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 32 #include <sys/cdefs.h> 33 #include "opt_param.h" 34 #include "opt_mbuf_stress_test.h" 35 #include "opt_mbuf_profiling.h" 36 37 #include <sys/param.h> 38 #include <sys/systm.h> 39 #include <sys/kernel.h> 40 #include <sys/limits.h> 41 #include <sys/lock.h> 42 #include <sys/malloc.h> 43 #include <sys/mbuf.h> 44 #include <sys/sysctl.h> 45 #include <sys/domain.h> 46 #include <sys/protosw.h> 47 #include <sys/uio.h> 48 #include <sys/vmmeter.h> 49 #include <sys/sbuf.h> 50 #include <sys/sdt.h> 51 #include <vm/vm.h> 52 #include <vm/vm_pageout.h> 53 #include <vm/vm_page.h> 54 55 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__init, 56 "struct mbuf *", "mbufinfo_t *", 57 "uint32_t", "uint32_t", 58 "uint16_t", "uint16_t", 59 "uint32_t", "uint32_t", 60 "uint32_t", "uint32_t"); 61 62 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__gethdr_raw, 63 "uint32_t", "uint32_t", 64 "uint16_t", "uint16_t", 65 "struct mbuf *", "mbufinfo_t *"); 66 67 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__gethdr, 68 "uint32_t", "uint32_t", 69 "uint16_t", "uint16_t", 70 "struct mbuf *", "mbufinfo_t *"); 71 72 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__get_raw, 73 "uint32_t", "uint32_t", 74 "uint16_t", "uint16_t", 75 "struct mbuf *", "mbufinfo_t *"); 76 77 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__get, 78 "uint32_t", "uint32_t", 79 "uint16_t", "uint16_t", 80 "struct mbuf *", "mbufinfo_t *"); 81 82 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__getcl, 83 "uint32_t", "uint32_t", 84 "uint16_t", "uint16_t", 85 "uint32_t", "uint32_t", 86 "struct mbuf *", "mbufinfo_t *"); 87 88 SDT_PROBE_DEFINE5_XLATE(sdt, , , m__getjcl, 89 "uint32_t", "uint32_t", 90 "uint16_t", "uint16_t", 91 "uint32_t", "uint32_t", 92 "uint32_t", "uint32_t", 93 "struct mbuf *", "mbufinfo_t *"); 94 95 SDT_PROBE_DEFINE3_XLATE(sdt, , , m__clget, 96 "struct mbuf *", "mbufinfo_t *", 97 "uint32_t", "uint32_t", 98 "uint32_t", "uint32_t"); 99 100 SDT_PROBE_DEFINE4_XLATE(sdt, , , m__cljget, 101 "struct mbuf *", "mbufinfo_t *", 102 "uint32_t", "uint32_t", 103 "uint32_t", "uint32_t", 104 "void*", "void*"); 105 106 SDT_PROBE_DEFINE(sdt, , , m__cljset); 107 108 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__free, 109 "struct mbuf *", "mbufinfo_t *"); 110 111 SDT_PROBE_DEFINE1_XLATE(sdt, , , m__freem, 112 "struct mbuf *", "mbufinfo_t *"); 113 114 #include <security/mac/mac_framework.h> 115 116 /* 117 * Provide minimum possible defaults for link and protocol header space, 118 * assuming IPv4 over Ethernet. Enabling IPv6, IEEE802.11 or some other 119 * protocol may grow these values. 120 */ 121 u_int max_linkhdr = 16; 122 u_int max_protohdr = 40; 123 u_int max_hdr = 16 + 40; 124 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RD, 125 &max_linkhdr, 16, "Size of largest link layer header"); 126 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RD, 127 &max_protohdr, 40, "Size of largest protocol layer header"); 128 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RD, 129 &max_hdr, 16 + 40, "Size of largest link plus protocol header"); 130 131 static void 132 max_hdr_grow(void) 133 { 134 135 max_hdr = max_linkhdr + max_protohdr; 136 MPASS(max_hdr <= MHLEN); 137 } 138 139 void 140 max_linkhdr_grow(u_int new) 141 { 142 143 if (new > max_linkhdr) { 144 max_linkhdr = new; 145 max_hdr_grow(); 146 } 147 } 148 149 void 150 max_protohdr_grow(u_int new) 151 { 152 153 if (new > max_protohdr) { 154 max_protohdr = new; 155 max_hdr_grow(); 156 } 157 } 158 159 #ifdef MBUF_STRESS_TEST 160 int m_defragpackets; 161 int m_defragbytes; 162 int m_defraguseless; 163 int m_defragfailure; 164 int m_defragrandomfailures; 165 166 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD, 167 &m_defragpackets, 0, ""); 168 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD, 169 &m_defragbytes, 0, ""); 170 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD, 171 &m_defraguseless, 0, ""); 172 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD, 173 &m_defragfailure, 0, ""); 174 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW, 175 &m_defragrandomfailures, 0, ""); 176 #endif 177 178 /* 179 * Ensure the correct size of various mbuf parameters. It could be off due 180 * to compiler-induced padding and alignment artifacts. 181 */ 182 CTASSERT(MSIZE - offsetof(struct mbuf, m_dat) == MLEN); 183 CTASSERT(MSIZE - offsetof(struct mbuf, m_pktdat) == MHLEN); 184 185 /* 186 * mbuf data storage should be 64-bit aligned regardless of architectural 187 * pointer size; check this is the case with and without a packet header. 188 */ 189 CTASSERT(offsetof(struct mbuf, m_dat) % 8 == 0); 190 CTASSERT(offsetof(struct mbuf, m_pktdat) % 8 == 0); 191 192 /* 193 * While the specific values here don't matter too much (i.e., +/- a few 194 * words), we do want to ensure that changes to these values are carefully 195 * reasoned about and properly documented. This is especially the case as 196 * network-protocol and device-driver modules encode these layouts, and must 197 * be recompiled if the structures change. Check these values at compile time 198 * against the ones documented in comments in mbuf.h. 199 * 200 * NB: Possibly they should be documented there via #define's and not just 201 * comments. 202 */ 203 #if defined(__LP64__) 204 CTASSERT(offsetof(struct mbuf, m_dat) == 32); 205 CTASSERT(sizeof(struct pkthdr) == 64); 206 CTASSERT(sizeof(struct m_ext) == 160); 207 #else 208 CTASSERT(offsetof(struct mbuf, m_dat) == 24); 209 CTASSERT(sizeof(struct pkthdr) == 56); 210 #if defined(__powerpc__) && defined(BOOKE) 211 /* PowerPC booke has 64-bit physical pointers. */ 212 CTASSERT(sizeof(struct m_ext) == 176); 213 #else 214 CTASSERT(sizeof(struct m_ext) == 172); 215 #endif 216 #endif 217 218 /* 219 * Assert that the queue(3) macros produce code of the same size as an old 220 * plain pointer does. 221 */ 222 #ifdef INVARIANTS 223 static struct mbuf __used m_assertbuf; 224 CTASSERT(sizeof(m_assertbuf.m_slist) == sizeof(m_assertbuf.m_next)); 225 CTASSERT(sizeof(m_assertbuf.m_stailq) == sizeof(m_assertbuf.m_next)); 226 CTASSERT(sizeof(m_assertbuf.m_slistpkt) == sizeof(m_assertbuf.m_nextpkt)); 227 CTASSERT(sizeof(m_assertbuf.m_stailqpkt) == sizeof(m_assertbuf.m_nextpkt)); 228 #endif 229 230 /* 231 * Attach the cluster from *m to *n, set up m_ext in *n 232 * and bump the refcount of the cluster. 233 */ 234 void 235 mb_dupcl(struct mbuf *n, struct mbuf *m) 236 { 237 volatile u_int *refcnt; 238 239 KASSERT(m->m_flags & (M_EXT | M_EXTPG), 240 ("%s: M_EXT | M_EXTPG not set on %p", __func__, m)); 241 KASSERT(!(n->m_flags & (M_EXT | M_EXTPG)), 242 ("%s: M_EXT | M_EXTPG set on %p", __func__, n)); 243 244 /* 245 * Cache access optimization. 246 * 247 * o Regular M_EXT storage doesn't need full copy of m_ext, since 248 * the holder of the 'ext_count' is responsible to carry the free 249 * routine and its arguments. 250 * o M_EXTPG data is split between main part of mbuf and m_ext, the 251 * main part is copied in full, the m_ext part is similar to M_EXT. 252 * o EXT_EXTREF, where 'ext_cnt' doesn't point into mbuf at all, is 253 * special - it needs full copy of m_ext into each mbuf, since any 254 * copy could end up as the last to free. 255 */ 256 if (m->m_flags & M_EXTPG) { 257 bcopy(&m->m_epg_startcopy, &n->m_epg_startcopy, 258 __rangeof(struct mbuf, m_epg_startcopy, m_epg_endcopy)); 259 bcopy(&m->m_ext, &n->m_ext, m_epg_ext_copylen); 260 } else if (m->m_ext.ext_type == EXT_EXTREF) 261 bcopy(&m->m_ext, &n->m_ext, sizeof(struct m_ext)); 262 else 263 bcopy(&m->m_ext, &n->m_ext, m_ext_copylen); 264 265 n->m_flags |= m->m_flags & (M_RDONLY | M_EXT | M_EXTPG); 266 267 /* See if this is the mbuf that holds the embedded refcount. */ 268 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) { 269 refcnt = n->m_ext.ext_cnt = &m->m_ext.ext_count; 270 n->m_ext.ext_flags &= ~EXT_FLAG_EMBREF; 271 } else { 272 KASSERT(m->m_ext.ext_cnt != NULL, 273 ("%s: no refcounting pointer on %p", __func__, m)); 274 refcnt = m->m_ext.ext_cnt; 275 } 276 277 if (*refcnt == 1) 278 *refcnt += 1; 279 else 280 atomic_add_int(refcnt, 1); 281 } 282 283 void 284 m_demote_pkthdr(struct mbuf *m) 285 { 286 287 M_ASSERTPKTHDR(m); 288 M_ASSERT_NO_SND_TAG(m); 289 290 m_tag_delete_chain(m, NULL); 291 m->m_flags &= ~M_PKTHDR; 292 bzero(&m->m_pkthdr, sizeof(struct pkthdr)); 293 } 294 295 /* 296 * Clean up mbuf (chain) from any tags and packet headers. 297 * If "all" is set then the first mbuf in the chain will be 298 * cleaned too. 299 */ 300 void 301 m_demote(struct mbuf *m0, int all, int flags) 302 { 303 struct mbuf *m; 304 305 flags |= M_DEMOTEFLAGS; 306 307 for (m = all ? m0 : m0->m_next; m != NULL; m = m->m_next) { 308 KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt in m %p, m0 %p", 309 __func__, m, m0)); 310 if (m->m_flags & M_PKTHDR) 311 m_demote_pkthdr(m); 312 m->m_flags &= flags; 313 } 314 } 315 316 /* 317 * Sanity checks on mbuf (chain) for use in KASSERT() and general 318 * debugging. 319 * Returns 0 or panics when bad and 1 on all tests passed. 320 * Sanitize, 0 to run M_SANITY_ACTION, 1 to garble things so they 321 * blow up later. 322 */ 323 int 324 m_sanity(struct mbuf *m0, int sanitize) 325 { 326 struct mbuf *m; 327 caddr_t a, b; 328 int pktlen = 0; 329 330 #ifdef INVARIANTS 331 #define M_SANITY_ACTION(s) panic("mbuf %p: " s, m) 332 #else 333 #define M_SANITY_ACTION(s) printf("mbuf %p: " s, m) 334 #endif 335 336 for (m = m0; m != NULL; m = m->m_next) { 337 /* 338 * Basic pointer checks. If any of these fails then some 339 * unrelated kernel memory before or after us is trashed. 340 * No way to recover from that. 341 */ 342 a = M_START(m); 343 b = a + M_SIZE(m); 344 if ((caddr_t)m->m_data < a) 345 M_SANITY_ACTION("m_data outside mbuf data range left"); 346 if ((caddr_t)m->m_data > b) 347 M_SANITY_ACTION("m_data outside mbuf data range right"); 348 if ((caddr_t)m->m_data + m->m_len > b) 349 M_SANITY_ACTION("m_data + m_len exeeds mbuf space"); 350 351 /* m->m_nextpkt may only be set on first mbuf in chain. */ 352 if (m != m0 && m->m_nextpkt != NULL) { 353 if (sanitize) { 354 m_freem(m->m_nextpkt); 355 m->m_nextpkt = (struct mbuf *)0xDEADC0DE; 356 } else 357 M_SANITY_ACTION("m->m_nextpkt on in-chain mbuf"); 358 } 359 360 /* packet length (not mbuf length!) calculation */ 361 if (m0->m_flags & M_PKTHDR) 362 pktlen += m->m_len; 363 364 /* m_tags may only be attached to first mbuf in chain. */ 365 if (m != m0 && m->m_flags & M_PKTHDR && 366 !SLIST_EMPTY(&m->m_pkthdr.tags)) { 367 if (sanitize) { 368 m_tag_delete_chain(m, NULL); 369 /* put in 0xDEADC0DE perhaps? */ 370 } else 371 M_SANITY_ACTION("m_tags on in-chain mbuf"); 372 } 373 374 /* M_PKTHDR may only be set on first mbuf in chain */ 375 if (m != m0 && m->m_flags & M_PKTHDR) { 376 if (sanitize) { 377 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr)); 378 m->m_flags &= ~M_PKTHDR; 379 /* put in 0xDEADCODE and leave hdr flag in */ 380 } else 381 M_SANITY_ACTION("M_PKTHDR on in-chain mbuf"); 382 } 383 } 384 m = m0; 385 if (pktlen && pktlen != m->m_pkthdr.len) { 386 if (sanitize) 387 m->m_pkthdr.len = 0; 388 else 389 M_SANITY_ACTION("m_pkthdr.len != mbuf chain length"); 390 } 391 return 1; 392 393 #undef M_SANITY_ACTION 394 } 395 396 /* 397 * Non-inlined part of m_init(). 398 */ 399 int 400 m_pkthdr_init(struct mbuf *m, int how) 401 { 402 #ifdef MAC 403 int error; 404 #endif 405 m->m_data = m->m_pktdat; 406 bzero(&m->m_pkthdr, sizeof(m->m_pkthdr)); 407 #ifdef NUMA 408 m->m_pkthdr.numa_domain = M_NODOM; 409 #endif 410 #ifdef MAC 411 /* If the label init fails, fail the alloc */ 412 error = mac_mbuf_init(m, how); 413 if (error) 414 return (error); 415 #endif 416 417 return (0); 418 } 419 420 /* 421 * "Move" mbuf pkthdr from "from" to "to". 422 * "from" must have M_PKTHDR set, and "to" must be empty. 423 */ 424 void 425 m_move_pkthdr(struct mbuf *to, struct mbuf *from) 426 { 427 428 #if 0 429 /* see below for why these are not enabled */ 430 M_ASSERTPKTHDR(to); 431 /* Note: with MAC, this may not be a good assertion. */ 432 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags), 433 ("m_move_pkthdr: to has tags")); 434 #endif 435 #ifdef MAC 436 /* 437 * XXXMAC: It could be this should also occur for non-MAC? 438 */ 439 if (to->m_flags & M_PKTHDR) 440 m_tag_delete_chain(to, NULL); 441 #endif 442 to->m_flags = (from->m_flags & M_COPYFLAGS) | 443 (to->m_flags & (M_EXT | M_EXTPG)); 444 if ((to->m_flags & M_EXT) == 0) 445 to->m_data = to->m_pktdat; 446 to->m_pkthdr = from->m_pkthdr; /* especially tags */ 447 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */ 448 from->m_flags &= ~M_PKTHDR; 449 if (from->m_pkthdr.csum_flags & CSUM_SND_TAG) { 450 from->m_pkthdr.csum_flags &= ~CSUM_SND_TAG; 451 from->m_pkthdr.snd_tag = NULL; 452 } 453 } 454 455 /* 456 * Duplicate "from"'s mbuf pkthdr in "to". 457 * "from" must have M_PKTHDR set, and "to" must be empty. 458 * In particular, this does a deep copy of the packet tags. 459 */ 460 int 461 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how) 462 { 463 464 #if 0 465 /* 466 * The mbuf allocator only initializes the pkthdr 467 * when the mbuf is allocated with m_gethdr(). Many users 468 * (e.g. m_copy*, m_prepend) use m_get() and then 469 * smash the pkthdr as needed causing these 470 * assertions to trip. For now just disable them. 471 */ 472 M_ASSERTPKTHDR(to); 473 /* Note: with MAC, this may not be a good assertion. */ 474 KASSERT(SLIST_EMPTY(&to->m_pkthdr.tags), ("m_dup_pkthdr: to has tags")); 475 #endif 476 MBUF_CHECKSLEEP(how); 477 #ifdef MAC 478 if (to->m_flags & M_PKTHDR) 479 m_tag_delete_chain(to, NULL); 480 #endif 481 to->m_flags = (from->m_flags & M_COPYFLAGS) | 482 (to->m_flags & (M_EXT | M_EXTPG)); 483 if ((to->m_flags & M_EXT) == 0) 484 to->m_data = to->m_pktdat; 485 to->m_pkthdr = from->m_pkthdr; 486 if (from->m_pkthdr.csum_flags & CSUM_SND_TAG) 487 m_snd_tag_ref(from->m_pkthdr.snd_tag); 488 SLIST_INIT(&to->m_pkthdr.tags); 489 return (m_tag_copy_chain(to, from, how)); 490 } 491 492 /* 493 * Lesser-used path for M_PREPEND: 494 * allocate new mbuf to prepend to chain, 495 * copy junk along. 496 */ 497 struct mbuf * 498 m_prepend(struct mbuf *m, int len, int how) 499 { 500 struct mbuf *mn; 501 502 if (m->m_flags & M_PKTHDR) 503 mn = m_gethdr(how, m->m_type); 504 else 505 mn = m_get(how, m->m_type); 506 if (mn == NULL) { 507 m_freem(m); 508 return (NULL); 509 } 510 if (m->m_flags & M_PKTHDR) 511 m_move_pkthdr(mn, m); 512 mn->m_next = m; 513 m = mn; 514 if (len < M_SIZE(m)) 515 M_ALIGN(m, len); 516 m->m_len = len; 517 return (m); 518 } 519 520 /* 521 * Make a copy of an mbuf chain starting "off0" bytes from the beginning, 522 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf. 523 * The wait parameter is a choice of M_WAITOK/M_NOWAIT from caller. 524 * Note that the copy is read-only, because clusters are not copied, 525 * only their reference counts are incremented. 526 */ 527 struct mbuf * 528 m_copym(struct mbuf *m, int off0, int len, int wait) 529 { 530 struct mbuf *n, **np; 531 int off = off0; 532 struct mbuf *top; 533 int copyhdr = 0; 534 535 KASSERT(off >= 0, ("m_copym, negative off %d", off)); 536 KASSERT(len >= 0, ("m_copym, negative len %d", len)); 537 MBUF_CHECKSLEEP(wait); 538 if (off == 0 && m->m_flags & M_PKTHDR) 539 copyhdr = 1; 540 while (off > 0) { 541 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain")); 542 if (off < m->m_len) 543 break; 544 off -= m->m_len; 545 m = m->m_next; 546 } 547 np = ⊤ 548 top = NULL; 549 while (len > 0) { 550 if (m == NULL) { 551 KASSERT(len == M_COPYALL, 552 ("m_copym, length > size of mbuf chain")); 553 break; 554 } 555 if (copyhdr) 556 n = m_gethdr(wait, m->m_type); 557 else 558 n = m_get(wait, m->m_type); 559 *np = n; 560 if (n == NULL) 561 goto nospace; 562 if (copyhdr) { 563 if (!m_dup_pkthdr(n, m, wait)) 564 goto nospace; 565 if (len == M_COPYALL) 566 n->m_pkthdr.len -= off0; 567 else 568 n->m_pkthdr.len = len; 569 copyhdr = 0; 570 } 571 n->m_len = min(len, m->m_len - off); 572 if (m->m_flags & (M_EXT | M_EXTPG)) { 573 n->m_data = m->m_data + off; 574 mb_dupcl(n, m); 575 } else 576 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), 577 (u_int)n->m_len); 578 if (len != M_COPYALL) 579 len -= n->m_len; 580 off = 0; 581 m = m->m_next; 582 np = &n->m_next; 583 } 584 585 return (top); 586 nospace: 587 m_freem(top); 588 return (NULL); 589 } 590 591 /* 592 * Copy an entire packet, including header (which must be present). 593 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'. 594 * Note that the copy is read-only, because clusters are not copied, 595 * only their reference counts are incremented. 596 * Preserve alignment of the first mbuf so if the creator has left 597 * some room at the beginning (e.g. for inserting protocol headers) 598 * the copies still have the room available. 599 */ 600 struct mbuf * 601 m_copypacket(struct mbuf *m, int how) 602 { 603 struct mbuf *top, *n, *o; 604 605 MBUF_CHECKSLEEP(how); 606 n = m_get(how, m->m_type); 607 top = n; 608 if (n == NULL) 609 goto nospace; 610 611 if (!m_dup_pkthdr(n, m, how)) 612 goto nospace; 613 n->m_len = m->m_len; 614 if (m->m_flags & (M_EXT | M_EXTPG)) { 615 n->m_data = m->m_data; 616 mb_dupcl(n, m); 617 } else { 618 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat ); 619 bcopy(mtod(m, char *), mtod(n, char *), n->m_len); 620 } 621 622 m = m->m_next; 623 while (m) { 624 o = m_get(how, m->m_type); 625 if (o == NULL) 626 goto nospace; 627 628 n->m_next = o; 629 n = n->m_next; 630 631 n->m_len = m->m_len; 632 if (m->m_flags & (M_EXT | M_EXTPG)) { 633 n->m_data = m->m_data; 634 mb_dupcl(n, m); 635 } else { 636 bcopy(mtod(m, char *), mtod(n, char *), n->m_len); 637 } 638 639 m = m->m_next; 640 } 641 return top; 642 nospace: 643 m_freem(top); 644 return (NULL); 645 } 646 647 static void 648 m_copyfromunmapped(const struct mbuf *m, int off, int len, caddr_t cp) 649 { 650 struct iovec iov; 651 struct uio uio; 652 int error __diagused; 653 654 KASSERT(off >= 0, ("m_copyfromunmapped: negative off %d", off)); 655 KASSERT(len >= 0, ("m_copyfromunmapped: negative len %d", len)); 656 KASSERT(off < m->m_len, 657 ("m_copyfromunmapped: len exceeds mbuf length")); 658 iov.iov_base = cp; 659 iov.iov_len = len; 660 uio.uio_resid = len; 661 uio.uio_iov = &iov; 662 uio.uio_segflg = UIO_SYSSPACE; 663 uio.uio_iovcnt = 1; 664 uio.uio_offset = 0; 665 uio.uio_rw = UIO_READ; 666 error = m_unmapped_uiomove(m, off, &uio, len); 667 KASSERT(error == 0, ("m_unmapped_uiomove failed: off %d, len %d", off, 668 len)); 669 } 670 671 /* 672 * Copy data from an mbuf chain starting "off" bytes from the beginning, 673 * continuing for "len" bytes, into the indicated buffer. 674 */ 675 void 676 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp) 677 { 678 u_int count; 679 680 KASSERT(off >= 0, ("m_copydata, negative off %d", off)); 681 KASSERT(len >= 0, ("m_copydata, negative len %d", len)); 682 while (off > 0) { 683 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain")); 684 if (off < m->m_len) 685 break; 686 off -= m->m_len; 687 m = m->m_next; 688 } 689 while (len > 0) { 690 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain")); 691 count = min(m->m_len - off, len); 692 if ((m->m_flags & M_EXTPG) != 0) 693 m_copyfromunmapped(m, off, count, cp); 694 else 695 bcopy(mtod(m, caddr_t) + off, cp, count); 696 len -= count; 697 cp += count; 698 off = 0; 699 m = m->m_next; 700 } 701 } 702 703 /* 704 * Copy a packet header mbuf chain into a completely new chain, including 705 * copying any mbuf clusters. Use this instead of m_copypacket() when 706 * you need a writable copy of an mbuf chain. 707 */ 708 struct mbuf * 709 m_dup(const struct mbuf *m, int how) 710 { 711 struct mbuf **p, *top = NULL; 712 int remain, moff, nsize; 713 714 MBUF_CHECKSLEEP(how); 715 /* Sanity check */ 716 if (m == NULL) 717 return (NULL); 718 M_ASSERTPKTHDR(m); 719 720 /* While there's more data, get a new mbuf, tack it on, and fill it */ 721 remain = m->m_pkthdr.len; 722 moff = 0; 723 p = ⊤ 724 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */ 725 struct mbuf *n; 726 727 /* Get the next new mbuf */ 728 if (remain >= MINCLSIZE) { 729 n = m_getcl(how, m->m_type, 0); 730 nsize = MCLBYTES; 731 } else { 732 n = m_get(how, m->m_type); 733 nsize = MLEN; 734 } 735 if (n == NULL) 736 goto nospace; 737 738 if (top == NULL) { /* First one, must be PKTHDR */ 739 if (!m_dup_pkthdr(n, m, how)) { 740 m_free(n); 741 goto nospace; 742 } 743 if ((n->m_flags & M_EXT) == 0) 744 nsize = MHLEN; 745 n->m_flags &= ~M_RDONLY; 746 } 747 n->m_len = 0; 748 749 /* Link it into the new chain */ 750 *p = n; 751 p = &n->m_next; 752 753 /* Copy data from original mbuf(s) into new mbuf */ 754 while (n->m_len < nsize && m != NULL) { 755 int chunk = min(nsize - n->m_len, m->m_len - moff); 756 757 m_copydata(m, moff, chunk, n->m_data + n->m_len); 758 moff += chunk; 759 n->m_len += chunk; 760 remain -= chunk; 761 if (moff == m->m_len) { 762 m = m->m_next; 763 moff = 0; 764 } 765 } 766 767 /* Check correct total mbuf length */ 768 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL), 769 ("%s: bogus m_pkthdr.len", __func__)); 770 } 771 return (top); 772 773 nospace: 774 m_freem(top); 775 return (NULL); 776 } 777 778 /* 779 * Concatenate mbuf chain n to m. 780 * Both chains must be of the same type (e.g. MT_DATA). 781 * Any m_pkthdr is not updated. 782 */ 783 void 784 m_cat(struct mbuf *m, struct mbuf *n) 785 { 786 while (m->m_next) 787 m = m->m_next; 788 while (n) { 789 if (!M_WRITABLE(m) || 790 (n->m_flags & M_EXTPG) != 0 || 791 M_TRAILINGSPACE(m) < n->m_len) { 792 /* just join the two chains */ 793 m->m_next = n; 794 return; 795 } 796 /* splat the data from one into the other */ 797 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len, 798 (u_int)n->m_len); 799 m->m_len += n->m_len; 800 n = m_free(n); 801 } 802 } 803 804 /* 805 * Concatenate two pkthdr mbuf chains. 806 */ 807 void 808 m_catpkt(struct mbuf *m, struct mbuf *n) 809 { 810 811 M_ASSERTPKTHDR(m); 812 M_ASSERTPKTHDR(n); 813 814 m->m_pkthdr.len += n->m_pkthdr.len; 815 m_demote(n, 1, 0); 816 817 m_cat(m, n); 818 } 819 820 void 821 m_adj(struct mbuf *mp, int req_len) 822 { 823 int len = req_len; 824 struct mbuf *m; 825 int count; 826 827 if ((m = mp) == NULL) 828 return; 829 if (len >= 0) { 830 /* 831 * Trim from head. 832 */ 833 while (m != NULL && len > 0) { 834 if (m->m_len <= len) { 835 len -= m->m_len; 836 m->m_len = 0; 837 m = m->m_next; 838 } else { 839 m->m_len -= len; 840 m->m_data += len; 841 len = 0; 842 } 843 } 844 if (mp->m_flags & M_PKTHDR) 845 mp->m_pkthdr.len -= (req_len - len); 846 } else { 847 /* 848 * Trim from tail. Scan the mbuf chain, 849 * calculating its length and finding the last mbuf. 850 * If the adjustment only affects this mbuf, then just 851 * adjust and return. Otherwise, rescan and truncate 852 * after the remaining size. 853 */ 854 len = -len; 855 count = 0; 856 for (;;) { 857 count += m->m_len; 858 if (m->m_next == (struct mbuf *)0) 859 break; 860 m = m->m_next; 861 } 862 if (m->m_len >= len) { 863 m->m_len -= len; 864 if (mp->m_flags & M_PKTHDR) 865 mp->m_pkthdr.len -= len; 866 return; 867 } 868 count -= len; 869 if (count < 0) 870 count = 0; 871 /* 872 * Correct length for chain is "count". 873 * Find the mbuf with last data, adjust its length, 874 * and toss data from remaining mbufs on chain. 875 */ 876 m = mp; 877 if (m->m_flags & M_PKTHDR) 878 m->m_pkthdr.len = count; 879 for (; m; m = m->m_next) { 880 if (m->m_len >= count) { 881 m->m_len = count; 882 if (m->m_next != NULL) { 883 m_freem(m->m_next); 884 m->m_next = NULL; 885 } 886 break; 887 } 888 count -= m->m_len; 889 } 890 } 891 } 892 893 void 894 m_adj_decap(struct mbuf *mp, int len) 895 { 896 uint8_t rsstype; 897 898 m_adj(mp, len); 899 if ((mp->m_flags & M_PKTHDR) != 0) { 900 /* 901 * If flowid was calculated by card from the inner 902 * headers, move flowid to the decapsulated mbuf 903 * chain, otherwise clear. This depends on the 904 * internals of m_adj, which keeps pkthdr as is, in 905 * particular not changing rsstype and flowid. 906 */ 907 rsstype = mp->m_pkthdr.rsstype; 908 if ((rsstype & M_HASHTYPE_INNER) != 0) { 909 M_HASHTYPE_SET(mp, rsstype & ~M_HASHTYPE_INNER); 910 } else { 911 M_HASHTYPE_CLEAR(mp); 912 } 913 } 914 } 915 916 /* 917 * Rearange an mbuf chain so that len bytes are contiguous 918 * and in the data area of an mbuf (so that mtod will work 919 * for a structure of size len). Returns the resulting 920 * mbuf chain on success, frees it and returns null on failure. 921 * If there is room, it will add up to max_protohdr-len extra bytes to the 922 * contiguous region in an attempt to avoid being called next time. 923 */ 924 struct mbuf * 925 m_pullup(struct mbuf *n, int len) 926 { 927 struct mbuf *m; 928 int count; 929 int space; 930 931 KASSERT((n->m_flags & M_EXTPG) == 0, 932 ("%s: unmapped mbuf %p", __func__, n)); 933 934 /* 935 * If first mbuf has no cluster, and has room for len bytes 936 * without shifting current data, pullup into it, 937 * otherwise allocate a new mbuf to prepend to the chain. 938 */ 939 if ((n->m_flags & M_EXT) == 0 && 940 n->m_data + len < &n->m_dat[MLEN] && n->m_next) { 941 if (n->m_len >= len) 942 return (n); 943 m = n; 944 n = n->m_next; 945 len -= m->m_len; 946 } else { 947 if (len > MHLEN) 948 goto bad; 949 m = m_get(M_NOWAIT, n->m_type); 950 if (m == NULL) 951 goto bad; 952 if (n->m_flags & M_PKTHDR) 953 m_move_pkthdr(m, n); 954 } 955 space = &m->m_dat[MLEN] - (m->m_data + m->m_len); 956 do { 957 count = min(min(max(len, max_protohdr), space), n->m_len); 958 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len, 959 (u_int)count); 960 len -= count; 961 m->m_len += count; 962 n->m_len -= count; 963 space -= count; 964 if (n->m_len) 965 n->m_data += count; 966 else 967 n = m_free(n); 968 } while (len > 0 && n); 969 if (len > 0) { 970 (void) m_free(m); 971 goto bad; 972 } 973 m->m_next = n; 974 return (m); 975 bad: 976 m_freem(n); 977 return (NULL); 978 } 979 980 /* 981 * Like m_pullup(), except a new mbuf is always allocated, and we allow 982 * the amount of empty space before the data in the new mbuf to be specified 983 * (in the event that the caller expects to prepend later). 984 */ 985 struct mbuf * 986 m_copyup(struct mbuf *n, int len, int dstoff) 987 { 988 struct mbuf *m; 989 int count, space; 990 991 if (len > (MHLEN - dstoff)) 992 goto bad; 993 m = m_get(M_NOWAIT, n->m_type); 994 if (m == NULL) 995 goto bad; 996 if (n->m_flags & M_PKTHDR) 997 m_move_pkthdr(m, n); 998 m->m_data += dstoff; 999 space = &m->m_dat[MLEN] - (m->m_data + m->m_len); 1000 do { 1001 count = min(min(max(len, max_protohdr), space), n->m_len); 1002 memcpy(mtod(m, caddr_t) + m->m_len, mtod(n, caddr_t), 1003 (unsigned)count); 1004 len -= count; 1005 m->m_len += count; 1006 n->m_len -= count; 1007 space -= count; 1008 if (n->m_len) 1009 n->m_data += count; 1010 else 1011 n = m_free(n); 1012 } while (len > 0 && n); 1013 if (len > 0) { 1014 (void) m_free(m); 1015 goto bad; 1016 } 1017 m->m_next = n; 1018 return (m); 1019 bad: 1020 m_freem(n); 1021 return (NULL); 1022 } 1023 1024 /* 1025 * Partition an mbuf chain in two pieces, returning the tail -- 1026 * all but the first len0 bytes. In case of failure, it returns NULL and 1027 * attempts to restore the chain to its original state. 1028 * 1029 * Note that the resulting mbufs might be read-only, because the new 1030 * mbuf can end up sharing an mbuf cluster with the original mbuf if 1031 * the "breaking point" happens to lie within a cluster mbuf. Use the 1032 * M_WRITABLE() macro to check for this case. 1033 */ 1034 struct mbuf * 1035 m_split(struct mbuf *m0, int len0, int wait) 1036 { 1037 struct mbuf *m, *n; 1038 u_int len = len0, remain; 1039 1040 MBUF_CHECKSLEEP(wait); 1041 for (m = m0; m && len > m->m_len; m = m->m_next) 1042 len -= m->m_len; 1043 if (m == NULL) 1044 return (NULL); 1045 remain = m->m_len - len; 1046 if (m0->m_flags & M_PKTHDR && remain == 0) { 1047 n = m_gethdr(wait, m0->m_type); 1048 if (n == NULL) 1049 return (NULL); 1050 n->m_next = m->m_next; 1051 m->m_next = NULL; 1052 if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG) { 1053 n->m_pkthdr.snd_tag = 1054 m_snd_tag_ref(m0->m_pkthdr.snd_tag); 1055 n->m_pkthdr.csum_flags |= CSUM_SND_TAG; 1056 } else 1057 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif; 1058 n->m_pkthdr.len = m0->m_pkthdr.len - len0; 1059 m0->m_pkthdr.len = len0; 1060 return (n); 1061 } else if (m0->m_flags & M_PKTHDR) { 1062 n = m_gethdr(wait, m0->m_type); 1063 if (n == NULL) 1064 return (NULL); 1065 if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG) { 1066 n->m_pkthdr.snd_tag = 1067 m_snd_tag_ref(m0->m_pkthdr.snd_tag); 1068 n->m_pkthdr.csum_flags |= CSUM_SND_TAG; 1069 } else 1070 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif; 1071 n->m_pkthdr.len = m0->m_pkthdr.len - len0; 1072 m0->m_pkthdr.len = len0; 1073 if (m->m_flags & (M_EXT | M_EXTPG)) 1074 goto extpacket; 1075 if (remain > MHLEN) { 1076 /* m can't be the lead packet */ 1077 M_ALIGN(n, 0); 1078 n->m_next = m_split(m, len, wait); 1079 if (n->m_next == NULL) { 1080 (void) m_free(n); 1081 return (NULL); 1082 } else { 1083 n->m_len = 0; 1084 return (n); 1085 } 1086 } else 1087 M_ALIGN(n, remain); 1088 } else if (remain == 0) { 1089 n = m->m_next; 1090 m->m_next = NULL; 1091 return (n); 1092 } else { 1093 n = m_get(wait, m->m_type); 1094 if (n == NULL) 1095 return (NULL); 1096 M_ALIGN(n, remain); 1097 } 1098 extpacket: 1099 if (m->m_flags & (M_EXT | M_EXTPG)) { 1100 n->m_data = m->m_data + len; 1101 mb_dupcl(n, m); 1102 } else { 1103 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain); 1104 } 1105 n->m_len = remain; 1106 m->m_len = len; 1107 n->m_next = m->m_next; 1108 m->m_next = NULL; 1109 return (n); 1110 } 1111 /* 1112 * Routine to copy from device local memory into mbufs. 1113 * Note that `off' argument is offset into first mbuf of target chain from 1114 * which to begin copying the data to. 1115 */ 1116 struct mbuf * 1117 m_devget(char *buf, int totlen, int off, struct ifnet *ifp, 1118 void (*copy)(char *from, caddr_t to, u_int len)) 1119 { 1120 struct mbuf *m; 1121 struct mbuf *top = NULL, **mp = ⊤ 1122 int len; 1123 1124 if (off < 0 || off > MHLEN) 1125 return (NULL); 1126 1127 while (totlen > 0) { 1128 if (top == NULL) { /* First one, must be PKTHDR */ 1129 if (totlen + off >= MINCLSIZE) { 1130 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 1131 len = MCLBYTES; 1132 } else { 1133 m = m_gethdr(M_NOWAIT, MT_DATA); 1134 len = MHLEN; 1135 1136 /* Place initial small packet/header at end of mbuf */ 1137 if (m && totlen + off + max_linkhdr <= MHLEN) { 1138 m->m_data += max_linkhdr; 1139 len -= max_linkhdr; 1140 } 1141 } 1142 if (m == NULL) 1143 return NULL; 1144 m->m_pkthdr.rcvif = ifp; 1145 m->m_pkthdr.len = totlen; 1146 } else { 1147 if (totlen + off >= MINCLSIZE) { 1148 m = m_getcl(M_NOWAIT, MT_DATA, 0); 1149 len = MCLBYTES; 1150 } else { 1151 m = m_get(M_NOWAIT, MT_DATA); 1152 len = MLEN; 1153 } 1154 if (m == NULL) { 1155 m_freem(top); 1156 return NULL; 1157 } 1158 } 1159 if (off) { 1160 m->m_data += off; 1161 len -= off; 1162 off = 0; 1163 } 1164 m->m_len = len = min(totlen, len); 1165 if (copy) 1166 copy(buf, mtod(m, caddr_t), (u_int)len); 1167 else 1168 bcopy(buf, mtod(m, caddr_t), (u_int)len); 1169 buf += len; 1170 *mp = m; 1171 mp = &m->m_next; 1172 totlen -= len; 1173 } 1174 return (top); 1175 } 1176 1177 static void 1178 m_copytounmapped(const struct mbuf *m, int off, int len, c_caddr_t cp) 1179 { 1180 struct iovec iov; 1181 struct uio uio; 1182 int error __diagused; 1183 1184 KASSERT(off >= 0, ("m_copytounmapped: negative off %d", off)); 1185 KASSERT(len >= 0, ("m_copytounmapped: negative len %d", len)); 1186 KASSERT(off < m->m_len, ("m_copytounmapped: len exceeds mbuf length")); 1187 iov.iov_base = __DECONST(caddr_t, cp); 1188 iov.iov_len = len; 1189 uio.uio_resid = len; 1190 uio.uio_iov = &iov; 1191 uio.uio_segflg = UIO_SYSSPACE; 1192 uio.uio_iovcnt = 1; 1193 uio.uio_offset = 0; 1194 uio.uio_rw = UIO_WRITE; 1195 error = m_unmapped_uiomove(m, off, &uio, len); 1196 KASSERT(error == 0, ("m_unmapped_uiomove failed: off %d, len %d", off, 1197 len)); 1198 } 1199 1200 /* 1201 * Copy data from a buffer back into the indicated mbuf chain, 1202 * starting "off" bytes from the beginning, extending the mbuf 1203 * chain if necessary. 1204 */ 1205 void 1206 m_copyback(struct mbuf *m0, int off, int len, c_caddr_t cp) 1207 { 1208 int mlen; 1209 struct mbuf *m = m0, *n; 1210 int totlen = 0; 1211 1212 if (m0 == NULL) 1213 return; 1214 while (off > (mlen = m->m_len)) { 1215 off -= mlen; 1216 totlen += mlen; 1217 if (m->m_next == NULL) { 1218 n = m_get(M_NOWAIT, m->m_type); 1219 if (n == NULL) 1220 goto out; 1221 bzero(mtod(n, caddr_t), MLEN); 1222 n->m_len = min(MLEN, len + off); 1223 m->m_next = n; 1224 } 1225 m = m->m_next; 1226 } 1227 while (len > 0) { 1228 if (m->m_next == NULL && (len > m->m_len - off)) { 1229 m->m_len += min(len - (m->m_len - off), 1230 M_TRAILINGSPACE(m)); 1231 } 1232 mlen = min (m->m_len - off, len); 1233 if ((m->m_flags & M_EXTPG) != 0) 1234 m_copytounmapped(m, off, mlen, cp); 1235 else 1236 bcopy(cp, off + mtod(m, caddr_t), (u_int)mlen); 1237 cp += mlen; 1238 len -= mlen; 1239 mlen += off; 1240 off = 0; 1241 totlen += mlen; 1242 if (len == 0) 1243 break; 1244 if (m->m_next == NULL) { 1245 n = m_get(M_NOWAIT, m->m_type); 1246 if (n == NULL) 1247 break; 1248 n->m_len = min(MLEN, len); 1249 m->m_next = n; 1250 } 1251 m = m->m_next; 1252 } 1253 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen)) 1254 m->m_pkthdr.len = totlen; 1255 } 1256 1257 /* 1258 * Append the specified data to the indicated mbuf chain, 1259 * Extend the mbuf chain if the new data does not fit in 1260 * existing space. 1261 * 1262 * Return 1 if able to complete the job; otherwise 0. 1263 */ 1264 int 1265 m_append(struct mbuf *m0, int len, c_caddr_t cp) 1266 { 1267 struct mbuf *m, *n; 1268 int remainder, space; 1269 1270 for (m = m0; m->m_next != NULL; m = m->m_next) 1271 ; 1272 remainder = len; 1273 space = M_TRAILINGSPACE(m); 1274 if (space > 0) { 1275 /* 1276 * Copy into available space. 1277 */ 1278 if (space > remainder) 1279 space = remainder; 1280 bcopy(cp, mtod(m, caddr_t) + m->m_len, space); 1281 m->m_len += space; 1282 cp += space, remainder -= space; 1283 } 1284 while (remainder > 0) { 1285 /* 1286 * Allocate a new mbuf; could check space 1287 * and allocate a cluster instead. 1288 */ 1289 n = m_get(M_NOWAIT, m->m_type); 1290 if (n == NULL) 1291 break; 1292 n->m_len = min(MLEN, remainder); 1293 bcopy(cp, mtod(n, caddr_t), n->m_len); 1294 cp += n->m_len, remainder -= n->m_len; 1295 m->m_next = n; 1296 m = n; 1297 } 1298 if (m0->m_flags & M_PKTHDR) 1299 m0->m_pkthdr.len += len - remainder; 1300 return (remainder == 0); 1301 } 1302 1303 static int 1304 m_apply_extpg_one(struct mbuf *m, int off, int len, 1305 int (*f)(void *, void *, u_int), void *arg) 1306 { 1307 void *p; 1308 u_int i, count, pgoff, pglen; 1309 int rval; 1310 1311 KASSERT(PMAP_HAS_DMAP, 1312 ("m_apply_extpg_one does not support unmapped mbufs")); 1313 off += mtod(m, vm_offset_t); 1314 if (off < m->m_epg_hdrlen) { 1315 count = min(m->m_epg_hdrlen - off, len); 1316 rval = f(arg, m->m_epg_hdr + off, count); 1317 if (rval) 1318 return (rval); 1319 len -= count; 1320 off = 0; 1321 } else 1322 off -= m->m_epg_hdrlen; 1323 pgoff = m->m_epg_1st_off; 1324 for (i = 0; i < m->m_epg_npgs && len > 0; i++) { 1325 pglen = m_epg_pagelen(m, i, pgoff); 1326 if (off < pglen) { 1327 count = min(pglen - off, len); 1328 p = (void *)PHYS_TO_DMAP(m->m_epg_pa[i] + pgoff + off); 1329 rval = f(arg, p, count); 1330 if (rval) 1331 return (rval); 1332 len -= count; 1333 off = 0; 1334 } else 1335 off -= pglen; 1336 pgoff = 0; 1337 } 1338 if (len > 0) { 1339 KASSERT(off < m->m_epg_trllen, 1340 ("m_apply_extpg_one: offset beyond trailer")); 1341 KASSERT(len <= m->m_epg_trllen - off, 1342 ("m_apply_extpg_one: length beyond trailer")); 1343 return (f(arg, m->m_epg_trail + off, len)); 1344 } 1345 return (0); 1346 } 1347 1348 /* Apply function f to the data in a single mbuf. */ 1349 static int 1350 m_apply_one(struct mbuf *m, int off, int len, 1351 int (*f)(void *, void *, u_int), void *arg) 1352 { 1353 if ((m->m_flags & M_EXTPG) != 0) 1354 return (m_apply_extpg_one(m, off, len, f, arg)); 1355 else 1356 return (f(arg, mtod(m, caddr_t) + off, len)); 1357 } 1358 1359 /* 1360 * Apply function f to the data in an mbuf chain starting "off" bytes from 1361 * the beginning, continuing for "len" bytes. 1362 */ 1363 int 1364 m_apply(struct mbuf *m, int off, int len, 1365 int (*f)(void *, void *, u_int), void *arg) 1366 { 1367 u_int count; 1368 int rval; 1369 1370 KASSERT(off >= 0, ("m_apply, negative off %d", off)); 1371 KASSERT(len >= 0, ("m_apply, negative len %d", len)); 1372 while (off > 0) { 1373 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain")); 1374 if (off < m->m_len) 1375 break; 1376 off -= m->m_len; 1377 m = m->m_next; 1378 } 1379 while (len > 0) { 1380 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain")); 1381 count = min(m->m_len - off, len); 1382 rval = m_apply_one(m, off, count, f, arg); 1383 if (rval) 1384 return (rval); 1385 len -= count; 1386 off = 0; 1387 m = m->m_next; 1388 } 1389 return (0); 1390 } 1391 1392 /* 1393 * Return a pointer to mbuf/offset of location in mbuf chain. 1394 */ 1395 struct mbuf * 1396 m_getptr(struct mbuf *m, int loc, int *off) 1397 { 1398 1399 while (loc >= 0) { 1400 /* Normal end of search. */ 1401 if (m->m_len > loc) { 1402 *off = loc; 1403 return (m); 1404 } else { 1405 loc -= m->m_len; 1406 if (m->m_next == NULL) { 1407 if (loc == 0) { 1408 /* Point at the end of valid data. */ 1409 *off = m->m_len; 1410 return (m); 1411 } 1412 return (NULL); 1413 } 1414 m = m->m_next; 1415 } 1416 } 1417 return (NULL); 1418 } 1419 1420 void 1421 m_print(const struct mbuf *m, int maxlen) 1422 { 1423 int len; 1424 int pdata; 1425 const struct mbuf *m2; 1426 1427 if (m == NULL) { 1428 printf("mbuf: %p\n", m); 1429 return; 1430 } 1431 1432 if (m->m_flags & M_PKTHDR) 1433 len = m->m_pkthdr.len; 1434 else 1435 len = -1; 1436 m2 = m; 1437 while (m2 != NULL && (len == -1 || len)) { 1438 pdata = m2->m_len; 1439 if (maxlen != -1 && pdata > maxlen) 1440 pdata = maxlen; 1441 printf("mbuf: %p len: %d, next: %p, %b%s", m2, m2->m_len, 1442 m2->m_next, m2->m_flags, "\20\20freelist\17skipfw" 1443 "\11proto5\10proto4\7proto3\6proto2\5proto1\4rdonly" 1444 "\3eor\2pkthdr\1ext", pdata ? "" : "\n"); 1445 if (pdata) 1446 printf(", %*D\n", pdata, (u_char *)m2->m_data, "-"); 1447 if (len != -1) 1448 len -= m2->m_len; 1449 m2 = m2->m_next; 1450 } 1451 if (len > 0) 1452 printf("%d bytes unaccounted for.\n", len); 1453 return; 1454 } 1455 1456 u_int 1457 m_fixhdr(struct mbuf *m0) 1458 { 1459 u_int len; 1460 1461 len = m_length(m0, NULL); 1462 m0->m_pkthdr.len = len; 1463 return (len); 1464 } 1465 1466 u_int 1467 m_length(struct mbuf *m0, struct mbuf **last) 1468 { 1469 struct mbuf *m; 1470 u_int len; 1471 1472 len = 0; 1473 for (m = m0; m != NULL; m = m->m_next) { 1474 len += m->m_len; 1475 if (m->m_next == NULL) 1476 break; 1477 } 1478 if (last != NULL) 1479 *last = m; 1480 return (len); 1481 } 1482 1483 /* 1484 * Defragment a mbuf chain, returning the shortest possible 1485 * chain of mbufs and clusters. If allocation fails and 1486 * this cannot be completed, NULL will be returned, but 1487 * the passed in chain will be unchanged. Upon success, 1488 * the original chain will be freed, and the new chain 1489 * will be returned. 1490 * 1491 * If a non-packet header is passed in, the original 1492 * mbuf (chain?) will be returned unharmed. 1493 */ 1494 struct mbuf * 1495 m_defrag(struct mbuf *m0, int how) 1496 { 1497 struct mbuf *m_new = NULL, *m_final = NULL; 1498 int progress = 0, length; 1499 1500 MBUF_CHECKSLEEP(how); 1501 if (!(m0->m_flags & M_PKTHDR)) 1502 return (m0); 1503 1504 m_fixhdr(m0); /* Needed sanity check */ 1505 1506 #ifdef MBUF_STRESS_TEST 1507 if (m_defragrandomfailures) { 1508 int temp = arc4random() & 0xff; 1509 if (temp == 0xba) 1510 goto nospace; 1511 } 1512 #endif 1513 1514 if (m0->m_pkthdr.len > MHLEN) 1515 m_final = m_getcl(how, MT_DATA, M_PKTHDR); 1516 else 1517 m_final = m_gethdr(how, MT_DATA); 1518 1519 if (m_final == NULL) 1520 goto nospace; 1521 1522 if (m_dup_pkthdr(m_final, m0, how) == 0) 1523 goto nospace; 1524 1525 m_new = m_final; 1526 1527 while (progress < m0->m_pkthdr.len) { 1528 length = m0->m_pkthdr.len - progress; 1529 if (length > MCLBYTES) 1530 length = MCLBYTES; 1531 1532 if (m_new == NULL) { 1533 if (length > MLEN) 1534 m_new = m_getcl(how, MT_DATA, 0); 1535 else 1536 m_new = m_get(how, MT_DATA); 1537 if (m_new == NULL) 1538 goto nospace; 1539 } 1540 1541 m_copydata(m0, progress, length, mtod(m_new, caddr_t)); 1542 progress += length; 1543 m_new->m_len = length; 1544 if (m_new != m_final) 1545 m_cat(m_final, m_new); 1546 m_new = NULL; 1547 } 1548 #ifdef MBUF_STRESS_TEST 1549 if (m0->m_next == NULL) 1550 m_defraguseless++; 1551 #endif 1552 m_freem(m0); 1553 m0 = m_final; 1554 #ifdef MBUF_STRESS_TEST 1555 m_defragpackets++; 1556 m_defragbytes += m0->m_pkthdr.len; 1557 #endif 1558 return (m0); 1559 nospace: 1560 #ifdef MBUF_STRESS_TEST 1561 m_defragfailure++; 1562 #endif 1563 if (m_final) 1564 m_freem(m_final); 1565 return (NULL); 1566 } 1567 1568 /* 1569 * Return the number of fragments an mbuf will use. This is usually 1570 * used as a proxy for the number of scatter/gather elements needed by 1571 * a DMA engine to access an mbuf. In general mapped mbufs are 1572 * assumed to be backed by physically contiguous buffers that only 1573 * need a single fragment. Unmapped mbufs, on the other hand, can 1574 * span disjoint physical pages. 1575 */ 1576 static int 1577 frags_per_mbuf(struct mbuf *m) 1578 { 1579 int frags; 1580 1581 if ((m->m_flags & M_EXTPG) == 0) 1582 return (1); 1583 1584 /* 1585 * The header and trailer are counted as a single fragment 1586 * each when present. 1587 * 1588 * XXX: This overestimates the number of fragments by assuming 1589 * all the backing physical pages are disjoint. 1590 */ 1591 frags = 0; 1592 if (m->m_epg_hdrlen != 0) 1593 frags++; 1594 frags += m->m_epg_npgs; 1595 if (m->m_epg_trllen != 0) 1596 frags++; 1597 1598 return (frags); 1599 } 1600 1601 /* 1602 * Defragment an mbuf chain, returning at most maxfrags separate 1603 * mbufs+clusters. If this is not possible NULL is returned and 1604 * the original mbuf chain is left in its present (potentially 1605 * modified) state. We use two techniques: collapsing consecutive 1606 * mbufs and replacing consecutive mbufs by a cluster. 1607 * 1608 * NB: this should really be named m_defrag but that name is taken 1609 */ 1610 struct mbuf * 1611 m_collapse(struct mbuf *m0, int how, int maxfrags) 1612 { 1613 struct mbuf *m, *n, *n2, **prev; 1614 u_int curfrags; 1615 1616 /* 1617 * Calculate the current number of frags. 1618 */ 1619 curfrags = 0; 1620 for (m = m0; m != NULL; m = m->m_next) 1621 curfrags += frags_per_mbuf(m); 1622 /* 1623 * First, try to collapse mbufs. Note that we always collapse 1624 * towards the front so we don't need to deal with moving the 1625 * pkthdr. This may be suboptimal if the first mbuf has much 1626 * less data than the following. 1627 */ 1628 m = m0; 1629 again: 1630 for (;;) { 1631 n = m->m_next; 1632 if (n == NULL) 1633 break; 1634 if (M_WRITABLE(m) && 1635 n->m_len < M_TRAILINGSPACE(m)) { 1636 m_copydata(n, 0, n->m_len, 1637 mtod(m, char *) + m->m_len); 1638 m->m_len += n->m_len; 1639 m->m_next = n->m_next; 1640 curfrags -= frags_per_mbuf(n); 1641 m_free(n); 1642 if (curfrags <= maxfrags) 1643 return m0; 1644 } else 1645 m = n; 1646 } 1647 KASSERT(maxfrags > 1, 1648 ("maxfrags %u, but normal collapse failed", maxfrags)); 1649 /* 1650 * Collapse consecutive mbufs to a cluster. 1651 */ 1652 prev = &m0->m_next; /* NB: not the first mbuf */ 1653 while ((n = *prev) != NULL) { 1654 if ((n2 = n->m_next) != NULL && 1655 n->m_len + n2->m_len < MCLBYTES) { 1656 m = m_getcl(how, MT_DATA, 0); 1657 if (m == NULL) 1658 goto bad; 1659 m_copydata(n, 0, n->m_len, mtod(m, char *)); 1660 m_copydata(n2, 0, n2->m_len, 1661 mtod(m, char *) + n->m_len); 1662 m->m_len = n->m_len + n2->m_len; 1663 m->m_next = n2->m_next; 1664 *prev = m; 1665 curfrags += 1; /* For the new cluster */ 1666 curfrags -= frags_per_mbuf(n); 1667 curfrags -= frags_per_mbuf(n2); 1668 m_free(n); 1669 m_free(n2); 1670 if (curfrags <= maxfrags) 1671 return m0; 1672 /* 1673 * Still not there, try the normal collapse 1674 * again before we allocate another cluster. 1675 */ 1676 goto again; 1677 } 1678 prev = &n->m_next; 1679 } 1680 /* 1681 * No place where we can collapse to a cluster; punt. 1682 * This can occur if, for example, you request 2 frags 1683 * but the packet requires that both be clusters (we 1684 * never reallocate the first mbuf to avoid moving the 1685 * packet header). 1686 */ 1687 bad: 1688 return NULL; 1689 } 1690 1691 #ifdef MBUF_STRESS_TEST 1692 1693 /* 1694 * Fragment an mbuf chain. There's no reason you'd ever want to do 1695 * this in normal usage, but it's great for stress testing various 1696 * mbuf consumers. 1697 * 1698 * If fragmentation is not possible, the original chain will be 1699 * returned. 1700 * 1701 * Possible length values: 1702 * 0 no fragmentation will occur 1703 * > 0 each fragment will be of the specified length 1704 * -1 each fragment will be the same random value in length 1705 * -2 each fragment's length will be entirely random 1706 * (Random values range from 1 to 256) 1707 */ 1708 struct mbuf * 1709 m_fragment(struct mbuf *m0, int how, int length) 1710 { 1711 struct mbuf *m_first, *m_last; 1712 int divisor = 255, progress = 0, fraglen; 1713 1714 if (!(m0->m_flags & M_PKTHDR)) 1715 return (m0); 1716 1717 if (length == 0 || length < -2) 1718 return (m0); 1719 if (length > MCLBYTES) 1720 length = MCLBYTES; 1721 if (length < 0 && divisor > MCLBYTES) 1722 divisor = MCLBYTES; 1723 if (length == -1) 1724 length = 1 + (arc4random() % divisor); 1725 if (length > 0) 1726 fraglen = length; 1727 1728 m_fixhdr(m0); /* Needed sanity check */ 1729 1730 m_first = m_getcl(how, MT_DATA, M_PKTHDR); 1731 if (m_first == NULL) 1732 goto nospace; 1733 1734 if (m_dup_pkthdr(m_first, m0, how) == 0) 1735 goto nospace; 1736 1737 m_last = m_first; 1738 1739 while (progress < m0->m_pkthdr.len) { 1740 if (length == -2) 1741 fraglen = 1 + (arc4random() % divisor); 1742 if (fraglen > m0->m_pkthdr.len - progress) 1743 fraglen = m0->m_pkthdr.len - progress; 1744 1745 if (progress != 0) { 1746 struct mbuf *m_new = m_getcl(how, MT_DATA, 0); 1747 if (m_new == NULL) 1748 goto nospace; 1749 1750 m_last->m_next = m_new; 1751 m_last = m_new; 1752 } 1753 1754 m_copydata(m0, progress, fraglen, mtod(m_last, caddr_t)); 1755 progress += fraglen; 1756 m_last->m_len = fraglen; 1757 } 1758 m_freem(m0); 1759 m0 = m_first; 1760 return (m0); 1761 nospace: 1762 if (m_first) 1763 m_freem(m_first); 1764 /* Return the original chain on failure */ 1765 return (m0); 1766 } 1767 1768 #endif 1769 1770 /* 1771 * Free pages from mbuf_ext_pgs, assuming they were allocated via 1772 * vm_page_alloc() and aren't associated with any object. Complement 1773 * to allocator from m_uiotombuf_nomap(). 1774 */ 1775 void 1776 mb_free_mext_pgs(struct mbuf *m) 1777 { 1778 vm_page_t pg; 1779 1780 M_ASSERTEXTPG(m); 1781 for (int i = 0; i < m->m_epg_npgs; i++) { 1782 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]); 1783 vm_page_unwire_noq(pg); 1784 vm_page_free(pg); 1785 } 1786 } 1787 1788 static struct mbuf * 1789 m_uiotombuf_nomap(struct uio *uio, int how, int len, int maxseg, int flags) 1790 { 1791 struct mbuf *m, *mb, *prev; 1792 vm_page_t pg_array[MBUF_PEXT_MAX_PGS]; 1793 int error, length, i, needed; 1794 ssize_t total; 1795 int pflags = malloc2vm_flags(how) | VM_ALLOC_NODUMP | VM_ALLOC_WIRED; 1796 1797 MPASS((flags & M_PKTHDR) == 0); 1798 MPASS((how & M_ZERO) == 0); 1799 1800 /* 1801 * len can be zero or an arbitrary large value bound by 1802 * the total data supplied by the uio. 1803 */ 1804 if (len > 0) 1805 total = MIN(uio->uio_resid, len); 1806 else 1807 total = uio->uio_resid; 1808 1809 if (maxseg == 0) 1810 maxseg = MBUF_PEXT_MAX_PGS * PAGE_SIZE; 1811 1812 /* 1813 * If total is zero, return an empty mbuf. This can occur 1814 * for TLS 1.0 connections which send empty fragments as 1815 * a countermeasure against the known-IV weakness in CBC 1816 * ciphersuites. 1817 */ 1818 if (__predict_false(total == 0)) { 1819 mb = mb_alloc_ext_pgs(how, mb_free_mext_pgs); 1820 if (mb == NULL) 1821 return (NULL); 1822 mb->m_epg_flags = EPG_FLAG_ANON; 1823 return (mb); 1824 } 1825 1826 /* 1827 * Allocate the pages 1828 */ 1829 m = NULL; 1830 while (total > 0) { 1831 mb = mb_alloc_ext_pgs(how, mb_free_mext_pgs); 1832 if (mb == NULL) 1833 goto failed; 1834 if (m == NULL) 1835 m = mb; 1836 else 1837 prev->m_next = mb; 1838 prev = mb; 1839 mb->m_epg_flags = EPG_FLAG_ANON; 1840 needed = length = MIN(maxseg, total); 1841 for (i = 0; needed > 0; i++, needed -= PAGE_SIZE) { 1842 retry_page: 1843 pg_array[i] = vm_page_alloc_noobj(pflags); 1844 if (pg_array[i] == NULL) { 1845 if (how & M_NOWAIT) { 1846 goto failed; 1847 } else { 1848 vm_wait(NULL); 1849 goto retry_page; 1850 } 1851 } 1852 mb->m_epg_pa[i] = VM_PAGE_TO_PHYS(pg_array[i]); 1853 mb->m_epg_npgs++; 1854 } 1855 mb->m_epg_last_len = length - PAGE_SIZE * (mb->m_epg_npgs - 1); 1856 MBUF_EXT_PGS_ASSERT_SANITY(mb); 1857 total -= length; 1858 error = uiomove_fromphys(pg_array, 0, length, uio); 1859 if (error != 0) 1860 goto failed; 1861 mb->m_len = length; 1862 mb->m_ext.ext_size += PAGE_SIZE * mb->m_epg_npgs; 1863 if (flags & M_PKTHDR) 1864 m->m_pkthdr.len += length; 1865 } 1866 return (m); 1867 1868 failed: 1869 m_freem(m); 1870 return (NULL); 1871 } 1872 1873 /* 1874 * Copy the contents of uio into a properly sized mbuf chain. 1875 */ 1876 struct mbuf * 1877 m_uiotombuf(struct uio *uio, int how, int len, int align, int flags) 1878 { 1879 struct mbuf *m, *mb; 1880 int error, length; 1881 ssize_t total; 1882 int progress = 0; 1883 1884 if (flags & M_EXTPG) 1885 return (m_uiotombuf_nomap(uio, how, len, align, flags)); 1886 1887 /* 1888 * len can be zero or an arbitrary large value bound by 1889 * the total data supplied by the uio. 1890 */ 1891 if (len > 0) 1892 total = (uio->uio_resid < len) ? uio->uio_resid : len; 1893 else 1894 total = uio->uio_resid; 1895 1896 /* 1897 * The smallest unit returned by m_getm2() is a single mbuf 1898 * with pkthdr. We can't align past it. 1899 */ 1900 if (align >= MHLEN) 1901 return (NULL); 1902 1903 /* 1904 * Give us the full allocation or nothing. 1905 * If len is zero return the smallest empty mbuf. 1906 */ 1907 m = m_getm2(NULL, max(total + align, 1), how, MT_DATA, flags); 1908 if (m == NULL) 1909 return (NULL); 1910 m->m_data += align; 1911 1912 /* Fill all mbufs with uio data and update header information. */ 1913 for (mb = m; mb != NULL; mb = mb->m_next) { 1914 length = min(M_TRAILINGSPACE(mb), total - progress); 1915 1916 error = uiomove(mtod(mb, void *), length, uio); 1917 if (error) { 1918 m_freem(m); 1919 return (NULL); 1920 } 1921 1922 mb->m_len = length; 1923 progress += length; 1924 if (flags & M_PKTHDR) { 1925 m->m_pkthdr.len += length; 1926 m->m_pkthdr.memlen += MSIZE; 1927 if (mb->m_flags & M_EXT) 1928 m->m_pkthdr.memlen += mb->m_ext.ext_size; 1929 } 1930 } 1931 KASSERT(progress == total, ("%s: progress != total", __func__)); 1932 1933 return (m); 1934 } 1935 1936 /* 1937 * Copy data to/from an unmapped mbuf into a uio limited by len if set. 1938 */ 1939 int 1940 m_unmapped_uiomove(const struct mbuf *m, int m_off, struct uio *uio, int len) 1941 { 1942 vm_page_t pg; 1943 int error, i, off, pglen, pgoff, seglen, segoff; 1944 1945 M_ASSERTEXTPG(m); 1946 error = 0; 1947 1948 /* Skip over any data removed from the front. */ 1949 off = mtod(m, vm_offset_t); 1950 1951 off += m_off; 1952 if (m->m_epg_hdrlen != 0) { 1953 if (off >= m->m_epg_hdrlen) { 1954 off -= m->m_epg_hdrlen; 1955 } else { 1956 seglen = m->m_epg_hdrlen - off; 1957 segoff = off; 1958 seglen = min(seglen, len); 1959 off = 0; 1960 len -= seglen; 1961 error = uiomove(__DECONST(void *, 1962 &m->m_epg_hdr[segoff]), seglen, uio); 1963 } 1964 } 1965 pgoff = m->m_epg_1st_off; 1966 for (i = 0; i < m->m_epg_npgs && error == 0 && len > 0; i++) { 1967 pglen = m_epg_pagelen(m, i, pgoff); 1968 if (off >= pglen) { 1969 off -= pglen; 1970 pgoff = 0; 1971 continue; 1972 } 1973 seglen = pglen - off; 1974 segoff = pgoff + off; 1975 off = 0; 1976 seglen = min(seglen, len); 1977 len -= seglen; 1978 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]); 1979 error = uiomove_fromphys(&pg, segoff, seglen, uio); 1980 pgoff = 0; 1981 }; 1982 if (len != 0 && error == 0) { 1983 KASSERT((off + len) <= m->m_epg_trllen, 1984 ("off + len > trail (%d + %d > %d, m_off = %d)", off, len, 1985 m->m_epg_trllen, m_off)); 1986 error = uiomove(__DECONST(void *, &m->m_epg_trail[off]), 1987 len, uio); 1988 } 1989 return (error); 1990 } 1991 1992 /* 1993 * Copy an mbuf chain into a uio limited by len if set. 1994 */ 1995 int 1996 m_mbuftouio(struct uio *uio, const struct mbuf *m, int len) 1997 { 1998 int error, length, total; 1999 int progress = 0; 2000 2001 if (len > 0) 2002 total = min(uio->uio_resid, len); 2003 else 2004 total = uio->uio_resid; 2005 2006 /* Fill the uio with data from the mbufs. */ 2007 for (; m != NULL; m = m->m_next) { 2008 length = min(m->m_len, total - progress); 2009 2010 if ((m->m_flags & M_EXTPG) != 0) 2011 error = m_unmapped_uiomove(m, 0, uio, length); 2012 else 2013 error = uiomove(mtod(m, void *), length, uio); 2014 if (error) 2015 return (error); 2016 2017 progress += length; 2018 } 2019 2020 return (0); 2021 } 2022 2023 /* 2024 * Create a writable copy of the mbuf chain. While doing this 2025 * we compact the chain with a goal of producing a chain with 2026 * at most two mbufs. The second mbuf in this chain is likely 2027 * to be a cluster. The primary purpose of this work is to create 2028 * a writable packet for encryption, compression, etc. The 2029 * secondary goal is to linearize the data so the data can be 2030 * passed to crypto hardware in the most efficient manner possible. 2031 */ 2032 struct mbuf * 2033 m_unshare(struct mbuf *m0, int how) 2034 { 2035 struct mbuf *m, *mprev; 2036 struct mbuf *n, *mfirst, *mlast; 2037 int len, off; 2038 2039 mprev = NULL; 2040 for (m = m0; m != NULL; m = mprev->m_next) { 2041 /* 2042 * Regular mbufs are ignored unless there's a cluster 2043 * in front of it that we can use to coalesce. We do 2044 * the latter mainly so later clusters can be coalesced 2045 * also w/o having to handle them specially (i.e. convert 2046 * mbuf+cluster -> cluster). This optimization is heavily 2047 * influenced by the assumption that we're running over 2048 * Ethernet where MCLBYTES is large enough that the max 2049 * packet size will permit lots of coalescing into a 2050 * single cluster. This in turn permits efficient 2051 * crypto operations, especially when using hardware. 2052 */ 2053 if ((m->m_flags & M_EXT) == 0) { 2054 if (mprev && (mprev->m_flags & M_EXT) && 2055 m->m_len <= M_TRAILINGSPACE(mprev)) { 2056 /* XXX: this ignores mbuf types */ 2057 memcpy(mtod(mprev, caddr_t) + mprev->m_len, 2058 mtod(m, caddr_t), m->m_len); 2059 mprev->m_len += m->m_len; 2060 mprev->m_next = m->m_next; /* unlink from chain */ 2061 m_free(m); /* reclaim mbuf */ 2062 } else { 2063 mprev = m; 2064 } 2065 continue; 2066 } 2067 /* 2068 * Writable mbufs are left alone (for now). 2069 */ 2070 if (M_WRITABLE(m)) { 2071 mprev = m; 2072 continue; 2073 } 2074 2075 /* 2076 * Not writable, replace with a copy or coalesce with 2077 * the previous mbuf if possible (since we have to copy 2078 * it anyway, we try to reduce the number of mbufs and 2079 * clusters so that future work is easier). 2080 */ 2081 KASSERT(m->m_flags & M_EXT, ("m_flags 0x%x", m->m_flags)); 2082 /* NB: we only coalesce into a cluster or larger */ 2083 if (mprev != NULL && (mprev->m_flags & M_EXT) && 2084 m->m_len <= M_TRAILINGSPACE(mprev)) { 2085 /* XXX: this ignores mbuf types */ 2086 memcpy(mtod(mprev, caddr_t) + mprev->m_len, 2087 mtod(m, caddr_t), m->m_len); 2088 mprev->m_len += m->m_len; 2089 mprev->m_next = m->m_next; /* unlink from chain */ 2090 m_free(m); /* reclaim mbuf */ 2091 continue; 2092 } 2093 2094 /* 2095 * Allocate new space to hold the copy and copy the data. 2096 * We deal with jumbo mbufs (i.e. m_len > MCLBYTES) by 2097 * splitting them into clusters. We could just malloc a 2098 * buffer and make it external but too many device drivers 2099 * don't know how to break up the non-contiguous memory when 2100 * doing DMA. 2101 */ 2102 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS); 2103 if (n == NULL) { 2104 m_freem(m0); 2105 return (NULL); 2106 } 2107 if (m->m_flags & M_PKTHDR) { 2108 KASSERT(mprev == NULL, ("%s: m0 %p, m %p has M_PKTHDR", 2109 __func__, m0, m)); 2110 m_move_pkthdr(n, m); 2111 } 2112 len = m->m_len; 2113 off = 0; 2114 mfirst = n; 2115 mlast = NULL; 2116 for (;;) { 2117 int cc = min(len, MCLBYTES); 2118 memcpy(mtod(n, caddr_t), mtod(m, caddr_t) + off, cc); 2119 n->m_len = cc; 2120 if (mlast != NULL) 2121 mlast->m_next = n; 2122 mlast = n; 2123 #if 0 2124 newipsecstat.ips_clcopied++; 2125 #endif 2126 2127 len -= cc; 2128 if (len <= 0) 2129 break; 2130 off += cc; 2131 2132 n = m_getcl(how, m->m_type, m->m_flags & M_COPYFLAGS); 2133 if (n == NULL) { 2134 m_freem(mfirst); 2135 m_freem(m0); 2136 return (NULL); 2137 } 2138 } 2139 n->m_next = m->m_next; 2140 if (mprev == NULL) 2141 m0 = mfirst; /* new head of chain */ 2142 else 2143 mprev->m_next = mfirst; /* replace old mbuf */ 2144 m_free(m); /* release old mbuf */ 2145 mprev = mfirst; 2146 } 2147 return (m0); 2148 } 2149 2150 #ifdef MBUF_PROFILING 2151 2152 #define MP_BUCKETS 32 /* don't just change this as things may overflow.*/ 2153 struct mbufprofile { 2154 uintmax_t wasted[MP_BUCKETS]; 2155 uintmax_t used[MP_BUCKETS]; 2156 uintmax_t segments[MP_BUCKETS]; 2157 } mbprof; 2158 2159 void 2160 m_profile(struct mbuf *m) 2161 { 2162 int segments = 0; 2163 int used = 0; 2164 int wasted = 0; 2165 2166 while (m) { 2167 segments++; 2168 used += m->m_len; 2169 if (m->m_flags & M_EXT) { 2170 wasted += MHLEN - sizeof(m->m_ext) + 2171 m->m_ext.ext_size - m->m_len; 2172 } else { 2173 if (m->m_flags & M_PKTHDR) 2174 wasted += MHLEN - m->m_len; 2175 else 2176 wasted += MLEN - m->m_len; 2177 } 2178 m = m->m_next; 2179 } 2180 /* be paranoid.. it helps */ 2181 if (segments > MP_BUCKETS - 1) 2182 segments = MP_BUCKETS - 1; 2183 if (used > 100000) 2184 used = 100000; 2185 if (wasted > 100000) 2186 wasted = 100000; 2187 /* store in the appropriate bucket */ 2188 /* don't bother locking. if it's slightly off, so what? */ 2189 mbprof.segments[segments]++; 2190 mbprof.used[fls(used)]++; 2191 mbprof.wasted[fls(wasted)]++; 2192 } 2193 2194 static int 2195 mbprof_handler(SYSCTL_HANDLER_ARGS) 2196 { 2197 char buf[256]; 2198 struct sbuf sb; 2199 int error; 2200 uint64_t *p; 2201 2202 sbuf_new_for_sysctl(&sb, buf, sizeof(buf), req); 2203 2204 p = &mbprof.wasted[0]; 2205 sbuf_printf(&sb, 2206 "wasted:\n" 2207 "%ju %ju %ju %ju %ju %ju %ju %ju " 2208 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 2209 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 2210 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 2211 #ifdef BIG_ARRAY 2212 p = &mbprof.wasted[16]; 2213 sbuf_printf(&sb, 2214 "%ju %ju %ju %ju %ju %ju %ju %ju " 2215 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 2216 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 2217 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 2218 #endif 2219 p = &mbprof.used[0]; 2220 sbuf_printf(&sb, 2221 "used:\n" 2222 "%ju %ju %ju %ju %ju %ju %ju %ju " 2223 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 2224 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 2225 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 2226 #ifdef BIG_ARRAY 2227 p = &mbprof.used[16]; 2228 sbuf_printf(&sb, 2229 "%ju %ju %ju %ju %ju %ju %ju %ju " 2230 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 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 p = &mbprof.segments[0]; 2235 sbuf_printf(&sb, 2236 "segments:\n" 2237 "%ju %ju %ju %ju %ju %ju %ju %ju " 2238 "%ju %ju %ju %ju %ju %ju %ju %ju\n", 2239 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 2240 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 2241 #ifdef BIG_ARRAY 2242 p = &mbprof.segments[16]; 2243 sbuf_printf(&sb, 2244 "%ju %ju %ju %ju %ju %ju %ju %ju " 2245 "%ju %ju %ju %ju %ju %ju %ju %jju", 2246 p[0], p[1], p[2], p[3], p[4], p[5], p[6], p[7], 2247 p[8], p[9], p[10], p[11], p[12], p[13], p[14], p[15]); 2248 #endif 2249 2250 error = sbuf_finish(&sb); 2251 sbuf_delete(&sb); 2252 return (error); 2253 } 2254 2255 static int 2256 mbprof_clr_handler(SYSCTL_HANDLER_ARGS) 2257 { 2258 int clear, error; 2259 2260 clear = 0; 2261 error = sysctl_handle_int(oidp, &clear, 0, req); 2262 if (error || !req->newptr) 2263 return (error); 2264 2265 if (clear) { 2266 bzero(&mbprof, sizeof(mbprof)); 2267 } 2268 2269 return (error); 2270 } 2271 2272 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofile, 2273 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, 2274 mbprof_handler, "A", 2275 "mbuf profiling statistics"); 2276 2277 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbufprofileclr, 2278 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0, 2279 mbprof_clr_handler, "I", 2280 "clear mbuf profiling statistics"); 2281 #endif 2282