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>
52 #include <sys/vmmeter.h>
56 #include <vm/vm_pageout.h>
57 #include <vm/vm_page.h>
59 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__init,
60 "struct mbuf *", "mbufinfo_t *",
61 "uint32_t", "uint32_t",
62 "uint16_t", "uint16_t",
63 "uint32_t", "uint32_t",
64 "uint32_t", "uint32_t");
66 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__gethdr,
67 "uint32_t", "uint32_t",
68 "uint16_t", "uint16_t",
69 "struct mbuf *", "mbufinfo_t *");
71 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__get,
72 "uint32_t", "uint32_t",
73 "uint16_t", "uint16_t",
74 "struct mbuf *", "mbufinfo_t *");
76 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__getcl,
77 "uint32_t", "uint32_t",
78 "uint16_t", "uint16_t",
79 "uint32_t", "uint32_t",
80 "struct mbuf *", "mbufinfo_t *");
82 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__getjcl,
83 "uint32_t", "uint32_t",
84 "uint16_t", "uint16_t",
85 "uint32_t", "uint32_t",
86 "uint32_t", "uint32_t",
87 "struct mbuf *", "mbufinfo_t *");
89 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__clget,
90 "struct mbuf *", "mbufinfo_t *",
91 "uint32_t", "uint32_t",
92 "uint32_t", "uint32_t");
94 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__cljget,
95 "struct mbuf *", "mbufinfo_t *",
96 "uint32_t", "uint32_t",
97 "uint32_t", "uint32_t",
100 SDT_PROBE_DEFINE(sdt, , , m__cljset);
102 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__free,
103 "struct mbuf *", "mbufinfo_t *");
105 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__freem,
106 "struct mbuf *", "mbufinfo_t *");
108 #include <security/mac/mac_framework.h>
114 #ifdef MBUF_STRESS_TEST
119 int m_defragrandomfailures;
123 * sysctl(8) exported objects
125 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RD,
126 &max_linkhdr, 0, "Size of largest link layer header");
127 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RD,
128 &max_protohdr, 0, "Size of largest protocol layer header");
129 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RD,
130 &max_hdr, 0, "Size of largest link plus protocol header");
131 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RD,
132 &max_datalen, 0, "Minimum space left in mbuf after max_hdr");
133 #ifdef MBUF_STRESS_TEST
134 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
135 &m_defragpackets, 0, "");
136 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
137 &m_defragbytes, 0, "");
138 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
139 &m_defraguseless, 0, "");
140 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
141 &m_defragfailure, 0, "");
142 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
143 &m_defragrandomfailures, 0, "");
147 * Ensure the correct size of various mbuf parameters. It could be off due
148 * to compiler-induced padding and alignment artifacts.
150 CTASSERT(MSIZE - offsetof(struct mbuf, m_dat) == MLEN);
151 CTASSERT(MSIZE - offsetof(struct mbuf, m_pktdat) == MHLEN);
154 * mbuf data storage should be 64-bit aligned regardless of architectural
155 * pointer size; check this is the case with and without a packet header.
157 CTASSERT(offsetof(struct mbuf, m_dat) % 8 == 0);
158 CTASSERT(offsetof(struct mbuf, m_pktdat) % 8 == 0);
161 * While the specific values here don't matter too much (i.e., +/- a few
162 * words), we do want to ensure that changes to these values are carefully
163 * reasoned about and properly documented. This is especially the case as
164 * network-protocol and device-driver modules encode these layouts, and must
165 * be recompiled if the structures change. Check these values at compile time
166 * against the ones documented in comments in mbuf.h.
168 * NB: Possibly they should be documented there via #define's and not just
171 #if defined(__LP64__)
172 CTASSERT(offsetof(struct mbuf, m_dat) == 32);
173 CTASSERT(sizeof(struct pkthdr) == 56);
174 CTASSERT(sizeof(struct m_ext) == 160);
176 CTASSERT(offsetof(struct mbuf, m_dat) == 24);
177 CTASSERT(sizeof(struct pkthdr) == 48);
178 #if defined(__powerpc__) && defined(BOOKE)
179 /* PowerPC booke has 64-bit physical pointers. */
180 CTASSERT(sizeof(struct m_ext) == 184);
182 CTASSERT(sizeof(struct m_ext) == 180);
187 * Assert that the queue(3) macros produce code of the same size as an old
188 * plain pointer does.
191 static struct mbuf __used m_assertbuf;
192 CTASSERT(sizeof(m_assertbuf.m_slist) == sizeof(m_assertbuf.m_next));
193 CTASSERT(sizeof(m_assertbuf.m_stailq) == sizeof(m_assertbuf.m_next));
194 CTASSERT(sizeof(m_assertbuf.m_slistpkt) == sizeof(m_assertbuf.m_nextpkt));
195 CTASSERT(sizeof(m_assertbuf.m_stailqpkt) == sizeof(m_assertbuf.m_nextpkt));
199 * Attach the cluster from *m to *n, set up m_ext in *n
200 * and bump the refcount of the cluster.
203 mb_dupcl(struct mbuf *n, struct mbuf *m)
205 volatile u_int *refcnt;
207 KASSERT(m->m_flags & (M_EXT|M_EXTPG),
208 ("%s: M_EXT|M_EXTPG not set on %p", __func__, m));
209 KASSERT(!(n->m_flags & (M_EXT|M_EXTPG)),
210 ("%s: M_EXT|M_EXTPG set on %p", __func__, n));
213 * Cache access optimization.
215 * o Regular M_EXT storage doesn't need full copy of m_ext, since
216 * the holder of the 'ext_count' is responsible to carry the free
217 * routine and its arguments.
218 * o M_EXTPG data is split between main part of mbuf and m_ext, the
219 * main part is copied in full, the m_ext part is similar to M_EXT.
220 * o EXT_EXTREF, where 'ext_cnt' doesn't point into mbuf at all, is
221 * special - it needs full copy of m_ext into each mbuf, since any
222 * copy could end up as the last to free.
224 if (m->m_flags & M_EXTPG) {
225 bcopy(&m->m_epg_startcopy, &n->m_epg_startcopy,
226 __rangeof(struct mbuf, m_epg_startcopy, m_epg_endcopy));
227 bcopy(&m->m_ext, &n->m_ext, m_epg_ext_copylen);
228 } else if (m->m_ext.ext_type == EXT_EXTREF)
229 bcopy(&m->m_ext, &n->m_ext, sizeof(struct m_ext));
231 bcopy(&m->m_ext, &n->m_ext, m_ext_copylen);
233 n->m_flags |= m->m_flags & (M_RDONLY | M_EXT | M_EXTPG);
235 /* See if this is the mbuf that holds the embedded refcount. */
236 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
237 refcnt = n->m_ext.ext_cnt = &m->m_ext.ext_count;
238 n->m_ext.ext_flags &= ~EXT_FLAG_EMBREF;
240 KASSERT(m->m_ext.ext_cnt != NULL,
241 ("%s: no refcounting pointer on %p", __func__, m));
242 refcnt = m->m_ext.ext_cnt;
248 atomic_add_int(refcnt, 1);
252 m_demote_pkthdr(struct mbuf *m)
257 m_tag_delete_chain(m, NULL);
258 m->m_flags &= ~M_PKTHDR;
259 bzero(&m->m_pkthdr, sizeof(struct pkthdr));
263 * Clean up mbuf (chain) from any tags and packet headers.
264 * If "all" is set then the first mbuf in the chain will be
268 m_demote(struct mbuf *m0, int all, int flags)
272 flags |= M_DEMOTEFLAGS;
274 for (m = all ? m0 : m0->m_next; m != NULL; m = m->m_next) {
275 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt in m %p, m0 %p",
277 if (m->m_flags & M_PKTHDR)
284 * Sanity checks on mbuf (chain) for use in KASSERT() and general
286 * Returns 0 or panics when bad and 1 on all tests passed.
287 * Sanitize, 0 to run M_SANITY_ACTION, 1 to garble things so they
291 m_sanity(struct mbuf *m0, int sanitize)
298 #define M_SANITY_ACTION(s) panic("mbuf %p: " s, m)
300 #define M_SANITY_ACTION(s) printf("mbuf %p: " s, m)
303 for (m = m0; m != NULL; m = m->m_next) {
305 * Basic pointer checks. If any of these fails then some
306 * unrelated kernel memory before or after us is trashed.
307 * No way to recover from that.
311 if ((caddr_t)m->m_data < a)
312 M_SANITY_ACTION("m_data outside mbuf data range left");
313 if ((caddr_t)m->m_data > b)
314 M_SANITY_ACTION("m_data outside mbuf data range right");
315 if ((caddr_t)m->m_data + m->m_len > b)
316 M_SANITY_ACTION("m_data + m_len exeeds mbuf space");
318 /* m->m_nextpkt may only be set on first mbuf in chain. */
319 if (m != m0 && m->m_nextpkt != NULL) {
321 m_freem(m->m_nextpkt);
322 m->m_nextpkt = (struct mbuf *)0xDEADC0DE;
324 M_SANITY_ACTION("m->m_nextpkt on in-chain mbuf");
327 /* packet length (not mbuf length!) calculation */
328 if (m0->m_flags & M_PKTHDR)
331 /* m_tags may only be attached to first mbuf in chain. */
332 if (m != m0 && m->m_flags & M_PKTHDR &&
333 !SLIST_EMPTY(&m->m_pkthdr.tags)) {
335 m_tag_delete_chain(m, NULL);
336 /* put in 0xDEADC0DE perhaps? */
338 M_SANITY_ACTION("m_tags on in-chain mbuf");
341 /* M_PKTHDR may only be set on first mbuf in chain */
342 if (m != m0 && m->m_flags & M_PKTHDR) {
344 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
345 m->m_flags &= ~M_PKTHDR;
346 /* put in 0xDEADCODE and leave hdr flag in */
348 M_SANITY_ACTION("M_PKTHDR on in-chain mbuf");
352 if (pktlen && pktlen != m->m_pkthdr.len) {
356 M_SANITY_ACTION("m_pkthdr.len != mbuf chain length");
360 #undef M_SANITY_ACTION
364 * Non-inlined part of m_init().
367 m_pkthdr_init(struct mbuf *m, int how)
372 m->m_data = m->m_pktdat;
373 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
375 m->m_pkthdr.numa_domain = M_NODOM;
378 /* If the label init fails, fail the alloc */
379 error = mac_mbuf_init(m, how);
388 * "Move" mbuf pkthdr from "from" to "to".
389 * "from" must have M_PKTHDR set, and "to" must be empty.
392 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
396 /* see below for why these are not enabled */
398 /* Note: with MAC, this may not be a good assertion. */
399 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags),
400 ("m_move_pkthdr: to has tags"));
404 * XXXMAC: It could be this should also occur for non-MAC?
406 if (to->m_flags & M_PKTHDR)
407 m_tag_delete_chain(to, NULL);
409 to->m_flags = (from->m_flags & M_COPYFLAGS) |
410 (to->m_flags & (M_EXT | M_EXTPG));
411 if ((to->m_flags & M_EXT) == 0)
412 to->m_data = to->m_pktdat;
413 to->m_pkthdr = from->m_pkthdr; /* especially tags */
414 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */
415 from->m_flags &= ~M_PKTHDR;
416 if (from->m_pkthdr.csum_flags & CSUM_SND_TAG) {
417 from->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
418 from->m_pkthdr.snd_tag = NULL;
423 * Duplicate "from"'s mbuf pkthdr in "to".
424 * "from" must have M_PKTHDR set, and "to" must be empty.
425 * In particular, this does a deep copy of the packet tags.
428 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
433 * The mbuf allocator only initializes the pkthdr
434 * when the mbuf is allocated with m_gethdr(). Many users
435 * (e.g. m_copy*, m_prepend) use m_get() and then
436 * smash the pkthdr as needed causing these
437 * assertions to trip. For now just disable them.
440 /* Note: with MAC, this may not be a good assertion. */
441 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags), ("m_dup_pkthdr: to has tags"));
443 MBUF_CHECKSLEEP(how);
445 if (to->m_flags & M_PKTHDR)
446 m_tag_delete_chain(to, NULL);
448 to->m_flags = (from->m_flags & M_COPYFLAGS) |
449 (to->m_flags & (M_EXT | M_EXTPG));
450 if ((to->m_flags & M_EXT) == 0)
451 to->m_data = to->m_pktdat;
452 to->m_pkthdr = from->m_pkthdr;
453 if (from->m_pkthdr.csum_flags & CSUM_SND_TAG)
454 m_snd_tag_ref(from->m_pkthdr.snd_tag);
455 SLIST_INIT(&to->m_pkthdr.tags);
456 return (m_tag_copy_chain(to, from, how));
460 * Lesser-used path for M_PREPEND:
461 * allocate new mbuf to prepend to chain,
465 m_prepend(struct mbuf *m, int len, int how)
469 if (m->m_flags & M_PKTHDR)
470 mn = m_gethdr(how, m->m_type);
472 mn = m_get(how, m->m_type);
477 if (m->m_flags & M_PKTHDR)
478 m_move_pkthdr(mn, m);
488 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
489 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
490 * The wait parameter is a choice of M_WAITOK/M_NOWAIT from caller.
491 * Note that the copy is read-only, because clusters are not copied,
492 * only their reference counts are incremented.
495 m_copym(struct mbuf *m, int off0, int len, int wait)
497 struct mbuf *n, **np;
502 KASSERT(off >= 0, ("m_copym, negative off %d", off));
503 KASSERT(len >= 0, ("m_copym, negative len %d", len));
504 MBUF_CHECKSLEEP(wait);
505 if (off == 0 && m->m_flags & M_PKTHDR)
508 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
518 KASSERT(len == M_COPYALL,
519 ("m_copym, length > size of mbuf chain"));
523 n = m_gethdr(wait, m->m_type);
525 n = m_get(wait, m->m_type);
530 if (!m_dup_pkthdr(n, m, wait))
532 if (len == M_COPYALL)
533 n->m_pkthdr.len -= off0;
535 n->m_pkthdr.len = len;
538 n->m_len = min(len, m->m_len - off);
539 if (m->m_flags & (M_EXT|M_EXTPG)) {
540 n->m_data = m->m_data + off;
543 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
545 if (len != M_COPYALL)
559 * Copy an entire packet, including header (which must be present).
560 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
561 * Note that the copy is read-only, because clusters are not copied,
562 * only their reference counts are incremented.
563 * Preserve alignment of the first mbuf so if the creator has left
564 * some room at the beginning (e.g. for inserting protocol headers)
565 * the copies still have the room available.
568 m_copypacket(struct mbuf *m, int how)
570 struct mbuf *top, *n, *o;
572 MBUF_CHECKSLEEP(how);
573 n = m_get(how, m->m_type);
578 if (!m_dup_pkthdr(n, m, how))
581 if (m->m_flags & (M_EXT|M_EXTPG)) {
582 n->m_data = m->m_data;
585 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
586 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
591 o = m_get(how, m->m_type);
599 if (m->m_flags & (M_EXT|M_EXTPG)) {
600 n->m_data = m->m_data;
603 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
615 m_copyfromunmapped(const struct mbuf *m, int off, int len, caddr_t cp)
621 KASSERT(off >= 0, ("m_copyfromunmapped: negative off %d", off));
622 KASSERT(len >= 0, ("m_copyfromunmapped: negative len %d", len));
623 KASSERT(off < m->m_len,
624 ("m_copyfromunmapped: len exceeds mbuf length"));
629 uio.uio_segflg = UIO_SYSSPACE;
632 uio.uio_rw = UIO_READ;
633 error = m_unmapped_uiomove(m, off, &uio, len);
634 KASSERT(error == 0, ("m_unmapped_uiomove failed: off %d, len %d", off,
639 * Copy data from an mbuf chain starting "off" bytes from the beginning,
640 * continuing for "len" bytes, into the indicated buffer.
643 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
647 KASSERT(off >= 0, ("m_copydata, negative off %d", off));
648 KASSERT(len >= 0, ("m_copydata, negative len %d", len));
650 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
657 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
658 count = min(m->m_len - off, len);
659 if ((m->m_flags & M_EXTPG) != 0)
660 m_copyfromunmapped(m, off, count, cp);
662 bcopy(mtod(m, caddr_t) + off, cp, count);
671 * Copy a packet header mbuf chain into a completely new chain, including
672 * copying any mbuf clusters. Use this instead of m_copypacket() when
673 * you need a writable copy of an mbuf chain.
676 m_dup(const struct mbuf *m, int how)
678 struct mbuf **p, *top = NULL;
679 int remain, moff, nsize;
681 MBUF_CHECKSLEEP(how);
687 /* While there's more data, get a new mbuf, tack it on, and fill it */
688 remain = m->m_pkthdr.len;
691 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */
694 /* Get the next new mbuf */
695 if (remain >= MINCLSIZE) {
696 n = m_getcl(how, m->m_type, 0);
699 n = m_get(how, m->m_type);
705 if (top == NULL) { /* First one, must be PKTHDR */
706 if (!m_dup_pkthdr(n, m, how)) {
710 if ((n->m_flags & M_EXT) == 0)
712 n->m_flags &= ~M_RDONLY;
716 /* Link it into the new chain */
720 /* Copy data from original mbuf(s) into new mbuf */
721 while (n->m_len < nsize && m != NULL) {
722 int chunk = min(nsize - n->m_len, m->m_len - moff);
724 m_copydata(m, moff, chunk, n->m_data + n->m_len);
728 if (moff == m->m_len) {
734 /* Check correct total mbuf length */
735 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
736 ("%s: bogus m_pkthdr.len", __func__));
746 * Concatenate mbuf chain n to m.
747 * Both chains must be of the same type (e.g. MT_DATA).
748 * Any m_pkthdr is not updated.
751 m_cat(struct mbuf *m, struct mbuf *n)
756 if (!M_WRITABLE(m) ||
757 (n->m_flags & M_EXTPG) != 0 ||
758 M_TRAILINGSPACE(m) < n->m_len) {
759 /* just join the two chains */
763 /* splat the data from one into the other */
764 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
766 m->m_len += n->m_len;
772 * Concatenate two pkthdr mbuf chains.
775 m_catpkt(struct mbuf *m, struct mbuf *n)
781 m->m_pkthdr.len += n->m_pkthdr.len;
788 m_adj(struct mbuf *mp, int req_len)
794 if ((m = mp) == NULL)
800 while (m != NULL && len > 0) {
801 if (m->m_len <= len) {
811 if (mp->m_flags & M_PKTHDR)
812 mp->m_pkthdr.len -= (req_len - len);
815 * Trim from tail. Scan the mbuf chain,
816 * calculating its length and finding the last mbuf.
817 * If the adjustment only affects this mbuf, then just
818 * adjust and return. Otherwise, rescan and truncate
819 * after the remaining size.
825 if (m->m_next == (struct mbuf *)0)
829 if (m->m_len >= len) {
831 if (mp->m_flags & M_PKTHDR)
832 mp->m_pkthdr.len -= len;
839 * Correct length for chain is "count".
840 * Find the mbuf with last data, adjust its length,
841 * and toss data from remaining mbufs on chain.
844 if (m->m_flags & M_PKTHDR)
845 m->m_pkthdr.len = count;
846 for (; m; m = m->m_next) {
847 if (m->m_len >= count) {
849 if (m->m_next != NULL) {
861 m_adj_decap(struct mbuf *mp, int len)
866 if ((mp->m_flags & M_PKTHDR) != 0) {
868 * If flowid was calculated by card from the inner
869 * headers, move flowid to the decapsulated mbuf
870 * chain, otherwise clear. This depends on the
871 * internals of m_adj, which keeps pkthdr as is, in
872 * particular not changing rsstype and flowid.
874 rsstype = mp->m_pkthdr.rsstype;
875 if ((rsstype & M_HASHTYPE_INNER) != 0) {
876 M_HASHTYPE_SET(mp, rsstype & ~M_HASHTYPE_INNER);
878 M_HASHTYPE_CLEAR(mp);
884 * Rearange an mbuf chain so that len bytes are contiguous
885 * and in the data area of an mbuf (so that mtod will work
886 * for a structure of size len). Returns the resulting
887 * mbuf chain on success, frees it and returns null on failure.
888 * If there is room, it will add up to max_protohdr-len extra bytes to the
889 * contiguous region in an attempt to avoid being called next time.
892 m_pullup(struct mbuf *n, int len)
898 KASSERT((n->m_flags & M_EXTPG) == 0,
899 ("%s: unmapped mbuf %p", __func__, n));
902 * If first mbuf has no cluster, and has room for len bytes
903 * without shifting current data, pullup into it,
904 * otherwise allocate a new mbuf to prepend to the chain.
906 if ((n->m_flags & M_EXT) == 0 &&
907 n->m_data + len < &n->m_dat[MLEN] && n->m_next) {
916 m = m_get(M_NOWAIT, n->m_type);
919 if (n->m_flags & M_PKTHDR)
922 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
924 count = min(min(max(len, max_protohdr), space), n->m_len);
925 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
935 } while (len > 0 && n);
948 * Like m_pullup(), except a new mbuf is always allocated, and we allow
949 * the amount of empty space before the data in the new mbuf to be specified
950 * (in the event that the caller expects to prepend later).
953 m_copyup(struct mbuf *n, int len, int dstoff)
958 if (len > (MHLEN - dstoff))
960 m = m_get(M_NOWAIT, n->m_type);
963 if (n->m_flags & M_PKTHDR)
966 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
968 count = min(min(max(len, max_protohdr), space), n->m_len);
969 memcpy(mtod(m, caddr_t) + m->m_len, mtod(n, caddr_t),
979 } while (len > 0 && n);
992 * Partition an mbuf chain in two pieces, returning the tail --
993 * all but the first len0 bytes. In case of failure, it returns NULL and
994 * attempts to restore the chain to its original state.
996 * Note that the resulting mbufs might be read-only, because the new
997 * mbuf can end up sharing an mbuf cluster with the original mbuf if
998 * the "breaking point" happens to lie within a cluster mbuf. Use the
999 * M_WRITABLE() macro to check for this case.
1002 m_split(struct mbuf *m0, int len0, int wait)
1005 u_int len = len0, remain;
1007 MBUF_CHECKSLEEP(wait);
1008 for (m = m0; m && len > m->m_len; m = m->m_next)
1012 remain = m->m_len - len;
1013 if (m0->m_flags & M_PKTHDR && remain == 0) {
1014 n = m_gethdr(wait, m0->m_type);
1017 n->m_next = m->m_next;
1019 if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG) {
1020 n->m_pkthdr.snd_tag =
1021 m_snd_tag_ref(m0->m_pkthdr.snd_tag);
1022 n->m_pkthdr.csum_flags |= CSUM_SND_TAG;
1024 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1025 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1026 m0->m_pkthdr.len = len0;
1028 } else if (m0->m_flags & M_PKTHDR) {
1029 n = m_gethdr(wait, m0->m_type);
1032 if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG) {
1033 n->m_pkthdr.snd_tag =
1034 m_snd_tag_ref(m0->m_pkthdr.snd_tag);
1035 n->m_pkthdr.csum_flags |= CSUM_SND_TAG;
1037 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1038 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1039 m0->m_pkthdr.len = len0;
1040 if (m->m_flags & (M_EXT|M_EXTPG))
1042 if (remain > MHLEN) {
1043 /* m can't be the lead packet */
1045 n->m_next = m_split(m, len, wait);
1046 if (n->m_next == NULL) {
1055 } else if (remain == 0) {
1060 n = m_get(wait, m->m_type);
1066 if (m->m_flags & (M_EXT|M_EXTPG)) {
1067 n->m_data = m->m_data + len;
1070 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1074 n->m_next = m->m_next;
1079 * Routine to copy from device local memory into mbufs.
1080 * Note that `off' argument is offset into first mbuf of target chain from
1081 * which to begin copying the data to.
1084 m_devget(char *buf, int totlen, int off, struct ifnet *ifp,
1085 void (*copy)(char *from, caddr_t to, u_int len))
1088 struct mbuf *top = NULL, **mp = ⊤
1091 if (off < 0 || off > MHLEN)
1094 while (totlen > 0) {
1095 if (top == NULL) { /* First one, must be PKTHDR */
1096 if (totlen + off >= MINCLSIZE) {
1097 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1100 m = m_gethdr(M_NOWAIT, MT_DATA);
1103 /* Place initial small packet/header at end of mbuf */
1104 if (m && totlen + off + max_linkhdr <= MHLEN) {
1105 m->m_data += max_linkhdr;
1111 m->m_pkthdr.rcvif = ifp;
1112 m->m_pkthdr.len = totlen;
1114 if (totlen + off >= MINCLSIZE) {
1115 m = m_getcl(M_NOWAIT, MT_DATA, 0);
1118 m = m_get(M_NOWAIT, MT_DATA);
1131 m->m_len = len = min(totlen, len);
1133 copy(buf, mtod(m, caddr_t), (u_int)len);
1135 bcopy(buf, mtod(m, caddr_t), (u_int)len);
1145 m_copytounmapped(const struct mbuf *m, int off, int len, c_caddr_t cp)
1151 KASSERT(off >= 0, ("m_copytounmapped: negative off %d", off));
1152 KASSERT(len >= 0, ("m_copytounmapped: negative len %d", len));
1153 KASSERT(off < m->m_len, ("m_copytounmapped: len exceeds mbuf length"));
1154 iov.iov_base = __DECONST(caddr_t, cp);
1156 uio.uio_resid = len;
1158 uio.uio_segflg = UIO_SYSSPACE;
1161 uio.uio_rw = UIO_WRITE;
1162 error = m_unmapped_uiomove(m, off, &uio, len);
1163 KASSERT(error == 0, ("m_unmapped_uiomove failed: off %d, len %d", off,
1168 * Copy data from a buffer back into the indicated mbuf chain,
1169 * starting "off" bytes from the beginning, extending the mbuf
1170 * chain if necessary.
1173 m_copyback(struct mbuf *m0, int off, int len, c_caddr_t cp)
1176 struct mbuf *m = m0, *n;
1181 while (off > (mlen = m->m_len)) {
1184 if (m->m_next == NULL) {
1185 n = m_get(M_NOWAIT, m->m_type);
1188 bzero(mtod(n, caddr_t), MLEN);
1189 n->m_len = min(MLEN, len + off);
1195 if (m->m_next == NULL && (len > m->m_len - off)) {
1196 m->m_len += min(len - (m->m_len - off),
1197 M_TRAILINGSPACE(m));
1199 mlen = min (m->m_len - off, len);
1200 if ((m->m_flags & M_EXTPG) != 0)
1201 m_copytounmapped(m, off, mlen, cp);
1203 bcopy(cp, off + mtod(m, caddr_t), (u_int)mlen);
1211 if (m->m_next == NULL) {
1212 n = m_get(M_NOWAIT, m->m_type);
1215 n->m_len = min(MLEN, len);
1220 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1221 m->m_pkthdr.len = totlen;
1225 * Append the specified data to the indicated mbuf chain,
1226 * Extend the mbuf chain if the new data does not fit in
1229 * Return 1 if able to complete the job; otherwise 0.
1232 m_append(struct mbuf *m0, int len, c_caddr_t cp)
1235 int remainder, space;
1237 for (m = m0; m->m_next != NULL; m = m->m_next)
1240 space = M_TRAILINGSPACE(m);
1243 * Copy into available space.
1245 if (space > remainder)
1247 bcopy(cp, mtod(m, caddr_t) + m->m_len, space);
1249 cp += space, remainder -= space;
1251 while (remainder > 0) {
1253 * Allocate a new mbuf; could check space
1254 * and allocate a cluster instead.
1256 n = m_get(M_NOWAIT, m->m_type);
1259 n->m_len = min(MLEN, remainder);
1260 bcopy(cp, mtod(n, caddr_t), n->m_len);
1261 cp += n->m_len, remainder -= n->m_len;
1265 if (m0->m_flags & M_PKTHDR)
1266 m0->m_pkthdr.len += len - remainder;
1267 return (remainder == 0);
1271 m_apply_extpg_one(struct mbuf *m, int off, int len,
1272 int (*f)(void *, void *, u_int), void *arg)
1275 u_int i, count, pgoff, pglen;
1278 KASSERT(PMAP_HAS_DMAP,
1279 ("m_apply_extpg_one does not support unmapped mbufs"));
1280 off += mtod(m, vm_offset_t);
1281 if (off < m->m_epg_hdrlen) {
1282 count = min(m->m_epg_hdrlen - off, len);
1283 rval = f(arg, m->m_epg_hdr + off, count);
1289 off -= m->m_epg_hdrlen;
1290 pgoff = m->m_epg_1st_off;
1291 for (i = 0; i < m->m_epg_npgs && len > 0; i++) {
1292 pglen = m_epg_pagelen(m, i, pgoff);
1294 count = min(pglen - off, len);
1295 p = (void *)PHYS_TO_DMAP(m->m_epg_pa[i] + pgoff);
1296 rval = f(arg, p, count);
1306 KASSERT(off < m->m_epg_trllen,
1307 ("m_apply_extpg_one: offset beyond trailer"));
1308 KASSERT(len <= m->m_epg_trllen - off,
1309 ("m_apply_extpg_one: length beyond trailer"));
1310 return (f(arg, m->m_epg_trail + off, len));
1315 /* Apply function f to the data in a single mbuf. */
1317 m_apply_one(struct mbuf *m, int off, int len,
1318 int (*f)(void *, void *, u_int), void *arg)
1320 if ((m->m_flags & M_EXTPG) != 0)
1321 return (m_apply_extpg_one(m, off, len, f, arg));
1323 return (f(arg, mtod(m, caddr_t) + off, len));
1327 * Apply function f to the data in an mbuf chain starting "off" bytes from
1328 * the beginning, continuing for "len" bytes.
1331 m_apply(struct mbuf *m, int off, int len,
1332 int (*f)(void *, void *, u_int), void *arg)
1337 KASSERT(off >= 0, ("m_apply, negative off %d", off));
1338 KASSERT(len >= 0, ("m_apply, negative len %d", len));
1340 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1347 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1348 count = min(m->m_len - off, len);
1349 rval = m_apply_one(m, off, count, f, arg);
1360 * Return a pointer to mbuf/offset of location in mbuf chain.
1363 m_getptr(struct mbuf *m, int loc, int *off)
1367 /* Normal end of search. */
1368 if (m->m_len > loc) {
1373 if (m->m_next == NULL) {
1375 /* Point at the end of valid data. */
1388 m_print(const struct mbuf *m, int maxlen)
1392 const struct mbuf *m2;
1395 printf("mbuf: %p\n", m);
1399 if (m->m_flags & M_PKTHDR)
1400 len = m->m_pkthdr.len;
1404 while (m2 != NULL && (len == -1 || len)) {
1406 if (maxlen != -1 && pdata > maxlen)
1408 printf("mbuf: %p len: %d, next: %p, %b%s", m2, m2->m_len,
1409 m2->m_next, m2->m_flags, "\20\20freelist\17skipfw"
1410 "\11proto5\10proto4\7proto3\6proto2\5proto1\4rdonly"
1411 "\3eor\2pkthdr\1ext", pdata ? "" : "\n");
1413 printf(", %*D\n", pdata, (u_char *)m2->m_data, "-");
1419 printf("%d bytes unaccounted for.\n", len);
1424 m_fixhdr(struct mbuf *m0)
1428 len = m_length(m0, NULL);
1429 m0->m_pkthdr.len = len;
1434 m_length(struct mbuf *m0, struct mbuf **last)
1440 for (m = m0; m != NULL; m = m->m_next) {
1442 if (m->m_next == NULL)
1451 * Defragment a mbuf chain, returning the shortest possible
1452 * chain of mbufs and clusters. If allocation fails and
1453 * this cannot be completed, NULL will be returned, but
1454 * the passed in chain will be unchanged. Upon success,
1455 * the original chain will be freed, and the new chain
1458 * If a non-packet header is passed in, the original
1459 * mbuf (chain?) will be returned unharmed.
1462 m_defrag(struct mbuf *m0, int how)
1464 struct mbuf *m_new = NULL, *m_final = NULL;
1465 int progress = 0, length;
1467 MBUF_CHECKSLEEP(how);
1468 if (!(m0->m_flags & M_PKTHDR))
1471 m_fixhdr(m0); /* Needed sanity check */
1473 #ifdef MBUF_STRESS_TEST
1474 if (m_defragrandomfailures) {
1475 int temp = arc4random() & 0xff;
1481 if (m0->m_pkthdr.len > MHLEN)
1482 m_final = m_getcl(how, MT_DATA, M_PKTHDR);
1484 m_final = m_gethdr(how, MT_DATA);
1486 if (m_final == NULL)
1489 if (m_dup_pkthdr(m_final, m0, how) == 0)
1494 while (progress < m0->m_pkthdr.len) {
1495 length = m0->m_pkthdr.len - progress;
1496 if (length > MCLBYTES)
1499 if (m_new == NULL) {
1501 m_new = m_getcl(how, MT_DATA, 0);
1503 m_new = m_get(how, MT_DATA);
1508 m_copydata(m0, progress, length, mtod(m_new, caddr_t));
1510 m_new->m_len = length;
1511 if (m_new != m_final)
1512 m_cat(m_final, m_new);
1515 #ifdef MBUF_STRESS_TEST
1516 if (m0->m_next == NULL)
1521 #ifdef MBUF_STRESS_TEST
1523 m_defragbytes += m0->m_pkthdr.len;
1527 #ifdef MBUF_STRESS_TEST
1536 * Return the number of fragments an mbuf will use. This is usually
1537 * used as a proxy for the number of scatter/gather elements needed by
1538 * a DMA engine to access an mbuf. In general mapped mbufs are
1539 * assumed to be backed by physically contiguous buffers that only
1540 * need a single fragment. Unmapped mbufs, on the other hand, can
1541 * span disjoint physical pages.
1544 frags_per_mbuf(struct mbuf *m)
1548 if ((m->m_flags & M_EXTPG) == 0)
1552 * The header and trailer are counted as a single fragment
1553 * each when present.
1555 * XXX: This overestimates the number of fragments by assuming
1556 * all the backing physical pages are disjoint.
1559 if (m->m_epg_hdrlen != 0)
1561 frags += m->m_epg_npgs;
1562 if (m->m_epg_trllen != 0)
1569 * Defragment an mbuf chain, returning at most maxfrags separate
1570 * mbufs+clusters. If this is not possible NULL is returned and
1571 * the original mbuf chain is left in its present (potentially
1572 * modified) state. We use two techniques: collapsing consecutive
1573 * mbufs and replacing consecutive mbufs by a cluster.
1575 * NB: this should really be named m_defrag but that name is taken
1578 m_collapse(struct mbuf *m0, int how, int maxfrags)
1580 struct mbuf *m, *n, *n2, **prev;
1584 * Calculate the current number of frags.
1587 for (m = m0; m != NULL; m = m->m_next)
1588 curfrags += frags_per_mbuf(m);
1590 * First, try to collapse mbufs. Note that we always collapse
1591 * towards the front so we don't need to deal with moving the
1592 * pkthdr. This may be suboptimal if the first mbuf has much
1593 * less data than the following.
1601 if (M_WRITABLE(m) &&
1602 n->m_len < M_TRAILINGSPACE(m)) {
1603 m_copydata(n, 0, n->m_len,
1604 mtod(m, char *) + m->m_len);
1605 m->m_len += n->m_len;
1606 m->m_next = n->m_next;
1607 curfrags -= frags_per_mbuf(n);
1609 if (curfrags <= maxfrags)
1614 KASSERT(maxfrags > 1,
1615 ("maxfrags %u, but normal collapse failed", maxfrags));
1617 * Collapse consecutive mbufs to a cluster.
1619 prev = &m0->m_next; /* NB: not the first mbuf */
1620 while ((n = *prev) != NULL) {
1621 if ((n2 = n->m_next) != NULL &&
1622 n->m_len + n2->m_len < MCLBYTES) {
1623 m = m_getcl(how, MT_DATA, 0);
1626 m_copydata(n, 0, n->m_len, mtod(m, char *));
1627 m_copydata(n2, 0, n2->m_len,
1628 mtod(m, char *) + n->m_len);
1629 m->m_len = n->m_len + n2->m_len;
1630 m->m_next = n2->m_next;
1632 curfrags += 1; /* For the new cluster */
1633 curfrags -= frags_per_mbuf(n);
1634 curfrags -= frags_per_mbuf(n2);
1637 if (curfrags <= maxfrags)
1640 * Still not there, try the normal collapse
1641 * again before we allocate another cluster.
1648 * No place where we can collapse to a cluster; punt.
1649 * This can occur if, for example, you request 2 frags
1650 * but the packet requires that both be clusters (we
1651 * never reallocate the first mbuf to avoid moving the
1658 #ifdef MBUF_STRESS_TEST
1661 * Fragment an mbuf chain. There's no reason you'd ever want to do
1662 * this in normal usage, but it's great for stress testing various
1665 * If fragmentation is not possible, the original chain will be
1668 * Possible length values:
1669 * 0 no fragmentation will occur
1670 * > 0 each fragment will be of the specified length
1671 * -1 each fragment will be the same random value in length
1672 * -2 each fragment's length will be entirely random
1673 * (Random values range from 1 to 256)
1676 m_fragment(struct mbuf *m0, int how, int length)
1678 struct mbuf *m_first, *m_last;
1679 int divisor = 255, progress = 0, fraglen;
1681 if (!(m0->m_flags & M_PKTHDR))
1684 if (length == 0 || length < -2)
1686 if (length > MCLBYTES)
1688 if (length < 0 && divisor > MCLBYTES)
1691 length = 1 + (arc4random() % divisor);
1695 m_fixhdr(m0); /* Needed sanity check */
1697 m_first = m_getcl(how, MT_DATA, M_PKTHDR);
1698 if (m_first == NULL)
1701 if (m_dup_pkthdr(m_first, m0, how) == 0)
1706 while (progress < m0->m_pkthdr.len) {
1708 fraglen = 1 + (arc4random() % divisor);
1709 if (fraglen > m0->m_pkthdr.len - progress)
1710 fraglen = m0->m_pkthdr.len - progress;
1712 if (progress != 0) {
1713 struct mbuf *m_new = m_getcl(how, MT_DATA, 0);
1717 m_last->m_next = m_new;
1721 m_copydata(m0, progress, fraglen, mtod(m_last, caddr_t));
1722 progress += fraglen;
1723 m_last->m_len = fraglen;
1731 /* Return the original chain on failure */
1738 * Free pages from mbuf_ext_pgs, assuming they were allocated via
1739 * vm_page_alloc() and aren't associated with any object. Complement
1740 * to allocator from m_uiotombuf_nomap().
1743 mb_free_mext_pgs(struct mbuf *m)
1748 for (int i = 0; i < m->m_epg_npgs; i++) {
1749 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
1750 vm_page_unwire_noq(pg);
1755 static struct mbuf *
1756 m_uiotombuf_nomap(struct uio *uio, int how, int len, int maxseg, int flags)
1758 struct mbuf *m, *mb, *prev;
1759 vm_page_t pg_array[MBUF_PEXT_MAX_PGS];
1760 int error, length, i, needed;
1762 int pflags = malloc2vm_flags(how) | VM_ALLOC_NODUMP | VM_ALLOC_WIRED;
1764 MPASS((flags & M_PKTHDR) == 0);
1765 MPASS((how & M_ZERO) == 0);
1768 * len can be zero or an arbitrary large value bound by
1769 * the total data supplied by the uio.
1772 total = MIN(uio->uio_resid, len);
1774 total = uio->uio_resid;
1777 maxseg = MBUF_PEXT_MAX_PGS * PAGE_SIZE;
1780 * If total is zero, return an empty mbuf. This can occur
1781 * for TLS 1.0 connections which send empty fragments as
1782 * a countermeasure against the known-IV weakness in CBC
1785 if (__predict_false(total == 0)) {
1786 mb = mb_alloc_ext_pgs(how, mb_free_mext_pgs);
1789 mb->m_epg_flags = EPG_FLAG_ANON;
1794 * Allocate the pages
1798 mb = mb_alloc_ext_pgs(how, mb_free_mext_pgs);
1806 mb->m_epg_flags = EPG_FLAG_ANON;
1807 needed = length = MIN(maxseg, total);
1808 for (i = 0; needed > 0; i++, needed -= PAGE_SIZE) {
1810 pg_array[i] = vm_page_alloc_noobj(pflags);
1811 if (pg_array[i] == NULL) {
1812 if (how & M_NOWAIT) {
1819 mb->m_epg_pa[i] = VM_PAGE_TO_PHYS(pg_array[i]);
1822 mb->m_epg_last_len = length - PAGE_SIZE * (mb->m_epg_npgs - 1);
1823 MBUF_EXT_PGS_ASSERT_SANITY(mb);
1825 error = uiomove_fromphys(pg_array, 0, length, uio);
1829 mb->m_ext.ext_size += PAGE_SIZE * mb->m_epg_npgs;
1830 if (flags & M_PKTHDR)
1831 m->m_pkthdr.len += length;
1841 * Copy the contents of uio into a properly sized mbuf chain.
1844 m_uiotombuf(struct uio *uio, int how, int len, int align, int flags)
1846 struct mbuf *m, *mb;
1851 if (flags & M_EXTPG)
1852 return (m_uiotombuf_nomap(uio, how, len, align, flags));
1855 * len can be zero or an arbitrary large value bound by
1856 * the total data supplied by the uio.
1859 total = (uio->uio_resid < len) ? uio->uio_resid : len;
1861 total = uio->uio_resid;
1864 * The smallest unit returned by m_getm2() is a single mbuf
1865 * with pkthdr. We can't align past it.
1871 * Give us the full allocation or nothing.
1872 * If len is zero return the smallest empty mbuf.
1874 m = m_getm2(NULL, max(total + align, 1), how, MT_DATA, flags);
1879 /* Fill all mbufs with uio data and update header information. */
1880 for (mb = m; mb != NULL; mb = mb->m_next) {
1881 length = min(M_TRAILINGSPACE(mb), total - progress);
1883 error = uiomove(mtod(mb, void *), length, uio);
1891 if (flags & M_PKTHDR)
1892 m->m_pkthdr.len += length;
1894 KASSERT(progress == total, ("%s: progress != total", __func__));
1900 * Copy data to/from an unmapped mbuf into a uio limited by len if set.
1903 m_unmapped_uiomove(const struct mbuf *m, int m_off, struct uio *uio, int len)
1906 int error, i, off, pglen, pgoff, seglen, segoff;
1911 /* Skip over any data removed from the front. */
1912 off = mtod(m, vm_offset_t);
1915 if (m->m_epg_hdrlen != 0) {
1916 if (off >= m->m_epg_hdrlen) {
1917 off -= m->m_epg_hdrlen;
1919 seglen = m->m_epg_hdrlen - off;
1921 seglen = min(seglen, len);
1924 error = uiomove(__DECONST(void *,
1925 &m->m_epg_hdr[segoff]), seglen, uio);
1928 pgoff = m->m_epg_1st_off;
1929 for (i = 0; i < m->m_epg_npgs && error == 0 && len > 0; i++) {
1930 pglen = m_epg_pagelen(m, i, pgoff);
1936 seglen = pglen - off;
1937 segoff = pgoff + off;
1939 seglen = min(seglen, len);
1941 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
1942 error = uiomove_fromphys(&pg, segoff, seglen, uio);
1945 if (len != 0 && error == 0) {
1946 KASSERT((off + len) <= m->m_epg_trllen,
1947 ("off + len > trail (%d + %d > %d, m_off = %d)", off, len,
1948 m->m_epg_trllen, m_off));
1949 error = uiomove(__DECONST(void *, &m->m_epg_trail[off]),
1956 * Copy an mbuf chain into a uio limited by len if set.
1959 m_mbuftouio(struct uio *uio, const struct mbuf *m, int len)
1961 int error, length, total;
1965 total = min(uio->uio_resid, len);
1967 total = uio->uio_resid;
1969 /* Fill the uio with data from the mbufs. */
1970 for (; m != NULL; m = m->m_next) {
1971 length = min(m->m_len, total - progress);
1973 if ((m->m_flags & M_EXTPG) != 0)
1974 error = m_unmapped_uiomove(m, 0, uio, length);
1976 error = uiomove(mtod(m, void *), length, uio);
1987 * Create a writable copy of the mbuf chain. While doing this
1988 * we compact the chain with a goal of producing a chain with
1989 * at most two mbufs. The second mbuf in this chain is likely
1990 * to be a cluster. The primary purpose of this work is to create
1991 * a writable packet for encryption, compression, etc. The
1992 * secondary goal is to linearize the data so the data can be
1993 * passed to crypto hardware in the most efficient manner possible.
1996 m_unshare(struct mbuf *m0, int how)
1998 struct mbuf *m, *mprev;
1999 struct mbuf *n, *mfirst, *mlast;
2003 for (m = m0; m != NULL; m = mprev->m_next) {
2005 * Regular mbufs are ignored unless there's a cluster
2006 * in front of it that we can use to coalesce. We do
2007 * the latter mainly so later clusters can be coalesced
2008 * also w/o having to handle them specially (i.e. convert
2009 * mbuf+cluster -> cluster). This optimization is heavily
2010 * influenced by the assumption that we're running over
2011 * Ethernet where MCLBYTES is large enough that the max
2012 * packet size will permit lots of coalescing into a
2013 * single cluster. This in turn permits efficient
2014 * crypto operations, especially when using hardware.
2016 if ((m->m_flags & M_EXT) == 0) {
2017 if (mprev && (mprev->m_flags & M_EXT) &&
2018 m->m_len <= M_TRAILINGSPACE(mprev)) {
2019 /* XXX: this ignores mbuf types */
2020 memcpy(mtod(mprev, caddr_t) + mprev->m_len,
2021 mtod(m, caddr_t), m->m_len);
2022 mprev->m_len += m->m_len;
2023 mprev->m_next = m->m_next; /* unlink from chain */
2024 m_free(m); /* reclaim mbuf */
2031 * Writable mbufs are left alone (for now).
2033 if (M_WRITABLE(m)) {
2039 * Not writable, replace with a copy or coalesce with
2040 * the previous mbuf if possible (since we have to copy
2041 * it anyway, we try to reduce the number of mbufs and
2042 * clusters so that future work is easier).
2044 KASSERT(m->m_flags & M_EXT, ("m_flags 0x%x", m->m_flags));
2045 /* NB: we only coalesce into a cluster or larger */
2046 if (mprev != NULL && (mprev->m_flags & M_EXT) &&
2047 m->m_len <= M_TRAILINGSPACE(mprev)) {
2048 /* XXX: this ignores mbuf types */
2049 memcpy(mtod(mprev, caddr_t) + mprev->m_len,
2050 mtod(m, caddr_t), m->m_len);
2051 mprev->m_len += m->m_len;
2052 mprev->m_next = m->m_next; /* unlink from chain */
2053 m_free(m); /* reclaim mbuf */
2058 * Allocate new space to hold the copy and copy the data.
2059 * We deal with jumbo mbufs (i.e. m_len > MCLBYTES) by
2060 * splitting them into clusters. We could just malloc a
2061 * buffer and make it external but too many device drivers
2062 * don't know how to break up the non-contiguous memory when
2065 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS);
2070 if (m->m_flags & M_PKTHDR) {
2071 KASSERT(mprev == NULL, ("%s: m0 %p, m %p has M_PKTHDR",
2073 m_move_pkthdr(n, m);
2080 int cc = min(len, MCLBYTES);
2081 memcpy(mtod(n, caddr_t), mtod(m, caddr_t) + off, cc);
2087 newipsecstat.ips_clcopied++;
2095 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS);
2102 n->m_next = m->m_next;
2104 m0 = mfirst; /* new head of chain */
2106 mprev->m_next = mfirst; /* replace old mbuf */
2107 m_free(m); /* release old mbuf */
2113 #ifdef MBUF_PROFILING
2115 #define MP_BUCKETS 32 /* don't just change this as things may overflow.*/
2116 struct mbufprofile {
2117 uintmax_t wasted[MP_BUCKETS];
2118 uintmax_t used[MP_BUCKETS];
2119 uintmax_t segments[MP_BUCKETS];
2123 m_profile(struct mbuf *m)
2132 if (m->m_flags & M_EXT) {
2133 wasted += MHLEN - sizeof(m->m_ext) +
2134 m->m_ext.ext_size - m->m_len;
2136 if (m->m_flags & M_PKTHDR)
2137 wasted += MHLEN - m->m_len;
2139 wasted += MLEN - m->m_len;
2143 /* be paranoid.. it helps */
2144 if (segments > MP_BUCKETS - 1)
2145 segments = MP_BUCKETS - 1;
2148 if (wasted > 100000)
2150 /* store in the appropriate bucket */
2151 /* don't bother locking. if it's slightly off, so what? */
2152 mbprof.segments[segments]++;
2153 mbprof.used[fls(used)]++;
2154 mbprof.wasted[fls(wasted)]++;
2158 mbprof_handler(SYSCTL_HANDLER_ARGS)
2165 sbuf_new_for_sysctl(&sb, buf, sizeof(buf), req);
2167 p = &mbprof.wasted[0];
2170 "%ju %ju %ju %ju %ju %ju %ju %ju "
2171 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
2172 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
2173 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
2175 p = &mbprof.wasted[16];
2177 "%ju %ju %ju %ju %ju %ju %ju %ju "
2178 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
2179 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
2180 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
2182 p = &mbprof.used[0];
2185 "%ju %ju %ju %ju %ju %ju %ju %ju "
2186 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
2187 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
2188 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
2190 p = &mbprof.used[16];
2192 "%ju %ju %ju %ju %ju %ju %ju %ju "
2193 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
2194 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
2195 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
2197 p = &mbprof.segments[0];
2200 "%ju %ju %ju %ju %ju %ju %ju %ju "
2201 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
2202 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
2203 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
2205 p = &mbprof.segments[16];
2207 "%ju %ju %ju %ju %ju %ju %ju %ju "
2208 "%ju %ju %ju %ju %ju %ju %ju %jju",
2209 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
2210 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
2213 error = sbuf_finish(&sb);
2219 mbprof_clr_handler(SYSCTL_HANDLER_ARGS)
2224 error = sysctl_handle_int(oidp, &clear, 0, req);
2225 if (error || !req->newptr)
2229 bzero(&mbprof, sizeof(mbprof));
2235 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofile,
2236 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
2237 mbprof_handler, "A",
2238 "mbuf profiling statistics");
2240 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofileclr,
2241 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
2242 mbprof_clr_handler, "I",
2243 "clear mbuf profiling statistics");