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 #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>
50 #include <sys/vmmeter.h>
54 #include <vm/vm_pageout.h>
55 #include <vm/vm_page.h>
57 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__init,
58 "struct mbuf *", "mbufinfo_t *",
59 "uint32_t", "uint32_t",
60 "uint16_t", "uint16_t",
61 "uint32_t", "uint32_t",
62 "uint32_t", "uint32_t");
64 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__gethdr,
65 "uint32_t", "uint32_t",
66 "uint16_t", "uint16_t",
67 "struct mbuf *", "mbufinfo_t *");
69 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__get,
70 "uint32_t", "uint32_t",
71 "uint16_t", "uint16_t",
72 "struct mbuf *", "mbufinfo_t *");
74 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__getcl,
75 "uint32_t", "uint32_t",
76 "uint16_t", "uint16_t",
77 "uint32_t", "uint32_t",
78 "struct mbuf *", "mbufinfo_t *");
80 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__getjcl,
81 "uint32_t", "uint32_t",
82 "uint16_t", "uint16_t",
83 "uint32_t", "uint32_t",
84 "uint32_t", "uint32_t",
85 "struct mbuf *", "mbufinfo_t *");
87 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__clget,
88 "struct mbuf *", "mbufinfo_t *",
89 "uint32_t", "uint32_t",
90 "uint32_t", "uint32_t");
92 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__cljget,
93 "struct mbuf *", "mbufinfo_t *",
94 "uint32_t", "uint32_t",
95 "uint32_t", "uint32_t",
98 SDT_PROBE_DEFINE(sdt, , , m__cljset);
100 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__free,
101 "struct mbuf *", "mbufinfo_t *");
103 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__freem,
104 "struct mbuf *", "mbufinfo_t *");
106 #include <security/mac/mac_framework.h>
112 #ifdef MBUF_STRESS_TEST
117 int m_defragrandomfailures;
121 * sysctl(8) exported objects
123 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RD,
124 &max_linkhdr, 0, "Size of largest link layer header");
125 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RD,
126 &max_protohdr, 0, "Size of largest protocol layer header");
127 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RD,
128 &max_hdr, 0, "Size of largest link plus protocol header");
129 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RD,
130 &max_datalen, 0, "Minimum space left in mbuf after max_hdr");
131 #ifdef MBUF_STRESS_TEST
132 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
133 &m_defragpackets, 0, "");
134 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
135 &m_defragbytes, 0, "");
136 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
137 &m_defraguseless, 0, "");
138 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
139 &m_defragfailure, 0, "");
140 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
141 &m_defragrandomfailures, 0, "");
145 * Ensure the correct size of various mbuf parameters. It could be off due
146 * to compiler-induced padding and alignment artifacts.
148 CTASSERT(MSIZE - offsetof(struct mbuf, m_dat) == MLEN);
149 CTASSERT(MSIZE - offsetof(struct mbuf, m_pktdat) == MHLEN);
152 * mbuf data storage should be 64-bit aligned regardless of architectural
153 * pointer size; check this is the case with and without a packet header.
155 CTASSERT(offsetof(struct mbuf, m_dat) % 8 == 0);
156 CTASSERT(offsetof(struct mbuf, m_pktdat) % 8 == 0);
159 * While the specific values here don't matter too much (i.e., +/- a few
160 * words), we do want to ensure that changes to these values are carefully
161 * reasoned about and properly documented. This is especially the case as
162 * network-protocol and device-driver modules encode these layouts, and must
163 * be recompiled if the structures change. Check these values at compile time
164 * against the ones documented in comments in mbuf.h.
166 * NB: Possibly they should be documented there via #define's and not just
169 #if defined(__LP64__)
170 CTASSERT(offsetof(struct mbuf, m_dat) == 32);
171 CTASSERT(sizeof(struct pkthdr) == 56);
172 CTASSERT(sizeof(struct m_ext) == 160);
174 CTASSERT(offsetof(struct mbuf, m_dat) == 24);
175 CTASSERT(sizeof(struct pkthdr) == 48);
176 #if defined(__powerpc__) && defined(BOOKE)
177 /* PowerPC booke has 64-bit physical pointers. */
178 CTASSERT(sizeof(struct m_ext) == 184);
180 CTASSERT(sizeof(struct m_ext) == 180);
185 * Assert that the queue(3) macros produce code of the same size as an old
186 * plain pointer does.
189 static struct mbuf __used m_assertbuf;
190 CTASSERT(sizeof(m_assertbuf.m_slist) == sizeof(m_assertbuf.m_next));
191 CTASSERT(sizeof(m_assertbuf.m_stailq) == sizeof(m_assertbuf.m_next));
192 CTASSERT(sizeof(m_assertbuf.m_slistpkt) == sizeof(m_assertbuf.m_nextpkt));
193 CTASSERT(sizeof(m_assertbuf.m_stailqpkt) == sizeof(m_assertbuf.m_nextpkt));
197 * Attach the cluster from *m to *n, set up m_ext in *n
198 * and bump the refcount of the cluster.
201 mb_dupcl(struct mbuf *n, struct mbuf *m)
203 volatile u_int *refcnt;
205 KASSERT(m->m_flags & (M_EXT|M_EXTPG),
206 ("%s: M_EXT|M_EXTPG not set on %p", __func__, m));
207 KASSERT(!(n->m_flags & (M_EXT|M_EXTPG)),
208 ("%s: M_EXT|M_EXTPG set on %p", __func__, n));
211 * Cache access optimization.
213 * o Regular M_EXT storage doesn't need full copy of m_ext, since
214 * the holder of the 'ext_count' is responsible to carry the free
215 * routine and its arguments.
216 * o M_EXTPG data is split between main part of mbuf and m_ext, the
217 * main part is copied in full, the m_ext part is similar to M_EXT.
218 * o EXT_EXTREF, where 'ext_cnt' doesn't point into mbuf at all, is
219 * special - it needs full copy of m_ext into each mbuf, since any
220 * copy could end up as the last to free.
222 if (m->m_flags & M_EXTPG) {
223 bcopy(&m->m_epg_startcopy, &n->m_epg_startcopy,
224 __rangeof(struct mbuf, m_epg_startcopy, m_epg_endcopy));
225 bcopy(&m->m_ext, &n->m_ext, m_epg_ext_copylen);
226 } else if (m->m_ext.ext_type == EXT_EXTREF)
227 bcopy(&m->m_ext, &n->m_ext, sizeof(struct m_ext));
229 bcopy(&m->m_ext, &n->m_ext, m_ext_copylen);
231 n->m_flags |= m->m_flags & (M_RDONLY | M_EXT | M_EXTPG);
233 /* See if this is the mbuf that holds the embedded refcount. */
234 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
235 refcnt = n->m_ext.ext_cnt = &m->m_ext.ext_count;
236 n->m_ext.ext_flags &= ~EXT_FLAG_EMBREF;
238 KASSERT(m->m_ext.ext_cnt != NULL,
239 ("%s: no refcounting pointer on %p", __func__, m));
240 refcnt = m->m_ext.ext_cnt;
246 atomic_add_int(refcnt, 1);
250 m_demote_pkthdr(struct mbuf *m)
254 M_ASSERT_NO_SND_TAG(m);
256 m_tag_delete_chain(m, NULL);
257 m->m_flags &= ~M_PKTHDR;
258 bzero(&m->m_pkthdr, sizeof(struct pkthdr));
262 * Clean up mbuf (chain) from any tags and packet headers.
263 * If "all" is set then the first mbuf in the chain will be
267 m_demote(struct mbuf *m0, int all, int flags)
271 flags |= M_DEMOTEFLAGS;
273 for (m = all ? m0 : m0->m_next; m != NULL; m = m->m_next) {
274 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt in m %p, m0 %p",
276 if (m->m_flags & M_PKTHDR)
283 * Sanity checks on mbuf (chain) for use in KASSERT() and general
285 * Returns 0 or panics when bad and 1 on all tests passed.
286 * Sanitize, 0 to run M_SANITY_ACTION, 1 to garble things so they
290 m_sanity(struct mbuf *m0, int sanitize)
297 #define M_SANITY_ACTION(s) panic("mbuf %p: " s, m)
299 #define M_SANITY_ACTION(s) printf("mbuf %p: " s, m)
302 for (m = m0; m != NULL; m = m->m_next) {
304 * Basic pointer checks. If any of these fails then some
305 * unrelated kernel memory before or after us is trashed.
306 * No way to recover from that.
310 if ((caddr_t)m->m_data < a)
311 M_SANITY_ACTION("m_data outside mbuf data range left");
312 if ((caddr_t)m->m_data > b)
313 M_SANITY_ACTION("m_data outside mbuf data range right");
314 if ((caddr_t)m->m_data + m->m_len > b)
315 M_SANITY_ACTION("m_data + m_len exeeds mbuf space");
317 /* m->m_nextpkt may only be set on first mbuf in chain. */
318 if (m != m0 && m->m_nextpkt != NULL) {
320 m_freem(m->m_nextpkt);
321 m->m_nextpkt = (struct mbuf *)0xDEADC0DE;
323 M_SANITY_ACTION("m->m_nextpkt on in-chain mbuf");
326 /* packet length (not mbuf length!) calculation */
327 if (m0->m_flags & M_PKTHDR)
330 /* m_tags may only be attached to first mbuf in chain. */
331 if (m != m0 && m->m_flags & M_PKTHDR &&
332 !SLIST_EMPTY(&m->m_pkthdr.tags)) {
334 m_tag_delete_chain(m, NULL);
335 /* put in 0xDEADC0DE perhaps? */
337 M_SANITY_ACTION("m_tags on in-chain mbuf");
340 /* M_PKTHDR may only be set on first mbuf in chain */
341 if (m != m0 && m->m_flags & M_PKTHDR) {
343 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
344 m->m_flags &= ~M_PKTHDR;
345 /* put in 0xDEADCODE and leave hdr flag in */
347 M_SANITY_ACTION("M_PKTHDR on in-chain mbuf");
351 if (pktlen && pktlen != m->m_pkthdr.len) {
355 M_SANITY_ACTION("m_pkthdr.len != mbuf chain length");
359 #undef M_SANITY_ACTION
363 * Non-inlined part of m_init().
366 m_pkthdr_init(struct mbuf *m, int how)
371 m->m_data = m->m_pktdat;
372 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
374 m->m_pkthdr.numa_domain = M_NODOM;
377 /* If the label init fails, fail the alloc */
378 error = mac_mbuf_init(m, how);
387 * "Move" mbuf pkthdr from "from" to "to".
388 * "from" must have M_PKTHDR set, and "to" must be empty.
391 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
395 /* see below for why these are not enabled */
397 /* Note: with MAC, this may not be a good assertion. */
398 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags),
399 ("m_move_pkthdr: to has tags"));
403 * XXXMAC: It could be this should also occur for non-MAC?
405 if (to->m_flags & M_PKTHDR)
406 m_tag_delete_chain(to, NULL);
408 to->m_flags = (from->m_flags & M_COPYFLAGS) |
409 (to->m_flags & (M_EXT | M_EXTPG));
410 if ((to->m_flags & M_EXT) == 0)
411 to->m_data = to->m_pktdat;
412 to->m_pkthdr = from->m_pkthdr; /* especially tags */
413 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */
414 from->m_flags &= ~M_PKTHDR;
415 if (from->m_pkthdr.csum_flags & CSUM_SND_TAG) {
416 from->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
417 from->m_pkthdr.snd_tag = NULL;
422 * Duplicate "from"'s mbuf pkthdr in "to".
423 * "from" must have M_PKTHDR set, and "to" must be empty.
424 * In particular, this does a deep copy of the packet tags.
427 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
432 * The mbuf allocator only initializes the pkthdr
433 * when the mbuf is allocated with m_gethdr(). Many users
434 * (e.g. m_copy*, m_prepend) use m_get() and then
435 * smash the pkthdr as needed causing these
436 * assertions to trip. For now just disable them.
439 /* Note: with MAC, this may not be a good assertion. */
440 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags), ("m_dup_pkthdr: to has tags"));
442 MBUF_CHECKSLEEP(how);
444 if (to->m_flags & M_PKTHDR)
445 m_tag_delete_chain(to, NULL);
447 to->m_flags = (from->m_flags & M_COPYFLAGS) |
448 (to->m_flags & (M_EXT | M_EXTPG));
449 if ((to->m_flags & M_EXT) == 0)
450 to->m_data = to->m_pktdat;
451 to->m_pkthdr = from->m_pkthdr;
452 if (from->m_pkthdr.csum_flags & CSUM_SND_TAG)
453 m_snd_tag_ref(from->m_pkthdr.snd_tag);
454 SLIST_INIT(&to->m_pkthdr.tags);
455 return (m_tag_copy_chain(to, from, how));
459 * Lesser-used path for M_PREPEND:
460 * allocate new mbuf to prepend to chain,
464 m_prepend(struct mbuf *m, int len, int how)
468 if (m->m_flags & M_PKTHDR)
469 mn = m_gethdr(how, m->m_type);
471 mn = m_get(how, m->m_type);
476 if (m->m_flags & M_PKTHDR)
477 m_move_pkthdr(mn, m);
487 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
488 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
489 * The wait parameter is a choice of M_WAITOK/M_NOWAIT from caller.
490 * Note that the copy is read-only, because clusters are not copied,
491 * only their reference counts are incremented.
494 m_copym(struct mbuf *m, int off0, int len, int wait)
496 struct mbuf *n, **np;
501 KASSERT(off >= 0, ("m_copym, negative off %d", off));
502 KASSERT(len >= 0, ("m_copym, negative len %d", len));
503 MBUF_CHECKSLEEP(wait);
504 if (off == 0 && m->m_flags & M_PKTHDR)
507 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
517 KASSERT(len == M_COPYALL,
518 ("m_copym, length > size of mbuf chain"));
522 n = m_gethdr(wait, m->m_type);
524 n = m_get(wait, m->m_type);
529 if (!m_dup_pkthdr(n, m, wait))
531 if (len == M_COPYALL)
532 n->m_pkthdr.len -= off0;
534 n->m_pkthdr.len = len;
537 n->m_len = min(len, m->m_len - off);
538 if (m->m_flags & (M_EXT|M_EXTPG)) {
539 n->m_data = m->m_data + off;
542 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
544 if (len != M_COPYALL)
558 * Copy an entire packet, including header (which must be present).
559 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
560 * Note that the copy is read-only, because clusters are not copied,
561 * only their reference counts are incremented.
562 * Preserve alignment of the first mbuf so if the creator has left
563 * some room at the beginning (e.g. for inserting protocol headers)
564 * the copies still have the room available.
567 m_copypacket(struct mbuf *m, int how)
569 struct mbuf *top, *n, *o;
571 MBUF_CHECKSLEEP(how);
572 n = m_get(how, m->m_type);
577 if (!m_dup_pkthdr(n, m, how))
580 if (m->m_flags & (M_EXT|M_EXTPG)) {
581 n->m_data = m->m_data;
584 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
585 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
590 o = m_get(how, m->m_type);
598 if (m->m_flags & (M_EXT|M_EXTPG)) {
599 n->m_data = m->m_data;
602 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
614 m_copyfromunmapped(const struct mbuf *m, int off, int len, caddr_t cp)
620 KASSERT(off >= 0, ("m_copyfromunmapped: negative off %d", off));
621 KASSERT(len >= 0, ("m_copyfromunmapped: negative len %d", len));
622 KASSERT(off < m->m_len,
623 ("m_copyfromunmapped: len exceeds mbuf length"));
628 uio.uio_segflg = UIO_SYSSPACE;
631 uio.uio_rw = UIO_READ;
632 error = m_unmapped_uiomove(m, off, &uio, len);
633 KASSERT(error == 0, ("m_unmapped_uiomove failed: off %d, len %d", off,
638 * Copy data from an mbuf chain starting "off" bytes from the beginning,
639 * continuing for "len" bytes, into the indicated buffer.
642 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
646 KASSERT(off >= 0, ("m_copydata, negative off %d", off));
647 KASSERT(len >= 0, ("m_copydata, negative len %d", len));
649 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
656 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
657 count = min(m->m_len - off, len);
658 if ((m->m_flags & M_EXTPG) != 0)
659 m_copyfromunmapped(m, off, count, cp);
661 bcopy(mtod(m, caddr_t) + off, cp, count);
670 * Copy a packet header mbuf chain into a completely new chain, including
671 * copying any mbuf clusters. Use this instead of m_copypacket() when
672 * you need a writable copy of an mbuf chain.
675 m_dup(const struct mbuf *m, int how)
677 struct mbuf **p, *top = NULL;
678 int remain, moff, nsize;
680 MBUF_CHECKSLEEP(how);
686 /* While there's more data, get a new mbuf, tack it on, and fill it */
687 remain = m->m_pkthdr.len;
690 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */
693 /* Get the next new mbuf */
694 if (remain >= MINCLSIZE) {
695 n = m_getcl(how, m->m_type, 0);
698 n = m_get(how, m->m_type);
704 if (top == NULL) { /* First one, must be PKTHDR */
705 if (!m_dup_pkthdr(n, m, how)) {
709 if ((n->m_flags & M_EXT) == 0)
711 n->m_flags &= ~M_RDONLY;
715 /* Link it into the new chain */
719 /* Copy data from original mbuf(s) into new mbuf */
720 while (n->m_len < nsize && m != NULL) {
721 int chunk = min(nsize - n->m_len, m->m_len - moff);
723 m_copydata(m, moff, chunk, n->m_data + n->m_len);
727 if (moff == m->m_len) {
733 /* Check correct total mbuf length */
734 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
735 ("%s: bogus m_pkthdr.len", __func__));
745 * Concatenate mbuf chain n to m.
746 * Both chains must be of the same type (e.g. MT_DATA).
747 * Any m_pkthdr is not updated.
750 m_cat(struct mbuf *m, struct mbuf *n)
755 if (!M_WRITABLE(m) ||
756 (n->m_flags & M_EXTPG) != 0 ||
757 M_TRAILINGSPACE(m) < n->m_len) {
758 /* just join the two chains */
762 /* splat the data from one into the other */
763 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
765 m->m_len += n->m_len;
771 * Concatenate two pkthdr mbuf chains.
774 m_catpkt(struct mbuf *m, struct mbuf *n)
780 m->m_pkthdr.len += n->m_pkthdr.len;
787 m_adj(struct mbuf *mp, int req_len)
793 if ((m = mp) == NULL)
799 while (m != NULL && len > 0) {
800 if (m->m_len <= len) {
810 if (mp->m_flags & M_PKTHDR)
811 mp->m_pkthdr.len -= (req_len - len);
814 * Trim from tail. Scan the mbuf chain,
815 * calculating its length and finding the last mbuf.
816 * If the adjustment only affects this mbuf, then just
817 * adjust and return. Otherwise, rescan and truncate
818 * after the remaining size.
824 if (m->m_next == (struct mbuf *)0)
828 if (m->m_len >= len) {
830 if (mp->m_flags & M_PKTHDR)
831 mp->m_pkthdr.len -= len;
838 * Correct length for chain is "count".
839 * Find the mbuf with last data, adjust its length,
840 * and toss data from remaining mbufs on chain.
843 if (m->m_flags & M_PKTHDR)
844 m->m_pkthdr.len = count;
845 for (; m; m = m->m_next) {
846 if (m->m_len >= count) {
848 if (m->m_next != NULL) {
860 m_adj_decap(struct mbuf *mp, int len)
865 if ((mp->m_flags & M_PKTHDR) != 0) {
867 * If flowid was calculated by card from the inner
868 * headers, move flowid to the decapsulated mbuf
869 * chain, otherwise clear. This depends on the
870 * internals of m_adj, which keeps pkthdr as is, in
871 * particular not changing rsstype and flowid.
873 rsstype = mp->m_pkthdr.rsstype;
874 if ((rsstype & M_HASHTYPE_INNER) != 0) {
875 M_HASHTYPE_SET(mp, rsstype & ~M_HASHTYPE_INNER);
877 M_HASHTYPE_CLEAR(mp);
883 * Rearange an mbuf chain so that len bytes are contiguous
884 * and in the data area of an mbuf (so that mtod will work
885 * for a structure of size len). Returns the resulting
886 * mbuf chain on success, frees it and returns null on failure.
887 * If there is room, it will add up to max_protohdr-len extra bytes to the
888 * contiguous region in an attempt to avoid being called next time.
891 m_pullup(struct mbuf *n, int len)
897 KASSERT((n->m_flags & M_EXTPG) == 0,
898 ("%s: unmapped mbuf %p", __func__, n));
901 * If first mbuf has no cluster, and has room for len bytes
902 * without shifting current data, pullup into it,
903 * otherwise allocate a new mbuf to prepend to the chain.
905 if ((n->m_flags & M_EXT) == 0 &&
906 n->m_data + len < &n->m_dat[MLEN] && n->m_next) {
915 m = m_get(M_NOWAIT, n->m_type);
918 if (n->m_flags & M_PKTHDR)
921 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
923 count = min(min(max(len, max_protohdr), space), n->m_len);
924 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
934 } while (len > 0 && n);
947 * Like m_pullup(), except a new mbuf is always allocated, and we allow
948 * the amount of empty space before the data in the new mbuf to be specified
949 * (in the event that the caller expects to prepend later).
952 m_copyup(struct mbuf *n, int len, int dstoff)
957 if (len > (MHLEN - dstoff))
959 m = m_get(M_NOWAIT, n->m_type);
962 if (n->m_flags & M_PKTHDR)
965 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
967 count = min(min(max(len, max_protohdr), space), n->m_len);
968 memcpy(mtod(m, caddr_t) + m->m_len, mtod(n, caddr_t),
978 } while (len > 0 && n);
991 * Partition an mbuf chain in two pieces, returning the tail --
992 * all but the first len0 bytes. In case of failure, it returns NULL and
993 * attempts to restore the chain to its original state.
995 * Note that the resulting mbufs might be read-only, because the new
996 * mbuf can end up sharing an mbuf cluster with the original mbuf if
997 * the "breaking point" happens to lie within a cluster mbuf. Use the
998 * M_WRITABLE() macro to check for this case.
1001 m_split(struct mbuf *m0, int len0, int wait)
1004 u_int len = len0, remain;
1006 MBUF_CHECKSLEEP(wait);
1007 for (m = m0; m && len > m->m_len; m = m->m_next)
1011 remain = m->m_len - len;
1012 if (m0->m_flags & M_PKTHDR && remain == 0) {
1013 n = m_gethdr(wait, m0->m_type);
1016 n->m_next = m->m_next;
1018 if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG) {
1019 n->m_pkthdr.snd_tag =
1020 m_snd_tag_ref(m0->m_pkthdr.snd_tag);
1021 n->m_pkthdr.csum_flags |= CSUM_SND_TAG;
1023 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1024 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1025 m0->m_pkthdr.len = len0;
1027 } else if (m0->m_flags & M_PKTHDR) {
1028 n = m_gethdr(wait, m0->m_type);
1031 if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG) {
1032 n->m_pkthdr.snd_tag =
1033 m_snd_tag_ref(m0->m_pkthdr.snd_tag);
1034 n->m_pkthdr.csum_flags |= CSUM_SND_TAG;
1036 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
1037 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
1038 m0->m_pkthdr.len = len0;
1039 if (m->m_flags & (M_EXT|M_EXTPG))
1041 if (remain > MHLEN) {
1042 /* m can't be the lead packet */
1044 n->m_next = m_split(m, len, wait);
1045 if (n->m_next == NULL) {
1054 } else if (remain == 0) {
1059 n = m_get(wait, m->m_type);
1065 if (m->m_flags & (M_EXT|M_EXTPG)) {
1066 n->m_data = m->m_data + len;
1069 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
1073 n->m_next = m->m_next;
1078 * Routine to copy from device local memory into mbufs.
1079 * Note that `off' argument is offset into first mbuf of target chain from
1080 * which to begin copying the data to.
1083 m_devget(char *buf, int totlen, int off, struct ifnet *ifp,
1084 void (*copy)(char *from, caddr_t to, u_int len))
1087 struct mbuf *top = NULL, **mp = ⊤
1090 if (off < 0 || off > MHLEN)
1093 while (totlen > 0) {
1094 if (top == NULL) { /* First one, must be PKTHDR */
1095 if (totlen + off >= MINCLSIZE) {
1096 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1099 m = m_gethdr(M_NOWAIT, MT_DATA);
1102 /* Place initial small packet/header at end of mbuf */
1103 if (m && totlen + off + max_linkhdr <= MHLEN) {
1104 m->m_data += max_linkhdr;
1110 m->m_pkthdr.rcvif = ifp;
1111 m->m_pkthdr.len = totlen;
1113 if (totlen + off >= MINCLSIZE) {
1114 m = m_getcl(M_NOWAIT, MT_DATA, 0);
1117 m = m_get(M_NOWAIT, MT_DATA);
1130 m->m_len = len = min(totlen, len);
1132 copy(buf, mtod(m, caddr_t), (u_int)len);
1134 bcopy(buf, mtod(m, caddr_t), (u_int)len);
1144 m_copytounmapped(const struct mbuf *m, int off, int len, c_caddr_t cp)
1150 KASSERT(off >= 0, ("m_copytounmapped: negative off %d", off));
1151 KASSERT(len >= 0, ("m_copytounmapped: negative len %d", len));
1152 KASSERT(off < m->m_len, ("m_copytounmapped: len exceeds mbuf length"));
1153 iov.iov_base = __DECONST(caddr_t, cp);
1155 uio.uio_resid = len;
1157 uio.uio_segflg = UIO_SYSSPACE;
1160 uio.uio_rw = UIO_WRITE;
1161 error = m_unmapped_uiomove(m, off, &uio, len);
1162 KASSERT(error == 0, ("m_unmapped_uiomove failed: off %d, len %d", off,
1167 * Copy data from a buffer back into the indicated mbuf chain,
1168 * starting "off" bytes from the beginning, extending the mbuf
1169 * chain if necessary.
1172 m_copyback(struct mbuf *m0, int off, int len, c_caddr_t cp)
1175 struct mbuf *m = m0, *n;
1180 while (off > (mlen = m->m_len)) {
1183 if (m->m_next == NULL) {
1184 n = m_get(M_NOWAIT, m->m_type);
1187 bzero(mtod(n, caddr_t), MLEN);
1188 n->m_len = min(MLEN, len + off);
1194 if (m->m_next == NULL && (len > m->m_len - off)) {
1195 m->m_len += min(len - (m->m_len - off),
1196 M_TRAILINGSPACE(m));
1198 mlen = min (m->m_len - off, len);
1199 if ((m->m_flags & M_EXTPG) != 0)
1200 m_copytounmapped(m, off, mlen, cp);
1202 bcopy(cp, off + mtod(m, caddr_t), (u_int)mlen);
1210 if (m->m_next == NULL) {
1211 n = m_get(M_NOWAIT, m->m_type);
1214 n->m_len = min(MLEN, len);
1219 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1220 m->m_pkthdr.len = totlen;
1224 * Append the specified data to the indicated mbuf chain,
1225 * Extend the mbuf chain if the new data does not fit in
1228 * Return 1 if able to complete the job; otherwise 0.
1231 m_append(struct mbuf *m0, int len, c_caddr_t cp)
1234 int remainder, space;
1236 for (m = m0; m->m_next != NULL; m = m->m_next)
1239 space = M_TRAILINGSPACE(m);
1242 * Copy into available space.
1244 if (space > remainder)
1246 bcopy(cp, mtod(m, caddr_t) + m->m_len, space);
1248 cp += space, remainder -= space;
1250 while (remainder > 0) {
1252 * Allocate a new mbuf; could check space
1253 * and allocate a cluster instead.
1255 n = m_get(M_NOWAIT, m->m_type);
1258 n->m_len = min(MLEN, remainder);
1259 bcopy(cp, mtod(n, caddr_t), n->m_len);
1260 cp += n->m_len, remainder -= n->m_len;
1264 if (m0->m_flags & M_PKTHDR)
1265 m0->m_pkthdr.len += len - remainder;
1266 return (remainder == 0);
1270 m_apply_extpg_one(struct mbuf *m, int off, int len,
1271 int (*f)(void *, void *, u_int), void *arg)
1274 u_int i, count, pgoff, pglen;
1277 KASSERT(PMAP_HAS_DMAP,
1278 ("m_apply_extpg_one does not support unmapped mbufs"));
1279 off += mtod(m, vm_offset_t);
1280 if (off < m->m_epg_hdrlen) {
1281 count = min(m->m_epg_hdrlen - off, len);
1282 rval = f(arg, m->m_epg_hdr + off, count);
1288 off -= m->m_epg_hdrlen;
1289 pgoff = m->m_epg_1st_off;
1290 for (i = 0; i < m->m_epg_npgs && len > 0; i++) {
1291 pglen = m_epg_pagelen(m, i, pgoff);
1293 count = min(pglen - off, len);
1294 p = (void *)PHYS_TO_DMAP(m->m_epg_pa[i] + pgoff + off);
1295 rval = f(arg, p, count);
1305 KASSERT(off < m->m_epg_trllen,
1306 ("m_apply_extpg_one: offset beyond trailer"));
1307 KASSERT(len <= m->m_epg_trllen - off,
1308 ("m_apply_extpg_one: length beyond trailer"));
1309 return (f(arg, m->m_epg_trail + off, len));
1314 /* Apply function f to the data in a single mbuf. */
1316 m_apply_one(struct mbuf *m, int off, int len,
1317 int (*f)(void *, void *, u_int), void *arg)
1319 if ((m->m_flags & M_EXTPG) != 0)
1320 return (m_apply_extpg_one(m, off, len, f, arg));
1322 return (f(arg, mtod(m, caddr_t) + off, len));
1326 * Apply function f to the data in an mbuf chain starting "off" bytes from
1327 * the beginning, continuing for "len" bytes.
1330 m_apply(struct mbuf *m, int off, int len,
1331 int (*f)(void *, void *, u_int), void *arg)
1336 KASSERT(off >= 0, ("m_apply, negative off %d", off));
1337 KASSERT(len >= 0, ("m_apply, negative len %d", len));
1339 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1346 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1347 count = min(m->m_len - off, len);
1348 rval = m_apply_one(m, off, count, f, arg);
1359 * Return a pointer to mbuf/offset of location in mbuf chain.
1362 m_getptr(struct mbuf *m, int loc, int *off)
1366 /* Normal end of search. */
1367 if (m->m_len > loc) {
1372 if (m->m_next == NULL) {
1374 /* Point at the end of valid data. */
1387 m_print(const struct mbuf *m, int maxlen)
1391 const struct mbuf *m2;
1394 printf("mbuf: %p\n", m);
1398 if (m->m_flags & M_PKTHDR)
1399 len = m->m_pkthdr.len;
1403 while (m2 != NULL && (len == -1 || len)) {
1405 if (maxlen != -1 && pdata > maxlen)
1407 printf("mbuf: %p len: %d, next: %p, %b%s", m2, m2->m_len,
1408 m2->m_next, m2->m_flags, "\20\20freelist\17skipfw"
1409 "\11proto5\10proto4\7proto3\6proto2\5proto1\4rdonly"
1410 "\3eor\2pkthdr\1ext", pdata ? "" : "\n");
1412 printf(", %*D\n", pdata, (u_char *)m2->m_data, "-");
1418 printf("%d bytes unaccounted for.\n", len);
1423 m_fixhdr(struct mbuf *m0)
1427 len = m_length(m0, NULL);
1428 m0->m_pkthdr.len = len;
1433 m_length(struct mbuf *m0, struct mbuf **last)
1439 for (m = m0; m != NULL; m = m->m_next) {
1441 if (m->m_next == NULL)
1450 * Defragment a mbuf chain, returning the shortest possible
1451 * chain of mbufs and clusters. If allocation fails and
1452 * this cannot be completed, NULL will be returned, but
1453 * the passed in chain will be unchanged. Upon success,
1454 * the original chain will be freed, and the new chain
1457 * If a non-packet header is passed in, the original
1458 * mbuf (chain?) will be returned unharmed.
1461 m_defrag(struct mbuf *m0, int how)
1463 struct mbuf *m_new = NULL, *m_final = NULL;
1464 int progress = 0, length;
1466 MBUF_CHECKSLEEP(how);
1467 if (!(m0->m_flags & M_PKTHDR))
1470 m_fixhdr(m0); /* Needed sanity check */
1472 #ifdef MBUF_STRESS_TEST
1473 if (m_defragrandomfailures) {
1474 int temp = arc4random() & 0xff;
1480 if (m0->m_pkthdr.len > MHLEN)
1481 m_final = m_getcl(how, MT_DATA, M_PKTHDR);
1483 m_final = m_gethdr(how, MT_DATA);
1485 if (m_final == NULL)
1488 if (m_dup_pkthdr(m_final, m0, how) == 0)
1493 while (progress < m0->m_pkthdr.len) {
1494 length = m0->m_pkthdr.len - progress;
1495 if (length > MCLBYTES)
1498 if (m_new == NULL) {
1500 m_new = m_getcl(how, MT_DATA, 0);
1502 m_new = m_get(how, MT_DATA);
1507 m_copydata(m0, progress, length, mtod(m_new, caddr_t));
1509 m_new->m_len = length;
1510 if (m_new != m_final)
1511 m_cat(m_final, m_new);
1514 #ifdef MBUF_STRESS_TEST
1515 if (m0->m_next == NULL)
1520 #ifdef MBUF_STRESS_TEST
1522 m_defragbytes += m0->m_pkthdr.len;
1526 #ifdef MBUF_STRESS_TEST
1535 * Return the number of fragments an mbuf will use. This is usually
1536 * used as a proxy for the number of scatter/gather elements needed by
1537 * a DMA engine to access an mbuf. In general mapped mbufs are
1538 * assumed to be backed by physically contiguous buffers that only
1539 * need a single fragment. Unmapped mbufs, on the other hand, can
1540 * span disjoint physical pages.
1543 frags_per_mbuf(struct mbuf *m)
1547 if ((m->m_flags & M_EXTPG) == 0)
1551 * The header and trailer are counted as a single fragment
1552 * each when present.
1554 * XXX: This overestimates the number of fragments by assuming
1555 * all the backing physical pages are disjoint.
1558 if (m->m_epg_hdrlen != 0)
1560 frags += m->m_epg_npgs;
1561 if (m->m_epg_trllen != 0)
1568 * Defragment an mbuf chain, returning at most maxfrags separate
1569 * mbufs+clusters. If this is not possible NULL is returned and
1570 * the original mbuf chain is left in its present (potentially
1571 * modified) state. We use two techniques: collapsing consecutive
1572 * mbufs and replacing consecutive mbufs by a cluster.
1574 * NB: this should really be named m_defrag but that name is taken
1577 m_collapse(struct mbuf *m0, int how, int maxfrags)
1579 struct mbuf *m, *n, *n2, **prev;
1583 * Calculate the current number of frags.
1586 for (m = m0; m != NULL; m = m->m_next)
1587 curfrags += frags_per_mbuf(m);
1589 * First, try to collapse mbufs. Note that we always collapse
1590 * towards the front so we don't need to deal with moving the
1591 * pkthdr. This may be suboptimal if the first mbuf has much
1592 * less data than the following.
1600 if (M_WRITABLE(m) &&
1601 n->m_len < M_TRAILINGSPACE(m)) {
1602 m_copydata(n, 0, n->m_len,
1603 mtod(m, char *) + m->m_len);
1604 m->m_len += n->m_len;
1605 m->m_next = n->m_next;
1606 curfrags -= frags_per_mbuf(n);
1608 if (curfrags <= maxfrags)
1613 KASSERT(maxfrags > 1,
1614 ("maxfrags %u, but normal collapse failed", maxfrags));
1616 * Collapse consecutive mbufs to a cluster.
1618 prev = &m0->m_next; /* NB: not the first mbuf */
1619 while ((n = *prev) != NULL) {
1620 if ((n2 = n->m_next) != NULL &&
1621 n->m_len + n2->m_len < MCLBYTES) {
1622 m = m_getcl(how, MT_DATA, 0);
1625 m_copydata(n, 0, n->m_len, mtod(m, char *));
1626 m_copydata(n2, 0, n2->m_len,
1627 mtod(m, char *) + n->m_len);
1628 m->m_len = n->m_len + n2->m_len;
1629 m->m_next = n2->m_next;
1631 curfrags += 1; /* For the new cluster */
1632 curfrags -= frags_per_mbuf(n);
1633 curfrags -= frags_per_mbuf(n2);
1636 if (curfrags <= maxfrags)
1639 * Still not there, try the normal collapse
1640 * again before we allocate another cluster.
1647 * No place where we can collapse to a cluster; punt.
1648 * This can occur if, for example, you request 2 frags
1649 * but the packet requires that both be clusters (we
1650 * never reallocate the first mbuf to avoid moving the
1657 #ifdef MBUF_STRESS_TEST
1660 * Fragment an mbuf chain. There's no reason you'd ever want to do
1661 * this in normal usage, but it's great for stress testing various
1664 * If fragmentation is not possible, the original chain will be
1667 * Possible length values:
1668 * 0 no fragmentation will occur
1669 * > 0 each fragment will be of the specified length
1670 * -1 each fragment will be the same random value in length
1671 * -2 each fragment's length will be entirely random
1672 * (Random values range from 1 to 256)
1675 m_fragment(struct mbuf *m0, int how, int length)
1677 struct mbuf *m_first, *m_last;
1678 int divisor = 255, progress = 0, fraglen;
1680 if (!(m0->m_flags & M_PKTHDR))
1683 if (length == 0 || length < -2)
1685 if (length > MCLBYTES)
1687 if (length < 0 && divisor > MCLBYTES)
1690 length = 1 + (arc4random() % divisor);
1694 m_fixhdr(m0); /* Needed sanity check */
1696 m_first = m_getcl(how, MT_DATA, M_PKTHDR);
1697 if (m_first == NULL)
1700 if (m_dup_pkthdr(m_first, m0, how) == 0)
1705 while (progress < m0->m_pkthdr.len) {
1707 fraglen = 1 + (arc4random() % divisor);
1708 if (fraglen > m0->m_pkthdr.len - progress)
1709 fraglen = m0->m_pkthdr.len - progress;
1711 if (progress != 0) {
1712 struct mbuf *m_new = m_getcl(how, MT_DATA, 0);
1716 m_last->m_next = m_new;
1720 m_copydata(m0, progress, fraglen, mtod(m_last, caddr_t));
1721 progress += fraglen;
1722 m_last->m_len = fraglen;
1730 /* Return the original chain on failure */
1737 * Free pages from mbuf_ext_pgs, assuming they were allocated via
1738 * vm_page_alloc() and aren't associated with any object. Complement
1739 * to allocator from m_uiotombuf_nomap().
1742 mb_free_mext_pgs(struct mbuf *m)
1747 for (int i = 0; i < m->m_epg_npgs; i++) {
1748 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
1749 vm_page_unwire_noq(pg);
1754 static struct mbuf *
1755 m_uiotombuf_nomap(struct uio *uio, int how, int len, int maxseg, int flags)
1757 struct mbuf *m, *mb, *prev;
1758 vm_page_t pg_array[MBUF_PEXT_MAX_PGS];
1759 int error, length, i, needed;
1761 int pflags = malloc2vm_flags(how) | VM_ALLOC_NODUMP | VM_ALLOC_WIRED;
1763 MPASS((flags & M_PKTHDR) == 0);
1764 MPASS((how & M_ZERO) == 0);
1767 * len can be zero or an arbitrary large value bound by
1768 * the total data supplied by the uio.
1771 total = MIN(uio->uio_resid, len);
1773 total = uio->uio_resid;
1776 maxseg = MBUF_PEXT_MAX_PGS * PAGE_SIZE;
1779 * If total is zero, return an empty mbuf. This can occur
1780 * for TLS 1.0 connections which send empty fragments as
1781 * a countermeasure against the known-IV weakness in CBC
1784 if (__predict_false(total == 0)) {
1785 mb = mb_alloc_ext_pgs(how, mb_free_mext_pgs);
1788 mb->m_epg_flags = EPG_FLAG_ANON;
1793 * Allocate the pages
1797 mb = mb_alloc_ext_pgs(how, mb_free_mext_pgs);
1805 mb->m_epg_flags = EPG_FLAG_ANON;
1806 needed = length = MIN(maxseg, total);
1807 for (i = 0; needed > 0; i++, needed -= PAGE_SIZE) {
1809 pg_array[i] = vm_page_alloc_noobj(pflags);
1810 if (pg_array[i] == NULL) {
1811 if (how & M_NOWAIT) {
1818 mb->m_epg_pa[i] = VM_PAGE_TO_PHYS(pg_array[i]);
1821 mb->m_epg_last_len = length - PAGE_SIZE * (mb->m_epg_npgs - 1);
1822 MBUF_EXT_PGS_ASSERT_SANITY(mb);
1824 error = uiomove_fromphys(pg_array, 0, length, uio);
1828 mb->m_ext.ext_size += PAGE_SIZE * mb->m_epg_npgs;
1829 if (flags & M_PKTHDR)
1830 m->m_pkthdr.len += length;
1840 * Copy the contents of uio into a properly sized mbuf chain.
1843 m_uiotombuf(struct uio *uio, int how, int len, int align, int flags)
1845 struct mbuf *m, *mb;
1850 if (flags & M_EXTPG)
1851 return (m_uiotombuf_nomap(uio, how, len, align, flags));
1854 * len can be zero or an arbitrary large value bound by
1855 * the total data supplied by the uio.
1858 total = (uio->uio_resid < len) ? uio->uio_resid : len;
1860 total = uio->uio_resid;
1863 * The smallest unit returned by m_getm2() is a single mbuf
1864 * with pkthdr. We can't align past it.
1870 * Give us the full allocation or nothing.
1871 * If len is zero return the smallest empty mbuf.
1873 m = m_getm2(NULL, max(total + align, 1), how, MT_DATA, flags);
1878 /* Fill all mbufs with uio data and update header information. */
1879 for (mb = m; mb != NULL; mb = mb->m_next) {
1880 length = min(M_TRAILINGSPACE(mb), total - progress);
1882 error = uiomove(mtod(mb, void *), length, uio);
1890 if (flags & M_PKTHDR)
1891 m->m_pkthdr.len += length;
1893 KASSERT(progress == total, ("%s: progress != total", __func__));
1899 * Copy data to/from an unmapped mbuf into a uio limited by len if set.
1902 m_unmapped_uiomove(const struct mbuf *m, int m_off, struct uio *uio, int len)
1905 int error, i, off, pglen, pgoff, seglen, segoff;
1910 /* Skip over any data removed from the front. */
1911 off = mtod(m, vm_offset_t);
1914 if (m->m_epg_hdrlen != 0) {
1915 if (off >= m->m_epg_hdrlen) {
1916 off -= m->m_epg_hdrlen;
1918 seglen = m->m_epg_hdrlen - off;
1920 seglen = min(seglen, len);
1923 error = uiomove(__DECONST(void *,
1924 &m->m_epg_hdr[segoff]), seglen, uio);
1927 pgoff = m->m_epg_1st_off;
1928 for (i = 0; i < m->m_epg_npgs && error == 0 && len > 0; i++) {
1929 pglen = m_epg_pagelen(m, i, pgoff);
1935 seglen = pglen - off;
1936 segoff = pgoff + off;
1938 seglen = min(seglen, len);
1940 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
1941 error = uiomove_fromphys(&pg, segoff, seglen, uio);
1944 if (len != 0 && error == 0) {
1945 KASSERT((off + len) <= m->m_epg_trllen,
1946 ("off + len > trail (%d + %d > %d, m_off = %d)", off, len,
1947 m->m_epg_trllen, m_off));
1948 error = uiomove(__DECONST(void *, &m->m_epg_trail[off]),
1955 * Copy an mbuf chain into a uio limited by len if set.
1958 m_mbuftouio(struct uio *uio, const struct mbuf *m, int len)
1960 int error, length, total;
1964 total = min(uio->uio_resid, len);
1966 total = uio->uio_resid;
1968 /* Fill the uio with data from the mbufs. */
1969 for (; m != NULL; m = m->m_next) {
1970 length = min(m->m_len, total - progress);
1972 if ((m->m_flags & M_EXTPG) != 0)
1973 error = m_unmapped_uiomove(m, 0, uio, length);
1975 error = uiomove(mtod(m, void *), length, uio);
1986 * Create a writable copy of the mbuf chain. While doing this
1987 * we compact the chain with a goal of producing a chain with
1988 * at most two mbufs. The second mbuf in this chain is likely
1989 * to be a cluster. The primary purpose of this work is to create
1990 * a writable packet for encryption, compression, etc. The
1991 * secondary goal is to linearize the data so the data can be
1992 * passed to crypto hardware in the most efficient manner possible.
1995 m_unshare(struct mbuf *m0, int how)
1997 struct mbuf *m, *mprev;
1998 struct mbuf *n, *mfirst, *mlast;
2002 for (m = m0; m != NULL; m = mprev->m_next) {
2004 * Regular mbufs are ignored unless there's a cluster
2005 * in front of it that we can use to coalesce. We do
2006 * the latter mainly so later clusters can be coalesced
2007 * also w/o having to handle them specially (i.e. convert
2008 * mbuf+cluster -> cluster). This optimization is heavily
2009 * influenced by the assumption that we're running over
2010 * Ethernet where MCLBYTES is large enough that the max
2011 * packet size will permit lots of coalescing into a
2012 * single cluster. This in turn permits efficient
2013 * crypto operations, especially when using hardware.
2015 if ((m->m_flags & M_EXT) == 0) {
2016 if (mprev && (mprev->m_flags & M_EXT) &&
2017 m->m_len <= M_TRAILINGSPACE(mprev)) {
2018 /* XXX: this ignores mbuf types */
2019 memcpy(mtod(mprev, caddr_t) + mprev->m_len,
2020 mtod(m, caddr_t), m->m_len);
2021 mprev->m_len += m->m_len;
2022 mprev->m_next = m->m_next; /* unlink from chain */
2023 m_free(m); /* reclaim mbuf */
2030 * Writable mbufs are left alone (for now).
2032 if (M_WRITABLE(m)) {
2038 * Not writable, replace with a copy or coalesce with
2039 * the previous mbuf if possible (since we have to copy
2040 * it anyway, we try to reduce the number of mbufs and
2041 * clusters so that future work is easier).
2043 KASSERT(m->m_flags & M_EXT, ("m_flags 0x%x", m->m_flags));
2044 /* NB: we only coalesce into a cluster or larger */
2045 if (mprev != NULL && (mprev->m_flags & M_EXT) &&
2046 m->m_len <= M_TRAILINGSPACE(mprev)) {
2047 /* XXX: this ignores mbuf types */
2048 memcpy(mtod(mprev, caddr_t) + mprev->m_len,
2049 mtod(m, caddr_t), m->m_len);
2050 mprev->m_len += m->m_len;
2051 mprev->m_next = m->m_next; /* unlink from chain */
2052 m_free(m); /* reclaim mbuf */
2057 * Allocate new space to hold the copy and copy the data.
2058 * We deal with jumbo mbufs (i.e. m_len > MCLBYTES) by
2059 * splitting them into clusters. We could just malloc a
2060 * buffer and make it external but too many device drivers
2061 * don't know how to break up the non-contiguous memory when
2064 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS);
2069 if (m->m_flags & M_PKTHDR) {
2070 KASSERT(mprev == NULL, ("%s: m0 %p, m %p has M_PKTHDR",
2072 m_move_pkthdr(n, m);
2079 int cc = min(len, MCLBYTES);
2080 memcpy(mtod(n, caddr_t), mtod(m, caddr_t) + off, cc);
2086 newipsecstat.ips_clcopied++;
2094 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS);
2101 n->m_next = m->m_next;
2103 m0 = mfirst; /* new head of chain */
2105 mprev->m_next = mfirst; /* replace old mbuf */
2106 m_free(m); /* release old mbuf */
2112 #ifdef MBUF_PROFILING
2114 #define MP_BUCKETS 32 /* don't just change this as things may overflow.*/
2115 struct mbufprofile {
2116 uintmax_t wasted[MP_BUCKETS];
2117 uintmax_t used[MP_BUCKETS];
2118 uintmax_t segments[MP_BUCKETS];
2122 m_profile(struct mbuf *m)
2131 if (m->m_flags & M_EXT) {
2132 wasted += MHLEN - sizeof(m->m_ext) +
2133 m->m_ext.ext_size - m->m_len;
2135 if (m->m_flags & M_PKTHDR)
2136 wasted += MHLEN - m->m_len;
2138 wasted += MLEN - m->m_len;
2142 /* be paranoid.. it helps */
2143 if (segments > MP_BUCKETS - 1)
2144 segments = MP_BUCKETS - 1;
2147 if (wasted > 100000)
2149 /* store in the appropriate bucket */
2150 /* don't bother locking. if it's slightly off, so what? */
2151 mbprof.segments[segments]++;
2152 mbprof.used[fls(used)]++;
2153 mbprof.wasted[fls(wasted)]++;
2157 mbprof_handler(SYSCTL_HANDLER_ARGS)
2164 sbuf_new_for_sysctl(&sb, buf, sizeof(buf), req);
2166 p = &mbprof.wasted[0];
2169 "%ju %ju %ju %ju %ju %ju %ju %ju "
2170 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
2171 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
2172 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
2174 p = &mbprof.wasted[16];
2176 "%ju %ju %ju %ju %ju %ju %ju %ju "
2177 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
2178 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
2179 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
2181 p = &mbprof.used[0];
2184 "%ju %ju %ju %ju %ju %ju %ju %ju "
2185 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
2186 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
2187 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
2189 p = &mbprof.used[16];
2191 "%ju %ju %ju %ju %ju %ju %ju %ju "
2192 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
2193 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
2194 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
2196 p = &mbprof.segments[0];
2199 "%ju %ju %ju %ju %ju %ju %ju %ju "
2200 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
2201 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
2202 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
2204 p = &mbprof.segments[16];
2206 "%ju %ju %ju %ju %ju %ju %ju %ju "
2207 "%ju %ju %ju %ju %ju %ju %ju %jju",
2208 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
2209 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
2212 error = sbuf_finish(&sb);
2218 mbprof_clr_handler(SYSCTL_HANDLER_ARGS)
2223 error = sysctl_handle_int(oidp, &clear, 0, req);
2224 if (error || !req->newptr)
2228 bzero(&mbprof, sizeof(mbprof));
2234 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofile,
2235 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
2236 mbprof_handler, "A",
2237 "mbuf profiling statistics");
2239 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofileclr,
2240 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
2241 mbprof_clr_handler, "I",
2242 "clear mbuf profiling statistics");