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/domainset.h> 63 #include <sys/eventhandler.h> 64 #include <sys/kernel.h> 65 #include <sys/types.h> 66 #include <sys/limits.h> 67 #include <sys/queue.h> 68 #include <sys/malloc.h> 69 #include <sys/ktr.h> 70 #include <sys/lock.h> 71 #include <sys/sysctl.h> 72 #include <sys/mutex.h> 73 #include <sys/proc.h> 74 #include <sys/random.h> 75 #include <sys/rwlock.h> 76 #include <sys/sbuf.h> 77 #include <sys/sched.h> 78 #include <sys/smp.h> 79 #include <sys/taskqueue.h> 80 #include <sys/vmmeter.h> 81 82 #include <vm/vm.h> 83 #include <vm/vm_domainset.h> 84 #include <vm/vm_object.h> 85 #include <vm/vm_page.h> 86 #include <vm/vm_pageout.h> 87 #include <vm/vm_param.h> 88 #include <vm/vm_phys.h> 89 #include <vm/vm_pagequeue.h> 90 #include <vm/vm_map.h> 91 #include <vm/vm_kern.h> 92 #include <vm/vm_extern.h> 93 #include <vm/uma.h> 94 #include <vm/uma_int.h> 95 #include <vm/uma_dbg.h> 96 97 #include <ddb/ddb.h> 98 99 #ifdef DEBUG_MEMGUARD 100 #include <vm/memguard.h> 101 #endif 102 103 /* 104 * This is the zone and keg from which all zones are spawned. 105 */ 106 static uma_zone_t kegs; 107 static uma_zone_t zones; 108 109 /* This is the zone from which all offpage uma_slab_ts are allocated. */ 110 static uma_zone_t slabzone; 111 112 /* 113 * The initial hash tables come out of this zone so they can be allocated 114 * prior to malloc coming up. 115 */ 116 static uma_zone_t hashzone; 117 118 /* The boot-time adjusted value for cache line alignment. */ 119 int uma_align_cache = 64 - 1; 120 121 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 122 123 /* 124 * Are we allowed to allocate buckets? 125 */ 126 static int bucketdisable = 1; 127 128 /* Linked list of all kegs in the system */ 129 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 130 131 /* Linked list of all cache-only zones in the system */ 132 static LIST_HEAD(,uma_zone) uma_cachezones = 133 LIST_HEAD_INITIALIZER(uma_cachezones); 134 135 /* This RW lock protects the keg list */ 136 static struct rwlock_padalign __exclusive_cache_line uma_rwlock; 137 138 /* 139 * Pointer and counter to pool of pages, that is preallocated at 140 * startup to bootstrap UMA. 141 */ 142 static char *bootmem; 143 static int boot_pages; 144 145 static struct sx uma_drain_lock; 146 147 /* kmem soft limit. */ 148 static unsigned long uma_kmem_limit = LONG_MAX; 149 static volatile unsigned long uma_kmem_total; 150 151 /* Is the VM done starting up? */ 152 static enum { BOOT_COLD = 0, BOOT_STRAPPED, BOOT_PAGEALLOC, BOOT_BUCKETS, 153 BOOT_RUNNING } booted = BOOT_COLD; 154 155 /* 156 * This is the handle used to schedule events that need to happen 157 * outside of the allocation fast path. 158 */ 159 static struct callout uma_callout; 160 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 161 162 /* 163 * This structure is passed as the zone ctor arg so that I don't have to create 164 * a special allocation function just for zones. 165 */ 166 struct uma_zctor_args { 167 const char *name; 168 size_t size; 169 uma_ctor ctor; 170 uma_dtor dtor; 171 uma_init uminit; 172 uma_fini fini; 173 uma_import import; 174 uma_release release; 175 void *arg; 176 uma_keg_t keg; 177 int align; 178 uint32_t flags; 179 }; 180 181 struct uma_kctor_args { 182 uma_zone_t zone; 183 size_t size; 184 uma_init uminit; 185 uma_fini fini; 186 int align; 187 uint32_t flags; 188 }; 189 190 struct uma_bucket_zone { 191 uma_zone_t ubz_zone; 192 char *ubz_name; 193 int ubz_entries; /* Number of items it can hold. */ 194 int ubz_maxsize; /* Maximum allocation size per-item. */ 195 }; 196 197 /* 198 * Compute the actual number of bucket entries to pack them in power 199 * of two sizes for more efficient space utilization. 200 */ 201 #define BUCKET_SIZE(n) \ 202 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 203 204 #define BUCKET_MAX BUCKET_SIZE(256) 205 206 struct uma_bucket_zone bucket_zones[] = { 207 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 208 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 209 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 210 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 211 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 212 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 213 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 214 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 215 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 216 { NULL, NULL, 0} 217 }; 218 219 /* 220 * Flags and enumerations to be passed to internal functions. 221 */ 222 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; 223 224 #define UMA_ANYDOMAIN -1 /* Special value for domain search. */ 225 226 /* Prototypes.. */ 227 228 int uma_startup_count(int); 229 void uma_startup(void *, int); 230 void uma_startup1(void); 231 void uma_startup2(void); 232 233 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 234 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 235 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 236 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 237 static void page_free(void *, vm_size_t, uint8_t); 238 static void pcpu_page_free(void *, vm_size_t, uint8_t); 239 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int); 240 static void cache_drain(uma_zone_t); 241 static void bucket_drain(uma_zone_t, uma_bucket_t); 242 static void bucket_cache_drain(uma_zone_t zone); 243 static int keg_ctor(void *, int, void *, int); 244 static void keg_dtor(void *, int, void *); 245 static int zone_ctor(void *, int, void *, int); 246 static void zone_dtor(void *, int, void *); 247 static int zero_init(void *, int, int); 248 static void keg_small_init(uma_keg_t keg); 249 static void keg_large_init(uma_keg_t keg); 250 static void zone_foreach(void (*zfunc)(uma_zone_t)); 251 static void zone_timeout(uma_zone_t zone); 252 static int hash_alloc(struct uma_hash *); 253 static int hash_expand(struct uma_hash *, struct uma_hash *); 254 static void hash_free(struct uma_hash *hash); 255 static void uma_timeout(void *); 256 static void uma_startup3(void); 257 static void *zone_alloc_item(uma_zone_t, void *, int, int); 258 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 259 static void bucket_enable(void); 260 static void bucket_init(void); 261 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 262 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 263 static void bucket_zone_drain(void); 264 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int); 265 static uma_slab_t zone_fetch_slab(uma_zone_t, uma_keg_t, int, int); 266 static uma_slab_t zone_fetch_slab_multi(uma_zone_t, uma_keg_t, int, int); 267 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 268 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); 269 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 270 uma_fini fini, int align, uint32_t flags); 271 static int zone_import(uma_zone_t, void **, int, int, int); 272 static void zone_release(uma_zone_t, void **, int); 273 static void uma_zero_item(void *, uma_zone_t); 274 275 void uma_print_zone(uma_zone_t); 276 void uma_print_stats(void); 277 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 278 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 279 280 #ifdef INVARIANTS 281 static bool uma_dbg_kskip(uma_keg_t keg, void *mem); 282 static bool uma_dbg_zskip(uma_zone_t zone, void *mem); 283 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item); 284 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item); 285 286 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0, 287 "Memory allocation debugging"); 288 289 static u_int dbg_divisor = 1; 290 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor, 291 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0, 292 "Debug & thrash every this item in memory allocator"); 293 294 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER; 295 static counter_u64_t uma_skip_cnt = EARLY_COUNTER; 296 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD, 297 &uma_dbg_cnt, "memory items debugged"); 298 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD, 299 &uma_skip_cnt, "memory items skipped, not debugged"); 300 #endif 301 302 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 303 304 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 305 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 306 307 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 308 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 309 310 static int zone_warnings = 1; 311 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, 312 "Warn when UMA zones becomes full"); 313 314 /* Adjust bytes under management by UMA. */ 315 static inline void 316 uma_total_dec(unsigned long size) 317 { 318 319 atomic_subtract_long(&uma_kmem_total, size); 320 } 321 322 static inline void 323 uma_total_inc(unsigned long size) 324 { 325 326 if (atomic_fetchadd_long(&uma_kmem_total, size) > uma_kmem_limit) 327 uma_reclaim_wakeup(); 328 } 329 330 /* 331 * This routine checks to see whether or not it's safe to enable buckets. 332 */ 333 static void 334 bucket_enable(void) 335 { 336 bucketdisable = vm_page_count_min(); 337 } 338 339 /* 340 * Initialize bucket_zones, the array of zones of buckets of various sizes. 341 * 342 * For each zone, calculate the memory required for each bucket, consisting 343 * of the header and an array of pointers. 344 */ 345 static void 346 bucket_init(void) 347 { 348 struct uma_bucket_zone *ubz; 349 int size; 350 351 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 352 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 353 size += sizeof(void *) * ubz->ubz_entries; 354 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 355 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 356 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET | UMA_ZONE_NUMA); 357 } 358 } 359 360 /* 361 * Given a desired number of entries for a bucket, return the zone from which 362 * to allocate the bucket. 363 */ 364 static struct uma_bucket_zone * 365 bucket_zone_lookup(int entries) 366 { 367 struct uma_bucket_zone *ubz; 368 369 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 370 if (ubz->ubz_entries >= entries) 371 return (ubz); 372 ubz--; 373 return (ubz); 374 } 375 376 static int 377 bucket_select(int size) 378 { 379 struct uma_bucket_zone *ubz; 380 381 ubz = &bucket_zones[0]; 382 if (size > ubz->ubz_maxsize) 383 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 384 385 for (; ubz->ubz_entries != 0; ubz++) 386 if (ubz->ubz_maxsize < size) 387 break; 388 ubz--; 389 return (ubz->ubz_entries); 390 } 391 392 static uma_bucket_t 393 bucket_alloc(uma_zone_t zone, void *udata, int flags) 394 { 395 struct uma_bucket_zone *ubz; 396 uma_bucket_t bucket; 397 398 /* 399 * This is to stop us from allocating per cpu buckets while we're 400 * running out of vm.boot_pages. Otherwise, we would exhaust the 401 * boot pages. This also prevents us from allocating buckets in 402 * low memory situations. 403 */ 404 if (bucketdisable) 405 return (NULL); 406 /* 407 * To limit bucket recursion we store the original zone flags 408 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 409 * NOVM flag to persist even through deep recursions. We also 410 * store ZFLAG_BUCKET once we have recursed attempting to allocate 411 * a bucket for a bucket zone so we do not allow infinite bucket 412 * recursion. This cookie will even persist to frees of unused 413 * buckets via the allocation path or bucket allocations in the 414 * free path. 415 */ 416 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 417 udata = (void *)(uintptr_t)zone->uz_flags; 418 else { 419 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 420 return (NULL); 421 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 422 } 423 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 424 flags |= M_NOVM; 425 ubz = bucket_zone_lookup(zone->uz_count); 426 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 427 ubz++; 428 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 429 if (bucket) { 430 #ifdef INVARIANTS 431 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 432 #endif 433 bucket->ub_cnt = 0; 434 bucket->ub_entries = ubz->ubz_entries; 435 } 436 437 return (bucket); 438 } 439 440 static void 441 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 442 { 443 struct uma_bucket_zone *ubz; 444 445 KASSERT(bucket->ub_cnt == 0, 446 ("bucket_free: Freeing a non free bucket.")); 447 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 448 udata = (void *)(uintptr_t)zone->uz_flags; 449 ubz = bucket_zone_lookup(bucket->ub_entries); 450 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 451 } 452 453 static void 454 bucket_zone_drain(void) 455 { 456 struct uma_bucket_zone *ubz; 457 458 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 459 zone_drain(ubz->ubz_zone); 460 } 461 462 static void 463 zone_log_warning(uma_zone_t zone) 464 { 465 static const struct timeval warninterval = { 300, 0 }; 466 467 if (!zone_warnings || zone->uz_warning == NULL) 468 return; 469 470 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 471 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 472 } 473 474 static inline void 475 zone_maxaction(uma_zone_t zone) 476 { 477 478 if (zone->uz_maxaction.ta_func != NULL) 479 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); 480 } 481 482 static void 483 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 484 { 485 uma_klink_t klink; 486 487 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 488 kegfn(klink->kl_keg); 489 } 490 491 /* 492 * Routine called by timeout which is used to fire off some time interval 493 * based calculations. (stats, hash size, etc.) 494 * 495 * Arguments: 496 * arg Unused 497 * 498 * Returns: 499 * Nothing 500 */ 501 static void 502 uma_timeout(void *unused) 503 { 504 bucket_enable(); 505 zone_foreach(zone_timeout); 506 507 /* Reschedule this event */ 508 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 509 } 510 511 /* 512 * Routine to perform timeout driven calculations. This expands the 513 * hashes and does per cpu statistics aggregation. 514 * 515 * Returns nothing. 516 */ 517 static void 518 keg_timeout(uma_keg_t keg) 519 { 520 521 KEG_LOCK(keg); 522 /* 523 * Expand the keg hash table. 524 * 525 * This is done if the number of slabs is larger than the hash size. 526 * What I'm trying to do here is completely reduce collisions. This 527 * may be a little aggressive. Should I allow for two collisions max? 528 */ 529 if (keg->uk_flags & UMA_ZONE_HASH && 530 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 531 struct uma_hash newhash; 532 struct uma_hash oldhash; 533 int ret; 534 535 /* 536 * This is so involved because allocating and freeing 537 * while the keg lock is held will lead to deadlock. 538 * I have to do everything in stages and check for 539 * races. 540 */ 541 newhash = keg->uk_hash; 542 KEG_UNLOCK(keg); 543 ret = hash_alloc(&newhash); 544 KEG_LOCK(keg); 545 if (ret) { 546 if (hash_expand(&keg->uk_hash, &newhash)) { 547 oldhash = keg->uk_hash; 548 keg->uk_hash = newhash; 549 } else 550 oldhash = newhash; 551 552 KEG_UNLOCK(keg); 553 hash_free(&oldhash); 554 return; 555 } 556 } 557 KEG_UNLOCK(keg); 558 } 559 560 static void 561 zone_timeout(uma_zone_t zone) 562 { 563 564 zone_foreach_keg(zone, &keg_timeout); 565 } 566 567 /* 568 * Allocate and zero fill the next sized hash table from the appropriate 569 * backing store. 570 * 571 * Arguments: 572 * hash A new hash structure with the old hash size in uh_hashsize 573 * 574 * Returns: 575 * 1 on success and 0 on failure. 576 */ 577 static int 578 hash_alloc(struct uma_hash *hash) 579 { 580 int oldsize; 581 int alloc; 582 583 oldsize = hash->uh_hashsize; 584 585 /* We're just going to go to a power of two greater */ 586 if (oldsize) { 587 hash->uh_hashsize = oldsize * 2; 588 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 589 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 590 M_UMAHASH, M_NOWAIT); 591 } else { 592 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 593 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 594 UMA_ANYDOMAIN, M_WAITOK); 595 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 596 } 597 if (hash->uh_slab_hash) { 598 bzero(hash->uh_slab_hash, alloc); 599 hash->uh_hashmask = hash->uh_hashsize - 1; 600 return (1); 601 } 602 603 return (0); 604 } 605 606 /* 607 * Expands the hash table for HASH zones. This is done from zone_timeout 608 * to reduce collisions. This must not be done in the regular allocation 609 * path, otherwise, we can recurse on the vm while allocating pages. 610 * 611 * Arguments: 612 * oldhash The hash you want to expand 613 * newhash The hash structure for the new table 614 * 615 * Returns: 616 * Nothing 617 * 618 * Discussion: 619 */ 620 static int 621 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 622 { 623 uma_slab_t slab; 624 int hval; 625 int i; 626 627 if (!newhash->uh_slab_hash) 628 return (0); 629 630 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 631 return (0); 632 633 /* 634 * I need to investigate hash algorithms for resizing without a 635 * full rehash. 636 */ 637 638 for (i = 0; i < oldhash->uh_hashsize; i++) 639 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 640 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 641 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 642 hval = UMA_HASH(newhash, slab->us_data); 643 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 644 slab, us_hlink); 645 } 646 647 return (1); 648 } 649 650 /* 651 * Free the hash bucket to the appropriate backing store. 652 * 653 * Arguments: 654 * slab_hash The hash bucket we're freeing 655 * hashsize The number of entries in that hash bucket 656 * 657 * Returns: 658 * Nothing 659 */ 660 static void 661 hash_free(struct uma_hash *hash) 662 { 663 if (hash->uh_slab_hash == NULL) 664 return; 665 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 666 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 667 else 668 free(hash->uh_slab_hash, M_UMAHASH); 669 } 670 671 /* 672 * Frees all outstanding items in a bucket 673 * 674 * Arguments: 675 * zone The zone to free to, must be unlocked. 676 * bucket The free/alloc bucket with items, cpu queue must be locked. 677 * 678 * Returns: 679 * Nothing 680 */ 681 682 static void 683 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 684 { 685 int i; 686 687 if (bucket == NULL) 688 return; 689 690 if (zone->uz_fini) 691 for (i = 0; i < bucket->ub_cnt; i++) 692 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 693 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 694 bucket->ub_cnt = 0; 695 } 696 697 /* 698 * Drains the per cpu caches for a zone. 699 * 700 * NOTE: This may only be called while the zone is being turn down, and not 701 * during normal operation. This is necessary in order that we do not have 702 * to migrate CPUs to drain the per-CPU caches. 703 * 704 * Arguments: 705 * zone The zone to drain, must be unlocked. 706 * 707 * Returns: 708 * Nothing 709 */ 710 static void 711 cache_drain(uma_zone_t zone) 712 { 713 uma_cache_t cache; 714 int cpu; 715 716 /* 717 * XXX: It is safe to not lock the per-CPU caches, because we're 718 * tearing down the zone anyway. I.e., there will be no further use 719 * of the caches at this point. 720 * 721 * XXX: It would good to be able to assert that the zone is being 722 * torn down to prevent improper use of cache_drain(). 723 * 724 * XXX: We lock the zone before passing into bucket_cache_drain() as 725 * it is used elsewhere. Should the tear-down path be made special 726 * there in some form? 727 */ 728 CPU_FOREACH(cpu) { 729 cache = &zone->uz_cpu[cpu]; 730 bucket_drain(zone, cache->uc_allocbucket); 731 bucket_drain(zone, cache->uc_freebucket); 732 if (cache->uc_allocbucket != NULL) 733 bucket_free(zone, cache->uc_allocbucket, NULL); 734 if (cache->uc_freebucket != NULL) 735 bucket_free(zone, cache->uc_freebucket, NULL); 736 cache->uc_allocbucket = cache->uc_freebucket = NULL; 737 } 738 ZONE_LOCK(zone); 739 bucket_cache_drain(zone); 740 ZONE_UNLOCK(zone); 741 } 742 743 static void 744 cache_shrink(uma_zone_t zone) 745 { 746 747 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 748 return; 749 750 ZONE_LOCK(zone); 751 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; 752 ZONE_UNLOCK(zone); 753 } 754 755 static void 756 cache_drain_safe_cpu(uma_zone_t zone) 757 { 758 uma_cache_t cache; 759 uma_bucket_t b1, b2; 760 int domain; 761 762 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 763 return; 764 765 b1 = b2 = NULL; 766 ZONE_LOCK(zone); 767 critical_enter(); 768 if (zone->uz_flags & UMA_ZONE_NUMA) 769 domain = PCPU_GET(domain); 770 else 771 domain = 0; 772 cache = &zone->uz_cpu[curcpu]; 773 if (cache->uc_allocbucket) { 774 if (cache->uc_allocbucket->ub_cnt != 0) 775 LIST_INSERT_HEAD(&zone->uz_domain[domain].uzd_buckets, 776 cache->uc_allocbucket, ub_link); 777 else 778 b1 = cache->uc_allocbucket; 779 cache->uc_allocbucket = NULL; 780 } 781 if (cache->uc_freebucket) { 782 if (cache->uc_freebucket->ub_cnt != 0) 783 LIST_INSERT_HEAD(&zone->uz_domain[domain].uzd_buckets, 784 cache->uc_freebucket, ub_link); 785 else 786 b2 = cache->uc_freebucket; 787 cache->uc_freebucket = NULL; 788 } 789 critical_exit(); 790 ZONE_UNLOCK(zone); 791 if (b1) 792 bucket_free(zone, b1, NULL); 793 if (b2) 794 bucket_free(zone, b2, NULL); 795 } 796 797 /* 798 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 799 * This is an expensive call because it needs to bind to all CPUs 800 * one by one and enter a critical section on each of them in order 801 * to safely access their cache buckets. 802 * Zone lock must not be held on call this function. 803 */ 804 static void 805 cache_drain_safe(uma_zone_t zone) 806 { 807 int cpu; 808 809 /* 810 * Polite bucket sizes shrinking was not enouth, shrink aggressively. 811 */ 812 if (zone) 813 cache_shrink(zone); 814 else 815 zone_foreach(cache_shrink); 816 817 CPU_FOREACH(cpu) { 818 thread_lock(curthread); 819 sched_bind(curthread, cpu); 820 thread_unlock(curthread); 821 822 if (zone) 823 cache_drain_safe_cpu(zone); 824 else 825 zone_foreach(cache_drain_safe_cpu); 826 } 827 thread_lock(curthread); 828 sched_unbind(curthread); 829 thread_unlock(curthread); 830 } 831 832 /* 833 * Drain the cached buckets from a zone. Expects a locked zone on entry. 834 */ 835 static void 836 bucket_cache_drain(uma_zone_t zone) 837 { 838 uma_zone_domain_t zdom; 839 uma_bucket_t bucket; 840 int i; 841 842 /* 843 * Drain the bucket queues and free the buckets. 844 */ 845 for (i = 0; i < vm_ndomains; i++) { 846 zdom = &zone->uz_domain[i]; 847 while ((bucket = LIST_FIRST(&zdom->uzd_buckets)) != NULL) { 848 LIST_REMOVE(bucket, ub_link); 849 ZONE_UNLOCK(zone); 850 bucket_drain(zone, bucket); 851 bucket_free(zone, bucket, NULL); 852 ZONE_LOCK(zone); 853 } 854 } 855 856 /* 857 * Shrink further bucket sizes. Price of single zone lock collision 858 * is probably lower then price of global cache drain. 859 */ 860 if (zone->uz_count > zone->uz_count_min) 861 zone->uz_count--; 862 } 863 864 static void 865 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 866 { 867 uint8_t *mem; 868 int i; 869 uint8_t flags; 870 871 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes", 872 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera); 873 874 mem = slab->us_data; 875 flags = slab->us_flags; 876 i = start; 877 if (keg->uk_fini != NULL) { 878 for (i--; i > -1; i--) 879 #ifdef INVARIANTS 880 /* 881 * trash_fini implies that dtor was trash_dtor. trash_fini 882 * would check that memory hasn't been modified since free, 883 * which executed trash_dtor. 884 * That's why we need to run uma_dbg_kskip() check here, 885 * albeit we don't make skip check for other init/fini 886 * invocations. 887 */ 888 if (!uma_dbg_kskip(keg, slab->us_data + (keg->uk_rsize * i)) || 889 keg->uk_fini != trash_fini) 890 #endif 891 keg->uk_fini(slab->us_data + (keg->uk_rsize * i), 892 keg->uk_size); 893 } 894 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 895 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 896 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 897 uma_total_dec(PAGE_SIZE * keg->uk_ppera); 898 } 899 900 /* 901 * Frees pages from a keg back to the system. This is done on demand from 902 * the pageout daemon. 903 * 904 * Returns nothing. 905 */ 906 static void 907 keg_drain(uma_keg_t keg) 908 { 909 struct slabhead freeslabs = { 0 }; 910 uma_domain_t dom; 911 uma_slab_t slab, tmp; 912 int i; 913 914 /* 915 * We don't want to take pages from statically allocated kegs at this 916 * time 917 */ 918 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 919 return; 920 921 CTR3(KTR_UMA, "keg_drain %s(%p) free items: %u", 922 keg->uk_name, keg, keg->uk_free); 923 KEG_LOCK(keg); 924 if (keg->uk_free == 0) 925 goto finished; 926 927 for (i = 0; i < vm_ndomains; i++) { 928 dom = &keg->uk_domain[i]; 929 LIST_FOREACH_SAFE(slab, &dom->ud_free_slab, us_link, tmp) { 930 /* We have nowhere to free these to. */ 931 if (slab->us_flags & UMA_SLAB_BOOT) 932 continue; 933 934 LIST_REMOVE(slab, us_link); 935 keg->uk_pages -= keg->uk_ppera; 936 keg->uk_free -= keg->uk_ipers; 937 938 if (keg->uk_flags & UMA_ZONE_HASH) 939 UMA_HASH_REMOVE(&keg->uk_hash, slab, 940 slab->us_data); 941 942 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 943 } 944 } 945 946 finished: 947 KEG_UNLOCK(keg); 948 949 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 950 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 951 keg_free_slab(keg, slab, keg->uk_ipers); 952 } 953 } 954 955 static void 956 zone_drain_wait(uma_zone_t zone, int waitok) 957 { 958 959 /* 960 * Set draining to interlock with zone_dtor() so we can release our 961 * locks as we go. Only dtor() should do a WAITOK call since it 962 * is the only call that knows the structure will still be available 963 * when it wakes up. 964 */ 965 ZONE_LOCK(zone); 966 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 967 if (waitok == M_NOWAIT) 968 goto out; 969 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 970 } 971 zone->uz_flags |= UMA_ZFLAG_DRAINING; 972 bucket_cache_drain(zone); 973 ZONE_UNLOCK(zone); 974 /* 975 * The DRAINING flag protects us from being freed while 976 * we're running. Normally the uma_rwlock would protect us but we 977 * must be able to release and acquire the right lock for each keg. 978 */ 979 zone_foreach_keg(zone, &keg_drain); 980 ZONE_LOCK(zone); 981 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 982 wakeup(zone); 983 out: 984 ZONE_UNLOCK(zone); 985 } 986 987 void 988 zone_drain(uma_zone_t zone) 989 { 990 991 zone_drain_wait(zone, M_NOWAIT); 992 } 993 994 /* 995 * Allocate a new slab for a keg. This does not insert the slab onto a list. 996 * If the allocation was successful, the keg lock will be held upon return, 997 * otherwise the keg will be left unlocked. 998 * 999 * Arguments: 1000 * wait Shall we wait? 1001 * 1002 * Returns: 1003 * The slab that was allocated or NULL if there is no memory and the 1004 * caller specified M_NOWAIT. 1005 */ 1006 static uma_slab_t 1007 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int wait) 1008 { 1009 uma_alloc allocf; 1010 uma_slab_t slab; 1011 unsigned long size; 1012 uint8_t *mem; 1013 uint8_t flags; 1014 int i; 1015 1016 KASSERT(domain >= 0 && domain < vm_ndomains, 1017 ("keg_alloc_slab: domain %d out of range", domain)); 1018 mtx_assert(&keg->uk_lock, MA_OWNED); 1019 1020 allocf = keg->uk_allocf; 1021 KEG_UNLOCK(keg); 1022 1023 slab = NULL; 1024 mem = NULL; 1025 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1026 slab = zone_alloc_item(keg->uk_slabzone, NULL, domain, wait); 1027 if (slab == NULL) 1028 goto out; 1029 } 1030 1031 /* 1032 * This reproduces the old vm_zone behavior of zero filling pages the 1033 * first time they are added to a zone. 1034 * 1035 * Malloced items are zeroed in uma_zalloc. 1036 */ 1037 1038 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 1039 wait |= M_ZERO; 1040 else 1041 wait &= ~M_ZERO; 1042 1043 if (keg->uk_flags & UMA_ZONE_NODUMP) 1044 wait |= M_NODUMP; 1045 1046 /* zone is passed for legacy reasons. */ 1047 size = keg->uk_ppera * PAGE_SIZE; 1048 mem = allocf(zone, size, domain, &flags, wait); 1049 if (mem == NULL) { 1050 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1051 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 1052 slab = NULL; 1053 goto out; 1054 } 1055 uma_total_inc(size); 1056 1057 /* Point the slab into the allocated memory */ 1058 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 1059 slab = (uma_slab_t )(mem + keg->uk_pgoff); 1060 1061 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 1062 for (i = 0; i < keg->uk_ppera; i++) 1063 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 1064 1065 slab->us_keg = keg; 1066 slab->us_data = mem; 1067 slab->us_freecount = keg->uk_ipers; 1068 slab->us_flags = flags; 1069 slab->us_domain = domain; 1070 BIT_FILL(SLAB_SETSIZE, &slab->us_free); 1071 #ifdef INVARIANTS 1072 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); 1073 #endif 1074 1075 if (keg->uk_init != NULL) { 1076 for (i = 0; i < keg->uk_ipers; i++) 1077 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 1078 keg->uk_size, wait) != 0) 1079 break; 1080 if (i != keg->uk_ipers) { 1081 keg_free_slab(keg, slab, i); 1082 slab = NULL; 1083 goto out; 1084 } 1085 } 1086 KEG_LOCK(keg); 1087 1088 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)", 1089 slab, keg->uk_name, keg); 1090 1091 if (keg->uk_flags & UMA_ZONE_HASH) 1092 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1093 1094 keg->uk_pages += keg->uk_ppera; 1095 keg->uk_free += keg->uk_ipers; 1096 1097 out: 1098 return (slab); 1099 } 1100 1101 /* 1102 * This function is intended to be used early on in place of page_alloc() so 1103 * that we may use the boot time page cache to satisfy allocations before 1104 * the VM is ready. 1105 */ 1106 static void * 1107 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1108 int wait) 1109 { 1110 uma_keg_t keg; 1111 void *mem; 1112 int pages; 1113 1114 keg = zone_first_keg(zone); 1115 1116 /* 1117 * If we are in BOOT_BUCKETS or higher, than switch to real 1118 * allocator. Zones with page sized slabs switch at BOOT_PAGEALLOC. 1119 */ 1120 switch (booted) { 1121 case BOOT_COLD: 1122 case BOOT_STRAPPED: 1123 break; 1124 case BOOT_PAGEALLOC: 1125 if (keg->uk_ppera > 1) 1126 break; 1127 case BOOT_BUCKETS: 1128 case BOOT_RUNNING: 1129 #ifdef UMA_MD_SMALL_ALLOC 1130 keg->uk_allocf = (keg->uk_ppera > 1) ? 1131 page_alloc : uma_small_alloc; 1132 #else 1133 keg->uk_allocf = page_alloc; 1134 #endif 1135 return keg->uk_allocf(zone, bytes, domain, pflag, wait); 1136 } 1137 1138 /* 1139 * Check our small startup cache to see if it has pages remaining. 1140 */ 1141 pages = howmany(bytes, PAGE_SIZE); 1142 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__)); 1143 if (pages > boot_pages) 1144 panic("UMA zone \"%s\": Increase vm.boot_pages", zone->uz_name); 1145 #ifdef DIAGNOSTIC 1146 printf("%s from \"%s\", %d boot pages left\n", __func__, zone->uz_name, 1147 boot_pages); 1148 #endif 1149 mem = bootmem; 1150 boot_pages -= pages; 1151 bootmem += pages * PAGE_SIZE; 1152 *pflag = UMA_SLAB_BOOT; 1153 1154 return (mem); 1155 } 1156 1157 /* 1158 * Allocates a number of pages from the system 1159 * 1160 * Arguments: 1161 * bytes The number of bytes requested 1162 * wait Shall we wait? 1163 * 1164 * Returns: 1165 * A pointer to the alloced memory or possibly 1166 * NULL if M_NOWAIT is set. 1167 */ 1168 static void * 1169 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1170 int wait) 1171 { 1172 void *p; /* Returned page */ 1173 1174 *pflag = UMA_SLAB_KERNEL; 1175 p = (void *) kmem_malloc_domain(domain, bytes, wait); 1176 1177 return (p); 1178 } 1179 1180 static void * 1181 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1182 int wait) 1183 { 1184 struct pglist alloctail; 1185 vm_offset_t addr, zkva; 1186 int cpu, flags; 1187 vm_page_t p, p_next; 1188 #ifdef NUMA 1189 struct pcpu *pc; 1190 #endif 1191 1192 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE); 1193 1194 TAILQ_INIT(&alloctail); 1195 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1196 malloc2vm_flags(wait); 1197 *pflag = UMA_SLAB_KERNEL; 1198 for (cpu = 0; cpu <= mp_maxid; cpu++) { 1199 if (CPU_ABSENT(cpu)) { 1200 p = vm_page_alloc(NULL, 0, flags); 1201 } else { 1202 #ifndef NUMA 1203 p = vm_page_alloc(NULL, 0, flags); 1204 #else 1205 pc = pcpu_find(cpu); 1206 p = vm_page_alloc_domain(NULL, 0, pc->pc_domain, flags); 1207 if (__predict_false(p == NULL)) 1208 p = vm_page_alloc(NULL, 0, flags); 1209 #endif 1210 } 1211 if (__predict_false(p == NULL)) 1212 goto fail; 1213 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1214 } 1215 if ((addr = kva_alloc(bytes)) == 0) 1216 goto fail; 1217 zkva = addr; 1218 TAILQ_FOREACH(p, &alloctail, listq) { 1219 pmap_qenter(zkva, &p, 1); 1220 zkva += PAGE_SIZE; 1221 } 1222 return ((void*)addr); 1223 fail: 1224 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1225 vm_page_unwire(p, PQ_NONE); 1226 vm_page_free(p); 1227 } 1228 return (NULL); 1229 } 1230 1231 /* 1232 * Allocates a number of pages from within an object 1233 * 1234 * Arguments: 1235 * bytes The number of bytes requested 1236 * wait Shall we wait? 1237 * 1238 * Returns: 1239 * A pointer to the alloced memory or possibly 1240 * NULL if M_NOWAIT is set. 1241 */ 1242 static void * 1243 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, 1244 int wait) 1245 { 1246 TAILQ_HEAD(, vm_page) alloctail; 1247 u_long npages; 1248 vm_offset_t retkva, zkva; 1249 vm_page_t p, p_next; 1250 uma_keg_t keg; 1251 1252 TAILQ_INIT(&alloctail); 1253 keg = zone_first_keg(zone); 1254 1255 npages = howmany(bytes, PAGE_SIZE); 1256 while (npages > 0) { 1257 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT | 1258 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1259 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK : 1260 VM_ALLOC_NOWAIT)); 1261 if (p != NULL) { 1262 /* 1263 * Since the page does not belong to an object, its 1264 * listq is unused. 1265 */ 1266 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1267 npages--; 1268 continue; 1269 } 1270 /* 1271 * Page allocation failed, free intermediate pages and 1272 * exit. 1273 */ 1274 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1275 vm_page_unwire(p, PQ_NONE); 1276 vm_page_free(p); 1277 } 1278 return (NULL); 1279 } 1280 *flags = UMA_SLAB_PRIV; 1281 zkva = keg->uk_kva + 1282 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1283 retkva = zkva; 1284 TAILQ_FOREACH(p, &alloctail, listq) { 1285 pmap_qenter(zkva, &p, 1); 1286 zkva += PAGE_SIZE; 1287 } 1288 1289 return ((void *)retkva); 1290 } 1291 1292 /* 1293 * Frees a number of pages to the system 1294 * 1295 * Arguments: 1296 * mem A pointer to the memory to be freed 1297 * size The size of the memory being freed 1298 * flags The original p->us_flags field 1299 * 1300 * Returns: 1301 * Nothing 1302 */ 1303 static void 1304 page_free(void *mem, vm_size_t size, uint8_t flags) 1305 { 1306 1307 if ((flags & UMA_SLAB_KERNEL) == 0) 1308 panic("UMA: page_free used with invalid flags %x", flags); 1309 1310 kmem_free((vm_offset_t)mem, size); 1311 } 1312 1313 /* 1314 * Frees pcpu zone allocations 1315 * 1316 * Arguments: 1317 * mem A pointer to the memory to be freed 1318 * size The size of the memory being freed 1319 * flags The original p->us_flags field 1320 * 1321 * Returns: 1322 * Nothing 1323 */ 1324 static void 1325 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags) 1326 { 1327 vm_offset_t sva, curva; 1328 vm_paddr_t paddr; 1329 vm_page_t m; 1330 1331 MPASS(size == (mp_maxid+1)*PAGE_SIZE); 1332 sva = (vm_offset_t)mem; 1333 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) { 1334 paddr = pmap_kextract(curva); 1335 m = PHYS_TO_VM_PAGE(paddr); 1336 vm_page_unwire(m, PQ_NONE); 1337 vm_page_free(m); 1338 } 1339 pmap_qremove(sva, size >> PAGE_SHIFT); 1340 kva_free(sva, size); 1341 } 1342 1343 1344 /* 1345 * Zero fill initializer 1346 * 1347 * Arguments/Returns follow uma_init specifications 1348 */ 1349 static int 1350 zero_init(void *mem, int size, int flags) 1351 { 1352 bzero(mem, size); 1353 return (0); 1354 } 1355 1356 /* 1357 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1358 * 1359 * Arguments 1360 * keg The zone we should initialize 1361 * 1362 * Returns 1363 * Nothing 1364 */ 1365 static void 1366 keg_small_init(uma_keg_t keg) 1367 { 1368 u_int rsize; 1369 u_int memused; 1370 u_int wastedspace; 1371 u_int shsize; 1372 u_int slabsize; 1373 1374 if (keg->uk_flags & UMA_ZONE_PCPU) { 1375 u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU; 1376 1377 slabsize = UMA_PCPU_ALLOC_SIZE; 1378 keg->uk_ppera = ncpus; 1379 } else { 1380 slabsize = UMA_SLAB_SIZE; 1381 keg->uk_ppera = 1; 1382 } 1383 1384 /* 1385 * Calculate the size of each allocation (rsize) according to 1386 * alignment. If the requested size is smaller than we have 1387 * allocation bits for we round it up. 1388 */ 1389 rsize = keg->uk_size; 1390 if (rsize < slabsize / SLAB_SETSIZE) 1391 rsize = slabsize / SLAB_SETSIZE; 1392 if (rsize & keg->uk_align) 1393 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1394 keg->uk_rsize = rsize; 1395 1396 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1397 keg->uk_rsize < UMA_PCPU_ALLOC_SIZE, 1398 ("%s: size %u too large", __func__, keg->uk_rsize)); 1399 1400 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1401 shsize = 0; 1402 else 1403 shsize = sizeof(struct uma_slab); 1404 1405 if (rsize <= slabsize - shsize) 1406 keg->uk_ipers = (slabsize - shsize) / rsize; 1407 else { 1408 /* Handle special case when we have 1 item per slab, so 1409 * alignment requirement can be relaxed. */ 1410 KASSERT(keg->uk_size <= slabsize - shsize, 1411 ("%s: size %u greater than slab", __func__, keg->uk_size)); 1412 keg->uk_ipers = 1; 1413 } 1414 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1415 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1416 1417 memused = keg->uk_ipers * rsize + shsize; 1418 wastedspace = slabsize - memused; 1419 1420 /* 1421 * We can't do OFFPAGE if we're internal or if we've been 1422 * asked to not go to the VM for buckets. If we do this we 1423 * may end up going to the VM for slabs which we do not 1424 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1425 * of UMA_ZONE_VM, which clearly forbids it. 1426 */ 1427 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1428 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1429 return; 1430 1431 /* 1432 * See if using an OFFPAGE slab will limit our waste. Only do 1433 * this if it permits more items per-slab. 1434 * 1435 * XXX We could try growing slabsize to limit max waste as well. 1436 * Historically this was not done because the VM could not 1437 * efficiently handle contiguous allocations. 1438 */ 1439 if ((wastedspace >= slabsize / UMA_MAX_WASTE) && 1440 (keg->uk_ipers < (slabsize / keg->uk_rsize))) { 1441 keg->uk_ipers = slabsize / keg->uk_rsize; 1442 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1443 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1444 CTR6(KTR_UMA, "UMA decided we need offpage slab headers for " 1445 "keg: %s(%p), calculated wastedspace = %d, " 1446 "maximum wasted space allowed = %d, " 1447 "calculated ipers = %d, " 1448 "new wasted space = %d\n", keg->uk_name, keg, wastedspace, 1449 slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1450 slabsize - keg->uk_ipers * keg->uk_rsize); 1451 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1452 } 1453 1454 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1455 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1456 keg->uk_flags |= UMA_ZONE_HASH; 1457 } 1458 1459 /* 1460 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1461 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1462 * more complicated. 1463 * 1464 * Arguments 1465 * keg The keg we should initialize 1466 * 1467 * Returns 1468 * Nothing 1469 */ 1470 static void 1471 keg_large_init(uma_keg_t keg) 1472 { 1473 u_int shsize; 1474 1475 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1476 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1477 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1478 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1479 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1480 1481 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1482 keg->uk_ipers = 1; 1483 keg->uk_rsize = keg->uk_size; 1484 1485 /* Check whether we have enough space to not do OFFPAGE. */ 1486 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) { 1487 shsize = sizeof(struct uma_slab); 1488 if (shsize & UMA_ALIGN_PTR) 1489 shsize = (shsize & ~UMA_ALIGN_PTR) + 1490 (UMA_ALIGN_PTR + 1); 1491 1492 if (PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < shsize) { 1493 /* 1494 * We can't do OFFPAGE if we're internal, in which case 1495 * we need an extra page per allocation to contain the 1496 * slab header. 1497 */ 1498 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0) 1499 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1500 else 1501 keg->uk_ppera++; 1502 } 1503 } 1504 1505 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1506 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1507 keg->uk_flags |= UMA_ZONE_HASH; 1508 } 1509 1510 static void 1511 keg_cachespread_init(uma_keg_t keg) 1512 { 1513 int alignsize; 1514 int trailer; 1515 int pages; 1516 int rsize; 1517 1518 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1519 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1520 1521 alignsize = keg->uk_align + 1; 1522 rsize = keg->uk_size; 1523 /* 1524 * We want one item to start on every align boundary in a page. To 1525 * do this we will span pages. We will also extend the item by the 1526 * size of align if it is an even multiple of align. Otherwise, it 1527 * would fall on the same boundary every time. 1528 */ 1529 if (rsize & keg->uk_align) 1530 rsize = (rsize & ~keg->uk_align) + alignsize; 1531 if ((rsize & alignsize) == 0) 1532 rsize += alignsize; 1533 trailer = rsize - keg->uk_size; 1534 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1535 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1536 keg->uk_rsize = rsize; 1537 keg->uk_ppera = pages; 1538 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1539 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1540 KASSERT(keg->uk_ipers <= SLAB_SETSIZE, 1541 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1542 keg->uk_ipers)); 1543 } 1544 1545 /* 1546 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1547 * the keg onto the global keg list. 1548 * 1549 * Arguments/Returns follow uma_ctor specifications 1550 * udata Actually uma_kctor_args 1551 */ 1552 static int 1553 keg_ctor(void *mem, int size, void *udata, int flags) 1554 { 1555 struct uma_kctor_args *arg = udata; 1556 uma_keg_t keg = mem; 1557 uma_zone_t zone; 1558 1559 bzero(keg, size); 1560 keg->uk_size = arg->size; 1561 keg->uk_init = arg->uminit; 1562 keg->uk_fini = arg->fini; 1563 keg->uk_align = arg->align; 1564 keg->uk_free = 0; 1565 keg->uk_reserve = 0; 1566 keg->uk_pages = 0; 1567 keg->uk_flags = arg->flags; 1568 keg->uk_slabzone = NULL; 1569 1570 /* 1571 * We use a global round-robin policy by default. Zones with 1572 * UMA_ZONE_NUMA set will use first-touch instead, in which case the 1573 * iterator is never run. 1574 */ 1575 keg->uk_dr.dr_policy = DOMAINSET_RR(); 1576 keg->uk_dr.dr_iter = 0; 1577 1578 /* 1579 * The master zone is passed to us at keg-creation time. 1580 */ 1581 zone = arg->zone; 1582 keg->uk_name = zone->uz_name; 1583 1584 if (arg->flags & UMA_ZONE_VM) 1585 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1586 1587 if (arg->flags & UMA_ZONE_ZINIT) 1588 keg->uk_init = zero_init; 1589 1590 if (arg->flags & UMA_ZONE_MALLOC) 1591 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1592 1593 if (arg->flags & UMA_ZONE_PCPU) 1594 #ifdef SMP 1595 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1596 #else 1597 keg->uk_flags &= ~UMA_ZONE_PCPU; 1598 #endif 1599 1600 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1601 keg_cachespread_init(keg); 1602 } else { 1603 if (keg->uk_size > UMA_SLAB_SPACE) 1604 keg_large_init(keg); 1605 else 1606 keg_small_init(keg); 1607 } 1608 1609 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1610 keg->uk_slabzone = slabzone; 1611 1612 /* 1613 * If we haven't booted yet we need allocations to go through the 1614 * startup cache until the vm is ready. 1615 */ 1616 if (booted < BOOT_PAGEALLOC) 1617 keg->uk_allocf = startup_alloc; 1618 #ifdef UMA_MD_SMALL_ALLOC 1619 else if (keg->uk_ppera == 1) 1620 keg->uk_allocf = uma_small_alloc; 1621 #endif 1622 else if (keg->uk_flags & UMA_ZONE_PCPU) 1623 keg->uk_allocf = pcpu_page_alloc; 1624 else 1625 keg->uk_allocf = page_alloc; 1626 #ifdef UMA_MD_SMALL_ALLOC 1627 if (keg->uk_ppera == 1) 1628 keg->uk_freef = uma_small_free; 1629 else 1630 #endif 1631 if (keg->uk_flags & UMA_ZONE_PCPU) 1632 keg->uk_freef = pcpu_page_free; 1633 else 1634 keg->uk_freef = page_free; 1635 1636 /* 1637 * Initialize keg's lock 1638 */ 1639 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1640 1641 /* 1642 * If we're putting the slab header in the actual page we need to 1643 * figure out where in each page it goes. This calculates a right 1644 * justified offset into the memory on an ALIGN_PTR boundary. 1645 */ 1646 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1647 u_int totsize; 1648 1649 /* Size of the slab struct and free list */ 1650 totsize = sizeof(struct uma_slab); 1651 1652 if (totsize & UMA_ALIGN_PTR) 1653 totsize = (totsize & ~UMA_ALIGN_PTR) + 1654 (UMA_ALIGN_PTR + 1); 1655 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1656 1657 /* 1658 * The only way the following is possible is if with our 1659 * UMA_ALIGN_PTR adjustments we are now bigger than 1660 * UMA_SLAB_SIZE. I haven't checked whether this is 1661 * mathematically possible for all cases, so we make 1662 * sure here anyway. 1663 */ 1664 totsize = keg->uk_pgoff + sizeof(struct uma_slab); 1665 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1666 printf("zone %s ipers %d rsize %d size %d\n", 1667 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1668 keg->uk_size); 1669 panic("UMA slab won't fit."); 1670 } 1671 } 1672 1673 if (keg->uk_flags & UMA_ZONE_HASH) 1674 hash_alloc(&keg->uk_hash); 1675 1676 CTR5(KTR_UMA, "keg_ctor %p zone %s(%p) out %d free %d\n", 1677 keg, zone->uz_name, zone, 1678 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, 1679 keg->uk_free); 1680 1681 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1682 1683 rw_wlock(&uma_rwlock); 1684 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1685 rw_wunlock(&uma_rwlock); 1686 return (0); 1687 } 1688 1689 /* 1690 * Zone header ctor. This initializes all fields, locks, etc. 1691 * 1692 * Arguments/Returns follow uma_ctor specifications 1693 * udata Actually uma_zctor_args 1694 */ 1695 static int 1696 zone_ctor(void *mem, int size, void *udata, int flags) 1697 { 1698 struct uma_zctor_args *arg = udata; 1699 uma_zone_t zone = mem; 1700 uma_zone_t z; 1701 uma_keg_t keg; 1702 1703 bzero(zone, size); 1704 zone->uz_name = arg->name; 1705 zone->uz_ctor = arg->ctor; 1706 zone->uz_dtor = arg->dtor; 1707 zone->uz_slab = zone_fetch_slab; 1708 zone->uz_init = NULL; 1709 zone->uz_fini = NULL; 1710 zone->uz_allocs = 0; 1711 zone->uz_frees = 0; 1712 zone->uz_fails = 0; 1713 zone->uz_sleeps = 0; 1714 zone->uz_count = 0; 1715 zone->uz_count_min = 0; 1716 zone->uz_flags = 0; 1717 zone->uz_warning = NULL; 1718 /* The domain structures follow the cpu structures. */ 1719 zone->uz_domain = (struct uma_zone_domain *)&zone->uz_cpu[mp_ncpus]; 1720 timevalclear(&zone->uz_ratecheck); 1721 keg = arg->keg; 1722 1723 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 1724 1725 /* 1726 * This is a pure cache zone, no kegs. 1727 */ 1728 if (arg->import) { 1729 if (arg->flags & UMA_ZONE_VM) 1730 arg->flags |= UMA_ZFLAG_CACHEONLY; 1731 zone->uz_flags = arg->flags; 1732 zone->uz_size = arg->size; 1733 zone->uz_import = arg->import; 1734 zone->uz_release = arg->release; 1735 zone->uz_arg = arg->arg; 1736 zone->uz_lockptr = &zone->uz_lock; 1737 rw_wlock(&uma_rwlock); 1738 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 1739 rw_wunlock(&uma_rwlock); 1740 goto out; 1741 } 1742 1743 /* 1744 * Use the regular zone/keg/slab allocator. 1745 */ 1746 zone->uz_import = (uma_import)zone_import; 1747 zone->uz_release = (uma_release)zone_release; 1748 zone->uz_arg = zone; 1749 1750 if (arg->flags & UMA_ZONE_SECONDARY) { 1751 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1752 zone->uz_init = arg->uminit; 1753 zone->uz_fini = arg->fini; 1754 zone->uz_lockptr = &keg->uk_lock; 1755 zone->uz_flags |= UMA_ZONE_SECONDARY; 1756 rw_wlock(&uma_rwlock); 1757 ZONE_LOCK(zone); 1758 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1759 if (LIST_NEXT(z, uz_link) == NULL) { 1760 LIST_INSERT_AFTER(z, zone, uz_link); 1761 break; 1762 } 1763 } 1764 ZONE_UNLOCK(zone); 1765 rw_wunlock(&uma_rwlock); 1766 } else if (keg == NULL) { 1767 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1768 arg->align, arg->flags)) == NULL) 1769 return (ENOMEM); 1770 } else { 1771 struct uma_kctor_args karg; 1772 int error; 1773 1774 /* We should only be here from uma_startup() */ 1775 karg.size = arg->size; 1776 karg.uminit = arg->uminit; 1777 karg.fini = arg->fini; 1778 karg.align = arg->align; 1779 karg.flags = arg->flags; 1780 karg.zone = zone; 1781 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1782 flags); 1783 if (error) 1784 return (error); 1785 } 1786 1787 /* 1788 * Link in the first keg. 1789 */ 1790 zone->uz_klink.kl_keg = keg; 1791 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1792 zone->uz_lockptr = &keg->uk_lock; 1793 zone->uz_size = keg->uk_size; 1794 zone->uz_flags |= (keg->uk_flags & 1795 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1796 1797 /* 1798 * Some internal zones don't have room allocated for the per cpu 1799 * caches. If we're internal, bail out here. 1800 */ 1801 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1802 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1803 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1804 return (0); 1805 } 1806 1807 out: 1808 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) != 1809 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET), 1810 ("Invalid zone flag combination")); 1811 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0) 1812 zone->uz_count = BUCKET_MAX; 1813 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0) 1814 zone->uz_count = 0; 1815 else 1816 zone->uz_count = bucket_select(zone->uz_size); 1817 zone->uz_count_min = zone->uz_count; 1818 1819 return (0); 1820 } 1821 1822 /* 1823 * Keg header dtor. This frees all data, destroys locks, frees the hash 1824 * table and removes the keg from the global list. 1825 * 1826 * Arguments/Returns follow uma_dtor specifications 1827 * udata unused 1828 */ 1829 static void 1830 keg_dtor(void *arg, int size, void *udata) 1831 { 1832 uma_keg_t keg; 1833 1834 keg = (uma_keg_t)arg; 1835 KEG_LOCK(keg); 1836 if (keg->uk_free != 0) { 1837 printf("Freed UMA keg (%s) was not empty (%d items). " 1838 " Lost %d pages of memory.\n", 1839 keg->uk_name ? keg->uk_name : "", 1840 keg->uk_free, keg->uk_pages); 1841 } 1842 KEG_UNLOCK(keg); 1843 1844 hash_free(&keg->uk_hash); 1845 1846 KEG_LOCK_FINI(keg); 1847 } 1848 1849 /* 1850 * Zone header dtor. 1851 * 1852 * Arguments/Returns follow uma_dtor specifications 1853 * udata unused 1854 */ 1855 static void 1856 zone_dtor(void *arg, int size, void *udata) 1857 { 1858 uma_klink_t klink; 1859 uma_zone_t zone; 1860 uma_keg_t keg; 1861 1862 zone = (uma_zone_t)arg; 1863 keg = zone_first_keg(zone); 1864 1865 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1866 cache_drain(zone); 1867 1868 rw_wlock(&uma_rwlock); 1869 LIST_REMOVE(zone, uz_link); 1870 rw_wunlock(&uma_rwlock); 1871 /* 1872 * XXX there are some races here where 1873 * the zone can be drained but zone lock 1874 * released and then refilled before we 1875 * remove it... we dont care for now 1876 */ 1877 zone_drain_wait(zone, M_WAITOK); 1878 /* 1879 * Unlink all of our kegs. 1880 */ 1881 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1882 klink->kl_keg = NULL; 1883 LIST_REMOVE(klink, kl_link); 1884 if (klink == &zone->uz_klink) 1885 continue; 1886 free(klink, M_TEMP); 1887 } 1888 /* 1889 * We only destroy kegs from non secondary zones. 1890 */ 1891 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1892 rw_wlock(&uma_rwlock); 1893 LIST_REMOVE(keg, uk_link); 1894 rw_wunlock(&uma_rwlock); 1895 zone_free_item(kegs, keg, NULL, SKIP_NONE); 1896 } 1897 ZONE_LOCK_FINI(zone); 1898 } 1899 1900 /* 1901 * Traverses every zone in the system and calls a callback 1902 * 1903 * Arguments: 1904 * zfunc A pointer to a function which accepts a zone 1905 * as an argument. 1906 * 1907 * Returns: 1908 * Nothing 1909 */ 1910 static void 1911 zone_foreach(void (*zfunc)(uma_zone_t)) 1912 { 1913 uma_keg_t keg; 1914 uma_zone_t zone; 1915 1916 rw_rlock(&uma_rwlock); 1917 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1918 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1919 zfunc(zone); 1920 } 1921 rw_runlock(&uma_rwlock); 1922 } 1923 1924 /* 1925 * Count how many pages do we need to bootstrap. VM supplies 1926 * its need in early zones in the argument, we add up our zones, 1927 * which consist of: UMA Slabs, UMA Hash and 9 Bucket zones. The 1928 * zone of zones and zone of kegs are accounted separately. 1929 */ 1930 #define UMA_BOOT_ZONES 11 1931 /* Zone of zones and zone of kegs have arbitrary alignment. */ 1932 #define UMA_BOOT_ALIGN 32 1933 static int zsize, ksize; 1934 int 1935 uma_startup_count(int vm_zones) 1936 { 1937 int zones, pages; 1938 1939 ksize = sizeof(struct uma_keg) + 1940 (sizeof(struct uma_domain) * vm_ndomains); 1941 zsize = sizeof(struct uma_zone) + 1942 (sizeof(struct uma_cache) * (mp_maxid + 1)) + 1943 (sizeof(struct uma_zone_domain) * vm_ndomains); 1944 1945 /* 1946 * Memory for the zone of kegs and its keg, 1947 * and for zone of zones. 1948 */ 1949 pages = howmany(roundup(zsize, CACHE_LINE_SIZE) * 2 + 1950 roundup(ksize, CACHE_LINE_SIZE), PAGE_SIZE); 1951 1952 #ifdef UMA_MD_SMALL_ALLOC 1953 zones = UMA_BOOT_ZONES; 1954 #else 1955 zones = UMA_BOOT_ZONES + vm_zones; 1956 vm_zones = 0; 1957 #endif 1958 1959 /* Memory for the rest of startup zones, UMA and VM, ... */ 1960 if (zsize > UMA_SLAB_SPACE) 1961 pages += (zones + vm_zones) * 1962 howmany(roundup2(zsize, UMA_BOOT_ALIGN), UMA_SLAB_SIZE); 1963 else if (roundup2(zsize, UMA_BOOT_ALIGN) > UMA_SLAB_SPACE) 1964 pages += zones; 1965 else 1966 pages += howmany(zones, 1967 UMA_SLAB_SPACE / roundup2(zsize, UMA_BOOT_ALIGN)); 1968 1969 /* ... and their kegs. Note that zone of zones allocates a keg! */ 1970 pages += howmany(zones + 1, 1971 UMA_SLAB_SPACE / roundup2(ksize, UMA_BOOT_ALIGN)); 1972 1973 /* 1974 * Most of startup zones are not going to be offpages, that's 1975 * why we use UMA_SLAB_SPACE instead of UMA_SLAB_SIZE in all 1976 * calculations. Some large bucket zones will be offpage, and 1977 * thus will allocate hashes. We take conservative approach 1978 * and assume that all zones may allocate hash. This may give 1979 * us some positive inaccuracy, usually an extra single page. 1980 */ 1981 pages += howmany(zones, UMA_SLAB_SPACE / 1982 (sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT)); 1983 1984 return (pages); 1985 } 1986 1987 void 1988 uma_startup(void *mem, int npages) 1989 { 1990 struct uma_zctor_args args; 1991 uma_keg_t masterkeg; 1992 uintptr_t m; 1993 1994 #ifdef DIAGNOSTIC 1995 printf("Entering %s with %d boot pages configured\n", __func__, npages); 1996 #endif 1997 1998 rw_init(&uma_rwlock, "UMA lock"); 1999 2000 /* Use bootpages memory for the zone of zones and zone of kegs. */ 2001 m = (uintptr_t)mem; 2002 zones = (uma_zone_t)m; 2003 m += roundup(zsize, CACHE_LINE_SIZE); 2004 kegs = (uma_zone_t)m; 2005 m += roundup(zsize, CACHE_LINE_SIZE); 2006 masterkeg = (uma_keg_t)m; 2007 m += roundup(ksize, CACHE_LINE_SIZE); 2008 m = roundup(m, PAGE_SIZE); 2009 npages -= (m - (uintptr_t)mem) / PAGE_SIZE; 2010 mem = (void *)m; 2011 2012 /* "manually" create the initial zone */ 2013 memset(&args, 0, sizeof(args)); 2014 args.name = "UMA Kegs"; 2015 args.size = ksize; 2016 args.ctor = keg_ctor; 2017 args.dtor = keg_dtor; 2018 args.uminit = zero_init; 2019 args.fini = NULL; 2020 args.keg = masterkeg; 2021 args.align = UMA_BOOT_ALIGN - 1; 2022 args.flags = UMA_ZFLAG_INTERNAL; 2023 zone_ctor(kegs, zsize, &args, M_WAITOK); 2024 2025 bootmem = mem; 2026 boot_pages = npages; 2027 2028 args.name = "UMA Zones"; 2029 args.size = zsize; 2030 args.ctor = zone_ctor; 2031 args.dtor = zone_dtor; 2032 args.uminit = zero_init; 2033 args.fini = NULL; 2034 args.keg = NULL; 2035 args.align = UMA_BOOT_ALIGN - 1; 2036 args.flags = UMA_ZFLAG_INTERNAL; 2037 zone_ctor(zones, zsize, &args, M_WAITOK); 2038 2039 /* Now make a zone for slab headers */ 2040 slabzone = uma_zcreate("UMA Slabs", 2041 sizeof(struct uma_slab), 2042 NULL, NULL, NULL, NULL, 2043 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2044 2045 hashzone = uma_zcreate("UMA Hash", 2046 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 2047 NULL, NULL, NULL, NULL, 2048 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2049 2050 bucket_init(); 2051 2052 booted = BOOT_STRAPPED; 2053 } 2054 2055 void 2056 uma_startup1(void) 2057 { 2058 2059 #ifdef DIAGNOSTIC 2060 printf("Entering %s with %d boot pages left\n", __func__, boot_pages); 2061 #endif 2062 booted = BOOT_PAGEALLOC; 2063 } 2064 2065 void 2066 uma_startup2(void) 2067 { 2068 2069 #ifdef DIAGNOSTIC 2070 printf("Entering %s with %d boot pages left\n", __func__, boot_pages); 2071 #endif 2072 booted = BOOT_BUCKETS; 2073 sx_init(&uma_drain_lock, "umadrain"); 2074 bucket_enable(); 2075 } 2076 2077 /* 2078 * Initialize our callout handle 2079 * 2080 */ 2081 static void 2082 uma_startup3(void) 2083 { 2084 2085 #ifdef INVARIANTS 2086 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor); 2087 uma_dbg_cnt = counter_u64_alloc(M_WAITOK); 2088 uma_skip_cnt = counter_u64_alloc(M_WAITOK); 2089 #endif 2090 callout_init(&uma_callout, 1); 2091 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 2092 booted = BOOT_RUNNING; 2093 } 2094 2095 static uma_keg_t 2096 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 2097 int align, uint32_t flags) 2098 { 2099 struct uma_kctor_args args; 2100 2101 args.size = size; 2102 args.uminit = uminit; 2103 args.fini = fini; 2104 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 2105 args.flags = flags; 2106 args.zone = zone; 2107 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK)); 2108 } 2109 2110 /* Public functions */ 2111 /* See uma.h */ 2112 void 2113 uma_set_align(int align) 2114 { 2115 2116 if (align != UMA_ALIGN_CACHE) 2117 uma_align_cache = align; 2118 } 2119 2120 /* See uma.h */ 2121 uma_zone_t 2122 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 2123 uma_init uminit, uma_fini fini, int align, uint32_t flags) 2124 2125 { 2126 struct uma_zctor_args args; 2127 uma_zone_t res; 2128 bool locked; 2129 2130 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"", 2131 align, name)); 2132 2133 /* This stuff is essential for the zone ctor */ 2134 memset(&args, 0, sizeof(args)); 2135 args.name = name; 2136 args.size = size; 2137 args.ctor = ctor; 2138 args.dtor = dtor; 2139 args.uminit = uminit; 2140 args.fini = fini; 2141 #ifdef INVARIANTS 2142 /* 2143 * If a zone is being created with an empty constructor and 2144 * destructor, pass UMA constructor/destructor which checks for 2145 * memory use after free. 2146 */ 2147 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) && 2148 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) { 2149 args.ctor = trash_ctor; 2150 args.dtor = trash_dtor; 2151 args.uminit = trash_init; 2152 args.fini = trash_fini; 2153 } 2154 #endif 2155 args.align = align; 2156 args.flags = flags; 2157 args.keg = NULL; 2158 2159 if (booted < BOOT_BUCKETS) { 2160 locked = false; 2161 } else { 2162 sx_slock(&uma_drain_lock); 2163 locked = true; 2164 } 2165 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2166 if (locked) 2167 sx_sunlock(&uma_drain_lock); 2168 return (res); 2169 } 2170 2171 /* See uma.h */ 2172 uma_zone_t 2173 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 2174 uma_init zinit, uma_fini zfini, uma_zone_t master) 2175 { 2176 struct uma_zctor_args args; 2177 uma_keg_t keg; 2178 uma_zone_t res; 2179 bool locked; 2180 2181 keg = zone_first_keg(master); 2182 memset(&args, 0, sizeof(args)); 2183 args.name = name; 2184 args.size = keg->uk_size; 2185 args.ctor = ctor; 2186 args.dtor = dtor; 2187 args.uminit = zinit; 2188 args.fini = zfini; 2189 args.align = keg->uk_align; 2190 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 2191 args.keg = keg; 2192 2193 if (booted < BOOT_BUCKETS) { 2194 locked = false; 2195 } else { 2196 sx_slock(&uma_drain_lock); 2197 locked = true; 2198 } 2199 /* XXX Attaches only one keg of potentially many. */ 2200 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2201 if (locked) 2202 sx_sunlock(&uma_drain_lock); 2203 return (res); 2204 } 2205 2206 /* See uma.h */ 2207 uma_zone_t 2208 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 2209 uma_init zinit, uma_fini zfini, uma_import zimport, 2210 uma_release zrelease, void *arg, int flags) 2211 { 2212 struct uma_zctor_args args; 2213 2214 memset(&args, 0, sizeof(args)); 2215 args.name = name; 2216 args.size = size; 2217 args.ctor = ctor; 2218 args.dtor = dtor; 2219 args.uminit = zinit; 2220 args.fini = zfini; 2221 args.import = zimport; 2222 args.release = zrelease; 2223 args.arg = arg; 2224 args.align = 0; 2225 args.flags = flags; 2226 2227 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK)); 2228 } 2229 2230 static void 2231 zone_lock_pair(uma_zone_t a, uma_zone_t b) 2232 { 2233 if (a < b) { 2234 ZONE_LOCK(a); 2235 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); 2236 } else { 2237 ZONE_LOCK(b); 2238 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); 2239 } 2240 } 2241 2242 static void 2243 zone_unlock_pair(uma_zone_t a, uma_zone_t b) 2244 { 2245 2246 ZONE_UNLOCK(a); 2247 ZONE_UNLOCK(b); 2248 } 2249 2250 int 2251 uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 2252 { 2253 uma_klink_t klink; 2254 uma_klink_t kl; 2255 int error; 2256 2257 error = 0; 2258 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 2259 2260 zone_lock_pair(zone, master); 2261 /* 2262 * zone must use vtoslab() to resolve objects and must already be 2263 * a secondary. 2264 */ 2265 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 2266 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 2267 error = EINVAL; 2268 goto out; 2269 } 2270 /* 2271 * The new master must also use vtoslab(). 2272 */ 2273 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 2274 error = EINVAL; 2275 goto out; 2276 } 2277 2278 /* 2279 * The underlying object must be the same size. rsize 2280 * may be different. 2281 */ 2282 if (master->uz_size != zone->uz_size) { 2283 error = E2BIG; 2284 goto out; 2285 } 2286 /* 2287 * Put it at the end of the list. 2288 */ 2289 klink->kl_keg = zone_first_keg(master); 2290 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 2291 if (LIST_NEXT(kl, kl_link) == NULL) { 2292 LIST_INSERT_AFTER(kl, klink, kl_link); 2293 break; 2294 } 2295 } 2296 klink = NULL; 2297 zone->uz_flags |= UMA_ZFLAG_MULTI; 2298 zone->uz_slab = zone_fetch_slab_multi; 2299 2300 out: 2301 zone_unlock_pair(zone, master); 2302 if (klink != NULL) 2303 free(klink, M_TEMP); 2304 2305 return (error); 2306 } 2307 2308 2309 /* See uma.h */ 2310 void 2311 uma_zdestroy(uma_zone_t zone) 2312 { 2313 2314 sx_slock(&uma_drain_lock); 2315 zone_free_item(zones, zone, NULL, SKIP_NONE); 2316 sx_sunlock(&uma_drain_lock); 2317 } 2318 2319 void 2320 uma_zwait(uma_zone_t zone) 2321 { 2322 void *item; 2323 2324 item = uma_zalloc_arg(zone, NULL, M_WAITOK); 2325 uma_zfree(zone, item); 2326 } 2327 2328 void * 2329 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags) 2330 { 2331 void *item; 2332 #ifdef SMP 2333 int i; 2334 2335 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2336 #endif 2337 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO); 2338 if (item != NULL && (flags & M_ZERO)) { 2339 #ifdef SMP 2340 for (i = 0; i <= mp_maxid; i++) 2341 bzero(zpcpu_get_cpu(item, i), zone->uz_size); 2342 #else 2343 bzero(item, zone->uz_size); 2344 #endif 2345 } 2346 return (item); 2347 } 2348 2349 /* 2350 * A stub while both regular and pcpu cases are identical. 2351 */ 2352 void 2353 uma_zfree_pcpu_arg(uma_zone_t zone, void *item, void *udata) 2354 { 2355 2356 #ifdef SMP 2357 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2358 #endif 2359 uma_zfree_arg(zone, item, udata); 2360 } 2361 2362 /* See uma.h */ 2363 void * 2364 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2365 { 2366 uma_zone_domain_t zdom; 2367 uma_bucket_t bucket; 2368 uma_cache_t cache; 2369 void *item; 2370 int cpu, domain, lockfail; 2371 #ifdef INVARIANTS 2372 bool skipdbg; 2373 #endif 2374 2375 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2376 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 2377 2378 /* This is the fast path allocation */ 2379 CTR4(KTR_UMA, "uma_zalloc_arg thread %x zone %s(%p) flags %d", 2380 curthread, zone->uz_name, zone, flags); 2381 2382 if (flags & M_WAITOK) { 2383 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2384 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2385 } 2386 KASSERT((flags & M_EXEC) == 0, ("uma_zalloc_arg: called with M_EXEC")); 2387 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2388 ("uma_zalloc_arg: called with spinlock or critical section held")); 2389 if (zone->uz_flags & UMA_ZONE_PCPU) 2390 KASSERT((flags & M_ZERO) == 0, ("allocating from a pcpu zone " 2391 "with M_ZERO passed")); 2392 2393 #ifdef DEBUG_MEMGUARD 2394 if (memguard_cmp_zone(zone)) { 2395 item = memguard_alloc(zone->uz_size, flags); 2396 if (item != NULL) { 2397 if (zone->uz_init != NULL && 2398 zone->uz_init(item, zone->uz_size, flags) != 0) 2399 return (NULL); 2400 if (zone->uz_ctor != NULL && 2401 zone->uz_ctor(item, zone->uz_size, udata, 2402 flags) != 0) { 2403 zone->uz_fini(item, zone->uz_size); 2404 return (NULL); 2405 } 2406 return (item); 2407 } 2408 /* This is unfortunate but should not be fatal. */ 2409 } 2410 #endif 2411 /* 2412 * If possible, allocate from the per-CPU cache. There are two 2413 * requirements for safe access to the per-CPU cache: (1) the thread 2414 * accessing the cache must not be preempted or yield during access, 2415 * and (2) the thread must not migrate CPUs without switching which 2416 * cache it accesses. We rely on a critical section to prevent 2417 * preemption and migration. We release the critical section in 2418 * order to acquire the zone mutex if we are unable to allocate from 2419 * the current cache; when we re-acquire the critical section, we 2420 * must detect and handle migration if it has occurred. 2421 */ 2422 zalloc_restart: 2423 critical_enter(); 2424 cpu = curcpu; 2425 cache = &zone->uz_cpu[cpu]; 2426 2427 zalloc_start: 2428 bucket = cache->uc_allocbucket; 2429 if (bucket != NULL && bucket->ub_cnt > 0) { 2430 bucket->ub_cnt--; 2431 item = bucket->ub_bucket[bucket->ub_cnt]; 2432 #ifdef INVARIANTS 2433 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2434 #endif 2435 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2436 cache->uc_allocs++; 2437 critical_exit(); 2438 #ifdef INVARIANTS 2439 skipdbg = uma_dbg_zskip(zone, item); 2440 #endif 2441 if (zone->uz_ctor != NULL && 2442 #ifdef INVARIANTS 2443 (!skipdbg || zone->uz_ctor != trash_ctor || 2444 zone->uz_dtor != trash_dtor) && 2445 #endif 2446 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2447 atomic_add_long(&zone->uz_fails, 1); 2448 zone_free_item(zone, item, udata, SKIP_DTOR); 2449 return (NULL); 2450 } 2451 #ifdef INVARIANTS 2452 if (!skipdbg) 2453 uma_dbg_alloc(zone, NULL, item); 2454 #endif 2455 if (flags & M_ZERO) 2456 uma_zero_item(item, zone); 2457 return (item); 2458 } 2459 2460 /* 2461 * We have run out of items in our alloc bucket. 2462 * See if we can switch with our free bucket. 2463 */ 2464 bucket = cache->uc_freebucket; 2465 if (bucket != NULL && bucket->ub_cnt > 0) { 2466 CTR2(KTR_UMA, 2467 "uma_zalloc: zone %s(%p) swapping empty with alloc", 2468 zone->uz_name, zone); 2469 cache->uc_freebucket = cache->uc_allocbucket; 2470 cache->uc_allocbucket = bucket; 2471 goto zalloc_start; 2472 } 2473 2474 /* 2475 * Discard any empty allocation bucket while we hold no locks. 2476 */ 2477 bucket = cache->uc_allocbucket; 2478 cache->uc_allocbucket = NULL; 2479 critical_exit(); 2480 if (bucket != NULL) 2481 bucket_free(zone, bucket, udata); 2482 2483 if (zone->uz_flags & UMA_ZONE_NUMA) { 2484 domain = PCPU_GET(domain); 2485 if (VM_DOMAIN_EMPTY(domain)) 2486 domain = UMA_ANYDOMAIN; 2487 } else 2488 domain = UMA_ANYDOMAIN; 2489 2490 /* Short-circuit for zones without buckets and low memory. */ 2491 if (zone->uz_count == 0 || bucketdisable) 2492 goto zalloc_item; 2493 2494 /* 2495 * Attempt to retrieve the item from the per-CPU cache has failed, so 2496 * we must go back to the zone. This requires the zone lock, so we 2497 * must drop the critical section, then re-acquire it when we go back 2498 * to the cache. Since the critical section is released, we may be 2499 * preempted or migrate. As such, make sure not to maintain any 2500 * thread-local state specific to the cache from prior to releasing 2501 * the critical section. 2502 */ 2503 lockfail = 0; 2504 if (ZONE_TRYLOCK(zone) == 0) { 2505 /* Record contention to size the buckets. */ 2506 ZONE_LOCK(zone); 2507 lockfail = 1; 2508 } 2509 critical_enter(); 2510 cpu = curcpu; 2511 cache = &zone->uz_cpu[cpu]; 2512 2513 /* See if we lost the race to fill the cache. */ 2514 if (cache->uc_allocbucket != NULL) { 2515 ZONE_UNLOCK(zone); 2516 goto zalloc_start; 2517 } 2518 2519 /* 2520 * Check the zone's cache of buckets. 2521 */ 2522 if (domain == UMA_ANYDOMAIN) 2523 zdom = &zone->uz_domain[0]; 2524 else 2525 zdom = &zone->uz_domain[domain]; 2526 if ((bucket = LIST_FIRST(&zdom->uzd_buckets)) != NULL) { 2527 KASSERT(bucket->ub_cnt != 0, 2528 ("uma_zalloc_arg: Returning an empty bucket.")); 2529 2530 LIST_REMOVE(bucket, ub_link); 2531 cache->uc_allocbucket = bucket; 2532 ZONE_UNLOCK(zone); 2533 goto zalloc_start; 2534 } 2535 /* We are no longer associated with this CPU. */ 2536 critical_exit(); 2537 2538 /* 2539 * We bump the uz count when the cache size is insufficient to 2540 * handle the working set. 2541 */ 2542 if (lockfail && zone->uz_count < BUCKET_MAX) 2543 zone->uz_count++; 2544 ZONE_UNLOCK(zone); 2545 2546 /* 2547 * Now lets just fill a bucket and put it on the free list. If that 2548 * works we'll restart the allocation from the beginning and it 2549 * will use the just filled bucket. 2550 */ 2551 bucket = zone_alloc_bucket(zone, udata, domain, flags); 2552 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p", 2553 zone->uz_name, zone, bucket); 2554 if (bucket != NULL) { 2555 ZONE_LOCK(zone); 2556 critical_enter(); 2557 cpu = curcpu; 2558 cache = &zone->uz_cpu[cpu]; 2559 /* 2560 * See if we lost the race or were migrated. Cache the 2561 * initialized bucket to make this less likely or claim 2562 * the memory directly. 2563 */ 2564 if (cache->uc_allocbucket == NULL && 2565 ((zone->uz_flags & UMA_ZONE_NUMA) == 0 || 2566 domain == PCPU_GET(domain))) { 2567 cache->uc_allocbucket = bucket; 2568 } else if ((zone->uz_flags & UMA_ZONE_NOBUCKETCACHE) != 0) { 2569 critical_exit(); 2570 ZONE_UNLOCK(zone); 2571 bucket_drain(zone, bucket); 2572 bucket_free(zone, bucket, udata); 2573 goto zalloc_restart; 2574 } else 2575 LIST_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link); 2576 ZONE_UNLOCK(zone); 2577 goto zalloc_start; 2578 } 2579 2580 /* 2581 * We may not be able to get a bucket so return an actual item. 2582 */ 2583 zalloc_item: 2584 item = zone_alloc_item(zone, udata, domain, flags); 2585 2586 return (item); 2587 } 2588 2589 void * 2590 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags) 2591 { 2592 2593 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2594 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 2595 2596 /* This is the fast path allocation */ 2597 CTR5(KTR_UMA, 2598 "uma_zalloc_domain thread %x zone %s(%p) domain %d flags %d", 2599 curthread, zone->uz_name, zone, domain, flags); 2600 2601 if (flags & M_WAITOK) { 2602 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2603 "uma_zalloc_domain: zone \"%s\"", zone->uz_name); 2604 } 2605 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2606 ("uma_zalloc_domain: called with spinlock or critical section held")); 2607 2608 return (zone_alloc_item(zone, udata, domain, flags)); 2609 } 2610 2611 /* 2612 * Find a slab with some space. Prefer slabs that are partially used over those 2613 * that are totally full. This helps to reduce fragmentation. 2614 * 2615 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check 2616 * only 'domain'. 2617 */ 2618 static uma_slab_t 2619 keg_first_slab(uma_keg_t keg, int domain, bool rr) 2620 { 2621 uma_domain_t dom; 2622 uma_slab_t slab; 2623 int start; 2624 2625 KASSERT(domain >= 0 && domain < vm_ndomains, 2626 ("keg_first_slab: domain %d out of range", domain)); 2627 2628 slab = NULL; 2629 start = domain; 2630 do { 2631 dom = &keg->uk_domain[domain]; 2632 if (!LIST_EMPTY(&dom->ud_part_slab)) 2633 return (LIST_FIRST(&dom->ud_part_slab)); 2634 if (!LIST_EMPTY(&dom->ud_free_slab)) { 2635 slab = LIST_FIRST(&dom->ud_free_slab); 2636 LIST_REMOVE(slab, us_link); 2637 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 2638 return (slab); 2639 } 2640 if (rr) 2641 domain = (domain + 1) % vm_ndomains; 2642 } while (domain != start); 2643 2644 return (NULL); 2645 } 2646 2647 static uma_slab_t 2648 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags) 2649 { 2650 uint32_t reserve; 2651 2652 mtx_assert(&keg->uk_lock, MA_OWNED); 2653 2654 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve; 2655 if (keg->uk_free <= reserve) 2656 return (NULL); 2657 return (keg_first_slab(keg, domain, rr)); 2658 } 2659 2660 static uma_slab_t 2661 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags) 2662 { 2663 struct vm_domainset_iter di; 2664 uma_domain_t dom; 2665 uma_slab_t slab; 2666 int aflags, domain; 2667 bool rr; 2668 2669 restart: 2670 mtx_assert(&keg->uk_lock, MA_OWNED); 2671 2672 /* 2673 * Use the keg's policy if upper layers haven't already specified a 2674 * domain (as happens with first-touch zones). 2675 * 2676 * To avoid races we run the iterator with the keg lock held, but that 2677 * means that we cannot allow the vm_domainset layer to sleep. Thus, 2678 * clear M_WAITOK and handle low memory conditions locally. 2679 */ 2680 rr = rdomain == UMA_ANYDOMAIN; 2681 if (rr) { 2682 aflags = (flags & ~M_WAITOK) | M_NOWAIT; 2683 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 2684 &aflags); 2685 } else { 2686 aflags = flags; 2687 domain = rdomain; 2688 } 2689 2690 for (;;) { 2691 slab = keg_fetch_free_slab(keg, domain, rr, flags); 2692 if (slab != NULL) { 2693 MPASS(slab->us_keg == keg); 2694 return (slab); 2695 } 2696 2697 /* 2698 * M_NOVM means don't ask at all! 2699 */ 2700 if (flags & M_NOVM) 2701 break; 2702 2703 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2704 keg->uk_flags |= UMA_ZFLAG_FULL; 2705 /* 2706 * If this is not a multi-zone, set the FULL bit. 2707 * Otherwise slab_multi() takes care of it. 2708 */ 2709 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2710 zone->uz_flags |= UMA_ZFLAG_FULL; 2711 zone_log_warning(zone); 2712 zone_maxaction(zone); 2713 } 2714 if (flags & M_NOWAIT) 2715 return (NULL); 2716 zone->uz_sleeps++; 2717 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2718 continue; 2719 } 2720 slab = keg_alloc_slab(keg, zone, domain, aflags); 2721 /* 2722 * If we got a slab here it's safe to mark it partially used 2723 * and return. We assume that the caller is going to remove 2724 * at least one item. 2725 */ 2726 if (slab) { 2727 MPASS(slab->us_keg == keg); 2728 dom = &keg->uk_domain[slab->us_domain]; 2729 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 2730 return (slab); 2731 } 2732 KEG_LOCK(keg); 2733 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) { 2734 if ((flags & M_WAITOK) != 0) { 2735 KEG_UNLOCK(keg); 2736 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask); 2737 KEG_LOCK(keg); 2738 goto restart; 2739 } 2740 break; 2741 } 2742 } 2743 2744 /* 2745 * We might not have been able to get a slab but another cpu 2746 * could have while we were unlocked. Check again before we 2747 * fail. 2748 */ 2749 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL) { 2750 MPASS(slab->us_keg == keg); 2751 return (slab); 2752 } 2753 return (NULL); 2754 } 2755 2756 static uma_slab_t 2757 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int domain, int flags) 2758 { 2759 uma_slab_t slab; 2760 2761 if (keg == NULL) { 2762 keg = zone_first_keg(zone); 2763 KEG_LOCK(keg); 2764 } 2765 2766 for (;;) { 2767 slab = keg_fetch_slab(keg, zone, domain, flags); 2768 if (slab) 2769 return (slab); 2770 if (flags & (M_NOWAIT | M_NOVM)) 2771 break; 2772 } 2773 KEG_UNLOCK(keg); 2774 return (NULL); 2775 } 2776 2777 /* 2778 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2779 * with the keg locked. On NULL no lock is held. 2780 * 2781 * The last pointer is used to seed the search. It is not required. 2782 */ 2783 static uma_slab_t 2784 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int domain, int rflags) 2785 { 2786 uma_klink_t klink; 2787 uma_slab_t slab; 2788 uma_keg_t keg; 2789 int flags; 2790 int empty; 2791 int full; 2792 2793 /* 2794 * Don't wait on the first pass. This will skip limit tests 2795 * as well. We don't want to block if we can find a provider 2796 * without blocking. 2797 */ 2798 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2799 /* 2800 * Use the last slab allocated as a hint for where to start 2801 * the search. 2802 */ 2803 if (last != NULL) { 2804 slab = keg_fetch_slab(last, zone, domain, flags); 2805 if (slab) 2806 return (slab); 2807 KEG_UNLOCK(last); 2808 } 2809 /* 2810 * Loop until we have a slab incase of transient failures 2811 * while M_WAITOK is specified. I'm not sure this is 100% 2812 * required but we've done it for so long now. 2813 */ 2814 for (;;) { 2815 empty = 0; 2816 full = 0; 2817 /* 2818 * Search the available kegs for slabs. Be careful to hold the 2819 * correct lock while calling into the keg layer. 2820 */ 2821 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2822 keg = klink->kl_keg; 2823 KEG_LOCK(keg); 2824 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2825 slab = keg_fetch_slab(keg, zone, domain, flags); 2826 if (slab) 2827 return (slab); 2828 } 2829 if (keg->uk_flags & UMA_ZFLAG_FULL) 2830 full++; 2831 else 2832 empty++; 2833 KEG_UNLOCK(keg); 2834 } 2835 if (rflags & (M_NOWAIT | M_NOVM)) 2836 break; 2837 flags = rflags; 2838 /* 2839 * All kegs are full. XXX We can't atomically check all kegs 2840 * and sleep so just sleep for a short period and retry. 2841 */ 2842 if (full && !empty) { 2843 ZONE_LOCK(zone); 2844 zone->uz_flags |= UMA_ZFLAG_FULL; 2845 zone->uz_sleeps++; 2846 zone_log_warning(zone); 2847 zone_maxaction(zone); 2848 msleep(zone, zone->uz_lockptr, PVM, 2849 "zonelimit", hz/100); 2850 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2851 ZONE_UNLOCK(zone); 2852 continue; 2853 } 2854 } 2855 return (NULL); 2856 } 2857 2858 static void * 2859 slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 2860 { 2861 uma_domain_t dom; 2862 void *item; 2863 uint8_t freei; 2864 2865 MPASS(keg == slab->us_keg); 2866 mtx_assert(&keg->uk_lock, MA_OWNED); 2867 2868 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; 2869 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); 2870 item = slab->us_data + (keg->uk_rsize * freei); 2871 slab->us_freecount--; 2872 keg->uk_free--; 2873 2874 /* Move this slab to the full list */ 2875 if (slab->us_freecount == 0) { 2876 LIST_REMOVE(slab, us_link); 2877 dom = &keg->uk_domain[slab->us_domain]; 2878 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link); 2879 } 2880 2881 return (item); 2882 } 2883 2884 static int 2885 zone_import(uma_zone_t zone, void **bucket, int max, int domain, int flags) 2886 { 2887 uma_slab_t slab; 2888 uma_keg_t keg; 2889 #ifdef NUMA 2890 int stripe; 2891 #endif 2892 int i; 2893 2894 slab = NULL; 2895 keg = NULL; 2896 /* Try to keep the buckets totally full */ 2897 for (i = 0; i < max; ) { 2898 if ((slab = zone->uz_slab(zone, keg, domain, flags)) == NULL) 2899 break; 2900 keg = slab->us_keg; 2901 #ifdef NUMA 2902 stripe = howmany(max, vm_ndomains); 2903 #endif 2904 while (slab->us_freecount && i < max) { 2905 bucket[i++] = slab_alloc_item(keg, slab); 2906 if (keg->uk_free <= keg->uk_reserve) 2907 break; 2908 #ifdef NUMA 2909 /* 2910 * If the zone is striped we pick a new slab for every 2911 * N allocations. Eliminating this conditional will 2912 * instead pick a new domain for each bucket rather 2913 * than stripe within each bucket. The current option 2914 * produces more fragmentation and requires more cpu 2915 * time but yields better distribution. 2916 */ 2917 if ((zone->uz_flags & UMA_ZONE_NUMA) == 0 && 2918 vm_ndomains > 1 && --stripe == 0) 2919 break; 2920 #endif 2921 } 2922 /* Don't block if we allocated any successfully. */ 2923 flags &= ~M_WAITOK; 2924 flags |= M_NOWAIT; 2925 } 2926 if (slab != NULL) 2927 KEG_UNLOCK(keg); 2928 2929 return i; 2930 } 2931 2932 static uma_bucket_t 2933 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags) 2934 { 2935 uma_bucket_t bucket; 2936 int max; 2937 2938 CTR1(KTR_UMA, "zone_alloc:_bucket domain %d)", domain); 2939 2940 /* Don't wait for buckets, preserve caller's NOVM setting. */ 2941 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 2942 if (bucket == NULL) 2943 return (NULL); 2944 2945 max = MIN(bucket->ub_entries, zone->uz_count); 2946 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 2947 max, domain, flags); 2948 2949 /* 2950 * Initialize the memory if necessary. 2951 */ 2952 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 2953 int i; 2954 2955 for (i = 0; i < bucket->ub_cnt; i++) 2956 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2957 flags) != 0) 2958 break; 2959 /* 2960 * If we couldn't initialize the whole bucket, put the 2961 * rest back onto the freelist. 2962 */ 2963 if (i != bucket->ub_cnt) { 2964 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 2965 bucket->ub_cnt - i); 2966 #ifdef INVARIANTS 2967 bzero(&bucket->ub_bucket[i], 2968 sizeof(void *) * (bucket->ub_cnt - i)); 2969 #endif 2970 bucket->ub_cnt = i; 2971 } 2972 } 2973 2974 if (bucket->ub_cnt == 0) { 2975 bucket_free(zone, bucket, udata); 2976 atomic_add_long(&zone->uz_fails, 1); 2977 return (NULL); 2978 } 2979 2980 return (bucket); 2981 } 2982 2983 /* 2984 * Allocates a single item from a zone. 2985 * 2986 * Arguments 2987 * zone The zone to alloc for. 2988 * udata The data to be passed to the constructor. 2989 * domain The domain to allocate from or UMA_ANYDOMAIN. 2990 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2991 * 2992 * Returns 2993 * NULL if there is no memory and M_NOWAIT is set 2994 * An item if successful 2995 */ 2996 2997 static void * 2998 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags) 2999 { 3000 void *item; 3001 #ifdef INVARIANTS 3002 bool skipdbg; 3003 #endif 3004 3005 item = NULL; 3006 3007 if (domain != UMA_ANYDOMAIN) { 3008 /* avoid allocs targeting empty domains */ 3009 if (VM_DOMAIN_EMPTY(domain)) 3010 domain = UMA_ANYDOMAIN; 3011 } 3012 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1) 3013 goto fail; 3014 atomic_add_long(&zone->uz_allocs, 1); 3015 3016 #ifdef INVARIANTS 3017 skipdbg = uma_dbg_zskip(zone, item); 3018 #endif 3019 /* 3020 * We have to call both the zone's init (not the keg's init) 3021 * and the zone's ctor. This is because the item is going from 3022 * a keg slab directly to the user, and the user is expecting it 3023 * to be both zone-init'd as well as zone-ctor'd. 3024 */ 3025 if (zone->uz_init != NULL) { 3026 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 3027 zone_free_item(zone, item, udata, SKIP_FINI); 3028 goto fail; 3029 } 3030 } 3031 if (zone->uz_ctor != NULL && 3032 #ifdef INVARIANTS 3033 (!skipdbg || zone->uz_ctor != trash_ctor || 3034 zone->uz_dtor != trash_dtor) && 3035 #endif 3036 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 3037 zone_free_item(zone, item, udata, SKIP_DTOR); 3038 goto fail; 3039 } 3040 #ifdef INVARIANTS 3041 if (!skipdbg) 3042 uma_dbg_alloc(zone, NULL, item); 3043 #endif 3044 if (flags & M_ZERO) 3045 uma_zero_item(item, zone); 3046 3047 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item, 3048 zone->uz_name, zone); 3049 3050 return (item); 3051 3052 fail: 3053 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)", 3054 zone->uz_name, zone); 3055 atomic_add_long(&zone->uz_fails, 1); 3056 return (NULL); 3057 } 3058 3059 /* See uma.h */ 3060 void 3061 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 3062 { 3063 uma_cache_t cache; 3064 uma_bucket_t bucket; 3065 uma_zone_domain_t zdom; 3066 int cpu, domain, lockfail; 3067 #ifdef INVARIANTS 3068 bool skipdbg; 3069 #endif 3070 3071 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3072 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3073 3074 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 3075 zone->uz_name); 3076 3077 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3078 ("uma_zfree_arg: called with spinlock or critical section held")); 3079 3080 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 3081 if (item == NULL) 3082 return; 3083 #ifdef DEBUG_MEMGUARD 3084 if (is_memguard_addr(item)) { 3085 if (zone->uz_dtor != NULL) 3086 zone->uz_dtor(item, zone->uz_size, udata); 3087 if (zone->uz_fini != NULL) 3088 zone->uz_fini(item, zone->uz_size); 3089 memguard_free(item); 3090 return; 3091 } 3092 #endif 3093 #ifdef INVARIANTS 3094 skipdbg = uma_dbg_zskip(zone, item); 3095 if (skipdbg == false) { 3096 if (zone->uz_flags & UMA_ZONE_MALLOC) 3097 uma_dbg_free(zone, udata, item); 3098 else 3099 uma_dbg_free(zone, NULL, item); 3100 } 3101 if (zone->uz_dtor != NULL && (!skipdbg || 3102 zone->uz_dtor != trash_dtor || zone->uz_ctor != trash_ctor)) 3103 #else 3104 if (zone->uz_dtor != NULL) 3105 #endif 3106 zone->uz_dtor(item, zone->uz_size, udata); 3107 3108 /* 3109 * The race here is acceptable. If we miss it we'll just have to wait 3110 * a little longer for the limits to be reset. 3111 */ 3112 if (zone->uz_flags & UMA_ZFLAG_FULL) 3113 goto zfree_item; 3114 3115 /* 3116 * If possible, free to the per-CPU cache. There are two 3117 * requirements for safe access to the per-CPU cache: (1) the thread 3118 * accessing the cache must not be preempted or yield during access, 3119 * and (2) the thread must not migrate CPUs without switching which 3120 * cache it accesses. We rely on a critical section to prevent 3121 * preemption and migration. We release the critical section in 3122 * order to acquire the zone mutex if we are unable to free to the 3123 * current cache; when we re-acquire the critical section, we must 3124 * detect and handle migration if it has occurred. 3125 */ 3126 zfree_restart: 3127 critical_enter(); 3128 cpu = curcpu; 3129 cache = &zone->uz_cpu[cpu]; 3130 3131 zfree_start: 3132 /* 3133 * Try to free into the allocbucket first to give LIFO ordering 3134 * for cache-hot datastructures. Spill over into the freebucket 3135 * if necessary. Alloc will swap them if one runs dry. 3136 */ 3137 bucket = cache->uc_allocbucket; 3138 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 3139 bucket = cache->uc_freebucket; 3140 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 3141 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 3142 ("uma_zfree: Freeing to non free bucket index.")); 3143 bucket->ub_bucket[bucket->ub_cnt] = item; 3144 bucket->ub_cnt++; 3145 cache->uc_frees++; 3146 critical_exit(); 3147 return; 3148 } 3149 3150 /* 3151 * We must go back the zone, which requires acquiring the zone lock, 3152 * which in turn means we must release and re-acquire the critical 3153 * section. Since the critical section is released, we may be 3154 * preempted or migrate. As such, make sure not to maintain any 3155 * thread-local state specific to the cache from prior to releasing 3156 * the critical section. 3157 */ 3158 critical_exit(); 3159 if (zone->uz_count == 0 || bucketdisable) 3160 goto zfree_item; 3161 3162 lockfail = 0; 3163 if (ZONE_TRYLOCK(zone) == 0) { 3164 /* Record contention to size the buckets. */ 3165 ZONE_LOCK(zone); 3166 lockfail = 1; 3167 } 3168 critical_enter(); 3169 cpu = curcpu; 3170 cache = &zone->uz_cpu[cpu]; 3171 3172 bucket = cache->uc_freebucket; 3173 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 3174 ZONE_UNLOCK(zone); 3175 goto zfree_start; 3176 } 3177 cache->uc_freebucket = NULL; 3178 /* We are no longer associated with this CPU. */ 3179 critical_exit(); 3180 3181 if ((zone->uz_flags & UMA_ZONE_NUMA) != 0) { 3182 domain = PCPU_GET(domain); 3183 if (VM_DOMAIN_EMPTY(domain)) 3184 domain = UMA_ANYDOMAIN; 3185 } else 3186 domain = 0; 3187 zdom = &zone->uz_domain[0]; 3188 3189 /* Can we throw this on the zone full list? */ 3190 if (bucket != NULL) { 3191 CTR3(KTR_UMA, 3192 "uma_zfree: zone %s(%p) putting bucket %p on free list", 3193 zone->uz_name, zone, bucket); 3194 /* ub_cnt is pointing to the last free item */ 3195 KASSERT(bucket->ub_cnt != 0, 3196 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 3197 if ((zone->uz_flags & UMA_ZONE_NOBUCKETCACHE) != 0) { 3198 ZONE_UNLOCK(zone); 3199 bucket_drain(zone, bucket); 3200 bucket_free(zone, bucket, udata); 3201 goto zfree_restart; 3202 } else 3203 LIST_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link); 3204 } 3205 3206 /* 3207 * We bump the uz count when the cache size is insufficient to 3208 * handle the working set. 3209 */ 3210 if (lockfail && zone->uz_count < BUCKET_MAX) 3211 zone->uz_count++; 3212 ZONE_UNLOCK(zone); 3213 3214 bucket = bucket_alloc(zone, udata, M_NOWAIT); 3215 CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p", 3216 zone->uz_name, zone, bucket); 3217 if (bucket) { 3218 critical_enter(); 3219 cpu = curcpu; 3220 cache = &zone->uz_cpu[cpu]; 3221 if (cache->uc_freebucket == NULL && 3222 ((zone->uz_flags & UMA_ZONE_NUMA) == 0 || 3223 domain == PCPU_GET(domain))) { 3224 cache->uc_freebucket = bucket; 3225 goto zfree_start; 3226 } 3227 /* 3228 * We lost the race, start over. We have to drop our 3229 * critical section to free the bucket. 3230 */ 3231 critical_exit(); 3232 bucket_free(zone, bucket, udata); 3233 goto zfree_restart; 3234 } 3235 3236 /* 3237 * If nothing else caught this, we'll just do an internal free. 3238 */ 3239 zfree_item: 3240 zone_free_item(zone, item, udata, SKIP_DTOR); 3241 3242 return; 3243 } 3244 3245 void 3246 uma_zfree_domain(uma_zone_t zone, void *item, void *udata) 3247 { 3248 3249 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3250 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3251 3252 CTR2(KTR_UMA, "uma_zfree_domain thread %x zone %s", curthread, 3253 zone->uz_name); 3254 3255 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3256 ("uma_zfree_domain: called with spinlock or critical section held")); 3257 3258 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 3259 if (item == NULL) 3260 return; 3261 zone_free_item(zone, item, udata, SKIP_NONE); 3262 } 3263 3264 static void 3265 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) 3266 { 3267 uma_domain_t dom; 3268 uint8_t freei; 3269 3270 mtx_assert(&keg->uk_lock, MA_OWNED); 3271 MPASS(keg == slab->us_keg); 3272 3273 dom = &keg->uk_domain[slab->us_domain]; 3274 3275 /* Do we need to remove from any lists? */ 3276 if (slab->us_freecount+1 == keg->uk_ipers) { 3277 LIST_REMOVE(slab, us_link); 3278 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); 3279 } else if (slab->us_freecount == 0) { 3280 LIST_REMOVE(slab, us_link); 3281 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 3282 } 3283 3284 /* Slab management. */ 3285 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3286 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); 3287 slab->us_freecount++; 3288 3289 /* Keg statistics. */ 3290 keg->uk_free++; 3291 } 3292 3293 static void 3294 zone_release(uma_zone_t zone, void **bucket, int cnt) 3295 { 3296 void *item; 3297 uma_slab_t slab; 3298 uma_keg_t keg; 3299 uint8_t *mem; 3300 int clearfull; 3301 int i; 3302 3303 clearfull = 0; 3304 keg = zone_first_keg(zone); 3305 KEG_LOCK(keg); 3306 for (i = 0; i < cnt; i++) { 3307 item = bucket[i]; 3308 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 3309 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 3310 if (zone->uz_flags & UMA_ZONE_HASH) { 3311 slab = hash_sfind(&keg->uk_hash, mem); 3312 } else { 3313 mem += keg->uk_pgoff; 3314 slab = (uma_slab_t)mem; 3315 } 3316 } else { 3317 slab = vtoslab((vm_offset_t)item); 3318 if (slab->us_keg != keg) { 3319 KEG_UNLOCK(keg); 3320 keg = slab->us_keg; 3321 KEG_LOCK(keg); 3322 } 3323 } 3324 slab_free_item(keg, slab, item); 3325 if (keg->uk_flags & UMA_ZFLAG_FULL) { 3326 if (keg->uk_pages < keg->uk_maxpages) { 3327 keg->uk_flags &= ~UMA_ZFLAG_FULL; 3328 clearfull = 1; 3329 } 3330 3331 /* 3332 * We can handle one more allocation. Since we're 3333 * clearing ZFLAG_FULL, wake up all procs blocked 3334 * on pages. This should be uncommon, so keeping this 3335 * simple for now (rather than adding count of blocked 3336 * threads etc). 3337 */ 3338 wakeup(keg); 3339 } 3340 } 3341 KEG_UNLOCK(keg); 3342 if (clearfull) { 3343 ZONE_LOCK(zone); 3344 zone->uz_flags &= ~UMA_ZFLAG_FULL; 3345 wakeup(zone); 3346 ZONE_UNLOCK(zone); 3347 } 3348 3349 } 3350 3351 /* 3352 * Frees a single item to any zone. 3353 * 3354 * Arguments: 3355 * zone The zone to free to 3356 * item The item we're freeing 3357 * udata User supplied data for the dtor 3358 * skip Skip dtors and finis 3359 */ 3360 static void 3361 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 3362 { 3363 #ifdef INVARIANTS 3364 bool skipdbg; 3365 3366 skipdbg = uma_dbg_zskip(zone, item); 3367 if (skip == SKIP_NONE && !skipdbg) { 3368 if (zone->uz_flags & UMA_ZONE_MALLOC) 3369 uma_dbg_free(zone, udata, item); 3370 else 3371 uma_dbg_free(zone, NULL, item); 3372 } 3373 3374 if (skip < SKIP_DTOR && zone->uz_dtor != NULL && 3375 (!skipdbg || zone->uz_dtor != trash_dtor || 3376 zone->uz_ctor != trash_ctor)) 3377 #else 3378 if (skip < SKIP_DTOR && zone->uz_dtor != NULL) 3379 #endif 3380 zone->uz_dtor(item, zone->uz_size, udata); 3381 3382 if (skip < SKIP_FINI && zone->uz_fini) 3383 zone->uz_fini(item, zone->uz_size); 3384 3385 atomic_add_long(&zone->uz_frees, 1); 3386 zone->uz_release(zone->uz_arg, &item, 1); 3387 } 3388 3389 /* See uma.h */ 3390 int 3391 uma_zone_set_max(uma_zone_t zone, int nitems) 3392 { 3393 uma_keg_t keg; 3394 3395 keg = zone_first_keg(zone); 3396 if (keg == NULL) 3397 return (0); 3398 KEG_LOCK(keg); 3399 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 3400 if (keg->uk_maxpages * keg->uk_ipers < nitems) 3401 keg->uk_maxpages += keg->uk_ppera; 3402 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers; 3403 KEG_UNLOCK(keg); 3404 3405 return (nitems); 3406 } 3407 3408 /* See uma.h */ 3409 int 3410 uma_zone_get_max(uma_zone_t zone) 3411 { 3412 int nitems; 3413 uma_keg_t keg; 3414 3415 keg = zone_first_keg(zone); 3416 if (keg == NULL) 3417 return (0); 3418 KEG_LOCK(keg); 3419 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers; 3420 KEG_UNLOCK(keg); 3421 3422 return (nitems); 3423 } 3424 3425 /* See uma.h */ 3426 void 3427 uma_zone_set_warning(uma_zone_t zone, const char *warning) 3428 { 3429 3430 ZONE_LOCK(zone); 3431 zone->uz_warning = warning; 3432 ZONE_UNLOCK(zone); 3433 } 3434 3435 /* See uma.h */ 3436 void 3437 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) 3438 { 3439 3440 ZONE_LOCK(zone); 3441 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); 3442 ZONE_UNLOCK(zone); 3443 } 3444 3445 /* See uma.h */ 3446 int 3447 uma_zone_get_cur(uma_zone_t zone) 3448 { 3449 int64_t nitems; 3450 u_int i; 3451 3452 ZONE_LOCK(zone); 3453 nitems = zone->uz_allocs - zone->uz_frees; 3454 CPU_FOREACH(i) { 3455 /* 3456 * See the comment in sysctl_vm_zone_stats() regarding the 3457 * safety of accessing the per-cpu caches. With the zone lock 3458 * held, it is safe, but can potentially result in stale data. 3459 */ 3460 nitems += zone->uz_cpu[i].uc_allocs - 3461 zone->uz_cpu[i].uc_frees; 3462 } 3463 ZONE_UNLOCK(zone); 3464 3465 return (nitems < 0 ? 0 : nitems); 3466 } 3467 3468 /* See uma.h */ 3469 void 3470 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 3471 { 3472 uma_keg_t keg; 3473 3474 keg = zone_first_keg(zone); 3475 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 3476 KEG_LOCK(keg); 3477 KASSERT(keg->uk_pages == 0, 3478 ("uma_zone_set_init on non-empty keg")); 3479 keg->uk_init = uminit; 3480 KEG_UNLOCK(keg); 3481 } 3482 3483 /* See uma.h */ 3484 void 3485 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 3486 { 3487 uma_keg_t keg; 3488 3489 keg = zone_first_keg(zone); 3490 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type")); 3491 KEG_LOCK(keg); 3492 KASSERT(keg->uk_pages == 0, 3493 ("uma_zone_set_fini on non-empty keg")); 3494 keg->uk_fini = fini; 3495 KEG_UNLOCK(keg); 3496 } 3497 3498 /* See uma.h */ 3499 void 3500 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 3501 { 3502 3503 ZONE_LOCK(zone); 3504 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3505 ("uma_zone_set_zinit on non-empty keg")); 3506 zone->uz_init = zinit; 3507 ZONE_UNLOCK(zone); 3508 } 3509 3510 /* See uma.h */ 3511 void 3512 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 3513 { 3514 3515 ZONE_LOCK(zone); 3516 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3517 ("uma_zone_set_zfini on non-empty keg")); 3518 zone->uz_fini = zfini; 3519 ZONE_UNLOCK(zone); 3520 } 3521 3522 /* See uma.h */ 3523 /* XXX uk_freef is not actually used with the zone locked */ 3524 void 3525 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 3526 { 3527 uma_keg_t keg; 3528 3529 keg = zone_first_keg(zone); 3530 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type")); 3531 KEG_LOCK(keg); 3532 keg->uk_freef = freef; 3533 KEG_UNLOCK(keg); 3534 } 3535 3536 /* See uma.h */ 3537 /* XXX uk_allocf is not actually used with the zone locked */ 3538 void 3539 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3540 { 3541 uma_keg_t keg; 3542 3543 keg = zone_first_keg(zone); 3544 KEG_LOCK(keg); 3545 keg->uk_allocf = allocf; 3546 KEG_UNLOCK(keg); 3547 } 3548 3549 /* See uma.h */ 3550 void 3551 uma_zone_reserve(uma_zone_t zone, int items) 3552 { 3553 uma_keg_t keg; 3554 3555 keg = zone_first_keg(zone); 3556 if (keg == NULL) 3557 return; 3558 KEG_LOCK(keg); 3559 keg->uk_reserve = items; 3560 KEG_UNLOCK(keg); 3561 3562 return; 3563 } 3564 3565 /* See uma.h */ 3566 int 3567 uma_zone_reserve_kva(uma_zone_t zone, int count) 3568 { 3569 uma_keg_t keg; 3570 vm_offset_t kva; 3571 u_int pages; 3572 3573 keg = zone_first_keg(zone); 3574 if (keg == NULL) 3575 return (0); 3576 pages = count / keg->uk_ipers; 3577 3578 if (pages * keg->uk_ipers < count) 3579 pages++; 3580 pages *= keg->uk_ppera; 3581 3582 #ifdef UMA_MD_SMALL_ALLOC 3583 if (keg->uk_ppera > 1) { 3584 #else 3585 if (1) { 3586 #endif 3587 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE); 3588 if (kva == 0) 3589 return (0); 3590 } else 3591 kva = 0; 3592 KEG_LOCK(keg); 3593 keg->uk_kva = kva; 3594 keg->uk_offset = 0; 3595 keg->uk_maxpages = pages; 3596 #ifdef UMA_MD_SMALL_ALLOC 3597 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3598 #else 3599 keg->uk_allocf = noobj_alloc; 3600 #endif 3601 keg->uk_flags |= UMA_ZONE_NOFREE; 3602 KEG_UNLOCK(keg); 3603 3604 return (1); 3605 } 3606 3607 /* See uma.h */ 3608 void 3609 uma_prealloc(uma_zone_t zone, int items) 3610 { 3611 struct vm_domainset_iter di; 3612 uma_domain_t dom; 3613 uma_slab_t slab; 3614 uma_keg_t keg; 3615 int domain, flags, slabs; 3616 3617 keg = zone_first_keg(zone); 3618 if (keg == NULL) 3619 return; 3620 KEG_LOCK(keg); 3621 slabs = items / keg->uk_ipers; 3622 if (slabs * keg->uk_ipers < items) 3623 slabs++; 3624 flags = M_WAITOK; 3625 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, &flags); 3626 while (slabs-- > 0) { 3627 slab = keg_alloc_slab(keg, zone, domain, flags); 3628 if (slab == NULL) 3629 return; 3630 MPASS(slab->us_keg == keg); 3631 dom = &keg->uk_domain[slab->us_domain]; 3632 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); 3633 if (vm_domainset_iter_policy(&di, &domain) != 0) 3634 break; 3635 } 3636 KEG_UNLOCK(keg); 3637 } 3638 3639 /* See uma.h */ 3640 static void 3641 uma_reclaim_locked(bool kmem_danger) 3642 { 3643 3644 CTR0(KTR_UMA, "UMA: vm asked us to release pages!"); 3645 sx_assert(&uma_drain_lock, SA_XLOCKED); 3646 bucket_enable(); 3647 zone_foreach(zone_drain); 3648 if (vm_page_count_min() || kmem_danger) { 3649 cache_drain_safe(NULL); 3650 zone_foreach(zone_drain); 3651 } 3652 /* 3653 * Some slabs may have been freed but this zone will be visited early 3654 * we visit again so that we can free pages that are empty once other 3655 * zones are drained. We have to do the same for buckets. 3656 */ 3657 zone_drain(slabzone); 3658 bucket_zone_drain(); 3659 } 3660 3661 void 3662 uma_reclaim(void) 3663 { 3664 3665 sx_xlock(&uma_drain_lock); 3666 uma_reclaim_locked(false); 3667 sx_xunlock(&uma_drain_lock); 3668 } 3669 3670 static volatile int uma_reclaim_needed; 3671 3672 void 3673 uma_reclaim_wakeup(void) 3674 { 3675 3676 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0) 3677 wakeup(uma_reclaim); 3678 } 3679 3680 void 3681 uma_reclaim_worker(void *arg __unused) 3682 { 3683 3684 for (;;) { 3685 sx_xlock(&uma_drain_lock); 3686 while (atomic_load_int(&uma_reclaim_needed) == 0) 3687 sx_sleep(uma_reclaim, &uma_drain_lock, PVM, "umarcl", 3688 hz); 3689 sx_xunlock(&uma_drain_lock); 3690 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); 3691 sx_xlock(&uma_drain_lock); 3692 uma_reclaim_locked(true); 3693 atomic_store_int(&uma_reclaim_needed, 0); 3694 sx_xunlock(&uma_drain_lock); 3695 /* Don't fire more than once per-second. */ 3696 pause("umarclslp", hz); 3697 } 3698 } 3699 3700 /* See uma.h */ 3701 int 3702 uma_zone_exhausted(uma_zone_t zone) 3703 { 3704 int full; 3705 3706 ZONE_LOCK(zone); 3707 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3708 ZONE_UNLOCK(zone); 3709 return (full); 3710 } 3711 3712 int 3713 uma_zone_exhausted_nolock(uma_zone_t zone) 3714 { 3715 return (zone->uz_flags & UMA_ZFLAG_FULL); 3716 } 3717 3718 void * 3719 uma_large_malloc_domain(vm_size_t size, int domain, int wait) 3720 { 3721 vm_offset_t addr; 3722 uma_slab_t slab; 3723 3724 if (domain != UMA_ANYDOMAIN) { 3725 /* avoid allocs targeting empty domains */ 3726 if (VM_DOMAIN_EMPTY(domain)) 3727 domain = UMA_ANYDOMAIN; 3728 } 3729 slab = zone_alloc_item(slabzone, NULL, domain, wait); 3730 if (slab == NULL) 3731 return (NULL); 3732 if (domain == UMA_ANYDOMAIN) 3733 addr = kmem_malloc(size, wait); 3734 else 3735 addr = kmem_malloc_domain(domain, size, wait); 3736 if (addr != 0) { 3737 vsetslab(addr, slab); 3738 slab->us_data = (void *)addr; 3739 slab->us_flags = UMA_SLAB_KERNEL | UMA_SLAB_MALLOC; 3740 slab->us_size = size; 3741 slab->us_domain = vm_phys_domain(PHYS_TO_VM_PAGE( 3742 pmap_kextract(addr))); 3743 uma_total_inc(size); 3744 } else { 3745 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3746 } 3747 3748 return ((void *)addr); 3749 } 3750 3751 void * 3752 uma_large_malloc(vm_size_t size, int wait) 3753 { 3754 3755 return uma_large_malloc_domain(size, UMA_ANYDOMAIN, wait); 3756 } 3757 3758 void 3759 uma_large_free(uma_slab_t slab) 3760 { 3761 3762 KASSERT((slab->us_flags & UMA_SLAB_KERNEL) != 0, 3763 ("uma_large_free: Memory not allocated with uma_large_malloc.")); 3764 kmem_free((vm_offset_t)slab->us_data, slab->us_size); 3765 uma_total_dec(slab->us_size); 3766 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3767 } 3768 3769 static void 3770 uma_zero_item(void *item, uma_zone_t zone) 3771 { 3772 3773 bzero(item, zone->uz_size); 3774 } 3775 3776 unsigned long 3777 uma_limit(void) 3778 { 3779 3780 return (uma_kmem_limit); 3781 } 3782 3783 void 3784 uma_set_limit(unsigned long limit) 3785 { 3786 3787 uma_kmem_limit = limit; 3788 } 3789 3790 unsigned long 3791 uma_size(void) 3792 { 3793 3794 return (uma_kmem_total); 3795 } 3796 3797 long 3798 uma_avail(void) 3799 { 3800 3801 return (uma_kmem_limit - uma_kmem_total); 3802 } 3803 3804 void 3805 uma_print_stats(void) 3806 { 3807 zone_foreach(uma_print_zone); 3808 } 3809 3810 static void 3811 slab_print(uma_slab_t slab) 3812 { 3813 printf("slab: keg %p, data %p, freecount %d\n", 3814 slab->us_keg, slab->us_data, slab->us_freecount); 3815 } 3816 3817 static void 3818 cache_print(uma_cache_t cache) 3819 { 3820 printf("alloc: %p(%d), free: %p(%d)\n", 3821 cache->uc_allocbucket, 3822 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3823 cache->uc_freebucket, 3824 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3825 } 3826 3827 static void 3828 uma_print_keg(uma_keg_t keg) 3829 { 3830 uma_domain_t dom; 3831 uma_slab_t slab; 3832 int i; 3833 3834 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3835 "out %d free %d limit %d\n", 3836 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3837 keg->uk_ipers, keg->uk_ppera, 3838 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, 3839 keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3840 for (i = 0; i < vm_ndomains; i++) { 3841 dom = &keg->uk_domain[i]; 3842 printf("Part slabs:\n"); 3843 LIST_FOREACH(slab, &dom->ud_part_slab, us_link) 3844 slab_print(slab); 3845 printf("Free slabs:\n"); 3846 LIST_FOREACH(slab, &dom->ud_free_slab, us_link) 3847 slab_print(slab); 3848 printf("Full slabs:\n"); 3849 LIST_FOREACH(slab, &dom->ud_full_slab, us_link) 3850 slab_print(slab); 3851 } 3852 } 3853 3854 void 3855 uma_print_zone(uma_zone_t zone) 3856 { 3857 uma_cache_t cache; 3858 uma_klink_t kl; 3859 int i; 3860 3861 printf("zone: %s(%p) size %d flags %#x\n", 3862 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3863 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3864 uma_print_keg(kl->kl_keg); 3865 CPU_FOREACH(i) { 3866 cache = &zone->uz_cpu[i]; 3867 printf("CPU %d Cache:\n", i); 3868 cache_print(cache); 3869 } 3870 } 3871 3872 #ifdef DDB 3873 /* 3874 * Generate statistics across both the zone and its per-cpu cache's. Return 3875 * desired statistics if the pointer is non-NULL for that statistic. 3876 * 3877 * Note: does not update the zone statistics, as it can't safely clear the 3878 * per-CPU cache statistic. 3879 * 3880 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3881 * safe from off-CPU; we should modify the caches to track this information 3882 * directly so that we don't have to. 3883 */ 3884 static void 3885 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3886 uint64_t *freesp, uint64_t *sleepsp) 3887 { 3888 uma_cache_t cache; 3889 uint64_t allocs, frees, sleeps; 3890 int cachefree, cpu; 3891 3892 allocs = frees = sleeps = 0; 3893 cachefree = 0; 3894 CPU_FOREACH(cpu) { 3895 cache = &z->uz_cpu[cpu]; 3896 if (cache->uc_allocbucket != NULL) 3897 cachefree += cache->uc_allocbucket->ub_cnt; 3898 if (cache->uc_freebucket != NULL) 3899 cachefree += cache->uc_freebucket->ub_cnt; 3900 allocs += cache->uc_allocs; 3901 frees += cache->uc_frees; 3902 } 3903 allocs += z->uz_allocs; 3904 frees += z->uz_frees; 3905 sleeps += z->uz_sleeps; 3906 if (cachefreep != NULL) 3907 *cachefreep = cachefree; 3908 if (allocsp != NULL) 3909 *allocsp = allocs; 3910 if (freesp != NULL) 3911 *freesp = frees; 3912 if (sleepsp != NULL) 3913 *sleepsp = sleeps; 3914 } 3915 #endif /* DDB */ 3916 3917 static int 3918 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3919 { 3920 uma_keg_t kz; 3921 uma_zone_t z; 3922 int count; 3923 3924 count = 0; 3925 rw_rlock(&uma_rwlock); 3926 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3927 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3928 count++; 3929 } 3930 rw_runlock(&uma_rwlock); 3931 return (sysctl_handle_int(oidp, &count, 0, req)); 3932 } 3933 3934 static int 3935 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3936 { 3937 struct uma_stream_header ush; 3938 struct uma_type_header uth; 3939 struct uma_percpu_stat *ups; 3940 uma_bucket_t bucket; 3941 uma_zone_domain_t zdom; 3942 struct sbuf sbuf; 3943 uma_cache_t cache; 3944 uma_klink_t kl; 3945 uma_keg_t kz; 3946 uma_zone_t z; 3947 uma_keg_t k; 3948 int count, error, i; 3949 3950 error = sysctl_wire_old_buffer(req, 0); 3951 if (error != 0) 3952 return (error); 3953 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3954 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); 3955 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK); 3956 3957 count = 0; 3958 rw_rlock(&uma_rwlock); 3959 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3960 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3961 count++; 3962 } 3963 3964 /* 3965 * Insert stream header. 3966 */ 3967 bzero(&ush, sizeof(ush)); 3968 ush.ush_version = UMA_STREAM_VERSION; 3969 ush.ush_maxcpus = (mp_maxid + 1); 3970 ush.ush_count = count; 3971 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3972 3973 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3974 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3975 bzero(&uth, sizeof(uth)); 3976 ZONE_LOCK(z); 3977 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3978 uth.uth_align = kz->uk_align; 3979 uth.uth_size = kz->uk_size; 3980 uth.uth_rsize = kz->uk_rsize; 3981 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3982 k = kl->kl_keg; 3983 uth.uth_maxpages += k->uk_maxpages; 3984 uth.uth_pages += k->uk_pages; 3985 uth.uth_keg_free += k->uk_free; 3986 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3987 * k->uk_ipers; 3988 } 3989 3990 /* 3991 * A zone is secondary is it is not the first entry 3992 * on the keg's zone list. 3993 */ 3994 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3995 (LIST_FIRST(&kz->uk_zones) != z)) 3996 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3997 3998 for (i = 0; i < vm_ndomains; i++) { 3999 zdom = &z->uz_domain[i]; 4000 LIST_FOREACH(bucket, &zdom->uzd_buckets, 4001 ub_link) 4002 uth.uth_zone_free += bucket->ub_cnt; 4003 } 4004 uth.uth_allocs = z->uz_allocs; 4005 uth.uth_frees = z->uz_frees; 4006 uth.uth_fails = z->uz_fails; 4007 uth.uth_sleeps = z->uz_sleeps; 4008 /* 4009 * While it is not normally safe to access the cache 4010 * bucket pointers while not on the CPU that owns the 4011 * cache, we only allow the pointers to be exchanged 4012 * without the zone lock held, not invalidated, so 4013 * accept the possible race associated with bucket 4014 * exchange during monitoring. 4015 */ 4016 for (i = 0; i < mp_maxid + 1; i++) { 4017 bzero(&ups[i], sizeof(*ups)); 4018 if (kz->uk_flags & UMA_ZFLAG_INTERNAL || 4019 CPU_ABSENT(i)) 4020 continue; 4021 cache = &z->uz_cpu[i]; 4022 if (cache->uc_allocbucket != NULL) 4023 ups[i].ups_cache_free += 4024 cache->uc_allocbucket->ub_cnt; 4025 if (cache->uc_freebucket != NULL) 4026 ups[i].ups_cache_free += 4027 cache->uc_freebucket->ub_cnt; 4028 ups[i].ups_allocs = cache->uc_allocs; 4029 ups[i].ups_frees = cache->uc_frees; 4030 } 4031 ZONE_UNLOCK(z); 4032 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 4033 for (i = 0; i < mp_maxid + 1; i++) 4034 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 4035 } 4036 } 4037 rw_runlock(&uma_rwlock); 4038 error = sbuf_finish(&sbuf); 4039 sbuf_delete(&sbuf); 4040 free(ups, M_TEMP); 4041 return (error); 4042 } 4043 4044 int 4045 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 4046 { 4047 uma_zone_t zone = *(uma_zone_t *)arg1; 4048 int error, max; 4049 4050 max = uma_zone_get_max(zone); 4051 error = sysctl_handle_int(oidp, &max, 0, req); 4052 if (error || !req->newptr) 4053 return (error); 4054 4055 uma_zone_set_max(zone, max); 4056 4057 return (0); 4058 } 4059 4060 int 4061 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 4062 { 4063 uma_zone_t zone = *(uma_zone_t *)arg1; 4064 int cur; 4065 4066 cur = uma_zone_get_cur(zone); 4067 return (sysctl_handle_int(oidp, &cur, 0, req)); 4068 } 4069 4070 #ifdef INVARIANTS 4071 static uma_slab_t 4072 uma_dbg_getslab(uma_zone_t zone, void *item) 4073 { 4074 uma_slab_t slab; 4075 uma_keg_t keg; 4076 uint8_t *mem; 4077 4078 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 4079 if (zone->uz_flags & UMA_ZONE_VTOSLAB) { 4080 slab = vtoslab((vm_offset_t)mem); 4081 } else { 4082 /* 4083 * It is safe to return the slab here even though the 4084 * zone is unlocked because the item's allocation state 4085 * essentially holds a reference. 4086 */ 4087 ZONE_LOCK(zone); 4088 keg = LIST_FIRST(&zone->uz_kegs)->kl_keg; 4089 if (keg->uk_flags & UMA_ZONE_HASH) 4090 slab = hash_sfind(&keg->uk_hash, mem); 4091 else 4092 slab = (uma_slab_t)(mem + keg->uk_pgoff); 4093 ZONE_UNLOCK(zone); 4094 } 4095 4096 return (slab); 4097 } 4098 4099 static bool 4100 uma_dbg_zskip(uma_zone_t zone, void *mem) 4101 { 4102 uma_keg_t keg; 4103 4104 if ((keg = zone_first_keg(zone)) == NULL) 4105 return (true); 4106 4107 return (uma_dbg_kskip(keg, mem)); 4108 } 4109 4110 static bool 4111 uma_dbg_kskip(uma_keg_t keg, void *mem) 4112 { 4113 uintptr_t idx; 4114 4115 if (dbg_divisor == 0) 4116 return (true); 4117 4118 if (dbg_divisor == 1) 4119 return (false); 4120 4121 idx = (uintptr_t)mem >> PAGE_SHIFT; 4122 if (keg->uk_ipers > 1) { 4123 idx *= keg->uk_ipers; 4124 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize; 4125 } 4126 4127 if ((idx / dbg_divisor) * dbg_divisor != idx) { 4128 counter_u64_add(uma_skip_cnt, 1); 4129 return (true); 4130 } 4131 counter_u64_add(uma_dbg_cnt, 1); 4132 4133 return (false); 4134 } 4135 4136 /* 4137 * Set up the slab's freei data such that uma_dbg_free can function. 4138 * 4139 */ 4140 static void 4141 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) 4142 { 4143 uma_keg_t keg; 4144 int freei; 4145 4146 if (slab == NULL) { 4147 slab = uma_dbg_getslab(zone, item); 4148 if (slab == NULL) 4149 panic("uma: item %p did not belong to zone %s\n", 4150 item, zone->uz_name); 4151 } 4152 keg = slab->us_keg; 4153 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 4154 4155 if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) 4156 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n", 4157 item, zone, zone->uz_name, slab, freei); 4158 BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); 4159 4160 return; 4161 } 4162 4163 /* 4164 * Verifies freed addresses. Checks for alignment, valid slab membership 4165 * and duplicate frees. 4166 * 4167 */ 4168 static void 4169 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) 4170 { 4171 uma_keg_t keg; 4172 int freei; 4173 4174 if (slab == NULL) { 4175 slab = uma_dbg_getslab(zone, item); 4176 if (slab == NULL) 4177 panic("uma: Freed item %p did not belong to zone %s\n", 4178 item, zone->uz_name); 4179 } 4180 keg = slab->us_keg; 4181 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 4182 4183 if (freei >= keg->uk_ipers) 4184 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n", 4185 item, zone, zone->uz_name, slab, freei); 4186 4187 if (((freei * keg->uk_rsize) + slab->us_data) != item) 4188 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n", 4189 item, zone, zone->uz_name, slab, freei); 4190 4191 if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) 4192 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n", 4193 item, zone, zone->uz_name, slab, freei); 4194 4195 BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); 4196 } 4197 #endif /* INVARIANTS */ 4198 4199 #ifdef DDB 4200 DB_SHOW_COMMAND(uma, db_show_uma) 4201 { 4202 uma_bucket_t bucket; 4203 uma_keg_t kz; 4204 uma_zone_t z; 4205 uma_zone_domain_t zdom; 4206 uint64_t allocs, frees, sleeps; 4207 int cachefree, i; 4208 4209 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used", 4210 "Free", "Requests", "Sleeps", "Bucket"); 4211 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4212 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 4213 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 4214 allocs = z->uz_allocs; 4215 frees = z->uz_frees; 4216 sleeps = z->uz_sleeps; 4217 cachefree = 0; 4218 } else 4219 uma_zone_sumstat(z, &cachefree, &allocs, 4220 &frees, &sleeps); 4221 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 4222 (LIST_FIRST(&kz->uk_zones) != z))) 4223 cachefree += kz->uk_free; 4224 for (i = 0; i < vm_ndomains; i++) { 4225 zdom = &z->uz_domain[i]; 4226 LIST_FOREACH(bucket, &zdom->uzd_buckets, 4227 ub_link) 4228 cachefree += bucket->ub_cnt; 4229 } 4230 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n", 4231 z->uz_name, (uintmax_t)kz->uk_size, 4232 (intmax_t)(allocs - frees), cachefree, 4233 (uintmax_t)allocs, sleeps, z->uz_count); 4234 if (db_pager_quit) 4235 return; 4236 } 4237 } 4238 } 4239 4240 DB_SHOW_COMMAND(umacache, db_show_umacache) 4241 { 4242 uma_bucket_t bucket; 4243 uma_zone_t z; 4244 uma_zone_domain_t zdom; 4245 uint64_t allocs, frees; 4246 int cachefree, i; 4247 4248 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 4249 "Requests", "Bucket"); 4250 LIST_FOREACH(z, &uma_cachezones, uz_link) { 4251 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL); 4252 for (i = 0; i < vm_ndomains; i++) { 4253 zdom = &z->uz_domain[i]; 4254 LIST_FOREACH(bucket, &zdom->uzd_buckets, ub_link) 4255 cachefree += bucket->ub_cnt; 4256 } 4257 db_printf("%18s %8ju %8jd %8d %12ju %8u\n", 4258 z->uz_name, (uintmax_t)z->uz_size, 4259 (intmax_t)(allocs - frees), cachefree, 4260 (uintmax_t)allocs, z->uz_count); 4261 if (db_pager_quit) 4262 return; 4263 } 4264 } 4265 #endif /* DDB */ 4266