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/vmmeter.h> 53 #include <sys/sdt.h> 54 #include <vm/vm.h> 55 #include <vm/vm_pageout.h> 56 #include <vm/vm_page.h> 57 58 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__init, 59 "struct mbuf *", "mbufinfo_t *", 60 "uint32_t", "uint32_t", 61 "uint16_t", "uint16_t", 62 "uint32_t", "uint32_t", 63 "uint32_t", "uint32_t"); 64 65 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__gethdr_raw, 66 "uint32_t", "uint32_t", 67 "uint16_t", "uint16_t", 68 "struct mbuf *", "mbufinfo_t *"); 69 70 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__gethdr, 71 "uint32_t", "uint32_t", 72 "uint16_t", "uint16_t", 73 "struct mbuf *", "mbufinfo_t *"); 74 75 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__get_raw, 76 "uint32_t", "uint32_t", 77 "uint16_t", "uint16_t", 78 "struct mbuf *", "mbufinfo_t *"); 79 80 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__get, 81 "uint32_t", "uint32_t", 82 "uint16_t", "uint16_t", 83 "struct mbuf *", "mbufinfo_t *"); 84 85 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__getcl, 86 "uint32_t", "uint32_t", 87 "uint16_t", "uint16_t", 88 "uint32_t", "uint32_t", 89 "struct mbuf *", "mbufinfo_t *"); 90 91 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__getjcl, 92 "uint32_t", "uint32_t", 93 "uint16_t", "uint16_t", 94 "uint32_t", "uint32_t", 95 "uint32_t", "uint32_t", 96 "struct mbuf *", "mbufinfo_t *"); 97 98 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__clget, 99 "struct mbuf *", "mbufinfo_t *", 100 "uint32_t", "uint32_t", 101 "uint32_t", "uint32_t"); 102 103 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__cljget, 104 "struct mbuf *", "mbufinfo_t *", 105 "uint32_t", "uint32_t", 106 "uint32_t", "uint32_t", 107 "void*", "void*"); 108 109 SDT_PROBE_DEFINE(sdt, , , m__cljset); 110 111 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__free, 112 "struct mbuf *", "mbufinfo_t *"); 113 114 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__freem, 115 "struct mbuf *", "mbufinfo_t *"); 116 117 #include <security/mac/mac_framework.h> 118 119 int max_linkhdr; 120 int max_protohdr; 121 int max_hdr; 122 int max_datalen; 123 #ifdef MBUF_STRESS_TEST 124 int m_defragpackets; 125 int m_defragbytes; 126 int m_defraguseless; 127 int m_defragfailure; 128 int m_defragrandomfailures; 129 #endif 130 131 /* 132 * sysctl(8) exported objects 133 */ 134 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RD, 135 &max_linkhdr, 0, "Size of largest link layer header"); 136 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RD, 137 &max_protohdr, 0, "Size of largest protocol layer header"); 138 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RD, 139 &max_hdr, 0, "Size of largest link plus protocol header"); 140 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RD, 141 &max_datalen, 0, "Minimum space left in mbuf after max_hdr"); 142 #ifdef MBUF_STRESS_TEST 143 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD, 144 &m_defragpackets, 0, ""); 145 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD, 146 &m_defragbytes, 0, ""); 147 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD, 148 &m_defraguseless, 0, ""); 149 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD, 150 &m_defragfailure, 0, ""); 151 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW, 152 &m_defragrandomfailures, 0, ""); 153 #endif 154 155 /* 156 * Ensure the correct size of various mbuf parameters. It could be off due 157 * to compiler-induced padding and alignment artifacts. 158 */ 159 CTASSERT(MSIZE - offsetof(struct mbuf, m_dat) == MLEN); 160 CTASSERT(MSIZE - offsetof(struct mbuf, m_pktdat) == MHLEN); 161 162 /* 163 * mbuf data storage should be 64-bit aligned regardless of architectural 164 * pointer size; check this is the case with and without a packet header. 165 */ 166 CTASSERT(offsetof(struct mbuf, m_dat) % 8 == 0); 167 CTASSERT(offsetof(struct mbuf, m_pktdat) % 8 == 0); 168 169 /* 170 * While the specific values here don't matter too much (i.e., +/- a few 171 * words), we do want to ensure that changes to these values are carefully 172 * reasoned about and properly documented. This is especially the case as 173 * network-protocol and device-driver modules encode these layouts, and must 174 * be recompiled if the structures change. Check these values at compile time 175 * against the ones documented in comments in mbuf.h. 176 * 177 * NB: Possibly they should be documented there via #define's and not just 178 * comments. 179 */ 180 #if defined(__LP64__) 181 CTASSERT(offsetof(struct mbuf, m_dat) == 32); 182 CTASSERT(sizeof(struct pkthdr) == 56); 183 CTASSERT(sizeof(struct m_ext) == 160); 184 #else 185 CTASSERT(offsetof(struct mbuf, m_dat) == 24); 186 CTASSERT(sizeof(struct pkthdr) == 48); 187 #if defined(__powerpc__) && defined(BOOKE) 188 /* PowerPC booke has 64-bit physical pointers. */ 189 CTASSERT(sizeof(struct m_ext) == 184); 190 #else 191 CTASSERT(sizeof(struct m_ext) == 180); 192 #endif 193 #endif 194 195 /* 196 * Assert that the queue(3) macros produce code of the same size as an old 197 * plain pointer does. 198 */ 199 #ifdef INVARIANTS 200 static struct mbuf __used m_assertbuf; 201 CTASSERT(sizeof(m_assertbuf.m_slist) == sizeof(m_assertbuf.m_next)); 202 CTASSERT(sizeof(m_assertbuf.m_stailq) == sizeof(m_assertbuf.m_next)); 203 CTASSERT(sizeof(m_assertbuf.m_slistpkt) == sizeof(m_assertbuf.m_nextpkt)); 204 CTASSERT(sizeof(m_assertbuf.m_stailqpkt) == sizeof(m_assertbuf.m_nextpkt)); 205 #endif 206 207 /* 208 * Attach the cluster from *m to *n, set up m_ext in *n 209 * and bump the refcount of the cluster. 210 */ 211 void 212 mb_dupcl(struct mbuf *n, struct mbuf *m) 213 { 214 volatile u_int *refcnt; 215 216 KASSERT(m->m_flags & (M_EXT|M_EXTPG), 217 ("%s: M_EXT|M_EXTPG not set on %p", __func__, m)); 218 KASSERT(!(n->m_flags & (M_EXT|M_EXTPG)), 219 ("%s: M_EXT|M_EXTPG set on %p", __func__, n)); 220 221 /* 222 * Cache access optimization. 223 * 224 * o Regular M_EXT storage doesn't need full copy of m_ext, since 225 * the holder of the 'ext_count' is responsible to carry the free 226 * routine and its arguments. 227 * o M_EXTPG data is split between main part of mbuf and m_ext, the 228 * main part is copied in full, the m_ext part is similar to M_EXT. 229 * o EXT_EXTREF, where 'ext_cnt' doesn't point into mbuf at all, is 230 * special - it needs full copy of m_ext into each mbuf, since any 231 * copy could end up as the last to free. 232 */ 233 if (m->m_flags & M_EXTPG) { 234 bcopy(&m->m_epg_startcopy, &n->m_epg_startcopy, 235 __rangeof(struct mbuf, m_epg_startcopy, m_epg_endcopy)); 236 bcopy(&m->m_ext, &n->m_ext, m_epg_ext_copylen); 237 } else if (m->m_ext.ext_type == EXT_EXTREF) 238 bcopy(&m->m_ext, &n->m_ext, sizeof(struct m_ext)); 239 else 240 bcopy(&m->m_ext, &n->m_ext, m_ext_copylen); 241 242 n->m_flags |= m->m_flags & (M_RDONLY | M_EXT | M_EXTPG); 243 244 /* See if this is the mbuf that holds the embedded refcount. */ 245 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) { 246 refcnt = n->m_ext.ext_cnt = &m->m_ext.ext_count; 247 n->m_ext.ext_flags &= ~EXT_FLAG_EMBREF; 248 } else { 249 KASSERT(m->m_ext.ext_cnt != NULL, 250 ("%s: no refcounting pointer on %p", __func__, m)); 251 refcnt = m->m_ext.ext_cnt; 252 } 253 254 if (*refcnt == 1) 255 *refcnt += 1; 256 else 257 atomic_add_int(refcnt, 1); 258 } 259 260 void 261 m_demote_pkthdr(struct mbuf *m) 262 { 263 264 M_ASSERTPKTHDR(m); 265 266 m_tag_delete_chain(m, NULL); 267 m->m_flags &= ~M_PKTHDR; 268 bzero(&m->m_pkthdr, sizeof(struct pkthdr)); 269 } 270 271 /* 272 * Clean up mbuf (chain) from any tags and packet headers. 273 * If "all" is set then the first mbuf in the chain will be 274 * cleaned too. 275 */ 276 void 277 m_demote(struct mbuf *m0, int all, int flags) 278 { 279 struct mbuf *m; 280 281 flags |= M_DEMOTEFLAGS; 282 283 for (m = all ? m0 : m0->m_next; m != NULL; m = m->m_next) { 284 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt in m %p, m0 %p", 285 __func__, m, m0)); 286 if (m->m_flags & M_PKTHDR) 287 m_demote_pkthdr(m); 288 m->m_flags &= flags; 289 } 290 } 291 292 /* 293 * Sanity checks on mbuf (chain) for use in KASSERT() and general 294 * debugging. 295 * Returns 0 or panics when bad and 1 on all tests passed. 296 * Sanitize, 0 to run M_SANITY_ACTION, 1 to garble things so they 297 * blow up later. 298 */ 299 int 300 m_sanity(struct mbuf *m0, int sanitize) 301 { 302 struct mbuf *m; 303 caddr_t a, b; 304 int pktlen = 0; 305 306 #ifdef INVARIANTS 307 #define M_SANITY_ACTION(s) panic("mbuf %p: " s, m) 308 #else 309 #define M_SANITY_ACTION(s) printf("mbuf %p: " s, m) 310 #endif 311 312 for (m = m0; m != NULL; m = m->m_next) { 313 /* 314 * Basic pointer checks. If any of these fails then some 315 * unrelated kernel memory before or after us is trashed. 316 * No way to recover from that. 317 */ 318 a = M_START(m); 319 b = a + M_SIZE(m); 320 if ((caddr_t)m->m_data < a) 321 M_SANITY_ACTION("m_data outside mbuf data range left"); 322 if ((caddr_t)m->m_data > b) 323 M_SANITY_ACTION("m_data outside mbuf data range right"); 324 if ((caddr_t)m->m_data + m->m_len > b) 325 M_SANITY_ACTION("m_data + m_len exeeds mbuf space"); 326 327 /* m->m_nextpkt may only be set on first mbuf in chain. */ 328 if (m != m0 && m->m_nextpkt != NULL) { 329 if (sanitize) { 330 m_freem(m->m_nextpkt); 331 m->m_nextpkt = (struct mbuf *)0xDEADC0DE; 332 } else 333 M_SANITY_ACTION("m->m_nextpkt on in-chain mbuf"); 334 } 335 336 /* packet length (not mbuf length!) calculation */ 337 if (m0->m_flags & M_PKTHDR) 338 pktlen += m->m_len; 339 340 /* m_tags may only be attached to first mbuf in chain. */ 341 if (m != m0 && m->m_flags & M_PKTHDR && 342 !SLIST_EMPTY(&m->m_pkthdr.tags)) { 343 if (sanitize) { 344 m_tag_delete_chain(m, NULL); 345 /* put in 0xDEADC0DE perhaps? */ 346 } else 347 M_SANITY_ACTION("m_tags on in-chain mbuf"); 348 } 349 350 /* M_PKTHDR may only be set on first mbuf in chain */ 351 if (m != m0 && m->m_flags & M_PKTHDR) { 352 if (sanitize) { 353 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr)); 354 m->m_flags &= ~M_PKTHDR; 355 /* put in 0xDEADCODE and leave hdr flag in */ 356 } else 357 M_SANITY_ACTION("M_PKTHDR on in-chain mbuf"); 358 } 359 } 360 m = m0; 361 if (pktlen && pktlen != m->m_pkthdr.len) { 362 if (sanitize) 363 m->m_pkthdr.len = 0; 364 else 365 M_SANITY_ACTION("m_pkthdr.len != mbuf chain length"); 366 } 367 return 1; 368 369 #undef M_SANITY_ACTION 370 } 371 372 /* 373 * Non-inlined part of m_init(). 374 */ 375 int 376 m_pkthdr_init(struct mbuf *m, int how) 377 { 378 #ifdef MAC 379 int error; 380 #endif 381 m->m_data = m->m_pktdat; 382 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr)); 383 #ifdef NUMA 384 m->m_pkthdr.numa_domain = M_NODOM; 385 #endif 386 #ifdef MAC 387 /* If the label init fails, fail the alloc */ 388 error = mac_mbuf_init(m, how); 389 if (error) 390 return (error); 391 #endif 392 393 return (0); 394 } 395 396 /* 397 * "Move" mbuf pkthdr from "from" to "to". 398 * "from" must have M_PKTHDR set, and "to" must be empty. 399 */ 400 void 401 m_move_pkthdr(struct mbuf *to, struct mbuf *from) 402 { 403 404 #if 0 405 /* see below for why these are not enabled */ 406 M_ASSERTPKTHDR(to); 407 /* Note: with MAC, this may not be a good assertion. */ 408 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags), 409 ("m_move_pkthdr: to has tags")); 410 #endif 411 #ifdef MAC 412 /* 413 * XXXMAC: It could be this should also occur for non-MAC? 414 */ 415 if (to->m_flags & M_PKTHDR) 416 m_tag_delete_chain(to, NULL); 417 #endif 418 to->m_flags = (from->m_flags & M_COPYFLAGS) | 419 (to->m_flags & (M_EXT | M_EXTPG)); 420 if ((to->m_flags & M_EXT) == 0) 421 to->m_data = to->m_pktdat; 422 to->m_pkthdr = from->m_pkthdr; /* especially tags */ 423 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */ 424 from->m_flags &= ~M_PKTHDR; 425 if (from->m_pkthdr.csum_flags & CSUM_SND_TAG) { 426 from->m_pkthdr.csum_flags &= ~CSUM_SND_TAG; 427 from->m_pkthdr.snd_tag = NULL; 428 } 429 } 430 431 /* 432 * Duplicate "from"'s mbuf pkthdr in "to". 433 * "from" must have M_PKTHDR set, and "to" must be empty. 434 * In particular, this does a deep copy of the packet tags. 435 */ 436 int 437 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how) 438 { 439 440 #if 0 441 /* 442 * The mbuf allocator only initializes the pkthdr 443 * when the mbuf is allocated with m_gethdr(). Many users 444 * (e.g. m_copy*, m_prepend) use m_get() and then 445 * smash the pkthdr as needed causing these 446 * assertions to trip. For now just disable them. 447 */ 448 M_ASSERTPKTHDR(to); 449 /* Note: with MAC, this may not be a good assertion. */ 450 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags), ("m_dup_pkthdr: to has tags")); 451 #endif 452 MBUF_CHECKSLEEP(how); 453 #ifdef MAC 454 if (to->m_flags & M_PKTHDR) 455 m_tag_delete_chain(to, NULL); 456 #endif 457 to->m_flags = (from->m_flags & M_COPYFLAGS) | 458 (to->m_flags & (M_EXT | M_EXTPG)); 459 if ((to->m_flags & M_EXT) == 0) 460 to->m_data = to->m_pktdat; 461 to->m_pkthdr = from->m_pkthdr; 462 if (from->m_pkthdr.csum_flags & CSUM_SND_TAG) 463 m_snd_tag_ref(from->m_pkthdr.snd_tag); 464 SLIST_INIT(&to->m_pkthdr.tags); 465 return (m_tag_copy_chain(to, from, how)); 466 } 467 468 /* 469 * Lesser-used path for M_PREPEND: 470 * allocate new mbuf to prepend to chain, 471 * copy junk along. 472 */ 473 struct mbuf * 474 m_prepend(struct mbuf *m, int len, int how) 475 { 476 struct mbuf *mn; 477 478 if (m->m_flags & M_PKTHDR) 479 mn = m_gethdr(how, m->m_type); 480 else 481 mn = m_get(how, m->m_type); 482 if (mn == NULL) { 483 m_freem(m); 484 return (NULL); 485 } 486 if (m->m_flags & M_PKTHDR) 487 m_move_pkthdr(mn, m); 488 mn->m_next = m; 489 m = mn; 490 if (len < M_SIZE(m)) 491 M_ALIGN(m, len); 492 m->m_len = len; 493 return (m); 494 } 495 496 /* 497 * Make a copy of an mbuf chain starting "off0" bytes from the beginning, 498 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf. 499 * The wait parameter is a choice of M_WAITOK/M_NOWAIT from caller. 500 * Note that the copy is read-only, because clusters are not copied, 501 * only their reference counts are incremented. 502 */ 503 struct mbuf * 504 m_copym(struct mbuf *m, int off0, int len, int wait) 505 { 506 struct mbuf *n, **np; 507 int off = off0; 508 struct mbuf *top; 509 int copyhdr = 0; 510 511 KASSERT(off >= 0, ("m_copym, negative off %d", off)); 512 KASSERT(len >= 0, ("m_copym, negative len %d", len)); 513 MBUF_CHECKSLEEP(wait); 514 if (off == 0 && m->m_flags & M_PKTHDR) 515 copyhdr = 1; 516 while (off > 0) { 517 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain")); 518 if (off < m->m_len) 519 break; 520 off -= m->m_len; 521 m = m->m_next; 522 } 523 np = ⊤ 524 top = NULL; 525 while (len > 0) { 526 if (m == NULL) { 527 KASSERT(len == M_COPYALL, 528 ("m_copym, length > size of mbuf chain")); 529 break; 530 } 531 if (copyhdr) 532 n = m_gethdr(wait, m->m_type); 533 else 534 n = m_get(wait, m->m_type); 535 *np = n; 536 if (n == NULL) 537 goto nospace; 538 if (copyhdr) { 539 if (!m_dup_pkthdr(n, m, wait)) 540 goto nospace; 541 if (len == M_COPYALL) 542 n->m_pkthdr.len -= off0; 543 else 544 n->m_pkthdr.len = len; 545 copyhdr = 0; 546 } 547 n->m_len = min(len, m->m_len - off); 548 if (m->m_flags & (M_EXT|M_EXTPG)) { 549 n->m_data = m->m_data + off; 550 mb_dupcl(n, m); 551 } else 552 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), 553 (u_int)n->m_len); 554 if (len != M_COPYALL) 555 len -= n->m_len; 556 off = 0; 557 m = m->m_next; 558 np = &n->m_next; 559 } 560 561 return (top); 562 nospace: 563 m_freem(top); 564 return (NULL); 565 } 566 567 /* 568 * Copy an entire packet, including header (which must be present). 569 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'. 570 * Note that the copy is read-only, because clusters are not copied, 571 * only their reference counts are incremented. 572 * Preserve alignment of the first mbuf so if the creator has left 573 * some room at the beginning (e.g. for inserting protocol headers) 574 * the copies still have the room available. 575 */ 576 struct mbuf * 577 m_copypacket(struct mbuf *m, int how) 578 { 579 struct mbuf *top, *n, *o; 580 581 MBUF_CHECKSLEEP(how); 582 n = m_get(how, m->m_type); 583 top = n; 584 if (n == NULL) 585 goto nospace; 586 587 if (!m_dup_pkthdr(n, m, how)) 588 goto nospace; 589 n->m_len = m->m_len; 590 if (m->m_flags & (M_EXT|M_EXTPG)) { 591 n->m_data = m->m_data; 592 mb_dupcl(n, m); 593 } else { 594 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat ); 595 bcopy(mtod(m, char *), mtod(n, char *), n->m_len); 596 } 597 598 m = m->m_next; 599 while (m) { 600 o = m_get(how, m->m_type); 601 if (o == NULL) 602 goto nospace; 603 604 n->m_next = o; 605 n = n->m_next; 606 607 n->m_len = m->m_len; 608 if (m->m_flags & (M_EXT|M_EXTPG)) { 609 n->m_data = m->m_data; 610 mb_dupcl(n, m); 611 } else { 612 bcopy(mtod(m, char *), mtod(n, char *), n->m_len); 613 } 614 615 m = m->m_next; 616 } 617 return top; 618 nospace: 619 m_freem(top); 620 return (NULL); 621 } 622 623 static void 624 m_copyfromunmapped(const struct mbuf *m, int off, int len, caddr_t cp) 625 { 626 struct iovec iov; 627 struct uio uio; 628 int error; 629 630 KASSERT(off >= 0, ("m_copyfromunmapped: negative off %d", off)); 631 KASSERT(len >= 0, ("m_copyfromunmapped: negative len %d", len)); 632 KASSERT(off < m->m_len, 633 ("m_copyfromunmapped: len exceeds mbuf length")); 634 iov.iov_base = cp; 635 iov.iov_len = len; 636 uio.uio_resid = len; 637 uio.uio_iov = &iov; 638 uio.uio_segflg = UIO_SYSSPACE; 639 uio.uio_iovcnt = 1; 640 uio.uio_offset = 0; 641 uio.uio_rw = UIO_READ; 642 error = m_unmapped_uiomove(m, off, &uio, len); 643 KASSERT(error == 0, ("m_unmapped_uiomove failed: off %d, len %d", off, 644 len)); 645 } 646 647 /* 648 * Copy data from an mbuf chain starting "off" bytes from the beginning, 649 * continuing for "len" bytes, into the indicated buffer. 650 */ 651 void 652 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp) 653 { 654 u_int count; 655 656 KASSERT(off >= 0, ("m_copydata, negative off %d", off)); 657 KASSERT(len >= 0, ("m_copydata, negative len %d", len)); 658 while (off > 0) { 659 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain")); 660 if (off < m->m_len) 661 break; 662 off -= m->m_len; 663 m = m->m_next; 664 } 665 while (len > 0) { 666 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain")); 667 count = min(m->m_len - off, len); 668 if ((m->m_flags & M_EXTPG) != 0) 669 m_copyfromunmapped(m, off, count, cp); 670 else 671 bcopy(mtod(m, caddr_t) + off, cp, count); 672 len -= count; 673 cp += count; 674 off = 0; 675 m = m->m_next; 676 } 677 } 678 679 /* 680 * Copy a packet header mbuf chain into a completely new chain, including 681 * copying any mbuf clusters. Use this instead of m_copypacket() when 682 * you need a writable copy of an mbuf chain. 683 */ 684 struct mbuf * 685 m_dup(const struct mbuf *m, int how) 686 { 687 struct mbuf **p, *top = NULL; 688 int remain, moff, nsize; 689 690 MBUF_CHECKSLEEP(how); 691 /* Sanity check */ 692 if (m == NULL) 693 return (NULL); 694 M_ASSERTPKTHDR(m); 695 696 /* While there's more data, get a new mbuf, tack it on, and fill it */ 697 remain = m->m_pkthdr.len; 698 moff = 0; 699 p = ⊤ 700 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */ 701 struct mbuf *n; 702 703 /* Get the next new mbuf */ 704 if (remain >= MINCLSIZE) { 705 n = m_getcl(how, m->m_type, 0); 706 nsize = MCLBYTES; 707 } else { 708 n = m_get(how, m->m_type); 709 nsize = MLEN; 710 } 711 if (n == NULL) 712 goto nospace; 713 714 if (top == NULL) { /* First one, must be PKTHDR */ 715 if (!m_dup_pkthdr(n, m, how)) { 716 m_free(n); 717 goto nospace; 718 } 719 if ((n->m_flags & M_EXT) == 0) 720 nsize = MHLEN; 721 n->m_flags &= ~M_RDONLY; 722 } 723 n->m_len = 0; 724 725 /* Link it into the new chain */ 726 *p = n; 727 p = &n->m_next; 728 729 /* Copy data from original mbuf(s) into new mbuf */ 730 while (n->m_len < nsize && m != NULL) { 731 int chunk = min(nsize - n->m_len, m->m_len - moff); 732 733 m_copydata(m, moff, chunk, n->m_data + n->m_len); 734 moff += chunk; 735 n->m_len += chunk; 736 remain -= chunk; 737 if (moff == m->m_len) { 738 m = m->m_next; 739 moff = 0; 740 } 741 } 742 743 /* Check correct total mbuf length */ 744 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL), 745 ("%s: bogus m_pkthdr.len", __func__)); 746 } 747 return (top); 748 749 nospace: 750 m_freem(top); 751 return (NULL); 752 } 753 754 /* 755 * Concatenate mbuf chain n to m. 756 * Both chains must be of the same type (e.g. MT_DATA). 757 * Any m_pkthdr is not updated. 758 */ 759 void 760 m_cat(struct mbuf *m, struct mbuf *n) 761 { 762 while (m->m_next) 763 m = m->m_next; 764 while (n) { 765 if (!M_WRITABLE(m) || 766 (n->m_flags & M_EXTPG) != 0 || 767 M_TRAILINGSPACE(m) < n->m_len) { 768 /* just join the two chains */ 769 m->m_next = n; 770 return; 771 } 772 /* splat the data from one into the other */ 773 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len, 774 (u_int)n->m_len); 775 m->m_len += n->m_len; 776 n = m_free(n); 777 } 778 } 779 780 /* 781 * Concatenate two pkthdr mbuf chains. 782 */ 783 void 784 m_catpkt(struct mbuf *m, struct mbuf *n) 785 { 786 787 M_ASSERTPKTHDR(m); 788 M_ASSERTPKTHDR(n); 789 790 m->m_pkthdr.len += n->m_pkthdr.len; 791 m_demote(n, 1, 0); 792 793 m_cat(m, n); 794 } 795 796 void 797 m_adj(struct mbuf *mp, int req_len) 798 { 799 int len = req_len; 800 struct mbuf *m; 801 int count; 802 803 if ((m = mp) == NULL) 804 return; 805 if (len >= 0) { 806 /* 807 * Trim from head. 808 */ 809 while (m != NULL && len > 0) { 810 if (m->m_len <= len) { 811 len -= m->m_len; 812 m->m_len = 0; 813 m = m->m_next; 814 } else { 815 m->m_len -= len; 816 m->m_data += len; 817 len = 0; 818 } 819 } 820 if (mp->m_flags & M_PKTHDR) 821 mp->m_pkthdr.len -= (req_len - len); 822 } else { 823 /* 824 * Trim from tail. Scan the mbuf chain, 825 * calculating its length and finding the last mbuf. 826 * If the adjustment only affects this mbuf, then just 827 * adjust and return. Otherwise, rescan and truncate 828 * after the remaining size. 829 */ 830 len = -len; 831 count = 0; 832 for (;;) { 833 count += m->m_len; 834 if (m->m_next == (struct mbuf *)0) 835 break; 836 m = m->m_next; 837 } 838 if (m->m_len >= len) { 839 m->m_len -= len; 840 if (mp->m_flags & M_PKTHDR) 841 mp->m_pkthdr.len -= len; 842 return; 843 } 844 count -= len; 845 if (count < 0) 846 count = 0; 847 /* 848 * Correct length for chain is "count". 849 * Find the mbuf with last data, adjust its length, 850 * and toss data from remaining mbufs on chain. 851 */ 852 m = mp; 853 if (m->m_flags & M_PKTHDR) 854 m->m_pkthdr.len = count; 855 for (; m; m = m->m_next) { 856 if (m->m_len >= count) { 857 m->m_len = count; 858 if (m->m_next != NULL) { 859 m_freem(m->m_next); 860 m->m_next = NULL; 861 } 862 break; 863 } 864 count -= m->m_len; 865 } 866 } 867 } 868 869 void 870 m_adj_decap(struct mbuf *mp, int len) 871 { 872 uint8_t rsstype; 873 874 m_adj(mp, len); 875 if ((mp->m_flags & M_PKTHDR) != 0) { 876 /* 877 * If flowid was calculated by card from the inner 878 * headers, move flowid to the decapsulated mbuf 879 * chain, otherwise clear. This depends on the 880 * internals of m_adj, which keeps pkthdr as is, in 881 * particular not changing rsstype and flowid. 882 */ 883 rsstype = mp->m_pkthdr.rsstype; 884 if ((rsstype & M_HASHTYPE_INNER) != 0) { 885 M_HASHTYPE_SET(mp, rsstype & ~M_HASHTYPE_INNER); 886 } else { 887 M_HASHTYPE_CLEAR(mp); 888 } 889 } 890 } 891 892 /* 893 * Rearange an mbuf chain so that len bytes are contiguous 894 * and in the data area of an mbuf (so that mtod will work 895 * for a structure of size len). Returns the resulting 896 * mbuf chain on success, frees it and returns null on failure. 897 * If there is room, it will add up to max_protohdr-len extra bytes to the 898 * contiguous region in an attempt to avoid being called next time. 899 */ 900 struct mbuf * 901 m_pullup(struct mbuf *n, int len) 902 { 903 struct mbuf *m; 904 int count; 905 int space; 906 907 KASSERT((n->m_flags & M_EXTPG) == 0, 908 ("%s: unmapped mbuf %p", __func__, n)); 909 910 /* 911 * If first mbuf has no cluster, and has room for len bytes 912 * without shifting current data, pullup into it, 913 * otherwise allocate a new mbuf to prepend to the chain. 914 */ 915 if ((n->m_flags & M_EXT) == 0 && 916 n->m_data + len < &n->m_dat[MLEN] && n->m_next) { 917 if (n->m_len >= len) 918 return (n); 919 m = n; 920 n = n->m_next; 921 len -= m->m_len; 922 } else { 923 if (len > MHLEN) 924 goto bad; 925 m = m_get(M_NOWAIT, n->m_type); 926 if (m == NULL) 927 goto bad; 928 if (n->m_flags & M_PKTHDR) 929 m_move_pkthdr(m, n); 930 } 931 space = &m->m_dat[MLEN] - (m->m_data + m->m_len); 932 do { 933 count = min(min(max(len, max_protohdr), space), n->m_len); 934 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len, 935 (u_int)count); 936 len -= count; 937 m->m_len += count; 938 n->m_len -= count; 939 space -= count; 940 if (n->m_len) 941 n->m_data += count; 942 else 943 n = m_free(n); 944 } while (len > 0 && n); 945 if (len > 0) { 946 (void) m_free(m); 947 goto bad; 948 } 949 m->m_next = n; 950 return (m); 951 bad: 952 m_freem(n); 953 return (NULL); 954 } 955 956 /* 957 * Like m_pullup(), except a new mbuf is always allocated, and we allow 958 * the amount of empty space before the data in the new mbuf to be specified 959 * (in the event that the caller expects to prepend later). 960 */ 961 struct mbuf * 962 m_copyup(struct mbuf *n, int len, int dstoff) 963 { 964 struct mbuf *m; 965 int count, space; 966 967 if (len > (MHLEN - dstoff)) 968 goto bad; 969 m = m_get(M_NOWAIT, n->m_type); 970 if (m == NULL) 971 goto bad; 972 if (n->m_flags & M_PKTHDR) 973 m_move_pkthdr(m, n); 974 m->m_data += dstoff; 975 space = &m->m_dat[MLEN] - (m->m_data + m->m_len); 976 do { 977 count = min(min(max(len, max_protohdr), space), n->m_len); 978 memcpy(mtod(m, caddr_t) + m->m_len, mtod(n, caddr_t), 979 (unsigned)count); 980 len -= count; 981 m->m_len += count; 982 n->m_len -= count; 983 space -= count; 984 if (n->m_len) 985 n->m_data += count; 986 else 987 n = m_free(n); 988 } while (len > 0 && n); 989 if (len > 0) { 990 (void) m_free(m); 991 goto bad; 992 } 993 m->m_next = n; 994 return (m); 995 bad: 996 m_freem(n); 997 return (NULL); 998 } 999 1000 /* 1001 * Partition an mbuf chain in two pieces, returning the tail -- 1002 * all but the first len0 bytes. In case of failure, it returns NULL and 1003 * attempts to restore the chain to its original state. 1004 * 1005 * Note that the resulting mbufs might be read-only, because the new 1006 * mbuf can end up sharing an mbuf cluster with the original mbuf if 1007 * the "breaking point" happens to lie within a cluster mbuf. Use the 1008 * M_WRITABLE() macro to check for this case. 1009 */ 1010 struct mbuf * 1011 m_split(struct mbuf *m0, int len0, int wait) 1012 { 1013 struct mbuf *m, *n; 1014 u_int len = len0, remain; 1015 1016 MBUF_CHECKSLEEP(wait); 1017 for (m = m0; m && len > m->m_len; m = m->m_next) 1018 len -= m->m_len; 1019 if (m == NULL) 1020 return (NULL); 1021 remain = m->m_len - len; 1022 if (m0->m_flags & M_PKTHDR && remain == 0) { 1023 n = m_gethdr(wait, m0->m_type); 1024 if (n == NULL) 1025 return (NULL); 1026 n->m_next = m->m_next; 1027 m->m_next = NULL; 1028 if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG) { 1029 n->m_pkthdr.snd_tag = 1030 m_snd_tag_ref(m0->m_pkthdr.snd_tag); 1031 n->m_pkthdr.csum_flags |= CSUM_SND_TAG; 1032 } else 1033 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif; 1034 n->m_pkthdr.len = m0->m_pkthdr.len - len0; 1035 m0->m_pkthdr.len = len0; 1036 return (n); 1037 } else if (m0->m_flags & M_PKTHDR) { 1038 n = m_gethdr(wait, m0->m_type); 1039 if (n == NULL) 1040 return (NULL); 1041 if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG) { 1042 n->m_pkthdr.snd_tag = 1043 m_snd_tag_ref(m0->m_pkthdr.snd_tag); 1044 n->m_pkthdr.csum_flags |= CSUM_SND_TAG; 1045 } else 1046 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif; 1047 n->m_pkthdr.len = m0->m_pkthdr.len - len0; 1048 m0->m_pkthdr.len = len0; 1049 if (m->m_flags & (M_EXT|M_EXTPG)) 1050 goto extpacket; 1051 if (remain > MHLEN) { 1052 /* m can't be the lead packet */ 1053 M_ALIGN(n, 0); 1054 n->m_next = m_split(m, len, wait); 1055 if (n->m_next == NULL) { 1056 (void) m_free(n); 1057 return (NULL); 1058 } else { 1059 n->m_len = 0; 1060 return (n); 1061 } 1062 } else 1063 M_ALIGN(n, remain); 1064 } else if (remain == 0) { 1065 n = m->m_next; 1066 m->m_next = NULL; 1067 return (n); 1068 } else { 1069 n = m_get(wait, m->m_type); 1070 if (n == NULL) 1071 return (NULL); 1072 M_ALIGN(n, remain); 1073 } 1074 extpacket: 1075 if (m->m_flags & (M_EXT|M_EXTPG)) { 1076 n->m_data = m->m_data + len; 1077 mb_dupcl(n, m); 1078 } else { 1079 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain); 1080 } 1081 n->m_len = remain; 1082 m->m_len = len; 1083 n->m_next = m->m_next; 1084 m->m_next = NULL; 1085 return (n); 1086 } 1087 /* 1088 * Routine to copy from device local memory into mbufs. 1089 * Note that `off' argument is offset into first mbuf of target chain from 1090 * which to begin copying the data to. 1091 */ 1092 struct mbuf * 1093 m_devget(char *buf, int totlen, int off, struct ifnet *ifp, 1094 void (*copy)(char *from, caddr_t to, u_int len)) 1095 { 1096 struct mbuf *m; 1097 struct mbuf *top = NULL, **mp = ⊤ 1098 int len; 1099 1100 if (off < 0 || off > MHLEN) 1101 return (NULL); 1102 1103 while (totlen > 0) { 1104 if (top == NULL) { /* First one, must be PKTHDR */ 1105 if (totlen + off >= MINCLSIZE) { 1106 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 1107 len = MCLBYTES; 1108 } else { 1109 m = m_gethdr(M_NOWAIT, MT_DATA); 1110 len = MHLEN; 1111 1112 /* Place initial small packet/header at end of mbuf */ 1113 if (m && totlen + off + max_linkhdr <= MHLEN) { 1114 m->m_data += max_linkhdr; 1115 len -= max_linkhdr; 1116 } 1117 } 1118 if (m == NULL) 1119 return NULL; 1120 m->m_pkthdr.rcvif = ifp; 1121 m->m_pkthdr.len = totlen; 1122 } else { 1123 if (totlen + off >= MINCLSIZE) { 1124 m = m_getcl(M_NOWAIT, MT_DATA, 0); 1125 len = MCLBYTES; 1126 } else { 1127 m = m_get(M_NOWAIT, MT_DATA); 1128 len = MLEN; 1129 } 1130 if (m == NULL) { 1131 m_freem(top); 1132 return NULL; 1133 } 1134 } 1135 if (off) { 1136 m->m_data += off; 1137 len -= off; 1138 off = 0; 1139 } 1140 m->m_len = len = min(totlen, len); 1141 if (copy) 1142 copy(buf, mtod(m, caddr_t), (u_int)len); 1143 else 1144 bcopy(buf, mtod(m, caddr_t), (u_int)len); 1145 buf += len; 1146 *mp = m; 1147 mp = &m->m_next; 1148 totlen -= len; 1149 } 1150 return (top); 1151 } 1152 1153 static void 1154 m_copytounmapped(const struct mbuf *m, int off, int len, c_caddr_t cp) 1155 { 1156 struct iovec iov; 1157 struct uio uio; 1158 int error; 1159 1160 KASSERT(off >= 0, ("m_copytounmapped: negative off %d", off)); 1161 KASSERT(len >= 0, ("m_copytounmapped: negative len %d", len)); 1162 KASSERT(off < m->m_len, ("m_copytounmapped: len exceeds mbuf length")); 1163 iov.iov_base = __DECONST(caddr_t, cp); 1164 iov.iov_len = len; 1165 uio.uio_resid = len; 1166 uio.uio_iov = &iov; 1167 uio.uio_segflg = UIO_SYSSPACE; 1168 uio.uio_iovcnt = 1; 1169 uio.uio_offset = 0; 1170 uio.uio_rw = UIO_WRITE; 1171 error = m_unmapped_uiomove(m, off, &uio, len); 1172 KASSERT(error == 0, ("m_unmapped_uiomove failed: off %d, len %d", off, 1173 len)); 1174 } 1175 1176 /* 1177 * Copy data from a buffer back into the indicated mbuf chain, 1178 * starting "off" bytes from the beginning, extending the mbuf 1179 * chain if necessary. 1180 */ 1181 void 1182 m_copyback(struct mbuf *m0, int off, int len, c_caddr_t cp) 1183 { 1184 int mlen; 1185 struct mbuf *m = m0, *n; 1186 int totlen = 0; 1187 1188 if (m0 == NULL) 1189 return; 1190 while (off > (mlen = m->m_len)) { 1191 off -= mlen; 1192 totlen += mlen; 1193 if (m->m_next == NULL) { 1194 n = m_get(M_NOWAIT, m->m_type); 1195 if (n == NULL) 1196 goto out; 1197 bzero(mtod(n, caddr_t), MLEN); 1198 n->m_len = min(MLEN, len + off); 1199 m->m_next = n; 1200 } 1201 m = m->m_next; 1202 } 1203 while (len > 0) { 1204 if (m->m_next == NULL && (len > m->m_len - off)) { 1205 m->m_len += min(len - (m->m_len - off), 1206 M_TRAILINGSPACE(m)); 1207 } 1208 mlen = min (m->m_len - off, len); 1209 if ((m->m_flags & M_EXTPG) != 0) 1210 m_copytounmapped(m, off, mlen, cp); 1211 else 1212 bcopy(cp, off + mtod(m, caddr_t), (u_int)mlen); 1213 cp += mlen; 1214 len -= mlen; 1215 mlen += off; 1216 off = 0; 1217 totlen += mlen; 1218 if (len == 0) 1219 break; 1220 if (m->m_next == NULL) { 1221 n = m_get(M_NOWAIT, m->m_type); 1222 if (n == NULL) 1223 break; 1224 n->m_len = min(MLEN, len); 1225 m->m_next = n; 1226 } 1227 m = m->m_next; 1228 } 1229 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen)) 1230 m->m_pkthdr.len = totlen; 1231 } 1232 1233 /* 1234 * Append the specified data to the indicated mbuf chain, 1235 * Extend the mbuf chain if the new data does not fit in 1236 * existing space. 1237 * 1238 * Return 1 if able to complete the job; otherwise 0. 1239 */ 1240 int 1241 m_append(struct mbuf *m0, int len, c_caddr_t cp) 1242 { 1243 struct mbuf *m, *n; 1244 int remainder, space; 1245 1246 for (m = m0; m->m_next != NULL; m = m->m_next) 1247 ; 1248 remainder = len; 1249 space = M_TRAILINGSPACE(m); 1250 if (space > 0) { 1251 /* 1252 * Copy into available space. 1253 */ 1254 if (space > remainder) 1255 space = remainder; 1256 bcopy(cp, mtod(m, caddr_t) + m->m_len, space); 1257 m->m_len += space; 1258 cp += space, remainder -= space; 1259 } 1260 while (remainder > 0) { 1261 /* 1262 * Allocate a new mbuf; could check space 1263 * and allocate a cluster instead. 1264 */ 1265 n = m_get(M_NOWAIT, m->m_type); 1266 if (n == NULL) 1267 break; 1268 n->m_len = min(MLEN, remainder); 1269 bcopy(cp, mtod(n, caddr_t), n->m_len); 1270 cp += n->m_len, remainder -= n->m_len; 1271 m->m_next = n; 1272 m = n; 1273 } 1274 if (m0->m_flags & M_PKTHDR) 1275 m0->m_pkthdr.len += len - remainder; 1276 return (remainder == 0); 1277 } 1278 1279 static int 1280 m_apply_extpg_one(struct mbuf *m, int off, int len, 1281 int (*f)(void *, void *, u_int), void *arg) 1282 { 1283 void *p; 1284 u_int i, count, pgoff, pglen; 1285 int rval; 1286 1287 KASSERT(PMAP_HAS_DMAP, 1288 ("m_apply_extpg_one does not support unmapped mbufs")); 1289 off += mtod(m, vm_offset_t); 1290 if (off < m->m_epg_hdrlen) { 1291 count = min(m->m_epg_hdrlen - off, len); 1292 rval = f(arg, m->m_epg_hdr + off, count); 1293 if (rval) 1294 return (rval); 1295 len -= count; 1296 off = 0; 1297 } else 1298 off -= m->m_epg_hdrlen; 1299 pgoff = m->m_epg_1st_off; 1300 for (i = 0; i < m->m_epg_npgs && len > 0; i++) { 1301 pglen = m_epg_pagelen(m, i, pgoff); 1302 if (off < pglen) { 1303 count = min(pglen - off, len); 1304 p = (void *)PHYS_TO_DMAP(m->m_epg_pa[i] + pgoff + off); 1305 rval = f(arg, p, count); 1306 if (rval) 1307 return (rval); 1308 len -= count; 1309 off = 0; 1310 } else 1311 off -= pglen; 1312 pgoff = 0; 1313 } 1314 if (len > 0) { 1315 KASSERT(off < m->m_epg_trllen, 1316 ("m_apply_extpg_one: offset beyond trailer")); 1317 KASSERT(len <= m->m_epg_trllen - off, 1318 ("m_apply_extpg_one: length beyond trailer")); 1319 return (f(arg, m->m_epg_trail + off, len)); 1320 } 1321 return (0); 1322 } 1323 1324 /* Apply function f to the data in a single mbuf. */ 1325 static int 1326 m_apply_one(struct mbuf *m, int off, int len, 1327 int (*f)(void *, void *, u_int), void *arg) 1328 { 1329 if ((m->m_flags & M_EXTPG) != 0) 1330 return (m_apply_extpg_one(m, off, len, f, arg)); 1331 else 1332 return (f(arg, mtod(m, caddr_t) + off, len)); 1333 } 1334 1335 /* 1336 * Apply function f to the data in an mbuf chain starting "off" bytes from 1337 * the beginning, continuing for "len" bytes. 1338 */ 1339 int 1340 m_apply(struct mbuf *m, int off, int len, 1341 int (*f)(void *, void *, u_int), void *arg) 1342 { 1343 u_int count; 1344 int rval; 1345 1346 KASSERT(off >= 0, ("m_apply, negative off %d", off)); 1347 KASSERT(len >= 0, ("m_apply, negative len %d", len)); 1348 while (off > 0) { 1349 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain")); 1350 if (off < m->m_len) 1351 break; 1352 off -= m->m_len; 1353 m = m->m_next; 1354 } 1355 while (len > 0) { 1356 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain")); 1357 count = min(m->m_len - off, len); 1358 rval = m_apply_one(m, off, count, f, arg); 1359 if (rval) 1360 return (rval); 1361 len -= count; 1362 off = 0; 1363 m = m->m_next; 1364 } 1365 return (0); 1366 } 1367 1368 /* 1369 * Return a pointer to mbuf/offset of location in mbuf chain. 1370 */ 1371 struct mbuf * 1372 m_getptr(struct mbuf *m, int loc, int *off) 1373 { 1374 1375 while (loc >= 0) { 1376 /* Normal end of search. */ 1377 if (m->m_len > loc) { 1378 *off = loc; 1379 return (m); 1380 } else { 1381 loc -= m->m_len; 1382 if (m->m_next == NULL) { 1383 if (loc == 0) { 1384 /* Point at the end of valid data. */ 1385 *off = m->m_len; 1386 return (m); 1387 } 1388 return (NULL); 1389 } 1390 m = m->m_next; 1391 } 1392 } 1393 return (NULL); 1394 } 1395 1396 void 1397 m_print(const struct mbuf *m, int maxlen) 1398 { 1399 int len; 1400 int pdata; 1401 const struct mbuf *m2; 1402 1403 if (m == NULL) { 1404 printf("mbuf: %p\n", m); 1405 return; 1406 } 1407 1408 if (m->m_flags & M_PKTHDR) 1409 len = m->m_pkthdr.len; 1410 else 1411 len = -1; 1412 m2 = m; 1413 while (m2 != NULL && (len == -1 || len)) { 1414 pdata = m2->m_len; 1415 if (maxlen != -1 && pdata > maxlen) 1416 pdata = maxlen; 1417 printf("mbuf: %p len: %d, next: %p, %b%s", m2, m2->m_len, 1418 m2->m_next, m2->m_flags, "\20\20freelist\17skipfw" 1419 "\11proto5\10proto4\7proto3\6proto2\5proto1\4rdonly" 1420 "\3eor\2pkthdr\1ext", pdata ? "" : "\n"); 1421 if (pdata) 1422 printf(", %*D\n", pdata, (u_char *)m2->m_data, "-"); 1423 if (len != -1) 1424 len -= m2->m_len; 1425 m2 = m2->m_next; 1426 } 1427 if (len > 0) 1428 printf("%d bytes unaccounted for.\n", len); 1429 return; 1430 } 1431 1432 u_int 1433 m_fixhdr(struct mbuf *m0) 1434 { 1435 u_int len; 1436 1437 len = m_length(m0, NULL); 1438 m0->m_pkthdr.len = len; 1439 return (len); 1440 } 1441 1442 u_int 1443 m_length(struct mbuf *m0, struct mbuf **last) 1444 { 1445 struct mbuf *m; 1446 u_int len; 1447 1448 len = 0; 1449 for (m = m0; m != NULL; m = m->m_next) { 1450 len += m->m_len; 1451 if (m->m_next == NULL) 1452 break; 1453 } 1454 if (last != NULL) 1455 *last = m; 1456 return (len); 1457 } 1458 1459 /* 1460 * Defragment a mbuf chain, returning the shortest possible 1461 * chain of mbufs and clusters. If allocation fails and 1462 * this cannot be completed, NULL will be returned, but 1463 * the passed in chain will be unchanged. Upon success, 1464 * the original chain will be freed, and the new chain 1465 * will be returned. 1466 * 1467 * If a non-packet header is passed in, the original 1468 * mbuf (chain?) will be returned unharmed. 1469 */ 1470 struct mbuf * 1471 m_defrag(struct mbuf *m0, int how) 1472 { 1473 struct mbuf *m_new = NULL, *m_final = NULL; 1474 int progress = 0, length; 1475 1476 MBUF_CHECKSLEEP(how); 1477 if (!(m0->m_flags & M_PKTHDR)) 1478 return (m0); 1479 1480 m_fixhdr(m0); /* Needed sanity check */ 1481 1482 #ifdef MBUF_STRESS_TEST 1483 if (m_defragrandomfailures) { 1484 int temp = arc4random() & 0xff; 1485 if (temp == 0xba) 1486 goto nospace; 1487 } 1488 #endif 1489 1490 if (m0->m_pkthdr.len > MHLEN) 1491 m_final = m_getcl(how, MT_DATA, M_PKTHDR); 1492 else 1493 m_final = m_gethdr(how, MT_DATA); 1494 1495 if (m_final == NULL) 1496 goto nospace; 1497 1498 if (m_dup_pkthdr(m_final, m0, how) == 0) 1499 goto nospace; 1500 1501 m_new = m_final; 1502 1503 while (progress < m0->m_pkthdr.len) { 1504 length = m0->m_pkthdr.len - progress; 1505 if (length > MCLBYTES) 1506 length = MCLBYTES; 1507 1508 if (m_new == NULL) { 1509 if (length > MLEN) 1510 m_new = m_getcl(how, MT_DATA, 0); 1511 else 1512 m_new = m_get(how, MT_DATA); 1513 if (m_new == NULL) 1514 goto nospace; 1515 } 1516 1517 m_copydata(m0, progress, length, mtod(m_new, caddr_t)); 1518 progress += length; 1519 m_new->m_len = length; 1520 if (m_new != m_final) 1521 m_cat(m_final, m_new); 1522 m_new = NULL; 1523 } 1524 #ifdef MBUF_STRESS_TEST 1525 if (m0->m_next == NULL) 1526 m_defraguseless++; 1527 #endif 1528 m_freem(m0); 1529 m0 = m_final; 1530 #ifdef MBUF_STRESS_TEST 1531 m_defragpackets++; 1532 m_defragbytes += m0->m_pkthdr.len; 1533 #endif 1534 return (m0); 1535 nospace: 1536 #ifdef MBUF_STRESS_TEST 1537 m_defragfailure++; 1538 #endif 1539 if (m_final) 1540 m_freem(m_final); 1541 return (NULL); 1542 } 1543 1544 /* 1545 * Return the number of fragments an mbuf will use. This is usually 1546 * used as a proxy for the number of scatter/gather elements needed by 1547 * a DMA engine to access an mbuf. In general mapped mbufs are 1548 * assumed to be backed by physically contiguous buffers that only 1549 * need a single fragment. Unmapped mbufs, on the other hand, can 1550 * span disjoint physical pages. 1551 */ 1552 static int 1553 frags_per_mbuf(struct mbuf *m) 1554 { 1555 int frags; 1556 1557 if ((m->m_flags & M_EXTPG) == 0) 1558 return (1); 1559 1560 /* 1561 * The header and trailer are counted as a single fragment 1562 * each when present. 1563 * 1564 * XXX: This overestimates the number of fragments by assuming 1565 * all the backing physical pages are disjoint. 1566 */ 1567 frags = 0; 1568 if (m->m_epg_hdrlen != 0) 1569 frags++; 1570 frags += m->m_epg_npgs; 1571 if (m->m_epg_trllen != 0) 1572 frags++; 1573 1574 return (frags); 1575 } 1576 1577 /* 1578 * Defragment an mbuf chain, returning at most maxfrags separate 1579 * mbufs+clusters. If this is not possible NULL is returned and 1580 * the original mbuf chain is left in its present (potentially 1581 * modified) state. We use two techniques: collapsing consecutive 1582 * mbufs and replacing consecutive mbufs by a cluster. 1583 * 1584 * NB: this should really be named m_defrag but that name is taken 1585 */ 1586 struct mbuf * 1587 m_collapse(struct mbuf *m0, int how, int maxfrags) 1588 { 1589 struct mbuf *m, *n, *n2, **prev; 1590 u_int curfrags; 1591 1592 /* 1593 * Calculate the current number of frags. 1594 */ 1595 curfrags = 0; 1596 for (m = m0; m != NULL; m = m->m_next) 1597 curfrags += frags_per_mbuf(m); 1598 /* 1599 * First, try to collapse mbufs. Note that we always collapse 1600 * towards the front so we don't need to deal with moving the 1601 * pkthdr. This may be suboptimal if the first mbuf has much 1602 * less data than the following. 1603 */ 1604 m = m0; 1605 again: 1606 for (;;) { 1607 n = m->m_next; 1608 if (n == NULL) 1609 break; 1610 if (M_WRITABLE(m) && 1611 n->m_len < M_TRAILINGSPACE(m)) { 1612 m_copydata(n, 0, n->m_len, 1613 mtod(m, char *) + m->m_len); 1614 m->m_len += n->m_len; 1615 m->m_next = n->m_next; 1616 curfrags -= frags_per_mbuf(n); 1617 m_free(n); 1618 if (curfrags <= maxfrags) 1619 return m0; 1620 } else 1621 m = n; 1622 } 1623 KASSERT(maxfrags > 1, 1624 ("maxfrags %u, but normal collapse failed", maxfrags)); 1625 /* 1626 * Collapse consecutive mbufs to a cluster. 1627 */ 1628 prev = &m0->m_next; /* NB: not the first mbuf */ 1629 while ((n = *prev) != NULL) { 1630 if ((n2 = n->m_next) != NULL && 1631 n->m_len + n2->m_len < MCLBYTES) { 1632 m = m_getcl(how, MT_DATA, 0); 1633 if (m == NULL) 1634 goto bad; 1635 m_copydata(n, 0, n->m_len, mtod(m, char *)); 1636 m_copydata(n2, 0, n2->m_len, 1637 mtod(m, char *) + n->m_len); 1638 m->m_len = n->m_len + n2->m_len; 1639 m->m_next = n2->m_next; 1640 *prev = m; 1641 curfrags += 1; /* For the new cluster */ 1642 curfrags -= frags_per_mbuf(n); 1643 curfrags -= frags_per_mbuf(n2); 1644 m_free(n); 1645 m_free(n2); 1646 if (curfrags <= maxfrags) 1647 return m0; 1648 /* 1649 * Still not there, try the normal collapse 1650 * again before we allocate another cluster. 1651 */ 1652 goto again; 1653 } 1654 prev = &n->m_next; 1655 } 1656 /* 1657 * No place where we can collapse to a cluster; punt. 1658 * This can occur if, for example, you request 2 frags 1659 * but the packet requires that both be clusters (we 1660 * never reallocate the first mbuf to avoid moving the 1661 * packet header). 1662 */ 1663 bad: 1664 return NULL; 1665 } 1666 1667 #ifdef MBUF_STRESS_TEST 1668 1669 /* 1670 * Fragment an mbuf chain. There's no reason you'd ever want to do 1671 * this in normal usage, but it's great for stress testing various 1672 * mbuf consumers. 1673 * 1674 * If fragmentation is not possible, the original chain will be 1675 * returned. 1676 * 1677 * Possible length values: 1678 * 0 no fragmentation will occur 1679 * > 0 each fragment will be of the specified length 1680 * -1 each fragment will be the same random value in length 1681 * -2 each fragment's length will be entirely random 1682 * (Random values range from 1 to 256) 1683 */ 1684 struct mbuf * 1685 m_fragment(struct mbuf *m0, int how, int length) 1686 { 1687 struct mbuf *m_first, *m_last; 1688 int divisor = 255, progress = 0, fraglen; 1689 1690 if (!(m0->m_flags & M_PKTHDR)) 1691 return (m0); 1692 1693 if (length == 0 || length < -2) 1694 return (m0); 1695 if (length > MCLBYTES) 1696 length = MCLBYTES; 1697 if (length < 0 && divisor > MCLBYTES) 1698 divisor = MCLBYTES; 1699 if (length == -1) 1700 length = 1 + (arc4random() % divisor); 1701 if (length > 0) 1702 fraglen = length; 1703 1704 m_fixhdr(m0); /* Needed sanity check */ 1705 1706 m_first = m_getcl(how, MT_DATA, M_PKTHDR); 1707 if (m_first == NULL) 1708 goto nospace; 1709 1710 if (m_dup_pkthdr(m_first, m0, how) == 0) 1711 goto nospace; 1712 1713 m_last = m_first; 1714 1715 while (progress < m0->m_pkthdr.len) { 1716 if (length == -2) 1717 fraglen = 1 + (arc4random() % divisor); 1718 if (fraglen > m0->m_pkthdr.len - progress) 1719 fraglen = m0->m_pkthdr.len - progress; 1720 1721 if (progress != 0) { 1722 struct mbuf *m_new = m_getcl(how, MT_DATA, 0); 1723 if (m_new == NULL) 1724 goto nospace; 1725 1726 m_last->m_next = m_new; 1727 m_last = m_new; 1728 } 1729 1730 m_copydata(m0, progress, fraglen, mtod(m_last, caddr_t)); 1731 progress += fraglen; 1732 m_last->m_len = fraglen; 1733 } 1734 m_freem(m0); 1735 m0 = m_first; 1736 return (m0); 1737 nospace: 1738 if (m_first) 1739 m_freem(m_first); 1740 /* Return the original chain on failure */ 1741 return (m0); 1742 } 1743 1744 #endif 1745 1746 /* 1747 * Free pages from mbuf_ext_pgs, assuming they were allocated via 1748 * vm_page_alloc() and aren't associated with any object. Complement 1749 * to allocator from m_uiotombuf_nomap(). 1750 */ 1751 void 1752 mb_free_mext_pgs(struct mbuf *m) 1753 { 1754 vm_page_t pg; 1755 1756 M_ASSERTEXTPG(m); 1757 for (int i = 0; i < m->m_epg_npgs; i++) { 1758 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]); 1759 vm_page_unwire_noq(pg); 1760 vm_page_free(pg); 1761 } 1762 } 1763 1764 static struct mbuf * 1765 m_uiotombuf_nomap(struct uio *uio, int how, int len, int maxseg, int flags) 1766 { 1767 struct mbuf *m, *mb, *prev; 1768 vm_page_t pg_array[MBUF_PEXT_MAX_PGS]; 1769 int error, length, i, needed; 1770 ssize_t total; 1771 int pflags = malloc2vm_flags(how) | VM_ALLOC_NODUMP | VM_ALLOC_WIRED; 1772 1773 MPASS((flags & M_PKTHDR) == 0); 1774 MPASS((how & M_ZERO) == 0); 1775 1776 /* 1777 * len can be zero or an arbitrary large value bound by 1778 * the total data supplied by the uio. 1779 */ 1780 if (len > 0) 1781 total = MIN(uio->uio_resid, len); 1782 else 1783 total = uio->uio_resid; 1784 1785 if (maxseg == 0) 1786 maxseg = MBUF_PEXT_MAX_PGS * PAGE_SIZE; 1787 1788 /* 1789 * If total is zero, return an empty mbuf. This can occur 1790 * for TLS 1.0 connections which send empty fragments as 1791 * a countermeasure against the known-IV weakness in CBC 1792 * ciphersuites. 1793 */ 1794 if (__predict_false(total == 0)) { 1795 mb = mb_alloc_ext_pgs(how, mb_free_mext_pgs); 1796 if (mb == NULL) 1797 return (NULL); 1798 mb->m_epg_flags = EPG_FLAG_ANON; 1799 return (mb); 1800 } 1801 1802 /* 1803 * Allocate the pages 1804 */ 1805 m = NULL; 1806 while (total > 0) { 1807 mb = mb_alloc_ext_pgs(how, mb_free_mext_pgs); 1808 if (mb == NULL) 1809 goto failed; 1810 if (m == NULL) 1811 m = mb; 1812 else 1813 prev->m_next = mb; 1814 prev = mb; 1815 mb->m_epg_flags = EPG_FLAG_ANON; 1816 needed = length = MIN(maxseg, total); 1817 for (i = 0; needed > 0; i++, needed -= PAGE_SIZE) { 1818 retry_page: 1819 pg_array[i] = vm_page_alloc_noobj(pflags); 1820 if (pg_array[i] == NULL) { 1821 if (how & M_NOWAIT) { 1822 goto failed; 1823 } else { 1824 vm_wait(NULL); 1825 goto retry_page; 1826 } 1827 } 1828 mb->m_epg_pa[i] = VM_PAGE_TO_PHYS(pg_array[i]); 1829 mb->m_epg_npgs++; 1830 } 1831 mb->m_epg_last_len = length - PAGE_SIZE * (mb->m_epg_npgs - 1); 1832 MBUF_EXT_PGS_ASSERT_SANITY(mb); 1833 total -= length; 1834 error = uiomove_fromphys(pg_array, 0, length, uio); 1835 if (error != 0) 1836 goto failed; 1837 mb->m_len = length; 1838 mb->m_ext.ext_size += PAGE_SIZE * mb->m_epg_npgs; 1839 if (flags & M_PKTHDR) 1840 m->m_pkthdr.len += length; 1841 } 1842 return (m); 1843 1844 failed: 1845 m_freem(m); 1846 return (NULL); 1847 } 1848 1849 /* 1850 * Copy the contents of uio into a properly sized mbuf chain. 1851 */ 1852 struct mbuf * 1853 m_uiotombuf(struct uio *uio, int how, int len, int align, int flags) 1854 { 1855 struct mbuf *m, *mb; 1856 int error, length; 1857 ssize_t total; 1858 int progress = 0; 1859 1860 if (flags & M_EXTPG) 1861 return (m_uiotombuf_nomap(uio, how, len, align, flags)); 1862 1863 /* 1864 * len can be zero or an arbitrary large value bound by 1865 * the total data supplied by the uio. 1866 */ 1867 if (len > 0) 1868 total = (uio->uio_resid < len) ? uio->uio_resid : len; 1869 else 1870 total = uio->uio_resid; 1871 1872 /* 1873 * The smallest unit returned by m_getm2() is a single mbuf 1874 * with pkthdr. We can't align past it. 1875 */ 1876 if (align >= MHLEN) 1877 return (NULL); 1878 1879 /* 1880 * Give us the full allocation or nothing. 1881 * If len is zero return the smallest empty mbuf. 1882 */ 1883 m = m_getm2(NULL, max(total + align, 1), how, MT_DATA, flags); 1884 if (m == NULL) 1885 return (NULL); 1886 m->m_data += align; 1887 1888 /* Fill all mbufs with uio data and update header information. */ 1889 for (mb = m; mb != NULL; mb = mb->m_next) { 1890 length = min(M_TRAILINGSPACE(mb), total - progress); 1891 1892 error = uiomove(mtod(mb, void *), length, uio); 1893 if (error) { 1894 m_freem(m); 1895 return (NULL); 1896 } 1897 1898 mb->m_len = length; 1899 progress += length; 1900 if (flags & M_PKTHDR) 1901 m->m_pkthdr.len += length; 1902 } 1903 KASSERT(progress == total, ("%s: progress != total", __func__)); 1904 1905 return (m); 1906 } 1907 1908 /* 1909 * Copy data to/from an unmapped mbuf into a uio limited by len if set. 1910 */ 1911 int 1912 m_unmapped_uiomove(const struct mbuf *m, int m_off, struct uio *uio, int len) 1913 { 1914 vm_page_t pg; 1915 int error, i, off, pglen, pgoff, seglen, segoff; 1916 1917 M_ASSERTEXTPG(m); 1918 error = 0; 1919 1920 /* Skip over any data removed from the front. */ 1921 off = mtod(m, vm_offset_t); 1922 1923 off += m_off; 1924 if (m->m_epg_hdrlen != 0) { 1925 if (off >= m->m_epg_hdrlen) { 1926 off -= m->m_epg_hdrlen; 1927 } else { 1928 seglen = m->m_epg_hdrlen - off; 1929 segoff = off; 1930 seglen = min(seglen, len); 1931 off = 0; 1932 len -= seglen; 1933 error = uiomove(__DECONST(void *, 1934 &m->m_epg_hdr[segoff]), seglen, uio); 1935 } 1936 } 1937 pgoff = m->m_epg_1st_off; 1938 for (i = 0; i < m->m_epg_npgs && error == 0 && len > 0; i++) { 1939 pglen = m_epg_pagelen(m, i, pgoff); 1940 if (off >= pglen) { 1941 off -= pglen; 1942 pgoff = 0; 1943 continue; 1944 } 1945 seglen = pglen - off; 1946 segoff = pgoff + off; 1947 off = 0; 1948 seglen = min(seglen, len); 1949 len -= seglen; 1950 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]); 1951 error = uiomove_fromphys(&pg, segoff, seglen, uio); 1952 pgoff = 0; 1953 }; 1954 if (len != 0 && error == 0) { 1955 KASSERT((off + len) <= m->m_epg_trllen, 1956 ("off + len > trail (%d + %d > %d, m_off = %d)", off, len, 1957 m->m_epg_trllen, m_off)); 1958 error = uiomove(__DECONST(void *, &m->m_epg_trail[off]), 1959 len, uio); 1960 } 1961 return (error); 1962 } 1963 1964 /* 1965 * Copy an mbuf chain into a uio limited by len if set. 1966 */ 1967 int 1968 m_mbuftouio(struct uio *uio, const struct mbuf *m, int len) 1969 { 1970 int error, length, total; 1971 int progress = 0; 1972 1973 if (len > 0) 1974 total = min(uio->uio_resid, len); 1975 else 1976 total = uio->uio_resid; 1977 1978 /* Fill the uio with data from the mbufs. */ 1979 for (; m != NULL; m = m->m_next) { 1980 length = min(m->m_len, total - progress); 1981 1982 if ((m->m_flags & M_EXTPG) != 0) 1983 error = m_unmapped_uiomove(m, 0, uio, length); 1984 else 1985 error = uiomove(mtod(m, void *), length, uio); 1986 if (error) 1987 return (error); 1988 1989 progress += length; 1990 } 1991 1992 return (0); 1993 } 1994 1995 /* 1996 * Create a writable copy of the mbuf chain. While doing this 1997 * we compact the chain with a goal of producing a chain with 1998 * at most two mbufs. The second mbuf in this chain is likely 1999 * to be a cluster. The primary purpose of this work is to create 2000 * a writable packet for encryption, compression, etc. The 2001 * secondary goal is to linearize the data so the data can be 2002 * passed to crypto hardware in the most efficient manner possible. 2003 */ 2004 struct mbuf * 2005 m_unshare(struct mbuf *m0, int how) 2006 { 2007 struct mbuf *m, *mprev; 2008 struct mbuf *n, *mfirst, *mlast; 2009 int len, off; 2010 2011 mprev = NULL; 2012 for (m = m0; m != NULL; m = mprev->m_next) { 2013 /* 2014 * Regular mbufs are ignored unless there's a cluster 2015 * in front of it that we can use to coalesce. We do 2016 * the latter mainly so later clusters can be coalesced 2017 * also w/o having to handle them specially (i.e. convert 2018 * mbuf+cluster -> cluster). This optimization is heavily 2019 * influenced by the assumption that we're running over 2020 * Ethernet where MCLBYTES is large enough that the max 2021 * packet size will permit lots of coalescing into a 2022 * single cluster. This in turn permits efficient 2023 * crypto operations, especially when using hardware. 2024 */ 2025 if ((m->m_flags & M_EXT) == 0) { 2026 if (mprev && (mprev->m_flags & M_EXT) && 2027 m->m_len <= M_TRAILINGSPACE(mprev)) { 2028 /* XXX: this ignores mbuf types */ 2029 memcpy(mtod(mprev, caddr_t) + mprev->m_len, 2030 mtod(m, caddr_t), m->m_len); 2031 mprev->m_len += m->m_len; 2032 mprev->m_next = m->m_next; /* unlink from chain */ 2033 m_free(m); /* reclaim mbuf */ 2034 } else { 2035 mprev = m; 2036 } 2037 continue; 2038 } 2039 /* 2040 * Writable mbufs are left alone (for now). 2041 */ 2042 if (M_WRITABLE(m)) { 2043 mprev = m; 2044 continue; 2045 } 2046 2047 /* 2048 * Not writable, replace with a copy or coalesce with 2049 * the previous mbuf if possible (since we have to copy 2050 * it anyway, we try to reduce the number of mbufs and 2051 * clusters so that future work is easier). 2052 */ 2053 KASSERT(m->m_flags & M_EXT, ("m_flags 0x%x", m->m_flags)); 2054 /* NB: we only coalesce into a cluster or larger */ 2055 if (mprev != NULL && (mprev->m_flags & M_EXT) && 2056 m->m_len <= M_TRAILINGSPACE(mprev)) { 2057 /* XXX: this ignores mbuf types */ 2058 memcpy(mtod(mprev, caddr_t) + mprev->m_len, 2059 mtod(m, caddr_t), m->m_len); 2060 mprev->m_len += m->m_len; 2061 mprev->m_next = m->m_next; /* unlink from chain */ 2062 m_free(m); /* reclaim mbuf */ 2063 continue; 2064 } 2065 2066 /* 2067 * Allocate new space to hold the copy and copy the data. 2068 * We deal with jumbo mbufs (i.e. m_len > MCLBYTES) by 2069 * splitting them into clusters. We could just malloc a 2070 * buffer and make it external but too many device drivers 2071 * don't know how to break up the non-contiguous memory when 2072 * doing DMA. 2073 */ 2074 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS); 2075 if (n == NULL) { 2076 m_freem(m0); 2077 return (NULL); 2078 } 2079 if (m->m_flags & M_PKTHDR) { 2080 KASSERT(mprev == NULL, ("%s: m0 %p, m %p has M_PKTHDR", 2081 __func__, m0, m)); 2082 m_move_pkthdr(n, m); 2083 } 2084 len = m->m_len; 2085 off = 0; 2086 mfirst = n; 2087 mlast = NULL; 2088 for (;;) { 2089 int cc = min(len, MCLBYTES); 2090 memcpy(mtod(n, caddr_t), mtod(m, caddr_t) + off, cc); 2091 n->m_len = cc; 2092 if (mlast != NULL) 2093 mlast->m_next = n; 2094 mlast = n; 2095 #if 0 2096 newipsecstat.ips_clcopied++; 2097 #endif 2098 2099 len -= cc; 2100 if (len <= 0) 2101 break; 2102 off += cc; 2103 2104 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS); 2105 if (n == NULL) { 2106 m_freem(mfirst); 2107 m_freem(m0); 2108 return (NULL); 2109 } 2110 } 2111 n->m_next = m->m_next; 2112 if (mprev == NULL) 2113 m0 = mfirst; /* new head of chain */ 2114 else 2115 mprev->m_next = mfirst; /* replace old mbuf */ 2116 m_free(m); /* release old mbuf */ 2117 mprev = mfirst; 2118 } 2119 return (m0); 2120 } 2121 2122 #ifdef MBUF_PROFILING 2123 2124 #define MP_BUCKETS 32 /* don't just change this as things may overflow.*/ 2125 struct mbufprofile { 2126 uintmax_t wasted[MP_BUCKETS]; 2127 uintmax_t used[MP_BUCKETS]; 2128 uintmax_t segments[MP_BUCKETS]; 2129 } mbprof; 2130 2131 #define MP_MAXDIGITS 21 /* strlen("16,000,000,000,000,000,000") == 21 */ 2132 #define MP_NUMLINES 6 2133 #define MP_NUMSPERLINE 16 2134 #define MP_EXTRABYTES 64 /* > strlen("used:\nwasted:\nsegments:\n") */ 2135 /* work out max space needed and add a bit of spare space too */ 2136 #define MP_MAXLINE ((MP_MAXDIGITS+1) * MP_NUMSPERLINE) 2137 #define MP_BUFSIZE ((MP_MAXLINE * MP_NUMLINES) + 1 + MP_EXTRABYTES) 2138 2139 char mbprofbuf[MP_BUFSIZE]; 2140 2141 void 2142 m_profile(struct mbuf *m) 2143 { 2144 int segments = 0; 2145 int used = 0; 2146 int wasted = 0; 2147 2148 while (m) { 2149 segments++; 2150 used += m->m_len; 2151 if (m->m_flags & M_EXT) { 2152 wasted += MHLEN - sizeof(m->m_ext) + 2153 m->m_ext.ext_size - m->m_len; 2154 } else { 2155 if (m->m_flags & M_PKTHDR) 2156 wasted += MHLEN - m->m_len; 2157 else 2158 wasted += MLEN - m->m_len; 2159 } 2160 m = m->m_next; 2161 } 2162 /* be paranoid.. it helps */ 2163 if (segments > MP_BUCKETS - 1) 2164 segments = MP_BUCKETS - 1; 2165 if (used > 100000) 2166 used = 100000; 2167 if (wasted > 100000) 2168 wasted = 100000; 2169 /* store in the appropriate bucket */ 2170 /* don't bother locking. if it's slightly off, so what? */ 2171 mbprof.segments[segments]++; 2172 mbprof.used[fls(used)]++; 2173 mbprof.wasted[fls(wasted)]++; 2174 } 2175 2176 static void 2177 mbprof_textify(void) 2178 { 2179 int offset; 2180 char *c; 2181 uint64_t *p; 2182 2183 p = &mbprof.wasted[0]; 2184 c = mbprofbuf; 2185 offset = snprintf(c, MP_MAXLINE + 10, 2186 "wasted:\n" 2187 "%ju %ju %ju %ju %ju %ju %ju %ju " 2188 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 2189 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 2190 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 2191 #ifdef BIG_ARRAY 2192 p = &mbprof.wasted[16]; 2193 c += offset; 2194 offset = snprintf(c, MP_MAXLINE, 2195 "%ju %ju %ju %ju %ju %ju %ju %ju " 2196 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 2197 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 2198 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 2199 #endif 2200 p = &mbprof.used[0]; 2201 c += offset; 2202 offset = snprintf(c, MP_MAXLINE + 10, 2203 "used:\n" 2204 "%ju %ju %ju %ju %ju %ju %ju %ju " 2205 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 2206 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 2207 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 2208 #ifdef BIG_ARRAY 2209 p = &mbprof.used[16]; 2210 c += offset; 2211 offset = snprintf(c, MP_MAXLINE, 2212 "%ju %ju %ju %ju %ju %ju %ju %ju " 2213 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 2214 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 2215 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 2216 #endif 2217 p = &mbprof.segments[0]; 2218 c += offset; 2219 offset = snprintf(c, MP_MAXLINE + 10, 2220 "segments:\n" 2221 "%ju %ju %ju %ju %ju %ju %ju %ju " 2222 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 2223 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 2224 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 2225 #ifdef BIG_ARRAY 2226 p = &mbprof.segments[16]; 2227 c += offset; 2228 offset = snprintf(c, MP_MAXLINE, 2229 "%ju %ju %ju %ju %ju %ju %ju %ju " 2230 "%ju %ju %ju %ju %ju %ju %ju %jju", 2231 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 2232 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 2233 #endif 2234 } 2235 2236 static int 2237 mbprof_handler(SYSCTL_HANDLER_ARGS) 2238 { 2239 int error; 2240 2241 mbprof_textify(); 2242 error = SYSCTL_OUT(req, mbprofbuf, strlen(mbprofbuf) + 1); 2243 return (error); 2244 } 2245 2246 static int 2247 mbprof_clr_handler(SYSCTL_HANDLER_ARGS) 2248 { 2249 int clear, error; 2250 2251 clear = 0; 2252 error = sysctl_handle_int(oidp, &clear, 0, req); 2253 if (error || !req->newptr) 2254 return (error); 2255 2256 if (clear) { 2257 bzero(&mbprof, sizeof(mbprof)); 2258 } 2259 2260 return (error); 2261 } 2262 2263 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofile, 2264 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0, 2265 mbprof_handler, "A", 2266 "mbuf profiling statistics"); 2267 2268 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofileclr, 2269 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0, 2270 mbprof_clr_handler, "I", 2271 "clear mbuf profiling statistics"); 2272 #endif 2273