FreeBSD/Linux Kernel Cross Reference
sys/kern/uipc_mbuf.c
1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1982, 1986, 1988, 1991, 1993
5 * The Regents of the University of California. All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
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.
18 *
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
29 * SUCH DAMAGE.
30 *
31 * @(#)uipc_mbuf.c 8.2 (Berkeley) 1/4/94
32 */
33
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36
37 #include "opt_param.h"
38 #include "opt_mbuf_stress_test.h"
39 #include "opt_mbuf_profiling.h"
40
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/limits.h>
45 #include <sys/lock.h>
46 #include <sys/malloc.h>
47 #include <sys/mbuf.h>
48 #include <sys/sysctl.h>
49 #include <sys/domain.h>
50 #include <sys/protosw.h>
51 #include <sys/uio.h>
52 #include <sys/sdt.h>
53
54 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__init,
55 "struct mbuf *", "mbufinfo_t *",
56 "uint32_t", "uint32_t",
57 "uint16_t", "uint16_t",
58 "uint32_t", "uint32_t",
59 "uint32_t", "uint32_t");
60
61 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__gethdr,
62 "uint32_t", "uint32_t",
63 "uint16_t", "uint16_t",
64 "struct mbuf *", "mbufinfo_t *");
65
66 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__get,
67 "uint32_t", "uint32_t",
68 "uint16_t", "uint16_t",
69 "struct mbuf *", "mbufinfo_t *");
70
71 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__getcl,
72 "uint32_t", "uint32_t",
73 "uint16_t", "uint16_t",
74 "uint32_t", "uint32_t",
75 "struct mbuf *", "mbufinfo_t *");
76
77 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__getjcl,
78 "uint32_t", "uint32_t",
79 "uint16_t", "uint16_t",
80 "uint32_t", "uint32_t",
81 "uint32_t", "uint32_t",
82 "struct mbuf *", "mbufinfo_t *");
83
84 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__clget,
85 "struct mbuf *", "mbufinfo_t *",
86 "uint32_t", "uint32_t",
87 "uint32_t", "uint32_t");
88
89 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__cljget,
90 "struct mbuf *", "mbufinfo_t *",
91 "uint32_t", "uint32_t",
92 "uint32_t", "uint32_t",
93 "void*", "void*");
94
95 SDT_PROBE_DEFINE(sdt, , , m__cljset);
96
97 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__free,
98 "struct mbuf *", "mbufinfo_t *");
99
100 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__freem,
101 "struct mbuf *", "mbufinfo_t *");
102
103 #include <security/mac/mac_framework.h>
104
105 int max_linkhdr;
106 int max_protohdr;
107 int max_hdr;
108 int max_datalen;
109 #ifdef MBUF_STRESS_TEST
110 int m_defragpackets;
111 int m_defragbytes;
112 int m_defraguseless;
113 int m_defragfailure;
114 int m_defragrandomfailures;
115 #endif
116
117 /*
118 * sysctl(8) exported objects
119 */
120 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RD,
121 &max_linkhdr, 0, "Size of largest link layer header");
122 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RD,
123 &max_protohdr, 0, "Size of largest protocol layer header");
124 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RD,
125 &max_hdr, 0, "Size of largest link plus protocol header");
126 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RD,
127 &max_datalen, 0, "Minimum space left in mbuf after max_hdr");
128 #ifdef MBUF_STRESS_TEST
129 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD,
130 &m_defragpackets, 0, "");
131 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD,
132 &m_defragbytes, 0, "");
133 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD,
134 &m_defraguseless, 0, "");
135 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD,
136 &m_defragfailure, 0, "");
137 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW,
138 &m_defragrandomfailures, 0, "");
139 #endif
140
141 /*
142 * Ensure the correct size of various mbuf parameters. It could be off due
143 * to compiler-induced padding and alignment artifacts.
144 */
145 CTASSERT(MSIZE - offsetof(struct mbuf, m_dat) == MLEN);
146 CTASSERT(MSIZE - offsetof(struct mbuf, m_pktdat) == MHLEN);
147
148 /*
149 * mbuf data storage should be 64-bit aligned regardless of architectural
150 * pointer size; check this is the case with and without a packet header.
151 */
152 CTASSERT(offsetof(struct mbuf, m_dat) % 8 == 0);
153 CTASSERT(offsetof(struct mbuf, m_pktdat) % 8 == 0);
154
155 /*
156 * While the specific values here don't matter too much (i.e., +/- a few
157 * words), we do want to ensure that changes to these values are carefully
158 * reasoned about and properly documented. This is especially the case as
159 * network-protocol and device-driver modules encode these layouts, and must
160 * be recompiled if the structures change. Check these values at compile time
161 * against the ones documented in comments in mbuf.h.
162 *
163 * NB: Possibly they should be documented there via #define's and not just
164 * comments.
165 */
166 #if defined(__LP64__)
167 CTASSERT(offsetof(struct mbuf, m_dat) == 32);
168 CTASSERT(sizeof(struct pkthdr) == 56);
169 CTASSERT(sizeof(struct m_ext) == 48);
170 #else
171 CTASSERT(offsetof(struct mbuf, m_dat) == 24);
172 CTASSERT(sizeof(struct pkthdr) == 48);
173 CTASSERT(sizeof(struct m_ext) == 28);
174 #endif
175
176 /*
177 * Assert that the queue(3) macros produce code of the same size as an old
178 * plain pointer does.
179 */
180 #ifdef INVARIANTS
181 static struct mbuf __used m_assertbuf;
182 CTASSERT(sizeof(m_assertbuf.m_slist) == sizeof(m_assertbuf.m_next));
183 CTASSERT(sizeof(m_assertbuf.m_stailq) == sizeof(m_assertbuf.m_next));
184 CTASSERT(sizeof(m_assertbuf.m_slistpkt) == sizeof(m_assertbuf.m_nextpkt));
185 CTASSERT(sizeof(m_assertbuf.m_stailqpkt) == sizeof(m_assertbuf.m_nextpkt));
186 #endif
187
188 /*
189 * Attach the cluster from *m to *n, set up m_ext in *n
190 * and bump the refcount of the cluster.
191 */
192 void
193 mb_dupcl(struct mbuf *n, struct mbuf *m)
194 {
195 volatile u_int *refcnt;
196
197 KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
198 KASSERT(!(n->m_flags & M_EXT), ("%s: M_EXT set on %p", __func__, n));
199
200 /*
201 * Cache access optimization. For most kinds of external
202 * storage we don't need full copy of m_ext, since the
203 * holder of the 'ext_count' is responsible to carry the
204 * free routine and its arguments. Exclusion is EXT_EXTREF,
205 * where 'ext_cnt' doesn't point into mbuf at all.
206 */
207 if (m->m_ext.ext_type == EXT_EXTREF)
208 bcopy(&m->m_ext, &n->m_ext, sizeof(struct m_ext));
209 else
210 bcopy(&m->m_ext, &n->m_ext, m_ext_copylen);
211 n->m_flags |= M_EXT;
212 n->m_flags |= m->m_flags & M_RDONLY;
213
214 /* See if this is the mbuf that holds the embedded refcount. */
215 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
216 refcnt = n->m_ext.ext_cnt = &m->m_ext.ext_count;
217 n->m_ext.ext_flags &= ~EXT_FLAG_EMBREF;
218 } else {
219 KASSERT(m->m_ext.ext_cnt != NULL,
220 ("%s: no refcounting pointer on %p", __func__, m));
221 refcnt = m->m_ext.ext_cnt;
222 }
223
224 if (*refcnt == 1)
225 *refcnt += 1;
226 else
227 atomic_add_int(refcnt, 1);
228 }
229
230 void
231 m_demote_pkthdr(struct mbuf *m)
232 {
233
234 M_ASSERTPKTHDR(m);
235
236 m_tag_delete_chain(m, NULL);
237 m->m_flags &= ~M_PKTHDR;
238 bzero(&m->m_pkthdr, sizeof(struct pkthdr));
239 }
240
241 /*
242 * Clean up mbuf (chain) from any tags and packet headers.
243 * If "all" is set then the first mbuf in the chain will be
244 * cleaned too.
245 */
246 void
247 m_demote(struct mbuf *m0, int all, int flags)
248 {
249 struct mbuf *m;
250
251 for (m = all ? m0 : m0->m_next; m != NULL; m = m->m_next) {
252 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt in m %p, m0 %p",
253 __func__, m, m0));
254 if (m->m_flags & M_PKTHDR)
255 m_demote_pkthdr(m);
256 m->m_flags = m->m_flags & (M_EXT | M_RDONLY | M_NOFREE | flags);
257 }
258 }
259
260 /*
261 * Sanity checks on mbuf (chain) for use in KASSERT() and general
262 * debugging.
263 * Returns 0 or panics when bad and 1 on all tests passed.
264 * Sanitize, 0 to run M_SANITY_ACTION, 1 to garble things so they
265 * blow up later.
266 */
267 int
268 m_sanity(struct mbuf *m0, int sanitize)
269 {
270 struct mbuf *m;
271 caddr_t a, b;
272 int pktlen = 0;
273
274 #ifdef INVARIANTS
275 #define M_SANITY_ACTION(s) panic("mbuf %p: " s, m)
276 #else
277 #define M_SANITY_ACTION(s) printf("mbuf %p: " s, m)
278 #endif
279
280 for (m = m0; m != NULL; m = m->m_next) {
281 /*
282 * Basic pointer checks. If any of these fails then some
283 * unrelated kernel memory before or after us is trashed.
284 * No way to recover from that.
285 */
286 a = M_START(m);
287 b = a + M_SIZE(m);
288 if ((caddr_t)m->m_data < a)
289 M_SANITY_ACTION("m_data outside mbuf data range left");
290 if ((caddr_t)m->m_data > b)
291 M_SANITY_ACTION("m_data outside mbuf data range right");
292 if ((caddr_t)m->m_data + m->m_len > b)
293 M_SANITY_ACTION("m_data + m_len exeeds mbuf space");
294
295 /* m->m_nextpkt may only be set on first mbuf in chain. */
296 if (m != m0 && m->m_nextpkt != NULL) {
297 if (sanitize) {
298 m_freem(m->m_nextpkt);
299 m->m_nextpkt = (struct mbuf *)0xDEADC0DE;
300 } else
301 M_SANITY_ACTION("m->m_nextpkt on in-chain mbuf");
302 }
303
304 /* packet length (not mbuf length!) calculation */
305 if (m0->m_flags & M_PKTHDR)
306 pktlen += m->m_len;
307
308 /* m_tags may only be attached to first mbuf in chain. */
309 if (m != m0 && m->m_flags & M_PKTHDR &&
310 !SLIST_EMPTY(&m->m_pkthdr.tags)) {
311 if (sanitize) {
312 m_tag_delete_chain(m, NULL);
313 /* put in 0xDEADC0DE perhaps? */
314 } else
315 M_SANITY_ACTION("m_tags on in-chain mbuf");
316 }
317
318 /* M_PKTHDR may only be set on first mbuf in chain */
319 if (m != m0 && m->m_flags & M_PKTHDR) {
320 if (sanitize) {
321 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
322 m->m_flags &= ~M_PKTHDR;
323 /* put in 0xDEADCODE and leave hdr flag in */
324 } else
325 M_SANITY_ACTION("M_PKTHDR on in-chain mbuf");
326 }
327 }
328 m = m0;
329 if (pktlen && pktlen != m->m_pkthdr.len) {
330 if (sanitize)
331 m->m_pkthdr.len = 0;
332 else
333 M_SANITY_ACTION("m_pkthdr.len != mbuf chain length");
334 }
335 return 1;
336
337 #undef M_SANITY_ACTION
338 }
339
340 /*
341 * Non-inlined part of m_init().
342 */
343 int
344 m_pkthdr_init(struct mbuf *m, int how)
345 {
346 #ifdef MAC
347 int error;
348 #endif
349 m->m_data = m->m_pktdat;
350 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
351 #ifdef MAC
352 /* If the label init fails, fail the alloc */
353 error = mac_mbuf_init(m, how);
354 if (error)
355 return (error);
356 #endif
357
358 return (0);
359 }
360
361 /*
362 * "Move" mbuf pkthdr from "from" to "to".
363 * "from" must have M_PKTHDR set, and "to" must be empty.
364 */
365 void
366 m_move_pkthdr(struct mbuf *to, struct mbuf *from)
367 {
368
369 #if 0
370 /* see below for why these are not enabled */
371 M_ASSERTPKTHDR(to);
372 /* Note: with MAC, this may not be a good assertion. */
373 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags),
374 ("m_move_pkthdr: to has tags"));
375 #endif
376 #ifdef MAC
377 /*
378 * XXXMAC: It could be this should also occur for non-MAC?
379 */
380 if (to->m_flags & M_PKTHDR)
381 m_tag_delete_chain(to, NULL);
382 #endif
383 to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT);
384 if ((to->m_flags & M_EXT) == 0)
385 to->m_data = to->m_pktdat;
386 to->m_pkthdr = from->m_pkthdr; /* especially tags */
387 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */
388 from->m_flags &= ~M_PKTHDR;
389 }
390
391 /*
392 * Duplicate "from"'s mbuf pkthdr in "to".
393 * "from" must have M_PKTHDR set, and "to" must be empty.
394 * In particular, this does a deep copy of the packet tags.
395 */
396 int
397 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how)
398 {
399
400 #if 0
401 /*
402 * The mbuf allocator only initializes the pkthdr
403 * when the mbuf is allocated with m_gethdr(). Many users
404 * (e.g. m_copy*, m_prepend) use m_get() and then
405 * smash the pkthdr as needed causing these
406 * assertions to trip. For now just disable them.
407 */
408 M_ASSERTPKTHDR(to);
409 /* Note: with MAC, this may not be a good assertion. */
410 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags), ("m_dup_pkthdr: to has tags"));
411 #endif
412 MBUF_CHECKSLEEP(how);
413 #ifdef MAC
414 if (to->m_flags & M_PKTHDR)
415 m_tag_delete_chain(to, NULL);
416 #endif
417 to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT);
418 if ((to->m_flags & M_EXT) == 0)
419 to->m_data = to->m_pktdat;
420 to->m_pkthdr = from->m_pkthdr;
421 SLIST_INIT(&to->m_pkthdr.tags);
422 return (m_tag_copy_chain(to, from, how));
423 }
424
425 /*
426 * Lesser-used path for M_PREPEND:
427 * allocate new mbuf to prepend to chain,
428 * copy junk along.
429 */
430 struct mbuf *
431 m_prepend(struct mbuf *m, int len, int how)
432 {
433 struct mbuf *mn;
434
435 if (m->m_flags & M_PKTHDR)
436 mn = m_gethdr(how, m->m_type);
437 else
438 mn = m_get(how, m->m_type);
439 if (mn == NULL) {
440 m_freem(m);
441 return (NULL);
442 }
443 if (m->m_flags & M_PKTHDR)
444 m_move_pkthdr(mn, m);
445 mn->m_next = m;
446 m = mn;
447 if (len < M_SIZE(m))
448 M_ALIGN(m, len);
449 m->m_len = len;
450 return (m);
451 }
452
453 /*
454 * Make a copy of an mbuf chain starting "off0" bytes from the beginning,
455 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf.
456 * The wait parameter is a choice of M_WAITOK/M_NOWAIT from caller.
457 * Note that the copy is read-only, because clusters are not copied,
458 * only their reference counts are incremented.
459 */
460 struct mbuf *
461 m_copym(struct mbuf *m, int off0, int len, int wait)
462 {
463 struct mbuf *n, **np;
464 int off = off0;
465 struct mbuf *top;
466 int copyhdr = 0;
467
468 KASSERT(off >= 0, ("m_copym, negative off %d", off));
469 KASSERT(len >= 0, ("m_copym, negative len %d", len));
470 MBUF_CHECKSLEEP(wait);
471 if (off == 0 && m->m_flags & M_PKTHDR)
472 copyhdr = 1;
473 while (off > 0) {
474 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain"));
475 if (off < m->m_len)
476 break;
477 off -= m->m_len;
478 m = m->m_next;
479 }
480 np = ⊤
481 top = NULL;
482 while (len > 0) {
483 if (m == NULL) {
484 KASSERT(len == M_COPYALL,
485 ("m_copym, length > size of mbuf chain"));
486 break;
487 }
488 if (copyhdr)
489 n = m_gethdr(wait, m->m_type);
490 else
491 n = m_get(wait, m->m_type);
492 *np = n;
493 if (n == NULL)
494 goto nospace;
495 if (copyhdr) {
496 if (!m_dup_pkthdr(n, m, wait))
497 goto nospace;
498 if (len == M_COPYALL)
499 n->m_pkthdr.len -= off0;
500 else
501 n->m_pkthdr.len = len;
502 copyhdr = 0;
503 }
504 n->m_len = min(len, m->m_len - off);
505 if (m->m_flags & M_EXT) {
506 n->m_data = m->m_data + off;
507 mb_dupcl(n, m);
508 } else
509 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
510 (u_int)n->m_len);
511 if (len != M_COPYALL)
512 len -= n->m_len;
513 off = 0;
514 m = m->m_next;
515 np = &n->m_next;
516 }
517
518 return (top);
519 nospace:
520 m_freem(top);
521 return (NULL);
522 }
523
524 /*
525 * Copy an entire packet, including header (which must be present).
526 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'.
527 * Note that the copy is read-only, because clusters are not copied,
528 * only their reference counts are incremented.
529 * Preserve alignment of the first mbuf so if the creator has left
530 * some room at the beginning (e.g. for inserting protocol headers)
531 * the copies still have the room available.
532 */
533 struct mbuf *
534 m_copypacket(struct mbuf *m, int how)
535 {
536 struct mbuf *top, *n, *o;
537
538 MBUF_CHECKSLEEP(how);
539 n = m_get(how, m->m_type);
540 top = n;
541 if (n == NULL)
542 goto nospace;
543
544 if (!m_dup_pkthdr(n, m, how))
545 goto nospace;
546 n->m_len = m->m_len;
547 if (m->m_flags & M_EXT) {
548 n->m_data = m->m_data;
549 mb_dupcl(n, m);
550 } else {
551 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat );
552 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
553 }
554
555 m = m->m_next;
556 while (m) {
557 o = m_get(how, m->m_type);
558 if (o == NULL)
559 goto nospace;
560
561 n->m_next = o;
562 n = n->m_next;
563
564 n->m_len = m->m_len;
565 if (m->m_flags & M_EXT) {
566 n->m_data = m->m_data;
567 mb_dupcl(n, m);
568 } else {
569 bcopy(mtod(m, char *), mtod(n, char *), n->m_len);
570 }
571
572 m = m->m_next;
573 }
574 return top;
575 nospace:
576 m_freem(top);
577 return (NULL);
578 }
579
580 /*
581 * Copy data from an mbuf chain starting "off" bytes from the beginning,
582 * continuing for "len" bytes, into the indicated buffer.
583 */
584 void
585 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp)
586 {
587 u_int count;
588
589 KASSERT(off >= 0, ("m_copydata, negative off %d", off));
590 KASSERT(len >= 0, ("m_copydata, negative len %d", len));
591 while (off > 0) {
592 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain"));
593 if (off < m->m_len)
594 break;
595 off -= m->m_len;
596 m = m->m_next;
597 }
598 while (len > 0) {
599 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain"));
600 count = min(m->m_len - off, len);
601 bcopy(mtod(m, caddr_t) + off, cp, count);
602 len -= count;
603 cp += count;
604 off = 0;
605 m = m->m_next;
606 }
607 }
608
609 /*
610 * Copy a packet header mbuf chain into a completely new chain, including
611 * copying any mbuf clusters. Use this instead of m_copypacket() when
612 * you need a writable copy of an mbuf chain.
613 */
614 struct mbuf *
615 m_dup(const struct mbuf *m, int how)
616 {
617 struct mbuf **p, *top = NULL;
618 int remain, moff, nsize;
619
620 MBUF_CHECKSLEEP(how);
621 /* Sanity check */
622 if (m == NULL)
623 return (NULL);
624 M_ASSERTPKTHDR(m);
625
626 /* While there's more data, get a new mbuf, tack it on, and fill it */
627 remain = m->m_pkthdr.len;
628 moff = 0;
629 p = ⊤
630 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */
631 struct mbuf *n;
632
633 /* Get the next new mbuf */
634 if (remain >= MINCLSIZE) {
635 n = m_getcl(how, m->m_type, 0);
636 nsize = MCLBYTES;
637 } else {
638 n = m_get(how, m->m_type);
639 nsize = MLEN;
640 }
641 if (n == NULL)
642 goto nospace;
643
644 if (top == NULL) { /* First one, must be PKTHDR */
645 if (!m_dup_pkthdr(n, m, how)) {
646 m_free(n);
647 goto nospace;
648 }
649 if ((n->m_flags & M_EXT) == 0)
650 nsize = MHLEN;
651 n->m_flags &= ~M_RDONLY;
652 }
653 n->m_len = 0;
654
655 /* Link it into the new chain */
656 *p = n;
657 p = &n->m_next;
658
659 /* Copy data from original mbuf(s) into new mbuf */
660 while (n->m_len < nsize && m != NULL) {
661 int chunk = min(nsize - n->m_len, m->m_len - moff);
662
663 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk);
664 moff += chunk;
665 n->m_len += chunk;
666 remain -= chunk;
667 if (moff == m->m_len) {
668 m = m->m_next;
669 moff = 0;
670 }
671 }
672
673 /* Check correct total mbuf length */
674 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL),
675 ("%s: bogus m_pkthdr.len", __func__));
676 }
677 return (top);
678
679 nospace:
680 m_freem(top);
681 return (NULL);
682 }
683
684 /*
685 * Concatenate mbuf chain n to m.
686 * Both chains must be of the same type (e.g. MT_DATA).
687 * Any m_pkthdr is not updated.
688 */
689 void
690 m_cat(struct mbuf *m, struct mbuf *n)
691 {
692 while (m->m_next)
693 m = m->m_next;
694 while (n) {
695 if (!M_WRITABLE(m) ||
696 M_TRAILINGSPACE(m) < n->m_len) {
697 /* just join the two chains */
698 m->m_next = n;
699 return;
700 }
701 /* splat the data from one into the other */
702 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
703 (u_int)n->m_len);
704 m->m_len += n->m_len;
705 n = m_free(n);
706 }
707 }
708
709 /*
710 * Concatenate two pkthdr mbuf chains.
711 */
712 void
713 m_catpkt(struct mbuf *m, struct mbuf *n)
714 {
715
716 M_ASSERTPKTHDR(m);
717 M_ASSERTPKTHDR(n);
718
719 m->m_pkthdr.len += n->m_pkthdr.len;
720 m_demote(n, 1, 0);
721
722 m_cat(m, n);
723 }
724
725 void
726 m_adj(struct mbuf *mp, int req_len)
727 {
728 int len = req_len;
729 struct mbuf *m;
730 int count;
731
732 if ((m = mp) == NULL)
733 return;
734 if (len >= 0) {
735 /*
736 * Trim from head.
737 */
738 while (m != NULL && len > 0) {
739 if (m->m_len <= len) {
740 len -= m->m_len;
741 m->m_len = 0;
742 m = m->m_next;
743 } else {
744 m->m_len -= len;
745 m->m_data += len;
746 len = 0;
747 }
748 }
749 if (mp->m_flags & M_PKTHDR)
750 mp->m_pkthdr.len -= (req_len - len);
751 } else {
752 /*
753 * Trim from tail. Scan the mbuf chain,
754 * calculating its length and finding the last mbuf.
755 * If the adjustment only affects this mbuf, then just
756 * adjust and return. Otherwise, rescan and truncate
757 * after the remaining size.
758 */
759 len = -len;
760 count = 0;
761 for (;;) {
762 count += m->m_len;
763 if (m->m_next == (struct mbuf *)0)
764 break;
765 m = m->m_next;
766 }
767 if (m->m_len >= len) {
768 m->m_len -= len;
769 if (mp->m_flags & M_PKTHDR)
770 mp->m_pkthdr.len -= len;
771 return;
772 }
773 count -= len;
774 if (count < 0)
775 count = 0;
776 /*
777 * Correct length for chain is "count".
778 * Find the mbuf with last data, adjust its length,
779 * and toss data from remaining mbufs on chain.
780 */
781 m = mp;
782 if (m->m_flags & M_PKTHDR)
783 m->m_pkthdr.len = count;
784 for (; m; m = m->m_next) {
785 if (m->m_len >= count) {
786 m->m_len = count;
787 if (m->m_next != NULL) {
788 m_freem(m->m_next);
789 m->m_next = NULL;
790 }
791 break;
792 }
793 count -= m->m_len;
794 }
795 }
796 }
797
798 void
799 m_adj_decap(struct mbuf *mp, int len)
800 {
801 uint8_t rsstype;
802
803 m_adj(mp, len);
804 if ((mp->m_flags & M_PKTHDR) != 0) {
805 /*
806 * If flowid was calculated by card from the inner
807 * headers, move flowid to the decapsulated mbuf
808 * chain, otherwise clear. This depends on the
809 * internals of m_adj, which keeps pkthdr as is, in
810 * particular not changing rsstype and flowid.
811 */
812 rsstype = mp->m_pkthdr.rsstype;
813 if ((rsstype & M_HASHTYPE_INNER) != 0) {
814 M_HASHTYPE_SET(mp, rsstype & ~M_HASHTYPE_INNER);
815 } else {
816 M_HASHTYPE_CLEAR(mp);
817 }
818 }
819 }
820
821 /*
822 * Rearange an mbuf chain so that len bytes are contiguous
823 * and in the data area of an mbuf (so that mtod will work
824 * for a structure of size len). Returns the resulting
825 * mbuf chain on success, frees it and returns null on failure.
826 * If there is room, it will add up to max_protohdr-len extra bytes to the
827 * contiguous region in an attempt to avoid being called next time.
828 */
829 struct mbuf *
830 m_pullup(struct mbuf *n, int len)
831 {
832 struct mbuf *m;
833 int count;
834 int space;
835
836 /*
837 * If first mbuf has no cluster, and has room for len bytes
838 * without shifting current data, pullup into it,
839 * otherwise allocate a new mbuf to prepend to the chain.
840 */
841 if ((n->m_flags & M_EXT) == 0 &&
842 n->m_data + len < &n->m_dat[MLEN] && n->m_next) {
843 if (n->m_len >= len)
844 return (n);
845 m = n;
846 n = n->m_next;
847 len -= m->m_len;
848 } else {
849 if (len > MHLEN)
850 goto bad;
851 m = m_get(M_NOWAIT, n->m_type);
852 if (m == NULL)
853 goto bad;
854 if (n->m_flags & M_PKTHDR)
855 m_move_pkthdr(m, n);
856 }
857 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
858 do {
859 count = min(min(max(len, max_protohdr), space), n->m_len);
860 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len,
861 (u_int)count);
862 len -= count;
863 m->m_len += count;
864 n->m_len -= count;
865 space -= count;
866 if (n->m_len)
867 n->m_data += count;
868 else
869 n = m_free(n);
870 } while (len > 0 && n);
871 if (len > 0) {
872 (void) m_free(m);
873 goto bad;
874 }
875 m->m_next = n;
876 return (m);
877 bad:
878 m_freem(n);
879 return (NULL);
880 }
881
882 /*
883 * Like m_pullup(), except a new mbuf is always allocated, and we allow
884 * the amount of empty space before the data in the new mbuf to be specified
885 * (in the event that the caller expects to prepend later).
886 */
887 struct mbuf *
888 m_copyup(struct mbuf *n, int len, int dstoff)
889 {
890 struct mbuf *m;
891 int count, space;
892
893 if (len > (MHLEN - dstoff))
894 goto bad;
895 m = m_get(M_NOWAIT, n->m_type);
896 if (m == NULL)
897 goto bad;
898 if (n->m_flags & M_PKTHDR)
899 m_move_pkthdr(m, n);
900 m->m_data += dstoff;
901 space = &m->m_dat[MLEN] - (m->m_data + m->m_len);
902 do {
903 count = min(min(max(len, max_protohdr), space), n->m_len);
904 memcpy(mtod(m, caddr_t) + m->m_len, mtod(n, caddr_t),
905 (unsigned)count);
906 len -= count;
907 m->m_len += count;
908 n->m_len -= count;
909 space -= count;
910 if (n->m_len)
911 n->m_data += count;
912 else
913 n = m_free(n);
914 } while (len > 0 && n);
915 if (len > 0) {
916 (void) m_free(m);
917 goto bad;
918 }
919 m->m_next = n;
920 return (m);
921 bad:
922 m_freem(n);
923 return (NULL);
924 }
925
926 /*
927 * Partition an mbuf chain in two pieces, returning the tail --
928 * all but the first len0 bytes. In case of failure, it returns NULL and
929 * attempts to restore the chain to its original state.
930 *
931 * Note that the resulting mbufs might be read-only, because the new
932 * mbuf can end up sharing an mbuf cluster with the original mbuf if
933 * the "breaking point" happens to lie within a cluster mbuf. Use the
934 * M_WRITABLE() macro to check for this case.
935 */
936 struct mbuf *
937 m_split(struct mbuf *m0, int len0, int wait)
938 {
939 struct mbuf *m, *n;
940 u_int len = len0, remain;
941
942 MBUF_CHECKSLEEP(wait);
943 for (m = m0; m && len > m->m_len; m = m->m_next)
944 len -= m->m_len;
945 if (m == NULL)
946 return (NULL);
947 remain = m->m_len - len;
948 if (m0->m_flags & M_PKTHDR && remain == 0) {
949 n = m_gethdr(wait, m0->m_type);
950 if (n == NULL)
951 return (NULL);
952 n->m_next = m->m_next;
953 m->m_next = NULL;
954 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
955 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
956 m0->m_pkthdr.len = len0;
957 return (n);
958 } else if (m0->m_flags & M_PKTHDR) {
959 n = m_gethdr(wait, m0->m_type);
960 if (n == NULL)
961 return (NULL);
962 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif;
963 n->m_pkthdr.len = m0->m_pkthdr.len - len0;
964 m0->m_pkthdr.len = len0;
965 if (m->m_flags & M_EXT)
966 goto extpacket;
967 if (remain > MHLEN) {
968 /* m can't be the lead packet */
969 M_ALIGN(n, 0);
970 n->m_next = m_split(m, len, wait);
971 if (n->m_next == NULL) {
972 (void) m_free(n);
973 return (NULL);
974 } else {
975 n->m_len = 0;
976 return (n);
977 }
978 } else
979 M_ALIGN(n, remain);
980 } else if (remain == 0) {
981 n = m->m_next;
982 m->m_next = NULL;
983 return (n);
984 } else {
985 n = m_get(wait, m->m_type);
986 if (n == NULL)
987 return (NULL);
988 M_ALIGN(n, remain);
989 }
990 extpacket:
991 if (m->m_flags & M_EXT) {
992 n->m_data = m->m_data + len;
993 mb_dupcl(n, m);
994 } else {
995 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain);
996 }
997 n->m_len = remain;
998 m->m_len = len;
999 n->m_next = m->m_next;
1000 m->m_next = NULL;
1001 return (n);
1002 }
1003 /*
1004 * Routine to copy from device local memory into mbufs.
1005 * Note that `off' argument is offset into first mbuf of target chain from
1006 * which to begin copying the data to.
1007 */
1008 struct mbuf *
1009 m_devget(char *buf, int totlen, int off, struct ifnet *ifp,
1010 void (*copy)(char *from, caddr_t to, u_int len))
1011 {
1012 struct mbuf *m;
1013 struct mbuf *top = NULL, **mp = ⊤
1014 int len;
1015
1016 if (off < 0 || off > MHLEN)
1017 return (NULL);
1018
1019 while (totlen > 0) {
1020 if (top == NULL) { /* First one, must be PKTHDR */
1021 if (totlen + off >= MINCLSIZE) {
1022 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1023 len = MCLBYTES;
1024 } else {
1025 m = m_gethdr(M_NOWAIT, MT_DATA);
1026 len = MHLEN;
1027
1028 /* Place initial small packet/header at end of mbuf */
1029 if (m && totlen + off + max_linkhdr <= MHLEN) {
1030 m->m_data += max_linkhdr;
1031 len -= max_linkhdr;
1032 }
1033 }
1034 if (m == NULL)
1035 return NULL;
1036 m->m_pkthdr.rcvif = ifp;
1037 m->m_pkthdr.len = totlen;
1038 } else {
1039 if (totlen + off >= MINCLSIZE) {
1040 m = m_getcl(M_NOWAIT, MT_DATA, 0);
1041 len = MCLBYTES;
1042 } else {
1043 m = m_get(M_NOWAIT, MT_DATA);
1044 len = MLEN;
1045 }
1046 if (m == NULL) {
1047 m_freem(top);
1048 return NULL;
1049 }
1050 }
1051 if (off) {
1052 m->m_data += off;
1053 len -= off;
1054 off = 0;
1055 }
1056 m->m_len = len = min(totlen, len);
1057 if (copy)
1058 copy(buf, mtod(m, caddr_t), (u_int)len);
1059 else
1060 bcopy(buf, mtod(m, caddr_t), (u_int)len);
1061 buf += len;
1062 *mp = m;
1063 mp = &m->m_next;
1064 totlen -= len;
1065 }
1066 return (top);
1067 }
1068
1069 /*
1070 * Copy data from a buffer back into the indicated mbuf chain,
1071 * starting "off" bytes from the beginning, extending the mbuf
1072 * chain if necessary.
1073 */
1074 void
1075 m_copyback(struct mbuf *m0, int off, int len, c_caddr_t cp)
1076 {
1077 int mlen;
1078 struct mbuf *m = m0, *n;
1079 int totlen = 0;
1080
1081 if (m0 == NULL)
1082 return;
1083 while (off > (mlen = m->m_len)) {
1084 off -= mlen;
1085 totlen += mlen;
1086 if (m->m_next == NULL) {
1087 n = m_get(M_NOWAIT, m->m_type);
1088 if (n == NULL)
1089 goto out;
1090 bzero(mtod(n, caddr_t), MLEN);
1091 n->m_len = min(MLEN, len + off);
1092 m->m_next = n;
1093 }
1094 m = m->m_next;
1095 }
1096 while (len > 0) {
1097 if (m->m_next == NULL && (len > m->m_len - off)) {
1098 m->m_len += min(len - (m->m_len - off),
1099 M_TRAILINGSPACE(m));
1100 }
1101 mlen = min (m->m_len - off, len);
1102 bcopy(cp, off + mtod(m, caddr_t), (u_int)mlen);
1103 cp += mlen;
1104 len -= mlen;
1105 mlen += off;
1106 off = 0;
1107 totlen += mlen;
1108 if (len == 0)
1109 break;
1110 if (m->m_next == NULL) {
1111 n = m_get(M_NOWAIT, m->m_type);
1112 if (n == NULL)
1113 break;
1114 n->m_len = min(MLEN, len);
1115 m->m_next = n;
1116 }
1117 m = m->m_next;
1118 }
1119 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen))
1120 m->m_pkthdr.len = totlen;
1121 }
1122
1123 /*
1124 * Append the specified data to the indicated mbuf chain,
1125 * Extend the mbuf chain if the new data does not fit in
1126 * existing space.
1127 *
1128 * Return 1 if able to complete the job; otherwise 0.
1129 */
1130 int
1131 m_append(struct mbuf *m0, int len, c_caddr_t cp)
1132 {
1133 struct mbuf *m, *n;
1134 int remainder, space;
1135
1136 for (m = m0; m->m_next != NULL; m = m->m_next)
1137 ;
1138 remainder = len;
1139 space = M_TRAILINGSPACE(m);
1140 if (space > 0) {
1141 /*
1142 * Copy into available space.
1143 */
1144 if (space > remainder)
1145 space = remainder;
1146 bcopy(cp, mtod(m, caddr_t) + m->m_len, space);
1147 m->m_len += space;
1148 cp += space, remainder -= space;
1149 }
1150 while (remainder > 0) {
1151 /*
1152 * Allocate a new mbuf; could check space
1153 * and allocate a cluster instead.
1154 */
1155 n = m_get(M_NOWAIT, m->m_type);
1156 if (n == NULL)
1157 break;
1158 n->m_len = min(MLEN, remainder);
1159 bcopy(cp, mtod(n, caddr_t), n->m_len);
1160 cp += n->m_len, remainder -= n->m_len;
1161 m->m_next = n;
1162 m = n;
1163 }
1164 if (m0->m_flags & M_PKTHDR)
1165 m0->m_pkthdr.len += len - remainder;
1166 return (remainder == 0);
1167 }
1168
1169 /*
1170 * Apply function f to the data in an mbuf chain starting "off" bytes from
1171 * the beginning, continuing for "len" bytes.
1172 */
1173 int
1174 m_apply(struct mbuf *m, int off, int len,
1175 int (*f)(void *, void *, u_int), void *arg)
1176 {
1177 u_int count;
1178 int rval;
1179
1180 KASSERT(off >= 0, ("m_apply, negative off %d", off));
1181 KASSERT(len >= 0, ("m_apply, negative len %d", len));
1182 while (off > 0) {
1183 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1184 if (off < m->m_len)
1185 break;
1186 off -= m->m_len;
1187 m = m->m_next;
1188 }
1189 while (len > 0) {
1190 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain"));
1191 count = min(m->m_len - off, len);
1192 rval = (*f)(arg, mtod(m, caddr_t) + off, count);
1193 if (rval)
1194 return (rval);
1195 len -= count;
1196 off = 0;
1197 m = m->m_next;
1198 }
1199 return (0);
1200 }
1201
1202 /*
1203 * Return a pointer to mbuf/offset of location in mbuf chain.
1204 */
1205 struct mbuf *
1206 m_getptr(struct mbuf *m, int loc, int *off)
1207 {
1208
1209 while (loc >= 0) {
1210 /* Normal end of search. */
1211 if (m->m_len > loc) {
1212 *off = loc;
1213 return (m);
1214 } else {
1215 loc -= m->m_len;
1216 if (m->m_next == NULL) {
1217 if (loc == 0) {
1218 /* Point at the end of valid data. */
1219 *off = m->m_len;
1220 return (m);
1221 }
1222 return (NULL);
1223 }
1224 m = m->m_next;
1225 }
1226 }
1227 return (NULL);
1228 }
1229
1230 void
1231 m_print(const struct mbuf *m, int maxlen)
1232 {
1233 int len;
1234 int pdata;
1235 const struct mbuf *m2;
1236
1237 if (m == NULL) {
1238 printf("mbuf: %p\n", m);
1239 return;
1240 }
1241
1242 if (m->m_flags & M_PKTHDR)
1243 len = m->m_pkthdr.len;
1244 else
1245 len = -1;
1246 m2 = m;
1247 while (m2 != NULL && (len == -1 || len)) {
1248 pdata = m2->m_len;
1249 if (maxlen != -1 && pdata > maxlen)
1250 pdata = maxlen;
1251 printf("mbuf: %p len: %d, next: %p, %b%s", m2, m2->m_len,
1252 m2->m_next, m2->m_flags, "\2\20freelist\17skipfw"
1253 "\11proto5\10proto4\7proto3\6proto2\5proto1\4rdonly"
1254 "\3eor\2pkthdr\1ext", pdata ? "" : "\n");
1255 if (pdata)
1256 printf(", %*D\n", pdata, (u_char *)m2->m_data, "-");
1257 if (len != -1)
1258 len -= m2->m_len;
1259 m2 = m2->m_next;
1260 }
1261 if (len > 0)
1262 printf("%d bytes unaccounted for.\n", len);
1263 return;
1264 }
1265
1266 u_int
1267 m_fixhdr(struct mbuf *m0)
1268 {
1269 u_int len;
1270
1271 len = m_length(m0, NULL);
1272 m0->m_pkthdr.len = len;
1273 return (len);
1274 }
1275
1276 u_int
1277 m_length(struct mbuf *m0, struct mbuf **last)
1278 {
1279 struct mbuf *m;
1280 u_int len;
1281
1282 len = 0;
1283 for (m = m0; m != NULL; m = m->m_next) {
1284 len += m->m_len;
1285 if (m->m_next == NULL)
1286 break;
1287 }
1288 if (last != NULL)
1289 *last = m;
1290 return (len);
1291 }
1292
1293 /*
1294 * Defragment a mbuf chain, returning the shortest possible
1295 * chain of mbufs and clusters. If allocation fails and
1296 * this cannot be completed, NULL will be returned, but
1297 * the passed in chain will be unchanged. Upon success,
1298 * the original chain will be freed, and the new chain
1299 * will be returned.
1300 *
1301 * If a non-packet header is passed in, the original
1302 * mbuf (chain?) will be returned unharmed.
1303 */
1304 struct mbuf *
1305 m_defrag(struct mbuf *m0, int how)
1306 {
1307 struct mbuf *m_new = NULL, *m_final = NULL;
1308 int progress = 0, length;
1309
1310 MBUF_CHECKSLEEP(how);
1311 if (!(m0->m_flags & M_PKTHDR))
1312 return (m0);
1313
1314 m_fixhdr(m0); /* Needed sanity check */
1315
1316 #ifdef MBUF_STRESS_TEST
1317 if (m_defragrandomfailures) {
1318 int temp = arc4random() & 0xff;
1319 if (temp == 0xba)
1320 goto nospace;
1321 }
1322 #endif
1323
1324 if (m0->m_pkthdr.len > MHLEN)
1325 m_final = m_getcl(how, MT_DATA, M_PKTHDR);
1326 else
1327 m_final = m_gethdr(how, MT_DATA);
1328
1329 if (m_final == NULL)
1330 goto nospace;
1331
1332 if (m_dup_pkthdr(m_final, m0, how) == 0)
1333 goto nospace;
1334
1335 m_new = m_final;
1336
1337 while (progress < m0->m_pkthdr.len) {
1338 length = m0->m_pkthdr.len - progress;
1339 if (length > MCLBYTES)
1340 length = MCLBYTES;
1341
1342 if (m_new == NULL) {
1343 if (length > MLEN)
1344 m_new = m_getcl(how, MT_DATA, 0);
1345 else
1346 m_new = m_get(how, MT_DATA);
1347 if (m_new == NULL)
1348 goto nospace;
1349 }
1350
1351 m_copydata(m0, progress, length, mtod(m_new, caddr_t));
1352 progress += length;
1353 m_new->m_len = length;
1354 if (m_new != m_final)
1355 m_cat(m_final, m_new);
1356 m_new = NULL;
1357 }
1358 #ifdef MBUF_STRESS_TEST
1359 if (m0->m_next == NULL)
1360 m_defraguseless++;
1361 #endif
1362 m_freem(m0);
1363 m0 = m_final;
1364 #ifdef MBUF_STRESS_TEST
1365 m_defragpackets++;
1366 m_defragbytes += m0->m_pkthdr.len;
1367 #endif
1368 return (m0);
1369 nospace:
1370 #ifdef MBUF_STRESS_TEST
1371 m_defragfailure++;
1372 #endif
1373 if (m_final)
1374 m_freem(m_final);
1375 return (NULL);
1376 }
1377
1378 /*
1379 * Defragment an mbuf chain, returning at most maxfrags separate
1380 * mbufs+clusters. If this is not possible NULL is returned and
1381 * the original mbuf chain is left in its present (potentially
1382 * modified) state. We use two techniques: collapsing consecutive
1383 * mbufs and replacing consecutive mbufs by a cluster.
1384 *
1385 * NB: this should really be named m_defrag but that name is taken
1386 */
1387 struct mbuf *
1388 m_collapse(struct mbuf *m0, int how, int maxfrags)
1389 {
1390 struct mbuf *m, *n, *n2, **prev;
1391 u_int curfrags;
1392
1393 /*
1394 * Calculate the current number of frags.
1395 */
1396 curfrags = 0;
1397 for (m = m0; m != NULL; m = m->m_next)
1398 curfrags++;
1399 /*
1400 * First, try to collapse mbufs. Note that we always collapse
1401 * towards the front so we don't need to deal with moving the
1402 * pkthdr. This may be suboptimal if the first mbuf has much
1403 * less data than the following.
1404 */
1405 m = m0;
1406 again:
1407 for (;;) {
1408 n = m->m_next;
1409 if (n == NULL)
1410 break;
1411 if (M_WRITABLE(m) &&
1412 n->m_len < M_TRAILINGSPACE(m)) {
1413 bcopy(mtod(n, void *), mtod(m, char *) + m->m_len,
1414 n->m_len);
1415 m->m_len += n->m_len;
1416 m->m_next = n->m_next;
1417 m_free(n);
1418 if (--curfrags <= maxfrags)
1419 return m0;
1420 } else
1421 m = n;
1422 }
1423 KASSERT(maxfrags > 1,
1424 ("maxfrags %u, but normal collapse failed", maxfrags));
1425 /*
1426 * Collapse consecutive mbufs to a cluster.
1427 */
1428 prev = &m0->m_next; /* NB: not the first mbuf */
1429 while ((n = *prev) != NULL) {
1430 if ((n2 = n->m_next) != NULL &&
1431 n->m_len + n2->m_len < MCLBYTES) {
1432 m = m_getcl(how, MT_DATA, 0);
1433 if (m == NULL)
1434 goto bad;
1435 bcopy(mtod(n, void *), mtod(m, void *), n->m_len);
1436 bcopy(mtod(n2, void *), mtod(m, char *) + n->m_len,
1437 n2->m_len);
1438 m->m_len = n->m_len + n2->m_len;
1439 m->m_next = n2->m_next;
1440 *prev = m;
1441 m_free(n);
1442 m_free(n2);
1443 if (--curfrags <= maxfrags) /* +1 cl -2 mbufs */
1444 return m0;
1445 /*
1446 * Still not there, try the normal collapse
1447 * again before we allocate another cluster.
1448 */
1449 goto again;
1450 }
1451 prev = &n->m_next;
1452 }
1453 /*
1454 * No place where we can collapse to a cluster; punt.
1455 * This can occur if, for example, you request 2 frags
1456 * but the packet requires that both be clusters (we
1457 * never reallocate the first mbuf to avoid moving the
1458 * packet header).
1459 */
1460 bad:
1461 return NULL;
1462 }
1463
1464 #ifdef MBUF_STRESS_TEST
1465
1466 /*
1467 * Fragment an mbuf chain. There's no reason you'd ever want to do
1468 * this in normal usage, but it's great for stress testing various
1469 * mbuf consumers.
1470 *
1471 * If fragmentation is not possible, the original chain will be
1472 * returned.
1473 *
1474 * Possible length values:
1475 * 0 no fragmentation will occur
1476 * > 0 each fragment will be of the specified length
1477 * -1 each fragment will be the same random value in length
1478 * -2 each fragment's length will be entirely random
1479 * (Random values range from 1 to 256)
1480 */
1481 struct mbuf *
1482 m_fragment(struct mbuf *m0, int how, int length)
1483 {
1484 struct mbuf *m_first, *m_last;
1485 int divisor = 255, progress = 0, fraglen;
1486
1487 if (!(m0->m_flags & M_PKTHDR))
1488 return (m0);
1489
1490 if (length == 0 || length < -2)
1491 return (m0);
1492 if (length > MCLBYTES)
1493 length = MCLBYTES;
1494 if (length < 0 && divisor > MCLBYTES)
1495 divisor = MCLBYTES;
1496 if (length == -1)
1497 length = 1 + (arc4random() % divisor);
1498 if (length > 0)
1499 fraglen = length;
1500
1501 m_fixhdr(m0); /* Needed sanity check */
1502
1503 m_first = m_getcl(how, MT_DATA, M_PKTHDR);
1504 if (m_first == NULL)
1505 goto nospace;
1506
1507 if (m_dup_pkthdr(m_first, m0, how) == 0)
1508 goto nospace;
1509
1510 m_last = m_first;
1511
1512 while (progress < m0->m_pkthdr.len) {
1513 if (length == -2)
1514 fraglen = 1 + (arc4random() % divisor);
1515 if (fraglen > m0->m_pkthdr.len - progress)
1516 fraglen = m0->m_pkthdr.len - progress;
1517
1518 if (progress != 0) {
1519 struct mbuf *m_new = m_getcl(how, MT_DATA, 0);
1520 if (m_new == NULL)
1521 goto nospace;
1522
1523 m_last->m_next = m_new;
1524 m_last = m_new;
1525 }
1526
1527 m_copydata(m0, progress, fraglen, mtod(m_last, caddr_t));
1528 progress += fraglen;
1529 m_last->m_len = fraglen;
1530 }
1531 m_freem(m0);
1532 m0 = m_first;
1533 return (m0);
1534 nospace:
1535 if (m_first)
1536 m_freem(m_first);
1537 /* Return the original chain on failure */
1538 return (m0);
1539 }
1540
1541 #endif
1542
1543 /*
1544 * Copy the contents of uio into a properly sized mbuf chain.
1545 */
1546 struct mbuf *
1547 m_uiotombuf(struct uio *uio, int how, int len, int align, int flags)
1548 {
1549 struct mbuf *m, *mb;
1550 int error, length;
1551 ssize_t total;
1552 int progress = 0;
1553
1554 /*
1555 * len can be zero or an arbitrary large value bound by
1556 * the total data supplied by the uio.
1557 */
1558 if (len > 0)
1559 total = (uio->uio_resid < len) ? uio->uio_resid : len;
1560 else
1561 total = uio->uio_resid;
1562
1563 /*
1564 * The smallest unit returned by m_getm2() is a single mbuf
1565 * with pkthdr. We can't align past it.
1566 */
1567 if (align >= MHLEN)
1568 return (NULL);
1569
1570 /*
1571 * Give us the full allocation or nothing.
1572 * If len is zero return the smallest empty mbuf.
1573 */
1574 m = m_getm2(NULL, max(total + align, 1), how, MT_DATA, flags);
1575 if (m == NULL)
1576 return (NULL);
1577 m->m_data += align;
1578
1579 /* Fill all mbufs with uio data and update header information. */
1580 for (mb = m; mb != NULL; mb = mb->m_next) {
1581 length = min(M_TRAILINGSPACE(mb), total - progress);
1582
1583 error = uiomove(mtod(mb, void *), length, uio);
1584 if (error) {
1585 m_freem(m);
1586 return (NULL);
1587 }
1588
1589 mb->m_len = length;
1590 progress += length;
1591 if (flags & M_PKTHDR)
1592 m->m_pkthdr.len += length;
1593 }
1594 KASSERT(progress == total, ("%s: progress != total", __func__));
1595
1596 return (m);
1597 }
1598
1599 /*
1600 * Copy an mbuf chain into a uio limited by len if set.
1601 */
1602 int
1603 m_mbuftouio(struct uio *uio, const struct mbuf *m, int len)
1604 {
1605 int error, length, total;
1606 int progress = 0;
1607
1608 if (len > 0)
1609 total = min(uio->uio_resid, len);
1610 else
1611 total = uio->uio_resid;
1612
1613 /* Fill the uio with data from the mbufs. */
1614 for (; m != NULL; m = m->m_next) {
1615 length = min(m->m_len, total - progress);
1616
1617 error = uiomove(mtod(m, void *), length, uio);
1618 if (error)
1619 return (error);
1620
1621 progress += length;
1622 }
1623
1624 return (0);
1625 }
1626
1627 /*
1628 * Create a writable copy of the mbuf chain. While doing this
1629 * we compact the chain with a goal of producing a chain with
1630 * at most two mbufs. The second mbuf in this chain is likely
1631 * to be a cluster. The primary purpose of this work is to create
1632 * a writable packet for encryption, compression, etc. The
1633 * secondary goal is to linearize the data so the data can be
1634 * passed to crypto hardware in the most efficient manner possible.
1635 */
1636 struct mbuf *
1637 m_unshare(struct mbuf *m0, int how)
1638 {
1639 struct mbuf *m, *mprev;
1640 struct mbuf *n, *mfirst, *mlast;
1641 int len, off;
1642
1643 mprev = NULL;
1644 for (m = m0; m != NULL; m = mprev->m_next) {
1645 /*
1646 * Regular mbufs are ignored unless there's a cluster
1647 * in front of it that we can use to coalesce. We do
1648 * the latter mainly so later clusters can be coalesced
1649 * also w/o having to handle them specially (i.e. convert
1650 * mbuf+cluster -> cluster). This optimization is heavily
1651 * influenced by the assumption that we're running over
1652 * Ethernet where MCLBYTES is large enough that the max
1653 * packet size will permit lots of coalescing into a
1654 * single cluster. This in turn permits efficient
1655 * crypto operations, especially when using hardware.
1656 */
1657 if ((m->m_flags & M_EXT) == 0) {
1658 if (mprev && (mprev->m_flags & M_EXT) &&
1659 m->m_len <= M_TRAILINGSPACE(mprev)) {
1660 /* XXX: this ignores mbuf types */
1661 memcpy(mtod(mprev, caddr_t) + mprev->m_len,
1662 mtod(m, caddr_t), m->m_len);
1663 mprev->m_len += m->m_len;
1664 mprev->m_next = m->m_next; /* unlink from chain */
1665 m_free(m); /* reclaim mbuf */
1666 } else {
1667 mprev = m;
1668 }
1669 continue;
1670 }
1671 /*
1672 * Writable mbufs are left alone (for now).
1673 */
1674 if (M_WRITABLE(m)) {
1675 mprev = m;
1676 continue;
1677 }
1678
1679 /*
1680 * Not writable, replace with a copy or coalesce with
1681 * the previous mbuf if possible (since we have to copy
1682 * it anyway, we try to reduce the number of mbufs and
1683 * clusters so that future work is easier).
1684 */
1685 KASSERT(m->m_flags & M_EXT, ("m_flags 0x%x", m->m_flags));
1686 /* NB: we only coalesce into a cluster or larger */
1687 if (mprev != NULL && (mprev->m_flags & M_EXT) &&
1688 m->m_len <= M_TRAILINGSPACE(mprev)) {
1689 /* XXX: this ignores mbuf types */
1690 memcpy(mtod(mprev, caddr_t) + mprev->m_len,
1691 mtod(m, caddr_t), m->m_len);
1692 mprev->m_len += m->m_len;
1693 mprev->m_next = m->m_next; /* unlink from chain */
1694 m_free(m); /* reclaim mbuf */
1695 continue;
1696 }
1697
1698 /*
1699 * Allocate new space to hold the copy and copy the data.
1700 * We deal with jumbo mbufs (i.e. m_len > MCLBYTES) by
1701 * splitting them into clusters. We could just malloc a
1702 * buffer and make it external but too many device drivers
1703 * don't know how to break up the non-contiguous memory when
1704 * doing DMA.
1705 */
1706 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS);
1707 if (n == NULL) {
1708 m_freem(m0);
1709 return (NULL);
1710 }
1711 if (m->m_flags & M_PKTHDR) {
1712 KASSERT(mprev == NULL, ("%s: m0 %p, m %p has M_PKTHDR",
1713 __func__, m0, m));
1714 m_move_pkthdr(n, m);
1715 }
1716 len = m->m_len;
1717 off = 0;
1718 mfirst = n;
1719 mlast = NULL;
1720 for (;;) {
1721 int cc = min(len, MCLBYTES);
1722 memcpy(mtod(n, caddr_t), mtod(m, caddr_t) + off, cc);
1723 n->m_len = cc;
1724 if (mlast != NULL)
1725 mlast->m_next = n;
1726 mlast = n;
1727 #if 0
1728 newipsecstat.ips_clcopied++;
1729 #endif
1730
1731 len -= cc;
1732 if (len <= 0)
1733 break;
1734 off += cc;
1735
1736 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS);
1737 if (n == NULL) {
1738 m_freem(mfirst);
1739 m_freem(m0);
1740 return (NULL);
1741 }
1742 }
1743 n->m_next = m->m_next;
1744 if (mprev == NULL)
1745 m0 = mfirst; /* new head of chain */
1746 else
1747 mprev->m_next = mfirst; /* replace old mbuf */
1748 m_free(m); /* release old mbuf */
1749 mprev = mfirst;
1750 }
1751 return (m0);
1752 }
1753
1754 #ifdef MBUF_PROFILING
1755
1756 #define MP_BUCKETS 32 /* don't just change this as things may overflow.*/
1757 struct mbufprofile {
1758 uintmax_t wasted[MP_BUCKETS];
1759 uintmax_t used[MP_BUCKETS];
1760 uintmax_t segments[MP_BUCKETS];
1761 } mbprof;
1762
1763 #define MP_MAXDIGITS 21 /* strlen("16,000,000,000,000,000,000") == 21 */
1764 #define MP_NUMLINES 6
1765 #define MP_NUMSPERLINE 16
1766 #define MP_EXTRABYTES 64 /* > strlen("used:\nwasted:\nsegments:\n") */
1767 /* work out max space needed and add a bit of spare space too */
1768 #define MP_MAXLINE ((MP_MAXDIGITS+1) * MP_NUMSPERLINE)
1769 #define MP_BUFSIZE ((MP_MAXLINE * MP_NUMLINES) + 1 + MP_EXTRABYTES)
1770
1771 char mbprofbuf[MP_BUFSIZE];
1772
1773 void
1774 m_profile(struct mbuf *m)
1775 {
1776 int segments = 0;
1777 int used = 0;
1778 int wasted = 0;
1779
1780 while (m) {
1781 segments++;
1782 used += m->m_len;
1783 if (m->m_flags & M_EXT) {
1784 wasted += MHLEN - sizeof(m->m_ext) +
1785 m->m_ext.ext_size - m->m_len;
1786 } else {
1787 if (m->m_flags & M_PKTHDR)
1788 wasted += MHLEN - m->m_len;
1789 else
1790 wasted += MLEN - m->m_len;
1791 }
1792 m = m->m_next;
1793 }
1794 /* be paranoid.. it helps */
1795 if (segments > MP_BUCKETS - 1)
1796 segments = MP_BUCKETS - 1;
1797 if (used > 100000)
1798 used = 100000;
1799 if (wasted > 100000)
1800 wasted = 100000;
1801 /* store in the appropriate bucket */
1802 /* don't bother locking. if it's slightly off, so what? */
1803 mbprof.segments[segments]++;
1804 mbprof.used[fls(used)]++;
1805 mbprof.wasted[fls(wasted)]++;
1806 }
1807
1808 static void
1809 mbprof_textify(void)
1810 {
1811 int offset;
1812 char *c;
1813 uint64_t *p;
1814
1815 p = &mbprof.wasted[0];
1816 c = mbprofbuf;
1817 offset = snprintf(c, MP_MAXLINE + 10,
1818 "wasted:\n"
1819 "%ju %ju %ju %ju %ju %ju %ju %ju "
1820 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1821 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1822 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1823 #ifdef BIG_ARRAY
1824 p = &mbprof.wasted[16];
1825 c += offset;
1826 offset = snprintf(c, MP_MAXLINE,
1827 "%ju %ju %ju %ju %ju %ju %ju %ju "
1828 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1829 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1830 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1831 #endif
1832 p = &mbprof.used[0];
1833 c += offset;
1834 offset = snprintf(c, MP_MAXLINE + 10,
1835 "used:\n"
1836 "%ju %ju %ju %ju %ju %ju %ju %ju "
1837 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1838 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1839 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1840 #ifdef BIG_ARRAY
1841 p = &mbprof.used[16];
1842 c += offset;
1843 offset = snprintf(c, MP_MAXLINE,
1844 "%ju %ju %ju %ju %ju %ju %ju %ju "
1845 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1846 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1847 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1848 #endif
1849 p = &mbprof.segments[0];
1850 c += offset;
1851 offset = snprintf(c, MP_MAXLINE + 10,
1852 "segments:\n"
1853 "%ju %ju %ju %ju %ju %ju %ju %ju "
1854 "%ju %ju %ju %ju %ju %ju %ju %ju\n",
1855 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1856 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1857 #ifdef BIG_ARRAY
1858 p = &mbprof.segments[16];
1859 c += offset;
1860 offset = snprintf(c, MP_MAXLINE,
1861 "%ju %ju %ju %ju %ju %ju %ju %ju "
1862 "%ju %ju %ju %ju %ju %ju %ju %jju",
1863 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7],
1864 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]);
1865 #endif
1866 }
1867
1868 static int
1869 mbprof_handler(SYSCTL_HANDLER_ARGS)
1870 {
1871 int error;
1872
1873 mbprof_textify();
1874 error = SYSCTL_OUT(req, mbprofbuf, strlen(mbprofbuf) + 1);
1875 return (error);
1876 }
1877
1878 static int
1879 mbprof_clr_handler(SYSCTL_HANDLER_ARGS)
1880 {
1881 int clear, error;
1882
1883 clear = 0;
1884 error = sysctl_handle_int(oidp, &clear, 0, req);
1885 if (error || !req->newptr)
1886 return (error);
1887
1888 if (clear) {
1889 bzero(&mbprof, sizeof(mbprof));
1890 }
1891
1892 return (error);
1893 }
1894
1895
1896 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofile, CTLTYPE_STRING|CTLFLAG_RD,
1897 NULL, 0, mbprof_handler, "A", "mbuf profiling statistics");
1898
1899 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofileclr, CTLTYPE_INT|CTLFLAG_RW,
1900 NULL, 0, mbprof_clr_handler, "I", "clear mbuf profiling statistics");
1901 #endif
1902
Cache object: aad955d54a3f59ef87c86b9946fad4c0
|