123static int 124sysctl_nmbclusters(SYSCTL_HANDLER_ARGS) 125{ 126 int error, newnmbclusters; 127 128 newnmbclusters = nmbclusters; 129 error = sysctl_handle_int(oidp, &newnmbclusters, 0, req); 130 if (error == 0 && req->newptr) { 131 if (newnmbclusters > nmbclusters) { 132 nmbclusters = newnmbclusters; 133 uma_zone_set_max(zone_clust, nmbclusters); 134 EVENTHANDLER_INVOKE(nmbclusters_change); 135 } else 136 error = EINVAL; 137 } 138 return (error); 139} 140SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters, CTLTYPE_INT|CTLFLAG_RW, 141&nmbclusters, 0, sysctl_nmbclusters, "IU", 142 "Maximum number of mbuf clusters allowed"); 143 144static int 145sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS) 146{ 147 int error, newnmbjumbop; 148 149 newnmbjumbop = nmbjumbop; 150 error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req); 151 if (error == 0 && req->newptr) { 152 if (newnmbjumbop> nmbjumbop) { 153 nmbjumbop = newnmbjumbop; 154 uma_zone_set_max(zone_jumbop, nmbjumbop); 155 } else 156 error = EINVAL; 157 } 158 return (error); 159} 160SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop, CTLTYPE_INT|CTLFLAG_RW, 161&nmbjumbop, 0, sysctl_nmbjumbop, "IU", 162 "Maximum number of mbuf page size jumbo clusters allowed"); 163 164 165static int 166sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS) 167{ 168 int error, newnmbjumbo9; 169 170 newnmbjumbo9 = nmbjumbo9; 171 error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req); 172 if (error == 0 && req->newptr) { 173 if (newnmbjumbo9> nmbjumbo9) { 174 nmbjumbo9 = newnmbjumbo9; 175 uma_zone_set_max(zone_jumbo9, nmbjumbo9); 176 } else 177 error = EINVAL; 178 } 179 return (error); 180} 181SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9, CTLTYPE_INT|CTLFLAG_RW, 182&nmbjumbo9, 0, sysctl_nmbjumbo9, "IU", 183 "Maximum number of mbuf 9k jumbo clusters allowed"); 184 185static int 186sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS) 187{ 188 int error, newnmbjumbo16; 189 190 newnmbjumbo16 = nmbjumbo16; 191 error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req); 192 if (error == 0 && req->newptr) { 193 if (newnmbjumbo16> nmbjumbo16) { 194 nmbjumbo16 = newnmbjumbo16; 195 uma_zone_set_max(zone_jumbo16, nmbjumbo16); 196 } else 197 error = EINVAL; 198 } 199 return (error); 200} 201SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16, CTLTYPE_INT|CTLFLAG_RW, 202&nmbjumbo16, 0, sysctl_nmbjumbo16, "IU", 203 "Maximum number of mbuf 16k jumbo clusters allowed"); 204 205 206 207SYSCTL_STRUCT(_kern_ipc, OID_AUTO, mbstat, CTLFLAG_RD, &mbstat, mbstat, 208 "Mbuf general information and statistics"); 209 210/* 211 * Zones from which we allocate. 212 */ 213uma_zone_t zone_mbuf; 214uma_zone_t zone_clust; 215uma_zone_t zone_pack; 216uma_zone_t zone_jumbop; 217uma_zone_t zone_jumbo9; 218uma_zone_t zone_jumbo16; 219uma_zone_t zone_ext_refcnt; 220 221/* 222 * Local prototypes. 223 */ 224static int mb_ctor_mbuf(void *, int, void *, int); 225static int mb_ctor_clust(void *, int, void *, int); 226static int mb_ctor_pack(void *, int, void *, int); 227static void mb_dtor_mbuf(void *, int, void *); 228static void mb_dtor_clust(void *, int, void *); 229static void mb_dtor_pack(void *, int, void *); 230static int mb_zinit_pack(void *, int, int); 231static void mb_zfini_pack(void *, int); 232 233static void mb_reclaim(void *); 234static void mbuf_init(void *); 235static void *mbuf_jumbo_alloc(uma_zone_t, int, u_int8_t *, int); 236static void mbuf_jumbo_free(void *, int, u_int8_t); 237 238static MALLOC_DEFINE(M_JUMBOFRAME, "jumboframes", "mbuf jumbo frame buffers"); 239 240/* Ensure that MSIZE doesn't break dtom() - it must be a power of 2 */ 241CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE); 242 243/* 244 * Initialize FreeBSD Network buffer allocation. 245 */ 246SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL); 247static void 248mbuf_init(void *dummy) 249{ 250 251 /* 252 * Configure UMA zones for Mbufs, Clusters, and Packets. 253 */ 254 zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE, 255 mb_ctor_mbuf, mb_dtor_mbuf, 256#ifdef INVARIANTS 257 trash_init, trash_fini, 258#else 259 NULL, NULL, 260#endif 261 MSIZE - 1, UMA_ZONE_MAXBUCKET); 262 263 zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES, 264 mb_ctor_clust, mb_dtor_clust, 265#ifdef INVARIANTS 266 trash_init, trash_fini, 267#else 268 NULL, NULL, 269#endif 270 UMA_ALIGN_PTR, UMA_ZONE_REFCNT); 271 if (nmbclusters > 0) 272 uma_zone_set_max(zone_clust, nmbclusters); 273 274 zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack, 275 mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf); 276 277 /* Make jumbo frame zone too. Page size, 9k and 16k. */ 278 zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE, 279 mb_ctor_clust, mb_dtor_clust, 280#ifdef INVARIANTS 281 trash_init, trash_fini, 282#else 283 NULL, NULL, 284#endif 285 UMA_ALIGN_PTR, UMA_ZONE_REFCNT); 286 if (nmbjumbop > 0) 287 uma_zone_set_max(zone_jumbop, nmbjumbop); 288 289 zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES, 290 mb_ctor_clust, mb_dtor_clust, 291#ifdef INVARIANTS 292 trash_init, trash_fini, 293#else 294 NULL, NULL, 295#endif 296 UMA_ALIGN_PTR, UMA_ZONE_REFCNT); 297 if (nmbjumbo9 > 0) 298 uma_zone_set_max(zone_jumbo9, nmbjumbo9); 299 uma_zone_set_allocf(zone_jumbo9, mbuf_jumbo_alloc); 300 uma_zone_set_freef(zone_jumbo9, mbuf_jumbo_free); 301 302 zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES, 303 mb_ctor_clust, mb_dtor_clust, 304#ifdef INVARIANTS 305 trash_init, trash_fini, 306#else 307 NULL, NULL, 308#endif 309 UMA_ALIGN_PTR, UMA_ZONE_REFCNT); 310 if (nmbjumbo16 > 0) 311 uma_zone_set_max(zone_jumbo16, nmbjumbo16); 312 uma_zone_set_allocf(zone_jumbo16, mbuf_jumbo_alloc); 313 uma_zone_set_freef(zone_jumbo16, mbuf_jumbo_free); 314 315 zone_ext_refcnt = uma_zcreate(MBUF_EXTREFCNT_MEM_NAME, sizeof(u_int), 316 NULL, NULL, 317 NULL, NULL, 318 UMA_ALIGN_PTR, UMA_ZONE_ZINIT); 319 320 /* uma_prealloc() goes here... */ 321 322 /* 323 * Hook event handler for low-memory situation, used to 324 * drain protocols and push data back to the caches (UMA 325 * later pushes it back to VM). 326 */ 327 EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL, 328 EVENTHANDLER_PRI_FIRST); 329 330 /* 331 * [Re]set counters and local statistics knobs. 332 * XXX Some of these should go and be replaced, but UMA stat 333 * gathering needs to be revised. 334 */ 335 mbstat.m_mbufs = 0; 336 mbstat.m_mclusts = 0; 337 mbstat.m_drain = 0; 338 mbstat.m_msize = MSIZE; 339 mbstat.m_mclbytes = MCLBYTES; 340 mbstat.m_minclsize = MINCLSIZE; 341 mbstat.m_mlen = MLEN; 342 mbstat.m_mhlen = MHLEN; 343 mbstat.m_numtypes = MT_NTYPES; 344 345 mbstat.m_mcfail = mbstat.m_mpfail = 0; 346 mbstat.sf_iocnt = 0; 347 mbstat.sf_allocwait = mbstat.sf_allocfail = 0; 348} 349 350/* 351 * UMA backend page allocator for the jumbo frame zones. 352 * 353 * Allocates kernel virtual memory that is backed by contiguous physical 354 * pages. 355 */ 356static void * 357mbuf_jumbo_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 358{ 359 360 /* Inform UMA that this allocator uses kernel_map/object. */ 361 *flags = UMA_SLAB_KERNEL; 362 return (contigmalloc(bytes, M_JUMBOFRAME, wait, (vm_paddr_t)0, 363 ~(vm_paddr_t)0, 1, 0)); 364} 365 366/* 367 * UMA backend page deallocator for the jumbo frame zones. 368 */ 369static void 370mbuf_jumbo_free(void *mem, int size, u_int8_t flags) 371{ 372 373 contigfree(mem, size, M_JUMBOFRAME); 374} 375 376/* 377 * Constructor for Mbuf master zone. 378 * 379 * The 'arg' pointer points to a mb_args structure which 380 * contains call-specific information required to support the 381 * mbuf allocation API. See mbuf.h. 382 */ 383static int 384mb_ctor_mbuf(void *mem, int size, void *arg, int how) 385{ 386 struct mbuf *m; 387 struct mb_args *args; 388#ifdef MAC 389 int error; 390#endif 391 int flags; 392 short type; 393 394#ifdef INVARIANTS 395 trash_ctor(mem, size, arg, how); 396#endif 397 m = (struct mbuf *)mem; 398 args = (struct mb_args *)arg; 399 flags = args->flags; 400 type = args->type; 401 402 /* 403 * The mbuf is initialized later. The caller has the 404 * responsibility to set up any MAC labels too. 405 */ 406 if (type == MT_NOINIT) 407 return (0); 408 409 m->m_next = NULL; 410 m->m_nextpkt = NULL; 411 m->m_len = 0; 412 m->m_flags = flags; 413 m->m_type = type; 414 if (flags & M_PKTHDR) { 415 m->m_data = m->m_pktdat; 416 m->m_pkthdr.rcvif = NULL; 417 m->m_pkthdr.header = NULL; 418 m->m_pkthdr.len = 0; 419 m->m_pkthdr.csum_flags = 0; 420 m->m_pkthdr.csum_data = 0; 421 m->m_pkthdr.tso_segsz = 0; 422 m->m_pkthdr.ether_vtag = 0; 423 SLIST_INIT(&m->m_pkthdr.tags); 424#ifdef MAC 425 /* If the label init fails, fail the alloc */ 426 error = mac_mbuf_init(m, how); 427 if (error) 428 return (error); 429#endif 430 } else 431 m->m_data = m->m_dat; 432 return (0); 433} 434 435/* 436 * The Mbuf master zone destructor. 437 */ 438static void 439mb_dtor_mbuf(void *mem, int size, void *arg) 440{ 441 struct mbuf *m; 442 unsigned long flags; 443 444 m = (struct mbuf *)mem; 445 flags = (unsigned long)arg; 446 447 if ((flags & MB_NOTAGS) == 0 && (m->m_flags & M_PKTHDR) != 0) 448 m_tag_delete_chain(m, NULL); 449 KASSERT((m->m_flags & M_EXT) == 0, ("%s: M_EXT set", __func__)); 450 KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__)); 451#ifdef INVARIANTS 452 trash_dtor(mem, size, arg); 453#endif 454} 455 456/* 457 * The Mbuf Packet zone destructor. 458 */ 459static void 460mb_dtor_pack(void *mem, int size, void *arg) 461{ 462 struct mbuf *m; 463 464 m = (struct mbuf *)mem; 465 if ((m->m_flags & M_PKTHDR) != 0) 466 m_tag_delete_chain(m, NULL); 467 468 /* Make sure we've got a clean cluster back. */ 469 KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__)); 470 KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__)); 471 KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__)); 472 KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__)); 473 KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__)); 474 KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__)); 475 KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__)); 476 KASSERT(*m->m_ext.ref_cnt == 1, ("%s: ref_cnt != 1", __func__)); 477#ifdef INVARIANTS 478 trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg); 479#endif 480 /* 481 * If there are processes blocked on zone_clust, waiting for pages 482 * to be freed up, * cause them to be woken up by draining the 483 * packet zone. We are exposed to a race here * (in the check for 484 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that 485 * is deliberate. We don't want to acquire the zone lock for every 486 * mbuf free. 487 */ 488 if (uma_zone_exhausted_nolock(zone_clust)) 489 zone_drain(zone_pack); 490} 491 492/* 493 * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor. 494 * 495 * Here the 'arg' pointer points to the Mbuf which we 496 * are configuring cluster storage for. If 'arg' is 497 * empty we allocate just the cluster without setting 498 * the mbuf to it. See mbuf.h. 499 */ 500static int 501mb_ctor_clust(void *mem, int size, void *arg, int how) 502{ 503 struct mbuf *m; 504 u_int *refcnt; 505 int type; 506 uma_zone_t zone; 507 508#ifdef INVARIANTS 509 trash_ctor(mem, size, arg, how); 510#endif 511 switch (size) { 512 case MCLBYTES: 513 type = EXT_CLUSTER; 514 zone = zone_clust; 515 break; 516#if MJUMPAGESIZE != MCLBYTES 517 case MJUMPAGESIZE: 518 type = EXT_JUMBOP; 519 zone = zone_jumbop; 520 break; 521#endif 522 case MJUM9BYTES: 523 type = EXT_JUMBO9; 524 zone = zone_jumbo9; 525 break; 526 case MJUM16BYTES: 527 type = EXT_JUMBO16; 528 zone = zone_jumbo16; 529 break; 530 default: 531 panic("unknown cluster size"); 532 break; 533 } 534 535 m = (struct mbuf *)arg; 536 refcnt = uma_find_refcnt(zone, mem); 537 *refcnt = 1; 538 if (m != NULL) { 539 m->m_ext.ext_buf = (caddr_t)mem; 540 m->m_data = m->m_ext.ext_buf; 541 m->m_flags |= M_EXT; 542 m->m_ext.ext_free = NULL; 543 m->m_ext.ext_arg1 = NULL; 544 m->m_ext.ext_arg2 = NULL; 545 m->m_ext.ext_size = size; 546 m->m_ext.ext_type = type; 547 m->m_ext.ref_cnt = refcnt; 548 } 549 550 return (0); 551} 552 553/* 554 * The Mbuf Cluster zone destructor. 555 */ 556static void 557mb_dtor_clust(void *mem, int size, void *arg) 558{ 559#ifdef INVARIANTS 560 uma_zone_t zone; 561 562 zone = m_getzone(size); 563 KASSERT(*(uma_find_refcnt(zone, mem)) <= 1, 564 ("%s: refcnt incorrect %u", __func__, 565 *(uma_find_refcnt(zone, mem))) ); 566 567 trash_dtor(mem, size, arg); 568#endif 569} 570 571/* 572 * The Packet secondary zone's init routine, executed on the 573 * object's transition from mbuf keg slab to zone cache. 574 */ 575static int 576mb_zinit_pack(void *mem, int size, int how) 577{ 578 struct mbuf *m; 579 580 m = (struct mbuf *)mem; /* m is virgin. */ 581 if (uma_zalloc_arg(zone_clust, m, how) == NULL || 582 m->m_ext.ext_buf == NULL) 583 return (ENOMEM); 584 m->m_ext.ext_type = EXT_PACKET; /* Override. */ 585#ifdef INVARIANTS 586 trash_init(m->m_ext.ext_buf, MCLBYTES, how); 587#endif 588 return (0); 589} 590 591/* 592 * The Packet secondary zone's fini routine, executed on the 593 * object's transition from zone cache to keg slab. 594 */ 595static void 596mb_zfini_pack(void *mem, int size) 597{ 598 struct mbuf *m; 599 600 m = (struct mbuf *)mem; 601#ifdef INVARIANTS 602 trash_fini(m->m_ext.ext_buf, MCLBYTES); 603#endif 604 uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL); 605#ifdef INVARIANTS 606 trash_dtor(mem, size, NULL); 607#endif 608} 609 610/* 611 * The "packet" keg constructor. 612 */ 613static int 614mb_ctor_pack(void *mem, int size, void *arg, int how) 615{ 616 struct mbuf *m; 617 struct mb_args *args; 618#ifdef MAC 619 int error; 620#endif 621 int flags; 622 short type; 623 624 m = (struct mbuf *)mem; 625 args = (struct mb_args *)arg; 626 flags = args->flags; 627 type = args->type; 628 629#ifdef INVARIANTS 630 trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how); 631#endif 632 m->m_next = NULL; 633 m->m_nextpkt = NULL; 634 m->m_data = m->m_ext.ext_buf; 635 m->m_len = 0; 636 m->m_flags = (flags | M_EXT); 637 m->m_type = type; 638 639 if (flags & M_PKTHDR) { 640 m->m_pkthdr.rcvif = NULL; 641 m->m_pkthdr.len = 0; 642 m->m_pkthdr.header = NULL; 643 m->m_pkthdr.csum_flags = 0; 644 m->m_pkthdr.csum_data = 0; 645 m->m_pkthdr.tso_segsz = 0; 646 m->m_pkthdr.ether_vtag = 0; 647 SLIST_INIT(&m->m_pkthdr.tags); 648#ifdef MAC 649 /* If the label init fails, fail the alloc */ 650 error = mac_mbuf_init(m, how); 651 if (error) 652 return (error); 653#endif 654 } 655 /* m_ext is already initialized. */ 656 657 return (0); 658} 659 660/* 661 * This is the protocol drain routine. 662 * 663 * No locks should be held when this is called. The drain routines have to 664 * presently acquire some locks which raises the possibility of lock order 665 * reversal. 666 */ 667static void 668mb_reclaim(void *junk) 669{ 670 struct domain *dp; 671 struct protosw *pr; 672 673 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL, 674 "mb_reclaim()"); 675 676 for (dp = domains; dp != NULL; dp = dp->dom_next) 677 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++) 678 if (pr->pr_drain != NULL) 679 (*pr->pr_drain)(); 680}
|