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 1.65 2007/08/12 01:46:26 dillon Exp $ 69 */ 70 71 #include "opt_param.h" 72 #include "opt_ddb.h" 73 #include "opt_mbuf_stress_test.h" 74 #include <sys/param.h> 75 #include <sys/systm.h> 76 #include <sys/malloc.h> 77 #include <sys/mbuf.h> 78 #include <sys/kernel.h> 79 #include <sys/sysctl.h> 80 #include <sys/domain.h> 81 #include <sys/objcache.h> 82 #include <sys/tree.h> 83 #include <sys/protosw.h> 84 #include <sys/uio.h> 85 #include <sys/thread.h> 86 #include <sys/globaldata.h> 87 #include <sys/thread2.h> 88 89 #include <machine/atomic.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 return (FALSE); 482 m->m_flags |= M_CLCACHE; 483 linkcluster(m, cl); 484 return (TRUE); 485 } 486 487 static boolean_t 488 mbufcluster_ctor(void *obj, void *private, int ocflags) 489 { 490 struct mbuf *m = obj; 491 struct mbcluster *cl; 492 493 mbuf_ctor(obj, private, ocflags); 494 cl = objcache_get(mclmeta_cache, ocflags); 495 if (cl == NULL) 496 return (FALSE); 497 m->m_flags |= M_CLCACHE; 498 linkcluster(m, cl); 499 return (TRUE); 500 } 501 502 /* 503 * Used for both the cluster and cluster PHDR caches. 504 * 505 * The mbuf may have lost its cluster due to sharing, deal 506 * with the situation by checking M_EXT. 507 */ 508 static void 509 mbufcluster_dtor(void *obj, void *private) 510 { 511 struct mbuf *m = obj; 512 struct mbcluster *mcl; 513 514 if (m->m_flags & M_EXT) { 515 KKASSERT((m->m_flags & M_EXT_CLUSTER) != 0); 516 mcl = m->m_ext.ext_arg; 517 KKASSERT(mcl->mcl_refs == 1); 518 mcl->mcl_refs = 0; 519 objcache_put(mclmeta_cache, mcl); 520 } 521 } 522 523 struct objcache_malloc_args mbuf_malloc_args = { MSIZE, M_MBUF }; 524 struct objcache_malloc_args mclmeta_malloc_args = 525 { sizeof(struct mbcluster), M_MCLMETA }; 526 527 /* ARGSUSED*/ 528 static void 529 mbinit(void *dummy) 530 { 531 int i; 532 533 for (i = 0; i < ncpus; i++) 534 { 535 atomic_set_long_nonlocked(&mbstat[i].m_msize, MSIZE); 536 atomic_set_long_nonlocked(&mbstat[i].m_mclbytes, MCLBYTES); 537 atomic_set_long_nonlocked(&mbstat[i].m_minclsize, MINCLSIZE); 538 atomic_set_long_nonlocked(&mbstat[i].m_mlen, MLEN); 539 atomic_set_long_nonlocked(&mbstat[i].m_mhlen, MHLEN); 540 } 541 542 mbuf_cache = objcache_create("mbuf", nmbufs, 0, 543 mbuf_ctor, NULL, NULL, 544 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 545 mbufphdr_cache = objcache_create("mbuf pkt hdr", nmbufs, 64, 546 mbufphdr_ctor, NULL, NULL, 547 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 548 mclmeta_cache = objcache_create("cluster mbuf", nmbclusters , 0, 549 mclmeta_ctor, mclmeta_dtor, NULL, 550 objcache_malloc_alloc, objcache_malloc_free, &mclmeta_malloc_args); 551 mbufcluster_cache = objcache_create("mbuf + cluster", nmbclusters, 0, 552 mbufcluster_ctor, mbufcluster_dtor, NULL, 553 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 554 mbufphdrcluster_cache = objcache_create("mbuf pkt hdr + cluster", 555 nmbclusters, 64, mbufphdrcluster_ctor, mbufcluster_dtor, NULL, 556 objcache_malloc_alloc, objcache_malloc_free, &mbuf_malloc_args); 557 return; 558 } 559 560 /* 561 * Return the number of references to this mbuf's data. 0 is returned 562 * if the mbuf is not M_EXT, a reference count is returned if it is 563 * M_EXT | M_EXT_CLUSTER, and 99 is returned if it is a special M_EXT. 564 */ 565 int 566 m_sharecount(struct mbuf *m) 567 { 568 switch (m->m_flags & (M_EXT | M_EXT_CLUSTER)) { 569 case 0: 570 return (0); 571 case M_EXT: 572 return (99); 573 case M_EXT | M_EXT_CLUSTER: 574 return (((struct mbcluster *)m->m_ext.ext_arg)->mcl_refs); 575 } 576 /* NOTREACHED */ 577 return (0); /* to shut up compiler */ 578 } 579 580 /* 581 * change mbuf to new type 582 */ 583 void 584 m_chtype(struct mbuf *m, int type) 585 { 586 struct globaldata *gd = mycpu; 587 588 atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1); 589 atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1); 590 atomic_set_short_nonlocked(&m->m_type, type); 591 } 592 593 static void 594 m_reclaim(void) 595 { 596 struct domain *dp; 597 struct protosw *pr; 598 599 crit_enter(); 600 SLIST_FOREACH(dp, &domains, dom_next) { 601 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) { 602 if (pr->pr_drain) 603 (*pr->pr_drain)(); 604 } 605 } 606 crit_exit(); 607 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_drain, 1); 608 } 609 610 static void __inline 611 updatestats(struct mbuf *m, int type) 612 { 613 struct globaldata *gd = mycpu; 614 m->m_type = type; 615 616 mbuftrack(m); 617 618 atomic_add_long_nonlocked(&mbtypes[gd->gd_cpuid][type], 1); 619 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1); 620 621 } 622 623 /* 624 * Allocate an mbuf. 625 */ 626 struct mbuf * 627 m_get(int how, int type) 628 { 629 struct mbuf *m; 630 int ntries = 0; 631 int ocf = MBTOM(how); 632 633 retryonce: 634 635 m = objcache_get(mbuf_cache, ocf); 636 637 if (m == NULL) { 638 if ((how & MB_TRYWAIT) && ntries++ == 0) { 639 struct objcache *reclaimlist[] = { 640 mbufphdr_cache, 641 mbufcluster_cache, mbufphdrcluster_cache 642 }; 643 const int nreclaims = __arysize(reclaimlist); 644 645 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf)) 646 m_reclaim(); 647 goto retryonce; 648 } 649 return (NULL); 650 } 651 652 updatestats(m, type); 653 return (m); 654 } 655 656 struct mbuf * 657 m_gethdr(int how, int type) 658 { 659 struct mbuf *m; 660 int ocf = MBTOM(how); 661 int ntries = 0; 662 663 retryonce: 664 665 m = objcache_get(mbufphdr_cache, ocf); 666 667 if (m == NULL) { 668 if ((how & MB_TRYWAIT) && ntries++ == 0) { 669 struct objcache *reclaimlist[] = { 670 mbuf_cache, 671 mbufcluster_cache, mbufphdrcluster_cache 672 }; 673 const int nreclaims = __arysize(reclaimlist); 674 675 if (!objcache_reclaimlist(reclaimlist, nreclaims, ocf)) 676 m_reclaim(); 677 goto retryonce; 678 } 679 return (NULL); 680 } 681 682 updatestats(m, type); 683 return (m); 684 } 685 686 /* 687 * Get a mbuf (not a mbuf cluster!) and zero it. 688 * Deprecated. 689 */ 690 struct mbuf * 691 m_getclr(int how, int type) 692 { 693 struct mbuf *m; 694 695 m = m_get(how, type); 696 if (m != NULL) 697 bzero(m->m_data, MLEN); 698 return (m); 699 } 700 701 /* 702 * Returns an mbuf with an attached cluster. 703 * Because many network drivers use this kind of buffers a lot, it is 704 * convenient to keep a small pool of free buffers of this kind. 705 * Even a small size such as 10 gives about 10% improvement in the 706 * forwarding rate in a bridge or router. 707 */ 708 struct mbuf * 709 m_getcl(int how, short type, int flags) 710 { 711 struct mbuf *m; 712 int ocflags = MBTOM(how); 713 int ntries = 0; 714 715 retryonce: 716 717 if (flags & M_PKTHDR) 718 m = objcache_get(mbufphdrcluster_cache, ocflags); 719 else 720 m = objcache_get(mbufcluster_cache, ocflags); 721 722 if (m == NULL) { 723 if ((how & MB_TRYWAIT) && ntries++ == 0) { 724 struct objcache *reclaimlist[1]; 725 726 if (flags & M_PKTHDR) 727 reclaimlist[0] = mbufcluster_cache; 728 else 729 reclaimlist[0] = mbufphdrcluster_cache; 730 if (!objcache_reclaimlist(reclaimlist, 1, ocflags)) 731 m_reclaim(); 732 goto retryonce; 733 } 734 return (NULL); 735 } 736 737 m->m_type = type; 738 739 mbuftrack(m); 740 741 atomic_add_long_nonlocked(&mbtypes[mycpu->gd_cpuid][type], 1); 742 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1); 743 return (m); 744 } 745 746 /* 747 * Allocate chain of requested length. 748 */ 749 struct mbuf * 750 m_getc(int len, int how, int type) 751 { 752 struct mbuf *n, *nfirst = NULL, **ntail = &nfirst; 753 int nsize; 754 755 while (len > 0) { 756 n = m_getl(len, how, type, 0, &nsize); 757 if (n == NULL) 758 goto failed; 759 n->m_len = 0; 760 *ntail = n; 761 ntail = &n->m_next; 762 len -= nsize; 763 } 764 return (nfirst); 765 766 failed: 767 m_freem(nfirst); 768 return (NULL); 769 } 770 771 /* 772 * Allocate len-worth of mbufs and/or mbuf clusters (whatever fits best) 773 * and return a pointer to the head of the allocated chain. If m0 is 774 * non-null, then we assume that it is a single mbuf or an mbuf chain to 775 * which we want len bytes worth of mbufs and/or clusters attached, and so 776 * if we succeed in allocating it, we will just return a pointer to m0. 777 * 778 * If we happen to fail at any point during the allocation, we will free 779 * up everything we have already allocated and return NULL. 780 * 781 * Deprecated. Use m_getc() and m_cat() instead. 782 */ 783 struct mbuf * 784 m_getm(struct mbuf *m0, int len, int type, int how) 785 { 786 struct mbuf *nfirst; 787 788 nfirst = m_getc(len, how, type); 789 790 if (m0 != NULL) { 791 m_last(m0)->m_next = nfirst; 792 return (m0); 793 } 794 795 return (nfirst); 796 } 797 798 /* 799 * Adds a cluster to a normal mbuf, M_EXT is set on success. 800 * Deprecated. Use m_getcl() instead. 801 */ 802 void 803 m_mclget(struct mbuf *m, int how) 804 { 805 struct mbcluster *mcl; 806 807 KKASSERT((m->m_flags & M_EXT) == 0); 808 mcl = objcache_get(mclmeta_cache, MBTOM(how)); 809 if (mcl != NULL) { 810 linkcluster(m, mcl); 811 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1); 812 } 813 } 814 815 /* 816 * Updates to mbcluster must be MPSAFE. Only an entity which already has 817 * a reference to the cluster can ref it, so we are in no danger of 818 * racing an add with a subtract. But the operation must still be atomic 819 * since multiple entities may have a reference on the cluster. 820 * 821 * m_mclfree() is almost the same but it must contend with two entities 822 * freeing the cluster at the same time. If there is only one reference 823 * count we are the only entity referencing the cluster and no further 824 * locking is required. Otherwise we must protect against a race to 0 825 * with the serializer. 826 */ 827 static void 828 m_mclref(void *arg) 829 { 830 struct mbcluster *mcl = arg; 831 832 atomic_add_int(&mcl->mcl_refs, 1); 833 } 834 835 /* 836 * When dereferencing a cluster we have to deal with a N->0 race, where 837 * N entities free their references simultaniously. To do this we use 838 * atomic_cmpset_int(). 839 */ 840 static void 841 m_mclfree(void *arg) 842 { 843 struct mbcluster *mcl = arg; 844 int refs; 845 846 do { 847 refs = mcl->mcl_refs; 848 } while (atomic_cmpset_int(&mcl->mcl_refs, refs, refs - 1) == 0); 849 if (refs == 1) 850 objcache_put(mclmeta_cache, mcl); 851 } 852 853 extern void db_print_backtrace(void); 854 855 /* 856 * Free a single mbuf and any associated external storage. The successor, 857 * if any, is returned. 858 * 859 * We do need to check non-first mbuf for m_aux, since some of existing 860 * code does not call M_PREPEND properly. 861 * (example: call to bpf_mtap from drivers) 862 */ 863 struct mbuf * 864 m_free(struct mbuf *m) 865 { 866 struct mbuf *n; 867 struct globaldata *gd = mycpu; 868 869 KASSERT(m->m_type != MT_FREE, ("freeing free mbuf %p", m)); 870 atomic_subtract_long_nonlocked(&mbtypes[gd->gd_cpuid][m->m_type], 1); 871 872 n = m->m_next; 873 874 /* 875 * Make sure the mbuf is in constructed state before returning it 876 * to the objcache. 877 */ 878 m->m_next = NULL; 879 mbufuntrack(m); 880 #ifdef notyet 881 KKASSERT(m->m_nextpkt == NULL); 882 #else 883 if (m->m_nextpkt != NULL) { 884 #ifdef DDB 885 static int afewtimes = 10; 886 887 if (afewtimes-- > 0) { 888 kprintf("mfree: m->m_nextpkt != NULL\n"); 889 db_print_backtrace(); 890 } 891 #endif 892 m->m_nextpkt = NULL; 893 } 894 #endif 895 if (m->m_flags & M_PKTHDR) { 896 m_tag_delete_chain(m); /* eliminate XXX JH */ 897 } 898 899 m->m_flags &= (M_EXT | M_EXT_CLUSTER | M_CLCACHE | M_PHCACHE); 900 901 /* 902 * Clean the M_PKTHDR state so we can return the mbuf to its original 903 * cache. This is based on the PHCACHE flag which tells us whether 904 * the mbuf was originally allocated out of a packet-header cache 905 * or a non-packet-header cache. 906 */ 907 if (m->m_flags & M_PHCACHE) { 908 m->m_flags |= M_PKTHDR; 909 m->m_pkthdr.rcvif = NULL; /* eliminate XXX JH */ 910 m->m_pkthdr.csum_flags = 0; /* eliminate XXX JH */ 911 m->m_pkthdr.fw_flags = 0; /* eliminate XXX JH */ 912 SLIST_INIT(&m->m_pkthdr.tags); 913 } 914 915 /* 916 * Handle remaining flags combinations. M_CLCACHE tells us whether 917 * the mbuf was originally allocated from a cluster cache or not, 918 * and is totally separate from whether the mbuf is currently 919 * associated with a cluster. 920 */ 921 crit_enter(); 922 switch(m->m_flags & (M_CLCACHE | M_EXT | M_EXT_CLUSTER)) { 923 case M_CLCACHE | M_EXT | M_EXT_CLUSTER: 924 /* 925 * mbuf+cluster cache case. The mbuf was allocated from the 926 * combined mbuf_cluster cache and can be returned to the 927 * cache if the cluster hasn't been shared. 928 */ 929 if (m_sharecount(m) == 1) { 930 /* 931 * The cluster has not been shared, we can just 932 * reset the data pointer and return the mbuf 933 * to the cluster cache. Note that the reference 934 * count is left intact (it is still associated with 935 * an mbuf). 936 */ 937 m->m_data = m->m_ext.ext_buf; 938 if (m->m_flags & M_PHCACHE) 939 objcache_put(mbufphdrcluster_cache, m); 940 else 941 objcache_put(mbufcluster_cache, m); 942 atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1); 943 } else { 944 /* 945 * Hell. Someone else has a ref on this cluster, 946 * we have to disconnect it which means we can't 947 * put it back into the mbufcluster_cache, we 948 * have to destroy the mbuf. 949 * 950 * Other mbuf references to the cluster will typically 951 * be M_EXT | M_EXT_CLUSTER but without M_CLCACHE. 952 * 953 * XXX we could try to connect another cluster to 954 * it. 955 */ 956 m->m_ext.ext_free(m->m_ext.ext_arg); 957 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER); 958 if (m->m_flags & M_PHCACHE) 959 objcache_dtor(mbufphdrcluster_cache, m); 960 else 961 objcache_dtor(mbufcluster_cache, m); 962 } 963 break; 964 case M_EXT | M_EXT_CLUSTER: 965 /* 966 * Normal cluster associated with an mbuf that was allocated 967 * from the normal mbuf pool rather then the cluster pool. 968 * The cluster has to be independantly disassociated from the 969 * mbuf. 970 */ 971 if (m_sharecount(m) == 1) 972 atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_clusters, 1); 973 /* fall through */ 974 case M_EXT: 975 /* 976 * Normal cluster association case, disconnect the cluster from 977 * the mbuf. The cluster may or may not be custom. 978 */ 979 m->m_ext.ext_free(m->m_ext.ext_arg); 980 m->m_flags &= ~(M_EXT | M_EXT_CLUSTER); 981 /* fall through */ 982 case 0: 983 /* 984 * return the mbuf to the mbuf cache. 985 */ 986 if (m->m_flags & M_PHCACHE) { 987 m->m_data = m->m_pktdat; 988 objcache_put(mbufphdr_cache, m); 989 } else { 990 m->m_data = m->m_dat; 991 objcache_put(mbuf_cache, m); 992 } 993 atomic_subtract_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mbufs, 1); 994 break; 995 default: 996 if (!panicstr) 997 panic("bad mbuf flags %p %08x\n", m, m->m_flags); 998 break; 999 } 1000 crit_exit(); 1001 return (n); 1002 } 1003 1004 void 1005 m_freem(struct mbuf *m) 1006 { 1007 crit_enter(); 1008 while (m) 1009 m = m_free(m); 1010 crit_exit(); 1011 } 1012 1013 /* 1014 * mbuf utility routines 1015 */ 1016 1017 /* 1018 * Lesser-used path for M_PREPEND: allocate new mbuf to prepend to chain and 1019 * copy junk along. 1020 */ 1021 struct mbuf * 1022 m_prepend(struct mbuf *m, int len, int how) 1023 { 1024 struct mbuf *mn; 1025 1026 if (m->m_flags & M_PKTHDR) 1027 mn = m_gethdr(how, m->m_type); 1028 else 1029 mn = m_get(how, m->m_type); 1030 if (mn == NULL) { 1031 m_freem(m); 1032 return (NULL); 1033 } 1034 if (m->m_flags & M_PKTHDR) 1035 M_MOVE_PKTHDR(mn, m); 1036 mn->m_next = m; 1037 m = mn; 1038 if (len < MHLEN) 1039 MH_ALIGN(m, len); 1040 m->m_len = len; 1041 return (m); 1042 } 1043 1044 /* 1045 * Make a copy of an mbuf chain starting "off0" bytes from the beginning, 1046 * continuing for "len" bytes. If len is M_COPYALL, copy to end of mbuf. 1047 * The wait parameter is a choice of MB_WAIT/MB_DONTWAIT from caller. 1048 * Note that the copy is read-only, because clusters are not copied, 1049 * only their reference counts are incremented. 1050 */ 1051 struct mbuf * 1052 m_copym(const struct mbuf *m, int off0, int len, int wait) 1053 { 1054 struct mbuf *n, **np; 1055 int off = off0; 1056 struct mbuf *top; 1057 int copyhdr = 0; 1058 1059 KASSERT(off >= 0, ("m_copym, negative off %d", off)); 1060 KASSERT(len >= 0, ("m_copym, negative len %d", len)); 1061 if (off == 0 && m->m_flags & M_PKTHDR) 1062 copyhdr = 1; 1063 while (off > 0) { 1064 KASSERT(m != NULL, ("m_copym, offset > size of mbuf chain")); 1065 if (off < m->m_len) 1066 break; 1067 off -= m->m_len; 1068 m = m->m_next; 1069 } 1070 np = ⊤ 1071 top = 0; 1072 while (len > 0) { 1073 if (m == NULL) { 1074 KASSERT(len == M_COPYALL, 1075 ("m_copym, length > size of mbuf chain")); 1076 break; 1077 } 1078 /* 1079 * Because we are sharing any cluster attachment below, 1080 * be sure to get an mbuf that does not have a cluster 1081 * associated with it. 1082 */ 1083 if (copyhdr) 1084 n = m_gethdr(wait, m->m_type); 1085 else 1086 n = m_get(wait, m->m_type); 1087 *np = n; 1088 if (n == NULL) 1089 goto nospace; 1090 if (copyhdr) { 1091 if (!m_dup_pkthdr(n, m, wait)) 1092 goto nospace; 1093 if (len == M_COPYALL) 1094 n->m_pkthdr.len -= off0; 1095 else 1096 n->m_pkthdr.len = len; 1097 copyhdr = 0; 1098 } 1099 n->m_len = min(len, m->m_len - off); 1100 if (m->m_flags & M_EXT) { 1101 KKASSERT((n->m_flags & M_EXT) == 0); 1102 n->m_data = m->m_data + off; 1103 m->m_ext.ext_ref(m->m_ext.ext_arg); 1104 n->m_ext = m->m_ext; 1105 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER); 1106 } else { 1107 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), 1108 (unsigned)n->m_len); 1109 } 1110 if (len != M_COPYALL) 1111 len -= n->m_len; 1112 off = 0; 1113 m = m->m_next; 1114 np = &n->m_next; 1115 } 1116 if (top == NULL) 1117 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1); 1118 return (top); 1119 nospace: 1120 m_freem(top); 1121 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1); 1122 return (NULL); 1123 } 1124 1125 /* 1126 * Copy an entire packet, including header (which must be present). 1127 * An optimization of the common case `m_copym(m, 0, M_COPYALL, how)'. 1128 * Note that the copy is read-only, because clusters are not copied, 1129 * only their reference counts are incremented. 1130 * Preserve alignment of the first mbuf so if the creator has left 1131 * some room at the beginning (e.g. for inserting protocol headers) 1132 * the copies also have the room available. 1133 */ 1134 struct mbuf * 1135 m_copypacket(struct mbuf *m, int how) 1136 { 1137 struct mbuf *top, *n, *o; 1138 1139 n = m_gethdr(how, m->m_type); 1140 top = n; 1141 if (!n) 1142 goto nospace; 1143 1144 if (!m_dup_pkthdr(n, m, how)) 1145 goto nospace; 1146 n->m_len = m->m_len; 1147 if (m->m_flags & M_EXT) { 1148 KKASSERT((n->m_flags & M_EXT) == 0); 1149 n->m_data = m->m_data; 1150 m->m_ext.ext_ref(m->m_ext.ext_arg); 1151 n->m_ext = m->m_ext; 1152 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER); 1153 } else { 1154 n->m_data = n->m_pktdat + (m->m_data - m->m_pktdat ); 1155 bcopy(mtod(m, char *), mtod(n, char *), n->m_len); 1156 } 1157 1158 m = m->m_next; 1159 while (m) { 1160 o = m_get(how, m->m_type); 1161 if (!o) 1162 goto nospace; 1163 1164 n->m_next = o; 1165 n = n->m_next; 1166 1167 n->m_len = m->m_len; 1168 if (m->m_flags & M_EXT) { 1169 KKASSERT((n->m_flags & M_EXT) == 0); 1170 n->m_data = m->m_data; 1171 m->m_ext.ext_ref(m->m_ext.ext_arg); 1172 n->m_ext = m->m_ext; 1173 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER); 1174 } else { 1175 bcopy(mtod(m, char *), mtod(n, char *), n->m_len); 1176 } 1177 1178 m = m->m_next; 1179 } 1180 return top; 1181 nospace: 1182 m_freem(top); 1183 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1); 1184 return (NULL); 1185 } 1186 1187 /* 1188 * Copy data from an mbuf chain starting "off" bytes from the beginning, 1189 * continuing for "len" bytes, into the indicated buffer. 1190 */ 1191 void 1192 m_copydata(const struct mbuf *m, int off, int len, caddr_t cp) 1193 { 1194 unsigned count; 1195 1196 KASSERT(off >= 0, ("m_copydata, negative off %d", off)); 1197 KASSERT(len >= 0, ("m_copydata, negative len %d", len)); 1198 while (off > 0) { 1199 KASSERT(m != NULL, ("m_copydata, offset > size of mbuf chain")); 1200 if (off < m->m_len) 1201 break; 1202 off -= m->m_len; 1203 m = m->m_next; 1204 } 1205 while (len > 0) { 1206 KASSERT(m != NULL, ("m_copydata, length > size of mbuf chain")); 1207 count = min(m->m_len - off, len); 1208 bcopy(mtod(m, caddr_t) + off, cp, count); 1209 len -= count; 1210 cp += count; 1211 off = 0; 1212 m = m->m_next; 1213 } 1214 } 1215 1216 /* 1217 * Copy a packet header mbuf chain into a completely new chain, including 1218 * copying any mbuf clusters. Use this instead of m_copypacket() when 1219 * you need a writable copy of an mbuf chain. 1220 */ 1221 struct mbuf * 1222 m_dup(struct mbuf *m, int how) 1223 { 1224 struct mbuf **p, *top = NULL; 1225 int remain, moff, nsize; 1226 1227 /* Sanity check */ 1228 if (m == NULL) 1229 return (NULL); 1230 KASSERT((m->m_flags & M_PKTHDR) != 0, ("%s: !PKTHDR", __func__)); 1231 1232 /* While there's more data, get a new mbuf, tack it on, and fill it */ 1233 remain = m->m_pkthdr.len; 1234 moff = 0; 1235 p = ⊤ 1236 while (remain > 0 || top == NULL) { /* allow m->m_pkthdr.len == 0 */ 1237 struct mbuf *n; 1238 1239 /* Get the next new mbuf */ 1240 n = m_getl(remain, how, m->m_type, top == NULL ? M_PKTHDR : 0, 1241 &nsize); 1242 if (n == NULL) 1243 goto nospace; 1244 if (top == NULL) 1245 if (!m_dup_pkthdr(n, m, how)) 1246 goto nospace0; 1247 1248 /* Link it into the new chain */ 1249 *p = n; 1250 p = &n->m_next; 1251 1252 /* Copy data from original mbuf(s) into new mbuf */ 1253 n->m_len = 0; 1254 while (n->m_len < nsize && m != NULL) { 1255 int chunk = min(nsize - n->m_len, m->m_len - moff); 1256 1257 bcopy(m->m_data + moff, n->m_data + n->m_len, chunk); 1258 moff += chunk; 1259 n->m_len += chunk; 1260 remain -= chunk; 1261 if (moff == m->m_len) { 1262 m = m->m_next; 1263 moff = 0; 1264 } 1265 } 1266 1267 /* Check correct total mbuf length */ 1268 KASSERT((remain > 0 && m != NULL) || (remain == 0 && m == NULL), 1269 ("%s: bogus m_pkthdr.len", __func__)); 1270 } 1271 return (top); 1272 1273 nospace: 1274 m_freem(top); 1275 nospace0: 1276 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1); 1277 return (NULL); 1278 } 1279 1280 /* 1281 * Concatenate mbuf chain n to m. 1282 * Both chains must be of the same type (e.g. MT_DATA). 1283 * Any m_pkthdr is not updated. 1284 */ 1285 void 1286 m_cat(struct mbuf *m, struct mbuf *n) 1287 { 1288 m = m_last(m); 1289 while (n) { 1290 if (m->m_flags & M_EXT || 1291 m->m_data + m->m_len + n->m_len >= &m->m_dat[MLEN]) { 1292 /* just join the two chains */ 1293 m->m_next = n; 1294 return; 1295 } 1296 /* splat the data from one into the other */ 1297 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len, 1298 (u_int)n->m_len); 1299 m->m_len += n->m_len; 1300 n = m_free(n); 1301 } 1302 } 1303 1304 void 1305 m_adj(struct mbuf *mp, int req_len) 1306 { 1307 int len = req_len; 1308 struct mbuf *m; 1309 int count; 1310 1311 if ((m = mp) == NULL) 1312 return; 1313 if (len >= 0) { 1314 /* 1315 * Trim from head. 1316 */ 1317 while (m != NULL && len > 0) { 1318 if (m->m_len <= len) { 1319 len -= m->m_len; 1320 m->m_len = 0; 1321 m = m->m_next; 1322 } else { 1323 m->m_len -= len; 1324 m->m_data += len; 1325 len = 0; 1326 } 1327 } 1328 m = mp; 1329 if (mp->m_flags & M_PKTHDR) 1330 m->m_pkthdr.len -= (req_len - len); 1331 } else { 1332 /* 1333 * Trim from tail. Scan the mbuf chain, 1334 * calculating its length and finding the last mbuf. 1335 * If the adjustment only affects this mbuf, then just 1336 * adjust and return. Otherwise, rescan and truncate 1337 * after the remaining size. 1338 */ 1339 len = -len; 1340 count = 0; 1341 for (;;) { 1342 count += m->m_len; 1343 if (m->m_next == (struct mbuf *)0) 1344 break; 1345 m = m->m_next; 1346 } 1347 if (m->m_len >= len) { 1348 m->m_len -= len; 1349 if (mp->m_flags & M_PKTHDR) 1350 mp->m_pkthdr.len -= len; 1351 return; 1352 } 1353 count -= len; 1354 if (count < 0) 1355 count = 0; 1356 /* 1357 * Correct length for chain is "count". 1358 * Find the mbuf with last data, adjust its length, 1359 * and toss data from remaining mbufs on chain. 1360 */ 1361 m = mp; 1362 if (m->m_flags & M_PKTHDR) 1363 m->m_pkthdr.len = count; 1364 for (; m; m = m->m_next) { 1365 if (m->m_len >= count) { 1366 m->m_len = count; 1367 break; 1368 } 1369 count -= m->m_len; 1370 } 1371 while (m->m_next) 1372 (m = m->m_next) ->m_len = 0; 1373 } 1374 } 1375 1376 /* 1377 * Rearrange an mbuf chain so that len bytes are contiguous 1378 * and in the data area of an mbuf (so that mtod will work for a structure 1379 * of size len). Returns the resulting mbuf chain on success, frees it and 1380 * returns null on failure. If there is room, it will add up to 1381 * max_protohdr-len extra bytes to the contiguous region in an attempt to 1382 * avoid being called next time. 1383 */ 1384 struct mbuf * 1385 m_pullup(struct mbuf *n, int len) 1386 { 1387 struct mbuf *m; 1388 int count; 1389 int space; 1390 1391 /* 1392 * If first mbuf has no cluster, and has room for len bytes 1393 * without shifting current data, pullup into it, 1394 * otherwise allocate a new mbuf to prepend to the chain. 1395 */ 1396 if (!(n->m_flags & M_EXT) && 1397 n->m_data + len < &n->m_dat[MLEN] && 1398 n->m_next) { 1399 if (n->m_len >= len) 1400 return (n); 1401 m = n; 1402 n = n->m_next; 1403 len -= m->m_len; 1404 } else { 1405 if (len > MHLEN) 1406 goto bad; 1407 if (n->m_flags & M_PKTHDR) 1408 m = m_gethdr(MB_DONTWAIT, n->m_type); 1409 else 1410 m = m_get(MB_DONTWAIT, n->m_type); 1411 if (m == NULL) 1412 goto bad; 1413 m->m_len = 0; 1414 if (n->m_flags & M_PKTHDR) 1415 M_MOVE_PKTHDR(m, n); 1416 } 1417 space = &m->m_dat[MLEN] - (m->m_data + m->m_len); 1418 do { 1419 count = min(min(max(len, max_protohdr), space), n->m_len); 1420 bcopy(mtod(n, caddr_t), mtod(m, caddr_t) + m->m_len, 1421 (unsigned)count); 1422 len -= count; 1423 m->m_len += count; 1424 n->m_len -= count; 1425 space -= count; 1426 if (n->m_len) 1427 n->m_data += count; 1428 else 1429 n = m_free(n); 1430 } while (len > 0 && n); 1431 if (len > 0) { 1432 m_free(m); 1433 goto bad; 1434 } 1435 m->m_next = n; 1436 return (m); 1437 bad: 1438 m_freem(n); 1439 atomic_add_long_nonlocked(&mbstat[mycpu->gd_cpuid].m_mcfail, 1); 1440 return (NULL); 1441 } 1442 1443 /* 1444 * Partition an mbuf chain in two pieces, returning the tail -- 1445 * all but the first len0 bytes. In case of failure, it returns NULL and 1446 * attempts to restore the chain to its original state. 1447 * 1448 * Note that the resulting mbufs might be read-only, because the new 1449 * mbuf can end up sharing an mbuf cluster with the original mbuf if 1450 * the "breaking point" happens to lie within a cluster mbuf. Use the 1451 * M_WRITABLE() macro to check for this case. 1452 */ 1453 struct mbuf * 1454 m_split(struct mbuf *m0, int len0, int wait) 1455 { 1456 struct mbuf *m, *n; 1457 unsigned len = len0, remain; 1458 1459 for (m = m0; m && len > m->m_len; m = m->m_next) 1460 len -= m->m_len; 1461 if (m == NULL) 1462 return (NULL); 1463 remain = m->m_len - len; 1464 if (m0->m_flags & M_PKTHDR) { 1465 n = m_gethdr(wait, m0->m_type); 1466 if (n == NULL) 1467 return (NULL); 1468 n->m_pkthdr.rcvif = m0->m_pkthdr.rcvif; 1469 n->m_pkthdr.len = m0->m_pkthdr.len - len0; 1470 m0->m_pkthdr.len = len0; 1471 if (m->m_flags & M_EXT) 1472 goto extpacket; 1473 if (remain > MHLEN) { 1474 /* m can't be the lead packet */ 1475 MH_ALIGN(n, 0); 1476 n->m_next = m_split(m, len, wait); 1477 if (n->m_next == NULL) { 1478 m_free(n); 1479 return (NULL); 1480 } else { 1481 n->m_len = 0; 1482 return (n); 1483 } 1484 } else 1485 MH_ALIGN(n, remain); 1486 } else if (remain == 0) { 1487 n = m->m_next; 1488 m->m_next = 0; 1489 return (n); 1490 } else { 1491 n = m_get(wait, m->m_type); 1492 if (n == NULL) 1493 return (NULL); 1494 M_ALIGN(n, remain); 1495 } 1496 extpacket: 1497 if (m->m_flags & M_EXT) { 1498 KKASSERT((n->m_flags & M_EXT) == 0); 1499 n->m_data = m->m_data + len; 1500 m->m_ext.ext_ref(m->m_ext.ext_arg); 1501 n->m_ext = m->m_ext; 1502 n->m_flags |= m->m_flags & (M_EXT | M_EXT_CLUSTER); 1503 } else { 1504 bcopy(mtod(m, caddr_t) + len, mtod(n, caddr_t), remain); 1505 } 1506 n->m_len = remain; 1507 m->m_len = len; 1508 n->m_next = m->m_next; 1509 m->m_next = 0; 1510 return (n); 1511 } 1512 1513 /* 1514 * Routine to copy from device local memory into mbufs. 1515 * Note: "offset" is ill-defined and always called as 0, so ignore it. 1516 */ 1517 struct mbuf * 1518 m_devget(char *buf, int len, int offset, struct ifnet *ifp, 1519 void (*copy)(volatile const void *from, volatile void *to, size_t length)) 1520 { 1521 struct mbuf *m, *mfirst = NULL, **mtail; 1522 int nsize, flags; 1523 1524 if (copy == NULL) 1525 copy = bcopy; 1526 mtail = &mfirst; 1527 flags = M_PKTHDR; 1528 1529 while (len > 0) { 1530 m = m_getl(len, MB_DONTWAIT, MT_DATA, flags, &nsize); 1531 if (m == NULL) { 1532 m_freem(mfirst); 1533 return (NULL); 1534 } 1535 m->m_len = min(len, nsize); 1536 1537 if (flags & M_PKTHDR) { 1538 if (len + max_linkhdr <= nsize) 1539 m->m_data += max_linkhdr; 1540 m->m_pkthdr.rcvif = ifp; 1541 m->m_pkthdr.len = len; 1542 flags = 0; 1543 } 1544 1545 copy(buf, m->m_data, (unsigned)m->m_len); 1546 buf += m->m_len; 1547 len -= m->m_len; 1548 *mtail = m; 1549 mtail = &m->m_next; 1550 } 1551 1552 return (mfirst); 1553 } 1554 1555 /* 1556 * Copy data from a buffer back into the indicated mbuf chain, 1557 * starting "off" bytes from the beginning, extending the mbuf 1558 * chain if necessary. 1559 */ 1560 void 1561 m_copyback(struct mbuf *m0, int off, int len, caddr_t cp) 1562 { 1563 int mlen; 1564 struct mbuf *m = m0, *n; 1565 int totlen = 0; 1566 1567 if (m0 == NULL) 1568 return; 1569 while (off > (mlen = m->m_len)) { 1570 off -= mlen; 1571 totlen += mlen; 1572 if (m->m_next == NULL) { 1573 n = m_getclr(MB_DONTWAIT, m->m_type); 1574 if (n == NULL) 1575 goto out; 1576 n->m_len = min(MLEN, len + off); 1577 m->m_next = n; 1578 } 1579 m = m->m_next; 1580 } 1581 while (len > 0) { 1582 mlen = min (m->m_len - off, len); 1583 bcopy(cp, off + mtod(m, caddr_t), (unsigned)mlen); 1584 cp += mlen; 1585 len -= mlen; 1586 mlen += off; 1587 off = 0; 1588 totlen += mlen; 1589 if (len == 0) 1590 break; 1591 if (m->m_next == NULL) { 1592 n = m_get(MB_DONTWAIT, m->m_type); 1593 if (n == NULL) 1594 break; 1595 n->m_len = min(MLEN, len); 1596 m->m_next = n; 1597 } 1598 m = m->m_next; 1599 } 1600 out: if (((m = m0)->m_flags & M_PKTHDR) && (m->m_pkthdr.len < totlen)) 1601 m->m_pkthdr.len = totlen; 1602 } 1603 1604 void 1605 m_print(const struct mbuf *m) 1606 { 1607 int len; 1608 const struct mbuf *m2; 1609 1610 len = m->m_pkthdr.len; 1611 m2 = m; 1612 while (len) { 1613 kprintf("%p %*D\n", m2, m2->m_len, (u_char *)m2->m_data, "-"); 1614 len -= m2->m_len; 1615 m2 = m2->m_next; 1616 } 1617 return; 1618 } 1619 1620 /* 1621 * "Move" mbuf pkthdr from "from" to "to". 1622 * "from" must have M_PKTHDR set, and "to" must be empty. 1623 */ 1624 void 1625 m_move_pkthdr(struct mbuf *to, struct mbuf *from) 1626 { 1627 KASSERT((to->m_flags & M_PKTHDR), ("m_move_pkthdr: not packet header")); 1628 1629 to->m_flags |= from->m_flags & M_COPYFLAGS; 1630 to->m_pkthdr = from->m_pkthdr; /* especially tags */ 1631 SLIST_INIT(&from->m_pkthdr.tags); /* purge tags from src */ 1632 } 1633 1634 /* 1635 * Duplicate "from"'s mbuf pkthdr in "to". 1636 * "from" must have M_PKTHDR set, and "to" must be empty. 1637 * In particular, this does a deep copy of the packet tags. 1638 */ 1639 int 1640 m_dup_pkthdr(struct mbuf *to, const struct mbuf *from, int how) 1641 { 1642 KASSERT((to->m_flags & M_PKTHDR), ("m_dup_pkthdr: not packet header")); 1643 1644 to->m_flags = (from->m_flags & M_COPYFLAGS) | 1645 (to->m_flags & ~M_COPYFLAGS); 1646 to->m_pkthdr = from->m_pkthdr; 1647 SLIST_INIT(&to->m_pkthdr.tags); 1648 return (m_tag_copy_chain(to, from, how)); 1649 } 1650 1651 /* 1652 * Defragment a mbuf chain, returning the shortest possible 1653 * chain of mbufs and clusters. If allocation fails and 1654 * this cannot be completed, NULL will be returned, but 1655 * the passed in chain will be unchanged. Upon success, 1656 * the original chain will be freed, and the new chain 1657 * will be returned. 1658 * 1659 * If a non-packet header is passed in, the original 1660 * mbuf (chain?) will be returned unharmed. 1661 * 1662 * m_defrag_nofree doesn't free the passed in mbuf. 1663 */ 1664 struct mbuf * 1665 m_defrag(struct mbuf *m0, int how) 1666 { 1667 struct mbuf *m_new; 1668 1669 if ((m_new = m_defrag_nofree(m0, how)) == NULL) 1670 return (NULL); 1671 if (m_new != m0) 1672 m_freem(m0); 1673 return (m_new); 1674 } 1675 1676 struct mbuf * 1677 m_defrag_nofree(struct mbuf *m0, int how) 1678 { 1679 struct mbuf *m_new = NULL, *m_final = NULL; 1680 int progress = 0, length, nsize; 1681 1682 if (!(m0->m_flags & M_PKTHDR)) 1683 return (m0); 1684 1685 #ifdef MBUF_STRESS_TEST 1686 if (m_defragrandomfailures) { 1687 int temp = karc4random() & 0xff; 1688 if (temp == 0xba) 1689 goto nospace; 1690 } 1691 #endif 1692 1693 m_final = m_getl(m0->m_pkthdr.len, how, MT_DATA, M_PKTHDR, &nsize); 1694 if (m_final == NULL) 1695 goto nospace; 1696 m_final->m_len = 0; /* in case m0->m_pkthdr.len is zero */ 1697 1698 if (m_dup_pkthdr(m_final, m0, how) == NULL) 1699 goto nospace; 1700 1701 m_new = m_final; 1702 1703 while (progress < m0->m_pkthdr.len) { 1704 length = m0->m_pkthdr.len - progress; 1705 if (length > MCLBYTES) 1706 length = MCLBYTES; 1707 1708 if (m_new == NULL) { 1709 m_new = m_getl(length, how, MT_DATA, 0, &nsize); 1710 if (m_new == NULL) 1711 goto nospace; 1712 } 1713 1714 m_copydata(m0, progress, length, mtod(m_new, caddr_t)); 1715 progress += length; 1716 m_new->m_len = length; 1717 if (m_new != m_final) 1718 m_cat(m_final, m_new); 1719 m_new = NULL; 1720 } 1721 if (m0->m_next == NULL) 1722 m_defraguseless++; 1723 m_defragpackets++; 1724 m_defragbytes += m_final->m_pkthdr.len; 1725 return (m_final); 1726 nospace: 1727 m_defragfailure++; 1728 if (m_new) 1729 m_free(m_new); 1730 m_freem(m_final); 1731 return (NULL); 1732 } 1733 1734 /* 1735 * Move data from uio into mbufs. 1736 */ 1737 struct mbuf * 1738 m_uiomove(struct uio *uio) 1739 { 1740 struct mbuf *m; /* current working mbuf */ 1741 struct mbuf *head = NULL; /* result mbuf chain */ 1742 struct mbuf **mp = &head; 1743 int resid = uio->uio_resid, nsize, flags = M_PKTHDR, error; 1744 1745 do { 1746 m = m_getl(resid, MB_WAIT, MT_DATA, flags, &nsize); 1747 if (flags) { 1748 m->m_pkthdr.len = 0; 1749 /* Leave room for protocol headers. */ 1750 if (resid < MHLEN) 1751 MH_ALIGN(m, resid); 1752 flags = 0; 1753 } 1754 m->m_len = min(nsize, resid); 1755 error = uiomove(mtod(m, caddr_t), m->m_len, uio); 1756 if (error) { 1757 m_free(m); 1758 goto failed; 1759 } 1760 *mp = m; 1761 mp = &m->m_next; 1762 head->m_pkthdr.len += m->m_len; 1763 resid -= m->m_len; 1764 } while (resid > 0); 1765 1766 return (head); 1767 1768 failed: 1769 m_freem(head); 1770 return (NULL); 1771 } 1772 1773 struct mbuf * 1774 m_last(struct mbuf *m) 1775 { 1776 while (m->m_next) 1777 m = m->m_next; 1778 return (m); 1779 } 1780 1781 /* 1782 * Return the number of bytes in an mbuf chain. 1783 * If lastm is not NULL, also return the last mbuf. 1784 */ 1785 u_int 1786 m_lengthm(struct mbuf *m, struct mbuf **lastm) 1787 { 1788 u_int len = 0; 1789 struct mbuf *prev = m; 1790 1791 while (m) { 1792 len += m->m_len; 1793 prev = m; 1794 m = m->m_next; 1795 } 1796 if (lastm != NULL) 1797 *lastm = prev; 1798 return (len); 1799 } 1800 1801 /* 1802 * Like m_lengthm(), except also keep track of mbuf usage. 1803 */ 1804 u_int 1805 m_countm(struct mbuf *m, struct mbuf **lastm, u_int *pmbcnt) 1806 { 1807 u_int len = 0, mbcnt = 0; 1808 struct mbuf *prev = m; 1809 1810 while (m) { 1811 len += m->m_len; 1812 mbcnt += MSIZE; 1813 if (m->m_flags & M_EXT) 1814 mbcnt += m->m_ext.ext_size; 1815 prev = m; 1816 m = m->m_next; 1817 } 1818 if (lastm != NULL) 1819 *lastm = prev; 1820 *pmbcnt = mbcnt; 1821 return (len); 1822 } 1823