uma_core.c revision 327404
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 * efficient. 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/11/sys/vm/uma_core.c 327404 2017-12-31 03:06:29Z mjg $"); 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/eventhandler.h> 68#include <sys/kernel.h> 69#include <sys/types.h> 70#include <sys/queue.h> 71#include <sys/malloc.h> 72#include <sys/ktr.h> 73#include <sys/lock.h> 74#include <sys/sysctl.h> 75#include <sys/mutex.h> 76#include <sys/proc.h> 77#include <sys/random.h> 78#include <sys/rwlock.h> 79#include <sys/sbuf.h> 80#include <sys/sched.h> 81#include <sys/smp.h> 82#include <sys/taskqueue.h> 83#include <sys/vmmeter.h> 84 85#include <vm/vm.h> 86#include <vm/vm_object.h> 87#include <vm/vm_page.h> 88#include <vm/vm_pageout.h> 89#include <vm/vm_param.h> 90#include <vm/vm_map.h> 91#include <vm/vm_kern.h> 92#include <vm/vm_extern.h> 93#include <vm/uma.h> 94#include <vm/uma_int.h> 95#include <vm/uma_dbg.h> 96 97#include <ddb/ddb.h> 98 99#ifdef DEBUG_MEMGUARD 100#include <vm/memguard.h> 101#endif 102 103/* 104 * This is the zone and keg from which all zones are spawned. The idea is that 105 * even the zone & keg heads are allocated from the allocator, so we use the 106 * bss section to bootstrap us. 107 */ 108static struct uma_keg masterkeg; 109static struct uma_zone masterzone_k; 110static struct uma_zone masterzone_z; 111static uma_zone_t kegs = &masterzone_k; 112static uma_zone_t zones = &masterzone_z; 113 114/* This is the zone from which all of uma_slab_t's are allocated. */ 115static uma_zone_t slabzone; 116 117/* 118 * The initial hash tables come out of this zone so they can be allocated 119 * prior to malloc coming up. 120 */ 121static uma_zone_t hashzone; 122 123/* The boot-time adjusted value for cache line alignment. */ 124int uma_align_cache = 64 - 1; 125 126static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 127 128/* 129 * Are we allowed to allocate buckets? 130 */ 131static int bucketdisable = 1; 132 133/* Linked list of all kegs in the system */ 134static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 135 136/* Linked list of all cache-only zones in the system */ 137static LIST_HEAD(,uma_zone) uma_cachezones = 138 LIST_HEAD_INITIALIZER(uma_cachezones); 139 140/* This RW lock protects the keg list */ 141static struct rwlock_padalign __exclusive_cache_line uma_rwlock; 142 143/* Linked list of boot time pages */ 144static LIST_HEAD(,uma_slab) uma_boot_pages = 145 LIST_HEAD_INITIALIZER(uma_boot_pages); 146 147/* This mutex protects the boot time pages list */ 148static struct mtx_padalign uma_boot_pages_mtx; 149 150static struct sx uma_drain_lock; 151 152/* Is the VM done starting up? */ 153static int booted = 0; 154#define UMA_STARTUP 1 155#define UMA_STARTUP2 2 156 157/* 158 * This is the handle used to schedule events that need to happen 159 * outside of the allocation fast path. 160 */ 161static struct callout uma_callout; 162#define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 163 164/* 165 * This structure is passed as the zone ctor arg so that I don't have to create 166 * a special allocation function just for zones. 167 */ 168struct uma_zctor_args { 169 const char *name; 170 size_t size; 171 uma_ctor ctor; 172 uma_dtor dtor; 173 uma_init uminit; 174 uma_fini fini; 175 uma_import import; 176 uma_release release; 177 void *arg; 178 uma_keg_t keg; 179 int align; 180 uint32_t flags; 181}; 182 183struct uma_kctor_args { 184 uma_zone_t zone; 185 size_t size; 186 uma_init uminit; 187 uma_fini fini; 188 int align; 189 uint32_t flags; 190}; 191 192struct uma_bucket_zone { 193 uma_zone_t ubz_zone; 194 char *ubz_name; 195 int ubz_entries; /* Number of items it can hold. */ 196 int ubz_maxsize; /* Maximum allocation size per-item. */ 197}; 198 199/* 200 * Compute the actual number of bucket entries to pack them in power 201 * of two sizes for more efficient space utilization. 202 */ 203#define BUCKET_SIZE(n) \ 204 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 205 206#define BUCKET_MAX BUCKET_SIZE(256) 207 208struct uma_bucket_zone bucket_zones[] = { 209 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 210 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 211 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 212 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 213 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 214 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 215 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 216 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 217 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 218 { NULL, NULL, 0} 219}; 220 221/* 222 * Flags and enumerations to be passed to internal functions. 223 */ 224enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; 225 226/* Prototypes.. */ 227 228static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 229static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 230static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 231static void page_free(void *, vm_size_t, uint8_t); 232static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); 233static void cache_drain(uma_zone_t); 234static void bucket_drain(uma_zone_t, uma_bucket_t); 235static void bucket_cache_drain(uma_zone_t zone); 236static int keg_ctor(void *, int, void *, int); 237static void keg_dtor(void *, int, void *); 238static int zone_ctor(void *, int, void *, int); 239static void zone_dtor(void *, int, void *); 240static int zero_init(void *, int, int); 241static void keg_small_init(uma_keg_t keg); 242static void keg_large_init(uma_keg_t keg); 243static void zone_foreach(void (*zfunc)(uma_zone_t)); 244static void zone_timeout(uma_zone_t zone); 245static int hash_alloc(struct uma_hash *); 246static int hash_expand(struct uma_hash *, struct uma_hash *); 247static void hash_free(struct uma_hash *hash); 248static void uma_timeout(void *); 249static void uma_startup3(void); 250static void *zone_alloc_item(uma_zone_t, void *, int); 251static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 252static void bucket_enable(void); 253static void bucket_init(void); 254static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 255static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 256static void bucket_zone_drain(void); 257static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags); 258static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); 259static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); 260static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 261static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); 262static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 263 uma_fini fini, int align, uint32_t flags); 264static int zone_import(uma_zone_t zone, void **bucket, int max, int flags); 265static void zone_release(uma_zone_t zone, void **bucket, int cnt); 266static void uma_zero_item(void *item, uma_zone_t zone); 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); 272 273#ifdef INVARIANTS 274static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item); 275static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item); 276#endif 277 278SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 279 280SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 281 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 282 283SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 284 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 285 286static int zone_warnings = 1; 287SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, 288 "Warn when UMA zones becomes full"); 289 290/* 291 * This routine checks to see whether or not it's safe to enable buckets. 292 */ 293static void 294bucket_enable(void) 295{ 296 bucketdisable = vm_page_count_min(); 297} 298 299/* 300 * Initialize bucket_zones, the array of zones of buckets of various sizes. 301 * 302 * For each zone, calculate the memory required for each bucket, consisting 303 * of the header and an array of pointers. 304 */ 305static void 306bucket_init(void) 307{ 308 struct uma_bucket_zone *ubz; 309 int size; 310 311 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 312 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 313 size += sizeof(void *) * ubz->ubz_entries; 314 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 315 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 316 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET); 317 } 318} 319 320/* 321 * Given a desired number of entries for a bucket, return the zone from which 322 * to allocate the bucket. 323 */ 324static struct uma_bucket_zone * 325bucket_zone_lookup(int entries) 326{ 327 struct uma_bucket_zone *ubz; 328 329 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 330 if (ubz->ubz_entries >= entries) 331 return (ubz); 332 ubz--; 333 return (ubz); 334} 335 336static int 337bucket_select(int size) 338{ 339 struct uma_bucket_zone *ubz; 340 341 ubz = &bucket_zones[0]; 342 if (size > ubz->ubz_maxsize) 343 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 344 345 for (; ubz->ubz_entries != 0; ubz++) 346 if (ubz->ubz_maxsize < size) 347 break; 348 ubz--; 349 return (ubz->ubz_entries); 350} 351 352static uma_bucket_t 353bucket_alloc(uma_zone_t zone, void *udata, int flags) 354{ 355 struct uma_bucket_zone *ubz; 356 uma_bucket_t bucket; 357 358 /* 359 * This is to stop us from allocating per cpu buckets while we're 360 * running out of vm.boot_pages. Otherwise, we would exhaust the 361 * boot pages. This also prevents us from allocating buckets in 362 * low memory situations. 363 */ 364 if (bucketdisable) 365 return (NULL); 366 /* 367 * To limit bucket recursion we store the original zone flags 368 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 369 * NOVM flag to persist even through deep recursions. We also 370 * store ZFLAG_BUCKET once we have recursed attempting to allocate 371 * a bucket for a bucket zone so we do not allow infinite bucket 372 * recursion. This cookie will even persist to frees of unused 373 * buckets via the allocation path or bucket allocations in the 374 * free path. 375 */ 376 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 377 udata = (void *)(uintptr_t)zone->uz_flags; 378 else { 379 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 380 return (NULL); 381 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 382 } 383 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 384 flags |= M_NOVM; 385 ubz = bucket_zone_lookup(zone->uz_count); 386 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 387 ubz++; 388 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 389 if (bucket) { 390#ifdef INVARIANTS 391 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 392#endif 393 bucket->ub_cnt = 0; 394 bucket->ub_entries = ubz->ubz_entries; 395 } 396 397 return (bucket); 398} 399 400static void 401bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 402{ 403 struct uma_bucket_zone *ubz; 404 405 KASSERT(bucket->ub_cnt == 0, 406 ("bucket_free: Freeing a non free bucket.")); 407 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 408 udata = (void *)(uintptr_t)zone->uz_flags; 409 ubz = bucket_zone_lookup(bucket->ub_entries); 410 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 411} 412 413static void 414bucket_zone_drain(void) 415{ 416 struct uma_bucket_zone *ubz; 417 418 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 419 zone_drain(ubz->ubz_zone); 420} 421 422static void 423zone_log_warning(uma_zone_t zone) 424{ 425 static const struct timeval warninterval = { 300, 0 }; 426 427 if (!zone_warnings || zone->uz_warning == NULL) 428 return; 429 430 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 431 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 432} 433 434static inline void 435zone_maxaction(uma_zone_t zone) 436{ 437 438 if (zone->uz_maxaction.ta_func != NULL) 439 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); 440} 441 442static void 443zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 444{ 445 uma_klink_t klink; 446 447 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 448 kegfn(klink->kl_keg); 449} 450 451/* 452 * Routine called by timeout which is used to fire off some time interval 453 * based calculations. (stats, hash size, etc.) 454 * 455 * Arguments: 456 * arg Unused 457 * 458 * Returns: 459 * Nothing 460 */ 461static void 462uma_timeout(void *unused) 463{ 464 bucket_enable(); 465 zone_foreach(zone_timeout); 466 467 /* Reschedule this event */ 468 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 469} 470 471/* 472 * Routine to perform timeout driven calculations. This expands the 473 * hashes and does per cpu statistics aggregation. 474 * 475 * Returns nothing. 476 */ 477static void 478keg_timeout(uma_keg_t keg) 479{ 480 481 KEG_LOCK(keg); 482 /* 483 * Expand the keg hash table. 484 * 485 * This is done if the number of slabs is larger than the hash size. 486 * What I'm trying to do here is completely reduce collisions. This 487 * may be a little aggressive. Should I allow for two collisions max? 488 */ 489 if (keg->uk_flags & UMA_ZONE_HASH && 490 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 491 struct uma_hash newhash; 492 struct uma_hash oldhash; 493 int ret; 494 495 /* 496 * This is so involved because allocating and freeing 497 * while the keg lock is held will lead to deadlock. 498 * I have to do everything in stages and check for 499 * races. 500 */ 501 newhash = keg->uk_hash; 502 KEG_UNLOCK(keg); 503 ret = hash_alloc(&newhash); 504 KEG_LOCK(keg); 505 if (ret) { 506 if (hash_expand(&keg->uk_hash, &newhash)) { 507 oldhash = keg->uk_hash; 508 keg->uk_hash = newhash; 509 } else 510 oldhash = newhash; 511 512 KEG_UNLOCK(keg); 513 hash_free(&oldhash); 514 return; 515 } 516 } 517 KEG_UNLOCK(keg); 518} 519 520static void 521zone_timeout(uma_zone_t zone) 522{ 523 524 zone_foreach_keg(zone, &keg_timeout); 525} 526 527/* 528 * Allocate and zero fill the next sized hash table from the appropriate 529 * backing store. 530 * 531 * Arguments: 532 * hash A new hash structure with the old hash size in uh_hashsize 533 * 534 * Returns: 535 * 1 on success and 0 on failure. 536 */ 537static int 538hash_alloc(struct uma_hash *hash) 539{ 540 int oldsize; 541 int alloc; 542 543 oldsize = hash->uh_hashsize; 544 545 /* We're just going to go to a power of two greater */ 546 if (oldsize) { 547 hash->uh_hashsize = oldsize * 2; 548 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 549 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 550 M_UMAHASH, M_NOWAIT); 551 } else { 552 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 553 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 554 M_WAITOK); 555 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 556 } 557 if (hash->uh_slab_hash) { 558 bzero(hash->uh_slab_hash, alloc); 559 hash->uh_hashmask = hash->uh_hashsize - 1; 560 return (1); 561 } 562 563 return (0); 564} 565 566/* 567 * Expands the hash table for HASH zones. This is done from zone_timeout 568 * to reduce collisions. This must not be done in the regular allocation 569 * path, otherwise, we can recurse on the vm while allocating pages. 570 * 571 * Arguments: 572 * oldhash The hash you want to expand 573 * newhash The hash structure for the new table 574 * 575 * Returns: 576 * Nothing 577 * 578 * Discussion: 579 */ 580static int 581hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 582{ 583 uma_slab_t slab; 584 int hval; 585 int i; 586 587 if (!newhash->uh_slab_hash) 588 return (0); 589 590 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 591 return (0); 592 593 /* 594 * I need to investigate hash algorithms for resizing without a 595 * full rehash. 596 */ 597 598 for (i = 0; i < oldhash->uh_hashsize; i++) 599 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 600 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 601 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 602 hval = UMA_HASH(newhash, slab->us_data); 603 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 604 slab, us_hlink); 605 } 606 607 return (1); 608} 609 610/* 611 * Free the hash bucket to the appropriate backing store. 612 * 613 * Arguments: 614 * slab_hash The hash bucket we're freeing 615 * hashsize The number of entries in that hash bucket 616 * 617 * Returns: 618 * Nothing 619 */ 620static void 621hash_free(struct uma_hash *hash) 622{ 623 if (hash->uh_slab_hash == NULL) 624 return; 625 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 626 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 627 else 628 free(hash->uh_slab_hash, M_UMAHASH); 629} 630 631/* 632 * Frees all outstanding items in a bucket 633 * 634 * Arguments: 635 * zone The zone to free to, must be unlocked. 636 * bucket The free/alloc bucket with items, cpu queue must be locked. 637 * 638 * Returns: 639 * Nothing 640 */ 641 642static void 643bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 644{ 645 int i; 646 647 if (bucket == NULL) 648 return; 649 650 if (zone->uz_fini) 651 for (i = 0; i < bucket->ub_cnt; i++) 652 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 653 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 654 bucket->ub_cnt = 0; 655} 656 657/* 658 * Drains the per cpu caches for a zone. 659 * 660 * NOTE: This may only be called while the zone is being turn down, and not 661 * during normal operation. This is necessary in order that we do not have 662 * to migrate CPUs to drain the per-CPU caches. 663 * 664 * Arguments: 665 * zone The zone to drain, must be unlocked. 666 * 667 * Returns: 668 * Nothing 669 */ 670static void 671cache_drain(uma_zone_t zone) 672{ 673 uma_cache_t cache; 674 int cpu; 675 676 /* 677 * XXX: It is safe to not lock the per-CPU caches, because we're 678 * tearing down the zone anyway. I.e., there will be no further use 679 * of the caches at this point. 680 * 681 * XXX: It would good to be able to assert that the zone is being 682 * torn down to prevent improper use of cache_drain(). 683 * 684 * XXX: We lock the zone before passing into bucket_cache_drain() as 685 * it is used elsewhere. Should the tear-down path be made special 686 * there in some form? 687 */ 688 CPU_FOREACH(cpu) { 689 cache = &zone->uz_cpu[cpu]; 690 bucket_drain(zone, cache->uc_allocbucket); 691 bucket_drain(zone, cache->uc_freebucket); 692 if (cache->uc_allocbucket != NULL) 693 bucket_free(zone, cache->uc_allocbucket, NULL); 694 if (cache->uc_freebucket != NULL) 695 bucket_free(zone, cache->uc_freebucket, NULL); 696 cache->uc_allocbucket = cache->uc_freebucket = NULL; 697 } 698 ZONE_LOCK(zone); 699 bucket_cache_drain(zone); 700 ZONE_UNLOCK(zone); 701} 702 703static void 704cache_shrink(uma_zone_t zone) 705{ 706 707 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 708 return; 709 710 ZONE_LOCK(zone); 711 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; 712 ZONE_UNLOCK(zone); 713} 714 715static void 716cache_drain_safe_cpu(uma_zone_t zone) 717{ 718 uma_cache_t cache; 719 uma_bucket_t b1, b2; 720 721 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 722 return; 723 724 b1 = b2 = NULL; 725 ZONE_LOCK(zone); 726 critical_enter(); 727 cache = &zone->uz_cpu[curcpu]; 728 if (cache->uc_allocbucket) { 729 if (cache->uc_allocbucket->ub_cnt != 0) 730 LIST_INSERT_HEAD(&zone->uz_buckets, 731 cache->uc_allocbucket, ub_link); 732 else 733 b1 = cache->uc_allocbucket; 734 cache->uc_allocbucket = NULL; 735 } 736 if (cache->uc_freebucket) { 737 if (cache->uc_freebucket->ub_cnt != 0) 738 LIST_INSERT_HEAD(&zone->uz_buckets, 739 cache->uc_freebucket, ub_link); 740 else 741 b2 = cache->uc_freebucket; 742 cache->uc_freebucket = NULL; 743 } 744 critical_exit(); 745 ZONE_UNLOCK(zone); 746 if (b1) 747 bucket_free(zone, b1, NULL); 748 if (b2) 749 bucket_free(zone, b2, NULL); 750} 751 752/* 753 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 754 * This is an expensive call because it needs to bind to all CPUs 755 * one by one and enter a critical section on each of them in order 756 * to safely access their cache buckets. 757 * Zone lock must not be held on call this function. 758 */ 759static void 760cache_drain_safe(uma_zone_t zone) 761{ 762 int cpu; 763 764 /* 765 * Polite bucket sizes shrinking was not enouth, shrink aggressively. 766 */ 767 if (zone) 768 cache_shrink(zone); 769 else 770 zone_foreach(cache_shrink); 771 772 CPU_FOREACH(cpu) { 773 thread_lock(curthread); 774 sched_bind(curthread, cpu); 775 thread_unlock(curthread); 776 777 if (zone) 778 cache_drain_safe_cpu(zone); 779 else 780 zone_foreach(cache_drain_safe_cpu); 781 } 782 thread_lock(curthread); 783 sched_unbind(curthread); 784 thread_unlock(curthread); 785} 786 787/* 788 * Drain the cached buckets from a zone. Expects a locked zone on entry. 789 */ 790static void 791bucket_cache_drain(uma_zone_t zone) 792{ 793 uma_bucket_t bucket; 794 795 /* 796 * Drain the bucket queues and free the buckets, we just keep two per 797 * cpu (alloc/free). 798 */ 799 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 800 LIST_REMOVE(bucket, ub_link); 801 ZONE_UNLOCK(zone); 802 bucket_drain(zone, bucket); 803 bucket_free(zone, bucket, NULL); 804 ZONE_LOCK(zone); 805 } 806 807 /* 808 * Shrink further bucket sizes. Price of single zone lock collision 809 * is probably lower then price of global cache drain. 810 */ 811 if (zone->uz_count > zone->uz_count_min) 812 zone->uz_count--; 813} 814 815static void 816keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 817{ 818 uint8_t *mem; 819 int i; 820 uint8_t flags; 821 822 mem = slab->us_data; 823 flags = slab->us_flags; 824 i = start; 825 if (keg->uk_fini != NULL) { 826 for (i--; i > -1; i--) 827 keg->uk_fini(slab->us_data + (keg->uk_rsize * i), 828 keg->uk_size); 829 } 830 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 831 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 832#ifdef UMA_DEBUG 833 printf("%s: Returning %d bytes.\n", keg->uk_name, 834 PAGE_SIZE * keg->uk_ppera); 835#endif 836 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 837} 838 839/* 840 * Frees pages from a keg back to the system. This is done on demand from 841 * the pageout daemon. 842 * 843 * Returns nothing. 844 */ 845static void 846keg_drain(uma_keg_t keg) 847{ 848 struct slabhead freeslabs = { 0 }; 849 uma_slab_t slab, tmp; 850 851 /* 852 * We don't want to take pages from statically allocated kegs at this 853 * time 854 */ 855 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 856 return; 857 858#ifdef UMA_DEBUG 859 printf("%s free items: %u\n", keg->uk_name, keg->uk_free); 860#endif 861 KEG_LOCK(keg); 862 if (keg->uk_free == 0) 863 goto finished; 864 865 LIST_FOREACH_SAFE(slab, &keg->uk_free_slab, us_link, tmp) { 866 /* We have nowhere to free these to. */ 867 if (slab->us_flags & UMA_SLAB_BOOT) 868 continue; 869 870 LIST_REMOVE(slab, us_link); 871 keg->uk_pages -= keg->uk_ppera; 872 keg->uk_free -= keg->uk_ipers; 873 874 if (keg->uk_flags & UMA_ZONE_HASH) 875 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); 876 877 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 878 } 879finished: 880 KEG_UNLOCK(keg); 881 882 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 883 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 884 keg_free_slab(keg, slab, keg->uk_ipers); 885 } 886} 887 888static void 889zone_drain_wait(uma_zone_t zone, int waitok) 890{ 891 892 /* 893 * Set draining to interlock with zone_dtor() so we can release our 894 * locks as we go. Only dtor() should do a WAITOK call since it 895 * is the only call that knows the structure will still be available 896 * when it wakes up. 897 */ 898 ZONE_LOCK(zone); 899 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 900 if (waitok == M_NOWAIT) 901 goto out; 902 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 903 } 904 zone->uz_flags |= UMA_ZFLAG_DRAINING; 905 bucket_cache_drain(zone); 906 ZONE_UNLOCK(zone); 907 /* 908 * The DRAINING flag protects us from being freed while 909 * we're running. Normally the uma_rwlock would protect us but we 910 * must be able to release and acquire the right lock for each keg. 911 */ 912 zone_foreach_keg(zone, &keg_drain); 913 ZONE_LOCK(zone); 914 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 915 wakeup(zone); 916out: 917 ZONE_UNLOCK(zone); 918} 919 920void 921zone_drain(uma_zone_t zone) 922{ 923 924 zone_drain_wait(zone, M_NOWAIT); 925} 926 927/* 928 * Allocate a new slab for a keg. This does not insert the slab onto a list. 929 * 930 * Arguments: 931 * wait Shall we wait? 932 * 933 * Returns: 934 * The slab that was allocated or NULL if there is no memory and the 935 * caller specified M_NOWAIT. 936 */ 937static uma_slab_t 938keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) 939{ 940 uma_alloc allocf; 941 uma_slab_t slab; 942 uint8_t *mem; 943 uint8_t flags; 944 int i; 945 946 mtx_assert(&keg->uk_lock, MA_OWNED); 947 slab = NULL; 948 mem = NULL; 949 950#ifdef UMA_DEBUG 951 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name); 952#endif 953 allocf = keg->uk_allocf; 954 KEG_UNLOCK(keg); 955 956 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 957 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); 958 if (slab == NULL) 959 goto out; 960 } 961 962 /* 963 * This reproduces the old vm_zone behavior of zero filling pages the 964 * first time they are added to a zone. 965 * 966 * Malloced items are zeroed in uma_zalloc. 967 */ 968 969 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 970 wait |= M_ZERO; 971 else 972 wait &= ~M_ZERO; 973 974 if (keg->uk_flags & UMA_ZONE_NODUMP) 975 wait |= M_NODUMP; 976 977 /* zone is passed for legacy reasons. */ 978 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait); 979 if (mem == NULL) { 980 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 981 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 982 slab = NULL; 983 goto out; 984 } 985 986 /* Point the slab into the allocated memory */ 987 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 988 slab = (uma_slab_t )(mem + keg->uk_pgoff); 989 990 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 991 for (i = 0; i < keg->uk_ppera; i++) 992 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 993 994 slab->us_keg = keg; 995 slab->us_data = mem; 996 slab->us_freecount = keg->uk_ipers; 997 slab->us_flags = flags; 998 BIT_FILL(SLAB_SETSIZE, &slab->us_free); 999#ifdef INVARIANTS 1000 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); 1001#endif 1002 1003 if (keg->uk_init != NULL) { 1004 for (i = 0; i < keg->uk_ipers; i++) 1005 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 1006 keg->uk_size, wait) != 0) 1007 break; 1008 if (i != keg->uk_ipers) { 1009 keg_free_slab(keg, slab, i); 1010 slab = NULL; 1011 goto out; 1012 } 1013 } 1014out: 1015 KEG_LOCK(keg); 1016 1017 if (slab != NULL) { 1018 if (keg->uk_flags & UMA_ZONE_HASH) 1019 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1020 1021 keg->uk_pages += keg->uk_ppera; 1022 keg->uk_free += keg->uk_ipers; 1023 } 1024 1025 return (slab); 1026} 1027 1028/* 1029 * This function is intended to be used early on in place of page_alloc() so 1030 * that we may use the boot time page cache to satisfy allocations before 1031 * the VM is ready. 1032 */ 1033static void * 1034startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) 1035{ 1036 uma_keg_t keg; 1037 uma_slab_t tmps; 1038 int pages, check_pages; 1039 1040 keg = zone_first_keg(zone); 1041 pages = howmany(bytes, PAGE_SIZE); 1042 check_pages = pages - 1; 1043 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n")); 1044 1045 /* 1046 * Check our small startup cache to see if it has pages remaining. 1047 */ 1048 mtx_lock(&uma_boot_pages_mtx); 1049 1050 /* First check if we have enough room. */ 1051 tmps = LIST_FIRST(&uma_boot_pages); 1052 while (tmps != NULL && check_pages-- > 0) 1053 tmps = LIST_NEXT(tmps, us_link); 1054 if (tmps != NULL) { 1055 /* 1056 * It's ok to lose tmps references. The last one will 1057 * have tmps->us_data pointing to the start address of 1058 * "pages" contiguous pages of memory. 1059 */ 1060 while (pages-- > 0) { 1061 tmps = LIST_FIRST(&uma_boot_pages); 1062 LIST_REMOVE(tmps, us_link); 1063 } 1064 mtx_unlock(&uma_boot_pages_mtx); 1065 *pflag = tmps->us_flags; 1066 return (tmps->us_data); 1067 } 1068 mtx_unlock(&uma_boot_pages_mtx); 1069 if (booted < UMA_STARTUP2) 1070 panic("UMA: Increase vm.boot_pages"); 1071 /* 1072 * Now that we've booted reset these users to their real allocator. 1073 */ 1074#ifdef UMA_MD_SMALL_ALLOC 1075 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; 1076#else 1077 keg->uk_allocf = page_alloc; 1078#endif 1079 return keg->uk_allocf(zone, bytes, pflag, wait); 1080} 1081 1082/* 1083 * Allocates a number of pages from the system 1084 * 1085 * Arguments: 1086 * bytes The number of bytes requested 1087 * wait Shall we wait? 1088 * 1089 * Returns: 1090 * A pointer to the alloced memory or possibly 1091 * NULL if M_NOWAIT is set. 1092 */ 1093static void * 1094page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) 1095{ 1096 void *p; /* Returned page */ 1097 1098 *pflag = UMA_SLAB_KMEM; 1099 p = (void *) kmem_malloc(kmem_arena, bytes, wait); 1100 1101 return (p); 1102} 1103 1104/* 1105 * Allocates a number of pages from within an object 1106 * 1107 * Arguments: 1108 * bytes The number of bytes requested 1109 * wait Shall we wait? 1110 * 1111 * Returns: 1112 * A pointer to the alloced memory or possibly 1113 * NULL if M_NOWAIT is set. 1114 */ 1115static void * 1116noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait) 1117{ 1118 TAILQ_HEAD(, vm_page) alloctail; 1119 u_long npages; 1120 vm_offset_t retkva, zkva; 1121 vm_page_t p, p_next; 1122 uma_keg_t keg; 1123 1124 TAILQ_INIT(&alloctail); 1125 keg = zone_first_keg(zone); 1126 1127 npages = howmany(bytes, PAGE_SIZE); 1128 while (npages > 0) { 1129 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT | 1130 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ); 1131 if (p != NULL) { 1132 /* 1133 * Since the page does not belong to an object, its 1134 * listq is unused. 1135 */ 1136 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1137 npages--; 1138 continue; 1139 } 1140 if (wait & M_WAITOK) { 1141 VM_WAIT; 1142 continue; 1143 } 1144 1145 /* 1146 * Page allocation failed, free intermediate pages and 1147 * exit. 1148 */ 1149 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1150 vm_page_unwire(p, PQ_NONE); 1151 vm_page_free(p); 1152 } 1153 return (NULL); 1154 } 1155 *flags = UMA_SLAB_PRIV; 1156 zkva = keg->uk_kva + 1157 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1158 retkva = zkva; 1159 TAILQ_FOREACH(p, &alloctail, listq) { 1160 pmap_qenter(zkva, &p, 1); 1161 zkva += PAGE_SIZE; 1162 } 1163 1164 return ((void *)retkva); 1165} 1166 1167/* 1168 * Frees a number of pages to the system 1169 * 1170 * Arguments: 1171 * mem A pointer to the memory to be freed 1172 * size The size of the memory being freed 1173 * flags The original p->us_flags field 1174 * 1175 * Returns: 1176 * Nothing 1177 */ 1178static void 1179page_free(void *mem, vm_size_t size, uint8_t flags) 1180{ 1181 struct vmem *vmem; 1182 1183 if (flags & UMA_SLAB_KMEM) 1184 vmem = kmem_arena; 1185 else if (flags & UMA_SLAB_KERNEL) 1186 vmem = kernel_arena; 1187 else 1188 panic("UMA: page_free used with invalid flags %d", flags); 1189 1190 kmem_free(vmem, (vm_offset_t)mem, size); 1191} 1192 1193/* 1194 * Zero fill initializer 1195 * 1196 * Arguments/Returns follow uma_init specifications 1197 */ 1198static int 1199zero_init(void *mem, int size, int flags) 1200{ 1201 bzero(mem, size); 1202 return (0); 1203} 1204 1205/* 1206 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1207 * 1208 * Arguments 1209 * keg The zone we should initialize 1210 * 1211 * Returns 1212 * Nothing 1213 */ 1214static void 1215keg_small_init(uma_keg_t keg) 1216{ 1217 u_int rsize; 1218 u_int memused; 1219 u_int wastedspace; 1220 u_int shsize; 1221 u_int slabsize; 1222 1223 if (keg->uk_flags & UMA_ZONE_PCPU) { 1224 u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU; 1225 1226 slabsize = sizeof(struct pcpu); 1227 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu), 1228 PAGE_SIZE); 1229 } else { 1230 slabsize = UMA_SLAB_SIZE; 1231 keg->uk_ppera = 1; 1232 } 1233 1234 /* 1235 * Calculate the size of each allocation (rsize) according to 1236 * alignment. If the requested size is smaller than we have 1237 * allocation bits for we round it up. 1238 */ 1239 rsize = keg->uk_size; 1240 if (rsize < slabsize / SLAB_SETSIZE) 1241 rsize = slabsize / SLAB_SETSIZE; 1242 if (rsize & keg->uk_align) 1243 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1244 keg->uk_rsize = rsize; 1245 1246 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1247 keg->uk_rsize < sizeof(struct pcpu), 1248 ("%s: size %u too large", __func__, keg->uk_rsize)); 1249 1250 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1251 shsize = 0; 1252 else 1253 shsize = sizeof(struct uma_slab); 1254 1255 keg->uk_ipers = (slabsize - shsize) / rsize; 1256 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1257 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1258 1259 memused = keg->uk_ipers * rsize + shsize; 1260 wastedspace = slabsize - memused; 1261 1262 /* 1263 * We can't do OFFPAGE if we're internal or if we've been 1264 * asked to not go to the VM for buckets. If we do this we 1265 * may end up going to the VM for slabs which we do not 1266 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1267 * of UMA_ZONE_VM, which clearly forbids it. 1268 */ 1269 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1270 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1271 return; 1272 1273 /* 1274 * See if using an OFFPAGE slab will limit our waste. Only do 1275 * this if it permits more items per-slab. 1276 * 1277 * XXX We could try growing slabsize to limit max waste as well. 1278 * Historically this was not done because the VM could not 1279 * efficiently handle contiguous allocations. 1280 */ 1281 if ((wastedspace >= slabsize / UMA_MAX_WASTE) && 1282 (keg->uk_ipers < (slabsize / keg->uk_rsize))) { 1283 keg->uk_ipers = slabsize / keg->uk_rsize; 1284 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1285 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1286#ifdef UMA_DEBUG 1287 printf("UMA decided we need offpage slab headers for " 1288 "keg: %s, calculated wastedspace = %d, " 1289 "maximum wasted space allowed = %d, " 1290 "calculated ipers = %d, " 1291 "new wasted space = %d\n", keg->uk_name, wastedspace, 1292 slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1293 slabsize - keg->uk_ipers * keg->uk_rsize); 1294#endif 1295 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1296 } 1297 1298 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1299 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1300 keg->uk_flags |= UMA_ZONE_HASH; 1301} 1302 1303/* 1304 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1305 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1306 * more complicated. 1307 * 1308 * Arguments 1309 * keg The keg we should initialize 1310 * 1311 * Returns 1312 * Nothing 1313 */ 1314static void 1315keg_large_init(uma_keg_t keg) 1316{ 1317 u_int shsize; 1318 1319 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1320 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1321 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1322 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1323 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1324 1325 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1326 keg->uk_ipers = 1; 1327 keg->uk_rsize = keg->uk_size; 1328 1329 /* Check whether we have enough space to not do OFFPAGE. */ 1330 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) { 1331 shsize = sizeof(struct uma_slab); 1332 if (shsize & UMA_ALIGN_PTR) 1333 shsize = (shsize & ~UMA_ALIGN_PTR) + 1334 (UMA_ALIGN_PTR + 1); 1335 1336 if (PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < shsize) { 1337 /* 1338 * We can't do OFFPAGE if we're internal, in which case 1339 * we need an extra page per allocation to contain the 1340 * slab header. 1341 */ 1342 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0) 1343 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1344 else 1345 keg->uk_ppera++; 1346 } 1347 } 1348 1349 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1350 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1351 keg->uk_flags |= UMA_ZONE_HASH; 1352} 1353 1354static void 1355keg_cachespread_init(uma_keg_t keg) 1356{ 1357 int alignsize; 1358 int trailer; 1359 int pages; 1360 int rsize; 1361 1362 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1363 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1364 1365 alignsize = keg->uk_align + 1; 1366 rsize = keg->uk_size; 1367 /* 1368 * We want one item to start on every align boundary in a page. To 1369 * do this we will span pages. We will also extend the item by the 1370 * size of align if it is an even multiple of align. Otherwise, it 1371 * would fall on the same boundary every time. 1372 */ 1373 if (rsize & keg->uk_align) 1374 rsize = (rsize & ~keg->uk_align) + alignsize; 1375 if ((rsize & alignsize) == 0) 1376 rsize += alignsize; 1377 trailer = rsize - keg->uk_size; 1378 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1379 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1380 keg->uk_rsize = rsize; 1381 keg->uk_ppera = pages; 1382 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1383 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1384 KASSERT(keg->uk_ipers <= SLAB_SETSIZE, 1385 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1386 keg->uk_ipers)); 1387} 1388 1389/* 1390 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1391 * the keg onto the global keg list. 1392 * 1393 * Arguments/Returns follow uma_ctor specifications 1394 * udata Actually uma_kctor_args 1395 */ 1396static int 1397keg_ctor(void *mem, int size, void *udata, int flags) 1398{ 1399 struct uma_kctor_args *arg = udata; 1400 uma_keg_t keg = mem; 1401 uma_zone_t zone; 1402 1403 bzero(keg, size); 1404 keg->uk_size = arg->size; 1405 keg->uk_init = arg->uminit; 1406 keg->uk_fini = arg->fini; 1407 keg->uk_align = arg->align; 1408 keg->uk_free = 0; 1409 keg->uk_reserve = 0; 1410 keg->uk_pages = 0; 1411 keg->uk_flags = arg->flags; 1412 keg->uk_allocf = page_alloc; 1413 keg->uk_freef = page_free; 1414 keg->uk_slabzone = NULL; 1415 1416 /* 1417 * The master zone is passed to us at keg-creation time. 1418 */ 1419 zone = arg->zone; 1420 keg->uk_name = zone->uz_name; 1421 1422 if (arg->flags & UMA_ZONE_VM) 1423 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1424 1425 if (arg->flags & UMA_ZONE_ZINIT) 1426 keg->uk_init = zero_init; 1427 1428 if (arg->flags & UMA_ZONE_MALLOC) 1429 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1430 1431 if (arg->flags & UMA_ZONE_PCPU) 1432#ifdef SMP 1433 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1434#else 1435 keg->uk_flags &= ~UMA_ZONE_PCPU; 1436#endif 1437 1438 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1439 keg_cachespread_init(keg); 1440 } else { 1441 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab))) 1442 keg_large_init(keg); 1443 else 1444 keg_small_init(keg); 1445 } 1446 1447 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1448 keg->uk_slabzone = slabzone; 1449 1450 /* 1451 * If we haven't booted yet we need allocations to go through the 1452 * startup cache until the vm is ready. 1453 */ 1454 if (keg->uk_ppera == 1) { 1455#ifdef UMA_MD_SMALL_ALLOC 1456 keg->uk_allocf = uma_small_alloc; 1457 keg->uk_freef = uma_small_free; 1458 1459 if (booted < UMA_STARTUP) 1460 keg->uk_allocf = startup_alloc; 1461#else 1462 if (booted < UMA_STARTUP2) 1463 keg->uk_allocf = startup_alloc; 1464#endif 1465 } else if (booted < UMA_STARTUP2 && 1466 (keg->uk_flags & UMA_ZFLAG_INTERNAL)) 1467 keg->uk_allocf = startup_alloc; 1468 1469 /* 1470 * Initialize keg's lock 1471 */ 1472 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1473 1474 /* 1475 * If we're putting the slab header in the actual page we need to 1476 * figure out where in each page it goes. This calculates a right 1477 * justified offset into the memory on an ALIGN_PTR boundary. 1478 */ 1479 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1480 u_int totsize; 1481 1482 /* Size of the slab struct and free list */ 1483 totsize = sizeof(struct uma_slab); 1484 1485 if (totsize & UMA_ALIGN_PTR) 1486 totsize = (totsize & ~UMA_ALIGN_PTR) + 1487 (UMA_ALIGN_PTR + 1); 1488 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1489 1490 /* 1491 * The only way the following is possible is if with our 1492 * UMA_ALIGN_PTR adjustments we are now bigger than 1493 * UMA_SLAB_SIZE. I haven't checked whether this is 1494 * mathematically possible for all cases, so we make 1495 * sure here anyway. 1496 */ 1497 totsize = keg->uk_pgoff + sizeof(struct uma_slab); 1498 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1499 printf("zone %s ipers %d rsize %d size %d\n", 1500 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1501 keg->uk_size); 1502 panic("UMA slab won't fit."); 1503 } 1504 } 1505 1506 if (keg->uk_flags & UMA_ZONE_HASH) 1507 hash_alloc(&keg->uk_hash); 1508 1509#ifdef UMA_DEBUG 1510 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n", 1511 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags, 1512 keg->uk_ipers, keg->uk_ppera, 1513 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, 1514 keg->uk_free); 1515#endif 1516 1517 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1518 1519 rw_wlock(&uma_rwlock); 1520 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1521 rw_wunlock(&uma_rwlock); 1522 return (0); 1523} 1524 1525/* 1526 * Zone header ctor. This initializes all fields, locks, etc. 1527 * 1528 * Arguments/Returns follow uma_ctor specifications 1529 * udata Actually uma_zctor_args 1530 */ 1531static int 1532zone_ctor(void *mem, int size, void *udata, int flags) 1533{ 1534 struct uma_zctor_args *arg = udata; 1535 uma_zone_t zone = mem; 1536 uma_zone_t z; 1537 uma_keg_t keg; 1538 1539 bzero(zone, size); 1540 zone->uz_name = arg->name; 1541 zone->uz_ctor = arg->ctor; 1542 zone->uz_dtor = arg->dtor; 1543 zone->uz_slab = zone_fetch_slab; 1544 zone->uz_init = NULL; 1545 zone->uz_fini = NULL; 1546 zone->uz_allocs = 0; 1547 zone->uz_frees = 0; 1548 zone->uz_fails = 0; 1549 zone->uz_sleeps = 0; 1550 zone->uz_count = 0; 1551 zone->uz_count_min = 0; 1552 zone->uz_flags = 0; 1553 zone->uz_warning = NULL; 1554 timevalclear(&zone->uz_ratecheck); 1555 keg = arg->keg; 1556 1557 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 1558 1559 /* 1560 * This is a pure cache zone, no kegs. 1561 */ 1562 if (arg->import) { 1563 if (arg->flags & UMA_ZONE_VM) 1564 arg->flags |= UMA_ZFLAG_CACHEONLY; 1565 zone->uz_flags = arg->flags; 1566 zone->uz_size = arg->size; 1567 zone->uz_import = arg->import; 1568 zone->uz_release = arg->release; 1569 zone->uz_arg = arg->arg; 1570 zone->uz_lockptr = &zone->uz_lock; 1571 rw_wlock(&uma_rwlock); 1572 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 1573 rw_wunlock(&uma_rwlock); 1574 goto out; 1575 } 1576 1577 /* 1578 * Use the regular zone/keg/slab allocator. 1579 */ 1580 zone->uz_import = (uma_import)zone_import; 1581 zone->uz_release = (uma_release)zone_release; 1582 zone->uz_arg = zone; 1583 1584 if (arg->flags & UMA_ZONE_SECONDARY) { 1585 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1586 zone->uz_init = arg->uminit; 1587 zone->uz_fini = arg->fini; 1588 zone->uz_lockptr = &keg->uk_lock; 1589 zone->uz_flags |= UMA_ZONE_SECONDARY; 1590 rw_wlock(&uma_rwlock); 1591 ZONE_LOCK(zone); 1592 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1593 if (LIST_NEXT(z, uz_link) == NULL) { 1594 LIST_INSERT_AFTER(z, zone, uz_link); 1595 break; 1596 } 1597 } 1598 ZONE_UNLOCK(zone); 1599 rw_wunlock(&uma_rwlock); 1600 } else if (keg == NULL) { 1601 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1602 arg->align, arg->flags)) == NULL) 1603 return (ENOMEM); 1604 } else { 1605 struct uma_kctor_args karg; 1606 int error; 1607 1608 /* We should only be here from uma_startup() */ 1609 karg.size = arg->size; 1610 karg.uminit = arg->uminit; 1611 karg.fini = arg->fini; 1612 karg.align = arg->align; 1613 karg.flags = arg->flags; 1614 karg.zone = zone; 1615 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1616 flags); 1617 if (error) 1618 return (error); 1619 } 1620 1621 /* 1622 * Link in the first keg. 1623 */ 1624 zone->uz_klink.kl_keg = keg; 1625 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1626 zone->uz_lockptr = &keg->uk_lock; 1627 zone->uz_size = keg->uk_size; 1628 zone->uz_flags |= (keg->uk_flags & 1629 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1630 1631 /* 1632 * Some internal zones don't have room allocated for the per cpu 1633 * caches. If we're internal, bail out here. 1634 */ 1635 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1636 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1637 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1638 return (0); 1639 } 1640 1641out: 1642 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0) 1643 zone->uz_count = bucket_select(zone->uz_size); 1644 else 1645 zone->uz_count = BUCKET_MAX; 1646 zone->uz_count_min = zone->uz_count; 1647 1648 return (0); 1649} 1650 1651/* 1652 * Keg header dtor. This frees all data, destroys locks, frees the hash 1653 * table and removes the keg from the global list. 1654 * 1655 * Arguments/Returns follow uma_dtor specifications 1656 * udata unused 1657 */ 1658static void 1659keg_dtor(void *arg, int size, void *udata) 1660{ 1661 uma_keg_t keg; 1662 1663 keg = (uma_keg_t)arg; 1664 KEG_LOCK(keg); 1665 if (keg->uk_free != 0) { 1666 printf("Freed UMA keg (%s) was not empty (%d items). " 1667 " Lost %d pages of memory.\n", 1668 keg->uk_name ? keg->uk_name : "", 1669 keg->uk_free, keg->uk_pages); 1670 } 1671 KEG_UNLOCK(keg); 1672 1673 hash_free(&keg->uk_hash); 1674 1675 KEG_LOCK_FINI(keg); 1676} 1677 1678/* 1679 * Zone header dtor. 1680 * 1681 * Arguments/Returns follow uma_dtor specifications 1682 * udata unused 1683 */ 1684static void 1685zone_dtor(void *arg, int size, void *udata) 1686{ 1687 uma_klink_t klink; 1688 uma_zone_t zone; 1689 uma_keg_t keg; 1690 1691 zone = (uma_zone_t)arg; 1692 keg = zone_first_keg(zone); 1693 1694 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1695 cache_drain(zone); 1696 1697 rw_wlock(&uma_rwlock); 1698 LIST_REMOVE(zone, uz_link); 1699 rw_wunlock(&uma_rwlock); 1700 /* 1701 * XXX there are some races here where 1702 * the zone can be drained but zone lock 1703 * released and then refilled before we 1704 * remove it... we dont care for now 1705 */ 1706 zone_drain_wait(zone, M_WAITOK); 1707 /* 1708 * Unlink all of our kegs. 1709 */ 1710 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1711 klink->kl_keg = NULL; 1712 LIST_REMOVE(klink, kl_link); 1713 if (klink == &zone->uz_klink) 1714 continue; 1715 free(klink, M_TEMP); 1716 } 1717 /* 1718 * We only destroy kegs from non secondary zones. 1719 */ 1720 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1721 rw_wlock(&uma_rwlock); 1722 LIST_REMOVE(keg, uk_link); 1723 rw_wunlock(&uma_rwlock); 1724 zone_free_item(kegs, keg, NULL, SKIP_NONE); 1725 } 1726 ZONE_LOCK_FINI(zone); 1727} 1728 1729/* 1730 * Traverses every zone in the system and calls a callback 1731 * 1732 * Arguments: 1733 * zfunc A pointer to a function which accepts a zone 1734 * as an argument. 1735 * 1736 * Returns: 1737 * Nothing 1738 */ 1739static void 1740zone_foreach(void (*zfunc)(uma_zone_t)) 1741{ 1742 uma_keg_t keg; 1743 uma_zone_t zone; 1744 1745 rw_rlock(&uma_rwlock); 1746 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1747 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1748 zfunc(zone); 1749 } 1750 rw_runlock(&uma_rwlock); 1751} 1752 1753/* Public functions */ 1754/* See uma.h */ 1755void 1756uma_startup(void *bootmem, int boot_pages) 1757{ 1758 struct uma_zctor_args args; 1759 uma_slab_t slab; 1760 int i; 1761 1762#ifdef UMA_DEBUG 1763 printf("Creating uma keg headers zone and keg.\n"); 1764#endif 1765 rw_init(&uma_rwlock, "UMA lock"); 1766 1767 /* "manually" create the initial zone */ 1768 memset(&args, 0, sizeof(args)); 1769 args.name = "UMA Kegs"; 1770 args.size = sizeof(struct uma_keg); 1771 args.ctor = keg_ctor; 1772 args.dtor = keg_dtor; 1773 args.uminit = zero_init; 1774 args.fini = NULL; 1775 args.keg = &masterkeg; 1776 args.align = 32 - 1; 1777 args.flags = UMA_ZFLAG_INTERNAL; 1778 /* The initial zone has no Per cpu queues so it's smaller */ 1779 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); 1780 1781#ifdef UMA_DEBUG 1782 printf("Filling boot free list.\n"); 1783#endif 1784 for (i = 0; i < boot_pages; i++) { 1785 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE)); 1786 slab->us_data = (uint8_t *)slab; 1787 slab->us_flags = UMA_SLAB_BOOT; 1788 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); 1789 } 1790 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); 1791 1792#ifdef UMA_DEBUG 1793 printf("Creating uma zone headers zone and keg.\n"); 1794#endif 1795 args.name = "UMA Zones"; 1796 args.size = sizeof(struct uma_zone) + 1797 (sizeof(struct uma_cache) * (mp_maxid + 1)); 1798 args.ctor = zone_ctor; 1799 args.dtor = zone_dtor; 1800 args.uminit = zero_init; 1801 args.fini = NULL; 1802 args.keg = NULL; 1803 args.align = 32 - 1; 1804 args.flags = UMA_ZFLAG_INTERNAL; 1805 /* The initial zone has no Per cpu queues so it's smaller */ 1806 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); 1807 1808#ifdef UMA_DEBUG 1809 printf("Creating slab and hash zones.\n"); 1810#endif 1811 1812 /* Now make a zone for slab headers */ 1813 slabzone = uma_zcreate("UMA Slabs", 1814 sizeof(struct uma_slab), 1815 NULL, NULL, NULL, NULL, 1816 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1817 1818 hashzone = uma_zcreate("UMA Hash", 1819 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 1820 NULL, NULL, NULL, NULL, 1821 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1822 1823 bucket_init(); 1824 1825 booted = UMA_STARTUP; 1826 1827#ifdef UMA_DEBUG 1828 printf("UMA startup complete.\n"); 1829#endif 1830} 1831 1832/* see uma.h */ 1833void 1834uma_startup2(void) 1835{ 1836 booted = UMA_STARTUP2; 1837 bucket_enable(); 1838 sx_init(&uma_drain_lock, "umadrain"); 1839#ifdef UMA_DEBUG 1840 printf("UMA startup2 complete.\n"); 1841#endif 1842} 1843 1844/* 1845 * Initialize our callout handle 1846 * 1847 */ 1848 1849static void 1850uma_startup3(void) 1851{ 1852#ifdef UMA_DEBUG 1853 printf("Starting callout.\n"); 1854#endif 1855 callout_init(&uma_callout, 1); 1856 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1857#ifdef UMA_DEBUG 1858 printf("UMA startup3 complete.\n"); 1859#endif 1860} 1861 1862static uma_keg_t 1863uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 1864 int align, uint32_t flags) 1865{ 1866 struct uma_kctor_args args; 1867 1868 args.size = size; 1869 args.uminit = uminit; 1870 args.fini = fini; 1871 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 1872 args.flags = flags; 1873 args.zone = zone; 1874 return (zone_alloc_item(kegs, &args, M_WAITOK)); 1875} 1876 1877/* See uma.h */ 1878void 1879uma_set_align(int align) 1880{ 1881 1882 if (align != UMA_ALIGN_CACHE) 1883 uma_align_cache = align; 1884} 1885 1886/* See uma.h */ 1887uma_zone_t 1888uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 1889 uma_init uminit, uma_fini fini, int align, uint32_t flags) 1890 1891{ 1892 struct uma_zctor_args args; 1893 uma_zone_t res; 1894 bool locked; 1895 1896 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"", 1897 align, name)); 1898 1899 /* This stuff is essential for the zone ctor */ 1900 memset(&args, 0, sizeof(args)); 1901 args.name = name; 1902 args.size = size; 1903 args.ctor = ctor; 1904 args.dtor = dtor; 1905 args.uminit = uminit; 1906 args.fini = fini; 1907#ifdef INVARIANTS 1908 /* 1909 * If a zone is being created with an empty constructor and 1910 * destructor, pass UMA constructor/destructor which checks for 1911 * memory use after free. 1912 */ 1913 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) && 1914 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) { 1915 args.ctor = trash_ctor; 1916 args.dtor = trash_dtor; 1917 args.uminit = trash_init; 1918 args.fini = trash_fini; 1919 } 1920#endif 1921 args.align = align; 1922 args.flags = flags; 1923 args.keg = NULL; 1924 1925 if (booted < UMA_STARTUP2) { 1926 locked = false; 1927 } else { 1928 sx_slock(&uma_drain_lock); 1929 locked = true; 1930 } 1931 res = zone_alloc_item(zones, &args, M_WAITOK); 1932 if (locked) 1933 sx_sunlock(&uma_drain_lock); 1934 return (res); 1935} 1936 1937/* See uma.h */ 1938uma_zone_t 1939uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 1940 uma_init zinit, uma_fini zfini, uma_zone_t master) 1941{ 1942 struct uma_zctor_args args; 1943 uma_keg_t keg; 1944 uma_zone_t res; 1945 bool locked; 1946 1947 keg = zone_first_keg(master); 1948 memset(&args, 0, sizeof(args)); 1949 args.name = name; 1950 args.size = keg->uk_size; 1951 args.ctor = ctor; 1952 args.dtor = dtor; 1953 args.uminit = zinit; 1954 args.fini = zfini; 1955 args.align = keg->uk_align; 1956 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 1957 args.keg = keg; 1958 1959 if (booted < UMA_STARTUP2) { 1960 locked = false; 1961 } else { 1962 sx_slock(&uma_drain_lock); 1963 locked = true; 1964 } 1965 /* XXX Attaches only one keg of potentially many. */ 1966 res = zone_alloc_item(zones, &args, M_WAITOK); 1967 if (locked) 1968 sx_sunlock(&uma_drain_lock); 1969 return (res); 1970} 1971 1972/* See uma.h */ 1973uma_zone_t 1974uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 1975 uma_init zinit, uma_fini zfini, uma_import zimport, 1976 uma_release zrelease, void *arg, int flags) 1977{ 1978 struct uma_zctor_args args; 1979 1980 memset(&args, 0, sizeof(args)); 1981 args.name = name; 1982 args.size = size; 1983 args.ctor = ctor; 1984 args.dtor = dtor; 1985 args.uminit = zinit; 1986 args.fini = zfini; 1987 args.import = zimport; 1988 args.release = zrelease; 1989 args.arg = arg; 1990 args.align = 0; 1991 args.flags = flags; 1992 1993 return (zone_alloc_item(zones, &args, M_WAITOK)); 1994} 1995 1996static void 1997zone_lock_pair(uma_zone_t a, uma_zone_t b) 1998{ 1999 if (a < b) { 2000 ZONE_LOCK(a); 2001 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); 2002 } else { 2003 ZONE_LOCK(b); 2004 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); 2005 } 2006} 2007 2008static void 2009zone_unlock_pair(uma_zone_t a, uma_zone_t b) 2010{ 2011 2012 ZONE_UNLOCK(a); 2013 ZONE_UNLOCK(b); 2014} 2015 2016int 2017uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 2018{ 2019 uma_klink_t klink; 2020 uma_klink_t kl; 2021 int error; 2022 2023 error = 0; 2024 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 2025 2026 zone_lock_pair(zone, master); 2027 /* 2028 * zone must use vtoslab() to resolve objects and must already be 2029 * a secondary. 2030 */ 2031 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 2032 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 2033 error = EINVAL; 2034 goto out; 2035 } 2036 /* 2037 * The new master must also use vtoslab(). 2038 */ 2039 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 2040 error = EINVAL; 2041 goto out; 2042 } 2043 2044 /* 2045 * The underlying object must be the same size. rsize 2046 * may be different. 2047 */ 2048 if (master->uz_size != zone->uz_size) { 2049 error = E2BIG; 2050 goto out; 2051 } 2052 /* 2053 * Put it at the end of the list. 2054 */ 2055 klink->kl_keg = zone_first_keg(master); 2056 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 2057 if (LIST_NEXT(kl, kl_link) == NULL) { 2058 LIST_INSERT_AFTER(kl, klink, kl_link); 2059 break; 2060 } 2061 } 2062 klink = NULL; 2063 zone->uz_flags |= UMA_ZFLAG_MULTI; 2064 zone->uz_slab = zone_fetch_slab_multi; 2065 2066out: 2067 zone_unlock_pair(zone, master); 2068 if (klink != NULL) 2069 free(klink, M_TEMP); 2070 2071 return (error); 2072} 2073 2074 2075/* See uma.h */ 2076void 2077uma_zdestroy(uma_zone_t zone) 2078{ 2079 2080 sx_slock(&uma_drain_lock); 2081 zone_free_item(zones, zone, NULL, SKIP_NONE); 2082 sx_sunlock(&uma_drain_lock); 2083} 2084 2085/* See uma.h */ 2086void * 2087uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2088{ 2089 void *item; 2090 uma_cache_t cache; 2091 uma_bucket_t bucket; 2092 int lockfail; 2093 int cpu; 2094 2095 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2096 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA); 2097 2098 /* This is the fast path allocation */ 2099#ifdef UMA_DEBUG_ALLOC_1 2100 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); 2101#endif 2102 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread, 2103 zone->uz_name, flags); 2104 2105 if (flags & M_WAITOK) { 2106 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2107 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2108 } 2109 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2110 ("uma_zalloc_arg: called with spinlock or critical section held")); 2111 2112#ifdef DEBUG_MEMGUARD 2113 if (memguard_cmp_zone(zone)) { 2114 item = memguard_alloc(zone->uz_size, flags); 2115 if (item != NULL) { 2116 if (zone->uz_init != NULL && 2117 zone->uz_init(item, zone->uz_size, flags) != 0) 2118 return (NULL); 2119 if (zone->uz_ctor != NULL && 2120 zone->uz_ctor(item, zone->uz_size, udata, 2121 flags) != 0) { 2122 zone->uz_fini(item, zone->uz_size); 2123 return (NULL); 2124 } 2125 return (item); 2126 } 2127 /* This is unfortunate but should not be fatal. */ 2128 } 2129#endif 2130 /* 2131 * If possible, allocate from the per-CPU cache. There are two 2132 * requirements for safe access to the per-CPU cache: (1) the thread 2133 * accessing the cache must not be preempted or yield during access, 2134 * and (2) the thread must not migrate CPUs without switching which 2135 * cache it accesses. We rely on a critical section to prevent 2136 * preemption and migration. We release the critical section in 2137 * order to acquire the zone mutex if we are unable to allocate from 2138 * the current cache; when we re-acquire the critical section, we 2139 * must detect and handle migration if it has occurred. 2140 */ 2141 critical_enter(); 2142 cpu = curcpu; 2143 cache = &zone->uz_cpu[cpu]; 2144 2145zalloc_start: 2146 bucket = cache->uc_allocbucket; 2147 if (bucket != NULL && bucket->ub_cnt > 0) { 2148 bucket->ub_cnt--; 2149 item = bucket->ub_bucket[bucket->ub_cnt]; 2150#ifdef INVARIANTS 2151 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2152#endif 2153 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2154 cache->uc_allocs++; 2155 critical_exit(); 2156 if (zone->uz_ctor != NULL && 2157 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2158 atomic_add_long(&zone->uz_fails, 1); 2159 zone_free_item(zone, item, udata, SKIP_DTOR); 2160 return (NULL); 2161 } 2162#ifdef INVARIANTS 2163 uma_dbg_alloc(zone, NULL, item); 2164#endif 2165 if (flags & M_ZERO) 2166 uma_zero_item(item, zone); 2167 return (item); 2168 } 2169 2170 /* 2171 * We have run out of items in our alloc bucket. 2172 * See if we can switch with our free bucket. 2173 */ 2174 bucket = cache->uc_freebucket; 2175 if (bucket != NULL && bucket->ub_cnt > 0) { 2176#ifdef UMA_DEBUG_ALLOC 2177 printf("uma_zalloc: Swapping empty with alloc.\n"); 2178#endif 2179 cache->uc_freebucket = cache->uc_allocbucket; 2180 cache->uc_allocbucket = bucket; 2181 goto zalloc_start; 2182 } 2183 2184 /* 2185 * Discard any empty allocation bucket while we hold no locks. 2186 */ 2187 bucket = cache->uc_allocbucket; 2188 cache->uc_allocbucket = NULL; 2189 critical_exit(); 2190 if (bucket != NULL) 2191 bucket_free(zone, bucket, udata); 2192 2193 /* Short-circuit for zones without buckets and low memory. */ 2194 if (zone->uz_count == 0 || bucketdisable) 2195 goto zalloc_item; 2196 2197 /* 2198 * Attempt to retrieve the item from the per-CPU cache has failed, so 2199 * we must go back to the zone. This requires the zone lock, so we 2200 * must drop the critical section, then re-acquire it when we go back 2201 * to the cache. Since the critical section is released, we may be 2202 * preempted or migrate. As such, make sure not to maintain any 2203 * thread-local state specific to the cache from prior to releasing 2204 * the critical section. 2205 */ 2206 lockfail = 0; 2207 if (ZONE_TRYLOCK(zone) == 0) { 2208 /* Record contention to size the buckets. */ 2209 ZONE_LOCK(zone); 2210 lockfail = 1; 2211 } 2212 critical_enter(); 2213 cpu = curcpu; 2214 cache = &zone->uz_cpu[cpu]; 2215 2216 /* 2217 * Since we have locked the zone we may as well send back our stats. 2218 */ 2219 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2220 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2221 cache->uc_allocs = 0; 2222 cache->uc_frees = 0; 2223 2224 /* See if we lost the race to fill the cache. */ 2225 if (cache->uc_allocbucket != NULL) { 2226 ZONE_UNLOCK(zone); 2227 goto zalloc_start; 2228 } 2229 2230 /* 2231 * Check the zone's cache of buckets. 2232 */ 2233 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 2234 KASSERT(bucket->ub_cnt != 0, 2235 ("uma_zalloc_arg: Returning an empty bucket.")); 2236 2237 LIST_REMOVE(bucket, ub_link); 2238 cache->uc_allocbucket = bucket; 2239 ZONE_UNLOCK(zone); 2240 goto zalloc_start; 2241 } 2242 /* We are no longer associated with this CPU. */ 2243 critical_exit(); 2244 2245 /* 2246 * We bump the uz count when the cache size is insufficient to 2247 * handle the working set. 2248 */ 2249 if (lockfail && zone->uz_count < BUCKET_MAX) 2250 zone->uz_count++; 2251 ZONE_UNLOCK(zone); 2252 2253 /* 2254 * Now lets just fill a bucket and put it on the free list. If that 2255 * works we'll restart the allocation from the beginning and it 2256 * will use the just filled bucket. 2257 */ 2258 bucket = zone_alloc_bucket(zone, udata, flags); 2259 if (bucket != NULL) { 2260 ZONE_LOCK(zone); 2261 critical_enter(); 2262 cpu = curcpu; 2263 cache = &zone->uz_cpu[cpu]; 2264 /* 2265 * See if we lost the race or were migrated. Cache the 2266 * initialized bucket to make this less likely or claim 2267 * the memory directly. 2268 */ 2269 if (cache->uc_allocbucket == NULL) 2270 cache->uc_allocbucket = bucket; 2271 else 2272 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2273 ZONE_UNLOCK(zone); 2274 goto zalloc_start; 2275 } 2276 2277 /* 2278 * We may not be able to get a bucket so return an actual item. 2279 */ 2280#ifdef UMA_DEBUG 2281 printf("uma_zalloc_arg: Bucketzone returned NULL\n"); 2282#endif 2283 2284zalloc_item: 2285 item = zone_alloc_item(zone, udata, flags); 2286 2287 return (item); 2288} 2289 2290static uma_slab_t 2291keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) 2292{ 2293 uma_slab_t slab; 2294 int reserve; 2295 2296 mtx_assert(&keg->uk_lock, MA_OWNED); 2297 slab = NULL; 2298 reserve = 0; 2299 if ((flags & M_USE_RESERVE) == 0) 2300 reserve = keg->uk_reserve; 2301 2302 for (;;) { 2303 /* 2304 * Find a slab with some space. Prefer slabs that are partially 2305 * used over those that are totally full. This helps to reduce 2306 * fragmentation. 2307 */ 2308 if (keg->uk_free > reserve) { 2309 if (!LIST_EMPTY(&keg->uk_part_slab)) { 2310 slab = LIST_FIRST(&keg->uk_part_slab); 2311 } else { 2312 slab = LIST_FIRST(&keg->uk_free_slab); 2313 LIST_REMOVE(slab, us_link); 2314 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, 2315 us_link); 2316 } 2317 MPASS(slab->us_keg == keg); 2318 return (slab); 2319 } 2320 2321 /* 2322 * M_NOVM means don't ask at all! 2323 */ 2324 if (flags & M_NOVM) 2325 break; 2326 2327 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2328 keg->uk_flags |= UMA_ZFLAG_FULL; 2329 /* 2330 * If this is not a multi-zone, set the FULL bit. 2331 * Otherwise slab_multi() takes care of it. 2332 */ 2333 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2334 zone->uz_flags |= UMA_ZFLAG_FULL; 2335 zone_log_warning(zone); 2336 zone_maxaction(zone); 2337 } 2338 if (flags & M_NOWAIT) 2339 break; 2340 zone->uz_sleeps++; 2341 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2342 continue; 2343 } 2344 slab = keg_alloc_slab(keg, zone, flags); 2345 /* 2346 * If we got a slab here it's safe to mark it partially used 2347 * and return. We assume that the caller is going to remove 2348 * at least one item. 2349 */ 2350 if (slab) { 2351 MPASS(slab->us_keg == keg); 2352 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2353 return (slab); 2354 } 2355 /* 2356 * We might not have been able to get a slab but another cpu 2357 * could have while we were unlocked. Check again before we 2358 * fail. 2359 */ 2360 flags |= M_NOVM; 2361 } 2362 return (slab); 2363} 2364 2365static uma_slab_t 2366zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) 2367{ 2368 uma_slab_t slab; 2369 2370 if (keg == NULL) { 2371 keg = zone_first_keg(zone); 2372 KEG_LOCK(keg); 2373 } 2374 2375 for (;;) { 2376 slab = keg_fetch_slab(keg, zone, flags); 2377 if (slab) 2378 return (slab); 2379 if (flags & (M_NOWAIT | M_NOVM)) 2380 break; 2381 } 2382 KEG_UNLOCK(keg); 2383 return (NULL); 2384} 2385 2386/* 2387 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2388 * with the keg locked. On NULL no lock is held. 2389 * 2390 * The last pointer is used to seed the search. It is not required. 2391 */ 2392static uma_slab_t 2393zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) 2394{ 2395 uma_klink_t klink; 2396 uma_slab_t slab; 2397 uma_keg_t keg; 2398 int flags; 2399 int empty; 2400 int full; 2401 2402 /* 2403 * Don't wait on the first pass. This will skip limit tests 2404 * as well. We don't want to block if we can find a provider 2405 * without blocking. 2406 */ 2407 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2408 /* 2409 * Use the last slab allocated as a hint for where to start 2410 * the search. 2411 */ 2412 if (last != NULL) { 2413 slab = keg_fetch_slab(last, zone, flags); 2414 if (slab) 2415 return (slab); 2416 KEG_UNLOCK(last); 2417 } 2418 /* 2419 * Loop until we have a slab incase of transient failures 2420 * while M_WAITOK is specified. I'm not sure this is 100% 2421 * required but we've done it for so long now. 2422 */ 2423 for (;;) { 2424 empty = 0; 2425 full = 0; 2426 /* 2427 * Search the available kegs for slabs. Be careful to hold the 2428 * correct lock while calling into the keg layer. 2429 */ 2430 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2431 keg = klink->kl_keg; 2432 KEG_LOCK(keg); 2433 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2434 slab = keg_fetch_slab(keg, zone, flags); 2435 if (slab) 2436 return (slab); 2437 } 2438 if (keg->uk_flags & UMA_ZFLAG_FULL) 2439 full++; 2440 else 2441 empty++; 2442 KEG_UNLOCK(keg); 2443 } 2444 if (rflags & (M_NOWAIT | M_NOVM)) 2445 break; 2446 flags = rflags; 2447 /* 2448 * All kegs are full. XXX We can't atomically check all kegs 2449 * and sleep so just sleep for a short period and retry. 2450 */ 2451 if (full && !empty) { 2452 ZONE_LOCK(zone); 2453 zone->uz_flags |= UMA_ZFLAG_FULL; 2454 zone->uz_sleeps++; 2455 zone_log_warning(zone); 2456 zone_maxaction(zone); 2457 msleep(zone, zone->uz_lockptr, PVM, 2458 "zonelimit", hz/100); 2459 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2460 ZONE_UNLOCK(zone); 2461 continue; 2462 } 2463 } 2464 return (NULL); 2465} 2466 2467static void * 2468slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 2469{ 2470 void *item; 2471 uint8_t freei; 2472 2473 MPASS(keg == slab->us_keg); 2474 mtx_assert(&keg->uk_lock, MA_OWNED); 2475 2476 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; 2477 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); 2478 item = slab->us_data + (keg->uk_rsize * freei); 2479 slab->us_freecount--; 2480 keg->uk_free--; 2481 2482 /* Move this slab to the full list */ 2483 if (slab->us_freecount == 0) { 2484 LIST_REMOVE(slab, us_link); 2485 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); 2486 } 2487 2488 return (item); 2489} 2490 2491static int 2492zone_import(uma_zone_t zone, void **bucket, int max, int flags) 2493{ 2494 uma_slab_t slab; 2495 uma_keg_t keg; 2496 int i; 2497 2498 slab = NULL; 2499 keg = NULL; 2500 /* Try to keep the buckets totally full */ 2501 for (i = 0; i < max; ) { 2502 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL) 2503 break; 2504 keg = slab->us_keg; 2505 while (slab->us_freecount && i < max) { 2506 bucket[i++] = slab_alloc_item(keg, slab); 2507 if (keg->uk_free <= keg->uk_reserve) 2508 break; 2509 } 2510 /* Don't grab more than one slab at a time. */ 2511 flags &= ~M_WAITOK; 2512 flags |= M_NOWAIT; 2513 } 2514 if (slab != NULL) 2515 KEG_UNLOCK(keg); 2516 2517 return i; 2518} 2519 2520static uma_bucket_t 2521zone_alloc_bucket(uma_zone_t zone, void *udata, int flags) 2522{ 2523 uma_bucket_t bucket; 2524 int max; 2525 2526 /* Don't wait for buckets, preserve caller's NOVM setting. */ 2527 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 2528 if (bucket == NULL) 2529 return (NULL); 2530 2531 max = MIN(bucket->ub_entries, zone->uz_count); 2532 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 2533 max, flags); 2534 2535 /* 2536 * Initialize the memory if necessary. 2537 */ 2538 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 2539 int i; 2540 2541 for (i = 0; i < bucket->ub_cnt; i++) 2542 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2543 flags) != 0) 2544 break; 2545 /* 2546 * If we couldn't initialize the whole bucket, put the 2547 * rest back onto the freelist. 2548 */ 2549 if (i != bucket->ub_cnt) { 2550 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 2551 bucket->ub_cnt - i); 2552#ifdef INVARIANTS 2553 bzero(&bucket->ub_bucket[i], 2554 sizeof(void *) * (bucket->ub_cnt - i)); 2555#endif 2556 bucket->ub_cnt = i; 2557 } 2558 } 2559 2560 if (bucket->ub_cnt == 0) { 2561 bucket_free(zone, bucket, udata); 2562 atomic_add_long(&zone->uz_fails, 1); 2563 return (NULL); 2564 } 2565 2566 return (bucket); 2567} 2568 2569/* 2570 * Allocates a single item from a zone. 2571 * 2572 * Arguments 2573 * zone The zone to alloc for. 2574 * udata The data to be passed to the constructor. 2575 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2576 * 2577 * Returns 2578 * NULL if there is no memory and M_NOWAIT is set 2579 * An item if successful 2580 */ 2581 2582static void * 2583zone_alloc_item(uma_zone_t zone, void *udata, int flags) 2584{ 2585 void *item; 2586 2587 item = NULL; 2588 2589#ifdef UMA_DEBUG_ALLOC 2590 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); 2591#endif 2592 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1) 2593 goto fail; 2594 atomic_add_long(&zone->uz_allocs, 1); 2595 2596 /* 2597 * We have to call both the zone's init (not the keg's init) 2598 * and the zone's ctor. This is because the item is going from 2599 * a keg slab directly to the user, and the user is expecting it 2600 * to be both zone-init'd as well as zone-ctor'd. 2601 */ 2602 if (zone->uz_init != NULL) { 2603 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 2604 zone_free_item(zone, item, udata, SKIP_FINI); 2605 goto fail; 2606 } 2607 } 2608 if (zone->uz_ctor != NULL) { 2609 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2610 zone_free_item(zone, item, udata, SKIP_DTOR); 2611 goto fail; 2612 } 2613 } 2614#ifdef INVARIANTS 2615 uma_dbg_alloc(zone, NULL, item); 2616#endif 2617 if (flags & M_ZERO) 2618 uma_zero_item(item, zone); 2619 2620 return (item); 2621 2622fail: 2623 atomic_add_long(&zone->uz_fails, 1); 2624 return (NULL); 2625} 2626 2627/* See uma.h */ 2628void 2629uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 2630{ 2631 uma_cache_t cache; 2632 uma_bucket_t bucket; 2633 int lockfail; 2634 int cpu; 2635 2636 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2637 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA); 2638 2639#ifdef UMA_DEBUG_ALLOC_1 2640 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); 2641#endif 2642 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 2643 zone->uz_name); 2644 2645 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2646 ("uma_zfree_arg: called with spinlock or critical section held")); 2647 2648 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 2649 if (item == NULL) 2650 return; 2651#ifdef DEBUG_MEMGUARD 2652 if (is_memguard_addr(item)) { 2653 if (zone->uz_dtor != NULL) 2654 zone->uz_dtor(item, zone->uz_size, udata); 2655 if (zone->uz_fini != NULL) 2656 zone->uz_fini(item, zone->uz_size); 2657 memguard_free(item); 2658 return; 2659 } 2660#endif 2661#ifdef INVARIANTS 2662 if (zone->uz_flags & UMA_ZONE_MALLOC) 2663 uma_dbg_free(zone, udata, item); 2664 else 2665 uma_dbg_free(zone, NULL, item); 2666#endif 2667 if (zone->uz_dtor != NULL) 2668 zone->uz_dtor(item, zone->uz_size, udata); 2669 2670 /* 2671 * The race here is acceptable. If we miss it we'll just have to wait 2672 * a little longer for the limits to be reset. 2673 */ 2674 if (zone->uz_flags & UMA_ZFLAG_FULL) 2675 goto zfree_item; 2676 2677 /* 2678 * If possible, free to the per-CPU cache. There are two 2679 * requirements for safe access to the per-CPU cache: (1) the thread 2680 * accessing the cache must not be preempted or yield during access, 2681 * and (2) the thread must not migrate CPUs without switching which 2682 * cache it accesses. We rely on a critical section to prevent 2683 * preemption and migration. We release the critical section in 2684 * order to acquire the zone mutex if we are unable to free to the 2685 * current cache; when we re-acquire the critical section, we must 2686 * detect and handle migration if it has occurred. 2687 */ 2688zfree_restart: 2689 critical_enter(); 2690 cpu = curcpu; 2691 cache = &zone->uz_cpu[cpu]; 2692 2693zfree_start: 2694 /* 2695 * Try to free into the allocbucket first to give LIFO ordering 2696 * for cache-hot datastructures. Spill over into the freebucket 2697 * if necessary. Alloc will swap them if one runs dry. 2698 */ 2699 bucket = cache->uc_allocbucket; 2700 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 2701 bucket = cache->uc_freebucket; 2702 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2703 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 2704 ("uma_zfree: Freeing to non free bucket index.")); 2705 bucket->ub_bucket[bucket->ub_cnt] = item; 2706 bucket->ub_cnt++; 2707 cache->uc_frees++; 2708 critical_exit(); 2709 return; 2710 } 2711 2712 /* 2713 * We must go back the zone, which requires acquiring the zone lock, 2714 * which in turn means we must release and re-acquire the critical 2715 * section. Since the critical section is released, we may be 2716 * preempted or migrate. As such, make sure not to maintain any 2717 * thread-local state specific to the cache from prior to releasing 2718 * the critical section. 2719 */ 2720 critical_exit(); 2721 if (zone->uz_count == 0 || bucketdisable) 2722 goto zfree_item; 2723 2724 lockfail = 0; 2725 if (ZONE_TRYLOCK(zone) == 0) { 2726 /* Record contention to size the buckets. */ 2727 ZONE_LOCK(zone); 2728 lockfail = 1; 2729 } 2730 critical_enter(); 2731 cpu = curcpu; 2732 cache = &zone->uz_cpu[cpu]; 2733 2734 /* 2735 * Since we have locked the zone we may as well send back our stats. 2736 */ 2737 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2738 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2739 cache->uc_allocs = 0; 2740 cache->uc_frees = 0; 2741 2742 bucket = cache->uc_freebucket; 2743 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2744 ZONE_UNLOCK(zone); 2745 goto zfree_start; 2746 } 2747 cache->uc_freebucket = NULL; 2748 /* We are no longer associated with this CPU. */ 2749 critical_exit(); 2750 2751 /* Can we throw this on the zone full list? */ 2752 if (bucket != NULL) { 2753#ifdef UMA_DEBUG_ALLOC 2754 printf("uma_zfree: Putting old bucket on the free list.\n"); 2755#endif 2756 /* ub_cnt is pointing to the last free item */ 2757 KASSERT(bucket->ub_cnt != 0, 2758 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 2759 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2760 } 2761 2762 /* 2763 * We bump the uz count when the cache size is insufficient to 2764 * handle the working set. 2765 */ 2766 if (lockfail && zone->uz_count < BUCKET_MAX) 2767 zone->uz_count++; 2768 ZONE_UNLOCK(zone); 2769 2770#ifdef UMA_DEBUG_ALLOC 2771 printf("uma_zfree: Allocating new free bucket.\n"); 2772#endif 2773 bucket = bucket_alloc(zone, udata, M_NOWAIT); 2774 if (bucket) { 2775 critical_enter(); 2776 cpu = curcpu; 2777 cache = &zone->uz_cpu[cpu]; 2778 if (cache->uc_freebucket == NULL) { 2779 cache->uc_freebucket = bucket; 2780 goto zfree_start; 2781 } 2782 /* 2783 * We lost the race, start over. We have to drop our 2784 * critical section to free the bucket. 2785 */ 2786 critical_exit(); 2787 bucket_free(zone, bucket, udata); 2788 goto zfree_restart; 2789 } 2790 2791 /* 2792 * If nothing else caught this, we'll just do an internal free. 2793 */ 2794zfree_item: 2795 zone_free_item(zone, item, udata, SKIP_DTOR); 2796 2797 return; 2798} 2799 2800static void 2801slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) 2802{ 2803 uint8_t freei; 2804 2805 mtx_assert(&keg->uk_lock, MA_OWNED); 2806 MPASS(keg == slab->us_keg); 2807 2808 /* Do we need to remove from any lists? */ 2809 if (slab->us_freecount+1 == keg->uk_ipers) { 2810 LIST_REMOVE(slab, us_link); 2811 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 2812 } else if (slab->us_freecount == 0) { 2813 LIST_REMOVE(slab, us_link); 2814 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2815 } 2816 2817 /* Slab management. */ 2818 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 2819 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); 2820 slab->us_freecount++; 2821 2822 /* Keg statistics. */ 2823 keg->uk_free++; 2824} 2825 2826static void 2827zone_release(uma_zone_t zone, void **bucket, int cnt) 2828{ 2829 void *item; 2830 uma_slab_t slab; 2831 uma_keg_t keg; 2832 uint8_t *mem; 2833 int clearfull; 2834 int i; 2835 2836 clearfull = 0; 2837 keg = zone_first_keg(zone); 2838 KEG_LOCK(keg); 2839 for (i = 0; i < cnt; i++) { 2840 item = bucket[i]; 2841 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 2842 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 2843 if (zone->uz_flags & UMA_ZONE_HASH) { 2844 slab = hash_sfind(&keg->uk_hash, mem); 2845 } else { 2846 mem += keg->uk_pgoff; 2847 slab = (uma_slab_t)mem; 2848 } 2849 } else { 2850 slab = vtoslab((vm_offset_t)item); 2851 if (slab->us_keg != keg) { 2852 KEG_UNLOCK(keg); 2853 keg = slab->us_keg; 2854 KEG_LOCK(keg); 2855 } 2856 } 2857 slab_free_item(keg, slab, item); 2858 if (keg->uk_flags & UMA_ZFLAG_FULL) { 2859 if (keg->uk_pages < keg->uk_maxpages) { 2860 keg->uk_flags &= ~UMA_ZFLAG_FULL; 2861 clearfull = 1; 2862 } 2863 2864 /* 2865 * We can handle one more allocation. Since we're 2866 * clearing ZFLAG_FULL, wake up all procs blocked 2867 * on pages. This should be uncommon, so keeping this 2868 * simple for now (rather than adding count of blocked 2869 * threads etc). 2870 */ 2871 wakeup(keg); 2872 } 2873 } 2874 KEG_UNLOCK(keg); 2875 if (clearfull) { 2876 ZONE_LOCK(zone); 2877 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2878 wakeup(zone); 2879 ZONE_UNLOCK(zone); 2880 } 2881 2882} 2883 2884/* 2885 * Frees a single item to any zone. 2886 * 2887 * Arguments: 2888 * zone The zone to free to 2889 * item The item we're freeing 2890 * udata User supplied data for the dtor 2891 * skip Skip dtors and finis 2892 */ 2893static void 2894zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 2895{ 2896 2897#ifdef INVARIANTS 2898 if (skip == SKIP_NONE) { 2899 if (zone->uz_flags & UMA_ZONE_MALLOC) 2900 uma_dbg_free(zone, udata, item); 2901 else 2902 uma_dbg_free(zone, NULL, item); 2903 } 2904#endif 2905 if (skip < SKIP_DTOR && zone->uz_dtor) 2906 zone->uz_dtor(item, zone->uz_size, udata); 2907 2908 if (skip < SKIP_FINI && zone->uz_fini) 2909 zone->uz_fini(item, zone->uz_size); 2910 2911 atomic_add_long(&zone->uz_frees, 1); 2912 zone->uz_release(zone->uz_arg, &item, 1); 2913} 2914 2915/* See uma.h */ 2916int 2917uma_zone_set_max(uma_zone_t zone, int nitems) 2918{ 2919 uma_keg_t keg; 2920 2921 keg = zone_first_keg(zone); 2922 if (keg == NULL) 2923 return (0); 2924 KEG_LOCK(keg); 2925 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 2926 if (keg->uk_maxpages * keg->uk_ipers < nitems) 2927 keg->uk_maxpages += keg->uk_ppera; 2928 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers; 2929 KEG_UNLOCK(keg); 2930 2931 return (nitems); 2932} 2933 2934/* See uma.h */ 2935int 2936uma_zone_get_max(uma_zone_t zone) 2937{ 2938 int nitems; 2939 uma_keg_t keg; 2940 2941 keg = zone_first_keg(zone); 2942 if (keg == NULL) 2943 return (0); 2944 KEG_LOCK(keg); 2945 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers; 2946 KEG_UNLOCK(keg); 2947 2948 return (nitems); 2949} 2950 2951/* See uma.h */ 2952void 2953uma_zone_set_warning(uma_zone_t zone, const char *warning) 2954{ 2955 2956 ZONE_LOCK(zone); 2957 zone->uz_warning = warning; 2958 ZONE_UNLOCK(zone); 2959} 2960 2961/* See uma.h */ 2962void 2963uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) 2964{ 2965 2966 ZONE_LOCK(zone); 2967 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); 2968 ZONE_UNLOCK(zone); 2969} 2970 2971/* See uma.h */ 2972int 2973uma_zone_get_cur(uma_zone_t zone) 2974{ 2975 int64_t nitems; 2976 u_int i; 2977 2978 ZONE_LOCK(zone); 2979 nitems = zone->uz_allocs - zone->uz_frees; 2980 CPU_FOREACH(i) { 2981 /* 2982 * See the comment in sysctl_vm_zone_stats() regarding the 2983 * safety of accessing the per-cpu caches. With the zone lock 2984 * held, it is safe, but can potentially result in stale data. 2985 */ 2986 nitems += zone->uz_cpu[i].uc_allocs - 2987 zone->uz_cpu[i].uc_frees; 2988 } 2989 ZONE_UNLOCK(zone); 2990 2991 return (nitems < 0 ? 0 : nitems); 2992} 2993 2994/* See uma.h */ 2995void 2996uma_zone_set_init(uma_zone_t zone, uma_init uminit) 2997{ 2998 uma_keg_t keg; 2999 3000 keg = zone_first_keg(zone); 3001 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 3002 KEG_LOCK(keg); 3003 KASSERT(keg->uk_pages == 0, 3004 ("uma_zone_set_init on non-empty keg")); 3005 keg->uk_init = uminit; 3006 KEG_UNLOCK(keg); 3007} 3008 3009/* See uma.h */ 3010void 3011uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 3012{ 3013 uma_keg_t keg; 3014 3015 keg = zone_first_keg(zone); 3016 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type")); 3017 KEG_LOCK(keg); 3018 KASSERT(keg->uk_pages == 0, 3019 ("uma_zone_set_fini on non-empty keg")); 3020 keg->uk_fini = fini; 3021 KEG_UNLOCK(keg); 3022} 3023 3024/* See uma.h */ 3025void 3026uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 3027{ 3028 3029 ZONE_LOCK(zone); 3030 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3031 ("uma_zone_set_zinit on non-empty keg")); 3032 zone->uz_init = zinit; 3033 ZONE_UNLOCK(zone); 3034} 3035 3036/* See uma.h */ 3037void 3038uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 3039{ 3040 3041 ZONE_LOCK(zone); 3042 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3043 ("uma_zone_set_zfini on non-empty keg")); 3044 zone->uz_fini = zfini; 3045 ZONE_UNLOCK(zone); 3046} 3047 3048/* See uma.h */ 3049/* XXX uk_freef is not actually used with the zone locked */ 3050void 3051uma_zone_set_freef(uma_zone_t zone, uma_free freef) 3052{ 3053 uma_keg_t keg; 3054 3055 keg = zone_first_keg(zone); 3056 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type")); 3057 KEG_LOCK(keg); 3058 keg->uk_freef = freef; 3059 KEG_UNLOCK(keg); 3060} 3061 3062/* See uma.h */ 3063/* XXX uk_allocf is not actually used with the zone locked */ 3064void 3065uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3066{ 3067 uma_keg_t keg; 3068 3069 keg = zone_first_keg(zone); 3070 KEG_LOCK(keg); 3071 keg->uk_allocf = allocf; 3072 KEG_UNLOCK(keg); 3073} 3074 3075/* See uma.h */ 3076void 3077uma_zone_reserve(uma_zone_t zone, int items) 3078{ 3079 uma_keg_t keg; 3080 3081 keg = zone_first_keg(zone); 3082 if (keg == NULL) 3083 return; 3084 KEG_LOCK(keg); 3085 keg->uk_reserve = items; 3086 KEG_UNLOCK(keg); 3087 3088 return; 3089} 3090 3091/* See uma.h */ 3092int 3093uma_zone_reserve_kva(uma_zone_t zone, int count) 3094{ 3095 uma_keg_t keg; 3096 vm_offset_t kva; 3097 u_int pages; 3098 3099 keg = zone_first_keg(zone); 3100 if (keg == NULL) 3101 return (0); 3102 pages = count / keg->uk_ipers; 3103 3104 if (pages * keg->uk_ipers < count) 3105 pages++; 3106 pages *= keg->uk_ppera; 3107 3108#ifdef UMA_MD_SMALL_ALLOC 3109 if (keg->uk_ppera > 1) { 3110#else 3111 if (1) { 3112#endif 3113 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE); 3114 if (kva == 0) 3115 return (0); 3116 } else 3117 kva = 0; 3118 KEG_LOCK(keg); 3119 keg->uk_kva = kva; 3120 keg->uk_offset = 0; 3121 keg->uk_maxpages = pages; 3122#ifdef UMA_MD_SMALL_ALLOC 3123 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3124#else 3125 keg->uk_allocf = noobj_alloc; 3126#endif 3127 keg->uk_flags |= UMA_ZONE_NOFREE; 3128 KEG_UNLOCK(keg); 3129 3130 return (1); 3131} 3132 3133/* See uma.h */ 3134void 3135uma_prealloc(uma_zone_t zone, int items) 3136{ 3137 int slabs; 3138 uma_slab_t slab; 3139 uma_keg_t keg; 3140 3141 keg = zone_first_keg(zone); 3142 if (keg == NULL) 3143 return; 3144 KEG_LOCK(keg); 3145 slabs = items / keg->uk_ipers; 3146 if (slabs * keg->uk_ipers < items) 3147 slabs++; 3148 while (slabs > 0) { 3149 slab = keg_alloc_slab(keg, zone, M_WAITOK); 3150 if (slab == NULL) 3151 break; 3152 MPASS(slab->us_keg == keg); 3153 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 3154 slabs--; 3155 } 3156 KEG_UNLOCK(keg); 3157} 3158 3159/* See uma.h */ 3160static void 3161uma_reclaim_locked(bool kmem_danger) 3162{ 3163 3164#ifdef UMA_DEBUG 3165 printf("UMA: vm asked us to release pages!\n"); 3166#endif 3167 sx_assert(&uma_drain_lock, SA_XLOCKED); 3168 bucket_enable(); 3169 zone_foreach(zone_drain); 3170 if (vm_page_count_min() || kmem_danger) { 3171 cache_drain_safe(NULL); 3172 zone_foreach(zone_drain); 3173 } 3174 /* 3175 * Some slabs may have been freed but this zone will be visited early 3176 * we visit again so that we can free pages that are empty once other 3177 * zones are drained. We have to do the same for buckets. 3178 */ 3179 zone_drain(slabzone); 3180 bucket_zone_drain(); 3181} 3182 3183void 3184uma_reclaim(void) 3185{ 3186 3187 sx_xlock(&uma_drain_lock); 3188 uma_reclaim_locked(false); 3189 sx_xunlock(&uma_drain_lock); 3190} 3191 3192static int uma_reclaim_needed; 3193 3194void 3195uma_reclaim_wakeup(void) 3196{ 3197 3198 uma_reclaim_needed = 1; 3199 wakeup(&uma_reclaim_needed); 3200} 3201 3202void 3203uma_reclaim_worker(void *arg __unused) 3204{ 3205 3206 sx_xlock(&uma_drain_lock); 3207 for (;;) { 3208 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM, 3209 "umarcl", 0); 3210 if (uma_reclaim_needed) { 3211 uma_reclaim_needed = 0; 3212 sx_xunlock(&uma_drain_lock); 3213 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); 3214 sx_xlock(&uma_drain_lock); 3215 uma_reclaim_locked(true); 3216 } 3217 } 3218} 3219 3220/* See uma.h */ 3221int 3222uma_zone_exhausted(uma_zone_t zone) 3223{ 3224 int full; 3225 3226 ZONE_LOCK(zone); 3227 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3228 ZONE_UNLOCK(zone); 3229 return (full); 3230} 3231 3232int 3233uma_zone_exhausted_nolock(uma_zone_t zone) 3234{ 3235 return (zone->uz_flags & UMA_ZFLAG_FULL); 3236} 3237 3238void * 3239uma_large_malloc(vm_size_t size, int wait) 3240{ 3241 void *mem; 3242 uma_slab_t slab; 3243 uint8_t flags; 3244 3245 slab = zone_alloc_item(slabzone, NULL, wait); 3246 if (slab == NULL) 3247 return (NULL); 3248 mem = page_alloc(NULL, size, &flags, wait); 3249 if (mem) { 3250 vsetslab((vm_offset_t)mem, slab); 3251 slab->us_data = mem; 3252 slab->us_flags = flags | UMA_SLAB_MALLOC; 3253 slab->us_size = size; 3254 } else { 3255 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3256 } 3257 3258 return (mem); 3259} 3260 3261void 3262uma_large_free(uma_slab_t slab) 3263{ 3264 3265 page_free(slab->us_data, slab->us_size, slab->us_flags); 3266 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3267} 3268 3269static void 3270uma_zero_item(void *item, uma_zone_t zone) 3271{ 3272 int i; 3273 3274 if (zone->uz_flags & UMA_ZONE_PCPU) { 3275 CPU_FOREACH(i) 3276 bzero(zpcpu_get_cpu(item, i), zone->uz_size); 3277 } else 3278 bzero(item, zone->uz_size); 3279} 3280 3281void 3282uma_print_stats(void) 3283{ 3284 zone_foreach(uma_print_zone); 3285} 3286 3287static void 3288slab_print(uma_slab_t slab) 3289{ 3290 printf("slab: keg %p, data %p, freecount %d\n", 3291 slab->us_keg, slab->us_data, slab->us_freecount); 3292} 3293 3294static void 3295cache_print(uma_cache_t cache) 3296{ 3297 printf("alloc: %p(%d), free: %p(%d)\n", 3298 cache->uc_allocbucket, 3299 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3300 cache->uc_freebucket, 3301 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3302} 3303 3304static void 3305uma_print_keg(uma_keg_t keg) 3306{ 3307 uma_slab_t slab; 3308 3309 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3310 "out %d free %d limit %d\n", 3311 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3312 keg->uk_ipers, keg->uk_ppera, 3313 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, 3314 keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3315 printf("Part slabs:\n"); 3316 LIST_FOREACH(slab, &keg->uk_part_slab, us_link) 3317 slab_print(slab); 3318 printf("Free slabs:\n"); 3319 LIST_FOREACH(slab, &keg->uk_free_slab, us_link) 3320 slab_print(slab); 3321 printf("Full slabs:\n"); 3322 LIST_FOREACH(slab, &keg->uk_full_slab, us_link) 3323 slab_print(slab); 3324} 3325 3326void 3327uma_print_zone(uma_zone_t zone) 3328{ 3329 uma_cache_t cache; 3330 uma_klink_t kl; 3331 int i; 3332 3333 printf("zone: %s(%p) size %d flags %#x\n", 3334 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3335 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3336 uma_print_keg(kl->kl_keg); 3337 CPU_FOREACH(i) { 3338 cache = &zone->uz_cpu[i]; 3339 printf("CPU %d Cache:\n", i); 3340 cache_print(cache); 3341 } 3342} 3343 3344#ifdef DDB 3345/* 3346 * Generate statistics across both the zone and its per-cpu cache's. Return 3347 * desired statistics if the pointer is non-NULL for that statistic. 3348 * 3349 * Note: does not update the zone statistics, as it can't safely clear the 3350 * per-CPU cache statistic. 3351 * 3352 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3353 * safe from off-CPU; we should modify the caches to track this information 3354 * directly so that we don't have to. 3355 */ 3356static void 3357uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3358 uint64_t *freesp, uint64_t *sleepsp) 3359{ 3360 uma_cache_t cache; 3361 uint64_t allocs, frees, sleeps; 3362 int cachefree, cpu; 3363 3364 allocs = frees = sleeps = 0; 3365 cachefree = 0; 3366 CPU_FOREACH(cpu) { 3367 cache = &z->uz_cpu[cpu]; 3368 if (cache->uc_allocbucket != NULL) 3369 cachefree += cache->uc_allocbucket->ub_cnt; 3370 if (cache->uc_freebucket != NULL) 3371 cachefree += cache->uc_freebucket->ub_cnt; 3372 allocs += cache->uc_allocs; 3373 frees += cache->uc_frees; 3374 } 3375 allocs += z->uz_allocs; 3376 frees += z->uz_frees; 3377 sleeps += z->uz_sleeps; 3378 if (cachefreep != NULL) 3379 *cachefreep = cachefree; 3380 if (allocsp != NULL) 3381 *allocsp = allocs; 3382 if (freesp != NULL) 3383 *freesp = frees; 3384 if (sleepsp != NULL) 3385 *sleepsp = sleeps; 3386} 3387#endif /* DDB */ 3388 3389static int 3390sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3391{ 3392 uma_keg_t kz; 3393 uma_zone_t z; 3394 int count; 3395 3396 count = 0; 3397 rw_rlock(&uma_rwlock); 3398 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3399 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3400 count++; 3401 } 3402 rw_runlock(&uma_rwlock); 3403 return (sysctl_handle_int(oidp, &count, 0, req)); 3404} 3405 3406static int 3407sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3408{ 3409 struct uma_stream_header ush; 3410 struct uma_type_header uth; 3411 struct uma_percpu_stat ups; 3412 uma_bucket_t bucket; 3413 struct sbuf sbuf; 3414 uma_cache_t cache; 3415 uma_klink_t kl; 3416 uma_keg_t kz; 3417 uma_zone_t z; 3418 uma_keg_t k; 3419 int count, error, i; 3420 3421 error = sysctl_wire_old_buffer(req, 0); 3422 if (error != 0) 3423 return (error); 3424 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3425 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); 3426 3427 count = 0; 3428 rw_rlock(&uma_rwlock); 3429 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3430 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3431 count++; 3432 } 3433 3434 /* 3435 * Insert stream header. 3436 */ 3437 bzero(&ush, sizeof(ush)); 3438 ush.ush_version = UMA_STREAM_VERSION; 3439 ush.ush_maxcpus = (mp_maxid + 1); 3440 ush.ush_count = count; 3441 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3442 3443 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3444 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3445 bzero(&uth, sizeof(uth)); 3446 ZONE_LOCK(z); 3447 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3448 uth.uth_align = kz->uk_align; 3449 uth.uth_size = kz->uk_size; 3450 uth.uth_rsize = kz->uk_rsize; 3451 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3452 k = kl->kl_keg; 3453 uth.uth_maxpages += k->uk_maxpages; 3454 uth.uth_pages += k->uk_pages; 3455 uth.uth_keg_free += k->uk_free; 3456 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3457 * k->uk_ipers; 3458 } 3459 3460 /* 3461 * A zone is secondary is it is not the first entry 3462 * on the keg's zone list. 3463 */ 3464 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3465 (LIST_FIRST(&kz->uk_zones) != z)) 3466 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3467 3468 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3469 uth.uth_zone_free += bucket->ub_cnt; 3470 uth.uth_allocs = z->uz_allocs; 3471 uth.uth_frees = z->uz_frees; 3472 uth.uth_fails = z->uz_fails; 3473 uth.uth_sleeps = z->uz_sleeps; 3474 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 3475 /* 3476 * While it is not normally safe to access the cache 3477 * bucket pointers while not on the CPU that owns the 3478 * cache, we only allow the pointers to be exchanged 3479 * without the zone lock held, not invalidated, so 3480 * accept the possible race associated with bucket 3481 * exchange during monitoring. 3482 */ 3483 for (i = 0; i < (mp_maxid + 1); i++) { 3484 bzero(&ups, sizeof(ups)); 3485 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) 3486 goto skip; 3487 if (CPU_ABSENT(i)) 3488 goto skip; 3489 cache = &z->uz_cpu[i]; 3490 if (cache->uc_allocbucket != NULL) 3491 ups.ups_cache_free += 3492 cache->uc_allocbucket->ub_cnt; 3493 if (cache->uc_freebucket != NULL) 3494 ups.ups_cache_free += 3495 cache->uc_freebucket->ub_cnt; 3496 ups.ups_allocs = cache->uc_allocs; 3497 ups.ups_frees = cache->uc_frees; 3498skip: 3499 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups)); 3500 } 3501 ZONE_UNLOCK(z); 3502 } 3503 } 3504 rw_runlock(&uma_rwlock); 3505 error = sbuf_finish(&sbuf); 3506 sbuf_delete(&sbuf); 3507 return (error); 3508} 3509 3510int 3511sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 3512{ 3513 uma_zone_t zone = *(uma_zone_t *)arg1; 3514 int error, max; 3515 3516 max = uma_zone_get_max(zone); 3517 error = sysctl_handle_int(oidp, &max, 0, req); 3518 if (error || !req->newptr) 3519 return (error); 3520 3521 uma_zone_set_max(zone, max); 3522 3523 return (0); 3524} 3525 3526int 3527sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 3528{ 3529 uma_zone_t zone = *(uma_zone_t *)arg1; 3530 int cur; 3531 3532 cur = uma_zone_get_cur(zone); 3533 return (sysctl_handle_int(oidp, &cur, 0, req)); 3534} 3535 3536#ifdef INVARIANTS 3537static uma_slab_t 3538uma_dbg_getslab(uma_zone_t zone, void *item) 3539{ 3540 uma_slab_t slab; 3541 uma_keg_t keg; 3542 uint8_t *mem; 3543 3544 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 3545 if (zone->uz_flags & UMA_ZONE_VTOSLAB) { 3546 slab = vtoslab((vm_offset_t)mem); 3547 } else { 3548 /* 3549 * It is safe to return the slab here even though the 3550 * zone is unlocked because the item's allocation state 3551 * essentially holds a reference. 3552 */ 3553 ZONE_LOCK(zone); 3554 keg = LIST_FIRST(&zone->uz_kegs)->kl_keg; 3555 if (keg->uk_flags & UMA_ZONE_HASH) 3556 slab = hash_sfind(&keg->uk_hash, mem); 3557 else 3558 slab = (uma_slab_t)(mem + keg->uk_pgoff); 3559 ZONE_UNLOCK(zone); 3560 } 3561 3562 return (slab); 3563} 3564 3565/* 3566 * Set up the slab's freei data such that uma_dbg_free can function. 3567 * 3568 */ 3569static void 3570uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) 3571{ 3572 uma_keg_t keg; 3573 int freei; 3574 3575 if (zone_first_keg(zone) == NULL) 3576 return; 3577 if (slab == NULL) { 3578 slab = uma_dbg_getslab(zone, item); 3579 if (slab == NULL) 3580 panic("uma: item %p did not belong to zone %s\n", 3581 item, zone->uz_name); 3582 } 3583 keg = slab->us_keg; 3584 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3585 3586 if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) 3587 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n", 3588 item, zone, zone->uz_name, slab, freei); 3589 BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); 3590 3591 return; 3592} 3593 3594/* 3595 * Verifies freed addresses. Checks for alignment, valid slab membership 3596 * and duplicate frees. 3597 * 3598 */ 3599static void 3600uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) 3601{ 3602 uma_keg_t keg; 3603 int freei; 3604 3605 if (zone_first_keg(zone) == NULL) 3606 return; 3607 if (slab == NULL) { 3608 slab = uma_dbg_getslab(zone, item); 3609 if (slab == NULL) 3610 panic("uma: Freed item %p did not belong to zone %s\n", 3611 item, zone->uz_name); 3612 } 3613 keg = slab->us_keg; 3614 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3615 3616 if (freei >= keg->uk_ipers) 3617 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n", 3618 item, zone, zone->uz_name, slab, freei); 3619 3620 if (((freei * keg->uk_rsize) + slab->us_data) != item) 3621 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n", 3622 item, zone, zone->uz_name, slab, freei); 3623 3624 if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) 3625 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n", 3626 item, zone, zone->uz_name, slab, freei); 3627 3628 BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); 3629} 3630#endif /* INVARIANTS */ 3631 3632#ifdef DDB 3633DB_SHOW_COMMAND(uma, db_show_uma) 3634{ 3635 uint64_t allocs, frees, sleeps; 3636 uma_bucket_t bucket; 3637 uma_keg_t kz; 3638 uma_zone_t z; 3639 int cachefree; 3640 3641 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used", 3642 "Free", "Requests", "Sleeps", "Bucket"); 3643 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3644 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3645 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 3646 allocs = z->uz_allocs; 3647 frees = z->uz_frees; 3648 sleeps = z->uz_sleeps; 3649 cachefree = 0; 3650 } else 3651 uma_zone_sumstat(z, &cachefree, &allocs, 3652 &frees, &sleeps); 3653 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 3654 (LIST_FIRST(&kz->uk_zones) != z))) 3655 cachefree += kz->uk_free; 3656 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3657 cachefree += bucket->ub_cnt; 3658 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n", 3659 z->uz_name, (uintmax_t)kz->uk_size, 3660 (intmax_t)(allocs - frees), cachefree, 3661 (uintmax_t)allocs, sleeps, z->uz_count); 3662 if (db_pager_quit) 3663 return; 3664 } 3665 } 3666} 3667 3668DB_SHOW_COMMAND(umacache, db_show_umacache) 3669{ 3670 uint64_t allocs, frees; 3671 uma_bucket_t bucket; 3672 uma_zone_t z; 3673 int cachefree; 3674 3675 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 3676 "Requests", "Bucket"); 3677 LIST_FOREACH(z, &uma_cachezones, uz_link) { 3678 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL); 3679 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3680 cachefree += bucket->ub_cnt; 3681 db_printf("%18s %8ju %8jd %8d %12ju %8u\n", 3682 z->uz_name, (uintmax_t)z->uz_size, 3683 (intmax_t)(allocs - frees), cachefree, 3684 (uintmax_t)allocs, z->uz_count); 3685 if (db_pager_quit) 3686 return; 3687 } 3688} 3689#endif /* DDB */ 3690