uma_core.c revision 314667
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 314667 2017-03-04 13:03:31Z avg $"); 52 53/* I should really use ktr.. */ 54/* 55#define UMA_DEBUG 1 56#define UMA_DEBUG_ALLOC 1 57#define UMA_DEBUG_ALLOC_1 1 58*/ 59 60#include "opt_ddb.h" 61#include "opt_param.h" 62#include "opt_vm.h" 63 64#include <sys/param.h> 65#include <sys/systm.h> 66#include <sys/bitset.h> 67#include <sys/eventhandler.h> 68#include <sys/kernel.h> 69#include <sys/types.h> 70#include <sys/queue.h> 71#include <sys/malloc.h> 72#include <sys/ktr.h> 73#include <sys/lock.h> 74#include <sys/sysctl.h> 75#include <sys/mutex.h> 76#include <sys/proc.h> 77#include <sys/rwlock.h> 78#include <sys/sbuf.h> 79#include <sys/sched.h> 80#include <sys/smp.h> 81#include <sys/vmmeter.h> 82 83#include <vm/vm.h> 84#include <vm/vm_object.h> 85#include <vm/vm_page.h> 86#include <vm/vm_pageout.h> 87#include <vm/vm_param.h> 88#include <vm/vm_map.h> 89#include <vm/vm_kern.h> 90#include <vm/vm_extern.h> 91#include <vm/uma.h> 92#include <vm/uma_int.h> 93#include <vm/uma_dbg.h> 94 95#include <ddb/ddb.h> 96 97#ifdef DEBUG_MEMGUARD 98#include <vm/memguard.h> 99#endif 100 101/* 102 * This is the zone and keg from which all zones are spawned. The idea is that 103 * even the zone & keg heads are allocated from the allocator, so we use the 104 * bss section to bootstrap us. 105 */ 106static struct uma_keg masterkeg; 107static struct uma_zone masterzone_k; 108static struct uma_zone masterzone_z; 109static uma_zone_t kegs = &masterzone_k; 110static uma_zone_t zones = &masterzone_z; 111 112/* This is the zone from which all of uma_slab_t's are allocated. */ 113static uma_zone_t slabzone; 114static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */ 115 116/* 117 * The initial hash tables come out of this zone so they can be allocated 118 * prior to malloc coming up. 119 */ 120static uma_zone_t hashzone; 121 122/* The boot-time adjusted value for cache line alignment. */ 123int uma_align_cache = 64 - 1; 124 125static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 126 127/* 128 * Are we allowed to allocate buckets? 129 */ 130static int bucketdisable = 1; 131 132/* Linked list of all kegs in the system */ 133static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 134 135/* Linked list of all cache-only zones in the system */ 136static LIST_HEAD(,uma_zone) uma_cachezones = 137 LIST_HEAD_INITIALIZER(uma_cachezones); 138 139/* This RW lock protects the keg list */ 140static struct rwlock_padalign uma_rwlock; 141 142/* Linked list of boot time pages */ 143static LIST_HEAD(,uma_slab) uma_boot_pages = 144 LIST_HEAD_INITIALIZER(uma_boot_pages); 145 146/* This mutex protects the boot time pages list */ 147static struct mtx_padalign uma_boot_pages_mtx; 148 149static struct sx uma_drain_lock; 150 151/* Is the VM done starting up? */ 152static int booted = 0; 153#define UMA_STARTUP 1 154#define UMA_STARTUP2 2 155 156/* 157 * Only mbuf clusters use ref zones. Just provide enough references 158 * to support the one user. New code should not use the ref facility. 159 */ 160static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES; 161 162/* 163 * This is the handle used to schedule events that need to happen 164 * outside of the allocation fast path. 165 */ 166static struct callout uma_callout; 167#define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 168 169/* 170 * This structure is passed as the zone ctor arg so that I don't have to create 171 * a special allocation function just for zones. 172 */ 173struct uma_zctor_args { 174 const char *name; 175 size_t size; 176 uma_ctor ctor; 177 uma_dtor dtor; 178 uma_init uminit; 179 uma_fini fini; 180 uma_import import; 181 uma_release release; 182 void *arg; 183 uma_keg_t keg; 184 int align; 185 uint32_t flags; 186}; 187 188struct uma_kctor_args { 189 uma_zone_t zone; 190 size_t size; 191 uma_init uminit; 192 uma_fini fini; 193 int align; 194 uint32_t flags; 195}; 196 197struct uma_bucket_zone { 198 uma_zone_t ubz_zone; 199 char *ubz_name; 200 int ubz_entries; /* Number of items it can hold. */ 201 int ubz_maxsize; /* Maximum allocation size per-item. */ 202}; 203 204/* 205 * Compute the actual number of bucket entries to pack them in power 206 * of two sizes for more efficient space utilization. 207 */ 208#define BUCKET_SIZE(n) \ 209 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 210 211#define BUCKET_MAX BUCKET_SIZE(256) 212 213struct uma_bucket_zone bucket_zones[] = { 214 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 215 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 216 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 217 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 218 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 219 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 220 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 221 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 222 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 223 { NULL, NULL, 0} 224}; 225 226/* 227 * Flags and enumerations to be passed to internal functions. 228 */ 229enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; 230 231/* Prototypes.. */ 232 233static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 234static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 235static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 236static void page_free(void *, vm_size_t, uint8_t); 237static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); 238static void cache_drain(uma_zone_t); 239static void bucket_drain(uma_zone_t, uma_bucket_t); 240static void bucket_cache_drain(uma_zone_t zone); 241static int keg_ctor(void *, int, void *, int); 242static void keg_dtor(void *, int, void *); 243static int zone_ctor(void *, int, void *, int); 244static void zone_dtor(void *, int, void *); 245static int zero_init(void *, int, int); 246static void keg_small_init(uma_keg_t keg); 247static void keg_large_init(uma_keg_t keg); 248static void zone_foreach(void (*zfunc)(uma_zone_t)); 249static void zone_timeout(uma_zone_t zone); 250static int hash_alloc(struct uma_hash *); 251static int hash_expand(struct uma_hash *, struct uma_hash *); 252static void hash_free(struct uma_hash *hash); 253static void uma_timeout(void *); 254static void uma_startup3(void); 255static void *zone_alloc_item(uma_zone_t, void *, int); 256static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 257static void bucket_enable(void); 258static void bucket_init(void); 259static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 260static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 261static void bucket_zone_drain(void); 262static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags); 263static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); 264static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); 265static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 266static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); 267static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 268 uma_fini fini, int align, uint32_t flags); 269static int zone_import(uma_zone_t zone, void **bucket, int max, int flags); 270static void zone_release(uma_zone_t zone, void **bucket, int cnt); 271static void uma_zero_item(void *item, uma_zone_t zone); 272 273void uma_print_zone(uma_zone_t); 274void uma_print_stats(void); 275static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 276static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 277 278SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 279 280SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 281 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 282 283SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 284 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 285 286static int zone_warnings = 1; 287TUNABLE_INT("vm.zone_warnings", &zone_warnings); 288SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0, 289 "Warn when UMA zones becomes full"); 290 291/* 292 * This routine checks to see whether or not it's safe to enable buckets. 293 */ 294static void 295bucket_enable(void) 296{ 297 bucketdisable = vm_page_count_min(); 298} 299 300/* 301 * Initialize bucket_zones, the array of zones of buckets of various sizes. 302 * 303 * For each zone, calculate the memory required for each bucket, consisting 304 * of the header and an array of pointers. 305 */ 306static void 307bucket_init(void) 308{ 309 struct uma_bucket_zone *ubz; 310 int size; 311 int i; 312 313 for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 314 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 315 size += sizeof(void *) * ubz->ubz_entries; 316 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 317 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 318 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET); 319 } 320} 321 322/* 323 * Given a desired number of entries for a bucket, return the zone from which 324 * to allocate the bucket. 325 */ 326static struct uma_bucket_zone * 327bucket_zone_lookup(int entries) 328{ 329 struct uma_bucket_zone *ubz; 330 331 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 332 if (ubz->ubz_entries >= entries) 333 return (ubz); 334 ubz--; 335 return (ubz); 336} 337 338static int 339bucket_select(int size) 340{ 341 struct uma_bucket_zone *ubz; 342 343 ubz = &bucket_zones[0]; 344 if (size > ubz->ubz_maxsize) 345 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 346 347 for (; ubz->ubz_entries != 0; ubz++) 348 if (ubz->ubz_maxsize < size) 349 break; 350 ubz--; 351 return (ubz->ubz_entries); 352} 353 354static uma_bucket_t 355bucket_alloc(uma_zone_t zone, void *udata, int flags) 356{ 357 struct uma_bucket_zone *ubz; 358 uma_bucket_t bucket; 359 360 /* 361 * This is to stop us from allocating per cpu buckets while we're 362 * running out of vm.boot_pages. Otherwise, we would exhaust the 363 * boot pages. This also prevents us from allocating buckets in 364 * low memory situations. 365 */ 366 if (bucketdisable) 367 return (NULL); 368 /* 369 * To limit bucket recursion we store the original zone flags 370 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 371 * NOVM flag to persist even through deep recursions. We also 372 * store ZFLAG_BUCKET once we have recursed attempting to allocate 373 * a bucket for a bucket zone so we do not allow infinite bucket 374 * recursion. This cookie will even persist to frees of unused 375 * buckets via the allocation path or bucket allocations in the 376 * free path. 377 */ 378 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 379 udata = (void *)(uintptr_t)zone->uz_flags; 380 else { 381 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 382 return (NULL); 383 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 384 } 385 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 386 flags |= M_NOVM; 387 ubz = bucket_zone_lookup(zone->uz_count); 388 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 389 ubz++; 390 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 391 if (bucket) { 392#ifdef INVARIANTS 393 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 394#endif 395 bucket->ub_cnt = 0; 396 bucket->ub_entries = ubz->ubz_entries; 397 } 398 399 return (bucket); 400} 401 402static void 403bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 404{ 405 struct uma_bucket_zone *ubz; 406 407 KASSERT(bucket->ub_cnt == 0, 408 ("bucket_free: Freeing a non free bucket.")); 409 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 410 udata = (void *)(uintptr_t)zone->uz_flags; 411 ubz = bucket_zone_lookup(bucket->ub_entries); 412 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 413} 414 415static void 416bucket_zone_drain(void) 417{ 418 struct uma_bucket_zone *ubz; 419 420 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 421 zone_drain(ubz->ubz_zone); 422} 423 424static void 425zone_log_warning(uma_zone_t zone) 426{ 427 static const struct timeval warninterval = { 300, 0 }; 428 429 if (!zone_warnings || zone->uz_warning == NULL) 430 return; 431 432 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 433 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 434} 435 436static void 437zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 438{ 439 uma_klink_t klink; 440 441 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 442 kegfn(klink->kl_keg); 443} 444 445/* 446 * Routine called by timeout which is used to fire off some time interval 447 * based calculations. (stats, hash size, etc.) 448 * 449 * Arguments: 450 * arg Unused 451 * 452 * Returns: 453 * Nothing 454 */ 455static void 456uma_timeout(void *unused) 457{ 458 bucket_enable(); 459 zone_foreach(zone_timeout); 460 461 /* Reschedule this event */ 462 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 463} 464 465/* 466 * Routine to perform timeout driven calculations. This expands the 467 * hashes and does per cpu statistics aggregation. 468 * 469 * Returns nothing. 470 */ 471static void 472keg_timeout(uma_keg_t keg) 473{ 474 475 KEG_LOCK(keg); 476 /* 477 * Expand the keg hash table. 478 * 479 * This is done if the number of slabs is larger than the hash size. 480 * What I'm trying to do here is completely reduce collisions. This 481 * may be a little aggressive. Should I allow for two collisions max? 482 */ 483 if (keg->uk_flags & UMA_ZONE_HASH && 484 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 485 struct uma_hash newhash; 486 struct uma_hash oldhash; 487 int ret; 488 489 /* 490 * This is so involved because allocating and freeing 491 * while the keg lock is held will lead to deadlock. 492 * I have to do everything in stages and check for 493 * races. 494 */ 495 newhash = keg->uk_hash; 496 KEG_UNLOCK(keg); 497 ret = hash_alloc(&newhash); 498 KEG_LOCK(keg); 499 if (ret) { 500 if (hash_expand(&keg->uk_hash, &newhash)) { 501 oldhash = keg->uk_hash; 502 keg->uk_hash = newhash; 503 } else 504 oldhash = newhash; 505 506 KEG_UNLOCK(keg); 507 hash_free(&oldhash); 508 return; 509 } 510 } 511 KEG_UNLOCK(keg); 512} 513 514static void 515zone_timeout(uma_zone_t zone) 516{ 517 518 zone_foreach_keg(zone, &keg_timeout); 519} 520 521/* 522 * Allocate and zero fill the next sized hash table from the appropriate 523 * backing store. 524 * 525 * Arguments: 526 * hash A new hash structure with the old hash size in uh_hashsize 527 * 528 * Returns: 529 * 1 on sucess and 0 on failure. 530 */ 531static int 532hash_alloc(struct uma_hash *hash) 533{ 534 int oldsize; 535 int alloc; 536 537 oldsize = hash->uh_hashsize; 538 539 /* We're just going to go to a power of two greater */ 540 if (oldsize) { 541 hash->uh_hashsize = oldsize * 2; 542 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 543 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 544 M_UMAHASH, M_NOWAIT); 545 } else { 546 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 547 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 548 M_WAITOK); 549 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 550 } 551 if (hash->uh_slab_hash) { 552 bzero(hash->uh_slab_hash, alloc); 553 hash->uh_hashmask = hash->uh_hashsize - 1; 554 return (1); 555 } 556 557 return (0); 558} 559 560/* 561 * Expands the hash table for HASH zones. This is done from zone_timeout 562 * to reduce collisions. This must not be done in the regular allocation 563 * path, otherwise, we can recurse on the vm while allocating pages. 564 * 565 * Arguments: 566 * oldhash The hash you want to expand 567 * newhash The hash structure for the new table 568 * 569 * Returns: 570 * Nothing 571 * 572 * Discussion: 573 */ 574static int 575hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 576{ 577 uma_slab_t slab; 578 int hval; 579 int i; 580 581 if (!newhash->uh_slab_hash) 582 return (0); 583 584 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 585 return (0); 586 587 /* 588 * I need to investigate hash algorithms for resizing without a 589 * full rehash. 590 */ 591 592 for (i = 0; i < oldhash->uh_hashsize; i++) 593 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 594 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 595 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 596 hval = UMA_HASH(newhash, slab->us_data); 597 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 598 slab, us_hlink); 599 } 600 601 return (1); 602} 603 604/* 605 * Free the hash bucket to the appropriate backing store. 606 * 607 * Arguments: 608 * slab_hash The hash bucket we're freeing 609 * hashsize The number of entries in that hash bucket 610 * 611 * Returns: 612 * Nothing 613 */ 614static void 615hash_free(struct uma_hash *hash) 616{ 617 if (hash->uh_slab_hash == NULL) 618 return; 619 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 620 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 621 else 622 free(hash->uh_slab_hash, M_UMAHASH); 623} 624 625/* 626 * Frees all outstanding items in a bucket 627 * 628 * Arguments: 629 * zone The zone to free to, must be unlocked. 630 * bucket The free/alloc bucket with items, cpu queue must be locked. 631 * 632 * Returns: 633 * Nothing 634 */ 635 636static void 637bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 638{ 639 int i; 640 641 if (bucket == NULL) 642 return; 643 644 if (zone->uz_fini) 645 for (i = 0; i < bucket->ub_cnt; i++) 646 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 647 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 648 bucket->ub_cnt = 0; 649} 650 651/* 652 * Drains the per cpu caches for a zone. 653 * 654 * NOTE: This may only be called while the zone is being turn down, and not 655 * during normal operation. This is necessary in order that we do not have 656 * to migrate CPUs to drain the per-CPU caches. 657 * 658 * Arguments: 659 * zone The zone to drain, must be unlocked. 660 * 661 * Returns: 662 * Nothing 663 */ 664static void 665cache_drain(uma_zone_t zone) 666{ 667 uma_cache_t cache; 668 int cpu; 669 670 /* 671 * XXX: It is safe to not lock the per-CPU caches, because we're 672 * tearing down the zone anyway. I.e., there will be no further use 673 * of the caches at this point. 674 * 675 * XXX: It would good to be able to assert that the zone is being 676 * torn down to prevent improper use of cache_drain(). 677 * 678 * XXX: We lock the zone before passing into bucket_cache_drain() as 679 * it is used elsewhere. Should the tear-down path be made special 680 * there in some form? 681 */ 682 CPU_FOREACH(cpu) { 683 cache = &zone->uz_cpu[cpu]; 684 bucket_drain(zone, cache->uc_allocbucket); 685 bucket_drain(zone, cache->uc_freebucket); 686 if (cache->uc_allocbucket != NULL) 687 bucket_free(zone, cache->uc_allocbucket, NULL); 688 if (cache->uc_freebucket != NULL) 689 bucket_free(zone, cache->uc_freebucket, NULL); 690 cache->uc_allocbucket = cache->uc_freebucket = NULL; 691 } 692 ZONE_LOCK(zone); 693 bucket_cache_drain(zone); 694 ZONE_UNLOCK(zone); 695} 696 697static void 698cache_shrink(uma_zone_t zone) 699{ 700 701 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 702 return; 703 704 ZONE_LOCK(zone); 705 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; 706 ZONE_UNLOCK(zone); 707} 708 709static void 710cache_drain_safe_cpu(uma_zone_t zone) 711{ 712 uma_cache_t cache; 713 uma_bucket_t b1, b2; 714 715 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 716 return; 717 718 b1 = b2 = NULL; 719 ZONE_LOCK(zone); 720 critical_enter(); 721 cache = &zone->uz_cpu[curcpu]; 722 if (cache->uc_allocbucket) { 723 if (cache->uc_allocbucket->ub_cnt != 0) 724 LIST_INSERT_HEAD(&zone->uz_buckets, 725 cache->uc_allocbucket, ub_link); 726 else 727 b1 = cache->uc_allocbucket; 728 cache->uc_allocbucket = NULL; 729 } 730 if (cache->uc_freebucket) { 731 if (cache->uc_freebucket->ub_cnt != 0) 732 LIST_INSERT_HEAD(&zone->uz_buckets, 733 cache->uc_freebucket, ub_link); 734 else 735 b2 = cache->uc_freebucket; 736 cache->uc_freebucket = NULL; 737 } 738 critical_exit(); 739 ZONE_UNLOCK(zone); 740 if (b1) 741 bucket_free(zone, b1, NULL); 742 if (b2) 743 bucket_free(zone, b2, NULL); 744} 745 746/* 747 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 748 * This is an expensive call because it needs to bind to all CPUs 749 * one by one and enter a critical section on each of them in order 750 * to safely access their cache buckets. 751 * Zone lock must not be held on call this function. 752 */ 753static void 754cache_drain_safe(uma_zone_t zone) 755{ 756 int cpu; 757 758 /* 759 * Polite bucket sizes shrinking was not enouth, shrink aggressively. 760 */ 761 if (zone) 762 cache_shrink(zone); 763 else 764 zone_foreach(cache_shrink); 765 766 CPU_FOREACH(cpu) { 767 thread_lock(curthread); 768 sched_bind(curthread, cpu); 769 thread_unlock(curthread); 770 771 if (zone) 772 cache_drain_safe_cpu(zone); 773 else 774 zone_foreach(cache_drain_safe_cpu); 775 } 776 thread_lock(curthread); 777 sched_unbind(curthread); 778 thread_unlock(curthread); 779} 780 781/* 782 * Drain the cached buckets from a zone. Expects a locked zone on entry. 783 */ 784static void 785bucket_cache_drain(uma_zone_t zone) 786{ 787 uma_bucket_t bucket; 788 789 /* 790 * Drain the bucket queues and free the buckets, we just keep two per 791 * cpu (alloc/free). 792 */ 793 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 794 LIST_REMOVE(bucket, ub_link); 795 ZONE_UNLOCK(zone); 796 bucket_drain(zone, bucket); 797 bucket_free(zone, bucket, NULL); 798 ZONE_LOCK(zone); 799 } 800 801 /* 802 * Shrink further bucket sizes. Price of single zone lock collision 803 * is probably lower then price of global cache drain. 804 */ 805 if (zone->uz_count > zone->uz_count_min) 806 zone->uz_count--; 807} 808 809static void 810keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 811{ 812 uint8_t *mem; 813 int i; 814 uint8_t flags; 815 816 mem = slab->us_data; 817 flags = slab->us_flags; 818 i = start; 819 if (keg->uk_fini != NULL) { 820 for (i--; i > -1; i--) 821 keg->uk_fini(slab->us_data + (keg->uk_rsize * i), 822 keg->uk_size); 823 } 824 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 825 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 826#ifdef UMA_DEBUG 827 printf("%s: Returning %d bytes.\n", keg->uk_name, 828 PAGE_SIZE * keg->uk_ppera); 829#endif 830 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 831} 832 833/* 834 * Frees pages from a keg back to the system. This is done on demand from 835 * the pageout daemon. 836 * 837 * Returns nothing. 838 */ 839static void 840keg_drain(uma_keg_t keg) 841{ 842 struct slabhead freeslabs = { 0 }; 843 uma_slab_t slab; 844 uma_slab_t n; 845 846 /* 847 * We don't want to take pages from statically allocated kegs at this 848 * time 849 */ 850 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 851 return; 852 853#ifdef UMA_DEBUG 854 printf("%s free items: %u\n", keg->uk_name, keg->uk_free); 855#endif 856 KEG_LOCK(keg); 857 if (keg->uk_free == 0) 858 goto finished; 859 860 slab = LIST_FIRST(&keg->uk_free_slab); 861 while (slab) { 862 n = LIST_NEXT(slab, us_link); 863 864 /* We have no where to free these to */ 865 if (slab->us_flags & UMA_SLAB_BOOT) { 866 slab = n; 867 continue; 868 } 869 870 LIST_REMOVE(slab, us_link); 871 keg->uk_pages -= keg->uk_ppera; 872 keg->uk_free -= keg->uk_ipers; 873 874 if (keg->uk_flags & UMA_ZONE_HASH) 875 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); 876 877 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 878 879 slab = n; 880 } 881finished: 882 KEG_UNLOCK(keg); 883 884 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 885 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 886 keg_free_slab(keg, slab, keg->uk_ipers); 887 } 888} 889 890static void 891zone_drain_wait(uma_zone_t zone, int waitok) 892{ 893 894 /* 895 * Set draining to interlock with zone_dtor() so we can release our 896 * locks as we go. Only dtor() should do a WAITOK call since it 897 * is the only call that knows the structure will still be available 898 * when it wakes up. 899 */ 900 ZONE_LOCK(zone); 901 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 902 if (waitok == M_NOWAIT) 903 goto out; 904 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 905 } 906 zone->uz_flags |= UMA_ZFLAG_DRAINING; 907 bucket_cache_drain(zone); 908 ZONE_UNLOCK(zone); 909 /* 910 * The DRAINING flag protects us from being freed while 911 * we're running. Normally the uma_rwlock would protect us but we 912 * must be able to release and acquire the right lock for each keg. 913 */ 914 zone_foreach_keg(zone, &keg_drain); 915 ZONE_LOCK(zone); 916 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 917 wakeup(zone); 918out: 919 ZONE_UNLOCK(zone); 920} 921 922void 923zone_drain(uma_zone_t zone) 924{ 925 926 zone_drain_wait(zone, M_NOWAIT); 927} 928 929/* 930 * Allocate a new slab for a keg. This does not insert the slab onto a list. 931 * 932 * Arguments: 933 * wait Shall we wait? 934 * 935 * Returns: 936 * The slab that was allocated or NULL if there is no memory and the 937 * caller specified M_NOWAIT. 938 */ 939static uma_slab_t 940keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) 941{ 942 uma_slabrefcnt_t slabref; 943 uma_alloc allocf; 944 uma_slab_t slab; 945 uint8_t *mem; 946 uint8_t flags; 947 int i; 948 949 mtx_assert(&keg->uk_lock, MA_OWNED); 950 slab = NULL; 951 mem = NULL; 952 953#ifdef UMA_DEBUG 954 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name); 955#endif 956 allocf = keg->uk_allocf; 957 KEG_UNLOCK(keg); 958 959 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 960 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); 961 if (slab == NULL) 962 goto out; 963 } 964 965 /* 966 * This reproduces the old vm_zone behavior of zero filling pages the 967 * first time they are added to a zone. 968 * 969 * Malloced items are zeroed in uma_zalloc. 970 */ 971 972 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 973 wait |= M_ZERO; 974 else 975 wait &= ~M_ZERO; 976 977 if (keg->uk_flags & UMA_ZONE_NODUMP) 978 wait |= M_NODUMP; 979 980 /* zone is passed for legacy reasons. */ 981 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait); 982 if (mem == NULL) { 983 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 984 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 985 slab = NULL; 986 goto out; 987 } 988 989 /* Point the slab into the allocated memory */ 990 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 991 slab = (uma_slab_t )(mem + keg->uk_pgoff); 992 993 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 994 for (i = 0; i < keg->uk_ppera; i++) 995 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 996 997 slab->us_keg = keg; 998 slab->us_data = mem; 999 slab->us_freecount = keg->uk_ipers; 1000 slab->us_flags = flags; 1001 BIT_FILL(SLAB_SETSIZE, &slab->us_free); 1002#ifdef INVARIANTS 1003 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); 1004#endif 1005 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1006 slabref = (uma_slabrefcnt_t)slab; 1007 for (i = 0; i < keg->uk_ipers; i++) 1008 slabref->us_refcnt[i] = 0; 1009 } 1010 1011 if (keg->uk_init != NULL) { 1012 for (i = 0; i < keg->uk_ipers; i++) 1013 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 1014 keg->uk_size, wait) != 0) 1015 break; 1016 if (i != keg->uk_ipers) { 1017 keg_free_slab(keg, slab, i); 1018 slab = NULL; 1019 goto out; 1020 } 1021 } 1022out: 1023 KEG_LOCK(keg); 1024 1025 if (slab != NULL) { 1026 if (keg->uk_flags & UMA_ZONE_HASH) 1027 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1028 1029 keg->uk_pages += keg->uk_ppera; 1030 keg->uk_free += keg->uk_ipers; 1031 } 1032 1033 return (slab); 1034} 1035 1036/* 1037 * This function is intended to be used early on in place of page_alloc() so 1038 * that we may use the boot time page cache to satisfy allocations before 1039 * the VM is ready. 1040 */ 1041static void * 1042startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) 1043{ 1044 uma_keg_t keg; 1045 uma_slab_t tmps; 1046 int pages, check_pages; 1047 1048 keg = zone_first_keg(zone); 1049 pages = howmany(bytes, PAGE_SIZE); 1050 check_pages = pages - 1; 1051 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n")); 1052 1053 /* 1054 * Check our small startup cache to see if it has pages remaining. 1055 */ 1056 mtx_lock(&uma_boot_pages_mtx); 1057 1058 /* First check if we have enough room. */ 1059 tmps = LIST_FIRST(&uma_boot_pages); 1060 while (tmps != NULL && check_pages-- > 0) 1061 tmps = LIST_NEXT(tmps, us_link); 1062 if (tmps != NULL) { 1063 /* 1064 * It's ok to lose tmps references. The last one will 1065 * have tmps->us_data pointing to the start address of 1066 * "pages" contiguous pages of memory. 1067 */ 1068 while (pages-- > 0) { 1069 tmps = LIST_FIRST(&uma_boot_pages); 1070 LIST_REMOVE(tmps, us_link); 1071 } 1072 mtx_unlock(&uma_boot_pages_mtx); 1073 *pflag = tmps->us_flags; 1074 return (tmps->us_data); 1075 } 1076 mtx_unlock(&uma_boot_pages_mtx); 1077 if (booted < UMA_STARTUP2) 1078 panic("UMA: Increase vm.boot_pages"); 1079 /* 1080 * Now that we've booted reset these users to their real allocator. 1081 */ 1082#ifdef UMA_MD_SMALL_ALLOC 1083 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; 1084#else 1085 keg->uk_allocf = page_alloc; 1086#endif 1087 return keg->uk_allocf(zone, bytes, pflag, wait); 1088} 1089 1090/* 1091 * Allocates a number of pages from the system 1092 * 1093 * Arguments: 1094 * bytes The number of bytes requested 1095 * wait Shall we wait? 1096 * 1097 * Returns: 1098 * A pointer to the alloced memory or possibly 1099 * NULL if M_NOWAIT is set. 1100 */ 1101static void * 1102page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) 1103{ 1104 void *p; /* Returned page */ 1105 1106 *pflag = UMA_SLAB_KMEM; 1107 p = (void *) kmem_malloc(kmem_arena, bytes, wait); 1108 1109 return (p); 1110} 1111 1112/* 1113 * Allocates a number of pages from within an object 1114 * 1115 * Arguments: 1116 * bytes The number of bytes requested 1117 * wait Shall we wait? 1118 * 1119 * Returns: 1120 * A pointer to the alloced memory or possibly 1121 * NULL if M_NOWAIT is set. 1122 */ 1123static void * 1124noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait) 1125{ 1126 TAILQ_HEAD(, vm_page) alloctail; 1127 u_long npages; 1128 vm_offset_t retkva, zkva; 1129 vm_page_t p, p_next; 1130 uma_keg_t keg; 1131 1132 TAILQ_INIT(&alloctail); 1133 keg = zone_first_keg(zone); 1134 1135 npages = howmany(bytes, PAGE_SIZE); 1136 while (npages > 0) { 1137 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT | 1138 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ); 1139 if (p != NULL) { 1140 /* 1141 * Since the page does not belong to an object, its 1142 * listq is unused. 1143 */ 1144 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1145 npages--; 1146 continue; 1147 } 1148 if (wait & M_WAITOK) { 1149 VM_WAIT; 1150 continue; 1151 } 1152 1153 /* 1154 * Page allocation failed, free intermediate pages and 1155 * exit. 1156 */ 1157 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1158 vm_page_unwire(p, 0); 1159 vm_page_free(p); 1160 } 1161 return (NULL); 1162 } 1163 *flags = UMA_SLAB_PRIV; 1164 zkva = keg->uk_kva + 1165 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1166 retkva = zkva; 1167 TAILQ_FOREACH(p, &alloctail, listq) { 1168 pmap_qenter(zkva, &p, 1); 1169 zkva += PAGE_SIZE; 1170 } 1171 1172 return ((void *)retkva); 1173} 1174 1175/* 1176 * Frees a number of pages to the system 1177 * 1178 * Arguments: 1179 * mem A pointer to the memory to be freed 1180 * size The size of the memory being freed 1181 * flags The original p->us_flags field 1182 * 1183 * Returns: 1184 * Nothing 1185 */ 1186static void 1187page_free(void *mem, vm_size_t size, uint8_t flags) 1188{ 1189 struct vmem *vmem; 1190 1191 if (flags & UMA_SLAB_KMEM) 1192 vmem = kmem_arena; 1193 else if (flags & UMA_SLAB_KERNEL) 1194 vmem = kernel_arena; 1195 else 1196 panic("UMA: page_free used with invalid flags %d", flags); 1197 1198 kmem_free(vmem, (vm_offset_t)mem, size); 1199} 1200 1201/* 1202 * Zero fill initializer 1203 * 1204 * Arguments/Returns follow uma_init specifications 1205 */ 1206static int 1207zero_init(void *mem, int size, int flags) 1208{ 1209 bzero(mem, size); 1210 return (0); 1211} 1212 1213/* 1214 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1215 * 1216 * Arguments 1217 * keg The zone we should initialize 1218 * 1219 * Returns 1220 * Nothing 1221 */ 1222static void 1223keg_small_init(uma_keg_t keg) 1224{ 1225 u_int rsize; 1226 u_int memused; 1227 u_int wastedspace; 1228 u_int shsize; 1229 1230 if (keg->uk_flags & UMA_ZONE_PCPU) { 1231 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU; 1232 1233 keg->uk_slabsize = sizeof(struct pcpu); 1234 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu), 1235 PAGE_SIZE); 1236 } else { 1237 keg->uk_slabsize = UMA_SLAB_SIZE; 1238 keg->uk_ppera = 1; 1239 } 1240 1241 /* 1242 * Calculate the size of each allocation (rsize) according to 1243 * alignment. If the requested size is smaller than we have 1244 * allocation bits for we round it up. 1245 */ 1246 rsize = keg->uk_size; 1247 if (rsize < keg->uk_slabsize / SLAB_SETSIZE) 1248 rsize = keg->uk_slabsize / SLAB_SETSIZE; 1249 if (rsize & keg->uk_align) 1250 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1251 keg->uk_rsize = rsize; 1252 1253 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1254 keg->uk_rsize < sizeof(struct pcpu), 1255 ("%s: size %u too large", __func__, keg->uk_rsize)); 1256 1257 if (keg->uk_flags & UMA_ZONE_REFCNT) 1258 rsize += sizeof(uint32_t); 1259 1260 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1261 shsize = 0; 1262 else 1263 shsize = sizeof(struct uma_slab); 1264 1265 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize; 1266 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1267 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1268 1269 memused = keg->uk_ipers * rsize + shsize; 1270 wastedspace = keg->uk_slabsize - memused; 1271 1272 /* 1273 * We can't do OFFPAGE if we're internal or if we've been 1274 * asked to not go to the VM for buckets. If we do this we 1275 * may end up going to the VM for slabs which we do not 1276 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1277 * of UMA_ZONE_VM, which clearly forbids it. 1278 */ 1279 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1280 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1281 return; 1282 1283 /* 1284 * See if using an OFFPAGE slab will limit our waste. Only do 1285 * this if it permits more items per-slab. 1286 * 1287 * XXX We could try growing slabsize to limit max waste as well. 1288 * Historically this was not done because the VM could not 1289 * efficiently handle contiguous allocations. 1290 */ 1291 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) && 1292 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) { 1293 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize; 1294 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1295 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1296#ifdef UMA_DEBUG 1297 printf("UMA decided we need offpage slab headers for " 1298 "keg: %s, calculated wastedspace = %d, " 1299 "maximum wasted space allowed = %d, " 1300 "calculated ipers = %d, " 1301 "new wasted space = %d\n", keg->uk_name, wastedspace, 1302 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1303 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize); 1304#endif 1305 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1306 } 1307 1308 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1309 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1310 keg->uk_flags |= UMA_ZONE_HASH; 1311} 1312 1313/* 1314 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1315 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1316 * more complicated. 1317 * 1318 * Arguments 1319 * keg The keg we should initialize 1320 * 1321 * Returns 1322 * Nothing 1323 */ 1324static void 1325keg_large_init(uma_keg_t keg) 1326{ 1327 u_int shsize; 1328 1329 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1330 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1331 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1332 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1333 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1334 1335 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1336 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE; 1337 keg->uk_ipers = 1; 1338 keg->uk_rsize = keg->uk_size; 1339 1340 /* We can't do OFFPAGE if we're internal, bail out here. */ 1341 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) 1342 return; 1343 1344 /* Check whether we have enough space to not do OFFPAGE. */ 1345 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) { 1346 shsize = sizeof(struct uma_slab); 1347 if (keg->uk_flags & UMA_ZONE_REFCNT) 1348 shsize += keg->uk_ipers * sizeof(uint32_t); 1349 if (shsize & UMA_ALIGN_PTR) 1350 shsize = (shsize & ~UMA_ALIGN_PTR) + 1351 (UMA_ALIGN_PTR + 1); 1352 1353 if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize) 1354 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1355 } 1356 1357 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1358 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1359 keg->uk_flags |= UMA_ZONE_HASH; 1360} 1361 1362static void 1363keg_cachespread_init(uma_keg_t keg) 1364{ 1365 int alignsize; 1366 int trailer; 1367 int pages; 1368 int rsize; 1369 1370 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1371 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1372 1373 alignsize = keg->uk_align + 1; 1374 rsize = keg->uk_size; 1375 /* 1376 * We want one item to start on every align boundary in a page. To 1377 * do this we will span pages. We will also extend the item by the 1378 * size of align if it is an even multiple of align. Otherwise, it 1379 * would fall on the same boundary every time. 1380 */ 1381 if (rsize & keg->uk_align) 1382 rsize = (rsize & ~keg->uk_align) + alignsize; 1383 if ((rsize & alignsize) == 0) 1384 rsize += alignsize; 1385 trailer = rsize - keg->uk_size; 1386 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1387 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1388 keg->uk_rsize = rsize; 1389 keg->uk_ppera = pages; 1390 keg->uk_slabsize = UMA_SLAB_SIZE; 1391 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1392 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1393 KASSERT(keg->uk_ipers <= SLAB_SETSIZE, 1394 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1395 keg->uk_ipers)); 1396} 1397 1398/* 1399 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1400 * the keg onto the global keg list. 1401 * 1402 * Arguments/Returns follow uma_ctor specifications 1403 * udata Actually uma_kctor_args 1404 */ 1405static int 1406keg_ctor(void *mem, int size, void *udata, int flags) 1407{ 1408 struct uma_kctor_args *arg = udata; 1409 uma_keg_t keg = mem; 1410 uma_zone_t zone; 1411 1412 bzero(keg, size); 1413 keg->uk_size = arg->size; 1414 keg->uk_init = arg->uminit; 1415 keg->uk_fini = arg->fini; 1416 keg->uk_align = arg->align; 1417 keg->uk_free = 0; 1418 keg->uk_reserve = 0; 1419 keg->uk_pages = 0; 1420 keg->uk_flags = arg->flags; 1421 keg->uk_allocf = page_alloc; 1422 keg->uk_freef = page_free; 1423 keg->uk_slabzone = NULL; 1424 1425 /* 1426 * The master zone is passed to us at keg-creation time. 1427 */ 1428 zone = arg->zone; 1429 keg->uk_name = zone->uz_name; 1430 1431 if (arg->flags & UMA_ZONE_VM) 1432 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1433 1434 if (arg->flags & UMA_ZONE_ZINIT) 1435 keg->uk_init = zero_init; 1436 1437 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC) 1438 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1439 1440 if (arg->flags & UMA_ZONE_PCPU) 1441#ifdef SMP 1442 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1443#else 1444 keg->uk_flags &= ~UMA_ZONE_PCPU; 1445#endif 1446 1447 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1448 keg_cachespread_init(keg); 1449 } else if (keg->uk_flags & UMA_ZONE_REFCNT) { 1450 if (keg->uk_size > 1451 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - 1452 sizeof(uint32_t))) 1453 keg_large_init(keg); 1454 else 1455 keg_small_init(keg); 1456 } else { 1457 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab))) 1458 keg_large_init(keg); 1459 else 1460 keg_small_init(keg); 1461 } 1462 1463 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1464 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1465 if (keg->uk_ipers > uma_max_ipers_ref) 1466 panic("Too many ref items per zone: %d > %d\n", 1467 keg->uk_ipers, uma_max_ipers_ref); 1468 keg->uk_slabzone = slabrefzone; 1469 } else 1470 keg->uk_slabzone = slabzone; 1471 } 1472 1473 /* 1474 * If we haven't booted yet we need allocations to go through the 1475 * startup cache until the vm is ready. 1476 */ 1477 if (keg->uk_ppera == 1) { 1478#ifdef UMA_MD_SMALL_ALLOC 1479 keg->uk_allocf = uma_small_alloc; 1480 keg->uk_freef = uma_small_free; 1481 1482 if (booted < UMA_STARTUP) 1483 keg->uk_allocf = startup_alloc; 1484#else 1485 if (booted < UMA_STARTUP2) 1486 keg->uk_allocf = startup_alloc; 1487#endif 1488 } else if (booted < UMA_STARTUP2 && 1489 (keg->uk_flags & UMA_ZFLAG_INTERNAL)) 1490 keg->uk_allocf = startup_alloc; 1491 1492 /* 1493 * Initialize keg's lock 1494 */ 1495 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1496 1497 /* 1498 * If we're putting the slab header in the actual page we need to 1499 * figure out where in each page it goes. This calculates a right 1500 * justified offset into the memory on an ALIGN_PTR boundary. 1501 */ 1502 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1503 u_int totsize; 1504 1505 /* Size of the slab struct and free list */ 1506 totsize = sizeof(struct uma_slab); 1507 1508 /* Size of the reference counts. */ 1509 if (keg->uk_flags & UMA_ZONE_REFCNT) 1510 totsize += keg->uk_ipers * sizeof(uint32_t); 1511 1512 if (totsize & UMA_ALIGN_PTR) 1513 totsize = (totsize & ~UMA_ALIGN_PTR) + 1514 (UMA_ALIGN_PTR + 1); 1515 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1516 1517 /* 1518 * The only way the following is possible is if with our 1519 * UMA_ALIGN_PTR adjustments we are now bigger than 1520 * UMA_SLAB_SIZE. I haven't checked whether this is 1521 * mathematically possible for all cases, so we make 1522 * sure here anyway. 1523 */ 1524 totsize = keg->uk_pgoff + sizeof(struct uma_slab); 1525 if (keg->uk_flags & UMA_ZONE_REFCNT) 1526 totsize += keg->uk_ipers * sizeof(uint32_t); 1527 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1528 printf("zone %s ipers %d rsize %d size %d\n", 1529 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1530 keg->uk_size); 1531 panic("UMA slab won't fit."); 1532 } 1533 } 1534 1535 if (keg->uk_flags & UMA_ZONE_HASH) 1536 hash_alloc(&keg->uk_hash); 1537 1538#ifdef UMA_DEBUG 1539 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n", 1540 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags, 1541 keg->uk_ipers, keg->uk_ppera, 1542 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free); 1543#endif 1544 1545 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1546 1547 rw_wlock(&uma_rwlock); 1548 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1549 rw_wunlock(&uma_rwlock); 1550 return (0); 1551} 1552 1553/* 1554 * Zone header ctor. This initializes all fields, locks, etc. 1555 * 1556 * Arguments/Returns follow uma_ctor specifications 1557 * udata Actually uma_zctor_args 1558 */ 1559static int 1560zone_ctor(void *mem, int size, void *udata, int flags) 1561{ 1562 struct uma_zctor_args *arg = udata; 1563 uma_zone_t zone = mem; 1564 uma_zone_t z; 1565 uma_keg_t keg; 1566 1567 bzero(zone, size); 1568 zone->uz_name = arg->name; 1569 zone->uz_ctor = arg->ctor; 1570 zone->uz_dtor = arg->dtor; 1571 zone->uz_slab = zone_fetch_slab; 1572 zone->uz_init = NULL; 1573 zone->uz_fini = NULL; 1574 zone->uz_allocs = 0; 1575 zone->uz_frees = 0; 1576 zone->uz_fails = 0; 1577 zone->uz_sleeps = 0; 1578 zone->uz_count = 0; 1579 zone->uz_count_min = 0; 1580 zone->uz_flags = 0; 1581 zone->uz_warning = NULL; 1582 timevalclear(&zone->uz_ratecheck); 1583 keg = arg->keg; 1584 1585 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 1586 1587 /* 1588 * This is a pure cache zone, no kegs. 1589 */ 1590 if (arg->import) { 1591 if (arg->flags & UMA_ZONE_VM) 1592 arg->flags |= UMA_ZFLAG_CACHEONLY; 1593 zone->uz_flags = arg->flags; 1594 zone->uz_size = arg->size; 1595 zone->uz_import = arg->import; 1596 zone->uz_release = arg->release; 1597 zone->uz_arg = arg->arg; 1598 zone->uz_lockptr = &zone->uz_lock; 1599 rw_wlock(&uma_rwlock); 1600 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 1601 rw_wunlock(&uma_rwlock); 1602 goto out; 1603 } 1604 1605 /* 1606 * Use the regular zone/keg/slab allocator. 1607 */ 1608 zone->uz_import = (uma_import)zone_import; 1609 zone->uz_release = (uma_release)zone_release; 1610 zone->uz_arg = zone; 1611 1612 if (arg->flags & UMA_ZONE_SECONDARY) { 1613 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1614 zone->uz_init = arg->uminit; 1615 zone->uz_fini = arg->fini; 1616 zone->uz_lockptr = &keg->uk_lock; 1617 zone->uz_flags |= UMA_ZONE_SECONDARY; 1618 rw_wlock(&uma_rwlock); 1619 ZONE_LOCK(zone); 1620 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1621 if (LIST_NEXT(z, uz_link) == NULL) { 1622 LIST_INSERT_AFTER(z, zone, uz_link); 1623 break; 1624 } 1625 } 1626 ZONE_UNLOCK(zone); 1627 rw_wunlock(&uma_rwlock); 1628 } else if (keg == NULL) { 1629 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1630 arg->align, arg->flags)) == NULL) 1631 return (ENOMEM); 1632 } else { 1633 struct uma_kctor_args karg; 1634 int error; 1635 1636 /* We should only be here from uma_startup() */ 1637 karg.size = arg->size; 1638 karg.uminit = arg->uminit; 1639 karg.fini = arg->fini; 1640 karg.align = arg->align; 1641 karg.flags = arg->flags; 1642 karg.zone = zone; 1643 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1644 flags); 1645 if (error) 1646 return (error); 1647 } 1648 1649 /* 1650 * Link in the first keg. 1651 */ 1652 zone->uz_klink.kl_keg = keg; 1653 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1654 zone->uz_lockptr = &keg->uk_lock; 1655 zone->uz_size = keg->uk_size; 1656 zone->uz_flags |= (keg->uk_flags & 1657 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1658 1659 /* 1660 * Some internal zones don't have room allocated for the per cpu 1661 * caches. If we're internal, bail out here. 1662 */ 1663 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1664 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1665 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1666 return (0); 1667 } 1668 1669out: 1670 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0) 1671 zone->uz_count = bucket_select(zone->uz_size); 1672 else 1673 zone->uz_count = BUCKET_MAX; 1674 zone->uz_count_min = zone->uz_count; 1675 1676 return (0); 1677} 1678 1679/* 1680 * Keg header dtor. This frees all data, destroys locks, frees the hash 1681 * table and removes the keg from the global list. 1682 * 1683 * Arguments/Returns follow uma_dtor specifications 1684 * udata unused 1685 */ 1686static void 1687keg_dtor(void *arg, int size, void *udata) 1688{ 1689 uma_keg_t keg; 1690 1691 keg = (uma_keg_t)arg; 1692 KEG_LOCK(keg); 1693 if (keg->uk_free != 0) { 1694 printf("Freed UMA keg (%s) was not empty (%d items). " 1695 " Lost %d pages of memory.\n", 1696 keg->uk_name ? keg->uk_name : "", 1697 keg->uk_free, keg->uk_pages); 1698 } 1699 KEG_UNLOCK(keg); 1700 1701 hash_free(&keg->uk_hash); 1702 1703 KEG_LOCK_FINI(keg); 1704} 1705 1706/* 1707 * Zone header dtor. 1708 * 1709 * Arguments/Returns follow uma_dtor specifications 1710 * udata unused 1711 */ 1712static void 1713zone_dtor(void *arg, int size, void *udata) 1714{ 1715 uma_klink_t klink; 1716 uma_zone_t zone; 1717 uma_keg_t keg; 1718 1719 zone = (uma_zone_t)arg; 1720 keg = zone_first_keg(zone); 1721 1722 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1723 cache_drain(zone); 1724 1725 rw_wlock(&uma_rwlock); 1726 LIST_REMOVE(zone, uz_link); 1727 rw_wunlock(&uma_rwlock); 1728 /* 1729 * XXX there are some races here where 1730 * the zone can be drained but zone lock 1731 * released and then refilled before we 1732 * remove it... we dont care for now 1733 */ 1734 zone_drain_wait(zone, M_WAITOK); 1735 /* 1736 * Unlink all of our kegs. 1737 */ 1738 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1739 klink->kl_keg = NULL; 1740 LIST_REMOVE(klink, kl_link); 1741 if (klink == &zone->uz_klink) 1742 continue; 1743 free(klink, M_TEMP); 1744 } 1745 /* 1746 * We only destroy kegs from non secondary zones. 1747 */ 1748 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1749 rw_wlock(&uma_rwlock); 1750 LIST_REMOVE(keg, uk_link); 1751 rw_wunlock(&uma_rwlock); 1752 zone_free_item(kegs, keg, NULL, SKIP_NONE); 1753 } 1754 ZONE_LOCK_FINI(zone); 1755} 1756 1757/* 1758 * Traverses every zone in the system and calls a callback 1759 * 1760 * Arguments: 1761 * zfunc A pointer to a function which accepts a zone 1762 * as an argument. 1763 * 1764 * Returns: 1765 * Nothing 1766 */ 1767static void 1768zone_foreach(void (*zfunc)(uma_zone_t)) 1769{ 1770 uma_keg_t keg; 1771 uma_zone_t zone; 1772 1773 rw_rlock(&uma_rwlock); 1774 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1775 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1776 zfunc(zone); 1777 } 1778 rw_runlock(&uma_rwlock); 1779} 1780 1781/* Public functions */ 1782/* See uma.h */ 1783void 1784uma_startup(void *bootmem, int boot_pages) 1785{ 1786 struct uma_zctor_args args; 1787 uma_slab_t slab; 1788 u_int slabsize; 1789 int i; 1790 1791#ifdef UMA_DEBUG 1792 printf("Creating uma keg headers zone and keg.\n"); 1793#endif 1794 rw_init(&uma_rwlock, "UMA lock"); 1795 1796 /* "manually" create the initial zone */ 1797 memset(&args, 0, sizeof(args)); 1798 args.name = "UMA Kegs"; 1799 args.size = sizeof(struct uma_keg); 1800 args.ctor = keg_ctor; 1801 args.dtor = keg_dtor; 1802 args.uminit = zero_init; 1803 args.fini = NULL; 1804 args.keg = &masterkeg; 1805 args.align = 32 - 1; 1806 args.flags = UMA_ZFLAG_INTERNAL; 1807 /* The initial zone has no Per cpu queues so it's smaller */ 1808 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); 1809 1810#ifdef UMA_DEBUG 1811 printf("Filling boot free list.\n"); 1812#endif 1813 for (i = 0; i < boot_pages; i++) { 1814 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE)); 1815 slab->us_data = (uint8_t *)slab; 1816 slab->us_flags = UMA_SLAB_BOOT; 1817 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); 1818 } 1819 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); 1820 1821#ifdef UMA_DEBUG 1822 printf("Creating uma zone headers zone and keg.\n"); 1823#endif 1824 args.name = "UMA Zones"; 1825 args.size = sizeof(struct uma_zone) + 1826 (sizeof(struct uma_cache) * (mp_maxid + 1)); 1827 args.ctor = zone_ctor; 1828 args.dtor = zone_dtor; 1829 args.uminit = zero_init; 1830 args.fini = NULL; 1831 args.keg = NULL; 1832 args.align = 32 - 1; 1833 args.flags = UMA_ZFLAG_INTERNAL; 1834 /* The initial zone has no Per cpu queues so it's smaller */ 1835 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); 1836 1837#ifdef UMA_DEBUG 1838 printf("Initializing pcpu cache locks.\n"); 1839#endif 1840#ifdef UMA_DEBUG 1841 printf("Creating slab and hash zones.\n"); 1842#endif 1843 1844 /* Now make a zone for slab headers */ 1845 slabzone = uma_zcreate("UMA Slabs", 1846 sizeof(struct uma_slab), 1847 NULL, NULL, NULL, NULL, 1848 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1849 1850 /* 1851 * We also create a zone for the bigger slabs with reference 1852 * counts in them, to accomodate UMA_ZONE_REFCNT zones. 1853 */ 1854 slabsize = sizeof(struct uma_slab_refcnt); 1855 slabsize += uma_max_ipers_ref * sizeof(uint32_t); 1856 slabrefzone = uma_zcreate("UMA RCntSlabs", 1857 slabsize, 1858 NULL, NULL, NULL, NULL, 1859 UMA_ALIGN_PTR, 1860 UMA_ZFLAG_INTERNAL); 1861 1862 hashzone = uma_zcreate("UMA Hash", 1863 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 1864 NULL, NULL, NULL, NULL, 1865 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1866 1867 bucket_init(); 1868 1869 booted = UMA_STARTUP; 1870 1871#ifdef UMA_DEBUG 1872 printf("UMA startup complete.\n"); 1873#endif 1874} 1875 1876/* see uma.h */ 1877void 1878uma_startup2(void) 1879{ 1880 booted = UMA_STARTUP2; 1881 bucket_enable(); 1882 sx_init(&uma_drain_lock, "umadrain"); 1883#ifdef UMA_DEBUG 1884 printf("UMA startup2 complete.\n"); 1885#endif 1886} 1887 1888/* 1889 * Initialize our callout handle 1890 * 1891 */ 1892 1893static void 1894uma_startup3(void) 1895{ 1896#ifdef UMA_DEBUG 1897 printf("Starting callout.\n"); 1898#endif 1899 callout_init(&uma_callout, 1); 1900 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1901#ifdef UMA_DEBUG 1902 printf("UMA startup3 complete.\n"); 1903#endif 1904} 1905 1906static uma_keg_t 1907uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 1908 int align, uint32_t flags) 1909{ 1910 struct uma_kctor_args args; 1911 1912 args.size = size; 1913 args.uminit = uminit; 1914 args.fini = fini; 1915 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 1916 args.flags = flags; 1917 args.zone = zone; 1918 return (zone_alloc_item(kegs, &args, M_WAITOK)); 1919} 1920 1921/* See uma.h */ 1922void 1923uma_set_align(int align) 1924{ 1925 1926 if (align != UMA_ALIGN_CACHE) 1927 uma_align_cache = align; 1928} 1929 1930/* See uma.h */ 1931uma_zone_t 1932uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 1933 uma_init uminit, uma_fini fini, int align, uint32_t flags) 1934 1935{ 1936 struct uma_zctor_args args; 1937 uma_zone_t res; 1938 bool locked; 1939 1940 /* This stuff is essential for the zone ctor */ 1941 memset(&args, 0, sizeof(args)); 1942 args.name = name; 1943 args.size = size; 1944 args.ctor = ctor; 1945 args.dtor = dtor; 1946 args.uminit = uminit; 1947 args.fini = fini; 1948 args.align = align; 1949 args.flags = flags; 1950 args.keg = NULL; 1951 1952 if (booted < UMA_STARTUP2) { 1953 locked = false; 1954 } else { 1955 sx_slock(&uma_drain_lock); 1956 locked = true; 1957 } 1958 res = zone_alloc_item(zones, &args, M_WAITOK); 1959 if (locked) 1960 sx_sunlock(&uma_drain_lock); 1961 return (res); 1962} 1963 1964/* See uma.h */ 1965uma_zone_t 1966uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 1967 uma_init zinit, uma_fini zfini, uma_zone_t master) 1968{ 1969 struct uma_zctor_args args; 1970 uma_keg_t keg; 1971 uma_zone_t res; 1972 bool locked; 1973 1974 keg = zone_first_keg(master); 1975 memset(&args, 0, sizeof(args)); 1976 args.name = name; 1977 args.size = keg->uk_size; 1978 args.ctor = ctor; 1979 args.dtor = dtor; 1980 args.uminit = zinit; 1981 args.fini = zfini; 1982 args.align = keg->uk_align; 1983 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 1984 args.keg = keg; 1985 1986 if (booted < UMA_STARTUP2) { 1987 locked = false; 1988 } else { 1989 sx_slock(&uma_drain_lock); 1990 locked = true; 1991 } 1992 /* XXX Attaches only one keg of potentially many. */ 1993 res = zone_alloc_item(zones, &args, M_WAITOK); 1994 if (locked) 1995 sx_sunlock(&uma_drain_lock); 1996 return (res); 1997} 1998 1999/* See uma.h */ 2000uma_zone_t 2001uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 2002 uma_init zinit, uma_fini zfini, uma_import zimport, 2003 uma_release zrelease, void *arg, int flags) 2004{ 2005 struct uma_zctor_args args; 2006 2007 memset(&args, 0, sizeof(args)); 2008 args.name = name; 2009 args.size = size; 2010 args.ctor = ctor; 2011 args.dtor = dtor; 2012 args.uminit = zinit; 2013 args.fini = zfini; 2014 args.import = zimport; 2015 args.release = zrelease; 2016 args.arg = arg; 2017 args.align = 0; 2018 args.flags = flags; 2019 2020 return (zone_alloc_item(zones, &args, M_WAITOK)); 2021} 2022 2023static void 2024zone_lock_pair(uma_zone_t a, uma_zone_t b) 2025{ 2026 if (a < b) { 2027 ZONE_LOCK(a); 2028 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); 2029 } else { 2030 ZONE_LOCK(b); 2031 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); 2032 } 2033} 2034 2035static void 2036zone_unlock_pair(uma_zone_t a, uma_zone_t b) 2037{ 2038 2039 ZONE_UNLOCK(a); 2040 ZONE_UNLOCK(b); 2041} 2042 2043int 2044uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 2045{ 2046 uma_klink_t klink; 2047 uma_klink_t kl; 2048 int error; 2049 2050 error = 0; 2051 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 2052 2053 zone_lock_pair(zone, master); 2054 /* 2055 * zone must use vtoslab() to resolve objects and must already be 2056 * a secondary. 2057 */ 2058 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 2059 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 2060 error = EINVAL; 2061 goto out; 2062 } 2063 /* 2064 * The new master must also use vtoslab(). 2065 */ 2066 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 2067 error = EINVAL; 2068 goto out; 2069 } 2070 /* 2071 * Both must either be refcnt, or not be refcnt. 2072 */ 2073 if ((zone->uz_flags & UMA_ZONE_REFCNT) != 2074 (master->uz_flags & UMA_ZONE_REFCNT)) { 2075 error = EINVAL; 2076 goto out; 2077 } 2078 /* 2079 * The underlying object must be the same size. rsize 2080 * may be different. 2081 */ 2082 if (master->uz_size != zone->uz_size) { 2083 error = E2BIG; 2084 goto out; 2085 } 2086 /* 2087 * Put it at the end of the list. 2088 */ 2089 klink->kl_keg = zone_first_keg(master); 2090 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 2091 if (LIST_NEXT(kl, kl_link) == NULL) { 2092 LIST_INSERT_AFTER(kl, klink, kl_link); 2093 break; 2094 } 2095 } 2096 klink = NULL; 2097 zone->uz_flags |= UMA_ZFLAG_MULTI; 2098 zone->uz_slab = zone_fetch_slab_multi; 2099 2100out: 2101 zone_unlock_pair(zone, master); 2102 if (klink != NULL) 2103 free(klink, M_TEMP); 2104 2105 return (error); 2106} 2107 2108 2109/* See uma.h */ 2110void 2111uma_zdestroy(uma_zone_t zone) 2112{ 2113 2114 sx_slock(&uma_drain_lock); 2115 zone_free_item(zones, zone, NULL, SKIP_NONE); 2116 sx_sunlock(&uma_drain_lock); 2117} 2118 2119/* See uma.h */ 2120void * 2121uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2122{ 2123 void *item; 2124 uma_cache_t cache; 2125 uma_bucket_t bucket; 2126 int lockfail; 2127 int cpu; 2128 2129 /* This is the fast path allocation */ 2130#ifdef UMA_DEBUG_ALLOC_1 2131 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); 2132#endif 2133 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread, 2134 zone->uz_name, flags); 2135 2136 if (flags & M_WAITOK) { 2137 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2138 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2139 } 2140#ifdef DEBUG_MEMGUARD 2141 if (memguard_cmp_zone(zone)) { 2142 item = memguard_alloc(zone->uz_size, flags); 2143 if (item != NULL) { 2144 /* 2145 * Avoid conflict with the use-after-free 2146 * protecting infrastructure from INVARIANTS. 2147 */ 2148 if (zone->uz_init != NULL && 2149 zone->uz_init != mtrash_init && 2150 zone->uz_init(item, zone->uz_size, flags) != 0) 2151 return (NULL); 2152 if (zone->uz_ctor != NULL && 2153 zone->uz_ctor != mtrash_ctor && 2154 zone->uz_ctor(item, zone->uz_size, udata, 2155 flags) != 0) { 2156 zone->uz_fini(item, zone->uz_size); 2157 return (NULL); 2158 } 2159 return (item); 2160 } 2161 /* This is unfortunate but should not be fatal. */ 2162 } 2163#endif 2164 /* 2165 * If possible, allocate from the per-CPU cache. There are two 2166 * requirements for safe access to the per-CPU cache: (1) the thread 2167 * accessing the cache must not be preempted or yield during access, 2168 * and (2) the thread must not migrate CPUs without switching which 2169 * cache it accesses. We rely on a critical section to prevent 2170 * preemption and migration. We release the critical section in 2171 * order to acquire the zone mutex if we are unable to allocate from 2172 * the current cache; when we re-acquire the critical section, we 2173 * must detect and handle migration if it has occurred. 2174 */ 2175 critical_enter(); 2176 cpu = curcpu; 2177 cache = &zone->uz_cpu[cpu]; 2178 2179zalloc_start: 2180 bucket = cache->uc_allocbucket; 2181 if (bucket != NULL && bucket->ub_cnt > 0) { 2182 bucket->ub_cnt--; 2183 item = bucket->ub_bucket[bucket->ub_cnt]; 2184#ifdef INVARIANTS 2185 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2186#endif 2187 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2188 cache->uc_allocs++; 2189 critical_exit(); 2190 if (zone->uz_ctor != NULL && 2191 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2192 atomic_add_long(&zone->uz_fails, 1); 2193 zone_free_item(zone, item, udata, SKIP_DTOR); 2194 return (NULL); 2195 } 2196#ifdef INVARIANTS 2197 uma_dbg_alloc(zone, NULL, item); 2198#endif 2199 if (flags & M_ZERO) 2200 uma_zero_item(item, zone); 2201 return (item); 2202 } 2203 2204 /* 2205 * We have run out of items in our alloc bucket. 2206 * See if we can switch with our free bucket. 2207 */ 2208 bucket = cache->uc_freebucket; 2209 if (bucket != NULL && bucket->ub_cnt > 0) { 2210#ifdef UMA_DEBUG_ALLOC 2211 printf("uma_zalloc: Swapping empty with alloc.\n"); 2212#endif 2213 cache->uc_freebucket = cache->uc_allocbucket; 2214 cache->uc_allocbucket = bucket; 2215 goto zalloc_start; 2216 } 2217 2218 /* 2219 * Discard any empty allocation bucket while we hold no locks. 2220 */ 2221 bucket = cache->uc_allocbucket; 2222 cache->uc_allocbucket = NULL; 2223 critical_exit(); 2224 if (bucket != NULL) 2225 bucket_free(zone, bucket, udata); 2226 2227 /* Short-circuit for zones without buckets and low memory. */ 2228 if (zone->uz_count == 0 || bucketdisable) 2229 goto zalloc_item; 2230 2231 /* 2232 * Attempt to retrieve the item from the per-CPU cache has failed, so 2233 * we must go back to the zone. This requires the zone lock, so we 2234 * must drop the critical section, then re-acquire it when we go back 2235 * to the cache. Since the critical section is released, we may be 2236 * preempted or migrate. As such, make sure not to maintain any 2237 * thread-local state specific to the cache from prior to releasing 2238 * the critical section. 2239 */ 2240 lockfail = 0; 2241 if (ZONE_TRYLOCK(zone) == 0) { 2242 /* Record contention to size the buckets. */ 2243 ZONE_LOCK(zone); 2244 lockfail = 1; 2245 } 2246 critical_enter(); 2247 cpu = curcpu; 2248 cache = &zone->uz_cpu[cpu]; 2249 2250 /* 2251 * Since we have locked the zone we may as well send back our stats. 2252 */ 2253 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2254 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2255 cache->uc_allocs = 0; 2256 cache->uc_frees = 0; 2257 2258 /* See if we lost the race to fill the cache. */ 2259 if (cache->uc_allocbucket != NULL) { 2260 ZONE_UNLOCK(zone); 2261 goto zalloc_start; 2262 } 2263 2264 /* 2265 * Check the zone's cache of buckets. 2266 */ 2267 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 2268 KASSERT(bucket->ub_cnt != 0, 2269 ("uma_zalloc_arg: Returning an empty bucket.")); 2270 2271 LIST_REMOVE(bucket, ub_link); 2272 cache->uc_allocbucket = bucket; 2273 ZONE_UNLOCK(zone); 2274 goto zalloc_start; 2275 } 2276 /* We are no longer associated with this CPU. */ 2277 critical_exit(); 2278 2279 /* 2280 * We bump the uz count when the cache size is insufficient to 2281 * handle the working set. 2282 */ 2283 if (lockfail && zone->uz_count < BUCKET_MAX) 2284 zone->uz_count++; 2285 ZONE_UNLOCK(zone); 2286 2287 /* 2288 * Now lets just fill a bucket and put it on the free list. If that 2289 * works we'll restart the allocation from the begining and it 2290 * will use the just filled bucket. 2291 */ 2292 bucket = zone_alloc_bucket(zone, udata, flags); 2293 if (bucket != NULL) { 2294 ZONE_LOCK(zone); 2295 critical_enter(); 2296 cpu = curcpu; 2297 cache = &zone->uz_cpu[cpu]; 2298 /* 2299 * See if we lost the race or were migrated. Cache the 2300 * initialized bucket to make this less likely or claim 2301 * the memory directly. 2302 */ 2303 if (cache->uc_allocbucket == NULL) 2304 cache->uc_allocbucket = bucket; 2305 else 2306 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2307 ZONE_UNLOCK(zone); 2308 goto zalloc_start; 2309 } 2310 2311 /* 2312 * We may not be able to get a bucket so return an actual item. 2313 */ 2314#ifdef UMA_DEBUG 2315 printf("uma_zalloc_arg: Bucketzone returned NULL\n"); 2316#endif 2317 2318zalloc_item: 2319 item = zone_alloc_item(zone, udata, flags); 2320 2321 return (item); 2322} 2323 2324static uma_slab_t 2325keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) 2326{ 2327 uma_slab_t slab; 2328 int reserve; 2329 2330 mtx_assert(&keg->uk_lock, MA_OWNED); 2331 slab = NULL; 2332 reserve = 0; 2333 if ((flags & M_USE_RESERVE) == 0) 2334 reserve = keg->uk_reserve; 2335 2336 for (;;) { 2337 /* 2338 * Find a slab with some space. Prefer slabs that are partially 2339 * used over those that are totally full. This helps to reduce 2340 * fragmentation. 2341 */ 2342 if (keg->uk_free > reserve) { 2343 if (!LIST_EMPTY(&keg->uk_part_slab)) { 2344 slab = LIST_FIRST(&keg->uk_part_slab); 2345 } else { 2346 slab = LIST_FIRST(&keg->uk_free_slab); 2347 LIST_REMOVE(slab, us_link); 2348 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, 2349 us_link); 2350 } 2351 MPASS(slab->us_keg == keg); 2352 return (slab); 2353 } 2354 2355 /* 2356 * M_NOVM means don't ask at all! 2357 */ 2358 if (flags & M_NOVM) 2359 break; 2360 2361 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2362 keg->uk_flags |= UMA_ZFLAG_FULL; 2363 /* 2364 * If this is not a multi-zone, set the FULL bit. 2365 * Otherwise slab_multi() takes care of it. 2366 */ 2367 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2368 zone->uz_flags |= UMA_ZFLAG_FULL; 2369 zone_log_warning(zone); 2370 } 2371 if (flags & M_NOWAIT) 2372 break; 2373 zone->uz_sleeps++; 2374 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2375 continue; 2376 } 2377 slab = keg_alloc_slab(keg, zone, flags); 2378 /* 2379 * If we got a slab here it's safe to mark it partially used 2380 * and return. We assume that the caller is going to remove 2381 * at least one item. 2382 */ 2383 if (slab) { 2384 MPASS(slab->us_keg == keg); 2385 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2386 return (slab); 2387 } 2388 /* 2389 * We might not have been able to get a slab but another cpu 2390 * could have while we were unlocked. Check again before we 2391 * fail. 2392 */ 2393 flags |= M_NOVM; 2394 } 2395 return (slab); 2396} 2397 2398static uma_slab_t 2399zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) 2400{ 2401 uma_slab_t slab; 2402 2403 if (keg == NULL) { 2404 keg = zone_first_keg(zone); 2405 KEG_LOCK(keg); 2406 } 2407 2408 for (;;) { 2409 slab = keg_fetch_slab(keg, zone, flags); 2410 if (slab) 2411 return (slab); 2412 if (flags & (M_NOWAIT | M_NOVM)) 2413 break; 2414 } 2415 KEG_UNLOCK(keg); 2416 return (NULL); 2417} 2418 2419/* 2420 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2421 * with the keg locked. On NULL no lock is held. 2422 * 2423 * The last pointer is used to seed the search. It is not required. 2424 */ 2425static uma_slab_t 2426zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) 2427{ 2428 uma_klink_t klink; 2429 uma_slab_t slab; 2430 uma_keg_t keg; 2431 int flags; 2432 int empty; 2433 int full; 2434 2435 /* 2436 * Don't wait on the first pass. This will skip limit tests 2437 * as well. We don't want to block if we can find a provider 2438 * without blocking. 2439 */ 2440 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2441 /* 2442 * Use the last slab allocated as a hint for where to start 2443 * the search. 2444 */ 2445 if (last != NULL) { 2446 slab = keg_fetch_slab(last, zone, flags); 2447 if (slab) 2448 return (slab); 2449 KEG_UNLOCK(last); 2450 } 2451 /* 2452 * Loop until we have a slab incase of transient failures 2453 * while M_WAITOK is specified. I'm not sure this is 100% 2454 * required but we've done it for so long now. 2455 */ 2456 for (;;) { 2457 empty = 0; 2458 full = 0; 2459 /* 2460 * Search the available kegs for slabs. Be careful to hold the 2461 * correct lock while calling into the keg layer. 2462 */ 2463 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2464 keg = klink->kl_keg; 2465 KEG_LOCK(keg); 2466 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2467 slab = keg_fetch_slab(keg, zone, flags); 2468 if (slab) 2469 return (slab); 2470 } 2471 if (keg->uk_flags & UMA_ZFLAG_FULL) 2472 full++; 2473 else 2474 empty++; 2475 KEG_UNLOCK(keg); 2476 } 2477 if (rflags & (M_NOWAIT | M_NOVM)) 2478 break; 2479 flags = rflags; 2480 /* 2481 * All kegs are full. XXX We can't atomically check all kegs 2482 * and sleep so just sleep for a short period and retry. 2483 */ 2484 if (full && !empty) { 2485 ZONE_LOCK(zone); 2486 zone->uz_flags |= UMA_ZFLAG_FULL; 2487 zone->uz_sleeps++; 2488 zone_log_warning(zone); 2489 msleep(zone, zone->uz_lockptr, PVM, 2490 "zonelimit", hz/100); 2491 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2492 ZONE_UNLOCK(zone); 2493 continue; 2494 } 2495 } 2496 return (NULL); 2497} 2498 2499static void * 2500slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 2501{ 2502 void *item; 2503 uint8_t freei; 2504 2505 MPASS(keg == slab->us_keg); 2506 mtx_assert(&keg->uk_lock, MA_OWNED); 2507 2508 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; 2509 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); 2510 item = slab->us_data + (keg->uk_rsize * freei); 2511 slab->us_freecount--; 2512 keg->uk_free--; 2513 2514 /* Move this slab to the full list */ 2515 if (slab->us_freecount == 0) { 2516 LIST_REMOVE(slab, us_link); 2517 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); 2518 } 2519 2520 return (item); 2521} 2522 2523static int 2524zone_import(uma_zone_t zone, void **bucket, int max, int flags) 2525{ 2526 uma_slab_t slab; 2527 uma_keg_t keg; 2528 int i; 2529 2530 slab = NULL; 2531 keg = NULL; 2532 /* Try to keep the buckets totally full */ 2533 for (i = 0; i < max; ) { 2534 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL) 2535 break; 2536 keg = slab->us_keg; 2537 while (slab->us_freecount && i < max) { 2538 bucket[i++] = slab_alloc_item(keg, slab); 2539 if (keg->uk_free <= keg->uk_reserve) 2540 break; 2541 } 2542 /* Don't grab more than one slab at a time. */ 2543 flags &= ~M_WAITOK; 2544 flags |= M_NOWAIT; 2545 } 2546 if (slab != NULL) 2547 KEG_UNLOCK(keg); 2548 2549 return i; 2550} 2551 2552static uma_bucket_t 2553zone_alloc_bucket(uma_zone_t zone, void *udata, int flags) 2554{ 2555 uma_bucket_t bucket; 2556 int max; 2557 2558 /* Don't wait for buckets, preserve caller's NOVM setting. */ 2559 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 2560 if (bucket == NULL) 2561 return (NULL); 2562 2563 max = MIN(bucket->ub_entries, zone->uz_count); 2564 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 2565 max, flags); 2566 2567 /* 2568 * Initialize the memory if necessary. 2569 */ 2570 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 2571 int i; 2572 2573 for (i = 0; i < bucket->ub_cnt; i++) 2574 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2575 flags) != 0) 2576 break; 2577 /* 2578 * If we couldn't initialize the whole bucket, put the 2579 * rest back onto the freelist. 2580 */ 2581 if (i != bucket->ub_cnt) { 2582 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 2583 bucket->ub_cnt - i); 2584#ifdef INVARIANTS 2585 bzero(&bucket->ub_bucket[i], 2586 sizeof(void *) * (bucket->ub_cnt - i)); 2587#endif 2588 bucket->ub_cnt = i; 2589 } 2590 } 2591 2592 if (bucket->ub_cnt == 0) { 2593 bucket_free(zone, bucket, udata); 2594 atomic_add_long(&zone->uz_fails, 1); 2595 return (NULL); 2596 } 2597 2598 return (bucket); 2599} 2600 2601/* 2602 * Allocates a single item from a zone. 2603 * 2604 * Arguments 2605 * zone The zone to alloc for. 2606 * udata The data to be passed to the constructor. 2607 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2608 * 2609 * Returns 2610 * NULL if there is no memory and M_NOWAIT is set 2611 * An item if successful 2612 */ 2613 2614static void * 2615zone_alloc_item(uma_zone_t zone, void *udata, int flags) 2616{ 2617 void *item; 2618 2619 item = NULL; 2620 2621#ifdef UMA_DEBUG_ALLOC 2622 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); 2623#endif 2624 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1) 2625 goto fail; 2626 atomic_add_long(&zone->uz_allocs, 1); 2627 2628 /* 2629 * We have to call both the zone's init (not the keg's init) 2630 * and the zone's ctor. This is because the item is going from 2631 * a keg slab directly to the user, and the user is expecting it 2632 * to be both zone-init'd as well as zone-ctor'd. 2633 */ 2634 if (zone->uz_init != NULL) { 2635 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 2636 zone_free_item(zone, item, udata, SKIP_FINI); 2637 goto fail; 2638 } 2639 } 2640 if (zone->uz_ctor != NULL) { 2641 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2642 zone_free_item(zone, item, udata, SKIP_DTOR); 2643 goto fail; 2644 } 2645 } 2646#ifdef INVARIANTS 2647 uma_dbg_alloc(zone, NULL, item); 2648#endif 2649 if (flags & M_ZERO) 2650 uma_zero_item(item, zone); 2651 2652 return (item); 2653 2654fail: 2655 atomic_add_long(&zone->uz_fails, 1); 2656 return (NULL); 2657} 2658 2659/* See uma.h */ 2660void 2661uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 2662{ 2663 uma_cache_t cache; 2664 uma_bucket_t bucket; 2665 int lockfail; 2666 int cpu; 2667 2668#ifdef UMA_DEBUG_ALLOC_1 2669 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); 2670#endif 2671 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 2672 zone->uz_name); 2673 2674 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 2675 if (item == NULL) 2676 return; 2677#ifdef DEBUG_MEMGUARD 2678 if (is_memguard_addr(item)) { 2679 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor) 2680 zone->uz_dtor(item, zone->uz_size, udata); 2681 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini) 2682 zone->uz_fini(item, zone->uz_size); 2683 memguard_free(item); 2684 return; 2685 } 2686#endif 2687#ifdef INVARIANTS 2688 if (zone->uz_flags & UMA_ZONE_MALLOC) 2689 uma_dbg_free(zone, udata, item); 2690 else 2691 uma_dbg_free(zone, NULL, item); 2692#endif 2693 if (zone->uz_dtor != NULL) 2694 zone->uz_dtor(item, zone->uz_size, udata); 2695 2696 /* 2697 * The race here is acceptable. If we miss it we'll just have to wait 2698 * a little longer for the limits to be reset. 2699 */ 2700 if (zone->uz_flags & UMA_ZFLAG_FULL) 2701 goto zfree_item; 2702 2703 /* 2704 * If possible, free to the per-CPU cache. There are two 2705 * requirements for safe access to the per-CPU cache: (1) the thread 2706 * accessing the cache must not be preempted or yield during access, 2707 * and (2) the thread must not migrate CPUs without switching which 2708 * cache it accesses. We rely on a critical section to prevent 2709 * preemption and migration. We release the critical section in 2710 * order to acquire the zone mutex if we are unable to free to the 2711 * current cache; when we re-acquire the critical section, we must 2712 * detect and handle migration if it has occurred. 2713 */ 2714zfree_restart: 2715 critical_enter(); 2716 cpu = curcpu; 2717 cache = &zone->uz_cpu[cpu]; 2718 2719zfree_start: 2720 /* 2721 * Try to free into the allocbucket first to give LIFO ordering 2722 * for cache-hot datastructures. Spill over into the freebucket 2723 * if necessary. Alloc will swap them if one runs dry. 2724 */ 2725 bucket = cache->uc_allocbucket; 2726 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 2727 bucket = cache->uc_freebucket; 2728 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2729 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 2730 ("uma_zfree: Freeing to non free bucket index.")); 2731 bucket->ub_bucket[bucket->ub_cnt] = item; 2732 bucket->ub_cnt++; 2733 cache->uc_frees++; 2734 critical_exit(); 2735 return; 2736 } 2737 2738 /* 2739 * We must go back the zone, which requires acquiring the zone lock, 2740 * which in turn means we must release and re-acquire the critical 2741 * section. Since the critical section is released, we may be 2742 * preempted or migrate. As such, make sure not to maintain any 2743 * thread-local state specific to the cache from prior to releasing 2744 * the critical section. 2745 */ 2746 critical_exit(); 2747 if (zone->uz_count == 0 || bucketdisable) 2748 goto zfree_item; 2749 2750 lockfail = 0; 2751 if (ZONE_TRYLOCK(zone) == 0) { 2752 /* Record contention to size the buckets. */ 2753 ZONE_LOCK(zone); 2754 lockfail = 1; 2755 } 2756 critical_enter(); 2757 cpu = curcpu; 2758 cache = &zone->uz_cpu[cpu]; 2759 2760 /* 2761 * Since we have locked the zone we may as well send back our stats. 2762 */ 2763 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2764 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2765 cache->uc_allocs = 0; 2766 cache->uc_frees = 0; 2767 2768 bucket = cache->uc_freebucket; 2769 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2770 ZONE_UNLOCK(zone); 2771 goto zfree_start; 2772 } 2773 cache->uc_freebucket = NULL; 2774 /* We are no longer associated with this CPU. */ 2775 critical_exit(); 2776 2777 /* Can we throw this on the zone full list? */ 2778 if (bucket != NULL) { 2779#ifdef UMA_DEBUG_ALLOC 2780 printf("uma_zfree: Putting old bucket on the free list.\n"); 2781#endif 2782 /* ub_cnt is pointing to the last free item */ 2783 KASSERT(bucket->ub_cnt != 0, 2784 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 2785 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2786 } 2787 2788 /* 2789 * We bump the uz count when the cache size is insufficient to 2790 * handle the working set. 2791 */ 2792 if (lockfail && zone->uz_count < BUCKET_MAX) 2793 zone->uz_count++; 2794 ZONE_UNLOCK(zone); 2795 2796#ifdef UMA_DEBUG_ALLOC 2797 printf("uma_zfree: Allocating new free bucket.\n"); 2798#endif 2799 bucket = bucket_alloc(zone, udata, M_NOWAIT); 2800 if (bucket) { 2801 critical_enter(); 2802 cpu = curcpu; 2803 cache = &zone->uz_cpu[cpu]; 2804 if (cache->uc_freebucket == NULL) { 2805 cache->uc_freebucket = bucket; 2806 goto zfree_start; 2807 } 2808 /* 2809 * We lost the race, start over. We have to drop our 2810 * critical section to free the bucket. 2811 */ 2812 critical_exit(); 2813 bucket_free(zone, bucket, udata); 2814 goto zfree_restart; 2815 } 2816 2817 /* 2818 * If nothing else caught this, we'll just do an internal free. 2819 */ 2820zfree_item: 2821 zone_free_item(zone, item, udata, SKIP_DTOR); 2822 2823 return; 2824} 2825 2826static void 2827slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) 2828{ 2829 uint8_t freei; 2830 2831 mtx_assert(&keg->uk_lock, MA_OWNED); 2832 MPASS(keg == slab->us_keg); 2833 2834 /* Do we need to remove from any lists? */ 2835 if (slab->us_freecount+1 == keg->uk_ipers) { 2836 LIST_REMOVE(slab, us_link); 2837 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 2838 } else if (slab->us_freecount == 0) { 2839 LIST_REMOVE(slab, us_link); 2840 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2841 } 2842 2843 /* Slab management. */ 2844 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 2845 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); 2846 slab->us_freecount++; 2847 2848 /* Keg statistics. */ 2849 keg->uk_free++; 2850} 2851 2852static void 2853zone_release(uma_zone_t zone, void **bucket, int cnt) 2854{ 2855 void *item; 2856 uma_slab_t slab; 2857 uma_keg_t keg; 2858 uint8_t *mem; 2859 int clearfull; 2860 int i; 2861 2862 clearfull = 0; 2863 keg = zone_first_keg(zone); 2864 KEG_LOCK(keg); 2865 for (i = 0; i < cnt; i++) { 2866 item = bucket[i]; 2867 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 2868 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 2869 if (zone->uz_flags & UMA_ZONE_HASH) { 2870 slab = hash_sfind(&keg->uk_hash, mem); 2871 } else { 2872 mem += keg->uk_pgoff; 2873 slab = (uma_slab_t)mem; 2874 } 2875 } else { 2876 slab = vtoslab((vm_offset_t)item); 2877 if (slab->us_keg != keg) { 2878 KEG_UNLOCK(keg); 2879 keg = slab->us_keg; 2880 KEG_LOCK(keg); 2881 } 2882 } 2883 slab_free_item(keg, slab, item); 2884 if (keg->uk_flags & UMA_ZFLAG_FULL) { 2885 if (keg->uk_pages < keg->uk_maxpages) { 2886 keg->uk_flags &= ~UMA_ZFLAG_FULL; 2887 clearfull = 1; 2888 } 2889 2890 /* 2891 * We can handle one more allocation. Since we're 2892 * clearing ZFLAG_FULL, wake up all procs blocked 2893 * on pages. This should be uncommon, so keeping this 2894 * simple for now (rather than adding count of blocked 2895 * threads etc). 2896 */ 2897 wakeup(keg); 2898 } 2899 } 2900 KEG_UNLOCK(keg); 2901 if (clearfull) { 2902 ZONE_LOCK(zone); 2903 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2904 wakeup(zone); 2905 ZONE_UNLOCK(zone); 2906 } 2907 2908} 2909 2910/* 2911 * Frees a single item to any zone. 2912 * 2913 * Arguments: 2914 * zone The zone to free to 2915 * item The item we're freeing 2916 * udata User supplied data for the dtor 2917 * skip Skip dtors and finis 2918 */ 2919static void 2920zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 2921{ 2922 2923#ifdef INVARIANTS 2924 if (skip == SKIP_NONE) { 2925 if (zone->uz_flags & UMA_ZONE_MALLOC) 2926 uma_dbg_free(zone, udata, item); 2927 else 2928 uma_dbg_free(zone, NULL, item); 2929 } 2930#endif 2931 if (skip < SKIP_DTOR && zone->uz_dtor) 2932 zone->uz_dtor(item, zone->uz_size, udata); 2933 2934 if (skip < SKIP_FINI && zone->uz_fini) 2935 zone->uz_fini(item, zone->uz_size); 2936 2937 atomic_add_long(&zone->uz_frees, 1); 2938 zone->uz_release(zone->uz_arg, &item, 1); 2939} 2940 2941/* See uma.h */ 2942int 2943uma_zone_set_max(uma_zone_t zone, int nitems) 2944{ 2945 uma_keg_t keg; 2946 2947 keg = zone_first_keg(zone); 2948 if (keg == NULL) 2949 return (0); 2950 KEG_LOCK(keg); 2951 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 2952 if (keg->uk_maxpages * keg->uk_ipers < nitems) 2953 keg->uk_maxpages += keg->uk_ppera; 2954 nitems = keg->uk_maxpages * keg->uk_ipers; 2955 KEG_UNLOCK(keg); 2956 2957 return (nitems); 2958} 2959 2960/* See uma.h */ 2961int 2962uma_zone_get_max(uma_zone_t zone) 2963{ 2964 int nitems; 2965 uma_keg_t keg; 2966 2967 keg = zone_first_keg(zone); 2968 if (keg == NULL) 2969 return (0); 2970 KEG_LOCK(keg); 2971 nitems = keg->uk_maxpages * keg->uk_ipers; 2972 KEG_UNLOCK(keg); 2973 2974 return (nitems); 2975} 2976 2977/* See uma.h */ 2978void 2979uma_zone_set_warning(uma_zone_t zone, const char *warning) 2980{ 2981 2982 ZONE_LOCK(zone); 2983 zone->uz_warning = warning; 2984 ZONE_UNLOCK(zone); 2985} 2986 2987/* See uma.h */ 2988int 2989uma_zone_get_cur(uma_zone_t zone) 2990{ 2991 int64_t nitems; 2992 u_int i; 2993 2994 ZONE_LOCK(zone); 2995 nitems = zone->uz_allocs - zone->uz_frees; 2996 CPU_FOREACH(i) { 2997 /* 2998 * See the comment in sysctl_vm_zone_stats() regarding the 2999 * safety of accessing the per-cpu caches. With the zone lock 3000 * held, it is safe, but can potentially result in stale data. 3001 */ 3002 nitems += zone->uz_cpu[i].uc_allocs - 3003 zone->uz_cpu[i].uc_frees; 3004 } 3005 ZONE_UNLOCK(zone); 3006 3007 return (nitems < 0 ? 0 : nitems); 3008} 3009 3010/* See uma.h */ 3011void 3012uma_zone_set_init(uma_zone_t zone, uma_init uminit) 3013{ 3014 uma_keg_t keg; 3015 3016 keg = zone_first_keg(zone); 3017 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 3018 KEG_LOCK(keg); 3019 KASSERT(keg->uk_pages == 0, 3020 ("uma_zone_set_init on non-empty keg")); 3021 keg->uk_init = uminit; 3022 KEG_UNLOCK(keg); 3023} 3024 3025/* See uma.h */ 3026void 3027uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 3028{ 3029 uma_keg_t keg; 3030 3031 keg = zone_first_keg(zone); 3032 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type")); 3033 KEG_LOCK(keg); 3034 KASSERT(keg->uk_pages == 0, 3035 ("uma_zone_set_fini on non-empty keg")); 3036 keg->uk_fini = fini; 3037 KEG_UNLOCK(keg); 3038} 3039 3040/* See uma.h */ 3041void 3042uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 3043{ 3044 3045 ZONE_LOCK(zone); 3046 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3047 ("uma_zone_set_zinit on non-empty keg")); 3048 zone->uz_init = zinit; 3049 ZONE_UNLOCK(zone); 3050} 3051 3052/* See uma.h */ 3053void 3054uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 3055{ 3056 3057 ZONE_LOCK(zone); 3058 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3059 ("uma_zone_set_zfini on non-empty keg")); 3060 zone->uz_fini = zfini; 3061 ZONE_UNLOCK(zone); 3062} 3063 3064/* See uma.h */ 3065/* XXX uk_freef is not actually used with the zone locked */ 3066void 3067uma_zone_set_freef(uma_zone_t zone, uma_free freef) 3068{ 3069 uma_keg_t keg; 3070 3071 keg = zone_first_keg(zone); 3072 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type")); 3073 KEG_LOCK(keg); 3074 keg->uk_freef = freef; 3075 KEG_UNLOCK(keg); 3076} 3077 3078/* See uma.h */ 3079/* XXX uk_allocf is not actually used with the zone locked */ 3080void 3081uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3082{ 3083 uma_keg_t keg; 3084 3085 keg = zone_first_keg(zone); 3086 KEG_LOCK(keg); 3087 keg->uk_allocf = allocf; 3088 KEG_UNLOCK(keg); 3089} 3090 3091/* See uma.h */ 3092void 3093uma_zone_reserve(uma_zone_t zone, int items) 3094{ 3095 uma_keg_t keg; 3096 3097 keg = zone_first_keg(zone); 3098 if (keg == NULL) 3099 return; 3100 KEG_LOCK(keg); 3101 keg->uk_reserve = items; 3102 KEG_UNLOCK(keg); 3103 3104 return; 3105} 3106 3107/* See uma.h */ 3108int 3109uma_zone_reserve_kva(uma_zone_t zone, int count) 3110{ 3111 uma_keg_t keg; 3112 vm_offset_t kva; 3113 int pages; 3114 3115 keg = zone_first_keg(zone); 3116 if (keg == NULL) 3117 return (0); 3118 pages = count / keg->uk_ipers; 3119 3120 if (pages * keg->uk_ipers < count) 3121 pages++; 3122 3123#ifdef UMA_MD_SMALL_ALLOC 3124 if (keg->uk_ppera > 1) { 3125#else 3126 if (1) { 3127#endif 3128 kva = kva_alloc(pages * UMA_SLAB_SIZE); 3129 if (kva == 0) 3130 return (0); 3131 } else 3132 kva = 0; 3133 KEG_LOCK(keg); 3134 keg->uk_kva = kva; 3135 keg->uk_offset = 0; 3136 keg->uk_maxpages = pages; 3137#ifdef UMA_MD_SMALL_ALLOC 3138 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3139#else 3140 keg->uk_allocf = noobj_alloc; 3141#endif 3142 keg->uk_flags |= UMA_ZONE_NOFREE; 3143 KEG_UNLOCK(keg); 3144 3145 return (1); 3146} 3147 3148/* See uma.h */ 3149void 3150uma_prealloc(uma_zone_t zone, int items) 3151{ 3152 int slabs; 3153 uma_slab_t slab; 3154 uma_keg_t keg; 3155 3156 keg = zone_first_keg(zone); 3157 if (keg == NULL) 3158 return; 3159 KEG_LOCK(keg); 3160 slabs = items / keg->uk_ipers; 3161 if (slabs * keg->uk_ipers < items) 3162 slabs++; 3163 while (slabs > 0) { 3164 slab = keg_alloc_slab(keg, zone, M_WAITOK); 3165 if (slab == NULL) 3166 break; 3167 MPASS(slab->us_keg == keg); 3168 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 3169 slabs--; 3170 } 3171 KEG_UNLOCK(keg); 3172} 3173 3174/* See uma.h */ 3175uint32_t * 3176uma_find_refcnt(uma_zone_t zone, void *item) 3177{ 3178 uma_slabrefcnt_t slabref; 3179 uma_slab_t slab; 3180 uma_keg_t keg; 3181 uint32_t *refcnt; 3182 int idx; 3183 3184 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK)); 3185 slabref = (uma_slabrefcnt_t)slab; 3186 keg = slab->us_keg; 3187 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT, 3188 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT")); 3189 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3190 refcnt = &slabref->us_refcnt[idx]; 3191 return refcnt; 3192} 3193 3194/* See uma.h */ 3195static void 3196uma_reclaim_locked(bool kmem_danger) 3197{ 3198 3199#ifdef UMA_DEBUG 3200 printf("UMA: vm asked us to release pages!\n"); 3201#endif 3202 sx_assert(&uma_drain_lock, SA_XLOCKED); 3203 bucket_enable(); 3204 zone_foreach(zone_drain); 3205 if (vm_page_count_min() || kmem_danger) { 3206 cache_drain_safe(NULL); 3207 zone_foreach(zone_drain); 3208 } 3209 /* 3210 * Some slabs may have been freed but this zone will be visited early 3211 * we visit again so that we can free pages that are empty once other 3212 * zones are drained. We have to do the same for buckets. 3213 */ 3214 zone_drain(slabzone); 3215 zone_drain(slabrefzone); 3216 bucket_zone_drain(); 3217} 3218 3219void 3220uma_reclaim(void) 3221{ 3222 3223 sx_xlock(&uma_drain_lock); 3224 uma_reclaim_locked(false); 3225 sx_xunlock(&uma_drain_lock); 3226} 3227 3228static int uma_reclaim_needed; 3229 3230void 3231uma_reclaim_wakeup(void) 3232{ 3233 3234 uma_reclaim_needed = 1; 3235 wakeup(&uma_reclaim_needed); 3236} 3237 3238void 3239uma_reclaim_worker(void *arg __unused) 3240{ 3241 3242 sx_xlock(&uma_drain_lock); 3243 for (;;) { 3244 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM, 3245 "umarcl", 0); 3246 if (uma_reclaim_needed) { 3247 uma_reclaim_needed = 0; 3248 sx_xunlock(&uma_drain_lock); 3249 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); 3250 sx_xlock(&uma_drain_lock); 3251 uma_reclaim_locked(true); 3252 } 3253 } 3254} 3255 3256/* See uma.h */ 3257int 3258uma_zone_exhausted(uma_zone_t zone) 3259{ 3260 int full; 3261 3262 ZONE_LOCK(zone); 3263 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3264 ZONE_UNLOCK(zone); 3265 return (full); 3266} 3267 3268int 3269uma_zone_exhausted_nolock(uma_zone_t zone) 3270{ 3271 return (zone->uz_flags & UMA_ZFLAG_FULL); 3272} 3273 3274void * 3275uma_large_malloc(vm_size_t size, int wait) 3276{ 3277 void *mem; 3278 uma_slab_t slab; 3279 uint8_t flags; 3280 3281 slab = zone_alloc_item(slabzone, NULL, wait); 3282 if (slab == NULL) 3283 return (NULL); 3284 mem = page_alloc(NULL, size, &flags, wait); 3285 if (mem) { 3286 vsetslab((vm_offset_t)mem, slab); 3287 slab->us_data = mem; 3288 slab->us_flags = flags | UMA_SLAB_MALLOC; 3289 slab->us_size = size; 3290 } else { 3291 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3292 } 3293 3294 return (mem); 3295} 3296 3297void 3298uma_large_free(uma_slab_t slab) 3299{ 3300 3301 page_free(slab->us_data, slab->us_size, slab->us_flags); 3302 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3303} 3304 3305static void 3306uma_zero_item(void *item, uma_zone_t zone) 3307{ 3308 3309 if (zone->uz_flags & UMA_ZONE_PCPU) { 3310 for (int i = 0; i < mp_ncpus; i++) 3311 bzero(zpcpu_get_cpu(item, i), zone->uz_size); 3312 } else 3313 bzero(item, zone->uz_size); 3314} 3315 3316void 3317uma_print_stats(void) 3318{ 3319 zone_foreach(uma_print_zone); 3320} 3321 3322static void 3323slab_print(uma_slab_t slab) 3324{ 3325 printf("slab: keg %p, data %p, freecount %d\n", 3326 slab->us_keg, slab->us_data, slab->us_freecount); 3327} 3328 3329static void 3330cache_print(uma_cache_t cache) 3331{ 3332 printf("alloc: %p(%d), free: %p(%d)\n", 3333 cache->uc_allocbucket, 3334 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3335 cache->uc_freebucket, 3336 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3337} 3338 3339static void 3340uma_print_keg(uma_keg_t keg) 3341{ 3342 uma_slab_t slab; 3343 3344 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3345 "out %d free %d limit %d\n", 3346 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3347 keg->uk_ipers, keg->uk_ppera, 3348 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free, 3349 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3350 printf("Part slabs:\n"); 3351 LIST_FOREACH(slab, &keg->uk_part_slab, us_link) 3352 slab_print(slab); 3353 printf("Free slabs:\n"); 3354 LIST_FOREACH(slab, &keg->uk_free_slab, us_link) 3355 slab_print(slab); 3356 printf("Full slabs:\n"); 3357 LIST_FOREACH(slab, &keg->uk_full_slab, us_link) 3358 slab_print(slab); 3359} 3360 3361void 3362uma_print_zone(uma_zone_t zone) 3363{ 3364 uma_cache_t cache; 3365 uma_klink_t kl; 3366 int i; 3367 3368 printf("zone: %s(%p) size %d flags %#x\n", 3369 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3370 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3371 uma_print_keg(kl->kl_keg); 3372 CPU_FOREACH(i) { 3373 cache = &zone->uz_cpu[i]; 3374 printf("CPU %d Cache:\n", i); 3375 cache_print(cache); 3376 } 3377} 3378 3379#ifdef DDB 3380/* 3381 * Generate statistics across both the zone and its per-cpu cache's. Return 3382 * desired statistics if the pointer is non-NULL for that statistic. 3383 * 3384 * Note: does not update the zone statistics, as it can't safely clear the 3385 * per-CPU cache statistic. 3386 * 3387 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3388 * safe from off-CPU; we should modify the caches to track this information 3389 * directly so that we don't have to. 3390 */ 3391static void 3392uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3393 uint64_t *freesp, uint64_t *sleepsp) 3394{ 3395 uma_cache_t cache; 3396 uint64_t allocs, frees, sleeps; 3397 int cachefree, cpu; 3398 3399 allocs = frees = sleeps = 0; 3400 cachefree = 0; 3401 CPU_FOREACH(cpu) { 3402 cache = &z->uz_cpu[cpu]; 3403 if (cache->uc_allocbucket != NULL) 3404 cachefree += cache->uc_allocbucket->ub_cnt; 3405 if (cache->uc_freebucket != NULL) 3406 cachefree += cache->uc_freebucket->ub_cnt; 3407 allocs += cache->uc_allocs; 3408 frees += cache->uc_frees; 3409 } 3410 allocs += z->uz_allocs; 3411 frees += z->uz_frees; 3412 sleeps += z->uz_sleeps; 3413 if (cachefreep != NULL) 3414 *cachefreep = cachefree; 3415 if (allocsp != NULL) 3416 *allocsp = allocs; 3417 if (freesp != NULL) 3418 *freesp = frees; 3419 if (sleepsp != NULL) 3420 *sleepsp = sleeps; 3421} 3422#endif /* DDB */ 3423 3424static int 3425sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3426{ 3427 uma_keg_t kz; 3428 uma_zone_t z; 3429 int count; 3430 3431 count = 0; 3432 rw_rlock(&uma_rwlock); 3433 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3434 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3435 count++; 3436 } 3437 rw_runlock(&uma_rwlock); 3438 return (sysctl_handle_int(oidp, &count, 0, req)); 3439} 3440 3441static int 3442sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3443{ 3444 struct uma_stream_header ush; 3445 struct uma_type_header uth; 3446 struct uma_percpu_stat ups; 3447 uma_bucket_t bucket; 3448 struct sbuf sbuf; 3449 uma_cache_t cache; 3450 uma_klink_t kl; 3451 uma_keg_t kz; 3452 uma_zone_t z; 3453 uma_keg_t k; 3454 int count, error, i; 3455 3456 error = sysctl_wire_old_buffer(req, 0); 3457 if (error != 0) 3458 return (error); 3459 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3460 3461 count = 0; 3462 rw_rlock(&uma_rwlock); 3463 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3464 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3465 count++; 3466 } 3467 3468 /* 3469 * Insert stream header. 3470 */ 3471 bzero(&ush, sizeof(ush)); 3472 ush.ush_version = UMA_STREAM_VERSION; 3473 ush.ush_maxcpus = (mp_maxid + 1); 3474 ush.ush_count = count; 3475 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3476 3477 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3478 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3479 bzero(&uth, sizeof(uth)); 3480 ZONE_LOCK(z); 3481 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3482 uth.uth_align = kz->uk_align; 3483 uth.uth_size = kz->uk_size; 3484 uth.uth_rsize = kz->uk_rsize; 3485 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3486 k = kl->kl_keg; 3487 uth.uth_maxpages += k->uk_maxpages; 3488 uth.uth_pages += k->uk_pages; 3489 uth.uth_keg_free += k->uk_free; 3490 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3491 * k->uk_ipers; 3492 } 3493 3494 /* 3495 * A zone is secondary is it is not the first entry 3496 * on the keg's zone list. 3497 */ 3498 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3499 (LIST_FIRST(&kz->uk_zones) != z)) 3500 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3501 3502 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3503 uth.uth_zone_free += bucket->ub_cnt; 3504 uth.uth_allocs = z->uz_allocs; 3505 uth.uth_frees = z->uz_frees; 3506 uth.uth_fails = z->uz_fails; 3507 uth.uth_sleeps = z->uz_sleeps; 3508 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 3509 /* 3510 * While it is not normally safe to access the cache 3511 * bucket pointers while not on the CPU that owns the 3512 * cache, we only allow the pointers to be exchanged 3513 * without the zone lock held, not invalidated, so 3514 * accept the possible race associated with bucket 3515 * exchange during monitoring. 3516 */ 3517 for (i = 0; i < (mp_maxid + 1); i++) { 3518 bzero(&ups, sizeof(ups)); 3519 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) 3520 goto skip; 3521 if (CPU_ABSENT(i)) 3522 goto skip; 3523 cache = &z->uz_cpu[i]; 3524 if (cache->uc_allocbucket != NULL) 3525 ups.ups_cache_free += 3526 cache->uc_allocbucket->ub_cnt; 3527 if (cache->uc_freebucket != NULL) 3528 ups.ups_cache_free += 3529 cache->uc_freebucket->ub_cnt; 3530 ups.ups_allocs = cache->uc_allocs; 3531 ups.ups_frees = cache->uc_frees; 3532skip: 3533 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups)); 3534 } 3535 ZONE_UNLOCK(z); 3536 } 3537 } 3538 rw_runlock(&uma_rwlock); 3539 error = sbuf_finish(&sbuf); 3540 sbuf_delete(&sbuf); 3541 return (error); 3542} 3543 3544int 3545sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 3546{ 3547 uma_zone_t zone = *(uma_zone_t *)arg1; 3548 int error, max, old; 3549 3550 old = max = uma_zone_get_max(zone); 3551 error = sysctl_handle_int(oidp, &max, 0, req); 3552 if (error || !req->newptr) 3553 return (error); 3554 3555 if (max < old) 3556 return (EINVAL); 3557 3558 uma_zone_set_max(zone, max); 3559 3560 return (0); 3561} 3562 3563int 3564sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 3565{ 3566 uma_zone_t zone = *(uma_zone_t *)arg1; 3567 int cur; 3568 3569 cur = uma_zone_get_cur(zone); 3570 return (sysctl_handle_int(oidp, &cur, 0, req)); 3571} 3572 3573#ifdef DDB 3574DB_SHOW_COMMAND(uma, db_show_uma) 3575{ 3576 uint64_t allocs, frees, sleeps; 3577 uma_bucket_t bucket; 3578 uma_keg_t kz; 3579 uma_zone_t z; 3580 int cachefree; 3581 3582 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used", 3583 "Free", "Requests", "Sleeps", "Bucket"); 3584 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3585 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3586 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 3587 allocs = z->uz_allocs; 3588 frees = z->uz_frees; 3589 sleeps = z->uz_sleeps; 3590 cachefree = 0; 3591 } else 3592 uma_zone_sumstat(z, &cachefree, &allocs, 3593 &frees, &sleeps); 3594 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 3595 (LIST_FIRST(&kz->uk_zones) != z))) 3596 cachefree += kz->uk_free; 3597 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3598 cachefree += bucket->ub_cnt; 3599 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n", 3600 z->uz_name, (uintmax_t)kz->uk_size, 3601 (intmax_t)(allocs - frees), cachefree, 3602 (uintmax_t)allocs, sleeps, z->uz_count); 3603 if (db_pager_quit) 3604 return; 3605 } 3606 } 3607} 3608 3609DB_SHOW_COMMAND(umacache, db_show_umacache) 3610{ 3611 uint64_t allocs, frees; 3612 uma_bucket_t bucket; 3613 uma_zone_t z; 3614 int cachefree; 3615 3616 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 3617 "Requests", "Bucket"); 3618 LIST_FOREACH(z, &uma_cachezones, uz_link) { 3619 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL); 3620 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3621 cachefree += bucket->ub_cnt; 3622 db_printf("%18s %8ju %8jd %8d %12ju %8u\n", 3623 z->uz_name, (uintmax_t)z->uz_size, 3624 (intmax_t)(allocs - frees), cachefree, 3625 (uintmax_t)allocs, z->uz_count); 3626 if (db_pager_quit) 3627 return; 3628 } 3629} 3630#endif 3631