uma_core.c revision 260303
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 260303 2014-01-04 23:39:39Z mav $"); 52 53/* I should really use ktr.. */ 54/* 55#define UMA_DEBUG 1 56#define UMA_DEBUG_ALLOC 1 57#define UMA_DEBUG_ALLOC_1 1 58*/ 59 60#include "opt_ddb.h" 61#include "opt_param.h" 62#include "opt_vm.h" 63 64#include <sys/param.h> 65#include <sys/systm.h> 66#include <sys/bitset.h> 67#include <sys/kernel.h> 68#include <sys/types.h> 69#include <sys/queue.h> 70#include <sys/malloc.h> 71#include <sys/ktr.h> 72#include <sys/lock.h> 73#include <sys/sysctl.h> 74#include <sys/mutex.h> 75#include <sys/proc.h> 76#include <sys/rwlock.h> 77#include <sys/sbuf.h> 78#include <sys/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/* This mutex protects the keg list */ 135static struct mtx_padalign uma_mtx; 136 137/* Linked list of boot time pages */ 138static LIST_HEAD(,uma_slab) uma_boot_pages = 139 LIST_HEAD_INITIALIZER(uma_boot_pages); 140 141/* This mutex protects the boot time pages list */ 142static struct mtx_padalign uma_boot_pages_mtx; 143 144/* Is the VM done starting up? */ 145static int booted = 0; 146#define UMA_STARTUP 1 147#define UMA_STARTUP2 2 148 149/* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */ 150static const u_int uma_max_ipers = SLAB_SETSIZE; 151 152/* 153 * Only mbuf clusters use ref zones. Just provide enough references 154 * to support the one user. New code should not use the ref facility. 155 */ 156static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES; 157 158/* 159 * This is the handle used to schedule events that need to happen 160 * outside of the allocation fast path. 161 */ 162static struct callout uma_callout; 163#define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 164 165/* 166 * This structure is passed as the zone ctor arg so that I don't have to create 167 * a special allocation function just for zones. 168 */ 169struct uma_zctor_args { 170 const char *name; 171 size_t size; 172 uma_ctor ctor; 173 uma_dtor dtor; 174 uma_init uminit; 175 uma_fini fini; 176 uma_import import; 177 uma_release release; 178 void *arg; 179 uma_keg_t keg; 180 int align; 181 uint32_t flags; 182}; 183 184struct uma_kctor_args { 185 uma_zone_t zone; 186 size_t size; 187 uma_init uminit; 188 uma_fini fini; 189 int align; 190 uint32_t flags; 191}; 192 193struct uma_bucket_zone { 194 uma_zone_t ubz_zone; 195 char *ubz_name; 196 int ubz_entries; /* Number of items it can hold. */ 197 int ubz_maxsize; /* Maximum allocation size per-item. */ 198}; 199 200/* 201 * Compute the actual number of bucket entries to pack them in power 202 * of two sizes for more efficient space utilization. 203 */ 204#define BUCKET_SIZE(n) \ 205 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 206 207#define BUCKET_MAX BUCKET_SIZE(128) 208 209struct uma_bucket_zone bucket_zones[] = { 210 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 211 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 212 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 213 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 214 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 215 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 216 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 217 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 218 { NULL, NULL, 0} 219}; 220 221/* 222 * Flags and enumerations to be passed to internal functions. 223 */ 224enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; 225 226/* Prototypes.. */ 227 228static void *noobj_alloc(uma_zone_t, int, uint8_t *, int); 229static void *page_alloc(uma_zone_t, int, uint8_t *, int); 230static void *startup_alloc(uma_zone_t, int, uint8_t *, int); 231static void page_free(void *, int, uint8_t); 232static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); 233static void cache_drain(uma_zone_t); 234static void bucket_drain(uma_zone_t, uma_bucket_t); 235static void bucket_cache_drain(uma_zone_t zone); 236static int keg_ctor(void *, int, void *, int); 237static void keg_dtor(void *, int, void *); 238static int zone_ctor(void *, int, void *, int); 239static void zone_dtor(void *, int, void *); 240static int zero_init(void *, int, int); 241static void keg_small_init(uma_keg_t keg); 242static void keg_large_init(uma_keg_t keg); 243static void zone_foreach(void (*zfunc)(uma_zone_t)); 244static void zone_timeout(uma_zone_t zone); 245static int hash_alloc(struct uma_hash *); 246static int hash_expand(struct uma_hash *, struct uma_hash *); 247static void hash_free(struct uma_hash *hash); 248static void uma_timeout(void *); 249static void uma_startup3(void); 250static void *zone_alloc_item(uma_zone_t, void *, int); 251static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 252static void bucket_enable(void); 253static void bucket_init(void); 254static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 255static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 256static void bucket_zone_drain(void); 257static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags); 258static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); 259static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); 260static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 261static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); 262static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 263 uma_fini fini, int align, uint32_t flags); 264static int zone_import(uma_zone_t zone, void **bucket, int max, int flags); 265static void zone_release(uma_zone_t zone, void **bucket, int cnt); 266 267void uma_print_zone(uma_zone_t); 268void uma_print_stats(void); 269static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 270static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 271 272SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 273 274SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 275 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 276 277SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 278 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 279 280static int zone_warnings = 1; 281TUNABLE_INT("vm.zone_warnings", &zone_warnings); 282SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0, 283 "Warn when UMA zones becomes full"); 284 285/* 286 * This routine checks to see whether or not it's safe to enable buckets. 287 */ 288static void 289bucket_enable(void) 290{ 291 bucketdisable = vm_page_count_min(); 292} 293 294/* 295 * Initialize bucket_zones, the array of zones of buckets of various sizes. 296 * 297 * For each zone, calculate the memory required for each bucket, consisting 298 * of the header and an array of pointers. 299 */ 300static void 301bucket_init(void) 302{ 303 struct uma_bucket_zone *ubz; 304 int size; 305 int i; 306 307 for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 308 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 309 size += sizeof(void *) * ubz->ubz_entries; 310 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 311 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 312 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET); 313 } 314} 315 316/* 317 * Given a desired number of entries for a bucket, return the zone from which 318 * to allocate the bucket. 319 */ 320static struct uma_bucket_zone * 321bucket_zone_lookup(int entries) 322{ 323 struct uma_bucket_zone *ubz; 324 325 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 326 if (ubz->ubz_entries >= entries) 327 return (ubz); 328 ubz--; 329 return (ubz); 330} 331 332static int 333bucket_select(int size) 334{ 335 struct uma_bucket_zone *ubz; 336 337 ubz = &bucket_zones[0]; 338 if (size > ubz->ubz_maxsize) 339 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 340 341 for (; ubz->ubz_entries != 0; ubz++) 342 if (ubz->ubz_maxsize < size) 343 break; 344 ubz--; 345 return (ubz->ubz_entries); 346} 347 348static uma_bucket_t 349bucket_alloc(uma_zone_t zone, void *udata, int flags) 350{ 351 struct uma_bucket_zone *ubz; 352 uma_bucket_t bucket; 353 354 /* 355 * This is to stop us from allocating per cpu buckets while we're 356 * running out of vm.boot_pages. Otherwise, we would exhaust the 357 * boot pages. This also prevents us from allocating buckets in 358 * low memory situations. 359 */ 360 if (bucketdisable) 361 return (NULL); 362 /* 363 * To limit bucket recursion we store the original zone flags 364 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 365 * NOVM flag to persist even through deep recursions. We also 366 * store ZFLAG_BUCKET once we have recursed attempting to allocate 367 * a bucket for a bucket zone so we do not allow infinite bucket 368 * recursion. This cookie will even persist to frees of unused 369 * buckets via the allocation path or bucket allocations in the 370 * free path. 371 */ 372 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 373 udata = (void *)(uintptr_t)zone->uz_flags; 374 else { 375 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 376 return (NULL); 377 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 378 } 379 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 380 flags |= M_NOVM; 381 ubz = bucket_zone_lookup(zone->uz_count); 382 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 383 if (bucket) { 384#ifdef INVARIANTS 385 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 386#endif 387 bucket->ub_cnt = 0; 388 bucket->ub_entries = ubz->ubz_entries; 389 } 390 391 return (bucket); 392} 393 394static void 395bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 396{ 397 struct uma_bucket_zone *ubz; 398 399 KASSERT(bucket->ub_cnt == 0, 400 ("bucket_free: Freeing a non free bucket.")); 401 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 402 udata = (void *)(uintptr_t)zone->uz_flags; 403 ubz = bucket_zone_lookup(bucket->ub_entries); 404 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 405} 406 407static void 408bucket_zone_drain(void) 409{ 410 struct uma_bucket_zone *ubz; 411 412 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 413 zone_drain(ubz->ubz_zone); 414} 415 416static void 417zone_log_warning(uma_zone_t zone) 418{ 419 static const struct timeval warninterval = { 300, 0 }; 420 421 if (!zone_warnings || zone->uz_warning == NULL) 422 return; 423 424 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 425 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 426} 427 428static void 429zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 430{ 431 uma_klink_t klink; 432 433 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 434 kegfn(klink->kl_keg); 435} 436 437/* 438 * Routine called by timeout which is used to fire off some time interval 439 * based calculations. (stats, hash size, etc.) 440 * 441 * Arguments: 442 * arg Unused 443 * 444 * Returns: 445 * Nothing 446 */ 447static void 448uma_timeout(void *unused) 449{ 450 bucket_enable(); 451 zone_foreach(zone_timeout); 452 453 /* Reschedule this event */ 454 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 455} 456 457/* 458 * Routine to perform timeout driven calculations. This expands the 459 * hashes and does per cpu statistics aggregation. 460 * 461 * Returns nothing. 462 */ 463static void 464keg_timeout(uma_keg_t keg) 465{ 466 467 KEG_LOCK(keg); 468 /* 469 * Expand the keg hash table. 470 * 471 * This is done if the number of slabs is larger than the hash size. 472 * What I'm trying to do here is completely reduce collisions. This 473 * may be a little aggressive. Should I allow for two collisions max? 474 */ 475 if (keg->uk_flags & UMA_ZONE_HASH && 476 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 477 struct uma_hash newhash; 478 struct uma_hash oldhash; 479 int ret; 480 481 /* 482 * This is so involved because allocating and freeing 483 * while the keg lock is held will lead to deadlock. 484 * I have to do everything in stages and check for 485 * races. 486 */ 487 newhash = keg->uk_hash; 488 KEG_UNLOCK(keg); 489 ret = hash_alloc(&newhash); 490 KEG_LOCK(keg); 491 if (ret) { 492 if (hash_expand(&keg->uk_hash, &newhash)) { 493 oldhash = keg->uk_hash; 494 keg->uk_hash = newhash; 495 } else 496 oldhash = newhash; 497 498 KEG_UNLOCK(keg); 499 hash_free(&oldhash); 500 return; 501 } 502 } 503 KEG_UNLOCK(keg); 504} 505 506static void 507zone_timeout(uma_zone_t zone) 508{ 509 510 zone_foreach_keg(zone, &keg_timeout); 511} 512 513/* 514 * Allocate and zero fill the next sized hash table from the appropriate 515 * backing store. 516 * 517 * Arguments: 518 * hash A new hash structure with the old hash size in uh_hashsize 519 * 520 * Returns: 521 * 1 on sucess and 0 on failure. 522 */ 523static int 524hash_alloc(struct uma_hash *hash) 525{ 526 int oldsize; 527 int alloc; 528 529 oldsize = hash->uh_hashsize; 530 531 /* We're just going to go to a power of two greater */ 532 if (oldsize) { 533 hash->uh_hashsize = oldsize * 2; 534 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 535 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 536 M_UMAHASH, M_NOWAIT); 537 } else { 538 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 539 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 540 M_WAITOK); 541 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 542 } 543 if (hash->uh_slab_hash) { 544 bzero(hash->uh_slab_hash, alloc); 545 hash->uh_hashmask = hash->uh_hashsize - 1; 546 return (1); 547 } 548 549 return (0); 550} 551 552/* 553 * Expands the hash table for HASH zones. This is done from zone_timeout 554 * to reduce collisions. This must not be done in the regular allocation 555 * path, otherwise, we can recurse on the vm while allocating pages. 556 * 557 * Arguments: 558 * oldhash The hash you want to expand 559 * newhash The hash structure for the new table 560 * 561 * Returns: 562 * Nothing 563 * 564 * Discussion: 565 */ 566static int 567hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 568{ 569 uma_slab_t slab; 570 int hval; 571 int i; 572 573 if (!newhash->uh_slab_hash) 574 return (0); 575 576 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 577 return (0); 578 579 /* 580 * I need to investigate hash algorithms for resizing without a 581 * full rehash. 582 */ 583 584 for (i = 0; i < oldhash->uh_hashsize; i++) 585 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 586 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 587 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 588 hval = UMA_HASH(newhash, slab->us_data); 589 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 590 slab, us_hlink); 591 } 592 593 return (1); 594} 595 596/* 597 * Free the hash bucket to the appropriate backing store. 598 * 599 * Arguments: 600 * slab_hash The hash bucket we're freeing 601 * hashsize The number of entries in that hash bucket 602 * 603 * Returns: 604 * Nothing 605 */ 606static void 607hash_free(struct uma_hash *hash) 608{ 609 if (hash->uh_slab_hash == NULL) 610 return; 611 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 612 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 613 else 614 free(hash->uh_slab_hash, M_UMAHASH); 615} 616 617/* 618 * Frees all outstanding items in a bucket 619 * 620 * Arguments: 621 * zone The zone to free to, must be unlocked. 622 * bucket The free/alloc bucket with items, cpu queue must be locked. 623 * 624 * Returns: 625 * Nothing 626 */ 627 628static void 629bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 630{ 631 int i; 632 633 if (bucket == NULL) 634 return; 635 636 if (zone->uz_fini) 637 for (i = 0; i < bucket->ub_cnt; i++) 638 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 639 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 640 bucket->ub_cnt = 0; 641} 642 643/* 644 * Drains the per cpu caches for a zone. 645 * 646 * NOTE: This may only be called while the zone is being turn down, and not 647 * during normal operation. This is necessary in order that we do not have 648 * to migrate CPUs to drain the per-CPU caches. 649 * 650 * Arguments: 651 * zone The zone to drain, must be unlocked. 652 * 653 * Returns: 654 * Nothing 655 */ 656static void 657cache_drain(uma_zone_t zone) 658{ 659 uma_cache_t cache; 660 int cpu; 661 662 /* 663 * XXX: It is safe to not lock the per-CPU caches, because we're 664 * tearing down the zone anyway. I.e., there will be no further use 665 * of the caches at this point. 666 * 667 * XXX: It would good to be able to assert that the zone is being 668 * torn down to prevent improper use of cache_drain(). 669 * 670 * XXX: We lock the zone before passing into bucket_cache_drain() as 671 * it is used elsewhere. Should the tear-down path be made special 672 * there in some form? 673 */ 674 CPU_FOREACH(cpu) { 675 cache = &zone->uz_cpu[cpu]; 676 bucket_drain(zone, cache->uc_allocbucket); 677 bucket_drain(zone, cache->uc_freebucket); 678 if (cache->uc_allocbucket != NULL) 679 bucket_free(zone, cache->uc_allocbucket, NULL); 680 if (cache->uc_freebucket != NULL) 681 bucket_free(zone, cache->uc_freebucket, NULL); 682 cache->uc_allocbucket = cache->uc_freebucket = NULL; 683 } 684 ZONE_LOCK(zone); 685 bucket_cache_drain(zone); 686 ZONE_UNLOCK(zone); 687} 688 689static void 690cache_shrink(uma_zone_t zone) 691{ 692 693 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 694 return; 695 696 ZONE_LOCK(zone); 697 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; 698 ZONE_UNLOCK(zone); 699} 700 701static void 702cache_drain_safe_cpu(uma_zone_t zone) 703{ 704 uma_cache_t cache; 705 uma_bucket_t b1, b2; 706 707 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 708 return; 709 710 b1 = b2 = NULL; 711 ZONE_LOCK(zone); 712 critical_enter(); 713 cache = &zone->uz_cpu[curcpu]; 714 if (cache->uc_allocbucket) { 715 if (cache->uc_allocbucket->ub_cnt != 0) 716 LIST_INSERT_HEAD(&zone->uz_buckets, 717 cache->uc_allocbucket, ub_link); 718 else 719 b1 = cache->uc_allocbucket; 720 cache->uc_allocbucket = NULL; 721 } 722 if (cache->uc_freebucket) { 723 if (cache->uc_freebucket->ub_cnt != 0) 724 LIST_INSERT_HEAD(&zone->uz_buckets, 725 cache->uc_freebucket, ub_link); 726 else 727 b2 = cache->uc_freebucket; 728 cache->uc_freebucket = NULL; 729 } 730 critical_exit(); 731 ZONE_UNLOCK(zone); 732 if (b1) 733 bucket_free(zone, b1, NULL); 734 if (b2) 735 bucket_free(zone, b2, NULL); 736} 737 738/* 739 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 740 * This is an expensive call because it needs to bind to all CPUs 741 * one by one and enter a critical section on each of them in order 742 * to safely access their cache buckets. 743 * Zone lock must not be held on call this function. 744 */ 745static void 746cache_drain_safe(uma_zone_t zone) 747{ 748 int cpu; 749 750 /* 751 * Polite bucket sizes shrinking was not enouth, shrink aggressively. 752 */ 753 if (zone) 754 cache_shrink(zone); 755 else 756 zone_foreach(cache_shrink); 757 758 CPU_FOREACH(cpu) { 759 thread_lock(curthread); 760 sched_bind(curthread, cpu); 761 thread_unlock(curthread); 762 763 if (zone) 764 cache_drain_safe_cpu(zone); 765 else 766 zone_foreach(cache_drain_safe_cpu); 767 } 768 thread_lock(curthread); 769 sched_unbind(curthread); 770 thread_unlock(curthread); 771} 772 773/* 774 * Drain the cached buckets from a zone. Expects a locked zone on entry. 775 */ 776static void 777bucket_cache_drain(uma_zone_t zone) 778{ 779 uma_bucket_t bucket; 780 781 /* 782 * Drain the bucket queues and free the buckets, we just keep two per 783 * cpu (alloc/free). 784 */ 785 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 786 LIST_REMOVE(bucket, ub_link); 787 ZONE_UNLOCK(zone); 788 bucket_drain(zone, bucket); 789 bucket_free(zone, bucket, NULL); 790 ZONE_LOCK(zone); 791 } 792 793 /* 794 * Shrink further bucket sizes. Price of single zone lock collision 795 * is probably lower then price of global cache drain. 796 */ 797 if (zone->uz_count > zone->uz_count_min) 798 zone->uz_count--; 799} 800 801static void 802keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 803{ 804 uint8_t *mem; 805 int i; 806 uint8_t flags; 807 808 mem = slab->us_data; 809 flags = slab->us_flags; 810 i = start; 811 if (keg->uk_fini != NULL) { 812 for (i--; i > -1; i--) 813 keg->uk_fini(slab->us_data + (keg->uk_rsize * i), 814 keg->uk_size); 815 } 816 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 817 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 818#ifdef UMA_DEBUG 819 printf("%s: Returning %d bytes.\n", keg->uk_name, 820 PAGE_SIZE * keg->uk_ppera); 821#endif 822 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 823} 824 825/* 826 * Frees pages from a keg back to the system. This is done on demand from 827 * the pageout daemon. 828 * 829 * Returns nothing. 830 */ 831static void 832keg_drain(uma_keg_t keg) 833{ 834 struct slabhead freeslabs = { 0 }; 835 uma_slab_t slab; 836 uma_slab_t n; 837 838 /* 839 * We don't want to take pages from statically allocated kegs at this 840 * time 841 */ 842 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 843 return; 844 845#ifdef UMA_DEBUG 846 printf("%s free items: %u\n", keg->uk_name, keg->uk_free); 847#endif 848 KEG_LOCK(keg); 849 if (keg->uk_free == 0) 850 goto finished; 851 852 slab = LIST_FIRST(&keg->uk_free_slab); 853 while (slab) { 854 n = LIST_NEXT(slab, us_link); 855 856 /* We have no where to free these to */ 857 if (slab->us_flags & UMA_SLAB_BOOT) { 858 slab = n; 859 continue; 860 } 861 862 LIST_REMOVE(slab, us_link); 863 keg->uk_pages -= keg->uk_ppera; 864 keg->uk_free -= keg->uk_ipers; 865 866 if (keg->uk_flags & UMA_ZONE_HASH) 867 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); 868 869 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 870 871 slab = n; 872 } 873finished: 874 KEG_UNLOCK(keg); 875 876 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 877 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 878 keg_free_slab(keg, slab, keg->uk_ipers); 879 } 880} 881 882static void 883zone_drain_wait(uma_zone_t zone, int waitok) 884{ 885 886 /* 887 * Set draining to interlock with zone_dtor() so we can release our 888 * locks as we go. Only dtor() should do a WAITOK call since it 889 * is the only call that knows the structure will still be available 890 * when it wakes up. 891 */ 892 ZONE_LOCK(zone); 893 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 894 if (waitok == M_NOWAIT) 895 goto out; 896 mtx_unlock(&uma_mtx); 897 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 898 mtx_lock(&uma_mtx); 899 } 900 zone->uz_flags |= UMA_ZFLAG_DRAINING; 901 bucket_cache_drain(zone); 902 ZONE_UNLOCK(zone); 903 /* 904 * The DRAINING flag protects us from being freed while 905 * we're running. Normally the uma_mtx would protect us but we 906 * must be able to release and acquire the right lock for each keg. 907 */ 908 zone_foreach_keg(zone, &keg_drain); 909 ZONE_LOCK(zone); 910 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 911 wakeup(zone); 912out: 913 ZONE_UNLOCK(zone); 914} 915 916void 917zone_drain(uma_zone_t zone) 918{ 919 920 zone_drain_wait(zone, M_NOWAIT); 921} 922 923/* 924 * Allocate a new slab for a keg. This does not insert the slab onto a list. 925 * 926 * Arguments: 927 * wait Shall we wait? 928 * 929 * Returns: 930 * The slab that was allocated or NULL if there is no memory and the 931 * caller specified M_NOWAIT. 932 */ 933static uma_slab_t 934keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) 935{ 936 uma_slabrefcnt_t slabref; 937 uma_alloc allocf; 938 uma_slab_t slab; 939 uint8_t *mem; 940 uint8_t flags; 941 int i; 942 943 mtx_assert(&keg->uk_lock, MA_OWNED); 944 slab = NULL; 945 mem = NULL; 946 947#ifdef UMA_DEBUG 948 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name); 949#endif 950 allocf = keg->uk_allocf; 951 KEG_UNLOCK(keg); 952 953 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 954 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); 955 if (slab == NULL) 956 goto out; 957 } 958 959 /* 960 * This reproduces the old vm_zone behavior of zero filling pages the 961 * first time they are added to a zone. 962 * 963 * Malloced items are zeroed in uma_zalloc. 964 */ 965 966 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 967 wait |= M_ZERO; 968 else 969 wait &= ~M_ZERO; 970 971 if (keg->uk_flags & UMA_ZONE_NODUMP) 972 wait |= M_NODUMP; 973 974 /* zone is passed for legacy reasons. */ 975 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait); 976 if (mem == NULL) { 977 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 978 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 979 slab = NULL; 980 goto out; 981 } 982 983 /* Point the slab into the allocated memory */ 984 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 985 slab = (uma_slab_t )(mem + keg->uk_pgoff); 986 987 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 988 for (i = 0; i < keg->uk_ppera; i++) 989 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 990 991 slab->us_keg = keg; 992 slab->us_data = mem; 993 slab->us_freecount = keg->uk_ipers; 994 slab->us_flags = flags; 995 BIT_FILL(SLAB_SETSIZE, &slab->us_free); 996#ifdef INVARIANTS 997 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); 998#endif 999 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1000 slabref = (uma_slabrefcnt_t)slab; 1001 for (i = 0; i < keg->uk_ipers; i++) 1002 slabref->us_refcnt[i] = 0; 1003 } 1004 1005 if (keg->uk_init != NULL) { 1006 for (i = 0; i < keg->uk_ipers; i++) 1007 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 1008 keg->uk_size, wait) != 0) 1009 break; 1010 if (i != keg->uk_ipers) { 1011 keg_free_slab(keg, slab, i); 1012 slab = NULL; 1013 goto out; 1014 } 1015 } 1016out: 1017 KEG_LOCK(keg); 1018 1019 if (slab != NULL) { 1020 if (keg->uk_flags & UMA_ZONE_HASH) 1021 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1022 1023 keg->uk_pages += keg->uk_ppera; 1024 keg->uk_free += keg->uk_ipers; 1025 } 1026 1027 return (slab); 1028} 1029 1030/* 1031 * This function is intended to be used early on in place of page_alloc() so 1032 * that we may use the boot time page cache to satisfy allocations before 1033 * the VM is ready. 1034 */ 1035static void * 1036startup_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait) 1037{ 1038 uma_keg_t keg; 1039 uma_slab_t tmps; 1040 int pages, check_pages; 1041 1042 keg = zone_first_keg(zone); 1043 pages = howmany(bytes, PAGE_SIZE); 1044 check_pages = pages - 1; 1045 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n")); 1046 1047 /* 1048 * Check our small startup cache to see if it has pages remaining. 1049 */ 1050 mtx_lock(&uma_boot_pages_mtx); 1051 1052 /* First check if we have enough room. */ 1053 tmps = LIST_FIRST(&uma_boot_pages); 1054 while (tmps != NULL && check_pages-- > 0) 1055 tmps = LIST_NEXT(tmps, us_link); 1056 if (tmps != NULL) { 1057 /* 1058 * It's ok to lose tmps references. The last one will 1059 * have tmps->us_data pointing to the start address of 1060 * "pages" contiguous pages of memory. 1061 */ 1062 while (pages-- > 0) { 1063 tmps = LIST_FIRST(&uma_boot_pages); 1064 LIST_REMOVE(tmps, us_link); 1065 } 1066 mtx_unlock(&uma_boot_pages_mtx); 1067 *pflag = tmps->us_flags; 1068 return (tmps->us_data); 1069 } 1070 mtx_unlock(&uma_boot_pages_mtx); 1071 if (booted < UMA_STARTUP2) 1072 panic("UMA: Increase vm.boot_pages"); 1073 /* 1074 * Now that we've booted reset these users to their real allocator. 1075 */ 1076#ifdef UMA_MD_SMALL_ALLOC 1077 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; 1078#else 1079 keg->uk_allocf = page_alloc; 1080#endif 1081 return keg->uk_allocf(zone, bytes, pflag, wait); 1082} 1083 1084/* 1085 * Allocates a number of pages from the system 1086 * 1087 * Arguments: 1088 * bytes The number of bytes requested 1089 * wait Shall we wait? 1090 * 1091 * Returns: 1092 * A pointer to the alloced memory or possibly 1093 * NULL if M_NOWAIT is set. 1094 */ 1095static void * 1096page_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait) 1097{ 1098 void *p; /* Returned page */ 1099 1100 *pflag = UMA_SLAB_KMEM; 1101 p = (void *) kmem_malloc(kmem_arena, bytes, wait); 1102 1103 return (p); 1104} 1105 1106/* 1107 * Allocates a number of pages from within an object 1108 * 1109 * Arguments: 1110 * bytes The number of bytes requested 1111 * wait Shall we wait? 1112 * 1113 * Returns: 1114 * A pointer to the alloced memory or possibly 1115 * NULL if M_NOWAIT is set. 1116 */ 1117static void * 1118noobj_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait) 1119{ 1120 TAILQ_HEAD(, vm_page) alloctail; 1121 u_long npages; 1122 vm_offset_t retkva, zkva; 1123 vm_page_t p, p_next; 1124 uma_keg_t keg; 1125 1126 TAILQ_INIT(&alloctail); 1127 keg = zone_first_keg(zone); 1128 1129 npages = howmany(bytes, PAGE_SIZE); 1130 while (npages > 0) { 1131 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT | 1132 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ); 1133 if (p != NULL) { 1134 /* 1135 * Since the page does not belong to an object, its 1136 * listq is unused. 1137 */ 1138 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1139 npages--; 1140 continue; 1141 } 1142 if (wait & M_WAITOK) { 1143 VM_WAIT; 1144 continue; 1145 } 1146 1147 /* 1148 * Page allocation failed, free intermediate pages and 1149 * exit. 1150 */ 1151 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1152 vm_page_unwire(p, 0); 1153 vm_page_free(p); 1154 } 1155 return (NULL); 1156 } 1157 *flags = UMA_SLAB_PRIV; 1158 zkva = keg->uk_kva + 1159 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1160 retkva = zkva; 1161 TAILQ_FOREACH(p, &alloctail, listq) { 1162 pmap_qenter(zkva, &p, 1); 1163 zkva += PAGE_SIZE; 1164 } 1165 1166 return ((void *)retkva); 1167} 1168 1169/* 1170 * Frees a number of pages to the system 1171 * 1172 * Arguments: 1173 * mem A pointer to the memory to be freed 1174 * size The size of the memory being freed 1175 * flags The original p->us_flags field 1176 * 1177 * Returns: 1178 * Nothing 1179 */ 1180static void 1181page_free(void *mem, int size, uint8_t flags) 1182{ 1183 struct vmem *vmem; 1184 1185 if (flags & UMA_SLAB_KMEM) 1186 vmem = kmem_arena; 1187 else if (flags & UMA_SLAB_KERNEL) 1188 vmem = kernel_arena; 1189 else 1190 panic("UMA: page_free used with invalid flags %d", flags); 1191 1192 kmem_free(vmem, (vm_offset_t)mem, size); 1193} 1194 1195/* 1196 * Zero fill initializer 1197 * 1198 * Arguments/Returns follow uma_init specifications 1199 */ 1200static int 1201zero_init(void *mem, int size, int flags) 1202{ 1203 bzero(mem, size); 1204 return (0); 1205} 1206 1207/* 1208 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1209 * 1210 * Arguments 1211 * keg The zone we should initialize 1212 * 1213 * Returns 1214 * Nothing 1215 */ 1216static void 1217keg_small_init(uma_keg_t keg) 1218{ 1219 u_int rsize; 1220 u_int memused; 1221 u_int wastedspace; 1222 u_int shsize; 1223 1224 if (keg->uk_flags & UMA_ZONE_PCPU) { 1225 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU; 1226 1227 keg->uk_slabsize = sizeof(struct pcpu); 1228 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu), 1229 PAGE_SIZE); 1230 } else { 1231 keg->uk_slabsize = UMA_SLAB_SIZE; 1232 keg->uk_ppera = 1; 1233 } 1234 1235 /* 1236 * Calculate the size of each allocation (rsize) according to 1237 * alignment. If the requested size is smaller than we have 1238 * allocation bits for we round it up. 1239 */ 1240 rsize = keg->uk_size; 1241 if (rsize < keg->uk_slabsize / SLAB_SETSIZE) 1242 rsize = keg->uk_slabsize / SLAB_SETSIZE; 1243 if (rsize & keg->uk_align) 1244 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1245 keg->uk_rsize = rsize; 1246 1247 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1248 keg->uk_rsize < sizeof(struct pcpu), 1249 ("%s: size %u too large", __func__, keg->uk_rsize)); 1250 1251 if (keg->uk_flags & UMA_ZONE_REFCNT) 1252 rsize += sizeof(uint32_t); 1253 1254 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1255 shsize = 0; 1256 else 1257 shsize = sizeof(struct uma_slab); 1258 1259 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize; 1260 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1261 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1262 1263 memused = keg->uk_ipers * rsize + shsize; 1264 wastedspace = keg->uk_slabsize - memused; 1265 1266 /* 1267 * We can't do OFFPAGE if we're internal or if we've been 1268 * asked to not go to the VM for buckets. If we do this we 1269 * may end up going to the VM for slabs which we do not 1270 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1271 * of UMA_ZONE_VM, which clearly forbids it. 1272 */ 1273 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1274 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1275 return; 1276 1277 /* 1278 * See if using an OFFPAGE slab will limit our waste. Only do 1279 * this if it permits more items per-slab. 1280 * 1281 * XXX We could try growing slabsize to limit max waste as well. 1282 * Historically this was not done because the VM could not 1283 * efficiently handle contiguous allocations. 1284 */ 1285 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) && 1286 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) { 1287 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize; 1288 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1289 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1290#ifdef UMA_DEBUG 1291 printf("UMA decided we need offpage slab headers for " 1292 "keg: %s, calculated wastedspace = %d, " 1293 "maximum wasted space allowed = %d, " 1294 "calculated ipers = %d, " 1295 "new wasted space = %d\n", keg->uk_name, wastedspace, 1296 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1297 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize); 1298#endif 1299 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1300 } 1301 1302 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1303 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1304 keg->uk_flags |= UMA_ZONE_HASH; 1305} 1306 1307/* 1308 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1309 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1310 * more complicated. 1311 * 1312 * Arguments 1313 * keg The keg we should initialize 1314 * 1315 * Returns 1316 * Nothing 1317 */ 1318static void 1319keg_large_init(uma_keg_t keg) 1320{ 1321 1322 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1323 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1324 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1325 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1326 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1327 1328 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1329 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE; 1330 keg->uk_ipers = 1; 1331 keg->uk_rsize = keg->uk_size; 1332 1333 /* We can't do OFFPAGE if we're internal, bail out here. */ 1334 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) 1335 return; 1336 1337 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1338 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1339 keg->uk_flags |= UMA_ZONE_HASH; 1340} 1341 1342static void 1343keg_cachespread_init(uma_keg_t keg) 1344{ 1345 int alignsize; 1346 int trailer; 1347 int pages; 1348 int rsize; 1349 1350 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1351 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1352 1353 alignsize = keg->uk_align + 1; 1354 rsize = keg->uk_size; 1355 /* 1356 * We want one item to start on every align boundary in a page. To 1357 * do this we will span pages. We will also extend the item by the 1358 * size of align if it is an even multiple of align. Otherwise, it 1359 * would fall on the same boundary every time. 1360 */ 1361 if (rsize & keg->uk_align) 1362 rsize = (rsize & ~keg->uk_align) + alignsize; 1363 if ((rsize & alignsize) == 0) 1364 rsize += alignsize; 1365 trailer = rsize - keg->uk_size; 1366 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1367 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1368 keg->uk_rsize = rsize; 1369 keg->uk_ppera = pages; 1370 keg->uk_slabsize = UMA_SLAB_SIZE; 1371 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1372 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1373 KASSERT(keg->uk_ipers <= uma_max_ipers, 1374 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1375 keg->uk_ipers)); 1376} 1377 1378/* 1379 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1380 * the keg onto the global keg list. 1381 * 1382 * Arguments/Returns follow uma_ctor specifications 1383 * udata Actually uma_kctor_args 1384 */ 1385static int 1386keg_ctor(void *mem, int size, void *udata, int flags) 1387{ 1388 struct uma_kctor_args *arg = udata; 1389 uma_keg_t keg = mem; 1390 uma_zone_t zone; 1391 1392 bzero(keg, size); 1393 keg->uk_size = arg->size; 1394 keg->uk_init = arg->uminit; 1395 keg->uk_fini = arg->fini; 1396 keg->uk_align = arg->align; 1397 keg->uk_free = 0; 1398 keg->uk_reserve = 0; 1399 keg->uk_pages = 0; 1400 keg->uk_flags = arg->flags; 1401 keg->uk_allocf = page_alloc; 1402 keg->uk_freef = page_free; 1403 keg->uk_slabzone = NULL; 1404 1405 /* 1406 * The master zone is passed to us at keg-creation time. 1407 */ 1408 zone = arg->zone; 1409 keg->uk_name = zone->uz_name; 1410 1411 if (arg->flags & UMA_ZONE_VM) 1412 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1413 1414 if (arg->flags & UMA_ZONE_ZINIT) 1415 keg->uk_init = zero_init; 1416 1417 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC) 1418 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1419 1420 if (arg->flags & UMA_ZONE_PCPU) 1421#ifdef SMP 1422 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1423#else 1424 keg->uk_flags &= ~UMA_ZONE_PCPU; 1425#endif 1426 1427 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1428 keg_cachespread_init(keg); 1429 } else if (keg->uk_flags & UMA_ZONE_REFCNT) { 1430 if (keg->uk_size > 1431 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - 1432 sizeof(uint32_t))) 1433 keg_large_init(keg); 1434 else 1435 keg_small_init(keg); 1436 } else { 1437 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab))) 1438 keg_large_init(keg); 1439 else 1440 keg_small_init(keg); 1441 } 1442 1443 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1444 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1445 if (keg->uk_ipers > uma_max_ipers_ref) 1446 panic("Too many ref items per zone: %d > %d\n", 1447 keg->uk_ipers, uma_max_ipers_ref); 1448 keg->uk_slabzone = slabrefzone; 1449 } else 1450 keg->uk_slabzone = slabzone; 1451 } 1452 1453 /* 1454 * If we haven't booted yet we need allocations to go through the 1455 * startup cache until the vm is ready. 1456 */ 1457 if (keg->uk_ppera == 1) { 1458#ifdef UMA_MD_SMALL_ALLOC 1459 keg->uk_allocf = uma_small_alloc; 1460 keg->uk_freef = uma_small_free; 1461 1462 if (booted < UMA_STARTUP) 1463 keg->uk_allocf = startup_alloc; 1464#else 1465 if (booted < UMA_STARTUP2) 1466 keg->uk_allocf = startup_alloc; 1467#endif 1468 } else if (booted < UMA_STARTUP2 && 1469 (keg->uk_flags & UMA_ZFLAG_INTERNAL)) 1470 keg->uk_allocf = startup_alloc; 1471 1472 /* 1473 * Initialize keg's lock 1474 */ 1475 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1476 1477 /* 1478 * If we're putting the slab header in the actual page we need to 1479 * figure out where in each page it goes. This calculates a right 1480 * justified offset into the memory on an ALIGN_PTR boundary. 1481 */ 1482 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1483 u_int totsize; 1484 1485 /* Size of the slab struct and free list */ 1486 totsize = sizeof(struct uma_slab); 1487 1488 /* Size of the reference counts. */ 1489 if (keg->uk_flags & UMA_ZONE_REFCNT) 1490 totsize += keg->uk_ipers * sizeof(uint32_t); 1491 1492 if (totsize & UMA_ALIGN_PTR) 1493 totsize = (totsize & ~UMA_ALIGN_PTR) + 1494 (UMA_ALIGN_PTR + 1); 1495 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1496 1497 /* 1498 * The only way the following is possible is if with our 1499 * UMA_ALIGN_PTR adjustments we are now bigger than 1500 * UMA_SLAB_SIZE. I haven't checked whether this is 1501 * mathematically possible for all cases, so we make 1502 * sure here anyway. 1503 */ 1504 totsize = keg->uk_pgoff + sizeof(struct uma_slab); 1505 if (keg->uk_flags & UMA_ZONE_REFCNT) 1506 totsize += keg->uk_ipers * sizeof(uint32_t); 1507 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1508 printf("zone %s ipers %d rsize %d size %d\n", 1509 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1510 keg->uk_size); 1511 panic("UMA slab won't fit."); 1512 } 1513 } 1514 1515 if (keg->uk_flags & UMA_ZONE_HASH) 1516 hash_alloc(&keg->uk_hash); 1517 1518#ifdef UMA_DEBUG 1519 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n", 1520 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags, 1521 keg->uk_ipers, keg->uk_ppera, 1522 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free); 1523#endif 1524 1525 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1526 1527 mtx_lock(&uma_mtx); 1528 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1529 mtx_unlock(&uma_mtx); 1530 return (0); 1531} 1532 1533/* 1534 * Zone header ctor. This initializes all fields, locks, etc. 1535 * 1536 * Arguments/Returns follow uma_ctor specifications 1537 * udata Actually uma_zctor_args 1538 */ 1539static int 1540zone_ctor(void *mem, int size, void *udata, int flags) 1541{ 1542 struct uma_zctor_args *arg = udata; 1543 uma_zone_t zone = mem; 1544 uma_zone_t z; 1545 uma_keg_t keg; 1546 1547 bzero(zone, size); 1548 zone->uz_name = arg->name; 1549 zone->uz_ctor = arg->ctor; 1550 zone->uz_dtor = arg->dtor; 1551 zone->uz_slab = zone_fetch_slab; 1552 zone->uz_init = NULL; 1553 zone->uz_fini = NULL; 1554 zone->uz_allocs = 0; 1555 zone->uz_frees = 0; 1556 zone->uz_fails = 0; 1557 zone->uz_sleeps = 0; 1558 zone->uz_count = 0; 1559 zone->uz_count_min = 0; 1560 zone->uz_flags = 0; 1561 zone->uz_warning = NULL; 1562 timevalclear(&zone->uz_ratecheck); 1563 keg = arg->keg; 1564 1565 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 1566 1567 /* 1568 * This is a pure cache zone, no kegs. 1569 */ 1570 if (arg->import) { 1571 if (arg->flags & UMA_ZONE_VM) 1572 arg->flags |= UMA_ZFLAG_CACHEONLY; 1573 zone->uz_flags = arg->flags; 1574 zone->uz_size = arg->size; 1575 zone->uz_import = arg->import; 1576 zone->uz_release = arg->release; 1577 zone->uz_arg = arg->arg; 1578 zone->uz_lockptr = &zone->uz_lock; 1579 goto out; 1580 } 1581 1582 /* 1583 * Use the regular zone/keg/slab allocator. 1584 */ 1585 zone->uz_import = (uma_import)zone_import; 1586 zone->uz_release = (uma_release)zone_release; 1587 zone->uz_arg = zone; 1588 1589 if (arg->flags & UMA_ZONE_SECONDARY) { 1590 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1591 zone->uz_init = arg->uminit; 1592 zone->uz_fini = arg->fini; 1593 zone->uz_lockptr = &keg->uk_lock; 1594 zone->uz_flags |= UMA_ZONE_SECONDARY; 1595 mtx_lock(&uma_mtx); 1596 ZONE_LOCK(zone); 1597 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1598 if (LIST_NEXT(z, uz_link) == NULL) { 1599 LIST_INSERT_AFTER(z, zone, uz_link); 1600 break; 1601 } 1602 } 1603 ZONE_UNLOCK(zone); 1604 mtx_unlock(&uma_mtx); 1605 } else if (keg == NULL) { 1606 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1607 arg->align, arg->flags)) == NULL) 1608 return (ENOMEM); 1609 } else { 1610 struct uma_kctor_args karg; 1611 int error; 1612 1613 /* We should only be here from uma_startup() */ 1614 karg.size = arg->size; 1615 karg.uminit = arg->uminit; 1616 karg.fini = arg->fini; 1617 karg.align = arg->align; 1618 karg.flags = arg->flags; 1619 karg.zone = zone; 1620 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1621 flags); 1622 if (error) 1623 return (error); 1624 } 1625 1626 /* 1627 * Link in the first keg. 1628 */ 1629 zone->uz_klink.kl_keg = keg; 1630 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1631 zone->uz_lockptr = &keg->uk_lock; 1632 zone->uz_size = keg->uk_size; 1633 zone->uz_flags |= (keg->uk_flags & 1634 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1635 1636 /* 1637 * Some internal zones don't have room allocated for the per cpu 1638 * caches. If we're internal, bail out here. 1639 */ 1640 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1641 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1642 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1643 return (0); 1644 } 1645 1646out: 1647 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0) 1648 zone->uz_count = bucket_select(zone->uz_size); 1649 else 1650 zone->uz_count = BUCKET_MAX; 1651 zone->uz_count_min = zone->uz_count; 1652 1653 return (0); 1654} 1655 1656/* 1657 * Keg header dtor. This frees all data, destroys locks, frees the hash 1658 * table and removes the keg from the global list. 1659 * 1660 * Arguments/Returns follow uma_dtor specifications 1661 * udata unused 1662 */ 1663static void 1664keg_dtor(void *arg, int size, void *udata) 1665{ 1666 uma_keg_t keg; 1667 1668 keg = (uma_keg_t)arg; 1669 KEG_LOCK(keg); 1670 if (keg->uk_free != 0) { 1671 printf("Freed UMA keg (%s) was not empty (%d items). " 1672 " Lost %d pages of memory.\n", 1673 keg->uk_name ? keg->uk_name : "", 1674 keg->uk_free, keg->uk_pages); 1675 } 1676 KEG_UNLOCK(keg); 1677 1678 hash_free(&keg->uk_hash); 1679 1680 KEG_LOCK_FINI(keg); 1681} 1682 1683/* 1684 * Zone header dtor. 1685 * 1686 * Arguments/Returns follow uma_dtor specifications 1687 * udata unused 1688 */ 1689static void 1690zone_dtor(void *arg, int size, void *udata) 1691{ 1692 uma_klink_t klink; 1693 uma_zone_t zone; 1694 uma_keg_t keg; 1695 1696 zone = (uma_zone_t)arg; 1697 keg = zone_first_keg(zone); 1698 1699 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1700 cache_drain(zone); 1701 1702 mtx_lock(&uma_mtx); 1703 LIST_REMOVE(zone, uz_link); 1704 mtx_unlock(&uma_mtx); 1705 /* 1706 * XXX there are some races here where 1707 * the zone can be drained but zone lock 1708 * released and then refilled before we 1709 * remove it... we dont care for now 1710 */ 1711 zone_drain_wait(zone, M_WAITOK); 1712 /* 1713 * Unlink all of our kegs. 1714 */ 1715 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1716 klink->kl_keg = NULL; 1717 LIST_REMOVE(klink, kl_link); 1718 if (klink == &zone->uz_klink) 1719 continue; 1720 free(klink, M_TEMP); 1721 } 1722 /* 1723 * We only destroy kegs from non secondary zones. 1724 */ 1725 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1726 mtx_lock(&uma_mtx); 1727 LIST_REMOVE(keg, uk_link); 1728 mtx_unlock(&uma_mtx); 1729 zone_free_item(kegs, keg, NULL, SKIP_NONE); 1730 } 1731 ZONE_LOCK_FINI(zone); 1732} 1733 1734/* 1735 * Traverses every zone in the system and calls a callback 1736 * 1737 * Arguments: 1738 * zfunc A pointer to a function which accepts a zone 1739 * as an argument. 1740 * 1741 * Returns: 1742 * Nothing 1743 */ 1744static void 1745zone_foreach(void (*zfunc)(uma_zone_t)) 1746{ 1747 uma_keg_t keg; 1748 uma_zone_t zone; 1749 1750 mtx_lock(&uma_mtx); 1751 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1752 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1753 zfunc(zone); 1754 } 1755 mtx_unlock(&uma_mtx); 1756} 1757 1758/* Public functions */ 1759/* See uma.h */ 1760void 1761uma_startup(void *bootmem, int boot_pages) 1762{ 1763 struct uma_zctor_args args; 1764 uma_slab_t slab; 1765 u_int slabsize; 1766 int i; 1767 1768#ifdef UMA_DEBUG 1769 printf("Creating uma keg headers zone and keg.\n"); 1770#endif 1771 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF); 1772 1773 /* "manually" create the initial zone */ 1774 memset(&args, 0, sizeof(args)); 1775 args.name = "UMA Kegs"; 1776 args.size = sizeof(struct uma_keg); 1777 args.ctor = keg_ctor; 1778 args.dtor = keg_dtor; 1779 args.uminit = zero_init; 1780 args.fini = NULL; 1781 args.keg = &masterkeg; 1782 args.align = 32 - 1; 1783 args.flags = UMA_ZFLAG_INTERNAL; 1784 /* The initial zone has no Per cpu queues so it's smaller */ 1785 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); 1786 1787#ifdef UMA_DEBUG 1788 printf("Filling boot free list.\n"); 1789#endif 1790 for (i = 0; i < boot_pages; i++) { 1791 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE)); 1792 slab->us_data = (uint8_t *)slab; 1793 slab->us_flags = UMA_SLAB_BOOT; 1794 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); 1795 } 1796 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); 1797 1798#ifdef UMA_DEBUG 1799 printf("Creating uma zone headers zone and keg.\n"); 1800#endif 1801 args.name = "UMA Zones"; 1802 args.size = sizeof(struct uma_zone) + 1803 (sizeof(struct uma_cache) * (mp_maxid + 1)); 1804 args.ctor = zone_ctor; 1805 args.dtor = zone_dtor; 1806 args.uminit = zero_init; 1807 args.fini = NULL; 1808 args.keg = NULL; 1809 args.align = 32 - 1; 1810 args.flags = UMA_ZFLAG_INTERNAL; 1811 /* The initial zone has no Per cpu queues so it's smaller */ 1812 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); 1813 1814#ifdef UMA_DEBUG 1815 printf("Initializing pcpu cache locks.\n"); 1816#endif 1817#ifdef UMA_DEBUG 1818 printf("Creating slab and hash zones.\n"); 1819#endif 1820 1821 /* Now make a zone for slab headers */ 1822 slabzone = uma_zcreate("UMA Slabs", 1823 sizeof(struct uma_slab), 1824 NULL, NULL, NULL, NULL, 1825 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1826 1827 /* 1828 * We also create a zone for the bigger slabs with reference 1829 * counts in them, to accomodate UMA_ZONE_REFCNT zones. 1830 */ 1831 slabsize = sizeof(struct uma_slab_refcnt); 1832 slabsize += uma_max_ipers_ref * sizeof(uint32_t); 1833 slabrefzone = uma_zcreate("UMA RCntSlabs", 1834 slabsize, 1835 NULL, NULL, NULL, NULL, 1836 UMA_ALIGN_PTR, 1837 UMA_ZFLAG_INTERNAL); 1838 1839 hashzone = uma_zcreate("UMA Hash", 1840 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 1841 NULL, NULL, NULL, NULL, 1842 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1843 1844 bucket_init(); 1845 1846 booted = UMA_STARTUP; 1847 1848#ifdef UMA_DEBUG 1849 printf("UMA startup complete.\n"); 1850#endif 1851} 1852 1853/* see uma.h */ 1854void 1855uma_startup2(void) 1856{ 1857 booted = UMA_STARTUP2; 1858 bucket_enable(); 1859#ifdef UMA_DEBUG 1860 printf("UMA startup2 complete.\n"); 1861#endif 1862} 1863 1864/* 1865 * Initialize our callout handle 1866 * 1867 */ 1868 1869static void 1870uma_startup3(void) 1871{ 1872#ifdef UMA_DEBUG 1873 printf("Starting callout.\n"); 1874#endif 1875 callout_init(&uma_callout, CALLOUT_MPSAFE); 1876 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1877#ifdef UMA_DEBUG 1878 printf("UMA startup3 complete.\n"); 1879#endif 1880} 1881 1882static uma_keg_t 1883uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 1884 int align, uint32_t flags) 1885{ 1886 struct uma_kctor_args args; 1887 1888 args.size = size; 1889 args.uminit = uminit; 1890 args.fini = fini; 1891 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 1892 args.flags = flags; 1893 args.zone = zone; 1894 return (zone_alloc_item(kegs, &args, M_WAITOK)); 1895} 1896 1897/* See uma.h */ 1898void 1899uma_set_align(int align) 1900{ 1901 1902 if (align != UMA_ALIGN_CACHE) 1903 uma_align_cache = align; 1904} 1905 1906/* See uma.h */ 1907uma_zone_t 1908uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 1909 uma_init uminit, uma_fini fini, int align, uint32_t flags) 1910 1911{ 1912 struct uma_zctor_args args; 1913 1914 /* This stuff is essential for the zone ctor */ 1915 memset(&args, 0, sizeof(args)); 1916 args.name = name; 1917 args.size = size; 1918 args.ctor = ctor; 1919 args.dtor = dtor; 1920 args.uminit = uminit; 1921 args.fini = fini; 1922 args.align = align; 1923 args.flags = flags; 1924 args.keg = NULL; 1925 1926 return (zone_alloc_item(zones, &args, M_WAITOK)); 1927} 1928 1929/* See uma.h */ 1930uma_zone_t 1931uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 1932 uma_init zinit, uma_fini zfini, uma_zone_t master) 1933{ 1934 struct uma_zctor_args args; 1935 uma_keg_t keg; 1936 1937 keg = zone_first_keg(master); 1938 memset(&args, 0, sizeof(args)); 1939 args.name = name; 1940 args.size = keg->uk_size; 1941 args.ctor = ctor; 1942 args.dtor = dtor; 1943 args.uminit = zinit; 1944 args.fini = zfini; 1945 args.align = keg->uk_align; 1946 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 1947 args.keg = keg; 1948 1949 /* XXX Attaches only one keg of potentially many. */ 1950 return (zone_alloc_item(zones, &args, M_WAITOK)); 1951} 1952 1953/* See uma.h */ 1954uma_zone_t 1955uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 1956 uma_init zinit, uma_fini zfini, uma_import zimport, 1957 uma_release zrelease, void *arg, int flags) 1958{ 1959 struct uma_zctor_args args; 1960 1961 memset(&args, 0, sizeof(args)); 1962 args.name = name; 1963 args.size = size; 1964 args.ctor = ctor; 1965 args.dtor = dtor; 1966 args.uminit = zinit; 1967 args.fini = zfini; 1968 args.import = zimport; 1969 args.release = zrelease; 1970 args.arg = arg; 1971 args.align = 0; 1972 args.flags = flags; 1973 1974 return (zone_alloc_item(zones, &args, M_WAITOK)); 1975} 1976 1977static void 1978zone_lock_pair(uma_zone_t a, uma_zone_t b) 1979{ 1980 if (a < b) { 1981 ZONE_LOCK(a); 1982 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); 1983 } else { 1984 ZONE_LOCK(b); 1985 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); 1986 } 1987} 1988 1989static void 1990zone_unlock_pair(uma_zone_t a, uma_zone_t b) 1991{ 1992 1993 ZONE_UNLOCK(a); 1994 ZONE_UNLOCK(b); 1995} 1996 1997int 1998uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 1999{ 2000 uma_klink_t klink; 2001 uma_klink_t kl; 2002 int error; 2003 2004 error = 0; 2005 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 2006 2007 zone_lock_pair(zone, master); 2008 /* 2009 * zone must use vtoslab() to resolve objects and must already be 2010 * a secondary. 2011 */ 2012 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 2013 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 2014 error = EINVAL; 2015 goto out; 2016 } 2017 /* 2018 * The new master must also use vtoslab(). 2019 */ 2020 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 2021 error = EINVAL; 2022 goto out; 2023 } 2024 /* 2025 * Both must either be refcnt, or not be refcnt. 2026 */ 2027 if ((zone->uz_flags & UMA_ZONE_REFCNT) != 2028 (master->uz_flags & UMA_ZONE_REFCNT)) { 2029 error = EINVAL; 2030 goto out; 2031 } 2032 /* 2033 * The underlying object must be the same size. rsize 2034 * may be different. 2035 */ 2036 if (master->uz_size != zone->uz_size) { 2037 error = E2BIG; 2038 goto out; 2039 } 2040 /* 2041 * Put it at the end of the list. 2042 */ 2043 klink->kl_keg = zone_first_keg(master); 2044 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 2045 if (LIST_NEXT(kl, kl_link) == NULL) { 2046 LIST_INSERT_AFTER(kl, klink, kl_link); 2047 break; 2048 } 2049 } 2050 klink = NULL; 2051 zone->uz_flags |= UMA_ZFLAG_MULTI; 2052 zone->uz_slab = zone_fetch_slab_multi; 2053 2054out: 2055 zone_unlock_pair(zone, master); 2056 if (klink != NULL) 2057 free(klink, M_TEMP); 2058 2059 return (error); 2060} 2061 2062 2063/* See uma.h */ 2064void 2065uma_zdestroy(uma_zone_t zone) 2066{ 2067 2068 zone_free_item(zones, zone, NULL, SKIP_NONE); 2069} 2070 2071/* See uma.h */ 2072void * 2073uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2074{ 2075 void *item; 2076 uma_cache_t cache; 2077 uma_bucket_t bucket; 2078 int lockfail; 2079 int cpu; 2080 2081 /* This is the fast path allocation */ 2082#ifdef UMA_DEBUG_ALLOC_1 2083 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); 2084#endif 2085 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread, 2086 zone->uz_name, flags); 2087 2088 if (flags & M_WAITOK) { 2089 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2090 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2091 } 2092#ifdef DEBUG_MEMGUARD 2093 if (memguard_cmp_zone(zone)) { 2094 item = memguard_alloc(zone->uz_size, flags); 2095 if (item != NULL) { 2096 /* 2097 * Avoid conflict with the use-after-free 2098 * protecting infrastructure from INVARIANTS. 2099 */ 2100 if (zone->uz_init != NULL && 2101 zone->uz_init != mtrash_init && 2102 zone->uz_init(item, zone->uz_size, flags) != 0) 2103 return (NULL); 2104 if (zone->uz_ctor != NULL && 2105 zone->uz_ctor != mtrash_ctor && 2106 zone->uz_ctor(item, zone->uz_size, udata, 2107 flags) != 0) { 2108 zone->uz_fini(item, zone->uz_size); 2109 return (NULL); 2110 } 2111 return (item); 2112 } 2113 /* This is unfortunate but should not be fatal. */ 2114 } 2115#endif 2116 /* 2117 * If possible, allocate from the per-CPU cache. There are two 2118 * requirements for safe access to the per-CPU cache: (1) the thread 2119 * accessing the cache must not be preempted or yield during access, 2120 * and (2) the thread must not migrate CPUs without switching which 2121 * cache it accesses. We rely on a critical section to prevent 2122 * preemption and migration. We release the critical section in 2123 * order to acquire the zone mutex if we are unable to allocate from 2124 * the current cache; when we re-acquire the critical section, we 2125 * must detect and handle migration if it has occurred. 2126 */ 2127 critical_enter(); 2128 cpu = curcpu; 2129 cache = &zone->uz_cpu[cpu]; 2130 2131zalloc_start: 2132 bucket = cache->uc_allocbucket; 2133 if (bucket != NULL && bucket->ub_cnt > 0) { 2134 bucket->ub_cnt--; 2135 item = bucket->ub_bucket[bucket->ub_cnt]; 2136#ifdef INVARIANTS 2137 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2138#endif 2139 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2140 cache->uc_allocs++; 2141 critical_exit(); 2142 if (zone->uz_ctor != NULL && 2143 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2144 atomic_add_long(&zone->uz_fails, 1); 2145 zone_free_item(zone, item, udata, SKIP_DTOR); 2146 return (NULL); 2147 } 2148#ifdef INVARIANTS 2149 uma_dbg_alloc(zone, NULL, item); 2150#endif 2151 if (flags & M_ZERO) 2152 bzero(item, zone->uz_size); 2153 return (item); 2154 } 2155 2156 /* 2157 * We have run out of items in our alloc bucket. 2158 * See if we can switch with our free bucket. 2159 */ 2160 bucket = cache->uc_freebucket; 2161 if (bucket != NULL && bucket->ub_cnt > 0) { 2162#ifdef UMA_DEBUG_ALLOC 2163 printf("uma_zalloc: Swapping empty with alloc.\n"); 2164#endif 2165 cache->uc_freebucket = cache->uc_allocbucket; 2166 cache->uc_allocbucket = bucket; 2167 goto zalloc_start; 2168 } 2169 2170 /* 2171 * Discard any empty allocation bucket while we hold no locks. 2172 */ 2173 bucket = cache->uc_allocbucket; 2174 cache->uc_allocbucket = NULL; 2175 critical_exit(); 2176 if (bucket != NULL) 2177 bucket_free(zone, bucket, udata); 2178 2179 /* Short-circuit for zones without buckets and low memory. */ 2180 if (zone->uz_count == 0 || bucketdisable) 2181 goto zalloc_item; 2182 2183 /* 2184 * Attempt to retrieve the item from the per-CPU cache has failed, so 2185 * we must go back to the zone. This requires the zone lock, so we 2186 * must drop the critical section, then re-acquire it when we go back 2187 * to the cache. Since the critical section is released, we may be 2188 * preempted or migrate. As such, make sure not to maintain any 2189 * thread-local state specific to the cache from prior to releasing 2190 * the critical section. 2191 */ 2192 lockfail = 0; 2193 if (ZONE_TRYLOCK(zone) == 0) { 2194 /* Record contention to size the buckets. */ 2195 ZONE_LOCK(zone); 2196 lockfail = 1; 2197 } 2198 critical_enter(); 2199 cpu = curcpu; 2200 cache = &zone->uz_cpu[cpu]; 2201 2202 /* 2203 * Since we have locked the zone we may as well send back our stats. 2204 */ 2205 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2206 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2207 cache->uc_allocs = 0; 2208 cache->uc_frees = 0; 2209 2210 /* See if we lost the race to fill the cache. */ 2211 if (cache->uc_allocbucket != NULL) { 2212 ZONE_UNLOCK(zone); 2213 goto zalloc_start; 2214 } 2215 2216 /* 2217 * Check the zone's cache of buckets. 2218 */ 2219 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 2220 KASSERT(bucket->ub_cnt != 0, 2221 ("uma_zalloc_arg: Returning an empty bucket.")); 2222 2223 LIST_REMOVE(bucket, ub_link); 2224 cache->uc_allocbucket = bucket; 2225 ZONE_UNLOCK(zone); 2226 goto zalloc_start; 2227 } 2228 /* We are no longer associated with this CPU. */ 2229 critical_exit(); 2230 2231 /* 2232 * We bump the uz count when the cache size is insufficient to 2233 * handle the working set. 2234 */ 2235 if (lockfail && zone->uz_count < BUCKET_MAX) 2236 zone->uz_count++; 2237 ZONE_UNLOCK(zone); 2238 2239 /* 2240 * Now lets just fill a bucket and put it on the free list. If that 2241 * works we'll restart the allocation from the begining and it 2242 * will use the just filled bucket. 2243 */ 2244 bucket = zone_alloc_bucket(zone, udata, flags); 2245 if (bucket != NULL) { 2246 ZONE_LOCK(zone); 2247 critical_enter(); 2248 cpu = curcpu; 2249 cache = &zone->uz_cpu[cpu]; 2250 /* 2251 * See if we lost the race or were migrated. Cache the 2252 * initialized bucket to make this less likely or claim 2253 * the memory directly. 2254 */ 2255 if (cache->uc_allocbucket == NULL) 2256 cache->uc_allocbucket = bucket; 2257 else 2258 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2259 ZONE_UNLOCK(zone); 2260 goto zalloc_start; 2261 } 2262 2263 /* 2264 * We may not be able to get a bucket so return an actual item. 2265 */ 2266#ifdef UMA_DEBUG 2267 printf("uma_zalloc_arg: Bucketzone returned NULL\n"); 2268#endif 2269 2270zalloc_item: 2271 item = zone_alloc_item(zone, udata, flags); 2272 2273 return (item); 2274} 2275 2276static uma_slab_t 2277keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) 2278{ 2279 uma_slab_t slab; 2280 int reserve; 2281 2282 mtx_assert(&keg->uk_lock, MA_OWNED); 2283 slab = NULL; 2284 reserve = 0; 2285 if ((flags & M_USE_RESERVE) == 0) 2286 reserve = keg->uk_reserve; 2287 2288 for (;;) { 2289 /* 2290 * Find a slab with some space. Prefer slabs that are partially 2291 * used over those that are totally full. This helps to reduce 2292 * fragmentation. 2293 */ 2294 if (keg->uk_free > reserve) { 2295 if (!LIST_EMPTY(&keg->uk_part_slab)) { 2296 slab = LIST_FIRST(&keg->uk_part_slab); 2297 } else { 2298 slab = LIST_FIRST(&keg->uk_free_slab); 2299 LIST_REMOVE(slab, us_link); 2300 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, 2301 us_link); 2302 } 2303 MPASS(slab->us_keg == keg); 2304 return (slab); 2305 } 2306 2307 /* 2308 * M_NOVM means don't ask at all! 2309 */ 2310 if (flags & M_NOVM) 2311 break; 2312 2313 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2314 keg->uk_flags |= UMA_ZFLAG_FULL; 2315 /* 2316 * If this is not a multi-zone, set the FULL bit. 2317 * Otherwise slab_multi() takes care of it. 2318 */ 2319 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2320 zone->uz_flags |= UMA_ZFLAG_FULL; 2321 zone_log_warning(zone); 2322 } 2323 if (flags & M_NOWAIT) 2324 break; 2325 zone->uz_sleeps++; 2326 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2327 continue; 2328 } 2329 slab = keg_alloc_slab(keg, zone, flags); 2330 /* 2331 * If we got a slab here it's safe to mark it partially used 2332 * and return. We assume that the caller is going to remove 2333 * at least one item. 2334 */ 2335 if (slab) { 2336 MPASS(slab->us_keg == keg); 2337 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2338 return (slab); 2339 } 2340 /* 2341 * We might not have been able to get a slab but another cpu 2342 * could have while we were unlocked. Check again before we 2343 * fail. 2344 */ 2345 flags |= M_NOVM; 2346 } 2347 return (slab); 2348} 2349 2350static uma_slab_t 2351zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) 2352{ 2353 uma_slab_t slab; 2354 2355 if (keg == NULL) { 2356 keg = zone_first_keg(zone); 2357 KEG_LOCK(keg); 2358 } 2359 2360 for (;;) { 2361 slab = keg_fetch_slab(keg, zone, flags); 2362 if (slab) 2363 return (slab); 2364 if (flags & (M_NOWAIT | M_NOVM)) 2365 break; 2366 } 2367 KEG_UNLOCK(keg); 2368 return (NULL); 2369} 2370 2371/* 2372 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2373 * with the keg locked. On NULL no lock is held. 2374 * 2375 * The last pointer is used to seed the search. It is not required. 2376 */ 2377static uma_slab_t 2378zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) 2379{ 2380 uma_klink_t klink; 2381 uma_slab_t slab; 2382 uma_keg_t keg; 2383 int flags; 2384 int empty; 2385 int full; 2386 2387 /* 2388 * Don't wait on the first pass. This will skip limit tests 2389 * as well. We don't want to block if we can find a provider 2390 * without blocking. 2391 */ 2392 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2393 /* 2394 * Use the last slab allocated as a hint for where to start 2395 * the search. 2396 */ 2397 if (last != NULL) { 2398 slab = keg_fetch_slab(last, zone, flags); 2399 if (slab) 2400 return (slab); 2401 KEG_UNLOCK(last); 2402 } 2403 /* 2404 * Loop until we have a slab incase of transient failures 2405 * while M_WAITOK is specified. I'm not sure this is 100% 2406 * required but we've done it for so long now. 2407 */ 2408 for (;;) { 2409 empty = 0; 2410 full = 0; 2411 /* 2412 * Search the available kegs for slabs. Be careful to hold the 2413 * correct lock while calling into the keg layer. 2414 */ 2415 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2416 keg = klink->kl_keg; 2417 KEG_LOCK(keg); 2418 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2419 slab = keg_fetch_slab(keg, zone, flags); 2420 if (slab) 2421 return (slab); 2422 } 2423 if (keg->uk_flags & UMA_ZFLAG_FULL) 2424 full++; 2425 else 2426 empty++; 2427 KEG_UNLOCK(keg); 2428 } 2429 if (rflags & (M_NOWAIT | M_NOVM)) 2430 break; 2431 flags = rflags; 2432 /* 2433 * All kegs are full. XXX We can't atomically check all kegs 2434 * and sleep so just sleep for a short period and retry. 2435 */ 2436 if (full && !empty) { 2437 ZONE_LOCK(zone); 2438 zone->uz_flags |= UMA_ZFLAG_FULL; 2439 zone->uz_sleeps++; 2440 zone_log_warning(zone); 2441 msleep(zone, zone->uz_lockptr, PVM, 2442 "zonelimit", hz/100); 2443 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2444 ZONE_UNLOCK(zone); 2445 continue; 2446 } 2447 } 2448 return (NULL); 2449} 2450 2451static void * 2452slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 2453{ 2454 void *item; 2455 uint8_t freei; 2456 2457 MPASS(keg == slab->us_keg); 2458 mtx_assert(&keg->uk_lock, MA_OWNED); 2459 2460 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; 2461 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); 2462 item = slab->us_data + (keg->uk_rsize * freei); 2463 slab->us_freecount--; 2464 keg->uk_free--; 2465 2466 /* Move this slab to the full list */ 2467 if (slab->us_freecount == 0) { 2468 LIST_REMOVE(slab, us_link); 2469 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); 2470 } 2471 2472 return (item); 2473} 2474 2475static int 2476zone_import(uma_zone_t zone, void **bucket, int max, int flags) 2477{ 2478 uma_slab_t slab; 2479 uma_keg_t keg; 2480 int i; 2481 2482 slab = NULL; 2483 keg = NULL; 2484 /* Try to keep the buckets totally full */ 2485 for (i = 0; i < max; ) { 2486 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL) 2487 break; 2488 keg = slab->us_keg; 2489 while (slab->us_freecount && i < max) { 2490 bucket[i++] = slab_alloc_item(keg, slab); 2491 if (keg->uk_free <= keg->uk_reserve) 2492 break; 2493 } 2494 /* Don't grab more than one slab at a time. */ 2495 flags &= ~M_WAITOK; 2496 flags |= M_NOWAIT; 2497 } 2498 if (slab != NULL) 2499 KEG_UNLOCK(keg); 2500 2501 return i; 2502} 2503 2504static uma_bucket_t 2505zone_alloc_bucket(uma_zone_t zone, void *udata, int flags) 2506{ 2507 uma_bucket_t bucket; 2508 int max; 2509 2510 /* Don't wait for buckets, preserve caller's NOVM setting. */ 2511 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 2512 if (bucket == NULL) 2513 goto out; 2514 2515 max = MIN(bucket->ub_entries, zone->uz_count); 2516 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 2517 max, flags); 2518 2519 /* 2520 * Initialize the memory if necessary. 2521 */ 2522 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 2523 int i; 2524 2525 for (i = 0; i < bucket->ub_cnt; i++) 2526 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2527 flags) != 0) 2528 break; 2529 /* 2530 * If we couldn't initialize the whole bucket, put the 2531 * rest back onto the freelist. 2532 */ 2533 if (i != bucket->ub_cnt) { 2534 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 2535 bucket->ub_cnt - i); 2536#ifdef INVARIANTS 2537 bzero(&bucket->ub_bucket[i], 2538 sizeof(void *) * (bucket->ub_cnt - i)); 2539#endif 2540 bucket->ub_cnt = i; 2541 } 2542 } 2543 2544out: 2545 if (bucket == NULL || bucket->ub_cnt == 0) { 2546 if (bucket != NULL) 2547 bucket_free(zone, bucket, udata); 2548 atomic_add_long(&zone->uz_fails, 1); 2549 return (NULL); 2550 } 2551 2552 return (bucket); 2553} 2554 2555/* 2556 * Allocates a single item from a zone. 2557 * 2558 * Arguments 2559 * zone The zone to alloc for. 2560 * udata The data to be passed to the constructor. 2561 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2562 * 2563 * Returns 2564 * NULL if there is no memory and M_NOWAIT is set 2565 * An item if successful 2566 */ 2567 2568static void * 2569zone_alloc_item(uma_zone_t zone, void *udata, int flags) 2570{ 2571 void *item; 2572 2573 item = NULL; 2574 2575#ifdef UMA_DEBUG_ALLOC 2576 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); 2577#endif 2578 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1) 2579 goto fail; 2580 atomic_add_long(&zone->uz_allocs, 1); 2581 2582 /* 2583 * We have to call both the zone's init (not the keg's init) 2584 * and the zone's ctor. This is because the item is going from 2585 * a keg slab directly to the user, and the user is expecting it 2586 * to be both zone-init'd as well as zone-ctor'd. 2587 */ 2588 if (zone->uz_init != NULL) { 2589 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 2590 zone_free_item(zone, item, udata, SKIP_FINI); 2591 goto fail; 2592 } 2593 } 2594 if (zone->uz_ctor != NULL) { 2595 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2596 zone_free_item(zone, item, udata, SKIP_DTOR); 2597 goto fail; 2598 } 2599 } 2600#ifdef INVARIANTS 2601 uma_dbg_alloc(zone, NULL, item); 2602#endif 2603 if (flags & M_ZERO) 2604 bzero(item, zone->uz_size); 2605 2606 return (item); 2607 2608fail: 2609 atomic_add_long(&zone->uz_fails, 1); 2610 return (NULL); 2611} 2612 2613/* See uma.h */ 2614void 2615uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 2616{ 2617 uma_cache_t cache; 2618 uma_bucket_t bucket; 2619 int lockfail; 2620 int cpu; 2621 2622#ifdef UMA_DEBUG_ALLOC_1 2623 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); 2624#endif 2625 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 2626 zone->uz_name); 2627 2628 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 2629 if (item == NULL) 2630 return; 2631#ifdef DEBUG_MEMGUARD 2632 if (is_memguard_addr(item)) { 2633 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor) 2634 zone->uz_dtor(item, zone->uz_size, udata); 2635 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini) 2636 zone->uz_fini(item, zone->uz_size); 2637 memguard_free(item); 2638 return; 2639 } 2640#endif 2641#ifdef INVARIANTS 2642 if (zone->uz_flags & UMA_ZONE_MALLOC) 2643 uma_dbg_free(zone, udata, item); 2644 else 2645 uma_dbg_free(zone, NULL, item); 2646#endif 2647 if (zone->uz_dtor != NULL) 2648 zone->uz_dtor(item, zone->uz_size, udata); 2649 2650 /* 2651 * The race here is acceptable. If we miss it we'll just have to wait 2652 * a little longer for the limits to be reset. 2653 */ 2654 if (zone->uz_flags & UMA_ZFLAG_FULL) 2655 goto zfree_item; 2656 2657 /* 2658 * If possible, free to the per-CPU cache. There are two 2659 * requirements for safe access to the per-CPU cache: (1) the thread 2660 * accessing the cache must not be preempted or yield during access, 2661 * and (2) the thread must not migrate CPUs without switching which 2662 * cache it accesses. We rely on a critical section to prevent 2663 * preemption and migration. We release the critical section in 2664 * order to acquire the zone mutex if we are unable to free to the 2665 * current cache; when we re-acquire the critical section, we must 2666 * detect and handle migration if it has occurred. 2667 */ 2668zfree_restart: 2669 critical_enter(); 2670 cpu = curcpu; 2671 cache = &zone->uz_cpu[cpu]; 2672 2673zfree_start: 2674 /* 2675 * Try to free into the allocbucket first to give LIFO ordering 2676 * for cache-hot datastructures. Spill over into the freebucket 2677 * if necessary. Alloc will swap them if one runs dry. 2678 */ 2679 bucket = cache->uc_allocbucket; 2680 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 2681 bucket = cache->uc_freebucket; 2682 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2683 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 2684 ("uma_zfree: Freeing to non free bucket index.")); 2685 bucket->ub_bucket[bucket->ub_cnt] = item; 2686 bucket->ub_cnt++; 2687 cache->uc_frees++; 2688 critical_exit(); 2689 return; 2690 } 2691 2692 /* 2693 * We must go back the zone, which requires acquiring the zone lock, 2694 * which in turn means we must release and re-acquire the critical 2695 * section. Since the critical section is released, we may be 2696 * preempted or migrate. As such, make sure not to maintain any 2697 * thread-local state specific to the cache from prior to releasing 2698 * the critical section. 2699 */ 2700 critical_exit(); 2701 if (zone->uz_count == 0 || bucketdisable) 2702 goto zfree_item; 2703 2704 lockfail = 0; 2705 if (ZONE_TRYLOCK(zone) == 0) { 2706 /* Record contention to size the buckets. */ 2707 ZONE_LOCK(zone); 2708 lockfail = 1; 2709 } 2710 critical_enter(); 2711 cpu = curcpu; 2712 cache = &zone->uz_cpu[cpu]; 2713 2714 /* 2715 * Since we have locked the zone we may as well send back our stats. 2716 */ 2717 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2718 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2719 cache->uc_allocs = 0; 2720 cache->uc_frees = 0; 2721 2722 bucket = cache->uc_freebucket; 2723 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2724 ZONE_UNLOCK(zone); 2725 goto zfree_start; 2726 } 2727 cache->uc_freebucket = NULL; 2728 2729 /* Can we throw this on the zone full list? */ 2730 if (bucket != NULL) { 2731#ifdef UMA_DEBUG_ALLOC 2732 printf("uma_zfree: Putting old bucket on the free list.\n"); 2733#endif 2734 /* ub_cnt is pointing to the last free item */ 2735 KASSERT(bucket->ub_cnt != 0, 2736 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 2737 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2738 } 2739 2740 /* We are no longer associated with this CPU. */ 2741 critical_exit(); 2742 2743 /* 2744 * We bump the uz count when the cache size is insufficient to 2745 * handle the working set. 2746 */ 2747 if (lockfail && zone->uz_count < BUCKET_MAX) 2748 zone->uz_count++; 2749 ZONE_UNLOCK(zone); 2750 2751#ifdef UMA_DEBUG_ALLOC 2752 printf("uma_zfree: Allocating new free bucket.\n"); 2753#endif 2754 bucket = bucket_alloc(zone, udata, M_NOWAIT); 2755 if (bucket) { 2756 critical_enter(); 2757 cpu = curcpu; 2758 cache = &zone->uz_cpu[cpu]; 2759 if (cache->uc_freebucket == NULL) { 2760 cache->uc_freebucket = bucket; 2761 goto zfree_start; 2762 } 2763 /* 2764 * We lost the race, start over. We have to drop our 2765 * critical section to free the bucket. 2766 */ 2767 critical_exit(); 2768 bucket_free(zone, bucket, udata); 2769 goto zfree_restart; 2770 } 2771 2772 /* 2773 * If nothing else caught this, we'll just do an internal free. 2774 */ 2775zfree_item: 2776 zone_free_item(zone, item, udata, SKIP_DTOR); 2777 2778 return; 2779} 2780 2781static void 2782slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) 2783{ 2784 uint8_t freei; 2785 2786 mtx_assert(&keg->uk_lock, MA_OWNED); 2787 MPASS(keg == slab->us_keg); 2788 2789 /* Do we need to remove from any lists? */ 2790 if (slab->us_freecount+1 == keg->uk_ipers) { 2791 LIST_REMOVE(slab, us_link); 2792 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 2793 } else if (slab->us_freecount == 0) { 2794 LIST_REMOVE(slab, us_link); 2795 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2796 } 2797 2798 /* Slab management. */ 2799 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 2800 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); 2801 slab->us_freecount++; 2802 2803 /* Keg statistics. */ 2804 keg->uk_free++; 2805} 2806 2807static void 2808zone_release(uma_zone_t zone, void **bucket, int cnt) 2809{ 2810 void *item; 2811 uma_slab_t slab; 2812 uma_keg_t keg; 2813 uint8_t *mem; 2814 int clearfull; 2815 int i; 2816 2817 clearfull = 0; 2818 keg = zone_first_keg(zone); 2819 KEG_LOCK(keg); 2820 for (i = 0; i < cnt; i++) { 2821 item = bucket[i]; 2822 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 2823 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 2824 if (zone->uz_flags & UMA_ZONE_HASH) { 2825 slab = hash_sfind(&keg->uk_hash, mem); 2826 } else { 2827 mem += keg->uk_pgoff; 2828 slab = (uma_slab_t)mem; 2829 } 2830 } else { 2831 slab = vtoslab((vm_offset_t)item); 2832 if (slab->us_keg != keg) { 2833 KEG_UNLOCK(keg); 2834 keg = slab->us_keg; 2835 KEG_LOCK(keg); 2836 } 2837 } 2838 slab_free_item(keg, slab, item); 2839 if (keg->uk_flags & UMA_ZFLAG_FULL) { 2840 if (keg->uk_pages < keg->uk_maxpages) { 2841 keg->uk_flags &= ~UMA_ZFLAG_FULL; 2842 clearfull = 1; 2843 } 2844 2845 /* 2846 * We can handle one more allocation. Since we're 2847 * clearing ZFLAG_FULL, wake up all procs blocked 2848 * on pages. This should be uncommon, so keeping this 2849 * simple for now (rather than adding count of blocked 2850 * threads etc). 2851 */ 2852 wakeup(keg); 2853 } 2854 } 2855 KEG_UNLOCK(keg); 2856 if (clearfull) { 2857 ZONE_LOCK(zone); 2858 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2859 wakeup(zone); 2860 ZONE_UNLOCK(zone); 2861 } 2862 2863} 2864 2865/* 2866 * Frees a single item to any zone. 2867 * 2868 * Arguments: 2869 * zone The zone to free to 2870 * item The item we're freeing 2871 * udata User supplied data for the dtor 2872 * skip Skip dtors and finis 2873 */ 2874static void 2875zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 2876{ 2877 2878#ifdef INVARIANTS 2879 if (skip == SKIP_NONE) { 2880 if (zone->uz_flags & UMA_ZONE_MALLOC) 2881 uma_dbg_free(zone, udata, item); 2882 else 2883 uma_dbg_free(zone, NULL, item); 2884 } 2885#endif 2886 if (skip < SKIP_DTOR && zone->uz_dtor) 2887 zone->uz_dtor(item, zone->uz_size, udata); 2888 2889 if (skip < SKIP_FINI && zone->uz_fini) 2890 zone->uz_fini(item, zone->uz_size); 2891 2892 atomic_add_long(&zone->uz_frees, 1); 2893 zone->uz_release(zone->uz_arg, &item, 1); 2894} 2895 2896/* See uma.h */ 2897int 2898uma_zone_set_max(uma_zone_t zone, int nitems) 2899{ 2900 uma_keg_t keg; 2901 2902 keg = zone_first_keg(zone); 2903 if (keg == NULL) 2904 return (0); 2905 KEG_LOCK(keg); 2906 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 2907 if (keg->uk_maxpages * keg->uk_ipers < nitems) 2908 keg->uk_maxpages += keg->uk_ppera; 2909 nitems = keg->uk_maxpages * keg->uk_ipers; 2910 KEG_UNLOCK(keg); 2911 2912 return (nitems); 2913} 2914 2915/* See uma.h */ 2916int 2917uma_zone_get_max(uma_zone_t zone) 2918{ 2919 int nitems; 2920 uma_keg_t keg; 2921 2922 keg = zone_first_keg(zone); 2923 if (keg == NULL) 2924 return (0); 2925 KEG_LOCK(keg); 2926 nitems = keg->uk_maxpages * keg->uk_ipers; 2927 KEG_UNLOCK(keg); 2928 2929 return (nitems); 2930} 2931 2932/* See uma.h */ 2933void 2934uma_zone_set_warning(uma_zone_t zone, const char *warning) 2935{ 2936 2937 ZONE_LOCK(zone); 2938 zone->uz_warning = warning; 2939 ZONE_UNLOCK(zone); 2940} 2941 2942/* See uma.h */ 2943int 2944uma_zone_get_cur(uma_zone_t zone) 2945{ 2946 int64_t nitems; 2947 u_int i; 2948 2949 ZONE_LOCK(zone); 2950 nitems = zone->uz_allocs - zone->uz_frees; 2951 CPU_FOREACH(i) { 2952 /* 2953 * See the comment in sysctl_vm_zone_stats() regarding the 2954 * safety of accessing the per-cpu caches. With the zone lock 2955 * held, it is safe, but can potentially result in stale data. 2956 */ 2957 nitems += zone->uz_cpu[i].uc_allocs - 2958 zone->uz_cpu[i].uc_frees; 2959 } 2960 ZONE_UNLOCK(zone); 2961 2962 return (nitems < 0 ? 0 : nitems); 2963} 2964 2965/* See uma.h */ 2966void 2967uma_zone_set_init(uma_zone_t zone, uma_init uminit) 2968{ 2969 uma_keg_t keg; 2970 2971 keg = zone_first_keg(zone); 2972 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 2973 KEG_LOCK(keg); 2974 KASSERT(keg->uk_pages == 0, 2975 ("uma_zone_set_init on non-empty keg")); 2976 keg->uk_init = uminit; 2977 KEG_UNLOCK(keg); 2978} 2979 2980/* See uma.h */ 2981void 2982uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 2983{ 2984 uma_keg_t keg; 2985 2986 keg = zone_first_keg(zone); 2987 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 2988 KEG_LOCK(keg); 2989 KASSERT(keg->uk_pages == 0, 2990 ("uma_zone_set_fini on non-empty keg")); 2991 keg->uk_fini = fini; 2992 KEG_UNLOCK(keg); 2993} 2994 2995/* See uma.h */ 2996void 2997uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 2998{ 2999 3000 ZONE_LOCK(zone); 3001 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3002 ("uma_zone_set_zinit on non-empty keg")); 3003 zone->uz_init = zinit; 3004 ZONE_UNLOCK(zone); 3005} 3006 3007/* See uma.h */ 3008void 3009uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 3010{ 3011 3012 ZONE_LOCK(zone); 3013 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3014 ("uma_zone_set_zfini on non-empty keg")); 3015 zone->uz_fini = zfini; 3016 ZONE_UNLOCK(zone); 3017} 3018 3019/* See uma.h */ 3020/* XXX uk_freef is not actually used with the zone locked */ 3021void 3022uma_zone_set_freef(uma_zone_t zone, uma_free freef) 3023{ 3024 uma_keg_t keg; 3025 3026 keg = zone_first_keg(zone); 3027 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 3028 KEG_LOCK(keg); 3029 keg->uk_freef = freef; 3030 KEG_UNLOCK(keg); 3031} 3032 3033/* See uma.h */ 3034/* XXX uk_allocf is not actually used with the zone locked */ 3035void 3036uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3037{ 3038 uma_keg_t keg; 3039 3040 keg = zone_first_keg(zone); 3041 KEG_LOCK(keg); 3042 keg->uk_allocf = allocf; 3043 KEG_UNLOCK(keg); 3044} 3045 3046/* See uma.h */ 3047void 3048uma_zone_reserve(uma_zone_t zone, int items) 3049{ 3050 uma_keg_t keg; 3051 3052 keg = zone_first_keg(zone); 3053 if (keg == NULL) 3054 return; 3055 KEG_LOCK(keg); 3056 keg->uk_reserve = items; 3057 KEG_UNLOCK(keg); 3058 3059 return; 3060} 3061 3062/* See uma.h */ 3063int 3064uma_zone_reserve_kva(uma_zone_t zone, int count) 3065{ 3066 uma_keg_t keg; 3067 vm_offset_t kva; 3068 int pages; 3069 3070 keg = zone_first_keg(zone); 3071 if (keg == NULL) 3072 return (0); 3073 pages = count / keg->uk_ipers; 3074 3075 if (pages * keg->uk_ipers < count) 3076 pages++; 3077 3078#ifdef UMA_MD_SMALL_ALLOC 3079 if (keg->uk_ppera > 1) { 3080#else 3081 if (1) { 3082#endif 3083 kva = kva_alloc(pages * UMA_SLAB_SIZE); 3084 if (kva == 0) 3085 return (0); 3086 } else 3087 kva = 0; 3088 KEG_LOCK(keg); 3089 keg->uk_kva = kva; 3090 keg->uk_offset = 0; 3091 keg->uk_maxpages = pages; 3092#ifdef UMA_MD_SMALL_ALLOC 3093 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3094#else 3095 keg->uk_allocf = noobj_alloc; 3096#endif 3097 keg->uk_flags |= UMA_ZONE_NOFREE; 3098 KEG_UNLOCK(keg); 3099 3100 return (1); 3101} 3102 3103/* See uma.h */ 3104void 3105uma_prealloc(uma_zone_t zone, int items) 3106{ 3107 int slabs; 3108 uma_slab_t slab; 3109 uma_keg_t keg; 3110 3111 keg = zone_first_keg(zone); 3112 if (keg == NULL) 3113 return; 3114 KEG_LOCK(keg); 3115 slabs = items / keg->uk_ipers; 3116 if (slabs * keg->uk_ipers < items) 3117 slabs++; 3118 while (slabs > 0) { 3119 slab = keg_alloc_slab(keg, zone, M_WAITOK); 3120 if (slab == NULL) 3121 break; 3122 MPASS(slab->us_keg == keg); 3123 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 3124 slabs--; 3125 } 3126 KEG_UNLOCK(keg); 3127} 3128 3129/* See uma.h */ 3130uint32_t * 3131uma_find_refcnt(uma_zone_t zone, void *item) 3132{ 3133 uma_slabrefcnt_t slabref; 3134 uma_slab_t slab; 3135 uma_keg_t keg; 3136 uint32_t *refcnt; 3137 int idx; 3138 3139 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK)); 3140 slabref = (uma_slabrefcnt_t)slab; 3141 keg = slab->us_keg; 3142 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT, 3143 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT")); 3144 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3145 refcnt = &slabref->us_refcnt[idx]; 3146 return refcnt; 3147} 3148 3149/* See uma.h */ 3150void 3151uma_reclaim(void) 3152{ 3153#ifdef UMA_DEBUG 3154 printf("UMA: vm asked us to release pages!\n"); 3155#endif 3156 bucket_enable(); 3157 zone_foreach(zone_drain); 3158 if (vm_page_count_min()) { 3159 cache_drain_safe(NULL); 3160 zone_foreach(zone_drain); 3161 } 3162 /* 3163 * Some slabs may have been freed but this zone will be visited early 3164 * we visit again so that we can free pages that are empty once other 3165 * zones are drained. We have to do the same for buckets. 3166 */ 3167 zone_drain(slabzone); 3168 zone_drain(slabrefzone); 3169 bucket_zone_drain(); 3170} 3171 3172/* See uma.h */ 3173int 3174uma_zone_exhausted(uma_zone_t zone) 3175{ 3176 int full; 3177 3178 ZONE_LOCK(zone); 3179 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3180 ZONE_UNLOCK(zone); 3181 return (full); 3182} 3183 3184int 3185uma_zone_exhausted_nolock(uma_zone_t zone) 3186{ 3187 return (zone->uz_flags & UMA_ZFLAG_FULL); 3188} 3189 3190void * 3191uma_large_malloc(int size, int wait) 3192{ 3193 void *mem; 3194 uma_slab_t slab; 3195 uint8_t flags; 3196 3197 slab = zone_alloc_item(slabzone, NULL, wait); 3198 if (slab == NULL) 3199 return (NULL); 3200 mem = page_alloc(NULL, size, &flags, wait); 3201 if (mem) { 3202 vsetslab((vm_offset_t)mem, slab); 3203 slab->us_data = mem; 3204 slab->us_flags = flags | UMA_SLAB_MALLOC; 3205 slab->us_size = size; 3206 } else { 3207 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3208 } 3209 3210 return (mem); 3211} 3212 3213void 3214uma_large_free(uma_slab_t slab) 3215{ 3216 3217 page_free(slab->us_data, slab->us_size, slab->us_flags); 3218 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3219} 3220 3221void 3222uma_print_stats(void) 3223{ 3224 zone_foreach(uma_print_zone); 3225} 3226 3227static void 3228slab_print(uma_slab_t slab) 3229{ 3230 printf("slab: keg %p, data %p, freecount %d\n", 3231 slab->us_keg, slab->us_data, slab->us_freecount); 3232} 3233 3234static void 3235cache_print(uma_cache_t cache) 3236{ 3237 printf("alloc: %p(%d), free: %p(%d)\n", 3238 cache->uc_allocbucket, 3239 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3240 cache->uc_freebucket, 3241 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3242} 3243 3244static void 3245uma_print_keg(uma_keg_t keg) 3246{ 3247 uma_slab_t slab; 3248 3249 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3250 "out %d free %d limit %d\n", 3251 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3252 keg->uk_ipers, keg->uk_ppera, 3253 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free, 3254 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3255 printf("Part slabs:\n"); 3256 LIST_FOREACH(slab, &keg->uk_part_slab, us_link) 3257 slab_print(slab); 3258 printf("Free slabs:\n"); 3259 LIST_FOREACH(slab, &keg->uk_free_slab, us_link) 3260 slab_print(slab); 3261 printf("Full slabs:\n"); 3262 LIST_FOREACH(slab, &keg->uk_full_slab, us_link) 3263 slab_print(slab); 3264} 3265 3266void 3267uma_print_zone(uma_zone_t zone) 3268{ 3269 uma_cache_t cache; 3270 uma_klink_t kl; 3271 int i; 3272 3273 printf("zone: %s(%p) size %d flags %#x\n", 3274 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3275 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3276 uma_print_keg(kl->kl_keg); 3277 CPU_FOREACH(i) { 3278 cache = &zone->uz_cpu[i]; 3279 printf("CPU %d Cache:\n", i); 3280 cache_print(cache); 3281 } 3282} 3283 3284#ifdef DDB 3285/* 3286 * Generate statistics across both the zone and its per-cpu cache's. Return 3287 * desired statistics if the pointer is non-NULL for that statistic. 3288 * 3289 * Note: does not update the zone statistics, as it can't safely clear the 3290 * per-CPU cache statistic. 3291 * 3292 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3293 * safe from off-CPU; we should modify the caches to track this information 3294 * directly so that we don't have to. 3295 */ 3296static void 3297uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3298 uint64_t *freesp, uint64_t *sleepsp) 3299{ 3300 uma_cache_t cache; 3301 uint64_t allocs, frees, sleeps; 3302 int cachefree, cpu; 3303 3304 allocs = frees = sleeps = 0; 3305 cachefree = 0; 3306 CPU_FOREACH(cpu) { 3307 cache = &z->uz_cpu[cpu]; 3308 if (cache->uc_allocbucket != NULL) 3309 cachefree += cache->uc_allocbucket->ub_cnt; 3310 if (cache->uc_freebucket != NULL) 3311 cachefree += cache->uc_freebucket->ub_cnt; 3312 allocs += cache->uc_allocs; 3313 frees += cache->uc_frees; 3314 } 3315 allocs += z->uz_allocs; 3316 frees += z->uz_frees; 3317 sleeps += z->uz_sleeps; 3318 if (cachefreep != NULL) 3319 *cachefreep = cachefree; 3320 if (allocsp != NULL) 3321 *allocsp = allocs; 3322 if (freesp != NULL) 3323 *freesp = frees; 3324 if (sleepsp != NULL) 3325 *sleepsp = sleeps; 3326} 3327#endif /* DDB */ 3328 3329static int 3330sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3331{ 3332 uma_keg_t kz; 3333 uma_zone_t z; 3334 int count; 3335 3336 count = 0; 3337 mtx_lock(&uma_mtx); 3338 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3339 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3340 count++; 3341 } 3342 mtx_unlock(&uma_mtx); 3343 return (sysctl_handle_int(oidp, &count, 0, req)); 3344} 3345 3346static int 3347sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3348{ 3349 struct uma_stream_header ush; 3350 struct uma_type_header uth; 3351 struct uma_percpu_stat ups; 3352 uma_bucket_t bucket; 3353 struct sbuf sbuf; 3354 uma_cache_t cache; 3355 uma_klink_t kl; 3356 uma_keg_t kz; 3357 uma_zone_t z; 3358 uma_keg_t k; 3359 int count, error, i; 3360 3361 error = sysctl_wire_old_buffer(req, 0); 3362 if (error != 0) 3363 return (error); 3364 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3365 3366 count = 0; 3367 mtx_lock(&uma_mtx); 3368 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3369 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3370 count++; 3371 } 3372 3373 /* 3374 * Insert stream header. 3375 */ 3376 bzero(&ush, sizeof(ush)); 3377 ush.ush_version = UMA_STREAM_VERSION; 3378 ush.ush_maxcpus = (mp_maxid + 1); 3379 ush.ush_count = count; 3380 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3381 3382 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3383 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3384 bzero(&uth, sizeof(uth)); 3385 ZONE_LOCK(z); 3386 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3387 uth.uth_align = kz->uk_align; 3388 uth.uth_size = kz->uk_size; 3389 uth.uth_rsize = kz->uk_rsize; 3390 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3391 k = kl->kl_keg; 3392 uth.uth_maxpages += k->uk_maxpages; 3393 uth.uth_pages += k->uk_pages; 3394 uth.uth_keg_free += k->uk_free; 3395 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3396 * k->uk_ipers; 3397 } 3398 3399 /* 3400 * A zone is secondary is it is not the first entry 3401 * on the keg's zone list. 3402 */ 3403 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3404 (LIST_FIRST(&kz->uk_zones) != z)) 3405 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3406 3407 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3408 uth.uth_zone_free += bucket->ub_cnt; 3409 uth.uth_allocs = z->uz_allocs; 3410 uth.uth_frees = z->uz_frees; 3411 uth.uth_fails = z->uz_fails; 3412 uth.uth_sleeps = z->uz_sleeps; 3413 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 3414 /* 3415 * While it is not normally safe to access the cache 3416 * bucket pointers while not on the CPU that owns the 3417 * cache, we only allow the pointers to be exchanged 3418 * without the zone lock held, not invalidated, so 3419 * accept the possible race associated with bucket 3420 * exchange during monitoring. 3421 */ 3422 for (i = 0; i < (mp_maxid + 1); i++) { 3423 bzero(&ups, sizeof(ups)); 3424 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) 3425 goto skip; 3426 if (CPU_ABSENT(i)) 3427 goto skip; 3428 cache = &z->uz_cpu[i]; 3429 if (cache->uc_allocbucket != NULL) 3430 ups.ups_cache_free += 3431 cache->uc_allocbucket->ub_cnt; 3432 if (cache->uc_freebucket != NULL) 3433 ups.ups_cache_free += 3434 cache->uc_freebucket->ub_cnt; 3435 ups.ups_allocs = cache->uc_allocs; 3436 ups.ups_frees = cache->uc_frees; 3437skip: 3438 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups)); 3439 } 3440 ZONE_UNLOCK(z); 3441 } 3442 } 3443 mtx_unlock(&uma_mtx); 3444 error = sbuf_finish(&sbuf); 3445 sbuf_delete(&sbuf); 3446 return (error); 3447} 3448 3449#ifdef DDB 3450DB_SHOW_COMMAND(uma, db_show_uma) 3451{ 3452 uint64_t allocs, frees, sleeps; 3453 uma_bucket_t bucket; 3454 uma_keg_t kz; 3455 uma_zone_t z; 3456 int cachefree; 3457 3458 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 3459 "Requests", "Sleeps"); 3460 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3461 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3462 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 3463 allocs = z->uz_allocs; 3464 frees = z->uz_frees; 3465 sleeps = z->uz_sleeps; 3466 cachefree = 0; 3467 } else 3468 uma_zone_sumstat(z, &cachefree, &allocs, 3469 &frees, &sleeps); 3470 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 3471 (LIST_FIRST(&kz->uk_zones) != z))) 3472 cachefree += kz->uk_free; 3473 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3474 cachefree += bucket->ub_cnt; 3475 db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name, 3476 (uintmax_t)kz->uk_size, 3477 (intmax_t)(allocs - frees), cachefree, 3478 (uintmax_t)allocs, sleeps); 3479 if (db_pager_quit) 3480 return; 3481 } 3482 } 3483} 3484#endif 3485