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_DEFINE3_XLATE(sdt, , , m__clget, 78 "struct mbuf *", "mbufinfo_t *", 79 "uint32_t", "uint32_t", 80 "uint32_t", "uint32_t"); 81 82 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__cljget, 83 "struct mbuf *", "mbufinfo_t *", 84 "uint32_t", "uint32_t", 85 "uint32_t", "uint32_t", 86 "void*", "void*"); 87 88 SDT_PROBE_DEFINE(sdt, , , m__cljset); 89 90 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__free, 91 "struct mbuf *", "mbufinfo_t *"); 92 93 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__freem, 94 "struct mbuf *", "mbufinfo_t *"); 95 96 #include <security/mac/mac_framework.h> 97 98 int max_linkhdr; 99 int max_protohdr; 100 int max_hdr; 101 int max_datalen; 102 #ifdef MBUF_STRESS_TEST 103 int m_defragpackets; 104 int m_defragbytes; 105 int m_defraguseless; 106 int m_defragfailure; 107 int m_defragrandomfailures; 108 #endif 109 110 /* 111 * sysctl(8) exported objects 112 */ 113 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RD, 114 &max_linkhdr, 0, "Size of largest link layer header"); 115 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RD, 116 &max_protohdr, 0, "Size of largest protocol layer header"); 117 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RD, 118 &max_hdr, 0, "Size of largest link plus protocol header"); 119 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RD, 120 &max_datalen, 0, "Minimum space left in mbuf after max_hdr"); 121 #ifdef MBUF_STRESS_TEST 122 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD, 123 &m_defragpackets, 0, ""); 124 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD, 125 &m_defragbytes, 0, ""); 126 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD, 127 &m_defraguseless, 0, ""); 128 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD, 129 &m_defragfailure, 0, ""); 130 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW, 131 &m_defragrandomfailures, 0, ""); 132 #endif 133 134 /* 135 * Ensure the correct size of various mbuf parameters. It could be off due 136 * to compiler-induced padding and alignment artifacts. 137 */ 138 CTASSERT(MSIZE - offsetof(struct mbuf, m_dat) == MLEN); 139 CTASSERT(MSIZE - offsetof(struct mbuf, m_pktdat) == MHLEN); 140 141 /* 142 * mbuf data storage should be 64-bit aligned regardless of architectural 143 * pointer size; check this is the case with and without a packet header. 144 */ 145 CTASSERT(offsetof(struct mbuf, m_dat) % 8 == 0); 146 CTASSERT(offsetof(struct mbuf, m_pktdat) % 8 == 0); 147 148 /* 149 * While the specific values here don't matter too much (i.e., +/- a few 150 * words), we do want to ensure that changes to these values are carefully 151 * reasoned about and properly documented. This is especially the case as 152 * network-protocol and device-driver modules encode these layouts, and must 153 * be recompiled if the structures change. Check these values at compile time 154 * against the ones documented in comments in mbuf.h. 155 * 156 * NB: Possibly they should be documented there via #define's and not just 157 * comments. 158 */ 159 #if defined(__LP64__) 160 CTASSERT(offsetof(struct mbuf, m_dat) == 32); 161 CTASSERT(sizeof(struct pkthdr) == 56); 162 CTASSERT(sizeof(struct m_ext) == 48); 163 #else 164 CTASSERT(offsetof(struct mbuf, m_dat) == 24); 165 CTASSERT(sizeof(struct pkthdr) == 48); 166 CTASSERT(sizeof(struct m_ext) == 28); 167 #endif 168 169 /* 170 * Assert that the queue(3) macros produce code of the same size as an old 171 * plain pointer does. 172 */ 173 #ifdef INVARIANTS 174 static struct mbuf __used m_assertbuf; 175 CTASSERT(sizeof(m_assertbuf.m_slist) == sizeof(m_assertbuf.m_next)); 176 CTASSERT(sizeof(m_assertbuf.m_stailq) == sizeof(m_assertbuf.m_next)); 177 CTASSERT(sizeof(m_assertbuf.m_slistpkt) == sizeof(m_assertbuf.m_nextpkt)); 178 CTASSERT(sizeof(m_assertbuf.m_stailqpkt) == sizeof(m_assertbuf.m_nextpkt)); 179 #endif 180 181 /* 182 * Attach the cluster from *m to *n, set up m_ext in *n 183 * and bump the refcount of the cluster. 184 */ 185 void 186 mb_dupcl(struct mbuf *n, struct mbuf *m) 187 { 188 volatile u_int *refcnt; 189 190 KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m)); 191 KASSERT(!(n->m_flags & M_EXT), ("%s: M_EXT set on %p", __func__, n)); 192 193 /* 194 * Cache access optimization. For most kinds of external 195 * storage we don't need full copy of m_ext, since the 196 * holder of the 'ext_count' is responsible to carry the 197 * free routine and its arguments. Exclusion is EXT_EXTREF, 198 * where 'ext_cnt' doesn't point into mbuf at all. 199 */ 200 if (m->m_ext.ext_type == EXT_EXTREF) 201 bcopy(&m->m_ext, &n->m_ext, sizeof(struct m_ext)); 202 else 203 bcopy(&m->m_ext, &n->m_ext, m_ext_copylen); 204 n->m_flags |= M_EXT; 205 n->m_flags |= m->m_flags & M_RDONLY; 206 207 /* See if this is the mbuf that holds the embedded refcount. */ 208 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) { 209 refcnt = n->m_ext.ext_cnt = &m->m_ext.ext_count; 210 n->m_ext.ext_flags &= ~EXT_FLAG_EMBREF; 211 } else { 212 KASSERT(m->m_ext.ext_cnt != NULL, 213 ("%s: no refcounting pointer on %p", __func__, m)); 214 refcnt = m->m_ext.ext_cnt; 215 } 216 217 if (*refcnt == 1) 218 *refcnt += 1; 219 else 220 atomic_add_int(refcnt, 1); 221 } 222 223 void 224 m_demote_pkthdr(struct mbuf *m) 225 { 226 227 M_ASSERTPKTHDR(m); 228 229 m_tag_delete_chain(m, NULL); 230 m->m_flags &= ~M_PKTHDR; 231 bzero(&m->m_pkthdr, sizeof(struct pkthdr)); 232 } 233 234 /* 235 * Clean up mbuf (chain) from any tags and packet headers. 236 * If "all" is set then the first mbuf in the chain will be 237 * cleaned too. 238 */ 239 void 240 m_demote(struct mbuf *m0, int all, int flags) 241 { 242 struct mbuf *m; 243 244 for (m = all ? m0 : m0->m_next; m != NULL; m = m->m_next) { 245 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt in m %p, m0 %p", 246 __func__, m, m0)); 247 if (m->m_flags & M_PKTHDR) 248 m_demote_pkthdr(m); 249 m->m_flags = m->m_flags & (M_EXT | M_RDONLY | M_NOFREE | flags); 250 } 251 } 252 253 /* 254 * Sanity checks on mbuf (chain) for use in KASSERT() and general 255 * debugging. 256 * Returns 0 or panics when bad and 1 on all tests passed. 257 * Sanitize, 0 to run M_SANITY_ACTION, 1 to garble things so they 258 * blow up later. 259 */ 260 int 261 m_sanity(struct mbuf *m0, int sanitize) 262 { 263 struct mbuf *m; 264 caddr_t a, b; 265 int pktlen = 0; 266 267 #ifdef INVARIANTS 268 #define M_SANITY_ACTION(s) panic("mbuf %p: " s, m) 269 #else 270 #define M_SANITY_ACTION(s) printf("mbuf %p: " s, m) 271 #endif 272 273 for (m = m0; m != NULL; m = m->m_next) { 274 /* 275 * Basic pointer checks. If any of these fails then some 276 * unrelated kernel memory before or after us is trashed. 277 * No way to recover from that. 278 */ 279 a = M_START(m); 280 b = a + M_SIZE(m); 281 if ((caddr_t)m->m_data < a) 282 M_SANITY_ACTION("m_data outside mbuf data range left"); 283 if ((caddr_t)m->m_data > b) 284 M_SANITY_ACTION("m_data outside mbuf data range right"); 285 if ((caddr_t)m->m_data + m->m_len > b) 286 M_SANITY_ACTION("m_data + m_len exeeds mbuf space"); 287 288 /* m->m_nextpkt may only be set on first mbuf in chain. */ 289 if (m != m0 && m->m_nextpkt != NULL) { 290 if (sanitize) { 291 m_freem(m->m_nextpkt); 292 m->m_nextpkt = (struct mbuf *)0xDEADC0DE; 293 } else 294 M_SANITY_ACTION("m->m_nextpkt on in-chain mbuf"); 295 } 296 297 /* packet length (not mbuf length!) calculation */ 298 if (m0->m_flags & M_PKTHDR) 299 pktlen += m->m_len; 300 301 /* m_tags may only be attached to first mbuf in chain. */ 302 if (m != m0 && m->m_flags & M_PKTHDR && 303 !SLIST_EMPTY(&m->m_pkthdr.tags)) { 304 if (sanitize) { 305 m_tag_delete_chain(m, NULL); 306 /* put in 0xDEADC0DE perhaps? */ 307 } else 308 M_SANITY_ACTION("m_tags on in-chain mbuf"); 309 } 310 311 /* M_PKTHDR may only be set on first mbuf in chain */ 312 if (m != m0 && m->m_flags & M_PKTHDR) { 313 if (sanitize) { 314 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr)); 315 m->m_flags &= ~M_PKTHDR; 316 /* put in 0xDEADCODE and leave hdr flag in */ 317 } else 318 M_SANITY_ACTION("M_PKTHDR on in-chain mbuf"); 319 } 320 } 321 m = m0; 322 if (pktlen && pktlen != m->m_pkthdr.len) { 323 if (sanitize) 324 m->m_pkthdr.len = 0; 325 else 326 M_SANITY_ACTION("m_pkthdr.len != mbuf chain length"); 327 } 328 return 1; 329 330 #undef M_SANITY_ACTION 331 } 332 333 /* 334 * Non-inlined part of m_init(). 335 */ 336 int 337 m_pkthdr_init(struct mbuf *m, int how) 338 { 339 #ifdef MAC 340 int error; 341 #endif 342 m->m_data = m->m_pktdat; 343 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr)); 344 #ifdef MAC 345 /* If the label init fails, fail the alloc */ 346 error = mac_mbuf_init(m, how); 347 if (error) 348 return (error); 349 #endif 350 351 return (0); 352 } 353 354 /* 355 * "Move" mbuf pkthdr from "from" to "to". 356 * "from" must have M_PKTHDR set, and "to" must be empty. 357 */ 358 void 359 m_move_pkthdr(struct mbuf *to, struct mbuf *from) 360 { 361 362 #if 0 363 /* see below for why these are not enabled */ 364 M_ASSERTPKTHDR(to); 365 /* Note: with MAC, this may not be a good assertion. */ 366 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags), 367 ("m_move_pkthdr: to has tags")); 368 #endif 369 #ifdef MAC 370 /* 371 * XXXMAC: It could be this should also occur for non-MAC? 372 */ 373 if (to->m_flags & M_PKTHDR) 374 m_tag_delete_chain(to, NULL); 375 #endif 376 to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT); 377 if ((to->m_flags & M_EXT) == 0) 378 to->m_data = to->m_pktdat; 379 to->m_pkthdr = from->m_pkthdr; /* especially tags */ 380 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */ 381 from->m_flags &= ~M_PKTHDR; 382 } 383 384 /* 385 * Duplicate "from"'s mbuf pkthdr in "to". 386 * "from" must have M_PKTHDR set, and "to" must be empty. 387 * In particular, this does a deep copy of the packet tags. 388 */ 389 int 390 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how) 391 { 392 393 #if 0 394 /* 395 * The mbuf allocator only initializes the pkthdr 396 * when the mbuf is allocated with m_gethdr(). Many users 397 * (e.g. m_copy*, m_prepend) use m_get() and then 398 * smash the pkthdr as needed causing these 399 * assertions to trip. For now just disable them. 400 */ 401 M_ASSERTPKTHDR(to); 402 /* Note: with MAC, this may not be a good assertion. */ 403 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags), ("m_dup_pkthdr: to has tags")); 404 #endif 405 MBUF_CHECKSLEEP(how); 406 #ifdef MAC 407 if (to->m_flags & M_PKTHDR) 408 m_tag_delete_chain(to, NULL); 409 #endif 410 to->m_flags = (from->m_flags & M_COPYFLAGS) | (to->m_flags & M_EXT); 411 if ((to->m_flags & M_EXT) == 0) 412 to->m_data = to->m_pktdat; 413 to->m_pkthdr = from->m_pkthdr; 414 SLIST_INIT(&to->m_pkthdr.tags); 415 return (m_tag_copy_chain(to, from, how)); 416 } 417 418 /* 419 * Lesser-used path for M_PREPEND: 420 * allocate new mbuf to prepend to chain, 421 * copy junk along. 422 */ 423 struct mbuf * 424 m_prepend(struct mbuf *m, int len, int how) 425 { 426 struct mbuf *mn; 427 428 if (m->m_flags & M_PKTHDR) 429 mn = m_gethdr(how, m->m_type); 430 else 431 mn = m_get(how, m->m_type); 432 if (mn == NULL) { 433 m_freem(m); 434 return (NULL); 435 } 436 if (m->m_flags & M_PKTHDR) 437 m_move_pkthdr(mn, m); 438 mn->m_next = m; 439 m = mn; 440 if (len < M_SIZE(m)) 441 M_ALIGN(m, len); 442 m->m_len = len; 443 return (m); 444 } 445 446 /* 447 * Make a copy of an mbuf chain starting "off0" bytes from the beginning, 448 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf. 449 * The wait parameter is a choice of M_WAITOK/M_NOWAIT from caller. 450 * Note that the copy is read-only, because clusters are not copied, 451 * only their reference counts are incremented. 452 */ 453 struct mbuf * 454 m_copym(struct mbuf *m, int off0, int len, int wait) 455 { 456 struct mbuf *n, **np; 457 int off = off0; 458 struct mbuf *top; 459 int copyhdr = 0; 460 461 KASSERT(off >= 0, ("m_copym, negative off %d", off)); 462 KASSERT(len >= 0, ("m_copym, negative len %d", len)); 463 MBUF_CHECKSLEEP(wait); 464 if (off == 0 && m->m_flags & M_PKTHDR) 465 copyhdr = 1; 466 while (off > 0) { 467 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain")); 468 if (off < m->m_len) 469 break; 470 off -= m->m_len; 471 m = m->m_next; 472 } 473 np = ⊤ 474 top = NULL; 475 while (len > 0) { 476 if (m == NULL) { 477 KASSERT(len == M_COPYALL, 478 ("m_copym, length > size of mbuf chain")); 479 break; 480 } 481 if (copyhdr) 482 n = m_gethdr(wait, m->m_type); 483 else 484 n = m_get(wait, m->m_type); 485 *np = n; 486 if (n == NULL) 487 goto nospace; 488 if (copyhdr) { 489 if (!m_dup_pkthdr(n, m, wait)) 490 goto nospace; 491 if (len == M_COPYALL) 492 n->m_pkthdr.len -= off0; 493 else 494 n->m_pkthdr.len = len; 495 copyhdr = 0; 496 } 497 n->m_len = min(len, m->m_len - off); 498 if (m->m_flags & M_EXT) { 499 n->m_data = m->m_data + off; 500 mb_dupcl(n, m); 501 } else 502 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), 503 (u_int)n->m_len); 504 if (len != M_COPYALL) 505 len -= n->m_len; 506 off = 0; 507 m = m->m_next; 508 np = &n->m_next; 509 } 510 511 return (top); 512 nospace: 513 m_freem(top); 514 return (NULL); 515 } 516 517 /* 518 * Copy an entire packet, including header (which must be present). 519 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'. 520 * Note that the copy is read-only, because clusters are not copied, 521 * only their reference counts are incremented. 522 * Preserve alignment of the first mbuf so if the creator has left 523 * some room at the beginning (e.g. for inserting protocol headers) 524 * the copies still have the room available. 525 */ 526 struct mbuf * 527 m_copypacket(struct mbuf *m, int how) 528 { 529 struct mbuf *top, *n, *o; 530 531 MBUF_CHECKSLEEP(how); 532 n = m_get(how, m->m_type); 533 top = n; 534 if (n == NULL) 535 goto nospace; 536 537 if (!m_dup_pkthdr(n, m, how)) 538 goto nospace; 539 n->m_len = m->m_len; 540 if (m->m_flags & M_EXT) { 541 n->m_data = m->m_data; 542 mb_dupcl(n, m); 543 } else { 544 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat ); 545 bcopy(mtod(m, char *), mtod(n, char *), n->m_len); 546 } 547 548 m = m->m_next; 549 while (m) { 550 o = m_get(how, m->m_type); 551 if (o == NULL) 552 goto nospace; 553 554 n->m_next = o; 555 n = n->m_next; 556 557 n->m_len = m->m_len; 558 if (m->m_flags & M_EXT) { 559 n->m_data = m->m_data; 560 mb_dupcl(n, m); 561 } else { 562 bcopy(mtod(m, char *), mtod(n, char *), n->m_len); 563 } 564 565 m = m->m_next; 566 } 567 return top; 568 nospace: 569 m_freem(top); 570 return (NULL); 571 } 572 573 /* 574 * Copy data from an mbuf chain starting "off" bytes from the beginning, 575 * continuing for "len" bytes, into the indicated buffer. 576 */ 577 void 578 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp) 579 { 580 u_int count; 581 582 KASSERT(off >= 0, ("m_copydata, negative off %d", off)); 583 KASSERT(len >= 0, ("m_copydata, negative len %d", len)); 584 while (off > 0) { 585 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain")); 586 if (off < m->m_len) 587 break; 588 off -= m->m_len; 589 m = m->m_next; 590 } 591 while (len > 0) { 592 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain")); 593 count = min(m->m_len - off, len); 594 bcopy(mtod(m, caddr_t) + off, cp, count); 595 len -= count; 596 cp += count; 597 off = 0; 598 m = m->m_next; 599 } 600 } 601 602 /* 603 * Copy a packet header mbuf chain into a completely new chain, including 604 * copying any mbuf clusters. Use this instead of m_copypacket() when 605 * you need a writable copy of an mbuf chain. 606 */ 607 struct mbuf * 608 m_dup(const struct mbuf *m, int how) 609 { 610 struct mbuf **p, *top = NULL; 611 int remain, moff, nsize; 612 613 MBUF_CHECKSLEEP(how); 614 /* Sanity check */ 615 if (m == NULL) 616 return (NULL); 617 M_ASSERTPKTHDR(m); 618 619 /* While there's more data, get a new mbuf, tack it on, and fill it */ 620 remain = m->m_pkthdr.len; 621 moff = 0; 622 p = ⊤ 623 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */ 624 struct mbuf *n; 625 626 /* Get the next new mbuf */ 627 if (remain >= MINCLSIZE) { 628 n = m_getcl(how, m->m_type, 0); 629 nsize = MCLBYTES; 630 } else { 631 n = m_get(how, m->m_type); 632 nsize = MLEN; 633 } 634 if (n == NULL) 635 goto nospace; 636 637 if (top == NULL) { /* First one, must be PKTHDR */ 638 if (!m_dup_pkthdr(n, m, how)) { 639 m_free(n); 640 goto nospace; 641 } 642 if ((n->m_flags & M_EXT) == 0) 643 nsize = MHLEN; 644 n->m_flags &= ~M_RDONLY; 645 } 646 n->m_len = 0; 647 648 /* Link it into the new chain */ 649 *p = n; 650 p = &n->m_next; 651 652 /* Copy data from original mbuf(s) into new mbuf */ 653 while (n->m_len < nsize && m != NULL) { 654 int chunk = min(nsize - n->m_len, m->m_len - moff); 655 656 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk); 657 moff += chunk; 658 n->m_len += chunk; 659 remain -= chunk; 660 if (moff == m->m_len) { 661 m = m->m_next; 662 moff = 0; 663 } 664 } 665 666 /* Check correct total mbuf length */ 667 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL), 668 ("%s: bogus m_pkthdr.len", __func__)); 669 } 670 return (top); 671 672 nospace: 673 m_freem(top); 674 return (NULL); 675 } 676 677 /* 678 * Concatenate mbuf chain n to m. 679 * Both chains must be of the same type (e.g. MT_DATA). 680 * Any m_pkthdr is not updated. 681 */ 682 void 683 m_cat(struct mbuf *m, struct mbuf *n) 684 { 685 while (m->m_next) 686 m = m->m_next; 687 while (n) { 688 if (!M_WRITABLE(m) || 689 M_TRAILINGSPACE(m) < n->m_len) { 690 /* just join the two chains */ 691 m->m_next = n; 692 return; 693 } 694 /* splat the data from one into the other */ 695 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len, 696 (u_int)n->m_len); 697 m->m_len += n->m_len; 698 n = m_free(n); 699 } 700 } 701 702 /* 703 * Concatenate two pkthdr mbuf chains. 704 */ 705 void 706 m_catpkt(struct mbuf *m, struct mbuf *n) 707 { 708 709 M_ASSERTPKTHDR(m); 710 M_ASSERTPKTHDR(n); 711 712 m->m_pkthdr.len += n->m_pkthdr.len; 713 m_demote(n, 1, 0); 714 715 m_cat(m, n); 716 } 717 718 void 719 m_adj(struct mbuf *mp, int req_len) 720 { 721 int len = req_len; 722 struct mbuf *m; 723 int count; 724 725 if ((m = mp) == NULL) 726 return; 727 if (len >= 0) { 728 /* 729 * Trim from head. 730 */ 731 while (m != NULL && len > 0) { 732 if (m->m_len <= len) { 733 len -= m->m_len; 734 m->m_len = 0; 735 m = m->m_next; 736 } else { 737 m->m_len -= len; 738 m->m_data += len; 739 len = 0; 740 } 741 } 742 if (mp->m_flags & M_PKTHDR) 743 mp->m_pkthdr.len -= (req_len - len); 744 } else { 745 /* 746 * Trim from tail. Scan the mbuf chain, 747 * calculating its length and finding the last mbuf. 748 * If the adjustment only affects this mbuf, then just 749 * adjust and return. Otherwise, rescan and truncate 750 * after the remaining size. 751 */ 752 len = -len; 753 count = 0; 754 for (;;) { 755 count += m->m_len; 756 if (m->m_next == (struct mbuf *)0) 757 break; 758 m = m->m_next; 759 } 760 if (m->m_len >= len) { 761 m->m_len -= len; 762 if (mp->m_flags & M_PKTHDR) 763 mp->m_pkthdr.len -= len; 764 return; 765 } 766 count -= len; 767 if (count < 0) 768 count = 0; 769 /* 770 * Correct length for chain is "count". 771 * Find the mbuf with last data, adjust its length, 772 * and toss data from remaining mbufs on chain. 773 */ 774 m = mp; 775 if (m->m_flags & M_PKTHDR) 776 m->m_pkthdr.len = count; 777 for (; m; m = m->m_next) { 778 if (m->m_len >= count) { 779 m->m_len = count; 780 if (m->m_next != NULL) { 781 m_freem(m->m_next); 782 m->m_next = NULL; 783 } 784 break; 785 } 786 count -= m->m_len; 787 } 788 } 789 } 790 791 /* 792 * Rearange an mbuf chain so that len bytes are contiguous 793 * and in the data area of an mbuf (so that mtod will work 794 * for a structure of size len). Returns the resulting 795 * mbuf chain on success, frees it and returns null on failure. 796 * If there is room, it will add up to max_protohdr-len extra bytes to the 797 * contiguous region in an attempt to avoid being called next time. 798 */ 799 struct mbuf * 800 m_pullup(struct mbuf *n, int len) 801 { 802 struct mbuf *m; 803 int count; 804 int space; 805 806 /* 807 * If first mbuf has no cluster, and has room for len bytes 808 * without shifting current data, pullup into it, 809 * otherwise allocate a new mbuf to prepend to the chain. 810 */ 811 if ((n->m_flags & M_EXT) == 0 && 812 n->m_data + len < &n->m_dat[MLEN] && n->m_next) { 813 if (n->m_len >= len) 814 return (n); 815 m = n; 816 n = n->m_next; 817 len -= m->m_len; 818 } else { 819 if (len > MHLEN) 820 goto bad; 821 m = m_get(M_NOWAIT, n->m_type); 822 if (m == NULL) 823 goto bad; 824 if (n->m_flags & M_PKTHDR) 825 m_move_pkthdr(m, n); 826 } 827 space = &m->m_dat[MLEN] - (m->m_data + m->m_len); 828 do { 829 count = min(min(max(len, max_protohdr), space), n->m_len); 830 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len, 831 (u_int)count); 832 len -= count; 833 m->m_len += count; 834 n->m_len -= count; 835 space -= count; 836 if (n->m_len) 837 n->m_data += count; 838 else 839 n = m_free(n); 840 } while (len > 0 && n); 841 if (len > 0) { 842 (void) m_free(m); 843 goto bad; 844 } 845 m->m_next = n; 846 return (m); 847 bad: 848 m_freem(n); 849 return (NULL); 850 } 851 852 /* 853 * Like m_pullup(), except a new mbuf is always allocated, and we allow 854 * the amount of empty space before the data in the new mbuf to be specified 855 * (in the event that the caller expects to prepend later). 856 */ 857 struct mbuf * 858 m_copyup(struct mbuf *n, int len, int dstoff) 859 { 860 struct mbuf *m; 861 int count, space; 862 863 if (len > (MHLEN - dstoff)) 864 goto bad; 865 m = m_get(M_NOWAIT, n->m_type); 866 if (m == NULL) 867 goto bad; 868 if (n->m_flags & M_PKTHDR) 869 m_move_pkthdr(m, n); 870 m->m_data += dstoff; 871 space = &m->m_dat[MLEN] - (m->m_data + m->m_len); 872 do { 873 count = min(min(max(len, max_protohdr), space), n->m_len); 874 memcpy(mtod(m, caddr_t) + m->m_len, mtod(n, caddr_t), 875 (unsigned)count); 876 len -= count; 877 m->m_len += count; 878 n->m_len -= count; 879 space -= count; 880 if (n->m_len) 881 n->m_data += count; 882 else 883 n = m_free(n); 884 } while (len > 0 && n); 885 if (len > 0) { 886 (void) m_free(m); 887 goto bad; 888 } 889 m->m_next = n; 890 return (m); 891 bad: 892 m_freem(n); 893 return (NULL); 894 } 895 896 /* 897 * Partition an mbuf chain in two pieces, returning the tail -- 898 * all but the first len0 bytes. In case of failure, it returns NULL and 899 * attempts to restore the chain to its original state. 900 * 901 * Note that the resulting mbufs might be read-only, because the new 902 * mbuf can end up sharing an mbuf cluster with the original mbuf if 903 * the "breaking point" happens to lie within a cluster mbuf. Use the 904 * M_WRITABLE() macro to check for this case. 905 */ 906 struct mbuf * 907 m_split(struct mbuf *m0, int len0, int wait) 908 { 909 struct mbuf *m, *n; 910 u_int len = len0, remain; 911 912 MBUF_CHECKSLEEP(wait); 913 for (m = m0; m && len > m->m_len; m = m->m_next) 914 len -= m->m_len; 915 if (m == NULL) 916 return (NULL); 917 remain = m->m_len - len; 918 if (m0->m_flags & M_PKTHDR && remain == 0) { 919 n = m_gethdr(wait, m0->m_type); 920 if (n == NULL) 921 return (NULL); 922 n->m_next = m->m_next; 923 m->m_next = NULL; 924 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif; 925 n->m_pkthdr.len = m0->m_pkthdr.len - len0; 926 m0->m_pkthdr.len = len0; 927 return (n); 928 } else if (m0->m_flags & M_PKTHDR) { 929 n = m_gethdr(wait, m0->m_type); 930 if (n == NULL) 931 return (NULL); 932 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif; 933 n->m_pkthdr.len = m0->m_pkthdr.len - len0; 934 m0->m_pkthdr.len = len0; 935 if (m->m_flags & M_EXT) 936 goto extpacket; 937 if (remain > MHLEN) { 938 /* m can't be the lead packet */ 939 M_ALIGN(n, 0); 940 n->m_next = m_split(m, len, wait); 941 if (n->m_next == NULL) { 942 (void) m_free(n); 943 return (NULL); 944 } else { 945 n->m_len = 0; 946 return (n); 947 } 948 } else 949 M_ALIGN(n, remain); 950 } else if (remain == 0) { 951 n = m->m_next; 952 m->m_next = NULL; 953 return (n); 954 } else { 955 n = m_get(wait, m->m_type); 956 if (n == NULL) 957 return (NULL); 958 M_ALIGN(n, remain); 959 } 960 extpacket: 961 if (m->m_flags & M_EXT) { 962 n->m_data = m->m_data + len; 963 mb_dupcl(n, m); 964 } else { 965 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain); 966 } 967 n->m_len = remain; 968 m->m_len = len; 969 n->m_next = m->m_next; 970 m->m_next = NULL; 971 return (n); 972 } 973 /* 974 * Routine to copy from device local memory into mbufs. 975 * Note that `off' argument is offset into first mbuf of target chain from 976 * which to begin copying the data to. 977 */ 978 struct mbuf * 979 m_devget(char *buf, int totlen, int off, struct ifnet *ifp, 980 void (*copy)(char *from, caddr_t to, u_int len)) 981 { 982 struct mbuf *m; 983 struct mbuf *top = NULL, **mp = ⊤ 984 int len; 985 986 if (off < 0 || off > MHLEN) 987 return (NULL); 988 989 while (totlen > 0) { 990 if (top == NULL) { /* First one, must be PKTHDR */ 991 if (totlen + off >= MINCLSIZE) { 992 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 993 len = MCLBYTES; 994 } else { 995 m = m_gethdr(M_NOWAIT, MT_DATA); 996 len = MHLEN; 997 998 /* Place initial small packet/header at end of mbuf */ 999 if (m && totlen + off + max_linkhdr <= MHLEN) { 1000 m->m_data += max_linkhdr; 1001 len -= max_linkhdr; 1002 } 1003 } 1004 if (m == NULL) 1005 return NULL; 1006 m->m_pkthdr.rcvif = ifp; 1007 m->m_pkthdr.len = totlen; 1008 } else { 1009 if (totlen + off >= MINCLSIZE) { 1010 m = m_getcl(M_NOWAIT, MT_DATA, 0); 1011 len = MCLBYTES; 1012 } else { 1013 m = m_get(M_NOWAIT, MT_DATA); 1014 len = MLEN; 1015 } 1016 if (m == NULL) { 1017 m_freem(top); 1018 return NULL; 1019 } 1020 } 1021 if (off) { 1022 m->m_data += off; 1023 len -= off; 1024 off = 0; 1025 } 1026 m->m_len = len = min(totlen, len); 1027 if (copy) 1028 copy(buf, mtod(m, caddr_t), (u_int)len); 1029 else 1030 bcopy(buf, mtod(m, caddr_t), (u_int)len); 1031 buf += len; 1032 *mp = m; 1033 mp = &m->m_next; 1034 totlen -= len; 1035 } 1036 return (top); 1037 } 1038 1039 /* 1040 * Copy data from a buffer back into the indicated mbuf chain, 1041 * starting "off" bytes from the beginning, extending the mbuf 1042 * chain if necessary. 1043 */ 1044 void 1045 m_copyback(struct mbuf *m0, int off, int len, c_caddr_t cp) 1046 { 1047 int mlen; 1048 struct mbuf *m = m0, *n; 1049 int totlen = 0; 1050 1051 if (m0 == NULL) 1052 return; 1053 while (off > (mlen = m->m_len)) { 1054 off -= mlen; 1055 totlen += mlen; 1056 if (m->m_next == NULL) { 1057 n = m_get(M_NOWAIT, m->m_type); 1058 if (n == NULL) 1059 goto out; 1060 bzero(mtod(n, caddr_t), MLEN); 1061 n->m_len = min(MLEN, len + off); 1062 m->m_next = n; 1063 } 1064 m = m->m_next; 1065 } 1066 while (len > 0) { 1067 if (m->m_next == NULL && (len > m->m_len - off)) { 1068 m->m_len += min(len - (m->m_len - off), 1069 M_TRAILINGSPACE(m)); 1070 } 1071 mlen = min (m->m_len - off, len); 1072 bcopy(cp, off + mtod(m, caddr_t), (u_int)mlen); 1073 cp += mlen; 1074 len -= mlen; 1075 mlen += off; 1076 off = 0; 1077 totlen += mlen; 1078 if (len == 0) 1079 break; 1080 if (m->m_next == NULL) { 1081 n = m_get(M_NOWAIT, m->m_type); 1082 if (n == NULL) 1083 break; 1084 n->m_len = min(MLEN, len); 1085 m->m_next = n; 1086 } 1087 m = m->m_next; 1088 } 1089 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen)) 1090 m->m_pkthdr.len = totlen; 1091 } 1092 1093 /* 1094 * Append the specified data to the indicated mbuf chain, 1095 * Extend the mbuf chain if the new data does not fit in 1096 * existing space. 1097 * 1098 * Return 1 if able to complete the job; otherwise 0. 1099 */ 1100 int 1101 m_append(struct mbuf *m0, int len, c_caddr_t cp) 1102 { 1103 struct mbuf *m, *n; 1104 int remainder, space; 1105 1106 for (m = m0; m->m_next != NULL; m = m->m_next) 1107 ; 1108 remainder = len; 1109 space = M_TRAILINGSPACE(m); 1110 if (space > 0) { 1111 /* 1112 * Copy into available space. 1113 */ 1114 if (space > remainder) 1115 space = remainder; 1116 bcopy(cp, mtod(m, caddr_t) + m->m_len, space); 1117 m->m_len += space; 1118 cp += space, remainder -= space; 1119 } 1120 while (remainder > 0) { 1121 /* 1122 * Allocate a new mbuf; could check space 1123 * and allocate a cluster instead. 1124 */ 1125 n = m_get(M_NOWAIT, m->m_type); 1126 if (n == NULL) 1127 break; 1128 n->m_len = min(MLEN, remainder); 1129 bcopy(cp, mtod(n, caddr_t), n->m_len); 1130 cp += n->m_len, remainder -= n->m_len; 1131 m->m_next = n; 1132 m = n; 1133 } 1134 if (m0->m_flags & M_PKTHDR) 1135 m0->m_pkthdr.len += len - remainder; 1136 return (remainder == 0); 1137 } 1138 1139 /* 1140 * Apply function f to the data in an mbuf chain starting "off" bytes from 1141 * the beginning, continuing for "len" bytes. 1142 */ 1143 int 1144 m_apply(struct mbuf *m, int off, int len, 1145 int (*f)(void *, void *, u_int), void *arg) 1146 { 1147 u_int count; 1148 int rval; 1149 1150 KASSERT(off >= 0, ("m_apply, negative off %d", off)); 1151 KASSERT(len >= 0, ("m_apply, negative len %d", len)); 1152 while (off > 0) { 1153 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain")); 1154 if (off < m->m_len) 1155 break; 1156 off -= m->m_len; 1157 m = m->m_next; 1158 } 1159 while (len > 0) { 1160 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain")); 1161 count = min(m->m_len - off, len); 1162 rval = (*f)(arg, mtod(m, caddr_t) + off, count); 1163 if (rval) 1164 return (rval); 1165 len -= count; 1166 off = 0; 1167 m = m->m_next; 1168 } 1169 return (0); 1170 } 1171 1172 /* 1173 * Return a pointer to mbuf/offset of location in mbuf chain. 1174 */ 1175 struct mbuf * 1176 m_getptr(struct mbuf *m, int loc, int *off) 1177 { 1178 1179 while (loc >= 0) { 1180 /* Normal end of search. */ 1181 if (m->m_len > loc) { 1182 *off = loc; 1183 return (m); 1184 } else { 1185 loc -= m->m_len; 1186 if (m->m_next == NULL) { 1187 if (loc == 0) { 1188 /* Point at the end of valid data. */ 1189 *off = m->m_len; 1190 return (m); 1191 } 1192 return (NULL); 1193 } 1194 m = m->m_next; 1195 } 1196 } 1197 return (NULL); 1198 } 1199 1200 void 1201 m_print(const struct mbuf *m, int maxlen) 1202 { 1203 int len; 1204 int pdata; 1205 const struct mbuf *m2; 1206 1207 if (m == NULL) { 1208 printf("mbuf: %p\n", m); 1209 return; 1210 } 1211 1212 if (m->m_flags & M_PKTHDR) 1213 len = m->m_pkthdr.len; 1214 else 1215 len = -1; 1216 m2 = m; 1217 while (m2 != NULL && (len == -1 || len)) { 1218 pdata = m2->m_len; 1219 if (maxlen != -1 && pdata > maxlen) 1220 pdata = maxlen; 1221 printf("mbuf: %p len: %d, next: %p, %b%s", m2, m2->m_len, 1222 m2->m_next, m2->m_flags, "\20\20freelist\17skipfw" 1223 "\11proto5\10proto4\7proto3\6proto2\5proto1\4rdonly" 1224 "\3eor\2pkthdr\1ext", pdata ? "" : "\n"); 1225 if (pdata) 1226 printf(", %*D\n", pdata, (u_char *)m2->m_data, "-"); 1227 if (len != -1) 1228 len -= m2->m_len; 1229 m2 = m2->m_next; 1230 } 1231 if (len > 0) 1232 printf("%d bytes unaccounted for.\n", len); 1233 return; 1234 } 1235 1236 u_int 1237 m_fixhdr(struct mbuf *m0) 1238 { 1239 u_int len; 1240 1241 len = m_length(m0, NULL); 1242 m0->m_pkthdr.len = len; 1243 return (len); 1244 } 1245 1246 u_int 1247 m_length(struct mbuf *m0, struct mbuf **last) 1248 { 1249 struct mbuf *m; 1250 u_int len; 1251 1252 len = 0; 1253 for (m = m0; m != NULL; m = m->m_next) { 1254 len += m->m_len; 1255 if (m->m_next == NULL) 1256 break; 1257 } 1258 if (last != NULL) 1259 *last = m; 1260 return (len); 1261 } 1262 1263 /* 1264 * Defragment a mbuf chain, returning the shortest possible 1265 * chain of mbufs and clusters. If allocation fails and 1266 * this cannot be completed, NULL will be returned, but 1267 * the passed in chain will be unchanged. Upon success, 1268 * the original chain will be freed, and the new chain 1269 * will be returned. 1270 * 1271 * If a non-packet header is passed in, the original 1272 * mbuf (chain?) will be returned unharmed. 1273 */ 1274 struct mbuf * 1275 m_defrag(struct mbuf *m0, int how) 1276 { 1277 struct mbuf *m_new = NULL, *m_final = NULL; 1278 int progress = 0, length; 1279 1280 MBUF_CHECKSLEEP(how); 1281 if (!(m0->m_flags & M_PKTHDR)) 1282 return (m0); 1283 1284 m_fixhdr(m0); /* Needed sanity check */ 1285 1286 #ifdef MBUF_STRESS_TEST 1287 if (m_defragrandomfailures) { 1288 int temp = arc4random() & 0xff; 1289 if (temp == 0xba) 1290 goto nospace; 1291 } 1292 #endif 1293 1294 if (m0->m_pkthdr.len > MHLEN) 1295 m_final = m_getcl(how, MT_DATA, M_PKTHDR); 1296 else 1297 m_final = m_gethdr(how, MT_DATA); 1298 1299 if (m_final == NULL) 1300 goto nospace; 1301 1302 if (m_dup_pkthdr(m_final, m0, how) == 0) 1303 goto nospace; 1304 1305 m_new = m_final; 1306 1307 while (progress < m0->m_pkthdr.len) { 1308 length = m0->m_pkthdr.len - progress; 1309 if (length > MCLBYTES) 1310 length = MCLBYTES; 1311 1312 if (m_new == NULL) { 1313 if (length > MLEN) 1314 m_new = m_getcl(how, MT_DATA, 0); 1315 else 1316 m_new = m_get(how, MT_DATA); 1317 if (m_new == NULL) 1318 goto nospace; 1319 } 1320 1321 m_copydata(m0, progress, length, mtod(m_new, caddr_t)); 1322 progress += length; 1323 m_new->m_len = length; 1324 if (m_new != m_final) 1325 m_cat(m_final, m_new); 1326 m_new = NULL; 1327 } 1328 #ifdef MBUF_STRESS_TEST 1329 if (m0->m_next == NULL) 1330 m_defraguseless++; 1331 #endif 1332 m_freem(m0); 1333 m0 = m_final; 1334 #ifdef MBUF_STRESS_TEST 1335 m_defragpackets++; 1336 m_defragbytes += m0->m_pkthdr.len; 1337 #endif 1338 return (m0); 1339 nospace: 1340 #ifdef MBUF_STRESS_TEST 1341 m_defragfailure++; 1342 #endif 1343 if (m_final) 1344 m_freem(m_final); 1345 return (NULL); 1346 } 1347 1348 /* 1349 * Defragment an mbuf chain, returning at most maxfrags separate 1350 * mbufs+clusters. If this is not possible NULL is returned and 1351 * the original mbuf chain is left in its present (potentially 1352 * modified) state. We use two techniques: collapsing consecutive 1353 * mbufs and replacing consecutive mbufs by a cluster. 1354 * 1355 * NB: this should really be named m_defrag but that name is taken 1356 */ 1357 struct mbuf * 1358 m_collapse(struct mbuf *m0, int how, int maxfrags) 1359 { 1360 struct mbuf *m, *n, *n2, **prev; 1361 u_int curfrags; 1362 1363 /* 1364 * Calculate the current number of frags. 1365 */ 1366 curfrags = 0; 1367 for (m = m0; m != NULL; m = m->m_next) 1368 curfrags++; 1369 /* 1370 * First, try to collapse mbufs. Note that we always collapse 1371 * towards the front so we don't need to deal with moving the 1372 * pkthdr. This may be suboptimal if the first mbuf has much 1373 * less data than the following. 1374 */ 1375 m = m0; 1376 again: 1377 for (;;) { 1378 n = m->m_next; 1379 if (n == NULL) 1380 break; 1381 if (M_WRITABLE(m) && 1382 n->m_len < M_TRAILINGSPACE(m)) { 1383 bcopy(mtod(n, void *), mtod(m, char *) + m->m_len, 1384 n->m_len); 1385 m->m_len += n->m_len; 1386 m->m_next = n->m_next; 1387 m_free(n); 1388 if (--curfrags <= maxfrags) 1389 return m0; 1390 } else 1391 m = n; 1392 } 1393 KASSERT(maxfrags > 1, 1394 ("maxfrags %u, but normal collapse failed", maxfrags)); 1395 /* 1396 * Collapse consecutive mbufs to a cluster. 1397 */ 1398 prev = &m0->m_next; /* NB: not the first mbuf */ 1399 while ((n = *prev) != NULL) { 1400 if ((n2 = n->m_next) != NULL && 1401 n->m_len + n2->m_len < MCLBYTES) { 1402 m = m_getcl(how, MT_DATA, 0); 1403 if (m == NULL) 1404 goto bad; 1405 bcopy(mtod(n, void *), mtod(m, void *), n->m_len); 1406 bcopy(mtod(n2, void *), mtod(m, char *) + n->m_len, 1407 n2->m_len); 1408 m->m_len = n->m_len + n2->m_len; 1409 m->m_next = n2->m_next; 1410 *prev = m; 1411 m_free(n); 1412 m_free(n2); 1413 if (--curfrags <= maxfrags) /* +1 cl -2 mbufs */ 1414 return m0; 1415 /* 1416 * Still not there, try the normal collapse 1417 * again before we allocate another cluster. 1418 */ 1419 goto again; 1420 } 1421 prev = &n->m_next; 1422 } 1423 /* 1424 * No place where we can collapse to a cluster; punt. 1425 * This can occur if, for example, you request 2 frags 1426 * but the packet requires that both be clusters (we 1427 * never reallocate the first mbuf to avoid moving the 1428 * packet header). 1429 */ 1430 bad: 1431 return NULL; 1432 } 1433 1434 #ifdef MBUF_STRESS_TEST 1435 1436 /* 1437 * Fragment an mbuf chain. There's no reason you'd ever want to do 1438 * this in normal usage, but it's great for stress testing various 1439 * mbuf consumers. 1440 * 1441 * If fragmentation is not possible, the original chain will be 1442 * returned. 1443 * 1444 * Possible length values: 1445 * 0 no fragmentation will occur 1446 * > 0 each fragment will be of the specified length 1447 * -1 each fragment will be the same random value in length 1448 * -2 each fragment's length will be entirely random 1449 * (Random values range from 1 to 256) 1450 */ 1451 struct mbuf * 1452 m_fragment(struct mbuf *m0, int how, int length) 1453 { 1454 struct mbuf *m_first, *m_last; 1455 int divisor = 255, progress = 0, fraglen; 1456 1457 if (!(m0->m_flags & M_PKTHDR)) 1458 return (m0); 1459 1460 if (length == 0 || length < -2) 1461 return (m0); 1462 if (length > MCLBYTES) 1463 length = MCLBYTES; 1464 if (length < 0 && divisor > MCLBYTES) 1465 divisor = MCLBYTES; 1466 if (length == -1) 1467 length = 1 + (arc4random() % divisor); 1468 if (length > 0) 1469 fraglen = length; 1470 1471 m_fixhdr(m0); /* Needed sanity check */ 1472 1473 m_first = m_getcl(how, MT_DATA, M_PKTHDR); 1474 if (m_first == NULL) 1475 goto nospace; 1476 1477 if (m_dup_pkthdr(m_first, m0, how) == 0) 1478 goto nospace; 1479 1480 m_last = m_first; 1481 1482 while (progress < m0->m_pkthdr.len) { 1483 if (length == -2) 1484 fraglen = 1 + (arc4random() % divisor); 1485 if (fraglen > m0->m_pkthdr.len - progress) 1486 fraglen = m0->m_pkthdr.len - progress; 1487 1488 if (progress != 0) { 1489 struct mbuf *m_new = m_getcl(how, MT_DATA, 0); 1490 if (m_new == NULL) 1491 goto nospace; 1492 1493 m_last->m_next = m_new; 1494 m_last = m_new; 1495 } 1496 1497 m_copydata(m0, progress, fraglen, mtod(m_last, caddr_t)); 1498 progress += fraglen; 1499 m_last->m_len = fraglen; 1500 } 1501 m_freem(m0); 1502 m0 = m_first; 1503 return (m0); 1504 nospace: 1505 if (m_first) 1506 m_freem(m_first); 1507 /* Return the original chain on failure */ 1508 return (m0); 1509 } 1510 1511 #endif 1512 1513 /* 1514 * Copy the contents of uio into a properly sized mbuf chain. 1515 */ 1516 struct mbuf * 1517 m_uiotombuf(struct uio *uio, int how, int len, int align, int flags) 1518 { 1519 struct mbuf *m, *mb; 1520 int error, length; 1521 ssize_t total; 1522 int progress = 0; 1523 1524 /* 1525 * len can be zero or an arbitrary large value bound by 1526 * the total data supplied by the uio. 1527 */ 1528 if (len > 0) 1529 total = (uio->uio_resid < len) ? uio->uio_resid : len; 1530 else 1531 total = uio->uio_resid; 1532 1533 /* 1534 * The smallest unit returned by m_getm2() is a single mbuf 1535 * with pkthdr. We can't align past it. 1536 */ 1537 if (align >= MHLEN) 1538 return (NULL); 1539 1540 /* 1541 * Give us the full allocation or nothing. 1542 * If len is zero return the smallest empty mbuf. 1543 */ 1544 m = m_getm2(NULL, max(total + align, 1), how, MT_DATA, flags); 1545 if (m == NULL) 1546 return (NULL); 1547 m->m_data += align; 1548 1549 /* Fill all mbufs with uio data and update header information. */ 1550 for (mb = m; mb != NULL; mb = mb->m_next) { 1551 length = min(M_TRAILINGSPACE(mb), total - progress); 1552 1553 error = uiomove(mtod(mb, void *), length, uio); 1554 if (error) { 1555 m_freem(m); 1556 return (NULL); 1557 } 1558 1559 mb->m_len = length; 1560 progress += length; 1561 if (flags & M_PKTHDR) 1562 m->m_pkthdr.len += length; 1563 } 1564 KASSERT(progress == total, ("%s: progress != total", __func__)); 1565 1566 return (m); 1567 } 1568 1569 /* 1570 * Copy an mbuf chain into a uio limited by len if set. 1571 */ 1572 int 1573 m_mbuftouio(struct uio *uio, const struct mbuf *m, int len) 1574 { 1575 int error, length, total; 1576 int progress = 0; 1577 1578 if (len > 0) 1579 total = min(uio->uio_resid, len); 1580 else 1581 total = uio->uio_resid; 1582 1583 /* Fill the uio with data from the mbufs. */ 1584 for (; m != NULL; m = m->m_next) { 1585 length = min(m->m_len, total - progress); 1586 1587 error = uiomove(mtod(m, void *), length, uio); 1588 if (error) 1589 return (error); 1590 1591 progress += length; 1592 } 1593 1594 return (0); 1595 } 1596 1597 /* 1598 * Create a writable copy of the mbuf chain. While doing this 1599 * we compact the chain with a goal of producing a chain with 1600 * at most two mbufs. The second mbuf in this chain is likely 1601 * to be a cluster. The primary purpose of this work is to create 1602 * a writable packet for encryption, compression, etc. The 1603 * secondary goal is to linearize the data so the data can be 1604 * passed to crypto hardware in the most efficient manner possible. 1605 */ 1606 struct mbuf * 1607 m_unshare(struct mbuf *m0, int how) 1608 { 1609 struct mbuf *m, *mprev; 1610 struct mbuf *n, *mfirst, *mlast; 1611 int len, off; 1612 1613 mprev = NULL; 1614 for (m = m0; m != NULL; m = mprev->m_next) { 1615 /* 1616 * Regular mbufs are ignored unless there's a cluster 1617 * in front of it that we can use to coalesce. We do 1618 * the latter mainly so later clusters can be coalesced 1619 * also w/o having to handle them specially (i.e. convert 1620 * mbuf+cluster -> cluster). This optimization is heavily 1621 * influenced by the assumption that we're running over 1622 * Ethernet where MCLBYTES is large enough that the max 1623 * packet size will permit lots of coalescing into a 1624 * single cluster. This in turn permits efficient 1625 * crypto operations, especially when using hardware. 1626 */ 1627 if ((m->m_flags & M_EXT) == 0) { 1628 if (mprev && (mprev->m_flags & M_EXT) && 1629 m->m_len <= M_TRAILINGSPACE(mprev)) { 1630 /* XXX: this ignores mbuf types */ 1631 memcpy(mtod(mprev, caddr_t) + mprev->m_len, 1632 mtod(m, caddr_t), m->m_len); 1633 mprev->m_len += m->m_len; 1634 mprev->m_next = m->m_next; /* unlink from chain */ 1635 m_free(m); /* reclaim mbuf */ 1636 #if 0 1637 newipsecstat.ips_mbcoalesced++; 1638 #endif 1639 } else { 1640 mprev = m; 1641 } 1642 continue; 1643 } 1644 /* 1645 * Writable mbufs are left alone (for now). 1646 */ 1647 if (M_WRITABLE(m)) { 1648 mprev = m; 1649 continue; 1650 } 1651 1652 /* 1653 * Not writable, replace with a copy or coalesce with 1654 * the previous mbuf if possible (since we have to copy 1655 * it anyway, we try to reduce the number of mbufs and 1656 * clusters so that future work is easier). 1657 */ 1658 KASSERT(m->m_flags & M_EXT, ("m_flags 0x%x", m->m_flags)); 1659 /* NB: we only coalesce into a cluster or larger */ 1660 if (mprev != NULL && (mprev->m_flags & M_EXT) && 1661 m->m_len <= M_TRAILINGSPACE(mprev)) { 1662 /* XXX: this ignores mbuf types */ 1663 memcpy(mtod(mprev, caddr_t) + mprev->m_len, 1664 mtod(m, caddr_t), m->m_len); 1665 mprev->m_len += m->m_len; 1666 mprev->m_next = m->m_next; /* unlink from chain */ 1667 m_free(m); /* reclaim mbuf */ 1668 #if 0 1669 newipsecstat.ips_clcoalesced++; 1670 #endif 1671 continue; 1672 } 1673 1674 /* 1675 * Allocate new space to hold the copy and copy the data. 1676 * We deal with jumbo mbufs (i.e. m_len > MCLBYTES) by 1677 * splitting them into clusters. We could just malloc a 1678 * buffer and make it external but too many device drivers 1679 * don't know how to break up the non-contiguous memory when 1680 * doing DMA. 1681 */ 1682 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS); 1683 if (n == NULL) { 1684 m_freem(m0); 1685 return (NULL); 1686 } 1687 if (m->m_flags & M_PKTHDR) { 1688 KASSERT(mprev == NULL, ("%s: m0 %p, m %p has M_PKTHDR", 1689 __func__, m0, m)); 1690 m_move_pkthdr(n, m); 1691 } 1692 len = m->m_len; 1693 off = 0; 1694 mfirst = n; 1695 mlast = NULL; 1696 for (;;) { 1697 int cc = min(len, MCLBYTES); 1698 memcpy(mtod(n, caddr_t), mtod(m, caddr_t) + off, cc); 1699 n->m_len = cc; 1700 if (mlast != NULL) 1701 mlast->m_next = n; 1702 mlast = n; 1703 #if 0 1704 newipsecstat.ips_clcopied++; 1705 #endif 1706 1707 len -= cc; 1708 if (len <= 0) 1709 break; 1710 off += cc; 1711 1712 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS); 1713 if (n == NULL) { 1714 m_freem(mfirst); 1715 m_freem(m0); 1716 return (NULL); 1717 } 1718 } 1719 n->m_next = m->m_next; 1720 if (mprev == NULL) 1721 m0 = mfirst; /* new head of chain */ 1722 else 1723 mprev->m_next = mfirst; /* replace old mbuf */ 1724 m_free(m); /* release old mbuf */ 1725 mprev = mfirst; 1726 } 1727 return (m0); 1728 } 1729 1730 #ifdef MBUF_PROFILING 1731 1732 #define MP_BUCKETS 32 /* don't just change this as things may overflow.*/ 1733 struct mbufprofile { 1734 uintmax_t wasted[MP_BUCKETS]; 1735 uintmax_t used[MP_BUCKETS]; 1736 uintmax_t segments[MP_BUCKETS]; 1737 } mbprof; 1738 1739 #define MP_MAXDIGITS 21 /* strlen("16,000,000,000,000,000,000") == 21 */ 1740 #define MP_NUMLINES 6 1741 #define MP_NUMSPERLINE 16 1742 #define MP_EXTRABYTES 64 /* > strlen("used:\nwasted:\nsegments:\n") */ 1743 /* work out max space needed and add a bit of spare space too */ 1744 #define MP_MAXLINE ((MP_MAXDIGITS+1) * MP_NUMSPERLINE) 1745 #define MP_BUFSIZE ((MP_MAXLINE * MP_NUMLINES) + 1 + MP_EXTRABYTES) 1746 1747 char mbprofbuf[MP_BUFSIZE]; 1748 1749 void 1750 m_profile(struct mbuf *m) 1751 { 1752 int segments = 0; 1753 int used = 0; 1754 int wasted = 0; 1755 1756 while (m) { 1757 segments++; 1758 used += m->m_len; 1759 if (m->m_flags & M_EXT) { 1760 wasted += MHLEN - sizeof(m->m_ext) + 1761 m->m_ext.ext_size - m->m_len; 1762 } else { 1763 if (m->m_flags & M_PKTHDR) 1764 wasted += MHLEN - m->m_len; 1765 else 1766 wasted += MLEN - m->m_len; 1767 } 1768 m = m->m_next; 1769 } 1770 /* be paranoid.. it helps */ 1771 if (segments > MP_BUCKETS - 1) 1772 segments = MP_BUCKETS - 1; 1773 if (used > 100000) 1774 used = 100000; 1775 if (wasted > 100000) 1776 wasted = 100000; 1777 /* store in the appropriate bucket */ 1778 /* don't bother locking. if it's slightly off, so what? */ 1779 mbprof.segments[segments]++; 1780 mbprof.used[fls(used)]++; 1781 mbprof.wasted[fls(wasted)]++; 1782 } 1783 1784 static void 1785 mbprof_textify(void) 1786 { 1787 int offset; 1788 char *c; 1789 uint64_t *p; 1790 1791 p = &mbprof.wasted[0]; 1792 c = mbprofbuf; 1793 offset = snprintf(c, MP_MAXLINE + 10, 1794 "wasted:\n" 1795 "%ju %ju %ju %ju %ju %ju %ju %ju " 1796 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 1797 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 1798 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 1799 #ifdef BIG_ARRAY 1800 p = &mbprof.wasted[16]; 1801 c += offset; 1802 offset = snprintf(c, MP_MAXLINE, 1803 "%ju %ju %ju %ju %ju %ju %ju %ju " 1804 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 1805 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 1806 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 1807 #endif 1808 p = &mbprof.used[0]; 1809 c += offset; 1810 offset = snprintf(c, MP_MAXLINE + 10, 1811 "used:\n" 1812 "%ju %ju %ju %ju %ju %ju %ju %ju " 1813 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 1814 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 1815 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 1816 #ifdef BIG_ARRAY 1817 p = &mbprof.used[16]; 1818 c += offset; 1819 offset = snprintf(c, MP_MAXLINE, 1820 "%ju %ju %ju %ju %ju %ju %ju %ju " 1821 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 1822 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 1823 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 1824 #endif 1825 p = &mbprof.segments[0]; 1826 c += offset; 1827 offset = snprintf(c, MP_MAXLINE + 10, 1828 "segments:\n" 1829 "%ju %ju %ju %ju %ju %ju %ju %ju " 1830 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 1831 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 1832 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 1833 #ifdef BIG_ARRAY 1834 p = &mbprof.segments[16]; 1835 c += offset; 1836 offset = snprintf(c, MP_MAXLINE, 1837 "%ju %ju %ju %ju %ju %ju %ju %ju " 1838 "%ju %ju %ju %ju %ju %ju %ju %jju", 1839 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 1840 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 1841 #endif 1842 } 1843 1844 static int 1845 mbprof_handler(SYSCTL_HANDLER_ARGS) 1846 { 1847 int error; 1848 1849 mbprof_textify(); 1850 error = SYSCTL_OUT(req, mbprofbuf, strlen(mbprofbuf) + 1); 1851 return (error); 1852 } 1853 1854 static int 1855 mbprof_clr_handler(SYSCTL_HANDLER_ARGS) 1856 { 1857 int clear, error; 1858 1859 clear = 0; 1860 error = sysctl_handle_int(oidp, &clear, 0, req); 1861 if (error || !req->newptr) 1862 return (error); 1863 1864 if (clear) { 1865 bzero(&mbprof, sizeof(mbprof)); 1866 } 1867 1868 return (error); 1869 } 1870 1871 1872 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofile, CTLTYPE_STRING|CTLFLAG_RD, 1873 NULL, 0, mbprof_handler, "A", "mbuf profiling statistics"); 1874 1875 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofileclr, CTLTYPE_INT|CTLFLAG_RW, 1876 NULL, 0, mbprof_clr_handler, "I", "clear mbuf profiling statistics"); 1877 #endif 1878 1879