uma_core.c revision 260301
1/*- 2 * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org> 3 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org> 4 * Copyright (c) 2004-2006 Robert N. M. Watson 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice unmodified, this list of conditions, and the following 12 * disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 */ 28 29/* 30 * uma_core.c Implementation of the Universal Memory allocator 31 * 32 * This allocator is intended to replace the multitude of similar object caches 33 * in the standard FreeBSD kernel. The intent is to be flexible as well as 34 * effecient. A primary design goal is to return unused memory to the rest of 35 * the system. This will make the system as a whole more flexible due to the 36 * ability to move memory to subsystems which most need it instead of leaving 37 * pools of reserved memory unused. 38 * 39 * The basic ideas stem from similar slab/zone based allocators whose algorithms 40 * are well known. 41 * 42 */ 43 44/* 45 * TODO: 46 * - Improve memory usage for large allocations 47 * - Investigate cache size adjustments 48 */ 49 50#include <sys/cdefs.h> 51__FBSDID("$FreeBSD: stable/10/sys/vm/uma_core.c 260301 2014-01-04 23:37:01Z mav $"); 52 53/* I should really use ktr.. */ 54/* 55#define UMA_DEBUG 1 56#define UMA_DEBUG_ALLOC 1 57#define UMA_DEBUG_ALLOC_1 1 58*/ 59 60#include "opt_ddb.h" 61#include "opt_param.h" 62#include "opt_vm.h" 63 64#include <sys/param.h> 65#include <sys/systm.h> 66#include <sys/bitset.h> 67#include <sys/kernel.h> 68#include <sys/types.h> 69#include <sys/queue.h> 70#include <sys/malloc.h> 71#include <sys/ktr.h> 72#include <sys/lock.h> 73#include <sys/sysctl.h> 74#include <sys/mutex.h> 75#include <sys/proc.h> 76#include <sys/rwlock.h> 77#include <sys/sbuf.h> 78#include <sys/smp.h> 79#include <sys/vmmeter.h> 80 81#include <vm/vm.h> 82#include <vm/vm_object.h> 83#include <vm/vm_page.h> 84#include <vm/vm_pageout.h> 85#include <vm/vm_param.h> 86#include <vm/vm_map.h> 87#include <vm/vm_kern.h> 88#include <vm/vm_extern.h> 89#include <vm/uma.h> 90#include <vm/uma_int.h> 91#include <vm/uma_dbg.h> 92 93#include <ddb/ddb.h> 94 95#ifdef DEBUG_MEMGUARD 96#include <vm/memguard.h> 97#endif 98 99/* 100 * This is the zone and keg from which all zones are spawned. The idea is that 101 * even the zone & keg heads are allocated from the allocator, so we use the 102 * bss section to bootstrap us. 103 */ 104static struct uma_keg masterkeg; 105static struct uma_zone masterzone_k; 106static struct uma_zone masterzone_z; 107static uma_zone_t kegs = &masterzone_k; 108static uma_zone_t zones = &masterzone_z; 109 110/* This is the zone from which all of uma_slab_t's are allocated. */ 111static uma_zone_t slabzone; 112static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */ 113 114/* 115 * The initial hash tables come out of this zone so they can be allocated 116 * prior to malloc coming up. 117 */ 118static uma_zone_t hashzone; 119 120/* The boot-time adjusted value for cache line alignment. */ 121int uma_align_cache = 64 - 1; 122 123static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 124 125/* 126 * Are we allowed to allocate buckets? 127 */ 128static int bucketdisable = 1; 129 130/* Linked list of all kegs in the system */ 131static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 132 133/* This mutex protects the keg list */ 134static struct mtx_padalign uma_mtx; 135 136/* Linked list of boot time pages */ 137static LIST_HEAD(,uma_slab) uma_boot_pages = 138 LIST_HEAD_INITIALIZER(uma_boot_pages); 139 140/* This mutex protects the boot time pages list */ 141static struct mtx_padalign uma_boot_pages_mtx; 142 143/* Is the VM done starting up? */ 144static int booted = 0; 145#define UMA_STARTUP 1 146#define UMA_STARTUP2 2 147 148/* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */ 149static const u_int uma_max_ipers = SLAB_SETSIZE; 150 151/* 152 * Only mbuf clusters use ref zones. Just provide enough references 153 * to support the one user. New code should not use the ref facility. 154 */ 155static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES; 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(128) 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, NULL, 0} 218}; 219 220/* 221 * Flags and enumerations to be passed to internal functions. 222 */ 223enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; 224 225/* Prototypes.. */ 226 227static void *noobj_alloc(uma_zone_t, int, uint8_t *, int); 228static void *page_alloc(uma_zone_t, int, uint8_t *, int); 229static void *startup_alloc(uma_zone_t, int, uint8_t *, int); 230static void page_free(void *, int, uint8_t); 231static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); 232static void cache_drain(uma_zone_t); 233static void bucket_drain(uma_zone_t, uma_bucket_t); 234static void bucket_cache_drain(uma_zone_t zone); 235static int keg_ctor(void *, int, void *, int); 236static void keg_dtor(void *, int, void *); 237static int zone_ctor(void *, int, void *, int); 238static void zone_dtor(void *, int, void *); 239static int zero_init(void *, int, int); 240static void keg_small_init(uma_keg_t keg); 241static void keg_large_init(uma_keg_t keg); 242static void zone_foreach(void (*zfunc)(uma_zone_t)); 243static void zone_timeout(uma_zone_t zone); 244static int hash_alloc(struct uma_hash *); 245static int hash_expand(struct uma_hash *, struct uma_hash *); 246static void hash_free(struct uma_hash *hash); 247static void uma_timeout(void *); 248static void uma_startup3(void); 249static void *zone_alloc_item(uma_zone_t, void *, int); 250static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 251static void bucket_enable(void); 252static void bucket_init(void); 253static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 254static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 255static void bucket_zone_drain(void); 256static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags); 257static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); 258static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); 259static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 260static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); 261static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 262 uma_fini fini, int align, uint32_t flags); 263static int zone_import(uma_zone_t zone, void **bucket, int max, int flags); 264static void zone_release(uma_zone_t zone, void **bucket, int cnt); 265 266void uma_print_zone(uma_zone_t); 267void uma_print_stats(void); 268static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 269static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 270 271SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 272 273SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 274 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 275 276SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 277 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 278 279static int zone_warnings = 1; 280TUNABLE_INT("vm.zone_warnings", &zone_warnings); 281SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0, 282 "Warn when UMA zones becomes full"); 283 284/* 285 * This routine checks to see whether or not it's safe to enable buckets. 286 */ 287static void 288bucket_enable(void) 289{ 290 bucketdisable = vm_page_count_min(); 291} 292 293/* 294 * Initialize bucket_zones, the array of zones of buckets of various sizes. 295 * 296 * For each zone, calculate the memory required for each bucket, consisting 297 * of the header and an array of pointers. 298 */ 299static void 300bucket_init(void) 301{ 302 struct uma_bucket_zone *ubz; 303 int size; 304 int i; 305 306 for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 307 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 308 size += sizeof(void *) * ubz->ubz_entries; 309 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 310 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 311 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET); 312 } 313} 314 315/* 316 * Given a desired number of entries for a bucket, return the zone from which 317 * to allocate the bucket. 318 */ 319static struct uma_bucket_zone * 320bucket_zone_lookup(int entries) 321{ 322 struct uma_bucket_zone *ubz; 323 324 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 325 if (ubz->ubz_entries >= entries) 326 return (ubz); 327 ubz--; 328 return (ubz); 329} 330 331static int 332bucket_select(int size) 333{ 334 struct uma_bucket_zone *ubz; 335 336 ubz = &bucket_zones[0]; 337 if (size > ubz->ubz_maxsize) 338 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 339 340 for (; ubz->ubz_entries != 0; ubz++) 341 if (ubz->ubz_maxsize < size) 342 break; 343 ubz--; 344 return (ubz->ubz_entries); 345} 346 347static uma_bucket_t 348bucket_alloc(uma_zone_t zone, void *udata, int flags) 349{ 350 struct uma_bucket_zone *ubz; 351 uma_bucket_t bucket; 352 353 /* 354 * This is to stop us from allocating per cpu buckets while we're 355 * running out of vm.boot_pages. Otherwise, we would exhaust the 356 * boot pages. This also prevents us from allocating buckets in 357 * low memory situations. 358 */ 359 if (bucketdisable) 360 return (NULL); 361 /* 362 * To limit bucket recursion we store the original zone flags 363 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 364 * NOVM flag to persist even through deep recursions. We also 365 * store ZFLAG_BUCKET once we have recursed attempting to allocate 366 * a bucket for a bucket zone so we do not allow infinite bucket 367 * recursion. This cookie will even persist to frees of unused 368 * buckets via the allocation path or bucket allocations in the 369 * free path. 370 */ 371 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 372 udata = (void *)(uintptr_t)zone->uz_flags; 373 else { 374 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 375 return (NULL); 376 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 377 } 378 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 379 flags |= M_NOVM; 380 ubz = bucket_zone_lookup(zone->uz_count); 381 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 382 if (bucket) { 383#ifdef INVARIANTS 384 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 385#endif 386 bucket->ub_cnt = 0; 387 bucket->ub_entries = ubz->ubz_entries; 388 } 389 390 return (bucket); 391} 392 393static void 394bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 395{ 396 struct uma_bucket_zone *ubz; 397 398 KASSERT(bucket->ub_cnt == 0, 399 ("bucket_free: Freeing a non free bucket.")); 400 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 401 udata = (void *)(uintptr_t)zone->uz_flags; 402 ubz = bucket_zone_lookup(bucket->ub_entries); 403 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 404} 405 406static void 407bucket_zone_drain(void) 408{ 409 struct uma_bucket_zone *ubz; 410 411 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 412 zone_drain(ubz->ubz_zone); 413} 414 415static void 416zone_log_warning(uma_zone_t zone) 417{ 418 static const struct timeval warninterval = { 300, 0 }; 419 420 if (!zone_warnings || zone->uz_warning == NULL) 421 return; 422 423 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 424 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 425} 426 427static void 428zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 429{ 430 uma_klink_t klink; 431 432 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 433 kegfn(klink->kl_keg); 434} 435 436/* 437 * Routine called by timeout which is used to fire off some time interval 438 * based calculations. (stats, hash size, etc.) 439 * 440 * Arguments: 441 * arg Unused 442 * 443 * Returns: 444 * Nothing 445 */ 446static void 447uma_timeout(void *unused) 448{ 449 bucket_enable(); 450 zone_foreach(zone_timeout); 451 452 /* Reschedule this event */ 453 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 454} 455 456/* 457 * Routine to perform timeout driven calculations. This expands the 458 * hashes and does per cpu statistics aggregation. 459 * 460 * Returns nothing. 461 */ 462static void 463keg_timeout(uma_keg_t keg) 464{ 465 466 KEG_LOCK(keg); 467 /* 468 * Expand the keg hash table. 469 * 470 * This is done if the number of slabs is larger than the hash size. 471 * What I'm trying to do here is completely reduce collisions. This 472 * may be a little aggressive. Should I allow for two collisions max? 473 */ 474 if (keg->uk_flags & UMA_ZONE_HASH && 475 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 476 struct uma_hash newhash; 477 struct uma_hash oldhash; 478 int ret; 479 480 /* 481 * This is so involved because allocating and freeing 482 * while the keg lock is held will lead to deadlock. 483 * I have to do everything in stages and check for 484 * races. 485 */ 486 newhash = keg->uk_hash; 487 KEG_UNLOCK(keg); 488 ret = hash_alloc(&newhash); 489 KEG_LOCK(keg); 490 if (ret) { 491 if (hash_expand(&keg->uk_hash, &newhash)) { 492 oldhash = keg->uk_hash; 493 keg->uk_hash = newhash; 494 } else 495 oldhash = newhash; 496 497 KEG_UNLOCK(keg); 498 hash_free(&oldhash); 499 return; 500 } 501 } 502 KEG_UNLOCK(keg); 503} 504 505static void 506zone_timeout(uma_zone_t zone) 507{ 508 509 zone_foreach_keg(zone, &keg_timeout); 510} 511 512/* 513 * Allocate and zero fill the next sized hash table from the appropriate 514 * backing store. 515 * 516 * Arguments: 517 * hash A new hash structure with the old hash size in uh_hashsize 518 * 519 * Returns: 520 * 1 on sucess and 0 on failure. 521 */ 522static int 523hash_alloc(struct uma_hash *hash) 524{ 525 int oldsize; 526 int alloc; 527 528 oldsize = hash->uh_hashsize; 529 530 /* We're just going to go to a power of two greater */ 531 if (oldsize) { 532 hash->uh_hashsize = oldsize * 2; 533 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 534 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 535 M_UMAHASH, M_NOWAIT); 536 } else { 537 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 538 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 539 M_WAITOK); 540 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 541 } 542 if (hash->uh_slab_hash) { 543 bzero(hash->uh_slab_hash, alloc); 544 hash->uh_hashmask = hash->uh_hashsize - 1; 545 return (1); 546 } 547 548 return (0); 549} 550 551/* 552 * Expands the hash table for HASH zones. This is done from zone_timeout 553 * to reduce collisions. This must not be done in the regular allocation 554 * path, otherwise, we can recurse on the vm while allocating pages. 555 * 556 * Arguments: 557 * oldhash The hash you want to expand 558 * newhash The hash structure for the new table 559 * 560 * Returns: 561 * Nothing 562 * 563 * Discussion: 564 */ 565static int 566hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 567{ 568 uma_slab_t slab; 569 int hval; 570 int i; 571 572 if (!newhash->uh_slab_hash) 573 return (0); 574 575 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 576 return (0); 577 578 /* 579 * I need to investigate hash algorithms for resizing without a 580 * full rehash. 581 */ 582 583 for (i = 0; i < oldhash->uh_hashsize; i++) 584 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 585 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 586 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 587 hval = UMA_HASH(newhash, slab->us_data); 588 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 589 slab, us_hlink); 590 } 591 592 return (1); 593} 594 595/* 596 * Free the hash bucket to the appropriate backing store. 597 * 598 * Arguments: 599 * slab_hash The hash bucket we're freeing 600 * hashsize The number of entries in that hash bucket 601 * 602 * Returns: 603 * Nothing 604 */ 605static void 606hash_free(struct uma_hash *hash) 607{ 608 if (hash->uh_slab_hash == NULL) 609 return; 610 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 611 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 612 else 613 free(hash->uh_slab_hash, M_UMAHASH); 614} 615 616/* 617 * Frees all outstanding items in a bucket 618 * 619 * Arguments: 620 * zone The zone to free to, must be unlocked. 621 * bucket The free/alloc bucket with items, cpu queue must be locked. 622 * 623 * Returns: 624 * Nothing 625 */ 626 627static void 628bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 629{ 630 int i; 631 632 if (bucket == NULL) 633 return; 634 635 if (zone->uz_fini) 636 for (i = 0; i < bucket->ub_cnt; i++) 637 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 638 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 639 bucket->ub_cnt = 0; 640} 641 642/* 643 * Drains the per cpu caches for a zone. 644 * 645 * NOTE: This may only be called while the zone is being turn down, and not 646 * during normal operation. This is necessary in order that we do not have 647 * to migrate CPUs to drain the per-CPU caches. 648 * 649 * Arguments: 650 * zone The zone to drain, must be unlocked. 651 * 652 * Returns: 653 * Nothing 654 */ 655static void 656cache_drain(uma_zone_t zone) 657{ 658 uma_cache_t cache; 659 int cpu; 660 661 /* 662 * XXX: It is safe to not lock the per-CPU caches, because we're 663 * tearing down the zone anyway. I.e., there will be no further use 664 * of the caches at this point. 665 * 666 * XXX: It would good to be able to assert that the zone is being 667 * torn down to prevent improper use of cache_drain(). 668 * 669 * XXX: We lock the zone before passing into bucket_cache_drain() as 670 * it is used elsewhere. Should the tear-down path be made special 671 * there in some form? 672 */ 673 CPU_FOREACH(cpu) { 674 cache = &zone->uz_cpu[cpu]; 675 bucket_drain(zone, cache->uc_allocbucket); 676 bucket_drain(zone, cache->uc_freebucket); 677 if (cache->uc_allocbucket != NULL) 678 bucket_free(zone, cache->uc_allocbucket, NULL); 679 if (cache->uc_freebucket != NULL) 680 bucket_free(zone, cache->uc_freebucket, NULL); 681 cache->uc_allocbucket = cache->uc_freebucket = NULL; 682 } 683 ZONE_LOCK(zone); 684 bucket_cache_drain(zone); 685 ZONE_UNLOCK(zone); 686} 687 688/* 689 * Drain the cached buckets from a zone. Expects a locked zone on entry. 690 */ 691static void 692bucket_cache_drain(uma_zone_t zone) 693{ 694 uma_bucket_t bucket; 695 696 /* 697 * Drain the bucket queues and free the buckets, we just keep two per 698 * cpu (alloc/free). 699 */ 700 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 701 LIST_REMOVE(bucket, ub_link); 702 ZONE_UNLOCK(zone); 703 bucket_drain(zone, bucket); 704 bucket_free(zone, bucket, NULL); 705 ZONE_LOCK(zone); 706 } 707 708 /* 709 * Shrink further bucket sizes. Price of single zone lock collision 710 * is probably lower then price of global cache drain. 711 */ 712 if (zone->uz_count > zone->uz_count_min) 713 zone->uz_count--; 714} 715 716static void 717keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 718{ 719 uint8_t *mem; 720 int i; 721 uint8_t flags; 722 723 mem = slab->us_data; 724 flags = slab->us_flags; 725 i = start; 726 if (keg->uk_fini != NULL) { 727 for (i--; i > -1; i--) 728 keg->uk_fini(slab->us_data + (keg->uk_rsize * i), 729 keg->uk_size); 730 } 731 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 732 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 733#ifdef UMA_DEBUG 734 printf("%s: Returning %d bytes.\n", keg->uk_name, 735 PAGE_SIZE * keg->uk_ppera); 736#endif 737 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 738} 739 740/* 741 * Frees pages from a keg back to the system. This is done on demand from 742 * the pageout daemon. 743 * 744 * Returns nothing. 745 */ 746static void 747keg_drain(uma_keg_t keg) 748{ 749 struct slabhead freeslabs = { 0 }; 750 uma_slab_t slab; 751 uma_slab_t n; 752 753 /* 754 * We don't want to take pages from statically allocated kegs at this 755 * time 756 */ 757 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 758 return; 759 760#ifdef UMA_DEBUG 761 printf("%s free items: %u\n", keg->uk_name, keg->uk_free); 762#endif 763 KEG_LOCK(keg); 764 if (keg->uk_free == 0) 765 goto finished; 766 767 slab = LIST_FIRST(&keg->uk_free_slab); 768 while (slab) { 769 n = LIST_NEXT(slab, us_link); 770 771 /* We have no where to free these to */ 772 if (slab->us_flags & UMA_SLAB_BOOT) { 773 slab = n; 774 continue; 775 } 776 777 LIST_REMOVE(slab, us_link); 778 keg->uk_pages -= keg->uk_ppera; 779 keg->uk_free -= keg->uk_ipers; 780 781 if (keg->uk_flags & UMA_ZONE_HASH) 782 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); 783 784 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 785 786 slab = n; 787 } 788finished: 789 KEG_UNLOCK(keg); 790 791 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 792 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 793 keg_free_slab(keg, slab, keg->uk_ipers); 794 } 795} 796 797static void 798zone_drain_wait(uma_zone_t zone, int waitok) 799{ 800 801 /* 802 * Set draining to interlock with zone_dtor() so we can release our 803 * locks as we go. Only dtor() should do a WAITOK call since it 804 * is the only call that knows the structure will still be available 805 * when it wakes up. 806 */ 807 ZONE_LOCK(zone); 808 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 809 if (waitok == M_NOWAIT) 810 goto out; 811 mtx_unlock(&uma_mtx); 812 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 813 mtx_lock(&uma_mtx); 814 } 815 zone->uz_flags |= UMA_ZFLAG_DRAINING; 816 bucket_cache_drain(zone); 817 ZONE_UNLOCK(zone); 818 /* 819 * The DRAINING flag protects us from being freed while 820 * we're running. Normally the uma_mtx would protect us but we 821 * must be able to release and acquire the right lock for each keg. 822 */ 823 zone_foreach_keg(zone, &keg_drain); 824 ZONE_LOCK(zone); 825 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 826 wakeup(zone); 827out: 828 ZONE_UNLOCK(zone); 829} 830 831void 832zone_drain(uma_zone_t zone) 833{ 834 835 zone_drain_wait(zone, M_NOWAIT); 836} 837 838/* 839 * Allocate a new slab for a keg. This does not insert the slab onto a list. 840 * 841 * Arguments: 842 * wait Shall we wait? 843 * 844 * Returns: 845 * The slab that was allocated or NULL if there is no memory and the 846 * caller specified M_NOWAIT. 847 */ 848static uma_slab_t 849keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) 850{ 851 uma_slabrefcnt_t slabref; 852 uma_alloc allocf; 853 uma_slab_t slab; 854 uint8_t *mem; 855 uint8_t flags; 856 int i; 857 858 mtx_assert(&keg->uk_lock, MA_OWNED); 859 slab = NULL; 860 mem = NULL; 861 862#ifdef UMA_DEBUG 863 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name); 864#endif 865 allocf = keg->uk_allocf; 866 KEG_UNLOCK(keg); 867 868 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 869 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); 870 if (slab == NULL) 871 goto out; 872 } 873 874 /* 875 * This reproduces the old vm_zone behavior of zero filling pages the 876 * first time they are added to a zone. 877 * 878 * Malloced items are zeroed in uma_zalloc. 879 */ 880 881 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 882 wait |= M_ZERO; 883 else 884 wait &= ~M_ZERO; 885 886 if (keg->uk_flags & UMA_ZONE_NODUMP) 887 wait |= M_NODUMP; 888 889 /* zone is passed for legacy reasons. */ 890 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait); 891 if (mem == NULL) { 892 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 893 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 894 slab = NULL; 895 goto out; 896 } 897 898 /* Point the slab into the allocated memory */ 899 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 900 slab = (uma_slab_t )(mem + keg->uk_pgoff); 901 902 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 903 for (i = 0; i < keg->uk_ppera; i++) 904 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 905 906 slab->us_keg = keg; 907 slab->us_data = mem; 908 slab->us_freecount = keg->uk_ipers; 909 slab->us_flags = flags; 910 BIT_FILL(SLAB_SETSIZE, &slab->us_free); 911#ifdef INVARIANTS 912 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); 913#endif 914 if (keg->uk_flags & UMA_ZONE_REFCNT) { 915 slabref = (uma_slabrefcnt_t)slab; 916 for (i = 0; i < keg->uk_ipers; i++) 917 slabref->us_refcnt[i] = 0; 918 } 919 920 if (keg->uk_init != NULL) { 921 for (i = 0; i < keg->uk_ipers; i++) 922 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 923 keg->uk_size, wait) != 0) 924 break; 925 if (i != keg->uk_ipers) { 926 keg_free_slab(keg, slab, i); 927 slab = NULL; 928 goto out; 929 } 930 } 931out: 932 KEG_LOCK(keg); 933 934 if (slab != NULL) { 935 if (keg->uk_flags & UMA_ZONE_HASH) 936 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 937 938 keg->uk_pages += keg->uk_ppera; 939 keg->uk_free += keg->uk_ipers; 940 } 941 942 return (slab); 943} 944 945/* 946 * This function is intended to be used early on in place of page_alloc() so 947 * that we may use the boot time page cache to satisfy allocations before 948 * the VM is ready. 949 */ 950static void * 951startup_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait) 952{ 953 uma_keg_t keg; 954 uma_slab_t tmps; 955 int pages, check_pages; 956 957 keg = zone_first_keg(zone); 958 pages = howmany(bytes, PAGE_SIZE); 959 check_pages = pages - 1; 960 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n")); 961 962 /* 963 * Check our small startup cache to see if it has pages remaining. 964 */ 965 mtx_lock(&uma_boot_pages_mtx); 966 967 /* First check if we have enough room. */ 968 tmps = LIST_FIRST(&uma_boot_pages); 969 while (tmps != NULL && check_pages-- > 0) 970 tmps = LIST_NEXT(tmps, us_link); 971 if (tmps != NULL) { 972 /* 973 * It's ok to lose tmps references. The last one will 974 * have tmps->us_data pointing to the start address of 975 * "pages" contiguous pages of memory. 976 */ 977 while (pages-- > 0) { 978 tmps = LIST_FIRST(&uma_boot_pages); 979 LIST_REMOVE(tmps, us_link); 980 } 981 mtx_unlock(&uma_boot_pages_mtx); 982 *pflag = tmps->us_flags; 983 return (tmps->us_data); 984 } 985 mtx_unlock(&uma_boot_pages_mtx); 986 if (booted < UMA_STARTUP2) 987 panic("UMA: Increase vm.boot_pages"); 988 /* 989 * Now that we've booted reset these users to their real allocator. 990 */ 991#ifdef UMA_MD_SMALL_ALLOC 992 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; 993#else 994 keg->uk_allocf = page_alloc; 995#endif 996 return keg->uk_allocf(zone, bytes, pflag, wait); 997} 998 999/* 1000 * Allocates a number of pages from the system 1001 * 1002 * Arguments: 1003 * bytes The number of bytes requested 1004 * wait Shall we wait? 1005 * 1006 * Returns: 1007 * A pointer to the alloced memory or possibly 1008 * NULL if M_NOWAIT is set. 1009 */ 1010static void * 1011page_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait) 1012{ 1013 void *p; /* Returned page */ 1014 1015 *pflag = UMA_SLAB_KMEM; 1016 p = (void *) kmem_malloc(kmem_arena, bytes, wait); 1017 1018 return (p); 1019} 1020 1021/* 1022 * Allocates a number of pages from within an object 1023 * 1024 * Arguments: 1025 * bytes The number of bytes requested 1026 * wait Shall we wait? 1027 * 1028 * Returns: 1029 * A pointer to the alloced memory or possibly 1030 * NULL if M_NOWAIT is set. 1031 */ 1032static void * 1033noobj_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait) 1034{ 1035 TAILQ_HEAD(, vm_page) alloctail; 1036 u_long npages; 1037 vm_offset_t retkva, zkva; 1038 vm_page_t p, p_next; 1039 uma_keg_t keg; 1040 1041 TAILQ_INIT(&alloctail); 1042 keg = zone_first_keg(zone); 1043 1044 npages = howmany(bytes, PAGE_SIZE); 1045 while (npages > 0) { 1046 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT | 1047 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ); 1048 if (p != NULL) { 1049 /* 1050 * Since the page does not belong to an object, its 1051 * listq is unused. 1052 */ 1053 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1054 npages--; 1055 continue; 1056 } 1057 if (wait & M_WAITOK) { 1058 VM_WAIT; 1059 continue; 1060 } 1061 1062 /* 1063 * Page allocation failed, free intermediate pages and 1064 * exit. 1065 */ 1066 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1067 vm_page_unwire(p, 0); 1068 vm_page_free(p); 1069 } 1070 return (NULL); 1071 } 1072 *flags = UMA_SLAB_PRIV; 1073 zkva = keg->uk_kva + 1074 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1075 retkva = zkva; 1076 TAILQ_FOREACH(p, &alloctail, listq) { 1077 pmap_qenter(zkva, &p, 1); 1078 zkva += PAGE_SIZE; 1079 } 1080 1081 return ((void *)retkva); 1082} 1083 1084/* 1085 * Frees a number of pages to the system 1086 * 1087 * Arguments: 1088 * mem A pointer to the memory to be freed 1089 * size The size of the memory being freed 1090 * flags The original p->us_flags field 1091 * 1092 * Returns: 1093 * Nothing 1094 */ 1095static void 1096page_free(void *mem, int size, uint8_t flags) 1097{ 1098 struct vmem *vmem; 1099 1100 if (flags & UMA_SLAB_KMEM) 1101 vmem = kmem_arena; 1102 else if (flags & UMA_SLAB_KERNEL) 1103 vmem = kernel_arena; 1104 else 1105 panic("UMA: page_free used with invalid flags %d", flags); 1106 1107 kmem_free(vmem, (vm_offset_t)mem, size); 1108} 1109 1110/* 1111 * Zero fill initializer 1112 * 1113 * Arguments/Returns follow uma_init specifications 1114 */ 1115static int 1116zero_init(void *mem, int size, int flags) 1117{ 1118 bzero(mem, size); 1119 return (0); 1120} 1121 1122/* 1123 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1124 * 1125 * Arguments 1126 * keg The zone we should initialize 1127 * 1128 * Returns 1129 * Nothing 1130 */ 1131static void 1132keg_small_init(uma_keg_t keg) 1133{ 1134 u_int rsize; 1135 u_int memused; 1136 u_int wastedspace; 1137 u_int shsize; 1138 1139 if (keg->uk_flags & UMA_ZONE_PCPU) { 1140 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU; 1141 1142 keg->uk_slabsize = sizeof(struct pcpu); 1143 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu), 1144 PAGE_SIZE); 1145 } else { 1146 keg->uk_slabsize = UMA_SLAB_SIZE; 1147 keg->uk_ppera = 1; 1148 } 1149 1150 /* 1151 * Calculate the size of each allocation (rsize) according to 1152 * alignment. If the requested size is smaller than we have 1153 * allocation bits for we round it up. 1154 */ 1155 rsize = keg->uk_size; 1156 if (rsize < keg->uk_slabsize / SLAB_SETSIZE) 1157 rsize = keg->uk_slabsize / SLAB_SETSIZE; 1158 if (rsize & keg->uk_align) 1159 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1160 keg->uk_rsize = rsize; 1161 1162 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1163 keg->uk_rsize < sizeof(struct pcpu), 1164 ("%s: size %u too large", __func__, keg->uk_rsize)); 1165 1166 if (keg->uk_flags & UMA_ZONE_REFCNT) 1167 rsize += sizeof(uint32_t); 1168 1169 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1170 shsize = 0; 1171 else 1172 shsize = sizeof(struct uma_slab); 1173 1174 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize; 1175 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1176 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1177 1178 memused = keg->uk_ipers * rsize + shsize; 1179 wastedspace = keg->uk_slabsize - memused; 1180 1181 /* 1182 * We can't do OFFPAGE if we're internal or if we've been 1183 * asked to not go to the VM for buckets. If we do this we 1184 * may end up going to the VM for slabs which we do not 1185 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1186 * of UMA_ZONE_VM, which clearly forbids it. 1187 */ 1188 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1189 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1190 return; 1191 1192 /* 1193 * See if using an OFFPAGE slab will limit our waste. Only do 1194 * this if it permits more items per-slab. 1195 * 1196 * XXX We could try growing slabsize to limit max waste as well. 1197 * Historically this was not done because the VM could not 1198 * efficiently handle contiguous allocations. 1199 */ 1200 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) && 1201 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) { 1202 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize; 1203 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1204 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1205#ifdef UMA_DEBUG 1206 printf("UMA decided we need offpage slab headers for " 1207 "keg: %s, calculated wastedspace = %d, " 1208 "maximum wasted space allowed = %d, " 1209 "calculated ipers = %d, " 1210 "new wasted space = %d\n", keg->uk_name, wastedspace, 1211 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1212 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize); 1213#endif 1214 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1215 } 1216 1217 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1218 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1219 keg->uk_flags |= UMA_ZONE_HASH; 1220} 1221 1222/* 1223 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1224 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1225 * more complicated. 1226 * 1227 * Arguments 1228 * keg The keg we should initialize 1229 * 1230 * Returns 1231 * Nothing 1232 */ 1233static void 1234keg_large_init(uma_keg_t keg) 1235{ 1236 1237 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1238 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1239 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1240 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1241 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1242 1243 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1244 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE; 1245 keg->uk_ipers = 1; 1246 keg->uk_rsize = keg->uk_size; 1247 1248 /* We can't do OFFPAGE if we're internal, bail out here. */ 1249 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) 1250 return; 1251 1252 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1253 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1254 keg->uk_flags |= UMA_ZONE_HASH; 1255} 1256 1257static void 1258keg_cachespread_init(uma_keg_t keg) 1259{ 1260 int alignsize; 1261 int trailer; 1262 int pages; 1263 int rsize; 1264 1265 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1266 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1267 1268 alignsize = keg->uk_align + 1; 1269 rsize = keg->uk_size; 1270 /* 1271 * We want one item to start on every align boundary in a page. To 1272 * do this we will span pages. We will also extend the item by the 1273 * size of align if it is an even multiple of align. Otherwise, it 1274 * would fall on the same boundary every time. 1275 */ 1276 if (rsize & keg->uk_align) 1277 rsize = (rsize & ~keg->uk_align) + alignsize; 1278 if ((rsize & alignsize) == 0) 1279 rsize += alignsize; 1280 trailer = rsize - keg->uk_size; 1281 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1282 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1283 keg->uk_rsize = rsize; 1284 keg->uk_ppera = pages; 1285 keg->uk_slabsize = UMA_SLAB_SIZE; 1286 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1287 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1288 KASSERT(keg->uk_ipers <= uma_max_ipers, 1289 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1290 keg->uk_ipers)); 1291} 1292 1293/* 1294 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1295 * the keg onto the global keg list. 1296 * 1297 * Arguments/Returns follow uma_ctor specifications 1298 * udata Actually uma_kctor_args 1299 */ 1300static int 1301keg_ctor(void *mem, int size, void *udata, int flags) 1302{ 1303 struct uma_kctor_args *arg = udata; 1304 uma_keg_t keg = mem; 1305 uma_zone_t zone; 1306 1307 bzero(keg, size); 1308 keg->uk_size = arg->size; 1309 keg->uk_init = arg->uminit; 1310 keg->uk_fini = arg->fini; 1311 keg->uk_align = arg->align; 1312 keg->uk_free = 0; 1313 keg->uk_reserve = 0; 1314 keg->uk_pages = 0; 1315 keg->uk_flags = arg->flags; 1316 keg->uk_allocf = page_alloc; 1317 keg->uk_freef = page_free; 1318 keg->uk_slabzone = NULL; 1319 1320 /* 1321 * The master zone is passed to us at keg-creation time. 1322 */ 1323 zone = arg->zone; 1324 keg->uk_name = zone->uz_name; 1325 1326 if (arg->flags & UMA_ZONE_VM) 1327 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1328 1329 if (arg->flags & UMA_ZONE_ZINIT) 1330 keg->uk_init = zero_init; 1331 1332 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC) 1333 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1334 1335 if (arg->flags & UMA_ZONE_PCPU) 1336#ifdef SMP 1337 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1338#else 1339 keg->uk_flags &= ~UMA_ZONE_PCPU; 1340#endif 1341 1342 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1343 keg_cachespread_init(keg); 1344 } else if (keg->uk_flags & UMA_ZONE_REFCNT) { 1345 if (keg->uk_size > 1346 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - 1347 sizeof(uint32_t))) 1348 keg_large_init(keg); 1349 else 1350 keg_small_init(keg); 1351 } else { 1352 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab))) 1353 keg_large_init(keg); 1354 else 1355 keg_small_init(keg); 1356 } 1357 1358 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1359 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1360 if (keg->uk_ipers > uma_max_ipers_ref) 1361 panic("Too many ref items per zone: %d > %d\n", 1362 keg->uk_ipers, uma_max_ipers_ref); 1363 keg->uk_slabzone = slabrefzone; 1364 } else 1365 keg->uk_slabzone = slabzone; 1366 } 1367 1368 /* 1369 * If we haven't booted yet we need allocations to go through the 1370 * startup cache until the vm is ready. 1371 */ 1372 if (keg->uk_ppera == 1) { 1373#ifdef UMA_MD_SMALL_ALLOC 1374 keg->uk_allocf = uma_small_alloc; 1375 keg->uk_freef = uma_small_free; 1376 1377 if (booted < UMA_STARTUP) 1378 keg->uk_allocf = startup_alloc; 1379#else 1380 if (booted < UMA_STARTUP2) 1381 keg->uk_allocf = startup_alloc; 1382#endif 1383 } else if (booted < UMA_STARTUP2 && 1384 (keg->uk_flags & UMA_ZFLAG_INTERNAL)) 1385 keg->uk_allocf = startup_alloc; 1386 1387 /* 1388 * Initialize keg's lock 1389 */ 1390 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1391 1392 /* 1393 * If we're putting the slab header in the actual page we need to 1394 * figure out where in each page it goes. This calculates a right 1395 * justified offset into the memory on an ALIGN_PTR boundary. 1396 */ 1397 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1398 u_int totsize; 1399 1400 /* Size of the slab struct and free list */ 1401 totsize = sizeof(struct uma_slab); 1402 1403 /* Size of the reference counts. */ 1404 if (keg->uk_flags & UMA_ZONE_REFCNT) 1405 totsize += keg->uk_ipers * sizeof(uint32_t); 1406 1407 if (totsize & UMA_ALIGN_PTR) 1408 totsize = (totsize & ~UMA_ALIGN_PTR) + 1409 (UMA_ALIGN_PTR + 1); 1410 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1411 1412 /* 1413 * The only way the following is possible is if with our 1414 * UMA_ALIGN_PTR adjustments we are now bigger than 1415 * UMA_SLAB_SIZE. I haven't checked whether this is 1416 * mathematically possible for all cases, so we make 1417 * sure here anyway. 1418 */ 1419 totsize = keg->uk_pgoff + sizeof(struct uma_slab); 1420 if (keg->uk_flags & UMA_ZONE_REFCNT) 1421 totsize += keg->uk_ipers * sizeof(uint32_t); 1422 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1423 printf("zone %s ipers %d rsize %d size %d\n", 1424 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1425 keg->uk_size); 1426 panic("UMA slab won't fit."); 1427 } 1428 } 1429 1430 if (keg->uk_flags & UMA_ZONE_HASH) 1431 hash_alloc(&keg->uk_hash); 1432 1433#ifdef UMA_DEBUG 1434 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n", 1435 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags, 1436 keg->uk_ipers, keg->uk_ppera, 1437 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free); 1438#endif 1439 1440 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1441 1442 mtx_lock(&uma_mtx); 1443 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1444 mtx_unlock(&uma_mtx); 1445 return (0); 1446} 1447 1448/* 1449 * Zone header ctor. This initializes all fields, locks, etc. 1450 * 1451 * Arguments/Returns follow uma_ctor specifications 1452 * udata Actually uma_zctor_args 1453 */ 1454static int 1455zone_ctor(void *mem, int size, void *udata, int flags) 1456{ 1457 struct uma_zctor_args *arg = udata; 1458 uma_zone_t zone = mem; 1459 uma_zone_t z; 1460 uma_keg_t keg; 1461 1462 bzero(zone, size); 1463 zone->uz_name = arg->name; 1464 zone->uz_ctor = arg->ctor; 1465 zone->uz_dtor = arg->dtor; 1466 zone->uz_slab = zone_fetch_slab; 1467 zone->uz_init = NULL; 1468 zone->uz_fini = NULL; 1469 zone->uz_allocs = 0; 1470 zone->uz_frees = 0; 1471 zone->uz_fails = 0; 1472 zone->uz_sleeps = 0; 1473 zone->uz_count = 0; 1474 zone->uz_count_min = 0; 1475 zone->uz_flags = 0; 1476 zone->uz_warning = NULL; 1477 timevalclear(&zone->uz_ratecheck); 1478 keg = arg->keg; 1479 1480 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 1481 1482 /* 1483 * This is a pure cache zone, no kegs. 1484 */ 1485 if (arg->import) { 1486 if (arg->flags & UMA_ZONE_VM) 1487 arg->flags |= UMA_ZFLAG_CACHEONLY; 1488 zone->uz_flags = arg->flags; 1489 zone->uz_size = arg->size; 1490 zone->uz_import = arg->import; 1491 zone->uz_release = arg->release; 1492 zone->uz_arg = arg->arg; 1493 zone->uz_lockptr = &zone->uz_lock; 1494 goto out; 1495 } 1496 1497 /* 1498 * Use the regular zone/keg/slab allocator. 1499 */ 1500 zone->uz_import = (uma_import)zone_import; 1501 zone->uz_release = (uma_release)zone_release; 1502 zone->uz_arg = zone; 1503 1504 if (arg->flags & UMA_ZONE_SECONDARY) { 1505 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1506 zone->uz_init = arg->uminit; 1507 zone->uz_fini = arg->fini; 1508 zone->uz_lockptr = &keg->uk_lock; 1509 zone->uz_flags |= UMA_ZONE_SECONDARY; 1510 mtx_lock(&uma_mtx); 1511 ZONE_LOCK(zone); 1512 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1513 if (LIST_NEXT(z, uz_link) == NULL) { 1514 LIST_INSERT_AFTER(z, zone, uz_link); 1515 break; 1516 } 1517 } 1518 ZONE_UNLOCK(zone); 1519 mtx_unlock(&uma_mtx); 1520 } else if (keg == NULL) { 1521 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1522 arg->align, arg->flags)) == NULL) 1523 return (ENOMEM); 1524 } else { 1525 struct uma_kctor_args karg; 1526 int error; 1527 1528 /* We should only be here from uma_startup() */ 1529 karg.size = arg->size; 1530 karg.uminit = arg->uminit; 1531 karg.fini = arg->fini; 1532 karg.align = arg->align; 1533 karg.flags = arg->flags; 1534 karg.zone = zone; 1535 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1536 flags); 1537 if (error) 1538 return (error); 1539 } 1540 1541 /* 1542 * Link in the first keg. 1543 */ 1544 zone->uz_klink.kl_keg = keg; 1545 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1546 zone->uz_lockptr = &keg->uk_lock; 1547 zone->uz_size = keg->uk_size; 1548 zone->uz_flags |= (keg->uk_flags & 1549 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1550 1551 /* 1552 * Some internal zones don't have room allocated for the per cpu 1553 * caches. If we're internal, bail out here. 1554 */ 1555 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1556 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1557 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1558 return (0); 1559 } 1560 1561out: 1562 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0) 1563 zone->uz_count = bucket_select(zone->uz_size); 1564 else 1565 zone->uz_count = BUCKET_MAX; 1566 zone->uz_count_min = zone->uz_count; 1567 1568 return (0); 1569} 1570 1571/* 1572 * Keg header dtor. This frees all data, destroys locks, frees the hash 1573 * table and removes the keg from the global list. 1574 * 1575 * Arguments/Returns follow uma_dtor specifications 1576 * udata unused 1577 */ 1578static void 1579keg_dtor(void *arg, int size, void *udata) 1580{ 1581 uma_keg_t keg; 1582 1583 keg = (uma_keg_t)arg; 1584 KEG_LOCK(keg); 1585 if (keg->uk_free != 0) { 1586 printf("Freed UMA keg (%s) was not empty (%d items). " 1587 " Lost %d pages of memory.\n", 1588 keg->uk_name ? keg->uk_name : "", 1589 keg->uk_free, keg->uk_pages); 1590 } 1591 KEG_UNLOCK(keg); 1592 1593 hash_free(&keg->uk_hash); 1594 1595 KEG_LOCK_FINI(keg); 1596} 1597 1598/* 1599 * Zone header dtor. 1600 * 1601 * Arguments/Returns follow uma_dtor specifications 1602 * udata unused 1603 */ 1604static void 1605zone_dtor(void *arg, int size, void *udata) 1606{ 1607 uma_klink_t klink; 1608 uma_zone_t zone; 1609 uma_keg_t keg; 1610 1611 zone = (uma_zone_t)arg; 1612 keg = zone_first_keg(zone); 1613 1614 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1615 cache_drain(zone); 1616 1617 mtx_lock(&uma_mtx); 1618 LIST_REMOVE(zone, uz_link); 1619 mtx_unlock(&uma_mtx); 1620 /* 1621 * XXX there are some races here where 1622 * the zone can be drained but zone lock 1623 * released and then refilled before we 1624 * remove it... we dont care for now 1625 */ 1626 zone_drain_wait(zone, M_WAITOK); 1627 /* 1628 * Unlink all of our kegs. 1629 */ 1630 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1631 klink->kl_keg = NULL; 1632 LIST_REMOVE(klink, kl_link); 1633 if (klink == &zone->uz_klink) 1634 continue; 1635 free(klink, M_TEMP); 1636 } 1637 /* 1638 * We only destroy kegs from non secondary zones. 1639 */ 1640 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1641 mtx_lock(&uma_mtx); 1642 LIST_REMOVE(keg, uk_link); 1643 mtx_unlock(&uma_mtx); 1644 zone_free_item(kegs, keg, NULL, SKIP_NONE); 1645 } 1646 ZONE_LOCK_FINI(zone); 1647} 1648 1649/* 1650 * Traverses every zone in the system and calls a callback 1651 * 1652 * Arguments: 1653 * zfunc A pointer to a function which accepts a zone 1654 * as an argument. 1655 * 1656 * Returns: 1657 * Nothing 1658 */ 1659static void 1660zone_foreach(void (*zfunc)(uma_zone_t)) 1661{ 1662 uma_keg_t keg; 1663 uma_zone_t zone; 1664 1665 mtx_lock(&uma_mtx); 1666 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1667 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1668 zfunc(zone); 1669 } 1670 mtx_unlock(&uma_mtx); 1671} 1672 1673/* Public functions */ 1674/* See uma.h */ 1675void 1676uma_startup(void *bootmem, int boot_pages) 1677{ 1678 struct uma_zctor_args args; 1679 uma_slab_t slab; 1680 u_int slabsize; 1681 int i; 1682 1683#ifdef UMA_DEBUG 1684 printf("Creating uma keg headers zone and keg.\n"); 1685#endif 1686 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF); 1687 1688 /* "manually" create the initial zone */ 1689 memset(&args, 0, sizeof(args)); 1690 args.name = "UMA Kegs"; 1691 args.size = sizeof(struct uma_keg); 1692 args.ctor = keg_ctor; 1693 args.dtor = keg_dtor; 1694 args.uminit = zero_init; 1695 args.fini = NULL; 1696 args.keg = &masterkeg; 1697 args.align = 32 - 1; 1698 args.flags = UMA_ZFLAG_INTERNAL; 1699 /* The initial zone has no Per cpu queues so it's smaller */ 1700 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); 1701 1702#ifdef UMA_DEBUG 1703 printf("Filling boot free list.\n"); 1704#endif 1705 for (i = 0; i < boot_pages; i++) { 1706 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE)); 1707 slab->us_data = (uint8_t *)slab; 1708 slab->us_flags = UMA_SLAB_BOOT; 1709 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); 1710 } 1711 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); 1712 1713#ifdef UMA_DEBUG 1714 printf("Creating uma zone headers zone and keg.\n"); 1715#endif 1716 args.name = "UMA Zones"; 1717 args.size = sizeof(struct uma_zone) + 1718 (sizeof(struct uma_cache) * (mp_maxid + 1)); 1719 args.ctor = zone_ctor; 1720 args.dtor = zone_dtor; 1721 args.uminit = zero_init; 1722 args.fini = NULL; 1723 args.keg = NULL; 1724 args.align = 32 - 1; 1725 args.flags = UMA_ZFLAG_INTERNAL; 1726 /* The initial zone has no Per cpu queues so it's smaller */ 1727 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); 1728 1729#ifdef UMA_DEBUG 1730 printf("Initializing pcpu cache locks.\n"); 1731#endif 1732#ifdef UMA_DEBUG 1733 printf("Creating slab and hash zones.\n"); 1734#endif 1735 1736 /* Now make a zone for slab headers */ 1737 slabzone = uma_zcreate("UMA Slabs", 1738 sizeof(struct uma_slab), 1739 NULL, NULL, NULL, NULL, 1740 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1741 1742 /* 1743 * We also create a zone for the bigger slabs with reference 1744 * counts in them, to accomodate UMA_ZONE_REFCNT zones. 1745 */ 1746 slabsize = sizeof(struct uma_slab_refcnt); 1747 slabsize += uma_max_ipers_ref * sizeof(uint32_t); 1748 slabrefzone = uma_zcreate("UMA RCntSlabs", 1749 slabsize, 1750 NULL, NULL, NULL, NULL, 1751 UMA_ALIGN_PTR, 1752 UMA_ZFLAG_INTERNAL); 1753 1754 hashzone = uma_zcreate("UMA Hash", 1755 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 1756 NULL, NULL, NULL, NULL, 1757 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1758 1759 bucket_init(); 1760 1761 booted = UMA_STARTUP; 1762 1763#ifdef UMA_DEBUG 1764 printf("UMA startup complete.\n"); 1765#endif 1766} 1767 1768/* see uma.h */ 1769void 1770uma_startup2(void) 1771{ 1772 booted = UMA_STARTUP2; 1773 bucket_enable(); 1774#ifdef UMA_DEBUG 1775 printf("UMA startup2 complete.\n"); 1776#endif 1777} 1778 1779/* 1780 * Initialize our callout handle 1781 * 1782 */ 1783 1784static void 1785uma_startup3(void) 1786{ 1787#ifdef UMA_DEBUG 1788 printf("Starting callout.\n"); 1789#endif 1790 callout_init(&uma_callout, CALLOUT_MPSAFE); 1791 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1792#ifdef UMA_DEBUG 1793 printf("UMA startup3 complete.\n"); 1794#endif 1795} 1796 1797static uma_keg_t 1798uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 1799 int align, uint32_t flags) 1800{ 1801 struct uma_kctor_args args; 1802 1803 args.size = size; 1804 args.uminit = uminit; 1805 args.fini = fini; 1806 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 1807 args.flags = flags; 1808 args.zone = zone; 1809 return (zone_alloc_item(kegs, &args, M_WAITOK)); 1810} 1811 1812/* See uma.h */ 1813void 1814uma_set_align(int align) 1815{ 1816 1817 if (align != UMA_ALIGN_CACHE) 1818 uma_align_cache = align; 1819} 1820 1821/* See uma.h */ 1822uma_zone_t 1823uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 1824 uma_init uminit, uma_fini fini, int align, uint32_t flags) 1825 1826{ 1827 struct uma_zctor_args args; 1828 1829 /* This stuff is essential for the zone ctor */ 1830 memset(&args, 0, sizeof(args)); 1831 args.name = name; 1832 args.size = size; 1833 args.ctor = ctor; 1834 args.dtor = dtor; 1835 args.uminit = uminit; 1836 args.fini = fini; 1837 args.align = align; 1838 args.flags = flags; 1839 args.keg = NULL; 1840 1841 return (zone_alloc_item(zones, &args, M_WAITOK)); 1842} 1843 1844/* See uma.h */ 1845uma_zone_t 1846uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 1847 uma_init zinit, uma_fini zfini, uma_zone_t master) 1848{ 1849 struct uma_zctor_args args; 1850 uma_keg_t keg; 1851 1852 keg = zone_first_keg(master); 1853 memset(&args, 0, sizeof(args)); 1854 args.name = name; 1855 args.size = keg->uk_size; 1856 args.ctor = ctor; 1857 args.dtor = dtor; 1858 args.uminit = zinit; 1859 args.fini = zfini; 1860 args.align = keg->uk_align; 1861 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 1862 args.keg = keg; 1863 1864 /* XXX Attaches only one keg of potentially many. */ 1865 return (zone_alloc_item(zones, &args, M_WAITOK)); 1866} 1867 1868/* See uma.h */ 1869uma_zone_t 1870uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 1871 uma_init zinit, uma_fini zfini, uma_import zimport, 1872 uma_release zrelease, void *arg, int flags) 1873{ 1874 struct uma_zctor_args args; 1875 1876 memset(&args, 0, sizeof(args)); 1877 args.name = name; 1878 args.size = size; 1879 args.ctor = ctor; 1880 args.dtor = dtor; 1881 args.uminit = zinit; 1882 args.fini = zfini; 1883 args.import = zimport; 1884 args.release = zrelease; 1885 args.arg = arg; 1886 args.align = 0; 1887 args.flags = flags; 1888 1889 return (zone_alloc_item(zones, &args, M_WAITOK)); 1890} 1891 1892static void 1893zone_lock_pair(uma_zone_t a, uma_zone_t b) 1894{ 1895 if (a < b) { 1896 ZONE_LOCK(a); 1897 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); 1898 } else { 1899 ZONE_LOCK(b); 1900 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); 1901 } 1902} 1903 1904static void 1905zone_unlock_pair(uma_zone_t a, uma_zone_t b) 1906{ 1907 1908 ZONE_UNLOCK(a); 1909 ZONE_UNLOCK(b); 1910} 1911 1912int 1913uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 1914{ 1915 uma_klink_t klink; 1916 uma_klink_t kl; 1917 int error; 1918 1919 error = 0; 1920 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 1921 1922 zone_lock_pair(zone, master); 1923 /* 1924 * zone must use vtoslab() to resolve objects and must already be 1925 * a secondary. 1926 */ 1927 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 1928 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 1929 error = EINVAL; 1930 goto out; 1931 } 1932 /* 1933 * The new master must also use vtoslab(). 1934 */ 1935 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 1936 error = EINVAL; 1937 goto out; 1938 } 1939 /* 1940 * Both must either be refcnt, or not be refcnt. 1941 */ 1942 if ((zone->uz_flags & UMA_ZONE_REFCNT) != 1943 (master->uz_flags & UMA_ZONE_REFCNT)) { 1944 error = EINVAL; 1945 goto out; 1946 } 1947 /* 1948 * The underlying object must be the same size. rsize 1949 * may be different. 1950 */ 1951 if (master->uz_size != zone->uz_size) { 1952 error = E2BIG; 1953 goto out; 1954 } 1955 /* 1956 * Put it at the end of the list. 1957 */ 1958 klink->kl_keg = zone_first_keg(master); 1959 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 1960 if (LIST_NEXT(kl, kl_link) == NULL) { 1961 LIST_INSERT_AFTER(kl, klink, kl_link); 1962 break; 1963 } 1964 } 1965 klink = NULL; 1966 zone->uz_flags |= UMA_ZFLAG_MULTI; 1967 zone->uz_slab = zone_fetch_slab_multi; 1968 1969out: 1970 zone_unlock_pair(zone, master); 1971 if (klink != NULL) 1972 free(klink, M_TEMP); 1973 1974 return (error); 1975} 1976 1977 1978/* See uma.h */ 1979void 1980uma_zdestroy(uma_zone_t zone) 1981{ 1982 1983 zone_free_item(zones, zone, NULL, SKIP_NONE); 1984} 1985 1986/* See uma.h */ 1987void * 1988uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 1989{ 1990 void *item; 1991 uma_cache_t cache; 1992 uma_bucket_t bucket; 1993 int lockfail; 1994 int cpu; 1995 1996 /* This is the fast path allocation */ 1997#ifdef UMA_DEBUG_ALLOC_1 1998 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); 1999#endif 2000 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread, 2001 zone->uz_name, flags); 2002 2003 if (flags & M_WAITOK) { 2004 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2005 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2006 } 2007#ifdef DEBUG_MEMGUARD 2008 if (memguard_cmp_zone(zone)) { 2009 item = memguard_alloc(zone->uz_size, flags); 2010 if (item != NULL) { 2011 /* 2012 * Avoid conflict with the use-after-free 2013 * protecting infrastructure from INVARIANTS. 2014 */ 2015 if (zone->uz_init != NULL && 2016 zone->uz_init != mtrash_init && 2017 zone->uz_init(item, zone->uz_size, flags) != 0) 2018 return (NULL); 2019 if (zone->uz_ctor != NULL && 2020 zone->uz_ctor != mtrash_ctor && 2021 zone->uz_ctor(item, zone->uz_size, udata, 2022 flags) != 0) { 2023 zone->uz_fini(item, zone->uz_size); 2024 return (NULL); 2025 } 2026 return (item); 2027 } 2028 /* This is unfortunate but should not be fatal. */ 2029 } 2030#endif 2031 /* 2032 * If possible, allocate from the per-CPU cache. There are two 2033 * requirements for safe access to the per-CPU cache: (1) the thread 2034 * accessing the cache must not be preempted or yield during access, 2035 * and (2) the thread must not migrate CPUs without switching which 2036 * cache it accesses. We rely on a critical section to prevent 2037 * preemption and migration. We release the critical section in 2038 * order to acquire the zone mutex if we are unable to allocate from 2039 * the current cache; when we re-acquire the critical section, we 2040 * must detect and handle migration if it has occurred. 2041 */ 2042 critical_enter(); 2043 cpu = curcpu; 2044 cache = &zone->uz_cpu[cpu]; 2045 2046zalloc_start: 2047 bucket = cache->uc_allocbucket; 2048 if (bucket != NULL && bucket->ub_cnt > 0) { 2049 bucket->ub_cnt--; 2050 item = bucket->ub_bucket[bucket->ub_cnt]; 2051#ifdef INVARIANTS 2052 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2053#endif 2054 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2055 cache->uc_allocs++; 2056 critical_exit(); 2057 if (zone->uz_ctor != NULL && 2058 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2059 atomic_add_long(&zone->uz_fails, 1); 2060 zone_free_item(zone, item, udata, SKIP_DTOR); 2061 return (NULL); 2062 } 2063#ifdef INVARIANTS 2064 uma_dbg_alloc(zone, NULL, item); 2065#endif 2066 if (flags & M_ZERO) 2067 bzero(item, zone->uz_size); 2068 return (item); 2069 } 2070 2071 /* 2072 * We have run out of items in our alloc bucket. 2073 * See if we can switch with our free bucket. 2074 */ 2075 bucket = cache->uc_freebucket; 2076 if (bucket != NULL && bucket->ub_cnt > 0) { 2077#ifdef UMA_DEBUG_ALLOC 2078 printf("uma_zalloc: Swapping empty with alloc.\n"); 2079#endif 2080 cache->uc_freebucket = cache->uc_allocbucket; 2081 cache->uc_allocbucket = bucket; 2082 goto zalloc_start; 2083 } 2084 2085 /* 2086 * Discard any empty allocation bucket while we hold no locks. 2087 */ 2088 bucket = cache->uc_allocbucket; 2089 cache->uc_allocbucket = NULL; 2090 critical_exit(); 2091 if (bucket != NULL) 2092 bucket_free(zone, bucket, udata); 2093 2094 /* Short-circuit for zones without buckets and low memory. */ 2095 if (zone->uz_count == 0 || bucketdisable) 2096 goto zalloc_item; 2097 2098 /* 2099 * Attempt to retrieve the item from the per-CPU cache has failed, so 2100 * we must go back to the zone. This requires the zone lock, so we 2101 * must drop the critical section, then re-acquire it when we go back 2102 * to the cache. Since the critical section is released, we may be 2103 * preempted or migrate. As such, make sure not to maintain any 2104 * thread-local state specific to the cache from prior to releasing 2105 * the critical section. 2106 */ 2107 lockfail = 0; 2108 if (ZONE_TRYLOCK(zone) == 0) { 2109 /* Record contention to size the buckets. */ 2110 ZONE_LOCK(zone); 2111 lockfail = 1; 2112 } 2113 critical_enter(); 2114 cpu = curcpu; 2115 cache = &zone->uz_cpu[cpu]; 2116 2117 /* 2118 * Since we have locked the zone we may as well send back our stats. 2119 */ 2120 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2121 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2122 cache->uc_allocs = 0; 2123 cache->uc_frees = 0; 2124 2125 /* See if we lost the race to fill the cache. */ 2126 if (cache->uc_allocbucket != NULL) { 2127 ZONE_UNLOCK(zone); 2128 goto zalloc_start; 2129 } 2130 2131 /* 2132 * Check the zone's cache of buckets. 2133 */ 2134 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 2135 KASSERT(bucket->ub_cnt != 0, 2136 ("uma_zalloc_arg: Returning an empty bucket.")); 2137 2138 LIST_REMOVE(bucket, ub_link); 2139 cache->uc_allocbucket = bucket; 2140 ZONE_UNLOCK(zone); 2141 goto zalloc_start; 2142 } 2143 /* We are no longer associated with this CPU. */ 2144 critical_exit(); 2145 2146 /* 2147 * We bump the uz count when the cache size is insufficient to 2148 * handle the working set. 2149 */ 2150 if (lockfail && zone->uz_count < BUCKET_MAX) 2151 zone->uz_count++; 2152 ZONE_UNLOCK(zone); 2153 2154 /* 2155 * Now lets just fill a bucket and put it on the free list. If that 2156 * works we'll restart the allocation from the begining and it 2157 * will use the just filled bucket. 2158 */ 2159 bucket = zone_alloc_bucket(zone, udata, flags); 2160 if (bucket != NULL) { 2161 ZONE_LOCK(zone); 2162 critical_enter(); 2163 cpu = curcpu; 2164 cache = &zone->uz_cpu[cpu]; 2165 /* 2166 * See if we lost the race or were migrated. Cache the 2167 * initialized bucket to make this less likely or claim 2168 * the memory directly. 2169 */ 2170 if (cache->uc_allocbucket == NULL) 2171 cache->uc_allocbucket = bucket; 2172 else 2173 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2174 ZONE_UNLOCK(zone); 2175 goto zalloc_start; 2176 } 2177 2178 /* 2179 * We may not be able to get a bucket so return an actual item. 2180 */ 2181#ifdef UMA_DEBUG 2182 printf("uma_zalloc_arg: Bucketzone returned NULL\n"); 2183#endif 2184 2185zalloc_item: 2186 item = zone_alloc_item(zone, udata, flags); 2187 2188 return (item); 2189} 2190 2191static uma_slab_t 2192keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) 2193{ 2194 uma_slab_t slab; 2195 int reserve; 2196 2197 mtx_assert(&keg->uk_lock, MA_OWNED); 2198 slab = NULL; 2199 reserve = 0; 2200 if ((flags & M_USE_RESERVE) == 0) 2201 reserve = keg->uk_reserve; 2202 2203 for (;;) { 2204 /* 2205 * Find a slab with some space. Prefer slabs that are partially 2206 * used over those that are totally full. This helps to reduce 2207 * fragmentation. 2208 */ 2209 if (keg->uk_free > reserve) { 2210 if (!LIST_EMPTY(&keg->uk_part_slab)) { 2211 slab = LIST_FIRST(&keg->uk_part_slab); 2212 } else { 2213 slab = LIST_FIRST(&keg->uk_free_slab); 2214 LIST_REMOVE(slab, us_link); 2215 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, 2216 us_link); 2217 } 2218 MPASS(slab->us_keg == keg); 2219 return (slab); 2220 } 2221 2222 /* 2223 * M_NOVM means don't ask at all! 2224 */ 2225 if (flags & M_NOVM) 2226 break; 2227 2228 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2229 keg->uk_flags |= UMA_ZFLAG_FULL; 2230 /* 2231 * If this is not a multi-zone, set the FULL bit. 2232 * Otherwise slab_multi() takes care of it. 2233 */ 2234 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2235 zone->uz_flags |= UMA_ZFLAG_FULL; 2236 zone_log_warning(zone); 2237 } 2238 if (flags & M_NOWAIT) 2239 break; 2240 zone->uz_sleeps++; 2241 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2242 continue; 2243 } 2244 slab = keg_alloc_slab(keg, zone, flags); 2245 /* 2246 * If we got a slab here it's safe to mark it partially used 2247 * and return. We assume that the caller is going to remove 2248 * at least one item. 2249 */ 2250 if (slab) { 2251 MPASS(slab->us_keg == keg); 2252 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2253 return (slab); 2254 } 2255 /* 2256 * We might not have been able to get a slab but another cpu 2257 * could have while we were unlocked. Check again before we 2258 * fail. 2259 */ 2260 flags |= M_NOVM; 2261 } 2262 return (slab); 2263} 2264 2265static uma_slab_t 2266zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) 2267{ 2268 uma_slab_t slab; 2269 2270 if (keg == NULL) { 2271 keg = zone_first_keg(zone); 2272 KEG_LOCK(keg); 2273 } 2274 2275 for (;;) { 2276 slab = keg_fetch_slab(keg, zone, flags); 2277 if (slab) 2278 return (slab); 2279 if (flags & (M_NOWAIT | M_NOVM)) 2280 break; 2281 } 2282 KEG_UNLOCK(keg); 2283 return (NULL); 2284} 2285 2286/* 2287 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2288 * with the keg locked. On NULL no lock is held. 2289 * 2290 * The last pointer is used to seed the search. It is not required. 2291 */ 2292static uma_slab_t 2293zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) 2294{ 2295 uma_klink_t klink; 2296 uma_slab_t slab; 2297 uma_keg_t keg; 2298 int flags; 2299 int empty; 2300 int full; 2301 2302 /* 2303 * Don't wait on the first pass. This will skip limit tests 2304 * as well. We don't want to block if we can find a provider 2305 * without blocking. 2306 */ 2307 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2308 /* 2309 * Use the last slab allocated as a hint for where to start 2310 * the search. 2311 */ 2312 if (last != NULL) { 2313 slab = keg_fetch_slab(last, zone, flags); 2314 if (slab) 2315 return (slab); 2316 KEG_UNLOCK(last); 2317 } 2318 /* 2319 * Loop until we have a slab incase of transient failures 2320 * while M_WAITOK is specified. I'm not sure this is 100% 2321 * required but we've done it for so long now. 2322 */ 2323 for (;;) { 2324 empty = 0; 2325 full = 0; 2326 /* 2327 * Search the available kegs for slabs. Be careful to hold the 2328 * correct lock while calling into the keg layer. 2329 */ 2330 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2331 keg = klink->kl_keg; 2332 KEG_LOCK(keg); 2333 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2334 slab = keg_fetch_slab(keg, zone, flags); 2335 if (slab) 2336 return (slab); 2337 } 2338 if (keg->uk_flags & UMA_ZFLAG_FULL) 2339 full++; 2340 else 2341 empty++; 2342 KEG_UNLOCK(keg); 2343 } 2344 if (rflags & (M_NOWAIT | M_NOVM)) 2345 break; 2346 flags = rflags; 2347 /* 2348 * All kegs are full. XXX We can't atomically check all kegs 2349 * and sleep so just sleep for a short period and retry. 2350 */ 2351 if (full && !empty) { 2352 ZONE_LOCK(zone); 2353 zone->uz_flags |= UMA_ZFLAG_FULL; 2354 zone->uz_sleeps++; 2355 zone_log_warning(zone); 2356 msleep(zone, zone->uz_lockptr, PVM, 2357 "zonelimit", hz/100); 2358 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2359 ZONE_UNLOCK(zone); 2360 continue; 2361 } 2362 } 2363 return (NULL); 2364} 2365 2366static void * 2367slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 2368{ 2369 void *item; 2370 uint8_t freei; 2371 2372 MPASS(keg == slab->us_keg); 2373 mtx_assert(&keg->uk_lock, MA_OWNED); 2374 2375 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; 2376 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); 2377 item = slab->us_data + (keg->uk_rsize * freei); 2378 slab->us_freecount--; 2379 keg->uk_free--; 2380 2381 /* Move this slab to the full list */ 2382 if (slab->us_freecount == 0) { 2383 LIST_REMOVE(slab, us_link); 2384 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); 2385 } 2386 2387 return (item); 2388} 2389 2390static int 2391zone_import(uma_zone_t zone, void **bucket, int max, int flags) 2392{ 2393 uma_slab_t slab; 2394 uma_keg_t keg; 2395 int i; 2396 2397 slab = NULL; 2398 keg = NULL; 2399 /* Try to keep the buckets totally full */ 2400 for (i = 0; i < max; ) { 2401 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL) 2402 break; 2403 keg = slab->us_keg; 2404 while (slab->us_freecount && i < max) { 2405 bucket[i++] = slab_alloc_item(keg, slab); 2406 if (keg->uk_free <= keg->uk_reserve) 2407 break; 2408 } 2409 /* Don't grab more than one slab at a time. */ 2410 flags &= ~M_WAITOK; 2411 flags |= M_NOWAIT; 2412 } 2413 if (slab != NULL) 2414 KEG_UNLOCK(keg); 2415 2416 return i; 2417} 2418 2419static uma_bucket_t 2420zone_alloc_bucket(uma_zone_t zone, void *udata, int flags) 2421{ 2422 uma_bucket_t bucket; 2423 int max; 2424 2425 /* Don't wait for buckets, preserve caller's NOVM setting. */ 2426 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 2427 if (bucket == NULL) 2428 goto out; 2429 2430 max = MIN(bucket->ub_entries, zone->uz_count); 2431 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 2432 max, flags); 2433 2434 /* 2435 * Initialize the memory if necessary. 2436 */ 2437 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 2438 int i; 2439 2440 for (i = 0; i < bucket->ub_cnt; i++) 2441 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2442 flags) != 0) 2443 break; 2444 /* 2445 * If we couldn't initialize the whole bucket, put the 2446 * rest back onto the freelist. 2447 */ 2448 if (i != bucket->ub_cnt) { 2449 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 2450 bucket->ub_cnt - i); 2451#ifdef INVARIANTS 2452 bzero(&bucket->ub_bucket[i], 2453 sizeof(void *) * (bucket->ub_cnt - i)); 2454#endif 2455 bucket->ub_cnt = i; 2456 } 2457 } 2458 2459out: 2460 if (bucket == NULL || bucket->ub_cnt == 0) { 2461 if (bucket != NULL) 2462 bucket_free(zone, bucket, udata); 2463 atomic_add_long(&zone->uz_fails, 1); 2464 return (NULL); 2465 } 2466 2467 return (bucket); 2468} 2469 2470/* 2471 * Allocates a single item from a zone. 2472 * 2473 * Arguments 2474 * zone The zone to alloc for. 2475 * udata The data to be passed to the constructor. 2476 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2477 * 2478 * Returns 2479 * NULL if there is no memory and M_NOWAIT is set 2480 * An item if successful 2481 */ 2482 2483static void * 2484zone_alloc_item(uma_zone_t zone, void *udata, int flags) 2485{ 2486 void *item; 2487 2488 item = NULL; 2489 2490#ifdef UMA_DEBUG_ALLOC 2491 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); 2492#endif 2493 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1) 2494 goto fail; 2495 atomic_add_long(&zone->uz_allocs, 1); 2496 2497 /* 2498 * We have to call both the zone's init (not the keg's init) 2499 * and the zone's ctor. This is because the item is going from 2500 * a keg slab directly to the user, and the user is expecting it 2501 * to be both zone-init'd as well as zone-ctor'd. 2502 */ 2503 if (zone->uz_init != NULL) { 2504 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 2505 zone_free_item(zone, item, udata, SKIP_FINI); 2506 goto fail; 2507 } 2508 } 2509 if (zone->uz_ctor != NULL) { 2510 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2511 zone_free_item(zone, item, udata, SKIP_DTOR); 2512 goto fail; 2513 } 2514 } 2515#ifdef INVARIANTS 2516 uma_dbg_alloc(zone, NULL, item); 2517#endif 2518 if (flags & M_ZERO) 2519 bzero(item, zone->uz_size); 2520 2521 return (item); 2522 2523fail: 2524 atomic_add_long(&zone->uz_fails, 1); 2525 return (NULL); 2526} 2527 2528/* See uma.h */ 2529void 2530uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 2531{ 2532 uma_cache_t cache; 2533 uma_bucket_t bucket; 2534 int cpu; 2535 2536#ifdef UMA_DEBUG_ALLOC_1 2537 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); 2538#endif 2539 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 2540 zone->uz_name); 2541 2542 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 2543 if (item == NULL) 2544 return; 2545#ifdef DEBUG_MEMGUARD 2546 if (is_memguard_addr(item)) { 2547 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor) 2548 zone->uz_dtor(item, zone->uz_size, udata); 2549 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini) 2550 zone->uz_fini(item, zone->uz_size); 2551 memguard_free(item); 2552 return; 2553 } 2554#endif 2555#ifdef INVARIANTS 2556 if (zone->uz_flags & UMA_ZONE_MALLOC) 2557 uma_dbg_free(zone, udata, item); 2558 else 2559 uma_dbg_free(zone, NULL, item); 2560#endif 2561 if (zone->uz_dtor != NULL) 2562 zone->uz_dtor(item, zone->uz_size, udata); 2563 2564 /* 2565 * The race here is acceptable. If we miss it we'll just have to wait 2566 * a little longer for the limits to be reset. 2567 */ 2568 if (zone->uz_flags & UMA_ZFLAG_FULL) 2569 goto zfree_item; 2570 2571 /* 2572 * If possible, free to the per-CPU cache. There are two 2573 * requirements for safe access to the per-CPU cache: (1) the thread 2574 * accessing the cache must not be preempted or yield during access, 2575 * and (2) the thread must not migrate CPUs without switching which 2576 * cache it accesses. We rely on a critical section to prevent 2577 * preemption and migration. We release the critical section in 2578 * order to acquire the zone mutex if we are unable to free to the 2579 * current cache; when we re-acquire the critical section, we must 2580 * detect and handle migration if it has occurred. 2581 */ 2582zfree_restart: 2583 critical_enter(); 2584 cpu = curcpu; 2585 cache = &zone->uz_cpu[cpu]; 2586 2587zfree_start: 2588 /* 2589 * Try to free into the allocbucket first to give LIFO ordering 2590 * for cache-hot datastructures. Spill over into the freebucket 2591 * if necessary. Alloc will swap them if one runs dry. 2592 */ 2593 bucket = cache->uc_allocbucket; 2594 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 2595 bucket = cache->uc_freebucket; 2596 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2597 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 2598 ("uma_zfree: Freeing to non free bucket index.")); 2599 bucket->ub_bucket[bucket->ub_cnt] = item; 2600 bucket->ub_cnt++; 2601 cache->uc_frees++; 2602 critical_exit(); 2603 return; 2604 } 2605 2606 /* 2607 * We must go back the zone, which requires acquiring the zone lock, 2608 * which in turn means we must release and re-acquire the critical 2609 * section. Since the critical section is released, we may be 2610 * preempted or migrate. As such, make sure not to maintain any 2611 * thread-local state specific to the cache from prior to releasing 2612 * the critical section. 2613 */ 2614 critical_exit(); 2615 if (zone->uz_count == 0 || bucketdisable) 2616 goto zfree_item; 2617 2618 ZONE_LOCK(zone); 2619 critical_enter(); 2620 cpu = curcpu; 2621 cache = &zone->uz_cpu[cpu]; 2622 2623 /* 2624 * Since we have locked the zone we may as well send back our stats. 2625 */ 2626 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2627 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2628 cache->uc_allocs = 0; 2629 cache->uc_frees = 0; 2630 2631 bucket = cache->uc_freebucket; 2632 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2633 ZONE_UNLOCK(zone); 2634 goto zfree_start; 2635 } 2636 cache->uc_freebucket = NULL; 2637 2638 /* Can we throw this on the zone full list? */ 2639 if (bucket != NULL) { 2640#ifdef UMA_DEBUG_ALLOC 2641 printf("uma_zfree: Putting old bucket on the free list.\n"); 2642#endif 2643 /* ub_cnt is pointing to the last free item */ 2644 KASSERT(bucket->ub_cnt != 0, 2645 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 2646 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2647 } 2648 2649 /* We are no longer associated with this CPU. */ 2650 critical_exit(); 2651 2652 /* And the zone.. */ 2653 ZONE_UNLOCK(zone); 2654 2655#ifdef UMA_DEBUG_ALLOC 2656 printf("uma_zfree: Allocating new free bucket.\n"); 2657#endif 2658 bucket = bucket_alloc(zone, udata, M_NOWAIT); 2659 if (bucket) { 2660 critical_enter(); 2661 cpu = curcpu; 2662 cache = &zone->uz_cpu[cpu]; 2663 if (cache->uc_freebucket == NULL) { 2664 cache->uc_freebucket = bucket; 2665 goto zfree_start; 2666 } 2667 /* 2668 * We lost the race, start over. We have to drop our 2669 * critical section to free the bucket. 2670 */ 2671 critical_exit(); 2672 bucket_free(zone, bucket, udata); 2673 goto zfree_restart; 2674 } 2675 2676 /* 2677 * If nothing else caught this, we'll just do an internal free. 2678 */ 2679zfree_item: 2680 zone_free_item(zone, item, udata, SKIP_DTOR); 2681 2682 return; 2683} 2684 2685static void 2686slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) 2687{ 2688 uint8_t freei; 2689 2690 mtx_assert(&keg->uk_lock, MA_OWNED); 2691 MPASS(keg == slab->us_keg); 2692 2693 /* Do we need to remove from any lists? */ 2694 if (slab->us_freecount+1 == keg->uk_ipers) { 2695 LIST_REMOVE(slab, us_link); 2696 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 2697 } else if (slab->us_freecount == 0) { 2698 LIST_REMOVE(slab, us_link); 2699 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2700 } 2701 2702 /* Slab management. */ 2703 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 2704 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); 2705 slab->us_freecount++; 2706 2707 /* Keg statistics. */ 2708 keg->uk_free++; 2709} 2710 2711static void 2712zone_release(uma_zone_t zone, void **bucket, int cnt) 2713{ 2714 void *item; 2715 uma_slab_t slab; 2716 uma_keg_t keg; 2717 uint8_t *mem; 2718 int clearfull; 2719 int i; 2720 2721 clearfull = 0; 2722 keg = zone_first_keg(zone); 2723 KEG_LOCK(keg); 2724 for (i = 0; i < cnt; i++) { 2725 item = bucket[i]; 2726 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 2727 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 2728 if (zone->uz_flags & UMA_ZONE_HASH) { 2729 slab = hash_sfind(&keg->uk_hash, mem); 2730 } else { 2731 mem += keg->uk_pgoff; 2732 slab = (uma_slab_t)mem; 2733 } 2734 } else { 2735 slab = vtoslab((vm_offset_t)item); 2736 if (slab->us_keg != keg) { 2737 KEG_UNLOCK(keg); 2738 keg = slab->us_keg; 2739 KEG_LOCK(keg); 2740 } 2741 } 2742 slab_free_item(keg, slab, item); 2743 if (keg->uk_flags & UMA_ZFLAG_FULL) { 2744 if (keg->uk_pages < keg->uk_maxpages) { 2745 keg->uk_flags &= ~UMA_ZFLAG_FULL; 2746 clearfull = 1; 2747 } 2748 2749 /* 2750 * We can handle one more allocation. Since we're 2751 * clearing ZFLAG_FULL, wake up all procs blocked 2752 * on pages. This should be uncommon, so keeping this 2753 * simple for now (rather than adding count of blocked 2754 * threads etc). 2755 */ 2756 wakeup(keg); 2757 } 2758 } 2759 KEG_UNLOCK(keg); 2760 if (clearfull) { 2761 ZONE_LOCK(zone); 2762 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2763 wakeup(zone); 2764 ZONE_UNLOCK(zone); 2765 } 2766 2767} 2768 2769/* 2770 * Frees a single item to any zone. 2771 * 2772 * Arguments: 2773 * zone The zone to free to 2774 * item The item we're freeing 2775 * udata User supplied data for the dtor 2776 * skip Skip dtors and finis 2777 */ 2778static void 2779zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 2780{ 2781 2782#ifdef INVARIANTS 2783 if (skip == SKIP_NONE) { 2784 if (zone->uz_flags & UMA_ZONE_MALLOC) 2785 uma_dbg_free(zone, udata, item); 2786 else 2787 uma_dbg_free(zone, NULL, item); 2788 } 2789#endif 2790 if (skip < SKIP_DTOR && zone->uz_dtor) 2791 zone->uz_dtor(item, zone->uz_size, udata); 2792 2793 if (skip < SKIP_FINI && zone->uz_fini) 2794 zone->uz_fini(item, zone->uz_size); 2795 2796 atomic_add_long(&zone->uz_frees, 1); 2797 zone->uz_release(zone->uz_arg, &item, 1); 2798} 2799 2800/* See uma.h */ 2801int 2802uma_zone_set_max(uma_zone_t zone, int nitems) 2803{ 2804 uma_keg_t keg; 2805 2806 keg = zone_first_keg(zone); 2807 if (keg == NULL) 2808 return (0); 2809 KEG_LOCK(keg); 2810 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 2811 if (keg->uk_maxpages * keg->uk_ipers < nitems) 2812 keg->uk_maxpages += keg->uk_ppera; 2813 nitems = keg->uk_maxpages * keg->uk_ipers; 2814 KEG_UNLOCK(keg); 2815 2816 return (nitems); 2817} 2818 2819/* See uma.h */ 2820int 2821uma_zone_get_max(uma_zone_t zone) 2822{ 2823 int nitems; 2824 uma_keg_t keg; 2825 2826 keg = zone_first_keg(zone); 2827 if (keg == NULL) 2828 return (0); 2829 KEG_LOCK(keg); 2830 nitems = keg->uk_maxpages * keg->uk_ipers; 2831 KEG_UNLOCK(keg); 2832 2833 return (nitems); 2834} 2835 2836/* See uma.h */ 2837void 2838uma_zone_set_warning(uma_zone_t zone, const char *warning) 2839{ 2840 2841 ZONE_LOCK(zone); 2842 zone->uz_warning = warning; 2843 ZONE_UNLOCK(zone); 2844} 2845 2846/* See uma.h */ 2847int 2848uma_zone_get_cur(uma_zone_t zone) 2849{ 2850 int64_t nitems; 2851 u_int i; 2852 2853 ZONE_LOCK(zone); 2854 nitems = zone->uz_allocs - zone->uz_frees; 2855 CPU_FOREACH(i) { 2856 /* 2857 * See the comment in sysctl_vm_zone_stats() regarding the 2858 * safety of accessing the per-cpu caches. With the zone lock 2859 * held, it is safe, but can potentially result in stale data. 2860 */ 2861 nitems += zone->uz_cpu[i].uc_allocs - 2862 zone->uz_cpu[i].uc_frees; 2863 } 2864 ZONE_UNLOCK(zone); 2865 2866 return (nitems < 0 ? 0 : nitems); 2867} 2868 2869/* See uma.h */ 2870void 2871uma_zone_set_init(uma_zone_t zone, uma_init uminit) 2872{ 2873 uma_keg_t keg; 2874 2875 keg = zone_first_keg(zone); 2876 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 2877 KEG_LOCK(keg); 2878 KASSERT(keg->uk_pages == 0, 2879 ("uma_zone_set_init on non-empty keg")); 2880 keg->uk_init = uminit; 2881 KEG_UNLOCK(keg); 2882} 2883 2884/* See uma.h */ 2885void 2886uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 2887{ 2888 uma_keg_t keg; 2889 2890 keg = zone_first_keg(zone); 2891 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 2892 KEG_LOCK(keg); 2893 KASSERT(keg->uk_pages == 0, 2894 ("uma_zone_set_fini on non-empty keg")); 2895 keg->uk_fini = fini; 2896 KEG_UNLOCK(keg); 2897} 2898 2899/* See uma.h */ 2900void 2901uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 2902{ 2903 2904 ZONE_LOCK(zone); 2905 KASSERT(zone_first_keg(zone)->uk_pages == 0, 2906 ("uma_zone_set_zinit on non-empty keg")); 2907 zone->uz_init = zinit; 2908 ZONE_UNLOCK(zone); 2909} 2910 2911/* See uma.h */ 2912void 2913uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 2914{ 2915 2916 ZONE_LOCK(zone); 2917 KASSERT(zone_first_keg(zone)->uk_pages == 0, 2918 ("uma_zone_set_zfini on non-empty keg")); 2919 zone->uz_fini = zfini; 2920 ZONE_UNLOCK(zone); 2921} 2922 2923/* See uma.h */ 2924/* XXX uk_freef is not actually used with the zone locked */ 2925void 2926uma_zone_set_freef(uma_zone_t zone, uma_free freef) 2927{ 2928 uma_keg_t keg; 2929 2930 keg = zone_first_keg(zone); 2931 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 2932 KEG_LOCK(keg); 2933 keg->uk_freef = freef; 2934 KEG_UNLOCK(keg); 2935} 2936 2937/* See uma.h */ 2938/* XXX uk_allocf is not actually used with the zone locked */ 2939void 2940uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 2941{ 2942 uma_keg_t keg; 2943 2944 keg = zone_first_keg(zone); 2945 KEG_LOCK(keg); 2946 keg->uk_allocf = allocf; 2947 KEG_UNLOCK(keg); 2948} 2949 2950/* See uma.h */ 2951void 2952uma_zone_reserve(uma_zone_t zone, int items) 2953{ 2954 uma_keg_t keg; 2955 2956 keg = zone_first_keg(zone); 2957 if (keg == NULL) 2958 return; 2959 KEG_LOCK(keg); 2960 keg->uk_reserve = items; 2961 KEG_UNLOCK(keg); 2962 2963 return; 2964} 2965 2966/* See uma.h */ 2967int 2968uma_zone_reserve_kva(uma_zone_t zone, int count) 2969{ 2970 uma_keg_t keg; 2971 vm_offset_t kva; 2972 int pages; 2973 2974 keg = zone_first_keg(zone); 2975 if (keg == NULL) 2976 return (0); 2977 pages = count / keg->uk_ipers; 2978 2979 if (pages * keg->uk_ipers < count) 2980 pages++; 2981 2982#ifdef UMA_MD_SMALL_ALLOC 2983 if (keg->uk_ppera > 1) { 2984#else 2985 if (1) { 2986#endif 2987 kva = kva_alloc(pages * UMA_SLAB_SIZE); 2988 if (kva == 0) 2989 return (0); 2990 } else 2991 kva = 0; 2992 KEG_LOCK(keg); 2993 keg->uk_kva = kva; 2994 keg->uk_offset = 0; 2995 keg->uk_maxpages = pages; 2996#ifdef UMA_MD_SMALL_ALLOC 2997 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 2998#else 2999 keg->uk_allocf = noobj_alloc; 3000#endif 3001 keg->uk_flags |= UMA_ZONE_NOFREE; 3002 KEG_UNLOCK(keg); 3003 3004 return (1); 3005} 3006 3007/* See uma.h */ 3008void 3009uma_prealloc(uma_zone_t zone, int items) 3010{ 3011 int slabs; 3012 uma_slab_t slab; 3013 uma_keg_t keg; 3014 3015 keg = zone_first_keg(zone); 3016 if (keg == NULL) 3017 return; 3018 KEG_LOCK(keg); 3019 slabs = items / keg->uk_ipers; 3020 if (slabs * keg->uk_ipers < items) 3021 slabs++; 3022 while (slabs > 0) { 3023 slab = keg_alloc_slab(keg, zone, M_WAITOK); 3024 if (slab == NULL) 3025 break; 3026 MPASS(slab->us_keg == keg); 3027 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 3028 slabs--; 3029 } 3030 KEG_UNLOCK(keg); 3031} 3032 3033/* See uma.h */ 3034uint32_t * 3035uma_find_refcnt(uma_zone_t zone, void *item) 3036{ 3037 uma_slabrefcnt_t slabref; 3038 uma_slab_t slab; 3039 uma_keg_t keg; 3040 uint32_t *refcnt; 3041 int idx; 3042 3043 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK)); 3044 slabref = (uma_slabrefcnt_t)slab; 3045 keg = slab->us_keg; 3046 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT, 3047 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT")); 3048 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3049 refcnt = &slabref->us_refcnt[idx]; 3050 return refcnt; 3051} 3052 3053/* See uma.h */ 3054void 3055uma_reclaim(void) 3056{ 3057#ifdef UMA_DEBUG 3058 printf("UMA: vm asked us to release pages!\n"); 3059#endif 3060 bucket_enable(); 3061 zone_foreach(zone_drain); 3062 /* 3063 * Some slabs may have been freed but this zone will be visited early 3064 * we visit again so that we can free pages that are empty once other 3065 * zones are drained. We have to do the same for buckets. 3066 */ 3067 zone_drain(slabzone); 3068 zone_drain(slabrefzone); 3069 bucket_zone_drain(); 3070} 3071 3072/* See uma.h */ 3073int 3074uma_zone_exhausted(uma_zone_t zone) 3075{ 3076 int full; 3077 3078 ZONE_LOCK(zone); 3079 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3080 ZONE_UNLOCK(zone); 3081 return (full); 3082} 3083 3084int 3085uma_zone_exhausted_nolock(uma_zone_t zone) 3086{ 3087 return (zone->uz_flags & UMA_ZFLAG_FULL); 3088} 3089 3090void * 3091uma_large_malloc(int size, int wait) 3092{ 3093 void *mem; 3094 uma_slab_t slab; 3095 uint8_t flags; 3096 3097 slab = zone_alloc_item(slabzone, NULL, wait); 3098 if (slab == NULL) 3099 return (NULL); 3100 mem = page_alloc(NULL, size, &flags, wait); 3101 if (mem) { 3102 vsetslab((vm_offset_t)mem, slab); 3103 slab->us_data = mem; 3104 slab->us_flags = flags | UMA_SLAB_MALLOC; 3105 slab->us_size = size; 3106 } else { 3107 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3108 } 3109 3110 return (mem); 3111} 3112 3113void 3114uma_large_free(uma_slab_t slab) 3115{ 3116 3117 page_free(slab->us_data, slab->us_size, slab->us_flags); 3118 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3119} 3120 3121void 3122uma_print_stats(void) 3123{ 3124 zone_foreach(uma_print_zone); 3125} 3126 3127static void 3128slab_print(uma_slab_t slab) 3129{ 3130 printf("slab: keg %p, data %p, freecount %d\n", 3131 slab->us_keg, slab->us_data, slab->us_freecount); 3132} 3133 3134static void 3135cache_print(uma_cache_t cache) 3136{ 3137 printf("alloc: %p(%d), free: %p(%d)\n", 3138 cache->uc_allocbucket, 3139 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3140 cache->uc_freebucket, 3141 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3142} 3143 3144static void 3145uma_print_keg(uma_keg_t keg) 3146{ 3147 uma_slab_t slab; 3148 3149 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3150 "out %d free %d limit %d\n", 3151 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3152 keg->uk_ipers, keg->uk_ppera, 3153 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free, 3154 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3155 printf("Part slabs:\n"); 3156 LIST_FOREACH(slab, &keg->uk_part_slab, us_link) 3157 slab_print(slab); 3158 printf("Free slabs:\n"); 3159 LIST_FOREACH(slab, &keg->uk_free_slab, us_link) 3160 slab_print(slab); 3161 printf("Full slabs:\n"); 3162 LIST_FOREACH(slab, &keg->uk_full_slab, us_link) 3163 slab_print(slab); 3164} 3165 3166void 3167uma_print_zone(uma_zone_t zone) 3168{ 3169 uma_cache_t cache; 3170 uma_klink_t kl; 3171 int i; 3172 3173 printf("zone: %s(%p) size %d flags %#x\n", 3174 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3175 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3176 uma_print_keg(kl->kl_keg); 3177 CPU_FOREACH(i) { 3178 cache = &zone->uz_cpu[i]; 3179 printf("CPU %d Cache:\n", i); 3180 cache_print(cache); 3181 } 3182} 3183 3184#ifdef DDB 3185/* 3186 * Generate statistics across both the zone and its per-cpu cache's. Return 3187 * desired statistics if the pointer is non-NULL for that statistic. 3188 * 3189 * Note: does not update the zone statistics, as it can't safely clear the 3190 * per-CPU cache statistic. 3191 * 3192 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3193 * safe from off-CPU; we should modify the caches to track this information 3194 * directly so that we don't have to. 3195 */ 3196static void 3197uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3198 uint64_t *freesp, uint64_t *sleepsp) 3199{ 3200 uma_cache_t cache; 3201 uint64_t allocs, frees, sleeps; 3202 int cachefree, cpu; 3203 3204 allocs = frees = sleeps = 0; 3205 cachefree = 0; 3206 CPU_FOREACH(cpu) { 3207 cache = &z->uz_cpu[cpu]; 3208 if (cache->uc_allocbucket != NULL) 3209 cachefree += cache->uc_allocbucket->ub_cnt; 3210 if (cache->uc_freebucket != NULL) 3211 cachefree += cache->uc_freebucket->ub_cnt; 3212 allocs += cache->uc_allocs; 3213 frees += cache->uc_frees; 3214 } 3215 allocs += z->uz_allocs; 3216 frees += z->uz_frees; 3217 sleeps += z->uz_sleeps; 3218 if (cachefreep != NULL) 3219 *cachefreep = cachefree; 3220 if (allocsp != NULL) 3221 *allocsp = allocs; 3222 if (freesp != NULL) 3223 *freesp = frees; 3224 if (sleepsp != NULL) 3225 *sleepsp = sleeps; 3226} 3227#endif /* DDB */ 3228 3229static int 3230sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3231{ 3232 uma_keg_t kz; 3233 uma_zone_t z; 3234 int count; 3235 3236 count = 0; 3237 mtx_lock(&uma_mtx); 3238 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3239 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3240 count++; 3241 } 3242 mtx_unlock(&uma_mtx); 3243 return (sysctl_handle_int(oidp, &count, 0, req)); 3244} 3245 3246static int 3247sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3248{ 3249 struct uma_stream_header ush; 3250 struct uma_type_header uth; 3251 struct uma_percpu_stat ups; 3252 uma_bucket_t bucket; 3253 struct sbuf sbuf; 3254 uma_cache_t cache; 3255 uma_klink_t kl; 3256 uma_keg_t kz; 3257 uma_zone_t z; 3258 uma_keg_t k; 3259 int count, error, i; 3260 3261 error = sysctl_wire_old_buffer(req, 0); 3262 if (error != 0) 3263 return (error); 3264 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3265 3266 count = 0; 3267 mtx_lock(&uma_mtx); 3268 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3269 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3270 count++; 3271 } 3272 3273 /* 3274 * Insert stream header. 3275 */ 3276 bzero(&ush, sizeof(ush)); 3277 ush.ush_version = UMA_STREAM_VERSION; 3278 ush.ush_maxcpus = (mp_maxid + 1); 3279 ush.ush_count = count; 3280 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3281 3282 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3283 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3284 bzero(&uth, sizeof(uth)); 3285 ZONE_LOCK(z); 3286 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3287 uth.uth_align = kz->uk_align; 3288 uth.uth_size = kz->uk_size; 3289 uth.uth_rsize = kz->uk_rsize; 3290 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3291 k = kl->kl_keg; 3292 uth.uth_maxpages += k->uk_maxpages; 3293 uth.uth_pages += k->uk_pages; 3294 uth.uth_keg_free += k->uk_free; 3295 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3296 * k->uk_ipers; 3297 } 3298 3299 /* 3300 * A zone is secondary is it is not the first entry 3301 * on the keg's zone list. 3302 */ 3303 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3304 (LIST_FIRST(&kz->uk_zones) != z)) 3305 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3306 3307 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3308 uth.uth_zone_free += bucket->ub_cnt; 3309 uth.uth_allocs = z->uz_allocs; 3310 uth.uth_frees = z->uz_frees; 3311 uth.uth_fails = z->uz_fails; 3312 uth.uth_sleeps = z->uz_sleeps; 3313 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 3314 /* 3315 * While it is not normally safe to access the cache 3316 * bucket pointers while not on the CPU that owns the 3317 * cache, we only allow the pointers to be exchanged 3318 * without the zone lock held, not invalidated, so 3319 * accept the possible race associated with bucket 3320 * exchange during monitoring. 3321 */ 3322 for (i = 0; i < (mp_maxid + 1); i++) { 3323 bzero(&ups, sizeof(ups)); 3324 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) 3325 goto skip; 3326 if (CPU_ABSENT(i)) 3327 goto skip; 3328 cache = &z->uz_cpu[i]; 3329 if (cache->uc_allocbucket != NULL) 3330 ups.ups_cache_free += 3331 cache->uc_allocbucket->ub_cnt; 3332 if (cache->uc_freebucket != NULL) 3333 ups.ups_cache_free += 3334 cache->uc_freebucket->ub_cnt; 3335 ups.ups_allocs = cache->uc_allocs; 3336 ups.ups_frees = cache->uc_frees; 3337skip: 3338 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups)); 3339 } 3340 ZONE_UNLOCK(z); 3341 } 3342 } 3343 mtx_unlock(&uma_mtx); 3344 error = sbuf_finish(&sbuf); 3345 sbuf_delete(&sbuf); 3346 return (error); 3347} 3348 3349#ifdef DDB 3350DB_SHOW_COMMAND(uma, db_show_uma) 3351{ 3352 uint64_t allocs, frees, sleeps; 3353 uma_bucket_t bucket; 3354 uma_keg_t kz; 3355 uma_zone_t z; 3356 int cachefree; 3357 3358 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 3359 "Requests", "Sleeps"); 3360 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3361 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3362 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 3363 allocs = z->uz_allocs; 3364 frees = z->uz_frees; 3365 sleeps = z->uz_sleeps; 3366 cachefree = 0; 3367 } else 3368 uma_zone_sumstat(z, &cachefree, &allocs, 3369 &frees, &sleeps); 3370 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 3371 (LIST_FIRST(&kz->uk_zones) != z))) 3372 cachefree += kz->uk_free; 3373 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3374 cachefree += bucket->ub_cnt; 3375 db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name, 3376 (uintmax_t)kz->uk_size, 3377 (intmax_t)(allocs - frees), cachefree, 3378 (uintmax_t)allocs, sleeps); 3379 if (db_pager_quit) 3380 return; 3381 } 3382 } 3383} 3384#endif 3385