1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 2004 Jeffrey M. Hsu. All rights reserved. 5 * Copyright (c) 2004 The DragonFly Project. All rights reserved. 6 * 7 * This code is derived from software contributed to The DragonFly Project 8 * by Jeffrey M. Hsu. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of The DragonFly Project nor the names of its 19 * contributors may be used to endorse or promote products derived 20 * from this software without specific, prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 */ 35 36 /* 37 * Copyright (c) 1982, 1986, 1988, 1991, 1993 38 * The Regents of the University of California. All rights reserved. 39 * 40 * Redistribution and use in source and binary forms, with or without 41 * modification, are permitted provided that the following conditions 42 * are met: 43 * 1. Redistributions of source code must retain the above copyright 44 * notice, this list of conditions and the following disclaimer. 45 * 2. Redistributions in binary form must reproduce the above copyright 46 * notice, this list of conditions and the following disclaimer in the 47 * documentation and/or other materials provided with the distribution. 48 * 3. All advertising materials mentioning features or use of this software 49 * must display the following acknowledgement: 50 * This product includes software developed by the University of 51 * California, Berkeley and its contributors. 52 * 4. Neither the name of the University nor the names of its contributors 53 * may be used to endorse or promote products derived from this software 54 * without specific prior written permission. 55 * 56 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 57 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 58 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 59 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 60 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 61 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 62 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 63 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 64 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 65 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 66 * SUCH DAMAGE. 67 * 68 * @(#)uipc_mbuf.c 8.2 (Berkeley) 1/4/94 69 * $FreeBSD: src/sys/kern/uipc_mbuf.c,v 1.51.2.24 2003/04/15 06:59:29 silby Exp $ 70 */ 71 72 #include "opt_param.h" 73 #include "opt_mbuf_stress_test.h" 74 #include <sys/param.h> 75 #include <sys/systm.h> 76 #include <sys/file.h> 77 #include <sys/malloc.h> 78 #include <sys/mbuf.h> 79 #include <sys/kernel.h> 80 #include <sys/sysctl.h> 81 #include <sys/domain.h> 82 #include <sys/objcache.h> 83 #include <sys/tree.h> 84 #include <sys/protosw.h> 85 #include <sys/uio.h> 86 #include <sys/thread.h> 87 #include <sys/globaldata.h> 88 89 #include <sys/thread2.h> 90 #include <sys/spinlock2.h> 91 92 #include <machine/atomic.h> 93 #include <machine/limits.h> 94 95 #include <vm/vm.h> 96 #include <vm/vm_kern.h> 97 #include <vm/vm_extern.h> 98 99 #ifdef INVARIANTS 100 #include <machine/cpu.h> 101 #endif 102 103 /* 104 * mbuf cluster meta-data 105 */ 106 struct mbcluster { 107 int32_t mcl_refs; 108 void *mcl_data; 109 }; 110 111 /* 112 * mbuf tracking for debugging purposes 113 */ 114 #ifdef MBUF_DEBUG 115 116 static MALLOC_DEFINE(M_MTRACK, "mtrack", "mtrack"); 117 118 struct mbctrack; 119 RB_HEAD(mbuf_rb_tree, mbtrack); 120 RB_PROTOTYPE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *); 121 122 struct mbtrack { 123 RB_ENTRY(mbtrack) rb_node; 124 int trackid; 125 struct mbuf *m; 126 }; 127 128 static int 129 mbtrack_cmp(struct mbtrack *mb1, struct mbtrack *mb2) 130 { 131 if (mb1->m < mb2->m) 132 return(-1); 133 if (mb1->m > mb2->m) 134 return(1); 135 return(0); 136 } 137 138 RB_GENERATE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *, m); 139 140 struct mbuf_rb_tree mbuf_track_root; 141 static struct spinlock mbuf_track_spin = SPINLOCK_INITIALIZER(mbuf_track_spin); 142 143 static void 144 mbuftrack(struct mbuf *m) 145 { 146 struct mbtrack *mbt; 147 148 mbt = kmalloc(sizeof(*mbt), M_MTRACK, M_INTWAIT|M_ZERO); 149 spin_lock(&mbuf_track_spin); 150 mbt->m = m; 151 if (mbuf_rb_tree_RB_INSERT(&mbuf_track_root, mbt)) { 152 spin_unlock(&mbuf_track_spin); 153 panic("mbuftrack: mbuf %p already being tracked", m); 154 } 155 spin_unlock(&mbuf_track_spin); 156 } 157 158 static void 159 mbufuntrack(struct mbuf *m) 160 { 161 struct mbtrack *mbt; 162 163 spin_lock(&mbuf_track_spin); 164 mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m); 165 if (mbt == NULL) { 166 spin_unlock(&mbuf_track_spin); 167 panic("mbufuntrack: mbuf %p was not tracked", m); 168 } else { 169 mbuf_rb_tree_RB_REMOVE(&mbuf_track_root, mbt); 170 spin_unlock(&mbuf_track_spin); 171 kfree(mbt, M_MTRACK); 172 } 173 } 174 175 void 176 mbuftrackid(struct mbuf *m, int trackid) 177 { 178 struct mbtrack *mbt; 179 struct mbuf *n; 180 181 spin_lock(&mbuf_track_spin); 182 while (m) { 183 n = m->m_nextpkt; 184 while (m) { 185 mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m); 186 if (mbt == NULL) { 187 spin_unlock(&mbuf_track_spin); 188 panic("mbuftrackid: mbuf %p not tracked", m); 189 } 190 mbt->trackid = trackid; 191 m = m->m_next; 192 } 193 m = n; 194 } 195 spin_unlock(&mbuf_track_spin); 196 } 197 198 static int 199 mbuftrack_callback(struct mbtrack *mbt, void *arg) 200 { 201 struct sysctl_req *req = arg; 202 char buf[64]; 203 int error; 204 205 ksnprintf(buf, sizeof(buf), "mbuf %p track %d\n", mbt->m, mbt->trackid); 206 207 spin_unlock(&mbuf_track_spin); 208 error = SYSCTL_OUT(req, buf, strlen(buf)); 209 spin_lock(&mbuf_track_spin); 210 if (error) 211 return(-error); 212 return(0); 213 } 214 215 static int 216 mbuftrack_show(SYSCTL_HANDLER_ARGS) 217 { 218 int error; 219 220 spin_lock(&mbuf_track_spin); 221 error = mbuf_rb_tree_RB_SCAN(&mbuf_track_root, NULL, 222 mbuftrack_callback, req); 223 spin_unlock(&mbuf_track_spin); 224 return (-error); 225 } 226 SYSCTL_PROC(_kern_ipc, OID_AUTO, showmbufs, CTLFLAG_RD|CTLTYPE_STRING, 227 0, 0, mbuftrack_show, "A", "Show all in-use mbufs"); 228 229 #else 230 231 #define mbuftrack(m) 232 #define mbufuntrack(m) 233 234 #endif 235 236 static void mbinit(void *); 237 SYSINIT(mbuf, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, mbinit, NULL) 238 239 static u_long mbtypes[SMP_MAXCPU][MT_NTYPES]; 240 241 static struct mbstat mbstat[SMP_MAXCPU]; 242 int max_linkhdr; 243 int max_protohdr; 244 int max_hdr; 245 int max_datalen; 246 int m_defragpackets; 247 int m_defragbytes; 248 int m_defraguseless; 249 int m_defragfailure; 250 #ifdef MBUF_STRESS_TEST 251 int m_defragrandomfailures; 252 #endif 253 254 struct objcache *mbuf_cache, *mbufphdr_cache; 255 struct objcache *mclmeta_cache, *mjclmeta_cache; 256 struct objcache *mbufcluster_cache, *mbufphdrcluster_cache; 257 struct objcache *mbufjcluster_cache, *mbufphdrjcluster_cache; 258 259 int nmbclusters; 260 int nmbufs; 261 262 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW, 263 &max_linkhdr, 0, "Max size of a link-level header"); 264 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW, 265 &max_protohdr, 0, "Max size of a protocol header"); 266 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0, 267 "Max size of link+protocol headers"); 268 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW, 269 &max_datalen, 0, "Max data payload size without headers"); 270 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW, 271 &mbuf_wait, 0, "Time in ticks to sleep after failed mbuf allocations"); 272 static int do_mbstat(SYSCTL_HANDLER_ARGS); 273 274 SYSCTL_PROC(_kern_ipc, KIPC_MBSTAT, mbstat, CTLTYPE_STRUCT|CTLFLAG_RD, 275 0, 0, do_mbstat, "S,mbstat", "mbuf usage statistics"); 276 277 static int do_mbtypes(SYSCTL_HANDLER_ARGS); 278 279 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbtypes, CTLTYPE_ULONG|CTLFLAG_RD, 280 0, 0, do_mbtypes, "LU", ""); 281 282 static int 283 do_mbstat(SYSCTL_HANDLER_ARGS) 284 { 285 struct mbstat mbstat_total; 286 struct mbstat *mbstat_totalp; 287 int i; 288 289 bzero(&mbstat_total, sizeof(mbstat_total)); 290 mbstat_totalp = &mbstat_total; 291 292 for (i = 0; i < ncpus; i++) 293 { 294 mbstat_total.m_mbufs += mbstat[i].m_mbufs; 295 mbstat_total.m_clusters += mbstat[i].m_clusters; 296 mbstat_total.m_spare += mbstat[i].m_spare; 297 mbstat_total.m_clfree += mbstat[i].m_clfree; 298 mbstat_total.m_drops += mbstat[i].m_drops; 299 mbstat_total.m_wait += mbstat[i].m_wait; 300 mbstat_total.m_drain += mbstat[i].m_drain; 301 mbstat_total.m_mcfail += mbstat[i].m_mcfail; 302 mbstat_total.m_mpfail += mbstat[i].m_mpfail; 303 304 } 305 /* 306 * The following fields are not cumulative fields so just 307 * get their values once. 308 */ 309 mbstat_total.m_msize = mbstat[0].m_msize; 310 mbstat_total.m_mclbytes = mbstat[0].m_mclbytes; 311 mbstat_total.m_minclsize = mbstat[0].m_minclsize; 312 mbstat_total.m_mlen = mbstat[0].m_mlen; 313 mbstat_total.m_mhlen = mbstat[0].m_mhlen; 314 315 return(sysctl_handle_opaque(oidp, mbstat_totalp, sizeof(mbstat_total), req)); 316 } 317 318 static int 319 do_mbtypes(SYSCTL_HANDLER_ARGS) 320 { 321 u_long totals[MT_NTYPES]; 322 int i, j; 323 324 for (i = 0; i < MT_NTYPES; i++) 325 totals[i] = 0; 326 327 for (i = 0; i < ncpus; i++) 328 { 329 for (j = 0; j < MT_NTYPES; j++) 330 totals[j] += mbtypes[i][j]; 331 } 332 333 return(sysctl_handle_opaque(oidp, totals, sizeof(totals), req)); 334 } 335 336 /* 337 * These are read-only because we do not currently have any code 338 * to adjust the objcache limits after the fact. The variables 339 * may only be set as boot-time tunables. 340 */ 341 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RD, 342 &nmbclusters, 0, "Maximum number of mbuf clusters available"); 343 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RD, &nmbufs, 0, 344 "Maximum number of mbufs available"); 345 346 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD, 347 &m_defragpackets, 0, "Number of defragment packets"); 348 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD, 349 &m_defragbytes, 0, "Number of defragment bytes"); 350 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD, 351 &m_defraguseless, 0, "Number of useless defragment mbuf chain operations"); 352 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD, 353 &m_defragfailure, 0, "Number of failed defragment mbuf chain operations"); 354 #ifdef MBUF_STRESS_TEST 355 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW, 356 &m_defragrandomfailures, 0, ""); 357 #endif 358 359 static MALLOC_DEFINE(M_MBUF, "mbuf", "mbuf"); 360 static MALLOC_DEFINE(M_MBUFCL, "mbufcl", "mbufcl"); 361 static MALLOC_DEFINE(M_MJBUFCL, "mbufcl", "mbufcl"); 362 static MALLOC_DEFINE(M_MCLMETA, "mclmeta", "mclmeta"); 363 static MALLOC_DEFINE(M_MJCLMETA, "mjclmeta", "mjclmeta"); 364 365 static void m_reclaim (void); 366 static void m_mclref(void *arg); 367 static void m_mclfree(void *arg); 368 369 /* 370 * NOTE: Default NMBUFS must take into account a possible DOS attack 371 * using fd passing on unix domain sockets. 372 */ 373 #ifndef NMBCLUSTERS 374 #define NMBCLUSTERS (512 + maxusers * 16) 375 #endif 376 #ifndef NMBUFS 377 #define NMBUFS (nmbclusters * 2 + maxfiles) 378 #endif 379 380 /* 381 * Perform sanity checks of tunables declared above. 382 */ 383 static void 384 tunable_mbinit(void *dummy) 385 { 386 /* 387 * This has to be done before VM init. 388 */ 389 nmbclusters = NMBCLUSTERS; 390 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters); 391 nmbufs = NMBUFS; 392 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs); 393 /* Sanity checks */ 394 if (nmbufs < nmbclusters * 2) 395 nmbufs = nmbclusters * 2; 396 } 397 SYSINIT(tunable_mbinit, SI_BOOT1_TUNABLES, SI_ORDER_ANY, 398 tunable_mbinit, NULL); 399 400 /* "number of clusters of pages" */ 401 #define NCL_INIT 1 402 403 #define NMB_INIT 16 404 405 /* 406 * The mbuf object cache only guarantees that m_next and m_nextpkt are 407 * NULL and that m_data points to the beginning of the data area. In 408 * particular, m_len and m_pkthdr.len are uninitialized. It is the 409 * responsibility of the caller to initialize those fields before use. 410 */ 411 412 static __inline boolean_t 413 mbuf_ctor(void *obj, void *private, int ocflags) 414 { 415 struct mbuf *m = obj; 416 417 m->m_next = NULL; 418 m->m_nextpkt = NULL; 419 m->m_data = m->m_dat; 420 m->m_flags = 0; 421 422 return (TRUE); 423 } 424 425 /* 426 * Initialize the mbuf and the packet header fields. 427 */ 428 static boolean_t 429 mbufphdr_ctor(void *obj, void *private, int ocflags) 430 { 431 struct mbuf *m = obj; 432 433 m->m_next = NULL; 434 m->m_nextpkt = NULL; 435 m->m_data = m->m_pktdat; 436 m->m_flags = M_PKTHDR | M_PHCACHE; 437 438 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */ 439 SLIST_INIT(&m->m_pkthdr.tags); 440 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */ 441 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */ 442 443 return (TRUE); 444 } 445 446 /* 447 * A mbcluster object consists of 2K (MCLBYTES) cluster and a refcount. 448 */ 449 static boolean_t 450 mclmeta_ctor(void *obj, void *private, int ocflags) 451 { 452 struct mbcluster *cl = obj; 453 void *buf; 454 455 if (ocflags & M_NOWAIT) 456 buf = kmalloc(MCLBYTES, M_MBUFCL, M_NOWAIT | M_ZERO); 457 else 458 buf = kmalloc(MCLBYTES, M_MBUFCL, M_INTWAIT | M_ZERO); 459 if (buf == NULL) 460 return (FALSE); 461 cl->mcl_refs = 0; 462 cl->mcl_data = buf; 463 return (TRUE); 464 } 465 466 static boolean_t 467 mjclmeta_ctor(void *obj, void *private, int ocflags) 468 { 469 struct mbcluster *cl = obj; 470 void *buf; 471 472 if (ocflags & M_NOWAIT) 473 buf = kmalloc(MJUMPAGESIZE, M_MBUFCL, M_NOWAIT | M_ZERO); 474 else 475 buf = kmalloc(MJUMPAGESIZE, M_MBUFCL, M_INTWAIT | M_ZERO); 476 if (buf == NULL) 477 return (FALSE); 478 cl->mcl_refs = 0; 479 cl->mcl_data = buf; 480 return (TRUE); 481 } 482 483 static void 484 mclmeta_dtor(void *obj, void *private) 485 { 486 struct mbcluster *mcl = obj; 487 488 KKASSERT(mcl->mcl_refs == 0); 489 kfree(mcl->mcl_data, M_MBUFCL); 490 } 491 492 static void 493 linkjcluster(struct mbuf *m, struct mbcluster *cl, uint size) 494 { 495 /* 496 * Add the cluster to the mbuf. The caller will detect that the 497 * mbuf now has an attached cluster. 498 */ 499 m->m_ext.ext_arg = cl; 500 m->m_ext.ext_buf = cl->mcl_data; 501 m->m_ext.ext_ref = m_mclref; 502 m->m_ext.ext_free = m_mclfree; 503 m->m_ext.ext_size = size; 504 atomic_add_int(&cl->mcl_refs, 1); 505 506 m->m_data = m->m_ext.ext_buf; 507 m->m_flags |= M_EXT | M_EXT_CLUSTER; 508 } 509 510 static void 511 linkcluster(struct mbuf *m, struct mbcluster *cl) 512 { 513 linkjcluster(m, cl, MCLBYTES); 514 } 515 516 static boolean_t 517 mbufphdrcluster_ctor(void *obj, void *private, int ocflags) 518 { 519 struct mbuf *m = obj; 520 struct mbcluster *cl; 521 522 mbufphdr_ctor(obj, private, ocflags); 523 cl = objcache_get(mclmeta_cache, ocflags); 524 if (cl == NULL) { 525 ++mbstat[mycpu->gd_cpuid].m_drops; 526 return (FALSE); 527 } 528 m->m_flags |= M_CLCACHE; 529 linkcluster(m, cl); 530 return (TRUE); 531 } 532 533 static boolean_t 534 mbufphdrjcluster_ctor(void *obj, void *private, int ocflags) 535 { 536 struct mbuf *m = obj; 537 struct mbcluster *cl; 538 539 mbufphdr_ctor(obj, private, ocflags); 540 cl = objcache_get(mjclmeta_cache, ocflags); 541 if (cl == NULL) { 542 ++mbstat[mycpu->gd_cpuid].m_drops; 543 return (FALSE); 544 } 545 m->m_flags |= M_CLCACHE; 546 linkjcluster(m, cl, MJUMPAGESIZE); 547 return (TRUE); 548 } 549 550 static boolean_t 551 mbufcluster_ctor(void *obj, void *private, int ocflags) 552 { 553 struct mbuf *m = obj; 554 struct mbcluster *cl; 555 556 mbuf_ctor(obj, private, ocflags); 557 cl = objcache_get(mclmeta_cache, ocflags); 558 if (cl == NULL) { 559 ++mbstat[mycpu->gd_cpuid].m_drops; 560 return (FALSE); 561 } 562 m->m_flags |= M_CLCACHE; 563 linkcluster(m, cl); 564 return (TRUE); 565 } 566 567 static boolean_t 568 mbufjcluster_ctor(void *obj, void *private, int ocflags) 569 { 570 struct mbuf *m = obj; 571 struct mbcluster *cl; 572 573 mbuf_ctor(obj, private, ocflags); 574 cl = objcache_get(mjclmeta_cache, ocflags); 575 if (cl == NULL) { 576 ++mbstat[mycpu->gd_cpuid].m_drops; 577 return (FALSE); 578 } 579 m->m_flags |= M_CLCACHE; 580 linkjcluster(m, cl, MJUMPAGESIZE); 581 return (TRUE); 582 } 583 584 /* 585 * Used for both the cluster and cluster PHDR caches. 586 * 587 * The mbuf may have lost its cluster due to sharing, deal 588 * with the situation by checking M_EXT. 589 */ 590 static void 591 mbufcluster_dtor(void *obj, void *private) 592 { 593 struct mbuf *m = obj; 594 struct mbcluster *mcl; 595 596 if (m->m_flags & M_EXT) { 597 KKASSERT((m->m_flags & M_EXT_CLUSTER) != 0); 598 mcl = m->m_ext.ext_arg; 599 KKASSERT(mcl->mcl_refs == 1); 600 mcl->mcl_refs = 0; 601 if (m->m_flags & M_EXT && m->m_ext.ext_size != MCLBYTES) 602 objcache_put(mjclmeta_cache, mcl); 603 else 604 objcache_put(mclmeta_cache, mcl); 605 } 606 } 607 608 struct objcache_malloc_args mbuf_malloc_args = { MSIZE, M_MBUF }; 609 struct objcache_malloc_args mclmeta_malloc_args = 610 { sizeof(struct mbcluster), M_MCLMETA }; 611 612 /* ARGSUSED*/ 613 static void 614 mbinit(void *dummy) 615 { 616 int mb_limit, cl_limit; 617 int limit; 618 int i; 619 620 /* 621 * Initialize statistics 622 */ 623 for (i = 0; i < ncpus; i++) { 624 mbstat[i].m_msize = MSIZE; 625 mbstat[i].m_mclbytes = MCLBYTES; 626 mbstat[i].m_mjumpagesize = MJUMPAGESIZE; 627 mbstat[i].m_minclsize = MINCLSIZE; 628 mbstat[i].m_mlen = MLEN; 629 mbstat[i].m_mhlen = MHLEN; 630 } 631 632 /* 633 * Create objtect caches and save cluster limits, which will 634 * be used to adjust backing kmalloc pools' limit later. 635 */ 636 637 mb_limit = cl_limit = 0; 638 639 limit = nmbufs; 640 mbuf_cache = objcache_create("mbuf", 641 &limit, 0, 642 mbuf_ctor, NULL, NULL, 643 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 644 mb_limit += limit; 645 646 limit = nmbufs; 647 mbufphdr_cache = objcache_create("mbuf pkt hdr", 648 &limit, nmbufs / 4, 649 mbufphdr_ctor, NULL, NULL, 650 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 651 mb_limit += limit; 652 653 cl_limit = nmbclusters; 654 mclmeta_cache = objcache_create("cluster mbuf", 655 &cl_limit, 0, 656 mclmeta_ctor, mclmeta_dtor, NULL, 657 objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args); 658 659 cl_limit = nmbclusters; 660 mjclmeta_cache = objcache_create("jcluster mbuf", 661 &cl_limit, 0, 662 mjclmeta_ctor, mclmeta_dtor, NULL, 663 objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args); 664 665 limit = nmbclusters; 666 mbufcluster_cache = objcache_create("mbuf + cluster", 667 &limit, 0, 668 mbufcluster_ctor, mbufcluster_dtor, NULL, 669 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 670 mb_limit += limit; 671 672 limit = nmbclusters; 673 mbufphdrcluster_cache = objcache_create("mbuf pkt hdr + cluster", 674 &limit, nmbclusters / 16, 675 mbufphdrcluster_ctor, mbufcluster_dtor, NULL, 676 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 677 mb_limit += limit; 678 679 limit = nmbclusters; 680 mbufjcluster_cache = objcache_create("mbuf + jcluster", 681 &limit, 0, 682 mbufjcluster_ctor, mbufcluster_dtor, NULL, 683 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 684 685 limit = nmbclusters; 686 mbufphdrjcluster_cache = objcache_create("mbuf pkt hdr + jcluster", 687 &limit, nmbclusters / 16, 688 mbufphdrjcluster_ctor, mbufcluster_dtor, NULL, 689 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 690 691 /* 692 * Adjust backing kmalloc pools' limit 693 * 694 * NOTE: We raise the limit by another 1/8 to take the effect 695 * of loosememuse into account. 696 */ 697 cl_limit += cl_limit / 8; 698 kmalloc_raise_limit(mclmeta_malloc_args.mtype, 699 mclmeta_malloc_args.objsize * (size_t)cl_limit); 700 kmalloc_raise_limit(M_MBUFCL, 701 ((MCLBYTES * (size_t)cl_limit * 3) / 4) + 702 ((MJUMPAGESIZE * (size_t)cl_limit) / 4)); 703 704 mb_limit += mb_limit / 8; 705 kmalloc_raise_limit(mbuf_malloc_args.mtype, 706 mbuf_malloc_args.objsize * (size_t)mb_limit); 707 } 708 709 /* 710 * Return the number of references to this mbuf's data. 0 is returned 711 * if the mbuf is not M_EXT, a reference count is returned if it is 712 * M_EXT | M_EXT_CLUSTER, and 99 is returned if it is a special M_EXT. 713 */ 714 int 715 m_sharecount(struct mbuf *m) 716 { 717 switch (m->m_flags & (M_EXT | M_EXT_CLUSTER)) { 718 case 0: 719 return (0); 720 case M_EXT: 721 return (99); 722 case M_EXT | M_EXT_CLUSTER: 723 return (((struct mbcluster *)m->m_ext.ext_arg)->mcl_refs); 724 } 725 /* NOTREACHED */ 726 return (0); /* to shut up compiler */ 727 } 728 729 /* 730 * change mbuf to new type 731 */ 732 void 733 m_chtype(struct mbuf *m, int type) 734 { 735 struct globaldata *gd = mycpu; 736 737 ++mbtypes[gd->gd_cpuid][type]; 738 --mbtypes[gd->gd_cpuid][m->m_type]; 739 m->m_type = type; 740 } 741 742 static void 743 m_reclaim(void) 744 { 745 struct domain *dp; 746 struct protosw *pr; 747 748 kprintf("Debug: m_reclaim() called\n"); 749 750 SLIST_FOREACH(dp, &domains, dom_next) { 751 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) { 752 if (pr->pr_drain) 753 (*pr->pr_drain)(); 754 } 755 } 756 ++mbstat[mycpu->gd_cpuid].m_drain; 757 } 758 759 static __inline void 760 updatestats(struct mbuf *m, int type) 761 { 762 struct globaldata *gd = mycpu; 763 764 m->m_type = type; 765 mbuftrack(m); 766 #ifdef MBUF_DEBUG 767 KASSERT(m->m_next == NULL, ("mbuf %p: bad m_next in get", m)); 768 KASSERT(m->m_nextpkt == NULL, ("mbuf %p: bad m_nextpkt in get", m)); 769 #endif 770 771 ++mbtypes[gd->gd_cpuid][type]; 772 ++mbstat[gd->gd_cpuid].m_mbufs; 773 774 } 775 776 /* 777 * Allocate an mbuf. 778 */ 779 struct mbuf * 780 m_get(int how, int type) 781 { 782 struct mbuf *m; 783 int ntries = 0; 784 int ocf = MBTOM(how); 785 786 retryonce: 787 788 m = objcache_get(mbuf_cache, ocf); 789 790 if (m == NULL) { 791 if ((how & MB_TRYWAIT) && ntries++ == 0) { 792 struct objcache *reclaimlist[] = { 793 mbufphdr_cache, 794 mbufcluster_cache, 795 mbufphdrcluster_cache, 796 mbufjcluster_cache, 797 mbufphdrjcluster_cache 798 }; 799 const int nreclaims = NELEM(reclaimlist); 800 801 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf)) 802 m_reclaim(); 803 goto retryonce; 804 } 805 ++mbstat[mycpu->gd_cpuid].m_drops; 806 return (NULL); 807 } 808 #ifdef MBUF_DEBUG 809 KASSERT(m->m_data == m->m_dat, ("mbuf %p: bad m_data in get", m)); 810 #endif 811 m->m_len = 0; 812 813 updatestats(m, type); 814 return (m); 815 } 816 817 struct mbuf * 818 m_gethdr(int how, int type) 819 { 820 struct mbuf *m; 821 int ocf = MBTOM(how); 822 int ntries = 0; 823 824 retryonce: 825 826 m = objcache_get(mbufphdr_cache, ocf); 827 828 if (m == NULL) { 829 if ((how & MB_TRYWAIT) && ntries++ == 0) { 830 struct objcache *reclaimlist[] = { 831 mbuf_cache, 832 mbufcluster_cache, mbufphdrcluster_cache, 833 mbufjcluster_cache, mbufphdrjcluster_cache 834 }; 835 const int nreclaims = NELEM(reclaimlist); 836 837 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf)) 838 m_reclaim(); 839 goto retryonce; 840 } 841 ++mbstat[mycpu->gd_cpuid].m_drops; 842 return (NULL); 843 } 844 #ifdef MBUF_DEBUG 845 KASSERT(m->m_data == m->m_pktdat, ("mbuf %p: bad m_data in get", m)); 846 #endif 847 m->m_len = 0; 848 m->m_pkthdr.len = 0; 849 850 updatestats(m, type); 851 return (m); 852 } 853 854 /* 855 * Get a mbuf (not a mbuf cluster!) and zero it. 856 * Deprecated. 857 */ 858 struct mbuf * 859 m_getclr(int how, int type) 860 { 861 struct mbuf *m; 862 863 m = m_get(how, type); 864 if (m != NULL) 865 bzero(m->m_data, MLEN); 866 return (m); 867 } 868 869 struct mbuf * 870 m_getjcl(int how, short type, int flags, size_t size) 871 { 872 struct mbuf *m = NULL; 873 struct objcache *mbclc, *mbphclc; 874 int ocflags = MBTOM(how); 875 int ntries = 0; 876 877 switch (size) { 878 case MCLBYTES: 879 mbclc = mbufcluster_cache; 880 mbphclc = mbufphdrcluster_cache; 881 break; 882 default: 883 mbclc = mbufjcluster_cache; 884 mbphclc = mbufphdrjcluster_cache; 885 break; 886 } 887 888 retryonce: 889 890 if (flags & M_PKTHDR) 891 m = objcache_get(mbphclc, ocflags); 892 else 893 m = objcache_get(mbclc, ocflags); 894 895 if (m == NULL) { 896 if ((how & MB_TRYWAIT) && ntries++ == 0) { 897 struct objcache *reclaimlist[1]; 898 899 if (flags & M_PKTHDR) 900 reclaimlist[0] = mbclc; 901 else 902 reclaimlist[0] = mbphclc; 903 if (!objcache_reclaimlist(reclaimlist, 1, ocflags)) 904 m_reclaim(); 905 goto retryonce; 906 } 907 ++mbstat[mycpu->gd_cpuid].m_drops; 908 return (NULL); 909 } 910 911 #ifdef MBUF_DEBUG 912 KASSERT(m->m_data == m->m_ext.ext_buf, 913 ("mbuf %p: bad m_data in get", m)); 914 #endif 915 m->m_type = type; 916 m->m_len = 0; 917 m->m_pkthdr.len = 0; /* just do it unconditonally */ 918 919 mbuftrack(m); 920 921 ++mbtypes[mycpu->gd_cpuid][type]; 922 ++mbstat[mycpu->gd_cpuid].m_clusters; 923 return (m); 924 } 925 926 /* 927 * Returns an mbuf with an attached cluster. 928 * Because many network drivers use this kind of buffers a lot, it is 929 * convenient to keep a small pool of free buffers of this kind. 930 * Even a small size such as 10 gives about 10% improvement in the 931 * forwarding rate in a bridge or router. 932 */ 933 struct mbuf * 934 m_getcl(int how, short type, int flags) 935 { 936 return (m_getjcl(how, type, flags, MCLBYTES)); 937 } 938 939 /* 940 * Allocate chain of requested length. 941 */ 942 struct mbuf * 943 m_getc(int len, int how, int type) 944 { 945 struct mbuf *n, *nfirst = NULL, **ntail = &nfirst; 946 int nsize; 947 948 while (len > 0) { 949 n = m_getl(len, how, type, 0, &nsize); 950 if (n == NULL) 951 goto failed; 952 n->m_len = 0; 953 *ntail = n; 954 ntail = &n->m_next; 955 len -= nsize; 956 } 957 return (nfirst); 958 959 failed: 960 m_freem(nfirst); 961 return (NULL); 962 } 963 964 /* 965 * Allocate len-worth of mbufs and/or mbuf clusters (whatever fits best) 966 * and return a pointer to the head of the allocated chain. If m0 is 967 * non-null, then we assume that it is a single mbuf or an mbuf chain to 968 * which we want len bytes worth of mbufs and/or clusters attached, and so 969 * if we succeed in allocating it, we will just return a pointer to m0. 970 * 971 * If we happen to fail at any point during the allocation, we will free 972 * up everything we have already allocated and return NULL. 973 * 974 * Deprecated. Use m_getc() and m_cat() instead. 975 */ 976 struct mbuf * 977 m_getm(struct mbuf *m0, int len, int type, int how) 978 { 979 struct mbuf *nfirst; 980 981 nfirst = m_getc(len, how, type); 982 983 if (m0 != NULL) { 984 m_last(m0)->m_next = nfirst; 985 return (m0); 986 } 987 988 return (nfirst); 989 } 990 991 /* 992 * Adds a cluster to a normal mbuf, M_EXT is set on success. 993 * Deprecated. Use m_getcl() instead. 994 */ 995 void 996 m_mclget(struct mbuf *m, int how) 997 { 998 struct mbcluster *mcl; 999 1000 KKASSERT((m->m_flags & M_EXT) == 0); 1001 mcl = objcache_get(mclmeta_cache, MBTOM(how)); 1002 if (mcl != NULL) { 1003 linkcluster(m, mcl); 1004 ++mbstat[mycpu->gd_cpuid].m_clusters; 1005 } else { 1006 ++mbstat[mycpu->gd_cpuid].m_drops; 1007 } 1008 } 1009 1010 /* 1011 * Updates to mbcluster must be MPSAFE. Only an entity which already has 1012 * a reference to the cluster can ref it, so we are in no danger of 1013 * racing an add with a subtract. But the operation must still be atomic 1014 * since multiple entities may have a reference on the cluster. 1015 * 1016 * m_mclfree() is almost the same but it must contend with two entities 1017 * freeing the cluster at the same time. 1018 */ 1019 static void 1020 m_mclref(void *arg) 1021 { 1022 struct mbcluster *mcl = arg; 1023 1024 atomic_add_int(&mcl->mcl_refs, 1); 1025 } 1026 1027 /* 1028 * When dereferencing a cluster we have to deal with a N->0 race, where 1029 * N entities free their references simultaniously. To do this we use 1030 * atomic_fetchadd_int(). 1031 */ 1032 static void 1033 m_mclfree(void *arg) 1034 { 1035 struct mbcluster *mcl = arg; 1036 1037 if (atomic_fetchadd_int(&mcl->mcl_refs, -1) == 1) { 1038 --mbstat[mycpu->gd_cpuid].m_clusters; 1039 objcache_put(mclmeta_cache, mcl); 1040 } 1041 } 1042 1043 /* 1044 * Free a single mbuf and any associated external storage. The successor, 1045 * if any, is returned. 1046 * 1047 * We do need to check non-first mbuf for m_aux, since some of existing 1048 * code does not call M_PREPEND properly. 1049 * (example: call to bpf_mtap from drivers) 1050 */ 1051 1052 #ifdef MBUF_DEBUG 1053 1054 struct mbuf * 1055 _m_free(struct mbuf *m, const char *func) 1056 1057 #else 1058 1059 struct mbuf * 1060 m_free(struct mbuf *m) 1061 1062 #endif 1063 { 1064 struct mbuf *n; 1065 struct globaldata *gd = mycpu; 1066 1067 KASSERT(m->m_type != MT_FREE, ("freeing free mbuf %p", m)); 1068 KASSERT(M_TRAILINGSPACE(m) >= 0, ("overflowed mbuf %p", m)); 1069 --mbtypes[gd->gd_cpuid][m->m_type]; 1070 1071 n = m->m_next; 1072 1073 /* 1074 * Make sure the mbuf is in constructed state before returning it 1075 * to the objcache. 1076 */ 1077 m->m_next = NULL; 1078 mbufuntrack(m); 1079 #ifdef MBUF_DEBUG 1080 m->m_hdr.mh_lastfunc = func; 1081 #endif 1082 #ifdef notyet 1083 KKASSERT(m->m_nextpkt == NULL); 1084 #else 1085 if (m->m_nextpkt != NULL) { 1086 static int afewtimes = 10; 1087 1088 if (afewtimes-- > 0) { 1089 kprintf("mfree: m->m_nextpkt != NULL\n"); 1090 print_backtrace(-1); 1091 } 1092 m->m_nextpkt = NULL; 1093 } 1094 #endif 1095 if (m->m_flags & M_PKTHDR) { 1096 m_tag_delete_chain(m); /* eliminate XXX JH */ 1097 } 1098 1099 m->m_flags &= (M_EXT | M_EXT_CLUSTER | M_CLCACHE | M_PHCACHE); 1100 1101 /* 1102 * Clean the M_PKTHDR state so we can return the mbuf to its original 1103 * cache. This is based on the PHCACHE flag which tells us whether 1104 * the mbuf was originally allocated out of a packet-header cache 1105 * or a non-packet-header cache. 1106 */ 1107 if (m->m_flags & M_PHCACHE) { 1108 m->m_flags |= M_PKTHDR; 1109 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */ 1110 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */ 1111 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */ 1112 SLIST_INIT(&m->m_pkthdr.tags); 1113 } 1114 1115 /* 1116 * Handle remaining flags combinations. M_CLCACHE tells us whether 1117 * the mbuf was originally allocated from a cluster cache or not, 1118 * and is totally separate from whether the mbuf is currently 1119 * associated with a cluster. 1120 */ 1121 switch(m->m_flags & (M_CLCACHE | M_EXT | M_EXT_CLUSTER)) { 1122 case M_CLCACHE | M_EXT | M_EXT_CLUSTER: 1123 /* 1124 * mbuf+cluster cache case. The mbuf was allocated from the 1125 * combined mbuf_cluster cache and can be returned to the 1126 * cache if the cluster hasn't been shared. 1127 */ 1128 if (m_sharecount(m) == 1) { 1129 /* 1130 * The cluster has not been shared, we can just 1131 * reset the data pointer and return the mbuf 1132 * to the cluster cache. Note that the reference 1133 * count is left intact (it is still associated with 1134 * an mbuf). 1135 */ 1136 m->m_data = m->m_ext.ext_buf; 1137 if (m->m_flags & M_EXT && m->m_ext.ext_size != MCLBYTES) { 1138 if (m->m_flags & M_PHCACHE) 1139 objcache_put(mbufphdrjcluster_cache, m); 1140 else 1141 objcache_put(mbufjcluster_cache, m); 1142 } else { 1143 if (m->m_flags & M_PHCACHE) 1144 objcache_put(mbufphdrcluster_cache, m); 1145 else 1146 objcache_put(mbufcluster_cache, m); 1147 } 1148 --mbstat[mycpu->gd_cpuid].m_clusters; 1149 } else { 1150 /* 1151 * Hell. Someone else has a ref on this cluster, 1152 * we have to disconnect it which means we can't 1153 * put it back into the mbufcluster_cache, we 1154 * have to destroy the mbuf. 1155 * 1156 * Other mbuf references to the cluster will typically 1157 * be M_EXT | M_EXT_CLUSTER but without M_CLCACHE. 1158 * 1159 * XXX we could try to connect another cluster to 1160 * it. 1161 */ 1162 m->m_ext.ext_free(m->m_ext.ext_arg); 1163 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER); 1164 if (m->m_ext.ext_size == MCLBYTES) { 1165 if (m->m_flags & M_PHCACHE) 1166 objcache_dtor(mbufphdrcluster_cache, m); 1167 else 1168 objcache_dtor(mbufcluster_cache, m); 1169 } else { 1170 if (m->m_flags & M_PHCACHE) 1171 objcache_dtor(mbufphdrjcluster_cache, m); 1172 else 1173 objcache_dtor(mbufjcluster_cache, m); 1174 } 1175 } 1176 break; 1177 case M_EXT | M_EXT_CLUSTER: 1178 case M_EXT: 1179 /* 1180 * Normal cluster association case, disconnect the cluster from 1181 * the mbuf. The cluster may or may not be custom. 1182 */ 1183 m->m_ext.ext_free(m->m_ext.ext_arg); 1184 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER); 1185 /* fall through */ 1186 case 0: 1187 /* 1188 * return the mbuf to the mbuf cache. 1189 */ 1190 if (m->m_flags & M_PHCACHE) { 1191 m->m_data = m->m_pktdat; 1192 objcache_put(mbufphdr_cache, m); 1193 } else { 1194 m->m_data = m->m_dat; 1195 objcache_put(mbuf_cache, m); 1196 } 1197 --mbstat[mycpu->gd_cpuid].m_mbufs; 1198 break; 1199 default: 1200 if (!panicstr) 1201 panic("bad mbuf flags %p %08x", m, m->m_flags); 1202 break; 1203 } 1204 return (n); 1205 } 1206 1207 #ifdef MBUF_DEBUG 1208 1209 void 1210 _m_freem(struct mbuf *m, const char *func) 1211 { 1212 while (m) 1213 m = _m_free(m, func); 1214 } 1215 1216 #else 1217 1218 void 1219 m_freem(struct mbuf *m) 1220 { 1221 while (m) 1222 m = m_free(m); 1223 } 1224 1225 #endif 1226 1227 void 1228 m_extadd(struct mbuf *m, caddr_t buf, u_int size, void (*reff)(void *), 1229 void (*freef)(void *), void *arg) 1230 { 1231 m->m_ext.ext_arg = arg; 1232 m->m_ext.ext_buf = buf; 1233 m->m_ext.ext_ref = reff; 1234 m->m_ext.ext_free = freef; 1235 m->m_ext.ext_size = size; 1236 reff(arg); 1237 m->m_data = buf; 1238 m->m_flags |= M_EXT; 1239 } 1240 1241 /* 1242 * mbuf utility routines 1243 */ 1244 1245 /* 1246 * Lesser-used path for M_PREPEND: allocate new mbuf to prepend to chain and 1247 * copy junk along. 1248 */ 1249 struct mbuf * 1250 m_prepend(struct mbuf *m, int len, int how) 1251 { 1252 struct mbuf *mn; 1253 1254 if (m->m_flags & M_PKTHDR) 1255 mn = m_gethdr(how, m->m_type); 1256 else 1257 mn = m_get(how, m->m_type); 1258 if (mn == NULL) { 1259 m_freem(m); 1260 return (NULL); 1261 } 1262 if (m->m_flags & M_PKTHDR) 1263 M_MOVE_PKTHDR(mn, m); 1264 mn->m_next = m; 1265 m = mn; 1266 if (len < MHLEN) 1267 MH_ALIGN(m, len); 1268 m->m_len = len; 1269 return (m); 1270 } 1271 1272 /* 1273 * Make a copy of an mbuf chain starting "off0" bytes from the beginning, 1274 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf. 1275 * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller. 1276 * Note that the copy is read-only, because clusters are not copied, 1277 * only their reference counts are incremented. 1278 */ 1279 struct mbuf * 1280 m_copym(const struct mbuf *m, int off0, int len, int wait) 1281 { 1282 struct mbuf *n, **np; 1283 int off = off0; 1284 struct mbuf *top; 1285 int copyhdr = 0; 1286 1287 KASSERT(off >= 0, ("m_copym, negative off %d", off)); 1288 KASSERT(len >= 0, ("m_copym, negative len %d", len)); 1289 if (off == 0 && (m->m_flags & M_PKTHDR)) 1290 copyhdr = 1; 1291 while (off > 0) { 1292 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain")); 1293 if (off < m->m_len) 1294 break; 1295 off -= m->m_len; 1296 m = m->m_next; 1297 } 1298 np = ⊤ 1299 top = NULL; 1300 while (len > 0) { 1301 if (m == NULL) { 1302 KASSERT(len == M_COPYALL, 1303 ("m_copym, length > size of mbuf chain")); 1304 break; 1305 } 1306 /* 1307 * Because we are sharing any cluster attachment below, 1308 * be sure to get an mbuf that does not have a cluster 1309 * associated with it. 1310 */ 1311 if (copyhdr) 1312 n = m_gethdr(wait, m->m_type); 1313 else 1314 n = m_get(wait, m->m_type); 1315 *np = n; 1316 if (n == NULL) 1317 goto nospace; 1318 if (copyhdr) { 1319 if (!m_dup_pkthdr(n, m, wait)) 1320 goto nospace; 1321 if (len == M_COPYALL) 1322 n->m_pkthdr.len -= off0; 1323 else 1324 n->m_pkthdr.len = len; 1325 copyhdr = 0; 1326 } 1327 n->m_len = min(len, m->m_len - off); 1328 if (m->m_flags & M_EXT) { 1329 KKASSERT((n->m_flags & M_EXT) == 0); 1330 n->m_data = m->m_data + off; 1331 m->m_ext.ext_ref(m->m_ext.ext_arg); 1332 n->m_ext = m->m_ext; 1333 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER); 1334 } else { 1335 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), 1336 (unsigned)n->m_len); 1337 } 1338 if (len != M_COPYALL) 1339 len -= n->m_len; 1340 off = 0; 1341 m = m->m_next; 1342 np = &n->m_next; 1343 } 1344 if (top == NULL) 1345 ++mbstat[mycpu->gd_cpuid].m_mcfail; 1346 return (top); 1347 nospace: 1348 m_freem(top); 1349 ++mbstat[mycpu->gd_cpuid].m_mcfail; 1350 return (NULL); 1351 } 1352 1353 /* 1354 * Copy an entire packet, including header (which must be present). 1355 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'. 1356 * Note that the copy is read-only, because clusters are not copied, 1357 * only their reference counts are incremented. 1358 * Preserve alignment of the first mbuf so if the creator has left 1359 * some room at the beginning (e.g. for inserting protocol headers) 1360 * the copies also have the room available. 1361 */ 1362 struct mbuf * 1363 m_copypacket(struct mbuf *m, int how) 1364 { 1365 struct mbuf *top, *n, *o; 1366 1367 n = m_gethdr(how, m->m_type); 1368 top = n; 1369 if (!n) 1370 goto nospace; 1371 1372 if (!m_dup_pkthdr(n, m, how)) 1373 goto nospace; 1374 n->m_len = m->m_len; 1375 if (m->m_flags & M_EXT) { 1376 KKASSERT((n->m_flags & M_EXT) == 0); 1377 n->m_data = m->m_data; 1378 m->m_ext.ext_ref(m->m_ext.ext_arg); 1379 n->m_ext = m->m_ext; 1380 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER); 1381 } else { 1382 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat ); 1383 bcopy(mtod(m, char *), mtod(n, char *), n->m_len); 1384 } 1385 1386 m = m->m_next; 1387 while (m) { 1388 o = m_get(how, m->m_type); 1389 if (!o) 1390 goto nospace; 1391 1392 n->m_next = o; 1393 n = n->m_next; 1394 1395 n->m_len = m->m_len; 1396 if (m->m_flags & M_EXT) { 1397 KKASSERT((n->m_flags & M_EXT) == 0); 1398 n->m_data = m->m_data; 1399 m->m_ext.ext_ref(m->m_ext.ext_arg); 1400 n->m_ext = m->m_ext; 1401 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER); 1402 } else { 1403 bcopy(mtod(m, char *), mtod(n, char *), n->m_len); 1404 } 1405 1406 m = m->m_next; 1407 } 1408 return top; 1409 nospace: 1410 m_freem(top); 1411 ++mbstat[mycpu->gd_cpuid].m_mcfail; 1412 return (NULL); 1413 } 1414 1415 /* 1416 * Copy data from an mbuf chain starting "off" bytes from the beginning, 1417 * continuing for "len" bytes, into the indicated buffer. 1418 */ 1419 void 1420 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp) 1421 { 1422 unsigned count; 1423 1424 KASSERT(off >= 0, ("m_copydata, negative off %d", off)); 1425 KASSERT(len >= 0, ("m_copydata, negative len %d", len)); 1426 while (off > 0) { 1427 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain")); 1428 if (off < m->m_len) 1429 break; 1430 off -= m->m_len; 1431 m = m->m_next; 1432 } 1433 while (len > 0) { 1434 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain")); 1435 count = min(m->m_len - off, len); 1436 bcopy(mtod(m, caddr_t) + off, cp, count); 1437 len -= count; 1438 cp += count; 1439 off = 0; 1440 m = m->m_next; 1441 } 1442 } 1443 1444 /* 1445 * Copy a packet header mbuf chain into a completely new chain, including 1446 * copying any mbuf clusters. Use this instead of m_copypacket() when 1447 * you need a writable copy of an mbuf chain. 1448 */ 1449 struct mbuf * 1450 m_dup(struct mbuf *m, int how) 1451 { 1452 struct mbuf **p, *top = NULL; 1453 int remain, moff, nsize; 1454 1455 /* Sanity check */ 1456 if (m == NULL) 1457 return (NULL); 1458 KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __func__)); 1459 1460 /* While there's more data, get a new mbuf, tack it on, and fill it */ 1461 remain = m->m_pkthdr.len; 1462 moff = 0; 1463 p = ⊤ 1464 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */ 1465 struct mbuf *n; 1466 1467 /* Get the next new mbuf */ 1468 n = m_getl(remain, how, m->m_type, top == NULL ? M_PKTHDR : 0, 1469 &nsize); 1470 if (n == NULL) 1471 goto nospace; 1472 if (top == NULL) 1473 if (!m_dup_pkthdr(n, m, how)) 1474 goto nospace0; 1475 1476 /* Link it into the new chain */ 1477 *p = n; 1478 p = &n->m_next; 1479 1480 /* Copy data from original mbuf(s) into new mbuf */ 1481 n->m_len = 0; 1482 while (n->m_len < nsize && m != NULL) { 1483 int chunk = min(nsize - n->m_len, m->m_len - moff); 1484 1485 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk); 1486 moff += chunk; 1487 n->m_len += chunk; 1488 remain -= chunk; 1489 if (moff == m->m_len) { 1490 m = m->m_next; 1491 moff = 0; 1492 } 1493 } 1494 1495 /* Check correct total mbuf length */ 1496 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL), 1497 ("%s: bogus m_pkthdr.len", __func__)); 1498 } 1499 return (top); 1500 1501 nospace: 1502 m_freem(top); 1503 nospace0: 1504 ++mbstat[mycpu->gd_cpuid].m_mcfail; 1505 return (NULL); 1506 } 1507 1508 /* 1509 * Copy the non-packet mbuf data chain into a new set of mbufs, including 1510 * copying any mbuf clusters. This is typically used to realign a data 1511 * chain by nfs_realign(). 1512 * 1513 * The original chain is left intact. how should be MB_WAIT or MB_DONTWAIT 1514 * and NULL can be returned if MB_DONTWAIT is passed. 1515 * 1516 * Be careful to use cluster mbufs, a large mbuf chain converted to non 1517 * cluster mbufs can exhaust our supply of mbufs. 1518 */ 1519 struct mbuf * 1520 m_dup_data(struct mbuf *m, int how) 1521 { 1522 struct mbuf **p, *n, *top = NULL; 1523 int mlen, moff, chunk, gsize, nsize; 1524 1525 /* 1526 * Degenerate case 1527 */ 1528 if (m == NULL) 1529 return (NULL); 1530 1531 /* 1532 * Optimize the mbuf allocation but do not get too carried away. 1533 */ 1534 if (m->m_next || m->m_len > MLEN) 1535 if (m->m_flags & M_EXT && m->m_ext.ext_size == MCLBYTES) 1536 gsize = MCLBYTES; 1537 else 1538 gsize = MJUMPAGESIZE; 1539 else 1540 gsize = MLEN; 1541 1542 /* Chain control */ 1543 p = ⊤ 1544 n = NULL; 1545 nsize = 0; 1546 1547 /* 1548 * Scan the mbuf chain until nothing is left, the new mbuf chain 1549 * will be allocated on the fly as needed. 1550 */ 1551 while (m) { 1552 mlen = m->m_len; 1553 moff = 0; 1554 1555 while (mlen) { 1556 KKASSERT(m->m_type == MT_DATA); 1557 if (n == NULL) { 1558 n = m_getl(gsize, how, MT_DATA, 0, &nsize); 1559 n->m_len = 0; 1560 if (n == NULL) 1561 goto nospace; 1562 *p = n; 1563 p = &n->m_next; 1564 } 1565 chunk = imin(mlen, nsize); 1566 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk); 1567 mlen -= chunk; 1568 moff += chunk; 1569 n->m_len += chunk; 1570 nsize -= chunk; 1571 if (nsize == 0) 1572 n = NULL; 1573 } 1574 m = m->m_next; 1575 } 1576 *p = NULL; 1577 return(top); 1578 nospace: 1579 *p = NULL; 1580 m_freem(top); 1581 ++mbstat[mycpu->gd_cpuid].m_mcfail; 1582 return (NULL); 1583 } 1584 1585 /* 1586 * Concatenate mbuf chain n to m. 1587 * Both chains must be of the same type (e.g. MT_DATA). 1588 * Any m_pkthdr is not updated. 1589 */ 1590 void 1591 m_cat(struct mbuf *m, struct mbuf *n) 1592 { 1593 m = m_last(m); 1594 while (n) { 1595 if (m->m_flags & M_EXT || 1596 m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) { 1597 /* just join the two chains */ 1598 m->m_next = n; 1599 return; 1600 } 1601 /* splat the data from one into the other */ 1602 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len, 1603 (u_int)n->m_len); 1604 m->m_len += n->m_len; 1605 n = m_free(n); 1606 } 1607 } 1608 1609 void 1610 m_adj(struct mbuf *mp, int req_len) 1611 { 1612 int len = req_len; 1613 struct mbuf *m; 1614 int count; 1615 1616 if ((m = mp) == NULL) 1617 return; 1618 if (len >= 0) { 1619 /* 1620 * Trim from head. 1621 */ 1622 while (m != NULL && len > 0) { 1623 if (m->m_len <= len) { 1624 len -= m->m_len; 1625 m->m_len = 0; 1626 m = m->m_next; 1627 } else { 1628 m->m_len -= len; 1629 m->m_data += len; 1630 len = 0; 1631 } 1632 } 1633 m = mp; 1634 if (mp->m_flags & M_PKTHDR) 1635 m->m_pkthdr.len -= (req_len - len); 1636 } else { 1637 /* 1638 * Trim from tail. Scan the mbuf chain, 1639 * calculating its length and finding the last mbuf. 1640 * If the adjustment only affects this mbuf, then just 1641 * adjust and return. Otherwise, rescan and truncate 1642 * after the remaining size. 1643 */ 1644 len = -len; 1645 count = 0; 1646 for (;;) { 1647 count += m->m_len; 1648 if (m->m_next == NULL) 1649 break; 1650 m = m->m_next; 1651 } 1652 if (m->m_len >= len) { 1653 m->m_len -= len; 1654 if (mp->m_flags & M_PKTHDR) 1655 mp->m_pkthdr.len -= len; 1656 return; 1657 } 1658 count -= len; 1659 if (count < 0) 1660 count = 0; 1661 /* 1662 * Correct length for chain is "count". 1663 * Find the mbuf with last data, adjust its length, 1664 * and toss data from remaining mbufs on chain. 1665 */ 1666 m = mp; 1667 if (m->m_flags & M_PKTHDR) 1668 m->m_pkthdr.len = count; 1669 for (; m; m = m->m_next) { 1670 if (m->m_len >= count) { 1671 m->m_len = count; 1672 break; 1673 } 1674 count -= m->m_len; 1675 } 1676 while (m->m_next) 1677 (m = m->m_next) ->m_len = 0; 1678 } 1679 } 1680 1681 /* 1682 * Set the m_data pointer of a newly-allocated mbuf 1683 * to place an object of the specified size at the 1684 * end of the mbuf, longword aligned. 1685 */ 1686 void 1687 m_align(struct mbuf *m, int len) 1688 { 1689 int adjust; 1690 1691 if (m->m_flags & M_EXT) 1692 adjust = m->m_ext.ext_size - len; 1693 else if (m->m_flags & M_PKTHDR) 1694 adjust = MHLEN - len; 1695 else 1696 adjust = MLEN - len; 1697 m->m_data += adjust &~ (sizeof(long)-1); 1698 } 1699 1700 /* 1701 * Create a writable copy of the mbuf chain. While doing this 1702 * we compact the chain with a goal of producing a chain with 1703 * at most two mbufs. The second mbuf in this chain is likely 1704 * to be a cluster. The primary purpose of this work is to create 1705 * a writable packet for encryption, compression, etc. The 1706 * secondary goal is to linearize the data so the data can be 1707 * passed to crypto hardware in the most efficient manner possible. 1708 */ 1709 struct mbuf * 1710 m_unshare(struct mbuf *m0, int how) 1711 { 1712 struct mbuf *m, *mprev; 1713 struct mbuf *n, *mfirst, *mlast; 1714 int len, off; 1715 1716 mprev = NULL; 1717 for (m = m0; m != NULL; m = mprev->m_next) { 1718 /* 1719 * Regular mbufs are ignored unless there's a cluster 1720 * in front of it that we can use to coalesce. We do 1721 * the latter mainly so later clusters can be coalesced 1722 * also w/o having to handle them specially (i.e. convert 1723 * mbuf+cluster -> cluster). This optimization is heavily 1724 * influenced by the assumption that we're running over 1725 * Ethernet where MCLBYTES is large enough that the max 1726 * packet size will permit lots of coalescing into a 1727 * single cluster. This in turn permits efficient 1728 * crypto operations, especially when using hardware. 1729 */ 1730 if ((m->m_flags & M_EXT) == 0) { 1731 if (mprev && (mprev->m_flags & M_EXT) && 1732 m->m_len <= M_TRAILINGSPACE(mprev)) { 1733 /* XXX: this ignores mbuf types */ 1734 memcpy(mtod(mprev, caddr_t) + mprev->m_len, 1735 mtod(m, caddr_t), m->m_len); 1736 mprev->m_len += m->m_len; 1737 mprev->m_next = m->m_next; /* unlink from chain */ 1738 m_free(m); /* reclaim mbuf */ 1739 } else { 1740 mprev = m; 1741 } 1742 continue; 1743 } 1744 /* 1745 * Writable mbufs are left alone (for now). 1746 */ 1747 if (M_WRITABLE(m)) { 1748 mprev = m; 1749 continue; 1750 } 1751 1752 /* 1753 * Not writable, replace with a copy or coalesce with 1754 * the previous mbuf if possible (since we have to copy 1755 * it anyway, we try to reduce the number of mbufs and 1756 * clusters so that future work is easier). 1757 */ 1758 KASSERT(m->m_flags & M_EXT, ("m_flags 0x%x", m->m_flags)); 1759 /* NB: we only coalesce into a cluster or larger */ 1760 if (mprev != NULL && (mprev->m_flags & M_EXT) && 1761 m->m_len <= M_TRAILINGSPACE(mprev)) { 1762 /* XXX: this ignores mbuf types */ 1763 memcpy(mtod(mprev, caddr_t) + mprev->m_len, 1764 mtod(m, caddr_t), m->m_len); 1765 mprev->m_len += m->m_len; 1766 mprev->m_next = m->m_next; /* unlink from chain */ 1767 m_free(m); /* reclaim mbuf */ 1768 continue; 1769 } 1770 1771 /* 1772 * Allocate new space to hold the copy... 1773 */ 1774 /* XXX why can M_PKTHDR be set past the first mbuf? */ 1775 if (mprev == NULL && (m->m_flags & M_PKTHDR)) { 1776 /* 1777 * NB: if a packet header is present we must 1778 * allocate the mbuf separately from any cluster 1779 * because M_MOVE_PKTHDR will smash the data 1780 * pointer and drop the M_EXT marker. 1781 */ 1782 MGETHDR(n, how, m->m_type); 1783 if (n == NULL) { 1784 m_freem(m0); 1785 return (NULL); 1786 } 1787 M_MOVE_PKTHDR(n, m); 1788 MCLGET(n, how); 1789 if ((n->m_flags & M_EXT) == 0) { 1790 m_free(n); 1791 m_freem(m0); 1792 return (NULL); 1793 } 1794 } else { 1795 n = m_getcl(how, m->m_type, m->m_flags); 1796 if (n == NULL) { 1797 m_freem(m0); 1798 return (NULL); 1799 } 1800 } 1801 /* 1802 * ... and copy the data. We deal with jumbo mbufs 1803 * (i.e. m_len > MCLBYTES) by splitting them into 1804 * clusters. We could just malloc a buffer and make 1805 * it external but too many device drivers don't know 1806 * how to break up the non-contiguous memory when 1807 * doing DMA. 1808 */ 1809 len = m->m_len; 1810 off = 0; 1811 mfirst = n; 1812 mlast = NULL; 1813 for (;;) { 1814 int cc = min(len, MCLBYTES); 1815 memcpy(mtod(n, caddr_t), mtod(m, caddr_t) + off, cc); 1816 n->m_len = cc; 1817 if (mlast != NULL) 1818 mlast->m_next = n; 1819 mlast = n; 1820 1821 len -= cc; 1822 if (len <= 0) 1823 break; 1824 off += cc; 1825 1826 n = m_getcl(how, m->m_type, m->m_flags); 1827 if (n == NULL) { 1828 m_freem(mfirst); 1829 m_freem(m0); 1830 return (NULL); 1831 } 1832 } 1833 n->m_next = m->m_next; 1834 if (mprev == NULL) 1835 m0 = mfirst; /* new head of chain */ 1836 else 1837 mprev->m_next = mfirst; /* replace old mbuf */ 1838 m_free(m); /* release old mbuf */ 1839 mprev = mfirst; 1840 } 1841 return (m0); 1842 } 1843 1844 /* 1845 * Rearrange an mbuf chain so that len bytes are contiguous 1846 * and in the data area of an mbuf (so that mtod will work for a structure 1847 * of size len). Returns the resulting mbuf chain on success, frees it and 1848 * returns null on failure. If there is room, it will add up to 1849 * max_protohdr-len extra bytes to the contiguous region in an attempt to 1850 * avoid being called next time. 1851 */ 1852 struct mbuf * 1853 m_pullup(struct mbuf *n, int len) 1854 { 1855 struct mbuf *m; 1856 int count; 1857 int space; 1858 1859 /* 1860 * If first mbuf has no cluster, and has room for len bytes 1861 * without shifting current data, pullup into it, 1862 * otherwise allocate a new mbuf to prepend to the chain. 1863 */ 1864 if (!(n->m_flags & M_EXT) && 1865 n->m_data + len < &n->m_dat[MLEN] && 1866 n->m_next) { 1867 if (n->m_len >= len) 1868 return (n); 1869 m = n; 1870 n = n->m_next; 1871 len -= m->m_len; 1872 } else { 1873 if (len > MHLEN) 1874 goto bad; 1875 if (n->m_flags & M_PKTHDR) 1876 m = m_gethdr(MB_DONTWAIT, n->m_type); 1877 else 1878 m = m_get(MB_DONTWAIT, n->m_type); 1879 if (m == NULL) 1880 goto bad; 1881 m->m_len = 0; 1882 if (n->m_flags & M_PKTHDR) 1883 M_MOVE_PKTHDR(m, n); 1884 } 1885 space = &m->m_dat[MLEN] - (m->m_data + m->m_len); 1886 do { 1887 count = min(min(max(len, max_protohdr), space), n->m_len); 1888 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len, 1889 (unsigned)count); 1890 len -= count; 1891 m->m_len += count; 1892 n->m_len -= count; 1893 space -= count; 1894 if (n->m_len) 1895 n->m_data += count; 1896 else 1897 n = m_free(n); 1898 } while (len > 0 && n); 1899 if (len > 0) { 1900 m_free(m); 1901 goto bad; 1902 } 1903 m->m_next = n; 1904 return (m); 1905 bad: 1906 m_freem(n); 1907 ++mbstat[mycpu->gd_cpuid].m_mcfail; 1908 return (NULL); 1909 } 1910 1911 /* 1912 * Partition an mbuf chain in two pieces, returning the tail -- 1913 * all but the first len0 bytes. In case of failure, it returns NULL and 1914 * attempts to restore the chain to its original state. 1915 * 1916 * Note that the resulting mbufs might be read-only, because the new 1917 * mbuf can end up sharing an mbuf cluster with the original mbuf if 1918 * the "breaking point" happens to lie within a cluster mbuf. Use the 1919 * M_WRITABLE() macro to check for this case. 1920 */ 1921 struct mbuf * 1922 m_split(struct mbuf *m0, int len0, int wait) 1923 { 1924 struct mbuf *m, *n; 1925 unsigned len = len0, remain; 1926 1927 for (m = m0; m && len > m->m_len; m = m->m_next) 1928 len -= m->m_len; 1929 if (m == NULL) 1930 return (NULL); 1931 remain = m->m_len - len; 1932 if (m0->m_flags & M_PKTHDR) { 1933 n = m_gethdr(wait, m0->m_type); 1934 if (n == NULL) 1935 return (NULL); 1936 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif; 1937 n->m_pkthdr.len = m0->m_pkthdr.len - len0; 1938 m0->m_pkthdr.len = len0; 1939 if (m->m_flags & M_EXT) 1940 goto extpacket; 1941 if (remain > MHLEN) { 1942 /* m can't be the lead packet */ 1943 MH_ALIGN(n, 0); 1944 n->m_next = m_split(m, len, wait); 1945 if (n->m_next == NULL) { 1946 m_free(n); 1947 return (NULL); 1948 } else { 1949 n->m_len = 0; 1950 return (n); 1951 } 1952 } else 1953 MH_ALIGN(n, remain); 1954 } else if (remain == 0) { 1955 n = m->m_next; 1956 m->m_next = 0; 1957 return (n); 1958 } else { 1959 n = m_get(wait, m->m_type); 1960 if (n == NULL) 1961 return (NULL); 1962 M_ALIGN(n, remain); 1963 } 1964 extpacket: 1965 if (m->m_flags & M_EXT) { 1966 KKASSERT((n->m_flags & M_EXT) == 0); 1967 n->m_data = m->m_data + len; 1968 m->m_ext.ext_ref(m->m_ext.ext_arg); 1969 n->m_ext = m->m_ext; 1970 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER); 1971 } else { 1972 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain); 1973 } 1974 n->m_len = remain; 1975 m->m_len = len; 1976 n->m_next = m->m_next; 1977 m->m_next = 0; 1978 return (n); 1979 } 1980 1981 /* 1982 * Routine to copy from device local memory into mbufs. 1983 * Note: "offset" is ill-defined and always called as 0, so ignore it. 1984 */ 1985 struct mbuf * 1986 m_devget(char *buf, int len, int offset, struct ifnet *ifp, 1987 void (*copy)(volatile const void *from, volatile void *to, size_t length)) 1988 { 1989 struct mbuf *m, *mfirst = NULL, **mtail; 1990 int nsize, flags; 1991 1992 if (copy == NULL) 1993 copy = bcopy; 1994 mtail = &mfirst; 1995 flags = M_PKTHDR; 1996 1997 while (len > 0) { 1998 m = m_getl(len, MB_DONTWAIT, MT_DATA, flags, &nsize); 1999 if (m == NULL) { 2000 m_freem(mfirst); 2001 return (NULL); 2002 } 2003 m->m_len = min(len, nsize); 2004 2005 if (flags & M_PKTHDR) { 2006 if (len + max_linkhdr <= nsize) 2007 m->m_data += max_linkhdr; 2008 m->m_pkthdr.rcvif = ifp; 2009 m->m_pkthdr.len = len; 2010 flags = 0; 2011 } 2012 2013 copy(buf, m->m_data, (unsigned)m->m_len); 2014 buf += m->m_len; 2015 len -= m->m_len; 2016 *mtail = m; 2017 mtail = &m->m_next; 2018 } 2019 2020 return (mfirst); 2021 } 2022 2023 /* 2024 * Routine to pad mbuf to the specified length 'padto'. 2025 */ 2026 int 2027 m_devpad(struct mbuf *m, int padto) 2028 { 2029 struct mbuf *last = NULL; 2030 int padlen; 2031 2032 if (padto <= m->m_pkthdr.len) 2033 return 0; 2034 2035 padlen = padto - m->m_pkthdr.len; 2036 2037 /* if there's only the packet-header and we can pad there, use it. */ 2038 if (m->m_pkthdr.len == m->m_len && M_TRAILINGSPACE(m) >= padlen) { 2039 last = m; 2040 } else { 2041 /* 2042 * Walk packet chain to find last mbuf. We will either 2043 * pad there, or append a new mbuf and pad it 2044 */ 2045 for (last = m; last->m_next != NULL; last = last->m_next) 2046 ; /* EMPTY */ 2047 2048 /* `last' now points to last in chain. */ 2049 if (M_TRAILINGSPACE(last) < padlen) { 2050 struct mbuf *n; 2051 2052 /* Allocate new empty mbuf, pad it. Compact later. */ 2053 MGET(n, MB_DONTWAIT, MT_DATA); 2054 if (n == NULL) 2055 return ENOBUFS; 2056 n->m_len = 0; 2057 last->m_next = n; 2058 last = n; 2059 } 2060 } 2061 KKASSERT(M_TRAILINGSPACE(last) >= padlen); 2062 KKASSERT(M_WRITABLE(last)); 2063 2064 /* Now zero the pad area */ 2065 bzero(mtod(last, char *) + last->m_len, padlen); 2066 last->m_len += padlen; 2067 m->m_pkthdr.len += padlen; 2068 return 0; 2069 } 2070 2071 /* 2072 * Copy data from a buffer back into the indicated mbuf chain, 2073 * starting "off" bytes from the beginning, extending the mbuf 2074 * chain if necessary. 2075 */ 2076 void 2077 m_copyback(struct mbuf *m0, int off, int len, caddr_t cp) 2078 { 2079 int mlen; 2080 struct mbuf *m = m0, *n; 2081 int totlen = 0; 2082 2083 if (m0 == NULL) 2084 return; 2085 while (off > (mlen = m->m_len)) { 2086 off -= mlen; 2087 totlen += mlen; 2088 if (m->m_next == NULL) { 2089 n = m_getclr(MB_DONTWAIT, m->m_type); 2090 if (n == NULL) 2091 goto out; 2092 n->m_len = min(MLEN, len + off); 2093 m->m_next = n; 2094 } 2095 m = m->m_next; 2096 } 2097 while (len > 0) { 2098 mlen = min (m->m_len - off, len); 2099 bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen); 2100 cp += mlen; 2101 len -= mlen; 2102 mlen += off; 2103 off = 0; 2104 totlen += mlen; 2105 if (len == 0) 2106 break; 2107 if (m->m_next == NULL) { 2108 n = m_get(MB_DONTWAIT, m->m_type); 2109 if (n == NULL) 2110 break; 2111 n->m_len = min(MLEN, len); 2112 m->m_next = n; 2113 } 2114 m = m->m_next; 2115 } 2116 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen)) 2117 m->m_pkthdr.len = totlen; 2118 } 2119 2120 /* 2121 * Append the specified data to the indicated mbuf chain, 2122 * Extend the mbuf chain if the new data does not fit in 2123 * existing space. 2124 * 2125 * Return 1 if able to complete the job; otherwise 0. 2126 */ 2127 int 2128 m_append(struct mbuf *m0, int len, c_caddr_t cp) 2129 { 2130 struct mbuf *m, *n; 2131 int remainder, space; 2132 2133 for (m = m0; m->m_next != NULL; m = m->m_next) 2134 ; 2135 remainder = len; 2136 space = M_TRAILINGSPACE(m); 2137 if (space > 0) { 2138 /* 2139 * Copy into available space. 2140 */ 2141 if (space > remainder) 2142 space = remainder; 2143 bcopy(cp, mtod(m, caddr_t) + m->m_len, space); 2144 m->m_len += space; 2145 cp += space, remainder -= space; 2146 } 2147 while (remainder > 0) { 2148 /* 2149 * Allocate a new mbuf; could check space 2150 * and allocate a cluster instead. 2151 */ 2152 n = m_get(MB_DONTWAIT, m->m_type); 2153 if (n == NULL) 2154 break; 2155 n->m_len = min(MLEN, remainder); 2156 bcopy(cp, mtod(n, caddr_t), n->m_len); 2157 cp += n->m_len, remainder -= n->m_len; 2158 m->m_next = n; 2159 m = n; 2160 } 2161 if (m0->m_flags & M_PKTHDR) 2162 m0->m_pkthdr.len += len - remainder; 2163 return (remainder == 0); 2164 } 2165 2166 /* 2167 * Apply function f to the data in an mbuf chain starting "off" bytes from 2168 * the beginning, continuing for "len" bytes. 2169 */ 2170 int 2171 m_apply(struct mbuf *m, int off, int len, 2172 int (*f)(void *, void *, u_int), void *arg) 2173 { 2174 u_int count; 2175 int rval; 2176 2177 KASSERT(off >= 0, ("m_apply, negative off %d", off)); 2178 KASSERT(len >= 0, ("m_apply, negative len %d", len)); 2179 while (off > 0) { 2180 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain")); 2181 if (off < m->m_len) 2182 break; 2183 off -= m->m_len; 2184 m = m->m_next; 2185 } 2186 while (len > 0) { 2187 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain")); 2188 count = min(m->m_len - off, len); 2189 rval = (*f)(arg, mtod(m, caddr_t) + off, count); 2190 if (rval) 2191 return (rval); 2192 len -= count; 2193 off = 0; 2194 m = m->m_next; 2195 } 2196 return (0); 2197 } 2198 2199 /* 2200 * Return a pointer to mbuf/offset of location in mbuf chain. 2201 */ 2202 struct mbuf * 2203 m_getptr(struct mbuf *m, int loc, int *off) 2204 { 2205 2206 while (loc >= 0) { 2207 /* Normal end of search. */ 2208 if (m->m_len > loc) { 2209 *off = loc; 2210 return (m); 2211 } else { 2212 loc -= m->m_len; 2213 if (m->m_next == NULL) { 2214 if (loc == 0) { 2215 /* Point at the end of valid data. */ 2216 *off = m->m_len; 2217 return (m); 2218 } 2219 return (NULL); 2220 } 2221 m = m->m_next; 2222 } 2223 } 2224 return (NULL); 2225 } 2226 2227 void 2228 m_print(const struct mbuf *m) 2229 { 2230 int len; 2231 const struct mbuf *m2; 2232 2233 len = m->m_pkthdr.len; 2234 m2 = m; 2235 while (len) { 2236 kprintf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-"); 2237 len -= m2->m_len; 2238 m2 = m2->m_next; 2239 } 2240 return; 2241 } 2242 2243 /* 2244 * "Move" mbuf pkthdr from "from" to "to". 2245 * "from" must have M_PKTHDR set, and "to" must be empty. 2246 */ 2247 void 2248 m_move_pkthdr(struct mbuf *to, struct mbuf *from) 2249 { 2250 KASSERT((to->m_flags & M_PKTHDR), ("m_move_pkthdr: not packet header")); 2251 2252 to->m_flags |= from->m_flags & M_COPYFLAGS; 2253 to->m_pkthdr = from->m_pkthdr; /* especially tags */ 2254 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */ 2255 } 2256 2257 /* 2258 * Duplicate "from"'s mbuf pkthdr in "to". 2259 * "from" must have M_PKTHDR set, and "to" must be empty. 2260 * In particular, this does a deep copy of the packet tags. 2261 */ 2262 int 2263 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how) 2264 { 2265 KASSERT((to->m_flags & M_PKTHDR), ("m_dup_pkthdr: not packet header")); 2266 2267 to->m_flags = (from->m_flags & M_COPYFLAGS) | 2268 (to->m_flags & ~M_COPYFLAGS); 2269 to->m_pkthdr = from->m_pkthdr; 2270 SLIST_INIT(&to->m_pkthdr.tags); 2271 return (m_tag_copy_chain(to, from, how)); 2272 } 2273 2274 /* 2275 * Defragment a mbuf chain, returning the shortest possible 2276 * chain of mbufs and clusters. If allocation fails and 2277 * this cannot be completed, NULL will be returned, but 2278 * the passed in chain will be unchanged. Upon success, 2279 * the original chain will be freed, and the new chain 2280 * will be returned. 2281 * 2282 * If a non-packet header is passed in, the original 2283 * mbuf (chain?) will be returned unharmed. 2284 * 2285 * m_defrag_nofree doesn't free the passed in mbuf. 2286 */ 2287 struct mbuf * 2288 m_defrag(struct mbuf *m0, int how) 2289 { 2290 struct mbuf *m_new; 2291 2292 if ((m_new = m_defrag_nofree(m0, how)) == NULL) 2293 return (NULL); 2294 if (m_new != m0) 2295 m_freem(m0); 2296 return (m_new); 2297 } 2298 2299 struct mbuf * 2300 m_defrag_nofree(struct mbuf *m0, int how) 2301 { 2302 struct mbuf *m_new = NULL, *m_final = NULL; 2303 int progress = 0, length, nsize; 2304 2305 if (!(m0->m_flags & M_PKTHDR)) 2306 return (m0); 2307 2308 #ifdef MBUF_STRESS_TEST 2309 if (m_defragrandomfailures) { 2310 int temp = karc4random() & 0xff; 2311 if (temp == 0xba) 2312 goto nospace; 2313 } 2314 #endif 2315 2316 m_final = m_getl(m0->m_pkthdr.len, how, MT_DATA, M_PKTHDR, &nsize); 2317 if (m_final == NULL) 2318 goto nospace; 2319 m_final->m_len = 0; /* in case m0->m_pkthdr.len is zero */ 2320 2321 if (m_dup_pkthdr(m_final, m0, how) == 0) 2322 goto nospace; 2323 2324 m_new = m_final; 2325 2326 while (progress < m0->m_pkthdr.len) { 2327 length = m0->m_pkthdr.len - progress; 2328 if (length > MCLBYTES) 2329 length = MCLBYTES; 2330 2331 if (m_new == NULL) { 2332 m_new = m_getl(length, how, MT_DATA, 0, &nsize); 2333 if (m_new == NULL) 2334 goto nospace; 2335 } 2336 2337 m_copydata(m0, progress, length, mtod(m_new, caddr_t)); 2338 progress += length; 2339 m_new->m_len = length; 2340 if (m_new != m_final) 2341 m_cat(m_final, m_new); 2342 m_new = NULL; 2343 } 2344 if (m0->m_next == NULL) 2345 m_defraguseless++; 2346 m_defragpackets++; 2347 m_defragbytes += m_final->m_pkthdr.len; 2348 return (m_final); 2349 nospace: 2350 m_defragfailure++; 2351 if (m_new) 2352 m_free(m_new); 2353 m_freem(m_final); 2354 return (NULL); 2355 } 2356 2357 /* 2358 * Move data from uio into mbufs. 2359 */ 2360 struct mbuf * 2361 m_uiomove(struct uio *uio) 2362 { 2363 struct mbuf *m; /* current working mbuf */ 2364 struct mbuf *head = NULL; /* result mbuf chain */ 2365 struct mbuf **mp = &head; 2366 int flags = M_PKTHDR; 2367 int nsize; 2368 int error; 2369 int resid; 2370 2371 do { 2372 if (uio->uio_resid > INT_MAX) 2373 resid = INT_MAX; 2374 else 2375 resid = (int)uio->uio_resid; 2376 m = m_getl(resid, MB_WAIT, MT_DATA, flags, &nsize); 2377 if (flags) { 2378 m->m_pkthdr.len = 0; 2379 /* Leave room for protocol headers. */ 2380 if (resid < MHLEN) 2381 MH_ALIGN(m, resid); 2382 flags = 0; 2383 } 2384 m->m_len = imin(nsize, resid); 2385 error = uiomove(mtod(m, caddr_t), m->m_len, uio); 2386 if (error) { 2387 m_free(m); 2388 goto failed; 2389 } 2390 *mp = m; 2391 mp = &m->m_next; 2392 head->m_pkthdr.len += m->m_len; 2393 } while (uio->uio_resid > 0); 2394 2395 return (head); 2396 2397 failed: 2398 m_freem(head); 2399 return (NULL); 2400 } 2401 2402 struct mbuf * 2403 m_last(struct mbuf *m) 2404 { 2405 while (m->m_next) 2406 m = m->m_next; 2407 return (m); 2408 } 2409 2410 /* 2411 * Return the number of bytes in an mbuf chain. 2412 * If lastm is not NULL, also return the last mbuf. 2413 */ 2414 u_int 2415 m_lengthm(struct mbuf *m, struct mbuf **lastm) 2416 { 2417 u_int len = 0; 2418 struct mbuf *prev = m; 2419 2420 while (m) { 2421 len += m->m_len; 2422 prev = m; 2423 m = m->m_next; 2424 } 2425 if (lastm != NULL) 2426 *lastm = prev; 2427 return (len); 2428 } 2429 2430 /* 2431 * Like m_lengthm(), except also keep track of mbuf usage. 2432 */ 2433 u_int 2434 m_countm(struct mbuf *m, struct mbuf **lastm, u_int *pmbcnt) 2435 { 2436 u_int len = 0, mbcnt = 0; 2437 struct mbuf *prev = m; 2438 2439 while (m) { 2440 len += m->m_len; 2441 mbcnt += MSIZE; 2442 if (m->m_flags & M_EXT) 2443 mbcnt += m->m_ext.ext_size; 2444 prev = m; 2445 m = m->m_next; 2446 } 2447 if (lastm != NULL) 2448 *lastm = prev; 2449 *pmbcnt = mbcnt; 2450 return (len); 2451 } 2452