uma_core.c revision 304094
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 304094 2016-08-14 19:02:57Z markj $"); 52 53/* I should really use ktr.. */ 54/* 55#define UMA_DEBUG 1 56#define UMA_DEBUG_ALLOC 1 57#define UMA_DEBUG_ALLOC_1 1 58*/ 59 60#include "opt_ddb.h" 61#include "opt_param.h" 62#include "opt_vm.h" 63 64#include <sys/param.h> 65#include <sys/systm.h> 66#include <sys/bitset.h> 67#include <sys/kernel.h> 68#include <sys/types.h> 69#include <sys/queue.h> 70#include <sys/malloc.h> 71#include <sys/ktr.h> 72#include <sys/lock.h> 73#include <sys/sysctl.h> 74#include <sys/mutex.h> 75#include <sys/proc.h> 76#include <sys/rwlock.h> 77#include <sys/sbuf.h> 78#include <sys/sched.h> 79#include <sys/smp.h> 80#include <sys/vmmeter.h> 81 82#include <vm/vm.h> 83#include <vm/vm_object.h> 84#include <vm/vm_page.h> 85#include <vm/vm_pageout.h> 86#include <vm/vm_param.h> 87#include <vm/vm_map.h> 88#include <vm/vm_kern.h> 89#include <vm/vm_extern.h> 90#include <vm/uma.h> 91#include <vm/uma_int.h> 92#include <vm/uma_dbg.h> 93 94#include <ddb/ddb.h> 95 96#ifdef DEBUG_MEMGUARD 97#include <vm/memguard.h> 98#endif 99 100/* 101 * This is the zone and keg from which all zones are spawned. The idea is that 102 * even the zone & keg heads are allocated from the allocator, so we use the 103 * bss section to bootstrap us. 104 */ 105static struct uma_keg masterkeg; 106static struct uma_zone masterzone_k; 107static struct uma_zone masterzone_z; 108static uma_zone_t kegs = &masterzone_k; 109static uma_zone_t zones = &masterzone_z; 110 111/* This is the zone from which all of uma_slab_t's are allocated. */ 112static uma_zone_t slabzone; 113static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */ 114 115/* 116 * The initial hash tables come out of this zone so they can be allocated 117 * prior to malloc coming up. 118 */ 119static uma_zone_t hashzone; 120 121/* The boot-time adjusted value for cache line alignment. */ 122int uma_align_cache = 64 - 1; 123 124static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 125 126/* 127 * Are we allowed to allocate buckets? 128 */ 129static int bucketdisable = 1; 130 131/* Linked list of all kegs in the system */ 132static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 133 134/* Linked list of all cache-only zones in the system */ 135static LIST_HEAD(,uma_zone) uma_cachezones = 136 LIST_HEAD_INITIALIZER(uma_cachezones); 137 138/* This RW lock protects the keg list */ 139static struct rwlock_padalign uma_rwlock; 140 141/* Linked list of boot time pages */ 142static LIST_HEAD(,uma_slab) uma_boot_pages = 143 LIST_HEAD_INITIALIZER(uma_boot_pages); 144 145/* This mutex protects the boot time pages list */ 146static struct mtx_padalign uma_boot_pages_mtx; 147 148static struct sx uma_drain_lock; 149 150/* Is the VM done starting up? */ 151static int booted = 0; 152#define UMA_STARTUP 1 153#define UMA_STARTUP2 2 154 155/* 156 * Only mbuf clusters use ref zones. Just provide enough references 157 * to support the one user. New code should not use the ref facility. 158 */ 159static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES; 160 161/* 162 * This is the handle used to schedule events that need to happen 163 * outside of the allocation fast path. 164 */ 165static struct callout uma_callout; 166#define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 167 168/* 169 * This structure is passed as the zone ctor arg so that I don't have to create 170 * a special allocation function just for zones. 171 */ 172struct uma_zctor_args { 173 const char *name; 174 size_t size; 175 uma_ctor ctor; 176 uma_dtor dtor; 177 uma_init uminit; 178 uma_fini fini; 179 uma_import import; 180 uma_release release; 181 void *arg; 182 uma_keg_t keg; 183 int align; 184 uint32_t flags; 185}; 186 187struct uma_kctor_args { 188 uma_zone_t zone; 189 size_t size; 190 uma_init uminit; 191 uma_fini fini; 192 int align; 193 uint32_t flags; 194}; 195 196struct uma_bucket_zone { 197 uma_zone_t ubz_zone; 198 char *ubz_name; 199 int ubz_entries; /* Number of items it can hold. */ 200 int ubz_maxsize; /* Maximum allocation size per-item. */ 201}; 202 203/* 204 * Compute the actual number of bucket entries to pack them in power 205 * of two sizes for more efficient space utilization. 206 */ 207#define BUCKET_SIZE(n) \ 208 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 209 210#define BUCKET_MAX BUCKET_SIZE(256) 211 212struct uma_bucket_zone bucket_zones[] = { 213 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 214 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 215 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 216 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 217 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 218 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 219 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 220 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 221 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 222 { NULL, NULL, 0} 223}; 224 225/* 226 * Flags and enumerations to be passed to internal functions. 227 */ 228enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; 229 230/* Prototypes.. */ 231 232static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 233static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 234static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 235static void page_free(void *, vm_size_t, uint8_t); 236static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); 237static void cache_drain(uma_zone_t); 238static void bucket_drain(uma_zone_t, uma_bucket_t); 239static void bucket_cache_drain(uma_zone_t zone); 240static int keg_ctor(void *, int, void *, int); 241static void keg_dtor(void *, int, void *); 242static int zone_ctor(void *, int, void *, int); 243static void zone_dtor(void *, int, void *); 244static int zero_init(void *, int, int); 245static void keg_small_init(uma_keg_t keg); 246static void keg_large_init(uma_keg_t keg); 247static void zone_foreach(void (*zfunc)(uma_zone_t)); 248static void zone_timeout(uma_zone_t zone); 249static int hash_alloc(struct uma_hash *); 250static int hash_expand(struct uma_hash *, struct uma_hash *); 251static void hash_free(struct uma_hash *hash); 252static void uma_timeout(void *); 253static void uma_startup3(void); 254static void *zone_alloc_item(uma_zone_t, void *, int); 255static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 256static void bucket_enable(void); 257static void bucket_init(void); 258static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 259static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 260static void bucket_zone_drain(void); 261static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags); 262static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); 263static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); 264static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 265static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); 266static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 267 uma_fini fini, int align, uint32_t flags); 268static int zone_import(uma_zone_t zone, void **bucket, int max, int flags); 269static void zone_release(uma_zone_t zone, void **bucket, int cnt); 270static void uma_zero_item(void *item, uma_zone_t zone); 271 272void uma_print_zone(uma_zone_t); 273void uma_print_stats(void); 274static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 275static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 276 277SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 278 279SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 280 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 281 282SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 283 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 284 285static int zone_warnings = 1; 286TUNABLE_INT("vm.zone_warnings", &zone_warnings); 287SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0, 288 "Warn when UMA zones becomes full"); 289 290/* 291 * This routine checks to see whether or not it's safe to enable buckets. 292 */ 293static void 294bucket_enable(void) 295{ 296 bucketdisable = vm_page_count_min(); 297} 298 299/* 300 * Initialize bucket_zones, the array of zones of buckets of various sizes. 301 * 302 * For each zone, calculate the memory required for each bucket, consisting 303 * of the header and an array of pointers. 304 */ 305static void 306bucket_init(void) 307{ 308 struct uma_bucket_zone *ubz; 309 int size; 310 int i; 311 312 for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 313 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 314 size += sizeof(void *) * ubz->ubz_entries; 315 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 316 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 317 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET); 318 } 319} 320 321/* 322 * Given a desired number of entries for a bucket, return the zone from which 323 * to allocate the bucket. 324 */ 325static struct uma_bucket_zone * 326bucket_zone_lookup(int entries) 327{ 328 struct uma_bucket_zone *ubz; 329 330 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 331 if (ubz->ubz_entries >= entries) 332 return (ubz); 333 ubz--; 334 return (ubz); 335} 336 337static int 338bucket_select(int size) 339{ 340 struct uma_bucket_zone *ubz; 341 342 ubz = &bucket_zones[0]; 343 if (size > ubz->ubz_maxsize) 344 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 345 346 for (; ubz->ubz_entries != 0; ubz++) 347 if (ubz->ubz_maxsize < size) 348 break; 349 ubz--; 350 return (ubz->ubz_entries); 351} 352 353static uma_bucket_t 354bucket_alloc(uma_zone_t zone, void *udata, int flags) 355{ 356 struct uma_bucket_zone *ubz; 357 uma_bucket_t bucket; 358 359 /* 360 * This is to stop us from allocating per cpu buckets while we're 361 * running out of vm.boot_pages. Otherwise, we would exhaust the 362 * boot pages. This also prevents us from allocating buckets in 363 * low memory situations. 364 */ 365 if (bucketdisable) 366 return (NULL); 367 /* 368 * To limit bucket recursion we store the original zone flags 369 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 370 * NOVM flag to persist even through deep recursions. We also 371 * store ZFLAG_BUCKET once we have recursed attempting to allocate 372 * a bucket for a bucket zone so we do not allow infinite bucket 373 * recursion. This cookie will even persist to frees of unused 374 * buckets via the allocation path or bucket allocations in the 375 * free path. 376 */ 377 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 378 udata = (void *)(uintptr_t)zone->uz_flags; 379 else { 380 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 381 return (NULL); 382 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 383 } 384 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 385 flags |= M_NOVM; 386 ubz = bucket_zone_lookup(zone->uz_count); 387 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 388 ubz++; 389 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 390 if (bucket) { 391#ifdef INVARIANTS 392 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 393#endif 394 bucket->ub_cnt = 0; 395 bucket->ub_entries = ubz->ubz_entries; 396 } 397 398 return (bucket); 399} 400 401static void 402bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 403{ 404 struct uma_bucket_zone *ubz; 405 406 KASSERT(bucket->ub_cnt == 0, 407 ("bucket_free: Freeing a non free bucket.")); 408 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 409 udata = (void *)(uintptr_t)zone->uz_flags; 410 ubz = bucket_zone_lookup(bucket->ub_entries); 411 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 412} 413 414static void 415bucket_zone_drain(void) 416{ 417 struct uma_bucket_zone *ubz; 418 419 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 420 zone_drain(ubz->ubz_zone); 421} 422 423static void 424zone_log_warning(uma_zone_t zone) 425{ 426 static const struct timeval warninterval = { 300, 0 }; 427 428 if (!zone_warnings || zone->uz_warning == NULL) 429 return; 430 431 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 432 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 433} 434 435static void 436zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 437{ 438 uma_klink_t klink; 439 440 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 441 kegfn(klink->kl_keg); 442} 443 444/* 445 * Routine called by timeout which is used to fire off some time interval 446 * based calculations. (stats, hash size, etc.) 447 * 448 * Arguments: 449 * arg Unused 450 * 451 * Returns: 452 * Nothing 453 */ 454static void 455uma_timeout(void *unused) 456{ 457 bucket_enable(); 458 zone_foreach(zone_timeout); 459 460 /* Reschedule this event */ 461 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 462} 463 464/* 465 * Routine to perform timeout driven calculations. This expands the 466 * hashes and does per cpu statistics aggregation. 467 * 468 * Returns nothing. 469 */ 470static void 471keg_timeout(uma_keg_t keg) 472{ 473 474 KEG_LOCK(keg); 475 /* 476 * Expand the keg hash table. 477 * 478 * This is done if the number of slabs is larger than the hash size. 479 * What I'm trying to do here is completely reduce collisions. This 480 * may be a little aggressive. Should I allow for two collisions max? 481 */ 482 if (keg->uk_flags & UMA_ZONE_HASH && 483 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 484 struct uma_hash newhash; 485 struct uma_hash oldhash; 486 int ret; 487 488 /* 489 * This is so involved because allocating and freeing 490 * while the keg lock is held will lead to deadlock. 491 * I have to do everything in stages and check for 492 * races. 493 */ 494 newhash = keg->uk_hash; 495 KEG_UNLOCK(keg); 496 ret = hash_alloc(&newhash); 497 KEG_LOCK(keg); 498 if (ret) { 499 if (hash_expand(&keg->uk_hash, &newhash)) { 500 oldhash = keg->uk_hash; 501 keg->uk_hash = newhash; 502 } else 503 oldhash = newhash; 504 505 KEG_UNLOCK(keg); 506 hash_free(&oldhash); 507 return; 508 } 509 } 510 KEG_UNLOCK(keg); 511} 512 513static void 514zone_timeout(uma_zone_t zone) 515{ 516 517 zone_foreach_keg(zone, &keg_timeout); 518} 519 520/* 521 * Allocate and zero fill the next sized hash table from the appropriate 522 * backing store. 523 * 524 * Arguments: 525 * hash A new hash structure with the old hash size in uh_hashsize 526 * 527 * Returns: 528 * 1 on sucess and 0 on failure. 529 */ 530static int 531hash_alloc(struct uma_hash *hash) 532{ 533 int oldsize; 534 int alloc; 535 536 oldsize = hash->uh_hashsize; 537 538 /* We're just going to go to a power of two greater */ 539 if (oldsize) { 540 hash->uh_hashsize = oldsize * 2; 541 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 542 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 543 M_UMAHASH, M_NOWAIT); 544 } else { 545 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 546 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 547 M_WAITOK); 548 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 549 } 550 if (hash->uh_slab_hash) { 551 bzero(hash->uh_slab_hash, alloc); 552 hash->uh_hashmask = hash->uh_hashsize - 1; 553 return (1); 554 } 555 556 return (0); 557} 558 559/* 560 * Expands the hash table for HASH zones. This is done from zone_timeout 561 * to reduce collisions. This must not be done in the regular allocation 562 * path, otherwise, we can recurse on the vm while allocating pages. 563 * 564 * Arguments: 565 * oldhash The hash you want to expand 566 * newhash The hash structure for the new table 567 * 568 * Returns: 569 * Nothing 570 * 571 * Discussion: 572 */ 573static int 574hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 575{ 576 uma_slab_t slab; 577 int hval; 578 int i; 579 580 if (!newhash->uh_slab_hash) 581 return (0); 582 583 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 584 return (0); 585 586 /* 587 * I need to investigate hash algorithms for resizing without a 588 * full rehash. 589 */ 590 591 for (i = 0; i < oldhash->uh_hashsize; i++) 592 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 593 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 594 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 595 hval = UMA_HASH(newhash, slab->us_data); 596 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 597 slab, us_hlink); 598 } 599 600 return (1); 601} 602 603/* 604 * Free the hash bucket to the appropriate backing store. 605 * 606 * Arguments: 607 * slab_hash The hash bucket we're freeing 608 * hashsize The number of entries in that hash bucket 609 * 610 * Returns: 611 * Nothing 612 */ 613static void 614hash_free(struct uma_hash *hash) 615{ 616 if (hash->uh_slab_hash == NULL) 617 return; 618 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 619 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 620 else 621 free(hash->uh_slab_hash, M_UMAHASH); 622} 623 624/* 625 * Frees all outstanding items in a bucket 626 * 627 * Arguments: 628 * zone The zone to free to, must be unlocked. 629 * bucket The free/alloc bucket with items, cpu queue must be locked. 630 * 631 * Returns: 632 * Nothing 633 */ 634 635static void 636bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 637{ 638 int i; 639 640 if (bucket == NULL) 641 return; 642 643 if (zone->uz_fini) 644 for (i = 0; i < bucket->ub_cnt; i++) 645 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 646 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 647 bucket->ub_cnt = 0; 648} 649 650/* 651 * Drains the per cpu caches for a zone. 652 * 653 * NOTE: This may only be called while the zone is being turn down, and not 654 * during normal operation. This is necessary in order that we do not have 655 * to migrate CPUs to drain the per-CPU caches. 656 * 657 * Arguments: 658 * zone The zone to drain, must be unlocked. 659 * 660 * Returns: 661 * Nothing 662 */ 663static void 664cache_drain(uma_zone_t zone) 665{ 666 uma_cache_t cache; 667 int cpu; 668 669 /* 670 * XXX: It is safe to not lock the per-CPU caches, because we're 671 * tearing down the zone anyway. I.e., there will be no further use 672 * of the caches at this point. 673 * 674 * XXX: It would good to be able to assert that the zone is being 675 * torn down to prevent improper use of cache_drain(). 676 * 677 * XXX: We lock the zone before passing into bucket_cache_drain() as 678 * it is used elsewhere. Should the tear-down path be made special 679 * there in some form? 680 */ 681 CPU_FOREACH(cpu) { 682 cache = &zone->uz_cpu[cpu]; 683 bucket_drain(zone, cache->uc_allocbucket); 684 bucket_drain(zone, cache->uc_freebucket); 685 if (cache->uc_allocbucket != NULL) 686 bucket_free(zone, cache->uc_allocbucket, NULL); 687 if (cache->uc_freebucket != NULL) 688 bucket_free(zone, cache->uc_freebucket, NULL); 689 cache->uc_allocbucket = cache->uc_freebucket = NULL; 690 } 691 ZONE_LOCK(zone); 692 bucket_cache_drain(zone); 693 ZONE_UNLOCK(zone); 694} 695 696static void 697cache_shrink(uma_zone_t zone) 698{ 699 700 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 701 return; 702 703 ZONE_LOCK(zone); 704 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; 705 ZONE_UNLOCK(zone); 706} 707 708static void 709cache_drain_safe_cpu(uma_zone_t zone) 710{ 711 uma_cache_t cache; 712 uma_bucket_t b1, b2; 713 714 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 715 return; 716 717 b1 = b2 = NULL; 718 ZONE_LOCK(zone); 719 critical_enter(); 720 cache = &zone->uz_cpu[curcpu]; 721 if (cache->uc_allocbucket) { 722 if (cache->uc_allocbucket->ub_cnt != 0) 723 LIST_INSERT_HEAD(&zone->uz_buckets, 724 cache->uc_allocbucket, ub_link); 725 else 726 b1 = cache->uc_allocbucket; 727 cache->uc_allocbucket = NULL; 728 } 729 if (cache->uc_freebucket) { 730 if (cache->uc_freebucket->ub_cnt != 0) 731 LIST_INSERT_HEAD(&zone->uz_buckets, 732 cache->uc_freebucket, ub_link); 733 else 734 b2 = cache->uc_freebucket; 735 cache->uc_freebucket = NULL; 736 } 737 critical_exit(); 738 ZONE_UNLOCK(zone); 739 if (b1) 740 bucket_free(zone, b1, NULL); 741 if (b2) 742 bucket_free(zone, b2, NULL); 743} 744 745/* 746 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 747 * This is an expensive call because it needs to bind to all CPUs 748 * one by one and enter a critical section on each of them in order 749 * to safely access their cache buckets. 750 * Zone lock must not be held on call this function. 751 */ 752static void 753cache_drain_safe(uma_zone_t zone) 754{ 755 int cpu; 756 757 /* 758 * Polite bucket sizes shrinking was not enouth, shrink aggressively. 759 */ 760 if (zone) 761 cache_shrink(zone); 762 else 763 zone_foreach(cache_shrink); 764 765 CPU_FOREACH(cpu) { 766 thread_lock(curthread); 767 sched_bind(curthread, cpu); 768 thread_unlock(curthread); 769 770 if (zone) 771 cache_drain_safe_cpu(zone); 772 else 773 zone_foreach(cache_drain_safe_cpu); 774 } 775 thread_lock(curthread); 776 sched_unbind(curthread); 777 thread_unlock(curthread); 778} 779 780/* 781 * Drain the cached buckets from a zone. Expects a locked zone on entry. 782 */ 783static void 784bucket_cache_drain(uma_zone_t zone) 785{ 786 uma_bucket_t bucket; 787 788 /* 789 * Drain the bucket queues and free the buckets, we just keep two per 790 * cpu (alloc/free). 791 */ 792 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 793 LIST_REMOVE(bucket, ub_link); 794 ZONE_UNLOCK(zone); 795 bucket_drain(zone, bucket); 796 bucket_free(zone, bucket, NULL); 797 ZONE_LOCK(zone); 798 } 799 800 /* 801 * Shrink further bucket sizes. Price of single zone lock collision 802 * is probably lower then price of global cache drain. 803 */ 804 if (zone->uz_count > zone->uz_count_min) 805 zone->uz_count--; 806} 807 808static void 809keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 810{ 811 uint8_t *mem; 812 int i; 813 uint8_t flags; 814 815 mem = slab->us_data; 816 flags = slab->us_flags; 817 i = start; 818 if (keg->uk_fini != NULL) { 819 for (i--; i > -1; i--) 820 keg->uk_fini(slab->us_data + (keg->uk_rsize * i), 821 keg->uk_size); 822 } 823 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 824 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 825#ifdef UMA_DEBUG 826 printf("%s: Returning %d bytes.\n", keg->uk_name, 827 PAGE_SIZE * keg->uk_ppera); 828#endif 829 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 830} 831 832/* 833 * Frees pages from a keg back to the system. This is done on demand from 834 * the pageout daemon. 835 * 836 * Returns nothing. 837 */ 838static void 839keg_drain(uma_keg_t keg) 840{ 841 struct slabhead freeslabs = { 0 }; 842 uma_slab_t slab; 843 uma_slab_t n; 844 845 /* 846 * We don't want to take pages from statically allocated kegs at this 847 * time 848 */ 849 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 850 return; 851 852#ifdef UMA_DEBUG 853 printf("%s free items: %u\n", keg->uk_name, keg->uk_free); 854#endif 855 KEG_LOCK(keg); 856 if (keg->uk_free == 0) 857 goto finished; 858 859 slab = LIST_FIRST(&keg->uk_free_slab); 860 while (slab) { 861 n = LIST_NEXT(slab, us_link); 862 863 /* We have no where to free these to */ 864 if (slab->us_flags & UMA_SLAB_BOOT) { 865 slab = n; 866 continue; 867 } 868 869 LIST_REMOVE(slab, us_link); 870 keg->uk_pages -= keg->uk_ppera; 871 keg->uk_free -= keg->uk_ipers; 872 873 if (keg->uk_flags & UMA_ZONE_HASH) 874 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); 875 876 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 877 878 slab = n; 879 } 880finished: 881 KEG_UNLOCK(keg); 882 883 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 884 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 885 keg_free_slab(keg, slab, keg->uk_ipers); 886 } 887} 888 889static void 890zone_drain_wait(uma_zone_t zone, int waitok) 891{ 892 893 /* 894 * Set draining to interlock with zone_dtor() so we can release our 895 * locks as we go. Only dtor() should do a WAITOK call since it 896 * is the only call that knows the structure will still be available 897 * when it wakes up. 898 */ 899 ZONE_LOCK(zone); 900 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 901 if (waitok == M_NOWAIT) 902 goto out; 903 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 904 } 905 zone->uz_flags |= UMA_ZFLAG_DRAINING; 906 bucket_cache_drain(zone); 907 ZONE_UNLOCK(zone); 908 /* 909 * The DRAINING flag protects us from being freed while 910 * we're running. Normally the uma_rwlock would protect us but we 911 * must be able to release and acquire the right lock for each keg. 912 */ 913 zone_foreach_keg(zone, &keg_drain); 914 ZONE_LOCK(zone); 915 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 916 wakeup(zone); 917out: 918 ZONE_UNLOCK(zone); 919} 920 921void 922zone_drain(uma_zone_t zone) 923{ 924 925 zone_drain_wait(zone, M_NOWAIT); 926} 927 928/* 929 * Allocate a new slab for a keg. This does not insert the slab onto a list. 930 * 931 * Arguments: 932 * wait Shall we wait? 933 * 934 * Returns: 935 * The slab that was allocated or NULL if there is no memory and the 936 * caller specified M_NOWAIT. 937 */ 938static uma_slab_t 939keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) 940{ 941 uma_slabrefcnt_t slabref; 942 uma_alloc allocf; 943 uma_slab_t slab; 944 uint8_t *mem; 945 uint8_t flags; 946 int i; 947 948 mtx_assert(&keg->uk_lock, MA_OWNED); 949 slab = NULL; 950 mem = NULL; 951 952#ifdef UMA_DEBUG 953 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name); 954#endif 955 allocf = keg->uk_allocf; 956 KEG_UNLOCK(keg); 957 958 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 959 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); 960 if (slab == NULL) 961 goto out; 962 } 963 964 /* 965 * This reproduces the old vm_zone behavior of zero filling pages the 966 * first time they are added to a zone. 967 * 968 * Malloced items are zeroed in uma_zalloc. 969 */ 970 971 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 972 wait |= M_ZERO; 973 else 974 wait &= ~M_ZERO; 975 976 if (keg->uk_flags & UMA_ZONE_NODUMP) 977 wait |= M_NODUMP; 978 979 /* zone is passed for legacy reasons. */ 980 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait); 981 if (mem == NULL) { 982 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 983 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 984 slab = NULL; 985 goto out; 986 } 987 988 /* Point the slab into the allocated memory */ 989 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 990 slab = (uma_slab_t )(mem + keg->uk_pgoff); 991 992 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 993 for (i = 0; i < keg->uk_ppera; i++) 994 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 995 996 slab->us_keg = keg; 997 slab->us_data = mem; 998 slab->us_freecount = keg->uk_ipers; 999 slab->us_flags = flags; 1000 BIT_FILL(SLAB_SETSIZE, &slab->us_free); 1001#ifdef INVARIANTS 1002 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); 1003#endif 1004 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1005 slabref = (uma_slabrefcnt_t)slab; 1006 for (i = 0; i < keg->uk_ipers; i++) 1007 slabref->us_refcnt[i] = 0; 1008 } 1009 1010 if (keg->uk_init != NULL) { 1011 for (i = 0; i < keg->uk_ipers; i++) 1012 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 1013 keg->uk_size, wait) != 0) 1014 break; 1015 if (i != keg->uk_ipers) { 1016 keg_free_slab(keg, slab, i); 1017 slab = NULL; 1018 goto out; 1019 } 1020 } 1021out: 1022 KEG_LOCK(keg); 1023 1024 if (slab != NULL) { 1025 if (keg->uk_flags & UMA_ZONE_HASH) 1026 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1027 1028 keg->uk_pages += keg->uk_ppera; 1029 keg->uk_free += keg->uk_ipers; 1030 } 1031 1032 return (slab); 1033} 1034 1035/* 1036 * This function is intended to be used early on in place of page_alloc() so 1037 * that we may use the boot time page cache to satisfy allocations before 1038 * the VM is ready. 1039 */ 1040static void * 1041startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) 1042{ 1043 uma_keg_t keg; 1044 uma_slab_t tmps; 1045 int pages, check_pages; 1046 1047 keg = zone_first_keg(zone); 1048 pages = howmany(bytes, PAGE_SIZE); 1049 check_pages = pages - 1; 1050 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n")); 1051 1052 /* 1053 * Check our small startup cache to see if it has pages remaining. 1054 */ 1055 mtx_lock(&uma_boot_pages_mtx); 1056 1057 /* First check if we have enough room. */ 1058 tmps = LIST_FIRST(&uma_boot_pages); 1059 while (tmps != NULL && check_pages-- > 0) 1060 tmps = LIST_NEXT(tmps, us_link); 1061 if (tmps != NULL) { 1062 /* 1063 * It's ok to lose tmps references. The last one will 1064 * have tmps->us_data pointing to the start address of 1065 * "pages" contiguous pages of memory. 1066 */ 1067 while (pages-- > 0) { 1068 tmps = LIST_FIRST(&uma_boot_pages); 1069 LIST_REMOVE(tmps, us_link); 1070 } 1071 mtx_unlock(&uma_boot_pages_mtx); 1072 *pflag = tmps->us_flags; 1073 return (tmps->us_data); 1074 } 1075 mtx_unlock(&uma_boot_pages_mtx); 1076 if (booted < UMA_STARTUP2) 1077 panic("UMA: Increase vm.boot_pages"); 1078 /* 1079 * Now that we've booted reset these users to their real allocator. 1080 */ 1081#ifdef UMA_MD_SMALL_ALLOC 1082 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; 1083#else 1084 keg->uk_allocf = page_alloc; 1085#endif 1086 return keg->uk_allocf(zone, bytes, pflag, wait); 1087} 1088 1089/* 1090 * Allocates a number of pages from the system 1091 * 1092 * Arguments: 1093 * bytes The number of bytes requested 1094 * wait Shall we wait? 1095 * 1096 * Returns: 1097 * A pointer to the alloced memory or possibly 1098 * NULL if M_NOWAIT is set. 1099 */ 1100static void * 1101page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) 1102{ 1103 void *p; /* Returned page */ 1104 1105 *pflag = UMA_SLAB_KMEM; 1106 p = (void *) kmem_malloc(kmem_arena, bytes, wait); 1107 1108 return (p); 1109} 1110 1111/* 1112 * Allocates a number of pages from within an object 1113 * 1114 * Arguments: 1115 * bytes The number of bytes requested 1116 * wait Shall we wait? 1117 * 1118 * Returns: 1119 * A pointer to the alloced memory or possibly 1120 * NULL if M_NOWAIT is set. 1121 */ 1122static void * 1123noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait) 1124{ 1125 TAILQ_HEAD(, vm_page) alloctail; 1126 u_long npages; 1127 vm_offset_t retkva, zkva; 1128 vm_page_t p, p_next; 1129 uma_keg_t keg; 1130 1131 TAILQ_INIT(&alloctail); 1132 keg = zone_first_keg(zone); 1133 1134 npages = howmany(bytes, PAGE_SIZE); 1135 while (npages > 0) { 1136 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT | 1137 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ); 1138 if (p != NULL) { 1139 /* 1140 * Since the page does not belong to an object, its 1141 * listq is unused. 1142 */ 1143 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1144 npages--; 1145 continue; 1146 } 1147 if (wait & M_WAITOK) { 1148 VM_WAIT; 1149 continue; 1150 } 1151 1152 /* 1153 * Page allocation failed, free intermediate pages and 1154 * exit. 1155 */ 1156 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1157 vm_page_unwire(p, 0); 1158 vm_page_free(p); 1159 } 1160 return (NULL); 1161 } 1162 *flags = UMA_SLAB_PRIV; 1163 zkva = keg->uk_kva + 1164 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1165 retkva = zkva; 1166 TAILQ_FOREACH(p, &alloctail, listq) { 1167 pmap_qenter(zkva, &p, 1); 1168 zkva += PAGE_SIZE; 1169 } 1170 1171 return ((void *)retkva); 1172} 1173 1174/* 1175 * Frees a number of pages to the system 1176 * 1177 * Arguments: 1178 * mem A pointer to the memory to be freed 1179 * size The size of the memory being freed 1180 * flags The original p->us_flags field 1181 * 1182 * Returns: 1183 * Nothing 1184 */ 1185static void 1186page_free(void *mem, vm_size_t size, uint8_t flags) 1187{ 1188 struct vmem *vmem; 1189 1190 if (flags & UMA_SLAB_KMEM) 1191 vmem = kmem_arena; 1192 else if (flags & UMA_SLAB_KERNEL) 1193 vmem = kernel_arena; 1194 else 1195 panic("UMA: page_free used with invalid flags %d", flags); 1196 1197 kmem_free(vmem, (vm_offset_t)mem, size); 1198} 1199 1200/* 1201 * Zero fill initializer 1202 * 1203 * Arguments/Returns follow uma_init specifications 1204 */ 1205static int 1206zero_init(void *mem, int size, int flags) 1207{ 1208 bzero(mem, size); 1209 return (0); 1210} 1211 1212/* 1213 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1214 * 1215 * Arguments 1216 * keg The zone we should initialize 1217 * 1218 * Returns 1219 * Nothing 1220 */ 1221static void 1222keg_small_init(uma_keg_t keg) 1223{ 1224 u_int rsize; 1225 u_int memused; 1226 u_int wastedspace; 1227 u_int shsize; 1228 1229 if (keg->uk_flags & UMA_ZONE_PCPU) { 1230 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU; 1231 1232 keg->uk_slabsize = sizeof(struct pcpu); 1233 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu), 1234 PAGE_SIZE); 1235 } else { 1236 keg->uk_slabsize = UMA_SLAB_SIZE; 1237 keg->uk_ppera = 1; 1238 } 1239 1240 /* 1241 * Calculate the size of each allocation (rsize) according to 1242 * alignment. If the requested size is smaller than we have 1243 * allocation bits for we round it up. 1244 */ 1245 rsize = keg->uk_size; 1246 if (rsize < keg->uk_slabsize / SLAB_SETSIZE) 1247 rsize = keg->uk_slabsize / SLAB_SETSIZE; 1248 if (rsize & keg->uk_align) 1249 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1250 keg->uk_rsize = rsize; 1251 1252 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1253 keg->uk_rsize < sizeof(struct pcpu), 1254 ("%s: size %u too large", __func__, keg->uk_rsize)); 1255 1256 if (keg->uk_flags & UMA_ZONE_REFCNT) 1257 rsize += sizeof(uint32_t); 1258 1259 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1260 shsize = 0; 1261 else 1262 shsize = sizeof(struct uma_slab); 1263 1264 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize; 1265 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1266 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1267 1268 memused = keg->uk_ipers * rsize + shsize; 1269 wastedspace = keg->uk_slabsize - memused; 1270 1271 /* 1272 * We can't do OFFPAGE if we're internal or if we've been 1273 * asked to not go to the VM for buckets. If we do this we 1274 * may end up going to the VM for slabs which we do not 1275 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1276 * of UMA_ZONE_VM, which clearly forbids it. 1277 */ 1278 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1279 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1280 return; 1281 1282 /* 1283 * See if using an OFFPAGE slab will limit our waste. Only do 1284 * this if it permits more items per-slab. 1285 * 1286 * XXX We could try growing slabsize to limit max waste as well. 1287 * Historically this was not done because the VM could not 1288 * efficiently handle contiguous allocations. 1289 */ 1290 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) && 1291 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) { 1292 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize; 1293 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1294 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1295#ifdef UMA_DEBUG 1296 printf("UMA decided we need offpage slab headers for " 1297 "keg: %s, calculated wastedspace = %d, " 1298 "maximum wasted space allowed = %d, " 1299 "calculated ipers = %d, " 1300 "new wasted space = %d\n", keg->uk_name, wastedspace, 1301 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1302 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize); 1303#endif 1304 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1305 } 1306 1307 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1308 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1309 keg->uk_flags |= UMA_ZONE_HASH; 1310} 1311 1312/* 1313 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1314 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1315 * more complicated. 1316 * 1317 * Arguments 1318 * keg The keg we should initialize 1319 * 1320 * Returns 1321 * Nothing 1322 */ 1323static void 1324keg_large_init(uma_keg_t keg) 1325{ 1326 u_int shsize; 1327 1328 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1329 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1330 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1331 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1332 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1333 1334 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1335 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE; 1336 keg->uk_ipers = 1; 1337 keg->uk_rsize = keg->uk_size; 1338 1339 /* We can't do OFFPAGE if we're internal, bail out here. */ 1340 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) 1341 return; 1342 1343 /* Check whether we have enough space to not do OFFPAGE. */ 1344 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) { 1345 shsize = sizeof(struct uma_slab); 1346 if (keg->uk_flags & UMA_ZONE_REFCNT) 1347 shsize += keg->uk_ipers * sizeof(uint32_t); 1348 if (shsize & UMA_ALIGN_PTR) 1349 shsize = (shsize & ~UMA_ALIGN_PTR) + 1350 (UMA_ALIGN_PTR + 1); 1351 1352 if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize) 1353 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1354 } 1355 1356 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1357 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1358 keg->uk_flags |= UMA_ZONE_HASH; 1359} 1360 1361static void 1362keg_cachespread_init(uma_keg_t keg) 1363{ 1364 int alignsize; 1365 int trailer; 1366 int pages; 1367 int rsize; 1368 1369 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1370 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1371 1372 alignsize = keg->uk_align + 1; 1373 rsize = keg->uk_size; 1374 /* 1375 * We want one item to start on every align boundary in a page. To 1376 * do this we will span pages. We will also extend the item by the 1377 * size of align if it is an even multiple of align. Otherwise, it 1378 * would fall on the same boundary every time. 1379 */ 1380 if (rsize & keg->uk_align) 1381 rsize = (rsize & ~keg->uk_align) + alignsize; 1382 if ((rsize & alignsize) == 0) 1383 rsize += alignsize; 1384 trailer = rsize - keg->uk_size; 1385 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1386 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1387 keg->uk_rsize = rsize; 1388 keg->uk_ppera = pages; 1389 keg->uk_slabsize = UMA_SLAB_SIZE; 1390 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1391 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1392 KASSERT(keg->uk_ipers <= SLAB_SETSIZE, 1393 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1394 keg->uk_ipers)); 1395} 1396 1397/* 1398 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1399 * the keg onto the global keg list. 1400 * 1401 * Arguments/Returns follow uma_ctor specifications 1402 * udata Actually uma_kctor_args 1403 */ 1404static int 1405keg_ctor(void *mem, int size, void *udata, int flags) 1406{ 1407 struct uma_kctor_args *arg = udata; 1408 uma_keg_t keg = mem; 1409 uma_zone_t zone; 1410 1411 bzero(keg, size); 1412 keg->uk_size = arg->size; 1413 keg->uk_init = arg->uminit; 1414 keg->uk_fini = arg->fini; 1415 keg->uk_align = arg->align; 1416 keg->uk_free = 0; 1417 keg->uk_reserve = 0; 1418 keg->uk_pages = 0; 1419 keg->uk_flags = arg->flags; 1420 keg->uk_allocf = page_alloc; 1421 keg->uk_freef = page_free; 1422 keg->uk_slabzone = NULL; 1423 1424 /* 1425 * The master zone is passed to us at keg-creation time. 1426 */ 1427 zone = arg->zone; 1428 keg->uk_name = zone->uz_name; 1429 1430 if (arg->flags & UMA_ZONE_VM) 1431 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1432 1433 if (arg->flags & UMA_ZONE_ZINIT) 1434 keg->uk_init = zero_init; 1435 1436 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC) 1437 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1438 1439 if (arg->flags & UMA_ZONE_PCPU) 1440#ifdef SMP 1441 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1442#else 1443 keg->uk_flags &= ~UMA_ZONE_PCPU; 1444#endif 1445 1446 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1447 keg_cachespread_init(keg); 1448 } else if (keg->uk_flags & UMA_ZONE_REFCNT) { 1449 if (keg->uk_size > 1450 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - 1451 sizeof(uint32_t))) 1452 keg_large_init(keg); 1453 else 1454 keg_small_init(keg); 1455 } else { 1456 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab))) 1457 keg_large_init(keg); 1458 else 1459 keg_small_init(keg); 1460 } 1461 1462 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1463 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1464 if (keg->uk_ipers > uma_max_ipers_ref) 1465 panic("Too many ref items per zone: %d > %d\n", 1466 keg->uk_ipers, uma_max_ipers_ref); 1467 keg->uk_slabzone = slabrefzone; 1468 } else 1469 keg->uk_slabzone = slabzone; 1470 } 1471 1472 /* 1473 * If we haven't booted yet we need allocations to go through the 1474 * startup cache until the vm is ready. 1475 */ 1476 if (keg->uk_ppera == 1) { 1477#ifdef UMA_MD_SMALL_ALLOC 1478 keg->uk_allocf = uma_small_alloc; 1479 keg->uk_freef = uma_small_free; 1480 1481 if (booted < UMA_STARTUP) 1482 keg->uk_allocf = startup_alloc; 1483#else 1484 if (booted < UMA_STARTUP2) 1485 keg->uk_allocf = startup_alloc; 1486#endif 1487 } else if (booted < UMA_STARTUP2 && 1488 (keg->uk_flags & UMA_ZFLAG_INTERNAL)) 1489 keg->uk_allocf = startup_alloc; 1490 1491 /* 1492 * Initialize keg's lock 1493 */ 1494 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1495 1496 /* 1497 * If we're putting the slab header in the actual page we need to 1498 * figure out where in each page it goes. This calculates a right 1499 * justified offset into the memory on an ALIGN_PTR boundary. 1500 */ 1501 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1502 u_int totsize; 1503 1504 /* Size of the slab struct and free list */ 1505 totsize = sizeof(struct uma_slab); 1506 1507 /* Size of the reference counts. */ 1508 if (keg->uk_flags & UMA_ZONE_REFCNT) 1509 totsize += keg->uk_ipers * sizeof(uint32_t); 1510 1511 if (totsize & UMA_ALIGN_PTR) 1512 totsize = (totsize & ~UMA_ALIGN_PTR) + 1513 (UMA_ALIGN_PTR + 1); 1514 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1515 1516 /* 1517 * The only way the following is possible is if with our 1518 * UMA_ALIGN_PTR adjustments we are now bigger than 1519 * UMA_SLAB_SIZE. I haven't checked whether this is 1520 * mathematically possible for all cases, so we make 1521 * sure here anyway. 1522 */ 1523 totsize = keg->uk_pgoff + sizeof(struct uma_slab); 1524 if (keg->uk_flags & UMA_ZONE_REFCNT) 1525 totsize += keg->uk_ipers * sizeof(uint32_t); 1526 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1527 printf("zone %s ipers %d rsize %d size %d\n", 1528 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1529 keg->uk_size); 1530 panic("UMA slab won't fit."); 1531 } 1532 } 1533 1534 if (keg->uk_flags & UMA_ZONE_HASH) 1535 hash_alloc(&keg->uk_hash); 1536 1537#ifdef UMA_DEBUG 1538 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n", 1539 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags, 1540 keg->uk_ipers, keg->uk_ppera, 1541 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free); 1542#endif 1543 1544 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1545 1546 rw_wlock(&uma_rwlock); 1547 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1548 rw_wunlock(&uma_rwlock); 1549 return (0); 1550} 1551 1552/* 1553 * Zone header ctor. This initializes all fields, locks, etc. 1554 * 1555 * Arguments/Returns follow uma_ctor specifications 1556 * udata Actually uma_zctor_args 1557 */ 1558static int 1559zone_ctor(void *mem, int size, void *udata, int flags) 1560{ 1561 struct uma_zctor_args *arg = udata; 1562 uma_zone_t zone = mem; 1563 uma_zone_t z; 1564 uma_keg_t keg; 1565 1566 bzero(zone, size); 1567 zone->uz_name = arg->name; 1568 zone->uz_ctor = arg->ctor; 1569 zone->uz_dtor = arg->dtor; 1570 zone->uz_slab = zone_fetch_slab; 1571 zone->uz_init = NULL; 1572 zone->uz_fini = NULL; 1573 zone->uz_allocs = 0; 1574 zone->uz_frees = 0; 1575 zone->uz_fails = 0; 1576 zone->uz_sleeps = 0; 1577 zone->uz_count = 0; 1578 zone->uz_count_min = 0; 1579 zone->uz_flags = 0; 1580 zone->uz_warning = NULL; 1581 timevalclear(&zone->uz_ratecheck); 1582 keg = arg->keg; 1583 1584 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 1585 1586 /* 1587 * This is a pure cache zone, no kegs. 1588 */ 1589 if (arg->import) { 1590 if (arg->flags & UMA_ZONE_VM) 1591 arg->flags |= UMA_ZFLAG_CACHEONLY; 1592 zone->uz_flags = arg->flags; 1593 zone->uz_size = arg->size; 1594 zone->uz_import = arg->import; 1595 zone->uz_release = arg->release; 1596 zone->uz_arg = arg->arg; 1597 zone->uz_lockptr = &zone->uz_lock; 1598 rw_wlock(&uma_rwlock); 1599 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 1600 rw_wunlock(&uma_rwlock); 1601 goto out; 1602 } 1603 1604 /* 1605 * Use the regular zone/keg/slab allocator. 1606 */ 1607 zone->uz_import = (uma_import)zone_import; 1608 zone->uz_release = (uma_release)zone_release; 1609 zone->uz_arg = zone; 1610 1611 if (arg->flags & UMA_ZONE_SECONDARY) { 1612 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1613 zone->uz_init = arg->uminit; 1614 zone->uz_fini = arg->fini; 1615 zone->uz_lockptr = &keg->uk_lock; 1616 zone->uz_flags |= UMA_ZONE_SECONDARY; 1617 rw_wlock(&uma_rwlock); 1618 ZONE_LOCK(zone); 1619 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1620 if (LIST_NEXT(z, uz_link) == NULL) { 1621 LIST_INSERT_AFTER(z, zone, uz_link); 1622 break; 1623 } 1624 } 1625 ZONE_UNLOCK(zone); 1626 rw_wunlock(&uma_rwlock); 1627 } else if (keg == NULL) { 1628 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1629 arg->align, arg->flags)) == NULL) 1630 return (ENOMEM); 1631 } else { 1632 struct uma_kctor_args karg; 1633 int error; 1634 1635 /* We should only be here from uma_startup() */ 1636 karg.size = arg->size; 1637 karg.uminit = arg->uminit; 1638 karg.fini = arg->fini; 1639 karg.align = arg->align; 1640 karg.flags = arg->flags; 1641 karg.zone = zone; 1642 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1643 flags); 1644 if (error) 1645 return (error); 1646 } 1647 1648 /* 1649 * Link in the first keg. 1650 */ 1651 zone->uz_klink.kl_keg = keg; 1652 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1653 zone->uz_lockptr = &keg->uk_lock; 1654 zone->uz_size = keg->uk_size; 1655 zone->uz_flags |= (keg->uk_flags & 1656 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1657 1658 /* 1659 * Some internal zones don't have room allocated for the per cpu 1660 * caches. If we're internal, bail out here. 1661 */ 1662 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1663 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1664 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1665 return (0); 1666 } 1667 1668out: 1669 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0) 1670 zone->uz_count = bucket_select(zone->uz_size); 1671 else 1672 zone->uz_count = BUCKET_MAX; 1673 zone->uz_count_min = zone->uz_count; 1674 1675 return (0); 1676} 1677 1678/* 1679 * Keg header dtor. This frees all data, destroys locks, frees the hash 1680 * table and removes the keg from the global list. 1681 * 1682 * Arguments/Returns follow uma_dtor specifications 1683 * udata unused 1684 */ 1685static void 1686keg_dtor(void *arg, int size, void *udata) 1687{ 1688 uma_keg_t keg; 1689 1690 keg = (uma_keg_t)arg; 1691 KEG_LOCK(keg); 1692 if (keg->uk_free != 0) { 1693 printf("Freed UMA keg (%s) was not empty (%d items). " 1694 " Lost %d pages of memory.\n", 1695 keg->uk_name ? keg->uk_name : "", 1696 keg->uk_free, keg->uk_pages); 1697 } 1698 KEG_UNLOCK(keg); 1699 1700 hash_free(&keg->uk_hash); 1701 1702 KEG_LOCK_FINI(keg); 1703} 1704 1705/* 1706 * Zone header dtor. 1707 * 1708 * Arguments/Returns follow uma_dtor specifications 1709 * udata unused 1710 */ 1711static void 1712zone_dtor(void *arg, int size, void *udata) 1713{ 1714 uma_klink_t klink; 1715 uma_zone_t zone; 1716 uma_keg_t keg; 1717 1718 zone = (uma_zone_t)arg; 1719 keg = zone_first_keg(zone); 1720 1721 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1722 cache_drain(zone); 1723 1724 rw_wlock(&uma_rwlock); 1725 LIST_REMOVE(zone, uz_link); 1726 rw_wunlock(&uma_rwlock); 1727 /* 1728 * XXX there are some races here where 1729 * the zone can be drained but zone lock 1730 * released and then refilled before we 1731 * remove it... we dont care for now 1732 */ 1733 zone_drain_wait(zone, M_WAITOK); 1734 /* 1735 * Unlink all of our kegs. 1736 */ 1737 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1738 klink->kl_keg = NULL; 1739 LIST_REMOVE(klink, kl_link); 1740 if (klink == &zone->uz_klink) 1741 continue; 1742 free(klink, M_TEMP); 1743 } 1744 /* 1745 * We only destroy kegs from non secondary zones. 1746 */ 1747 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1748 rw_wlock(&uma_rwlock); 1749 LIST_REMOVE(keg, uk_link); 1750 rw_wunlock(&uma_rwlock); 1751 zone_free_item(kegs, keg, NULL, SKIP_NONE); 1752 } 1753 ZONE_LOCK_FINI(zone); 1754} 1755 1756/* 1757 * Traverses every zone in the system and calls a callback 1758 * 1759 * Arguments: 1760 * zfunc A pointer to a function which accepts a zone 1761 * as an argument. 1762 * 1763 * Returns: 1764 * Nothing 1765 */ 1766static void 1767zone_foreach(void (*zfunc)(uma_zone_t)) 1768{ 1769 uma_keg_t keg; 1770 uma_zone_t zone; 1771 1772 rw_rlock(&uma_rwlock); 1773 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1774 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1775 zfunc(zone); 1776 } 1777 rw_runlock(&uma_rwlock); 1778} 1779 1780/* Public functions */ 1781/* See uma.h */ 1782void 1783uma_startup(void *bootmem, int boot_pages) 1784{ 1785 struct uma_zctor_args args; 1786 uma_slab_t slab; 1787 u_int slabsize; 1788 int i; 1789 1790#ifdef UMA_DEBUG 1791 printf("Creating uma keg headers zone and keg.\n"); 1792#endif 1793 rw_init(&uma_rwlock, "UMA lock"); 1794 1795 /* "manually" create the initial zone */ 1796 memset(&args, 0, sizeof(args)); 1797 args.name = "UMA Kegs"; 1798 args.size = sizeof(struct uma_keg); 1799 args.ctor = keg_ctor; 1800 args.dtor = keg_dtor; 1801 args.uminit = zero_init; 1802 args.fini = NULL; 1803 args.keg = &masterkeg; 1804 args.align = 32 - 1; 1805 args.flags = UMA_ZFLAG_INTERNAL; 1806 /* The initial zone has no Per cpu queues so it's smaller */ 1807 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); 1808 1809#ifdef UMA_DEBUG 1810 printf("Filling boot free list.\n"); 1811#endif 1812 for (i = 0; i < boot_pages; i++) { 1813 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE)); 1814 slab->us_data = (uint8_t *)slab; 1815 slab->us_flags = UMA_SLAB_BOOT; 1816 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); 1817 } 1818 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); 1819 1820#ifdef UMA_DEBUG 1821 printf("Creating uma zone headers zone and keg.\n"); 1822#endif 1823 args.name = "UMA Zones"; 1824 args.size = sizeof(struct uma_zone) + 1825 (sizeof(struct uma_cache) * (mp_maxid + 1)); 1826 args.ctor = zone_ctor; 1827 args.dtor = zone_dtor; 1828 args.uminit = zero_init; 1829 args.fini = NULL; 1830 args.keg = NULL; 1831 args.align = 32 - 1; 1832 args.flags = UMA_ZFLAG_INTERNAL; 1833 /* The initial zone has no Per cpu queues so it's smaller */ 1834 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); 1835 1836#ifdef UMA_DEBUG 1837 printf("Initializing pcpu cache locks.\n"); 1838#endif 1839#ifdef UMA_DEBUG 1840 printf("Creating slab and hash zones.\n"); 1841#endif 1842 1843 /* Now make a zone for slab headers */ 1844 slabzone = uma_zcreate("UMA Slabs", 1845 sizeof(struct uma_slab), 1846 NULL, NULL, NULL, NULL, 1847 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1848 1849 /* 1850 * We also create a zone for the bigger slabs with reference 1851 * counts in them, to accomodate UMA_ZONE_REFCNT zones. 1852 */ 1853 slabsize = sizeof(struct uma_slab_refcnt); 1854 slabsize += uma_max_ipers_ref * sizeof(uint32_t); 1855 slabrefzone = uma_zcreate("UMA RCntSlabs", 1856 slabsize, 1857 NULL, NULL, NULL, NULL, 1858 UMA_ALIGN_PTR, 1859 UMA_ZFLAG_INTERNAL); 1860 1861 hashzone = uma_zcreate("UMA Hash", 1862 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 1863 NULL, NULL, NULL, NULL, 1864 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1865 1866 bucket_init(); 1867 1868 booted = UMA_STARTUP; 1869 1870#ifdef UMA_DEBUG 1871 printf("UMA startup complete.\n"); 1872#endif 1873} 1874 1875/* see uma.h */ 1876void 1877uma_startup2(void) 1878{ 1879 booted = UMA_STARTUP2; 1880 bucket_enable(); 1881 sx_init(&uma_drain_lock, "umadrain"); 1882#ifdef UMA_DEBUG 1883 printf("UMA startup2 complete.\n"); 1884#endif 1885} 1886 1887/* 1888 * Initialize our callout handle 1889 * 1890 */ 1891 1892static void 1893uma_startup3(void) 1894{ 1895#ifdef UMA_DEBUG 1896 printf("Starting callout.\n"); 1897#endif 1898 callout_init(&uma_callout, CALLOUT_MPSAFE); 1899 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1900#ifdef UMA_DEBUG 1901 printf("UMA startup3 complete.\n"); 1902#endif 1903} 1904 1905static uma_keg_t 1906uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 1907 int align, uint32_t flags) 1908{ 1909 struct uma_kctor_args args; 1910 1911 args.size = size; 1912 args.uminit = uminit; 1913 args.fini = fini; 1914 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 1915 args.flags = flags; 1916 args.zone = zone; 1917 return (zone_alloc_item(kegs, &args, M_WAITOK)); 1918} 1919 1920/* See uma.h */ 1921void 1922uma_set_align(int align) 1923{ 1924 1925 if (align != UMA_ALIGN_CACHE) 1926 uma_align_cache = align; 1927} 1928 1929/* See uma.h */ 1930uma_zone_t 1931uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 1932 uma_init uminit, uma_fini fini, int align, uint32_t flags) 1933 1934{ 1935 struct uma_zctor_args args; 1936 uma_zone_t res; 1937 bool locked; 1938 1939 /* This stuff is essential for the zone ctor */ 1940 memset(&args, 0, sizeof(args)); 1941 args.name = name; 1942 args.size = size; 1943 args.ctor = ctor; 1944 args.dtor = dtor; 1945 args.uminit = uminit; 1946 args.fini = fini; 1947 args.align = align; 1948 args.flags = flags; 1949 args.keg = NULL; 1950 1951 if (booted < UMA_STARTUP2) { 1952 locked = false; 1953 } else { 1954 sx_slock(&uma_drain_lock); 1955 locked = true; 1956 } 1957 res = zone_alloc_item(zones, &args, M_WAITOK); 1958 if (locked) 1959 sx_sunlock(&uma_drain_lock); 1960 return (res); 1961} 1962 1963/* See uma.h */ 1964uma_zone_t 1965uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 1966 uma_init zinit, uma_fini zfini, uma_zone_t master) 1967{ 1968 struct uma_zctor_args args; 1969 uma_keg_t keg; 1970 uma_zone_t res; 1971 bool locked; 1972 1973 keg = zone_first_keg(master); 1974 memset(&args, 0, sizeof(args)); 1975 args.name = name; 1976 args.size = keg->uk_size; 1977 args.ctor = ctor; 1978 args.dtor = dtor; 1979 args.uminit = zinit; 1980 args.fini = zfini; 1981 args.align = keg->uk_align; 1982 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 1983 args.keg = keg; 1984 1985 if (booted < UMA_STARTUP2) { 1986 locked = false; 1987 } else { 1988 sx_slock(&uma_drain_lock); 1989 locked = true; 1990 } 1991 /* XXX Attaches only one keg of potentially many. */ 1992 res = zone_alloc_item(zones, &args, M_WAITOK); 1993 if (locked) 1994 sx_sunlock(&uma_drain_lock); 1995 return (res); 1996} 1997 1998/* See uma.h */ 1999uma_zone_t 2000uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 2001 uma_init zinit, uma_fini zfini, uma_import zimport, 2002 uma_release zrelease, void *arg, int flags) 2003{ 2004 struct uma_zctor_args args; 2005 2006 memset(&args, 0, sizeof(args)); 2007 args.name = name; 2008 args.size = size; 2009 args.ctor = ctor; 2010 args.dtor = dtor; 2011 args.uminit = zinit; 2012 args.fini = zfini; 2013 args.import = zimport; 2014 args.release = zrelease; 2015 args.arg = arg; 2016 args.align = 0; 2017 args.flags = flags; 2018 2019 return (zone_alloc_item(zones, &args, M_WAITOK)); 2020} 2021 2022static void 2023zone_lock_pair(uma_zone_t a, uma_zone_t b) 2024{ 2025 if (a < b) { 2026 ZONE_LOCK(a); 2027 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); 2028 } else { 2029 ZONE_LOCK(b); 2030 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); 2031 } 2032} 2033 2034static void 2035zone_unlock_pair(uma_zone_t a, uma_zone_t b) 2036{ 2037 2038 ZONE_UNLOCK(a); 2039 ZONE_UNLOCK(b); 2040} 2041 2042int 2043uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 2044{ 2045 uma_klink_t klink; 2046 uma_klink_t kl; 2047 int error; 2048 2049 error = 0; 2050 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 2051 2052 zone_lock_pair(zone, master); 2053 /* 2054 * zone must use vtoslab() to resolve objects and must already be 2055 * a secondary. 2056 */ 2057 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 2058 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 2059 error = EINVAL; 2060 goto out; 2061 } 2062 /* 2063 * The new master must also use vtoslab(). 2064 */ 2065 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 2066 error = EINVAL; 2067 goto out; 2068 } 2069 /* 2070 * Both must either be refcnt, or not be refcnt. 2071 */ 2072 if ((zone->uz_flags & UMA_ZONE_REFCNT) != 2073 (master->uz_flags & UMA_ZONE_REFCNT)) { 2074 error = EINVAL; 2075 goto out; 2076 } 2077 /* 2078 * The underlying object must be the same size. rsize 2079 * may be different. 2080 */ 2081 if (master->uz_size != zone->uz_size) { 2082 error = E2BIG; 2083 goto out; 2084 } 2085 /* 2086 * Put it at the end of the list. 2087 */ 2088 klink->kl_keg = zone_first_keg(master); 2089 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 2090 if (LIST_NEXT(kl, kl_link) == NULL) { 2091 LIST_INSERT_AFTER(kl, klink, kl_link); 2092 break; 2093 } 2094 } 2095 klink = NULL; 2096 zone->uz_flags |= UMA_ZFLAG_MULTI; 2097 zone->uz_slab = zone_fetch_slab_multi; 2098 2099out: 2100 zone_unlock_pair(zone, master); 2101 if (klink != NULL) 2102 free(klink, M_TEMP); 2103 2104 return (error); 2105} 2106 2107 2108/* See uma.h */ 2109void 2110uma_zdestroy(uma_zone_t zone) 2111{ 2112 2113 sx_slock(&uma_drain_lock); 2114 zone_free_item(zones, zone, NULL, SKIP_NONE); 2115 sx_sunlock(&uma_drain_lock); 2116} 2117 2118/* See uma.h */ 2119void * 2120uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2121{ 2122 void *item; 2123 uma_cache_t cache; 2124 uma_bucket_t bucket; 2125 int lockfail; 2126 int cpu; 2127 2128 /* This is the fast path allocation */ 2129#ifdef UMA_DEBUG_ALLOC_1 2130 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); 2131#endif 2132 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread, 2133 zone->uz_name, flags); 2134 2135 if (flags & M_WAITOK) { 2136 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2137 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2138 } 2139#ifdef DEBUG_MEMGUARD 2140 if (memguard_cmp_zone(zone)) { 2141 item = memguard_alloc(zone->uz_size, flags); 2142 if (item != NULL) { 2143 /* 2144 * Avoid conflict with the use-after-free 2145 * protecting infrastructure from INVARIANTS. 2146 */ 2147 if (zone->uz_init != NULL && 2148 zone->uz_init != mtrash_init && 2149 zone->uz_init(item, zone->uz_size, flags) != 0) 2150 return (NULL); 2151 if (zone->uz_ctor != NULL && 2152 zone->uz_ctor != mtrash_ctor && 2153 zone->uz_ctor(item, zone->uz_size, udata, 2154 flags) != 0) { 2155 zone->uz_fini(item, zone->uz_size); 2156 return (NULL); 2157 } 2158 return (item); 2159 } 2160 /* This is unfortunate but should not be fatal. */ 2161 } 2162#endif 2163 /* 2164 * If possible, allocate from the per-CPU cache. There are two 2165 * requirements for safe access to the per-CPU cache: (1) the thread 2166 * accessing the cache must not be preempted or yield during access, 2167 * and (2) the thread must not migrate CPUs without switching which 2168 * cache it accesses. We rely on a critical section to prevent 2169 * preemption and migration. We release the critical section in 2170 * order to acquire the zone mutex if we are unable to allocate from 2171 * the current cache; when we re-acquire the critical section, we 2172 * must detect and handle migration if it has occurred. 2173 */ 2174 critical_enter(); 2175 cpu = curcpu; 2176 cache = &zone->uz_cpu[cpu]; 2177 2178zalloc_start: 2179 bucket = cache->uc_allocbucket; 2180 if (bucket != NULL && bucket->ub_cnt > 0) { 2181 bucket->ub_cnt--; 2182 item = bucket->ub_bucket[bucket->ub_cnt]; 2183#ifdef INVARIANTS 2184 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2185#endif 2186 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2187 cache->uc_allocs++; 2188 critical_exit(); 2189 if (zone->uz_ctor != NULL && 2190 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2191 atomic_add_long(&zone->uz_fails, 1); 2192 zone_free_item(zone, item, udata, SKIP_DTOR); 2193 return (NULL); 2194 } 2195#ifdef INVARIANTS 2196 uma_dbg_alloc(zone, NULL, item); 2197#endif 2198 if (flags & M_ZERO) 2199 uma_zero_item(item, zone); 2200 return (item); 2201 } 2202 2203 /* 2204 * We have run out of items in our alloc bucket. 2205 * See if we can switch with our free bucket. 2206 */ 2207 bucket = cache->uc_freebucket; 2208 if (bucket != NULL && bucket->ub_cnt > 0) { 2209#ifdef UMA_DEBUG_ALLOC 2210 printf("uma_zalloc: Swapping empty with alloc.\n"); 2211#endif 2212 cache->uc_freebucket = cache->uc_allocbucket; 2213 cache->uc_allocbucket = bucket; 2214 goto zalloc_start; 2215 } 2216 2217 /* 2218 * Discard any empty allocation bucket while we hold no locks. 2219 */ 2220 bucket = cache->uc_allocbucket; 2221 cache->uc_allocbucket = NULL; 2222 critical_exit(); 2223 if (bucket != NULL) 2224 bucket_free(zone, bucket, udata); 2225 2226 /* Short-circuit for zones without buckets and low memory. */ 2227 if (zone->uz_count == 0 || bucketdisable) 2228 goto zalloc_item; 2229 2230 /* 2231 * Attempt to retrieve the item from the per-CPU cache has failed, so 2232 * we must go back to the zone. This requires the zone lock, so we 2233 * must drop the critical section, then re-acquire it when we go back 2234 * to the cache. Since the critical section is released, we may be 2235 * preempted or migrate. As such, make sure not to maintain any 2236 * thread-local state specific to the cache from prior to releasing 2237 * the critical section. 2238 */ 2239 lockfail = 0; 2240 if (ZONE_TRYLOCK(zone) == 0) { 2241 /* Record contention to size the buckets. */ 2242 ZONE_LOCK(zone); 2243 lockfail = 1; 2244 } 2245 critical_enter(); 2246 cpu = curcpu; 2247 cache = &zone->uz_cpu[cpu]; 2248 2249 /* 2250 * Since we have locked the zone we may as well send back our stats. 2251 */ 2252 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2253 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2254 cache->uc_allocs = 0; 2255 cache->uc_frees = 0; 2256 2257 /* See if we lost the race to fill the cache. */ 2258 if (cache->uc_allocbucket != NULL) { 2259 ZONE_UNLOCK(zone); 2260 goto zalloc_start; 2261 } 2262 2263 /* 2264 * Check the zone's cache of buckets. 2265 */ 2266 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 2267 KASSERT(bucket->ub_cnt != 0, 2268 ("uma_zalloc_arg: Returning an empty bucket.")); 2269 2270 LIST_REMOVE(bucket, ub_link); 2271 cache->uc_allocbucket = bucket; 2272 ZONE_UNLOCK(zone); 2273 goto zalloc_start; 2274 } 2275 /* We are no longer associated with this CPU. */ 2276 critical_exit(); 2277 2278 /* 2279 * We bump the uz count when the cache size is insufficient to 2280 * handle the working set. 2281 */ 2282 if (lockfail && zone->uz_count < BUCKET_MAX) 2283 zone->uz_count++; 2284 ZONE_UNLOCK(zone); 2285 2286 /* 2287 * Now lets just fill a bucket and put it on the free list. If that 2288 * works we'll restart the allocation from the begining and it 2289 * will use the just filled bucket. 2290 */ 2291 bucket = zone_alloc_bucket(zone, udata, flags); 2292 if (bucket != NULL) { 2293 ZONE_LOCK(zone); 2294 critical_enter(); 2295 cpu = curcpu; 2296 cache = &zone->uz_cpu[cpu]; 2297 /* 2298 * See if we lost the race or were migrated. Cache the 2299 * initialized bucket to make this less likely or claim 2300 * the memory directly. 2301 */ 2302 if (cache->uc_allocbucket == NULL) 2303 cache->uc_allocbucket = bucket; 2304 else 2305 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2306 ZONE_UNLOCK(zone); 2307 goto zalloc_start; 2308 } 2309 2310 /* 2311 * We may not be able to get a bucket so return an actual item. 2312 */ 2313#ifdef UMA_DEBUG 2314 printf("uma_zalloc_arg: Bucketzone returned NULL\n"); 2315#endif 2316 2317zalloc_item: 2318 item = zone_alloc_item(zone, udata, flags); 2319 2320 return (item); 2321} 2322 2323static uma_slab_t 2324keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) 2325{ 2326 uma_slab_t slab; 2327 int reserve; 2328 2329 mtx_assert(&keg->uk_lock, MA_OWNED); 2330 slab = NULL; 2331 reserve = 0; 2332 if ((flags & M_USE_RESERVE) == 0) 2333 reserve = keg->uk_reserve; 2334 2335 for (;;) { 2336 /* 2337 * Find a slab with some space. Prefer slabs that are partially 2338 * used over those that are totally full. This helps to reduce 2339 * fragmentation. 2340 */ 2341 if (keg->uk_free > reserve) { 2342 if (!LIST_EMPTY(&keg->uk_part_slab)) { 2343 slab = LIST_FIRST(&keg->uk_part_slab); 2344 } else { 2345 slab = LIST_FIRST(&keg->uk_free_slab); 2346 LIST_REMOVE(slab, us_link); 2347 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, 2348 us_link); 2349 } 2350 MPASS(slab->us_keg == keg); 2351 return (slab); 2352 } 2353 2354 /* 2355 * M_NOVM means don't ask at all! 2356 */ 2357 if (flags & M_NOVM) 2358 break; 2359 2360 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2361 keg->uk_flags |= UMA_ZFLAG_FULL; 2362 /* 2363 * If this is not a multi-zone, set the FULL bit. 2364 * Otherwise slab_multi() takes care of it. 2365 */ 2366 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2367 zone->uz_flags |= UMA_ZFLAG_FULL; 2368 zone_log_warning(zone); 2369 } 2370 if (flags & M_NOWAIT) 2371 break; 2372 zone->uz_sleeps++; 2373 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2374 continue; 2375 } 2376 slab = keg_alloc_slab(keg, zone, flags); 2377 /* 2378 * If we got a slab here it's safe to mark it partially used 2379 * and return. We assume that the caller is going to remove 2380 * at least one item. 2381 */ 2382 if (slab) { 2383 MPASS(slab->us_keg == keg); 2384 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2385 return (slab); 2386 } 2387 /* 2388 * We might not have been able to get a slab but another cpu 2389 * could have while we were unlocked. Check again before we 2390 * fail. 2391 */ 2392 flags |= M_NOVM; 2393 } 2394 return (slab); 2395} 2396 2397static uma_slab_t 2398zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) 2399{ 2400 uma_slab_t slab; 2401 2402 if (keg == NULL) { 2403 keg = zone_first_keg(zone); 2404 KEG_LOCK(keg); 2405 } 2406 2407 for (;;) { 2408 slab = keg_fetch_slab(keg, zone, flags); 2409 if (slab) 2410 return (slab); 2411 if (flags & (M_NOWAIT | M_NOVM)) 2412 break; 2413 } 2414 KEG_UNLOCK(keg); 2415 return (NULL); 2416} 2417 2418/* 2419 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2420 * with the keg locked. On NULL no lock is held. 2421 * 2422 * The last pointer is used to seed the search. It is not required. 2423 */ 2424static uma_slab_t 2425zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) 2426{ 2427 uma_klink_t klink; 2428 uma_slab_t slab; 2429 uma_keg_t keg; 2430 int flags; 2431 int empty; 2432 int full; 2433 2434 /* 2435 * Don't wait on the first pass. This will skip limit tests 2436 * as well. We don't want to block if we can find a provider 2437 * without blocking. 2438 */ 2439 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2440 /* 2441 * Use the last slab allocated as a hint for where to start 2442 * the search. 2443 */ 2444 if (last != NULL) { 2445 slab = keg_fetch_slab(last, zone, flags); 2446 if (slab) 2447 return (slab); 2448 KEG_UNLOCK(last); 2449 } 2450 /* 2451 * Loop until we have a slab incase of transient failures 2452 * while M_WAITOK is specified. I'm not sure this is 100% 2453 * required but we've done it for so long now. 2454 */ 2455 for (;;) { 2456 empty = 0; 2457 full = 0; 2458 /* 2459 * Search the available kegs for slabs. Be careful to hold the 2460 * correct lock while calling into the keg layer. 2461 */ 2462 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2463 keg = klink->kl_keg; 2464 KEG_LOCK(keg); 2465 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2466 slab = keg_fetch_slab(keg, zone, flags); 2467 if (slab) 2468 return (slab); 2469 } 2470 if (keg->uk_flags & UMA_ZFLAG_FULL) 2471 full++; 2472 else 2473 empty++; 2474 KEG_UNLOCK(keg); 2475 } 2476 if (rflags & (M_NOWAIT | M_NOVM)) 2477 break; 2478 flags = rflags; 2479 /* 2480 * All kegs are full. XXX We can't atomically check all kegs 2481 * and sleep so just sleep for a short period and retry. 2482 */ 2483 if (full && !empty) { 2484 ZONE_LOCK(zone); 2485 zone->uz_flags |= UMA_ZFLAG_FULL; 2486 zone->uz_sleeps++; 2487 zone_log_warning(zone); 2488 msleep(zone, zone->uz_lockptr, PVM, 2489 "zonelimit", hz/100); 2490 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2491 ZONE_UNLOCK(zone); 2492 continue; 2493 } 2494 } 2495 return (NULL); 2496} 2497 2498static void * 2499slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 2500{ 2501 void *item; 2502 uint8_t freei; 2503 2504 MPASS(keg == slab->us_keg); 2505 mtx_assert(&keg->uk_lock, MA_OWNED); 2506 2507 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; 2508 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); 2509 item = slab->us_data + (keg->uk_rsize * freei); 2510 slab->us_freecount--; 2511 keg->uk_free--; 2512 2513 /* Move this slab to the full list */ 2514 if (slab->us_freecount == 0) { 2515 LIST_REMOVE(slab, us_link); 2516 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); 2517 } 2518 2519 return (item); 2520} 2521 2522static int 2523zone_import(uma_zone_t zone, void **bucket, int max, int flags) 2524{ 2525 uma_slab_t slab; 2526 uma_keg_t keg; 2527 int i; 2528 2529 slab = NULL; 2530 keg = NULL; 2531 /* Try to keep the buckets totally full */ 2532 for (i = 0; i < max; ) { 2533 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL) 2534 break; 2535 keg = slab->us_keg; 2536 while (slab->us_freecount && i < max) { 2537 bucket[i++] = slab_alloc_item(keg, slab); 2538 if (keg->uk_free <= keg->uk_reserve) 2539 break; 2540 } 2541 /* Don't grab more than one slab at a time. */ 2542 flags &= ~M_WAITOK; 2543 flags |= M_NOWAIT; 2544 } 2545 if (slab != NULL) 2546 KEG_UNLOCK(keg); 2547 2548 return i; 2549} 2550 2551static uma_bucket_t 2552zone_alloc_bucket(uma_zone_t zone, void *udata, int flags) 2553{ 2554 uma_bucket_t bucket; 2555 int max; 2556 2557 /* Don't wait for buckets, preserve caller's NOVM setting. */ 2558 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 2559 if (bucket == NULL) 2560 return (NULL); 2561 2562 max = MIN(bucket->ub_entries, zone->uz_count); 2563 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 2564 max, flags); 2565 2566 /* 2567 * Initialize the memory if necessary. 2568 */ 2569 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 2570 int i; 2571 2572 for (i = 0; i < bucket->ub_cnt; i++) 2573 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2574 flags) != 0) 2575 break; 2576 /* 2577 * If we couldn't initialize the whole bucket, put the 2578 * rest back onto the freelist. 2579 */ 2580 if (i != bucket->ub_cnt) { 2581 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 2582 bucket->ub_cnt - i); 2583#ifdef INVARIANTS 2584 bzero(&bucket->ub_bucket[i], 2585 sizeof(void *) * (bucket->ub_cnt - i)); 2586#endif 2587 bucket->ub_cnt = i; 2588 } 2589 } 2590 2591 if (bucket->ub_cnt == 0) { 2592 bucket_free(zone, bucket, udata); 2593 atomic_add_long(&zone->uz_fails, 1); 2594 return (NULL); 2595 } 2596 2597 return (bucket); 2598} 2599 2600/* 2601 * Allocates a single item from a zone. 2602 * 2603 * Arguments 2604 * zone The zone to alloc for. 2605 * udata The data to be passed to the constructor. 2606 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2607 * 2608 * Returns 2609 * NULL if there is no memory and M_NOWAIT is set 2610 * An item if successful 2611 */ 2612 2613static void * 2614zone_alloc_item(uma_zone_t zone, void *udata, int flags) 2615{ 2616 void *item; 2617 2618 item = NULL; 2619 2620#ifdef UMA_DEBUG_ALLOC 2621 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); 2622#endif 2623 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1) 2624 goto fail; 2625 atomic_add_long(&zone->uz_allocs, 1); 2626 2627 /* 2628 * We have to call both the zone's init (not the keg's init) 2629 * and the zone's ctor. This is because the item is going from 2630 * a keg slab directly to the user, and the user is expecting it 2631 * to be both zone-init'd as well as zone-ctor'd. 2632 */ 2633 if (zone->uz_init != NULL) { 2634 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 2635 zone_free_item(zone, item, udata, SKIP_FINI); 2636 goto fail; 2637 } 2638 } 2639 if (zone->uz_ctor != NULL) { 2640 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2641 zone_free_item(zone, item, udata, SKIP_DTOR); 2642 goto fail; 2643 } 2644 } 2645#ifdef INVARIANTS 2646 uma_dbg_alloc(zone, NULL, item); 2647#endif 2648 if (flags & M_ZERO) 2649 uma_zero_item(item, zone); 2650 2651 return (item); 2652 2653fail: 2654 atomic_add_long(&zone->uz_fails, 1); 2655 return (NULL); 2656} 2657 2658/* See uma.h */ 2659void 2660uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 2661{ 2662 uma_cache_t cache; 2663 uma_bucket_t bucket; 2664 int lockfail; 2665 int cpu; 2666 2667#ifdef UMA_DEBUG_ALLOC_1 2668 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); 2669#endif 2670 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 2671 zone->uz_name); 2672 2673 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 2674 if (item == NULL) 2675 return; 2676#ifdef DEBUG_MEMGUARD 2677 if (is_memguard_addr(item)) { 2678 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor) 2679 zone->uz_dtor(item, zone->uz_size, udata); 2680 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini) 2681 zone->uz_fini(item, zone->uz_size); 2682 memguard_free(item); 2683 return; 2684 } 2685#endif 2686#ifdef INVARIANTS 2687 if (zone->uz_flags & UMA_ZONE_MALLOC) 2688 uma_dbg_free(zone, udata, item); 2689 else 2690 uma_dbg_free(zone, NULL, item); 2691#endif 2692 if (zone->uz_dtor != NULL) 2693 zone->uz_dtor(item, zone->uz_size, udata); 2694 2695 /* 2696 * The race here is acceptable. If we miss it we'll just have to wait 2697 * a little longer for the limits to be reset. 2698 */ 2699 if (zone->uz_flags & UMA_ZFLAG_FULL) 2700 goto zfree_item; 2701 2702 /* 2703 * If possible, free to the per-CPU cache. There are two 2704 * requirements for safe access to the per-CPU cache: (1) the thread 2705 * accessing the cache must not be preempted or yield during access, 2706 * and (2) the thread must not migrate CPUs without switching which 2707 * cache it accesses. We rely on a critical section to prevent 2708 * preemption and migration. We release the critical section in 2709 * order to acquire the zone mutex if we are unable to free to the 2710 * current cache; when we re-acquire the critical section, we must 2711 * detect and handle migration if it has occurred. 2712 */ 2713zfree_restart: 2714 critical_enter(); 2715 cpu = curcpu; 2716 cache = &zone->uz_cpu[cpu]; 2717 2718zfree_start: 2719 /* 2720 * Try to free into the allocbucket first to give LIFO ordering 2721 * for cache-hot datastructures. Spill over into the freebucket 2722 * if necessary. Alloc will swap them if one runs dry. 2723 */ 2724 bucket = cache->uc_allocbucket; 2725 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 2726 bucket = cache->uc_freebucket; 2727 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2728 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 2729 ("uma_zfree: Freeing to non free bucket index.")); 2730 bucket->ub_bucket[bucket->ub_cnt] = item; 2731 bucket->ub_cnt++; 2732 cache->uc_frees++; 2733 critical_exit(); 2734 return; 2735 } 2736 2737 /* 2738 * We must go back the zone, which requires acquiring the zone lock, 2739 * which in turn means we must release and re-acquire the critical 2740 * section. Since the critical section is released, we may be 2741 * preempted or migrate. As such, make sure not to maintain any 2742 * thread-local state specific to the cache from prior to releasing 2743 * the critical section. 2744 */ 2745 critical_exit(); 2746 if (zone->uz_count == 0 || bucketdisable) 2747 goto zfree_item; 2748 2749 lockfail = 0; 2750 if (ZONE_TRYLOCK(zone) == 0) { 2751 /* Record contention to size the buckets. */ 2752 ZONE_LOCK(zone); 2753 lockfail = 1; 2754 } 2755 critical_enter(); 2756 cpu = curcpu; 2757 cache = &zone->uz_cpu[cpu]; 2758 2759 /* 2760 * Since we have locked the zone we may as well send back our stats. 2761 */ 2762 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2763 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2764 cache->uc_allocs = 0; 2765 cache->uc_frees = 0; 2766 2767 bucket = cache->uc_freebucket; 2768 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2769 ZONE_UNLOCK(zone); 2770 goto zfree_start; 2771 } 2772 cache->uc_freebucket = NULL; 2773 /* We are no longer associated with this CPU. */ 2774 critical_exit(); 2775 2776 /* Can we throw this on the zone full list? */ 2777 if (bucket != NULL) { 2778#ifdef UMA_DEBUG_ALLOC 2779 printf("uma_zfree: Putting old bucket on the free list.\n"); 2780#endif 2781 /* ub_cnt is pointing to the last free item */ 2782 KASSERT(bucket->ub_cnt != 0, 2783 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 2784 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2785 } 2786 2787 /* 2788 * We bump the uz count when the cache size is insufficient to 2789 * handle the working set. 2790 */ 2791 if (lockfail && zone->uz_count < BUCKET_MAX) 2792 zone->uz_count++; 2793 ZONE_UNLOCK(zone); 2794 2795#ifdef UMA_DEBUG_ALLOC 2796 printf("uma_zfree: Allocating new free bucket.\n"); 2797#endif 2798 bucket = bucket_alloc(zone, udata, M_NOWAIT); 2799 if (bucket) { 2800 critical_enter(); 2801 cpu = curcpu; 2802 cache = &zone->uz_cpu[cpu]; 2803 if (cache->uc_freebucket == NULL) { 2804 cache->uc_freebucket = bucket; 2805 goto zfree_start; 2806 } 2807 /* 2808 * We lost the race, start over. We have to drop our 2809 * critical section to free the bucket. 2810 */ 2811 critical_exit(); 2812 bucket_free(zone, bucket, udata); 2813 goto zfree_restart; 2814 } 2815 2816 /* 2817 * If nothing else caught this, we'll just do an internal free. 2818 */ 2819zfree_item: 2820 zone_free_item(zone, item, udata, SKIP_DTOR); 2821 2822 return; 2823} 2824 2825static void 2826slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) 2827{ 2828 uint8_t freei; 2829 2830 mtx_assert(&keg->uk_lock, MA_OWNED); 2831 MPASS(keg == slab->us_keg); 2832 2833 /* Do we need to remove from any lists? */ 2834 if (slab->us_freecount+1 == keg->uk_ipers) { 2835 LIST_REMOVE(slab, us_link); 2836 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 2837 } else if (slab->us_freecount == 0) { 2838 LIST_REMOVE(slab, us_link); 2839 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2840 } 2841 2842 /* Slab management. */ 2843 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 2844 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); 2845 slab->us_freecount++; 2846 2847 /* Keg statistics. */ 2848 keg->uk_free++; 2849} 2850 2851static void 2852zone_release(uma_zone_t zone, void **bucket, int cnt) 2853{ 2854 void *item; 2855 uma_slab_t slab; 2856 uma_keg_t keg; 2857 uint8_t *mem; 2858 int clearfull; 2859 int i; 2860 2861 clearfull = 0; 2862 keg = zone_first_keg(zone); 2863 KEG_LOCK(keg); 2864 for (i = 0; i < cnt; i++) { 2865 item = bucket[i]; 2866 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 2867 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 2868 if (zone->uz_flags & UMA_ZONE_HASH) { 2869 slab = hash_sfind(&keg->uk_hash, mem); 2870 } else { 2871 mem += keg->uk_pgoff; 2872 slab = (uma_slab_t)mem; 2873 } 2874 } else { 2875 slab = vtoslab((vm_offset_t)item); 2876 if (slab->us_keg != keg) { 2877 KEG_UNLOCK(keg); 2878 keg = slab->us_keg; 2879 KEG_LOCK(keg); 2880 } 2881 } 2882 slab_free_item(keg, slab, item); 2883 if (keg->uk_flags & UMA_ZFLAG_FULL) { 2884 if (keg->uk_pages < keg->uk_maxpages) { 2885 keg->uk_flags &= ~UMA_ZFLAG_FULL; 2886 clearfull = 1; 2887 } 2888 2889 /* 2890 * We can handle one more allocation. Since we're 2891 * clearing ZFLAG_FULL, wake up all procs blocked 2892 * on pages. This should be uncommon, so keeping this 2893 * simple for now (rather than adding count of blocked 2894 * threads etc). 2895 */ 2896 wakeup(keg); 2897 } 2898 } 2899 KEG_UNLOCK(keg); 2900 if (clearfull) { 2901 ZONE_LOCK(zone); 2902 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2903 wakeup(zone); 2904 ZONE_UNLOCK(zone); 2905 } 2906 2907} 2908 2909/* 2910 * Frees a single item to any zone. 2911 * 2912 * Arguments: 2913 * zone The zone to free to 2914 * item The item we're freeing 2915 * udata User supplied data for the dtor 2916 * skip Skip dtors and finis 2917 */ 2918static void 2919zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 2920{ 2921 2922#ifdef INVARIANTS 2923 if (skip == SKIP_NONE) { 2924 if (zone->uz_flags & UMA_ZONE_MALLOC) 2925 uma_dbg_free(zone, udata, item); 2926 else 2927 uma_dbg_free(zone, NULL, item); 2928 } 2929#endif 2930 if (skip < SKIP_DTOR && zone->uz_dtor) 2931 zone->uz_dtor(item, zone->uz_size, udata); 2932 2933 if (skip < SKIP_FINI && zone->uz_fini) 2934 zone->uz_fini(item, zone->uz_size); 2935 2936 atomic_add_long(&zone->uz_frees, 1); 2937 zone->uz_release(zone->uz_arg, &item, 1); 2938} 2939 2940/* See uma.h */ 2941int 2942uma_zone_set_max(uma_zone_t zone, int nitems) 2943{ 2944 uma_keg_t keg; 2945 2946 keg = zone_first_keg(zone); 2947 if (keg == NULL) 2948 return (0); 2949 KEG_LOCK(keg); 2950 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 2951 if (keg->uk_maxpages * keg->uk_ipers < nitems) 2952 keg->uk_maxpages += keg->uk_ppera; 2953 nitems = keg->uk_maxpages * keg->uk_ipers; 2954 KEG_UNLOCK(keg); 2955 2956 return (nitems); 2957} 2958 2959/* See uma.h */ 2960int 2961uma_zone_get_max(uma_zone_t zone) 2962{ 2963 int nitems; 2964 uma_keg_t keg; 2965 2966 keg = zone_first_keg(zone); 2967 if (keg == NULL) 2968 return (0); 2969 KEG_LOCK(keg); 2970 nitems = keg->uk_maxpages * keg->uk_ipers; 2971 KEG_UNLOCK(keg); 2972 2973 return (nitems); 2974} 2975 2976/* See uma.h */ 2977void 2978uma_zone_set_warning(uma_zone_t zone, const char *warning) 2979{ 2980 2981 ZONE_LOCK(zone); 2982 zone->uz_warning = warning; 2983 ZONE_UNLOCK(zone); 2984} 2985 2986/* See uma.h */ 2987int 2988uma_zone_get_cur(uma_zone_t zone) 2989{ 2990 int64_t nitems; 2991 u_int i; 2992 2993 ZONE_LOCK(zone); 2994 nitems = zone->uz_allocs - zone->uz_frees; 2995 CPU_FOREACH(i) { 2996 /* 2997 * See the comment in sysctl_vm_zone_stats() regarding the 2998 * safety of accessing the per-cpu caches. With the zone lock 2999 * held, it is safe, but can potentially result in stale data. 3000 */ 3001 nitems += zone->uz_cpu[i].uc_allocs - 3002 zone->uz_cpu[i].uc_frees; 3003 } 3004 ZONE_UNLOCK(zone); 3005 3006 return (nitems < 0 ? 0 : nitems); 3007} 3008 3009/* See uma.h */ 3010void 3011uma_zone_set_init(uma_zone_t zone, uma_init uminit) 3012{ 3013 uma_keg_t keg; 3014 3015 keg = zone_first_keg(zone); 3016 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 3017 KEG_LOCK(keg); 3018 KASSERT(keg->uk_pages == 0, 3019 ("uma_zone_set_init on non-empty keg")); 3020 keg->uk_init = uminit; 3021 KEG_UNLOCK(keg); 3022} 3023 3024/* See uma.h */ 3025void 3026uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 3027{ 3028 uma_keg_t keg; 3029 3030 keg = zone_first_keg(zone); 3031 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type")); 3032 KEG_LOCK(keg); 3033 KASSERT(keg->uk_pages == 0, 3034 ("uma_zone_set_fini on non-empty keg")); 3035 keg->uk_fini = fini; 3036 KEG_UNLOCK(keg); 3037} 3038 3039/* See uma.h */ 3040void 3041uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 3042{ 3043 3044 ZONE_LOCK(zone); 3045 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3046 ("uma_zone_set_zinit on non-empty keg")); 3047 zone->uz_init = zinit; 3048 ZONE_UNLOCK(zone); 3049} 3050 3051/* See uma.h */ 3052void 3053uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 3054{ 3055 3056 ZONE_LOCK(zone); 3057 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3058 ("uma_zone_set_zfini on non-empty keg")); 3059 zone->uz_fini = zfini; 3060 ZONE_UNLOCK(zone); 3061} 3062 3063/* See uma.h */ 3064/* XXX uk_freef is not actually used with the zone locked */ 3065void 3066uma_zone_set_freef(uma_zone_t zone, uma_free freef) 3067{ 3068 uma_keg_t keg; 3069 3070 keg = zone_first_keg(zone); 3071 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type")); 3072 KEG_LOCK(keg); 3073 keg->uk_freef = freef; 3074 KEG_UNLOCK(keg); 3075} 3076 3077/* See uma.h */ 3078/* XXX uk_allocf is not actually used with the zone locked */ 3079void 3080uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3081{ 3082 uma_keg_t keg; 3083 3084 keg = zone_first_keg(zone); 3085 KEG_LOCK(keg); 3086 keg->uk_allocf = allocf; 3087 KEG_UNLOCK(keg); 3088} 3089 3090/* See uma.h */ 3091void 3092uma_zone_reserve(uma_zone_t zone, int items) 3093{ 3094 uma_keg_t keg; 3095 3096 keg = zone_first_keg(zone); 3097 if (keg == NULL) 3098 return; 3099 KEG_LOCK(keg); 3100 keg->uk_reserve = items; 3101 KEG_UNLOCK(keg); 3102 3103 return; 3104} 3105 3106/* See uma.h */ 3107int 3108uma_zone_reserve_kva(uma_zone_t zone, int count) 3109{ 3110 uma_keg_t keg; 3111 vm_offset_t kva; 3112 int pages; 3113 3114 keg = zone_first_keg(zone); 3115 if (keg == NULL) 3116 return (0); 3117 pages = count / keg->uk_ipers; 3118 3119 if (pages * keg->uk_ipers < count) 3120 pages++; 3121 3122#ifdef UMA_MD_SMALL_ALLOC 3123 if (keg->uk_ppera > 1) { 3124#else 3125 if (1) { 3126#endif 3127 kva = kva_alloc(pages * UMA_SLAB_SIZE); 3128 if (kva == 0) 3129 return (0); 3130 } else 3131 kva = 0; 3132 KEG_LOCK(keg); 3133 keg->uk_kva = kva; 3134 keg->uk_offset = 0; 3135 keg->uk_maxpages = pages; 3136#ifdef UMA_MD_SMALL_ALLOC 3137 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3138#else 3139 keg->uk_allocf = noobj_alloc; 3140#endif 3141 keg->uk_flags |= UMA_ZONE_NOFREE; 3142 KEG_UNLOCK(keg); 3143 3144 return (1); 3145} 3146 3147/* See uma.h */ 3148void 3149uma_prealloc(uma_zone_t zone, int items) 3150{ 3151 int slabs; 3152 uma_slab_t slab; 3153 uma_keg_t keg; 3154 3155 keg = zone_first_keg(zone); 3156 if (keg == NULL) 3157 return; 3158 KEG_LOCK(keg); 3159 slabs = items / keg->uk_ipers; 3160 if (slabs * keg->uk_ipers < items) 3161 slabs++; 3162 while (slabs > 0) { 3163 slab = keg_alloc_slab(keg, zone, M_WAITOK); 3164 if (slab == NULL) 3165 break; 3166 MPASS(slab->us_keg == keg); 3167 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 3168 slabs--; 3169 } 3170 KEG_UNLOCK(keg); 3171} 3172 3173/* See uma.h */ 3174uint32_t * 3175uma_find_refcnt(uma_zone_t zone, void *item) 3176{ 3177 uma_slabrefcnt_t slabref; 3178 uma_slab_t slab; 3179 uma_keg_t keg; 3180 uint32_t *refcnt; 3181 int idx; 3182 3183 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK)); 3184 slabref = (uma_slabrefcnt_t)slab; 3185 keg = slab->us_keg; 3186 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT, 3187 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT")); 3188 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3189 refcnt = &slabref->us_refcnt[idx]; 3190 return refcnt; 3191} 3192 3193/* See uma.h */ 3194static void 3195uma_reclaim_locked(bool kmem_danger) 3196{ 3197 3198#ifdef UMA_DEBUG 3199 printf("UMA: vm asked us to release pages!\n"); 3200#endif 3201 sx_assert(&uma_drain_lock, SA_XLOCKED); 3202 bucket_enable(); 3203 zone_foreach(zone_drain); 3204 if (vm_page_count_min() || kmem_danger) { 3205 cache_drain_safe(NULL); 3206 zone_foreach(zone_drain); 3207 } 3208 /* 3209 * Some slabs may have been freed but this zone will be visited early 3210 * we visit again so that we can free pages that are empty once other 3211 * zones are drained. We have to do the same for buckets. 3212 */ 3213 zone_drain(slabzone); 3214 zone_drain(slabrefzone); 3215 bucket_zone_drain(); 3216} 3217 3218void 3219uma_reclaim(void) 3220{ 3221 3222 sx_xlock(&uma_drain_lock); 3223 uma_reclaim_locked(false); 3224 sx_xunlock(&uma_drain_lock); 3225} 3226 3227static int uma_reclaim_needed; 3228 3229void 3230uma_reclaim_wakeup(void) 3231{ 3232 3233 uma_reclaim_needed = 1; 3234 wakeup(&uma_reclaim_needed); 3235} 3236 3237void 3238uma_reclaim_worker(void *arg __unused) 3239{ 3240 3241 sx_xlock(&uma_drain_lock); 3242 for (;;) { 3243 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM, 3244 "umarcl", 0); 3245 if (uma_reclaim_needed) { 3246 uma_reclaim_needed = 0; 3247 uma_reclaim_locked(true); 3248 } 3249 } 3250} 3251 3252/* See uma.h */ 3253int 3254uma_zone_exhausted(uma_zone_t zone) 3255{ 3256 int full; 3257 3258 ZONE_LOCK(zone); 3259 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3260 ZONE_UNLOCK(zone); 3261 return (full); 3262} 3263 3264int 3265uma_zone_exhausted_nolock(uma_zone_t zone) 3266{ 3267 return (zone->uz_flags & UMA_ZFLAG_FULL); 3268} 3269 3270void * 3271uma_large_malloc(vm_size_t size, int wait) 3272{ 3273 void *mem; 3274 uma_slab_t slab; 3275 uint8_t flags; 3276 3277 slab = zone_alloc_item(slabzone, NULL, wait); 3278 if (slab == NULL) 3279 return (NULL); 3280 mem = page_alloc(NULL, size, &flags, wait); 3281 if (mem) { 3282 vsetslab((vm_offset_t)mem, slab); 3283 slab->us_data = mem; 3284 slab->us_flags = flags | UMA_SLAB_MALLOC; 3285 slab->us_size = size; 3286 } else { 3287 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3288 } 3289 3290 return (mem); 3291} 3292 3293void 3294uma_large_free(uma_slab_t slab) 3295{ 3296 3297 page_free(slab->us_data, slab->us_size, slab->us_flags); 3298 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3299} 3300 3301static void 3302uma_zero_item(void *item, uma_zone_t zone) 3303{ 3304 3305 if (zone->uz_flags & UMA_ZONE_PCPU) { 3306 for (int i = 0; i < mp_ncpus; i++) 3307 bzero(zpcpu_get_cpu(item, i), zone->uz_size); 3308 } else 3309 bzero(item, zone->uz_size); 3310} 3311 3312void 3313uma_print_stats(void) 3314{ 3315 zone_foreach(uma_print_zone); 3316} 3317 3318static void 3319slab_print(uma_slab_t slab) 3320{ 3321 printf("slab: keg %p, data %p, freecount %d\n", 3322 slab->us_keg, slab->us_data, slab->us_freecount); 3323} 3324 3325static void 3326cache_print(uma_cache_t cache) 3327{ 3328 printf("alloc: %p(%d), free: %p(%d)\n", 3329 cache->uc_allocbucket, 3330 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3331 cache->uc_freebucket, 3332 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3333} 3334 3335static void 3336uma_print_keg(uma_keg_t keg) 3337{ 3338 uma_slab_t slab; 3339 3340 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3341 "out %d free %d limit %d\n", 3342 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3343 keg->uk_ipers, keg->uk_ppera, 3344 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free, 3345 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3346 printf("Part slabs:\n"); 3347 LIST_FOREACH(slab, &keg->uk_part_slab, us_link) 3348 slab_print(slab); 3349 printf("Free slabs:\n"); 3350 LIST_FOREACH(slab, &keg->uk_free_slab, us_link) 3351 slab_print(slab); 3352 printf("Full slabs:\n"); 3353 LIST_FOREACH(slab, &keg->uk_full_slab, us_link) 3354 slab_print(slab); 3355} 3356 3357void 3358uma_print_zone(uma_zone_t zone) 3359{ 3360 uma_cache_t cache; 3361 uma_klink_t kl; 3362 int i; 3363 3364 printf("zone: %s(%p) size %d flags %#x\n", 3365 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3366 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3367 uma_print_keg(kl->kl_keg); 3368 CPU_FOREACH(i) { 3369 cache = &zone->uz_cpu[i]; 3370 printf("CPU %d Cache:\n", i); 3371 cache_print(cache); 3372 } 3373} 3374 3375#ifdef DDB 3376/* 3377 * Generate statistics across both the zone and its per-cpu cache's. Return 3378 * desired statistics if the pointer is non-NULL for that statistic. 3379 * 3380 * Note: does not update the zone statistics, as it can't safely clear the 3381 * per-CPU cache statistic. 3382 * 3383 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3384 * safe from off-CPU; we should modify the caches to track this information 3385 * directly so that we don't have to. 3386 */ 3387static void 3388uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3389 uint64_t *freesp, uint64_t *sleepsp) 3390{ 3391 uma_cache_t cache; 3392 uint64_t allocs, frees, sleeps; 3393 int cachefree, cpu; 3394 3395 allocs = frees = sleeps = 0; 3396 cachefree = 0; 3397 CPU_FOREACH(cpu) { 3398 cache = &z->uz_cpu[cpu]; 3399 if (cache->uc_allocbucket != NULL) 3400 cachefree += cache->uc_allocbucket->ub_cnt; 3401 if (cache->uc_freebucket != NULL) 3402 cachefree += cache->uc_freebucket->ub_cnt; 3403 allocs += cache->uc_allocs; 3404 frees += cache->uc_frees; 3405 } 3406 allocs += z->uz_allocs; 3407 frees += z->uz_frees; 3408 sleeps += z->uz_sleeps; 3409 if (cachefreep != NULL) 3410 *cachefreep = cachefree; 3411 if (allocsp != NULL) 3412 *allocsp = allocs; 3413 if (freesp != NULL) 3414 *freesp = frees; 3415 if (sleepsp != NULL) 3416 *sleepsp = sleeps; 3417} 3418#endif /* DDB */ 3419 3420static int 3421sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3422{ 3423 uma_keg_t kz; 3424 uma_zone_t z; 3425 int count; 3426 3427 count = 0; 3428 rw_rlock(&uma_rwlock); 3429 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3430 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3431 count++; 3432 } 3433 rw_runlock(&uma_rwlock); 3434 return (sysctl_handle_int(oidp, &count, 0, req)); 3435} 3436 3437static int 3438sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3439{ 3440 struct uma_stream_header ush; 3441 struct uma_type_header uth; 3442 struct uma_percpu_stat ups; 3443 uma_bucket_t bucket; 3444 struct sbuf sbuf; 3445 uma_cache_t cache; 3446 uma_klink_t kl; 3447 uma_keg_t kz; 3448 uma_zone_t z; 3449 uma_keg_t k; 3450 int count, error, i; 3451 3452 error = sysctl_wire_old_buffer(req, 0); 3453 if (error != 0) 3454 return (error); 3455 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3456 3457 count = 0; 3458 rw_rlock(&uma_rwlock); 3459 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3460 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3461 count++; 3462 } 3463 3464 /* 3465 * Insert stream header. 3466 */ 3467 bzero(&ush, sizeof(ush)); 3468 ush.ush_version = UMA_STREAM_VERSION; 3469 ush.ush_maxcpus = (mp_maxid + 1); 3470 ush.ush_count = count; 3471 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3472 3473 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3474 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3475 bzero(&uth, sizeof(uth)); 3476 ZONE_LOCK(z); 3477 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3478 uth.uth_align = kz->uk_align; 3479 uth.uth_size = kz->uk_size; 3480 uth.uth_rsize = kz->uk_rsize; 3481 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3482 k = kl->kl_keg; 3483 uth.uth_maxpages += k->uk_maxpages; 3484 uth.uth_pages += k->uk_pages; 3485 uth.uth_keg_free += k->uk_free; 3486 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3487 * k->uk_ipers; 3488 } 3489 3490 /* 3491 * A zone is secondary is it is not the first entry 3492 * on the keg's zone list. 3493 */ 3494 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3495 (LIST_FIRST(&kz->uk_zones) != z)) 3496 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3497 3498 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3499 uth.uth_zone_free += bucket->ub_cnt; 3500 uth.uth_allocs = z->uz_allocs; 3501 uth.uth_frees = z->uz_frees; 3502 uth.uth_fails = z->uz_fails; 3503 uth.uth_sleeps = z->uz_sleeps; 3504 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 3505 /* 3506 * While it is not normally safe to access the cache 3507 * bucket pointers while not on the CPU that owns the 3508 * cache, we only allow the pointers to be exchanged 3509 * without the zone lock held, not invalidated, so 3510 * accept the possible race associated with bucket 3511 * exchange during monitoring. 3512 */ 3513 for (i = 0; i < (mp_maxid + 1); i++) { 3514 bzero(&ups, sizeof(ups)); 3515 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) 3516 goto skip; 3517 if (CPU_ABSENT(i)) 3518 goto skip; 3519 cache = &z->uz_cpu[i]; 3520 if (cache->uc_allocbucket != NULL) 3521 ups.ups_cache_free += 3522 cache->uc_allocbucket->ub_cnt; 3523 if (cache->uc_freebucket != NULL) 3524 ups.ups_cache_free += 3525 cache->uc_freebucket->ub_cnt; 3526 ups.ups_allocs = cache->uc_allocs; 3527 ups.ups_frees = cache->uc_frees; 3528skip: 3529 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups)); 3530 } 3531 ZONE_UNLOCK(z); 3532 } 3533 } 3534 rw_runlock(&uma_rwlock); 3535 error = sbuf_finish(&sbuf); 3536 sbuf_delete(&sbuf); 3537 return (error); 3538} 3539 3540int 3541sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 3542{ 3543 uma_zone_t zone = *(uma_zone_t *)arg1; 3544 int error, max, old; 3545 3546 old = max = uma_zone_get_max(zone); 3547 error = sysctl_handle_int(oidp, &max, 0, req); 3548 if (error || !req->newptr) 3549 return (error); 3550 3551 if (max < old) 3552 return (EINVAL); 3553 3554 uma_zone_set_max(zone, max); 3555 3556 return (0); 3557} 3558 3559int 3560sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 3561{ 3562 uma_zone_t zone = *(uma_zone_t *)arg1; 3563 int cur; 3564 3565 cur = uma_zone_get_cur(zone); 3566 return (sysctl_handle_int(oidp, &cur, 0, req)); 3567} 3568 3569#ifdef DDB 3570DB_SHOW_COMMAND(uma, db_show_uma) 3571{ 3572 uint64_t allocs, frees, sleeps; 3573 uma_bucket_t bucket; 3574 uma_keg_t kz; 3575 uma_zone_t z; 3576 int cachefree; 3577 3578 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used", 3579 "Free", "Requests", "Sleeps", "Bucket"); 3580 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3581 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3582 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 3583 allocs = z->uz_allocs; 3584 frees = z->uz_frees; 3585 sleeps = z->uz_sleeps; 3586 cachefree = 0; 3587 } else 3588 uma_zone_sumstat(z, &cachefree, &allocs, 3589 &frees, &sleeps); 3590 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 3591 (LIST_FIRST(&kz->uk_zones) != z))) 3592 cachefree += kz->uk_free; 3593 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3594 cachefree += bucket->ub_cnt; 3595 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n", 3596 z->uz_name, (uintmax_t)kz->uk_size, 3597 (intmax_t)(allocs - frees), cachefree, 3598 (uintmax_t)allocs, sleeps, z->uz_count); 3599 if (db_pager_quit) 3600 return; 3601 } 3602 } 3603} 3604 3605DB_SHOW_COMMAND(umacache, db_show_umacache) 3606{ 3607 uint64_t allocs, frees; 3608 uma_bucket_t bucket; 3609 uma_zone_t z; 3610 int cachefree; 3611 3612 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 3613 "Requests", "Bucket"); 3614 LIST_FOREACH(z, &uma_cachezones, uz_link) { 3615 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL); 3616 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3617 cachefree += bucket->ub_cnt; 3618 db_printf("%18s %8ju %8jd %8d %12ju %8u\n", 3619 z->uz_name, (uintmax_t)z->uz_size, 3620 (intmax_t)(allocs - frees), cachefree, 3621 (uintmax_t)allocs, z->uz_count); 3622 if (db_pager_quit) 3623 return; 3624 } 3625} 3626#endif 3627