uma_core.c revision 262127
1/*- 2 * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org> 3 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org> 4 * Copyright (c) 2004-2006 Robert N. M. Watson 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice unmodified, this list of conditions, and the following 12 * 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 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 */ 28 29/* 30 * uma_core.c Implementation of the Universal Memory allocator 31 * 32 * This allocator is intended to replace the multitude of similar object caches 33 * in the standard FreeBSD kernel. The intent is to be flexible as well as 34 * effecient. A primary design goal is to return unused memory to the rest of 35 * the system. This will make the system as a whole more flexible due to the 36 * ability to move memory to subsystems which most need it instead of leaving 37 * pools of reserved memory unused. 38 * 39 * The basic ideas stem from similar slab/zone based allocators whose algorithms 40 * are well known. 41 * 42 */ 43 44/* 45 * TODO: 46 * - Improve memory usage for large allocations 47 * - Investigate cache size adjustments 48 */ 49 50#include <sys/cdefs.h> 51__FBSDID("$FreeBSD: stable/10/sys/vm/uma_core.c 262127 2014-02-17 20:25:17Z dim $"); 52 53/* I should really use ktr.. */ 54/* 55#define UMA_DEBUG 1 56#define UMA_DEBUG_ALLOC 1 57#define UMA_DEBUG_ALLOC_1 1 58*/ 59 60#include "opt_ddb.h" 61#include "opt_param.h" 62#include "opt_vm.h" 63 64#include <sys/param.h> 65#include <sys/systm.h> 66#include <sys/bitset.h> 67#include <sys/kernel.h> 68#include <sys/types.h> 69#include <sys/queue.h> 70#include <sys/malloc.h> 71#include <sys/ktr.h> 72#include <sys/lock.h> 73#include <sys/sysctl.h> 74#include <sys/mutex.h> 75#include <sys/proc.h> 76#include <sys/rwlock.h> 77#include <sys/sbuf.h> 78#include <sys/sched.h> 79#include <sys/smp.h> 80#include <sys/vmmeter.h> 81 82#include <vm/vm.h> 83#include <vm/vm_object.h> 84#include <vm/vm_page.h> 85#include <vm/vm_pageout.h> 86#include <vm/vm_param.h> 87#include <vm/vm_map.h> 88#include <vm/vm_kern.h> 89#include <vm/vm_extern.h> 90#include <vm/uma.h> 91#include <vm/uma_int.h> 92#include <vm/uma_dbg.h> 93 94#include <ddb/ddb.h> 95 96#ifdef DEBUG_MEMGUARD 97#include <vm/memguard.h> 98#endif 99 100/* 101 * This is the zone and keg from which all zones are spawned. The idea is that 102 * even the zone & keg heads are allocated from the allocator, so we use the 103 * bss section to bootstrap us. 104 */ 105static struct uma_keg masterkeg; 106static struct uma_zone masterzone_k; 107static struct uma_zone masterzone_z; 108static uma_zone_t kegs = &masterzone_k; 109static uma_zone_t zones = &masterzone_z; 110 111/* This is the zone from which all of uma_slab_t's are allocated. */ 112static uma_zone_t slabzone; 113static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */ 114 115/* 116 * The initial hash tables come out of this zone so they can be allocated 117 * prior to malloc coming up. 118 */ 119static uma_zone_t hashzone; 120 121/* The boot-time adjusted value for cache line alignment. */ 122int uma_align_cache = 64 - 1; 123 124static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 125 126/* 127 * Are we allowed to allocate buckets? 128 */ 129static int bucketdisable = 1; 130 131/* Linked list of all kegs in the system */ 132static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 133 134/* Linked list of all cache-only zones in the system */ 135static LIST_HEAD(,uma_zone) uma_cachezones = 136 LIST_HEAD_INITIALIZER(uma_cachezones); 137 138/* This mutex protects the keg list */ 139static struct mtx_padalign uma_mtx; 140 141/* Linked list of boot time pages */ 142static LIST_HEAD(,uma_slab) uma_boot_pages = 143 LIST_HEAD_INITIALIZER(uma_boot_pages); 144 145/* This mutex protects the boot time pages list */ 146static struct mtx_padalign uma_boot_pages_mtx; 147 148/* Is the VM done starting up? */ 149static int booted = 0; 150#define UMA_STARTUP 1 151#define UMA_STARTUP2 2 152 153/* 154 * Only mbuf clusters use ref zones. Just provide enough references 155 * to support the one user. New code should not use the ref facility. 156 */ 157static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES; 158 159/* 160 * This is the handle used to schedule events that need to happen 161 * outside of the allocation fast path. 162 */ 163static struct callout uma_callout; 164#define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 165 166/* 167 * This structure is passed as the zone ctor arg so that I don't have to create 168 * a special allocation function just for zones. 169 */ 170struct uma_zctor_args { 171 const char *name; 172 size_t size; 173 uma_ctor ctor; 174 uma_dtor dtor; 175 uma_init uminit; 176 uma_fini fini; 177 uma_import import; 178 uma_release release; 179 void *arg; 180 uma_keg_t keg; 181 int align; 182 uint32_t flags; 183}; 184 185struct uma_kctor_args { 186 uma_zone_t zone; 187 size_t size; 188 uma_init uminit; 189 uma_fini fini; 190 int align; 191 uint32_t flags; 192}; 193 194struct uma_bucket_zone { 195 uma_zone_t ubz_zone; 196 char *ubz_name; 197 int ubz_entries; /* Number of items it can hold. */ 198 int ubz_maxsize; /* Maximum allocation size per-item. */ 199}; 200 201/* 202 * Compute the actual number of bucket entries to pack them in power 203 * of two sizes for more efficient space utilization. 204 */ 205#define BUCKET_SIZE(n) \ 206 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 207 208#define BUCKET_MAX BUCKET_SIZE(128) 209 210struct uma_bucket_zone bucket_zones[] = { 211 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 212 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 213 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 214 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 215 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 216 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 217 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 218 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 219 { NULL, NULL, 0} 220}; 221 222/* 223 * Flags and enumerations to be passed to internal functions. 224 */ 225enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; 226 227/* Prototypes.. */ 228 229static void *noobj_alloc(uma_zone_t, int, uint8_t *, int); 230static void *page_alloc(uma_zone_t, int, uint8_t *, int); 231static void *startup_alloc(uma_zone_t, int, uint8_t *, int); 232static void page_free(void *, int, uint8_t); 233static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); 234static void cache_drain(uma_zone_t); 235static void bucket_drain(uma_zone_t, uma_bucket_t); 236static void bucket_cache_drain(uma_zone_t zone); 237static int keg_ctor(void *, int, void *, int); 238static void keg_dtor(void *, int, void *); 239static int zone_ctor(void *, int, void *, int); 240static void zone_dtor(void *, int, void *); 241static int zero_init(void *, int, int); 242static void keg_small_init(uma_keg_t keg); 243static void keg_large_init(uma_keg_t keg); 244static void zone_foreach(void (*zfunc)(uma_zone_t)); 245static void zone_timeout(uma_zone_t zone); 246static int hash_alloc(struct uma_hash *); 247static int hash_expand(struct uma_hash *, struct uma_hash *); 248static void hash_free(struct uma_hash *hash); 249static void uma_timeout(void *); 250static void uma_startup3(void); 251static void *zone_alloc_item(uma_zone_t, void *, int); 252static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 253static void bucket_enable(void); 254static void bucket_init(void); 255static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 256static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 257static void bucket_zone_drain(void); 258static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags); 259static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); 260static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); 261static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 262static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); 263static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 264 uma_fini fini, int align, uint32_t flags); 265static int zone_import(uma_zone_t zone, void **bucket, int max, int flags); 266static void zone_release(uma_zone_t zone, void **bucket, int cnt); 267 268void uma_print_zone(uma_zone_t); 269void uma_print_stats(void); 270static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 271static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 272 273SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 274 275SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 276 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 277 278SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 279 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 280 281static int zone_warnings = 1; 282TUNABLE_INT("vm.zone_warnings", &zone_warnings); 283SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0, 284 "Warn when UMA zones becomes full"); 285 286/* 287 * This routine checks to see whether or not it's safe to enable buckets. 288 */ 289static void 290bucket_enable(void) 291{ 292 bucketdisable = vm_page_count_min(); 293} 294 295/* 296 * Initialize bucket_zones, the array of zones of buckets of various sizes. 297 * 298 * For each zone, calculate the memory required for each bucket, consisting 299 * of the header and an array of pointers. 300 */ 301static void 302bucket_init(void) 303{ 304 struct uma_bucket_zone *ubz; 305 int size; 306 int i; 307 308 for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 309 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 310 size += sizeof(void *) * ubz->ubz_entries; 311 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 312 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 313 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET); 314 } 315} 316 317/* 318 * Given a desired number of entries for a bucket, return the zone from which 319 * to allocate the bucket. 320 */ 321static struct uma_bucket_zone * 322bucket_zone_lookup(int entries) 323{ 324 struct uma_bucket_zone *ubz; 325 326 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 327 if (ubz->ubz_entries >= entries) 328 return (ubz); 329 ubz--; 330 return (ubz); 331} 332 333static int 334bucket_select(int size) 335{ 336 struct uma_bucket_zone *ubz; 337 338 ubz = &bucket_zones[0]; 339 if (size > ubz->ubz_maxsize) 340 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 341 342 for (; ubz->ubz_entries != 0; ubz++) 343 if (ubz->ubz_maxsize < size) 344 break; 345 ubz--; 346 return (ubz->ubz_entries); 347} 348 349static uma_bucket_t 350bucket_alloc(uma_zone_t zone, void *udata, int flags) 351{ 352 struct uma_bucket_zone *ubz; 353 uma_bucket_t bucket; 354 355 /* 356 * This is to stop us from allocating per cpu buckets while we're 357 * running out of vm.boot_pages. Otherwise, we would exhaust the 358 * boot pages. This also prevents us from allocating buckets in 359 * low memory situations. 360 */ 361 if (bucketdisable) 362 return (NULL); 363 /* 364 * To limit bucket recursion we store the original zone flags 365 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 366 * NOVM flag to persist even through deep recursions. We also 367 * store ZFLAG_BUCKET once we have recursed attempting to allocate 368 * a bucket for a bucket zone so we do not allow infinite bucket 369 * recursion. This cookie will even persist to frees of unused 370 * buckets via the allocation path or bucket allocations in the 371 * free path. 372 */ 373 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 374 udata = (void *)(uintptr_t)zone->uz_flags; 375 else { 376 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 377 return (NULL); 378 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 379 } 380 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 381 flags |= M_NOVM; 382 ubz = bucket_zone_lookup(zone->uz_count); 383 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 384 if (bucket) { 385#ifdef INVARIANTS 386 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 387#endif 388 bucket->ub_cnt = 0; 389 bucket->ub_entries = ubz->ubz_entries; 390 } 391 392 return (bucket); 393} 394 395static void 396bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 397{ 398 struct uma_bucket_zone *ubz; 399 400 KASSERT(bucket->ub_cnt == 0, 401 ("bucket_free: Freeing a non free bucket.")); 402 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 403 udata = (void *)(uintptr_t)zone->uz_flags; 404 ubz = bucket_zone_lookup(bucket->ub_entries); 405 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 406} 407 408static void 409bucket_zone_drain(void) 410{ 411 struct uma_bucket_zone *ubz; 412 413 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 414 zone_drain(ubz->ubz_zone); 415} 416 417static void 418zone_log_warning(uma_zone_t zone) 419{ 420 static const struct timeval warninterval = { 300, 0 }; 421 422 if (!zone_warnings || zone->uz_warning == NULL) 423 return; 424 425 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 426 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 427} 428 429static void 430zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 431{ 432 uma_klink_t klink; 433 434 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 435 kegfn(klink->kl_keg); 436} 437 438/* 439 * Routine called by timeout which is used to fire off some time interval 440 * based calculations. (stats, hash size, etc.) 441 * 442 * Arguments: 443 * arg Unused 444 * 445 * Returns: 446 * Nothing 447 */ 448static void 449uma_timeout(void *unused) 450{ 451 bucket_enable(); 452 zone_foreach(zone_timeout); 453 454 /* Reschedule this event */ 455 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 456} 457 458/* 459 * Routine to perform timeout driven calculations. This expands the 460 * hashes and does per cpu statistics aggregation. 461 * 462 * Returns nothing. 463 */ 464static void 465keg_timeout(uma_keg_t keg) 466{ 467 468 KEG_LOCK(keg); 469 /* 470 * Expand the keg hash table. 471 * 472 * This is done if the number of slabs is larger than the hash size. 473 * What I'm trying to do here is completely reduce collisions. This 474 * may be a little aggressive. Should I allow for two collisions max? 475 */ 476 if (keg->uk_flags & UMA_ZONE_HASH && 477 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 478 struct uma_hash newhash; 479 struct uma_hash oldhash; 480 int ret; 481 482 /* 483 * This is so involved because allocating and freeing 484 * while the keg lock is held will lead to deadlock. 485 * I have to do everything in stages and check for 486 * races. 487 */ 488 newhash = keg->uk_hash; 489 KEG_UNLOCK(keg); 490 ret = hash_alloc(&newhash); 491 KEG_LOCK(keg); 492 if (ret) { 493 if (hash_expand(&keg->uk_hash, &newhash)) { 494 oldhash = keg->uk_hash; 495 keg->uk_hash = newhash; 496 } else 497 oldhash = newhash; 498 499 KEG_UNLOCK(keg); 500 hash_free(&oldhash); 501 return; 502 } 503 } 504 KEG_UNLOCK(keg); 505} 506 507static void 508zone_timeout(uma_zone_t zone) 509{ 510 511 zone_foreach_keg(zone, &keg_timeout); 512} 513 514/* 515 * Allocate and zero fill the next sized hash table from the appropriate 516 * backing store. 517 * 518 * Arguments: 519 * hash A new hash structure with the old hash size in uh_hashsize 520 * 521 * Returns: 522 * 1 on sucess and 0 on failure. 523 */ 524static int 525hash_alloc(struct uma_hash *hash) 526{ 527 int oldsize; 528 int alloc; 529 530 oldsize = hash->uh_hashsize; 531 532 /* We're just going to go to a power of two greater */ 533 if (oldsize) { 534 hash->uh_hashsize = oldsize * 2; 535 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 536 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 537 M_UMAHASH, M_NOWAIT); 538 } else { 539 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 540 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 541 M_WAITOK); 542 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 543 } 544 if (hash->uh_slab_hash) { 545 bzero(hash->uh_slab_hash, alloc); 546 hash->uh_hashmask = hash->uh_hashsize - 1; 547 return (1); 548 } 549 550 return (0); 551} 552 553/* 554 * Expands the hash table for HASH zones. This is done from zone_timeout 555 * to reduce collisions. This must not be done in the regular allocation 556 * path, otherwise, we can recurse on the vm while allocating pages. 557 * 558 * Arguments: 559 * oldhash The hash you want to expand 560 * newhash The hash structure for the new table 561 * 562 * Returns: 563 * Nothing 564 * 565 * Discussion: 566 */ 567static int 568hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 569{ 570 uma_slab_t slab; 571 int hval; 572 int i; 573 574 if (!newhash->uh_slab_hash) 575 return (0); 576 577 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 578 return (0); 579 580 /* 581 * I need to investigate hash algorithms for resizing without a 582 * full rehash. 583 */ 584 585 for (i = 0; i < oldhash->uh_hashsize; i++) 586 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 587 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 588 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 589 hval = UMA_HASH(newhash, slab->us_data); 590 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 591 slab, us_hlink); 592 } 593 594 return (1); 595} 596 597/* 598 * Free the hash bucket to the appropriate backing store. 599 * 600 * Arguments: 601 * slab_hash The hash bucket we're freeing 602 * hashsize The number of entries in that hash bucket 603 * 604 * Returns: 605 * Nothing 606 */ 607static void 608hash_free(struct uma_hash *hash) 609{ 610 if (hash->uh_slab_hash == NULL) 611 return; 612 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 613 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 614 else 615 free(hash->uh_slab_hash, M_UMAHASH); 616} 617 618/* 619 * Frees all outstanding items in a bucket 620 * 621 * Arguments: 622 * zone The zone to free to, must be unlocked. 623 * bucket The free/alloc bucket with items, cpu queue must be locked. 624 * 625 * Returns: 626 * Nothing 627 */ 628 629static void 630bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 631{ 632 int i; 633 634 if (bucket == NULL) 635 return; 636 637 if (zone->uz_fini) 638 for (i = 0; i < bucket->ub_cnt; i++) 639 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 640 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 641 bucket->ub_cnt = 0; 642} 643 644/* 645 * Drains the per cpu caches for a zone. 646 * 647 * NOTE: This may only be called while the zone is being turn down, and not 648 * during normal operation. This is necessary in order that we do not have 649 * to migrate CPUs to drain the per-CPU caches. 650 * 651 * Arguments: 652 * zone The zone to drain, must be unlocked. 653 * 654 * Returns: 655 * Nothing 656 */ 657static void 658cache_drain(uma_zone_t zone) 659{ 660 uma_cache_t cache; 661 int cpu; 662 663 /* 664 * XXX: It is safe to not lock the per-CPU caches, because we're 665 * tearing down the zone anyway. I.e., there will be no further use 666 * of the caches at this point. 667 * 668 * XXX: It would good to be able to assert that the zone is being 669 * torn down to prevent improper use of cache_drain(). 670 * 671 * XXX: We lock the zone before passing into bucket_cache_drain() as 672 * it is used elsewhere. Should the tear-down path be made special 673 * there in some form? 674 */ 675 CPU_FOREACH(cpu) { 676 cache = &zone->uz_cpu[cpu]; 677 bucket_drain(zone, cache->uc_allocbucket); 678 bucket_drain(zone, cache->uc_freebucket); 679 if (cache->uc_allocbucket != NULL) 680 bucket_free(zone, cache->uc_allocbucket, NULL); 681 if (cache->uc_freebucket != NULL) 682 bucket_free(zone, cache->uc_freebucket, NULL); 683 cache->uc_allocbucket = cache->uc_freebucket = NULL; 684 } 685 ZONE_LOCK(zone); 686 bucket_cache_drain(zone); 687 ZONE_UNLOCK(zone); 688} 689 690static void 691cache_shrink(uma_zone_t zone) 692{ 693 694 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 695 return; 696 697 ZONE_LOCK(zone); 698 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; 699 ZONE_UNLOCK(zone); 700} 701 702static void 703cache_drain_safe_cpu(uma_zone_t zone) 704{ 705 uma_cache_t cache; 706 uma_bucket_t b1, b2; 707 708 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 709 return; 710 711 b1 = b2 = NULL; 712 ZONE_LOCK(zone); 713 critical_enter(); 714 cache = &zone->uz_cpu[curcpu]; 715 if (cache->uc_allocbucket) { 716 if (cache->uc_allocbucket->ub_cnt != 0) 717 LIST_INSERT_HEAD(&zone->uz_buckets, 718 cache->uc_allocbucket, ub_link); 719 else 720 b1 = cache->uc_allocbucket; 721 cache->uc_allocbucket = NULL; 722 } 723 if (cache->uc_freebucket) { 724 if (cache->uc_freebucket->ub_cnt != 0) 725 LIST_INSERT_HEAD(&zone->uz_buckets, 726 cache->uc_freebucket, ub_link); 727 else 728 b2 = cache->uc_freebucket; 729 cache->uc_freebucket = NULL; 730 } 731 critical_exit(); 732 ZONE_UNLOCK(zone); 733 if (b1) 734 bucket_free(zone, b1, NULL); 735 if (b2) 736 bucket_free(zone, b2, NULL); 737} 738 739/* 740 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 741 * This is an expensive call because it needs to bind to all CPUs 742 * one by one and enter a critical section on each of them in order 743 * to safely access their cache buckets. 744 * Zone lock must not be held on call this function. 745 */ 746static void 747cache_drain_safe(uma_zone_t zone) 748{ 749 int cpu; 750 751 /* 752 * Polite bucket sizes shrinking was not enouth, shrink aggressively. 753 */ 754 if (zone) 755 cache_shrink(zone); 756 else 757 zone_foreach(cache_shrink); 758 759 CPU_FOREACH(cpu) { 760 thread_lock(curthread); 761 sched_bind(curthread, cpu); 762 thread_unlock(curthread); 763 764 if (zone) 765 cache_drain_safe_cpu(zone); 766 else 767 zone_foreach(cache_drain_safe_cpu); 768 } 769 thread_lock(curthread); 770 sched_unbind(curthread); 771 thread_unlock(curthread); 772} 773 774/* 775 * Drain the cached buckets from a zone. Expects a locked zone on entry. 776 */ 777static void 778bucket_cache_drain(uma_zone_t zone) 779{ 780 uma_bucket_t bucket; 781 782 /* 783 * Drain the bucket queues and free the buckets, we just keep two per 784 * cpu (alloc/free). 785 */ 786 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 787 LIST_REMOVE(bucket, ub_link); 788 ZONE_UNLOCK(zone); 789 bucket_drain(zone, bucket); 790 bucket_free(zone, bucket, NULL); 791 ZONE_LOCK(zone); 792 } 793 794 /* 795 * Shrink further bucket sizes. Price of single zone lock collision 796 * is probably lower then price of global cache drain. 797 */ 798 if (zone->uz_count > zone->uz_count_min) 799 zone->uz_count--; 800} 801 802static void 803keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 804{ 805 uint8_t *mem; 806 int i; 807 uint8_t flags; 808 809 mem = slab->us_data; 810 flags = slab->us_flags; 811 i = start; 812 if (keg->uk_fini != NULL) { 813 for (i--; i > -1; i--) 814 keg->uk_fini(slab->us_data + (keg->uk_rsize * i), 815 keg->uk_size); 816 } 817 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 818 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 819#ifdef UMA_DEBUG 820 printf("%s: Returning %d bytes.\n", keg->uk_name, 821 PAGE_SIZE * keg->uk_ppera); 822#endif 823 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 824} 825 826/* 827 * Frees pages from a keg back to the system. This is done on demand from 828 * the pageout daemon. 829 * 830 * Returns nothing. 831 */ 832static void 833keg_drain(uma_keg_t keg) 834{ 835 struct slabhead freeslabs = { 0 }; 836 uma_slab_t slab; 837 uma_slab_t n; 838 839 /* 840 * We don't want to take pages from statically allocated kegs at this 841 * time 842 */ 843 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 844 return; 845 846#ifdef UMA_DEBUG 847 printf("%s free items: %u\n", keg->uk_name, keg->uk_free); 848#endif 849 KEG_LOCK(keg); 850 if (keg->uk_free == 0) 851 goto finished; 852 853 slab = LIST_FIRST(&keg->uk_free_slab); 854 while (slab) { 855 n = LIST_NEXT(slab, us_link); 856 857 /* We have no where to free these to */ 858 if (slab->us_flags & UMA_SLAB_BOOT) { 859 slab = n; 860 continue; 861 } 862 863 LIST_REMOVE(slab, us_link); 864 keg->uk_pages -= keg->uk_ppera; 865 keg->uk_free -= keg->uk_ipers; 866 867 if (keg->uk_flags & UMA_ZONE_HASH) 868 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); 869 870 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 871 872 slab = n; 873 } 874finished: 875 KEG_UNLOCK(keg); 876 877 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 878 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 879 keg_free_slab(keg, slab, keg->uk_ipers); 880 } 881} 882 883static void 884zone_drain_wait(uma_zone_t zone, int waitok) 885{ 886 887 /* 888 * Set draining to interlock with zone_dtor() so we can release our 889 * locks as we go. Only dtor() should do a WAITOK call since it 890 * is the only call that knows the structure will still be available 891 * when it wakes up. 892 */ 893 ZONE_LOCK(zone); 894 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 895 if (waitok == M_NOWAIT) 896 goto out; 897 mtx_unlock(&uma_mtx); 898 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 899 mtx_lock(&uma_mtx); 900 } 901 zone->uz_flags |= UMA_ZFLAG_DRAINING; 902 bucket_cache_drain(zone); 903 ZONE_UNLOCK(zone); 904 /* 905 * The DRAINING flag protects us from being freed while 906 * we're running. Normally the uma_mtx would protect us but we 907 * must be able to release and acquire the right lock for each keg. 908 */ 909 zone_foreach_keg(zone, &keg_drain); 910 ZONE_LOCK(zone); 911 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 912 wakeup(zone); 913out: 914 ZONE_UNLOCK(zone); 915} 916 917void 918zone_drain(uma_zone_t zone) 919{ 920 921 zone_drain_wait(zone, M_NOWAIT); 922} 923 924/* 925 * Allocate a new slab for a keg. This does not insert the slab onto a list. 926 * 927 * Arguments: 928 * wait Shall we wait? 929 * 930 * Returns: 931 * The slab that was allocated or NULL if there is no memory and the 932 * caller specified M_NOWAIT. 933 */ 934static uma_slab_t 935keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) 936{ 937 uma_slabrefcnt_t slabref; 938 uma_alloc allocf; 939 uma_slab_t slab; 940 uint8_t *mem; 941 uint8_t flags; 942 int i; 943 944 mtx_assert(&keg->uk_lock, MA_OWNED); 945 slab = NULL; 946 mem = NULL; 947 948#ifdef UMA_DEBUG 949 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name); 950#endif 951 allocf = keg->uk_allocf; 952 KEG_UNLOCK(keg); 953 954 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 955 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); 956 if (slab == NULL) 957 goto out; 958 } 959 960 /* 961 * This reproduces the old vm_zone behavior of zero filling pages the 962 * first time they are added to a zone. 963 * 964 * Malloced items are zeroed in uma_zalloc. 965 */ 966 967 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 968 wait |= M_ZERO; 969 else 970 wait &= ~M_ZERO; 971 972 if (keg->uk_flags & UMA_ZONE_NODUMP) 973 wait |= M_NODUMP; 974 975 /* zone is passed for legacy reasons. */ 976 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait); 977 if (mem == NULL) { 978 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 979 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 980 slab = NULL; 981 goto out; 982 } 983 984 /* Point the slab into the allocated memory */ 985 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 986 slab = (uma_slab_t )(mem + keg->uk_pgoff); 987 988 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 989 for (i = 0; i < keg->uk_ppera; i++) 990 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 991 992 slab->us_keg = keg; 993 slab->us_data = mem; 994 slab->us_freecount = keg->uk_ipers; 995 slab->us_flags = flags; 996 BIT_FILL(SLAB_SETSIZE, &slab->us_free); 997#ifdef INVARIANTS 998 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); 999#endif 1000 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1001 slabref = (uma_slabrefcnt_t)slab; 1002 for (i = 0; i < keg->uk_ipers; i++) 1003 slabref->us_refcnt[i] = 0; 1004 } 1005 1006 if (keg->uk_init != NULL) { 1007 for (i = 0; i < keg->uk_ipers; i++) 1008 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 1009 keg->uk_size, wait) != 0) 1010 break; 1011 if (i != keg->uk_ipers) { 1012 keg_free_slab(keg, slab, i); 1013 slab = NULL; 1014 goto out; 1015 } 1016 } 1017out: 1018 KEG_LOCK(keg); 1019 1020 if (slab != NULL) { 1021 if (keg->uk_flags & UMA_ZONE_HASH) 1022 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1023 1024 keg->uk_pages += keg->uk_ppera; 1025 keg->uk_free += keg->uk_ipers; 1026 } 1027 1028 return (slab); 1029} 1030 1031/* 1032 * This function is intended to be used early on in place of page_alloc() so 1033 * that we may use the boot time page cache to satisfy allocations before 1034 * the VM is ready. 1035 */ 1036static void * 1037startup_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait) 1038{ 1039 uma_keg_t keg; 1040 uma_slab_t tmps; 1041 int pages, check_pages; 1042 1043 keg = zone_first_keg(zone); 1044 pages = howmany(bytes, PAGE_SIZE); 1045 check_pages = pages - 1; 1046 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n")); 1047 1048 /* 1049 * Check our small startup cache to see if it has pages remaining. 1050 */ 1051 mtx_lock(&uma_boot_pages_mtx); 1052 1053 /* First check if we have enough room. */ 1054 tmps = LIST_FIRST(&uma_boot_pages); 1055 while (tmps != NULL && check_pages-- > 0) 1056 tmps = LIST_NEXT(tmps, us_link); 1057 if (tmps != NULL) { 1058 /* 1059 * It's ok to lose tmps references. The last one will 1060 * have tmps->us_data pointing to the start address of 1061 * "pages" contiguous pages of memory. 1062 */ 1063 while (pages-- > 0) { 1064 tmps = LIST_FIRST(&uma_boot_pages); 1065 LIST_REMOVE(tmps, us_link); 1066 } 1067 mtx_unlock(&uma_boot_pages_mtx); 1068 *pflag = tmps->us_flags; 1069 return (tmps->us_data); 1070 } 1071 mtx_unlock(&uma_boot_pages_mtx); 1072 if (booted < UMA_STARTUP2) 1073 panic("UMA: Increase vm.boot_pages"); 1074 /* 1075 * Now that we've booted reset these users to their real allocator. 1076 */ 1077#ifdef UMA_MD_SMALL_ALLOC 1078 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; 1079#else 1080 keg->uk_allocf = page_alloc; 1081#endif 1082 return keg->uk_allocf(zone, bytes, pflag, wait); 1083} 1084 1085/* 1086 * Allocates a number of pages from the system 1087 * 1088 * Arguments: 1089 * bytes The number of bytes requested 1090 * wait Shall we wait? 1091 * 1092 * Returns: 1093 * A pointer to the alloced memory or possibly 1094 * NULL if M_NOWAIT is set. 1095 */ 1096static void * 1097page_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait) 1098{ 1099 void *p; /* Returned page */ 1100 1101 *pflag = UMA_SLAB_KMEM; 1102 p = (void *) kmem_malloc(kmem_arena, bytes, wait); 1103 1104 return (p); 1105} 1106 1107/* 1108 * Allocates a number of pages from within an object 1109 * 1110 * Arguments: 1111 * bytes The number of bytes requested 1112 * wait Shall we wait? 1113 * 1114 * Returns: 1115 * A pointer to the alloced memory or possibly 1116 * NULL if M_NOWAIT is set. 1117 */ 1118static void * 1119noobj_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait) 1120{ 1121 TAILQ_HEAD(, vm_page) alloctail; 1122 u_long npages; 1123 vm_offset_t retkva, zkva; 1124 vm_page_t p, p_next; 1125 uma_keg_t keg; 1126 1127 TAILQ_INIT(&alloctail); 1128 keg = zone_first_keg(zone); 1129 1130 npages = howmany(bytes, PAGE_SIZE); 1131 while (npages > 0) { 1132 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT | 1133 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ); 1134 if (p != NULL) { 1135 /* 1136 * Since the page does not belong to an object, its 1137 * listq is unused. 1138 */ 1139 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1140 npages--; 1141 continue; 1142 } 1143 if (wait & M_WAITOK) { 1144 VM_WAIT; 1145 continue; 1146 } 1147 1148 /* 1149 * Page allocation failed, free intermediate pages and 1150 * exit. 1151 */ 1152 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1153 vm_page_unwire(p, 0); 1154 vm_page_free(p); 1155 } 1156 return (NULL); 1157 } 1158 *flags = UMA_SLAB_PRIV; 1159 zkva = keg->uk_kva + 1160 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1161 retkva = zkva; 1162 TAILQ_FOREACH(p, &alloctail, listq) { 1163 pmap_qenter(zkva, &p, 1); 1164 zkva += PAGE_SIZE; 1165 } 1166 1167 return ((void *)retkva); 1168} 1169 1170/* 1171 * Frees a number of pages to the system 1172 * 1173 * Arguments: 1174 * mem A pointer to the memory to be freed 1175 * size The size of the memory being freed 1176 * flags The original p->us_flags field 1177 * 1178 * Returns: 1179 * Nothing 1180 */ 1181static void 1182page_free(void *mem, int size, uint8_t flags) 1183{ 1184 struct vmem *vmem; 1185 1186 if (flags & UMA_SLAB_KMEM) 1187 vmem = kmem_arena; 1188 else if (flags & UMA_SLAB_KERNEL) 1189 vmem = kernel_arena; 1190 else 1191 panic("UMA: page_free used with invalid flags %d", flags); 1192 1193 kmem_free(vmem, (vm_offset_t)mem, size); 1194} 1195 1196/* 1197 * Zero fill initializer 1198 * 1199 * Arguments/Returns follow uma_init specifications 1200 */ 1201static int 1202zero_init(void *mem, int size, int flags) 1203{ 1204 bzero(mem, size); 1205 return (0); 1206} 1207 1208/* 1209 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1210 * 1211 * Arguments 1212 * keg The zone we should initialize 1213 * 1214 * Returns 1215 * Nothing 1216 */ 1217static void 1218keg_small_init(uma_keg_t keg) 1219{ 1220 u_int rsize; 1221 u_int memused; 1222 u_int wastedspace; 1223 u_int shsize; 1224 1225 if (keg->uk_flags & UMA_ZONE_PCPU) { 1226 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU; 1227 1228 keg->uk_slabsize = sizeof(struct pcpu); 1229 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu), 1230 PAGE_SIZE); 1231 } else { 1232 keg->uk_slabsize = UMA_SLAB_SIZE; 1233 keg->uk_ppera = 1; 1234 } 1235 1236 /* 1237 * Calculate the size of each allocation (rsize) according to 1238 * alignment. If the requested size is smaller than we have 1239 * allocation bits for we round it up. 1240 */ 1241 rsize = keg->uk_size; 1242 if (rsize < keg->uk_slabsize / SLAB_SETSIZE) 1243 rsize = keg->uk_slabsize / SLAB_SETSIZE; 1244 if (rsize & keg->uk_align) 1245 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1246 keg->uk_rsize = rsize; 1247 1248 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1249 keg->uk_rsize < sizeof(struct pcpu), 1250 ("%s: size %u too large", __func__, keg->uk_rsize)); 1251 1252 if (keg->uk_flags & UMA_ZONE_REFCNT) 1253 rsize += sizeof(uint32_t); 1254 1255 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1256 shsize = 0; 1257 else 1258 shsize = sizeof(struct uma_slab); 1259 1260 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize; 1261 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1262 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1263 1264 memused = keg->uk_ipers * rsize + shsize; 1265 wastedspace = keg->uk_slabsize - memused; 1266 1267 /* 1268 * We can't do OFFPAGE if we're internal or if we've been 1269 * asked to not go to the VM for buckets. If we do this we 1270 * may end up going to the VM for slabs which we do not 1271 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1272 * of UMA_ZONE_VM, which clearly forbids it. 1273 */ 1274 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1275 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1276 return; 1277 1278 /* 1279 * See if using an OFFPAGE slab will limit our waste. Only do 1280 * this if it permits more items per-slab. 1281 * 1282 * XXX We could try growing slabsize to limit max waste as well. 1283 * Historically this was not done because the VM could not 1284 * efficiently handle contiguous allocations. 1285 */ 1286 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) && 1287 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) { 1288 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize; 1289 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1290 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1291#ifdef UMA_DEBUG 1292 printf("UMA decided we need offpage slab headers for " 1293 "keg: %s, calculated wastedspace = %d, " 1294 "maximum wasted space allowed = %d, " 1295 "calculated ipers = %d, " 1296 "new wasted space = %d\n", keg->uk_name, wastedspace, 1297 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1298 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize); 1299#endif 1300 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1301 } 1302 1303 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1304 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1305 keg->uk_flags |= UMA_ZONE_HASH; 1306} 1307 1308/* 1309 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1310 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1311 * more complicated. 1312 * 1313 * Arguments 1314 * keg The keg we should initialize 1315 * 1316 * Returns 1317 * Nothing 1318 */ 1319static void 1320keg_large_init(uma_keg_t keg) 1321{ 1322 u_int shsize; 1323 1324 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1325 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1326 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1327 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1328 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1329 1330 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1331 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE; 1332 keg->uk_ipers = 1; 1333 keg->uk_rsize = keg->uk_size; 1334 1335 /* We can't do OFFPAGE if we're internal, bail out here. */ 1336 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) 1337 return; 1338 1339 /* Check whether we have enough space to not do OFFPAGE. */ 1340 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) { 1341 shsize = sizeof(struct uma_slab); 1342 if (keg->uk_flags & UMA_ZONE_REFCNT) 1343 shsize += keg->uk_ipers * sizeof(uint32_t); 1344 if (shsize & UMA_ALIGN_PTR) 1345 shsize = (shsize & ~UMA_ALIGN_PTR) + 1346 (UMA_ALIGN_PTR + 1); 1347 1348 if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize) 1349 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1350 } 1351 1352 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1353 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1354 keg->uk_flags |= UMA_ZONE_HASH; 1355} 1356 1357static void 1358keg_cachespread_init(uma_keg_t keg) 1359{ 1360 int alignsize; 1361 int trailer; 1362 int pages; 1363 int rsize; 1364 1365 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1366 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1367 1368 alignsize = keg->uk_align + 1; 1369 rsize = keg->uk_size; 1370 /* 1371 * We want one item to start on every align boundary in a page. To 1372 * do this we will span pages. We will also extend the item by the 1373 * size of align if it is an even multiple of align. Otherwise, it 1374 * would fall on the same boundary every time. 1375 */ 1376 if (rsize & keg->uk_align) 1377 rsize = (rsize & ~keg->uk_align) + alignsize; 1378 if ((rsize & alignsize) == 0) 1379 rsize += alignsize; 1380 trailer = rsize - keg->uk_size; 1381 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1382 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1383 keg->uk_rsize = rsize; 1384 keg->uk_ppera = pages; 1385 keg->uk_slabsize = UMA_SLAB_SIZE; 1386 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1387 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1388 KASSERT(keg->uk_ipers <= SLAB_SETSIZE, 1389 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1390 keg->uk_ipers)); 1391} 1392 1393/* 1394 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1395 * the keg onto the global keg list. 1396 * 1397 * Arguments/Returns follow uma_ctor specifications 1398 * udata Actually uma_kctor_args 1399 */ 1400static int 1401keg_ctor(void *mem, int size, void *udata, int flags) 1402{ 1403 struct uma_kctor_args *arg = udata; 1404 uma_keg_t keg = mem; 1405 uma_zone_t zone; 1406 1407 bzero(keg, size); 1408 keg->uk_size = arg->size; 1409 keg->uk_init = arg->uminit; 1410 keg->uk_fini = arg->fini; 1411 keg->uk_align = arg->align; 1412 keg->uk_free = 0; 1413 keg->uk_reserve = 0; 1414 keg->uk_pages = 0; 1415 keg->uk_flags = arg->flags; 1416 keg->uk_allocf = page_alloc; 1417 keg->uk_freef = page_free; 1418 keg->uk_slabzone = NULL; 1419 1420 /* 1421 * The master zone is passed to us at keg-creation time. 1422 */ 1423 zone = arg->zone; 1424 keg->uk_name = zone->uz_name; 1425 1426 if (arg->flags & UMA_ZONE_VM) 1427 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1428 1429 if (arg->flags & UMA_ZONE_ZINIT) 1430 keg->uk_init = zero_init; 1431 1432 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC) 1433 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1434 1435 if (arg->flags & UMA_ZONE_PCPU) 1436#ifdef SMP 1437 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1438#else 1439 keg->uk_flags &= ~UMA_ZONE_PCPU; 1440#endif 1441 1442 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1443 keg_cachespread_init(keg); 1444 } else if (keg->uk_flags & UMA_ZONE_REFCNT) { 1445 if (keg->uk_size > 1446 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - 1447 sizeof(uint32_t))) 1448 keg_large_init(keg); 1449 else 1450 keg_small_init(keg); 1451 } else { 1452 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab))) 1453 keg_large_init(keg); 1454 else 1455 keg_small_init(keg); 1456 } 1457 1458 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1459 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1460 if (keg->uk_ipers > uma_max_ipers_ref) 1461 panic("Too many ref items per zone: %d > %d\n", 1462 keg->uk_ipers, uma_max_ipers_ref); 1463 keg->uk_slabzone = slabrefzone; 1464 } else 1465 keg->uk_slabzone = slabzone; 1466 } 1467 1468 /* 1469 * If we haven't booted yet we need allocations to go through the 1470 * startup cache until the vm is ready. 1471 */ 1472 if (keg->uk_ppera == 1) { 1473#ifdef UMA_MD_SMALL_ALLOC 1474 keg->uk_allocf = uma_small_alloc; 1475 keg->uk_freef = uma_small_free; 1476 1477 if (booted < UMA_STARTUP) 1478 keg->uk_allocf = startup_alloc; 1479#else 1480 if (booted < UMA_STARTUP2) 1481 keg->uk_allocf = startup_alloc; 1482#endif 1483 } else if (booted < UMA_STARTUP2 && 1484 (keg->uk_flags & UMA_ZFLAG_INTERNAL)) 1485 keg->uk_allocf = startup_alloc; 1486 1487 /* 1488 * Initialize keg's lock 1489 */ 1490 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1491 1492 /* 1493 * If we're putting the slab header in the actual page we need to 1494 * figure out where in each page it goes. This calculates a right 1495 * justified offset into the memory on an ALIGN_PTR boundary. 1496 */ 1497 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1498 u_int totsize; 1499 1500 /* Size of the slab struct and free list */ 1501 totsize = sizeof(struct uma_slab); 1502 1503 /* Size of the reference counts. */ 1504 if (keg->uk_flags & UMA_ZONE_REFCNT) 1505 totsize += keg->uk_ipers * sizeof(uint32_t); 1506 1507 if (totsize & UMA_ALIGN_PTR) 1508 totsize = (totsize & ~UMA_ALIGN_PTR) + 1509 (UMA_ALIGN_PTR + 1); 1510 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1511 1512 /* 1513 * The only way the following is possible is if with our 1514 * UMA_ALIGN_PTR adjustments we are now bigger than 1515 * UMA_SLAB_SIZE. I haven't checked whether this is 1516 * mathematically possible for all cases, so we make 1517 * sure here anyway. 1518 */ 1519 totsize = keg->uk_pgoff + sizeof(struct uma_slab); 1520 if (keg->uk_flags & UMA_ZONE_REFCNT) 1521 totsize += keg->uk_ipers * sizeof(uint32_t); 1522 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1523 printf("zone %s ipers %d rsize %d size %d\n", 1524 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1525 keg->uk_size); 1526 panic("UMA slab won't fit."); 1527 } 1528 } 1529 1530 if (keg->uk_flags & UMA_ZONE_HASH) 1531 hash_alloc(&keg->uk_hash); 1532 1533#ifdef UMA_DEBUG 1534 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n", 1535 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags, 1536 keg->uk_ipers, keg->uk_ppera, 1537 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free); 1538#endif 1539 1540 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1541 1542 mtx_lock(&uma_mtx); 1543 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1544 mtx_unlock(&uma_mtx); 1545 return (0); 1546} 1547 1548/* 1549 * Zone header ctor. This initializes all fields, locks, etc. 1550 * 1551 * Arguments/Returns follow uma_ctor specifications 1552 * udata Actually uma_zctor_args 1553 */ 1554static int 1555zone_ctor(void *mem, int size, void *udata, int flags) 1556{ 1557 struct uma_zctor_args *arg = udata; 1558 uma_zone_t zone = mem; 1559 uma_zone_t z; 1560 uma_keg_t keg; 1561 1562 bzero(zone, size); 1563 zone->uz_name = arg->name; 1564 zone->uz_ctor = arg->ctor; 1565 zone->uz_dtor = arg->dtor; 1566 zone->uz_slab = zone_fetch_slab; 1567 zone->uz_init = NULL; 1568 zone->uz_fini = NULL; 1569 zone->uz_allocs = 0; 1570 zone->uz_frees = 0; 1571 zone->uz_fails = 0; 1572 zone->uz_sleeps = 0; 1573 zone->uz_count = 0; 1574 zone->uz_count_min = 0; 1575 zone->uz_flags = 0; 1576 zone->uz_warning = NULL; 1577 timevalclear(&zone->uz_ratecheck); 1578 keg = arg->keg; 1579 1580 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 1581 1582 /* 1583 * This is a pure cache zone, no kegs. 1584 */ 1585 if (arg->import) { 1586 if (arg->flags & UMA_ZONE_VM) 1587 arg->flags |= UMA_ZFLAG_CACHEONLY; 1588 zone->uz_flags = arg->flags; 1589 zone->uz_size = arg->size; 1590 zone->uz_import = arg->import; 1591 zone->uz_release = arg->release; 1592 zone->uz_arg = arg->arg; 1593 zone->uz_lockptr = &zone->uz_lock; 1594 mtx_lock(&uma_mtx); 1595 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 1596 mtx_unlock(&uma_mtx); 1597 goto out; 1598 } 1599 1600 /* 1601 * Use the regular zone/keg/slab allocator. 1602 */ 1603 zone->uz_import = (uma_import)zone_import; 1604 zone->uz_release = (uma_release)zone_release; 1605 zone->uz_arg = zone; 1606 1607 if (arg->flags & UMA_ZONE_SECONDARY) { 1608 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1609 zone->uz_init = arg->uminit; 1610 zone->uz_fini = arg->fini; 1611 zone->uz_lockptr = &keg->uk_lock; 1612 zone->uz_flags |= UMA_ZONE_SECONDARY; 1613 mtx_lock(&uma_mtx); 1614 ZONE_LOCK(zone); 1615 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1616 if (LIST_NEXT(z, uz_link) == NULL) { 1617 LIST_INSERT_AFTER(z, zone, uz_link); 1618 break; 1619 } 1620 } 1621 ZONE_UNLOCK(zone); 1622 mtx_unlock(&uma_mtx); 1623 } else if (keg == NULL) { 1624 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1625 arg->align, arg->flags)) == NULL) 1626 return (ENOMEM); 1627 } else { 1628 struct uma_kctor_args karg; 1629 int error; 1630 1631 /* We should only be here from uma_startup() */ 1632 karg.size = arg->size; 1633 karg.uminit = arg->uminit; 1634 karg.fini = arg->fini; 1635 karg.align = arg->align; 1636 karg.flags = arg->flags; 1637 karg.zone = zone; 1638 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1639 flags); 1640 if (error) 1641 return (error); 1642 } 1643 1644 /* 1645 * Link in the first keg. 1646 */ 1647 zone->uz_klink.kl_keg = keg; 1648 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1649 zone->uz_lockptr = &keg->uk_lock; 1650 zone->uz_size = keg->uk_size; 1651 zone->uz_flags |= (keg->uk_flags & 1652 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1653 1654 /* 1655 * Some internal zones don't have room allocated for the per cpu 1656 * caches. If we're internal, bail out here. 1657 */ 1658 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1659 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1660 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1661 return (0); 1662 } 1663 1664out: 1665 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0) 1666 zone->uz_count = bucket_select(zone->uz_size); 1667 else 1668 zone->uz_count = BUCKET_MAX; 1669 zone->uz_count_min = zone->uz_count; 1670 1671 return (0); 1672} 1673 1674/* 1675 * Keg header dtor. This frees all data, destroys locks, frees the hash 1676 * table and removes the keg from the global list. 1677 * 1678 * Arguments/Returns follow uma_dtor specifications 1679 * udata unused 1680 */ 1681static void 1682keg_dtor(void *arg, int size, void *udata) 1683{ 1684 uma_keg_t keg; 1685 1686 keg = (uma_keg_t)arg; 1687 KEG_LOCK(keg); 1688 if (keg->uk_free != 0) { 1689 printf("Freed UMA keg (%s) was not empty (%d items). " 1690 " Lost %d pages of memory.\n", 1691 keg->uk_name ? keg->uk_name : "", 1692 keg->uk_free, keg->uk_pages); 1693 } 1694 KEG_UNLOCK(keg); 1695 1696 hash_free(&keg->uk_hash); 1697 1698 KEG_LOCK_FINI(keg); 1699} 1700 1701/* 1702 * Zone header dtor. 1703 * 1704 * Arguments/Returns follow uma_dtor specifications 1705 * udata unused 1706 */ 1707static void 1708zone_dtor(void *arg, int size, void *udata) 1709{ 1710 uma_klink_t klink; 1711 uma_zone_t zone; 1712 uma_keg_t keg; 1713 1714 zone = (uma_zone_t)arg; 1715 keg = zone_first_keg(zone); 1716 1717 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1718 cache_drain(zone); 1719 1720 mtx_lock(&uma_mtx); 1721 LIST_REMOVE(zone, uz_link); 1722 mtx_unlock(&uma_mtx); 1723 /* 1724 * XXX there are some races here where 1725 * the zone can be drained but zone lock 1726 * released and then refilled before we 1727 * remove it... we dont care for now 1728 */ 1729 zone_drain_wait(zone, M_WAITOK); 1730 /* 1731 * Unlink all of our kegs. 1732 */ 1733 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1734 klink->kl_keg = NULL; 1735 LIST_REMOVE(klink, kl_link); 1736 if (klink == &zone->uz_klink) 1737 continue; 1738 free(klink, M_TEMP); 1739 } 1740 /* 1741 * We only destroy kegs from non secondary zones. 1742 */ 1743 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1744 mtx_lock(&uma_mtx); 1745 LIST_REMOVE(keg, uk_link); 1746 mtx_unlock(&uma_mtx); 1747 zone_free_item(kegs, keg, NULL, SKIP_NONE); 1748 } 1749 ZONE_LOCK_FINI(zone); 1750} 1751 1752/* 1753 * Traverses every zone in the system and calls a callback 1754 * 1755 * Arguments: 1756 * zfunc A pointer to a function which accepts a zone 1757 * as an argument. 1758 * 1759 * Returns: 1760 * Nothing 1761 */ 1762static void 1763zone_foreach(void (*zfunc)(uma_zone_t)) 1764{ 1765 uma_keg_t keg; 1766 uma_zone_t zone; 1767 1768 mtx_lock(&uma_mtx); 1769 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1770 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1771 zfunc(zone); 1772 } 1773 mtx_unlock(&uma_mtx); 1774} 1775 1776/* Public functions */ 1777/* See uma.h */ 1778void 1779uma_startup(void *bootmem, int boot_pages) 1780{ 1781 struct uma_zctor_args args; 1782 uma_slab_t slab; 1783 u_int slabsize; 1784 int i; 1785 1786#ifdef UMA_DEBUG 1787 printf("Creating uma keg headers zone and keg.\n"); 1788#endif 1789 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF); 1790 1791 /* "manually" create the initial zone */ 1792 memset(&args, 0, sizeof(args)); 1793 args.name = "UMA Kegs"; 1794 args.size = sizeof(struct uma_keg); 1795 args.ctor = keg_ctor; 1796 args.dtor = keg_dtor; 1797 args.uminit = zero_init; 1798 args.fini = NULL; 1799 args.keg = &masterkeg; 1800 args.align = 32 - 1; 1801 args.flags = UMA_ZFLAG_INTERNAL; 1802 /* The initial zone has no Per cpu queues so it's smaller */ 1803 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); 1804 1805#ifdef UMA_DEBUG 1806 printf("Filling boot free list.\n"); 1807#endif 1808 for (i = 0; i < boot_pages; i++) { 1809 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE)); 1810 slab->us_data = (uint8_t *)slab; 1811 slab->us_flags = UMA_SLAB_BOOT; 1812 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); 1813 } 1814 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); 1815 1816#ifdef UMA_DEBUG 1817 printf("Creating uma zone headers zone and keg.\n"); 1818#endif 1819 args.name = "UMA Zones"; 1820 args.size = sizeof(struct uma_zone) + 1821 (sizeof(struct uma_cache) * (mp_maxid + 1)); 1822 args.ctor = zone_ctor; 1823 args.dtor = zone_dtor; 1824 args.uminit = zero_init; 1825 args.fini = NULL; 1826 args.keg = NULL; 1827 args.align = 32 - 1; 1828 args.flags = UMA_ZFLAG_INTERNAL; 1829 /* The initial zone has no Per cpu queues so it's smaller */ 1830 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); 1831 1832#ifdef UMA_DEBUG 1833 printf("Initializing pcpu cache locks.\n"); 1834#endif 1835#ifdef UMA_DEBUG 1836 printf("Creating slab and hash zones.\n"); 1837#endif 1838 1839 /* Now make a zone for slab headers */ 1840 slabzone = uma_zcreate("UMA Slabs", 1841 sizeof(struct uma_slab), 1842 NULL, NULL, NULL, NULL, 1843 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1844 1845 /* 1846 * We also create a zone for the bigger slabs with reference 1847 * counts in them, to accomodate UMA_ZONE_REFCNT zones. 1848 */ 1849 slabsize = sizeof(struct uma_slab_refcnt); 1850 slabsize += uma_max_ipers_ref * sizeof(uint32_t); 1851 slabrefzone = uma_zcreate("UMA RCntSlabs", 1852 slabsize, 1853 NULL, NULL, NULL, NULL, 1854 UMA_ALIGN_PTR, 1855 UMA_ZFLAG_INTERNAL); 1856 1857 hashzone = uma_zcreate("UMA Hash", 1858 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 1859 NULL, NULL, NULL, NULL, 1860 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1861 1862 bucket_init(); 1863 1864 booted = UMA_STARTUP; 1865 1866#ifdef UMA_DEBUG 1867 printf("UMA startup complete.\n"); 1868#endif 1869} 1870 1871/* see uma.h */ 1872void 1873uma_startup2(void) 1874{ 1875 booted = UMA_STARTUP2; 1876 bucket_enable(); 1877#ifdef UMA_DEBUG 1878 printf("UMA startup2 complete.\n"); 1879#endif 1880} 1881 1882/* 1883 * Initialize our callout handle 1884 * 1885 */ 1886 1887static void 1888uma_startup3(void) 1889{ 1890#ifdef UMA_DEBUG 1891 printf("Starting callout.\n"); 1892#endif 1893 callout_init(&uma_callout, CALLOUT_MPSAFE); 1894 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1895#ifdef UMA_DEBUG 1896 printf("UMA startup3 complete.\n"); 1897#endif 1898} 1899 1900static uma_keg_t 1901uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 1902 int align, uint32_t flags) 1903{ 1904 struct uma_kctor_args args; 1905 1906 args.size = size; 1907 args.uminit = uminit; 1908 args.fini = fini; 1909 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 1910 args.flags = flags; 1911 args.zone = zone; 1912 return (zone_alloc_item(kegs, &args, M_WAITOK)); 1913} 1914 1915/* See uma.h */ 1916void 1917uma_set_align(int align) 1918{ 1919 1920 if (align != UMA_ALIGN_CACHE) 1921 uma_align_cache = align; 1922} 1923 1924/* See uma.h */ 1925uma_zone_t 1926uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 1927 uma_init uminit, uma_fini fini, int align, uint32_t flags) 1928 1929{ 1930 struct uma_zctor_args args; 1931 1932 /* This stuff is essential for the zone ctor */ 1933 memset(&args, 0, sizeof(args)); 1934 args.name = name; 1935 args.size = size; 1936 args.ctor = ctor; 1937 args.dtor = dtor; 1938 args.uminit = uminit; 1939 args.fini = fini; 1940 args.align = align; 1941 args.flags = flags; 1942 args.keg = NULL; 1943 1944 return (zone_alloc_item(zones, &args, M_WAITOK)); 1945} 1946 1947/* See uma.h */ 1948uma_zone_t 1949uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 1950 uma_init zinit, uma_fini zfini, uma_zone_t master) 1951{ 1952 struct uma_zctor_args args; 1953 uma_keg_t keg; 1954 1955 keg = zone_first_keg(master); 1956 memset(&args, 0, sizeof(args)); 1957 args.name = name; 1958 args.size = keg->uk_size; 1959 args.ctor = ctor; 1960 args.dtor = dtor; 1961 args.uminit = zinit; 1962 args.fini = zfini; 1963 args.align = keg->uk_align; 1964 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 1965 args.keg = keg; 1966 1967 /* XXX Attaches only one keg of potentially many. */ 1968 return (zone_alloc_item(zones, &args, M_WAITOK)); 1969} 1970 1971/* See uma.h */ 1972uma_zone_t 1973uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 1974 uma_init zinit, uma_fini zfini, uma_import zimport, 1975 uma_release zrelease, void *arg, int flags) 1976{ 1977 struct uma_zctor_args args; 1978 1979 memset(&args, 0, sizeof(args)); 1980 args.name = name; 1981 args.size = size; 1982 args.ctor = ctor; 1983 args.dtor = dtor; 1984 args.uminit = zinit; 1985 args.fini = zfini; 1986 args.import = zimport; 1987 args.release = zrelease; 1988 args.arg = arg; 1989 args.align = 0; 1990 args.flags = flags; 1991 1992 return (zone_alloc_item(zones, &args, M_WAITOK)); 1993} 1994 1995static void 1996zone_lock_pair(uma_zone_t a, uma_zone_t b) 1997{ 1998 if (a < b) { 1999 ZONE_LOCK(a); 2000 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); 2001 } else { 2002 ZONE_LOCK(b); 2003 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); 2004 } 2005} 2006 2007static void 2008zone_unlock_pair(uma_zone_t a, uma_zone_t b) 2009{ 2010 2011 ZONE_UNLOCK(a); 2012 ZONE_UNLOCK(b); 2013} 2014 2015int 2016uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 2017{ 2018 uma_klink_t klink; 2019 uma_klink_t kl; 2020 int error; 2021 2022 error = 0; 2023 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 2024 2025 zone_lock_pair(zone, master); 2026 /* 2027 * zone must use vtoslab() to resolve objects and must already be 2028 * a secondary. 2029 */ 2030 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 2031 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 2032 error = EINVAL; 2033 goto out; 2034 } 2035 /* 2036 * The new master must also use vtoslab(). 2037 */ 2038 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 2039 error = EINVAL; 2040 goto out; 2041 } 2042 /* 2043 * Both must either be refcnt, or not be refcnt. 2044 */ 2045 if ((zone->uz_flags & UMA_ZONE_REFCNT) != 2046 (master->uz_flags & UMA_ZONE_REFCNT)) { 2047 error = EINVAL; 2048 goto out; 2049 } 2050 /* 2051 * The underlying object must be the same size. rsize 2052 * may be different. 2053 */ 2054 if (master->uz_size != zone->uz_size) { 2055 error = E2BIG; 2056 goto out; 2057 } 2058 /* 2059 * Put it at the end of the list. 2060 */ 2061 klink->kl_keg = zone_first_keg(master); 2062 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 2063 if (LIST_NEXT(kl, kl_link) == NULL) { 2064 LIST_INSERT_AFTER(kl, klink, kl_link); 2065 break; 2066 } 2067 } 2068 klink = NULL; 2069 zone->uz_flags |= UMA_ZFLAG_MULTI; 2070 zone->uz_slab = zone_fetch_slab_multi; 2071 2072out: 2073 zone_unlock_pair(zone, master); 2074 if (klink != NULL) 2075 free(klink, M_TEMP); 2076 2077 return (error); 2078} 2079 2080 2081/* See uma.h */ 2082void 2083uma_zdestroy(uma_zone_t zone) 2084{ 2085 2086 zone_free_item(zones, zone, NULL, SKIP_NONE); 2087} 2088 2089/* See uma.h */ 2090void * 2091uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2092{ 2093 void *item; 2094 uma_cache_t cache; 2095 uma_bucket_t bucket; 2096 int lockfail; 2097 int cpu; 2098 2099 /* This is the fast path allocation */ 2100#ifdef UMA_DEBUG_ALLOC_1 2101 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); 2102#endif 2103 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread, 2104 zone->uz_name, flags); 2105 2106 if (flags & M_WAITOK) { 2107 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2108 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2109 } 2110#ifdef DEBUG_MEMGUARD 2111 if (memguard_cmp_zone(zone)) { 2112 item = memguard_alloc(zone->uz_size, flags); 2113 if (item != NULL) { 2114 /* 2115 * Avoid conflict with the use-after-free 2116 * protecting infrastructure from INVARIANTS. 2117 */ 2118 if (zone->uz_init != NULL && 2119 zone->uz_init != mtrash_init && 2120 zone->uz_init(item, zone->uz_size, flags) != 0) 2121 return (NULL); 2122 if (zone->uz_ctor != NULL && 2123 zone->uz_ctor != mtrash_ctor && 2124 zone->uz_ctor(item, zone->uz_size, udata, 2125 flags) != 0) { 2126 zone->uz_fini(item, zone->uz_size); 2127 return (NULL); 2128 } 2129 return (item); 2130 } 2131 /* This is unfortunate but should not be fatal. */ 2132 } 2133#endif 2134 /* 2135 * If possible, allocate from the per-CPU cache. There are two 2136 * requirements for safe access to the per-CPU cache: (1) the thread 2137 * accessing the cache must not be preempted or yield during access, 2138 * and (2) the thread must not migrate CPUs without switching which 2139 * cache it accesses. We rely on a critical section to prevent 2140 * preemption and migration. We release the critical section in 2141 * order to acquire the zone mutex if we are unable to allocate from 2142 * the current cache; when we re-acquire the critical section, we 2143 * must detect and handle migration if it has occurred. 2144 */ 2145 critical_enter(); 2146 cpu = curcpu; 2147 cache = &zone->uz_cpu[cpu]; 2148 2149zalloc_start: 2150 bucket = cache->uc_allocbucket; 2151 if (bucket != NULL && bucket->ub_cnt > 0) { 2152 bucket->ub_cnt--; 2153 item = bucket->ub_bucket[bucket->ub_cnt]; 2154#ifdef INVARIANTS 2155 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2156#endif 2157 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2158 cache->uc_allocs++; 2159 critical_exit(); 2160 if (zone->uz_ctor != NULL && 2161 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2162 atomic_add_long(&zone->uz_fails, 1); 2163 zone_free_item(zone, item, udata, SKIP_DTOR); 2164 return (NULL); 2165 } 2166#ifdef INVARIANTS 2167 uma_dbg_alloc(zone, NULL, item); 2168#endif 2169 if (flags & M_ZERO) 2170 bzero(item, zone->uz_size); 2171 return (item); 2172 } 2173 2174 /* 2175 * We have run out of items in our alloc bucket. 2176 * See if we can switch with our free bucket. 2177 */ 2178 bucket = cache->uc_freebucket; 2179 if (bucket != NULL && bucket->ub_cnt > 0) { 2180#ifdef UMA_DEBUG_ALLOC 2181 printf("uma_zalloc: Swapping empty with alloc.\n"); 2182#endif 2183 cache->uc_freebucket = cache->uc_allocbucket; 2184 cache->uc_allocbucket = bucket; 2185 goto zalloc_start; 2186 } 2187 2188 /* 2189 * Discard any empty allocation bucket while we hold no locks. 2190 */ 2191 bucket = cache->uc_allocbucket; 2192 cache->uc_allocbucket = NULL; 2193 critical_exit(); 2194 if (bucket != NULL) 2195 bucket_free(zone, bucket, udata); 2196 2197 /* Short-circuit for zones without buckets and low memory. */ 2198 if (zone->uz_count == 0 || bucketdisable) 2199 goto zalloc_item; 2200 2201 /* 2202 * Attempt to retrieve the item from the per-CPU cache has failed, so 2203 * we must go back to the zone. This requires the zone lock, so we 2204 * must drop the critical section, then re-acquire it when we go back 2205 * to the cache. Since the critical section is released, we may be 2206 * preempted or migrate. As such, make sure not to maintain any 2207 * thread-local state specific to the cache from prior to releasing 2208 * the critical section. 2209 */ 2210 lockfail = 0; 2211 if (ZONE_TRYLOCK(zone) == 0) { 2212 /* Record contention to size the buckets. */ 2213 ZONE_LOCK(zone); 2214 lockfail = 1; 2215 } 2216 critical_enter(); 2217 cpu = curcpu; 2218 cache = &zone->uz_cpu[cpu]; 2219 2220 /* 2221 * Since we have locked the zone we may as well send back our stats. 2222 */ 2223 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2224 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2225 cache->uc_allocs = 0; 2226 cache->uc_frees = 0; 2227 2228 /* See if we lost the race to fill the cache. */ 2229 if (cache->uc_allocbucket != NULL) { 2230 ZONE_UNLOCK(zone); 2231 goto zalloc_start; 2232 } 2233 2234 /* 2235 * Check the zone's cache of buckets. 2236 */ 2237 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 2238 KASSERT(bucket->ub_cnt != 0, 2239 ("uma_zalloc_arg: Returning an empty bucket.")); 2240 2241 LIST_REMOVE(bucket, ub_link); 2242 cache->uc_allocbucket = bucket; 2243 ZONE_UNLOCK(zone); 2244 goto zalloc_start; 2245 } 2246 /* We are no longer associated with this CPU. */ 2247 critical_exit(); 2248 2249 /* 2250 * We bump the uz count when the cache size is insufficient to 2251 * handle the working set. 2252 */ 2253 if (lockfail && zone->uz_count < BUCKET_MAX) 2254 zone->uz_count++; 2255 ZONE_UNLOCK(zone); 2256 2257 /* 2258 * Now lets just fill a bucket and put it on the free list. If that 2259 * works we'll restart the allocation from the begining and it 2260 * will use the just filled bucket. 2261 */ 2262 bucket = zone_alloc_bucket(zone, udata, flags); 2263 if (bucket != NULL) { 2264 ZONE_LOCK(zone); 2265 critical_enter(); 2266 cpu = curcpu; 2267 cache = &zone->uz_cpu[cpu]; 2268 /* 2269 * See if we lost the race or were migrated. Cache the 2270 * initialized bucket to make this less likely or claim 2271 * the memory directly. 2272 */ 2273 if (cache->uc_allocbucket == NULL) 2274 cache->uc_allocbucket = bucket; 2275 else 2276 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2277 ZONE_UNLOCK(zone); 2278 goto zalloc_start; 2279 } 2280 2281 /* 2282 * We may not be able to get a bucket so return an actual item. 2283 */ 2284#ifdef UMA_DEBUG 2285 printf("uma_zalloc_arg: Bucketzone returned NULL\n"); 2286#endif 2287 2288zalloc_item: 2289 item = zone_alloc_item(zone, udata, flags); 2290 2291 return (item); 2292} 2293 2294static uma_slab_t 2295keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) 2296{ 2297 uma_slab_t slab; 2298 int reserve; 2299 2300 mtx_assert(&keg->uk_lock, MA_OWNED); 2301 slab = NULL; 2302 reserve = 0; 2303 if ((flags & M_USE_RESERVE) == 0) 2304 reserve = keg->uk_reserve; 2305 2306 for (;;) { 2307 /* 2308 * Find a slab with some space. Prefer slabs that are partially 2309 * used over those that are totally full. This helps to reduce 2310 * fragmentation. 2311 */ 2312 if (keg->uk_free > reserve) { 2313 if (!LIST_EMPTY(&keg->uk_part_slab)) { 2314 slab = LIST_FIRST(&keg->uk_part_slab); 2315 } else { 2316 slab = LIST_FIRST(&keg->uk_free_slab); 2317 LIST_REMOVE(slab, us_link); 2318 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, 2319 us_link); 2320 } 2321 MPASS(slab->us_keg == keg); 2322 return (slab); 2323 } 2324 2325 /* 2326 * M_NOVM means don't ask at all! 2327 */ 2328 if (flags & M_NOVM) 2329 break; 2330 2331 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2332 keg->uk_flags |= UMA_ZFLAG_FULL; 2333 /* 2334 * If this is not a multi-zone, set the FULL bit. 2335 * Otherwise slab_multi() takes care of it. 2336 */ 2337 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2338 zone->uz_flags |= UMA_ZFLAG_FULL; 2339 zone_log_warning(zone); 2340 } 2341 if (flags & M_NOWAIT) 2342 break; 2343 zone->uz_sleeps++; 2344 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2345 continue; 2346 } 2347 slab = keg_alloc_slab(keg, zone, flags); 2348 /* 2349 * If we got a slab here it's safe to mark it partially used 2350 * and return. We assume that the caller is going to remove 2351 * at least one item. 2352 */ 2353 if (slab) { 2354 MPASS(slab->us_keg == keg); 2355 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2356 return (slab); 2357 } 2358 /* 2359 * We might not have been able to get a slab but another cpu 2360 * could have while we were unlocked. Check again before we 2361 * fail. 2362 */ 2363 flags |= M_NOVM; 2364 } 2365 return (slab); 2366} 2367 2368static uma_slab_t 2369zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) 2370{ 2371 uma_slab_t slab; 2372 2373 if (keg == NULL) { 2374 keg = zone_first_keg(zone); 2375 KEG_LOCK(keg); 2376 } 2377 2378 for (;;) { 2379 slab = keg_fetch_slab(keg, zone, flags); 2380 if (slab) 2381 return (slab); 2382 if (flags & (M_NOWAIT | M_NOVM)) 2383 break; 2384 } 2385 KEG_UNLOCK(keg); 2386 return (NULL); 2387} 2388 2389/* 2390 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2391 * with the keg locked. On NULL no lock is held. 2392 * 2393 * The last pointer is used to seed the search. It is not required. 2394 */ 2395static uma_slab_t 2396zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) 2397{ 2398 uma_klink_t klink; 2399 uma_slab_t slab; 2400 uma_keg_t keg; 2401 int flags; 2402 int empty; 2403 int full; 2404 2405 /* 2406 * Don't wait on the first pass. This will skip limit tests 2407 * as well. We don't want to block if we can find a provider 2408 * without blocking. 2409 */ 2410 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2411 /* 2412 * Use the last slab allocated as a hint for where to start 2413 * the search. 2414 */ 2415 if (last != NULL) { 2416 slab = keg_fetch_slab(last, zone, flags); 2417 if (slab) 2418 return (slab); 2419 KEG_UNLOCK(last); 2420 } 2421 /* 2422 * Loop until we have a slab incase of transient failures 2423 * while M_WAITOK is specified. I'm not sure this is 100% 2424 * required but we've done it for so long now. 2425 */ 2426 for (;;) { 2427 empty = 0; 2428 full = 0; 2429 /* 2430 * Search the available kegs for slabs. Be careful to hold the 2431 * correct lock while calling into the keg layer. 2432 */ 2433 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2434 keg = klink->kl_keg; 2435 KEG_LOCK(keg); 2436 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2437 slab = keg_fetch_slab(keg, zone, flags); 2438 if (slab) 2439 return (slab); 2440 } 2441 if (keg->uk_flags & UMA_ZFLAG_FULL) 2442 full++; 2443 else 2444 empty++; 2445 KEG_UNLOCK(keg); 2446 } 2447 if (rflags & (M_NOWAIT | M_NOVM)) 2448 break; 2449 flags = rflags; 2450 /* 2451 * All kegs are full. XXX We can't atomically check all kegs 2452 * and sleep so just sleep for a short period and retry. 2453 */ 2454 if (full && !empty) { 2455 ZONE_LOCK(zone); 2456 zone->uz_flags |= UMA_ZFLAG_FULL; 2457 zone->uz_sleeps++; 2458 zone_log_warning(zone); 2459 msleep(zone, zone->uz_lockptr, PVM, 2460 "zonelimit", hz/100); 2461 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2462 ZONE_UNLOCK(zone); 2463 continue; 2464 } 2465 } 2466 return (NULL); 2467} 2468 2469static void * 2470slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 2471{ 2472 void *item; 2473 uint8_t freei; 2474 2475 MPASS(keg == slab->us_keg); 2476 mtx_assert(&keg->uk_lock, MA_OWNED); 2477 2478 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; 2479 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); 2480 item = slab->us_data + (keg->uk_rsize * freei); 2481 slab->us_freecount--; 2482 keg->uk_free--; 2483 2484 /* Move this slab to the full list */ 2485 if (slab->us_freecount == 0) { 2486 LIST_REMOVE(slab, us_link); 2487 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); 2488 } 2489 2490 return (item); 2491} 2492 2493static int 2494zone_import(uma_zone_t zone, void **bucket, int max, int flags) 2495{ 2496 uma_slab_t slab; 2497 uma_keg_t keg; 2498 int i; 2499 2500 slab = NULL; 2501 keg = NULL; 2502 /* Try to keep the buckets totally full */ 2503 for (i = 0; i < max; ) { 2504 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL) 2505 break; 2506 keg = slab->us_keg; 2507 while (slab->us_freecount && i < max) { 2508 bucket[i++] = slab_alloc_item(keg, slab); 2509 if (keg->uk_free <= keg->uk_reserve) 2510 break; 2511 } 2512 /* Don't grab more than one slab at a time. */ 2513 flags &= ~M_WAITOK; 2514 flags |= M_NOWAIT; 2515 } 2516 if (slab != NULL) 2517 KEG_UNLOCK(keg); 2518 2519 return i; 2520} 2521 2522static uma_bucket_t 2523zone_alloc_bucket(uma_zone_t zone, void *udata, int flags) 2524{ 2525 uma_bucket_t bucket; 2526 int max; 2527 2528 /* Don't wait for buckets, preserve caller's NOVM setting. */ 2529 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 2530 if (bucket == NULL) 2531 return (NULL); 2532 2533 max = MIN(bucket->ub_entries, zone->uz_count); 2534 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 2535 max, flags); 2536 2537 /* 2538 * Initialize the memory if necessary. 2539 */ 2540 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 2541 int i; 2542 2543 for (i = 0; i < bucket->ub_cnt; i++) 2544 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2545 flags) != 0) 2546 break; 2547 /* 2548 * If we couldn't initialize the whole bucket, put the 2549 * rest back onto the freelist. 2550 */ 2551 if (i != bucket->ub_cnt) { 2552 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 2553 bucket->ub_cnt - i); 2554#ifdef INVARIANTS 2555 bzero(&bucket->ub_bucket[i], 2556 sizeof(void *) * (bucket->ub_cnt - i)); 2557#endif 2558 bucket->ub_cnt = i; 2559 } 2560 } 2561 2562 if (bucket->ub_cnt == 0) { 2563 bucket_free(zone, bucket, udata); 2564 atomic_add_long(&zone->uz_fails, 1); 2565 return (NULL); 2566 } 2567 2568 return (bucket); 2569} 2570 2571/* 2572 * Allocates a single item from a zone. 2573 * 2574 * Arguments 2575 * zone The zone to alloc for. 2576 * udata The data to be passed to the constructor. 2577 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2578 * 2579 * Returns 2580 * NULL if there is no memory and M_NOWAIT is set 2581 * An item if successful 2582 */ 2583 2584static void * 2585zone_alloc_item(uma_zone_t zone, void *udata, int flags) 2586{ 2587 void *item; 2588 2589 item = NULL; 2590 2591#ifdef UMA_DEBUG_ALLOC 2592 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); 2593#endif 2594 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1) 2595 goto fail; 2596 atomic_add_long(&zone->uz_allocs, 1); 2597 2598 /* 2599 * We have to call both the zone's init (not the keg's init) 2600 * and the zone's ctor. This is because the item is going from 2601 * a keg slab directly to the user, and the user is expecting it 2602 * to be both zone-init'd as well as zone-ctor'd. 2603 */ 2604 if (zone->uz_init != NULL) { 2605 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 2606 zone_free_item(zone, item, udata, SKIP_FINI); 2607 goto fail; 2608 } 2609 } 2610 if (zone->uz_ctor != NULL) { 2611 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2612 zone_free_item(zone, item, udata, SKIP_DTOR); 2613 goto fail; 2614 } 2615 } 2616#ifdef INVARIANTS 2617 uma_dbg_alloc(zone, NULL, item); 2618#endif 2619 if (flags & M_ZERO) 2620 bzero(item, zone->uz_size); 2621 2622 return (item); 2623 2624fail: 2625 atomic_add_long(&zone->uz_fails, 1); 2626 return (NULL); 2627} 2628 2629/* See uma.h */ 2630void 2631uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 2632{ 2633 uma_cache_t cache; 2634 uma_bucket_t bucket; 2635 int lockfail; 2636 int cpu; 2637 2638#ifdef UMA_DEBUG_ALLOC_1 2639 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); 2640#endif 2641 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 2642 zone->uz_name); 2643 2644 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 2645 if (item == NULL) 2646 return; 2647#ifdef DEBUG_MEMGUARD 2648 if (is_memguard_addr(item)) { 2649 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor) 2650 zone->uz_dtor(item, zone->uz_size, udata); 2651 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini) 2652 zone->uz_fini(item, zone->uz_size); 2653 memguard_free(item); 2654 return; 2655 } 2656#endif 2657#ifdef INVARIANTS 2658 if (zone->uz_flags & UMA_ZONE_MALLOC) 2659 uma_dbg_free(zone, udata, item); 2660 else 2661 uma_dbg_free(zone, NULL, item); 2662#endif 2663 if (zone->uz_dtor != NULL) 2664 zone->uz_dtor(item, zone->uz_size, udata); 2665 2666 /* 2667 * The race here is acceptable. If we miss it we'll just have to wait 2668 * a little longer for the limits to be reset. 2669 */ 2670 if (zone->uz_flags & UMA_ZFLAG_FULL) 2671 goto zfree_item; 2672 2673 /* 2674 * If possible, free to the per-CPU cache. There are two 2675 * requirements for safe access to the per-CPU cache: (1) the thread 2676 * accessing the cache must not be preempted or yield during access, 2677 * and (2) the thread must not migrate CPUs without switching which 2678 * cache it accesses. We rely on a critical section to prevent 2679 * preemption and migration. We release the critical section in 2680 * order to acquire the zone mutex if we are unable to free to the 2681 * current cache; when we re-acquire the critical section, we must 2682 * detect and handle migration if it has occurred. 2683 */ 2684zfree_restart: 2685 critical_enter(); 2686 cpu = curcpu; 2687 cache = &zone->uz_cpu[cpu]; 2688 2689zfree_start: 2690 /* 2691 * Try to free into the allocbucket first to give LIFO ordering 2692 * for cache-hot datastructures. Spill over into the freebucket 2693 * if necessary. Alloc will swap them if one runs dry. 2694 */ 2695 bucket = cache->uc_allocbucket; 2696 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 2697 bucket = cache->uc_freebucket; 2698 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2699 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 2700 ("uma_zfree: Freeing to non free bucket index.")); 2701 bucket->ub_bucket[bucket->ub_cnt] = item; 2702 bucket->ub_cnt++; 2703 cache->uc_frees++; 2704 critical_exit(); 2705 return; 2706 } 2707 2708 /* 2709 * We must go back the zone, which requires acquiring the zone lock, 2710 * which in turn means we must release and re-acquire the critical 2711 * section. Since the critical section is released, we may be 2712 * preempted or migrate. As such, make sure not to maintain any 2713 * thread-local state specific to the cache from prior to releasing 2714 * the critical section. 2715 */ 2716 critical_exit(); 2717 if (zone->uz_count == 0 || bucketdisable) 2718 goto zfree_item; 2719 2720 lockfail = 0; 2721 if (ZONE_TRYLOCK(zone) == 0) { 2722 /* Record contention to size the buckets. */ 2723 ZONE_LOCK(zone); 2724 lockfail = 1; 2725 } 2726 critical_enter(); 2727 cpu = curcpu; 2728 cache = &zone->uz_cpu[cpu]; 2729 2730 /* 2731 * Since we have locked the zone we may as well send back our stats. 2732 */ 2733 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2734 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2735 cache->uc_allocs = 0; 2736 cache->uc_frees = 0; 2737 2738 bucket = cache->uc_freebucket; 2739 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2740 ZONE_UNLOCK(zone); 2741 goto zfree_start; 2742 } 2743 cache->uc_freebucket = NULL; 2744 2745 /* Can we throw this on the zone full list? */ 2746 if (bucket != NULL) { 2747#ifdef UMA_DEBUG_ALLOC 2748 printf("uma_zfree: Putting old bucket on the free list.\n"); 2749#endif 2750 /* ub_cnt is pointing to the last free item */ 2751 KASSERT(bucket->ub_cnt != 0, 2752 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 2753 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2754 } 2755 2756 /* We are no longer associated with this CPU. */ 2757 critical_exit(); 2758 2759 /* 2760 * We bump the uz count when the cache size is insufficient to 2761 * handle the working set. 2762 */ 2763 if (lockfail && zone->uz_count < BUCKET_MAX) 2764 zone->uz_count++; 2765 ZONE_UNLOCK(zone); 2766 2767#ifdef UMA_DEBUG_ALLOC 2768 printf("uma_zfree: Allocating new free bucket.\n"); 2769#endif 2770 bucket = bucket_alloc(zone, udata, M_NOWAIT); 2771 if (bucket) { 2772 critical_enter(); 2773 cpu = curcpu; 2774 cache = &zone->uz_cpu[cpu]; 2775 if (cache->uc_freebucket == NULL) { 2776 cache->uc_freebucket = bucket; 2777 goto zfree_start; 2778 } 2779 /* 2780 * We lost the race, start over. We have to drop our 2781 * critical section to free the bucket. 2782 */ 2783 critical_exit(); 2784 bucket_free(zone, bucket, udata); 2785 goto zfree_restart; 2786 } 2787 2788 /* 2789 * If nothing else caught this, we'll just do an internal free. 2790 */ 2791zfree_item: 2792 zone_free_item(zone, item, udata, SKIP_DTOR); 2793 2794 return; 2795} 2796 2797static void 2798slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) 2799{ 2800 uint8_t freei; 2801 2802 mtx_assert(&keg->uk_lock, MA_OWNED); 2803 MPASS(keg == slab->us_keg); 2804 2805 /* Do we need to remove from any lists? */ 2806 if (slab->us_freecount+1 == keg->uk_ipers) { 2807 LIST_REMOVE(slab, us_link); 2808 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 2809 } else if (slab->us_freecount == 0) { 2810 LIST_REMOVE(slab, us_link); 2811 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2812 } 2813 2814 /* Slab management. */ 2815 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 2816 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); 2817 slab->us_freecount++; 2818 2819 /* Keg statistics. */ 2820 keg->uk_free++; 2821} 2822 2823static void 2824zone_release(uma_zone_t zone, void **bucket, int cnt) 2825{ 2826 void *item; 2827 uma_slab_t slab; 2828 uma_keg_t keg; 2829 uint8_t *mem; 2830 int clearfull; 2831 int i; 2832 2833 clearfull = 0; 2834 keg = zone_first_keg(zone); 2835 KEG_LOCK(keg); 2836 for (i = 0; i < cnt; i++) { 2837 item = bucket[i]; 2838 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 2839 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 2840 if (zone->uz_flags & UMA_ZONE_HASH) { 2841 slab = hash_sfind(&keg->uk_hash, mem); 2842 } else { 2843 mem += keg->uk_pgoff; 2844 slab = (uma_slab_t)mem; 2845 } 2846 } else { 2847 slab = vtoslab((vm_offset_t)item); 2848 if (slab->us_keg != keg) { 2849 KEG_UNLOCK(keg); 2850 keg = slab->us_keg; 2851 KEG_LOCK(keg); 2852 } 2853 } 2854 slab_free_item(keg, slab, item); 2855 if (keg->uk_flags & UMA_ZFLAG_FULL) { 2856 if (keg->uk_pages < keg->uk_maxpages) { 2857 keg->uk_flags &= ~UMA_ZFLAG_FULL; 2858 clearfull = 1; 2859 } 2860 2861 /* 2862 * We can handle one more allocation. Since we're 2863 * clearing ZFLAG_FULL, wake up all procs blocked 2864 * on pages. This should be uncommon, so keeping this 2865 * simple for now (rather than adding count of blocked 2866 * threads etc). 2867 */ 2868 wakeup(keg); 2869 } 2870 } 2871 KEG_UNLOCK(keg); 2872 if (clearfull) { 2873 ZONE_LOCK(zone); 2874 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2875 wakeup(zone); 2876 ZONE_UNLOCK(zone); 2877 } 2878 2879} 2880 2881/* 2882 * Frees a single item to any zone. 2883 * 2884 * Arguments: 2885 * zone The zone to free to 2886 * item The item we're freeing 2887 * udata User supplied data for the dtor 2888 * skip Skip dtors and finis 2889 */ 2890static void 2891zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 2892{ 2893 2894#ifdef INVARIANTS 2895 if (skip == SKIP_NONE) { 2896 if (zone->uz_flags & UMA_ZONE_MALLOC) 2897 uma_dbg_free(zone, udata, item); 2898 else 2899 uma_dbg_free(zone, NULL, item); 2900 } 2901#endif 2902 if (skip < SKIP_DTOR && zone->uz_dtor) 2903 zone->uz_dtor(item, zone->uz_size, udata); 2904 2905 if (skip < SKIP_FINI && zone->uz_fini) 2906 zone->uz_fini(item, zone->uz_size); 2907 2908 atomic_add_long(&zone->uz_frees, 1); 2909 zone->uz_release(zone->uz_arg, &item, 1); 2910} 2911 2912/* See uma.h */ 2913int 2914uma_zone_set_max(uma_zone_t zone, int nitems) 2915{ 2916 uma_keg_t keg; 2917 2918 keg = zone_first_keg(zone); 2919 if (keg == NULL) 2920 return (0); 2921 KEG_LOCK(keg); 2922 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 2923 if (keg->uk_maxpages * keg->uk_ipers < nitems) 2924 keg->uk_maxpages += keg->uk_ppera; 2925 nitems = keg->uk_maxpages * keg->uk_ipers; 2926 KEG_UNLOCK(keg); 2927 2928 return (nitems); 2929} 2930 2931/* See uma.h */ 2932int 2933uma_zone_get_max(uma_zone_t zone) 2934{ 2935 int nitems; 2936 uma_keg_t keg; 2937 2938 keg = zone_first_keg(zone); 2939 if (keg == NULL) 2940 return (0); 2941 KEG_LOCK(keg); 2942 nitems = keg->uk_maxpages * keg->uk_ipers; 2943 KEG_UNLOCK(keg); 2944 2945 return (nitems); 2946} 2947 2948/* See uma.h */ 2949void 2950uma_zone_set_warning(uma_zone_t zone, const char *warning) 2951{ 2952 2953 ZONE_LOCK(zone); 2954 zone->uz_warning = warning; 2955 ZONE_UNLOCK(zone); 2956} 2957 2958/* See uma.h */ 2959int 2960uma_zone_get_cur(uma_zone_t zone) 2961{ 2962 int64_t nitems; 2963 u_int i; 2964 2965 ZONE_LOCK(zone); 2966 nitems = zone->uz_allocs - zone->uz_frees; 2967 CPU_FOREACH(i) { 2968 /* 2969 * See the comment in sysctl_vm_zone_stats() regarding the 2970 * safety of accessing the per-cpu caches. With the zone lock 2971 * held, it is safe, but can potentially result in stale data. 2972 */ 2973 nitems += zone->uz_cpu[i].uc_allocs - 2974 zone->uz_cpu[i].uc_frees; 2975 } 2976 ZONE_UNLOCK(zone); 2977 2978 return (nitems < 0 ? 0 : nitems); 2979} 2980 2981/* See uma.h */ 2982void 2983uma_zone_set_init(uma_zone_t zone, uma_init uminit) 2984{ 2985 uma_keg_t keg; 2986 2987 keg = zone_first_keg(zone); 2988 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 2989 KEG_LOCK(keg); 2990 KASSERT(keg->uk_pages == 0, 2991 ("uma_zone_set_init on non-empty keg")); 2992 keg->uk_init = uminit; 2993 KEG_UNLOCK(keg); 2994} 2995 2996/* See uma.h */ 2997void 2998uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 2999{ 3000 uma_keg_t keg; 3001 3002 keg = zone_first_keg(zone); 3003 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 3004 KEG_LOCK(keg); 3005 KASSERT(keg->uk_pages == 0, 3006 ("uma_zone_set_fini on non-empty keg")); 3007 keg->uk_fini = fini; 3008 KEG_UNLOCK(keg); 3009} 3010 3011/* See uma.h */ 3012void 3013uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 3014{ 3015 3016 ZONE_LOCK(zone); 3017 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3018 ("uma_zone_set_zinit on non-empty keg")); 3019 zone->uz_init = zinit; 3020 ZONE_UNLOCK(zone); 3021} 3022 3023/* See uma.h */ 3024void 3025uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 3026{ 3027 3028 ZONE_LOCK(zone); 3029 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3030 ("uma_zone_set_zfini on non-empty keg")); 3031 zone->uz_fini = zfini; 3032 ZONE_UNLOCK(zone); 3033} 3034 3035/* See uma.h */ 3036/* XXX uk_freef is not actually used with the zone locked */ 3037void 3038uma_zone_set_freef(uma_zone_t zone, uma_free freef) 3039{ 3040 uma_keg_t keg; 3041 3042 keg = zone_first_keg(zone); 3043 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 3044 KEG_LOCK(keg); 3045 keg->uk_freef = freef; 3046 KEG_UNLOCK(keg); 3047} 3048 3049/* See uma.h */ 3050/* XXX uk_allocf is not actually used with the zone locked */ 3051void 3052uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3053{ 3054 uma_keg_t keg; 3055 3056 keg = zone_first_keg(zone); 3057 KEG_LOCK(keg); 3058 keg->uk_allocf = allocf; 3059 KEG_UNLOCK(keg); 3060} 3061 3062/* See uma.h */ 3063void 3064uma_zone_reserve(uma_zone_t zone, int items) 3065{ 3066 uma_keg_t keg; 3067 3068 keg = zone_first_keg(zone); 3069 if (keg == NULL) 3070 return; 3071 KEG_LOCK(keg); 3072 keg->uk_reserve = items; 3073 KEG_UNLOCK(keg); 3074 3075 return; 3076} 3077 3078/* See uma.h */ 3079int 3080uma_zone_reserve_kva(uma_zone_t zone, int count) 3081{ 3082 uma_keg_t keg; 3083 vm_offset_t kva; 3084 int pages; 3085 3086 keg = zone_first_keg(zone); 3087 if (keg == NULL) 3088 return (0); 3089 pages = count / keg->uk_ipers; 3090 3091 if (pages * keg->uk_ipers < count) 3092 pages++; 3093 3094#ifdef UMA_MD_SMALL_ALLOC 3095 if (keg->uk_ppera > 1) { 3096#else 3097 if (1) { 3098#endif 3099 kva = kva_alloc(pages * UMA_SLAB_SIZE); 3100 if (kva == 0) 3101 return (0); 3102 } else 3103 kva = 0; 3104 KEG_LOCK(keg); 3105 keg->uk_kva = kva; 3106 keg->uk_offset = 0; 3107 keg->uk_maxpages = pages; 3108#ifdef UMA_MD_SMALL_ALLOC 3109 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3110#else 3111 keg->uk_allocf = noobj_alloc; 3112#endif 3113 keg->uk_flags |= UMA_ZONE_NOFREE; 3114 KEG_UNLOCK(keg); 3115 3116 return (1); 3117} 3118 3119/* See uma.h */ 3120void 3121uma_prealloc(uma_zone_t zone, int items) 3122{ 3123 int slabs; 3124 uma_slab_t slab; 3125 uma_keg_t keg; 3126 3127 keg = zone_first_keg(zone); 3128 if (keg == NULL) 3129 return; 3130 KEG_LOCK(keg); 3131 slabs = items / keg->uk_ipers; 3132 if (slabs * keg->uk_ipers < items) 3133 slabs++; 3134 while (slabs > 0) { 3135 slab = keg_alloc_slab(keg, zone, M_WAITOK); 3136 if (slab == NULL) 3137 break; 3138 MPASS(slab->us_keg == keg); 3139 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 3140 slabs--; 3141 } 3142 KEG_UNLOCK(keg); 3143} 3144 3145/* See uma.h */ 3146uint32_t * 3147uma_find_refcnt(uma_zone_t zone, void *item) 3148{ 3149 uma_slabrefcnt_t slabref; 3150 uma_slab_t slab; 3151 uma_keg_t keg; 3152 uint32_t *refcnt; 3153 int idx; 3154 3155 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK)); 3156 slabref = (uma_slabrefcnt_t)slab; 3157 keg = slab->us_keg; 3158 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT, 3159 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT")); 3160 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3161 refcnt = &slabref->us_refcnt[idx]; 3162 return refcnt; 3163} 3164 3165/* See uma.h */ 3166void 3167uma_reclaim(void) 3168{ 3169#ifdef UMA_DEBUG 3170 printf("UMA: vm asked us to release pages!\n"); 3171#endif 3172 bucket_enable(); 3173 zone_foreach(zone_drain); 3174 if (vm_page_count_min()) { 3175 cache_drain_safe(NULL); 3176 zone_foreach(zone_drain); 3177 } 3178 /* 3179 * Some slabs may have been freed but this zone will be visited early 3180 * we visit again so that we can free pages that are empty once other 3181 * zones are drained. We have to do the same for buckets. 3182 */ 3183 zone_drain(slabzone); 3184 zone_drain(slabrefzone); 3185 bucket_zone_drain(); 3186} 3187 3188/* See uma.h */ 3189int 3190uma_zone_exhausted(uma_zone_t zone) 3191{ 3192 int full; 3193 3194 ZONE_LOCK(zone); 3195 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3196 ZONE_UNLOCK(zone); 3197 return (full); 3198} 3199 3200int 3201uma_zone_exhausted_nolock(uma_zone_t zone) 3202{ 3203 return (zone->uz_flags & UMA_ZFLAG_FULL); 3204} 3205 3206void * 3207uma_large_malloc(int size, int wait) 3208{ 3209 void *mem; 3210 uma_slab_t slab; 3211 uint8_t flags; 3212 3213 slab = zone_alloc_item(slabzone, NULL, wait); 3214 if (slab == NULL) 3215 return (NULL); 3216 mem = page_alloc(NULL, size, &flags, wait); 3217 if (mem) { 3218 vsetslab((vm_offset_t)mem, slab); 3219 slab->us_data = mem; 3220 slab->us_flags = flags | UMA_SLAB_MALLOC; 3221 slab->us_size = size; 3222 } else { 3223 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3224 } 3225 3226 return (mem); 3227} 3228 3229void 3230uma_large_free(uma_slab_t slab) 3231{ 3232 3233 page_free(slab->us_data, slab->us_size, slab->us_flags); 3234 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3235} 3236 3237void 3238uma_print_stats(void) 3239{ 3240 zone_foreach(uma_print_zone); 3241} 3242 3243static void 3244slab_print(uma_slab_t slab) 3245{ 3246 printf("slab: keg %p, data %p, freecount %d\n", 3247 slab->us_keg, slab->us_data, slab->us_freecount); 3248} 3249 3250static void 3251cache_print(uma_cache_t cache) 3252{ 3253 printf("alloc: %p(%d), free: %p(%d)\n", 3254 cache->uc_allocbucket, 3255 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3256 cache->uc_freebucket, 3257 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3258} 3259 3260static void 3261uma_print_keg(uma_keg_t keg) 3262{ 3263 uma_slab_t slab; 3264 3265 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3266 "out %d free %d limit %d\n", 3267 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3268 keg->uk_ipers, keg->uk_ppera, 3269 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free, 3270 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3271 printf("Part slabs:\n"); 3272 LIST_FOREACH(slab, &keg->uk_part_slab, us_link) 3273 slab_print(slab); 3274 printf("Free slabs:\n"); 3275 LIST_FOREACH(slab, &keg->uk_free_slab, us_link) 3276 slab_print(slab); 3277 printf("Full slabs:\n"); 3278 LIST_FOREACH(slab, &keg->uk_full_slab, us_link) 3279 slab_print(slab); 3280} 3281 3282void 3283uma_print_zone(uma_zone_t zone) 3284{ 3285 uma_cache_t cache; 3286 uma_klink_t kl; 3287 int i; 3288 3289 printf("zone: %s(%p) size %d flags %#x\n", 3290 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3291 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3292 uma_print_keg(kl->kl_keg); 3293 CPU_FOREACH(i) { 3294 cache = &zone->uz_cpu[i]; 3295 printf("CPU %d Cache:\n", i); 3296 cache_print(cache); 3297 } 3298} 3299 3300#ifdef DDB 3301/* 3302 * Generate statistics across both the zone and its per-cpu cache's. Return 3303 * desired statistics if the pointer is non-NULL for that statistic. 3304 * 3305 * Note: does not update the zone statistics, as it can't safely clear the 3306 * per-CPU cache statistic. 3307 * 3308 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3309 * safe from off-CPU; we should modify the caches to track this information 3310 * directly so that we don't have to. 3311 */ 3312static void 3313uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3314 uint64_t *freesp, uint64_t *sleepsp) 3315{ 3316 uma_cache_t cache; 3317 uint64_t allocs, frees, sleeps; 3318 int cachefree, cpu; 3319 3320 allocs = frees = sleeps = 0; 3321 cachefree = 0; 3322 CPU_FOREACH(cpu) { 3323 cache = &z->uz_cpu[cpu]; 3324 if (cache->uc_allocbucket != NULL) 3325 cachefree += cache->uc_allocbucket->ub_cnt; 3326 if (cache->uc_freebucket != NULL) 3327 cachefree += cache->uc_freebucket->ub_cnt; 3328 allocs += cache->uc_allocs; 3329 frees += cache->uc_frees; 3330 } 3331 allocs += z->uz_allocs; 3332 frees += z->uz_frees; 3333 sleeps += z->uz_sleeps; 3334 if (cachefreep != NULL) 3335 *cachefreep = cachefree; 3336 if (allocsp != NULL) 3337 *allocsp = allocs; 3338 if (freesp != NULL) 3339 *freesp = frees; 3340 if (sleepsp != NULL) 3341 *sleepsp = sleeps; 3342} 3343#endif /* DDB */ 3344 3345static int 3346sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3347{ 3348 uma_keg_t kz; 3349 uma_zone_t z; 3350 int count; 3351 3352 count = 0; 3353 mtx_lock(&uma_mtx); 3354 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3355 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3356 count++; 3357 } 3358 mtx_unlock(&uma_mtx); 3359 return (sysctl_handle_int(oidp, &count, 0, req)); 3360} 3361 3362static int 3363sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3364{ 3365 struct uma_stream_header ush; 3366 struct uma_type_header uth; 3367 struct uma_percpu_stat ups; 3368 uma_bucket_t bucket; 3369 struct sbuf sbuf; 3370 uma_cache_t cache; 3371 uma_klink_t kl; 3372 uma_keg_t kz; 3373 uma_zone_t z; 3374 uma_keg_t k; 3375 int count, error, i; 3376 3377 error = sysctl_wire_old_buffer(req, 0); 3378 if (error != 0) 3379 return (error); 3380 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3381 3382 count = 0; 3383 mtx_lock(&uma_mtx); 3384 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3385 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3386 count++; 3387 } 3388 3389 /* 3390 * Insert stream header. 3391 */ 3392 bzero(&ush, sizeof(ush)); 3393 ush.ush_version = UMA_STREAM_VERSION; 3394 ush.ush_maxcpus = (mp_maxid + 1); 3395 ush.ush_count = count; 3396 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3397 3398 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3399 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3400 bzero(&uth, sizeof(uth)); 3401 ZONE_LOCK(z); 3402 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3403 uth.uth_align = kz->uk_align; 3404 uth.uth_size = kz->uk_size; 3405 uth.uth_rsize = kz->uk_rsize; 3406 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3407 k = kl->kl_keg; 3408 uth.uth_maxpages += k->uk_maxpages; 3409 uth.uth_pages += k->uk_pages; 3410 uth.uth_keg_free += k->uk_free; 3411 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3412 * k->uk_ipers; 3413 } 3414 3415 /* 3416 * A zone is secondary is it is not the first entry 3417 * on the keg's zone list. 3418 */ 3419 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3420 (LIST_FIRST(&kz->uk_zones) != z)) 3421 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3422 3423 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3424 uth.uth_zone_free += bucket->ub_cnt; 3425 uth.uth_allocs = z->uz_allocs; 3426 uth.uth_frees = z->uz_frees; 3427 uth.uth_fails = z->uz_fails; 3428 uth.uth_sleeps = z->uz_sleeps; 3429 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 3430 /* 3431 * While it is not normally safe to access the cache 3432 * bucket pointers while not on the CPU that owns the 3433 * cache, we only allow the pointers to be exchanged 3434 * without the zone lock held, not invalidated, so 3435 * accept the possible race associated with bucket 3436 * exchange during monitoring. 3437 */ 3438 for (i = 0; i < (mp_maxid + 1); i++) { 3439 bzero(&ups, sizeof(ups)); 3440 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) 3441 goto skip; 3442 if (CPU_ABSENT(i)) 3443 goto skip; 3444 cache = &z->uz_cpu[i]; 3445 if (cache->uc_allocbucket != NULL) 3446 ups.ups_cache_free += 3447 cache->uc_allocbucket->ub_cnt; 3448 if (cache->uc_freebucket != NULL) 3449 ups.ups_cache_free += 3450 cache->uc_freebucket->ub_cnt; 3451 ups.ups_allocs = cache->uc_allocs; 3452 ups.ups_frees = cache->uc_frees; 3453skip: 3454 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups)); 3455 } 3456 ZONE_UNLOCK(z); 3457 } 3458 } 3459 mtx_unlock(&uma_mtx); 3460 error = sbuf_finish(&sbuf); 3461 sbuf_delete(&sbuf); 3462 return (error); 3463} 3464 3465#ifdef DDB 3466DB_SHOW_COMMAND(uma, db_show_uma) 3467{ 3468 uint64_t allocs, frees, sleeps; 3469 uma_bucket_t bucket; 3470 uma_keg_t kz; 3471 uma_zone_t z; 3472 int cachefree; 3473 3474 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used", 3475 "Free", "Requests", "Sleeps", "Bucket"); 3476 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3477 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3478 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 3479 allocs = z->uz_allocs; 3480 frees = z->uz_frees; 3481 sleeps = z->uz_sleeps; 3482 cachefree = 0; 3483 } else 3484 uma_zone_sumstat(z, &cachefree, &allocs, 3485 &frees, &sleeps); 3486 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 3487 (LIST_FIRST(&kz->uk_zones) != z))) 3488 cachefree += kz->uk_free; 3489 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3490 cachefree += bucket->ub_cnt; 3491 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n", 3492 z->uz_name, (uintmax_t)kz->uk_size, 3493 (intmax_t)(allocs - frees), cachefree, 3494 (uintmax_t)allocs, sleeps, z->uz_count); 3495 if (db_pager_quit) 3496 return; 3497 } 3498 } 3499} 3500 3501DB_SHOW_COMMAND(umacache, db_show_umacache) 3502{ 3503 uint64_t allocs, frees; 3504 uma_bucket_t bucket; 3505 uma_zone_t z; 3506 int cachefree; 3507 3508 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 3509 "Requests", "Bucket"); 3510 LIST_FOREACH(z, &uma_cachezones, uz_link) { 3511 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL); 3512 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3513 cachefree += bucket->ub_cnt; 3514 db_printf("%18s %8ju %8jd %8d %12ju %8u\n", 3515 z->uz_name, (uintmax_t)z->uz_size, 3516 (intmax_t)(allocs - frees), cachefree, 3517 (uintmax_t)allocs, z->uz_count); 3518 if (db_pager_quit) 3519 return; 3520 } 3521} 3522#endif 3523