4 * Copyright (c) 2004 Jeffrey M. Hsu. All rights reserved.
5 * Copyright (c) 2004 The DragonFly Project. All rights reserved.
7 * This code is derived from software contributed to The DragonFly Project
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of The DragonFly Project nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific, prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37 * Copyright (c) 1982, 1986, 1988, 1991, 1993
38 * The Regents of the University of California. All rights reserved.
40 * Redistribution and use in source and binary forms, with or without
41 * modification, are permitted provided that the following conditions
43 * 1. Redistributions of source code must retain the above copyright
44 * notice, this list of conditions and the following disclaimer.
45 * 2. Redistributions in binary form must reproduce the above copyright
46 * notice, this list of conditions and the following disclaimer in the
47 * documentation and/or other materials provided with the distribution.
48 * 3. All advertising materials mentioning features or use of this software
49 * must display the following acknowledgement:
50 * This product includes software developed by the University of
51 * California, Berkeley and its contributors.
52 * 4. Neither the name of the University nor the names of its contributors
53 * may be used to endorse or promote products derived from this software
54 * without specific prior written permission.
56 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
57 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
58 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
59 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
60 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
61 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
62 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
63 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
64 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
65 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
68 * @(#)uipc_mbuf.c 8.2 (Berkeley) 1/4/94
69 * $FreeBSD: src/sys/kern/uipc_mbuf.c,v 1.51.2.24 2003/04/15 06:59:29 silby Exp $
72 #include "opt_param.h"
73 #include "opt_mbuf_stress_test.h"
74 #include <sys/param.h>
75 #include <sys/systm.h>
77 #include <sys/malloc.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
81 #include <sys/domain.h>
82 #include <sys/objcache.h>
84 #include <sys/protosw.h>
86 #include <sys/thread.h>
87 #include <sys/globaldata.h>
89 #include <sys/thread2.h>
90 #include <sys/spinlock2.h>
92 #include <machine/atomic.h>
93 #include <machine/limits.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_extern.h>
100 #include <machine/cpu.h>
104 * mbuf cluster meta-data
112 * mbuf tracking for debugging purposes
116 static MALLOC_DEFINE(M_MTRACK, "mtrack", "mtrack");
119 RB_HEAD(mbuf_rb_tree, mbtrack);
120 RB_PROTOTYPE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *);
123 RB_ENTRY(mbtrack) rb_node;
129 mbtrack_cmp(struct mbtrack *mb1, struct mbtrack *mb2)
138 RB_GENERATE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *, m);
140 struct mbuf_rb_tree mbuf_track_root;
141 static struct spinlock mbuf_track_spin = SPINLOCK_INITIALIZER(mbuf_track_spin);
144 mbuftrack(struct mbuf *m)
148 mbt = kmalloc(sizeof(*mbt), M_MTRACK, M_INTWAIT|M_ZERO);
149 spin_lock(&mbuf_track_spin);
151 if (mbuf_rb_tree_RB_INSERT(&mbuf_track_root, mbt)) {
152 spin_unlock(&mbuf_track_spin);
153 panic("mbuftrack: mbuf %p already being tracked", m);
155 spin_unlock(&mbuf_track_spin);
159 mbufuntrack(struct mbuf *m)
163 spin_lock(&mbuf_track_spin);
164 mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m);
166 spin_unlock(&mbuf_track_spin);
167 panic("mbufuntrack: mbuf %p was not tracked", m);
169 mbuf_rb_tree_RB_REMOVE(&mbuf_track_root, mbt);
170 spin_unlock(&mbuf_track_spin);
171 kfree(mbt, M_MTRACK);
176 mbuftrackid(struct mbuf *m, int trackid)
181 spin_lock(&mbuf_track_spin);
185 mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m);
187 spin_unlock(&mbuf_track_spin);
188 panic("mbuftrackid: mbuf %p not tracked", m);
190 mbt->trackid = trackid;
195 spin_unlock(&mbuf_track_spin);
199 mbuftrack_callback(struct mbtrack *mbt, void *arg)
201 struct sysctl_req *req = arg;
205 ksnprintf(buf, sizeof(buf), "mbuf %p track %d\n", mbt->m, mbt->trackid);
207 spin_unlock(&mbuf_track_spin);
208 error = SYSCTL_OUT(req, buf, strlen(buf));
209 spin_lock(&mbuf_track_spin);
216 mbuftrack_show(SYSCTL_HANDLER_ARGS)
220 spin_lock(&mbuf_track_spin);
221 error = mbuf_rb_tree_RB_SCAN(&mbuf_track_root, NULL,
222 mbuftrack_callback, req);
223 spin_unlock(&mbuf_track_spin);
226 SYSCTL_PROC(_kern_ipc, OID_AUTO, showmbufs, CTLFLAG_RD|CTLTYPE_STRING,
227 0, 0, mbuftrack_show, "A", "Show all in-use mbufs");
232 #define mbufuntrack(m)
236 static void mbinit(void *);
237 SYSINIT(mbuf, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, mbinit, NULL)
239 static u_long mbtypes[SMP_MAXCPU][MT_NTYPES];
241 static struct mbstat mbstat[SMP_MAXCPU];
250 #ifdef MBUF_STRESS_TEST
251 int m_defragrandomfailures;
254 struct objcache *mbuf_cache, *mbufphdr_cache;
255 struct objcache *mclmeta_cache, *mjclmeta_cache;
256 struct objcache *mbufcluster_cache, *mbufphdrcluster_cache;
257 struct objcache *mbufjcluster_cache, *mbufphdrjcluster_cache;
262 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW,
263 &max_linkhdr, 0, "Max size of a link-level header");
264 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW,
265 &max_protohdr, 0, "Max size of a protocol header");
266 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0,
267 "Max size of link+protocol headers");
268 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW,
269 &max_datalen, 0, "Max data payload size without headers");
270 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW,
271 &mbuf_wait, 0, "Time in ticks to sleep after failed mbuf allocations");
272 static int do_mbstat(SYSCTL_HANDLER_ARGS);
274 SYSCTL_PROC(_kern_ipc, KIPC_MBSTAT, mbstat, CTLTYPE_STRUCT|CTLFLAG_RD,
275 0, 0, do_mbstat, "S,mbstat", "mbuf usage statistics");
277 static int do_mbtypes(SYSCTL_HANDLER_ARGS);
279 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbtypes, CTLTYPE_ULONG|CTLFLAG_RD,
280 0, 0, do_mbtypes, "LU", "");
283 do_mbstat(SYSCTL_HANDLER_ARGS)
285 struct mbstat mbstat_total;
286 struct mbstat *mbstat_totalp;
289 bzero(&mbstat_total, sizeof(mbstat_total));
290 mbstat_totalp = &mbstat_total;
292 for (i = 0; i < ncpus; i++)
294 mbstat_total.m_mbufs += mbstat[i].m_mbufs;
295 mbstat_total.m_clusters += mbstat[i].m_clusters;
296 mbstat_total.m_spare += mbstat[i].m_spare;
297 mbstat_total.m_clfree += mbstat[i].m_clfree;
298 mbstat_total.m_drops += mbstat[i].m_drops;
299 mbstat_total.m_wait += mbstat[i].m_wait;
300 mbstat_total.m_drain += mbstat[i].m_drain;
301 mbstat_total.m_mcfail += mbstat[i].m_mcfail;
302 mbstat_total.m_mpfail += mbstat[i].m_mpfail;
306 * The following fields are not cumulative fields so just
307 * get their values once.
309 mbstat_total.m_msize = mbstat[0].m_msize;
310 mbstat_total.m_mclbytes = mbstat[0].m_mclbytes;
311 mbstat_total.m_minclsize = mbstat[0].m_minclsize;
312 mbstat_total.m_mlen = mbstat[0].m_mlen;
313 mbstat_total.m_mhlen = mbstat[0].m_mhlen;
315 return(sysctl_handle_opaque(oidp, mbstat_totalp, sizeof(mbstat_total), req));
319 do_mbtypes(SYSCTL_HANDLER_ARGS)
321 u_long totals[MT_NTYPES];
324 for (i = 0; i < MT_NTYPES; i++)
327 for (i = 0; i < ncpus; i++)
329 for (j = 0; j < MT_NTYPES; j++)
330 totals[j] += mbtypes[i][j];
333 return(sysctl_handle_opaque(oidp, totals, sizeof(totals), req));
337 * These are read-only because we do not currently have any code
338 * to adjust the objcache limits after the fact. The variables
339 * may only be set as boot-time tunables.
341 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RD,
342 &nmbclusters, 0, "Maximum number of mbuf clusters available");
343 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RD, &nmbufs, 0,
344 "Maximum number of mbufs available");
346 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
347 &m_defragpackets, 0, "Number of defragment packets");
348 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
349 &m_defragbytes, 0, "Number of defragment bytes");
350 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
351 &m_defraguseless, 0, "Number of useless defragment mbuf chain operations");
352 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
353 &m_defragfailure, 0, "Number of failed defragment mbuf chain operations");
354 #ifdef MBUF_STRESS_TEST
355 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
356 &m_defragrandomfailures, 0, "");
359 static MALLOC_DEFINE(M_MBUF, "mbuf", "mbuf");
360 static MALLOC_DEFINE(M_MBUFCL, "mbufcl", "mbufcl");
361 static MALLOC_DEFINE(M_MCLMETA, "mclmeta", "mclmeta");
363 static void m_reclaim (void);
364 static void m_mclref(void *arg);
365 static void m_mclfree(void *arg);
368 * NOTE: Default NMBUFS must take into account a possible DOS attack
369 * using fd passing on unix domain sockets.
372 #define NMBCLUSTERS (512 + maxusers * 16)
375 #define NMBUFS (nmbclusters * 2 + maxfiles)
379 * Perform sanity checks of tunables declared above.
382 tunable_mbinit(void *dummy)
385 * This has to be done before VM init.
387 nmbclusters = NMBCLUSTERS;
388 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
390 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
392 if (nmbufs < nmbclusters * 2)
393 nmbufs = nmbclusters * 2;
395 SYSINIT(tunable_mbinit, SI_BOOT1_TUNABLES, SI_ORDER_ANY,
396 tunable_mbinit, NULL);
398 /* "number of clusters of pages" */
404 * The mbuf object cache only guarantees that m_next and m_nextpkt are
405 * NULL and that m_data points to the beginning of the data area. In
406 * particular, m_len and m_pkthdr.len are uninitialized. It is the
407 * responsibility of the caller to initialize those fields before use.
410 static __inline boolean_t
411 mbuf_ctor(void *obj, void *private, int ocflags)
413 struct mbuf *m = obj;
417 m->m_data = m->m_dat;
424 * Initialize the mbuf and the packet header fields.
427 mbufphdr_ctor(void *obj, void *private, int ocflags)
429 struct mbuf *m = obj;
433 m->m_data = m->m_pktdat;
434 m->m_flags = M_PKTHDR | M_PHCACHE;
436 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */
437 SLIST_INIT(&m->m_pkthdr.tags);
438 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */
439 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */
445 * A mbcluster object consists of 2K (MCLBYTES) cluster and a refcount.
448 mclmeta_ctor(void *obj, void *private, int ocflags)
450 struct mbcluster *cl = obj;
453 if (ocflags & M_NOWAIT)
454 buf = kmalloc(MCLBYTES, M_MBUFCL, M_NOWAIT | M_ZERO);
456 buf = kmalloc(MCLBYTES, M_MBUFCL, M_INTWAIT | M_ZERO);
465 mjclmeta_ctor(void *obj, void *private, int ocflags)
467 struct mbcluster *cl = obj;
470 if (ocflags & M_NOWAIT)
471 buf = kmalloc(MJUMPAGESIZE, M_MBUFCL, M_NOWAIT | M_ZERO);
473 buf = kmalloc(MJUMPAGESIZE, M_MBUFCL, M_INTWAIT | M_ZERO);
482 mclmeta_dtor(void *obj, void *private)
484 struct mbcluster *mcl = obj;
486 KKASSERT(mcl->mcl_refs == 0);
487 kfree(mcl->mcl_data, M_MBUFCL);
491 linkjcluster(struct mbuf *m, struct mbcluster *cl, uint size)
494 * Add the cluster to the mbuf. The caller will detect that the
495 * mbuf now has an attached cluster.
497 m->m_ext.ext_arg = cl;
498 m->m_ext.ext_buf = cl->mcl_data;
499 m->m_ext.ext_ref = m_mclref;
500 m->m_ext.ext_free = m_mclfree;
501 m->m_ext.ext_size = size;
502 atomic_add_int(&cl->mcl_refs, 1);
504 m->m_data = m->m_ext.ext_buf;
505 m->m_flags |= M_EXT | M_EXT_CLUSTER;
509 linkcluster(struct mbuf *m, struct mbcluster *cl)
511 linkjcluster(m, cl, MCLBYTES);
515 mbufphdrcluster_ctor(void *obj, void *private, int ocflags)
517 struct mbuf *m = obj;
518 struct mbcluster *cl;
520 mbufphdr_ctor(obj, private, ocflags);
521 cl = objcache_get(mclmeta_cache, ocflags);
523 ++mbstat[mycpu->gd_cpuid].m_drops;
526 m->m_flags |= M_CLCACHE;
532 mbufphdrjcluster_ctor(void *obj, void *private, int ocflags)
534 struct mbuf *m = obj;
535 struct mbcluster *cl;
537 mbufphdr_ctor(obj, private, ocflags);
538 cl = objcache_get(mjclmeta_cache, ocflags);
540 ++mbstat[mycpu->gd_cpuid].m_drops;
543 m->m_flags |= M_CLCACHE;
544 linkjcluster(m, cl, MJUMPAGESIZE);
549 mbufcluster_ctor(void *obj, void *private, int ocflags)
551 struct mbuf *m = obj;
552 struct mbcluster *cl;
554 mbuf_ctor(obj, private, ocflags);
555 cl = objcache_get(mclmeta_cache, ocflags);
557 ++mbstat[mycpu->gd_cpuid].m_drops;
560 m->m_flags |= M_CLCACHE;
566 mbufjcluster_ctor(void *obj, void *private, int ocflags)
568 struct mbuf *m = obj;
569 struct mbcluster *cl;
571 mbuf_ctor(obj, private, ocflags);
572 cl = objcache_get(mjclmeta_cache, ocflags);
574 ++mbstat[mycpu->gd_cpuid].m_drops;
577 m->m_flags |= M_CLCACHE;
578 linkjcluster(m, cl, MJUMPAGESIZE);
583 * Used for both the cluster and cluster PHDR caches.
585 * The mbuf may have lost its cluster due to sharing, deal
586 * with the situation by checking M_EXT.
589 mbufcluster_dtor(void *obj, void *private)
591 struct mbuf *m = obj;
592 struct mbcluster *mcl;
594 if (m->m_flags & M_EXT) {
595 KKASSERT((m->m_flags & M_EXT_CLUSTER) != 0);
596 mcl = m->m_ext.ext_arg;
597 KKASSERT(mcl->mcl_refs == 1);
599 if (m->m_flags & M_EXT && m->m_ext.ext_size != MCLBYTES)
600 objcache_put(mjclmeta_cache, mcl);
602 objcache_put(mclmeta_cache, mcl);
606 struct objcache_malloc_args mbuf_malloc_args = { MSIZE, M_MBUF };
607 struct objcache_malloc_args mclmeta_malloc_args =
608 { sizeof(struct mbcluster), M_MCLMETA };
614 int mb_limit, cl_limit;
619 * Initialize statistics
621 for (i = 0; i < ncpus; i++) {
622 mbstat[i].m_msize = MSIZE;
623 mbstat[i].m_mclbytes = MCLBYTES;
624 mbstat[i].m_mjumpagesize = MJUMPAGESIZE;
625 mbstat[i].m_minclsize = MINCLSIZE;
626 mbstat[i].m_mlen = MLEN;
627 mbstat[i].m_mhlen = MHLEN;
631 * Create objtect caches and save cluster limits, which will
632 * be used to adjust backing kmalloc pools' limit later.
635 mb_limit = cl_limit = 0;
638 mbuf_cache = objcache_create("mbuf",
640 mbuf_ctor, NULL, NULL,
641 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
645 mbufphdr_cache = objcache_create("mbuf pkt hdr",
647 mbufphdr_ctor, NULL, NULL,
648 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
651 cl_limit = nmbclusters;
652 mclmeta_cache = objcache_create("cluster mbuf",
654 mclmeta_ctor, mclmeta_dtor, NULL,
655 objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args);
657 cl_limit = nmbclusters;
658 mjclmeta_cache = objcache_create("jcluster mbuf",
660 mjclmeta_ctor, mclmeta_dtor, NULL,
661 objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args);
664 mbufcluster_cache = objcache_create("mbuf + cluster",
666 mbufcluster_ctor, mbufcluster_dtor, NULL,
667 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
671 mbufphdrcluster_cache = objcache_create("mbuf pkt hdr + cluster",
672 &limit, nmbclusters / 16,
673 mbufphdrcluster_ctor, mbufcluster_dtor, NULL,
674 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
678 mbufjcluster_cache = objcache_create("mbuf + jcluster",
680 mbufjcluster_ctor, mbufcluster_dtor, NULL,
681 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
684 mbufphdrjcluster_cache = objcache_create("mbuf pkt hdr + jcluster",
685 &limit, nmbclusters / 16,
686 mbufphdrjcluster_ctor, mbufcluster_dtor, NULL,
687 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
690 * Adjust backing kmalloc pools' limit
692 * NOTE: We raise the limit by another 1/8 to take the effect
693 * of loosememuse into account.
695 cl_limit += cl_limit / 8;
696 kmalloc_raise_limit(mclmeta_malloc_args.mtype,
697 mclmeta_malloc_args.objsize * (size_t)cl_limit);
698 kmalloc_raise_limit(M_MBUFCL,
699 ((MCLBYTES * (size_t)cl_limit * 3) / 4) +
700 ((MJUMPAGESIZE * (size_t)cl_limit) / 4));
702 mb_limit += mb_limit / 8;
703 kmalloc_raise_limit(mbuf_malloc_args.mtype,
704 mbuf_malloc_args.objsize * (size_t)mb_limit);
708 * Return the number of references to this mbuf's data. 0 is returned
709 * if the mbuf is not M_EXT, a reference count is returned if it is
710 * M_EXT | M_EXT_CLUSTER, and 99 is returned if it is a special M_EXT.
713 m_sharecount(struct mbuf *m)
715 switch (m->m_flags & (M_EXT | M_EXT_CLUSTER)) {
720 case M_EXT | M_EXT_CLUSTER:
721 return (((struct mbcluster *)m->m_ext.ext_arg)->mcl_refs);
724 return (0); /* to shut up compiler */
728 * change mbuf to new type
731 m_chtype(struct mbuf *m, int type)
733 struct globaldata *gd = mycpu;
735 ++mbtypes[gd->gd_cpuid][type];
736 --mbtypes[gd->gd_cpuid][m->m_type];
746 kprintf("Debug: m_reclaim() called\n");
748 SLIST_FOREACH(dp, &domains, dom_next) {
749 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) {
754 ++mbstat[mycpu->gd_cpuid].m_drain;
758 updatestats(struct mbuf *m, int type)
760 struct globaldata *gd = mycpu;
765 KASSERT(m->m_next == NULL, ("mbuf %p: bad m_next in get", m));
766 KASSERT(m->m_nextpkt == NULL, ("mbuf %p: bad m_nextpkt in get", m));
769 ++mbtypes[gd->gd_cpuid][type];
770 ++mbstat[gd->gd_cpuid].m_mbufs;
778 m_get(int how, int type)
782 int ocf = MBTOM(how);
786 m = objcache_get(mbuf_cache, ocf);
789 if ((how & MB_TRYWAIT) && ntries++ == 0) {
790 struct objcache *reclaimlist[] = {
793 mbufphdrcluster_cache,
795 mbufphdrjcluster_cache
797 const int nreclaims = NELEM(reclaimlist);
799 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
803 ++mbstat[mycpu->gd_cpuid].m_drops;
807 KASSERT(m->m_data == m->m_dat, ("mbuf %p: bad m_data in get", m));
811 updatestats(m, type);
816 m_gethdr(int how, int type)
819 int ocf = MBTOM(how);
824 m = objcache_get(mbufphdr_cache, ocf);
827 if ((how & MB_TRYWAIT) && ntries++ == 0) {
828 struct objcache *reclaimlist[] = {
830 mbufcluster_cache, mbufphdrcluster_cache,
831 mbufjcluster_cache, mbufphdrjcluster_cache
833 const int nreclaims = NELEM(reclaimlist);
835 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
839 ++mbstat[mycpu->gd_cpuid].m_drops;
843 KASSERT(m->m_data == m->m_pktdat, ("mbuf %p: bad m_data in get", m));
848 updatestats(m, type);
853 * Get a mbuf (not a mbuf cluster!) and zero it.
857 m_getclr(int how, int type)
861 m = m_get(how, type);
863 bzero(m->m_data, MLEN);
868 m_getjcl(int how, short type, int flags, size_t size)
870 struct mbuf *m = NULL;
871 struct objcache *mbclc, *mbphclc;
872 int ocflags = MBTOM(how);
877 mbclc = mbufcluster_cache;
878 mbphclc = mbufphdrcluster_cache;
881 mbclc = mbufjcluster_cache;
882 mbphclc = mbufphdrjcluster_cache;
888 if (flags & M_PKTHDR)
889 m = objcache_get(mbphclc, ocflags);
891 m = objcache_get(mbclc, ocflags);
894 if ((how & MB_TRYWAIT) && ntries++ == 0) {
895 struct objcache *reclaimlist[1];
897 if (flags & M_PKTHDR)
898 reclaimlist[0] = mbclc;
900 reclaimlist[0] = mbphclc;
901 if (!objcache_reclaimlist(reclaimlist, 1, ocflags))
905 ++mbstat[mycpu->gd_cpuid].m_drops;
910 KASSERT(m->m_data == m->m_ext.ext_buf,
911 ("mbuf %p: bad m_data in get", m));
915 m->m_pkthdr.len = 0; /* just do it unconditonally */
919 ++mbtypes[mycpu->gd_cpuid][type];
920 ++mbstat[mycpu->gd_cpuid].m_clusters;
925 * Returns an mbuf with an attached cluster.
926 * Because many network drivers use this kind of buffers a lot, it is
927 * convenient to keep a small pool of free buffers of this kind.
928 * Even a small size such as 10 gives about 10% improvement in the
929 * forwarding rate in a bridge or router.
932 m_getcl(int how, short type, int flags)
934 return (m_getjcl(how, type, flags, MCLBYTES));
938 * Allocate chain of requested length.
941 m_getc(int len, int how, int type)
943 struct mbuf *n, *nfirst = NULL, **ntail = &nfirst;
947 n = m_getl(len, how, type, 0, &nsize);
963 * Allocate len-worth of mbufs and/or mbuf clusters (whatever fits best)
964 * and return a pointer to the head of the allocated chain. If m0 is
965 * non-null, then we assume that it is a single mbuf or an mbuf chain to
966 * which we want len bytes worth of mbufs and/or clusters attached, and so
967 * if we succeed in allocating it, we will just return a pointer to m0.
969 * If we happen to fail at any point during the allocation, we will free
970 * up everything we have already allocated and return NULL.
972 * Deprecated. Use m_getc() and m_cat() instead.
975 m_getm(struct mbuf *m0, int len, int type, int how)
979 nfirst = m_getc(len, how, type);
982 m_last(m0)->m_next = nfirst;
990 * Adds a cluster to a normal mbuf, M_EXT is set on success.
991 * Deprecated. Use m_getcl() instead.
994 m_mclget(struct mbuf *m, int how)
996 struct mbcluster *mcl;
998 KKASSERT((m->m_flags & M_EXT) == 0);
999 mcl = objcache_get(mclmeta_cache, MBTOM(how));
1001 linkcluster(m, mcl);
1002 ++mbstat[mycpu->gd_cpuid].m_clusters;
1004 ++mbstat[mycpu->gd_cpuid].m_drops;
1009 * Updates to mbcluster must be MPSAFE. Only an entity which already has
1010 * a reference to the cluster can ref it, so we are in no danger of
1011 * racing an add with a subtract. But the operation must still be atomic
1012 * since multiple entities may have a reference on the cluster.
1014 * m_mclfree() is almost the same but it must contend with two entities
1015 * freeing the cluster at the same time.
1020 struct mbcluster *mcl = arg;
1022 atomic_add_int(&mcl->mcl_refs, 1);
1026 * When dereferencing a cluster we have to deal with a N->0 race, where
1027 * N entities free their references simultaniously. To do this we use
1028 * atomic_fetchadd_int().
1031 m_mclfree(void *arg)
1033 struct mbcluster *mcl = arg;
1035 if (atomic_fetchadd_int(&mcl->mcl_refs, -1) == 1) {
1036 --mbstat[mycpu->gd_cpuid].m_clusters;
1037 objcache_put(mclmeta_cache, mcl);
1042 * Free a single mbuf and any associated external storage. The successor,
1043 * if any, is returned.
1045 * We do need to check non-first mbuf for m_aux, since some of existing
1046 * code does not call M_PREPEND properly.
1047 * (example: call to bpf_mtap from drivers)
1053 _m_free(struct mbuf *m, const char *func)
1058 m_free(struct mbuf *m)
1063 struct globaldata *gd = mycpu;
1065 KASSERT(m->m_type != MT_FREE, ("freeing free mbuf %p", m));
1066 KASSERT(M_TRAILINGSPACE(m) >= 0, ("overflowed mbuf %p", m));
1067 --mbtypes[gd->gd_cpuid][m->m_type];
1072 * Make sure the mbuf is in constructed state before returning it
1078 m->m_hdr.mh_lastfunc = func;
1081 KKASSERT(m->m_nextpkt == NULL);
1083 if (m->m_nextpkt != NULL) {
1084 static int afewtimes = 10;
1086 if (afewtimes-- > 0) {
1087 kprintf("mfree: m->m_nextpkt != NULL\n");
1088 print_backtrace(-1);
1090 m->m_nextpkt = NULL;
1093 if (m->m_flags & M_PKTHDR) {
1094 m_tag_delete_chain(m); /* eliminate XXX JH */
1097 m->m_flags &= (M_EXT | M_EXT_CLUSTER | M_CLCACHE | M_PHCACHE);
1100 * Clean the M_PKTHDR state so we can return the mbuf to its original
1101 * cache. This is based on the PHCACHE flag which tells us whether
1102 * the mbuf was originally allocated out of a packet-header cache
1103 * or a non-packet-header cache.
1105 if (m->m_flags & M_PHCACHE) {
1106 m->m_flags |= M_PKTHDR;
1107 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */
1108 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */
1109 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */
1110 SLIST_INIT(&m->m_pkthdr.tags);
1114 * Handle remaining flags combinations. M_CLCACHE tells us whether
1115 * the mbuf was originally allocated from a cluster cache or not,
1116 * and is totally separate from whether the mbuf is currently
1117 * associated with a cluster.
1119 switch(m->m_flags & (M_CLCACHE | M_EXT | M_EXT_CLUSTER)) {
1120 case M_CLCACHE | M_EXT | M_EXT_CLUSTER:
1122 * mbuf+cluster cache case. The mbuf was allocated from the
1123 * combined mbuf_cluster cache and can be returned to the
1124 * cache if the cluster hasn't been shared.
1126 if (m_sharecount(m) == 1) {
1128 * The cluster has not been shared, we can just
1129 * reset the data pointer and return the mbuf
1130 * to the cluster cache. Note that the reference
1131 * count is left intact (it is still associated with
1134 m->m_data = m->m_ext.ext_buf;
1135 if (m->m_flags & M_EXT && m->m_ext.ext_size != MCLBYTES) {
1136 if (m->m_flags & M_PHCACHE)
1137 objcache_put(mbufphdrjcluster_cache, m);
1139 objcache_put(mbufjcluster_cache, m);
1141 if (m->m_flags & M_PHCACHE)
1142 objcache_put(mbufphdrcluster_cache, m);
1144 objcache_put(mbufcluster_cache, m);
1146 --mbstat[mycpu->gd_cpuid].m_clusters;
1149 * Hell. Someone else has a ref on this cluster,
1150 * we have to disconnect it which means we can't
1151 * put it back into the mbufcluster_cache, we
1152 * have to destroy the mbuf.
1154 * Other mbuf references to the cluster will typically
1155 * be M_EXT | M_EXT_CLUSTER but without M_CLCACHE.
1157 * XXX we could try to connect another cluster to
1160 m->m_ext.ext_free(m->m_ext.ext_arg);
1161 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
1162 if (m->m_ext.ext_size == MCLBYTES) {
1163 if (m->m_flags & M_PHCACHE)
1164 objcache_dtor(mbufphdrcluster_cache, m);
1166 objcache_dtor(mbufcluster_cache, m);
1168 if (m->m_flags & M_PHCACHE)
1169 objcache_dtor(mbufphdrjcluster_cache, m);
1171 objcache_dtor(mbufjcluster_cache, m);
1175 case M_EXT | M_EXT_CLUSTER:
1178 * Normal cluster association case, disconnect the cluster from
1179 * the mbuf. The cluster may or may not be custom.
1181 m->m_ext.ext_free(m->m_ext.ext_arg);
1182 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
1186 * return the mbuf to the mbuf cache.
1188 if (m->m_flags & M_PHCACHE) {
1189 m->m_data = m->m_pktdat;
1190 objcache_put(mbufphdr_cache, m);
1192 m->m_data = m->m_dat;
1193 objcache_put(mbuf_cache, m);
1195 --mbstat[mycpu->gd_cpuid].m_mbufs;
1199 panic("bad mbuf flags %p %08x", m, m->m_flags);
1208 _m_freem(struct mbuf *m, const char *func)
1211 m = _m_free(m, func);
1217 m_freem(struct mbuf *m)
1226 m_extadd(struct mbuf *m, caddr_t buf, u_int size, void (*reff)(void *),
1227 void (*freef)(void *), void *arg)
1229 m->m_ext.ext_arg = arg;
1230 m->m_ext.ext_buf = buf;
1231 m->m_ext.ext_ref = reff;
1232 m->m_ext.ext_free = freef;
1233 m->m_ext.ext_size = size;
1236 m->m_flags |= M_EXT;
1240 * mbuf utility routines
1244 * Lesser-used path for M_PREPEND: allocate new mbuf to prepend to chain and
1248 m_prepend(struct mbuf *m, int len, int how)
1252 if (m->m_flags & M_PKTHDR)
1253 mn = m_gethdr(how, m->m_type);
1255 mn = m_get(how, m->m_type);
1260 if (m->m_flags & M_PKTHDR)
1261 M_MOVE_PKTHDR(mn, m);
1271 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
1272 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
1273 * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller.
1274 * Note that the copy is read-only, because clusters are not copied,
1275 * only their reference counts are incremented.
1278 m_copym(const struct mbuf *m, int off0, int len, int wait)
1280 struct mbuf *n, **np;
1285 KASSERT(off >= 0, ("m_copym, negative off %d", off));
1286 KASSERT(len >= 0, ("m_copym, negative len %d", len));
1287 if (off == 0 && (m->m_flags & M_PKTHDR))
1290 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
1300 KASSERT(len == M_COPYALL,
1301 ("m_copym, length > size of mbuf chain"));
1305 * Because we are sharing any cluster attachment below,
1306 * be sure to get an mbuf that does not have a cluster
1307 * associated with it.
1310 n = m_gethdr(wait, m->m_type);
1312 n = m_get(wait, m->m_type);
1317 if (!m_dup_pkthdr(n, m, wait))
1319 if (len == M_COPYALL)
1320 n->m_pkthdr.len -= off0;
1322 n->m_pkthdr.len = len;
1325 n->m_len = min(len, m->m_len - off);
1326 if (m->m_flags & M_EXT) {
1327 KKASSERT((n->m_flags & M_EXT) == 0);
1328 n->m_data = m->m_data + off;
1329 m->m_ext.ext_ref(m->m_ext.ext_arg);
1330 n->m_ext = m->m_ext;
1331 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1333 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1334 (unsigned)n->m_len);
1336 if (len != M_COPYALL)
1343 ++mbstat[mycpu->gd_cpuid].m_mcfail;
1347 ++mbstat[mycpu->gd_cpuid].m_mcfail;
1352 * Copy an entire packet, including header (which must be present).
1353 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
1354 * Note that the copy is read-only, because clusters are not copied,
1355 * only their reference counts are incremented.
1356 * Preserve alignment of the first mbuf so if the creator has left
1357 * some room at the beginning (e.g. for inserting protocol headers)
1358 * the copies also have the room available.
1361 m_copypacket(struct mbuf *m, int how)
1363 struct mbuf *top, *n, *o;
1365 n = m_gethdr(how, m->m_type);
1370 if (!m_dup_pkthdr(n, m, how))
1372 n->m_len = m->m_len;
1373 if (m->m_flags & M_EXT) {
1374 KKASSERT((n->m_flags & M_EXT) == 0);
1375 n->m_data = m->m_data;
1376 m->m_ext.ext_ref(m->m_ext.ext_arg);
1377 n->m_ext = m->m_ext;
1378 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1380 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
1381 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1386 o = m_get(how, m->m_type);
1393 n->m_len = m->m_len;
1394 if (m->m_flags & M_EXT) {
1395 KKASSERT((n->m_flags & M_EXT) == 0);
1396 n->m_data = m->m_data;
1397 m->m_ext.ext_ref(m->m_ext.ext_arg);
1398 n->m_ext = m->m_ext;
1399 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1401 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1409 ++mbstat[mycpu->gd_cpuid].m_mcfail;
1414 * Copy data from an mbuf chain starting "off" bytes from the beginning,
1415 * continuing for "len" bytes, into the indicated buffer.
1418 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
1422 KASSERT(off >= 0, ("m_copydata, negative off %d", off));
1423 KASSERT(len >= 0, ("m_copydata, negative len %d", len));
1425 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
1432 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
1433 count = min(m->m_len - off, len);
1434 bcopy(mtod(m, caddr_t) + off, cp, count);
1443 * Copy a packet header mbuf chain into a completely new chain, including
1444 * copying any mbuf clusters. Use this instead of m_copypacket() when
1445 * you need a writable copy of an mbuf chain.
1448 m_dup(struct mbuf *m, int how)
1450 struct mbuf **p, *top = NULL;
1451 int remain, moff, nsize;
1456 KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __func__));
1458 /* While there's more data, get a new mbuf, tack it on, and fill it */
1459 remain = m->m_pkthdr.len;
1462 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */
1465 /* Get the next new mbuf */
1466 n = m_getl(remain, how, m->m_type, top == NULL ? M_PKTHDR : 0,
1471 if (!m_dup_pkthdr(n, m, how))
1474 /* Link it into the new chain */
1478 /* Copy data from original mbuf(s) into new mbuf */
1480 while (n->m_len < nsize && m != NULL) {
1481 int chunk = min(nsize - n->m_len, m->m_len - moff);
1483 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1487 if (moff == m->m_len) {
1493 /* Check correct total mbuf length */
1494 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
1495 ("%s: bogus m_pkthdr.len", __func__));
1502 ++mbstat[mycpu->gd_cpuid].m_mcfail;
1507 * Copy the non-packet mbuf data chain into a new set of mbufs, including
1508 * copying any mbuf clusters. This is typically used to realign a data
1509 * chain by nfs_realign().
1511 * The original chain is left intact. how should be MB_WAIT or MB_DONTWAIT
1512 * and NULL can be returned if MB_DONTWAIT is passed.
1514 * Be careful to use cluster mbufs, a large mbuf chain converted to non
1515 * cluster mbufs can exhaust our supply of mbufs.
1518 m_dup_data(struct mbuf *m, int how)
1520 struct mbuf **p, *n, *top = NULL;
1521 int mlen, moff, chunk, gsize, nsize;
1530 * Optimize the mbuf allocation but do not get too carried away.
1532 if (m->m_next || m->m_len > MLEN)
1533 if (m->m_flags & M_EXT && m->m_ext.ext_size == MCLBYTES)
1536 gsize = MJUMPAGESIZE;
1546 * Scan the mbuf chain until nothing is left, the new mbuf chain
1547 * will be allocated on the fly as needed.
1554 KKASSERT(m->m_type == MT_DATA);
1556 n = m_getl(gsize, how, MT_DATA, 0, &nsize);
1563 chunk = imin(mlen, nsize);
1564 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1579 ++mbstat[mycpu->gd_cpuid].m_mcfail;
1584 * Concatenate mbuf chain n to m.
1585 * Both chains must be of the same type (e.g. MT_DATA).
1586 * Any m_pkthdr is not updated.
1589 m_cat(struct mbuf *m, struct mbuf *n)
1593 if (m->m_flags & M_EXT ||
1594 m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) {
1595 /* just join the two chains */
1599 /* splat the data from one into the other */
1600 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1602 m->m_len += n->m_len;
1608 m_adj(struct mbuf *mp, int req_len)
1614 if ((m = mp) == NULL)
1620 while (m != NULL && len > 0) {
1621 if (m->m_len <= len) {
1632 if (mp->m_flags & M_PKTHDR)
1633 m->m_pkthdr.len -= (req_len - len);
1636 * Trim from tail. Scan the mbuf chain,
1637 * calculating its length and finding the last mbuf.
1638 * If the adjustment only affects this mbuf, then just
1639 * adjust and return. Otherwise, rescan and truncate
1640 * after the remaining size.
1646 if (m->m_next == NULL)
1650 if (m->m_len >= len) {
1652 if (mp->m_flags & M_PKTHDR)
1653 mp->m_pkthdr.len -= len;
1660 * Correct length for chain is "count".
1661 * Find the mbuf with last data, adjust its length,
1662 * and toss data from remaining mbufs on chain.
1665 if (m->m_flags & M_PKTHDR)
1666 m->m_pkthdr.len = count;
1667 for (; m; m = m->m_next) {
1668 if (m->m_len >= count) {
1675 (m = m->m_next) ->m_len = 0;
1680 * Set the m_data pointer of a newly-allocated mbuf
1681 * to place an object of the specified size at the
1682 * end of the mbuf, longword aligned.
1685 m_align(struct mbuf *m, int len)
1689 if (m->m_flags & M_EXT)
1690 adjust = m->m_ext.ext_size - len;
1691 else if (m->m_flags & M_PKTHDR)
1692 adjust = MHLEN - len;
1694 adjust = MLEN - len;
1695 m->m_data += adjust &~ (sizeof(long)-1);
1699 * Create a writable copy of the mbuf chain. While doing this
1700 * we compact the chain with a goal of producing a chain with
1701 * at most two mbufs. The second mbuf in this chain is likely
1702 * to be a cluster. The primary purpose of this work is to create
1703 * a writable packet for encryption, compression, etc. The
1704 * secondary goal is to linearize the data so the data can be
1705 * passed to crypto hardware in the most efficient manner possible.
1708 m_unshare(struct mbuf *m0, int how)
1710 struct mbuf *m, *mprev;
1711 struct mbuf *n, *mfirst, *mlast;
1715 for (m = m0; m != NULL; m = mprev->m_next) {
1717 * Regular mbufs are ignored unless there's a cluster
1718 * in front of it that we can use to coalesce. We do
1719 * the latter mainly so later clusters can be coalesced
1720 * also w/o having to handle them specially (i.e. convert
1721 * mbuf+cluster -> cluster). This optimization is heavily
1722 * influenced by the assumption that we're running over
1723 * Ethernet where MCLBYTES is large enough that the max
1724 * packet size will permit lots of coalescing into a
1725 * single cluster. This in turn permits efficient
1726 * crypto operations, especially when using hardware.
1728 if ((m->m_flags & M_EXT) == 0) {
1729 if (mprev && (mprev->m_flags & M_EXT) &&
1730 m->m_len <= M_TRAILINGSPACE(mprev)) {
1731 /* XXX: this ignores mbuf types */
1732 memcpy(mtod(mprev, caddr_t) + mprev->m_len,
1733 mtod(m, caddr_t), m->m_len);
1734 mprev->m_len += m->m_len;
1735 mprev->m_next = m->m_next; /* unlink from chain */
1736 m_free(m); /* reclaim mbuf */
1743 * Writable mbufs are left alone (for now).
1745 if (M_WRITABLE(m)) {
1751 * Not writable, replace with a copy or coalesce with
1752 * the previous mbuf if possible (since we have to copy
1753 * it anyway, we try to reduce the number of mbufs and
1754 * clusters so that future work is easier).
1756 KASSERT(m->m_flags & M_EXT, ("m_flags 0x%x", m->m_flags));
1757 /* NB: we only coalesce into a cluster or larger */
1758 if (mprev != NULL && (mprev->m_flags & M_EXT) &&
1759 m->m_len <= M_TRAILINGSPACE(mprev)) {
1760 /* XXX: this ignores mbuf types */
1761 memcpy(mtod(mprev, caddr_t) + mprev->m_len,
1762 mtod(m, caddr_t), m->m_len);
1763 mprev->m_len += m->m_len;
1764 mprev->m_next = m->m_next; /* unlink from chain */
1765 m_free(m); /* reclaim mbuf */
1770 * Allocate new space to hold the copy...
1772 /* XXX why can M_PKTHDR be set past the first mbuf? */
1773 if (mprev == NULL && (m->m_flags & M_PKTHDR)) {
1775 * NB: if a packet header is present we must
1776 * allocate the mbuf separately from any cluster
1777 * because M_MOVE_PKTHDR will smash the data
1778 * pointer and drop the M_EXT marker.
1780 MGETHDR(n, how, m->m_type);
1785 M_MOVE_PKTHDR(n, m);
1787 if ((n->m_flags & M_EXT) == 0) {
1793 n = m_getcl(how, m->m_type, m->m_flags);
1800 * ... and copy the data. We deal with jumbo mbufs
1801 * (i.e. m_len > MCLBYTES) by splitting them into
1802 * clusters. We could just malloc a buffer and make
1803 * it external but too many device drivers don't know
1804 * how to break up the non-contiguous memory when
1812 int cc = min(len, MCLBYTES);
1813 memcpy(mtod(n, caddr_t), mtod(m, caddr_t) + off, cc);
1824 n = m_getcl(how, m->m_type, m->m_flags);
1831 n->m_next = m->m_next;
1833 m0 = mfirst; /* new head of chain */
1835 mprev->m_next = mfirst; /* replace old mbuf */
1836 m_free(m); /* release old mbuf */
1843 * Rearrange an mbuf chain so that len bytes are contiguous
1844 * and in the data area of an mbuf (so that mtod will work for a structure
1845 * of size len). Returns the resulting mbuf chain on success, frees it and
1846 * returns null on failure. If there is room, it will add up to
1847 * max_protohdr-len extra bytes to the contiguous region in an attempt to
1848 * avoid being called next time.
1851 m_pullup(struct mbuf *n, int len)
1858 * If first mbuf has no cluster, and has room for len bytes
1859 * without shifting current data, pullup into it,
1860 * otherwise allocate a new mbuf to prepend to the chain.
1862 if (!(n->m_flags & M_EXT) &&
1863 n->m_data + len < &n->m_dat[MLEN] &&
1865 if (n->m_len >= len)
1873 if (n->m_flags & M_PKTHDR)
1874 m = m_gethdr(MB_DONTWAIT, n->m_type);
1876 m = m_get(MB_DONTWAIT, n->m_type);
1880 if (n->m_flags & M_PKTHDR)
1881 M_MOVE_PKTHDR(m, n);
1883 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
1885 count = min(min(max(len, max_protohdr), space), n->m_len);
1886 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1896 } while (len > 0 && n);
1905 ++mbstat[mycpu->gd_cpuid].m_mcfail;
1910 * Partition an mbuf chain in two pieces, returning the tail --
1911 * all but the first len0 bytes. In case of failure, it returns NULL and
1912 * attempts to restore the chain to its original state.
1914 * Note that the resulting mbufs might be read-only, because the new
1915 * mbuf can end up sharing an mbuf cluster with the original mbuf if
1916 * the "breaking point" happens to lie within a cluster mbuf. Use the
1917 * M_WRITABLE() macro to check for this case.
1920 m_split(struct mbuf *m0, int len0, int wait)
1923 unsigned len = len0, remain;
1925 for (m = m0; m && len > m->m_len; m = m->m_next)
1929 remain = m->m_len - len;
1930 if (m0->m_flags & M_PKTHDR) {
1931 n = m_gethdr(wait, m0->m_type);
1934 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1935 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1936 m0->m_pkthdr.len = len0;
1937 if (m->m_flags & M_EXT)
1939 if (remain > MHLEN) {
1940 /* m can't be the lead packet */
1942 n->m_next = m_split(m, len, wait);
1943 if (n->m_next == NULL) {
1951 MH_ALIGN(n, remain);
1952 } else if (remain == 0) {
1957 n = m_get(wait, m->m_type);
1963 if (m->m_flags & M_EXT) {
1964 KKASSERT((n->m_flags & M_EXT) == 0);
1965 n->m_data = m->m_data + len;
1966 m->m_ext.ext_ref(m->m_ext.ext_arg);
1967 n->m_ext = m->m_ext;
1968 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1970 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1974 n->m_next = m->m_next;
1980 * Routine to copy from device local memory into mbufs.
1981 * Note: "offset" is ill-defined and always called as 0, so ignore it.
1984 m_devget(char *buf, int len, int offset, struct ifnet *ifp,
1985 void (*copy)(volatile const void *from, volatile void *to, size_t length))
1987 struct mbuf *m, *mfirst = NULL, **mtail;
1996 m = m_getl(len, MB_DONTWAIT, MT_DATA, flags, &nsize);
2001 m->m_len = min(len, nsize);
2003 if (flags & M_PKTHDR) {
2004 if (len + max_linkhdr <= nsize)
2005 m->m_data += max_linkhdr;
2006 m->m_pkthdr.rcvif = ifp;
2007 m->m_pkthdr.len = len;
2011 copy(buf, m->m_data, (unsigned)m->m_len);
2022 * Routine to pad mbuf to the specified length 'padto'.
2025 m_devpad(struct mbuf *m, int padto)
2027 struct mbuf *last = NULL;
2030 if (padto <= m->m_pkthdr.len)
2033 padlen = padto - m->m_pkthdr.len;
2035 /* if there's only the packet-header and we can pad there, use it. */
2036 if (m->m_pkthdr.len == m->m_len && M_TRAILINGSPACE(m) >= padlen) {
2040 * Walk packet chain to find last mbuf. We will either
2041 * pad there, or append a new mbuf and pad it
2043 for (last = m; last->m_next != NULL; last = last->m_next)
2046 /* `last' now points to last in chain. */
2047 if (M_TRAILINGSPACE(last) < padlen) {
2050 /* Allocate new empty mbuf, pad it. Compact later. */
2051 MGET(n, MB_DONTWAIT, MT_DATA);
2059 KKASSERT(M_TRAILINGSPACE(last) >= padlen);
2060 KKASSERT(M_WRITABLE(last));
2062 /* Now zero the pad area */
2063 bzero(mtod(last, char *) + last->m_len, padlen);
2064 last->m_len += padlen;
2065 m->m_pkthdr.len += padlen;
2070 * Copy data from a buffer back into the indicated mbuf chain,
2071 * starting "off" bytes from the beginning, extending the mbuf
2072 * chain if necessary.
2075 m_copyback(struct mbuf *m0, int off, int len, caddr_t cp)
2078 struct mbuf *m = m0, *n;
2083 while (off > (mlen = m->m_len)) {
2086 if (m->m_next == NULL) {
2087 n = m_getclr(MB_DONTWAIT, m->m_type);
2090 n->m_len = min(MLEN, len + off);
2096 mlen = min (m->m_len - off, len);
2097 bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen);
2105 if (m->m_next == NULL) {
2106 n = m_get(MB_DONTWAIT, m->m_type);
2109 n->m_len = min(MLEN, len);
2114 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
2115 m->m_pkthdr.len = totlen;
2119 * Append the specified data to the indicated mbuf chain,
2120 * Extend the mbuf chain if the new data does not fit in
2123 * Return 1 if able to complete the job; otherwise 0.
2126 m_append(struct mbuf *m0, int len, c_caddr_t cp)
2129 int remainder, space;
2131 for (m = m0; m->m_next != NULL; m = m->m_next)
2134 space = M_TRAILINGSPACE(m);
2137 * Copy into available space.
2139 if (space > remainder)
2141 bcopy(cp, mtod(m, caddr_t) + m->m_len, space);
2143 cp += space, remainder -= space;
2145 while (remainder > 0) {
2147 * Allocate a new mbuf; could check space
2148 * and allocate a cluster instead.
2150 n = m_get(MB_DONTWAIT, m->m_type);
2153 n->m_len = min(MLEN, remainder);
2154 bcopy(cp, mtod(n, caddr_t), n->m_len);
2155 cp += n->m_len, remainder -= n->m_len;
2159 if (m0->m_flags & M_PKTHDR)
2160 m0->m_pkthdr.len += len - remainder;
2161 return (remainder == 0);
2165 * Apply function f to the data in an mbuf chain starting "off" bytes from
2166 * the beginning, continuing for "len" bytes.
2169 m_apply(struct mbuf *m, int off, int len,
2170 int (*f)(void *, void *, u_int), void *arg)
2175 KASSERT(off >= 0, ("m_apply, negative off %d", off));
2176 KASSERT(len >= 0, ("m_apply, negative len %d", len));
2178 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
2185 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
2186 count = min(m->m_len - off, len);
2187 rval = (*f)(arg, mtod(m, caddr_t) + off, count);
2198 * Return a pointer to mbuf/offset of location in mbuf chain.
2201 m_getptr(struct mbuf *m, int loc, int *off)
2205 /* Normal end of search. */
2206 if (m->m_len > loc) {
2211 if (m->m_next == NULL) {
2213 /* Point at the end of valid data. */
2226 m_print(const struct mbuf *m)
2229 const struct mbuf *m2;
2231 len = m->m_pkthdr.len;
2234 kprintf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-");
2242 * "Move" mbuf pkthdr from "from" to "to".
2243 * "from" must have M_PKTHDR set, and "to" must be empty.
2246 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
2248 KASSERT((to->m_flags & M_PKTHDR), ("m_move_pkthdr: not packet header"));
2250 to->m_flags |= from->m_flags & M_COPYFLAGS;
2251 to->m_pkthdr = from->m_pkthdr; /* especially tags */
2252 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */
2256 * Duplicate "from"'s mbuf pkthdr in "to".
2257 * "from" must have M_PKTHDR set, and "to" must be empty.
2258 * In particular, this does a deep copy of the packet tags.
2261 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
2263 KASSERT((to->m_flags & M_PKTHDR), ("m_dup_pkthdr: not packet header"));
2265 to->m_flags = (from->m_flags & M_COPYFLAGS) |
2266 (to->m_flags & ~M_COPYFLAGS);
2267 to->m_pkthdr = from->m_pkthdr;
2268 SLIST_INIT(&to->m_pkthdr.tags);
2269 return (m_tag_copy_chain(to, from, how));
2273 * Defragment a mbuf chain, returning the shortest possible
2274 * chain of mbufs and clusters. If allocation fails and
2275 * this cannot be completed, NULL will be returned, but
2276 * the passed in chain will be unchanged. Upon success,
2277 * the original chain will be freed, and the new chain
2280 * If a non-packet header is passed in, the original
2281 * mbuf (chain?) will be returned unharmed.
2283 * m_defrag_nofree doesn't free the passed in mbuf.
2286 m_defrag(struct mbuf *m0, int how)
2290 if ((m_new = m_defrag_nofree(m0, how)) == NULL)
2298 m_defrag_nofree(struct mbuf *m0, int how)
2300 struct mbuf *m_new = NULL, *m_final = NULL;
2301 int progress = 0, length, nsize;
2303 if (!(m0->m_flags & M_PKTHDR))
2306 #ifdef MBUF_STRESS_TEST
2307 if (m_defragrandomfailures) {
2308 int temp = karc4random() & 0xff;
2314 m_final = m_getl(m0->m_pkthdr.len, how, MT_DATA, M_PKTHDR, &nsize);
2315 if (m_final == NULL)
2317 m_final->m_len = 0; /* in case m0->m_pkthdr.len is zero */
2319 if (m_dup_pkthdr(m_final, m0, how) == 0)
2324 while (progress < m0->m_pkthdr.len) {
2325 length = m0->m_pkthdr.len - progress;
2326 if (length > MCLBYTES)
2329 if (m_new == NULL) {
2330 m_new = m_getl(length, how, MT_DATA, 0, &nsize);
2335 m_copydata(m0, progress, length, mtod(m_new, caddr_t));
2337 m_new->m_len = length;
2338 if (m_new != m_final)
2339 m_cat(m_final, m_new);
2342 if (m0->m_next == NULL)
2345 m_defragbytes += m_final->m_pkthdr.len;
2356 * Move data from uio into mbufs.
2359 m_uiomove(struct uio *uio)
2361 struct mbuf *m; /* current working mbuf */
2362 struct mbuf *head = NULL; /* result mbuf chain */
2363 struct mbuf **mp = &head;
2364 int flags = M_PKTHDR;
2370 if (uio->uio_resid > INT_MAX)
2373 resid = (int)uio->uio_resid;
2374 m = m_getl(resid, MB_WAIT, MT_DATA, flags, &nsize);
2376 m->m_pkthdr.len = 0;
2377 /* Leave room for protocol headers. */
2382 m->m_len = imin(nsize, resid);
2383 error = uiomove(mtod(m, caddr_t), m->m_len, uio);
2390 head->m_pkthdr.len += m->m_len;
2391 } while (uio->uio_resid > 0);
2401 m_last(struct mbuf *m)
2409 * Return the number of bytes in an mbuf chain.
2410 * If lastm is not NULL, also return the last mbuf.
2413 m_lengthm(struct mbuf *m, struct mbuf **lastm)
2416 struct mbuf *prev = m;
2429 * Like m_lengthm(), except also keep track of mbuf usage.
2432 m_countm(struct mbuf *m, struct mbuf **lastm, u_int *pmbcnt)
2434 u_int len = 0, mbcnt = 0;
2435 struct mbuf *prev = m;
2440 if (m->m_flags & M_EXT)
2441 mbcnt += m->m_ext.ext_size;