2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1988, 1991, 1993
5 * The Regents of the University of California. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. Neither the name of the University nor the names of its contributors
16 * may be used to endorse or promote products derived from this software
17 * without specific prior written permission.
19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * @(#)uipc_mbuf.c 8.2 (Berkeley) 1/4/94
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
37 #include "opt_param.h"
38 #include "opt_mbuf_stress_test.h"
39 #include "opt_mbuf_profiling.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/limits.h>
46 #include <sys/malloc.h>
48 #include <sys/sysctl.h>
49 #include <sys/domain.h>
50 #include <sys/protosw.h>
54 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__init,
55 "struct mbuf *", "mbufinfo_t *",
56 "uint32_t", "uint32_t",
57 "uint16_t", "uint16_t",
58 "uint32_t", "uint32_t",
59 "uint32_t", "uint32_t");
61 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__gethdr,
62 "uint32_t", "uint32_t",
63 "uint16_t", "uint16_t",
64 "struct mbuf *", "mbufinfo_t *");
66 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__get,
67 "uint32_t", "uint32_t",
68 "uint16_t", "uint16_t",
69 "struct mbuf *", "mbufinfo_t *");
71 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__getcl,
72 "uint32_t", "uint32_t",
73 "uint16_t", "uint16_t",
74 "uint32_t", "uint32_t",
75 "struct mbuf *", "mbufinfo_t *");
77 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__getjcl,
78 "uint32_t", "uint32_t",
79 "uint16_t", "uint16_t",
80 "uint32_t", "uint32_t",
81 "uint32_t", "uint32_t",
82 "struct mbuf *", "mbufinfo_t *");
84 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__clget,
85 "struct mbuf *", "mbufinfo_t *",
86 "uint32_t", "uint32_t",
87 "uint32_t", "uint32_t");
89 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__cljget,
90 "struct mbuf *", "mbufinfo_t *",
91 "uint32_t", "uint32_t",
92 "uint32_t", "uint32_t",
95 SDT_PROBE_DEFINE(sdt, , , m__cljset);
97 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__free,
98 "struct mbuf *", "mbufinfo_t *");
100 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__freem,
101 "struct mbuf *", "mbufinfo_t *");
103 #include <security/mac/mac_framework.h>
109 #ifdef MBUF_STRESS_TEST
114 int m_defragrandomfailures;
118 * sysctl(8) exported objects
120 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RD,
121 &max_linkhdr, 0, "Size of largest link layer header");
122 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RD,
123 &max_protohdr, 0, "Size of largest protocol layer header");
124 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RD,
125 &max_hdr, 0, "Size of largest link plus protocol header");
126 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RD,
127 &max_datalen, 0, "Minimum space left in mbuf after max_hdr");
128 #ifdef MBUF_STRESS_TEST
129 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
130 &m_defragpackets, 0, "");
131 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
132 &m_defragbytes, 0, "");
133 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
134 &m_defraguseless, 0, "");
135 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
136 &m_defragfailure, 0, "");
137 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
138 &m_defragrandomfailures, 0, "");
142 * Ensure the correct size of various mbuf parameters. It could be off due
143 * to compiler-induced padding and alignment artifacts.
145 CTASSERT(MSIZE - offsetof(struct mbuf, m_dat) == MLEN);
146 CTASSERT(MSIZE - offsetof(struct mbuf, m_pktdat) == MHLEN);
149 * mbuf data storage should be 64-bit aligned regardless of architectural
150 * pointer size; check this is the case with and without a packet header.
152 CTASSERT(offsetof(struct mbuf, m_dat) % 8 == 0);
153 CTASSERT(offsetof(struct mbuf, m_pktdat) % 8 == 0);
156 * While the specific values here don't matter too much (i.e., +/- a few
157 * words), we do want to ensure that changes to these values are carefully
158 * reasoned about and properly documented. This is especially the case as
159 * network-protocol and device-driver modules encode these layouts, and must
160 * be recompiled if the structures change. Check these values at compile time
161 * against the ones documented in comments in mbuf.h.
163 * NB: Possibly they should be documented there via #define's and not just
166 #if defined(__LP64__)
167 CTASSERT(offsetof(struct mbuf, m_dat) == 32);
168 CTASSERT(sizeof(struct pkthdr) == 56);
169 CTASSERT(sizeof(struct m_ext) == 48);
171 CTASSERT(offsetof(struct mbuf, m_dat) == 24);
172 CTASSERT(sizeof(struct pkthdr) == 48);
173 CTASSERT(sizeof(struct m_ext) == 28);
177 * Assert that the queue(3) macros produce code of the same size as an old
178 * plain pointer does.
181 static struct mbuf __used m_assertbuf;
182 CTASSERT(sizeof(m_assertbuf.m_slist) == sizeof(m_assertbuf.m_next));
183 CTASSERT(sizeof(m_assertbuf.m_stailq) == sizeof(m_assertbuf.m_next));
184 CTASSERT(sizeof(m_assertbuf.m_slistpkt) == sizeof(m_assertbuf.m_nextpkt));
185 CTASSERT(sizeof(m_assertbuf.m_stailqpkt) == sizeof(m_assertbuf.m_nextpkt));
189 * Attach the cluster from *m to *n, set up m_ext in *n
190 * and bump the refcount of the cluster.
193 mb_dupcl(struct mbuf *n, struct mbuf *m)
195 volatile u_int *refcnt;
197 KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
198 KASSERT(!(n->m_flags & M_EXT), ("%s: M_EXT set on %p", __func__, n));
201 * Cache access optimization. For most kinds of external
202 * storage we don't need full copy of m_ext, since the
203 * holder of the 'ext_count' is responsible to carry the
204 * free routine and its arguments. Exclusion is EXT_EXTREF,
205 * where 'ext_cnt' doesn't point into mbuf at all.
207 if (m->m_ext.ext_type == EXT_EXTREF)
208 bcopy(&m->m_ext, &n->m_ext, sizeof(struct m_ext));
210 bcopy(&m->m_ext, &n->m_ext, m_ext_copylen);
212 n->m_flags |= m->m_flags & M_RDONLY;
214 /* See if this is the mbuf that holds the embedded refcount. */
215 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
216 refcnt = n->m_ext.ext_cnt = &m->m_ext.ext_count;
217 n->m_ext.ext_flags &= ~EXT_FLAG_EMBREF;
219 KASSERT(m->m_ext.ext_cnt != NULL,
220 ("%s: no refcounting pointer on %p", __func__, m));
221 refcnt = m->m_ext.ext_cnt;
227 atomic_add_int(refcnt, 1);
231 m_demote_pkthdr(struct mbuf *m)
236 m_tag_delete_chain(m, NULL);
237 m->m_flags &= ~M_PKTHDR;
238 bzero(&m->m_pkthdr, sizeof(struct pkthdr));
242 * Clean up mbuf (chain) from any tags and packet headers.
243 * If "all" is set then the first mbuf in the chain will be
247 m_demote(struct mbuf *m0, int all, int flags)
251 for (m = all ? m0 : m0->m_next; m != NULL; m = m->m_next) {
252 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt in m %p, m0 %p",
254 if (m->m_flags & M_PKTHDR)
256 m->m_flags = m->m_flags & (M_EXT | M_RDONLY | M_NOFREE | flags);
261 * Sanity checks on mbuf (chain) for use in KASSERT() and general
263 * Returns 0 or panics when bad and 1 on all tests passed.
264 * Sanitize, 0 to run M_SANITY_ACTION, 1 to garble things so they
268 m_sanity(struct mbuf *m0, int sanitize)
275 #define M_SANITY_ACTION(s) panic("mbuf %p: " s, m)
277 #define M_SANITY_ACTION(s) printf("mbuf %p: " s, m)
280 for (m = m0; m != NULL; m = m->m_next) {
282 * Basic pointer checks. If any of these fails then some
283 * unrelated kernel memory before or after us is trashed.
284 * No way to recover from that.
288 if ((caddr_t)m->m_data < a)
289 M_SANITY_ACTION("m_data outside mbuf data range left");
290 if ((caddr_t)m->m_data > b)
291 M_SANITY_ACTION("m_data outside mbuf data range right");
292 if ((caddr_t)m->m_data + m->m_len > b)
293 M_SANITY_ACTION("m_data + m_len exeeds mbuf space");
295 /* m->m_nextpkt may only be set on first mbuf in chain. */
296 if (m != m0 && m->m_nextpkt != NULL) {
298 m_freem(m->m_nextpkt);
299 m->m_nextpkt = (struct mbuf *)0xDEADC0DE;
301 M_SANITY_ACTION("m->m_nextpkt on in-chain mbuf");
304 /* packet length (not mbuf length!) calculation */
305 if (m0->m_flags & M_PKTHDR)
308 /* m_tags may only be attached to first mbuf in chain. */
309 if (m != m0 && m->m_flags & M_PKTHDR &&
310 !SLIST_EMPTY(&m->m_pkthdr.tags)) {
312 m_tag_delete_chain(m, NULL);
313 /* put in 0xDEADC0DE perhaps? */
315 M_SANITY_ACTION("m_tags on in-chain mbuf");
318 /* M_PKTHDR may only be set on first mbuf in chain */
319 if (m != m0 && m->m_flags & M_PKTHDR) {
321 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
322 m->m_flags &= ~M_PKTHDR;
323 /* put in 0xDEADCODE and leave hdr flag in */
325 M_SANITY_ACTION("M_PKTHDR on in-chain mbuf");
329 if (pktlen && pktlen != m->m_pkthdr.len) {
333 M_SANITY_ACTION("m_pkthdr.len != mbuf chain length");
337 #undef M_SANITY_ACTION
341 * Non-inlined part of m_init().
344 m_pkthdr_init(struct mbuf *m, int how)
349 m->m_data = m->m_pktdat;
350 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
352 /* If the label init fails, fail the alloc */
353 error = mac_mbuf_init(m, how);
362 * "Move" mbuf pkthdr from "from" to "to".
363 * "from" must have M_PKTHDR set, and "to" must be empty.
366 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
370 /* see below for why these are not enabled */
372 /* Note: with MAC, this may not be a good assertion. */
373 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags),
374 ("m_move_pkthdr: to has tags"));
378 * XXXMAC: It could be this should also occur for non-MAC?
380 if (to->m_flags & M_PKTHDR)
381 m_tag_delete_chain(to, NULL);
383 to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT);
384 if ((to->m_flags & M_EXT) == 0)
385 to->m_data = to->m_pktdat;
386 to->m_pkthdr = from->m_pkthdr; /* especially tags */
387 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */
388 from->m_flags &= ~M_PKTHDR;
392 * Duplicate "from"'s mbuf pkthdr in "to".
393 * "from" must have M_PKTHDR set, and "to" must be empty.
394 * In particular, this does a deep copy of the packet tags.
397 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
402 * The mbuf allocator only initializes the pkthdr
403 * when the mbuf is allocated with m_gethdr(). Many users
404 * (e.g. m_copy*, m_prepend) use m_get() and then
405 * smash the pkthdr as needed causing these
406 * assertions to trip. For now just disable them.
409 /* Note: with MAC, this may not be a good assertion. */
410 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags), ("m_dup_pkthdr: to has tags"));
412 MBUF_CHECKSLEEP(how);
414 if (to->m_flags & M_PKTHDR)
415 m_tag_delete_chain(to, NULL);
417 to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT);
418 if ((to->m_flags & M_EXT) == 0)
419 to->m_data = to->m_pktdat;
420 to->m_pkthdr = from->m_pkthdr;
421 SLIST_INIT(&to->m_pkthdr.tags);
422 return (m_tag_copy_chain(to, from, how));
426 * Lesser-used path for M_PREPEND:
427 * allocate new mbuf to prepend to chain,
431 m_prepend(struct mbuf *m, int len, int how)
435 if (m->m_flags & M_PKTHDR)
436 mn = m_gethdr(how, m->m_type);
438 mn = m_get(how, m->m_type);
443 if (m->m_flags & M_PKTHDR)
444 m_move_pkthdr(mn, m);
454 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
455 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
456 * The wait parameter is a choice of M_WAITOK/M_NOWAIT from caller.
457 * Note that the copy is read-only, because clusters are not copied,
458 * only their reference counts are incremented.
461 m_copym(struct mbuf *m, int off0, int len, int wait)
463 struct mbuf *n, **np;
468 KASSERT(off >= 0, ("m_copym, negative off %d", off));
469 KASSERT(len >= 0, ("m_copym, negative len %d", len));
470 MBUF_CHECKSLEEP(wait);
471 if (off == 0 && m->m_flags & M_PKTHDR)
474 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
484 KASSERT(len == M_COPYALL,
485 ("m_copym, length > size of mbuf chain"));
489 n = m_gethdr(wait, m->m_type);
491 n = m_get(wait, m->m_type);
496 if (!m_dup_pkthdr(n, m, wait))
498 if (len == M_COPYALL)
499 n->m_pkthdr.len -= off0;
501 n->m_pkthdr.len = len;
504 n->m_len = min(len, m->m_len - off);
505 if (m->m_flags & M_EXT) {
506 n->m_data = m->m_data + off;
509 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
511 if (len != M_COPYALL)
525 * Copy an entire packet, including header (which must be present).
526 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
527 * Note that the copy is read-only, because clusters are not copied,
528 * only their reference counts are incremented.
529 * Preserve alignment of the first mbuf so if the creator has left
530 * some room at the beginning (e.g. for inserting protocol headers)
531 * the copies still have the room available.
534 m_copypacket(struct mbuf *m, int how)
536 struct mbuf *top, *n, *o;
538 MBUF_CHECKSLEEP(how);
539 n = m_get(how, m->m_type);
544 if (!m_dup_pkthdr(n, m, how))
547 if (m->m_flags & M_EXT) {
548 n->m_data = m->m_data;
551 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
552 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
557 o = m_get(how, m->m_type);
565 if (m->m_flags & M_EXT) {
566 n->m_data = m->m_data;
569 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
581 * Copy data from an mbuf chain starting "off" bytes from the beginning,
582 * continuing for "len" bytes, into the indicated buffer.
585 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
589 KASSERT(off >= 0, ("m_copydata, negative off %d", off));
590 KASSERT(len >= 0, ("m_copydata, negative len %d", len));
592 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
599 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
600 count = min(m->m_len - off, len);
601 bcopy(mtod(m, caddr_t) + off, cp, count);
610 * Copy a packet header mbuf chain into a completely new chain, including
611 * copying any mbuf clusters. Use this instead of m_copypacket() when
612 * you need a writable copy of an mbuf chain.
615 m_dup(const struct mbuf *m, int how)
617 struct mbuf **p, *top = NULL;
618 int remain, moff, nsize;
620 MBUF_CHECKSLEEP(how);
626 /* While there's more data, get a new mbuf, tack it on, and fill it */
627 remain = m->m_pkthdr.len;
630 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */
633 /* Get the next new mbuf */
634 if (remain >= MINCLSIZE) {
635 n = m_getcl(how, m->m_type, 0);
638 n = m_get(how, m->m_type);
644 if (top == NULL) { /* First one, must be PKTHDR */
645 if (!m_dup_pkthdr(n, m, how)) {
649 if ((n->m_flags & M_EXT) == 0)
651 n->m_flags &= ~M_RDONLY;
655 /* Link it into the new chain */
659 /* Copy data from original mbuf(s) into new mbuf */
660 while (n->m_len < nsize && m != NULL) {
661 int chunk = min(nsize - n->m_len, m->m_len - moff);
663 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
667 if (moff == m->m_len) {
673 /* Check correct total mbuf length */
674 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
675 ("%s: bogus m_pkthdr.len", __func__));
685 * Concatenate mbuf chain n to m.
686 * Both chains must be of the same type (e.g. MT_DATA).
687 * Any m_pkthdr is not updated.
690 m_cat(struct mbuf *m, struct mbuf *n)
695 if (!M_WRITABLE(m) ||
696 M_TRAILINGSPACE(m) < n->m_len) {
697 /* just join the two chains */
701 /* splat the data from one into the other */
702 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
704 m->m_len += n->m_len;
710 * Concatenate two pkthdr mbuf chains.
713 m_catpkt(struct mbuf *m, struct mbuf *n)
719 m->m_pkthdr.len += n->m_pkthdr.len;
726 m_adj(struct mbuf *mp, int req_len)
732 if ((m = mp) == NULL)
738 while (m != NULL && len > 0) {
739 if (m->m_len <= len) {
749 if (mp->m_flags & M_PKTHDR)
750 mp->m_pkthdr.len -= (req_len - len);
753 * Trim from tail. Scan the mbuf chain,
754 * calculating its length and finding the last mbuf.
755 * If the adjustment only affects this mbuf, then just
756 * adjust and return. Otherwise, rescan and truncate
757 * after the remaining size.
763 if (m->m_next == (struct mbuf *)0)
767 if (m->m_len >= len) {
769 if (mp->m_flags & M_PKTHDR)
770 mp->m_pkthdr.len -= len;
777 * Correct length for chain is "count".
778 * Find the mbuf with last data, adjust its length,
779 * and toss data from remaining mbufs on chain.
782 if (m->m_flags & M_PKTHDR)
783 m->m_pkthdr.len = count;
784 for (; m; m = m->m_next) {
785 if (m->m_len >= count) {
787 if (m->m_next != NULL) {
799 * Rearange an mbuf chain so that len bytes are contiguous
800 * and in the data area of an mbuf (so that mtod will work
801 * for a structure of size len). Returns the resulting
802 * mbuf chain on success, frees it and returns null on failure.
803 * If there is room, it will add up to max_protohdr-len extra bytes to the
804 * contiguous region in an attempt to avoid being called next time.
807 m_pullup(struct mbuf *n, int len)
814 * If first mbuf has no cluster, and has room for len bytes
815 * without shifting current data, pullup into it,
816 * otherwise allocate a new mbuf to prepend to the chain.
818 if ((n->m_flags & M_EXT) == 0 &&
819 n->m_data + len < &n->m_dat[MLEN] && n->m_next) {
828 m = m_get(M_NOWAIT, n->m_type);
831 if (n->m_flags & M_PKTHDR)
834 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
836 count = min(min(max(len, max_protohdr), space), n->m_len);
837 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
847 } while (len > 0 && n);
860 * Like m_pullup(), except a new mbuf is always allocated, and we allow
861 * the amount of empty space before the data in the new mbuf to be specified
862 * (in the event that the caller expects to prepend later).
865 m_copyup(struct mbuf *n, int len, int dstoff)
870 if (len > (MHLEN - dstoff))
872 m = m_get(M_NOWAIT, n->m_type);
875 if (n->m_flags & M_PKTHDR)
878 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
880 count = min(min(max(len, max_protohdr), space), n->m_len);
881 memcpy(mtod(m, caddr_t) + m->m_len, mtod(n, caddr_t),
891 } while (len > 0 && n);
904 * Partition an mbuf chain in two pieces, returning the tail --
905 * all but the first len0 bytes. In case of failure, it returns NULL and
906 * attempts to restore the chain to its original state.
908 * Note that the resulting mbufs might be read-only, because the new
909 * mbuf can end up sharing an mbuf cluster with the original mbuf if
910 * the "breaking point" happens to lie within a cluster mbuf. Use the
911 * M_WRITABLE() macro to check for this case.
914 m_split(struct mbuf *m0, int len0, int wait)
917 u_int len = len0, remain;
919 MBUF_CHECKSLEEP(wait);
920 for (m = m0; m && len > m->m_len; m = m->m_next)
924 remain = m->m_len - len;
925 if (m0->m_flags & M_PKTHDR && remain == 0) {
926 n = m_gethdr(wait, m0->m_type);
929 n->m_next = m->m_next;
931 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
932 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
933 m0->m_pkthdr.len = len0;
935 } else if (m0->m_flags & M_PKTHDR) {
936 n = m_gethdr(wait, m0->m_type);
939 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
940 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
941 m0->m_pkthdr.len = len0;
942 if (m->m_flags & M_EXT)
944 if (remain > MHLEN) {
945 /* m can't be the lead packet */
947 n->m_next = m_split(m, len, wait);
948 if (n->m_next == NULL) {
957 } else if (remain == 0) {
962 n = m_get(wait, m->m_type);
968 if (m->m_flags & M_EXT) {
969 n->m_data = m->m_data + len;
972 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
976 n->m_next = m->m_next;
981 * Routine to copy from device local memory into mbufs.
982 * Note that `off' argument is offset into first mbuf of target chain from
983 * which to begin copying the data to.
986 m_devget(char *buf, int totlen, int off, struct ifnet *ifp,
987 void (*copy)(char *from, caddr_t to, u_int len))
990 struct mbuf *top = NULL, **mp = ⊤
993 if (off < 0 || off > MHLEN)
997 if (top == NULL) { /* First one, must be PKTHDR */
998 if (totlen + off >= MINCLSIZE) {
999 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1002 m = m_gethdr(M_NOWAIT, MT_DATA);
1005 /* Place initial small packet/header at end of mbuf */
1006 if (m && totlen + off + max_linkhdr <= MHLEN) {
1007 m->m_data += max_linkhdr;
1013 m->m_pkthdr.rcvif = ifp;
1014 m->m_pkthdr.len = totlen;
1016 if (totlen + off >= MINCLSIZE) {
1017 m = m_getcl(M_NOWAIT, MT_DATA, 0);
1020 m = m_get(M_NOWAIT, MT_DATA);
1033 m->m_len = len = min(totlen, len);
1035 copy(buf, mtod(m, caddr_t), (u_int)len);
1037 bcopy(buf, mtod(m, caddr_t), (u_int)len);
1047 * Copy data from a buffer back into the indicated mbuf chain,
1048 * starting "off" bytes from the beginning, extending the mbuf
1049 * chain if necessary.
1052 m_copyback(struct mbuf *m0, int off, int len, c_caddr_t cp)
1055 struct mbuf *m = m0, *n;
1060 while (off > (mlen = m->m_len)) {
1063 if (m->m_next == NULL) {
1064 n = m_get(M_NOWAIT, m->m_type);
1067 bzero(mtod(n, caddr_t), MLEN);
1068 n->m_len = min(MLEN, len + off);
1074 if (m->m_next == NULL && (len > m->m_len - off)) {
1075 m->m_len += min(len - (m->m_len - off),
1076 M_TRAILINGSPACE(m));
1078 mlen = min (m->m_len - off, len);
1079 bcopy(cp, off + mtod(m, caddr_t), (u_int)mlen);
1087 if (m->m_next == NULL) {
1088 n = m_get(M_NOWAIT, m->m_type);
1091 n->m_len = min(MLEN, len);
1096 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1097 m->m_pkthdr.len = totlen;
1101 * Append the specified data to the indicated mbuf chain,
1102 * Extend the mbuf chain if the new data does not fit in
1105 * Return 1 if able to complete the job; otherwise 0.
1108 m_append(struct mbuf *m0, int len, c_caddr_t cp)
1111 int remainder, space;
1113 for (m = m0; m->m_next != NULL; m = m->m_next)
1116 space = M_TRAILINGSPACE(m);
1119 * Copy into available space.
1121 if (space > remainder)
1123 bcopy(cp, mtod(m, caddr_t) + m->m_len, space);
1125 cp += space, remainder -= space;
1127 while (remainder > 0) {
1129 * Allocate a new mbuf; could check space
1130 * and allocate a cluster instead.
1132 n = m_get(M_NOWAIT, m->m_type);
1135 n->m_len = min(MLEN, remainder);
1136 bcopy(cp, mtod(n, caddr_t), n->m_len);
1137 cp += n->m_len, remainder -= n->m_len;
1141 if (m0->m_flags & M_PKTHDR)
1142 m0->m_pkthdr.len += len - remainder;
1143 return (remainder == 0);
1147 * Apply function f to the data in an mbuf chain starting "off" bytes from
1148 * the beginning, continuing for "len" bytes.
1151 m_apply(struct mbuf *m, int off, int len,
1152 int (*f)(void *, void *, u_int), void *arg)
1157 KASSERT(off >= 0, ("m_apply, negative off %d", off));
1158 KASSERT(len >= 0, ("m_apply, negative len %d", len));
1160 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1167 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1168 count = min(m->m_len - off, len);
1169 rval = (*f)(arg, mtod(m, caddr_t) + off, count);
1180 * Return a pointer to mbuf/offset of location in mbuf chain.
1183 m_getptr(struct mbuf *m, int loc, int *off)
1187 /* Normal end of search. */
1188 if (m->m_len > loc) {
1193 if (m->m_next == NULL) {
1195 /* Point at the end of valid data. */
1208 m_print(const struct mbuf *m, int maxlen)
1212 const struct mbuf *m2;
1215 printf("mbuf: %p\n", m);
1219 if (m->m_flags & M_PKTHDR)
1220 len = m->m_pkthdr.len;
1224 while (m2 != NULL && (len == -1 || len)) {
1226 if (maxlen != -1 && pdata > maxlen)
1228 printf("mbuf: %p len: %d, next: %p, %b%s", m2, m2->m_len,
1229 m2->m_next, m2->m_flags, "\20\20freelist\17skipfw"
1230 "\11proto5\10proto4\7proto3\6proto2\5proto1\4rdonly"
1231 "\3eor\2pkthdr\1ext", pdata ? "" : "\n");
1233 printf(", %*D\n", pdata, (u_char *)m2->m_data, "-");
1239 printf("%d bytes unaccounted for.\n", len);
1244 m_fixhdr(struct mbuf *m0)
1248 len = m_length(m0, NULL);
1249 m0->m_pkthdr.len = len;
1254 m_length(struct mbuf *m0, struct mbuf **last)
1260 for (m = m0; m != NULL; m = m->m_next) {
1262 if (m->m_next == NULL)
1271 * Defragment a mbuf chain, returning the shortest possible
1272 * chain of mbufs and clusters. If allocation fails and
1273 * this cannot be completed, NULL will be returned, but
1274 * the passed in chain will be unchanged. Upon success,
1275 * the original chain will be freed, and the new chain
1278 * If a non-packet header is passed in, the original
1279 * mbuf (chain?) will be returned unharmed.
1282 m_defrag(struct mbuf *m0, int how)
1284 struct mbuf *m_new = NULL, *m_final = NULL;
1285 int progress = 0, length;
1287 MBUF_CHECKSLEEP(how);
1288 if (!(m0->m_flags & M_PKTHDR))
1291 m_fixhdr(m0); /* Needed sanity check */
1293 #ifdef MBUF_STRESS_TEST
1294 if (m_defragrandomfailures) {
1295 int temp = arc4random() & 0xff;
1301 if (m0->m_pkthdr.len > MHLEN)
1302 m_final = m_getcl(how, MT_DATA, M_PKTHDR);
1304 m_final = m_gethdr(how, MT_DATA);
1306 if (m_final == NULL)
1309 if (m_dup_pkthdr(m_final, m0, how) == 0)
1314 while (progress < m0->m_pkthdr.len) {
1315 length = m0->m_pkthdr.len - progress;
1316 if (length > MCLBYTES)
1319 if (m_new == NULL) {
1321 m_new = m_getcl(how, MT_DATA, 0);
1323 m_new = m_get(how, MT_DATA);
1328 m_copydata(m0, progress, length, mtod(m_new, caddr_t));
1330 m_new->m_len = length;
1331 if (m_new != m_final)
1332 m_cat(m_final, m_new);
1335 #ifdef MBUF_STRESS_TEST
1336 if (m0->m_next == NULL)
1341 #ifdef MBUF_STRESS_TEST
1343 m_defragbytes += m0->m_pkthdr.len;
1347 #ifdef MBUF_STRESS_TEST
1356 * Defragment an mbuf chain, returning at most maxfrags separate
1357 * mbufs+clusters. If this is not possible NULL is returned and
1358 * the original mbuf chain is left in its present (potentially
1359 * modified) state. We use two techniques: collapsing consecutive
1360 * mbufs and replacing consecutive mbufs by a cluster.
1362 * NB: this should really be named m_defrag but that name is taken
1365 m_collapse(struct mbuf *m0, int how, int maxfrags)
1367 struct mbuf *m, *n, *n2, **prev;
1371 * Calculate the current number of frags.
1374 for (m = m0; m != NULL; m = m->m_next)
1377 * First, try to collapse mbufs. Note that we always collapse
1378 * towards the front so we don't need to deal with moving the
1379 * pkthdr. This may be suboptimal if the first mbuf has much
1380 * less data than the following.
1388 if (M_WRITABLE(m) &&
1389 n->m_len < M_TRAILINGSPACE(m)) {
1390 bcopy(mtod(n, void *), mtod(m, char *) + m->m_len,
1392 m->m_len += n->m_len;
1393 m->m_next = n->m_next;
1395 if (--curfrags <= maxfrags)
1400 KASSERT(maxfrags > 1,
1401 ("maxfrags %u, but normal collapse failed", maxfrags));
1403 * Collapse consecutive mbufs to a cluster.
1405 prev = &m0->m_next; /* NB: not the first mbuf */
1406 while ((n = *prev) != NULL) {
1407 if ((n2 = n->m_next) != NULL &&
1408 n->m_len + n2->m_len < MCLBYTES) {
1409 m = m_getcl(how, MT_DATA, 0);
1412 bcopy(mtod(n, void *), mtod(m, void *), n->m_len);
1413 bcopy(mtod(n2, void *), mtod(m, char *) + n->m_len,
1415 m->m_len = n->m_len + n2->m_len;
1416 m->m_next = n2->m_next;
1420 if (--curfrags <= maxfrags) /* +1 cl -2 mbufs */
1423 * Still not there, try the normal collapse
1424 * again before we allocate another cluster.
1431 * No place where we can collapse to a cluster; punt.
1432 * This can occur if, for example, you request 2 frags
1433 * but the packet requires that both be clusters (we
1434 * never reallocate the first mbuf to avoid moving the
1441 #ifdef MBUF_STRESS_TEST
1444 * Fragment an mbuf chain. There's no reason you'd ever want to do
1445 * this in normal usage, but it's great for stress testing various
1448 * If fragmentation is not possible, the original chain will be
1451 * Possible length values:
1452 * 0 no fragmentation will occur
1453 * > 0 each fragment will be of the specified length
1454 * -1 each fragment will be the same random value in length
1455 * -2 each fragment's length will be entirely random
1456 * (Random values range from 1 to 256)
1459 m_fragment(struct mbuf *m0, int how, int length)
1461 struct mbuf *m_first, *m_last;
1462 int divisor = 255, progress = 0, fraglen;
1464 if (!(m0->m_flags & M_PKTHDR))
1467 if (length == 0 || length < -2)
1469 if (length > MCLBYTES)
1471 if (length < 0 && divisor > MCLBYTES)
1474 length = 1 + (arc4random() % divisor);
1478 m_fixhdr(m0); /* Needed sanity check */
1480 m_first = m_getcl(how, MT_DATA, M_PKTHDR);
1481 if (m_first == NULL)
1484 if (m_dup_pkthdr(m_first, m0, how) == 0)
1489 while (progress < m0->m_pkthdr.len) {
1491 fraglen = 1 + (arc4random() % divisor);
1492 if (fraglen > m0->m_pkthdr.len - progress)
1493 fraglen = m0->m_pkthdr.len - progress;
1495 if (progress != 0) {
1496 struct mbuf *m_new = m_getcl(how, MT_DATA, 0);
1500 m_last->m_next = m_new;
1504 m_copydata(m0, progress, fraglen, mtod(m_last, caddr_t));
1505 progress += fraglen;
1506 m_last->m_len = fraglen;
1514 /* Return the original chain on failure */
1521 * Copy the contents of uio into a properly sized mbuf chain.
1524 m_uiotombuf(struct uio *uio, int how, int len, int align, int flags)
1526 struct mbuf *m, *mb;
1532 * len can be zero or an arbitrary large value bound by
1533 * the total data supplied by the uio.
1536 total = (uio->uio_resid < len) ? uio->uio_resid : len;
1538 total = uio->uio_resid;
1541 * The smallest unit returned by m_getm2() is a single mbuf
1542 * with pkthdr. We can't align past it.
1548 * Give us the full allocation or nothing.
1549 * If len is zero return the smallest empty mbuf.
1551 m = m_getm2(NULL, max(total + align, 1), how, MT_DATA, flags);
1556 /* Fill all mbufs with uio data and update header information. */
1557 for (mb = m; mb != NULL; mb = mb->m_next) {
1558 length = min(M_TRAILINGSPACE(mb), total - progress);
1560 error = uiomove(mtod(mb, void *), length, uio);
1568 if (flags & M_PKTHDR)
1569 m->m_pkthdr.len += length;
1571 KASSERT(progress == total, ("%s: progress != total", __func__));
1577 * Copy an mbuf chain into a uio limited by len if set.
1580 m_mbuftouio(struct uio *uio, const struct mbuf *m, int len)
1582 int error, length, total;
1586 total = min(uio->uio_resid, len);
1588 total = uio->uio_resid;
1590 /* Fill the uio with data from the mbufs. */
1591 for (; m != NULL; m = m->m_next) {
1592 length = min(m->m_len, total - progress);
1594 error = uiomove(mtod(m, void *), length, uio);
1605 * Create a writable copy of the mbuf chain. While doing this
1606 * we compact the chain with a goal of producing a chain with
1607 * at most two mbufs. The second mbuf in this chain is likely
1608 * to be a cluster. The primary purpose of this work is to create
1609 * a writable packet for encryption, compression, etc. The
1610 * secondary goal is to linearize the data so the data can be
1611 * passed to crypto hardware in the most efficient manner possible.
1614 m_unshare(struct mbuf *m0, int how)
1616 struct mbuf *m, *mprev;
1617 struct mbuf *n, *mfirst, *mlast;
1621 for (m = m0; m != NULL; m = mprev->m_next) {
1623 * Regular mbufs are ignored unless there's a cluster
1624 * in front of it that we can use to coalesce. We do
1625 * the latter mainly so later clusters can be coalesced
1626 * also w/o having to handle them specially (i.e. convert
1627 * mbuf+cluster -> cluster). This optimization is heavily
1628 * influenced by the assumption that we're running over
1629 * Ethernet where MCLBYTES is large enough that the max
1630 * packet size will permit lots of coalescing into a
1631 * single cluster. This in turn permits efficient
1632 * crypto operations, especially when using hardware.
1634 if ((m->m_flags & M_EXT) == 0) {
1635 if (mprev && (mprev->m_flags & M_EXT) &&
1636 m->m_len <= M_TRAILINGSPACE(mprev)) {
1637 /* XXX: this ignores mbuf types */
1638 memcpy(mtod(mprev, caddr_t) + mprev->m_len,
1639 mtod(m, caddr_t), m->m_len);
1640 mprev->m_len += m->m_len;
1641 mprev->m_next = m->m_next; /* unlink from chain */
1642 m_free(m); /* reclaim mbuf */
1649 * Writable mbufs are left alone (for now).
1651 if (M_WRITABLE(m)) {
1657 * Not writable, replace with a copy or coalesce with
1658 * the previous mbuf if possible (since we have to copy
1659 * it anyway, we try to reduce the number of mbufs and
1660 * clusters so that future work is easier).
1662 KASSERT(m->m_flags & M_EXT, ("m_flags 0x%x", m->m_flags));
1663 /* NB: we only coalesce into a cluster or larger */
1664 if (mprev != NULL && (mprev->m_flags & M_EXT) &&
1665 m->m_len <= M_TRAILINGSPACE(mprev)) {
1666 /* XXX: this ignores mbuf types */
1667 memcpy(mtod(mprev, caddr_t) + mprev->m_len,
1668 mtod(m, caddr_t), m->m_len);
1669 mprev->m_len += m->m_len;
1670 mprev->m_next = m->m_next; /* unlink from chain */
1671 m_free(m); /* reclaim mbuf */
1676 * Allocate new space to hold the copy and copy the data.
1677 * We deal with jumbo mbufs (i.e. m_len > MCLBYTES) by
1678 * splitting them into clusters. We could just malloc a
1679 * buffer and make it external but too many device drivers
1680 * don't know how to break up the non-contiguous memory when
1683 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS);
1688 if (m->m_flags & M_PKTHDR) {
1689 KASSERT(mprev == NULL, ("%s: m0 %p, m %p has M_PKTHDR",
1691 m_move_pkthdr(n, m);
1698 int cc = min(len, MCLBYTES);
1699 memcpy(mtod(n, caddr_t), mtod(m, caddr_t) + off, cc);
1705 newipsecstat.ips_clcopied++;
1713 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS);
1720 n->m_next = m->m_next;
1722 m0 = mfirst; /* new head of chain */
1724 mprev->m_next = mfirst; /* replace old mbuf */
1725 m_free(m); /* release old mbuf */
1731 #ifdef MBUF_PROFILING
1733 #define MP_BUCKETS 32 /* don't just change this as things may overflow.*/
1734 struct mbufprofile {
1735 uintmax_t wasted[MP_BUCKETS];
1736 uintmax_t used[MP_BUCKETS];
1737 uintmax_t segments[MP_BUCKETS];
1740 #define MP_MAXDIGITS 21 /* strlen("16,000,000,000,000,000,000") == 21 */
1741 #define MP_NUMLINES 6
1742 #define MP_NUMSPERLINE 16
1743 #define MP_EXTRABYTES 64 /* > strlen("used:\nwasted:\nsegments:\n") */
1744 /* work out max space needed and add a bit of spare space too */
1745 #define MP_MAXLINE ((MP_MAXDIGITS+1) * MP_NUMSPERLINE)
1746 #define MP_BUFSIZE ((MP_MAXLINE * MP_NUMLINES) + 1 + MP_EXTRABYTES)
1748 char mbprofbuf[MP_BUFSIZE];
1751 m_profile(struct mbuf *m)
1760 if (m->m_flags & M_EXT) {
1761 wasted += MHLEN - sizeof(m->m_ext) +
1762 m->m_ext.ext_size - m->m_len;
1764 if (m->m_flags & M_PKTHDR)
1765 wasted += MHLEN - m->m_len;
1767 wasted += MLEN - m->m_len;
1771 /* be paranoid.. it helps */
1772 if (segments > MP_BUCKETS - 1)
1773 segments = MP_BUCKETS - 1;
1776 if (wasted > 100000)
1778 /* store in the appropriate bucket */
1779 /* don't bother locking. if it's slightly off, so what? */
1780 mbprof.segments[segments]++;
1781 mbprof.used[fls(used)]++;
1782 mbprof.wasted[fls(wasted)]++;
1786 mbprof_textify(void)
1792 p = &mbprof.wasted[0];
1794 offset = snprintf(c, MP_MAXLINE + 10,
1796 "%ju %ju %ju %ju %ju %ju %ju %ju "
1797 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1798 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1799 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1801 p = &mbprof.wasted[16];
1803 offset = snprintf(c, MP_MAXLINE,
1804 "%ju %ju %ju %ju %ju %ju %ju %ju "
1805 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1806 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1807 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1809 p = &mbprof.used[0];
1811 offset = snprintf(c, MP_MAXLINE + 10,
1813 "%ju %ju %ju %ju %ju %ju %ju %ju "
1814 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1815 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1816 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1818 p = &mbprof.used[16];
1820 offset = snprintf(c, MP_MAXLINE,
1821 "%ju %ju %ju %ju %ju %ju %ju %ju "
1822 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1823 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1824 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1826 p = &mbprof.segments[0];
1828 offset = snprintf(c, MP_MAXLINE + 10,
1830 "%ju %ju %ju %ju %ju %ju %ju %ju "
1831 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1832 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1833 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1835 p = &mbprof.segments[16];
1837 offset = snprintf(c, MP_MAXLINE,
1838 "%ju %ju %ju %ju %ju %ju %ju %ju "
1839 "%ju %ju %ju %ju %ju %ju %ju %jju",
1840 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1841 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1846 mbprof_handler(SYSCTL_HANDLER_ARGS)
1851 error = SYSCTL_OUT(req, mbprofbuf, strlen(mbprofbuf) + 1);
1856 mbprof_clr_handler(SYSCTL_HANDLER_ARGS)
1861 error = sysctl_handle_int(oidp, &clear, 0, req);
1862 if (error || !req->newptr)
1866 bzero(&mbprof, sizeof(mbprof));
1873 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofile, CTLTYPE_STRING|CTLFLAG_RD,
1874 NULL, 0, mbprof_handler, "A", "mbuf profiling statistics");
1876 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofileclr, CTLTYPE_INT|CTLFLAG_RW,
1877 NULL, 0, mbprof_clr_handler, "I", "clear mbuf profiling statistics");