2 * Copyright (c) 1982, 1986, 1988, 1991, 1993
3 * The Regents of the University of California. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 3. Neither the name of the University nor the names of its contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * @(#)uipc_mbuf.c 8.2 (Berkeley) 1/4/94
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
35 #include "opt_param.h"
36 #include "opt_mbuf_stress_test.h"
37 #include "opt_mbuf_profiling.h"
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/limits.h>
44 #include <sys/malloc.h>
46 #include <sys/sysctl.h>
47 #include <sys/domain.h>
48 #include <sys/protosw.h>
52 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__init,
53 "struct mbuf *", "mbufinfo_t *",
54 "uint32_t", "uint32_t",
55 "uint16_t", "uint16_t",
56 "uint32_t", "uint32_t",
57 "uint32_t", "uint32_t");
59 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__gethdr,
60 "uint32_t", "uint32_t",
61 "uint16_t", "uint16_t",
62 "struct mbuf *", "mbufinfo_t *");
64 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__get,
65 "uint32_t", "uint32_t",
66 "uint16_t", "uint16_t",
67 "struct mbuf *", "mbufinfo_t *");
69 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__getcl,
70 "uint32_t", "uint32_t",
71 "uint16_t", "uint16_t",
72 "uint32_t", "uint32_t",
73 "struct mbuf *", "mbufinfo_t *");
75 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__clget,
76 "struct mbuf *", "mbufinfo_t *",
77 "uint32_t", "uint32_t",
78 "uint32_t", "uint32_t");
80 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__cljget,
81 "struct mbuf *", "mbufinfo_t *",
82 "uint32_t", "uint32_t",
83 "uint32_t", "uint32_t",
86 SDT_PROBE_DEFINE(sdt, , , m__cljset);
88 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__free,
89 "struct mbuf *", "mbufinfo_t *");
91 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__freem,
92 "struct mbuf *", "mbufinfo_t *");
94 #include <security/mac/mac_framework.h>
100 #ifdef MBUF_STRESS_TEST
105 int m_defragrandomfailures;
109 * sysctl(8) exported objects
111 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RD,
112 &max_linkhdr, 0, "Size of largest link layer header");
113 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RD,
114 &max_protohdr, 0, "Size of largest protocol layer header");
115 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RD,
116 &max_hdr, 0, "Size of largest link plus protocol header");
117 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RD,
118 &max_datalen, 0, "Minimum space left in mbuf after max_hdr");
119 #ifdef MBUF_STRESS_TEST
120 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
121 &m_defragpackets, 0, "");
122 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
123 &m_defragbytes, 0, "");
124 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
125 &m_defraguseless, 0, "");
126 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
127 &m_defragfailure, 0, "");
128 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
129 &m_defragrandomfailures, 0, "");
133 * Ensure the correct size of various mbuf parameters. It could be off due
134 * to compiler-induced padding and alignment artifacts.
136 CTASSERT(MSIZE - offsetof(struct mbuf, m_dat) == MLEN);
137 CTASSERT(MSIZE - offsetof(struct mbuf, m_pktdat) == MHLEN);
140 * mbuf data storage should be 64-bit aligned regardless of architectural
141 * pointer size; check this is the case with and without a packet header.
143 CTASSERT(offsetof(struct mbuf, m_dat) % 8 == 0);
144 CTASSERT(offsetof(struct mbuf, m_pktdat) % 8 == 0);
147 * While the specific values here don't matter too much (i.e., +/- a few
148 * words), we do want to ensure that changes to these values are carefully
149 * reasoned about and properly documented. This is especially the case as
150 * network-protocol and device-driver modules encode these layouts, and must
151 * be recompiled if the structures change. Check these values at compile time
152 * against the ones documented in comments in mbuf.h.
154 * NB: Possibly they should be documented there via #define's and not just
157 #if defined(__LP64__)
158 CTASSERT(offsetof(struct mbuf, m_dat) == 32);
159 CTASSERT(sizeof(struct pkthdr) == 56);
160 CTASSERT(sizeof(struct m_ext) == 48);
162 CTASSERT(offsetof(struct mbuf, m_dat) == 24);
163 CTASSERT(sizeof(struct pkthdr) == 48);
164 CTASSERT(sizeof(struct m_ext) == 28);
168 * Assert that the queue(3) macros produce code of the same size as an old
169 * plain pointer does.
172 static struct mbuf __used m_assertbuf;
173 CTASSERT(sizeof(m_assertbuf.m_slist) == sizeof(m_assertbuf.m_next));
174 CTASSERT(sizeof(m_assertbuf.m_stailq) == sizeof(m_assertbuf.m_next));
175 CTASSERT(sizeof(m_assertbuf.m_slistpkt) == sizeof(m_assertbuf.m_nextpkt));
176 CTASSERT(sizeof(m_assertbuf.m_stailqpkt) == sizeof(m_assertbuf.m_nextpkt));
180 * Attach the cluster from *m to *n, set up m_ext in *n
181 * and bump the refcount of the cluster.
184 mb_dupcl(struct mbuf *n, struct mbuf *m)
186 volatile u_int *refcnt;
188 KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
189 KASSERT(!(n->m_flags & M_EXT), ("%s: M_EXT set on %p", __func__, n));
192 * Cache access optimization. For most kinds of external
193 * storage we don't need full copy of m_ext, since the
194 * holder of the 'ext_count' is responsible to carry the
195 * free routine and its arguments. Exclusion is EXT_EXTREF,
196 * where 'ext_cnt' doesn't point into mbuf at all.
198 if (m->m_ext.ext_type == EXT_EXTREF)
199 bcopy(&m->m_ext, &n->m_ext, sizeof(struct m_ext));
201 bcopy(&m->m_ext, &n->m_ext, m_ext_copylen);
203 n->m_flags |= m->m_flags & M_RDONLY;
205 /* See if this is the mbuf that holds the embedded refcount. */
206 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
207 refcnt = n->m_ext.ext_cnt = &m->m_ext.ext_count;
208 n->m_ext.ext_flags &= ~EXT_FLAG_EMBREF;
210 KASSERT(m->m_ext.ext_cnt != NULL,
211 ("%s: no refcounting pointer on %p", __func__, m));
212 refcnt = m->m_ext.ext_cnt;
218 atomic_add_int(refcnt, 1);
222 m_demote_pkthdr(struct mbuf *m)
227 m_tag_delete_chain(m, NULL);
228 m->m_flags &= ~M_PKTHDR;
229 bzero(&m->m_pkthdr, sizeof(struct pkthdr));
233 * Clean up mbuf (chain) from any tags and packet headers.
234 * If "all" is set then the first mbuf in the chain will be
238 m_demote(struct mbuf *m0, int all, int flags)
242 for (m = all ? m0 : m0->m_next; m != NULL; m = m->m_next) {
243 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt in m %p, m0 %p",
245 if (m->m_flags & M_PKTHDR)
247 m->m_flags = m->m_flags & (M_EXT | M_RDONLY | M_NOFREE | flags);
252 * Sanity checks on mbuf (chain) for use in KASSERT() and general
254 * Returns 0 or panics when bad and 1 on all tests passed.
255 * Sanitize, 0 to run M_SANITY_ACTION, 1 to garble things so they
259 m_sanity(struct mbuf *m0, int sanitize)
266 #define M_SANITY_ACTION(s) panic("mbuf %p: " s, m)
268 #define M_SANITY_ACTION(s) printf("mbuf %p: " s, m)
271 for (m = m0; m != NULL; m = m->m_next) {
273 * Basic pointer checks. If any of these fails then some
274 * unrelated kernel memory before or after us is trashed.
275 * No way to recover from that.
279 if ((caddr_t)m->m_data < a)
280 M_SANITY_ACTION("m_data outside mbuf data range left");
281 if ((caddr_t)m->m_data > b)
282 M_SANITY_ACTION("m_data outside mbuf data range right");
283 if ((caddr_t)m->m_data + m->m_len > b)
284 M_SANITY_ACTION("m_data + m_len exeeds mbuf space");
286 /* m->m_nextpkt may only be set on first mbuf in chain. */
287 if (m != m0 && m->m_nextpkt != NULL) {
289 m_freem(m->m_nextpkt);
290 m->m_nextpkt = (struct mbuf *)0xDEADC0DE;
292 M_SANITY_ACTION("m->m_nextpkt on in-chain mbuf");
295 /* packet length (not mbuf length!) calculation */
296 if (m0->m_flags & M_PKTHDR)
299 /* m_tags may only be attached to first mbuf in chain. */
300 if (m != m0 && m->m_flags & M_PKTHDR &&
301 !SLIST_EMPTY(&m->m_pkthdr.tags)) {
303 m_tag_delete_chain(m, NULL);
304 /* put in 0xDEADC0DE perhaps? */
306 M_SANITY_ACTION("m_tags on in-chain mbuf");
309 /* M_PKTHDR may only be set on first mbuf in chain */
310 if (m != m0 && m->m_flags & M_PKTHDR) {
312 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
313 m->m_flags &= ~M_PKTHDR;
314 /* put in 0xDEADCODE and leave hdr flag in */
316 M_SANITY_ACTION("M_PKTHDR on in-chain mbuf");
320 if (pktlen && pktlen != m->m_pkthdr.len) {
324 M_SANITY_ACTION("m_pkthdr.len != mbuf chain length");
328 #undef M_SANITY_ACTION
332 * Non-inlined part of m_init().
335 m_pkthdr_init(struct mbuf *m, int how)
340 m->m_data = m->m_pktdat;
341 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
343 /* If the label init fails, fail the alloc */
344 error = mac_mbuf_init(m, how);
353 * "Move" mbuf pkthdr from "from" to "to".
354 * "from" must have M_PKTHDR set, and "to" must be empty.
357 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
361 /* see below for why these are not enabled */
363 /* Note: with MAC, this may not be a good assertion. */
364 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags),
365 ("m_move_pkthdr: to has tags"));
369 * XXXMAC: It could be this should also occur for non-MAC?
371 if (to->m_flags & M_PKTHDR)
372 m_tag_delete_chain(to, NULL);
374 to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT);
375 if ((to->m_flags & M_EXT) == 0)
376 to->m_data = to->m_pktdat;
377 to->m_pkthdr = from->m_pkthdr; /* especially tags */
378 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */
379 from->m_flags &= ~M_PKTHDR;
383 * Duplicate "from"'s mbuf pkthdr in "to".
384 * "from" must have M_PKTHDR set, and "to" must be empty.
385 * In particular, this does a deep copy of the packet tags.
388 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
393 * The mbuf allocator only initializes the pkthdr
394 * when the mbuf is allocated with m_gethdr(). Many users
395 * (e.g. m_copy*, m_prepend) use m_get() and then
396 * smash the pkthdr as needed causing these
397 * assertions to trip. For now just disable them.
400 /* Note: with MAC, this may not be a good assertion. */
401 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags), ("m_dup_pkthdr: to has tags"));
403 MBUF_CHECKSLEEP(how);
405 if (to->m_flags & M_PKTHDR)
406 m_tag_delete_chain(to, NULL);
408 to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT);
409 if ((to->m_flags & M_EXT) == 0)
410 to->m_data = to->m_pktdat;
411 to->m_pkthdr = from->m_pkthdr;
412 SLIST_INIT(&to->m_pkthdr.tags);
413 return (m_tag_copy_chain(to, from, how));
417 * Lesser-used path for M_PREPEND:
418 * allocate new mbuf to prepend to chain,
422 m_prepend(struct mbuf *m, int len, int how)
426 if (m->m_flags & M_PKTHDR)
427 mn = m_gethdr(how, m->m_type);
429 mn = m_get(how, m->m_type);
434 if (m->m_flags & M_PKTHDR)
435 m_move_pkthdr(mn, m);
445 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
446 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
447 * The wait parameter is a choice of M_WAITOK/M_NOWAIT from caller.
448 * Note that the copy is read-only, because clusters are not copied,
449 * only their reference counts are incremented.
452 m_copym(struct mbuf *m, int off0, int len, int wait)
454 struct mbuf *n, **np;
459 KASSERT(off >= 0, ("m_copym, negative off %d", off));
460 KASSERT(len >= 0, ("m_copym, negative len %d", len));
461 MBUF_CHECKSLEEP(wait);
462 if (off == 0 && m->m_flags & M_PKTHDR)
465 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
475 KASSERT(len == M_COPYALL,
476 ("m_copym, length > size of mbuf chain"));
480 n = m_gethdr(wait, m->m_type);
482 n = m_get(wait, m->m_type);
487 if (!m_dup_pkthdr(n, m, wait))
489 if (len == M_COPYALL)
490 n->m_pkthdr.len -= off0;
492 n->m_pkthdr.len = len;
495 n->m_len = min(len, m->m_len - off);
496 if (m->m_flags & M_EXT) {
497 n->m_data = m->m_data + off;
500 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
502 if (len != M_COPYALL)
516 * Copy an entire packet, including header (which must be present).
517 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
518 * Note that the copy is read-only, because clusters are not copied,
519 * only their reference counts are incremented.
520 * Preserve alignment of the first mbuf so if the creator has left
521 * some room at the beginning (e.g. for inserting protocol headers)
522 * the copies still have the room available.
525 m_copypacket(struct mbuf *m, int how)
527 struct mbuf *top, *n, *o;
529 MBUF_CHECKSLEEP(how);
530 n = m_get(how, m->m_type);
535 if (!m_dup_pkthdr(n, m, how))
538 if (m->m_flags & M_EXT) {
539 n->m_data = m->m_data;
542 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
543 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
548 o = m_get(how, m->m_type);
556 if (m->m_flags & M_EXT) {
557 n->m_data = m->m_data;
560 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
572 * Copy data from an mbuf chain starting "off" bytes from the beginning,
573 * continuing for "len" bytes, into the indicated buffer.
576 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
580 KASSERT(off >= 0, ("m_copydata, negative off %d", off));
581 KASSERT(len >= 0, ("m_copydata, negative len %d", len));
583 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
590 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
591 count = min(m->m_len - off, len);
592 bcopy(mtod(m, caddr_t) + off, cp, count);
601 * Copy a packet header mbuf chain into a completely new chain, including
602 * copying any mbuf clusters. Use this instead of m_copypacket() when
603 * you need a writable copy of an mbuf chain.
606 m_dup(const struct mbuf *m, int how)
608 struct mbuf **p, *top = NULL;
609 int remain, moff, nsize;
611 MBUF_CHECKSLEEP(how);
617 /* While there's more data, get a new mbuf, tack it on, and fill it */
618 remain = m->m_pkthdr.len;
621 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */
624 /* Get the next new mbuf */
625 if (remain >= MINCLSIZE) {
626 n = m_getcl(how, m->m_type, 0);
629 n = m_get(how, m->m_type);
635 if (top == NULL) { /* First one, must be PKTHDR */
636 if (!m_dup_pkthdr(n, m, how)) {
640 if ((n->m_flags & M_EXT) == 0)
642 n->m_flags &= ~M_RDONLY;
646 /* Link it into the new chain */
650 /* Copy data from original mbuf(s) into new mbuf */
651 while (n->m_len < nsize && m != NULL) {
652 int chunk = min(nsize - n->m_len, m->m_len - moff);
654 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
658 if (moff == m->m_len) {
664 /* Check correct total mbuf length */
665 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
666 ("%s: bogus m_pkthdr.len", __func__));
676 * Concatenate mbuf chain n to m.
677 * Both chains must be of the same type (e.g. MT_DATA).
678 * Any m_pkthdr is not updated.
681 m_cat(struct mbuf *m, struct mbuf *n)
686 if (!M_WRITABLE(m) ||
687 M_TRAILINGSPACE(m) < n->m_len) {
688 /* just join the two chains */
692 /* splat the data from one into the other */
693 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
695 m->m_len += n->m_len;
701 * Concatenate two pkthdr mbuf chains.
704 m_catpkt(struct mbuf *m, struct mbuf *n)
710 m->m_pkthdr.len += n->m_pkthdr.len;
717 m_adj(struct mbuf *mp, int req_len)
723 if ((m = mp) == NULL)
729 while (m != NULL && len > 0) {
730 if (m->m_len <= len) {
740 if (mp->m_flags & M_PKTHDR)
741 mp->m_pkthdr.len -= (req_len - len);
744 * Trim from tail. Scan the mbuf chain,
745 * calculating its length and finding the last mbuf.
746 * If the adjustment only affects this mbuf, then just
747 * adjust and return. Otherwise, rescan and truncate
748 * after the remaining size.
754 if (m->m_next == (struct mbuf *)0)
758 if (m->m_len >= len) {
760 if (mp->m_flags & M_PKTHDR)
761 mp->m_pkthdr.len -= len;
768 * Correct length for chain is "count".
769 * Find the mbuf with last data, adjust its length,
770 * and toss data from remaining mbufs on chain.
773 if (m->m_flags & M_PKTHDR)
774 m->m_pkthdr.len = count;
775 for (; m; m = m->m_next) {
776 if (m->m_len >= count) {
778 if (m->m_next != NULL) {
790 * Rearange an mbuf chain so that len bytes are contiguous
791 * and in the data area of an mbuf (so that mtod will work
792 * for a structure of size len). Returns the resulting
793 * mbuf chain on success, frees it and returns null on failure.
794 * If there is room, it will add up to max_protohdr-len extra bytes to the
795 * contiguous region in an attempt to avoid being called next time.
798 m_pullup(struct mbuf *n, int len)
805 * If first mbuf has no cluster, and has room for len bytes
806 * without shifting current data, pullup into it,
807 * otherwise allocate a new mbuf to prepend to the chain.
809 if ((n->m_flags & M_EXT) == 0 &&
810 n->m_data + len < &n->m_dat[MLEN] && n->m_next) {
819 m = m_get(M_NOWAIT, n->m_type);
822 if (n->m_flags & M_PKTHDR)
825 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
827 count = min(min(max(len, max_protohdr), space), n->m_len);
828 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
838 } while (len > 0 && n);
851 * Like m_pullup(), except a new mbuf is always allocated, and we allow
852 * the amount of empty space before the data in the new mbuf to be specified
853 * (in the event that the caller expects to prepend later).
856 m_copyup(struct mbuf *n, int len, int dstoff)
861 if (len > (MHLEN - dstoff))
863 m = m_get(M_NOWAIT, n->m_type);
866 if (n->m_flags & M_PKTHDR)
869 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
871 count = min(min(max(len, max_protohdr), space), n->m_len);
872 memcpy(mtod(m, caddr_t) + m->m_len, mtod(n, caddr_t),
882 } while (len > 0 && n);
895 * Partition an mbuf chain in two pieces, returning the tail --
896 * all but the first len0 bytes. In case of failure, it returns NULL and
897 * attempts to restore the chain to its original state.
899 * Note that the resulting mbufs might be read-only, because the new
900 * mbuf can end up sharing an mbuf cluster with the original mbuf if
901 * the "breaking point" happens to lie within a cluster mbuf. Use the
902 * M_WRITABLE() macro to check for this case.
905 m_split(struct mbuf *m0, int len0, int wait)
908 u_int len = len0, remain;
910 MBUF_CHECKSLEEP(wait);
911 for (m = m0; m && len > m->m_len; m = m->m_next)
915 remain = m->m_len - len;
916 if (m0->m_flags & M_PKTHDR && remain == 0) {
917 n = m_gethdr(wait, m0->m_type);
920 n->m_next = m->m_next;
922 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
923 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
924 m0->m_pkthdr.len = len0;
926 } else if (m0->m_flags & M_PKTHDR) {
927 n = m_gethdr(wait, m0->m_type);
930 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
931 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
932 m0->m_pkthdr.len = len0;
933 if (m->m_flags & M_EXT)
935 if (remain > MHLEN) {
936 /* m can't be the lead packet */
938 n->m_next = m_split(m, len, wait);
939 if (n->m_next == NULL) {
948 } else if (remain == 0) {
953 n = m_get(wait, m->m_type);
959 if (m->m_flags & M_EXT) {
960 n->m_data = m->m_data + len;
963 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
967 n->m_next = m->m_next;
972 * Routine to copy from device local memory into mbufs.
973 * Note that `off' argument is offset into first mbuf of target chain from
974 * which to begin copying the data to.
977 m_devget(char *buf, int totlen, int off, struct ifnet *ifp,
978 void (*copy)(char *from, caddr_t to, u_int len))
981 struct mbuf *top = NULL, **mp = ⊤
984 if (off < 0 || off > MHLEN)
988 if (top == NULL) { /* First one, must be PKTHDR */
989 if (totlen + off >= MINCLSIZE) {
990 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
993 m = m_gethdr(M_NOWAIT, MT_DATA);
996 /* Place initial small packet/header at end of mbuf */
997 if (m && totlen + off + max_linkhdr <= MHLEN) {
998 m->m_data += max_linkhdr;
1004 m->m_pkthdr.rcvif = ifp;
1005 m->m_pkthdr.len = totlen;
1007 if (totlen + off >= MINCLSIZE) {
1008 m = m_getcl(M_NOWAIT, MT_DATA, 0);
1011 m = m_get(M_NOWAIT, MT_DATA);
1024 m->m_len = len = min(totlen, len);
1026 copy(buf, mtod(m, caddr_t), (u_int)len);
1028 bcopy(buf, mtod(m, caddr_t), (u_int)len);
1038 * Copy data from a buffer back into the indicated mbuf chain,
1039 * starting "off" bytes from the beginning, extending the mbuf
1040 * chain if necessary.
1043 m_copyback(struct mbuf *m0, int off, int len, c_caddr_t cp)
1046 struct mbuf *m = m0, *n;
1051 while (off > (mlen = m->m_len)) {
1054 if (m->m_next == NULL) {
1055 n = m_get(M_NOWAIT, m->m_type);
1058 bzero(mtod(n, caddr_t), MLEN);
1059 n->m_len = min(MLEN, len + off);
1065 if (m->m_next == NULL && (len > m->m_len - off)) {
1066 m->m_len += min(len - (m->m_len - off),
1067 M_TRAILINGSPACE(m));
1069 mlen = min (m->m_len - off, len);
1070 bcopy(cp, off + mtod(m, caddr_t), (u_int)mlen);
1078 if (m->m_next == NULL) {
1079 n = m_get(M_NOWAIT, m->m_type);
1082 n->m_len = min(MLEN, len);
1087 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1088 m->m_pkthdr.len = totlen;
1092 * Append the specified data to the indicated mbuf chain,
1093 * Extend the mbuf chain if the new data does not fit in
1096 * Return 1 if able to complete the job; otherwise 0.
1099 m_append(struct mbuf *m0, int len, c_caddr_t cp)
1102 int remainder, space;
1104 for (m = m0; m->m_next != NULL; m = m->m_next)
1107 space = M_TRAILINGSPACE(m);
1110 * Copy into available space.
1112 if (space > remainder)
1114 bcopy(cp, mtod(m, caddr_t) + m->m_len, space);
1116 cp += space, remainder -= space;
1118 while (remainder > 0) {
1120 * Allocate a new mbuf; could check space
1121 * and allocate a cluster instead.
1123 n = m_get(M_NOWAIT, m->m_type);
1126 n->m_len = min(MLEN, remainder);
1127 bcopy(cp, mtod(n, caddr_t), n->m_len);
1128 cp += n->m_len, remainder -= n->m_len;
1132 if (m0->m_flags & M_PKTHDR)
1133 m0->m_pkthdr.len += len - remainder;
1134 return (remainder == 0);
1138 * Apply function f to the data in an mbuf chain starting "off" bytes from
1139 * the beginning, continuing for "len" bytes.
1142 m_apply(struct mbuf *m, int off, int len,
1143 int (*f)(void *, void *, u_int), void *arg)
1148 KASSERT(off >= 0, ("m_apply, negative off %d", off));
1149 KASSERT(len >= 0, ("m_apply, negative len %d", len));
1151 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1158 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1159 count = min(m->m_len - off, len);
1160 rval = (*f)(arg, mtod(m, caddr_t) + off, count);
1171 * Return a pointer to mbuf/offset of location in mbuf chain.
1174 m_getptr(struct mbuf *m, int loc, int *off)
1178 /* Normal end of search. */
1179 if (m->m_len > loc) {
1184 if (m->m_next == NULL) {
1186 /* Point at the end of valid data. */
1199 m_print(const struct mbuf *m, int maxlen)
1203 const struct mbuf *m2;
1206 printf("mbuf: %p\n", m);
1210 if (m->m_flags & M_PKTHDR)
1211 len = m->m_pkthdr.len;
1215 while (m2 != NULL && (len == -1 || len)) {
1217 if (maxlen != -1 && pdata > maxlen)
1219 printf("mbuf: %p len: %d, next: %p, %b%s", m2, m2->m_len,
1220 m2->m_next, m2->m_flags, "\20\20freelist\17skipfw"
1221 "\11proto5\10proto4\7proto3\6proto2\5proto1\4rdonly"
1222 "\3eor\2pkthdr\1ext", pdata ? "" : "\n");
1224 printf(", %*D\n", pdata, (u_char *)m2->m_data, "-");
1230 printf("%d bytes unaccounted for.\n", len);
1235 m_fixhdr(struct mbuf *m0)
1239 len = m_length(m0, NULL);
1240 m0->m_pkthdr.len = len;
1245 m_length(struct mbuf *m0, struct mbuf **last)
1251 for (m = m0; m != NULL; m = m->m_next) {
1253 if (m->m_next == NULL)
1262 * Defragment a mbuf chain, returning the shortest possible
1263 * chain of mbufs and clusters. If allocation fails and
1264 * this cannot be completed, NULL will be returned, but
1265 * the passed in chain will be unchanged. Upon success,
1266 * the original chain will be freed, and the new chain
1269 * If a non-packet header is passed in, the original
1270 * mbuf (chain?) will be returned unharmed.
1273 m_defrag(struct mbuf *m0, int how)
1275 struct mbuf *m_new = NULL, *m_final = NULL;
1276 int progress = 0, length;
1278 MBUF_CHECKSLEEP(how);
1279 if (!(m0->m_flags & M_PKTHDR))
1282 m_fixhdr(m0); /* Needed sanity check */
1284 #ifdef MBUF_STRESS_TEST
1285 if (m_defragrandomfailures) {
1286 int temp = arc4random() & 0xff;
1292 if (m0->m_pkthdr.len > MHLEN)
1293 m_final = m_getcl(how, MT_DATA, M_PKTHDR);
1295 m_final = m_gethdr(how, MT_DATA);
1297 if (m_final == NULL)
1300 if (m_dup_pkthdr(m_final, m0, how) == 0)
1305 while (progress < m0->m_pkthdr.len) {
1306 length = m0->m_pkthdr.len - progress;
1307 if (length > MCLBYTES)
1310 if (m_new == NULL) {
1312 m_new = m_getcl(how, MT_DATA, 0);
1314 m_new = m_get(how, MT_DATA);
1319 m_copydata(m0, progress, length, mtod(m_new, caddr_t));
1321 m_new->m_len = length;
1322 if (m_new != m_final)
1323 m_cat(m_final, m_new);
1326 #ifdef MBUF_STRESS_TEST
1327 if (m0->m_next == NULL)
1332 #ifdef MBUF_STRESS_TEST
1334 m_defragbytes += m0->m_pkthdr.len;
1338 #ifdef MBUF_STRESS_TEST
1347 * Defragment an mbuf chain, returning at most maxfrags separate
1348 * mbufs+clusters. If this is not possible NULL is returned and
1349 * the original mbuf chain is left in its present (potentially
1350 * modified) state. We use two techniques: collapsing consecutive
1351 * mbufs and replacing consecutive mbufs by a cluster.
1353 * NB: this should really be named m_defrag but that name is taken
1356 m_collapse(struct mbuf *m0, int how, int maxfrags)
1358 struct mbuf *m, *n, *n2, **prev;
1362 * Calculate the current number of frags.
1365 for (m = m0; m != NULL; m = m->m_next)
1368 * First, try to collapse mbufs. Note that we always collapse
1369 * towards the front so we don't need to deal with moving the
1370 * pkthdr. This may be suboptimal if the first mbuf has much
1371 * less data than the following.
1379 if (M_WRITABLE(m) &&
1380 n->m_len < M_TRAILINGSPACE(m)) {
1381 bcopy(mtod(n, void *), mtod(m, char *) + m->m_len,
1383 m->m_len += n->m_len;
1384 m->m_next = n->m_next;
1386 if (--curfrags <= maxfrags)
1391 KASSERT(maxfrags > 1,
1392 ("maxfrags %u, but normal collapse failed", maxfrags));
1394 * Collapse consecutive mbufs to a cluster.
1396 prev = &m0->m_next; /* NB: not the first mbuf */
1397 while ((n = *prev) != NULL) {
1398 if ((n2 = n->m_next) != NULL &&
1399 n->m_len + n2->m_len < MCLBYTES) {
1400 m = m_getcl(how, MT_DATA, 0);
1403 bcopy(mtod(n, void *), mtod(m, void *), n->m_len);
1404 bcopy(mtod(n2, void *), mtod(m, char *) + n->m_len,
1406 m->m_len = n->m_len + n2->m_len;
1407 m->m_next = n2->m_next;
1411 if (--curfrags <= maxfrags) /* +1 cl -2 mbufs */
1414 * Still not there, try the normal collapse
1415 * again before we allocate another cluster.
1422 * No place where we can collapse to a cluster; punt.
1423 * This can occur if, for example, you request 2 frags
1424 * but the packet requires that both be clusters (we
1425 * never reallocate the first mbuf to avoid moving the
1432 #ifdef MBUF_STRESS_TEST
1435 * Fragment an mbuf chain. There's no reason you'd ever want to do
1436 * this in normal usage, but it's great for stress testing various
1439 * If fragmentation is not possible, the original chain will be
1442 * Possible length values:
1443 * 0 no fragmentation will occur
1444 * > 0 each fragment will be of the specified length
1445 * -1 each fragment will be the same random value in length
1446 * -2 each fragment's length will be entirely random
1447 * (Random values range from 1 to 256)
1450 m_fragment(struct mbuf *m0, int how, int length)
1452 struct mbuf *m_new = NULL, *m_final = NULL;
1455 if (!(m0->m_flags & M_PKTHDR))
1458 if ((length == 0) || (length < -2))
1461 m_fixhdr(m0); /* Needed sanity check */
1463 m_final = m_getcl(how, MT_DATA, M_PKTHDR);
1465 if (m_final == NULL)
1468 if (m_dup_pkthdr(m_final, m0, how) == 0)
1474 length = 1 + (arc4random() & 255);
1476 while (progress < m0->m_pkthdr.len) {
1482 fraglen = 1 + (arc4random() & 255);
1483 if (fraglen > m0->m_pkthdr.len - progress)
1484 fraglen = m0->m_pkthdr.len - progress;
1486 if (fraglen > MCLBYTES)
1489 if (m_new == NULL) {
1490 m_new = m_getcl(how, MT_DATA, 0);
1495 m_copydata(m0, progress, fraglen, mtod(m_new, caddr_t));
1496 progress += fraglen;
1497 m_new->m_len = fraglen;
1498 if (m_new != m_final)
1499 m_cat(m_final, m_new);
1508 /* Return the original chain on failure */
1515 * Copy the contents of uio into a properly sized mbuf chain.
1518 m_uiotombuf(struct uio *uio, int how, int len, int align, int flags)
1520 struct mbuf *m, *mb;
1526 * len can be zero or an arbitrary large value bound by
1527 * the total data supplied by the uio.
1530 total = (uio->uio_resid < len) ? uio->uio_resid : len;
1532 total = uio->uio_resid;
1535 * The smallest unit returned by m_getm2() is a single mbuf
1536 * with pkthdr. We can't align past it.
1542 * Give us the full allocation or nothing.
1543 * If len is zero return the smallest empty mbuf.
1545 m = m_getm2(NULL, max(total + align, 1), how, MT_DATA, flags);
1550 /* Fill all mbufs with uio data and update header information. */
1551 for (mb = m; mb != NULL; mb = mb->m_next) {
1552 length = min(M_TRAILINGSPACE(mb), total - progress);
1554 error = uiomove(mtod(mb, void *), length, uio);
1562 if (flags & M_PKTHDR)
1563 m->m_pkthdr.len += length;
1565 KASSERT(progress == total, ("%s: progress != total", __func__));
1571 * Copy an mbuf chain into a uio limited by len if set.
1574 m_mbuftouio(struct uio *uio, const struct mbuf *m, int len)
1576 int error, length, total;
1580 total = min(uio->uio_resid, len);
1582 total = uio->uio_resid;
1584 /* Fill the uio with data from the mbufs. */
1585 for (; m != NULL; m = m->m_next) {
1586 length = min(m->m_len, total - progress);
1588 error = uiomove(mtod(m, void *), length, uio);
1599 * Create a writable copy of the mbuf chain. While doing this
1600 * we compact the chain with a goal of producing a chain with
1601 * at most two mbufs. The second mbuf in this chain is likely
1602 * to be a cluster. The primary purpose of this work is to create
1603 * a writable packet for encryption, compression, etc. The
1604 * secondary goal is to linearize the data so the data can be
1605 * passed to crypto hardware in the most efficient manner possible.
1608 m_unshare(struct mbuf *m0, int how)
1610 struct mbuf *m, *mprev;
1611 struct mbuf *n, *mfirst, *mlast;
1615 for (m = m0; m != NULL; m = mprev->m_next) {
1617 * Regular mbufs are ignored unless there's a cluster
1618 * in front of it that we can use to coalesce. We do
1619 * the latter mainly so later clusters can be coalesced
1620 * also w/o having to handle them specially (i.e. convert
1621 * mbuf+cluster -> cluster). This optimization is heavily
1622 * influenced by the assumption that we're running over
1623 * Ethernet where MCLBYTES is large enough that the max
1624 * packet size will permit lots of coalescing into a
1625 * single cluster. This in turn permits efficient
1626 * crypto operations, especially when using hardware.
1628 if ((m->m_flags & M_EXT) == 0) {
1629 if (mprev && (mprev->m_flags & M_EXT) &&
1630 m->m_len <= M_TRAILINGSPACE(mprev)) {
1631 /* XXX: this ignores mbuf types */
1632 memcpy(mtod(mprev, caddr_t) + mprev->m_len,
1633 mtod(m, caddr_t), m->m_len);
1634 mprev->m_len += m->m_len;
1635 mprev->m_next = m->m_next; /* unlink from chain */
1636 m_free(m); /* reclaim mbuf */
1638 newipsecstat.ips_mbcoalesced++;
1646 * Writable mbufs are left alone (for now).
1648 if (M_WRITABLE(m)) {
1654 * Not writable, replace with a copy or coalesce with
1655 * the previous mbuf if possible (since we have to copy
1656 * it anyway, we try to reduce the number of mbufs and
1657 * clusters so that future work is easier).
1659 KASSERT(m->m_flags & M_EXT, ("m_flags 0x%x", m->m_flags));
1660 /* NB: we only coalesce into a cluster or larger */
1661 if (mprev != NULL && (mprev->m_flags & M_EXT) &&
1662 m->m_len <= M_TRAILINGSPACE(mprev)) {
1663 /* XXX: this ignores mbuf types */
1664 memcpy(mtod(mprev, caddr_t) + mprev->m_len,
1665 mtod(m, caddr_t), m->m_len);
1666 mprev->m_len += m->m_len;
1667 mprev->m_next = m->m_next; /* unlink from chain */
1668 m_free(m); /* reclaim mbuf */
1670 newipsecstat.ips_clcoalesced++;
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");