1 /* 2 * Copyright (c) 2004 Jeffrey M. Hsu. All rights reserved. 3 * Copyright (c) 2004 The DragonFly Project. All rights reserved. 4 * 5 * This code is derived from software contributed to The DragonFly Project 6 * by Jeffrey M. Hsu. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of The DragonFly Project nor the names of its 17 * contributors may be used to endorse or promote products derived 18 * from this software without specific, prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 /* 35 * Copyright (c) 1982, 1986, 1988, 1991, 1993 36 * The Regents of the University of California. All rights reserved. 37 * 38 * Redistribution and use in source and binary forms, with or without 39 * modification, are permitted provided that the following conditions 40 * are met: 41 * 1. Redistributions of source code must retain the above copyright 42 * notice, this list of conditions and the following disclaimer. 43 * 2. Redistributions in binary form must reproduce the above copyright 44 * notice, this list of conditions and the following disclaimer in the 45 * documentation and/or other materials provided with the distribution. 46 * 3. All advertising materials mentioning features or use of this software 47 * must display the following acknowledgement: 48 * This product includes software developed by the University of 49 * California, Berkeley and its contributors. 50 * 4. Neither the name of the University nor the names of its contributors 51 * may be used to endorse or promote products derived from this software 52 * without specific prior written permission. 53 * 54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 64 * SUCH DAMAGE. 65 * 66 * @(#)uipc_mbuf.c 8.2 (Berkeley) 1/4/94 67 * $FreeBSD: src/sys/kern/uipc_mbuf.c,v 1.51.2.24 2003/04/15 06:59:29 silby Exp $ 68 * $DragonFly: src/sys/kern/uipc_mbuf.c,v 1.70 2008/11/20 14:21:01 sephe Exp $ 69 */ 70 71 #include "opt_param.h" 72 #include "opt_mbuf_stress_test.h" 73 #include <sys/param.h> 74 #include <sys/systm.h> 75 #include <sys/malloc.h> 76 #include <sys/mbuf.h> 77 #include <sys/kernel.h> 78 #include <sys/sysctl.h> 79 #include <sys/domain.h> 80 #include <sys/objcache.h> 81 #include <sys/tree.h> 82 #include <sys/protosw.h> 83 #include <sys/uio.h> 84 #include <sys/thread.h> 85 #include <sys/globaldata.h> 86 #include <sys/thread2.h> 87 88 #include <machine/atomic.h> 89 #include <machine/limits.h> 90 91 #include <vm/vm.h> 92 #include <vm/vm_kern.h> 93 #include <vm/vm_extern.h> 94 95 #ifdef INVARIANTS 96 #include <machine/cpu.h> 97 #endif 98 99 /* 100 * mbuf cluster meta-data 101 */ 102 struct mbcluster { 103 int32_t mcl_refs; 104 void *mcl_data; 105 }; 106 107 /* 108 * mbuf tracking for debugging purposes 109 */ 110 #ifdef MBUF_DEBUG 111 112 static MALLOC_DEFINE(M_MTRACK, "mtrack", "mtrack"); 113 114 struct mbctrack; 115 RB_HEAD(mbuf_rb_tree, mbtrack); 116 RB_PROTOTYPE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *); 117 118 struct mbtrack { 119 RB_ENTRY(mbtrack) rb_node; 120 int trackid; 121 struct mbuf *m; 122 }; 123 124 static int 125 mbtrack_cmp(struct mbtrack *mb1, struct mbtrack *mb2) 126 { 127 if (mb1->m < mb2->m) 128 return(-1); 129 if (mb1->m > mb2->m) 130 return(1); 131 return(0); 132 } 133 134 RB_GENERATE2(mbuf_rb_tree, mbtrack, rb_node, mbtrack_cmp, struct mbuf *, m); 135 136 struct mbuf_rb_tree mbuf_track_root; 137 138 static void 139 mbuftrack(struct mbuf *m) 140 { 141 struct mbtrack *mbt; 142 143 crit_enter(); 144 mbt = kmalloc(sizeof(*mbt), M_MTRACK, M_INTWAIT|M_ZERO); 145 mbt->m = m; 146 if (mbuf_rb_tree_RB_INSERT(&mbuf_track_root, mbt)) 147 panic("mbuftrack: mbuf %p already being tracked\n", m); 148 crit_exit(); 149 } 150 151 static void 152 mbufuntrack(struct mbuf *m) 153 { 154 struct mbtrack *mbt; 155 156 crit_enter(); 157 mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m); 158 if (mbt == NULL) { 159 kprintf("mbufuntrack: mbuf %p was not tracked\n", m); 160 } else { 161 mbuf_rb_tree_RB_REMOVE(&mbuf_track_root, mbt); 162 kfree(mbt, M_MTRACK); 163 } 164 crit_exit(); 165 } 166 167 void 168 mbuftrackid(struct mbuf *m, int trackid) 169 { 170 struct mbtrack *mbt; 171 struct mbuf *n; 172 173 crit_enter(); 174 while (m) { 175 n = m->m_nextpkt; 176 while (m) { 177 mbt = mbuf_rb_tree_RB_LOOKUP(&mbuf_track_root, m); 178 if (mbt) 179 mbt->trackid = trackid; 180 m = m->m_next; 181 } 182 m = n; 183 } 184 crit_exit(); 185 } 186 187 static int 188 mbuftrack_callback(struct mbtrack *mbt, void *arg) 189 { 190 struct sysctl_req *req = arg; 191 char buf[64]; 192 int error; 193 194 ksnprintf(buf, sizeof(buf), "mbuf %p track %d\n", mbt->m, mbt->trackid); 195 196 error = SYSCTL_OUT(req, buf, strlen(buf)); 197 if (error) 198 return(-error); 199 return(0); 200 } 201 202 static int 203 mbuftrack_show(SYSCTL_HANDLER_ARGS) 204 { 205 int error; 206 207 crit_enter(); 208 error = mbuf_rb_tree_RB_SCAN(&mbuf_track_root, NULL, 209 mbuftrack_callback, req); 210 crit_exit(); 211 return (-error); 212 } 213 SYSCTL_PROC(_kern_ipc, OID_AUTO, showmbufs, CTLFLAG_RD|CTLTYPE_STRING, 214 0, 0, mbuftrack_show, "A", "Show all in-use mbufs"); 215 216 #else 217 218 #define mbuftrack(m) 219 #define mbufuntrack(m) 220 221 #endif 222 223 static void mbinit(void *); 224 SYSINIT(mbuf, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, mbinit, NULL) 225 226 static u_long mbtypes[SMP_MAXCPU][MT_NTYPES]; 227 228 static struct mbstat mbstat[SMP_MAXCPU]; 229 int max_linkhdr; 230 int max_protohdr; 231 int max_hdr; 232 int max_datalen; 233 int m_defragpackets; 234 int m_defragbytes; 235 int m_defraguseless; 236 int m_defragfailure; 237 #ifdef MBUF_STRESS_TEST 238 int m_defragrandomfailures; 239 #endif 240 241 struct objcache *mbuf_cache, *mbufphdr_cache; 242 struct objcache *mclmeta_cache; 243 struct objcache *mbufcluster_cache, *mbufphdrcluster_cache; 244 245 int nmbclusters; 246 int nmbufs; 247 248 SYSCTL_INT(_kern_ipc, KIPC_MAX_LINKHDR, max_linkhdr, CTLFLAG_RW, 249 &max_linkhdr, 0, ""); 250 SYSCTL_INT(_kern_ipc, KIPC_MAX_PROTOHDR, max_protohdr, CTLFLAG_RW, 251 &max_protohdr, 0, ""); 252 SYSCTL_INT(_kern_ipc, KIPC_MAX_HDR, max_hdr, CTLFLAG_RW, &max_hdr, 0, ""); 253 SYSCTL_INT(_kern_ipc, KIPC_MAX_DATALEN, max_datalen, CTLFLAG_RW, 254 &max_datalen, 0, ""); 255 SYSCTL_INT(_kern_ipc, OID_AUTO, mbuf_wait, CTLFLAG_RW, 256 &mbuf_wait, 0, ""); 257 static int do_mbstat(SYSCTL_HANDLER_ARGS); 258 259 SYSCTL_PROC(_kern_ipc, KIPC_MBSTAT, mbstat, CTLTYPE_STRUCT|CTLFLAG_RD, 260 0, 0, do_mbstat, "S,mbstat", ""); 261 262 static int do_mbtypes(SYSCTL_HANDLER_ARGS); 263 264 SYSCTL_PROC(_kern_ipc, OID_AUTO, mbtypes, CTLTYPE_ULONG|CTLFLAG_RD, 265 0, 0, do_mbtypes, "LU", ""); 266 267 static int 268 do_mbstat(SYSCTL_HANDLER_ARGS) 269 { 270 struct mbstat mbstat_total; 271 struct mbstat *mbstat_totalp; 272 int i; 273 274 bzero(&mbstat_total, sizeof(mbstat_total)); 275 mbstat_totalp = &mbstat_total; 276 277 for (i = 0; i < ncpus; i++) 278 { 279 mbstat_total.m_mbufs += mbstat[i].m_mbufs; 280 mbstat_total.m_clusters += mbstat[i].m_clusters; 281 mbstat_total.m_spare += mbstat[i].m_spare; 282 mbstat_total.m_clfree += mbstat[i].m_clfree; 283 mbstat_total.m_drops += mbstat[i].m_drops; 284 mbstat_total.m_wait += mbstat[i].m_wait; 285 mbstat_total.m_drain += mbstat[i].m_drain; 286 mbstat_total.m_mcfail += mbstat[i].m_mcfail; 287 mbstat_total.m_mpfail += mbstat[i].m_mpfail; 288 289 } 290 /* 291 * The following fields are not cumulative fields so just 292 * get their values once. 293 */ 294 mbstat_total.m_msize = mbstat[0].m_msize; 295 mbstat_total.m_mclbytes = mbstat[0].m_mclbytes; 296 mbstat_total.m_minclsize = mbstat[0].m_minclsize; 297 mbstat_total.m_mlen = mbstat[0].m_mlen; 298 mbstat_total.m_mhlen = mbstat[0].m_mhlen; 299 300 return(sysctl_handle_opaque(oidp, mbstat_totalp, sizeof(mbstat_total), req)); 301 } 302 303 static int 304 do_mbtypes(SYSCTL_HANDLER_ARGS) 305 { 306 u_long totals[MT_NTYPES]; 307 int i, j; 308 309 for (i = 0; i < MT_NTYPES; i++) 310 totals[i] = 0; 311 312 for (i = 0; i < ncpus; i++) 313 { 314 for (j = 0; j < MT_NTYPES; j++) 315 totals[j] += mbtypes[i][j]; 316 } 317 318 return(sysctl_handle_opaque(oidp, totals, sizeof(totals), req)); 319 } 320 321 /* 322 * These are read-only because we do not currently have any code 323 * to adjust the objcache limits after the fact. The variables 324 * may only be set as boot-time tunables. 325 */ 326 SYSCTL_INT(_kern_ipc, KIPC_NMBCLUSTERS, nmbclusters, CTLFLAG_RD, 327 &nmbclusters, 0, "Maximum number of mbuf clusters available"); 328 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbufs, CTLFLAG_RD, &nmbufs, 0, 329 "Maximum number of mbufs available"); 330 331 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragpackets, CTLFLAG_RD, 332 &m_defragpackets, 0, ""); 333 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragbytes, CTLFLAG_RD, 334 &m_defragbytes, 0, ""); 335 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defraguseless, CTLFLAG_RD, 336 &m_defraguseless, 0, ""); 337 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragfailure, CTLFLAG_RD, 338 &m_defragfailure, 0, ""); 339 #ifdef MBUF_STRESS_TEST 340 SYSCTL_INT(_kern_ipc, OID_AUTO, m_defragrandomfailures, CTLFLAG_RW, 341 &m_defragrandomfailures, 0, ""); 342 #endif 343 344 static MALLOC_DEFINE(M_MBUF, "mbuf", "mbuf"); 345 static MALLOC_DEFINE(M_MBUFCL, "mbufcl", "mbufcl"); 346 static MALLOC_DEFINE(M_MCLMETA, "mclmeta", "mclmeta"); 347 348 static void m_reclaim (void); 349 static void m_mclref(void *arg); 350 static void m_mclfree(void *arg); 351 352 #ifndef NMBCLUSTERS 353 #define NMBCLUSTERS (512 + maxusers * 16) 354 #endif 355 #ifndef NMBUFS 356 #define NMBUFS (nmbclusters * 2) 357 #endif 358 359 /* 360 * Perform sanity checks of tunables declared above. 361 */ 362 static void 363 tunable_mbinit(void *dummy) 364 { 365 /* 366 * This has to be done before VM init. 367 */ 368 nmbclusters = NMBCLUSTERS; 369 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters); 370 nmbufs = NMBUFS; 371 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs); 372 /* Sanity checks */ 373 if (nmbufs < nmbclusters * 2) 374 nmbufs = nmbclusters * 2; 375 } 376 SYSINIT(tunable_mbinit, SI_BOOT1_TUNABLES, SI_ORDER_ANY, 377 tunable_mbinit, NULL); 378 379 /* "number of clusters of pages" */ 380 #define NCL_INIT 1 381 382 #define NMB_INIT 16 383 384 /* 385 * The mbuf object cache only guarantees that m_next and m_nextpkt are 386 * NULL and that m_data points to the beginning of the data area. In 387 * particular, m_len and m_pkthdr.len are uninitialized. It is the 388 * responsibility of the caller to initialize those fields before use. 389 */ 390 391 static boolean_t __inline 392 mbuf_ctor(void *obj, void *private, int ocflags) 393 { 394 struct mbuf *m = obj; 395 396 m->m_next = NULL; 397 m->m_nextpkt = NULL; 398 m->m_data = m->m_dat; 399 m->m_flags = 0; 400 401 return (TRUE); 402 } 403 404 /* 405 * Initialize the mbuf and the packet header fields. 406 */ 407 static boolean_t 408 mbufphdr_ctor(void *obj, void *private, int ocflags) 409 { 410 struct mbuf *m = obj; 411 412 m->m_next = NULL; 413 m->m_nextpkt = NULL; 414 m->m_data = m->m_pktdat; 415 m->m_flags = M_PKTHDR | M_PHCACHE; 416 417 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */ 418 SLIST_INIT(&m->m_pkthdr.tags); 419 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */ 420 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */ 421 422 return (TRUE); 423 } 424 425 /* 426 * A mbcluster object consists of 2K (MCLBYTES) cluster and a refcount. 427 */ 428 static boolean_t 429 mclmeta_ctor(void *obj, void *private, int ocflags) 430 { 431 struct mbcluster *cl = obj; 432 void *buf; 433 434 if (ocflags & M_NOWAIT) 435 buf = kmalloc(MCLBYTES, M_MBUFCL, M_NOWAIT | M_ZERO); 436 else 437 buf = kmalloc(MCLBYTES, M_MBUFCL, M_INTWAIT | M_ZERO); 438 if (buf == NULL) 439 return (FALSE); 440 cl->mcl_refs = 0; 441 cl->mcl_data = buf; 442 return (TRUE); 443 } 444 445 static void 446 mclmeta_dtor(void *obj, void *private) 447 { 448 struct mbcluster *mcl = obj; 449 450 KKASSERT(mcl->mcl_refs == 0); 451 kfree(mcl->mcl_data, M_MBUFCL); 452 } 453 454 static void 455 linkcluster(struct mbuf *m, struct mbcluster *cl) 456 { 457 /* 458 * Add the cluster to the mbuf. The caller will detect that the 459 * mbuf now has an attached cluster. 460 */ 461 m->m_ext.ext_arg = cl; 462 m->m_ext.ext_buf = cl->mcl_data; 463 m->m_ext.ext_ref = m_mclref; 464 m->m_ext.ext_free = m_mclfree; 465 m->m_ext.ext_size = MCLBYTES; 466 atomic_add_int(&cl->mcl_refs, 1); 467 468 m->m_data = m->m_ext.ext_buf; 469 m->m_flags |= M_EXT | M_EXT_CLUSTER; 470 } 471 472 static boolean_t 473 mbufphdrcluster_ctor(void *obj, void *private, int ocflags) 474 { 475 struct mbuf *m = obj; 476 struct mbcluster *cl; 477 478 mbufphdr_ctor(obj, private, ocflags); 479 cl = objcache_get(mclmeta_cache, ocflags); 480 if (cl == NULL) { 481 ++mbstat[mycpu->gd_cpuid].m_drops; 482 return (FALSE); 483 } 484 m->m_flags |= M_CLCACHE; 485 linkcluster(m, cl); 486 return (TRUE); 487 } 488 489 static boolean_t 490 mbufcluster_ctor(void *obj, void *private, int ocflags) 491 { 492 struct mbuf *m = obj; 493 struct mbcluster *cl; 494 495 mbuf_ctor(obj, private, ocflags); 496 cl = objcache_get(mclmeta_cache, ocflags); 497 if (cl == NULL) { 498 ++mbstat[mycpu->gd_cpuid].m_drops; 499 return (FALSE); 500 } 501 m->m_flags |= M_CLCACHE; 502 linkcluster(m, cl); 503 return (TRUE); 504 } 505 506 /* 507 * Used for both the cluster and cluster PHDR caches. 508 * 509 * The mbuf may have lost its cluster due to sharing, deal 510 * with the situation by checking M_EXT. 511 */ 512 static void 513 mbufcluster_dtor(void *obj, void *private) 514 { 515 struct mbuf *m = obj; 516 struct mbcluster *mcl; 517 518 if (m->m_flags & M_EXT) { 519 KKASSERT((m->m_flags & M_EXT_CLUSTER) != 0); 520 mcl = m->m_ext.ext_arg; 521 KKASSERT(mcl->mcl_refs == 1); 522 mcl->mcl_refs = 0; 523 objcache_put(mclmeta_cache, mcl); 524 } 525 } 526 527 struct objcache_malloc_args mbuf_malloc_args = { MSIZE, M_MBUF }; 528 struct objcache_malloc_args mclmeta_malloc_args = 529 { sizeof(struct mbcluster), M_MCLMETA }; 530 531 /* ARGSUSED*/ 532 static void 533 mbinit(void *dummy) 534 { 535 int mb_limit, cl_limit; 536 int limit; 537 int i; 538 539 /* 540 * Initialize statistics 541 */ 542 for (i = 0; i < ncpus; i++) { 543 atomic_set_long_nonlocked(&mbstat[i].m_msize, MSIZE); 544 atomic_set_long_nonlocked(&mbstat[i].m_mclbytes, MCLBYTES); 545 atomic_set_long_nonlocked(&mbstat[i].m_minclsize, MINCLSIZE); 546 atomic_set_long_nonlocked(&mbstat[i].m_mlen, MLEN); 547 atomic_set_long_nonlocked(&mbstat[i].m_mhlen, MHLEN); 548 } 549 550 /* 551 * Create objtect caches and save cluster limits, which will 552 * be used to adjust backing kmalloc pools' limit later. 553 */ 554 555 mb_limit = cl_limit = 0; 556 557 limit = nmbufs; 558 mbuf_cache = objcache_create("mbuf", &limit, 0, 559 mbuf_ctor, NULL, NULL, 560 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 561 mb_limit += limit; 562 563 limit = nmbufs; 564 mbufphdr_cache = objcache_create("mbuf pkt hdr", &limit, 64, 565 mbufphdr_ctor, NULL, NULL, 566 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 567 mb_limit += limit; 568 569 cl_limit = nmbclusters; 570 mclmeta_cache = objcache_create("cluster mbuf", &cl_limit, 0, 571 mclmeta_ctor, mclmeta_dtor, NULL, 572 objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args); 573 574 limit = nmbclusters; 575 mbufcluster_cache = objcache_create("mbuf + cluster", &limit, 0, 576 mbufcluster_ctor, mbufcluster_dtor, NULL, 577 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 578 mb_limit += limit; 579 580 limit = nmbclusters; 581 mbufphdrcluster_cache = objcache_create("mbuf pkt hdr + cluster", 582 &limit, 64, mbufphdrcluster_ctor, mbufcluster_dtor, NULL, 583 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 584 mb_limit += limit; 585 586 /* 587 * Adjust backing kmalloc pools' limit 588 * 589 * NOTE: We raise the limit by another 1/8 to take the effect 590 * of loosememuse into account. 591 */ 592 cl_limit += cl_limit / 8; 593 kmalloc_raise_limit(mclmeta_malloc_args.mtype, 594 mclmeta_malloc_args.objsize * cl_limit); 595 kmalloc_raise_limit(M_MBUFCL, MCLBYTES * cl_limit); 596 597 mb_limit += mb_limit / 8; 598 kmalloc_raise_limit(mbuf_malloc_args.mtype, 599 mbuf_malloc_args.objsize * mb_limit); 600 } 601 602 /* 603 * Return the number of references to this mbuf's data. 0 is returned 604 * if the mbuf is not M_EXT, a reference count is returned if it is 605 * M_EXT | M_EXT_CLUSTER, and 99 is returned if it is a special M_EXT. 606 */ 607 int 608 m_sharecount(struct mbuf *m) 609 { 610 switch (m->m_flags & (M_EXT | M_EXT_CLUSTER)) { 611 case 0: 612 return (0); 613 case M_EXT: 614 return (99); 615 case M_EXT | M_EXT_CLUSTER: 616 return (((struct mbcluster *)m->m_ext.ext_arg)->mcl_refs); 617 } 618 /* NOTREACHED */ 619 return (0); /* to shut up compiler */ 620 } 621 622 /* 623 * change mbuf to new type 624 */ 625 void 626 m_chtype(struct mbuf *m, int type) 627 { 628 struct globaldata *gd = mycpu; 629 630 atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1); 631 atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1); 632 atomic_set_short_nonlocked(&m->m_type, type); 633 } 634 635 static void 636 m_reclaim(void) 637 { 638 struct domain *dp; 639 struct protosw *pr; 640 641 crit_enter(); 642 SLIST_FOREACH(dp, &domains, dom_next) { 643 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) { 644 if (pr->pr_drain) 645 (*pr->pr_drain)(); 646 } 647 } 648 crit_exit(); 649 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_drain, 1); 650 } 651 652 static void __inline 653 updatestats(struct mbuf *m, int type) 654 { 655 struct globaldata *gd = mycpu; 656 m->m_type = type; 657 658 mbuftrack(m); 659 660 atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1); 661 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1); 662 663 } 664 665 /* 666 * Allocate an mbuf. 667 */ 668 struct mbuf * 669 m_get(int how, int type) 670 { 671 struct mbuf *m; 672 int ntries = 0; 673 int ocf = MBTOM(how); 674 675 retryonce: 676 677 m = objcache_get(mbuf_cache, ocf); 678 679 if (m == NULL) { 680 if ((how & MB_TRYWAIT) && ntries++ == 0) { 681 struct objcache *reclaimlist[] = { 682 mbufphdr_cache, 683 mbufcluster_cache, mbufphdrcluster_cache 684 }; 685 const int nreclaims = __arysize(reclaimlist); 686 687 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf)) 688 m_reclaim(); 689 goto retryonce; 690 } 691 ++mbstat[mycpu->gd_cpuid].m_drops; 692 return (NULL); 693 } 694 695 updatestats(m, type); 696 return (m); 697 } 698 699 struct mbuf * 700 m_gethdr(int how, int type) 701 { 702 struct mbuf *m; 703 int ocf = MBTOM(how); 704 int ntries = 0; 705 706 retryonce: 707 708 m = objcache_get(mbufphdr_cache, ocf); 709 710 if (m == NULL) { 711 if ((how & MB_TRYWAIT) && ntries++ == 0) { 712 struct objcache *reclaimlist[] = { 713 mbuf_cache, 714 mbufcluster_cache, mbufphdrcluster_cache 715 }; 716 const int nreclaims = __arysize(reclaimlist); 717 718 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf)) 719 m_reclaim(); 720 goto retryonce; 721 } 722 ++mbstat[mycpu->gd_cpuid].m_drops; 723 return (NULL); 724 } 725 726 updatestats(m, type); 727 return (m); 728 } 729 730 /* 731 * Get a mbuf (not a mbuf cluster!) and zero it. 732 * Deprecated. 733 */ 734 struct mbuf * 735 m_getclr(int how, int type) 736 { 737 struct mbuf *m; 738 739 m = m_get(how, type); 740 if (m != NULL) 741 bzero(m->m_data, MLEN); 742 return (m); 743 } 744 745 /* 746 * Returns an mbuf with an attached cluster. 747 * Because many network drivers use this kind of buffers a lot, it is 748 * convenient to keep a small pool of free buffers of this kind. 749 * Even a small size such as 10 gives about 10% improvement in the 750 * forwarding rate in a bridge or router. 751 */ 752 struct mbuf * 753 m_getcl(int how, short type, int flags) 754 { 755 struct mbuf *m; 756 int ocflags = MBTOM(how); 757 int ntries = 0; 758 759 retryonce: 760 761 if (flags & M_PKTHDR) 762 m = objcache_get(mbufphdrcluster_cache, ocflags); 763 else 764 m = objcache_get(mbufcluster_cache, ocflags); 765 766 if (m == NULL) { 767 if ((how & MB_TRYWAIT) && ntries++ == 0) { 768 struct objcache *reclaimlist[1]; 769 770 if (flags & M_PKTHDR) 771 reclaimlist[0] = mbufcluster_cache; 772 else 773 reclaimlist[0] = mbufphdrcluster_cache; 774 if (!objcache_reclaimlist(reclaimlist, 1, ocflags)) 775 m_reclaim(); 776 goto retryonce; 777 } 778 ++mbstat[mycpu->gd_cpuid].m_drops; 779 return (NULL); 780 } 781 782 m->m_type = type; 783 784 mbuftrack(m); 785 786 atomic_add_long_nonlocked(&mbtypes[mycpu->gd_cpuid][type], 1); 787 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1); 788 return (m); 789 } 790 791 /* 792 * Allocate chain of requested length. 793 */ 794 struct mbuf * 795 m_getc(int len, int how, int type) 796 { 797 struct mbuf *n, *nfirst = NULL, **ntail = &nfirst; 798 int nsize; 799 800 while (len > 0) { 801 n = m_getl(len, how, type, 0, &nsize); 802 if (n == NULL) 803 goto failed; 804 n->m_len = 0; 805 *ntail = n; 806 ntail = &n->m_next; 807 len -= nsize; 808 } 809 return (nfirst); 810 811 failed: 812 m_freem(nfirst); 813 return (NULL); 814 } 815 816 /* 817 * Allocate len-worth of mbufs and/or mbuf clusters (whatever fits best) 818 * and return a pointer to the head of the allocated chain. If m0 is 819 * non-null, then we assume that it is a single mbuf or an mbuf chain to 820 * which we want len bytes worth of mbufs and/or clusters attached, and so 821 * if we succeed in allocating it, we will just return a pointer to m0. 822 * 823 * If we happen to fail at any point during the allocation, we will free 824 * up everything we have already allocated and return NULL. 825 * 826 * Deprecated. Use m_getc() and m_cat() instead. 827 */ 828 struct mbuf * 829 m_getm(struct mbuf *m0, int len, int type, int how) 830 { 831 struct mbuf *nfirst; 832 833 nfirst = m_getc(len, how, type); 834 835 if (m0 != NULL) { 836 m_last(m0)->m_next = nfirst; 837 return (m0); 838 } 839 840 return (nfirst); 841 } 842 843 /* 844 * Adds a cluster to a normal mbuf, M_EXT is set on success. 845 * Deprecated. Use m_getcl() instead. 846 */ 847 void 848 m_mclget(struct mbuf *m, int how) 849 { 850 struct mbcluster *mcl; 851 852 KKASSERT((m->m_flags & M_EXT) == 0); 853 mcl = objcache_get(mclmeta_cache, MBTOM(how)); 854 if (mcl != NULL) { 855 linkcluster(m, mcl); 856 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1); 857 } else { 858 ++mbstat[mycpu->gd_cpuid].m_drops; 859 } 860 } 861 862 /* 863 * Updates to mbcluster must be MPSAFE. Only an entity which already has 864 * a reference to the cluster can ref it, so we are in no danger of 865 * racing an add with a subtract. But the operation must still be atomic 866 * since multiple entities may have a reference on the cluster. 867 * 868 * m_mclfree() is almost the same but it must contend with two entities 869 * freeing the cluster at the same time. If there is only one reference 870 * count we are the only entity referencing the cluster and no further 871 * locking is required. Otherwise we must protect against a race to 0 872 * with the serializer. 873 */ 874 static void 875 m_mclref(void *arg) 876 { 877 struct mbcluster *mcl = arg; 878 879 atomic_add_int(&mcl->mcl_refs, 1); 880 } 881 882 /* 883 * When dereferencing a cluster we have to deal with a N->0 race, where 884 * N entities free their references simultaniously. To do this we use 885 * atomic_fetchadd_int(). 886 */ 887 static void 888 m_mclfree(void *arg) 889 { 890 struct mbcluster *mcl = arg; 891 892 if (atomic_fetchadd_int(&mcl->mcl_refs, -1) == 1) 893 objcache_put(mclmeta_cache, mcl); 894 } 895 896 /* 897 * Free a single mbuf and any associated external storage. The successor, 898 * if any, is returned. 899 * 900 * We do need to check non-first mbuf for m_aux, since some of existing 901 * code does not call M_PREPEND properly. 902 * (example: call to bpf_mtap from drivers) 903 */ 904 struct mbuf * 905 m_free(struct mbuf *m) 906 { 907 struct mbuf *n; 908 struct globaldata *gd = mycpu; 909 910 KASSERT(m->m_type != MT_FREE, ("freeing free mbuf %p", m)); 911 atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1); 912 913 n = m->m_next; 914 915 /* 916 * Make sure the mbuf is in constructed state before returning it 917 * to the objcache. 918 */ 919 m->m_next = NULL; 920 mbufuntrack(m); 921 #ifdef notyet 922 KKASSERT(m->m_nextpkt == NULL); 923 #else 924 if (m->m_nextpkt != NULL) { 925 static int afewtimes = 10; 926 927 if (afewtimes-- > 0) { 928 kprintf("mfree: m->m_nextpkt != NULL\n"); 929 print_backtrace(-1); 930 } 931 m->m_nextpkt = NULL; 932 } 933 #endif 934 if (m->m_flags & M_PKTHDR) { 935 m_tag_delete_chain(m); /* eliminate XXX JH */ 936 } 937 938 m->m_flags &= (M_EXT | M_EXT_CLUSTER | M_CLCACHE | M_PHCACHE); 939 940 /* 941 * Clean the M_PKTHDR state so we can return the mbuf to its original 942 * cache. This is based on the PHCACHE flag which tells us whether 943 * the mbuf was originally allocated out of a packet-header cache 944 * or a non-packet-header cache. 945 */ 946 if (m->m_flags & M_PHCACHE) { 947 m->m_flags |= M_PKTHDR; 948 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */ 949 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */ 950 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */ 951 SLIST_INIT(&m->m_pkthdr.tags); 952 } 953 954 /* 955 * Handle remaining flags combinations. M_CLCACHE tells us whether 956 * the mbuf was originally allocated from a cluster cache or not, 957 * and is totally separate from whether the mbuf is currently 958 * associated with a cluster. 959 */ 960 crit_enter(); 961 switch(m->m_flags & (M_CLCACHE | M_EXT | M_EXT_CLUSTER)) { 962 case M_CLCACHE | M_EXT | M_EXT_CLUSTER: 963 /* 964 * mbuf+cluster cache case. The mbuf was allocated from the 965 * combined mbuf_cluster cache and can be returned to the 966 * cache if the cluster hasn't been shared. 967 */ 968 if (m_sharecount(m) == 1) { 969 /* 970 * The cluster has not been shared, we can just 971 * reset the data pointer and return the mbuf 972 * to the cluster cache. Note that the reference 973 * count is left intact (it is still associated with 974 * an mbuf). 975 */ 976 m->m_data = m->m_ext.ext_buf; 977 if (m->m_flags & M_PHCACHE) 978 objcache_put(mbufphdrcluster_cache, m); 979 else 980 objcache_put(mbufcluster_cache, m); 981 atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1); 982 } else { 983 /* 984 * Hell. Someone else has a ref on this cluster, 985 * we have to disconnect it which means we can't 986 * put it back into the mbufcluster_cache, we 987 * have to destroy the mbuf. 988 * 989 * Other mbuf references to the cluster will typically 990 * be M_EXT | M_EXT_CLUSTER but without M_CLCACHE. 991 * 992 * XXX we could try to connect another cluster to 993 * it. 994 */ 995 m->m_ext.ext_free(m->m_ext.ext_arg); 996 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER); 997 if (m->m_flags & M_PHCACHE) 998 objcache_dtor(mbufphdrcluster_cache, m); 999 else 1000 objcache_dtor(mbufcluster_cache, m); 1001 } 1002 break; 1003 case M_EXT | M_EXT_CLUSTER: 1004 /* 1005 * Normal cluster associated with an mbuf that was allocated 1006 * from the normal mbuf pool rather then the cluster pool. 1007 * The cluster has to be independantly disassociated from the 1008 * mbuf. 1009 */ 1010 if (m_sharecount(m) == 1) 1011 atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1); 1012 /* fall through */ 1013 case M_EXT: 1014 /* 1015 * Normal cluster association case, disconnect the cluster from 1016 * the mbuf. The cluster may or may not be custom. 1017 */ 1018 m->m_ext.ext_free(m->m_ext.ext_arg); 1019 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER); 1020 /* fall through */ 1021 case 0: 1022 /* 1023 * return the mbuf to the mbuf cache. 1024 */ 1025 if (m->m_flags & M_PHCACHE) { 1026 m->m_data = m->m_pktdat; 1027 objcache_put(mbufphdr_cache, m); 1028 } else { 1029 m->m_data = m->m_dat; 1030 objcache_put(mbuf_cache, m); 1031 } 1032 atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1); 1033 break; 1034 default: 1035 if (!panicstr) 1036 panic("bad mbuf flags %p %08x\n", m, m->m_flags); 1037 break; 1038 } 1039 crit_exit(); 1040 return (n); 1041 } 1042 1043 void 1044 m_freem(struct mbuf *m) 1045 { 1046 crit_enter(); 1047 while (m) 1048 m = m_free(m); 1049 crit_exit(); 1050 } 1051 1052 /* 1053 * mbuf utility routines 1054 */ 1055 1056 /* 1057 * Lesser-used path for M_PREPEND: allocate new mbuf to prepend to chain and 1058 * copy junk along. 1059 */ 1060 struct mbuf * 1061 m_prepend(struct mbuf *m, int len, int how) 1062 { 1063 struct mbuf *mn; 1064 1065 if (m->m_flags & M_PKTHDR) 1066 mn = m_gethdr(how, m->m_type); 1067 else 1068 mn = m_get(how, m->m_type); 1069 if (mn == NULL) { 1070 m_freem(m); 1071 return (NULL); 1072 } 1073 if (m->m_flags & M_PKTHDR) 1074 M_MOVE_PKTHDR(mn, m); 1075 mn->m_next = m; 1076 m = mn; 1077 if (len < MHLEN) 1078 MH_ALIGN(m, len); 1079 m->m_len = len; 1080 return (m); 1081 } 1082 1083 /* 1084 * Make a copy of an mbuf chain starting "off0" bytes from the beginning, 1085 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf. 1086 * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller. 1087 * Note that the copy is read-only, because clusters are not copied, 1088 * only their reference counts are incremented. 1089 */ 1090 struct mbuf * 1091 m_copym(const struct mbuf *m, int off0, int len, int wait) 1092 { 1093 struct mbuf *n, **np; 1094 int off = off0; 1095 struct mbuf *top; 1096 int copyhdr = 0; 1097 1098 KASSERT(off >= 0, ("m_copym, negative off %d", off)); 1099 KASSERT(len >= 0, ("m_copym, negative len %d", len)); 1100 if (off == 0 && m->m_flags & M_PKTHDR) 1101 copyhdr = 1; 1102 while (off > 0) { 1103 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain")); 1104 if (off < m->m_len) 1105 break; 1106 off -= m->m_len; 1107 m = m->m_next; 1108 } 1109 np = ⊤ 1110 top = 0; 1111 while (len > 0) { 1112 if (m == NULL) { 1113 KASSERT(len == M_COPYALL, 1114 ("m_copym, length > size of mbuf chain")); 1115 break; 1116 } 1117 /* 1118 * Because we are sharing any cluster attachment below, 1119 * be sure to get an mbuf that does not have a cluster 1120 * associated with it. 1121 */ 1122 if (copyhdr) 1123 n = m_gethdr(wait, m->m_type); 1124 else 1125 n = m_get(wait, m->m_type); 1126 *np = n; 1127 if (n == NULL) 1128 goto nospace; 1129 if (copyhdr) { 1130 if (!m_dup_pkthdr(n, m, wait)) 1131 goto nospace; 1132 if (len == M_COPYALL) 1133 n->m_pkthdr.len -= off0; 1134 else 1135 n->m_pkthdr.len = len; 1136 copyhdr = 0; 1137 } 1138 n->m_len = min(len, m->m_len - off); 1139 if (m->m_flags & M_EXT) { 1140 KKASSERT((n->m_flags & M_EXT) == 0); 1141 n->m_data = m->m_data + off; 1142 m->m_ext.ext_ref(m->m_ext.ext_arg); 1143 n->m_ext = m->m_ext; 1144 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER); 1145 } else { 1146 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), 1147 (unsigned)n->m_len); 1148 } 1149 if (len != M_COPYALL) 1150 len -= n->m_len; 1151 off = 0; 1152 m = m->m_next; 1153 np = &n->m_next; 1154 } 1155 if (top == NULL) 1156 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1); 1157 return (top); 1158 nospace: 1159 m_freem(top); 1160 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1); 1161 return (NULL); 1162 } 1163 1164 /* 1165 * Copy an entire packet, including header (which must be present). 1166 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'. 1167 * Note that the copy is read-only, because clusters are not copied, 1168 * only their reference counts are incremented. 1169 * Preserve alignment of the first mbuf so if the creator has left 1170 * some room at the beginning (e.g. for inserting protocol headers) 1171 * the copies also have the room available. 1172 */ 1173 struct mbuf * 1174 m_copypacket(struct mbuf *m, int how) 1175 { 1176 struct mbuf *top, *n, *o; 1177 1178 n = m_gethdr(how, m->m_type); 1179 top = n; 1180 if (!n) 1181 goto nospace; 1182 1183 if (!m_dup_pkthdr(n, m, how)) 1184 goto nospace; 1185 n->m_len = m->m_len; 1186 if (m->m_flags & M_EXT) { 1187 KKASSERT((n->m_flags & M_EXT) == 0); 1188 n->m_data = m->m_data; 1189 m->m_ext.ext_ref(m->m_ext.ext_arg); 1190 n->m_ext = m->m_ext; 1191 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER); 1192 } else { 1193 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat ); 1194 bcopy(mtod(m, char *), mtod(n, char *), n->m_len); 1195 } 1196 1197 m = m->m_next; 1198 while (m) { 1199 o = m_get(how, m->m_type); 1200 if (!o) 1201 goto nospace; 1202 1203 n->m_next = o; 1204 n = n->m_next; 1205 1206 n->m_len = m->m_len; 1207 if (m->m_flags & M_EXT) { 1208 KKASSERT((n->m_flags & M_EXT) == 0); 1209 n->m_data = m->m_data; 1210 m->m_ext.ext_ref(m->m_ext.ext_arg); 1211 n->m_ext = m->m_ext; 1212 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER); 1213 } else { 1214 bcopy(mtod(m, char *), mtod(n, char *), n->m_len); 1215 } 1216 1217 m = m->m_next; 1218 } 1219 return top; 1220 nospace: 1221 m_freem(top); 1222 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1); 1223 return (NULL); 1224 } 1225 1226 /* 1227 * Copy data from an mbuf chain starting "off" bytes from the beginning, 1228 * continuing for "len" bytes, into the indicated buffer. 1229 */ 1230 void 1231 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp) 1232 { 1233 unsigned count; 1234 1235 KASSERT(off >= 0, ("m_copydata, negative off %d", off)); 1236 KASSERT(len >= 0, ("m_copydata, negative len %d", len)); 1237 while (off > 0) { 1238 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain")); 1239 if (off < m->m_len) 1240 break; 1241 off -= m->m_len; 1242 m = m->m_next; 1243 } 1244 while (len > 0) { 1245 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain")); 1246 count = min(m->m_len - off, len); 1247 bcopy(mtod(m, caddr_t) + off, cp, count); 1248 len -= count; 1249 cp += count; 1250 off = 0; 1251 m = m->m_next; 1252 } 1253 } 1254 1255 /* 1256 * Copy a packet header mbuf chain into a completely new chain, including 1257 * copying any mbuf clusters. Use this instead of m_copypacket() when 1258 * you need a writable copy of an mbuf chain. 1259 */ 1260 struct mbuf * 1261 m_dup(struct mbuf *m, int how) 1262 { 1263 struct mbuf **p, *top = NULL; 1264 int remain, moff, nsize; 1265 1266 /* Sanity check */ 1267 if (m == NULL) 1268 return (NULL); 1269 KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __func__)); 1270 1271 /* While there's more data, get a new mbuf, tack it on, and fill it */ 1272 remain = m->m_pkthdr.len; 1273 moff = 0; 1274 p = ⊤ 1275 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */ 1276 struct mbuf *n; 1277 1278 /* Get the next new mbuf */ 1279 n = m_getl(remain, how, m->m_type, top == NULL ? M_PKTHDR : 0, 1280 &nsize); 1281 if (n == NULL) 1282 goto nospace; 1283 if (top == NULL) 1284 if (!m_dup_pkthdr(n, m, how)) 1285 goto nospace0; 1286 1287 /* Link it into the new chain */ 1288 *p = n; 1289 p = &n->m_next; 1290 1291 /* Copy data from original mbuf(s) into new mbuf */ 1292 n->m_len = 0; 1293 while (n->m_len < nsize && m != NULL) { 1294 int chunk = min(nsize - n->m_len, m->m_len - moff); 1295 1296 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk); 1297 moff += chunk; 1298 n->m_len += chunk; 1299 remain -= chunk; 1300 if (moff == m->m_len) { 1301 m = m->m_next; 1302 moff = 0; 1303 } 1304 } 1305 1306 /* Check correct total mbuf length */ 1307 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL), 1308 ("%s: bogus m_pkthdr.len", __func__)); 1309 } 1310 return (top); 1311 1312 nospace: 1313 m_freem(top); 1314 nospace0: 1315 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1); 1316 return (NULL); 1317 } 1318 1319 /* 1320 * Concatenate mbuf chain n to m. 1321 * Both chains must be of the same type (e.g. MT_DATA). 1322 * Any m_pkthdr is not updated. 1323 */ 1324 void 1325 m_cat(struct mbuf *m, struct mbuf *n) 1326 { 1327 m = m_last(m); 1328 while (n) { 1329 if (m->m_flags & M_EXT || 1330 m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) { 1331 /* just join the two chains */ 1332 m->m_next = n; 1333 return; 1334 } 1335 /* splat the data from one into the other */ 1336 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len, 1337 (u_int)n->m_len); 1338 m->m_len += n->m_len; 1339 n = m_free(n); 1340 } 1341 } 1342 1343 void 1344 m_adj(struct mbuf *mp, int req_len) 1345 { 1346 int len = req_len; 1347 struct mbuf *m; 1348 int count; 1349 1350 if ((m = mp) == NULL) 1351 return; 1352 if (len >= 0) { 1353 /* 1354 * Trim from head. 1355 */ 1356 while (m != NULL && len > 0) { 1357 if (m->m_len <= len) { 1358 len -= m->m_len; 1359 m->m_len = 0; 1360 m = m->m_next; 1361 } else { 1362 m->m_len -= len; 1363 m->m_data += len; 1364 len = 0; 1365 } 1366 } 1367 m = mp; 1368 if (mp->m_flags & M_PKTHDR) 1369 m->m_pkthdr.len -= (req_len - len); 1370 } else { 1371 /* 1372 * Trim from tail. Scan the mbuf chain, 1373 * calculating its length and finding the last mbuf. 1374 * If the adjustment only affects this mbuf, then just 1375 * adjust and return. Otherwise, rescan and truncate 1376 * after the remaining size. 1377 */ 1378 len = -len; 1379 count = 0; 1380 for (;;) { 1381 count += m->m_len; 1382 if (m->m_next == NULL) 1383 break; 1384 m = m->m_next; 1385 } 1386 if (m->m_len >= len) { 1387 m->m_len -= len; 1388 if (mp->m_flags & M_PKTHDR) 1389 mp->m_pkthdr.len -= len; 1390 return; 1391 } 1392 count -= len; 1393 if (count < 0) 1394 count = 0; 1395 /* 1396 * Correct length for chain is "count". 1397 * Find the mbuf with last data, adjust its length, 1398 * and toss data from remaining mbufs on chain. 1399 */ 1400 m = mp; 1401 if (m->m_flags & M_PKTHDR) 1402 m->m_pkthdr.len = count; 1403 for (; m; m = m->m_next) { 1404 if (m->m_len >= count) { 1405 m->m_len = count; 1406 break; 1407 } 1408 count -= m->m_len; 1409 } 1410 while (m->m_next) 1411 (m = m->m_next) ->m_len = 0; 1412 } 1413 } 1414 1415 /* 1416 * Set the m_data pointer of a newly-allocated mbuf 1417 * to place an object of the specified size at the 1418 * end of the mbuf, longword aligned. 1419 */ 1420 void 1421 m_align(struct mbuf *m, int len) 1422 { 1423 int adjust; 1424 1425 if (m->m_flags & M_EXT) 1426 adjust = m->m_ext.ext_size - len; 1427 else if (m->m_flags & M_PKTHDR) 1428 adjust = MHLEN - len; 1429 else 1430 adjust = MLEN - len; 1431 m->m_data += adjust &~ (sizeof(long)-1); 1432 } 1433 1434 /* 1435 * Rearrange an mbuf chain so that len bytes are contiguous 1436 * and in the data area of an mbuf (so that mtod will work for a structure 1437 * of size len). Returns the resulting mbuf chain on success, frees it and 1438 * returns null on failure. If there is room, it will add up to 1439 * max_protohdr-len extra bytes to the contiguous region in an attempt to 1440 * avoid being called next time. 1441 */ 1442 struct mbuf * 1443 m_pullup(struct mbuf *n, int len) 1444 { 1445 struct mbuf *m; 1446 int count; 1447 int space; 1448 1449 /* 1450 * If first mbuf has no cluster, and has room for len bytes 1451 * without shifting current data, pullup into it, 1452 * otherwise allocate a new mbuf to prepend to the chain. 1453 */ 1454 if (!(n->m_flags & M_EXT) && 1455 n->m_data + len < &n->m_dat[MLEN] && 1456 n->m_next) { 1457 if (n->m_len >= len) 1458 return (n); 1459 m = n; 1460 n = n->m_next; 1461 len -= m->m_len; 1462 } else { 1463 if (len > MHLEN) 1464 goto bad; 1465 if (n->m_flags & M_PKTHDR) 1466 m = m_gethdr(MB_DONTWAIT, n->m_type); 1467 else 1468 m = m_get(MB_DONTWAIT, n->m_type); 1469 if (m == NULL) 1470 goto bad; 1471 m->m_len = 0; 1472 if (n->m_flags & M_PKTHDR) 1473 M_MOVE_PKTHDR(m, n); 1474 } 1475 space = &m->m_dat[MLEN] - (m->m_data + m->m_len); 1476 do { 1477 count = min(min(max(len, max_protohdr), space), n->m_len); 1478 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len, 1479 (unsigned)count); 1480 len -= count; 1481 m->m_len += count; 1482 n->m_len -= count; 1483 space -= count; 1484 if (n->m_len) 1485 n->m_data += count; 1486 else 1487 n = m_free(n); 1488 } while (len > 0 && n); 1489 if (len > 0) { 1490 m_free(m); 1491 goto bad; 1492 } 1493 m->m_next = n; 1494 return (m); 1495 bad: 1496 m_freem(n); 1497 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1); 1498 return (NULL); 1499 } 1500 1501 /* 1502 * Partition an mbuf chain in two pieces, returning the tail -- 1503 * all but the first len0 bytes. In case of failure, it returns NULL and 1504 * attempts to restore the chain to its original state. 1505 * 1506 * Note that the resulting mbufs might be read-only, because the new 1507 * mbuf can end up sharing an mbuf cluster with the original mbuf if 1508 * the "breaking point" happens to lie within a cluster mbuf. Use the 1509 * M_WRITABLE() macro to check for this case. 1510 */ 1511 struct mbuf * 1512 m_split(struct mbuf *m0, int len0, int wait) 1513 { 1514 struct mbuf *m, *n; 1515 unsigned len = len0, remain; 1516 1517 for (m = m0; m && len > m->m_len; m = m->m_next) 1518 len -= m->m_len; 1519 if (m == NULL) 1520 return (NULL); 1521 remain = m->m_len - len; 1522 if (m0->m_flags & M_PKTHDR) { 1523 n = m_gethdr(wait, m0->m_type); 1524 if (n == NULL) 1525 return (NULL); 1526 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif; 1527 n->m_pkthdr.len = m0->m_pkthdr.len - len0; 1528 m0->m_pkthdr.len = len0; 1529 if (m->m_flags & M_EXT) 1530 goto extpacket; 1531 if (remain > MHLEN) { 1532 /* m can't be the lead packet */ 1533 MH_ALIGN(n, 0); 1534 n->m_next = m_split(m, len, wait); 1535 if (n->m_next == NULL) { 1536 m_free(n); 1537 return (NULL); 1538 } else { 1539 n->m_len = 0; 1540 return (n); 1541 } 1542 } else 1543 MH_ALIGN(n, remain); 1544 } else if (remain == 0) { 1545 n = m->m_next; 1546 m->m_next = 0; 1547 return (n); 1548 } else { 1549 n = m_get(wait, m->m_type); 1550 if (n == NULL) 1551 return (NULL); 1552 M_ALIGN(n, remain); 1553 } 1554 extpacket: 1555 if (m->m_flags & M_EXT) { 1556 KKASSERT((n->m_flags & M_EXT) == 0); 1557 n->m_data = m->m_data + len; 1558 m->m_ext.ext_ref(m->m_ext.ext_arg); 1559 n->m_ext = m->m_ext; 1560 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER); 1561 } else { 1562 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain); 1563 } 1564 n->m_len = remain; 1565 m->m_len = len; 1566 n->m_next = m->m_next; 1567 m->m_next = 0; 1568 return (n); 1569 } 1570 1571 /* 1572 * Routine to copy from device local memory into mbufs. 1573 * Note: "offset" is ill-defined and always called as 0, so ignore it. 1574 */ 1575 struct mbuf * 1576 m_devget(char *buf, int len, int offset, struct ifnet *ifp, 1577 void (*copy)(volatile const void *from, volatile void *to, size_t length)) 1578 { 1579 struct mbuf *m, *mfirst = NULL, **mtail; 1580 int nsize, flags; 1581 1582 if (copy == NULL) 1583 copy = bcopy; 1584 mtail = &mfirst; 1585 flags = M_PKTHDR; 1586 1587 while (len > 0) { 1588 m = m_getl(len, MB_DONTWAIT, MT_DATA, flags, &nsize); 1589 if (m == NULL) { 1590 m_freem(mfirst); 1591 return (NULL); 1592 } 1593 m->m_len = min(len, nsize); 1594 1595 if (flags & M_PKTHDR) { 1596 if (len + max_linkhdr <= nsize) 1597 m->m_data += max_linkhdr; 1598 m->m_pkthdr.rcvif = ifp; 1599 m->m_pkthdr.len = len; 1600 flags = 0; 1601 } 1602 1603 copy(buf, m->m_data, (unsigned)m->m_len); 1604 buf += m->m_len; 1605 len -= m->m_len; 1606 *mtail = m; 1607 mtail = &m->m_next; 1608 } 1609 1610 return (mfirst); 1611 } 1612 1613 /* 1614 * Routine to pad mbuf to the specified length 'padto'. 1615 */ 1616 int 1617 m_devpad(struct mbuf *m, int padto) 1618 { 1619 struct mbuf *last = NULL; 1620 int padlen; 1621 1622 if (padto <= m->m_pkthdr.len) 1623 return 0; 1624 1625 padlen = padto - m->m_pkthdr.len; 1626 1627 /* if there's only the packet-header and we can pad there, use it. */ 1628 if (m->m_pkthdr.len == m->m_len && M_TRAILINGSPACE(m) >= padlen) { 1629 last = m; 1630 } else { 1631 /* 1632 * Walk packet chain to find last mbuf. We will either 1633 * pad there, or append a new mbuf and pad it 1634 */ 1635 for (last = m; last->m_next != NULL; last = last->m_next) 1636 ; /* EMPTY */ 1637 1638 /* `last' now points to last in chain. */ 1639 if (M_TRAILINGSPACE(last) < padlen) { 1640 struct mbuf *n; 1641 1642 /* Allocate new empty mbuf, pad it. Compact later. */ 1643 MGET(n, MB_DONTWAIT, MT_DATA); 1644 if (n == NULL) 1645 return ENOBUFS; 1646 n->m_len = 0; 1647 last->m_next = n; 1648 last = n; 1649 } 1650 } 1651 KKASSERT(M_TRAILINGSPACE(last) >= padlen); 1652 KKASSERT(M_WRITABLE(last)); 1653 1654 /* Now zero the pad area */ 1655 bzero(mtod(last, char *) + last->m_len, padlen); 1656 last->m_len += padlen; 1657 m->m_pkthdr.len += padlen; 1658 return 0; 1659 } 1660 1661 /* 1662 * Copy data from a buffer back into the indicated mbuf chain, 1663 * starting "off" bytes from the beginning, extending the mbuf 1664 * chain if necessary. 1665 */ 1666 void 1667 m_copyback(struct mbuf *m0, int off, int len, caddr_t cp) 1668 { 1669 int mlen; 1670 struct mbuf *m = m0, *n; 1671 int totlen = 0; 1672 1673 if (m0 == NULL) 1674 return; 1675 while (off > (mlen = m->m_len)) { 1676 off -= mlen; 1677 totlen += mlen; 1678 if (m->m_next == NULL) { 1679 n = m_getclr(MB_DONTWAIT, m->m_type); 1680 if (n == NULL) 1681 goto out; 1682 n->m_len = min(MLEN, len + off); 1683 m->m_next = n; 1684 } 1685 m = m->m_next; 1686 } 1687 while (len > 0) { 1688 mlen = min (m->m_len - off, len); 1689 bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen); 1690 cp += mlen; 1691 len -= mlen; 1692 mlen += off; 1693 off = 0; 1694 totlen += mlen; 1695 if (len == 0) 1696 break; 1697 if (m->m_next == NULL) { 1698 n = m_get(MB_DONTWAIT, m->m_type); 1699 if (n == NULL) 1700 break; 1701 n->m_len = min(MLEN, len); 1702 m->m_next = n; 1703 } 1704 m = m->m_next; 1705 } 1706 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen)) 1707 m->m_pkthdr.len = totlen; 1708 } 1709 1710 /* 1711 * Append the specified data to the indicated mbuf chain, 1712 * Extend the mbuf chain if the new data does not fit in 1713 * existing space. 1714 * 1715 * Return 1 if able to complete the job; otherwise 0. 1716 */ 1717 int 1718 m_append(struct mbuf *m0, int len, c_caddr_t cp) 1719 { 1720 struct mbuf *m, *n; 1721 int remainder, space; 1722 1723 for (m = m0; m->m_next != NULL; m = m->m_next) 1724 ; 1725 remainder = len; 1726 space = M_TRAILINGSPACE(m); 1727 if (space > 0) { 1728 /* 1729 * Copy into available space. 1730 */ 1731 if (space > remainder) 1732 space = remainder; 1733 bcopy(cp, mtod(m, caddr_t) + m->m_len, space); 1734 m->m_len += space; 1735 cp += space, remainder -= space; 1736 } 1737 while (remainder > 0) { 1738 /* 1739 * Allocate a new mbuf; could check space 1740 * and allocate a cluster instead. 1741 */ 1742 n = m_get(MB_DONTWAIT, m->m_type); 1743 if (n == NULL) 1744 break; 1745 n->m_len = min(MLEN, remainder); 1746 bcopy(cp, mtod(n, caddr_t), n->m_len); 1747 cp += n->m_len, remainder -= n->m_len; 1748 m->m_next = n; 1749 m = n; 1750 } 1751 if (m0->m_flags & M_PKTHDR) 1752 m0->m_pkthdr.len += len - remainder; 1753 return (remainder == 0); 1754 } 1755 1756 /* 1757 * Apply function f to the data in an mbuf chain starting "off" bytes from 1758 * the beginning, continuing for "len" bytes. 1759 */ 1760 int 1761 m_apply(struct mbuf *m, int off, int len, 1762 int (*f)(void *, void *, u_int), void *arg) 1763 { 1764 u_int count; 1765 int rval; 1766 1767 KASSERT(off >= 0, ("m_apply, negative off %d", off)); 1768 KASSERT(len >= 0, ("m_apply, negative len %d", len)); 1769 while (off > 0) { 1770 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain")); 1771 if (off < m->m_len) 1772 break; 1773 off -= m->m_len; 1774 m = m->m_next; 1775 } 1776 while (len > 0) { 1777 KASSERT(m != NULL, ("m_apply, offset > size of mbuf chain")); 1778 count = min(m->m_len - off, len); 1779 rval = (*f)(arg, mtod(m, caddr_t) + off, count); 1780 if (rval) 1781 return (rval); 1782 len -= count; 1783 off = 0; 1784 m = m->m_next; 1785 } 1786 return (0); 1787 } 1788 1789 /* 1790 * Return a pointer to mbuf/offset of location in mbuf chain. 1791 */ 1792 struct mbuf * 1793 m_getptr(struct mbuf *m, int loc, int *off) 1794 { 1795 1796 while (loc >= 0) { 1797 /* Normal end of search. */ 1798 if (m->m_len > loc) { 1799 *off = loc; 1800 return (m); 1801 } else { 1802 loc -= m->m_len; 1803 if (m->m_next == NULL) { 1804 if (loc == 0) { 1805 /* Point at the end of valid data. */ 1806 *off = m->m_len; 1807 return (m); 1808 } 1809 return (NULL); 1810 } 1811 m = m->m_next; 1812 } 1813 } 1814 return (NULL); 1815 } 1816 1817 void 1818 m_print(const struct mbuf *m) 1819 { 1820 int len; 1821 const struct mbuf *m2; 1822 1823 len = m->m_pkthdr.len; 1824 m2 = m; 1825 while (len) { 1826 kprintf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-"); 1827 len -= m2->m_len; 1828 m2 = m2->m_next; 1829 } 1830 return; 1831 } 1832 1833 /* 1834 * "Move" mbuf pkthdr from "from" to "to". 1835 * "from" must have M_PKTHDR set, and "to" must be empty. 1836 */ 1837 void 1838 m_move_pkthdr(struct mbuf *to, struct mbuf *from) 1839 { 1840 KASSERT((to->m_flags & M_PKTHDR), ("m_move_pkthdr: not packet header")); 1841 1842 to->m_flags |= from->m_flags & M_COPYFLAGS; 1843 to->m_pkthdr = from->m_pkthdr; /* especially tags */ 1844 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */ 1845 } 1846 1847 /* 1848 * Duplicate "from"'s mbuf pkthdr in "to". 1849 * "from" must have M_PKTHDR set, and "to" must be empty. 1850 * In particular, this does a deep copy of the packet tags. 1851 */ 1852 int 1853 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how) 1854 { 1855 KASSERT((to->m_flags & M_PKTHDR), ("m_dup_pkthdr: not packet header")); 1856 1857 to->m_flags = (from->m_flags & M_COPYFLAGS) | 1858 (to->m_flags & ~M_COPYFLAGS); 1859 to->m_pkthdr = from->m_pkthdr; 1860 SLIST_INIT(&to->m_pkthdr.tags); 1861 return (m_tag_copy_chain(to, from, how)); 1862 } 1863 1864 /* 1865 * Defragment a mbuf chain, returning the shortest possible 1866 * chain of mbufs and clusters. If allocation fails and 1867 * this cannot be completed, NULL will be returned, but 1868 * the passed in chain will be unchanged. Upon success, 1869 * the original chain will be freed, and the new chain 1870 * will be returned. 1871 * 1872 * If a non-packet header is passed in, the original 1873 * mbuf (chain?) will be returned unharmed. 1874 * 1875 * m_defrag_nofree doesn't free the passed in mbuf. 1876 */ 1877 struct mbuf * 1878 m_defrag(struct mbuf *m0, int how) 1879 { 1880 struct mbuf *m_new; 1881 1882 if ((m_new = m_defrag_nofree(m0, how)) == NULL) 1883 return (NULL); 1884 if (m_new != m0) 1885 m_freem(m0); 1886 return (m_new); 1887 } 1888 1889 struct mbuf * 1890 m_defrag_nofree(struct mbuf *m0, int how) 1891 { 1892 struct mbuf *m_new = NULL, *m_final = NULL; 1893 int progress = 0, length, nsize; 1894 1895 if (!(m0->m_flags & M_PKTHDR)) 1896 return (m0); 1897 1898 #ifdef MBUF_STRESS_TEST 1899 if (m_defragrandomfailures) { 1900 int temp = karc4random() & 0xff; 1901 if (temp == 0xba) 1902 goto nospace; 1903 } 1904 #endif 1905 1906 m_final = m_getl(m0->m_pkthdr.len, how, MT_DATA, M_PKTHDR, &nsize); 1907 if (m_final == NULL) 1908 goto nospace; 1909 m_final->m_len = 0; /* in case m0->m_pkthdr.len is zero */ 1910 1911 if (m_dup_pkthdr(m_final, m0, how) == 0) 1912 goto nospace; 1913 1914 m_new = m_final; 1915 1916 while (progress < m0->m_pkthdr.len) { 1917 length = m0->m_pkthdr.len - progress; 1918 if (length > MCLBYTES) 1919 length = MCLBYTES; 1920 1921 if (m_new == NULL) { 1922 m_new = m_getl(length, how, MT_DATA, 0, &nsize); 1923 if (m_new == NULL) 1924 goto nospace; 1925 } 1926 1927 m_copydata(m0, progress, length, mtod(m_new, caddr_t)); 1928 progress += length; 1929 m_new->m_len = length; 1930 if (m_new != m_final) 1931 m_cat(m_final, m_new); 1932 m_new = NULL; 1933 } 1934 if (m0->m_next == NULL) 1935 m_defraguseless++; 1936 m_defragpackets++; 1937 m_defragbytes += m_final->m_pkthdr.len; 1938 return (m_final); 1939 nospace: 1940 m_defragfailure++; 1941 if (m_new) 1942 m_free(m_new); 1943 m_freem(m_final); 1944 return (NULL); 1945 } 1946 1947 /* 1948 * Move data from uio into mbufs. 1949 */ 1950 struct mbuf * 1951 m_uiomove(struct uio *uio) 1952 { 1953 struct mbuf *m; /* current working mbuf */ 1954 struct mbuf *head = NULL; /* result mbuf chain */ 1955 struct mbuf **mp = &head; 1956 int flags = M_PKTHDR; 1957 int nsize; 1958 int error; 1959 int resid; 1960 1961 do { 1962 if (uio->uio_resid > INT_MAX) 1963 resid = INT_MAX; 1964 else 1965 resid = (int)uio->uio_resid; 1966 m = m_getl(resid, MB_WAIT, MT_DATA, flags, &nsize); 1967 if (flags) { 1968 m->m_pkthdr.len = 0; 1969 /* Leave room for protocol headers. */ 1970 if (resid < MHLEN) 1971 MH_ALIGN(m, resid); 1972 flags = 0; 1973 } 1974 m->m_len = imin(nsize, resid); 1975 error = uiomove(mtod(m, caddr_t), m->m_len, uio); 1976 if (error) { 1977 m_free(m); 1978 goto failed; 1979 } 1980 *mp = m; 1981 mp = &m->m_next; 1982 head->m_pkthdr.len += m->m_len; 1983 } while (uio->uio_resid > 0); 1984 1985 return (head); 1986 1987 failed: 1988 m_freem(head); 1989 return (NULL); 1990 } 1991 1992 struct mbuf * 1993 m_last(struct mbuf *m) 1994 { 1995 while (m->m_next) 1996 m = m->m_next; 1997 return (m); 1998 } 1999 2000 /* 2001 * Return the number of bytes in an mbuf chain. 2002 * If lastm is not NULL, also return the last mbuf. 2003 */ 2004 u_int 2005 m_lengthm(struct mbuf *m, struct mbuf **lastm) 2006 { 2007 u_int len = 0; 2008 struct mbuf *prev = m; 2009 2010 while (m) { 2011 len += m->m_len; 2012 prev = m; 2013 m = m->m_next; 2014 } 2015 if (lastm != NULL) 2016 *lastm = prev; 2017 return (len); 2018 } 2019 2020 /* 2021 * Like m_lengthm(), except also keep track of mbuf usage. 2022 */ 2023 u_int 2024 m_countm(struct mbuf *m, struct mbuf **lastm, u_int *pmbcnt) 2025 { 2026 u_int len = 0, mbcnt = 0; 2027 struct mbuf *prev = m; 2028 2029 while (m) { 2030 len += m->m_len; 2031 mbcnt += MSIZE; 2032 if (m->m_flags & M_EXT) 2033 mbcnt += m->m_ext.ext_size; 2034 prev = m; 2035 m = m->m_next; 2036 } 2037 if (lastm != NULL) 2038 *lastm = prev; 2039 *pmbcnt = mbcnt; 2040 return (len); 2041 } 2042