/vm_kern.h>
91 #include <vm/vm_extern.h>
94 #include <machine/cpu.h>
98 * mbuf cluster meta-data
103 struct lwkt_serialize mcl_serializer;
106 static void mbinit(void *);
107 SYSINIT(mbuf, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, mbinit, NULL)
109 static u_long mbtypes[MT_NTYPES];
111 struct mbstat mbstat;
120 #ifdef MBUF_STRESS_TEST
121 int m_defragrandomfailures;
124 struct objcache *mbuf_cache, *mbufphdr_cache;
125 struct objcache *mclmeta_cache;
126 struct objcache *mbufcluster_cache, *mbufphdrcluster_cache;
131 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW,
132 &max_linkhdr, 0, "");
133 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW,
134 &max_protohdr, 0, "");
135 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0, "");
136 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW,
137 &max_datalen, 0, "");
138 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW,
140 SYSCTL_STRUCT(_kern_ipc, KIPC_MBSTAT, mbstat, CTLFLAG_RW, &mbstat, mbstat, "");
141 SYSCTL_OPAQUE(_kern_ipc, OID_AUTO, mbtypes, CTLFLAG_RD, mbtypes,
142 sizeof(mbtypes), "LU", "");
145 * These are read-only because we do not currently have any code
146 * to adjust the objcache limits after the fact. The variables
147 * may only be set as boot-time tunables.
149 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RD,
150 &nmbclusters, 0, "Maximum number of mbuf clusters available");
151 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RD, &nmbufs, 0,
152 "Maximum number of mbufs available");
154 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
155 &m_defragpackets, 0, "");
156 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
157 &m_defragbytes, 0, "");
158 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
159 &m_defraguseless, 0, "");
160 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
161 &m_defragfailure, 0, "");
162 #ifdef MBUF_STRESS_TEST
163 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
164 &m_defragrandomfailures, 0, "");
167 static MALLOC_DEFINE(M_MBUF, "mbuf", "mbuf");
168 static MALLOC_DEFINE(M_MBUFCL, "mbufcl", "mbufcl");
169 static MALLOC_DEFINE(M_MCLMETA, "mclmeta", "mclmeta");
171 static void m_reclaim (void);
172 static void m_mclref(void *arg);
173 static void m_mclfree(void *arg);
176 #define NMBCLUSTERS (512 + maxusers * 16)
179 #define NMBUFS (nmbclusters * 2)
183 * Perform sanity checks of tunables declared above.
186 tunable_mbinit(void *dummy)
189 * This has to be done before VM init.
191 nmbclusters = NMBCLUSTERS;
192 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
194 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
196 if (nmbufs < nmbclusters * 2)
197 nmbufs = nmbclusters * 2;
199 SYSINIT(tunable_mbinit, SI_BOOT1_TUNABLES, SI_ORDER_ANY,
200 tunable_mbinit, NULL);
202 /* "number of clusters of pages" */
208 * The mbuf object cache only guarantees that m_next and m_nextpkt are
209 * NULL and that m_data points to the beginning of the data area. In
210 * particular, m_len and m_pkthdr.len are uninitialized. It is the
211 * responsibility of the caller to initialize those fields before use.
214 static boolean_t __inline
215 mbuf_ctor(void *obj, void *private, int ocflags)
217 struct mbuf *m = obj;
221 m->m_data = m->m_dat;
228 * Initialize the mbuf and the packet header fields.
231 mbufphdr_ctor(void *obj, void *private, int ocflags)
233 struct mbuf *m = obj;
237 m->m_data = m->m_pktdat;
238 m->m_flags = M_PKTHDR | M_PHCACHE;
240 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */
241 SLIST_INIT(&m->m_pkthdr.tags);
242 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */
243 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */
249 * A mbcluster object consists of 2K (MCLBYTES) cluster and a refcount.
252 mclmeta_ctor(void *obj, void *private, int ocflags)
254 struct mbcluster *cl = obj;
257 if (ocflags & M_NOWAIT)
258 buf = kmalloc(MCLBYTES, M_MBUFCL, M_NOWAIT | M_ZERO);
260 buf = kmalloc(MCLBYTES, M_MBUFCL, M_INTWAIT | M_ZERO);
265 lwkt_serialize_init(&cl->mcl_serializer);
270 mclmeta_dtor(void *obj, void *private)
272 struct mbcluster *mcl = obj;
274 KKASSERT(mcl->mcl_refs == 0);
275 kfree(mcl->mcl_data, M_MBUFCL);
279 linkcluster(struct mbuf *m, struct mbcluster *cl)
282 * Add the cluster to the mbuf. The caller will detect that the
283 * mbuf now has an attached cluster.
285 m->m_ext.ext_arg = cl;
286 m->m_ext.ext_buf = cl->mcl_data;
287 m->m_ext.ext_ref = m_mclref;
288 m->m_ext.ext_free = m_mclfree;
289 m->m_ext.ext_size = MCLBYTES;
290 atomic_add_int(&cl->mcl_refs, 1);
292 m->m_data = m->m_ext.ext_buf;
293 m->m_flags |= M_EXT | M_EXT_CLUSTER;
297 mbufphdrcluster_ctor(void *obj, void *private, int ocflags)
299 struct mbuf *m = obj;
300 struct mbcluster *cl;
302 mbufphdr_ctor(obj, private, ocflags);
303 cl = objcache_get(mclmeta_cache, ocflags);
306 m->m_flags |= M_CLCACHE;
312 mbufcluster_ctor(void *obj, void *private, int ocflags)
314 struct mbuf *m = obj;
315 struct mbcluster *cl;
317 mbuf_ctor(obj, private, ocflags);
318 cl = objcache_get(mclmeta_cache, ocflags);
321 m->m_flags |= M_CLCACHE;
327 * Used for both the cluster and cluster PHDR caches.
329 * The mbuf may have lost its cluster due to sharing, deal
330 * with the situation by checking M_EXT.
333 mbufcluster_dtor(void *obj, void *private)
335 struct mbuf *m = obj;
336 struct mbcluster *mcl;
338 if (m->m_flags & M_EXT) {
339 KKASSERT((m->m_flags & M_EXT_CLUSTER) != 0);
340 mcl = m->m_ext.ext_arg;
341 KKASSERT(mcl->mcl_refs == 1);
343 objcache_put(mclmeta_cache, mcl);
347 struct objcache_malloc_args mbuf_malloc_args = { MSIZE, M_MBUF };
348 struct objcache_malloc_args mclmeta_malloc_args =
349 { sizeof(struct mbcluster), M_MCLMETA };
355 mbstat.m_msize = MSIZE;
356 mbstat.m_mclbytes = MCLBYTES;
357 mbstat.m_minclsize = MINCLSIZE;
358 mbstat.m_mlen = MLEN;
359 mbstat.m_mhlen = MHLEN;
361 mbuf_cache = objcache_create("mbuf", nmbufs, 0,
362 mbuf_ctor, NULL, NULL,
363 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
364 mbufphdr_cache = objcache_create("mbuf pkt hdr", nmbufs, 64,
365 mbufphdr_ctor, NULL, NULL,
366 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
367 mclmeta_cache = objcache_create("cluster mbuf", nmbclusters , 0,
368 mclmeta_ctor, mclmeta_dtor, NULL,
369 objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args);
370 mbufcluster_cache = objcache_create("mbuf + cluster", nmbclusters, 0,
371 mbufcluster_ctor, mbufcluster_dtor, NULL,
372 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
373 mbufphdrcluster_cache = objcache_create("mbuf pkt hdr + cluster",
374 nmbclusters, 64, mbufphdrcluster_ctor, mbufcluster_dtor, NULL,
375 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
380 * Return the number of references to this mbuf's data. 0 is returned
381 * if the mbuf is not M_EXT, a reference count is returned if it is
382 * M_EXT | M_EXT_CLUSTER, and 99 is returned if it is a special M_EXT.
385 m_sharecount(struct mbuf *m)
387 switch (m->m_flags & (M_EXT | M_EXT_CLUSTER)) {
392 case M_EXT | M_EXT_CLUSTER:
393 return (((struct mbcluster *)m->m_ext.ext_arg)->mcl_refs);
396 return (0); /* to shut up compiler */
400 * change mbuf to new type
403 m_chtype(struct mbuf *m, int type)
407 --mbtypes[m->m_type];
419 SLIST_FOREACH(dp, &domains, dom_next) {
420 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) {
430 updatestats(struct mbuf *m, int type)
444 m_get(int how, int type)
448 int ocf = MBTOM(how);
452 m = objcache_get(mbuf_cache, ocf);
455 if ((how & MB_TRYWAIT) && ntries++ == 0) {
456 struct objcache *reclaimlist[] = {
458 mbufcluster_cache, mbufphdrcluster_cache
460 const int nreclaims = __arysize(reclaimlist);
462 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
469 updatestats(m, type);
474 m_gethdr(int how, int type)
477 int ocf = MBTOM(how);
482 m = objcache_get(mbufphdr_cache, ocf);
485 if ((how & MB_TRYWAIT) && ntries++ == 0) {
486 struct objcache *reclaimlist[] = {
488 mbufcluster_cache, mbufphdrcluster_cache
490 const int nreclaims = __arysize(reclaimlist);
492 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
499 updatestats(m, type);
504 * Get a mbuf (not a mbuf cluster!) and zero it.
508 m_getclr(int how, int type)
512 m = m_get(how, type);
514 bzero(m->m_data, MLEN);
519 * Returns an mbuf with an attached cluster.
520 * Because many network drivers use this kind of buffers a lot, it is
521 * convenient to keep a small pool of free buffers of this kind.
522 * Even a small size such as 10 gives about 10% improvement in the
523 * forwarding rate in a bridge or router.
526 m_getcl(int how, short type, int flags)
529 int ocflags = MBTOM(how);
534 if (flags & M_PKTHDR)
535 m = objcache_get(mbufphdrcluster_cache, ocflags);
537 m = objcache_get(mbufcluster_cache, ocflags);
540 if ((how & MB_TRYWAIT) && ntries++ == 0) {
541 struct objcache *reclaimlist[1];
543 if (flags & M_PKTHDR)
544 reclaimlist[0] = mbufcluster_cache;
546 reclaimlist[0] = mbufphdrcluster_cache;
547 if (!objcache_reclaimlist(reclaimlist, 1, ocflags))
564 * Allocate chain of requested length.
567 m_getc(int len, int how, int type)
569 struct mbuf *n, *nfirst = NULL, **ntail = &nfirst;
573 n = m_getl(len, how, type, 0, &nsize);
589 * Allocate len-worth of mbufs and/or mbuf clusters (whatever fits best)
590 * and return a pointer to the head of the allocated chain. If m0 is
591 * non-null, then we assume that it is a single mbuf or an mbuf chain to
592 * which we want len bytes worth of mbufs and/or clusters attached, and so
593 * if we succeed in allocating it, we will just return a pointer to m0.
595 * If we happen to fail at any point during the allocation, we will free
596 * up everything we have already allocated and return NULL.
598 * Deprecated. Use m_getc() and m_cat() instead.
601 m_getm(struct mbuf *m0, int len, int type, int how)
605 nfirst = m_getc(len, how, type);
608 m_last(m0)->m_next = nfirst;
616 * Adds a cluster to a normal mbuf, M_EXT is set on success.
617 * Deprecated. Use m_getcl() instead.
620 m_mclget(struct mbuf *m, int how)
622 struct mbcluster *mcl;
624 KKASSERT((m->m_flags & M_EXT) == 0);
625 mcl = objcache_get(mclmeta_cache, MBTOM(how));
630 /* leave the m_mbufs count intact for original mbuf */
636 * Updates to mbcluster must be MPSAFE. Only an entity which already has
637 * a reference to the cluster can ref it, so we are in no danger of
638 * racing an add with a subtract. But the operation must still be atomic
639 * since multiple entities may have a reference on the cluster.
641 * m_mclfree() is almost the same but it must contend with two entities
642 * freeing the cluster at the same time. If there is only one reference
643 * count we are the only entity referencing the cluster and no further
644 * locking is required. Otherwise we must protect against a race to 0
645 * with the serializer.
650 struct mbcluster *mcl = arg;
652 atomic_add_int(&mcl->mcl_refs, 1);
658 struct mbcluster *mcl = arg;
660 if (mcl->mcl_refs == 1) {
662 objcache_put(mclmeta_cache, mcl);
664 lwkt_serialize_enter(&mcl->mcl_serializer);
665 if (mcl->mcl_refs > 1) {
666 atomic_subtract_int(&mcl->mcl_refs, 1);
667 lwkt_serialize_exit(&mcl->mcl_serializer);
669 lwkt_serialize_exit(&mcl->mcl_serializer);
670 KKASSERT(mcl->mcl_refs == 1);
672 objcache_put(mclmeta_cache, mcl);
677 extern void db_print_backtrace(void);
680 * Free a single mbuf and any associated external storage. The successor,
681 * if any, is returned.
683 * We do need to check non-first mbuf for m_aux, since some of existing
684 * code does not call M_PREPEND properly.
685 * (example: call to bpf_mtap from drivers)
688 m_free(struct mbuf *m)
692 KASSERT(m->m_type != MT_FREE, ("freeing free mbuf %p", m));
693 --mbtypes[m->m_type];
698 * Make sure the mbuf is in constructed state before returning it
703 KKASSERT(m->m_nextpkt == NULL);
705 if (m->m_nextpkt != NULL) {
707 static int afewtimes = 10;
709 if (afewtimes-- > 0) {
710 kprintf("mfree: m->m_nextpkt != NULL\n");
711 db_print_backtrace();
717 if (m->m_flags & M_PKTHDR) {
718 m_tag_delete_chain(m); /* eliminate XXX JH */
721 m->m_flags &= (M_EXT | M_EXT_CLUSTER | M_CLCACHE | M_PHCACHE);
724 * Clean the M_PKTHDR state so we can return the mbuf to its original
725 * cache. This is based on the PHCACHE flag which tells us whether
726 * the mbuf was originally allocated out of a packet-header cache
727 * or a non-packet-header cache.
729 if (m->m_flags & M_PHCACHE) {
730 m->m_flags |= M_PKTHDR;
731 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */
732 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */
733 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */
734 SLIST_INIT(&m->m_pkthdr.tags);
738 * Handle remaining flags combinations. M_CLCACHE tells us whether
739 * the mbuf was originally allocated from a cluster cache or not,
740 * and is totally separate from whether the mbuf is currently
741 * associated with a cluster.
744 switch(m->m_flags & (M_CLCACHE | M_EXT | M_EXT_CLUSTER)) {
745 case M_CLCACHE | M_EXT | M_EXT_CLUSTER:
747 * mbuf+cluster cache case. The mbuf was allocated from the
748 * combined mbuf_cluster cache and can be returned to the
749 * cache if the cluster hasn't been shared.
751 if (m_sharecount(m) == 1) {
753 * The cluster has not been shared, we can just
754 * reset the data pointer and return the mbuf
755 * to the cluster cache. Note that the reference
756 * count is left intact (it is still associated with
759 m->m_data = m->m_ext.ext_buf;
760 if (m->m_flags & M_PHCACHE)
761 objcache_put(mbufphdrcluster_cache, m);
763 objcache_put(mbufcluster_cache, m);
767 * Hell. Someone else has a ref on this cluster,
768 * we have to disconnect it which means we can't
769 * put it back into the mbufcluster_cache, we
770 * have to destroy the mbuf.
772 * Other mbuf references to the cluster will typically
773 * be M_EXT | M_EXT_CLUSTER but without M_CLCACHE.
775 * XXX we could try to connect another cluster to
778 m->m_ext.ext_free(m->m_ext.ext_arg);
779 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
780 if (m->m_flags & M_PHCACHE)
781 objcache_dtor(mbufphdrcluster_cache, m);
783 objcache_dtor(mbufcluster_cache, m);
786 case M_EXT | M_EXT_CLUSTER:
788 * Normal cluster associated with an mbuf that was allocated
789 * from the normal mbuf pool rather then the cluster pool.
790 * The cluster has to be independantly disassociated from the
793 if (m_sharecount(m) == 1)
798 * Normal cluster association case, disconnect the cluster from
799 * the mbuf. The cluster may or may not be custom.
801 m->m_ext.ext_free(m->m_ext.ext_arg);
802 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
806 * return the mbuf to the mbuf cache.
808 if (m->m_flags & M_PHCACHE) {
809 m->m_data = m->m_pktdat;
810 objcache_put(mbufphdr_cache, m);
812 m->m_data = m->m_dat;
813 objcache_put(mbuf_cache, m);
819 panic("bad mbuf flags %p %08x\n", m, m->m_flags);
827 m_freem(struct mbuf *m)
836 * mbuf utility routines
840 * Lesser-used path for M_PREPEND: allocate new mbuf to prepend to chain and
844 m_prepend(struct mbuf *m, int len, int how)
848 if (m->m_flags & M_PKTHDR)
849 mn = m_gethdr(how, m->m_type);
851 mn = m_get(how, m->m_type);
856 if (m->m_flags & M_PKTHDR)
857 M_MOVE_PKTHDR(mn, m);
867 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
868 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
869 * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller.
870 * Note that the copy is read-only, because clusters are not copied,
871 * only their reference counts are incremented.
874 m_copym(const struct mbuf *m, int off0, int len, int wait)
876 struct mbuf *n, **np;
881 KASSERT(off >= 0, ("m_copym, negative off %d", off));
882 KASSERT(len >= 0, ("m_copym, negative len %d", len));
883 if (off == 0 && m->m_flags & M_PKTHDR)
886 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
896 KASSERT(len == M_COPYALL,
897 ("m_copym, length > size of mbuf chain"));
901 * Because we are sharing any cluster attachment below,
902 * be sure to get an mbuf that does not have a cluster
903 * associated with it.
906 n = m_gethdr(wait, m->m_type);
908 n = m_get(wait, m->m_type);
913 if (!m_dup_pkthdr(n, m, wait))
915 if (len == M_COPYALL)
916 n->m_pkthdr.len -= off0;
918 n->m_pkthdr.len = len;
921 n->m_len = min(len, m->m_len - off);
922 if (m->m_flags & M_EXT) {
923 KKASSERT((n->m_flags & M_EXT) == 0);
924 n->m_data = m->m_data + off;
925 m->m_ext.ext_ref(m->m_ext.ext_arg);
927 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
929 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
932 if (len != M_COPYALL)
948 * Copy an entire packet, including header (which must be present).
949 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
950 * Note that the copy is read-only, because clusters are not copied,
951 * only their reference counts are incremented.
952 * Preserve alignment of the first mbuf so if the creator has left
953 * some room at the beginning (e.g. for inserting protocol headers)
954 * the copies also have the room available.
957 m_copypacket(struct mbuf *m, int how)
959 struct mbuf *top, *n, *o;
961 n = m_gethdr(how, m->m_type);
966 if (!m_dup_pkthdr(n, m, how))
969 if (m->m_flags & M_EXT) {
970 KKASSERT((n->m_flags & M_EXT) == 0);
971 n->m_data = m->m_data;
972 m->m_ext.ext_ref(m->m_ext.ext_arg);
974 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
976 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
977 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
982 o = m_get(how, m->m_type);
990 if (m->m_flags & M_EXT) {
991 KKASSERT((n->m_flags & M_EXT) == 0);
992 n->m_data = m->m_data;
993 m->m_ext.ext_ref(m->m_ext.ext_arg);
995 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
997 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1010 * Copy data from an mbuf chain starting "off" bytes from the beginning,
1011 * continuing for "len" bytes, into the indicated buffer.
1014 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
1018 KASSERT(off >= 0, ("m_copydata, negative off %d", off));
1019 KASSERT(len >= 0, ("m_copydata, negative len %d", len));
1021 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
1028 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
1029 count = min(m->m_len - off, len);
1030 bcopy(mtod(m, caddr_t) + off, cp, count);
1039 * Copy a packet header mbuf chain into a completely new chain, including
1040 * copying any mbuf clusters. Use this instead of m_copypacket() when
1041 * you need a writable copy of an mbuf chain.
1044 m_dup(struct mbuf *m, int how)
1046 struct mbuf **p, *top = NULL;
1047 int remain, moff, nsize;
1052 KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __func__));
1054 /* While there's more data, get a new mbuf, tack it on, and fill it */
1055 remain = m->m_pkthdr.len;
1058 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */
1061 /* Get the next new mbuf */
1062 n = m_getl(remain, how, m->m_type, top == NULL ? M_PKTHDR : 0,
1067 if (!m_dup_pkthdr(n, m, how))
1070 /* Link it into the new chain */
1074 /* Copy data from original mbuf(s) into new mbuf */
1076 while (n->m_len < nsize && m != NULL) {
1077 int chunk = min(nsize - n->m_len, m->m_len - moff);
1079 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1083 if (moff == m->m_len) {
1089 /* Check correct total mbuf length */
1090 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
1091 ("%s: bogus m_pkthdr.len", __func__));
1103 * Concatenate mbuf chain n to m.
1104 * Both chains must be of the same type (e.g. MT_DATA).
1105 * Any m_pkthdr is not updated.
1108 m_cat(struct mbuf *m, struct mbuf *n)
1112 if (m->m_flags & M_EXT ||
1113 m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) {
1114 /* just join the two chains */
1118 /* splat the data from one into the other */
1119 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1121 m->m_len += n->m_len;
1127 m_adj(struct mbuf *mp, int req_len)
1133 if ((m = mp) == NULL)
1139 while (m != NULL && len > 0) {
1140 if (m->m_len <= len) {
1151 if (mp->m_flags & M_PKTHDR)
1152 m->m_pkthdr.len -= (req_len - len);
1155 * Trim from tail. Scan the mbuf chain,
1156 * calculating its length and finding the last mbuf.
1157 * If the adjustment only affects this mbuf, then just
1158 * adjust and return. Otherwise, rescan and truncate
1159 * after the remaining size.
1165 if (m->m_next == (struct mbuf *)0)
1169 if (m->m_len >= len) {
1171 if (mp->m_flags & M_PKTHDR)
1172 mp->m_pkthdr.len -= len;
1179 * Correct length for chain is "count".
1180 * Find the mbuf with last data, adjust its length,
1181 * and toss data from remaining mbufs on chain.
1184 if (m->m_flags & M_PKTHDR)
1185 m->m_pkthdr.len = count;
1186 for (; m; m = m->m_next) {
1187 if (m->m_len >= count) {
1194 (m = m->m_next) ->m_len = 0;
1199 * Rearrange an mbuf chain so that len bytes are contiguous
1200 * and in the data area of an mbuf (so that mtod will work for a structure
1201 * of size len). Returns the resulting mbuf chain on success, frees it and
1202 * returns null on failure. If there is room, it will add up to
1203 * max_protohdr-len extra bytes to the contiguous region in an attempt to
1204 * avoid being called next time.
1207 m_pullup(struct mbuf *n, int len)
1214 * If first mbuf has no cluster, and has room for len bytes
1215 * without shifting current data, pullup into it,
1216 * otherwise allocate a new mbuf to prepend to the chain.
1218 if (!(n->m_flags & M_EXT) &&
1219 n->m_data + len < &n->m_dat[MLEN] &&
1221 if (n->m_len >= len)
1229 if (n->m_flags & M_PKTHDR)
1230 m = m_gethdr(MB_DONTWAIT, n->m_type);
1232 m = m_get(MB_DONTWAIT, n->m_type);
1236 if (n->m_flags & M_PKTHDR)
1237 M_MOVE_PKTHDR(m, n);
1239 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
1241 count = min(min(max(len, max_protohdr), space), n->m_len);
1242 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1252 } while (len > 0 && n);
1266 * Partition an mbuf chain in two pieces, returning the tail --
1267 * all but the first len0 bytes. In case of failure, it returns NULL and
1268 * attempts to restore the chain to its original state.
1270 * Note that the resulting mbufs might be read-only, because the new
1271 * mbuf can end up sharing an mbuf cluster with the original mbuf if
1272 * the "breaking point" happens to lie within a cluster mbuf. Use the
1273 * M_WRITABLE() macro to check for this case.
1276 m_split(struct mbuf *m0, int len0, int wait)
1279 unsigned len = len0, remain;
1281 for (m = m0; m && len > m->m_len; m = m->m_next)
1285 remain = m->m_len - len;
1286 if (m0->m_flags & M_PKTHDR) {
1287 n = m_gethdr(wait, m0->m_type);
1290 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1291 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1292 m0->m_pkthdr.len = len0;
1293 if (m->m_flags & M_EXT)
1295 if (remain > MHLEN) {
1296 /* m can't be the lead packet */
1298 n->m_next = m_split(m, len, wait);
1299 if (n->m_next == NULL) {
1307 MH_ALIGN(n, remain);
1308 } else if (remain == 0) {
1313 n = m_get(wait, m->m_type);
1319 if (m->m_flags & M_EXT) {
1320 KKASSERT((n->m_flags & M_EXT) == 0);
1321 n->m_data = m->m_data + len;
1322 m->m_ext.ext_ref(m->m_ext.ext_arg);
1323 n->m_ext = m->m_ext;
1324 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1326 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1330 n->m_next = m->m_next;
1336 * Routine to copy from device local memory into mbufs.
1337 * Note: "offset" is ill-defined and always called as 0, so ignore it.
1340 m_devget(char *buf, int len, int offset, struct ifnet *ifp,
1341 void (*copy)(volatile const void *from, volatile void *to, size_t length))
1343 struct mbuf *m, *mfirst = NULL, **mtail;
1352 m = m_getl(len, MB_DONTWAIT, MT_DATA, flags, &nsize);
1357 m->m_len = min(len, nsize);
1359 if (flags & M_PKTHDR) {
1360 if (len + max_linkhdr <= nsize)
1361 m->m_data += max_linkhdr;
1362 m->m_pkthdr.rcvif = ifp;
1363 m->m_pkthdr.len = len;
1367 copy(buf, m->m_data, (unsigned)m->m_len);
1378 * Copy data from a buffer back into the indicated mbuf chain,
1379 * starting "off" bytes from the beginning, extending the mbuf
1380 * chain if necessary.
1383 m_copyback(struct mbuf *m0, int off, int len, caddr_t cp)
1386 struct mbuf *m = m0, *n;
1391 while (off > (mlen = m->m_len)) {
1394 if (m->m_next == NULL) {
1395 n = m_getclr(MB_DONTWAIT, m->m_type);
1398 n->m_len = min(MLEN, len + off);
1404 mlen = min (m->m_len - off, len);
1405 bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen);
1413 if (m->m_next == NULL) {
1414 n = m_get(MB_DONTWAIT, m->m_type);
1417 n->m_len = min(MLEN, len);
1422 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1423 m->m_pkthdr.len = totlen;
1427 m_print(const struct mbuf *m)
1430 const struct mbuf *m2;
1432 len = m->m_pkthdr.len;
1435 kprintf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-");
1443 * "Move" mbuf pkthdr from "from" to "to".
1444 * "from" must have M_PKTHDR set, and "to" must be empty.
1447 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
1449 KASSERT((to->m_flags & M_PKTHDR), ("m_move_pkthdr: not packet header"));
1451 to->m_flags |= from->m_flags & M_COPYFLAGS;
1452 to->m_pkthdr = from->m_pkthdr; /* especially tags */
1453 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */
1457 * Duplicate "from"'s mbuf pkthdr in "to".
1458 * "from" must have M_PKTHDR set, and "to" must be empty.
1459 * In particular, this does a deep copy of the packet tags.
1462 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
1464 KASSERT((to->m_flags & M_PKTHDR), ("m_dup_pkthdr: not packet header"));
1466 to->m_flags = (from->m_flags & M_COPYFLAGS) |
1467 (to->m_flags & ~M_COPYFLAGS);
1468 to->m_pkthdr = from->m_pkthdr;
1469 SLIST_INIT(&to->m_pkthdr.tags);
1470 return (m_tag_copy_chain(to, from, how));
1474 * Defragment a mbuf chain, returning the shortest possible
1475 * chain of mbufs and clusters. If allocation fails and
1476 * this cannot be completed, NULL will be returned, but
1477 * the passed in chain will be unchanged. Upon success,
1478 * the original chain will be freed, and the new chain
1481 * If a non-packet header is passed in, the original
1482 * mbuf (chain?) will be returned unharmed.
1484 * m_defrag_nofree doesn't free the passed in mbuf.
1487 m_defrag(struct mbuf *m0, int how)
1491 if ((m_new = m_defrag_nofree(m0, how)) == NULL)
1499 m_defrag_nofree(struct mbuf *m0, int how)
1501 struct mbuf *m_new = NULL, *m_final = NULL;
1502 int progress = 0, length, nsize;
1504 if (!(m0->m_flags & M_PKTHDR))
1507 #ifdef MBUF_STRESS_TEST
1508 if (m_defragrandomfailures) {
1509 int temp = karc4random() & 0xff;
1515 m_final = m_getl(m0->m_pkthdr.len, how, MT_DATA, M_PKTHDR, &nsize);
1516 if (m_final == NULL)
1518 m_final->m_len = 0; /* in case m0->m_pkthdr.len is zero */
1520 if (m_dup_pkthdr(m_final, m0, how) == NULL)
1525 while (progress < m0->m_pkthdr.len) {
1526 length = m0->m_pkthdr.len - progress;
1527 if (length > MCLBYTES)
1530 if (m_new == NULL) {
1531 m_new = m_getl(length, how, MT_DATA, 0, &nsize);
1536 m_copydata(m0, progress, length, mtod(m_new, caddr_t));
1538 m_new->m_len = length;
1539 if (m_new != m_final)
1540 m_cat(m_final, m_new);
1543 if (m0->m_next == NULL)
1546 m_defragbytes += m_final->m_pkthdr.len;
1557 * Move data from uio into mbufs.
1560 m_uiomove(struct uio *uio)
1562 struct mbuf *m; /* current working mbuf */
1563 struct mbuf *head = NULL; /* result mbuf chain */
1564 struct mbuf **mp = &head;
1565 int resid = uio->uio_resid, nsize, flags = M_PKTHDR, error;
1568 m = m_getl(resid, MB_WAIT, MT_DATA, flags, &nsize);
1570 m->m_pkthdr.len = 0;
1571 /* Leave room for protocol headers. */
1576 m->m_len = min(nsize, resid);
1577 error = uiomove(mtod(m, caddr_t), m->m_len, uio);
1584 head->m_pkthdr.len += m->m_len;
1586 } while (resid > 0);
1596 m_last(struct mbuf *m)
1604 * Return the number of bytes in an mbuf chain.
1605 * If lastm is not NULL, also return the last mbuf.
1608 m_lengthm(struct mbuf *m, struct mbuf **lastm)
1611 struct mbuf *prev = m;
1624 * Like m_lengthm(), except also keep track of mbuf usage.
1627 m_countm(struct mbuf *m, struct mbuf **lastm, u_int *pmbcnt)
1629 u_int len = 0, mbcnt = 0;
1630 struct mbuf *prev = m;
1635 if (m->m_flags & M_EXT)
1636 mbcnt += m->m_ext.ext_size;