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 * 4. Neither the name of the University nor the names of its contributors
49 * may be used to endorse or promote products derived from this software
50 * without specific prior written permission.
52 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
53 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
54 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
55 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
56 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
57 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
58 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
59 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
60 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
61 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64 * @(#)uipc_mbuf.c 8.2 (Berkeley) 1/4/94
65 * $FreeBSD: src/sys/kern/uipc_mbuf.c,v 1.51.2.24 2003/04/15 06:59:29 silby Exp $
68 #include "opt_param.h"
69 #include "opt_mbuf_stress_test.h"
70 #include <sys/param.h>
71 #include <sys/systm.h>
73 #include <sys/malloc.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/domain.h>
78 #include <sys/objcache.h>
80 #include <sys/protosw.h>
82 #include <sys/thread.h>
83 #include <sys/globaldata.h>
85 #include <sys/thread2.h>
86 #include <sys/spinlock2.h>
88 #include <machine/atomic.h>
89 #include <machine/limits.h>
92 #include <vm/vm_kern.h>
93 #include <vm/vm_extern.h>
96 #include <machine/cpu.h>
100 * mbuf cluster meta-data
108 * mbuf tracking for debugging purposes
112 static MALLOC_DEFINE(M_MTRACK
, "mtrack", "mtrack");
115 RB_HEAD(mbuf_rb_tree
, mbtrack
);
116 RB_PROTOTYPE2(mbuf_rb_tree
, mbtrack
, rb_node
, mbtrack_cmp
, struct mbuf
*);
119 RB_ENTRY(mbtrack
) rb_node
;
125 mbtrack_cmp(struct mbtrack
*mb1
, struct mbtrack
*mb2
)
134 RB_GENERATE2(mbuf_rb_tree
, mbtrack
, rb_node
, mbtrack_cmp
, struct mbuf
*, m
);
136 struct mbuf_rb_tree mbuf_track_root
;
137 static struct spinlock mbuf_track_spin
= SPINLOCK_INITIALIZER(mbuf_track_spin
);
140 mbuftrack(struct mbuf
*m
)
144 mbt
= kmalloc(sizeof(*mbt
), M_MTRACK
, M_INTWAIT
|M_ZERO
);
145 spin_lock(&mbuf_track_spin
);
147 if (mbuf_rb_tree_RB_INSERT(&mbuf_track_root
, mbt
)) {
148 spin_unlock(&mbuf_track_spin
);
149 panic("mbuftrack: mbuf %p already being tracked", m
);
151 spin_unlock(&mbuf_track_spin
);
155 mbufuntrack(struct mbuf
*m
)
159 spin_lock(&mbuf_track_spin
);
160 mbt
= mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root
, m
);
162 spin_unlock(&mbuf_track_spin
);
163 panic("mbufuntrack: mbuf %p was not tracked", m
);
165 mbuf_rb_tree_RB_REMOVE(&mbuf_track_root
, mbt
);
166 spin_unlock(&mbuf_track_spin
);
167 kfree(mbt
, M_MTRACK
);
172 mbuftrackid(struct mbuf
*m
, int trackid
)
177 spin_lock(&mbuf_track_spin
);
181 mbt
= mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root
, m
);
183 spin_unlock(&mbuf_track_spin
);
184 panic("mbuftrackid: mbuf %p not tracked", m
);
186 mbt
->trackid
= trackid
;
191 spin_unlock(&mbuf_track_spin
);
195 mbuftrack_callback(struct mbtrack
*mbt
, void *arg
)
197 struct sysctl_req
*req
= arg
;
201 ksnprintf(buf
, sizeof(buf
), "mbuf %p track %d\n", mbt
->m
, mbt
->trackid
);
203 spin_unlock(&mbuf_track_spin
);
204 error
= SYSCTL_OUT(req
, buf
, strlen(buf
));
205 spin_lock(&mbuf_track_spin
);
212 mbuftrack_show(SYSCTL_HANDLER_ARGS
)
216 spin_lock(&mbuf_track_spin
);
217 error
= mbuf_rb_tree_RB_SCAN(&mbuf_track_root
, NULL
,
218 mbuftrack_callback
, req
);
219 spin_unlock(&mbuf_track_spin
);
222 SYSCTL_PROC(_kern_ipc
, OID_AUTO
, showmbufs
, CTLFLAG_RD
|CTLTYPE_STRING
,
223 0, 0, mbuftrack_show
, "A", "Show all in-use mbufs");
228 #define mbufuntrack(m)
232 static void mbinit(void *);
233 SYSINIT(mbuf
, SI_BOOT2_MACHDEP
, SI_ORDER_FIRST
, mbinit
, NULL
)
235 static u_long mbtypes
[SMP_MAXCPU
][MT_NTYPES
];
237 static struct mbstat mbstat
[SMP_MAXCPU
];
246 #ifdef MBUF_STRESS_TEST
247 int m_defragrandomfailures
;
250 struct objcache
*mbuf_cache
, *mbufphdr_cache
;
251 struct objcache
*mclmeta_cache
, *mjclmeta_cache
;
252 struct objcache
*mbufcluster_cache
, *mbufphdrcluster_cache
;
253 struct objcache
*mbufjcluster_cache
, *mbufphdrjcluster_cache
;
256 static int nmbjclusters
;
259 static int mclph_cachefrac
;
260 static int mcl_cachefrac
;
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");
345 SYSCTL_INT(_kern_ipc
, OID_AUTO
, nmbjclusters
, CTLFLAG_RD
, &nmbjclusters
, 0,
346 "Maximum number of mbuf jclusters available");
347 SYSCTL_INT(_kern_ipc
, OID_AUTO
, mclph_cachefrac
, CTLFLAG_RD
,
349 "Fraction of cacheable mbuf clusters w/ pkthdr");
350 SYSCTL_INT(_kern_ipc
, OID_AUTO
, mcl_cachefrac
, CTLFLAG_RD
,
351 &mcl_cachefrac
, 0, "Fraction of cacheable mbuf clusters");
353 SYSCTL_INT(_kern_ipc
, OID_AUTO
, m_defragpackets
, CTLFLAG_RD
,
354 &m_defragpackets
, 0, "Number of defragment packets");
355 SYSCTL_INT(_kern_ipc
, OID_AUTO
, m_defragbytes
, CTLFLAG_RD
,
356 &m_defragbytes
, 0, "Number of defragment bytes");
357 SYSCTL_INT(_kern_ipc
, OID_AUTO
, m_defraguseless
, CTLFLAG_RD
,
358 &m_defraguseless
, 0, "Number of useless defragment mbuf chain operations");
359 SYSCTL_INT(_kern_ipc
, OID_AUTO
, m_defragfailure
, CTLFLAG_RD
,
360 &m_defragfailure
, 0, "Number of failed defragment mbuf chain operations");
361 #ifdef MBUF_STRESS_TEST
362 SYSCTL_INT(_kern_ipc
, OID_AUTO
, m_defragrandomfailures
, CTLFLAG_RW
,
363 &m_defragrandomfailures
, 0, "");
366 static MALLOC_DEFINE(M_MBUF
, "mbuf", "mbuf");
367 static MALLOC_DEFINE(M_MBUFCL
, "mbufcl", "mbufcl");
368 static MALLOC_DEFINE(M_MCLMETA
, "mclmeta", "mclmeta");
370 static void m_reclaim (void);
371 static void m_mclref(void *arg
);
372 static void m_mclfree(void *arg
);
375 * NOTE: Default NMBUFS must take into account a possible DOS attack
376 * using fd passing on unix domain sockets.
379 #define NMBCLUSTERS (512 + maxusers * 16)
381 #ifndef MCLPH_CACHEFRAC
382 #define MCLPH_CACHEFRAC 16
384 #ifndef MCL_CACHEFRAC
385 #define MCL_CACHEFRAC 4
388 #define NMBJCLUSTERS 2048
391 #define NMBUFS (nmbclusters * 2 + maxfiles)
395 * Perform sanity checks of tunables declared above.
398 tunable_mbinit(void *dummy
)
401 * This has to be done before VM init.
403 nmbclusters
= NMBCLUSTERS
;
404 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters
);
405 mclph_cachefrac
= MCLPH_CACHEFRAC
;
406 TUNABLE_INT_FETCH("kern.ipc.mclph_cachefrac", &mclph_cachefrac
);
407 mcl_cachefrac
= MCL_CACHEFRAC
;
408 TUNABLE_INT_FETCH("kern.ipc.mcl_cachefrac", &mcl_cachefrac
);
410 nmbjclusters
= NMBJCLUSTERS
;
411 TUNABLE_INT_FETCH("kern.ipc.nmbjclusters", &nmbjclusters
);
414 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs
);
417 if (nmbufs
< nmbclusters
* 2)
418 nmbufs
= nmbclusters
* 2;
420 SYSINIT(tunable_mbinit
, SI_BOOT1_TUNABLES
, SI_ORDER_ANY
,
421 tunable_mbinit
, NULL
);
423 /* "number of clusters of pages" */
429 * The mbuf object cache only guarantees that m_next and m_nextpkt are
430 * NULL and that m_data points to the beginning of the data area. In
431 * particular, m_len and m_pkthdr.len are uninitialized. It is the
432 * responsibility of the caller to initialize those fields before use.
435 static __inline boolean_t
436 mbuf_ctor(void *obj
, void *private, int ocflags
)
438 struct mbuf
*m
= obj
;
442 m
->m_data
= m
->m_dat
;
449 * Initialize the mbuf and the packet header fields.
452 mbufphdr_ctor(void *obj
, void *private, int ocflags
)
454 struct mbuf
*m
= obj
;
458 m
->m_data
= m
->m_pktdat
;
459 m
->m_flags
= M_PKTHDR
| M_PHCACHE
;
461 m
->m_pkthdr
.rcvif
= NULL
; /* eliminate XXX JH */
462 SLIST_INIT(&m
->m_pkthdr
.tags
);
463 m
->m_pkthdr
.csum_flags
= 0; /* eliminate XXX JH */
464 m
->m_pkthdr
.fw_flags
= 0; /* eliminate XXX JH */
470 * A mbcluster object consists of 2K (MCLBYTES) cluster and a refcount.
473 mclmeta_ctor(void *obj
, void *private, int ocflags
)
475 struct mbcluster
*cl
= obj
;
478 if (ocflags
& M_NOWAIT
)
479 buf
= kmalloc(MCLBYTES
, M_MBUFCL
, M_NOWAIT
| M_ZERO
);
481 buf
= kmalloc(MCLBYTES
, M_MBUFCL
, M_INTWAIT
| M_ZERO
);
490 mjclmeta_ctor(void *obj
, void *private, int ocflags
)
492 struct mbcluster
*cl
= obj
;
495 if (ocflags
& M_NOWAIT
)
496 buf
= kmalloc(MJUMPAGESIZE
, M_MBUFCL
, M_NOWAIT
| M_ZERO
);
498 buf
= kmalloc(MJUMPAGESIZE
, M_MBUFCL
, M_INTWAIT
| M_ZERO
);
507 mclmeta_dtor(void *obj
, void *private)
509 struct mbcluster
*mcl
= obj
;
511 KKASSERT(mcl
->mcl_refs
== 0);
512 kfree(mcl
->mcl_data
, M_MBUFCL
);
516 linkjcluster(struct mbuf
*m
, struct mbcluster
*cl
, uint size
)
519 * Add the cluster to the mbuf. The caller will detect that the
520 * mbuf now has an attached cluster.
522 m
->m_ext
.ext_arg
= cl
;
523 m
->m_ext
.ext_buf
= cl
->mcl_data
;
524 m
->m_ext
.ext_ref
= m_mclref
;
525 m
->m_ext
.ext_free
= m_mclfree
;
526 m
->m_ext
.ext_size
= size
;
527 atomic_add_int(&cl
->mcl_refs
, 1);
529 m
->m_data
= m
->m_ext
.ext_buf
;
530 m
->m_flags
|= M_EXT
| M_EXT_CLUSTER
;
534 linkcluster(struct mbuf
*m
, struct mbcluster
*cl
)
536 linkjcluster(m
, cl
, MCLBYTES
);
540 mbufphdrcluster_ctor(void *obj
, void *private, int ocflags
)
542 struct mbuf
*m
= obj
;
543 struct mbcluster
*cl
;
545 mbufphdr_ctor(obj
, private, ocflags
);
546 cl
= objcache_get(mclmeta_cache
, ocflags
);
548 ++mbstat
[mycpu
->gd_cpuid
].m_drops
;
551 m
->m_flags
|= M_CLCACHE
;
557 mbufphdrjcluster_ctor(void *obj
, void *private, int ocflags
)
559 struct mbuf
*m
= obj
;
560 struct mbcluster
*cl
;
562 mbufphdr_ctor(obj
, private, ocflags
);
563 cl
= objcache_get(mjclmeta_cache
, ocflags
);
565 ++mbstat
[mycpu
->gd_cpuid
].m_drops
;
568 m
->m_flags
|= M_CLCACHE
;
569 linkjcluster(m
, cl
, MJUMPAGESIZE
);
574 mbufcluster_ctor(void *obj
, void *private, int ocflags
)
576 struct mbuf
*m
= obj
;
577 struct mbcluster
*cl
;
579 mbuf_ctor(obj
, private, ocflags
);
580 cl
= objcache_get(mclmeta_cache
, ocflags
);
582 ++mbstat
[mycpu
->gd_cpuid
].m_drops
;
585 m
->m_flags
|= M_CLCACHE
;
591 mbufjcluster_ctor(void *obj
, void *private, int ocflags
)
593 struct mbuf
*m
= obj
;
594 struct mbcluster
*cl
;
596 mbuf_ctor(obj
, private, ocflags
);
597 cl
= objcache_get(mjclmeta_cache
, ocflags
);
599 ++mbstat
[mycpu
->gd_cpuid
].m_drops
;
602 m
->m_flags
|= M_CLCACHE
;
603 linkjcluster(m
, cl
, MJUMPAGESIZE
);
608 * Used for both the cluster and cluster PHDR caches.
610 * The mbuf may have lost its cluster due to sharing, deal
611 * with the situation by checking M_EXT.
614 mbufcluster_dtor(void *obj
, void *private)
616 struct mbuf
*m
= obj
;
617 struct mbcluster
*mcl
;
619 if (m
->m_flags
& M_EXT
) {
620 KKASSERT((m
->m_flags
& M_EXT_CLUSTER
) != 0);
621 mcl
= m
->m_ext
.ext_arg
;
622 KKASSERT(mcl
->mcl_refs
== 1);
624 if (m
->m_flags
& M_EXT
&& m
->m_ext
.ext_size
!= MCLBYTES
)
625 objcache_put(mjclmeta_cache
, mcl
);
627 objcache_put(mclmeta_cache
, mcl
);
631 struct objcache_malloc_args mbuf_malloc_args
= { MSIZE
, M_MBUF
};
632 struct objcache_malloc_args mclmeta_malloc_args
=
633 { sizeof(struct mbcluster
), M_MCLMETA
};
639 int mb_limit
, cl_limit
, ncl_limit
, jcl_limit
;
644 * Initialize statistics
646 for (i
= 0; i
< ncpus
; i
++) {
647 mbstat
[i
].m_msize
= MSIZE
;
648 mbstat
[i
].m_mclbytes
= MCLBYTES
;
649 mbstat
[i
].m_mjumpagesize
= MJUMPAGESIZE
;
650 mbstat
[i
].m_minclsize
= MINCLSIZE
;
651 mbstat
[i
].m_mlen
= MLEN
;
652 mbstat
[i
].m_mhlen
= MHLEN
;
656 * Create objtect caches and save cluster limits, which will
657 * be used to adjust backing kmalloc pools' limit later.
660 mb_limit
= cl_limit
= 0;
663 mbuf_cache
= objcache_create("mbuf",
665 mbuf_ctor
, NULL
, NULL
,
666 objcache_malloc_alloc
, objcache_malloc_free
, &mbuf_malloc_args
);
670 mbufphdr_cache
= objcache_create("mbuf pkt hdr",
672 mbufphdr_ctor
, NULL
, NULL
,
673 objcache_malloc_alloc
, objcache_malloc_free
, &mbuf_malloc_args
);
676 ncl_limit
= nmbclusters
;
677 mclmeta_cache
= objcache_create("cluster mbuf",
679 mclmeta_ctor
, mclmeta_dtor
, NULL
,
680 objcache_malloc_alloc
, objcache_malloc_free
, &mclmeta_malloc_args
);
681 cl_limit
+= ncl_limit
;
683 jcl_limit
= nmbjclusters
;
684 mjclmeta_cache
= objcache_create("jcluster mbuf",
686 mjclmeta_ctor
, mclmeta_dtor
, NULL
,
687 objcache_malloc_alloc
, objcache_malloc_free
, &mclmeta_malloc_args
);
688 cl_limit
+= jcl_limit
;
691 mbufcluster_cache
= objcache_create("mbuf + cluster",
692 limit
, nmbclusters
/ mcl_cachefrac
,
693 mbufcluster_ctor
, mbufcluster_dtor
, NULL
,
694 objcache_malloc_alloc
, objcache_malloc_free
, &mbuf_malloc_args
);
698 mbufphdrcluster_cache
= objcache_create("mbuf pkt hdr + cluster",
699 limit
, nmbclusters
/ mclph_cachefrac
,
700 mbufphdrcluster_ctor
, mbufcluster_dtor
, NULL
,
701 objcache_malloc_alloc
, objcache_malloc_free
, &mbuf_malloc_args
);
704 limit
= nmbjclusters
/ 4; /* XXX really rarely used */
705 mbufjcluster_cache
= objcache_create("mbuf + jcluster",
707 mbufjcluster_ctor
, mbufcluster_dtor
, NULL
,
708 objcache_malloc_alloc
, objcache_malloc_free
, &mbuf_malloc_args
);
711 limit
= nmbjclusters
;
712 mbufphdrjcluster_cache
= objcache_create("mbuf pkt hdr + jcluster",
713 limit
, nmbjclusters
/ 16,
714 mbufphdrjcluster_ctor
, mbufcluster_dtor
, NULL
,
715 objcache_malloc_alloc
, objcache_malloc_free
, &mbuf_malloc_args
);
719 * Adjust backing kmalloc pools' limit
721 * NOTE: We raise the limit by another 1/8 to take the effect
722 * of loosememuse into account.
724 cl_limit
+= cl_limit
/ 8;
725 kmalloc_raise_limit(mclmeta_malloc_args
.mtype
,
726 mclmeta_malloc_args
.objsize
* (size_t)cl_limit
);
727 kmalloc_raise_limit(M_MBUFCL
,
728 (MCLBYTES
* (size_t)ncl_limit
) +
729 (MJUMPAGESIZE
* (size_t)jcl_limit
));
731 mb_limit
+= mb_limit
/ 8;
732 kmalloc_raise_limit(mbuf_malloc_args
.mtype
,
733 mbuf_malloc_args
.objsize
* (size_t)mb_limit
);
737 * Return the number of references to this mbuf's data. 0 is returned
738 * if the mbuf is not M_EXT, a reference count is returned if it is
739 * M_EXT | M_EXT_CLUSTER, and 99 is returned if it is a special M_EXT.
742 m_sharecount(struct mbuf
*m
)
744 switch (m
->m_flags
& (M_EXT
| M_EXT_CLUSTER
)) {
749 case M_EXT
| M_EXT_CLUSTER
:
750 return (((struct mbcluster
*)m
->m_ext
.ext_arg
)->mcl_refs
);
753 return (0); /* to shut up compiler */
757 * change mbuf to new type
760 m_chtype(struct mbuf
*m
, int type
)
762 struct globaldata
*gd
= mycpu
;
764 ++mbtypes
[gd
->gd_cpuid
][type
];
765 --mbtypes
[gd
->gd_cpuid
][m
->m_type
];
775 kprintf("Debug: m_reclaim() called\n");
777 SLIST_FOREACH(dp
, &domains
, dom_next
) {
778 for (pr
= dp
->dom_protosw
; pr
< dp
->dom_protoswNPROTOSW
; pr
++) {
783 ++mbstat
[mycpu
->gd_cpuid
].m_drain
;
787 updatestats(struct mbuf
*m
, int type
)
789 struct globaldata
*gd
= mycpu
;
794 KASSERT(m
->m_next
== NULL
, ("mbuf %p: bad m_next in get", m
));
795 KASSERT(m
->m_nextpkt
== NULL
, ("mbuf %p: bad m_nextpkt in get", m
));
798 ++mbtypes
[gd
->gd_cpuid
][type
];
799 ++mbstat
[gd
->gd_cpuid
].m_mbufs
;
807 m_get(int how
, int type
)
811 int ocf
= MBTOM(how
);
815 m
= objcache_get(mbuf_cache
, ocf
);
818 if ((how
& MB_TRYWAIT
) && ntries
++ == 0) {
819 struct objcache
*reclaimlist
[] = {
822 mbufphdrcluster_cache
,
824 mbufphdrjcluster_cache
826 const int nreclaims
= NELEM(reclaimlist
);
828 if (!objcache_reclaimlist(reclaimlist
, nreclaims
, ocf
))
832 ++mbstat
[mycpu
->gd_cpuid
].m_drops
;
836 KASSERT(m
->m_data
== m
->m_dat
, ("mbuf %p: bad m_data in get", m
));
840 updatestats(m
, type
);
845 m_gethdr(int how
, int type
)
848 int ocf
= MBTOM(how
);
853 m
= objcache_get(mbufphdr_cache
, ocf
);
856 if ((how
& MB_TRYWAIT
) && ntries
++ == 0) {
857 struct objcache
*reclaimlist
[] = {
859 mbufcluster_cache
, mbufphdrcluster_cache
,
860 mbufjcluster_cache
, mbufphdrjcluster_cache
862 const int nreclaims
= NELEM(reclaimlist
);
864 if (!objcache_reclaimlist(reclaimlist
, nreclaims
, ocf
))
868 ++mbstat
[mycpu
->gd_cpuid
].m_drops
;
872 KASSERT(m
->m_data
== m
->m_pktdat
, ("mbuf %p: bad m_data in get", m
));
877 updatestats(m
, type
);
882 * Get a mbuf (not a mbuf cluster!) and zero it.
886 m_getclr(int how
, int type
)
890 m
= m_get(how
, type
);
892 bzero(m
->m_data
, MLEN
);
897 m_getcl_cache(int how
, short type
, int flags
, struct objcache
*mbclc
,
898 struct objcache
*mbphclc
)
900 struct mbuf
*m
= NULL
;
901 int ocflags
= MBTOM(how
);
906 if (flags
& M_PKTHDR
)
907 m
= objcache_get(mbphclc
, ocflags
);
909 m
= objcache_get(mbclc
, ocflags
);
912 if ((how
& MB_TRYWAIT
) && ntries
++ == 0) {
913 struct objcache
*reclaimlist
[1];
915 if (flags
& M_PKTHDR
)
916 reclaimlist
[0] = mbclc
;
918 reclaimlist
[0] = mbphclc
;
919 if (!objcache_reclaimlist(reclaimlist
, 1, ocflags
))
923 ++mbstat
[mycpu
->gd_cpuid
].m_drops
;
928 KASSERT(m
->m_data
== m
->m_ext
.ext_buf
,
929 ("mbuf %p: bad m_data in get", m
));
933 m
->m_pkthdr
.len
= 0; /* just do it unconditonally */
937 ++mbtypes
[mycpu
->gd_cpuid
][type
];
938 ++mbstat
[mycpu
->gd_cpuid
].m_clusters
;
943 m_getjcl(int how
, short type
, int flags
, size_t size
)
945 struct objcache
*mbclc
, *mbphclc
;
949 mbclc
= mbufcluster_cache
;
950 mbphclc
= mbufphdrcluster_cache
;
954 mbclc
= mbufjcluster_cache
;
955 mbphclc
= mbufphdrjcluster_cache
;
958 return m_getcl_cache(how
, type
, flags
, mbclc
, mbphclc
);
962 * Returns an mbuf with an attached cluster.
963 * Because many network drivers use this kind of buffers a lot, it is
964 * convenient to keep a small pool of free buffers of this kind.
965 * Even a small size such as 10 gives about 10% improvement in the
966 * forwarding rate in a bridge or router.
969 m_getcl(int how
, short type
, int flags
)
971 return m_getcl_cache(how
, type
, flags
,
972 mbufcluster_cache
, mbufphdrcluster_cache
);
976 * Allocate chain of requested length.
979 m_getc(int len
, int how
, int type
)
981 struct mbuf
*n
, *nfirst
= NULL
, **ntail
= &nfirst
;
985 n
= m_getl(len
, how
, type
, 0, &nsize
);
1001 * Allocate len-worth of mbufs and/or mbuf clusters (whatever fits best)
1002 * and return a pointer to the head of the allocated chain. If m0 is
1003 * non-null, then we assume that it is a single mbuf or an mbuf chain to
1004 * which we want len bytes worth of mbufs and/or clusters attached, and so
1005 * if we succeed in allocating it, we will just return a pointer to m0.
1007 * If we happen to fail at any point during the allocation, we will free
1008 * up everything we have already allocated and return NULL.
1010 * Deprecated. Use m_getc() and m_cat() instead.
1013 m_getm(struct mbuf
*m0
, int len
, int type
, int how
)
1015 struct mbuf
*nfirst
;
1017 nfirst
= m_getc(len
, how
, type
);
1020 m_last(m0
)->m_next
= nfirst
;
1028 * Adds a cluster to a normal mbuf, M_EXT is set on success.
1029 * Deprecated. Use m_getcl() instead.
1032 m_mclget(struct mbuf
*m
, int how
)
1034 struct mbcluster
*mcl
;
1036 KKASSERT((m
->m_flags
& M_EXT
) == 0);
1037 mcl
= objcache_get(mclmeta_cache
, MBTOM(how
));
1039 linkcluster(m
, mcl
);
1040 ++mbstat
[mycpu
->gd_cpuid
].m_clusters
;
1042 ++mbstat
[mycpu
->gd_cpuid
].m_drops
;
1047 * Updates to mbcluster must be MPSAFE. Only an entity which already has
1048 * a reference to the cluster can ref it, so we are in no danger of
1049 * racing an add with a subtract. But the operation must still be atomic
1050 * since multiple entities may have a reference on the cluster.
1052 * m_mclfree() is almost the same but it must contend with two entities
1053 * freeing the cluster at the same time.
1058 struct mbcluster
*mcl
= arg
;
1060 atomic_add_int(&mcl
->mcl_refs
, 1);
1064 * When dereferencing a cluster we have to deal with a N->0 race, where
1065 * N entities free their references simultaniously. To do this we use
1066 * atomic_fetchadd_int().
1069 m_mclfree(void *arg
)
1071 struct mbcluster
*mcl
= arg
;
1073 if (atomic_fetchadd_int(&mcl
->mcl_refs
, -1) == 1) {
1074 --mbstat
[mycpu
->gd_cpuid
].m_clusters
;
1075 objcache_put(mclmeta_cache
, mcl
);
1080 * Free a single mbuf and any associated external storage. The successor,
1081 * if any, is returned.
1083 * We do need to check non-first mbuf for m_aux, since some of existing
1084 * code does not call M_PREPEND properly.
1085 * (example: call to bpf_mtap from drivers)
1091 _m_free(struct mbuf
*m
, const char *func
)
1096 m_free(struct mbuf
*m
)
1101 struct globaldata
*gd
= mycpu
;
1103 KASSERT(m
->m_type
!= MT_FREE
, ("freeing free mbuf %p", m
));
1104 KASSERT(M_TRAILINGSPACE(m
) >= 0, ("overflowed mbuf %p", m
));
1105 --mbtypes
[gd
->gd_cpuid
][m
->m_type
];
1110 * Make sure the mbuf is in constructed state before returning it
1116 m
->m_hdr
.mh_lastfunc
= func
;
1119 KKASSERT(m
->m_nextpkt
== NULL
);
1121 if (m
->m_nextpkt
!= NULL
) {
1122 static int afewtimes
= 10;
1124 if (afewtimes
-- > 0) {
1125 kprintf("mfree: m->m_nextpkt != NULL\n");
1126 print_backtrace(-1);
1128 m
->m_nextpkt
= NULL
;
1131 if (m
->m_flags
& M_PKTHDR
) {
1132 m_tag_delete_chain(m
); /* eliminate XXX JH */
1135 m
->m_flags
&= (M_EXT
| M_EXT_CLUSTER
| M_CLCACHE
| M_PHCACHE
);
1138 * Clean the M_PKTHDR state so we can return the mbuf to its original
1139 * cache. This is based on the PHCACHE flag which tells us whether
1140 * the mbuf was originally allocated out of a packet-header cache
1141 * or a non-packet-header cache.
1143 if (m
->m_flags
& M_PHCACHE
) {
1144 m
->m_flags
|= M_PKTHDR
;
1145 m
->m_pkthdr
.rcvif
= NULL
; /* eliminate XXX JH */
1146 m
->m_pkthdr
.csum_flags
= 0; /* eliminate XXX JH */
1147 m
->m_pkthdr
.fw_flags
= 0; /* eliminate XXX JH */
1148 SLIST_INIT(&m
->m_pkthdr
.tags
);
1152 * Handle remaining flags combinations. M_CLCACHE tells us whether
1153 * the mbuf was originally allocated from a cluster cache or not,
1154 * and is totally separate from whether the mbuf is currently
1155 * associated with a cluster.
1157 switch(m
->m_flags
& (M_CLCACHE
| M_EXT
| M_EXT_CLUSTER
)) {
1158 case M_CLCACHE
| M_EXT
| M_EXT_CLUSTER
:
1160 * mbuf+cluster cache case. The mbuf was allocated from the
1161 * combined mbuf_cluster cache and can be returned to the
1162 * cache if the cluster hasn't been shared.
1164 if (m_sharecount(m
) == 1) {
1166 * The cluster has not been shared, we can just
1167 * reset the data pointer and return the mbuf
1168 * to the cluster cache. Note that the reference
1169 * count is left intact (it is still associated with
1172 m
->m_data
= m
->m_ext
.ext_buf
;
1173 if (m
->m_flags
& M_EXT
&& m
->m_ext
.ext_size
!= MCLBYTES
) {
1174 if (m
->m_flags
& M_PHCACHE
)
1175 objcache_put(mbufphdrjcluster_cache
, m
);
1177 objcache_put(mbufjcluster_cache
, m
);
1179 if (m
->m_flags
& M_PHCACHE
)
1180 objcache_put(mbufphdrcluster_cache
, m
);
1182 objcache_put(mbufcluster_cache
, m
);
1184 --mbstat
[mycpu
->gd_cpuid
].m_clusters
;
1187 * Hell. Someone else has a ref on this cluster,
1188 * we have to disconnect it which means we can't
1189 * put it back into the mbufcluster_cache, we
1190 * have to destroy the mbuf.
1192 * Other mbuf references to the cluster will typically
1193 * be M_EXT | M_EXT_CLUSTER but without M_CLCACHE.
1195 * XXX we could try to connect another cluster to
1198 m
->m_ext
.ext_free(m
->m_ext
.ext_arg
);
1199 m
->m_flags
&= ~(M_EXT
| M_EXT_CLUSTER
);
1200 if (m
->m_ext
.ext_size
== MCLBYTES
) {
1201 if (m
->m_flags
& M_PHCACHE
)
1202 objcache_dtor(mbufphdrcluster_cache
, m
);
1204 objcache_dtor(mbufcluster_cache
, m
);
1206 if (m
->m_flags
& M_PHCACHE
)
1207 objcache_dtor(mbufphdrjcluster_cache
, m
);
1209 objcache_dtor(mbufjcluster_cache
, m
);
1213 case M_EXT
| M_EXT_CLUSTER
:
1216 * Normal cluster association case, disconnect the cluster from
1217 * the mbuf. The cluster may or may not be custom.
1219 m
->m_ext
.ext_free(m
->m_ext
.ext_arg
);
1220 m
->m_flags
&= ~(M_EXT
| M_EXT_CLUSTER
);
1224 * return the mbuf to the mbuf cache.
1226 if (m
->m_flags
& M_PHCACHE
) {
1227 m
->m_data
= m
->m_pktdat
;
1228 objcache_put(mbufphdr_cache
, m
);
1230 m
->m_data
= m
->m_dat
;
1231 objcache_put(mbuf_cache
, m
);
1233 --mbstat
[mycpu
->gd_cpuid
].m_mbufs
;
1237 panic("bad mbuf flags %p %08x", m
, m
->m_flags
);
1246 _m_freem(struct mbuf
*m
, const char *func
)
1249 m
= _m_free(m
, func
);
1255 m_freem(struct mbuf
*m
)
1264 m_extadd(struct mbuf
*m
, caddr_t buf
, u_int size
, void (*reff
)(void *),
1265 void (*freef
)(void *), void *arg
)
1267 m
->m_ext
.ext_arg
= arg
;
1268 m
->m_ext
.ext_buf
= buf
;
1269 m
->m_ext
.ext_ref
= reff
;
1270 m
->m_ext
.ext_free
= freef
;
1271 m
->m_ext
.ext_size
= size
;
1274 m
->m_flags
|= M_EXT
;
1278 * mbuf utility routines
1282 * Lesser-used path for M_PREPEND: allocate new mbuf to prepend to chain and
1286 m_prepend(struct mbuf
*m
, int len
, int how
)
1290 if (m
->m_flags
& M_PKTHDR
)
1291 mn
= m_gethdr(how
, m
->m_type
);
1293 mn
= m_get(how
, m
->m_type
);
1298 if (m
->m_flags
& M_PKTHDR
)
1299 M_MOVE_PKTHDR(mn
, m
);
1309 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
1310 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
1311 * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller.
1312 * Note that the copy is read-only, because clusters are not copied,
1313 * only their reference counts are incremented.
1316 m_copym(const struct mbuf
*m
, int off0
, int len
, int wait
)
1318 struct mbuf
*n
, **np
;
1323 KASSERT(off
>= 0, ("m_copym, negative off %d", off
));
1324 KASSERT(len
>= 0, ("m_copym, negative len %d", len
));
1325 if (off
== 0 && (m
->m_flags
& M_PKTHDR
))
1328 KASSERT(m
!= NULL
, ("m_copym, offset > size of mbuf chain"));
1338 KASSERT(len
== M_COPYALL
,
1339 ("m_copym, length > size of mbuf chain"));
1343 * Because we are sharing any cluster attachment below,
1344 * be sure to get an mbuf that does not have a cluster
1345 * associated with it.
1348 n
= m_gethdr(wait
, m
->m_type
);
1350 n
= m_get(wait
, m
->m_type
);
1355 if (!m_dup_pkthdr(n
, m
, wait
))
1357 if (len
== M_COPYALL
)
1358 n
->m_pkthdr
.len
-= off0
;
1360 n
->m_pkthdr
.len
= len
;
1363 n
->m_len
= min(len
, m
->m_len
- off
);
1364 if (m
->m_flags
& M_EXT
) {
1365 KKASSERT((n
->m_flags
& M_EXT
) == 0);
1366 n
->m_data
= m
->m_data
+ off
;
1367 m
->m_ext
.ext_ref(m
->m_ext
.ext_arg
);
1368 n
->m_ext
= m
->m_ext
;
1369 n
->m_flags
|= m
->m_flags
& (M_EXT
| M_EXT_CLUSTER
);
1371 bcopy(mtod(m
, caddr_t
)+off
, mtod(n
, caddr_t
),
1372 (unsigned)n
->m_len
);
1374 if (len
!= M_COPYALL
)
1381 ++mbstat
[mycpu
->gd_cpuid
].m_mcfail
;
1385 ++mbstat
[mycpu
->gd_cpuid
].m_mcfail
;
1390 * Copy an entire packet, including header (which must be present).
1391 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
1392 * Note that the copy is read-only, because clusters are not copied,
1393 * only their reference counts are incremented.
1394 * Preserve alignment of the first mbuf so if the creator has left
1395 * some room at the beginning (e.g. for inserting protocol headers)
1396 * the copies also have the room available.
1399 m_copypacket(struct mbuf
*m
, int how
)
1401 struct mbuf
*top
, *n
, *o
;
1403 n
= m_gethdr(how
, m
->m_type
);
1408 if (!m_dup_pkthdr(n
, m
, how
))
1410 n
->m_len
= m
->m_len
;
1411 if (m
->m_flags
& M_EXT
) {
1412 KKASSERT((n
->m_flags
& M_EXT
) == 0);
1413 n
->m_data
= m
->m_data
;
1414 m
->m_ext
.ext_ref(m
->m_ext
.ext_arg
);
1415 n
->m_ext
= m
->m_ext
;
1416 n
->m_flags
|= m
->m_flags
& (M_EXT
| M_EXT_CLUSTER
);
1418 n
->m_data
= n
->m_pktdat
+ (m
->m_data
- m
->m_pktdat
);
1419 bcopy(mtod(m
, char *), mtod(n
, char *), n
->m_len
);
1424 o
= m_get(how
, m
->m_type
);
1431 n
->m_len
= m
->m_len
;
1432 if (m
->m_flags
& M_EXT
) {
1433 KKASSERT((n
->m_flags
& M_EXT
) == 0);
1434 n
->m_data
= m
->m_data
;
1435 m
->m_ext
.ext_ref(m
->m_ext
.ext_arg
);
1436 n
->m_ext
= m
->m_ext
;
1437 n
->m_flags
|= m
->m_flags
& (M_EXT
| M_EXT_CLUSTER
);
1439 bcopy(mtod(m
, char *), mtod(n
, char *), n
->m_len
);
1447 ++mbstat
[mycpu
->gd_cpuid
].m_mcfail
;
1452 * Copy data from an mbuf chain starting "off" bytes from the beginning,
1453 * continuing for "len" bytes, into the indicated buffer.
1456 m_copydata(const struct mbuf
*m
, int off
, int len
, caddr_t cp
)
1460 KASSERT(off
>= 0, ("m_copydata, negative off %d", off
));
1461 KASSERT(len
>= 0, ("m_copydata, negative len %d", len
));
1463 KASSERT(m
!= NULL
, ("m_copydata, offset > size of mbuf chain"));
1470 KASSERT(m
!= NULL
, ("m_copydata, length > size of mbuf chain"));
1471 count
= min(m
->m_len
- off
, len
);
1472 bcopy(mtod(m
, caddr_t
) + off
, cp
, count
);
1481 * Copy a packet header mbuf chain into a completely new chain, including
1482 * copying any mbuf clusters. Use this instead of m_copypacket() when
1483 * you need a writable copy of an mbuf chain.
1486 m_dup(struct mbuf
*m
, int how
)
1488 struct mbuf
**p
, *top
= NULL
;
1489 int remain
, moff
, nsize
;
1494 KASSERT((m
->m_flags
& M_PKTHDR
) != 0, ("%s: !PKTHDR", __func__
));
1496 /* While there's more data, get a new mbuf, tack it on, and fill it */
1497 remain
= m
->m_pkthdr
.len
;
1500 while (remain
> 0 || top
== NULL
) { /* allow m->m_pkthdr.len == 0 */
1503 /* Get the next new mbuf */
1504 n
= m_getl(remain
, how
, m
->m_type
, top
== NULL
? M_PKTHDR
: 0,
1509 if (!m_dup_pkthdr(n
, m
, how
))
1512 /* Link it into the new chain */
1516 /* Copy data from original mbuf(s) into new mbuf */
1518 while (n
->m_len
< nsize
&& m
!= NULL
) {
1519 int chunk
= min(nsize
- n
->m_len
, m
->m_len
- moff
);
1521 bcopy(m
->m_data
+ moff
, n
->m_data
+ n
->m_len
, chunk
);
1525 if (moff
== m
->m_len
) {
1531 /* Check correct total mbuf length */
1532 KASSERT((remain
> 0 && m
!= NULL
) || (remain
== 0 && m
== NULL
),
1533 ("%s: bogus m_pkthdr.len", __func__
));
1540 ++mbstat
[mycpu
->gd_cpuid
].m_mcfail
;
1545 * Copy the non-packet mbuf data chain into a new set of mbufs, including
1546 * copying any mbuf clusters. This is typically used to realign a data
1547 * chain by nfs_realign().
1549 * The original chain is left intact. how should be MB_WAIT or MB_DONTWAIT
1550 * and NULL can be returned if MB_DONTWAIT is passed.
1552 * Be careful to use cluster mbufs, a large mbuf chain converted to non
1553 * cluster mbufs can exhaust our supply of mbufs.
1556 m_dup_data(struct mbuf
*m
, int how
)
1558 struct mbuf
**p
, *n
, *top
= NULL
;
1559 int mlen
, moff
, chunk
, gsize
, nsize
;
1568 * Optimize the mbuf allocation but do not get too carried away.
1570 if (m
->m_next
|| m
->m_len
> MLEN
)
1571 if (m
->m_flags
& M_EXT
&& m
->m_ext
.ext_size
== MCLBYTES
)
1574 gsize
= MJUMPAGESIZE
;
1584 * Scan the mbuf chain until nothing is left, the new mbuf chain
1585 * will be allocated on the fly as needed.
1592 KKASSERT(m
->m_type
== MT_DATA
);
1594 n
= m_getl(gsize
, how
, MT_DATA
, 0, &nsize
);
1601 chunk
= imin(mlen
, nsize
);
1602 bcopy(m
->m_data
+ moff
, n
->m_data
+ n
->m_len
, chunk
);
1617 ++mbstat
[mycpu
->gd_cpuid
].m_mcfail
;
1622 * Concatenate mbuf chain n to m.
1623 * Both chains must be of the same type (e.g. MT_DATA).
1624 * Any m_pkthdr is not updated.
1627 m_cat(struct mbuf
*m
, struct mbuf
*n
)
1631 if (m
->m_flags
& M_EXT
||
1632 m
->m_data
+ m
->m_len
+ n
->m_len
>= &m
->m_dat
[MLEN
]) {
1633 /* just join the two chains */
1637 /* splat the data from one into the other */
1638 bcopy(mtod(n
, caddr_t
), mtod(m
, caddr_t
) + m
->m_len
,
1640 m
->m_len
+= n
->m_len
;
1646 m_adj(struct mbuf
*mp
, int req_len
)
1652 if ((m
= mp
) == NULL
)
1658 while (m
!= NULL
&& len
> 0) {
1659 if (m
->m_len
<= len
) {
1670 if (mp
->m_flags
& M_PKTHDR
)
1671 m
->m_pkthdr
.len
-= (req_len
- len
);
1674 * Trim from tail. Scan the mbuf chain,
1675 * calculating its length and finding the last mbuf.
1676 * If the adjustment only affects this mbuf, then just
1677 * adjust and return. Otherwise, rescan and truncate
1678 * after the remaining size.
1684 if (m
->m_next
== NULL
)
1688 if (m
->m_len
>= len
) {
1690 if (mp
->m_flags
& M_PKTHDR
)
1691 mp
->m_pkthdr
.len
-= len
;
1698 * Correct length for chain is "count".
1699 * Find the mbuf with last data, adjust its length,
1700 * and toss data from remaining mbufs on chain.
1703 if (m
->m_flags
& M_PKTHDR
)
1704 m
->m_pkthdr
.len
= count
;
1705 for (; m
; m
= m
->m_next
) {
1706 if (m
->m_len
>= count
) {
1713 (m
= m
->m_next
) ->m_len
= 0;
1718 * Set the m_data pointer of a newly-allocated mbuf
1719 * to place an object of the specified size at the
1720 * end of the mbuf, longword aligned.
1723 m_align(struct mbuf
*m
, int len
)
1727 if (m
->m_flags
& M_EXT
)
1728 adjust
= m
->m_ext
.ext_size
- len
;
1729 else if (m
->m_flags
& M_PKTHDR
)
1730 adjust
= MHLEN
- len
;
1732 adjust
= MLEN
- len
;
1733 m
->m_data
+= adjust
&~ (sizeof(long)-1);
1737 * Create a writable copy of the mbuf chain. While doing this
1738 * we compact the chain with a goal of producing a chain with
1739 * at most two mbufs. The second mbuf in this chain is likely
1740 * to be a cluster. The primary purpose of this work is to create
1741 * a writable packet for encryption, compression, etc. The
1742 * secondary goal is to linearize the data so the data can be
1743 * passed to crypto hardware in the most efficient manner possible.
1746 m_unshare(struct mbuf
*m0
, int how
)
1748 struct mbuf
*m
, *mprev
;
1749 struct mbuf
*n
, *mfirst
, *mlast
;
1753 for (m
= m0
; m
!= NULL
; m
= mprev
->m_next
) {
1755 * Regular mbufs are ignored unless there's a cluster
1756 * in front of it that we can use to coalesce. We do
1757 * the latter mainly so later clusters can be coalesced
1758 * also w/o having to handle them specially (i.e. convert
1759 * mbuf+cluster -> cluster). This optimization is heavily
1760 * influenced by the assumption that we're running over
1761 * Ethernet where MCLBYTES is large enough that the max
1762 * packet size will permit lots of coalescing into a
1763 * single cluster. This in turn permits efficient
1764 * crypto operations, especially when using hardware.
1766 if ((m
->m_flags
& M_EXT
) == 0) {
1767 if (mprev
&& (mprev
->m_flags
& M_EXT
) &&
1768 m
->m_len
<= M_TRAILINGSPACE(mprev
)) {
1769 /* XXX: this ignores mbuf types */
1770 memcpy(mtod(mprev
, caddr_t
) + mprev
->m_len
,
1771 mtod(m
, caddr_t
), m
->m_len
);
1772 mprev
->m_len
+= m
->m_len
;
1773 mprev
->m_next
= m
->m_next
; /* unlink from chain */
1774 m_free(m
); /* reclaim mbuf */
1781 * Writable mbufs are left alone (for now).
1783 if (M_WRITABLE(m
)) {
1789 * Not writable, replace with a copy or coalesce with
1790 * the previous mbuf if possible (since we have to copy
1791 * it anyway, we try to reduce the number of mbufs and
1792 * clusters so that future work is easier).
1794 KASSERT(m
->m_flags
& M_EXT
, ("m_flags 0x%x", m
->m_flags
));
1795 /* NB: we only coalesce into a cluster or larger */
1796 if (mprev
!= NULL
&& (mprev
->m_flags
& M_EXT
) &&
1797 m
->m_len
<= M_TRAILINGSPACE(mprev
)) {
1798 /* XXX: this ignores mbuf types */
1799 memcpy(mtod(mprev
, caddr_t
) + mprev
->m_len
,
1800 mtod(m
, caddr_t
), m
->m_len
);
1801 mprev
->m_len
+= m
->m_len
;
1802 mprev
->m_next
= m
->m_next
; /* unlink from chain */
1803 m_free(m
); /* reclaim mbuf */
1808 * Allocate new space to hold the copy...
1810 /* XXX why can M_PKTHDR be set past the first mbuf? */
1811 if (mprev
== NULL
&& (m
->m_flags
& M_PKTHDR
)) {
1813 * NB: if a packet header is present we must
1814 * allocate the mbuf separately from any cluster
1815 * because M_MOVE_PKTHDR will smash the data
1816 * pointer and drop the M_EXT marker.
1818 MGETHDR(n
, how
, m
->m_type
);
1823 M_MOVE_PKTHDR(n
, m
);
1825 if ((n
->m_flags
& M_EXT
) == 0) {
1831 n
= m_getcl(how
, m
->m_type
, m
->m_flags
);
1838 * ... and copy the data. We deal with jumbo mbufs
1839 * (i.e. m_len > MCLBYTES) by splitting them into
1840 * clusters. We could just malloc a buffer and make
1841 * it external but too many device drivers don't know
1842 * how to break up the non-contiguous memory when
1850 int cc
= min(len
, MCLBYTES
);
1851 memcpy(mtod(n
, caddr_t
), mtod(m
, caddr_t
) + off
, cc
);
1862 n
= m_getcl(how
, m
->m_type
, m
->m_flags
);
1869 n
->m_next
= m
->m_next
;
1871 m0
= mfirst
; /* new head of chain */
1873 mprev
->m_next
= mfirst
; /* replace old mbuf */
1874 m_free(m
); /* release old mbuf */
1881 * Rearrange an mbuf chain so that len bytes are contiguous
1882 * and in the data area of an mbuf (so that mtod will work for a structure
1883 * of size len). Returns the resulting mbuf chain on success, frees it and
1884 * returns null on failure. If there is room, it will add up to
1885 * max_protohdr-len extra bytes to the contiguous region in an attempt to
1886 * avoid being called next time.
1889 m_pullup(struct mbuf
*n
, int len
)
1896 * If first mbuf has no cluster, and has room for len bytes
1897 * without shifting current data, pullup into it,
1898 * otherwise allocate a new mbuf to prepend to the chain.
1900 if (!(n
->m_flags
& M_EXT
) &&
1901 n
->m_data
+ len
< &n
->m_dat
[MLEN
] &&
1903 if (n
->m_len
>= len
)
1911 if (n
->m_flags
& M_PKTHDR
)
1912 m
= m_gethdr(MB_DONTWAIT
, n
->m_type
);
1914 m
= m_get(MB_DONTWAIT
, n
->m_type
);
1918 if (n
->m_flags
& M_PKTHDR
)
1919 M_MOVE_PKTHDR(m
, n
);
1921 space
= &m
->m_dat
[MLEN
] - (m
->m_data
+ m
->m_len
);
1923 count
= min(min(max(len
, max_protohdr
), space
), n
->m_len
);
1924 bcopy(mtod(n
, caddr_t
), mtod(m
, caddr_t
) + m
->m_len
,
1934 } while (len
> 0 && n
);
1943 ++mbstat
[mycpu
->gd_cpuid
].m_mcfail
;
1948 * Partition an mbuf chain in two pieces, returning the tail --
1949 * all but the first len0 bytes. In case of failure, it returns NULL and
1950 * attempts to restore the chain to its original state.
1952 * Note that the resulting mbufs might be read-only, because the new
1953 * mbuf can end up sharing an mbuf cluster with the original mbuf if
1954 * the "breaking point" happens to lie within a cluster mbuf. Use the
1955 * M_WRITABLE() macro to check for this case.
1958 m_split(struct mbuf
*m0
, int len0
, int wait
)
1961 unsigned len
= len0
, remain
;
1963 for (m
= m0
; m
&& len
> m
->m_len
; m
= m
->m_next
)
1967 remain
= m
->m_len
- len
;
1968 if (m0
->m_flags
& M_PKTHDR
) {
1969 n
= m_gethdr(wait
, m0
->m_type
);
1972 n
->m_pkthdr
.rcvif
= m0
->m_pkthdr
.rcvif
;
1973 n
->m_pkthdr
.len
= m0
->m_pkthdr
.len
- len0
;
1974 m0
->m_pkthdr
.len
= len0
;
1975 if (m
->m_flags
& M_EXT
)
1977 if (remain
> MHLEN
) {
1978 /* m can't be the lead packet */
1980 n
->m_next
= m_split(m
, len
, wait
);
1981 if (n
->m_next
== NULL
) {
1989 MH_ALIGN(n
, remain
);
1990 } else if (remain
== 0) {
1995 n
= m_get(wait
, m
->m_type
);
2001 if (m
->m_flags
& M_EXT
) {
2002 KKASSERT((n
->m_flags
& M_EXT
) == 0);
2003 n
->m_data
= m
->m_data
+ len
;
2004 m
->m_ext
.ext_ref(m
->m_ext
.ext_arg
);
2005 n
->m_ext
= m
->m_ext
;
2006 n
->m_flags
|= m
->m_flags
& (M_EXT
| M_EXT_CLUSTER
);
2008 bcopy(mtod(m
, caddr_t
) + len
, mtod(n
, caddr_t
), remain
);
2012 n
->m_next
= m
->m_next
;
2018 * Routine to copy from device local memory into mbufs.
2019 * Note: "offset" is ill-defined and always called as 0, so ignore it.
2022 m_devget(char *buf
, int len
, int offset
, struct ifnet
*ifp
,
2023 void (*copy
)(volatile const void *from
, volatile void *to
, size_t length
))
2025 struct mbuf
*m
, *mfirst
= NULL
, **mtail
;
2034 m
= m_getl(len
, MB_DONTWAIT
, MT_DATA
, flags
, &nsize
);
2039 m
->m_len
= min(len
, nsize
);
2041 if (flags
& M_PKTHDR
) {
2042 if (len
+ max_linkhdr
<= nsize
)
2043 m
->m_data
+= max_linkhdr
;
2044 m
->m_pkthdr
.rcvif
= ifp
;
2045 m
->m_pkthdr
.len
= len
;
2049 copy(buf
, m
->m_data
, (unsigned)m
->m_len
);
2060 * Routine to pad mbuf to the specified length 'padto'.
2063 m_devpad(struct mbuf
*m
, int padto
)
2065 struct mbuf
*last
= NULL
;
2068 if (padto
<= m
->m_pkthdr
.len
)
2071 padlen
= padto
- m
->m_pkthdr
.len
;
2073 /* if there's only the packet-header and we can pad there, use it. */
2074 if (m
->m_pkthdr
.len
== m
->m_len
&& M_TRAILINGSPACE(m
) >= padlen
) {
2078 * Walk packet chain to find last mbuf. We will either
2079 * pad there, or append a new mbuf and pad it
2081 for (last
= m
; last
->m_next
!= NULL
; last
= last
->m_next
)
2084 /* `last' now points to last in chain. */
2085 if (M_TRAILINGSPACE(last
) < padlen
) {
2088 /* Allocate new empty mbuf, pad it. Compact later. */
2089 MGET(n
, MB_DONTWAIT
, MT_DATA
);
2097 KKASSERT(M_TRAILINGSPACE(last
) >= padlen
);
2098 KKASSERT(M_WRITABLE(last
));
2100 /* Now zero the pad area */
2101 bzero(mtod(last
, char *) + last
->m_len
, padlen
);
2102 last
->m_len
+= padlen
;
2103 m
->m_pkthdr
.len
+= padlen
;
2108 * Copy data from a buffer back into the indicated mbuf chain,
2109 * starting "off" bytes from the beginning, extending the mbuf
2110 * chain if necessary.
2113 m_copyback(struct mbuf
*m0
, int off
, int len
, caddr_t cp
)
2116 struct mbuf
*m
= m0
, *n
;
2121 while (off
> (mlen
= m
->m_len
)) {
2124 if (m
->m_next
== NULL
) {
2125 n
= m_getclr(MB_DONTWAIT
, m
->m_type
);
2128 n
->m_len
= min(MLEN
, len
+ off
);
2134 mlen
= min (m
->m_len
- off
, len
);
2135 bcopy(cp
, off
+ mtod(m
, caddr_t
), (unsigned)mlen
);
2143 if (m
->m_next
== NULL
) {
2144 n
= m_get(MB_DONTWAIT
, m
->m_type
);
2147 n
->m_len
= min(MLEN
, len
);
2152 out
: if (((m
= m0
)->m_flags
& M_PKTHDR
) && (m
->m_pkthdr
.len
< totlen
))
2153 m
->m_pkthdr
.len
= totlen
;
2157 * Append the specified data to the indicated mbuf chain,
2158 * Extend the mbuf chain if the new data does not fit in
2161 * Return 1 if able to complete the job; otherwise 0.
2164 m_append(struct mbuf
*m0
, int len
, c_caddr_t cp
)
2167 int remainder
, space
;
2169 for (m
= m0
; m
->m_next
!= NULL
; m
= m
->m_next
)
2172 space
= M_TRAILINGSPACE(m
);
2175 * Copy into available space.
2177 if (space
> remainder
)
2179 bcopy(cp
, mtod(m
, caddr_t
) + m
->m_len
, space
);
2181 cp
+= space
, remainder
-= space
;
2183 while (remainder
> 0) {
2185 * Allocate a new mbuf; could check space
2186 * and allocate a cluster instead.
2188 n
= m_get(MB_DONTWAIT
, m
->m_type
);
2191 n
->m_len
= min(MLEN
, remainder
);
2192 bcopy(cp
, mtod(n
, caddr_t
), n
->m_len
);
2193 cp
+= n
->m_len
, remainder
-= n
->m_len
;
2197 if (m0
->m_flags
& M_PKTHDR
)
2198 m0
->m_pkthdr
.len
+= len
- remainder
;
2199 return (remainder
== 0);
2203 * Apply function f to the data in an mbuf chain starting "off" bytes from
2204 * the beginning, continuing for "len" bytes.
2207 m_apply(struct mbuf
*m
, int off
, int len
,
2208 int (*f
)(void *, void *, u_int
), void *arg
)
2213 KASSERT(off
>= 0, ("m_apply, negative off %d", off
));
2214 KASSERT(len
>= 0, ("m_apply, negative len %d", len
));
2216 KASSERT(m
!= NULL
, ("m_apply, offset > size of mbuf chain"));
2223 KASSERT(m
!= NULL
, ("m_apply, offset > size of mbuf chain"));
2224 count
= min(m
->m_len
- off
, len
);
2225 rval
= (*f
)(arg
, mtod(m
, caddr_t
) + off
, count
);
2236 * Return a pointer to mbuf/offset of location in mbuf chain.
2239 m_getptr(struct mbuf
*m
, int loc
, int *off
)
2243 /* Normal end of search. */
2244 if (m
->m_len
> loc
) {
2249 if (m
->m_next
== NULL
) {
2251 /* Point at the end of valid data. */
2264 m_print(const struct mbuf
*m
)
2267 const struct mbuf
*m2
;
2270 len
= m
->m_pkthdr
.len
;
2272 hexstr
= kmalloc(HEX_NCPYLEN(len
), M_TEMP
, M_ZERO
| M_WAITOK
);
2274 kprintf("%p %s\n", m2
, hexncpy(m2
->m_data
, m2
->m_len
, hexstr
,
2275 HEX_NCPYLEN(m2
->m_len
), "-"));
2279 kfree(hexstr
, M_TEMP
);
2284 * "Move" mbuf pkthdr from "from" to "to".
2285 * "from" must have M_PKTHDR set, and "to" must be empty.
2288 m_move_pkthdr(struct mbuf
*to
, struct mbuf
*from
)
2290 KASSERT((to
->m_flags
& M_PKTHDR
), ("m_move_pkthdr: not packet header"));
2292 to
->m_flags
|= from
->m_flags
& M_COPYFLAGS
;
2293 to
->m_pkthdr
= from
->m_pkthdr
; /* especially tags */
2294 SLIST_INIT(&from
->m_pkthdr
.tags
); /* purge tags from src */
2298 * Duplicate "from"'s mbuf pkthdr in "to".
2299 * "from" must have M_PKTHDR set, and "to" must be empty.
2300 * In particular, this does a deep copy of the packet tags.
2303 m_dup_pkthdr(struct mbuf
*to
, const struct mbuf
*from
, int how
)
2305 KASSERT((to
->m_flags
& M_PKTHDR
), ("m_dup_pkthdr: not packet header"));
2307 to
->m_flags
= (from
->m_flags
& M_COPYFLAGS
) |
2308 (to
->m_flags
& ~M_COPYFLAGS
);
2309 to
->m_pkthdr
= from
->m_pkthdr
;
2310 SLIST_INIT(&to
->m_pkthdr
.tags
);
2311 return (m_tag_copy_chain(to
, from
, how
));
2315 * Defragment a mbuf chain, returning the shortest possible
2316 * chain of mbufs and clusters. If allocation fails and
2317 * this cannot be completed, NULL will be returned, but
2318 * the passed in chain will be unchanged. Upon success,
2319 * the original chain will be freed, and the new chain
2322 * If a non-packet header is passed in, the original
2323 * mbuf (chain?) will be returned unharmed.
2325 * m_defrag_nofree doesn't free the passed in mbuf.
2328 m_defrag(struct mbuf
*m0
, int how
)
2332 if ((m_new
= m_defrag_nofree(m0
, how
)) == NULL
)
2340 m_defrag_nofree(struct mbuf
*m0
, int how
)
2342 struct mbuf
*m_new
= NULL
, *m_final
= NULL
;
2343 int progress
= 0, length
, nsize
;
2345 if (!(m0
->m_flags
& M_PKTHDR
))
2348 #ifdef MBUF_STRESS_TEST
2349 if (m_defragrandomfailures
) {
2350 int temp
= karc4random() & 0xff;
2356 m_final
= m_getl(m0
->m_pkthdr
.len
, how
, MT_DATA
, M_PKTHDR
, &nsize
);
2357 if (m_final
== NULL
)
2359 m_final
->m_len
= 0; /* in case m0->m_pkthdr.len is zero */
2361 if (m_dup_pkthdr(m_final
, m0
, how
) == 0)
2366 while (progress
< m0
->m_pkthdr
.len
) {
2367 length
= m0
->m_pkthdr
.len
- progress
;
2368 if (length
> MCLBYTES
)
2371 if (m_new
== NULL
) {
2372 m_new
= m_getl(length
, how
, MT_DATA
, 0, &nsize
);
2377 m_copydata(m0
, progress
, length
, mtod(m_new
, caddr_t
));
2379 m_new
->m_len
= length
;
2380 if (m_new
!= m_final
)
2381 m_cat(m_final
, m_new
);
2384 if (m0
->m_next
== NULL
)
2387 m_defragbytes
+= m_final
->m_pkthdr
.len
;
2398 * Move data from uio into mbufs.
2401 m_uiomove(struct uio
*uio
)
2403 struct mbuf
*m
; /* current working mbuf */
2404 struct mbuf
*head
= NULL
; /* result mbuf chain */
2405 struct mbuf
**mp
= &head
;
2406 int flags
= M_PKTHDR
;
2412 if (uio
->uio_resid
> INT_MAX
)
2415 resid
= (int)uio
->uio_resid
;
2416 m
= m_getl(resid
, MB_WAIT
, MT_DATA
, flags
, &nsize
);
2418 m
->m_pkthdr
.len
= 0;
2419 /* Leave room for protocol headers. */
2424 m
->m_len
= imin(nsize
, resid
);
2425 error
= uiomove(mtod(m
, caddr_t
), m
->m_len
, uio
);
2432 head
->m_pkthdr
.len
+= m
->m_len
;
2433 } while (uio
->uio_resid
> 0);
2443 m_last(struct mbuf
*m
)
2451 * Return the number of bytes in an mbuf chain.
2452 * If lastm is not NULL, also return the last mbuf.
2455 m_lengthm(struct mbuf
*m
, struct mbuf
**lastm
)
2458 struct mbuf
*prev
= m
;
2471 * Like m_lengthm(), except also keep track of mbuf usage.
2474 m_countm(struct mbuf
*m
, struct mbuf
**lastm
, u_int
*pmbcnt
)
2476 u_int len
= 0, mbcnt
= 0;
2477 struct mbuf
*prev
= m
;
2482 if (m
->m_flags
& M_EXT
)
2483 mbcnt
+= m
->m_ext
.ext_size
;