kern_mbuf.c revision 174247
1/*- 2 * Copyright (c) 2004, 2005, 3 * Bosko Milekic <bmilekic@FreeBSD.org>. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice unmodified, this list of conditions and the following 10 * disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28#include <sys/cdefs.h> 29__FBSDID("$FreeBSD: head/sys/kern/kern_mbuf.c 174247 2007-12-04 07:06:08Z alc $"); 30 31#include "opt_mac.h" 32#include "opt_param.h" 33 34#include <sys/param.h> 35#include <sys/malloc.h> 36#include <sys/systm.h> 37#include <sys/mbuf.h> 38#include <sys/domain.h> 39#include <sys/eventhandler.h> 40#include <sys/kernel.h> 41#include <sys/protosw.h> 42#include <sys/smp.h> 43#include <sys/sysctl.h> 44 45#include <security/mac/mac_framework.h> 46 47#include <vm/vm.h> 48#include <vm/vm_page.h> 49#include <vm/uma.h> 50#include <vm/uma_int.h> 51#include <vm/uma_dbg.h> 52 53/* 54 * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA 55 * Zones. 56 * 57 * Mbuf Clusters (2K, contiguous) are allocated from the Cluster 58 * Zone. The Zone can be capped at kern.ipc.nmbclusters, if the 59 * administrator so desires. 60 * 61 * Mbufs are allocated from a UMA Master Zone called the Mbuf 62 * Zone. 63 * 64 * Additionally, FreeBSD provides a Packet Zone, which it 65 * configures as a Secondary Zone to the Mbuf Master Zone, 66 * thus sharing backend Slab kegs with the Mbuf Master Zone. 67 * 68 * Thus common-case allocations and locking are simplified: 69 * 70 * m_clget() m_getcl() 71 * | | 72 * | .------------>[(Packet Cache)] m_get(), m_gethdr() 73 * | | [ Packet ] | 74 * [(Cluster Cache)] [ Secondary ] [ (Mbuf Cache) ] 75 * [ Cluster Zone ] [ Zone ] [ Mbuf Master Zone ] 76 * | \________ | 77 * [ Cluster Keg ] \ / 78 * | [ Mbuf Keg ] 79 * [ Cluster Slabs ] | 80 * | [ Mbuf Slabs ] 81 * \____________(VM)_________________/ 82 * 83 * 84 * Whenever an object is allocated with uma_zalloc() out of 85 * one of the Zones its _ctor_ function is executed. The same 86 * for any deallocation through uma_zfree() the _dtor_ function 87 * is executed. 88 * 89 * Caches are per-CPU and are filled from the Master Zone. 90 * 91 * Whenever an object is allocated from the underlying global 92 * memory pool it gets pre-initialized with the _zinit_ functions. 93 * When the Keg's are overfull objects get decomissioned with 94 * _zfini_ functions and free'd back to the global memory pool. 95 * 96 */ 97 98int nmbclusters; /* limits number of mbuf clusters */ 99int nmbjumbop; /* limits number of page size jumbo clusters */ 100int nmbjumbo9; /* limits number of 9k jumbo clusters */ 101int nmbjumbo16; /* limits number of 16k jumbo clusters */ 102struct mbstat mbstat; 103 104static void 105tunable_mbinit(void *dummy) 106{ 107 108 /* This has to be done before VM init. */ 109 nmbclusters = 1024 + maxusers * 64; 110 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters); 111} 112SYSINIT(tunable_mbinit, SI_SUB_TUNABLES, SI_ORDER_ANY, tunable_mbinit, NULL); 113 114/* XXX: These should be tuneables. Can't change UMA limits on the fly. */ 115static int 116sysctl_nmbclusters(SYSCTL_HANDLER_ARGS) 117{ 118 int error, newnmbclusters; 119 120 newnmbclusters = nmbclusters; 121 error = sysctl_handle_int(oidp, &newnmbclusters, 0, req); 122 if (error == 0 && req->newptr) { 123 if (newnmbclusters > nmbclusters) { 124 nmbclusters = newnmbclusters; 125 uma_zone_set_max(zone_clust, nmbclusters); 126 EVENTHANDLER_INVOKE(nmbclusters_change); 127 } else 128 error = EINVAL; 129 } 130 return (error); 131} 132SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters, CTLTYPE_INT|CTLFLAG_RW, 133&nmbclusters, 0, sysctl_nmbclusters, "IU", 134 "Maximum number of mbuf clusters allowed"); 135SYSCTL_INT(_kern_ipc, OID_AUTO, nmbjumbop, CTLFLAG_RW, &nmbjumbop, 0, 136 "Maximum number of mbuf page size jumbo clusters allowed"); 137SYSCTL_INT(_kern_ipc, OID_AUTO, nmbjumbo9, CTLFLAG_RW, &nmbjumbo9, 0, 138 "Maximum number of mbuf 9k jumbo clusters allowed"); 139SYSCTL_INT(_kern_ipc, OID_AUTO, nmbjumbo16, CTLFLAG_RW, &nmbjumbo16, 0, 140 "Maximum number of mbuf 16k jumbo clusters allowed"); 141SYSCTL_STRUCT(_kern_ipc, OID_AUTO, mbstat, CTLFLAG_RD, &mbstat, mbstat, 142 "Mbuf general information and statistics"); 143 144/* 145 * Zones from which we allocate. 146 */ 147uma_zone_t zone_mbuf; 148uma_zone_t zone_clust; 149uma_zone_t zone_pack; 150uma_zone_t zone_jumbop; 151uma_zone_t zone_jumbo9; 152uma_zone_t zone_jumbo16; 153uma_zone_t zone_ext_refcnt; 154 155/* 156 * Local prototypes. 157 */ 158static int mb_ctor_mbuf(void *, int, void *, int); 159static int mb_ctor_clust(void *, int, void *, int); 160static int mb_ctor_pack(void *, int, void *, int); 161static void mb_dtor_mbuf(void *, int, void *); 162static void mb_dtor_clust(void *, int, void *); 163static void mb_dtor_pack(void *, int, void *); 164static int mb_zinit_pack(void *, int, int); 165static void mb_zfini_pack(void *, int); 166 167static void mb_reclaim(void *); 168static void mbuf_init(void *); 169static void *mbuf_jumbo_alloc(uma_zone_t, int, u_int8_t *, int); 170static void mbuf_jumbo_free(void *, int, u_int8_t); 171 172static MALLOC_DEFINE(M_JUMBOFRAME, "jumboframes", "mbuf jumbo frame buffers"); 173 174/* Ensure that MSIZE doesn't break dtom() - it must be a power of 2 */ 175CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE); 176 177/* 178 * Initialize FreeBSD Network buffer allocation. 179 */ 180SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL) 181static void 182mbuf_init(void *dummy) 183{ 184 185 /* 186 * Configure UMA zones for Mbufs, Clusters, and Packets. 187 */ 188 zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE, 189 mb_ctor_mbuf, mb_dtor_mbuf, 190#ifdef INVARIANTS 191 trash_init, trash_fini, 192#else 193 NULL, NULL, 194#endif 195 MSIZE - 1, UMA_ZONE_MAXBUCKET); 196 197 zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES, 198 mb_ctor_clust, mb_dtor_clust, 199#ifdef INVARIANTS 200 trash_init, trash_fini, 201#else 202 NULL, NULL, 203#endif 204 UMA_ALIGN_PTR, UMA_ZONE_REFCNT); 205 if (nmbclusters > 0) 206 uma_zone_set_max(zone_clust, nmbclusters); 207 208 zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack, 209 mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf); 210 211 /* Make jumbo frame zone too. Page size, 9k and 16k. */ 212 zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE, 213 mb_ctor_clust, mb_dtor_clust, 214#ifdef INVARIANTS 215 trash_init, trash_fini, 216#else 217 NULL, NULL, 218#endif 219 UMA_ALIGN_PTR, UMA_ZONE_REFCNT); 220 if (nmbjumbop > 0) 221 uma_zone_set_max(zone_jumbop, nmbjumbop); 222 223 zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES, 224 mb_ctor_clust, mb_dtor_clust, 225#ifdef INVARIANTS 226 trash_init, trash_fini, 227#else 228 NULL, NULL, 229#endif 230 UMA_ALIGN_PTR, UMA_ZONE_REFCNT); 231 if (nmbjumbo9 > 0) 232 uma_zone_set_max(zone_jumbo9, nmbjumbo9); 233 uma_zone_set_allocf(zone_jumbo9, mbuf_jumbo_alloc); 234 uma_zone_set_freef(zone_jumbo9, mbuf_jumbo_free); 235 236 zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES, 237 mb_ctor_clust, mb_dtor_clust, 238#ifdef INVARIANTS 239 trash_init, trash_fini, 240#else 241 NULL, NULL, 242#endif 243 UMA_ALIGN_PTR, UMA_ZONE_REFCNT); 244 if (nmbjumbo16 > 0) 245 uma_zone_set_max(zone_jumbo16, nmbjumbo16); 246 uma_zone_set_allocf(zone_jumbo16, mbuf_jumbo_alloc); 247 uma_zone_set_freef(zone_jumbo16, mbuf_jumbo_free); 248 249 zone_ext_refcnt = uma_zcreate(MBUF_EXTREFCNT_MEM_NAME, sizeof(u_int), 250 NULL, NULL, 251 NULL, NULL, 252 UMA_ALIGN_PTR, UMA_ZONE_ZINIT); 253 254 /* uma_prealloc() goes here... */ 255 256 /* 257 * Hook event handler for low-memory situation, used to 258 * drain protocols and push data back to the caches (UMA 259 * later pushes it back to VM). 260 */ 261 EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL, 262 EVENTHANDLER_PRI_FIRST); 263 264 /* 265 * [Re]set counters and local statistics knobs. 266 * XXX Some of these should go and be replaced, but UMA stat 267 * gathering needs to be revised. 268 */ 269 mbstat.m_mbufs = 0; 270 mbstat.m_mclusts = 0; 271 mbstat.m_drain = 0; 272 mbstat.m_msize = MSIZE; 273 mbstat.m_mclbytes = MCLBYTES; 274 mbstat.m_minclsize = MINCLSIZE; 275 mbstat.m_mlen = MLEN; 276 mbstat.m_mhlen = MHLEN; 277 mbstat.m_numtypes = MT_NTYPES; 278 279 mbstat.m_mcfail = mbstat.m_mpfail = 0; 280 mbstat.sf_iocnt = 0; 281 mbstat.sf_allocwait = mbstat.sf_allocfail = 0; 282} 283 284/* 285 * UMA backend page allocator for the jumbo frame zones. 286 * 287 * Allocates kernel virtual memory that is backed by contiguous physical 288 * pages. 289 */ 290static void * 291mbuf_jumbo_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 292{ 293 294 *flags = UMA_SLAB_PRIV; 295 return (contigmalloc(bytes, M_JUMBOFRAME, wait, (vm_paddr_t)0, 296 ~(vm_paddr_t)0, 1, 0)); 297} 298 299/* 300 * UMA backend page deallocator for the jumbo frame zones. 301 */ 302static void 303mbuf_jumbo_free(void *mem, int size, u_int8_t flags) 304{ 305 306 contigfree(mem, size, M_JUMBOFRAME); 307} 308 309/* 310 * Constructor for Mbuf master zone. 311 * 312 * The 'arg' pointer points to a mb_args structure which 313 * contains call-specific information required to support the 314 * mbuf allocation API. See mbuf.h. 315 */ 316static int 317mb_ctor_mbuf(void *mem, int size, void *arg, int how) 318{ 319 struct mbuf *m; 320 struct mb_args *args; 321#ifdef MAC 322 int error; 323#endif 324 int flags; 325 short type; 326 327#ifdef INVARIANTS 328 trash_ctor(mem, size, arg, how); 329#endif 330 m = (struct mbuf *)mem; 331 args = (struct mb_args *)arg; 332 flags = args->flags; 333 type = args->type; 334 335 /* 336 * The mbuf is initialized later. The caller has the 337 * responsibility to set up any MAC labels too. 338 */ 339 if (type == MT_NOINIT) 340 return (0); 341 342 m->m_next = NULL; 343 m->m_nextpkt = NULL; 344 m->m_len = 0; 345 m->m_flags = flags; 346 m->m_type = type; 347 if (flags & M_PKTHDR) { 348 m->m_data = m->m_pktdat; 349 m->m_pkthdr.rcvif = NULL; 350 m->m_pkthdr.len = 0; 351 m->m_pkthdr.header = NULL; 352 m->m_pkthdr.csum_flags = 0; 353 m->m_pkthdr.csum_data = 0; 354 m->m_pkthdr.tso_segsz = 0; 355 m->m_pkthdr.ether_vtag = 0; 356 SLIST_INIT(&m->m_pkthdr.tags); 357#ifdef MAC 358 /* If the label init fails, fail the alloc */ 359 error = mac_mbuf_init(m, how); 360 if (error) 361 return (error); 362#endif 363 } else 364 m->m_data = m->m_dat; 365 return (0); 366} 367 368/* 369 * The Mbuf master zone destructor. 370 */ 371static void 372mb_dtor_mbuf(void *mem, int size, void *arg) 373{ 374 struct mbuf *m; 375 unsigned long flags; 376 377 m = (struct mbuf *)mem; 378 flags = (unsigned long)arg; 379 380 if ((flags & MB_NOTAGS) == 0 && (m->m_flags & M_PKTHDR) != 0) 381 m_tag_delete_chain(m, NULL); 382 KASSERT((m->m_flags & M_EXT) == 0, ("%s: M_EXT set", __func__)); 383 KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__)); 384#ifdef INVARIANTS 385 trash_dtor(mem, size, arg); 386#endif 387} 388 389/* 390 * The Mbuf Packet zone destructor. 391 */ 392static void 393mb_dtor_pack(void *mem, int size, void *arg) 394{ 395 struct mbuf *m; 396 397 m = (struct mbuf *)mem; 398 if ((m->m_flags & M_PKTHDR) != 0) 399 m_tag_delete_chain(m, NULL); 400 401 /* Make sure we've got a clean cluster back. */ 402 KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__)); 403 KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__)); 404 KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__)); 405 KASSERT(m->m_ext.ext_args == NULL, ("%s: ext_args != NULL", __func__)); 406 KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__)); 407 KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__)); 408 KASSERT(*m->m_ext.ref_cnt == 1, ("%s: ref_cnt != 1", __func__)); 409#ifdef INVARIANTS 410 trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg); 411#endif 412 /* 413 * If there are processes blocked on zone_clust, waiting for pages 414 * to be freed up, * cause them to be woken up by draining the 415 * packet zone. We are exposed to a race here * (in the check for 416 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that 417 * is deliberate. We don't want to acquire the zone lock for every 418 * mbuf free. 419 */ 420 if (uma_zone_exhausted_nolock(zone_clust)) 421 zone_drain(zone_pack); 422} 423 424/* 425 * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor. 426 * 427 * Here the 'arg' pointer points to the Mbuf which we 428 * are configuring cluster storage for. If 'arg' is 429 * empty we allocate just the cluster without setting 430 * the mbuf to it. See mbuf.h. 431 */ 432static int 433mb_ctor_clust(void *mem, int size, void *arg, int how) 434{ 435 struct mbuf *m; 436 u_int *refcnt; 437 int type; 438 uma_zone_t zone; 439 440#ifdef INVARIANTS 441 trash_ctor(mem, size, arg, how); 442#endif 443 switch (size) { 444 case MCLBYTES: 445 type = EXT_CLUSTER; 446 zone = zone_clust; 447 break; 448#if MJUMPAGESIZE != MCLBYTES 449 case MJUMPAGESIZE: 450 type = EXT_JUMBOP; 451 zone = zone_jumbop; 452 break; 453#endif 454 case MJUM9BYTES: 455 type = EXT_JUMBO9; 456 zone = zone_jumbo9; 457 break; 458 case MJUM16BYTES: 459 type = EXT_JUMBO16; 460 zone = zone_jumbo16; 461 break; 462 default: 463 panic("unknown cluster size"); 464 break; 465 } 466 467 m = (struct mbuf *)arg; 468 refcnt = uma_find_refcnt(zone, mem); 469 *refcnt = 1; 470 if (m != NULL) { 471 m->m_ext.ext_buf = (caddr_t)mem; 472 m->m_data = m->m_ext.ext_buf; 473 m->m_flags |= M_EXT; 474 m->m_ext.ext_free = NULL; 475 m->m_ext.ext_args = NULL; 476 m->m_ext.ext_size = size; 477 m->m_ext.ext_type = type; 478 m->m_ext.ref_cnt = refcnt; 479 } 480 481 return (0); 482} 483 484/* 485 * The Mbuf Cluster zone destructor. 486 */ 487static void 488mb_dtor_clust(void *mem, int size, void *arg) 489{ 490#ifdef INVARIANTS 491 uma_zone_t zone; 492 493 zone = m_getzone(size); 494 KASSERT(*(uma_find_refcnt(zone, mem)) <= 1, 495 ("%s: refcnt incorrect %u", __func__, 496 *(uma_find_refcnt(zone, mem))) ); 497 498 trash_dtor(mem, size, arg); 499#endif 500} 501 502/* 503 * The Packet secondary zone's init routine, executed on the 504 * object's transition from mbuf keg slab to zone cache. 505 */ 506static int 507mb_zinit_pack(void *mem, int size, int how) 508{ 509 struct mbuf *m; 510 511 m = (struct mbuf *)mem; /* m is virgin. */ 512 if (uma_zalloc_arg(zone_clust, m, how) == NULL || 513 m->m_ext.ext_buf == NULL) 514 return (ENOMEM); 515 m->m_ext.ext_type = EXT_PACKET; /* Override. */ 516#ifdef INVARIANTS 517 trash_init(m->m_ext.ext_buf, MCLBYTES, how); 518#endif 519 return (0); 520} 521 522/* 523 * The Packet secondary zone's fini routine, executed on the 524 * object's transition from zone cache to keg slab. 525 */ 526static void 527mb_zfini_pack(void *mem, int size) 528{ 529 struct mbuf *m; 530 531 m = (struct mbuf *)mem; 532#ifdef INVARIANTS 533 trash_fini(m->m_ext.ext_buf, MCLBYTES); 534#endif 535 uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL); 536#ifdef INVARIANTS 537 trash_dtor(mem, size, NULL); 538#endif 539} 540 541/* 542 * The "packet" keg constructor. 543 */ 544static int 545mb_ctor_pack(void *mem, int size, void *arg, int how) 546{ 547 struct mbuf *m; 548 struct mb_args *args; 549#ifdef MAC 550 int error; 551#endif 552 int flags; 553 short type; 554 555 m = (struct mbuf *)mem; 556 args = (struct mb_args *)arg; 557 flags = args->flags; 558 type = args->type; 559 560#ifdef INVARIANTS 561 trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how); 562#endif 563 m->m_next = NULL; 564 m->m_nextpkt = NULL; 565 m->m_data = m->m_ext.ext_buf; 566 m->m_len = 0; 567 m->m_flags = (flags | M_EXT); 568 m->m_type = type; 569 570 if (flags & M_PKTHDR) { 571 m->m_pkthdr.rcvif = NULL; 572 m->m_pkthdr.len = 0; 573 m->m_pkthdr.header = NULL; 574 m->m_pkthdr.csum_flags = 0; 575 m->m_pkthdr.csum_data = 0; 576 m->m_pkthdr.tso_segsz = 0; 577 m->m_pkthdr.ether_vtag = 0; 578 SLIST_INIT(&m->m_pkthdr.tags); 579#ifdef MAC 580 /* If the label init fails, fail the alloc */ 581 error = mac_mbuf_init(m, how); 582 if (error) 583 return (error); 584#endif 585 } 586 /* m_ext is already initialized. */ 587 588 return (0); 589} 590 591/* 592 * This is the protocol drain routine. 593 * 594 * No locks should be held when this is called. The drain routines have to 595 * presently acquire some locks which raises the possibility of lock order 596 * reversal. 597 */ 598static void 599mb_reclaim(void *junk) 600{ 601 struct domain *dp; 602 struct protosw *pr; 603 604 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL, 605 "mb_reclaim()"); 606 607 for (dp = domains; dp != NULL; dp = dp->dom_next) 608 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) 609 if (pr->pr_drain != NULL) 610 (*pr->pr_drain)(); 611} 612