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) 2004 Jeffrey M. Hsu. All rights reserved.
37 * License terms: all terms for the DragonFly license above plus the following:
39 * 4. All advertising materials mentioning features or use of this software
40 * must display the following acknowledgement:
42 * This product includes software developed by Jeffrey M. Hsu
43 * for the DragonFly Project.
45 * This requirement may be waived with permission from Jeffrey Hsu.
46 * This requirement will sunset and may be removed on July 8 2005,
47 * after which the standard DragonFly license (as shown above) will
52 * Copyright (c) 1982, 1986, 1988, 1991, 1993
53 * The Regents of the University of California. All rights reserved.
55 * Redistribution and use in source and binary forms, with or without
56 * modification, are permitted provided that the following conditions
58 * 1. Redistributions of source code must retain the above copyright
59 * notice, this list of conditions and the following disclaimer.
60 * 2. Redistributions in binary form must reproduce the above copyright
61 * notice, this list of conditions and the following disclaimer in the
62 * documentation and/or other materials provided with the distribution.
63 * 3. All advertising materials mentioning features or use of this software
64 * must display the following acknowledgement:
65 * This product includes software developed by the University of
66 * California, Berkeley and its contributors.
67 * 4. Neither the name of the University nor the names of its contributors
68 * may be used to endorse or promote products derived from this software
69 * without specific prior written permission.
71 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
72 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
73 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
74 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
75 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
76 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
77 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
78 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
79 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
80 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
83 * @(#)uipc_mbuf.c 8.2 (Berkeley) 1/4/94
84 * $FreeBSD: src/sys/kern/uipc_mbuf.c,v 1.51.2.24 2003/04/15 06:59:29 silby Exp $
85 * $DragonFly: src/sys/kern/uipc_mbuf.c,v 1.28 2004/09/19 22:32:47 joerg Exp $
88 #include "opt_param.h"
89 #include "opt_mbuf_stress_test.h"
90 #include <sys/param.h>
91 #include <sys/systm.h>
92 #include <sys/malloc.h>
94 #include <sys/kernel.h>
95 #include <sys/sysctl.h>
96 #include <sys/domain.h>
97 #include <sys/protosw.h>
99 #include <sys/thread.h>
100 #include <sys/globaldata.h>
101 #include <sys/thread2.h>
104 #include <vm/vm_kern.h>
105 #include <vm/vm_extern.h>
108 #include <machine/cpu.h>
112 * mbuf cluster meta-data
114 typedef struct mbcluster {
115 struct mbcluster *mcl_next;
121 typedef struct mbuf *mbuf_t;
123 #define MCL_MAGIC 0x6d62636c
125 static void mbinit (void *);
126 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbinit, NULL)
128 static u_long mbtypes[MT_NTYPES];
130 struct mbstat mbstat;
139 #ifdef MBUF_STRESS_TEST
140 int m_defragrandomfailures;
145 u_int m_mballoc_wid = 0;
146 u_int m_clalloc_wid = 0;
148 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW,
149 &max_linkhdr, 0, "");
150 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW,
151 &max_protohdr, 0, "");
152 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0, "");
153 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW,
154 &max_datalen, 0, "");
155 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW,
157 SYSCTL_STRUCT(_kern_ipc, KIPC_MBSTAT, mbstat, CTLFLAG_RW, &mbstat, mbstat, "");
158 SYSCTL_OPAQUE(_kern_ipc, OID_AUTO, mbtypes, CTLFLAG_RD, mbtypes,
159 sizeof(mbtypes), "LU", "");
160 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RW,
161 &nmbclusters, 0, "Maximum number of mbuf clusters available");
162 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RW, &nmbufs, 0,
163 "Maximum number of mbufs available");
164 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
165 &m_defragpackets, 0, "");
166 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
167 &m_defragbytes, 0, "");
168 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
169 &m_defraguseless, 0, "");
170 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
171 &m_defragfailure, 0, "");
172 #ifdef MBUF_STRESS_TEST
173 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
174 &m_defragrandomfailures, 0, "");
177 static int mcl_pool_count;
178 static int mcl_pool_max = 20;
179 static int mcl_free_max = 1000;
180 static int mbuf_free_max = 5000;
182 SYSCTL_INT(_kern_ipc, OID_AUTO, mcl_pool_max, CTLFLAG_RW, &mcl_pool_max, 0,
183 "Maximum number of mbufs+cluster in free list");
184 SYSCTL_INT(_kern_ipc, OID_AUTO, mcl_pool_count, CTLFLAG_RD, &mcl_pool_count, 0,
185 "Current number of mbufs+cluster in free list");
186 SYSCTL_INT(_kern_ipc, OID_AUTO, mcl_free_max, CTLFLAG_RW, &mcl_free_max, 0,
187 "Maximum number of clusters on the free list");
188 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_free_max, CTLFLAG_RW, &mbuf_free_max, 0,
189 "Maximum number of mbufs on the free list");
191 static MALLOC_DEFINE(M_MBUF, "mbuf", "mbuf");
192 static MALLOC_DEFINE(M_MBUFCL, "mbufcl", "mbufcl");
194 static mbuf_t mmbfree;
195 static mbcluster_t mclfree;
196 static struct mbuf *mcl_pool;
198 static void m_reclaim (void);
199 static int m_mballoc(int nmb, int how);
200 static int m_clalloc(int ncl, int how);
201 static struct mbuf *m_mballoc_wait(int caller, int type);
202 static void m_mclref(void *arg);
203 static void m_mclfree(void *arg);
206 #define NMBCLUSTERS (512 + maxusers * 16)
209 #define NMBUFS (nmbclusters * 4)
213 * Perform sanity checks of tunables declared above.
216 tunable_mbinit(void *dummy)
220 * This has to be done before VM init.
222 nmbclusters = NMBCLUSTERS;
223 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
225 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
227 if (nmbufs < nmbclusters * 2)
228 nmbufs = nmbclusters * 2;
232 SYSINIT(tunable_mbinit, SI_SUB_TUNABLES, SI_ORDER_ANY, tunable_mbinit, NULL);
234 /* "number of clusters of pages" */
247 mbstat.m_msize = MSIZE;
248 mbstat.m_mclbytes = MCLBYTES;
249 mbstat.m_minclsize = MINCLSIZE;
250 mbstat.m_mlen = MLEN;
251 mbstat.m_mhlen = MHLEN;
254 if (m_mballoc(NMB_INIT, MB_DONTWAIT) == 0)
256 #if MCLBYTES <= PAGE_SIZE
257 if (m_clalloc(NCL_INIT, MB_DONTWAIT) == 0)
260 /* It's OK to call contigmalloc in this context. */
261 if (m_clalloc(16, MB_WAIT) == 0)
271 * Allocate at least nmb mbufs and place on mbuf free list.
272 * Returns the number of mbufs successfully allocated, 0 if none.
274 * Must be called at splimp.
277 m_mballoc(int nmb, int how)
283 * If we've hit the mbuf limit, stop allocating (or trying to)
284 * in order to avoid exhausting kernel memory entirely.
286 if ((nmb + mbstat.m_mbufs) > nmbufs)
290 * Attempt to allocate the requested number of mbufs, terminate when
291 * the allocation fails but if blocking is allowed allocate at least
294 for (i = 0; i < nmb; ++i) {
295 m = malloc(MSIZE, M_MBUF, M_NOWAIT|M_NULLOK|M_ZERO);
297 if (how == MB_WAIT) {
299 m = malloc(MSIZE, M_MBUF,
300 M_WAITOK|M_NULLOK|M_ZERO);
315 * Once mbuf memory has been exhausted and if the call to the allocation macros
316 * (or, in some cases, functions) is with MB_WAIT, then it is necessary to rely
317 * solely on reclaimed mbufs. Here we wait for an mbuf to be freed for a
318 * designated (mbuf_wait) time.
321 m_mballoc_wait(int caller, int type)
328 if ((tsleep(&m_mballoc_wid, 0, "mballc", mbuf_wait)) == EWOULDBLOCK)
333 * Now that we (think) that we've got something, we will redo an
334 * MGET, but avoid getting into another instance of m_mballoc_wait()
335 * XXX: We retry to fetch _even_ if the sleep timed out. This is left
336 * this way, purposely, in the [unlikely] case that an mbuf was
337 * freed but the sleep was not awakened in time.
342 MGET(m, MB_DONTWAIT, type);
345 MGETHDR(m, MB_DONTWAIT, type);
348 panic("m_mballoc_wait: invalid caller (%d)", caller);
352 if (m != NULL) { /* We waited and got something... */
354 /* Wake up another if we have more free. */
362 #if MCLBYTES > PAGE_SIZE
363 static int i_want_my_mcl;
373 tsleep(&i_want_my_mcl, 0, "mclalloc", 0);
375 while (i_want_my_mcl > 0) {
376 if (m_clalloc(1, MB_WAIT) == 0)
377 printf("m_clalloc failed even in thread context!\n");
385 static struct thread *mclallocthread;
386 static struct kproc_desc mclalloc_kp = {
391 SYSINIT(mclallocthread, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start,
396 * Allocate at least nmb mbuf clusters and place on mbuf free list.
397 * Returns the number of mbuf clusters successfully allocated, 0 if none.
399 * Must be called at splimp.
402 m_clalloc(int ncl, int how)
404 static int last_report;
410 * If we've hit the mbuf cluster limit, stop allocating (or trying to).
412 if ((ncl + mbstat.m_clusters) > nmbclusters)
416 * Attempt to allocate the requested number of mbuf clusters,
417 * terminate when the allocation fails but if blocking is allowed
418 * allocate at least one.
420 * We need to allocate two structures for each cluster... a
421 * ref counting / governing structure and the actual data. MCLBYTES
422 * should be a power of 2 which means that the slab allocator will
423 * return a buffer that does not cross a page boundary.
425 for (i = 0; i < ncl; ++i) {
429 mcl = malloc(sizeof(*mcl), M_MBUFCL, M_NOWAIT|M_NULLOK|M_ZERO);
430 if (mcl == NULL && how == MB_WAIT) {
432 mcl = malloc(sizeof(*mcl),
433 M_MBUFCL, M_WAITOK|M_NULLOK|M_ZERO);
437 * Physically contiguous data buffer.
439 #if MCLBYTES > PAGE_SIZE
440 if (how != MB_WAIT) {
441 i_want_my_mcl += ncl - i;
442 wakeup(&i_want_my_mcl);
446 data = contigmalloc_map(MCLBYTES, M_MBUFCL,
447 M_WAITOK, 0ul, ~0ul, PAGE_SIZE, 0, kernel_map);
450 data = malloc(MCLBYTES, M_MBUFCL, M_NOWAIT|M_NULLOK);
452 if (how == MB_WAIT) {
454 data = malloc(MCLBYTES, M_MBUFCL,
463 mcl->mcl_next = mclfree;
464 mcl->mcl_data = data;
465 mcl->mcl_magic = MCL_MAGIC;
474 * If we could not allocate any report failure no more often then
479 if (ticks < last_report || (ticks - last_report) >= hz) {
481 printf("All mbuf clusters exhausted, please see tuning(7).\n");
488 * Once cluster memory has been exhausted and the allocation is called with
489 * MB_WAIT, we rely on the mclfree pointers. If nothing is free, we will
490 * sleep for a designated amount of time (mbuf_wait) or until we're woken up
491 * due to sudden mcluster availability.
498 /* If in interrupt context, and INVARIANTS, maintain sanity and die. */
499 KASSERT(mycpu->gd_intr_nesting_level == 0,
500 ("CLALLOC: CANNOT WAIT IN INTERRUPT"));
503 * Sleep until something's available or until we expire.
506 if ((tsleep(&m_clalloc_wid, 0, "mclalc", mbuf_wait)) == EWOULDBLOCK)
510 * Try the allocation once more, and if we see mor then two
511 * free entries wake up others as well.
513 m_clalloc(1, MB_WAIT);
515 if (mclfree && mclfree->mcl_next) {
522 * Return the number of references to this mbuf's data. 0 is returned
523 * if the mbuf is not M_EXT, a reference count is returned if it is
524 * M_EXT|M_EXT_CLUSTER, and 99 is returned if it is a special M_EXT.
527 m_sharecount(struct mbuf *m)
531 switch(m->m_flags & (M_EXT|M_EXT_CLUSTER)) {
538 case M_EXT|M_EXT_CLUSTER:
539 count = ((mbcluster_t)m->m_ext.ext_arg)->mcl_refs;
542 panic("bad mbuf flags: %p", m);
549 * change mbuf to new type
552 m_chtype(struct mbuf *m, int type)
557 --mbtypes[m->m_type];
564 * When MGET fails, ask protocols to free space when short of memory,
565 * then re-attempt to allocate an mbuf.
568 m_retry(int how, int t)
574 * Must only do the reclaim if not in an interrupt context.
576 if (how == MB_WAIT) {
577 KASSERT(mycpu->gd_intr_nesting_level == 0,
578 ("MBALLOC: CANNOT WAIT IN INTERRUPT"));
593 m->m_data = m->m_dat;
598 static int last_report ; /* when we did that (in ticks) */
602 if (ticks < last_report || (ticks - last_report) >= hz) {
604 printf("All mbufs exhausted, please see tuning(7).\n");
611 * As above; retry an MGETHDR.
614 m_retryhdr(int how, int t)
620 * Must only do the reclaim if not in an interrupt context.
622 if (how == MB_WAIT) {
623 KASSERT(mycpu->gd_intr_nesting_level == 0,
624 ("MBALLOC: CANNOT WAIT IN INTERRUPT"));
639 m->m_data = m->m_pktdat;
640 m->m_flags = M_PKTHDR;
641 m->m_pkthdr.rcvif = NULL;
642 SLIST_INIT(&m->m_pkthdr.tags);
643 m->m_pkthdr.csum_flags = 0;
647 static int last_report ; /* when we did that (in ticks) */
651 if (ticks < last_report || (ticks - last_report) >= hz) {
653 printf("All mbufs exhausted, please see tuning(7).\n");
667 for (dp = domains; dp; dp = dp->dom_next) {
668 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) {
678 * Space allocation routines.
679 * These are also available as macros
680 * for critical paths.
683 m_get(int how, int type)
699 m->m_data = m->m_dat;
704 m = m_retry(how, type);
705 if (m == NULL && how == MB_WAIT)
706 m = m_mballoc_wait(MGET_C, type);
712 m_gethdr(int how, int type)
728 m->m_data = m->m_pktdat;
729 m->m_flags = M_PKTHDR;
730 m->m_pkthdr.rcvif = NULL;
731 SLIST_INIT(&m->m_pkthdr.tags);
732 m->m_pkthdr.csum_flags = 0;
733 m->m_pkthdr.pf_flags = 0;
737 m = m_retryhdr(how, type);
738 if (m == NULL && how == MB_WAIT)
739 m = m_mballoc_wait(MGETHDR_C, type);
745 m_getclr(int how, int type)
749 if ((m = m_get(how, type)) != NULL) {
750 bzero(mtod(m, caddr_t), MLEN);
756 * m_getcl() returns an mbuf with an attached cluster.
757 * Because many network drivers use this kind of buffers a lot, it is
758 * convenient to keep a small pool of free buffers of this kind.
759 * Even a small size such as 10 gives about 10% improvement in the
760 * forwarding rate in a bridge or router.
761 * The size of this free list is controlled by the sysctl variable
762 * mcl_pool_max. The list is populated on m_freem(), and used in
763 * m_getcl() if elements are available.
766 m_getcl(int how, short type, int flags)
772 if (flags & M_PKTHDR) {
773 if (type == MT_DATA && mcl_pool) {
775 mcl_pool = mp->m_nextpkt;
778 mp->m_nextpkt = NULL;
779 mp->m_data = mp->m_ext.ext_buf;
780 mp->m_flags = M_PKTHDR|M_EXT|M_EXT_CLUSTER;
781 mp->m_pkthdr.rcvif = NULL;
782 mp->m_pkthdr.csum_flags = 0;
785 MGETHDR(mp, how, type);
792 if ((mp->m_flags & M_EXT) == 0) {
803 * m_getm(m, len, how, type)
805 * This will allocate len-worth of mbufs and/or mbuf clusters (whatever fits
806 * best) and return a pointer to the top of the allocated chain. If m is
807 * non-null, then we assume that it is a single mbuf or an mbuf chain to
808 * which we want len bytes worth of mbufs and/or clusters attached, and so
809 * if we succeed in allocating it, we will just return a pointer to m.
811 * If we happen to fail at any point during the allocation, we will free
812 * up everything we have already allocated and return NULL.
816 m_getm(struct mbuf *m, int len, int how, int type)
818 struct mbuf *top, *tail, *mp, *mtail = NULL;
820 KASSERT(len >= 0, ("len is < 0 in m_getm"));
822 mp = m_get(how, type);
825 } else if (len > MINCLSIZE) {
827 if ((mp->m_flags & M_EXT) == 0) {
833 len -= M_TRAILINGSPACE(mp);
836 for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next)
844 mp = m_get(how, type);
850 if (len > MINCLSIZE) {
852 if ((mp->m_flags & M_EXT) == 0)
857 len -= M_TRAILINGSPACE(mp);
869 * m_mclget() - Adds a cluster to a normal mbuf, M_EXT is set on success.
872 m_mclget(struct mbuf *m, int how)
877 KKASSERT((m->m_flags & M_EXT_OLD) == 0);
880 if ((mcl = mclfree) == NULL) {
882 if ((mcl = mclfree) == NULL) {
883 if (how == MB_WAIT) {
891 * Possibly found a cluster, unlink it from the free list and
898 KKASSERT(mcl->mcl_refs == 0);
899 mclfree = mcl->mcl_next;
905 * Add the cluster to the mbuf.
907 m->m_ext.ext_arg = mcl;
908 m->m_ext.ext_buf = mcl->mcl_data;
909 m->m_ext.ext_nref.new = m_mclref;
910 m->m_ext.ext_nfree.new = m_mclfree;
911 m->m_ext.ext_size = MCLBYTES;
913 m->m_data = m->m_ext.ext_buf;
914 m->m_flags |= M_EXT | M_EXT_CLUSTER;
920 mbcluster_t mcl = arg;
922 KKASSERT(mcl->mcl_magic == MCL_MAGIC);
923 KKASSERT(mcl->mcl_refs > 0);
925 if (--mcl->mcl_refs == 0) {
926 if (mbstat.m_clfree < mcl_free_max) {
927 mcl->mcl_next = mclfree;
933 free(mcl->mcl_data, M_MBUFCL);
944 mbcluster_t mcl = arg;
946 KKASSERT(mcl->mcl_magic == MCL_MAGIC);
953 * Helper routines for M_EXT reference/free
956 m_extref(const struct mbuf *m)
960 KKASSERT(m->m_ext.ext_nfree.any != NULL);
962 if (m->m_flags & M_EXT_OLD)
963 m->m_ext.ext_nref.old(m->m_ext.ext_buf, m->m_ext.ext_size);
965 m->m_ext.ext_nref.new(m->m_ext.ext_arg);
972 * Free a single mbuf and any associated external storage. The successor,
973 * if any, is returned.
975 * We do need to check non-first mbuf for m_aux, since some of existing
976 * code does not call M_PREPEND properly.
977 * (example: call to bpf_mtap from drivers)
980 m_free(struct mbuf *m)
986 KASSERT(m->m_type != MT_FREE, ("freeing free mbuf %p", m));
989 * Adjust our type count and delete any attached chains if the
990 * mbuf is a packet header.
992 if ((m->m_flags & M_PKTHDR) != 0)
993 m_tag_delete_chain(m, NULL);
996 * Place the mbuf on the appropriate free list. Try to maintain a
997 * small cache of mbuf+cluster pairs.
1001 if (m->m_flags & M_EXT) {
1002 KKASSERT(m->m_ext.ext_nfree.any != NULL);
1003 if (mcl_pool_count < mcl_pool_max && m && m->m_next == NULL &&
1004 (m->m_flags & (M_PKTHDR|M_EXT_CLUSTER)) == (M_PKTHDR|M_EXT_CLUSTER) &&
1005 m->m_type == MT_DATA && M_EXT_WRITABLE(m) ) {
1006 KKASSERT(((mbcluster_t)m->m_ext.ext_arg)->mcl_magic == MCL_MAGIC);
1007 m->m_nextpkt = mcl_pool;
1012 if (m->m_flags & M_EXT_OLD)
1013 m->m_ext.ext_nfree.old(m->m_ext.ext_buf, m->m_ext.ext_size);
1015 m->m_ext.ext_nfree.new(m->m_ext.ext_arg);
1017 m->m_ext.ext_arg = NULL;
1018 m->m_ext.ext_nref.new = NULL;
1019 m->m_ext.ext_nfree.new = NULL;
1023 --mbtypes[m->m_type];
1024 if (mbtypes[MT_FREE] < mbuf_free_max) {
1025 m->m_type = MT_FREE;
1027 m->m_next = mmbfree;
1040 m_freem(struct mbuf *m)
1051 * Mbuffer utility routines.
1055 * Lesser-used path for M_PREPEND:
1056 * allocate new mbuf to prepend to chain,
1060 m_prepend(struct mbuf *m, int len, int how)
1064 MGET(mn, how, m->m_type);
1065 if (mn == (struct mbuf *)NULL) {
1067 return ((struct mbuf *)NULL);
1069 if (m->m_flags & M_PKTHDR)
1070 M_MOVE_PKTHDR(mn, m);
1080 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
1081 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
1082 * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller.
1083 * Note that the copy is read-only, because clusters are not copied,
1084 * only their reference counts are incremented.
1086 #define MCFail (mbstat.m_mcfail)
1089 m_copym(const struct mbuf *m, int off0, int len, int wait)
1091 struct mbuf *n, **np;
1096 KASSERT(off >= 0, ("m_copym, negative off %d", off));
1097 KASSERT(len >= 0, ("m_copym, negative len %d", len));
1098 if (off == 0 && m->m_flags & M_PKTHDR)
1101 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
1111 KASSERT(len == M_COPYALL,
1112 ("m_copym, length > size of mbuf chain"));
1115 MGET(n, wait, m->m_type);
1120 if (!m_dup_pkthdr(n, m, wait))
1122 if (len == M_COPYALL)
1123 n->m_pkthdr.len -= off0;
1125 n->m_pkthdr.len = len;
1128 n->m_len = min(len, m->m_len - off);
1129 if (m->m_flags & M_EXT) {
1130 n->m_data = m->m_data + off;
1132 n->m_ext = m->m_ext;
1133 n->m_flags |= m->m_flags &
1134 (M_EXT | M_EXT_OLD | M_EXT_CLUSTER);
1136 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1137 (unsigned)n->m_len);
1139 if (len != M_COPYALL)
1155 * Copy an entire packet, including header (which must be present).
1156 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
1157 * Note that the copy is read-only, because clusters are not copied,
1158 * only their reference counts are incremented.
1159 * Preserve alignment of the first mbuf so if the creator has left
1160 * some room at the beginning (e.g. for inserting protocol headers)
1161 * the copies also have the room available.
1164 m_copypacket(struct mbuf *m, int how)
1166 struct mbuf *top, *n, *o;
1168 MGET(n, how, m->m_type);
1173 if (!m_dup_pkthdr(n, m, how))
1175 n->m_len = m->m_len;
1176 if (m->m_flags & M_EXT) {
1177 n->m_data = m->m_data;
1179 n->m_ext = m->m_ext;
1180 n->m_flags |= m->m_flags & (M_EXT | M_EXT_OLD | M_EXT_CLUSTER);
1182 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
1183 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1188 MGET(o, how, m->m_type);
1195 n->m_len = m->m_len;
1196 if (m->m_flags & M_EXT) {
1197 n->m_data = m->m_data;
1199 n->m_ext = m->m_ext;
1200 n->m_flags |= m->m_flags &
1201 (M_EXT | M_EXT_OLD | M_EXT_CLUSTER);
1203 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1216 * Copy data from an mbuf chain starting "off" bytes from the beginning,
1217 * continuing for "len" bytes, into the indicated buffer.
1220 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
1224 KASSERT(off >= 0, ("m_copydata, negative off %d", off));
1225 KASSERT(len >= 0, ("m_copydata, negative len %d", len));
1227 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
1234 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
1235 count = min(m->m_len - off, len);
1236 bcopy(mtod(m, caddr_t) + off, cp, count);
1245 * Copy a packet header mbuf chain into a completely new chain, including
1246 * copying any mbuf clusters. Use this instead of m_copypacket() when
1247 * you need a writable copy of an mbuf chain.
1250 m_dup(struct mbuf *m, int how)
1252 struct mbuf **p, *top = NULL;
1253 int remain, moff, nsize;
1258 KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __FUNCTION__));
1260 /* While there's more data, get a new mbuf, tack it on, and fill it */
1261 remain = m->m_pkthdr.len;
1264 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */
1267 /* Get the next new mbuf */
1268 MGET(n, how, m->m_type);
1271 if (top == NULL) { /* first one, must be PKTHDR */
1272 if (!m_dup_pkthdr(n, m, how))
1275 } else /* not the first one */
1277 if (remain >= MINCLSIZE) {
1279 if ((n->m_flags & M_EXT) == 0) {
1287 /* Link it into the new chain */
1291 /* Copy data from original mbuf(s) into new mbuf */
1292 while (n->m_len < nsize && m != NULL) {
1293 int chunk = min(nsize - n->m_len, m->m_len - moff);
1295 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1299 if (moff == m->m_len) {
1305 /* Check correct total mbuf length */
1306 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
1307 ("%s: bogus m_pkthdr.len", __FUNCTION__));
1318 * Concatenate mbuf chain n to m.
1319 * Both chains must be of the same type (e.g. MT_DATA).
1320 * Any m_pkthdr is not updated.
1323 m_cat(struct mbuf *m, struct mbuf *n)
1328 if (m->m_flags & M_EXT ||
1329 m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) {
1330 /* just join the two chains */
1334 /* splat the data from one into the other */
1335 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1337 m->m_len += n->m_len;
1343 m_adj(struct mbuf *mp, int req_len)
1349 if ((m = mp) == NULL)
1355 while (m != NULL && len > 0) {
1356 if (m->m_len <= len) {
1367 if (mp->m_flags & M_PKTHDR)
1368 m->m_pkthdr.len -= (req_len - len);
1371 * Trim from tail. Scan the mbuf chain,
1372 * calculating its length and finding the last mbuf.
1373 * If the adjustment only affects this mbuf, then just
1374 * adjust and return. Otherwise, rescan and truncate
1375 * after the remaining size.
1381 if (m->m_next == (struct mbuf *)0)
1385 if (m->m_len >= len) {
1387 if (mp->m_flags & M_PKTHDR)
1388 mp->m_pkthdr.len -= len;
1395 * Correct length for chain is "count".
1396 * Find the mbuf with last data, adjust its length,
1397 * and toss data from remaining mbufs on chain.
1400 if (m->m_flags & M_PKTHDR)
1401 m->m_pkthdr.len = count;
1402 for (; m; m = m->m_next) {
1403 if (m->m_len >= count) {
1410 (m = m->m_next) ->m_len = 0;
1415 * Rearange an mbuf chain so that len bytes are contiguous
1416 * and in the data area of an mbuf (so that mtod will work for a structure
1417 * of size len). Returns the resulting mbuf chain on success, frees it and
1418 * returns null on failure. If there is room, it will add up to
1419 * max_protohdr-len extra bytes to the contiguous region in an attempt to
1420 * avoid being called next time.
1422 #define MPFail (mbstat.m_mpfail)
1425 m_pullup(struct mbuf *n, int len)
1432 * If first mbuf has no cluster, and has room for len bytes
1433 * without shifting current data, pullup into it,
1434 * otherwise allocate a new mbuf to prepend to the chain.
1436 if ((n->m_flags & M_EXT) == 0 &&
1437 n->m_data + len < &n->m_dat[MLEN] && n->m_next) {
1438 if (n->m_len >= len)
1446 MGET(m, MB_DONTWAIT, n->m_type);
1450 if (n->m_flags & M_PKTHDR)
1451 M_MOVE_PKTHDR(m, n);
1453 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
1455 count = min(min(max(len, max_protohdr), space), n->m_len);
1456 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1466 } while (len > 0 && n);
1480 * Partition an mbuf chain in two pieces, returning the tail --
1481 * all but the first len0 bytes. In case of failure, it returns NULL and
1482 * attempts to restore the chain to its original state.
1484 * Note that the resulting mbufs might be read-only, because the new
1485 * mbuf can end up sharing an mbuf cluster with the original mbuf if
1486 * the "breaking point" happens to lie within a cluster mbuf. Use the
1487 * M_WRITABLE() macro to check for this case.
1490 m_split(struct mbuf *m0, int len0, int wait)
1493 unsigned len = len0, remain;
1495 for (m = m0; m && len > m->m_len; m = m->m_next)
1499 remain = m->m_len - len;
1500 if (m0->m_flags & M_PKTHDR) {
1501 MGETHDR(n, wait, m0->m_type);
1504 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1505 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1506 m0->m_pkthdr.len = len0;
1507 if (m->m_flags & M_EXT)
1509 if (remain > MHLEN) {
1510 /* m can't be the lead packet */
1512 n->m_next = m_split(m, len, wait);
1513 if (n->m_next == 0) {
1521 MH_ALIGN(n, remain);
1522 } else if (remain == 0) {
1527 MGET(n, wait, m->m_type);
1533 if (m->m_flags & M_EXT) {
1534 n->m_data = m->m_data + len;
1536 n->m_ext = m->m_ext;
1537 n->m_flags |= m->m_flags & (M_EXT | M_EXT_OLD | M_EXT_CLUSTER);
1539 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1543 n->m_next = m->m_next;
1548 * Routine to copy from device local memory into mbufs.
1551 m_devget(char *buf, int totlen, int off0, struct ifnet *ifp,
1552 void (*copy) (char *from, caddr_t to, u_int len))
1555 struct mbuf *top = 0, **mp = ⊤
1556 int off = off0, len;
1563 cp += off + 2 * sizeof(u_short);
1564 totlen -= 2 * sizeof(u_short);
1566 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1569 m->m_pkthdr.rcvif = ifp;
1570 m->m_pkthdr.len = totlen;
1573 while (totlen > 0) {
1575 MGET(m, MB_DONTWAIT, MT_DATA);
1582 len = min(totlen, epkt - cp);
1583 if (len >= MINCLSIZE) {
1584 MCLGET(m, MB_DONTWAIT);
1585 if (m->m_flags & M_EXT)
1586 m->m_len = len = min(len, MCLBYTES);
1591 * Place initial small packet/header at end of mbuf.
1593 if (len < m->m_len) {
1594 if (top == 0 && len + max_linkhdr <= m->m_len)
1595 m->m_data += max_linkhdr;
1601 copy(cp, mtod(m, caddr_t), (unsigned)len);
1603 bcopy(cp, mtod(m, caddr_t), (unsigned)len);
1615 * Copy data from a buffer back into the indicated mbuf chain,
1616 * starting "off" bytes from the beginning, extending the mbuf
1617 * chain if necessary.
1620 m_copyback(struct mbuf *m0, int off, int len, caddr_t cp)
1623 struct mbuf *m = m0, *n;
1628 while (off > (mlen = m->m_len)) {
1631 if (m->m_next == 0) {
1632 n = m_getclr(MB_DONTWAIT, m->m_type);
1635 n->m_len = min(MLEN, len + off);
1641 mlen = min (m->m_len - off, len);
1642 bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen);
1650 if (m->m_next == 0) {
1651 n = m_get(MB_DONTWAIT, m->m_type);
1654 n->m_len = min(MLEN, len);
1659 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1660 m->m_pkthdr.len = totlen;
1664 m_print(const struct mbuf *m)
1667 const struct mbuf *m2;
1669 len = m->m_pkthdr.len;
1672 printf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-");
1680 * "Move" mbuf pkthdr from "from" to "to".
1681 * "from" must have M_PKTHDR set, and "to" must be empty.
1684 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
1686 KASSERT((to->m_flags & M_EXT) == 0, ("m_move_pkthdr: to has cluster"));
1688 to->m_flags = from->m_flags & M_COPYFLAGS;
1689 to->m_data = to->m_pktdat;
1690 to->m_pkthdr = from->m_pkthdr; /* especially tags */
1691 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */
1692 from->m_flags &= ~M_PKTHDR;
1696 * Duplicate "from"'s mbuf pkthdr in "to".
1697 * "from" must have M_PKTHDR set, and "to" must be empty.
1698 * In particular, this does a deep copy of the packet tags.
1701 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
1703 to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT);
1704 if ((to->m_flags & M_EXT) == 0)
1705 to->m_data = to->m_pktdat;
1706 to->m_pkthdr = from->m_pkthdr;
1707 SLIST_INIT(&to->m_pkthdr.tags);
1708 return (m_tag_copy_chain(to, from, how));
1712 * Defragment a mbuf chain, returning the shortest possible
1713 * chain of mbufs and clusters. If allocation fails and
1714 * this cannot be completed, NULL will be returned, but
1715 * the passed in chain will be unchanged. Upon success,
1716 * the original chain will be freed, and the new chain
1719 * If a non-packet header is passed in, the original
1720 * mbuf (chain?) will be returned unharmed.
1723 m_defrag(struct mbuf *m0, int how)
1725 struct mbuf *m_new = NULL, *m_final = NULL;
1726 int progress = 0, length;
1728 if (!(m0->m_flags & M_PKTHDR))
1731 #ifdef MBUF_STRESS_TEST
1732 if (m_defragrandomfailures) {
1733 int temp = arc4random() & 0xff;
1739 if (m0->m_pkthdr.len > MHLEN)
1740 m_final = m_getcl(how, MT_DATA, M_PKTHDR);
1742 m_final = m_gethdr(how, MT_DATA);
1744 if (m_final == NULL)
1747 if (m_dup_pkthdr(m_final, m0, how) == NULL)
1752 while (progress < m0->m_pkthdr.len) {
1753 length = m0->m_pkthdr.len - progress;
1754 if (length > MCLBYTES)
1757 if (m_new == NULL) {
1759 m_new = m_getcl(how, MT_DATA, 0);
1761 m_new = m_get(how, MT_DATA);
1766 m_copydata(m0, progress, length, mtod(m_new, caddr_t));
1768 m_new->m_len = length;
1769 if (m_new != m_final)
1770 m_cat(m_final, m_new);
1773 if (m0->m_next == NULL)
1778 m_defragbytes += m0->m_pkthdr.len;
1790 * Move data from uio into mbufs.
1791 * A length of zero means copy the whole uio.
1794 m_uiomove(struct uio *uio, int wait, int len0)
1796 struct mbuf *head; /* result mbuf chain */
1797 struct mbuf *m; /* current working mbuf */
1799 int resid, datalen, error;
1801 resid = (len0 == 0) ? uio->uio_resid : min(len0, uio->uio_resid);
1806 if (resid > MHLEN) {
1807 m = m_getcl(wait, MT_DATA, head == NULL ? M_PKTHDR : 0);
1810 if (m->m_flags & M_PKTHDR)
1811 m->m_pkthdr.len = 0;
1814 MGETHDR(m, wait, MT_DATA);
1817 m->m_pkthdr.len = 0;
1818 /* Leave room for protocol headers. */
1822 MGET(m, wait, MT_DATA);
1827 datalen = min(MCLBYTES, resid);
1828 error = uiomove(mtod(m, caddr_t), datalen, uio);
1836 head->m_pkthdr.len += datalen;
1838 } while (resid > 0);