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