1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org> 5 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org> 6 * Copyright (c) 2004-2006 Robert N. M. Watson 7 * All rights reserved. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice unmodified, this list of conditions, and the following 14 * disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 29 */ 30 31 /* 32 * uma_core.c Implementation of the Universal Memory allocator 33 * 34 * This allocator is intended to replace the multitude of similar object caches 35 * in the standard FreeBSD kernel. The intent is to be flexible as well as 36 * efficient. A primary design goal is to return unused memory to the rest of 37 * the system. This will make the system as a whole more flexible due to the 38 * ability to move memory to subsystems which most need it instead of leaving 39 * pools of reserved memory unused. 40 * 41 * The basic ideas stem from similar slab/zone based allocators whose algorithms 42 * are well known. 43 * 44 */ 45 46 /* 47 * TODO: 48 * - Improve memory usage for large allocations 49 * - Investigate cache size adjustments 50 */ 51 52 #include <sys/cdefs.h> 53 __FBSDID("$FreeBSD$"); 54 55 #include "opt_ddb.h" 56 #include "opt_param.h" 57 #include "opt_vm.h" 58 59 #include <sys/param.h> 60 #include <sys/systm.h> 61 #include <sys/bitset.h> 62 #include <sys/eventhandler.h> 63 #include <sys/kernel.h> 64 #include <sys/types.h> 65 #include <sys/limits.h> 66 #include <sys/queue.h> 67 #include <sys/malloc.h> 68 #include <sys/ktr.h> 69 #include <sys/lock.h> 70 #include <sys/sysctl.h> 71 #include <sys/mutex.h> 72 #include <sys/proc.h> 73 #include <sys/random.h> 74 #include <sys/rwlock.h> 75 #include <sys/sbuf.h> 76 #include <sys/sched.h> 77 #include <sys/smp.h> 78 #include <sys/taskqueue.h> 79 #include <sys/vmmeter.h> 80 81 #include <vm/vm.h> 82 #include <vm/vm_object.h> 83 #include <vm/vm_page.h> 84 #include <vm/vm_pageout.h> 85 #include <vm/vm_param.h> 86 #include <vm/vm_phys.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. 102 */ 103 static uma_zone_t kegs; 104 static uma_zone_t zones; 105 106 /* This is the zone from which all offpage uma_slab_ts are allocated. */ 107 static uma_zone_t slabzone; 108 109 /* 110 * The initial hash tables come out of this zone so they can be allocated 111 * prior to malloc coming up. 112 */ 113 static uma_zone_t hashzone; 114 115 /* The boot-time adjusted value for cache line alignment. */ 116 int uma_align_cache = 64 - 1; 117 118 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 119 120 /* 121 * Are we allowed to allocate buckets? 122 */ 123 static int bucketdisable = 1; 124 125 /* Linked list of all kegs in the system */ 126 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 127 128 /* Linked list of all cache-only zones in the system */ 129 static LIST_HEAD(,uma_zone) uma_cachezones = 130 LIST_HEAD_INITIALIZER(uma_cachezones); 131 132 /* This RW lock protects the keg list */ 133 static struct rwlock_padalign __exclusive_cache_line uma_rwlock; 134 135 /* 136 * Pointer and counter to pool of pages, that is preallocated at 137 * startup to bootstrap UMA. 138 */ 139 static char *bootmem; 140 static int boot_pages; 141 142 static struct sx uma_drain_lock; 143 144 /* kmem soft limit. */ 145 static unsigned long uma_kmem_limit = LONG_MAX; 146 static volatile unsigned long uma_kmem_total; 147 148 /* Is the VM done starting up? */ 149 static enum { BOOT_COLD = 0, BOOT_STRAPPED, BOOT_PAGEALLOC, BOOT_BUCKETS, 150 BOOT_RUNNING } booted = BOOT_COLD; 151 152 /* 153 * This is the handle used to schedule events that need to happen 154 * outside of the allocation fast path. 155 */ 156 static struct callout uma_callout; 157 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 158 159 /* 160 * This structure is passed as the zone ctor arg so that I don't have to create 161 * a special allocation function just for zones. 162 */ 163 struct uma_zctor_args { 164 const char *name; 165 size_t size; 166 uma_ctor ctor; 167 uma_dtor dtor; 168 uma_init uminit; 169 uma_fini fini; 170 uma_import import; 171 uma_release release; 172 void *arg; 173 uma_keg_t keg; 174 int align; 175 uint32_t flags; 176 }; 177 178 struct uma_kctor_args { 179 uma_zone_t zone; 180 size_t size; 181 uma_init uminit; 182 uma_fini fini; 183 int align; 184 uint32_t flags; 185 }; 186 187 struct uma_bucket_zone { 188 uma_zone_t ubz_zone; 189 char *ubz_name; 190 int ubz_entries; /* Number of items it can hold. */ 191 int ubz_maxsize; /* Maximum allocation size per-item. */ 192 }; 193 194 /* 195 * Compute the actual number of bucket entries to pack them in power 196 * of two sizes for more efficient space utilization. 197 */ 198 #define BUCKET_SIZE(n) \ 199 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 200 201 #define BUCKET_MAX BUCKET_SIZE(256) 202 203 struct uma_bucket_zone bucket_zones[] = { 204 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 205 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 206 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 207 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 208 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 209 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 210 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 211 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 212 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 213 { NULL, NULL, 0} 214 }; 215 216 /* 217 * Flags and enumerations to be passed to internal functions. 218 */ 219 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; 220 221 #define UMA_ANYDOMAIN -1 /* Special value for domain search. */ 222 223 /* Prototypes.. */ 224 225 int uma_startup_count(int); 226 void uma_startup(void *, int); 227 void uma_startup1(void); 228 void uma_startup2(void); 229 230 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 231 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 232 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 233 static void page_free(void *, vm_size_t, uint8_t); 234 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int); 235 static void cache_drain(uma_zone_t); 236 static void bucket_drain(uma_zone_t, uma_bucket_t); 237 static void bucket_cache_drain(uma_zone_t zone); 238 static int keg_ctor(void *, int, void *, int); 239 static void keg_dtor(void *, int, void *); 240 static int zone_ctor(void *, int, void *, int); 241 static void zone_dtor(void *, int, void *); 242 static int zero_init(void *, int, int); 243 static void keg_small_init(uma_keg_t keg); 244 static void keg_large_init(uma_keg_t keg); 245 static void zone_foreach(void (*zfunc)(uma_zone_t)); 246 static void zone_timeout(uma_zone_t zone); 247 static int hash_alloc(struct uma_hash *); 248 static int hash_expand(struct uma_hash *, struct uma_hash *); 249 static void hash_free(struct uma_hash *hash); 250 static void uma_timeout(void *); 251 static void uma_startup3(void); 252 static void *zone_alloc_item(uma_zone_t, void *, int, int); 253 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 254 static void bucket_enable(void); 255 static void bucket_init(void); 256 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 257 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 258 static void bucket_zone_drain(void); 259 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int); 260 static uma_slab_t zone_fetch_slab(uma_zone_t, uma_keg_t, int, int); 261 static uma_slab_t zone_fetch_slab_multi(uma_zone_t, uma_keg_t, int, int); 262 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 263 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); 264 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 265 uma_fini fini, int align, uint32_t flags); 266 static int zone_import(uma_zone_t, void **, int, int, int); 267 static void zone_release(uma_zone_t, void **, int); 268 static void uma_zero_item(void *, uma_zone_t); 269 270 void uma_print_zone(uma_zone_t); 271 void uma_print_stats(void); 272 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 273 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 274 275 #ifdef INVARIANTS 276 static bool uma_dbg_kskip(uma_keg_t keg, void *mem); 277 static bool uma_dbg_zskip(uma_zone_t zone, void *mem); 278 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item); 279 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item); 280 281 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0, 282 "Memory allocation debugging"); 283 284 static u_int dbg_divisor = 1; 285 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor, 286 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0, 287 "Debug & thrash every this item in memory allocator"); 288 289 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER; 290 static counter_u64_t uma_skip_cnt = EARLY_COUNTER; 291 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD, 292 &uma_dbg_cnt, "memory items debugged"); 293 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD, 294 &uma_skip_cnt, "memory items skipped, not debugged"); 295 #endif 296 297 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 298 299 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 300 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 301 302 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 303 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 304 305 static int zone_warnings = 1; 306 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, 307 "Warn when UMA zones becomes full"); 308 309 /* Adjust bytes under management by UMA. */ 310 static inline void 311 uma_total_dec(unsigned long size) 312 { 313 314 atomic_subtract_long(&uma_kmem_total, size); 315 } 316 317 static inline void 318 uma_total_inc(unsigned long size) 319 { 320 321 if (atomic_fetchadd_long(&uma_kmem_total, size) > uma_kmem_limit) 322 uma_reclaim_wakeup(); 323 } 324 325 /* 326 * This routine checks to see whether or not it's safe to enable buckets. 327 */ 328 static void 329 bucket_enable(void) 330 { 331 bucketdisable = vm_page_count_min(); 332 } 333 334 /* 335 * Initialize bucket_zones, the array of zones of buckets of various sizes. 336 * 337 * For each zone, calculate the memory required for each bucket, consisting 338 * of the header and an array of pointers. 339 */ 340 static void 341 bucket_init(void) 342 { 343 struct uma_bucket_zone *ubz; 344 int size; 345 346 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 347 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 348 size += sizeof(void *) * ubz->ubz_entries; 349 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 350 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 351 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET | UMA_ZONE_NUMA); 352 } 353 } 354 355 /* 356 * Given a desired number of entries for a bucket, return the zone from which 357 * to allocate the bucket. 358 */ 359 static struct uma_bucket_zone * 360 bucket_zone_lookup(int entries) 361 { 362 struct uma_bucket_zone *ubz; 363 364 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 365 if (ubz->ubz_entries >= entries) 366 return (ubz); 367 ubz--; 368 return (ubz); 369 } 370 371 static int 372 bucket_select(int size) 373 { 374 struct uma_bucket_zone *ubz; 375 376 ubz = &bucket_zones[0]; 377 if (size > ubz->ubz_maxsize) 378 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 379 380 for (; ubz->ubz_entries != 0; ubz++) 381 if (ubz->ubz_maxsize < size) 382 break; 383 ubz--; 384 return (ubz->ubz_entries); 385 } 386 387 static uma_bucket_t 388 bucket_alloc(uma_zone_t zone, void *udata, int flags) 389 { 390 struct uma_bucket_zone *ubz; 391 uma_bucket_t bucket; 392 393 /* 394 * This is to stop us from allocating per cpu buckets while we're 395 * running out of vm.boot_pages. Otherwise, we would exhaust the 396 * boot pages. This also prevents us from allocating buckets in 397 * low memory situations. 398 */ 399 if (bucketdisable) 400 return (NULL); 401 /* 402 * To limit bucket recursion we store the original zone flags 403 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 404 * NOVM flag to persist even through deep recursions. We also 405 * store ZFLAG_BUCKET once we have recursed attempting to allocate 406 * a bucket for a bucket zone so we do not allow infinite bucket 407 * recursion. This cookie will even persist to frees of unused 408 * buckets via the allocation path or bucket allocations in the 409 * free path. 410 */ 411 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 412 udata = (void *)(uintptr_t)zone->uz_flags; 413 else { 414 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 415 return (NULL); 416 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 417 } 418 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 419 flags |= M_NOVM; 420 ubz = bucket_zone_lookup(zone->uz_count); 421 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 422 ubz++; 423 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 424 if (bucket) { 425 #ifdef INVARIANTS 426 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 427 #endif 428 bucket->ub_cnt = 0; 429 bucket->ub_entries = ubz->ubz_entries; 430 } 431 432 return (bucket); 433 } 434 435 static void 436 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 437 { 438 struct uma_bucket_zone *ubz; 439 440 KASSERT(bucket->ub_cnt == 0, 441 ("bucket_free: Freeing a non free bucket.")); 442 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 443 udata = (void *)(uintptr_t)zone->uz_flags; 444 ubz = bucket_zone_lookup(bucket->ub_entries); 445 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 446 } 447 448 static void 449 bucket_zone_drain(void) 450 { 451 struct uma_bucket_zone *ubz; 452 453 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 454 zone_drain(ubz->ubz_zone); 455 } 456 457 static void 458 zone_log_warning(uma_zone_t zone) 459 { 460 static const struct timeval warninterval = { 300, 0 }; 461 462 if (!zone_warnings || zone->uz_warning == NULL) 463 return; 464 465 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 466 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 467 } 468 469 static inline void 470 zone_maxaction(uma_zone_t zone) 471 { 472 473 if (zone->uz_maxaction.ta_func != NULL) 474 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); 475 } 476 477 static void 478 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 479 { 480 uma_klink_t klink; 481 482 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 483 kegfn(klink->kl_keg); 484 } 485 486 /* 487 * Routine called by timeout which is used to fire off some time interval 488 * based calculations. (stats, hash size, etc.) 489 * 490 * Arguments: 491 * arg Unused 492 * 493 * Returns: 494 * Nothing 495 */ 496 static void 497 uma_timeout(void *unused) 498 { 499 bucket_enable(); 500 zone_foreach(zone_timeout); 501 502 /* Reschedule this event */ 503 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 504 } 505 506 /* 507 * Routine to perform timeout driven calculations. This expands the 508 * hashes and does per cpu statistics aggregation. 509 * 510 * Returns nothing. 511 */ 512 static void 513 keg_timeout(uma_keg_t keg) 514 { 515 516 KEG_LOCK(keg); 517 /* 518 * Expand the keg hash table. 519 * 520 * This is done if the number of slabs is larger than the hash size. 521 * What I'm trying to do here is completely reduce collisions. This 522 * may be a little aggressive. Should I allow for two collisions max? 523 */ 524 if (keg->uk_flags & UMA_ZONE_HASH && 525 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 526 struct uma_hash newhash; 527 struct uma_hash oldhash; 528 int ret; 529 530 /* 531 * This is so involved because allocating and freeing 532 * while the keg lock is held will lead to deadlock. 533 * I have to do everything in stages and check for 534 * races. 535 */ 536 newhash = keg->uk_hash; 537 KEG_UNLOCK(keg); 538 ret = hash_alloc(&newhash); 539 KEG_LOCK(keg); 540 if (ret) { 541 if (hash_expand(&keg->uk_hash, &newhash)) { 542 oldhash = keg->uk_hash; 543 keg->uk_hash = newhash; 544 } else 545 oldhash = newhash; 546 547 KEG_UNLOCK(keg); 548 hash_free(&oldhash); 549 return; 550 } 551 } 552 KEG_UNLOCK(keg); 553 } 554 555 static void 556 zone_timeout(uma_zone_t zone) 557 { 558 559 zone_foreach_keg(zone, &keg_timeout); 560 } 561 562 /* 563 * Allocate and zero fill the next sized hash table from the appropriate 564 * backing store. 565 * 566 * Arguments: 567 * hash A new hash structure with the old hash size in uh_hashsize 568 * 569 * Returns: 570 * 1 on success and 0 on failure. 571 */ 572 static int 573 hash_alloc(struct uma_hash *hash) 574 { 575 int oldsize; 576 int alloc; 577 578 oldsize = hash->uh_hashsize; 579 580 /* We're just going to go to a power of two greater */ 581 if (oldsize) { 582 hash->uh_hashsize = oldsize * 2; 583 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 584 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 585 M_UMAHASH, M_NOWAIT); 586 } else { 587 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 588 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 589 UMA_ANYDOMAIN, M_WAITOK); 590 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 591 } 592 if (hash->uh_slab_hash) { 593 bzero(hash->uh_slab_hash, alloc); 594 hash->uh_hashmask = hash->uh_hashsize - 1; 595 return (1); 596 } 597 598 return (0); 599 } 600 601 /* 602 * Expands the hash table for HASH zones. This is done from zone_timeout 603 * to reduce collisions. This must not be done in the regular allocation 604 * path, otherwise, we can recurse on the vm while allocating pages. 605 * 606 * Arguments: 607 * oldhash The hash you want to expand 608 * newhash The hash structure for the new table 609 * 610 * Returns: 611 * Nothing 612 * 613 * Discussion: 614 */ 615 static int 616 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 617 { 618 uma_slab_t slab; 619 int hval; 620 int i; 621 622 if (!newhash->uh_slab_hash) 623 return (0); 624 625 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 626 return (0); 627 628 /* 629 * I need to investigate hash algorithms for resizing without a 630 * full rehash. 631 */ 632 633 for (i = 0; i < oldhash->uh_hashsize; i++) 634 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 635 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 636 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 637 hval = UMA_HASH(newhash, slab->us_data); 638 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 639 slab, us_hlink); 640 } 641 642 return (1); 643 } 644 645 /* 646 * Free the hash bucket to the appropriate backing store. 647 * 648 * Arguments: 649 * slab_hash The hash bucket we're freeing 650 * hashsize The number of entries in that hash bucket 651 * 652 * Returns: 653 * Nothing 654 */ 655 static void 656 hash_free(struct uma_hash *hash) 657 { 658 if (hash->uh_slab_hash == NULL) 659 return; 660 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 661 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 662 else 663 free(hash->uh_slab_hash, M_UMAHASH); 664 } 665 666 /* 667 * Frees all outstanding items in a bucket 668 * 669 * Arguments: 670 * zone The zone to free to, must be unlocked. 671 * bucket The free/alloc bucket with items, cpu queue must be locked. 672 * 673 * Returns: 674 * Nothing 675 */ 676 677 static void 678 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 679 { 680 int i; 681 682 if (bucket == NULL) 683 return; 684 685 if (zone->uz_fini) 686 for (i = 0; i < bucket->ub_cnt; i++) 687 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 688 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 689 bucket->ub_cnt = 0; 690 } 691 692 /* 693 * Drains the per cpu caches for a zone. 694 * 695 * NOTE: This may only be called while the zone is being turn down, and not 696 * during normal operation. This is necessary in order that we do not have 697 * to migrate CPUs to drain the per-CPU caches. 698 * 699 * Arguments: 700 * zone The zone to drain, must be unlocked. 701 * 702 * Returns: 703 * Nothing 704 */ 705 static void 706 cache_drain(uma_zone_t zone) 707 { 708 uma_cache_t cache; 709 int cpu; 710 711 /* 712 * XXX: It is safe to not lock the per-CPU caches, because we're 713 * tearing down the zone anyway. I.e., there will be no further use 714 * of the caches at this point. 715 * 716 * XXX: It would good to be able to assert that the zone is being 717 * torn down to prevent improper use of cache_drain(). 718 * 719 * XXX: We lock the zone before passing into bucket_cache_drain() as 720 * it is used elsewhere. Should the tear-down path be made special 721 * there in some form? 722 */ 723 CPU_FOREACH(cpu) { 724 cache = &zone->uz_cpu[cpu]; 725 bucket_drain(zone, cache->uc_allocbucket); 726 bucket_drain(zone, cache->uc_freebucket); 727 if (cache->uc_allocbucket != NULL) 728 bucket_free(zone, cache->uc_allocbucket, NULL); 729 if (cache->uc_freebucket != NULL) 730 bucket_free(zone, cache->uc_freebucket, NULL); 731 cache->uc_allocbucket = cache->uc_freebucket = NULL; 732 } 733 ZONE_LOCK(zone); 734 bucket_cache_drain(zone); 735 ZONE_UNLOCK(zone); 736 } 737 738 static void 739 cache_shrink(uma_zone_t zone) 740 { 741 742 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 743 return; 744 745 ZONE_LOCK(zone); 746 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; 747 ZONE_UNLOCK(zone); 748 } 749 750 static void 751 cache_drain_safe_cpu(uma_zone_t zone) 752 { 753 uma_cache_t cache; 754 uma_bucket_t b1, b2; 755 int domain; 756 757 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 758 return; 759 760 b1 = b2 = NULL; 761 ZONE_LOCK(zone); 762 critical_enter(); 763 if (zone->uz_flags & UMA_ZONE_NUMA) 764 domain = PCPU_GET(domain); 765 else 766 domain = 0; 767 cache = &zone->uz_cpu[curcpu]; 768 if (cache->uc_allocbucket) { 769 if (cache->uc_allocbucket->ub_cnt != 0) 770 LIST_INSERT_HEAD(&zone->uz_domain[domain].uzd_buckets, 771 cache->uc_allocbucket, ub_link); 772 else 773 b1 = cache->uc_allocbucket; 774 cache->uc_allocbucket = NULL; 775 } 776 if (cache->uc_freebucket) { 777 if (cache->uc_freebucket->ub_cnt != 0) 778 LIST_INSERT_HEAD(&zone->uz_domain[domain].uzd_buckets, 779 cache->uc_freebucket, ub_link); 780 else 781 b2 = cache->uc_freebucket; 782 cache->uc_freebucket = NULL; 783 } 784 critical_exit(); 785 ZONE_UNLOCK(zone); 786 if (b1) 787 bucket_free(zone, b1, NULL); 788 if (b2) 789 bucket_free(zone, b2, NULL); 790 } 791 792 /* 793 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 794 * This is an expensive call because it needs to bind to all CPUs 795 * one by one and enter a critical section on each of them in order 796 * to safely access their cache buckets. 797 * Zone lock must not be held on call this function. 798 */ 799 static void 800 cache_drain_safe(uma_zone_t zone) 801 { 802 int cpu; 803 804 /* 805 * Polite bucket sizes shrinking was not enouth, shrink aggressively. 806 */ 807 if (zone) 808 cache_shrink(zone); 809 else 810 zone_foreach(cache_shrink); 811 812 CPU_FOREACH(cpu) { 813 thread_lock(curthread); 814 sched_bind(curthread, cpu); 815 thread_unlock(curthread); 816 817 if (zone) 818 cache_drain_safe_cpu(zone); 819 else 820 zone_foreach(cache_drain_safe_cpu); 821 } 822 thread_lock(curthread); 823 sched_unbind(curthread); 824 thread_unlock(curthread); 825 } 826 827 /* 828 * Drain the cached buckets from a zone. Expects a locked zone on entry. 829 */ 830 static void 831 bucket_cache_drain(uma_zone_t zone) 832 { 833 uma_zone_domain_t zdom; 834 uma_bucket_t bucket; 835 int i; 836 837 /* 838 * Drain the bucket queues and free the buckets. 839 */ 840 for (i = 0; i < vm_ndomains; i++) { 841 zdom = &zone->uz_domain[i]; 842 while ((bucket = LIST_FIRST(&zdom->uzd_buckets)) != NULL) { 843 LIST_REMOVE(bucket, ub_link); 844 ZONE_UNLOCK(zone); 845 bucket_drain(zone, bucket); 846 bucket_free(zone, bucket, NULL); 847 ZONE_LOCK(zone); 848 } 849 } 850 851 /* 852 * Shrink further bucket sizes. Price of single zone lock collision 853 * is probably lower then price of global cache drain. 854 */ 855 if (zone->uz_count > zone->uz_count_min) 856 zone->uz_count--; 857 } 858 859 static void 860 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 861 { 862 uint8_t *mem; 863 int i; 864 uint8_t flags; 865 866 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes", 867 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera); 868 869 mem = slab->us_data; 870 flags = slab->us_flags; 871 i = start; 872 if (keg->uk_fini != NULL) { 873 for (i--; i > -1; i--) 874 #ifdef INVARIANTS 875 /* 876 * trash_fini implies that dtor was trash_dtor. trash_fini 877 * would check that memory hasn't been modified since free, 878 * which executed trash_dtor. 879 * That's why we need to run uma_dbg_kskip() check here, 880 * albeit we don't make skip check for other init/fini 881 * invocations. 882 */ 883 if (!uma_dbg_kskip(keg, slab->us_data + (keg->uk_rsize * i)) || 884 keg->uk_fini != trash_fini) 885 #endif 886 keg->uk_fini(slab->us_data + (keg->uk_rsize * i), 887 keg->uk_size); 888 } 889 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 890 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 891 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 892 uma_total_dec(PAGE_SIZE * keg->uk_ppera); 893 } 894 895 /* 896 * Frees pages from a keg back to the system. This is done on demand from 897 * the pageout daemon. 898 * 899 * Returns nothing. 900 */ 901 static void 902 keg_drain(uma_keg_t keg) 903 { 904 struct slabhead freeslabs = { 0 }; 905 uma_domain_t dom; 906 uma_slab_t slab, tmp; 907 int i; 908 909 /* 910 * We don't want to take pages from statically allocated kegs at this 911 * time 912 */ 913 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 914 return; 915 916 CTR3(KTR_UMA, "keg_drain %s(%p) free items: %u", 917 keg->uk_name, keg, keg->uk_free); 918 KEG_LOCK(keg); 919 if (keg->uk_free == 0) 920 goto finished; 921 922 for (i = 0; i < vm_ndomains; i++) { 923 dom = &keg->uk_domain[i]; 924 LIST_FOREACH_SAFE(slab, &dom->ud_free_slab, us_link, tmp) { 925 /* We have nowhere to free these to. */ 926 if (slab->us_flags & UMA_SLAB_BOOT) 927 continue; 928 929 LIST_REMOVE(slab, us_link); 930 keg->uk_pages -= keg->uk_ppera; 931 keg->uk_free -= keg->uk_ipers; 932 933 if (keg->uk_flags & UMA_ZONE_HASH) 934 UMA_HASH_REMOVE(&keg->uk_hash, slab, 935 slab->us_data); 936 937 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 938 } 939 } 940 941 finished: 942 KEG_UNLOCK(keg); 943 944 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 945 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 946 keg_free_slab(keg, slab, keg->uk_ipers); 947 } 948 } 949 950 static void 951 zone_drain_wait(uma_zone_t zone, int waitok) 952 { 953 954 /* 955 * Set draining to interlock with zone_dtor() so we can release our 956 * locks as we go. Only dtor() should do a WAITOK call since it 957 * is the only call that knows the structure will still be available 958 * when it wakes up. 959 */ 960 ZONE_LOCK(zone); 961 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 962 if (waitok == M_NOWAIT) 963 goto out; 964 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 965 } 966 zone->uz_flags |= UMA_ZFLAG_DRAINING; 967 bucket_cache_drain(zone); 968 ZONE_UNLOCK(zone); 969 /* 970 * The DRAINING flag protects us from being freed while 971 * we're running. Normally the uma_rwlock would protect us but we 972 * must be able to release and acquire the right lock for each keg. 973 */ 974 zone_foreach_keg(zone, &keg_drain); 975 ZONE_LOCK(zone); 976 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 977 wakeup(zone); 978 out: 979 ZONE_UNLOCK(zone); 980 } 981 982 void 983 zone_drain(uma_zone_t zone) 984 { 985 986 zone_drain_wait(zone, M_NOWAIT); 987 } 988 989 /* 990 * Allocate a new slab for a keg. This does not insert the slab onto a list. 991 * 992 * Arguments: 993 * wait Shall we wait? 994 * 995 * Returns: 996 * The slab that was allocated or NULL if there is no memory and the 997 * caller specified M_NOWAIT. 998 */ 999 static uma_slab_t 1000 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int wait) 1001 { 1002 uma_alloc allocf; 1003 uma_slab_t slab; 1004 unsigned long size; 1005 uint8_t *mem; 1006 uint8_t flags; 1007 int i; 1008 1009 KASSERT(domain >= 0 && domain < vm_ndomains, 1010 ("keg_alloc_slab: domain %d out of range", domain)); 1011 mtx_assert(&keg->uk_lock, MA_OWNED); 1012 slab = NULL; 1013 mem = NULL; 1014 1015 allocf = keg->uk_allocf; 1016 KEG_UNLOCK(keg); 1017 size = keg->uk_ppera * PAGE_SIZE; 1018 1019 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1020 slab = zone_alloc_item(keg->uk_slabzone, NULL, domain, wait); 1021 if (slab == NULL) 1022 goto out; 1023 } 1024 1025 /* 1026 * This reproduces the old vm_zone behavior of zero filling pages the 1027 * first time they are added to a zone. 1028 * 1029 * Malloced items are zeroed in uma_zalloc. 1030 */ 1031 1032 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 1033 wait |= M_ZERO; 1034 else 1035 wait &= ~M_ZERO; 1036 1037 if (keg->uk_flags & UMA_ZONE_NODUMP) 1038 wait |= M_NODUMP; 1039 1040 /* zone is passed for legacy reasons. */ 1041 mem = allocf(zone, size, domain, &flags, wait); 1042 if (mem == NULL) { 1043 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1044 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 1045 slab = NULL; 1046 goto out; 1047 } 1048 uma_total_inc(size); 1049 1050 /* Point the slab into the allocated memory */ 1051 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 1052 slab = (uma_slab_t )(mem + keg->uk_pgoff); 1053 1054 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 1055 for (i = 0; i < keg->uk_ppera; i++) 1056 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 1057 1058 slab->us_keg = keg; 1059 slab->us_data = mem; 1060 slab->us_freecount = keg->uk_ipers; 1061 slab->us_flags = flags; 1062 slab->us_domain = domain; 1063 BIT_FILL(SLAB_SETSIZE, &slab->us_free); 1064 #ifdef INVARIANTS 1065 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); 1066 #endif 1067 1068 if (keg->uk_init != NULL) { 1069 for (i = 0; i < keg->uk_ipers; i++) 1070 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 1071 keg->uk_size, wait) != 0) 1072 break; 1073 if (i != keg->uk_ipers) { 1074 keg_free_slab(keg, slab, i); 1075 slab = NULL; 1076 goto out; 1077 } 1078 } 1079 out: 1080 KEG_LOCK(keg); 1081 1082 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)", 1083 slab, keg->uk_name, keg); 1084 1085 if (slab != NULL) { 1086 if (keg->uk_flags & UMA_ZONE_HASH) 1087 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1088 1089 keg->uk_pages += keg->uk_ppera; 1090 keg->uk_free += keg->uk_ipers; 1091 } 1092 1093 return (slab); 1094 } 1095 1096 /* 1097 * This function is intended to be used early on in place of page_alloc() so 1098 * that we may use the boot time page cache to satisfy allocations before 1099 * the VM is ready. 1100 */ 1101 static void * 1102 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1103 int wait) 1104 { 1105 uma_keg_t keg; 1106 void *mem; 1107 int pages; 1108 1109 keg = zone_first_keg(zone); 1110 1111 /* 1112 * If we are in BOOT_BUCKETS or higher, than switch to real 1113 * allocator. Zones with page sized slabs switch at BOOT_PAGEALLOC. 1114 */ 1115 switch (booted) { 1116 case BOOT_COLD: 1117 case BOOT_STRAPPED: 1118 break; 1119 case BOOT_PAGEALLOC: 1120 if (keg->uk_ppera > 1) 1121 break; 1122 case BOOT_BUCKETS: 1123 case BOOT_RUNNING: 1124 #ifdef UMA_MD_SMALL_ALLOC 1125 keg->uk_allocf = (keg->uk_ppera > 1) ? 1126 page_alloc : uma_small_alloc; 1127 #else 1128 keg->uk_allocf = page_alloc; 1129 #endif 1130 return keg->uk_allocf(zone, bytes, domain, pflag, wait); 1131 } 1132 1133 /* 1134 * Check our small startup cache to see if it has pages remaining. 1135 */ 1136 pages = howmany(bytes, PAGE_SIZE); 1137 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__)); 1138 if (pages > boot_pages) 1139 panic("UMA zone \"%s\": Increase vm.boot_pages", zone->uz_name); 1140 #ifdef DIAGNOSTIC 1141 printf("%s from \"%s\", %d boot pages left\n", __func__, zone->uz_name, 1142 boot_pages); 1143 #endif 1144 mem = bootmem; 1145 boot_pages -= pages; 1146 bootmem += pages * PAGE_SIZE; 1147 *pflag = UMA_SLAB_BOOT; 1148 1149 return (mem); 1150 } 1151 1152 /* 1153 * Allocates a number of pages from the system 1154 * 1155 * Arguments: 1156 * bytes The number of bytes requested 1157 * wait Shall we wait? 1158 * 1159 * Returns: 1160 * A pointer to the alloced memory or possibly 1161 * NULL if M_NOWAIT is set. 1162 */ 1163 static void * 1164 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1165 int wait) 1166 { 1167 void *p; /* Returned page */ 1168 1169 *pflag = UMA_SLAB_KERNEL; 1170 p = (void *) kmem_malloc_domain(domain, bytes, wait); 1171 1172 return (p); 1173 } 1174 1175 /* 1176 * Allocates a number of pages from within an object 1177 * 1178 * Arguments: 1179 * bytes The number of bytes requested 1180 * wait Shall we wait? 1181 * 1182 * Returns: 1183 * A pointer to the alloced memory or possibly 1184 * NULL if M_NOWAIT is set. 1185 */ 1186 static void * 1187 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, 1188 int wait) 1189 { 1190 TAILQ_HEAD(, vm_page) alloctail; 1191 u_long npages; 1192 vm_offset_t retkva, zkva; 1193 vm_page_t p, p_next; 1194 uma_keg_t keg; 1195 1196 TAILQ_INIT(&alloctail); 1197 keg = zone_first_keg(zone); 1198 1199 npages = howmany(bytes, PAGE_SIZE); 1200 while (npages > 0) { 1201 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT | 1202 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1203 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK : 1204 VM_ALLOC_NOWAIT)); 1205 if (p != NULL) { 1206 /* 1207 * Since the page does not belong to an object, its 1208 * listq is unused. 1209 */ 1210 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1211 npages--; 1212 continue; 1213 } 1214 /* 1215 * Page allocation failed, free intermediate pages and 1216 * exit. 1217 */ 1218 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1219 vm_page_unwire(p, PQ_NONE); 1220 vm_page_free(p); 1221 } 1222 return (NULL); 1223 } 1224 *flags = UMA_SLAB_PRIV; 1225 zkva = keg->uk_kva + 1226 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1227 retkva = zkva; 1228 TAILQ_FOREACH(p, &alloctail, listq) { 1229 pmap_qenter(zkva, &p, 1); 1230 zkva += PAGE_SIZE; 1231 } 1232 1233 return ((void *)retkva); 1234 } 1235 1236 /* 1237 * Frees a number of pages to the system 1238 * 1239 * Arguments: 1240 * mem A pointer to the memory to be freed 1241 * size The size of the memory being freed 1242 * flags The original p->us_flags field 1243 * 1244 * Returns: 1245 * Nothing 1246 */ 1247 static void 1248 page_free(void *mem, vm_size_t size, uint8_t flags) 1249 { 1250 struct vmem *vmem; 1251 1252 if (flags & UMA_SLAB_KERNEL) 1253 vmem = kernel_arena; 1254 else 1255 panic("UMA: page_free used with invalid flags %x", flags); 1256 1257 kmem_free(vmem, (vm_offset_t)mem, size); 1258 } 1259 1260 /* 1261 * Zero fill initializer 1262 * 1263 * Arguments/Returns follow uma_init specifications 1264 */ 1265 static int 1266 zero_init(void *mem, int size, int flags) 1267 { 1268 bzero(mem, size); 1269 return (0); 1270 } 1271 1272 /* 1273 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1274 * 1275 * Arguments 1276 * keg The zone we should initialize 1277 * 1278 * Returns 1279 * Nothing 1280 */ 1281 static void 1282 keg_small_init(uma_keg_t keg) 1283 { 1284 u_int rsize; 1285 u_int memused; 1286 u_int wastedspace; 1287 u_int shsize; 1288 u_int slabsize; 1289 1290 if (keg->uk_flags & UMA_ZONE_PCPU) { 1291 u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU; 1292 1293 slabsize = sizeof(struct pcpu); 1294 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu), 1295 PAGE_SIZE); 1296 } else { 1297 slabsize = UMA_SLAB_SIZE; 1298 keg->uk_ppera = 1; 1299 } 1300 1301 /* 1302 * Calculate the size of each allocation (rsize) according to 1303 * alignment. If the requested size is smaller than we have 1304 * allocation bits for we round it up. 1305 */ 1306 rsize = keg->uk_size; 1307 if (rsize < slabsize / SLAB_SETSIZE) 1308 rsize = slabsize / SLAB_SETSIZE; 1309 if (rsize & keg->uk_align) 1310 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1311 keg->uk_rsize = rsize; 1312 1313 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1314 keg->uk_rsize < sizeof(struct pcpu), 1315 ("%s: size %u too large", __func__, keg->uk_rsize)); 1316 1317 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1318 shsize = 0; 1319 else 1320 shsize = sizeof(struct uma_slab); 1321 1322 if (rsize <= slabsize - shsize) 1323 keg->uk_ipers = (slabsize - shsize) / rsize; 1324 else { 1325 /* Handle special case when we have 1 item per slab, so 1326 * alignment requirement can be relaxed. */ 1327 KASSERT(keg->uk_size <= slabsize - shsize, 1328 ("%s: size %u greater than slab", __func__, keg->uk_size)); 1329 keg->uk_ipers = 1; 1330 } 1331 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1332 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1333 1334 memused = keg->uk_ipers * rsize + shsize; 1335 wastedspace = slabsize - memused; 1336 1337 /* 1338 * We can't do OFFPAGE if we're internal or if we've been 1339 * asked to not go to the VM for buckets. If we do this we 1340 * may end up going to the VM for slabs which we do not 1341 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1342 * of UMA_ZONE_VM, which clearly forbids it. 1343 */ 1344 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1345 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1346 return; 1347 1348 /* 1349 * See if using an OFFPAGE slab will limit our waste. Only do 1350 * this if it permits more items per-slab. 1351 * 1352 * XXX We could try growing slabsize to limit max waste as well. 1353 * Historically this was not done because the VM could not 1354 * efficiently handle contiguous allocations. 1355 */ 1356 if ((wastedspace >= slabsize / UMA_MAX_WASTE) && 1357 (keg->uk_ipers < (slabsize / keg->uk_rsize))) { 1358 keg->uk_ipers = slabsize / keg->uk_rsize; 1359 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1360 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1361 CTR6(KTR_UMA, "UMA decided we need offpage slab headers for " 1362 "keg: %s(%p), calculated wastedspace = %d, " 1363 "maximum wasted space allowed = %d, " 1364 "calculated ipers = %d, " 1365 "new wasted space = %d\n", keg->uk_name, keg, wastedspace, 1366 slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1367 slabsize - keg->uk_ipers * keg->uk_rsize); 1368 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1369 } 1370 1371 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1372 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1373 keg->uk_flags |= UMA_ZONE_HASH; 1374 } 1375 1376 /* 1377 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1378 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1379 * more complicated. 1380 * 1381 * Arguments 1382 * keg The keg we should initialize 1383 * 1384 * Returns 1385 * Nothing 1386 */ 1387 static void 1388 keg_large_init(uma_keg_t keg) 1389 { 1390 u_int shsize; 1391 1392 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1393 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1394 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1395 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1396 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1397 1398 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1399 keg->uk_ipers = 1; 1400 keg->uk_rsize = keg->uk_size; 1401 1402 /* Check whether we have enough space to not do OFFPAGE. */ 1403 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) { 1404 shsize = sizeof(struct uma_slab); 1405 if (shsize & UMA_ALIGN_PTR) 1406 shsize = (shsize & ~UMA_ALIGN_PTR) + 1407 (UMA_ALIGN_PTR + 1); 1408 1409 if (PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < shsize) { 1410 /* 1411 * We can't do OFFPAGE if we're internal, in which case 1412 * we need an extra page per allocation to contain the 1413 * slab header. 1414 */ 1415 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0) 1416 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1417 else 1418 keg->uk_ppera++; 1419 } 1420 } 1421 1422 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1423 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1424 keg->uk_flags |= UMA_ZONE_HASH; 1425 } 1426 1427 static void 1428 keg_cachespread_init(uma_keg_t keg) 1429 { 1430 int alignsize; 1431 int trailer; 1432 int pages; 1433 int rsize; 1434 1435 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1436 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1437 1438 alignsize = keg->uk_align + 1; 1439 rsize = keg->uk_size; 1440 /* 1441 * We want one item to start on every align boundary in a page. To 1442 * do this we will span pages. We will also extend the item by the 1443 * size of align if it is an even multiple of align. Otherwise, it 1444 * would fall on the same boundary every time. 1445 */ 1446 if (rsize & keg->uk_align) 1447 rsize = (rsize & ~keg->uk_align) + alignsize; 1448 if ((rsize & alignsize) == 0) 1449 rsize += alignsize; 1450 trailer = rsize - keg->uk_size; 1451 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1452 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1453 keg->uk_rsize = rsize; 1454 keg->uk_ppera = pages; 1455 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1456 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1457 KASSERT(keg->uk_ipers <= SLAB_SETSIZE, 1458 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1459 keg->uk_ipers)); 1460 } 1461 1462 /* 1463 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1464 * the keg onto the global keg list. 1465 * 1466 * Arguments/Returns follow uma_ctor specifications 1467 * udata Actually uma_kctor_args 1468 */ 1469 static int 1470 keg_ctor(void *mem, int size, void *udata, int flags) 1471 { 1472 struct uma_kctor_args *arg = udata; 1473 uma_keg_t keg = mem; 1474 uma_zone_t zone; 1475 1476 bzero(keg, size); 1477 keg->uk_size = arg->size; 1478 keg->uk_init = arg->uminit; 1479 keg->uk_fini = arg->fini; 1480 keg->uk_align = arg->align; 1481 keg->uk_cursor = 0; 1482 keg->uk_free = 0; 1483 keg->uk_reserve = 0; 1484 keg->uk_pages = 0; 1485 keg->uk_flags = arg->flags; 1486 keg->uk_slabzone = NULL; 1487 1488 /* 1489 * The master zone is passed to us at keg-creation time. 1490 */ 1491 zone = arg->zone; 1492 keg->uk_name = zone->uz_name; 1493 1494 if (arg->flags & UMA_ZONE_VM) 1495 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1496 1497 if (arg->flags & UMA_ZONE_ZINIT) 1498 keg->uk_init = zero_init; 1499 1500 if (arg->flags & UMA_ZONE_MALLOC) 1501 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1502 1503 if (arg->flags & UMA_ZONE_PCPU) 1504 #ifdef SMP 1505 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1506 #else 1507 keg->uk_flags &= ~UMA_ZONE_PCPU; 1508 #endif 1509 1510 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1511 keg_cachespread_init(keg); 1512 } else { 1513 if (keg->uk_size > UMA_SLAB_SPACE) 1514 keg_large_init(keg); 1515 else 1516 keg_small_init(keg); 1517 } 1518 1519 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1520 keg->uk_slabzone = slabzone; 1521 1522 /* 1523 * If we haven't booted yet we need allocations to go through the 1524 * startup cache until the vm is ready. 1525 */ 1526 if (booted < BOOT_PAGEALLOC) 1527 keg->uk_allocf = startup_alloc; 1528 #ifdef UMA_MD_SMALL_ALLOC 1529 else if (keg->uk_ppera == 1) 1530 keg->uk_allocf = uma_small_alloc; 1531 #endif 1532 else 1533 keg->uk_allocf = page_alloc; 1534 #ifdef UMA_MD_SMALL_ALLOC 1535 if (keg->uk_ppera == 1) 1536 keg->uk_freef = uma_small_free; 1537 else 1538 #endif 1539 keg->uk_freef = page_free; 1540 1541 /* 1542 * Initialize keg's lock 1543 */ 1544 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1545 1546 /* 1547 * If we're putting the slab header in the actual page we need to 1548 * figure out where in each page it goes. This calculates a right 1549 * justified offset into the memory on an ALIGN_PTR boundary. 1550 */ 1551 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1552 u_int totsize; 1553 1554 /* Size of the slab struct and free list */ 1555 totsize = sizeof(struct uma_slab); 1556 1557 if (totsize & UMA_ALIGN_PTR) 1558 totsize = (totsize & ~UMA_ALIGN_PTR) + 1559 (UMA_ALIGN_PTR + 1); 1560 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1561 1562 /* 1563 * The only way the following is possible is if with our 1564 * UMA_ALIGN_PTR adjustments we are now bigger than 1565 * UMA_SLAB_SIZE. I haven't checked whether this is 1566 * mathematically possible for all cases, so we make 1567 * sure here anyway. 1568 */ 1569 totsize = keg->uk_pgoff + sizeof(struct uma_slab); 1570 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1571 printf("zone %s ipers %d rsize %d size %d\n", 1572 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1573 keg->uk_size); 1574 panic("UMA slab won't fit."); 1575 } 1576 } 1577 1578 if (keg->uk_flags & UMA_ZONE_HASH) 1579 hash_alloc(&keg->uk_hash); 1580 1581 CTR5(KTR_UMA, "keg_ctor %p zone %s(%p) out %d free %d\n", 1582 keg, zone->uz_name, zone, 1583 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, 1584 keg->uk_free); 1585 1586 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1587 1588 rw_wlock(&uma_rwlock); 1589 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1590 rw_wunlock(&uma_rwlock); 1591 return (0); 1592 } 1593 1594 /* 1595 * Zone header ctor. This initializes all fields, locks, etc. 1596 * 1597 * Arguments/Returns follow uma_ctor specifications 1598 * udata Actually uma_zctor_args 1599 */ 1600 static int 1601 zone_ctor(void *mem, int size, void *udata, int flags) 1602 { 1603 struct uma_zctor_args *arg = udata; 1604 uma_zone_t zone = mem; 1605 uma_zone_t z; 1606 uma_keg_t keg; 1607 1608 bzero(zone, size); 1609 zone->uz_name = arg->name; 1610 zone->uz_ctor = arg->ctor; 1611 zone->uz_dtor = arg->dtor; 1612 zone->uz_slab = zone_fetch_slab; 1613 zone->uz_init = NULL; 1614 zone->uz_fini = NULL; 1615 zone->uz_allocs = 0; 1616 zone->uz_frees = 0; 1617 zone->uz_fails = 0; 1618 zone->uz_sleeps = 0; 1619 zone->uz_count = 0; 1620 zone->uz_count_min = 0; 1621 zone->uz_flags = 0; 1622 zone->uz_warning = NULL; 1623 /* The domain structures follow the cpu structures. */ 1624 zone->uz_domain = (struct uma_zone_domain *)&zone->uz_cpu[mp_ncpus]; 1625 timevalclear(&zone->uz_ratecheck); 1626 keg = arg->keg; 1627 1628 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 1629 1630 /* 1631 * This is a pure cache zone, no kegs. 1632 */ 1633 if (arg->import) { 1634 if (arg->flags & UMA_ZONE_VM) 1635 arg->flags |= UMA_ZFLAG_CACHEONLY; 1636 zone->uz_flags = arg->flags; 1637 zone->uz_size = arg->size; 1638 zone->uz_import = arg->import; 1639 zone->uz_release = arg->release; 1640 zone->uz_arg = arg->arg; 1641 zone->uz_lockptr = &zone->uz_lock; 1642 rw_wlock(&uma_rwlock); 1643 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 1644 rw_wunlock(&uma_rwlock); 1645 goto out; 1646 } 1647 1648 /* 1649 * Use the regular zone/keg/slab allocator. 1650 */ 1651 zone->uz_import = (uma_import)zone_import; 1652 zone->uz_release = (uma_release)zone_release; 1653 zone->uz_arg = zone; 1654 1655 if (arg->flags & UMA_ZONE_SECONDARY) { 1656 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1657 zone->uz_init = arg->uminit; 1658 zone->uz_fini = arg->fini; 1659 zone->uz_lockptr = &keg->uk_lock; 1660 zone->uz_flags |= UMA_ZONE_SECONDARY; 1661 rw_wlock(&uma_rwlock); 1662 ZONE_LOCK(zone); 1663 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1664 if (LIST_NEXT(z, uz_link) == NULL) { 1665 LIST_INSERT_AFTER(z, zone, uz_link); 1666 break; 1667 } 1668 } 1669 ZONE_UNLOCK(zone); 1670 rw_wunlock(&uma_rwlock); 1671 } else if (keg == NULL) { 1672 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1673 arg->align, arg->flags)) == NULL) 1674 return (ENOMEM); 1675 } else { 1676 struct uma_kctor_args karg; 1677 int error; 1678 1679 /* We should only be here from uma_startup() */ 1680 karg.size = arg->size; 1681 karg.uminit = arg->uminit; 1682 karg.fini = arg->fini; 1683 karg.align = arg->align; 1684 karg.flags = arg->flags; 1685 karg.zone = zone; 1686 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1687 flags); 1688 if (error) 1689 return (error); 1690 } 1691 1692 /* 1693 * Link in the first keg. 1694 */ 1695 zone->uz_klink.kl_keg = keg; 1696 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1697 zone->uz_lockptr = &keg->uk_lock; 1698 zone->uz_size = keg->uk_size; 1699 zone->uz_flags |= (keg->uk_flags & 1700 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1701 1702 /* 1703 * Some internal zones don't have room allocated for the per cpu 1704 * caches. If we're internal, bail out here. 1705 */ 1706 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1707 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1708 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1709 return (0); 1710 } 1711 1712 out: 1713 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) != 1714 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET), 1715 ("Invalid zone flag combination")); 1716 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0) 1717 zone->uz_count = BUCKET_MAX; 1718 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0) 1719 zone->uz_count = 0; 1720 else 1721 zone->uz_count = bucket_select(zone->uz_size); 1722 zone->uz_count_min = zone->uz_count; 1723 1724 return (0); 1725 } 1726 1727 /* 1728 * Keg header dtor. This frees all data, destroys locks, frees the hash 1729 * table and removes the keg from the global list. 1730 * 1731 * Arguments/Returns follow uma_dtor specifications 1732 * udata unused 1733 */ 1734 static void 1735 keg_dtor(void *arg, int size, void *udata) 1736 { 1737 uma_keg_t keg; 1738 1739 keg = (uma_keg_t)arg; 1740 KEG_LOCK(keg); 1741 if (keg->uk_free != 0) { 1742 printf("Freed UMA keg (%s) was not empty (%d items). " 1743 " Lost %d pages of memory.\n", 1744 keg->uk_name ? keg->uk_name : "", 1745 keg->uk_free, keg->uk_pages); 1746 } 1747 KEG_UNLOCK(keg); 1748 1749 hash_free(&keg->uk_hash); 1750 1751 KEG_LOCK_FINI(keg); 1752 } 1753 1754 /* 1755 * Zone header dtor. 1756 * 1757 * Arguments/Returns follow uma_dtor specifications 1758 * udata unused 1759 */ 1760 static void 1761 zone_dtor(void *arg, int size, void *udata) 1762 { 1763 uma_klink_t klink; 1764 uma_zone_t zone; 1765 uma_keg_t keg; 1766 1767 zone = (uma_zone_t)arg; 1768 keg = zone_first_keg(zone); 1769 1770 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1771 cache_drain(zone); 1772 1773 rw_wlock(&uma_rwlock); 1774 LIST_REMOVE(zone, uz_link); 1775 rw_wunlock(&uma_rwlock); 1776 /* 1777 * XXX there are some races here where 1778 * the zone can be drained but zone lock 1779 * released and then refilled before we 1780 * remove it... we dont care for now 1781 */ 1782 zone_drain_wait(zone, M_WAITOK); 1783 /* 1784 * Unlink all of our kegs. 1785 */ 1786 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1787 klink->kl_keg = NULL; 1788 LIST_REMOVE(klink, kl_link); 1789 if (klink == &zone->uz_klink) 1790 continue; 1791 free(klink, M_TEMP); 1792 } 1793 /* 1794 * We only destroy kegs from non secondary zones. 1795 */ 1796 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1797 rw_wlock(&uma_rwlock); 1798 LIST_REMOVE(keg, uk_link); 1799 rw_wunlock(&uma_rwlock); 1800 zone_free_item(kegs, keg, NULL, SKIP_NONE); 1801 } 1802 ZONE_LOCK_FINI(zone); 1803 } 1804 1805 /* 1806 * Traverses every zone in the system and calls a callback 1807 * 1808 * Arguments: 1809 * zfunc A pointer to a function which accepts a zone 1810 * as an argument. 1811 * 1812 * Returns: 1813 * Nothing 1814 */ 1815 static void 1816 zone_foreach(void (*zfunc)(uma_zone_t)) 1817 { 1818 uma_keg_t keg; 1819 uma_zone_t zone; 1820 1821 rw_rlock(&uma_rwlock); 1822 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1823 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1824 zfunc(zone); 1825 } 1826 rw_runlock(&uma_rwlock); 1827 } 1828 1829 /* 1830 * Count how many pages do we need to bootstrap. VM supplies 1831 * its need in early zones in the argument, we add up our zones, 1832 * which consist of: UMA Slabs, UMA Hash and 9 Bucket zones. The 1833 * zone of zones and zone of kegs are accounted separately. 1834 */ 1835 #define UMA_BOOT_ZONES 11 1836 /* Zone of zones and zone of kegs have arbitrary alignment. */ 1837 #define UMA_BOOT_ALIGN 32 1838 static int zsize, ksize; 1839 int 1840 uma_startup_count(int vm_zones) 1841 { 1842 int zones, pages; 1843 1844 ksize = sizeof(struct uma_keg) + 1845 (sizeof(struct uma_domain) * vm_ndomains); 1846 zsize = sizeof(struct uma_zone) + 1847 (sizeof(struct uma_cache) * (mp_maxid + 1)) + 1848 (sizeof(struct uma_zone_domain) * vm_ndomains); 1849 1850 /* 1851 * Memory for the zone of kegs and its keg, 1852 * and for zone of zones. 1853 */ 1854 pages = howmany(roundup(zsize, CACHE_LINE_SIZE) * 2 + 1855 roundup(ksize, CACHE_LINE_SIZE), PAGE_SIZE); 1856 1857 #ifdef UMA_MD_SMALL_ALLOC 1858 zones = UMA_BOOT_ZONES; 1859 #else 1860 zones = UMA_BOOT_ZONES + vm_zones; 1861 vm_zones = 0; 1862 #endif 1863 1864 /* Memory for the rest of startup zones, UMA and VM, ... */ 1865 if (zsize > UMA_SLAB_SPACE) 1866 pages += (zones + vm_zones) * 1867 howmany(roundup2(zsize, UMA_BOOT_ALIGN), UMA_SLAB_SIZE); 1868 else if (roundup2(zsize, UMA_BOOT_ALIGN) > UMA_SLAB_SPACE) 1869 pages += zones; 1870 else 1871 pages += howmany(zones, 1872 UMA_SLAB_SPACE / roundup2(zsize, UMA_BOOT_ALIGN)); 1873 1874 /* ... and their kegs. Note that zone of zones allocates a keg! */ 1875 pages += howmany(zones + 1, 1876 UMA_SLAB_SPACE / roundup2(ksize, UMA_BOOT_ALIGN)); 1877 1878 /* 1879 * Most of startup zones are not going to be offpages, that's 1880 * why we use UMA_SLAB_SPACE instead of UMA_SLAB_SIZE in all 1881 * calculations. Some large bucket zones will be offpage, and 1882 * thus will allocate hashes. We take conservative approach 1883 * and assume that all zones may allocate hash. This may give 1884 * us some positive inaccuracy, usually an extra single page. 1885 */ 1886 pages += howmany(zones, UMA_SLAB_SPACE / 1887 (sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT)); 1888 1889 return (pages); 1890 } 1891 1892 void 1893 uma_startup(void *mem, int npages) 1894 { 1895 struct uma_zctor_args args; 1896 uma_keg_t masterkeg; 1897 uintptr_t m; 1898 1899 #ifdef DIAGNOSTIC 1900 printf("Entering %s with %d boot pages configured\n", __func__, npages); 1901 #endif 1902 1903 rw_init(&uma_rwlock, "UMA lock"); 1904 1905 /* Use bootpages memory for the zone of zones and zone of kegs. */ 1906 m = (uintptr_t)mem; 1907 zones = (uma_zone_t)m; 1908 m += roundup(zsize, CACHE_LINE_SIZE); 1909 kegs = (uma_zone_t)m; 1910 m += roundup(zsize, CACHE_LINE_SIZE); 1911 masterkeg = (uma_keg_t)m; 1912 m += roundup(ksize, CACHE_LINE_SIZE); 1913 m = roundup(m, PAGE_SIZE); 1914 npages -= (m - (uintptr_t)mem) / PAGE_SIZE; 1915 mem = (void *)m; 1916 1917 /* "manually" create the initial zone */ 1918 memset(&args, 0, sizeof(args)); 1919 args.name = "UMA Kegs"; 1920 args.size = ksize; 1921 args.ctor = keg_ctor; 1922 args.dtor = keg_dtor; 1923 args.uminit = zero_init; 1924 args.fini = NULL; 1925 args.keg = masterkeg; 1926 args.align = UMA_BOOT_ALIGN - 1; 1927 args.flags = UMA_ZFLAG_INTERNAL; 1928 zone_ctor(kegs, zsize, &args, M_WAITOK); 1929 1930 bootmem = mem; 1931 boot_pages = npages; 1932 1933 args.name = "UMA Zones"; 1934 args.size = zsize; 1935 args.ctor = zone_ctor; 1936 args.dtor = zone_dtor; 1937 args.uminit = zero_init; 1938 args.fini = NULL; 1939 args.keg = NULL; 1940 args.align = UMA_BOOT_ALIGN - 1; 1941 args.flags = UMA_ZFLAG_INTERNAL; 1942 zone_ctor(zones, zsize, &args, M_WAITOK); 1943 1944 /* Now make a zone for slab headers */ 1945 slabzone = uma_zcreate("UMA Slabs", 1946 sizeof(struct uma_slab), 1947 NULL, NULL, NULL, NULL, 1948 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1949 1950 hashzone = uma_zcreate("UMA Hash", 1951 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 1952 NULL, NULL, NULL, NULL, 1953 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1954 1955 bucket_init(); 1956 1957 booted = BOOT_STRAPPED; 1958 } 1959 1960 void 1961 uma_startup1(void) 1962 { 1963 1964 #ifdef DIAGNOSTIC 1965 printf("Entering %s with %d boot pages left\n", __func__, boot_pages); 1966 #endif 1967 booted = BOOT_PAGEALLOC; 1968 } 1969 1970 void 1971 uma_startup2(void) 1972 { 1973 1974 #ifdef DIAGNOSTIC 1975 printf("Entering %s with %d boot pages left\n", __func__, boot_pages); 1976 #endif 1977 booted = BOOT_BUCKETS; 1978 sx_init(&uma_drain_lock, "umadrain"); 1979 bucket_enable(); 1980 } 1981 1982 /* 1983 * Initialize our callout handle 1984 * 1985 */ 1986 static void 1987 uma_startup3(void) 1988 { 1989 1990 #ifdef INVARIANTS 1991 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor); 1992 uma_dbg_cnt = counter_u64_alloc(M_WAITOK); 1993 uma_skip_cnt = counter_u64_alloc(M_WAITOK); 1994 #endif 1995 callout_init(&uma_callout, 1); 1996 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1997 booted = BOOT_RUNNING; 1998 } 1999 2000 static uma_keg_t 2001 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 2002 int align, uint32_t flags) 2003 { 2004 struct uma_kctor_args args; 2005 2006 args.size = size; 2007 args.uminit = uminit; 2008 args.fini = fini; 2009 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 2010 args.flags = flags; 2011 args.zone = zone; 2012 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK)); 2013 } 2014 2015 /* Public functions */ 2016 /* See uma.h */ 2017 void 2018 uma_set_align(int align) 2019 { 2020 2021 if (align != UMA_ALIGN_CACHE) 2022 uma_align_cache = align; 2023 } 2024 2025 /* See uma.h */ 2026 uma_zone_t 2027 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 2028 uma_init uminit, uma_fini fini, int align, uint32_t flags) 2029 2030 { 2031 struct uma_zctor_args args; 2032 uma_zone_t res; 2033 bool locked; 2034 2035 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"", 2036 align, name)); 2037 2038 /* This stuff is essential for the zone ctor */ 2039 memset(&args, 0, sizeof(args)); 2040 args.name = name; 2041 args.size = size; 2042 args.ctor = ctor; 2043 args.dtor = dtor; 2044 args.uminit = uminit; 2045 args.fini = fini; 2046 #ifdef INVARIANTS 2047 /* 2048 * If a zone is being created with an empty constructor and 2049 * destructor, pass UMA constructor/destructor which checks for 2050 * memory use after free. 2051 */ 2052 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) && 2053 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) { 2054 args.ctor = trash_ctor; 2055 args.dtor = trash_dtor; 2056 args.uminit = trash_init; 2057 args.fini = trash_fini; 2058 } 2059 #endif 2060 args.align = align; 2061 args.flags = flags; 2062 args.keg = NULL; 2063 2064 if (booted < BOOT_BUCKETS) { 2065 locked = false; 2066 } else { 2067 sx_slock(&uma_drain_lock); 2068 locked = true; 2069 } 2070 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2071 if (locked) 2072 sx_sunlock(&uma_drain_lock); 2073 return (res); 2074 } 2075 2076 /* See uma.h */ 2077 uma_zone_t 2078 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 2079 uma_init zinit, uma_fini zfini, uma_zone_t master) 2080 { 2081 struct uma_zctor_args args; 2082 uma_keg_t keg; 2083 uma_zone_t res; 2084 bool locked; 2085 2086 keg = zone_first_keg(master); 2087 memset(&args, 0, sizeof(args)); 2088 args.name = name; 2089 args.size = keg->uk_size; 2090 args.ctor = ctor; 2091 args.dtor = dtor; 2092 args.uminit = zinit; 2093 args.fini = zfini; 2094 args.align = keg->uk_align; 2095 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 2096 args.keg = keg; 2097 2098 if (booted < BOOT_BUCKETS) { 2099 locked = false; 2100 } else { 2101 sx_slock(&uma_drain_lock); 2102 locked = true; 2103 } 2104 /* XXX Attaches only one keg of potentially many. */ 2105 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2106 if (locked) 2107 sx_sunlock(&uma_drain_lock); 2108 return (res); 2109 } 2110 2111 /* See uma.h */ 2112 uma_zone_t 2113 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 2114 uma_init zinit, uma_fini zfini, uma_import zimport, 2115 uma_release zrelease, void *arg, int flags) 2116 { 2117 struct uma_zctor_args args; 2118 2119 memset(&args, 0, sizeof(args)); 2120 args.name = name; 2121 args.size = size; 2122 args.ctor = ctor; 2123 args.dtor = dtor; 2124 args.uminit = zinit; 2125 args.fini = zfini; 2126 args.import = zimport; 2127 args.release = zrelease; 2128 args.arg = arg; 2129 args.align = 0; 2130 args.flags = flags; 2131 2132 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK)); 2133 } 2134 2135 static void 2136 zone_lock_pair(uma_zone_t a, uma_zone_t b) 2137 { 2138 if (a < b) { 2139 ZONE_LOCK(a); 2140 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); 2141 } else { 2142 ZONE_LOCK(b); 2143 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); 2144 } 2145 } 2146 2147 static void 2148 zone_unlock_pair(uma_zone_t a, uma_zone_t b) 2149 { 2150 2151 ZONE_UNLOCK(a); 2152 ZONE_UNLOCK(b); 2153 } 2154 2155 int 2156 uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 2157 { 2158 uma_klink_t klink; 2159 uma_klink_t kl; 2160 int error; 2161 2162 error = 0; 2163 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 2164 2165 zone_lock_pair(zone, master); 2166 /* 2167 * zone must use vtoslab() to resolve objects and must already be 2168 * a secondary. 2169 */ 2170 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 2171 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 2172 error = EINVAL; 2173 goto out; 2174 } 2175 /* 2176 * The new master must also use vtoslab(). 2177 */ 2178 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 2179 error = EINVAL; 2180 goto out; 2181 } 2182 2183 /* 2184 * The underlying object must be the same size. rsize 2185 * may be different. 2186 */ 2187 if (master->uz_size != zone->uz_size) { 2188 error = E2BIG; 2189 goto out; 2190 } 2191 /* 2192 * Put it at the end of the list. 2193 */ 2194 klink->kl_keg = zone_first_keg(master); 2195 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 2196 if (LIST_NEXT(kl, kl_link) == NULL) { 2197 LIST_INSERT_AFTER(kl, klink, kl_link); 2198 break; 2199 } 2200 } 2201 klink = NULL; 2202 zone->uz_flags |= UMA_ZFLAG_MULTI; 2203 zone->uz_slab = zone_fetch_slab_multi; 2204 2205 out: 2206 zone_unlock_pair(zone, master); 2207 if (klink != NULL) 2208 free(klink, M_TEMP); 2209 2210 return (error); 2211 } 2212 2213 2214 /* See uma.h */ 2215 void 2216 uma_zdestroy(uma_zone_t zone) 2217 { 2218 2219 sx_slock(&uma_drain_lock); 2220 zone_free_item(zones, zone, NULL, SKIP_NONE); 2221 sx_sunlock(&uma_drain_lock); 2222 } 2223 2224 void 2225 uma_zwait(uma_zone_t zone) 2226 { 2227 void *item; 2228 2229 item = uma_zalloc_arg(zone, NULL, M_WAITOK); 2230 uma_zfree(zone, item); 2231 } 2232 2233 void * 2234 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags) 2235 { 2236 void *item; 2237 int i; 2238 2239 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2240 item = uma_zalloc_arg(zone, udata, flags &~ M_ZERO); 2241 if (item != NULL && (flags & M_ZERO)) { 2242 CPU_FOREACH(i) 2243 bzero(zpcpu_get_cpu(item, i), zone->uz_size); 2244 } 2245 return (item); 2246 } 2247 2248 /* 2249 * A stub while both regular and pcpu cases are identical. 2250 */ 2251 void 2252 uma_zfree_pcpu_arg(uma_zone_t zone, void *item, void *udata) 2253 { 2254 2255 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2256 uma_zfree_arg(zone, item, udata); 2257 } 2258 2259 /* See uma.h */ 2260 void * 2261 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2262 { 2263 uma_zone_domain_t zdom; 2264 uma_bucket_t bucket; 2265 uma_cache_t cache; 2266 void *item; 2267 int cpu, domain, lockfail; 2268 #ifdef INVARIANTS 2269 bool skipdbg; 2270 #endif 2271 2272 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2273 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA); 2274 2275 /* This is the fast path allocation */ 2276 CTR4(KTR_UMA, "uma_zalloc_arg thread %x zone %s(%p) flags %d", 2277 curthread, zone->uz_name, zone, flags); 2278 2279 if (flags & M_WAITOK) { 2280 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2281 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2282 } 2283 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2284 ("uma_zalloc_arg: called with spinlock or critical section held")); 2285 if (zone->uz_flags & UMA_ZONE_PCPU) 2286 KASSERT((flags & M_ZERO) == 0, ("allocating from a pcpu zone " 2287 "with M_ZERO passed")); 2288 2289 #ifdef DEBUG_MEMGUARD 2290 if (memguard_cmp_zone(zone)) { 2291 item = memguard_alloc(zone->uz_size, flags); 2292 if (item != NULL) { 2293 if (zone->uz_init != NULL && 2294 zone->uz_init(item, zone->uz_size, flags) != 0) 2295 return (NULL); 2296 if (zone->uz_ctor != NULL && 2297 zone->uz_ctor(item, zone->uz_size, udata, 2298 flags) != 0) { 2299 zone->uz_fini(item, zone->uz_size); 2300 return (NULL); 2301 } 2302 return (item); 2303 } 2304 /* This is unfortunate but should not be fatal. */ 2305 } 2306 #endif 2307 /* 2308 * If possible, allocate from the per-CPU cache. There are two 2309 * requirements for safe access to the per-CPU cache: (1) the thread 2310 * accessing the cache must not be preempted or yield during access, 2311 * and (2) the thread must not migrate CPUs without switching which 2312 * cache it accesses. We rely on a critical section to prevent 2313 * preemption and migration. We release the critical section in 2314 * order to acquire the zone mutex if we are unable to allocate from 2315 * the current cache; when we re-acquire the critical section, we 2316 * must detect and handle migration if it has occurred. 2317 */ 2318 critical_enter(); 2319 cpu = curcpu; 2320 cache = &zone->uz_cpu[cpu]; 2321 2322 zalloc_start: 2323 bucket = cache->uc_allocbucket; 2324 if (bucket != NULL && bucket->ub_cnt > 0) { 2325 bucket->ub_cnt--; 2326 item = bucket->ub_bucket[bucket->ub_cnt]; 2327 #ifdef INVARIANTS 2328 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2329 #endif 2330 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2331 cache->uc_allocs++; 2332 critical_exit(); 2333 #ifdef INVARIANTS 2334 skipdbg = uma_dbg_zskip(zone, item); 2335 #endif 2336 if (zone->uz_ctor != NULL && 2337 #ifdef INVARIANTS 2338 (!skipdbg || zone->uz_ctor != trash_ctor || 2339 zone->uz_dtor != trash_dtor) && 2340 #endif 2341 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2342 atomic_add_long(&zone->uz_fails, 1); 2343 zone_free_item(zone, item, udata, SKIP_DTOR); 2344 return (NULL); 2345 } 2346 #ifdef INVARIANTS 2347 if (!skipdbg) 2348 uma_dbg_alloc(zone, NULL, item); 2349 #endif 2350 if (flags & M_ZERO) 2351 uma_zero_item(item, zone); 2352 return (item); 2353 } 2354 2355 /* 2356 * We have run out of items in our alloc bucket. 2357 * See if we can switch with our free bucket. 2358 */ 2359 bucket = cache->uc_freebucket; 2360 if (bucket != NULL && bucket->ub_cnt > 0) { 2361 CTR2(KTR_UMA, 2362 "uma_zalloc: zone %s(%p) swapping empty with alloc", 2363 zone->uz_name, zone); 2364 cache->uc_freebucket = cache->uc_allocbucket; 2365 cache->uc_allocbucket = bucket; 2366 goto zalloc_start; 2367 } 2368 2369 /* 2370 * Discard any empty allocation bucket while we hold no locks. 2371 */ 2372 bucket = cache->uc_allocbucket; 2373 cache->uc_allocbucket = NULL; 2374 critical_exit(); 2375 if (bucket != NULL) 2376 bucket_free(zone, bucket, udata); 2377 2378 if (zone->uz_flags & UMA_ZONE_NUMA) 2379 domain = PCPU_GET(domain); 2380 else 2381 domain = UMA_ANYDOMAIN; 2382 2383 /* Short-circuit for zones without buckets and low memory. */ 2384 if (zone->uz_count == 0 || bucketdisable) 2385 goto zalloc_item; 2386 2387 /* 2388 * Attempt to retrieve the item from the per-CPU cache has failed, so 2389 * we must go back to the zone. This requires the zone lock, so we 2390 * must drop the critical section, then re-acquire it when we go back 2391 * to the cache. Since the critical section is released, we may be 2392 * preempted or migrate. As such, make sure not to maintain any 2393 * thread-local state specific to the cache from prior to releasing 2394 * the critical section. 2395 */ 2396 lockfail = 0; 2397 if (ZONE_TRYLOCK(zone) == 0) { 2398 /* Record contention to size the buckets. */ 2399 ZONE_LOCK(zone); 2400 lockfail = 1; 2401 } 2402 critical_enter(); 2403 cpu = curcpu; 2404 cache = &zone->uz_cpu[cpu]; 2405 2406 /* See if we lost the race to fill the cache. */ 2407 if (cache->uc_allocbucket != NULL) { 2408 ZONE_UNLOCK(zone); 2409 goto zalloc_start; 2410 } 2411 2412 /* 2413 * Check the zone's cache of buckets. 2414 */ 2415 if (domain == UMA_ANYDOMAIN) 2416 zdom = &zone->uz_domain[0]; 2417 else 2418 zdom = &zone->uz_domain[domain]; 2419 if ((bucket = LIST_FIRST(&zdom->uzd_buckets)) != NULL) { 2420 KASSERT(bucket->ub_cnt != 0, 2421 ("uma_zalloc_arg: Returning an empty bucket.")); 2422 2423 LIST_REMOVE(bucket, ub_link); 2424 cache->uc_allocbucket = bucket; 2425 ZONE_UNLOCK(zone); 2426 goto zalloc_start; 2427 } 2428 /* We are no longer associated with this CPU. */ 2429 critical_exit(); 2430 2431 /* 2432 * We bump the uz count when the cache size is insufficient to 2433 * handle the working set. 2434 */ 2435 if (lockfail && zone->uz_count < BUCKET_MAX) 2436 zone->uz_count++; 2437 ZONE_UNLOCK(zone); 2438 2439 /* 2440 * Now lets just fill a bucket and put it on the free list. If that 2441 * works we'll restart the allocation from the beginning and it 2442 * will use the just filled bucket. 2443 */ 2444 bucket = zone_alloc_bucket(zone, udata, domain, flags); 2445 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p", 2446 zone->uz_name, zone, bucket); 2447 if (bucket != NULL) { 2448 ZONE_LOCK(zone); 2449 critical_enter(); 2450 cpu = curcpu; 2451 cache = &zone->uz_cpu[cpu]; 2452 /* 2453 * See if we lost the race or were migrated. Cache the 2454 * initialized bucket to make this less likely or claim 2455 * the memory directly. 2456 */ 2457 if (cache->uc_allocbucket != NULL || 2458 (zone->uz_flags & UMA_ZONE_NUMA && 2459 domain != PCPU_GET(domain))) 2460 LIST_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link); 2461 else 2462 cache->uc_allocbucket = bucket; 2463 ZONE_UNLOCK(zone); 2464 goto zalloc_start; 2465 } 2466 2467 /* 2468 * We may not be able to get a bucket so return an actual item. 2469 */ 2470 zalloc_item: 2471 item = zone_alloc_item(zone, udata, domain, flags); 2472 2473 return (item); 2474 } 2475 2476 void * 2477 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags) 2478 { 2479 2480 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2481 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA); 2482 2483 /* This is the fast path allocation */ 2484 CTR5(KTR_UMA, 2485 "uma_zalloc_domain thread %x zone %s(%p) domain %d flags %d", 2486 curthread, zone->uz_name, zone, domain, flags); 2487 2488 if (flags & M_WAITOK) { 2489 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2490 "uma_zalloc_domain: zone \"%s\"", zone->uz_name); 2491 } 2492 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2493 ("uma_zalloc_domain: called with spinlock or critical section held")); 2494 2495 return (zone_alloc_item(zone, udata, domain, flags)); 2496 } 2497 2498 /* 2499 * Find a slab with some space. Prefer slabs that are partially used over those 2500 * that are totally full. This helps to reduce fragmentation. 2501 * 2502 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check 2503 * only 'domain'. 2504 */ 2505 static uma_slab_t 2506 keg_first_slab(uma_keg_t keg, int domain, int rr) 2507 { 2508 uma_domain_t dom; 2509 uma_slab_t slab; 2510 int start; 2511 2512 KASSERT(domain >= 0 && domain < vm_ndomains, 2513 ("keg_first_slab: domain %d out of range", domain)); 2514 2515 slab = NULL; 2516 start = domain; 2517 do { 2518 dom = &keg->uk_domain[domain]; 2519 if (!LIST_EMPTY(&dom->ud_part_slab)) 2520 return (LIST_FIRST(&dom->ud_part_slab)); 2521 if (!LIST_EMPTY(&dom->ud_free_slab)) { 2522 slab = LIST_FIRST(&dom->ud_free_slab); 2523 LIST_REMOVE(slab, us_link); 2524 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 2525 return (slab); 2526 } 2527 if (rr) 2528 domain = (domain + 1) % vm_ndomains; 2529 } while (domain != start); 2530 2531 return (NULL); 2532 } 2533 2534 static uma_slab_t 2535 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, int flags) 2536 { 2537 uma_domain_t dom; 2538 uma_slab_t slab; 2539 int allocflags, domain, reserve, rr, start; 2540 2541 mtx_assert(&keg->uk_lock, MA_OWNED); 2542 slab = NULL; 2543 reserve = 0; 2544 allocflags = flags; 2545 if ((flags & M_USE_RESERVE) == 0) 2546 reserve = keg->uk_reserve; 2547 2548 /* 2549 * Round-robin for non first-touch zones when there is more than one 2550 * domain. 2551 */ 2552 if (vm_ndomains == 1) 2553 rdomain = 0; 2554 rr = rdomain == UMA_ANYDOMAIN; 2555 if (rr) { 2556 keg->uk_cursor = (keg->uk_cursor + 1) % vm_ndomains; 2557 domain = start = keg->uk_cursor; 2558 /* Only block on the second pass. */ 2559 if ((flags & (M_WAITOK | M_NOVM)) == M_WAITOK) 2560 allocflags = (allocflags & ~M_WAITOK) | M_NOWAIT; 2561 } else 2562 domain = start = rdomain; 2563 2564 again: 2565 do { 2566 if (keg->uk_free > reserve && 2567 (slab = keg_first_slab(keg, domain, rr)) != NULL) { 2568 MPASS(slab->us_keg == keg); 2569 return (slab); 2570 } 2571 2572 /* 2573 * M_NOVM means don't ask at all! 2574 */ 2575 if (flags & M_NOVM) 2576 break; 2577 2578 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2579 keg->uk_flags |= UMA_ZFLAG_FULL; 2580 /* 2581 * If this is not a multi-zone, set the FULL bit. 2582 * Otherwise slab_multi() takes care of it. 2583 */ 2584 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2585 zone->uz_flags |= UMA_ZFLAG_FULL; 2586 zone_log_warning(zone); 2587 zone_maxaction(zone); 2588 } 2589 if (flags & M_NOWAIT) 2590 return (NULL); 2591 zone->uz_sleeps++; 2592 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2593 continue; 2594 } 2595 slab = keg_alloc_slab(keg, zone, domain, allocflags); 2596 /* 2597 * If we got a slab here it's safe to mark it partially used 2598 * and return. We assume that the caller is going to remove 2599 * at least one item. 2600 */ 2601 if (slab) { 2602 MPASS(slab->us_keg == keg); 2603 dom = &keg->uk_domain[slab->us_domain]; 2604 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 2605 return (slab); 2606 } 2607 if (rr) { 2608 keg->uk_cursor = (keg->uk_cursor + 1) % vm_ndomains; 2609 domain = keg->uk_cursor; 2610 } 2611 } while (domain != start); 2612 2613 /* Retry domain scan with blocking. */ 2614 if (allocflags != flags) { 2615 allocflags = flags; 2616 goto again; 2617 } 2618 2619 /* 2620 * We might not have been able to get a slab but another cpu 2621 * could have while we were unlocked. Check again before we 2622 * fail. 2623 */ 2624 if (keg->uk_free > reserve && 2625 (slab = keg_first_slab(keg, domain, rr)) != NULL) { 2626 MPASS(slab->us_keg == keg); 2627 return (slab); 2628 } 2629 return (NULL); 2630 } 2631 2632 static uma_slab_t 2633 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int domain, int flags) 2634 { 2635 uma_slab_t slab; 2636 2637 if (keg == NULL) { 2638 keg = zone_first_keg(zone); 2639 KEG_LOCK(keg); 2640 } 2641 2642 for (;;) { 2643 slab = keg_fetch_slab(keg, zone, domain, flags); 2644 if (slab) 2645 return (slab); 2646 if (flags & (M_NOWAIT | M_NOVM)) 2647 break; 2648 } 2649 KEG_UNLOCK(keg); 2650 return (NULL); 2651 } 2652 2653 /* 2654 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2655 * with the keg locked. On NULL no lock is held. 2656 * 2657 * The last pointer is used to seed the search. It is not required. 2658 */ 2659 static uma_slab_t 2660 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int domain, int rflags) 2661 { 2662 uma_klink_t klink; 2663 uma_slab_t slab; 2664 uma_keg_t keg; 2665 int flags; 2666 int empty; 2667 int full; 2668 2669 /* 2670 * Don't wait on the first pass. This will skip limit tests 2671 * as well. We don't want to block if we can find a provider 2672 * without blocking. 2673 */ 2674 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2675 /* 2676 * Use the last slab allocated as a hint for where to start 2677 * the search. 2678 */ 2679 if (last != NULL) { 2680 slab = keg_fetch_slab(last, zone, domain, flags); 2681 if (slab) 2682 return (slab); 2683 KEG_UNLOCK(last); 2684 } 2685 /* 2686 * Loop until we have a slab incase of transient failures 2687 * while M_WAITOK is specified. I'm not sure this is 100% 2688 * required but we've done it for so long now. 2689 */ 2690 for (;;) { 2691 empty = 0; 2692 full = 0; 2693 /* 2694 * Search the available kegs for slabs. Be careful to hold the 2695 * correct lock while calling into the keg layer. 2696 */ 2697 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2698 keg = klink->kl_keg; 2699 KEG_LOCK(keg); 2700 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2701 slab = keg_fetch_slab(keg, zone, domain, flags); 2702 if (slab) 2703 return (slab); 2704 } 2705 if (keg->uk_flags & UMA_ZFLAG_FULL) 2706 full++; 2707 else 2708 empty++; 2709 KEG_UNLOCK(keg); 2710 } 2711 if (rflags & (M_NOWAIT | M_NOVM)) 2712 break; 2713 flags = rflags; 2714 /* 2715 * All kegs are full. XXX We can't atomically check all kegs 2716 * and sleep so just sleep for a short period and retry. 2717 */ 2718 if (full && !empty) { 2719 ZONE_LOCK(zone); 2720 zone->uz_flags |= UMA_ZFLAG_FULL; 2721 zone->uz_sleeps++; 2722 zone_log_warning(zone); 2723 zone_maxaction(zone); 2724 msleep(zone, zone->uz_lockptr, PVM, 2725 "zonelimit", hz/100); 2726 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2727 ZONE_UNLOCK(zone); 2728 continue; 2729 } 2730 } 2731 return (NULL); 2732 } 2733 2734 static void * 2735 slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 2736 { 2737 uma_domain_t dom; 2738 void *item; 2739 uint8_t freei; 2740 2741 MPASS(keg == slab->us_keg); 2742 mtx_assert(&keg->uk_lock, MA_OWNED); 2743 2744 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; 2745 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); 2746 item = slab->us_data + (keg->uk_rsize * freei); 2747 slab->us_freecount--; 2748 keg->uk_free--; 2749 2750 /* Move this slab to the full list */ 2751 if (slab->us_freecount == 0) { 2752 LIST_REMOVE(slab, us_link); 2753 dom = &keg->uk_domain[slab->us_domain]; 2754 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link); 2755 } 2756 2757 return (item); 2758 } 2759 2760 static int 2761 zone_import(uma_zone_t zone, void **bucket, int max, int domain, int flags) 2762 { 2763 uma_slab_t slab; 2764 uma_keg_t keg; 2765 int stripe; 2766 int i; 2767 2768 slab = NULL; 2769 keg = NULL; 2770 /* Try to keep the buckets totally full */ 2771 for (i = 0; i < max; ) { 2772 if ((slab = zone->uz_slab(zone, keg, domain, flags)) == NULL) 2773 break; 2774 keg = slab->us_keg; 2775 stripe = howmany(max, vm_ndomains); 2776 while (slab->us_freecount && i < max) { 2777 bucket[i++] = slab_alloc_item(keg, slab); 2778 if (keg->uk_free <= keg->uk_reserve) 2779 break; 2780 #ifdef NUMA 2781 /* 2782 * If the zone is striped we pick a new slab for every 2783 * N allocations. Eliminating this conditional will 2784 * instead pick a new domain for each bucket rather 2785 * than stripe within each bucket. The current option 2786 * produces more fragmentation and requires more cpu 2787 * time but yields better distribution. 2788 */ 2789 if ((zone->uz_flags & UMA_ZONE_NUMA) == 0 && 2790 vm_ndomains > 1 && --stripe == 0) 2791 break; 2792 #endif 2793 } 2794 /* Don't block if we allocated any successfully. */ 2795 flags &= ~M_WAITOK; 2796 flags |= M_NOWAIT; 2797 } 2798 if (slab != NULL) 2799 KEG_UNLOCK(keg); 2800 2801 return i; 2802 } 2803 2804 static uma_bucket_t 2805 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags) 2806 { 2807 uma_bucket_t bucket; 2808 int max; 2809 2810 /* Don't wait for buckets, preserve caller's NOVM setting. */ 2811 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 2812 if (bucket == NULL) 2813 return (NULL); 2814 2815 max = MIN(bucket->ub_entries, zone->uz_count); 2816 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 2817 max, domain, flags); 2818 2819 /* 2820 * Initialize the memory if necessary. 2821 */ 2822 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 2823 int i; 2824 2825 for (i = 0; i < bucket->ub_cnt; i++) 2826 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2827 flags) != 0) 2828 break; 2829 /* 2830 * If we couldn't initialize the whole bucket, put the 2831 * rest back onto the freelist. 2832 */ 2833 if (i != bucket->ub_cnt) { 2834 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 2835 bucket->ub_cnt - i); 2836 #ifdef INVARIANTS 2837 bzero(&bucket->ub_bucket[i], 2838 sizeof(void *) * (bucket->ub_cnt - i)); 2839 #endif 2840 bucket->ub_cnt = i; 2841 } 2842 } 2843 2844 if (bucket->ub_cnt == 0) { 2845 bucket_free(zone, bucket, udata); 2846 atomic_add_long(&zone->uz_fails, 1); 2847 return (NULL); 2848 } 2849 2850 return (bucket); 2851 } 2852 2853 /* 2854 * Allocates a single item from a zone. 2855 * 2856 * Arguments 2857 * zone The zone to alloc for. 2858 * udata The data to be passed to the constructor. 2859 * domain The domain to allocate from or UMA_ANYDOMAIN. 2860 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2861 * 2862 * Returns 2863 * NULL if there is no memory and M_NOWAIT is set 2864 * An item if successful 2865 */ 2866 2867 static void * 2868 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags) 2869 { 2870 void *item; 2871 #ifdef INVARIANTS 2872 bool skipdbg; 2873 #endif 2874 2875 item = NULL; 2876 2877 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1) 2878 goto fail; 2879 atomic_add_long(&zone->uz_allocs, 1); 2880 2881 #ifdef INVARIANTS 2882 skipdbg = uma_dbg_zskip(zone, item); 2883 #endif 2884 /* 2885 * We have to call both the zone's init (not the keg's init) 2886 * and the zone's ctor. This is because the item is going from 2887 * a keg slab directly to the user, and the user is expecting it 2888 * to be both zone-init'd as well as zone-ctor'd. 2889 */ 2890 if (zone->uz_init != NULL) { 2891 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 2892 zone_free_item(zone, item, udata, SKIP_FINI); 2893 goto fail; 2894 } 2895 } 2896 if (zone->uz_ctor != NULL && 2897 #ifdef INVARIANTS 2898 (!skipdbg || zone->uz_ctor != trash_ctor || 2899 zone->uz_dtor != trash_dtor) && 2900 #endif 2901 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2902 zone_free_item(zone, item, udata, SKIP_DTOR); 2903 goto fail; 2904 } 2905 #ifdef INVARIANTS 2906 if (!skipdbg) 2907 uma_dbg_alloc(zone, NULL, item); 2908 #endif 2909 if (flags & M_ZERO) 2910 uma_zero_item(item, zone); 2911 2912 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item, 2913 zone->uz_name, zone); 2914 2915 return (item); 2916 2917 fail: 2918 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)", 2919 zone->uz_name, zone); 2920 atomic_add_long(&zone->uz_fails, 1); 2921 return (NULL); 2922 } 2923 2924 /* See uma.h */ 2925 void 2926 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 2927 { 2928 uma_cache_t cache; 2929 uma_bucket_t bucket; 2930 uma_zone_domain_t zdom; 2931 int cpu, domain, lockfail; 2932 #ifdef INVARIANTS 2933 bool skipdbg; 2934 #endif 2935 2936 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2937 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA); 2938 2939 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 2940 zone->uz_name); 2941 2942 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2943 ("uma_zfree_arg: called with spinlock or critical section held")); 2944 2945 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 2946 if (item == NULL) 2947 return; 2948 #ifdef DEBUG_MEMGUARD 2949 if (is_memguard_addr(item)) { 2950 if (zone->uz_dtor != NULL) 2951 zone->uz_dtor(item, zone->uz_size, udata); 2952 if (zone->uz_fini != NULL) 2953 zone->uz_fini(item, zone->uz_size); 2954 memguard_free(item); 2955 return; 2956 } 2957 #endif 2958 #ifdef INVARIANTS 2959 skipdbg = uma_dbg_zskip(zone, item); 2960 if (skipdbg == false) { 2961 if (zone->uz_flags & UMA_ZONE_MALLOC) 2962 uma_dbg_free(zone, udata, item); 2963 else 2964 uma_dbg_free(zone, NULL, item); 2965 } 2966 if (zone->uz_dtor != NULL && (!skipdbg || 2967 zone->uz_dtor != trash_dtor || zone->uz_ctor != trash_ctor)) 2968 #else 2969 if (zone->uz_dtor != NULL) 2970 #endif 2971 zone->uz_dtor(item, zone->uz_size, udata); 2972 2973 /* 2974 * The race here is acceptable. If we miss it we'll just have to wait 2975 * a little longer for the limits to be reset. 2976 */ 2977 if (zone->uz_flags & UMA_ZFLAG_FULL) 2978 goto zfree_item; 2979 2980 /* 2981 * If possible, free to the per-CPU cache. There are two 2982 * requirements for safe access to the per-CPU cache: (1) the thread 2983 * accessing the cache must not be preempted or yield during access, 2984 * and (2) the thread must not migrate CPUs without switching which 2985 * cache it accesses. We rely on a critical section to prevent 2986 * preemption and migration. We release the critical section in 2987 * order to acquire the zone mutex if we are unable to free to the 2988 * current cache; when we re-acquire the critical section, we must 2989 * detect and handle migration if it has occurred. 2990 */ 2991 zfree_restart: 2992 critical_enter(); 2993 cpu = curcpu; 2994 cache = &zone->uz_cpu[cpu]; 2995 2996 zfree_start: 2997 /* 2998 * Try to free into the allocbucket first to give LIFO ordering 2999 * for cache-hot datastructures. Spill over into the freebucket 3000 * if necessary. Alloc will swap them if one runs dry. 3001 */ 3002 bucket = cache->uc_allocbucket; 3003 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 3004 bucket = cache->uc_freebucket; 3005 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 3006 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 3007 ("uma_zfree: Freeing to non free bucket index.")); 3008 bucket->ub_bucket[bucket->ub_cnt] = item; 3009 bucket->ub_cnt++; 3010 cache->uc_frees++; 3011 critical_exit(); 3012 return; 3013 } 3014 3015 /* 3016 * We must go back the zone, which requires acquiring the zone lock, 3017 * which in turn means we must release and re-acquire the critical 3018 * section. Since the critical section is released, we may be 3019 * preempted or migrate. As such, make sure not to maintain any 3020 * thread-local state specific to the cache from prior to releasing 3021 * the critical section. 3022 */ 3023 critical_exit(); 3024 if (zone->uz_count == 0 || bucketdisable) 3025 goto zfree_item; 3026 3027 lockfail = 0; 3028 if (ZONE_TRYLOCK(zone) == 0) { 3029 /* Record contention to size the buckets. */ 3030 ZONE_LOCK(zone); 3031 lockfail = 1; 3032 } 3033 critical_enter(); 3034 cpu = curcpu; 3035 cache = &zone->uz_cpu[cpu]; 3036 3037 /* 3038 * Since we have locked the zone we may as well send back our stats. 3039 */ 3040 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 3041 atomic_add_long(&zone->uz_frees, cache->uc_frees); 3042 cache->uc_allocs = 0; 3043 cache->uc_frees = 0; 3044 3045 bucket = cache->uc_freebucket; 3046 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 3047 ZONE_UNLOCK(zone); 3048 goto zfree_start; 3049 } 3050 cache->uc_freebucket = NULL; 3051 /* We are no longer associated with this CPU. */ 3052 critical_exit(); 3053 3054 if ((zone->uz_flags & UMA_ZONE_NUMA) != 0) 3055 domain = PCPU_GET(domain); 3056 else 3057 domain = 0; 3058 zdom = &zone->uz_domain[0]; 3059 3060 /* Can we throw this on the zone full list? */ 3061 if (bucket != NULL) { 3062 CTR3(KTR_UMA, 3063 "uma_zfree: zone %s(%p) putting bucket %p on free list", 3064 zone->uz_name, zone, bucket); 3065 /* ub_cnt is pointing to the last free item */ 3066 KASSERT(bucket->ub_cnt != 0, 3067 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 3068 if ((zone->uz_flags & UMA_ZONE_NOBUCKETCACHE) != 0) { 3069 ZONE_UNLOCK(zone); 3070 bucket_drain(zone, bucket); 3071 bucket_free(zone, bucket, udata); 3072 goto zfree_restart; 3073 } else 3074 LIST_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link); 3075 } 3076 3077 /* 3078 * We bump the uz count when the cache size is insufficient to 3079 * handle the working set. 3080 */ 3081 if (lockfail && zone->uz_count < BUCKET_MAX) 3082 zone->uz_count++; 3083 ZONE_UNLOCK(zone); 3084 3085 bucket = bucket_alloc(zone, udata, M_NOWAIT); 3086 CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p", 3087 zone->uz_name, zone, bucket); 3088 if (bucket) { 3089 critical_enter(); 3090 cpu = curcpu; 3091 cache = &zone->uz_cpu[cpu]; 3092 if (cache->uc_freebucket == NULL && 3093 ((zone->uz_flags & UMA_ZONE_NUMA) == 0 || 3094 domain == PCPU_GET(domain))) { 3095 cache->uc_freebucket = bucket; 3096 goto zfree_start; 3097 } 3098 /* 3099 * We lost the race, start over. We have to drop our 3100 * critical section to free the bucket. 3101 */ 3102 critical_exit(); 3103 bucket_free(zone, bucket, udata); 3104 goto zfree_restart; 3105 } 3106 3107 /* 3108 * If nothing else caught this, we'll just do an internal free. 3109 */ 3110 zfree_item: 3111 zone_free_item(zone, item, udata, SKIP_DTOR); 3112 3113 return; 3114 } 3115 3116 void 3117 uma_zfree_domain(uma_zone_t zone, void *item, void *udata) 3118 { 3119 3120 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3121 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA); 3122 3123 CTR2(KTR_UMA, "uma_zfree_domain thread %x zone %s", curthread, 3124 zone->uz_name); 3125 3126 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3127 ("uma_zfree_domain: called with spinlock or critical section held")); 3128 3129 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 3130 if (item == NULL) 3131 return; 3132 zone_free_item(zone, item, udata, SKIP_NONE); 3133 } 3134 3135 static void 3136 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) 3137 { 3138 uma_domain_t dom; 3139 uint8_t freei; 3140 3141 mtx_assert(&keg->uk_lock, MA_OWNED); 3142 MPASS(keg == slab->us_keg); 3143 3144 dom = &keg->uk_domain[slab->us_domain]; 3145 3146 /* Do we need to remove from any lists? */ 3147 if (slab->us_freecount+1 == keg->uk_ipers) { 3148 LIST_REMOVE(slab, us_link); 3149 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); 3150 } else if (slab->us_freecount == 0) { 3151 LIST_REMOVE(slab, us_link); 3152 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 3153 } 3154 3155 /* Slab management. */ 3156 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3157 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); 3158 slab->us_freecount++; 3159 3160 /* Keg statistics. */ 3161 keg->uk_free++; 3162 } 3163 3164 static void 3165 zone_release(uma_zone_t zone, void **bucket, int cnt) 3166 { 3167 void *item; 3168 uma_slab_t slab; 3169 uma_keg_t keg; 3170 uint8_t *mem; 3171 int clearfull; 3172 int i; 3173 3174 clearfull = 0; 3175 keg = zone_first_keg(zone); 3176 KEG_LOCK(keg); 3177 for (i = 0; i < cnt; i++) { 3178 item = bucket[i]; 3179 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 3180 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 3181 if (zone->uz_flags & UMA_ZONE_HASH) { 3182 slab = hash_sfind(&keg->uk_hash, mem); 3183 } else { 3184 mem += keg->uk_pgoff; 3185 slab = (uma_slab_t)mem; 3186 } 3187 } else { 3188 slab = vtoslab((vm_offset_t)item); 3189 if (slab->us_keg != keg) { 3190 KEG_UNLOCK(keg); 3191 keg = slab->us_keg; 3192 KEG_LOCK(keg); 3193 } 3194 } 3195 slab_free_item(keg, slab, item); 3196 if (keg->uk_flags & UMA_ZFLAG_FULL) { 3197 if (keg->uk_pages < keg->uk_maxpages) { 3198 keg->uk_flags &= ~UMA_ZFLAG_FULL; 3199 clearfull = 1; 3200 } 3201 3202 /* 3203 * We can handle one more allocation. Since we're 3204 * clearing ZFLAG_FULL, wake up all procs blocked 3205 * on pages. This should be uncommon, so keeping this 3206 * simple for now (rather than adding count of blocked 3207 * threads etc). 3208 */ 3209 wakeup(keg); 3210 } 3211 } 3212 KEG_UNLOCK(keg); 3213 if (clearfull) { 3214 ZONE_LOCK(zone); 3215 zone->uz_flags &= ~UMA_ZFLAG_FULL; 3216 wakeup(zone); 3217 ZONE_UNLOCK(zone); 3218 } 3219 3220 } 3221 3222 /* 3223 * Frees a single item to any zone. 3224 * 3225 * Arguments: 3226 * zone The zone to free to 3227 * item The item we're freeing 3228 * udata User supplied data for the dtor 3229 * skip Skip dtors and finis 3230 */ 3231 static void 3232 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 3233 { 3234 #ifdef INVARIANTS 3235 bool skipdbg; 3236 3237 skipdbg = uma_dbg_zskip(zone, item); 3238 if (skip == SKIP_NONE && !skipdbg) { 3239 if (zone->uz_flags & UMA_ZONE_MALLOC) 3240 uma_dbg_free(zone, udata, item); 3241 else 3242 uma_dbg_free(zone, NULL, item); 3243 } 3244 3245 if (skip < SKIP_DTOR && zone->uz_dtor != NULL && 3246 (!skipdbg || zone->uz_dtor != trash_dtor || 3247 zone->uz_ctor != trash_ctor)) 3248 #else 3249 if (skip < SKIP_DTOR && zone->uz_dtor != NULL) 3250 #endif 3251 zone->uz_dtor(item, zone->uz_size, udata); 3252 3253 if (skip < SKIP_FINI && zone->uz_fini) 3254 zone->uz_fini(item, zone->uz_size); 3255 3256 atomic_add_long(&zone->uz_frees, 1); 3257 zone->uz_release(zone->uz_arg, &item, 1); 3258 } 3259 3260 /* See uma.h */ 3261 int 3262 uma_zone_set_max(uma_zone_t zone, int nitems) 3263 { 3264 uma_keg_t keg; 3265 3266 keg = zone_first_keg(zone); 3267 if (keg == NULL) 3268 return (0); 3269 KEG_LOCK(keg); 3270 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 3271 if (keg->uk_maxpages * keg->uk_ipers < nitems) 3272 keg->uk_maxpages += keg->uk_ppera; 3273 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers; 3274 KEG_UNLOCK(keg); 3275 3276 return (nitems); 3277 } 3278 3279 /* See uma.h */ 3280 int 3281 uma_zone_get_max(uma_zone_t zone) 3282 { 3283 int nitems; 3284 uma_keg_t keg; 3285 3286 keg = zone_first_keg(zone); 3287 if (keg == NULL) 3288 return (0); 3289 KEG_LOCK(keg); 3290 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers; 3291 KEG_UNLOCK(keg); 3292 3293 return (nitems); 3294 } 3295 3296 /* See uma.h */ 3297 void 3298 uma_zone_set_warning(uma_zone_t zone, const char *warning) 3299 { 3300 3301 ZONE_LOCK(zone); 3302 zone->uz_warning = warning; 3303 ZONE_UNLOCK(zone); 3304 } 3305 3306 /* See uma.h */ 3307 void 3308 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) 3309 { 3310 3311 ZONE_LOCK(zone); 3312 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); 3313 ZONE_UNLOCK(zone); 3314 } 3315 3316 /* See uma.h */ 3317 int 3318 uma_zone_get_cur(uma_zone_t zone) 3319 { 3320 int64_t nitems; 3321 u_int i; 3322 3323 ZONE_LOCK(zone); 3324 nitems = zone->uz_allocs - zone->uz_frees; 3325 CPU_FOREACH(i) { 3326 /* 3327 * See the comment in sysctl_vm_zone_stats() regarding the 3328 * safety of accessing the per-cpu caches. With the zone lock 3329 * held, it is safe, but can potentially result in stale data. 3330 */ 3331 nitems += zone->uz_cpu[i].uc_allocs - 3332 zone->uz_cpu[i].uc_frees; 3333 } 3334 ZONE_UNLOCK(zone); 3335 3336 return (nitems < 0 ? 0 : nitems); 3337 } 3338 3339 /* See uma.h */ 3340 void 3341 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 3342 { 3343 uma_keg_t keg; 3344 3345 keg = zone_first_keg(zone); 3346 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 3347 KEG_LOCK(keg); 3348 KASSERT(keg->uk_pages == 0, 3349 ("uma_zone_set_init on non-empty keg")); 3350 keg->uk_init = uminit; 3351 KEG_UNLOCK(keg); 3352 } 3353 3354 /* See uma.h */ 3355 void 3356 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 3357 { 3358 uma_keg_t keg; 3359 3360 keg = zone_first_keg(zone); 3361 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type")); 3362 KEG_LOCK(keg); 3363 KASSERT(keg->uk_pages == 0, 3364 ("uma_zone_set_fini on non-empty keg")); 3365 keg->uk_fini = fini; 3366 KEG_UNLOCK(keg); 3367 } 3368 3369 /* See uma.h */ 3370 void 3371 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 3372 { 3373 3374 ZONE_LOCK(zone); 3375 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3376 ("uma_zone_set_zinit on non-empty keg")); 3377 zone->uz_init = zinit; 3378 ZONE_UNLOCK(zone); 3379 } 3380 3381 /* See uma.h */ 3382 void 3383 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 3384 { 3385 3386 ZONE_LOCK(zone); 3387 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3388 ("uma_zone_set_zfini on non-empty keg")); 3389 zone->uz_fini = zfini; 3390 ZONE_UNLOCK(zone); 3391 } 3392 3393 /* See uma.h */ 3394 /* XXX uk_freef is not actually used with the zone locked */ 3395 void 3396 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 3397 { 3398 uma_keg_t keg; 3399 3400 keg = zone_first_keg(zone); 3401 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type")); 3402 KEG_LOCK(keg); 3403 keg->uk_freef = freef; 3404 KEG_UNLOCK(keg); 3405 } 3406 3407 /* See uma.h */ 3408 /* XXX uk_allocf is not actually used with the zone locked */ 3409 void 3410 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3411 { 3412 uma_keg_t keg; 3413 3414 keg = zone_first_keg(zone); 3415 KEG_LOCK(keg); 3416 keg->uk_allocf = allocf; 3417 KEG_UNLOCK(keg); 3418 } 3419 3420 /* See uma.h */ 3421 void 3422 uma_zone_reserve(uma_zone_t zone, int items) 3423 { 3424 uma_keg_t keg; 3425 3426 keg = zone_first_keg(zone); 3427 if (keg == NULL) 3428 return; 3429 KEG_LOCK(keg); 3430 keg->uk_reserve = items; 3431 KEG_UNLOCK(keg); 3432 3433 return; 3434 } 3435 3436 /* See uma.h */ 3437 int 3438 uma_zone_reserve_kva(uma_zone_t zone, int count) 3439 { 3440 uma_keg_t keg; 3441 vm_offset_t kva; 3442 u_int pages; 3443 3444 keg = zone_first_keg(zone); 3445 if (keg == NULL) 3446 return (0); 3447 pages = count / keg->uk_ipers; 3448 3449 if (pages * keg->uk_ipers < count) 3450 pages++; 3451 pages *= keg->uk_ppera; 3452 3453 #ifdef UMA_MD_SMALL_ALLOC 3454 if (keg->uk_ppera > 1) { 3455 #else 3456 if (1) { 3457 #endif 3458 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE); 3459 if (kva == 0) 3460 return (0); 3461 } else 3462 kva = 0; 3463 KEG_LOCK(keg); 3464 keg->uk_kva = kva; 3465 keg->uk_offset = 0; 3466 keg->uk_maxpages = pages; 3467 #ifdef UMA_MD_SMALL_ALLOC 3468 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3469 #else 3470 keg->uk_allocf = noobj_alloc; 3471 #endif 3472 keg->uk_flags |= UMA_ZONE_NOFREE; 3473 KEG_UNLOCK(keg); 3474 3475 return (1); 3476 } 3477 3478 /* See uma.h */ 3479 void 3480 uma_prealloc(uma_zone_t zone, int items) 3481 { 3482 uma_domain_t dom; 3483 uma_slab_t slab; 3484 uma_keg_t keg; 3485 int domain, slabs; 3486 3487 keg = zone_first_keg(zone); 3488 if (keg == NULL) 3489 return; 3490 KEG_LOCK(keg); 3491 slabs = items / keg->uk_ipers; 3492 domain = 0; 3493 if (slabs * keg->uk_ipers < items) 3494 slabs++; 3495 while (slabs > 0) { 3496 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK); 3497 if (slab == NULL) 3498 break; 3499 MPASS(slab->us_keg == keg); 3500 dom = &keg->uk_domain[slab->us_domain]; 3501 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); 3502 slabs--; 3503 domain = (domain + 1) % vm_ndomains; 3504 } 3505 KEG_UNLOCK(keg); 3506 } 3507 3508 /* See uma.h */ 3509 static void 3510 uma_reclaim_locked(bool kmem_danger) 3511 { 3512 3513 CTR0(KTR_UMA, "UMA: vm asked us to release pages!"); 3514 sx_assert(&uma_drain_lock, SA_XLOCKED); 3515 bucket_enable(); 3516 zone_foreach(zone_drain); 3517 if (vm_page_count_min() || kmem_danger) { 3518 cache_drain_safe(NULL); 3519 zone_foreach(zone_drain); 3520 } 3521 /* 3522 * Some slabs may have been freed but this zone will be visited early 3523 * we visit again so that we can free pages that are empty once other 3524 * zones are drained. We have to do the same for buckets. 3525 */ 3526 zone_drain(slabzone); 3527 bucket_zone_drain(); 3528 } 3529 3530 void 3531 uma_reclaim(void) 3532 { 3533 3534 sx_xlock(&uma_drain_lock); 3535 uma_reclaim_locked(false); 3536 sx_xunlock(&uma_drain_lock); 3537 } 3538 3539 static volatile int uma_reclaim_needed; 3540 3541 void 3542 uma_reclaim_wakeup(void) 3543 { 3544 3545 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0) 3546 wakeup(uma_reclaim); 3547 } 3548 3549 void 3550 uma_reclaim_worker(void *arg __unused) 3551 { 3552 3553 for (;;) { 3554 sx_xlock(&uma_drain_lock); 3555 while (atomic_load_int(&uma_reclaim_needed) == 0) 3556 sx_sleep(uma_reclaim, &uma_drain_lock, PVM, "umarcl", 3557 hz); 3558 sx_xunlock(&uma_drain_lock); 3559 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); 3560 sx_xlock(&uma_drain_lock); 3561 uma_reclaim_locked(true); 3562 atomic_store_int(&uma_reclaim_needed, 0); 3563 sx_xunlock(&uma_drain_lock); 3564 /* Don't fire more than once per-second. */ 3565 pause("umarclslp", hz); 3566 } 3567 } 3568 3569 /* See uma.h */ 3570 int 3571 uma_zone_exhausted(uma_zone_t zone) 3572 { 3573 int full; 3574 3575 ZONE_LOCK(zone); 3576 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3577 ZONE_UNLOCK(zone); 3578 return (full); 3579 } 3580 3581 int 3582 uma_zone_exhausted_nolock(uma_zone_t zone) 3583 { 3584 return (zone->uz_flags & UMA_ZFLAG_FULL); 3585 } 3586 3587 void * 3588 uma_large_malloc_domain(vm_size_t size, int domain, int wait) 3589 { 3590 vm_offset_t addr; 3591 uma_slab_t slab; 3592 3593 slab = zone_alloc_item(slabzone, NULL, domain, wait); 3594 if (slab == NULL) 3595 return (NULL); 3596 if (domain == UMA_ANYDOMAIN) 3597 addr = kmem_malloc(kernel_arena, size, wait); 3598 else 3599 addr = kmem_malloc_domain(domain, size, wait); 3600 if (addr != 0) { 3601 vsetslab(addr, slab); 3602 slab->us_data = (void *)addr; 3603 slab->us_flags = UMA_SLAB_KERNEL | UMA_SLAB_MALLOC; 3604 slab->us_size = size; 3605 slab->us_domain = vm_phys_domain(PHYS_TO_VM_PAGE( 3606 pmap_kextract(addr))); 3607 uma_total_inc(size); 3608 } else { 3609 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3610 } 3611 3612 return ((void *)addr); 3613 } 3614 3615 void * 3616 uma_large_malloc(vm_size_t size, int wait) 3617 { 3618 3619 return uma_large_malloc_domain(size, UMA_ANYDOMAIN, wait); 3620 } 3621 3622 void 3623 uma_large_free(uma_slab_t slab) 3624 { 3625 3626 KASSERT((slab->us_flags & UMA_SLAB_KERNEL) != 0, 3627 ("uma_large_free: Memory not allocated with uma_large_malloc.")); 3628 kmem_free(kernel_arena, (vm_offset_t)slab->us_data, slab->us_size); 3629 uma_total_dec(slab->us_size); 3630 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3631 } 3632 3633 static void 3634 uma_zero_item(void *item, uma_zone_t zone) 3635 { 3636 3637 bzero(item, zone->uz_size); 3638 } 3639 3640 unsigned long 3641 uma_limit(void) 3642 { 3643 3644 return (uma_kmem_limit); 3645 } 3646 3647 void 3648 uma_set_limit(unsigned long limit) 3649 { 3650 3651 uma_kmem_limit = limit; 3652 } 3653 3654 unsigned long 3655 uma_size(void) 3656 { 3657 3658 return (uma_kmem_total); 3659 } 3660 3661 long 3662 uma_avail(void) 3663 { 3664 3665 return (uma_kmem_limit - uma_kmem_total); 3666 } 3667 3668 void 3669 uma_print_stats(void) 3670 { 3671 zone_foreach(uma_print_zone); 3672 } 3673 3674 static void 3675 slab_print(uma_slab_t slab) 3676 { 3677 printf("slab: keg %p, data %p, freecount %d\n", 3678 slab->us_keg, slab->us_data, slab->us_freecount); 3679 } 3680 3681 static void 3682 cache_print(uma_cache_t cache) 3683 { 3684 printf("alloc: %p(%d), free: %p(%d)\n", 3685 cache->uc_allocbucket, 3686 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3687 cache->uc_freebucket, 3688 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3689 } 3690 3691 static void 3692 uma_print_keg(uma_keg_t keg) 3693 { 3694 uma_domain_t dom; 3695 uma_slab_t slab; 3696 int i; 3697 3698 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3699 "out %d free %d limit %d\n", 3700 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3701 keg->uk_ipers, keg->uk_ppera, 3702 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, 3703 keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3704 for (i = 0; i < vm_ndomains; i++) { 3705 dom = &keg->uk_domain[i]; 3706 printf("Part slabs:\n"); 3707 LIST_FOREACH(slab, &dom->ud_part_slab, us_link) 3708 slab_print(slab); 3709 printf("Free slabs:\n"); 3710 LIST_FOREACH(slab, &dom->ud_free_slab, us_link) 3711 slab_print(slab); 3712 printf("Full slabs:\n"); 3713 LIST_FOREACH(slab, &dom->ud_full_slab, us_link) 3714 slab_print(slab); 3715 } 3716 } 3717 3718 void 3719 uma_print_zone(uma_zone_t zone) 3720 { 3721 uma_cache_t cache; 3722 uma_klink_t kl; 3723 int i; 3724 3725 printf("zone: %s(%p) size %d flags %#x\n", 3726 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3727 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3728 uma_print_keg(kl->kl_keg); 3729 CPU_FOREACH(i) { 3730 cache = &zone->uz_cpu[i]; 3731 printf("CPU %d Cache:\n", i); 3732 cache_print(cache); 3733 } 3734 } 3735 3736 #ifdef DDB 3737 /* 3738 * Generate statistics across both the zone and its per-cpu cache's. Return 3739 * desired statistics if the pointer is non-NULL for that statistic. 3740 * 3741 * Note: does not update the zone statistics, as it can't safely clear the 3742 * per-CPU cache statistic. 3743 * 3744 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3745 * safe from off-CPU; we should modify the caches to track this information 3746 * directly so that we don't have to. 3747 */ 3748 static void 3749 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3750 uint64_t *freesp, uint64_t *sleepsp) 3751 { 3752 uma_cache_t cache; 3753 uint64_t allocs, frees, sleeps; 3754 int cachefree, cpu; 3755 3756 allocs = frees = sleeps = 0; 3757 cachefree = 0; 3758 CPU_FOREACH(cpu) { 3759 cache = &z->uz_cpu[cpu]; 3760 if (cache->uc_allocbucket != NULL) 3761 cachefree += cache->uc_allocbucket->ub_cnt; 3762 if (cache->uc_freebucket != NULL) 3763 cachefree += cache->uc_freebucket->ub_cnt; 3764 allocs += cache->uc_allocs; 3765 frees += cache->uc_frees; 3766 } 3767 allocs += z->uz_allocs; 3768 frees += z->uz_frees; 3769 sleeps += z->uz_sleeps; 3770 if (cachefreep != NULL) 3771 *cachefreep = cachefree; 3772 if (allocsp != NULL) 3773 *allocsp = allocs; 3774 if (freesp != NULL) 3775 *freesp = frees; 3776 if (sleepsp != NULL) 3777 *sleepsp = sleeps; 3778 } 3779 #endif /* DDB */ 3780 3781 static int 3782 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3783 { 3784 uma_keg_t kz; 3785 uma_zone_t z; 3786 int count; 3787 3788 count = 0; 3789 rw_rlock(&uma_rwlock); 3790 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3791 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3792 count++; 3793 } 3794 rw_runlock(&uma_rwlock); 3795 return (sysctl_handle_int(oidp, &count, 0, req)); 3796 } 3797 3798 static int 3799 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3800 { 3801 struct uma_stream_header ush; 3802 struct uma_type_header uth; 3803 struct uma_percpu_stat *ups; 3804 uma_bucket_t bucket; 3805 uma_zone_domain_t zdom; 3806 struct sbuf sbuf; 3807 uma_cache_t cache; 3808 uma_klink_t kl; 3809 uma_keg_t kz; 3810 uma_zone_t z; 3811 uma_keg_t k; 3812 int count, error, i; 3813 3814 error = sysctl_wire_old_buffer(req, 0); 3815 if (error != 0) 3816 return (error); 3817 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3818 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); 3819 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK); 3820 3821 count = 0; 3822 rw_rlock(&uma_rwlock); 3823 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3824 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3825 count++; 3826 } 3827 3828 /* 3829 * Insert stream header. 3830 */ 3831 bzero(&ush, sizeof(ush)); 3832 ush.ush_version = UMA_STREAM_VERSION; 3833 ush.ush_maxcpus = (mp_maxid + 1); 3834 ush.ush_count = count; 3835 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3836 3837 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3838 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3839 bzero(&uth, sizeof(uth)); 3840 ZONE_LOCK(z); 3841 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3842 uth.uth_align = kz->uk_align; 3843 uth.uth_size = kz->uk_size; 3844 uth.uth_rsize = kz->uk_rsize; 3845 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3846 k = kl->kl_keg; 3847 uth.uth_maxpages += k->uk_maxpages; 3848 uth.uth_pages += k->uk_pages; 3849 uth.uth_keg_free += k->uk_free; 3850 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3851 * k->uk_ipers; 3852 } 3853 3854 /* 3855 * A zone is secondary is it is not the first entry 3856 * on the keg's zone list. 3857 */ 3858 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3859 (LIST_FIRST(&kz->uk_zones) != z)) 3860 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3861 3862 for (i = 0; i < vm_ndomains; i++) { 3863 zdom = &z->uz_domain[i]; 3864 LIST_FOREACH(bucket, &zdom->uzd_buckets, 3865 ub_link) 3866 uth.uth_zone_free += bucket->ub_cnt; 3867 } 3868 uth.uth_allocs = z->uz_allocs; 3869 uth.uth_frees = z->uz_frees; 3870 uth.uth_fails = z->uz_fails; 3871 uth.uth_sleeps = z->uz_sleeps; 3872 /* 3873 * While it is not normally safe to access the cache 3874 * bucket pointers while not on the CPU that owns the 3875 * cache, we only allow the pointers to be exchanged 3876 * without the zone lock held, not invalidated, so 3877 * accept the possible race associated with bucket 3878 * exchange during monitoring. 3879 */ 3880 for (i = 0; i < mp_maxid + 1; i++) { 3881 bzero(&ups[i], sizeof(*ups)); 3882 if (kz->uk_flags & UMA_ZFLAG_INTERNAL || 3883 CPU_ABSENT(i)) 3884 continue; 3885 cache = &z->uz_cpu[i]; 3886 if (cache->uc_allocbucket != NULL) 3887 ups[i].ups_cache_free += 3888 cache->uc_allocbucket->ub_cnt; 3889 if (cache->uc_freebucket != NULL) 3890 ups[i].ups_cache_free += 3891 cache->uc_freebucket->ub_cnt; 3892 ups[i].ups_allocs = cache->uc_allocs; 3893 ups[i].ups_frees = cache->uc_frees; 3894 } 3895 ZONE_UNLOCK(z); 3896 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 3897 for (i = 0; i < mp_maxid + 1; i++) 3898 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 3899 } 3900 } 3901 rw_runlock(&uma_rwlock); 3902 error = sbuf_finish(&sbuf); 3903 sbuf_delete(&sbuf); 3904 free(ups, M_TEMP); 3905 return (error); 3906 } 3907 3908 int 3909 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 3910 { 3911 uma_zone_t zone = *(uma_zone_t *)arg1; 3912 int error, max; 3913 3914 max = uma_zone_get_max(zone); 3915 error = sysctl_handle_int(oidp, &max, 0, req); 3916 if (error || !req->newptr) 3917 return (error); 3918 3919 uma_zone_set_max(zone, max); 3920 3921 return (0); 3922 } 3923 3924 int 3925 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 3926 { 3927 uma_zone_t zone = *(uma_zone_t *)arg1; 3928 int cur; 3929 3930 cur = uma_zone_get_cur(zone); 3931 return (sysctl_handle_int(oidp, &cur, 0, req)); 3932 } 3933 3934 #ifdef INVARIANTS 3935 static uma_slab_t 3936 uma_dbg_getslab(uma_zone_t zone, void *item) 3937 { 3938 uma_slab_t slab; 3939 uma_keg_t keg; 3940 uint8_t *mem; 3941 3942 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 3943 if (zone->uz_flags & UMA_ZONE_VTOSLAB) { 3944 slab = vtoslab((vm_offset_t)mem); 3945 } else { 3946 /* 3947 * It is safe to return the slab here even though the 3948 * zone is unlocked because the item's allocation state 3949 * essentially holds a reference. 3950 */ 3951 ZONE_LOCK(zone); 3952 keg = LIST_FIRST(&zone->uz_kegs)->kl_keg; 3953 if (keg->uk_flags & UMA_ZONE_HASH) 3954 slab = hash_sfind(&keg->uk_hash, mem); 3955 else 3956 slab = (uma_slab_t)(mem + keg->uk_pgoff); 3957 ZONE_UNLOCK(zone); 3958 } 3959 3960 return (slab); 3961 } 3962 3963 static bool 3964 uma_dbg_zskip(uma_zone_t zone, void *mem) 3965 { 3966 uma_keg_t keg; 3967 3968 if ((keg = zone_first_keg(zone)) == NULL) 3969 return (true); 3970 3971 return (uma_dbg_kskip(keg, mem)); 3972 } 3973 3974 static bool 3975 uma_dbg_kskip(uma_keg_t keg, void *mem) 3976 { 3977 uintptr_t idx; 3978 3979 if (dbg_divisor == 0) 3980 return (true); 3981 3982 if (dbg_divisor == 1) 3983 return (false); 3984 3985 idx = (uintptr_t)mem >> PAGE_SHIFT; 3986 if (keg->uk_ipers > 1) { 3987 idx *= keg->uk_ipers; 3988 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize; 3989 } 3990 3991 if ((idx / dbg_divisor) * dbg_divisor != idx) { 3992 counter_u64_add(uma_skip_cnt, 1); 3993 return (true); 3994 } 3995 counter_u64_add(uma_dbg_cnt, 1); 3996 3997 return (false); 3998 } 3999 4000 /* 4001 * Set up the slab's freei data such that uma_dbg_free can function. 4002 * 4003 */ 4004 static void 4005 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) 4006 { 4007 uma_keg_t keg; 4008 int freei; 4009 4010 if (slab == NULL) { 4011 slab = uma_dbg_getslab(zone, item); 4012 if (slab == NULL) 4013 panic("uma: item %p did not belong to zone %s\n", 4014 item, zone->uz_name); 4015 } 4016 keg = slab->us_keg; 4017 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 4018 4019 if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) 4020 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n", 4021 item, zone, zone->uz_name, slab, freei); 4022 BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); 4023 4024 return; 4025 } 4026 4027 /* 4028 * Verifies freed addresses. Checks for alignment, valid slab membership 4029 * and duplicate frees. 4030 * 4031 */ 4032 static void 4033 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) 4034 { 4035 uma_keg_t keg; 4036 int freei; 4037 4038 if (slab == NULL) { 4039 slab = uma_dbg_getslab(zone, item); 4040 if (slab == NULL) 4041 panic("uma: Freed item %p did not belong to zone %s\n", 4042 item, zone->uz_name); 4043 } 4044 keg = slab->us_keg; 4045 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 4046 4047 if (freei >= keg->uk_ipers) 4048 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n", 4049 item, zone, zone->uz_name, slab, freei); 4050 4051 if (((freei * keg->uk_rsize) + slab->us_data) != item) 4052 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n", 4053 item, zone, zone->uz_name, slab, freei); 4054 4055 if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) 4056 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n", 4057 item, zone, zone->uz_name, slab, freei); 4058 4059 BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); 4060 } 4061 #endif /* INVARIANTS */ 4062 4063 #ifdef DDB 4064 DB_SHOW_COMMAND(uma, db_show_uma) 4065 { 4066 uma_bucket_t bucket; 4067 uma_keg_t kz; 4068 uma_zone_t z; 4069 uma_zone_domain_t zdom; 4070 uint64_t allocs, frees, sleeps; 4071 int cachefree, i; 4072 4073 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used", 4074 "Free", "Requests", "Sleeps", "Bucket"); 4075 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4076 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 4077 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 4078 allocs = z->uz_allocs; 4079 frees = z->uz_frees; 4080 sleeps = z->uz_sleeps; 4081 cachefree = 0; 4082 } else 4083 uma_zone_sumstat(z, &cachefree, &allocs, 4084 &frees, &sleeps); 4085 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 4086 (LIST_FIRST(&kz->uk_zones) != z))) 4087 cachefree += kz->uk_free; 4088 for (i = 0; i < vm_ndomains; i++) { 4089 zdom = &z->uz_domain[i]; 4090 LIST_FOREACH(bucket, &zdom->uzd_buckets, 4091 ub_link) 4092 cachefree += bucket->ub_cnt; 4093 } 4094 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n", 4095 z->uz_name, (uintmax_t)kz->uk_size, 4096 (intmax_t)(allocs - frees), cachefree, 4097 (uintmax_t)allocs, sleeps, z->uz_count); 4098 if (db_pager_quit) 4099 return; 4100 } 4101 } 4102 } 4103 4104 DB_SHOW_COMMAND(umacache, db_show_umacache) 4105 { 4106 uma_bucket_t bucket; 4107 uma_zone_t z; 4108 uma_zone_domain_t zdom; 4109 uint64_t allocs, frees; 4110 int cachefree, i; 4111 4112 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 4113 "Requests", "Bucket"); 4114 LIST_FOREACH(z, &uma_cachezones, uz_link) { 4115 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL); 4116 for (i = 0; i < vm_ndomains; i++) { 4117 zdom = &z->uz_domain[i]; 4118 LIST_FOREACH(bucket, &zdom->uzd_buckets, ub_link) 4119 cachefree += bucket->ub_cnt; 4120 } 4121 db_printf("%18s %8ju %8jd %8d %12ju %8u\n", 4122 z->uz_name, (uintmax_t)z->uz_size, 4123 (intmax_t)(allocs - frees), cachefree, 4124 (uintmax_t)allocs, z->uz_count); 4125 if (db_pager_quit) 4126 return; 4127 } 4128 } 4129 #endif /* DDB */ 4130