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 $
70 * $DragonFly: src/sys/kern/uipc_mbuf.c,v 1.70 2008/11/20 14:21:01 sephe Exp $
73 #include "opt_param.h"
74 #include "opt_mbuf_stress_test.h"
75 #include <sys/param.h>
76 #include <sys/systm.h>
78 #include <sys/malloc.h>
80 #include <sys/kernel.h>
81 #include <sys/sysctl.h>
82 #include <sys/domain.h>
83 #include <sys/objcache.h>
85 #include <sys/protosw.h>
87 #include <sys/thread.h>
88 #include <sys/globaldata.h>
90 #include <sys/thread2.h>
91 #include <sys/spinlock2.h>
93 #include <machine/atomic.h>
94 #include <machine/limits.h>
97 #include <vm/vm_kern.h>
98 #include <vm/vm_extern.h>
101 #include <machine/cpu.h>
105 * mbuf cluster meta-data
113 * mbuf tracking for debugging purposes
117 static MALLOC_DEFINE(M_MTRACK, "mtrack", "mtrack");
120 RB_HEAD(mbuf_rb_tree, mbtrack);
121 RB_PROTOTYPE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *);
124 RB_ENTRY(mbtrack) rb_node;
130 mbtrack_cmp(struct mbtrack *mb1, struct mbtrack *mb2)
139 RB_GENERATE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *, m);
141 struct mbuf_rb_tree mbuf_track_root;
142 static struct spinlock mbuf_track_spin = SPINLOCK_INITIALIZER(mbuf_track_spin);
145 mbuftrack(struct mbuf *m)
149 mbt = kmalloc(sizeof(*mbt), M_MTRACK, M_INTWAIT|M_ZERO);
150 spin_lock(&mbuf_track_spin);
152 if (mbuf_rb_tree_RB_INSERT(&mbuf_track_root, mbt)) {
153 spin_unlock(&mbuf_track_spin);
154 panic("mbuftrack: mbuf %p already being tracked\n", m);
156 spin_unlock(&mbuf_track_spin);
160 mbufuntrack(struct mbuf *m)
164 spin_lock(&mbuf_track_spin);
165 mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m);
167 spin_unlock(&mbuf_track_spin);
168 panic("mbufuntrack: mbuf %p was not tracked\n", m);
170 mbuf_rb_tree_RB_REMOVE(&mbuf_track_root, mbt);
171 spin_unlock(&mbuf_track_spin);
172 kfree(mbt, M_MTRACK);
177 mbuftrackid(struct mbuf *m, int trackid)
182 spin_lock(&mbuf_track_spin);
186 mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m);
188 spin_unlock(&mbuf_track_spin);
189 panic("mbuftrackid: mbuf %p not tracked", m);
191 mbt->trackid = trackid;
196 spin_unlock(&mbuf_track_spin);
200 mbuftrack_callback(struct mbtrack *mbt, void *arg)
202 struct sysctl_req *req = arg;
206 ksnprintf(buf, sizeof(buf), "mbuf %p track %d\n", mbt->m, mbt->trackid);
208 spin_unlock(&mbuf_track_spin);
209 error = SYSCTL_OUT(req, buf, strlen(buf));
210 spin_lock(&mbuf_track_spin);
217 mbuftrack_show(SYSCTL_HANDLER_ARGS)
221 spin_lock(&mbuf_track_spin);
222 error = mbuf_rb_tree_RB_SCAN(&mbuf_track_root, NULL,
223 mbuftrack_callback, req);
224 spin_unlock(&mbuf_track_spin);
227 SYSCTL_PROC(_kern_ipc, OID_AUTO, showmbufs, CTLFLAG_RD|CTLTYPE_STRING,
228 0, 0, mbuftrack_show, "A", "Show all in-use mbufs");
233 #define mbufuntrack(m)
237 static void mbinit(void *);
238 SYSINIT(mbuf, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, mbinit, NULL)
240 static u_long mbtypes[SMP_MAXCPU][MT_NTYPES];
242 static struct mbstat mbstat[SMP_MAXCPU];
251 #ifdef MBUF_STRESS_TEST
252 int m_defragrandomfailures;
255 struct objcache *mbuf_cache, *mbufphdr_cache;
256 struct objcache *mclmeta_cache, *mjclmeta_cache;
257 struct objcache *mbufcluster_cache, *mbufphdrcluster_cache;
258 struct objcache *mbufjcluster_cache, *mbufphdrjcluster_cache;
263 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW,
264 &max_linkhdr, 0, "Max size of a link-level header");
265 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW,
266 &max_protohdr, 0, "Max size of a protocol header");
267 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0,
268 "Max size of link+protocol headers");
269 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW,
270 &max_datalen, 0, "Max data payload size without headers");
271 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW,
272 &mbuf_wait, 0, "Time in ticks to sleep after failed mbuf allocations");
273 static int do_mbstat(SYSCTL_HANDLER_ARGS);
275 SYSCTL_PROC(_kern_ipc, KIPC_MBSTAT, mbstat, CTLTYPE_STRUCT|CTLFLAG_RD,
276 0, 0, do_mbstat, "S,mbstat", "mbuf usage statistics");
278 static int do_mbtypes(SYSCTL_HANDLER_ARGS);
280 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbtypes, CTLTYPE_ULONG|CTLFLAG_RD,
281 0, 0, do_mbtypes, "LU", "");
284 do_mbstat(SYSCTL_HANDLER_ARGS)
286 struct mbstat mbstat_total;
287 struct mbstat *mbstat_totalp;
290 bzero(&mbstat_total, sizeof(mbstat_total));
291 mbstat_totalp = &mbstat_total;
293 for (i = 0; i < ncpus; i++)
295 mbstat_total.m_mbufs += mbstat[i].m_mbufs;
296 mbstat_total.m_clusters += mbstat[i].m_clusters;
297 mbstat_total.m_spare += mbstat[i].m_spare;
298 mbstat_total.m_clfree += mbstat[i].m_clfree;
299 mbstat_total.m_drops += mbstat[i].m_drops;
300 mbstat_total.m_wait += mbstat[i].m_wait;
301 mbstat_total.m_drain += mbstat[i].m_drain;
302 mbstat_total.m_mcfail += mbstat[i].m_mcfail;
303 mbstat_total.m_mpfail += mbstat[i].m_mpfail;
307 * The following fields are not cumulative fields so just
308 * get their values once.
310 mbstat_total.m_msize = mbstat[0].m_msize;
311 mbstat_total.m_mclbytes = mbstat[0].m_mclbytes;
312 mbstat_total.m_minclsize = mbstat[0].m_minclsize;
313 mbstat_total.m_mlen = mbstat[0].m_mlen;
314 mbstat_total.m_mhlen = mbstat[0].m_mhlen;
316 return(sysctl_handle_opaque(oidp, mbstat_totalp, sizeof(mbstat_total), req));
320 do_mbtypes(SYSCTL_HANDLER_ARGS)
322 u_long totals[MT_NTYPES];
325 for (i = 0; i < MT_NTYPES; i++)
328 for (i = 0; i < ncpus; i++)
330 for (j = 0; j < MT_NTYPES; j++)
331 totals[j] += mbtypes[i][j];
334 return(sysctl_handle_opaque(oidp, totals, sizeof(totals), req));
338 * These are read-only because we do not currently have any code
339 * to adjust the objcache limits after the fact. The variables
340 * may only be set as boot-time tunables.
342 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RD,
343 &nmbclusters, 0, "Maximum number of mbuf clusters available");
344 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RD, &nmbufs, 0,
345 "Maximum number of mbufs available");
347 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
348 &m_defragpackets, 0, "Number of defragment packets");
349 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
350 &m_defragbytes, 0, "Number of defragment bytes");
351 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
352 &m_defraguseless, 0, "Number of useless defragment mbuf chain operations");
353 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
354 &m_defragfailure, 0, "Number of failed defragment mbuf chain operations");
355 #ifdef MBUF_STRESS_TEST
356 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
357 &m_defragrandomfailures, 0, "");
360 static MALLOC_DEFINE(M_MBUF, "mbuf", "mbuf");
361 static MALLOC_DEFINE(M_MBUFCL, "mbufcl", "mbufcl");
362 static MALLOC_DEFINE(M_MJBUFCL, "mbufcl", "mbufcl");
363 static MALLOC_DEFINE(M_MCLMETA, "mclmeta", "mclmeta");
364 static MALLOC_DEFINE(M_MJCLMETA, "mjclmeta", "mjclmeta");
366 static void m_reclaim (void);
367 static void m_mclref(void *arg);
368 static void m_mclfree(void *arg);
371 * NOTE: Default NMBUFS must take into account a possible DOS attack
372 * using fd passing on unix domain sockets.
375 #define NMBCLUSTERS (512 + maxusers * 16)
378 #define NMBUFS (nmbclusters * 2 + maxfiles)
382 * Perform sanity checks of tunables declared above.
385 tunable_mbinit(void *dummy)
388 * This has to be done before VM init.
390 nmbclusters = NMBCLUSTERS;
391 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
393 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
395 if (nmbufs < nmbclusters * 2)
396 nmbufs = nmbclusters * 2;
398 SYSINIT(tunable_mbinit, SI_BOOT1_TUNABLES, SI_ORDER_ANY,
399 tunable_mbinit, NULL);
401 /* "number of clusters of pages" */
407 * The mbuf object cache only guarantees that m_next and m_nextpkt are
408 * NULL and that m_data points to the beginning of the data area. In
409 * particular, m_len and m_pkthdr.len are uninitialized. It is the
410 * responsibility of the caller to initialize those fields before use.
413 static boolean_t __inline
414 mbuf_ctor(void *obj, void *private, int ocflags)
416 struct mbuf *m = obj;
420 m->m_data = m->m_dat;
427 * Initialize the mbuf and the packet header fields.
430 mbufphdr_ctor(void *obj, void *private, int ocflags)
432 struct mbuf *m = obj;
436 m->m_data = m->m_pktdat;
437 m->m_flags = M_PKTHDR | M_PHCACHE;
439 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */
440 SLIST_INIT(&m->m_pkthdr.tags);
441 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */
442 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */
448 * A mbcluster object consists of 2K (MCLBYTES) cluster and a refcount.
451 mclmeta_ctor(void *obj, void *private, int ocflags)
453 struct mbcluster *cl = obj;
456 if (ocflags & M_NOWAIT)
457 buf = kmalloc(MCLBYTES, M_MBUFCL, M_NOWAIT | M_ZERO);
459 buf = kmalloc(MCLBYTES, M_MBUFCL, M_INTWAIT | M_ZERO);
468 mjclmeta_ctor(void *obj, void *private, int ocflags)
470 struct mbcluster *cl = obj;
473 if (ocflags & M_NOWAIT)
474 buf = kmalloc(MJUMPAGESIZE, M_MBUFCL, M_NOWAIT | M_ZERO);
476 buf = kmalloc(MJUMPAGESIZE, M_MBUFCL, M_INTWAIT | M_ZERO);
485 mclmeta_dtor(void *obj, void *private)
487 struct mbcluster *mcl = obj;
489 KKASSERT(mcl->mcl_refs == 0);
490 kfree(mcl->mcl_data, M_MBUFCL);
494 linkjcluster(struct mbuf *m, struct mbcluster *cl, uint size)
497 * Add the cluster to the mbuf. The caller will detect that the
498 * mbuf now has an attached cluster.
500 m->m_ext.ext_arg = cl;
501 m->m_ext.ext_buf = cl->mcl_data;
502 m->m_ext.ext_ref = m_mclref;
503 m->m_ext.ext_free = m_mclfree;
504 m->m_ext.ext_size = size;
505 atomic_add_int(&cl->mcl_refs, 1);
507 m->m_data = m->m_ext.ext_buf;
508 m->m_flags |= M_EXT | M_EXT_CLUSTER;
512 linkcluster(struct mbuf *m, struct mbcluster *cl)
514 linkjcluster(m, cl, MCLBYTES);
518 mbufphdrcluster_ctor(void *obj, void *private, int ocflags)
520 struct mbuf *m = obj;
521 struct mbcluster *cl;
523 mbufphdr_ctor(obj, private, ocflags);
524 cl = objcache_get(mclmeta_cache, ocflags);
526 ++mbstat[mycpu->gd_cpuid].m_drops;
529 m->m_flags |= M_CLCACHE;
535 mbufphdrjcluster_ctor(void *obj, void *private, int ocflags)
537 struct mbuf *m = obj;
538 struct mbcluster *cl;
540 mbufphdr_ctor(obj, private, ocflags);
541 cl = objcache_get(mjclmeta_cache, ocflags);
543 ++mbstat[mycpu->gd_cpuid].m_drops;
546 m->m_flags |= M_CLCACHE;
547 linkjcluster(m, cl, MJUMPAGESIZE);
552 mbufcluster_ctor(void *obj, void *private, int ocflags)
554 struct mbuf *m = obj;
555 struct mbcluster *cl;
557 mbuf_ctor(obj, private, ocflags);
558 cl = objcache_get(mclmeta_cache, ocflags);
560 ++mbstat[mycpu->gd_cpuid].m_drops;
563 m->m_flags |= M_CLCACHE;
569 mbufjcluster_ctor(void *obj, void *private, int ocflags)
571 struct mbuf *m = obj;
572 struct mbcluster *cl;
574 mbuf_ctor(obj, private, ocflags);
575 cl = objcache_get(mjclmeta_cache, ocflags);
577 ++mbstat[mycpu->gd_cpuid].m_drops;
580 m->m_flags |= M_CLCACHE;
581 linkjcluster(m, cl, MJUMPAGESIZE);
586 * Used for both the cluster and cluster PHDR caches.
588 * The mbuf may have lost its cluster due to sharing, deal
589 * with the situation by checking M_EXT.
592 mbufcluster_dtor(void *obj, void *private)
594 struct mbuf *m = obj;
595 struct mbcluster *mcl;
597 if (m->m_flags & M_EXT) {
598 KKASSERT((m->m_flags & M_EXT_CLUSTER) != 0);
599 mcl = m->m_ext.ext_arg;
600 KKASSERT(mcl->mcl_refs == 1);
602 if (m->m_flags & M_EXT && m->m_ext.ext_size != MCLBYTES)
603 objcache_put(mjclmeta_cache, mcl);
605 objcache_put(mclmeta_cache, mcl);
609 struct objcache_malloc_args mbuf_malloc_args = { MSIZE, M_MBUF };
610 struct objcache_malloc_args mclmeta_malloc_args =
611 { sizeof(struct mbcluster), M_MCLMETA };
617 int mb_limit, cl_limit;
622 * Initialize statistics
624 for (i = 0; i < ncpus; i++) {
625 atomic_set_long_nonlocked(&mbstat[i].m_msize, MSIZE);
626 atomic_set_long_nonlocked(&mbstat[i].m_mclbytes, MCLBYTES);
627 atomic_set_long_nonlocked(&mbstat[i].m_mjumpagesize, MJUMPAGESIZE);
628 atomic_set_long_nonlocked(&mbstat[i].m_minclsize, MINCLSIZE);
629 atomic_set_long_nonlocked(&mbstat[i].m_mlen, MLEN);
630 atomic_set_long_nonlocked(&mbstat[i].m_mhlen, MHLEN);
634 * Create objtect caches and save cluster limits, which will
635 * be used to adjust backing kmalloc pools' limit later.
638 mb_limit = cl_limit = 0;
641 mbuf_cache = objcache_create("mbuf", &limit, 0,
642 mbuf_ctor, NULL, NULL,
643 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
647 mbufphdr_cache = objcache_create("mbuf pkt hdr", &limit, 128,
648 mbufphdr_ctor, NULL, NULL,
649 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
652 cl_limit = nmbclusters;
653 mclmeta_cache = objcache_create("cluster mbuf", &cl_limit, 0,
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", &cl_limit, 0,
659 mjclmeta_ctor, mclmeta_dtor, NULL,
660 objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args);
663 mbufcluster_cache = objcache_create("mbuf + cluster", &limit, 0,
664 mbufcluster_ctor, mbufcluster_dtor, NULL,
665 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
669 mbufphdrcluster_cache = objcache_create("mbuf pkt hdr + cluster",
670 &limit, 128, mbufphdrcluster_ctor, mbufcluster_dtor, NULL,
671 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
675 mbufjcluster_cache = objcache_create("mbuf + jcluster", &limit, 0,
676 mbufjcluster_ctor, mbufcluster_dtor, NULL,
677 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
681 mbufphdrjcluster_cache = objcache_create("mbuf pkt hdr + jcluster",
682 &limit, 64, mbufphdrjcluster_ctor, mbufcluster_dtor, NULL,
683 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args);
687 * Adjust backing kmalloc pools' limit
689 * NOTE: We raise the limit by another 1/8 to take the effect
690 * of loosememuse into account.
692 cl_limit += cl_limit / 8;
693 kmalloc_raise_limit(mclmeta_malloc_args.mtype,
694 mclmeta_malloc_args.objsize * cl_limit);
695 kmalloc_raise_limit(M_MBUFCL, MCLBYTES * cl_limit * 3/4 + MJUMPAGESIZE * cl_limit / 4);
696 /*kmalloc_raise_limit(M_MBUFCL, MCLBYTES * cl_limit);*/
698 mb_limit += mb_limit / 8;
699 kmalloc_raise_limit(mbuf_malloc_args.mtype,
700 mbuf_malloc_args.objsize * mb_limit);
704 * Return the number of references to this mbuf's data. 0 is returned
705 * if the mbuf is not M_EXT, a reference count is returned if it is
706 * M_EXT | M_EXT_CLUSTER, and 99 is returned if it is a special M_EXT.
709 m_sharecount(struct mbuf *m)
711 switch (m->m_flags & (M_EXT | M_EXT_CLUSTER)) {
716 case M_EXT | M_EXT_CLUSTER:
717 return (((struct mbcluster *)m->m_ext.ext_arg)->mcl_refs);
720 return (0); /* to shut up compiler */
724 * change mbuf to new type
727 m_chtype(struct mbuf *m, int type)
729 struct globaldata *gd = mycpu;
731 atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1);
732 atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1);
733 atomic_set_short_nonlocked(&m->m_type, type);
742 kprintf("Debug: m_reclaim() called\n");
744 SLIST_FOREACH(dp, &domains, dom_next) {
745 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) {
750 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_drain, 1);
754 updatestats(struct mbuf *m, int type)
756 struct globaldata *gd = mycpu;
761 KASSERT(m->m_next == NULL, ("mbuf %p: bad m_next in get", m));
762 KASSERT(m->m_nextpkt == NULL, ("mbuf %p: bad m_nextpkt in get", m));
765 atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1);
766 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1);
774 m_get(int how, int type)
778 int ocf = MBTOM(how);
782 m = objcache_get(mbuf_cache, ocf);
785 if ((how & MB_TRYWAIT) && ntries++ == 0) {
786 struct objcache *reclaimlist[] = {
789 mbufphdrcluster_cache,
791 mbufphdrjcluster_cache
793 const int nreclaims = NELEM(reclaimlist);
795 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
799 ++mbstat[mycpu->gd_cpuid].m_drops;
803 KASSERT(m->m_data == m->m_dat, ("mbuf %p: bad m_data in get", m));
807 updatestats(m, type);
812 m_gethdr(int how, int type)
815 int ocf = MBTOM(how);
820 m = objcache_get(mbufphdr_cache, ocf);
823 if ((how & MB_TRYWAIT) && ntries++ == 0) {
824 struct objcache *reclaimlist[] = {
826 mbufcluster_cache, mbufphdrcluster_cache,
827 mbufjcluster_cache, mbufphdrjcluster_cache
829 const int nreclaims = NELEM(reclaimlist);
831 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf))
835 ++mbstat[mycpu->gd_cpuid].m_drops;
839 KASSERT(m->m_data == m->m_pktdat, ("mbuf %p: bad m_data in get", m));
844 updatestats(m, type);
849 * Get a mbuf (not a mbuf cluster!) and zero it.
853 m_getclr(int how, int type)
857 m = m_get(how, type);
859 bzero(m->m_data, MLEN);
864 m_getjcl(int how, short type, int flags, size_t size)
866 struct mbuf *m = NULL;
867 int ocflags = MBTOM(how);
872 if (flags & M_PKTHDR)
873 m = objcache_get(mbufphdrjcluster_cache, ocflags);
875 m = objcache_get(mbufjcluster_cache, ocflags);
878 if ((how & MB_TRYWAIT) && ntries++ == 0) {
879 struct objcache *reclaimlist[1];
881 if (flags & M_PKTHDR)
882 reclaimlist[0] = mbufjcluster_cache;
884 reclaimlist[0] = mbufphdrjcluster_cache;
885 if (!objcache_reclaimlist(reclaimlist, 1, ocflags))
889 ++mbstat[mycpu->gd_cpuid].m_drops;
894 KASSERT(m->m_data == m->m_ext.ext_buf,
895 ("mbuf %p: bad m_data in get", m));
899 m->m_pkthdr.len = 0; /* just do it unconditonally */
903 atomic_add_long_nonlocked(&mbtypes[mycpu->gd_cpuid][type], 1);
904 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
909 * Returns an mbuf with an attached cluster.
910 * Because many network drivers use this kind of buffers a lot, it is
911 * convenient to keep a small pool of free buffers of this kind.
912 * Even a small size such as 10 gives about 10% improvement in the
913 * forwarding rate in a bridge or router.
916 m_getcl(int how, short type, int flags)
919 int ocflags = MBTOM(how);
924 if (flags & M_PKTHDR)
925 m = objcache_get(mbufphdrcluster_cache, ocflags);
927 m = objcache_get(mbufcluster_cache, ocflags);
930 if ((how & MB_TRYWAIT) && ntries++ == 0) {
931 struct objcache *reclaimlist[1];
933 if (flags & M_PKTHDR)
934 reclaimlist[0] = mbufcluster_cache;
936 reclaimlist[0] = mbufphdrcluster_cache;
937 if (!objcache_reclaimlist(reclaimlist, 1, ocflags))
941 ++mbstat[mycpu->gd_cpuid].m_drops;
946 KASSERT(m->m_data == m->m_ext.ext_buf,
947 ("mbuf %p: bad m_data in get", m));
951 m->m_pkthdr.len = 0; /* just do it unconditonally */
955 atomic_add_long_nonlocked(&mbtypes[mycpu->gd_cpuid][type], 1);
956 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
961 * Allocate chain of requested length.
964 m_getc(int len, int how, int type)
966 struct mbuf *n, *nfirst = NULL, **ntail = &nfirst;
970 n = m_getl(len, how, type, 0, &nsize);
986 * Allocate len-worth of mbufs and/or mbuf clusters (whatever fits best)
987 * and return a pointer to the head of the allocated chain. If m0 is
988 * non-null, then we assume that it is a single mbuf or an mbuf chain to
989 * which we want len bytes worth of mbufs and/or clusters attached, and so
990 * if we succeed in allocating it, we will just return a pointer to m0.
992 * If we happen to fail at any point during the allocation, we will free
993 * up everything we have already allocated and return NULL.
995 * Deprecated. Use m_getc() and m_cat() instead.
998 m_getm(struct mbuf *m0, int len, int type, int how)
1000 struct mbuf *nfirst;
1002 nfirst = m_getc(len, how, type);
1005 m_last(m0)->m_next = nfirst;
1013 * Adds a cluster to a normal mbuf, M_EXT is set on success.
1014 * Deprecated. Use m_getcl() instead.
1017 m_mclget(struct mbuf *m, int how)
1019 struct mbcluster *mcl;
1021 KKASSERT((m->m_flags & M_EXT) == 0);
1022 mcl = objcache_get(mclmeta_cache, MBTOM(how));
1024 linkcluster(m, mcl);
1025 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters,
1028 ++mbstat[mycpu->gd_cpuid].m_drops;
1033 * Updates to mbcluster must be MPSAFE. Only an entity which already has
1034 * a reference to the cluster can ref it, so we are in no danger of
1035 * racing an add with a subtract. But the operation must still be atomic
1036 * since multiple entities may have a reference on the cluster.
1038 * m_mclfree() is almost the same but it must contend with two entities
1039 * freeing the cluster at the same time.
1044 struct mbcluster *mcl = arg;
1046 atomic_add_int(&mcl->mcl_refs, 1);
1050 * When dereferencing a cluster we have to deal with a N->0 race, where
1051 * N entities free their references simultaniously. To do this we use
1052 * atomic_fetchadd_int().
1055 m_mclfree(void *arg)
1057 struct mbcluster *mcl = arg;
1059 if (atomic_fetchadd_int(&mcl->mcl_refs, -1) == 1)
1060 objcache_put(mclmeta_cache, mcl);
1064 * Free a single mbuf and any associated external storage. The successor,
1065 * if any, is returned.
1067 * We do need to check non-first mbuf for m_aux, since some of existing
1068 * code does not call M_PREPEND properly.
1069 * (example: call to bpf_mtap from drivers)
1075 _m_free(struct mbuf *m, const char *func)
1080 m_free(struct mbuf *m)
1085 struct globaldata *gd = mycpu;
1087 KASSERT(m->m_type != MT_FREE, ("freeing free mbuf %p", m));
1088 KASSERT(M_TRAILINGSPACE(m) >= 0, ("overflowed mbuf %p", m));
1089 atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1);
1094 * Make sure the mbuf is in constructed state before returning it
1100 m->m_hdr.mh_lastfunc = func;
1103 KKASSERT(m->m_nextpkt == NULL);
1105 if (m->m_nextpkt != NULL) {
1106 static int afewtimes = 10;
1108 if (afewtimes-- > 0) {
1109 kprintf("mfree: m->m_nextpkt != NULL\n");
1110 print_backtrace(-1);
1112 m->m_nextpkt = NULL;
1115 if (m->m_flags & M_PKTHDR) {
1116 m_tag_delete_chain(m); /* eliminate XXX JH */
1119 m->m_flags &= (M_EXT | M_EXT_CLUSTER | M_CLCACHE | M_PHCACHE);
1122 * Clean the M_PKTHDR state so we can return the mbuf to its original
1123 * cache. This is based on the PHCACHE flag which tells us whether
1124 * the mbuf was originally allocated out of a packet-header cache
1125 * or a non-packet-header cache.
1127 if (m->m_flags & M_PHCACHE) {
1128 m->m_flags |= M_PKTHDR;
1129 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */
1130 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */
1131 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */
1132 SLIST_INIT(&m->m_pkthdr.tags);
1136 * Handle remaining flags combinations. M_CLCACHE tells us whether
1137 * the mbuf was originally allocated from a cluster cache or not,
1138 * and is totally separate from whether the mbuf is currently
1139 * associated with a cluster.
1141 switch(m->m_flags & (M_CLCACHE | M_EXT | M_EXT_CLUSTER)) {
1142 case M_CLCACHE | M_EXT | M_EXT_CLUSTER:
1144 * mbuf+cluster cache case. The mbuf was allocated from the
1145 * combined mbuf_cluster cache and can be returned to the
1146 * cache if the cluster hasn't been shared.
1148 if (m_sharecount(m) == 1) {
1150 * The cluster has not been shared, we can just
1151 * reset the data pointer and return the mbuf
1152 * to the cluster cache. Note that the reference
1153 * count is left intact (it is still associated with
1156 m->m_data = m->m_ext.ext_buf;
1157 if (m->m_flags & M_EXT && m->m_ext.ext_size != MCLBYTES) {
1158 if (m->m_flags & M_PHCACHE)
1159 objcache_put(mbufphdrjcluster_cache, m);
1161 objcache_put(mbufjcluster_cache, m);
1163 if (m->m_flags & M_PHCACHE)
1164 objcache_put(mbufphdrcluster_cache, m);
1166 objcache_put(mbufcluster_cache, m);
1168 atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
1171 * Hell. Someone else has a ref on this cluster,
1172 * we have to disconnect it which means we can't
1173 * put it back into the mbufcluster_cache, we
1174 * have to destroy the mbuf.
1176 * Other mbuf references to the cluster will typically
1177 * be M_EXT | M_EXT_CLUSTER but without M_CLCACHE.
1179 * XXX we could try to connect another cluster to
1183 m->m_ext.ext_free(m->m_ext.ext_arg);
1184 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
1185 if (m->m_ext.ext_size == MCLBYTES) {
1186 if (m->m_flags & M_PHCACHE)
1187 objcache_dtor(mbufphdrcluster_cache, m);
1189 objcache_dtor(mbufcluster_cache, m);
1191 if (m->m_flags & M_PHCACHE)
1192 objcache_dtor(mbufphdrjcluster_cache, m);
1194 objcache_dtor(mbufjcluster_cache, m);
1198 case M_EXT | M_EXT_CLUSTER:
1200 * Normal cluster associated with an mbuf that was allocated
1201 * from the normal mbuf pool rather then the cluster pool.
1202 * The cluster has to be independantly disassociated from the
1205 if (m_sharecount(m) == 1)
1206 atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1);
1210 * Normal cluster association case, disconnect the cluster from
1211 * the mbuf. The cluster may or may not be custom.
1213 m->m_ext.ext_free(m->m_ext.ext_arg);
1214 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER);
1218 * return the mbuf to the mbuf cache.
1220 if (m->m_flags & M_PHCACHE) {
1221 m->m_data = m->m_pktdat;
1222 objcache_put(mbufphdr_cache, m);
1224 m->m_data = m->m_dat;
1225 objcache_put(mbuf_cache, m);
1227 atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1);
1231 panic("bad mbuf flags %p %08x\n", m, m->m_flags);
1240 _m_freem(struct mbuf *m, const char *func)
1243 m = _m_free(m, func);
1249 m_freem(struct mbuf *m)
1258 * mbuf utility routines
1262 * Lesser-used path for M_PREPEND: allocate new mbuf to prepend to chain and
1266 m_prepend(struct mbuf *m, int len, int how)
1270 if (m->m_flags & M_PKTHDR)
1271 mn = m_gethdr(how, m->m_type);
1273 mn = m_get(how, m->m_type);
1278 if (m->m_flags & M_PKTHDR)
1279 M_MOVE_PKTHDR(mn, m);
1289 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
1290 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
1291 * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller.
1292 * Note that the copy is read-only, because clusters are not copied,
1293 * only their reference counts are incremented.
1296 m_copym(const struct mbuf *m, int off0, int len, int wait)
1298 struct mbuf *n, **np;
1303 KASSERT(off >= 0, ("m_copym, negative off %d", off));
1304 KASSERT(len >= 0, ("m_copym, negative len %d", len));
1305 if (off == 0 && (m->m_flags & M_PKTHDR))
1308 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
1318 KASSERT(len == M_COPYALL,
1319 ("m_copym, length > size of mbuf chain"));
1323 * Because we are sharing any cluster attachment below,
1324 * be sure to get an mbuf that does not have a cluster
1325 * associated with it.
1328 n = m_gethdr(wait, m->m_type);
1330 n = m_get(wait, m->m_type);
1335 if (!m_dup_pkthdr(n, m, wait))
1337 if (len == M_COPYALL)
1338 n->m_pkthdr.len -= off0;
1340 n->m_pkthdr.len = len;
1343 n->m_len = min(len, m->m_len - off);
1344 if (m->m_flags & M_EXT) {
1345 KKASSERT((n->m_flags & M_EXT) == 0);
1346 n->m_data = m->m_data + off;
1347 m->m_ext.ext_ref(m->m_ext.ext_arg);
1348 n->m_ext = m->m_ext;
1349 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1351 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1352 (unsigned)n->m_len);
1354 if (len != M_COPYALL)
1361 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1365 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1370 * Copy an entire packet, including header (which must be present).
1371 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
1372 * Note that the copy is read-only, because clusters are not copied,
1373 * only their reference counts are incremented.
1374 * Preserve alignment of the first mbuf so if the creator has left
1375 * some room at the beginning (e.g. for inserting protocol headers)
1376 * the copies also have the room available.
1379 m_copypacket(struct mbuf *m, int how)
1381 struct mbuf *top, *n, *o;
1383 n = m_gethdr(how, m->m_type);
1388 if (!m_dup_pkthdr(n, m, how))
1390 n->m_len = m->m_len;
1391 if (m->m_flags & M_EXT) {
1392 KKASSERT((n->m_flags & M_EXT) == 0);
1393 n->m_data = m->m_data;
1394 m->m_ext.ext_ref(m->m_ext.ext_arg);
1395 n->m_ext = m->m_ext;
1396 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1398 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
1399 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1404 o = m_get(how, m->m_type);
1411 n->m_len = m->m_len;
1412 if (m->m_flags & M_EXT) {
1413 KKASSERT((n->m_flags & M_EXT) == 0);
1414 n->m_data = m->m_data;
1415 m->m_ext.ext_ref(m->m_ext.ext_arg);
1416 n->m_ext = m->m_ext;
1417 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1419 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
1427 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1432 * Copy data from an mbuf chain starting "off" bytes from the beginning,
1433 * continuing for "len" bytes, into the indicated buffer.
1436 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
1440 KASSERT(off >= 0, ("m_copydata, negative off %d", off));
1441 KASSERT(len >= 0, ("m_copydata, negative len %d", len));
1443 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
1450 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
1451 count = min(m->m_len - off, len);
1452 bcopy(mtod(m, caddr_t) + off, cp, count);
1461 * Copy a packet header mbuf chain into a completely new chain, including
1462 * copying any mbuf clusters. Use this instead of m_copypacket() when
1463 * you need a writable copy of an mbuf chain.
1466 m_dup(struct mbuf *m, int how)
1468 struct mbuf **p, *top = NULL;
1469 int remain, moff, nsize;
1474 KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __func__));
1476 /* While there's more data, get a new mbuf, tack it on, and fill it */
1477 remain = m->m_pkthdr.len;
1480 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */
1483 /* Get the next new mbuf */
1484 n = m_getl(remain, how, m->m_type, top == NULL ? M_PKTHDR : 0,
1489 if (!m_dup_pkthdr(n, m, how))
1492 /* Link it into the new chain */
1496 /* Copy data from original mbuf(s) into new mbuf */
1498 while (n->m_len < nsize && m != NULL) {
1499 int chunk = min(nsize - n->m_len, m->m_len - moff);
1501 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1505 if (moff == m->m_len) {
1511 /* Check correct total mbuf length */
1512 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
1513 ("%s: bogus m_pkthdr.len", __func__));
1520 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1525 * Copy the non-packet mbuf data chain into a new set of mbufs, including
1526 * copying any mbuf clusters. This is typically used to realign a data
1527 * chain by nfs_realign().
1529 * The original chain is left intact. how should be MB_WAIT or MB_DONTWAIT
1530 * and NULL can be returned if MB_DONTWAIT is passed.
1532 * Be careful to use cluster mbufs, a large mbuf chain converted to non
1533 * cluster mbufs can exhaust our supply of mbufs.
1536 m_dup_data(struct mbuf *m, int how)
1538 struct mbuf **p, *n, *top = NULL;
1539 int mlen, moff, chunk, gsize, nsize;
1548 * Optimize the mbuf allocation but do not get too carried away.
1550 if (m->m_next || m->m_len > MLEN)
1551 if (m->m_flags & M_EXT && m->m_ext.ext_size == MCLBYTES)
1554 gsize = MJUMPAGESIZE;
1564 * Scan the mbuf chain until nothing is left, the new mbuf chain
1565 * will be allocated on the fly as needed.
1572 KKASSERT(m->m_type == MT_DATA);
1574 n = m_getl(gsize, how, MT_DATA, 0, &nsize);
1581 chunk = imin(mlen, nsize);
1582 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
1597 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1602 * Concatenate mbuf chain n to m.
1603 * Both chains must be of the same type (e.g. MT_DATA).
1604 * Any m_pkthdr is not updated.
1607 m_cat(struct mbuf *m, struct mbuf *n)
1611 if (m->m_flags & M_EXT ||
1612 m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) {
1613 /* just join the two chains */
1617 /* splat the data from one into the other */
1618 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1620 m->m_len += n->m_len;
1626 m_adj(struct mbuf *mp, int req_len)
1632 if ((m = mp) == NULL)
1638 while (m != NULL && len > 0) {
1639 if (m->m_len <= len) {
1650 if (mp->m_flags & M_PKTHDR)
1651 m->m_pkthdr.len -= (req_len - len);
1654 * Trim from tail. Scan the mbuf chain,
1655 * calculating its length and finding the last mbuf.
1656 * If the adjustment only affects this mbuf, then just
1657 * adjust and return. Otherwise, rescan and truncate
1658 * after the remaining size.
1664 if (m->m_next == NULL)
1668 if (m->m_len >= len) {
1670 if (mp->m_flags & M_PKTHDR)
1671 mp->m_pkthdr.len -= len;
1678 * Correct length for chain is "count".
1679 * Find the mbuf with last data, adjust its length,
1680 * and toss data from remaining mbufs on chain.
1683 if (m->m_flags & M_PKTHDR)
1684 m->m_pkthdr.len = count;
1685 for (; m; m = m->m_next) {
1686 if (m->m_len >= count) {
1693 (m = m->m_next) ->m_len = 0;
1698 * Set the m_data pointer of a newly-allocated mbuf
1699 * to place an object of the specified size at the
1700 * end of the mbuf, longword aligned.
1703 m_align(struct mbuf *m, int len)
1707 if (m->m_flags & M_EXT)
1708 adjust = m->m_ext.ext_size - len;
1709 else if (m->m_flags & M_PKTHDR)
1710 adjust = MHLEN - len;
1712 adjust = MLEN - len;
1713 m->m_data += adjust &~ (sizeof(long)-1);
1717 * Rearrange an mbuf chain so that len bytes are contiguous
1718 * and in the data area of an mbuf (so that mtod will work for a structure
1719 * of size len). Returns the resulting mbuf chain on success, frees it and
1720 * returns null on failure. If there is room, it will add up to
1721 * max_protohdr-len extra bytes to the contiguous region in an attempt to
1722 * avoid being called next time.
1725 m_pullup(struct mbuf *n, int len)
1732 * If first mbuf has no cluster, and has room for len bytes
1733 * without shifting current data, pullup into it,
1734 * otherwise allocate a new mbuf to prepend to the chain.
1736 if (!(n->m_flags & M_EXT) &&
1737 n->m_data + len < &n->m_dat[MLEN] &&
1739 if (n->m_len >= len)
1747 if (n->m_flags & M_PKTHDR)
1748 m = m_gethdr(MB_DONTWAIT, n->m_type);
1750 m = m_get(MB_DONTWAIT, n->m_type);
1754 if (n->m_flags & M_PKTHDR)
1755 M_MOVE_PKTHDR(m, n);
1757 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
1759 count = min(min(max(len, max_protohdr), space), n->m_len);
1760 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
1770 } while (len > 0 && n);
1779 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1);
1784 * Partition an mbuf chain in two pieces, returning the tail --
1785 * all but the first len0 bytes. In case of failure, it returns NULL and
1786 * attempts to restore the chain to its original state.
1788 * Note that the resulting mbufs might be read-only, because the new
1789 * mbuf can end up sharing an mbuf cluster with the original mbuf if
1790 * the "breaking point" happens to lie within a cluster mbuf. Use the
1791 * M_WRITABLE() macro to check for this case.
1794 m_split(struct mbuf *m0, int len0, int wait)
1797 unsigned len = len0, remain;
1799 for (m = m0; m && len > m->m_len; m = m->m_next)
1803 remain = m->m_len - len;
1804 if (m0->m_flags & M_PKTHDR) {
1805 n = m_gethdr(wait, m0->m_type);
1808 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1809 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1810 m0->m_pkthdr.len = len0;
1811 if (m->m_flags & M_EXT)
1813 if (remain > MHLEN) {
1814 /* m can't be the lead packet */
1816 n->m_next = m_split(m, len, wait);
1817 if (n->m_next == NULL) {
1825 MH_ALIGN(n, remain);
1826 } else if (remain == 0) {
1831 n = m_get(wait, m->m_type);
1837 if (m->m_flags & M_EXT) {
1838 KKASSERT((n->m_flags & M_EXT) == 0);
1839 n->m_data = m->m_data + len;
1840 m->m_ext.ext_ref(m->m_ext.ext_arg);
1841 n->m_ext = m->m_ext;
1842 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER);
1844 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1848 n->m_next = m->m_next;
1854 * Routine to copy from device local memory into mbufs.
1855 * Note: "offset" is ill-defined and always called as 0, so ignore it.
1858 m_devget(char *buf, int len, int offset, struct ifnet *ifp,
1859 void (*copy)(volatile const void *from, volatile void *to, size_t length))
1861 struct mbuf *m, *mfirst = NULL, **mtail;
1870 m = m_getl(len, MB_DONTWAIT, MT_DATA, flags, &nsize);
1875 m->m_len = min(len, nsize);
1877 if (flags & M_PKTHDR) {
1878 if (len + max_linkhdr <= nsize)
1879 m->m_data += max_linkhdr;
1880 m->m_pkthdr.rcvif = ifp;
1881 m->m_pkthdr.len = len;
1885 copy(buf, m->m_data, (unsigned)m->m_len);
1896 * Routine to pad mbuf to the specified length 'padto'.
1899 m_devpad(struct mbuf *m, int padto)
1901 struct mbuf *last = NULL;
1904 if (padto <= m->m_pkthdr.len)
1907 padlen = padto - m->m_pkthdr.len;
1909 /* if there's only the packet-header and we can pad there, use it. */
1910 if (m->m_pkthdr.len == m->m_len && M_TRAILINGSPACE(m) >= padlen) {
1914 * Walk packet chain to find last mbuf. We will either
1915 * pad there, or append a new mbuf and pad it
1917 for (last = m; last->m_next != NULL; last = last->m_next)
1920 /* `last' now points to last in chain. */
1921 if (M_TRAILINGSPACE(last) < padlen) {
1924 /* Allocate new empty mbuf, pad it. Compact later. */
1925 MGET(n, MB_DONTWAIT, MT_DATA);
1933 KKASSERT(M_TRAILINGSPACE(last) >= padlen);
1934 KKASSERT(M_WRITABLE(last));
1936 /* Now zero the pad area */
1937 bzero(mtod(last, char *) + last->m_len, padlen);
1938 last->m_len += padlen;
1939 m->m_pkthdr.len += padlen;
1944 * Copy data from a buffer back into the indicated mbuf chain,
1945 * starting "off" bytes from the beginning, extending the mbuf
1946 * chain if necessary.
1949 m_copyback(struct mbuf *m0, int off, int len, caddr_t cp)
1952 struct mbuf *m = m0, *n;
1957 while (off > (mlen = m->m_len)) {
1960 if (m->m_next == NULL) {
1961 n = m_getclr(MB_DONTWAIT, m->m_type);
1964 n->m_len = min(MLEN, len + off);
1970 mlen = min (m->m_len - off, len);
1971 bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen);
1979 if (m->m_next == NULL) {
1980 n = m_get(MB_DONTWAIT, m->m_type);
1983 n->m_len = min(MLEN, len);
1988 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1989 m->m_pkthdr.len = totlen;
1993 * Append the specified data to the indicated mbuf chain,
1994 * Extend the mbuf chain if the new data does not fit in
1997 * Return 1 if able to complete the job; otherwise 0.
2000 m_append(struct mbuf *m0, int len, c_caddr_t cp)
2003 int remainder, space;
2005 for (m = m0; m->m_next != NULL; m = m->m_next)
2008 space = M_TRAILINGSPACE(m);
2011 * Copy into available space.
2013 if (space > remainder)
2015 bcopy(cp, mtod(m, caddr_t) + m->m_len, space);
2017 cp += space, remainder -= space;
2019 while (remainder > 0) {
2021 * Allocate a new mbuf; could check space
2022 * and allocate a cluster instead.
2024 n = m_get(MB_DONTWAIT, m->m_type);
2027 n->m_len = min(MLEN, remainder);
2028 bcopy(cp, mtod(n, caddr_t), n->m_len);
2029 cp += n->m_len, remainder -= n->m_len;
2033 if (m0->m_flags & M_PKTHDR)
2034 m0->m_pkthdr.len += len - remainder;
2035 return (remainder == 0);
2039 * Apply function f to the data in an mbuf chain starting "off" bytes from
2040 * the beginning, continuing for "len" bytes.
2043 m_apply(struct mbuf *m, int off, int len,
2044 int (*f)(void *, void *, u_int), void *arg)
2049 KASSERT(off >= 0, ("m_apply, negative off %d", off));
2050 KASSERT(len >= 0, ("m_apply, negative len %d", len));
2052 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
2059 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
2060 count = min(m->m_len - off, len);
2061 rval = (*f)(arg, mtod(m, caddr_t) + off, count);
2072 * Return a pointer to mbuf/offset of location in mbuf chain.
2075 m_getptr(struct mbuf *m, int loc, int *off)
2079 /* Normal end of search. */
2080 if (m->m_len > loc) {
2085 if (m->m_next == NULL) {
2087 /* Point at the end of valid data. */
2100 m_print(const struct mbuf *m)
2103 const struct mbuf *m2;
2105 len = m->m_pkthdr.len;
2108 kprintf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-");
2116 * "Move" mbuf pkthdr from "from" to "to".
2117 * "from" must have M_PKTHDR set, and "to" must be empty.
2120 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
2122 KASSERT((to->m_flags & M_PKTHDR), ("m_move_pkthdr: not packet header"));
2124 to->m_flags |= from->m_flags & M_COPYFLAGS;
2125 to->m_pkthdr = from->m_pkthdr; /* especially tags */
2126 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */
2130 * Duplicate "from"'s mbuf pkthdr in "to".
2131 * "from" must have M_PKTHDR set, and "to" must be empty.
2132 * In particular, this does a deep copy of the packet tags.
2135 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
2137 KASSERT((to->m_flags & M_PKTHDR), ("m_dup_pkthdr: not packet header"));
2139 to->m_flags = (from->m_flags & M_COPYFLAGS) |
2140 (to->m_flags & ~M_COPYFLAGS);
2141 to->m_pkthdr = from->m_pkthdr;
2142 SLIST_INIT(&to->m_pkthdr.tags);
2143 return (m_tag_copy_chain(to, from, how));
2147 * Defragment a mbuf chain, returning the shortest possible
2148 * chain of mbufs and clusters. If allocation fails and
2149 * this cannot be completed, NULL will be returned, but
2150 * the passed in chain will be unchanged. Upon success,
2151 * the original chain will be freed, and the new chain
2154 * If a non-packet header is passed in, the original
2155 * mbuf (chain?) will be returned unharmed.
2157 * m_defrag_nofree doesn't free the passed in mbuf.
2160 m_defrag(struct mbuf *m0, int how)
2164 if ((m_new = m_defrag_nofree(m0, how)) == NULL)
2172 m_defrag_nofree(struct mbuf *m0, int how)
2174 struct mbuf *m_new = NULL, *m_final = NULL;
2175 int progress = 0, length, nsize;
2177 if (!(m0->m_flags & M_PKTHDR))
2180 #ifdef MBUF_STRESS_TEST
2181 if (m_defragrandomfailures) {
2182 int temp = karc4random() & 0xff;
2188 m_final = m_getl(m0->m_pkthdr.len, how, MT_DATA, M_PKTHDR, &nsize);
2189 if (m_final == NULL)
2191 m_final->m_len = 0; /* in case m0->m_pkthdr.len is zero */
2193 if (m_dup_pkthdr(m_final, m0, how) == 0)
2198 while (progress < m0->m_pkthdr.len) {
2199 length = m0->m_pkthdr.len - progress;
2200 if (length > MCLBYTES)
2203 if (m_new == NULL) {
2204 m_new = m_getl(length, how, MT_DATA, 0, &nsize);
2209 m_copydata(m0, progress, length, mtod(m_new, caddr_t));
2211 m_new->m_len = length;
2212 if (m_new != m_final)
2213 m_cat(m_final, m_new);
2216 if (m0->m_next == NULL)
2219 m_defragbytes += m_final->m_pkthdr.len;
2230 * Move data from uio into mbufs.
2233 m_uiomove(struct uio *uio)
2235 struct mbuf *m; /* current working mbuf */
2236 struct mbuf *head = NULL; /* result mbuf chain */
2237 struct mbuf **mp = &head;
2238 int flags = M_PKTHDR;
2244 if (uio->uio_resid > INT_MAX)
2247 resid = (int)uio->uio_resid;
2248 m = m_getl(resid, MB_WAIT, MT_DATA, flags, &nsize);
2250 m->m_pkthdr.len = 0;
2251 /* Leave room for protocol headers. */
2256 m->m_len = imin(nsize, resid);
2257 error = uiomove(mtod(m, caddr_t), m->m_len, uio);
2264 head->m_pkthdr.len += m->m_len;
2265 } while (uio->uio_resid > 0);
2275 m_last(struct mbuf *m)
2283 * Return the number of bytes in an mbuf chain.
2284 * If lastm is not NULL, also return the last mbuf.
2287 m_lengthm(struct mbuf *m, struct mbuf **lastm)
2290 struct mbuf *prev = m;
2303 * Like m_lengthm(), except also keep track of mbuf usage.
2306 m_countm(struct mbuf *m, struct mbuf **lastm, u_int *pmbcnt)
2308 u_int len = 0, mbcnt = 0;
2309 struct mbuf *prev = m;
2314 if (m->m_flags & M_EXT)
2315 mbcnt += m->m_ext.ext_size;