1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org> 5 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org> 6 * Copyright (c) 2004-2006 Robert N. M. Watson 7 * All rights reserved. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice unmodified, this list of conditions, and the following 14 * disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 29 */ 30 31 /* 32 * uma_core.c Implementation of the Universal Memory allocator 33 * 34 * This allocator is intended to replace the multitude of similar object caches 35 * in the standard FreeBSD kernel. The intent is to be flexible as well as 36 * efficient. A primary design goal is to return unused memory to the rest of 37 * the system. This will make the system as a whole more flexible due to the 38 * ability to move memory to subsystems which most need it instead of leaving 39 * pools of reserved memory unused. 40 * 41 * The basic ideas stem from similar slab/zone based allocators whose algorithms 42 * are well known. 43 * 44 */ 45 46 /* 47 * TODO: 48 * - Improve memory usage for large allocations 49 * - Investigate cache size adjustments 50 */ 51 52 #include <sys/cdefs.h> 53 __FBSDID("$FreeBSD$"); 54 55 #include "opt_ddb.h" 56 #include "opt_param.h" 57 #include "opt_vm.h" 58 59 #include <sys/param.h> 60 #include <sys/systm.h> 61 #include <sys/bitset.h> 62 #include <sys/eventhandler.h> 63 #include <sys/kernel.h> 64 #include <sys/types.h> 65 #include <sys/limits.h> 66 #include <sys/queue.h> 67 #include <sys/malloc.h> 68 #include <sys/ktr.h> 69 #include <sys/lock.h> 70 #include <sys/sysctl.h> 71 #include <sys/mutex.h> 72 #include <sys/proc.h> 73 #include <sys/random.h> 74 #include <sys/rwlock.h> 75 #include <sys/sbuf.h> 76 #include <sys/sched.h> 77 #include <sys/smp.h> 78 #include <sys/taskqueue.h> 79 #include <sys/vmmeter.h> 80 81 #include <vm/vm.h> 82 #include <vm/vm_object.h> 83 #include <vm/vm_page.h> 84 #include <vm/vm_pageout.h> 85 #include <vm/vm_param.h> 86 #include <vm/vm_phys.h> 87 #include <vm/vm_pagequeue.h> 88 #include <vm/vm_map.h> 89 #include <vm/vm_kern.h> 90 #include <vm/vm_extern.h> 91 #include <vm/uma.h> 92 #include <vm/uma_int.h> 93 #include <vm/uma_dbg.h> 94 95 #include <ddb/ddb.h> 96 97 #ifdef DEBUG_MEMGUARD 98 #include <vm/memguard.h> 99 #endif 100 101 /* 102 * This is the zone and keg from which all zones are spawned. 103 */ 104 static uma_zone_t kegs; 105 static uma_zone_t zones; 106 107 /* This is the zone from which all offpage uma_slab_ts are allocated. */ 108 static uma_zone_t slabzone; 109 110 /* 111 * The initial hash tables come out of this zone so they can be allocated 112 * prior to malloc coming up. 113 */ 114 static uma_zone_t hashzone; 115 116 /* The boot-time adjusted value for cache line alignment. */ 117 int uma_align_cache = 64 - 1; 118 119 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 120 121 /* 122 * Are we allowed to allocate buckets? 123 */ 124 static int bucketdisable = 1; 125 126 /* Linked list of all kegs in the system */ 127 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 128 129 /* Linked list of all cache-only zones in the system */ 130 static LIST_HEAD(,uma_zone) uma_cachezones = 131 LIST_HEAD_INITIALIZER(uma_cachezones); 132 133 /* This RW lock protects the keg list */ 134 static struct rwlock_padalign __exclusive_cache_line uma_rwlock; 135 136 /* 137 * Pointer and counter to pool of pages, that is preallocated at 138 * startup to bootstrap UMA. 139 */ 140 static char *bootmem; 141 static int boot_pages; 142 143 static struct sx uma_drain_lock; 144 145 /* kmem soft limit. */ 146 static unsigned long uma_kmem_limit = LONG_MAX; 147 static volatile unsigned long uma_kmem_total; 148 149 /* Is the VM done starting up? */ 150 static enum { BOOT_COLD = 0, BOOT_STRAPPED, BOOT_PAGEALLOC, BOOT_BUCKETS, 151 BOOT_RUNNING } booted = BOOT_COLD; 152 153 /* 154 * This is the handle used to schedule events that need to happen 155 * outside of the allocation fast path. 156 */ 157 static struct callout uma_callout; 158 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 159 160 /* 161 * This structure is passed as the zone ctor arg so that I don't have to create 162 * a special allocation function just for zones. 163 */ 164 struct uma_zctor_args { 165 const char *name; 166 size_t size; 167 uma_ctor ctor; 168 uma_dtor dtor; 169 uma_init uminit; 170 uma_fini fini; 171 uma_import import; 172 uma_release release; 173 void *arg; 174 uma_keg_t keg; 175 int align; 176 uint32_t flags; 177 }; 178 179 struct uma_kctor_args { 180 uma_zone_t zone; 181 size_t size; 182 uma_init uminit; 183 uma_fini fini; 184 int align; 185 uint32_t flags; 186 }; 187 188 struct uma_bucket_zone { 189 uma_zone_t ubz_zone; 190 char *ubz_name; 191 int ubz_entries; /* Number of items it can hold. */ 192 int ubz_maxsize; /* Maximum allocation size per-item. */ 193 }; 194 195 /* 196 * Compute the actual number of bucket entries to pack them in power 197 * of two sizes for more efficient space utilization. 198 */ 199 #define BUCKET_SIZE(n) \ 200 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 201 202 #define BUCKET_MAX BUCKET_SIZE(256) 203 204 struct uma_bucket_zone bucket_zones[] = { 205 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 206 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 207 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 208 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 209 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 210 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 211 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 212 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 213 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 214 { NULL, NULL, 0} 215 }; 216 217 /* 218 * Flags and enumerations to be passed to internal functions. 219 */ 220 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; 221 222 #define UMA_ANYDOMAIN -1 /* Special value for domain search. */ 223 224 /* Prototypes.. */ 225 226 int uma_startup_count(int); 227 void uma_startup(void *, int); 228 void uma_startup1(void); 229 void uma_startup2(void); 230 231 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 232 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 233 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 234 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 235 static void page_free(void *, vm_size_t, uint8_t); 236 static void pcpu_page_free(void *, vm_size_t, uint8_t); 237 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int); 238 static void cache_drain(uma_zone_t); 239 static void bucket_drain(uma_zone_t, uma_bucket_t); 240 static void bucket_cache_drain(uma_zone_t zone); 241 static int keg_ctor(void *, int, void *, int); 242 static void keg_dtor(void *, int, void *); 243 static int zone_ctor(void *, int, void *, int); 244 static void zone_dtor(void *, int, void *); 245 static int zero_init(void *, int, int); 246 static void keg_small_init(uma_keg_t keg); 247 static void keg_large_init(uma_keg_t keg); 248 static void zone_foreach(void (*zfunc)(uma_zone_t)); 249 static void zone_timeout(uma_zone_t zone); 250 static int hash_alloc(struct uma_hash *); 251 static int hash_expand(struct uma_hash *, struct uma_hash *); 252 static void hash_free(struct uma_hash *hash); 253 static void uma_timeout(void *); 254 static void uma_startup3(void); 255 static void *zone_alloc_item(uma_zone_t, void *, int, int); 256 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 257 static void bucket_enable(void); 258 static void bucket_init(void); 259 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 260 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 261 static void bucket_zone_drain(void); 262 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int); 263 static uma_slab_t zone_fetch_slab(uma_zone_t, uma_keg_t, int, int); 264 static uma_slab_t zone_fetch_slab_multi(uma_zone_t, uma_keg_t, int, int); 265 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 266 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); 267 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 268 uma_fini fini, int align, uint32_t flags); 269 static int zone_import(uma_zone_t, void **, int, int, int); 270 static void zone_release(uma_zone_t, void **, int); 271 static void uma_zero_item(void *, uma_zone_t); 272 273 void uma_print_zone(uma_zone_t); 274 void uma_print_stats(void); 275 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 276 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 277 278 #ifdef INVARIANTS 279 static bool uma_dbg_kskip(uma_keg_t keg, void *mem); 280 static bool uma_dbg_zskip(uma_zone_t zone, void *mem); 281 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item); 282 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item); 283 284 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0, 285 "Memory allocation debugging"); 286 287 static u_int dbg_divisor = 1; 288 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor, 289 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0, 290 "Debug & thrash every this item in memory allocator"); 291 292 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER; 293 static counter_u64_t uma_skip_cnt = EARLY_COUNTER; 294 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD, 295 &uma_dbg_cnt, "memory items debugged"); 296 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD, 297 &uma_skip_cnt, "memory items skipped, not debugged"); 298 #endif 299 300 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 301 302 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 303 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 304 305 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 306 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 307 308 static int zone_warnings = 1; 309 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, 310 "Warn when UMA zones becomes full"); 311 312 /* Adjust bytes under management by UMA. */ 313 static inline void 314 uma_total_dec(unsigned long size) 315 { 316 317 atomic_subtract_long(&uma_kmem_total, size); 318 } 319 320 static inline void 321 uma_total_inc(unsigned long size) 322 { 323 324 if (atomic_fetchadd_long(&uma_kmem_total, size) > uma_kmem_limit) 325 uma_reclaim_wakeup(); 326 } 327 328 /* 329 * This routine checks to see whether or not it's safe to enable buckets. 330 */ 331 static void 332 bucket_enable(void) 333 { 334 bucketdisable = vm_page_count_min(); 335 } 336 337 /* 338 * Initialize bucket_zones, the array of zones of buckets of various sizes. 339 * 340 * For each zone, calculate the memory required for each bucket, consisting 341 * of the header and an array of pointers. 342 */ 343 static void 344 bucket_init(void) 345 { 346 struct uma_bucket_zone *ubz; 347 int size; 348 349 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 350 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 351 size += sizeof(void *) * ubz->ubz_entries; 352 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 353 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 354 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET | UMA_ZONE_NUMA); 355 } 356 } 357 358 /* 359 * Given a desired number of entries for a bucket, return the zone from which 360 * to allocate the bucket. 361 */ 362 static struct uma_bucket_zone * 363 bucket_zone_lookup(int entries) 364 { 365 struct uma_bucket_zone *ubz; 366 367 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 368 if (ubz->ubz_entries >= entries) 369 return (ubz); 370 ubz--; 371 return (ubz); 372 } 373 374 static int 375 bucket_select(int size) 376 { 377 struct uma_bucket_zone *ubz; 378 379 ubz = &bucket_zones[0]; 380 if (size > ubz->ubz_maxsize) 381 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 382 383 for (; ubz->ubz_entries != 0; ubz++) 384 if (ubz->ubz_maxsize < size) 385 break; 386 ubz--; 387 return (ubz->ubz_entries); 388 } 389 390 static uma_bucket_t 391 bucket_alloc(uma_zone_t zone, void *udata, int flags) 392 { 393 struct uma_bucket_zone *ubz; 394 uma_bucket_t bucket; 395 396 /* 397 * This is to stop us from allocating per cpu buckets while we're 398 * running out of vm.boot_pages. Otherwise, we would exhaust the 399 * boot pages. This also prevents us from allocating buckets in 400 * low memory situations. 401 */ 402 if (bucketdisable) 403 return (NULL); 404 /* 405 * To limit bucket recursion we store the original zone flags 406 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 407 * NOVM flag to persist even through deep recursions. We also 408 * store ZFLAG_BUCKET once we have recursed attempting to allocate 409 * a bucket for a bucket zone so we do not allow infinite bucket 410 * recursion. This cookie will even persist to frees of unused 411 * buckets via the allocation path or bucket allocations in the 412 * free path. 413 */ 414 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 415 udata = (void *)(uintptr_t)zone->uz_flags; 416 else { 417 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 418 return (NULL); 419 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 420 } 421 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 422 flags |= M_NOVM; 423 ubz = bucket_zone_lookup(zone->uz_count); 424 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 425 ubz++; 426 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 427 if (bucket) { 428 #ifdef INVARIANTS 429 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 430 #endif 431 bucket->ub_cnt = 0; 432 bucket->ub_entries = ubz->ubz_entries; 433 } 434 435 return (bucket); 436 } 437 438 static void 439 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 440 { 441 struct uma_bucket_zone *ubz; 442 443 KASSERT(bucket->ub_cnt == 0, 444 ("bucket_free: Freeing a non free bucket.")); 445 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 446 udata = (void *)(uintptr_t)zone->uz_flags; 447 ubz = bucket_zone_lookup(bucket->ub_entries); 448 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 449 } 450 451 static void 452 bucket_zone_drain(void) 453 { 454 struct uma_bucket_zone *ubz; 455 456 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 457 zone_drain(ubz->ubz_zone); 458 } 459 460 static void 461 zone_log_warning(uma_zone_t zone) 462 { 463 static const struct timeval warninterval = { 300, 0 }; 464 465 if (!zone_warnings || zone->uz_warning == NULL) 466 return; 467 468 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 469 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 470 } 471 472 static inline void 473 zone_maxaction(uma_zone_t zone) 474 { 475 476 if (zone->uz_maxaction.ta_func != NULL) 477 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); 478 } 479 480 static void 481 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 482 { 483 uma_klink_t klink; 484 485 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 486 kegfn(klink->kl_keg); 487 } 488 489 /* 490 * Routine called by timeout which is used to fire off some time interval 491 * based calculations. (stats, hash size, etc.) 492 * 493 * Arguments: 494 * arg Unused 495 * 496 * Returns: 497 * Nothing 498 */ 499 static void 500 uma_timeout(void *unused) 501 { 502 bucket_enable(); 503 zone_foreach(zone_timeout); 504 505 /* Reschedule this event */ 506 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 507 } 508 509 /* 510 * Routine to perform timeout driven calculations. This expands the 511 * hashes and does per cpu statistics aggregation. 512 * 513 * Returns nothing. 514 */ 515 static void 516 keg_timeout(uma_keg_t keg) 517 { 518 519 KEG_LOCK(keg); 520 /* 521 * Expand the keg hash table. 522 * 523 * This is done if the number of slabs is larger than the hash size. 524 * What I'm trying to do here is completely reduce collisions. This 525 * may be a little aggressive. Should I allow for two collisions max? 526 */ 527 if (keg->uk_flags & UMA_ZONE_HASH && 528 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 529 struct uma_hash newhash; 530 struct uma_hash oldhash; 531 int ret; 532 533 /* 534 * This is so involved because allocating and freeing 535 * while the keg lock is held will lead to deadlock. 536 * I have to do everything in stages and check for 537 * races. 538 */ 539 newhash = keg->uk_hash; 540 KEG_UNLOCK(keg); 541 ret = hash_alloc(&newhash); 542 KEG_LOCK(keg); 543 if (ret) { 544 if (hash_expand(&keg->uk_hash, &newhash)) { 545 oldhash = keg->uk_hash; 546 keg->uk_hash = newhash; 547 } else 548 oldhash = newhash; 549 550 KEG_UNLOCK(keg); 551 hash_free(&oldhash); 552 return; 553 } 554 } 555 KEG_UNLOCK(keg); 556 } 557 558 static void 559 zone_timeout(uma_zone_t zone) 560 { 561 562 zone_foreach_keg(zone, &keg_timeout); 563 } 564 565 /* 566 * Allocate and zero fill the next sized hash table from the appropriate 567 * backing store. 568 * 569 * Arguments: 570 * hash A new hash structure with the old hash size in uh_hashsize 571 * 572 * Returns: 573 * 1 on success and 0 on failure. 574 */ 575 static int 576 hash_alloc(struct uma_hash *hash) 577 { 578 int oldsize; 579 int alloc; 580 581 oldsize = hash->uh_hashsize; 582 583 /* We're just going to go to a power of two greater */ 584 if (oldsize) { 585 hash->uh_hashsize = oldsize * 2; 586 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 587 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 588 M_UMAHASH, M_NOWAIT); 589 } else { 590 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 591 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 592 UMA_ANYDOMAIN, M_WAITOK); 593 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 594 } 595 if (hash->uh_slab_hash) { 596 bzero(hash->uh_slab_hash, alloc); 597 hash->uh_hashmask = hash->uh_hashsize - 1; 598 return (1); 599 } 600 601 return (0); 602 } 603 604 /* 605 * Expands the hash table for HASH zones. This is done from zone_timeout 606 * to reduce collisions. This must not be done in the regular allocation 607 * path, otherwise, we can recurse on the vm while allocating pages. 608 * 609 * Arguments: 610 * oldhash The hash you want to expand 611 * newhash The hash structure for the new table 612 * 613 * Returns: 614 * Nothing 615 * 616 * Discussion: 617 */ 618 static int 619 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 620 { 621 uma_slab_t slab; 622 int hval; 623 int i; 624 625 if (!newhash->uh_slab_hash) 626 return (0); 627 628 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 629 return (0); 630 631 /* 632 * I need to investigate hash algorithms for resizing without a 633 * full rehash. 634 */ 635 636 for (i = 0; i < oldhash->uh_hashsize; i++) 637 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 638 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 639 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 640 hval = UMA_HASH(newhash, slab->us_data); 641 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 642 slab, us_hlink); 643 } 644 645 return (1); 646 } 647 648 /* 649 * Free the hash bucket to the appropriate backing store. 650 * 651 * Arguments: 652 * slab_hash The hash bucket we're freeing 653 * hashsize The number of entries in that hash bucket 654 * 655 * Returns: 656 * Nothing 657 */ 658 static void 659 hash_free(struct uma_hash *hash) 660 { 661 if (hash->uh_slab_hash == NULL) 662 return; 663 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 664 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 665 else 666 free(hash->uh_slab_hash, M_UMAHASH); 667 } 668 669 /* 670 * Frees all outstanding items in a bucket 671 * 672 * Arguments: 673 * zone The zone to free to, must be unlocked. 674 * bucket The free/alloc bucket with items, cpu queue must be locked. 675 * 676 * Returns: 677 * Nothing 678 */ 679 680 static void 681 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 682 { 683 int i; 684 685 if (bucket == NULL) 686 return; 687 688 if (zone->uz_fini) 689 for (i = 0; i < bucket->ub_cnt; i++) 690 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 691 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 692 bucket->ub_cnt = 0; 693 } 694 695 /* 696 * Drains the per cpu caches for a zone. 697 * 698 * NOTE: This may only be called while the zone is being turn down, and not 699 * during normal operation. This is necessary in order that we do not have 700 * to migrate CPUs to drain the per-CPU caches. 701 * 702 * Arguments: 703 * zone The zone to drain, must be unlocked. 704 * 705 * Returns: 706 * Nothing 707 */ 708 static void 709 cache_drain(uma_zone_t zone) 710 { 711 uma_cache_t cache; 712 int cpu; 713 714 /* 715 * XXX: It is safe to not lock the per-CPU caches, because we're 716 * tearing down the zone anyway. I.e., there will be no further use 717 * of the caches at this point. 718 * 719 * XXX: It would good to be able to assert that the zone is being 720 * torn down to prevent improper use of cache_drain(). 721 * 722 * XXX: We lock the zone before passing into bucket_cache_drain() as 723 * it is used elsewhere. Should the tear-down path be made special 724 * there in some form? 725 */ 726 CPU_FOREACH(cpu) { 727 cache = &zone->uz_cpu[cpu]; 728 bucket_drain(zone, cache->uc_allocbucket); 729 bucket_drain(zone, cache->uc_freebucket); 730 if (cache->uc_allocbucket != NULL) 731 bucket_free(zone, cache->uc_allocbucket, NULL); 732 if (cache->uc_freebucket != NULL) 733 bucket_free(zone, cache->uc_freebucket, NULL); 734 cache->uc_allocbucket = cache->uc_freebucket = NULL; 735 } 736 ZONE_LOCK(zone); 737 bucket_cache_drain(zone); 738 ZONE_UNLOCK(zone); 739 } 740 741 static void 742 cache_shrink(uma_zone_t zone) 743 { 744 745 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 746 return; 747 748 ZONE_LOCK(zone); 749 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; 750 ZONE_UNLOCK(zone); 751 } 752 753 static void 754 cache_drain_safe_cpu(uma_zone_t zone) 755 { 756 uma_cache_t cache; 757 uma_bucket_t b1, b2; 758 int domain; 759 760 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 761 return; 762 763 b1 = b2 = NULL; 764 ZONE_LOCK(zone); 765 critical_enter(); 766 if (zone->uz_flags & UMA_ZONE_NUMA) 767 domain = PCPU_GET(domain); 768 else 769 domain = 0; 770 cache = &zone->uz_cpu[curcpu]; 771 if (cache->uc_allocbucket) { 772 if (cache->uc_allocbucket->ub_cnt != 0) 773 LIST_INSERT_HEAD(&zone->uz_domain[domain].uzd_buckets, 774 cache->uc_allocbucket, ub_link); 775 else 776 b1 = cache->uc_allocbucket; 777 cache->uc_allocbucket = NULL; 778 } 779 if (cache->uc_freebucket) { 780 if (cache->uc_freebucket->ub_cnt != 0) 781 LIST_INSERT_HEAD(&zone->uz_domain[domain].uzd_buckets, 782 cache->uc_freebucket, ub_link); 783 else 784 b2 = cache->uc_freebucket; 785 cache->uc_freebucket = NULL; 786 } 787 critical_exit(); 788 ZONE_UNLOCK(zone); 789 if (b1) 790 bucket_free(zone, b1, NULL); 791 if (b2) 792 bucket_free(zone, b2, NULL); 793 } 794 795 /* 796 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 797 * This is an expensive call because it needs to bind to all CPUs 798 * one by one and enter a critical section on each of them in order 799 * to safely access their cache buckets. 800 * Zone lock must not be held on call this function. 801 */ 802 static void 803 cache_drain_safe(uma_zone_t zone) 804 { 805 int cpu; 806 807 /* 808 * Polite bucket sizes shrinking was not enouth, shrink aggressively. 809 */ 810 if (zone) 811 cache_shrink(zone); 812 else 813 zone_foreach(cache_shrink); 814 815 CPU_FOREACH(cpu) { 816 thread_lock(curthread); 817 sched_bind(curthread, cpu); 818 thread_unlock(curthread); 819 820 if (zone) 821 cache_drain_safe_cpu(zone); 822 else 823 zone_foreach(cache_drain_safe_cpu); 824 } 825 thread_lock(curthread); 826 sched_unbind(curthread); 827 thread_unlock(curthread); 828 } 829 830 /* 831 * Drain the cached buckets from a zone. Expects a locked zone on entry. 832 */ 833 static void 834 bucket_cache_drain(uma_zone_t zone) 835 { 836 uma_zone_domain_t zdom; 837 uma_bucket_t bucket; 838 int i; 839 840 /* 841 * Drain the bucket queues and free the buckets. 842 */ 843 for (i = 0; i < vm_ndomains; i++) { 844 zdom = &zone->uz_domain[i]; 845 while ((bucket = LIST_FIRST(&zdom->uzd_buckets)) != NULL) { 846 LIST_REMOVE(bucket, ub_link); 847 ZONE_UNLOCK(zone); 848 bucket_drain(zone, bucket); 849 bucket_free(zone, bucket, NULL); 850 ZONE_LOCK(zone); 851 } 852 } 853 854 /* 855 * Shrink further bucket sizes. Price of single zone lock collision 856 * is probably lower then price of global cache drain. 857 */ 858 if (zone->uz_count > zone->uz_count_min) 859 zone->uz_count--; 860 } 861 862 static void 863 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 864 { 865 uint8_t *mem; 866 int i; 867 uint8_t flags; 868 869 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes", 870 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera); 871 872 mem = slab->us_data; 873 flags = slab->us_flags; 874 i = start; 875 if (keg->uk_fini != NULL) { 876 for (i--; i > -1; i--) 877 #ifdef INVARIANTS 878 /* 879 * trash_fini implies that dtor was trash_dtor. trash_fini 880 * would check that memory hasn't been modified since free, 881 * which executed trash_dtor. 882 * That's why we need to run uma_dbg_kskip() check here, 883 * albeit we don't make skip check for other init/fini 884 * invocations. 885 */ 886 if (!uma_dbg_kskip(keg, slab->us_data + (keg->uk_rsize * i)) || 887 keg->uk_fini != trash_fini) 888 #endif 889 keg->uk_fini(slab->us_data + (keg->uk_rsize * i), 890 keg->uk_size); 891 } 892 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 893 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 894 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 895 uma_total_dec(PAGE_SIZE * keg->uk_ppera); 896 } 897 898 /* 899 * Frees pages from a keg back to the system. This is done on demand from 900 * the pageout daemon. 901 * 902 * Returns nothing. 903 */ 904 static void 905 keg_drain(uma_keg_t keg) 906 { 907 struct slabhead freeslabs = { 0 }; 908 uma_domain_t dom; 909 uma_slab_t slab, tmp; 910 int i; 911 912 /* 913 * We don't want to take pages from statically allocated kegs at this 914 * time 915 */ 916 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 917 return; 918 919 CTR3(KTR_UMA, "keg_drain %s(%p) free items: %u", 920 keg->uk_name, keg, keg->uk_free); 921 KEG_LOCK(keg); 922 if (keg->uk_free == 0) 923 goto finished; 924 925 for (i = 0; i < vm_ndomains; i++) { 926 dom = &keg->uk_domain[i]; 927 LIST_FOREACH_SAFE(slab, &dom->ud_free_slab, us_link, tmp) { 928 /* We have nowhere to free these to. */ 929 if (slab->us_flags & UMA_SLAB_BOOT) 930 continue; 931 932 LIST_REMOVE(slab, us_link); 933 keg->uk_pages -= keg->uk_ppera; 934 keg->uk_free -= keg->uk_ipers; 935 936 if (keg->uk_flags & UMA_ZONE_HASH) 937 UMA_HASH_REMOVE(&keg->uk_hash, slab, 938 slab->us_data); 939 940 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 941 } 942 } 943 944 finished: 945 KEG_UNLOCK(keg); 946 947 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 948 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 949 keg_free_slab(keg, slab, keg->uk_ipers); 950 } 951 } 952 953 static void 954 zone_drain_wait(uma_zone_t zone, int waitok) 955 { 956 957 /* 958 * Set draining to interlock with zone_dtor() so we can release our 959 * locks as we go. Only dtor() should do a WAITOK call since it 960 * is the only call that knows the structure will still be available 961 * when it wakes up. 962 */ 963 ZONE_LOCK(zone); 964 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 965 if (waitok == M_NOWAIT) 966 goto out; 967 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 968 } 969 zone->uz_flags |= UMA_ZFLAG_DRAINING; 970 bucket_cache_drain(zone); 971 ZONE_UNLOCK(zone); 972 /* 973 * The DRAINING flag protects us from being freed while 974 * we're running. Normally the uma_rwlock would protect us but we 975 * must be able to release and acquire the right lock for each keg. 976 */ 977 zone_foreach_keg(zone, &keg_drain); 978 ZONE_LOCK(zone); 979 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 980 wakeup(zone); 981 out: 982 ZONE_UNLOCK(zone); 983 } 984 985 void 986 zone_drain(uma_zone_t zone) 987 { 988 989 zone_drain_wait(zone, M_NOWAIT); 990 } 991 992 /* 993 * Allocate a new slab for a keg. This does not insert the slab onto a list. 994 * 995 * Arguments: 996 * wait Shall we wait? 997 * 998 * Returns: 999 * The slab that was allocated or NULL if there is no memory and the 1000 * caller specified M_NOWAIT. 1001 */ 1002 static uma_slab_t 1003 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int wait) 1004 { 1005 uma_alloc allocf; 1006 uma_slab_t slab; 1007 unsigned long size; 1008 uint8_t *mem; 1009 uint8_t flags; 1010 int i; 1011 1012 KASSERT(domain >= 0 && domain < vm_ndomains, 1013 ("keg_alloc_slab: domain %d out of range", domain)); 1014 mtx_assert(&keg->uk_lock, MA_OWNED); 1015 slab = NULL; 1016 mem = NULL; 1017 1018 allocf = keg->uk_allocf; 1019 KEG_UNLOCK(keg); 1020 size = keg->uk_ppera * PAGE_SIZE; 1021 1022 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1023 slab = zone_alloc_item(keg->uk_slabzone, NULL, domain, wait); 1024 if (slab == NULL) 1025 goto out; 1026 } 1027 1028 /* 1029 * This reproduces the old vm_zone behavior of zero filling pages the 1030 * first time they are added to a zone. 1031 * 1032 * Malloced items are zeroed in uma_zalloc. 1033 */ 1034 1035 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 1036 wait |= M_ZERO; 1037 else 1038 wait &= ~M_ZERO; 1039 1040 if (keg->uk_flags & UMA_ZONE_NODUMP) 1041 wait |= M_NODUMP; 1042 1043 /* zone is passed for legacy reasons. */ 1044 mem = allocf(zone, size, domain, &flags, wait); 1045 if (mem == NULL) { 1046 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1047 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 1048 slab = NULL; 1049 goto out; 1050 } 1051 uma_total_inc(size); 1052 1053 /* Point the slab into the allocated memory */ 1054 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 1055 slab = (uma_slab_t )(mem + keg->uk_pgoff); 1056 1057 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 1058 for (i = 0; i < keg->uk_ppera; i++) 1059 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 1060 1061 slab->us_keg = keg; 1062 slab->us_data = mem; 1063 slab->us_freecount = keg->uk_ipers; 1064 slab->us_flags = flags; 1065 slab->us_domain = domain; 1066 BIT_FILL(SLAB_SETSIZE, &slab->us_free); 1067 #ifdef INVARIANTS 1068 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); 1069 #endif 1070 1071 if (keg->uk_init != NULL) { 1072 for (i = 0; i < keg->uk_ipers; i++) 1073 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 1074 keg->uk_size, wait) != 0) 1075 break; 1076 if (i != keg->uk_ipers) { 1077 keg_free_slab(keg, slab, i); 1078 slab = NULL; 1079 goto out; 1080 } 1081 } 1082 out: 1083 KEG_LOCK(keg); 1084 1085 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)", 1086 slab, keg->uk_name, keg); 1087 1088 if (slab != NULL) { 1089 if (keg->uk_flags & UMA_ZONE_HASH) 1090 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1091 1092 keg->uk_pages += keg->uk_ppera; 1093 keg->uk_free += keg->uk_ipers; 1094 } 1095 1096 return (slab); 1097 } 1098 1099 /* 1100 * This function is intended to be used early on in place of page_alloc() so 1101 * that we may use the boot time page cache to satisfy allocations before 1102 * the VM is ready. 1103 */ 1104 static void * 1105 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1106 int wait) 1107 { 1108 uma_keg_t keg; 1109 void *mem; 1110 int pages; 1111 1112 keg = zone_first_keg(zone); 1113 1114 /* 1115 * If we are in BOOT_BUCKETS or higher, than switch to real 1116 * allocator. Zones with page sized slabs switch at BOOT_PAGEALLOC. 1117 */ 1118 switch (booted) { 1119 case BOOT_COLD: 1120 case BOOT_STRAPPED: 1121 break; 1122 case BOOT_PAGEALLOC: 1123 if (keg->uk_ppera > 1) 1124 break; 1125 case BOOT_BUCKETS: 1126 case BOOT_RUNNING: 1127 #ifdef UMA_MD_SMALL_ALLOC 1128 keg->uk_allocf = (keg->uk_ppera > 1) ? 1129 page_alloc : uma_small_alloc; 1130 #else 1131 keg->uk_allocf = page_alloc; 1132 #endif 1133 return keg->uk_allocf(zone, bytes, domain, pflag, wait); 1134 } 1135 1136 /* 1137 * Check our small startup cache to see if it has pages remaining. 1138 */ 1139 pages = howmany(bytes, PAGE_SIZE); 1140 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__)); 1141 if (pages > boot_pages) 1142 panic("UMA zone \"%s\": Increase vm.boot_pages", zone->uz_name); 1143 #ifdef DIAGNOSTIC 1144 printf("%s from \"%s\", %d boot pages left\n", __func__, zone->uz_name, 1145 boot_pages); 1146 #endif 1147 mem = bootmem; 1148 boot_pages -= pages; 1149 bootmem += pages * PAGE_SIZE; 1150 *pflag = UMA_SLAB_BOOT; 1151 1152 return (mem); 1153 } 1154 1155 /* 1156 * Allocates a number of pages from the system 1157 * 1158 * Arguments: 1159 * bytes The number of bytes requested 1160 * wait Shall we wait? 1161 * 1162 * Returns: 1163 * A pointer to the alloced memory or possibly 1164 * NULL if M_NOWAIT is set. 1165 */ 1166 static void * 1167 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1168 int wait) 1169 { 1170 void *p; /* Returned page */ 1171 1172 *pflag = UMA_SLAB_KERNEL; 1173 p = (void *) kmem_malloc_domain(domain, bytes, wait); 1174 1175 return (p); 1176 } 1177 1178 static void * 1179 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1180 int wait) 1181 { 1182 struct pglist alloctail; 1183 vm_offset_t addr, zkva; 1184 int cpu, flags; 1185 vm_page_t p, p_next; 1186 #ifdef NUMA 1187 struct pcpu *pc; 1188 #endif 1189 1190 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE); 1191 1192 TAILQ_INIT(&alloctail); 1193 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1194 malloc2vm_flags(wait); 1195 *pflag = UMA_SLAB_KERNEL; 1196 for (cpu = 0; cpu <= mp_maxid; cpu++) { 1197 if (CPU_ABSENT(cpu)) { 1198 p = vm_page_alloc(NULL, 0, flags); 1199 } else { 1200 #ifndef NUMA 1201 p = vm_page_alloc(NULL, 0, flags); 1202 #else 1203 pc = pcpu_find(cpu); 1204 p = vm_page_alloc_domain(NULL, 0, pc->pc_domain, flags); 1205 if (__predict_false(p == NULL)) 1206 p = vm_page_alloc(NULL, 0, flags); 1207 #endif 1208 } 1209 if (__predict_false(p == NULL)) 1210 goto fail; 1211 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1212 } 1213 if ((addr = kva_alloc(bytes)) == 0) 1214 goto fail; 1215 zkva = addr; 1216 TAILQ_FOREACH(p, &alloctail, listq) { 1217 pmap_qenter(zkva, &p, 1); 1218 zkva += PAGE_SIZE; 1219 } 1220 return ((void*)addr); 1221 fail: 1222 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1223 vm_page_unwire(p, PQ_NONE); 1224 vm_page_free(p); 1225 } 1226 return (NULL); 1227 } 1228 1229 /* 1230 * Allocates a number of pages from within an object 1231 * 1232 * Arguments: 1233 * bytes The number of bytes requested 1234 * wait Shall we wait? 1235 * 1236 * Returns: 1237 * A pointer to the alloced memory or possibly 1238 * NULL if M_NOWAIT is set. 1239 */ 1240 static void * 1241 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, 1242 int wait) 1243 { 1244 TAILQ_HEAD(, vm_page) alloctail; 1245 u_long npages; 1246 vm_offset_t retkva, zkva; 1247 vm_page_t p, p_next; 1248 uma_keg_t keg; 1249 1250 TAILQ_INIT(&alloctail); 1251 keg = zone_first_keg(zone); 1252 1253 npages = howmany(bytes, PAGE_SIZE); 1254 while (npages > 0) { 1255 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT | 1256 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1257 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK : 1258 VM_ALLOC_NOWAIT)); 1259 if (p != NULL) { 1260 /* 1261 * Since the page does not belong to an object, its 1262 * listq is unused. 1263 */ 1264 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1265 npages--; 1266 continue; 1267 } 1268 /* 1269 * Page allocation failed, free intermediate pages and 1270 * exit. 1271 */ 1272 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1273 vm_page_unwire(p, PQ_NONE); 1274 vm_page_free(p); 1275 } 1276 return (NULL); 1277 } 1278 *flags = UMA_SLAB_PRIV; 1279 zkva = keg->uk_kva + 1280 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1281 retkva = zkva; 1282 TAILQ_FOREACH(p, &alloctail, listq) { 1283 pmap_qenter(zkva, &p, 1); 1284 zkva += PAGE_SIZE; 1285 } 1286 1287 return ((void *)retkva); 1288 } 1289 1290 /* 1291 * Frees a number of pages to the system 1292 * 1293 * Arguments: 1294 * mem A pointer to the memory to be freed 1295 * size The size of the memory being freed 1296 * flags The original p->us_flags field 1297 * 1298 * Returns: 1299 * Nothing 1300 */ 1301 static void 1302 page_free(void *mem, vm_size_t size, uint8_t flags) 1303 { 1304 1305 if ((flags & UMA_SLAB_KERNEL) == 0) 1306 panic("UMA: page_free used with invalid flags %x", flags); 1307 1308 kmem_free((vm_offset_t)mem, size); 1309 } 1310 1311 /* 1312 * Frees pcpu zone allocations 1313 * 1314 * Arguments: 1315 * mem A pointer to the memory to be freed 1316 * size The size of the memory being freed 1317 * flags The original p->us_flags field 1318 * 1319 * Returns: 1320 * Nothing 1321 */ 1322 static void 1323 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags) 1324 { 1325 vm_offset_t sva, curva; 1326 vm_paddr_t paddr; 1327 vm_page_t m; 1328 1329 MPASS(size == (mp_maxid+1)*PAGE_SIZE); 1330 sva = (vm_offset_t)mem; 1331 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) { 1332 paddr = pmap_kextract(curva); 1333 m = PHYS_TO_VM_PAGE(paddr); 1334 vm_page_unwire(m, PQ_NONE); 1335 vm_page_free(m); 1336 } 1337 pmap_qremove(sva, size >> PAGE_SHIFT); 1338 kva_free(sva, size); 1339 } 1340 1341 1342 /* 1343 * Zero fill initializer 1344 * 1345 * Arguments/Returns follow uma_init specifications 1346 */ 1347 static int 1348 zero_init(void *mem, int size, int flags) 1349 { 1350 bzero(mem, size); 1351 return (0); 1352 } 1353 1354 /* 1355 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1356 * 1357 * Arguments 1358 * keg The zone we should initialize 1359 * 1360 * Returns 1361 * Nothing 1362 */ 1363 static void 1364 keg_small_init(uma_keg_t keg) 1365 { 1366 u_int rsize; 1367 u_int memused; 1368 u_int wastedspace; 1369 u_int shsize; 1370 u_int slabsize; 1371 1372 if (keg->uk_flags & UMA_ZONE_PCPU) { 1373 u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU; 1374 1375 slabsize = UMA_PCPU_ALLOC_SIZE; 1376 keg->uk_ppera = ncpus; 1377 } else { 1378 slabsize = UMA_SLAB_SIZE; 1379 keg->uk_ppera = 1; 1380 } 1381 1382 /* 1383 * Calculate the size of each allocation (rsize) according to 1384 * alignment. If the requested size is smaller than we have 1385 * allocation bits for we round it up. 1386 */ 1387 rsize = keg->uk_size; 1388 if (rsize < slabsize / SLAB_SETSIZE) 1389 rsize = slabsize / SLAB_SETSIZE; 1390 if (rsize & keg->uk_align) 1391 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1392 keg->uk_rsize = rsize; 1393 1394 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1395 keg->uk_rsize < UMA_PCPU_ALLOC_SIZE, 1396 ("%s: size %u too large", __func__, keg->uk_rsize)); 1397 1398 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1399 shsize = 0; 1400 else 1401 shsize = sizeof(struct uma_slab); 1402 1403 if (rsize <= slabsize - shsize) 1404 keg->uk_ipers = (slabsize - shsize) / rsize; 1405 else { 1406 /* Handle special case when we have 1 item per slab, so 1407 * alignment requirement can be relaxed. */ 1408 KASSERT(keg->uk_size <= slabsize - shsize, 1409 ("%s: size %u greater than slab", __func__, keg->uk_size)); 1410 keg->uk_ipers = 1; 1411 } 1412 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1413 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1414 1415 memused = keg->uk_ipers * rsize + shsize; 1416 wastedspace = slabsize - memused; 1417 1418 /* 1419 * We can't do OFFPAGE if we're internal or if we've been 1420 * asked to not go to the VM for buckets. If we do this we 1421 * may end up going to the VM for slabs which we do not 1422 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1423 * of UMA_ZONE_VM, which clearly forbids it. 1424 */ 1425 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1426 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1427 return; 1428 1429 /* 1430 * See if using an OFFPAGE slab will limit our waste. Only do 1431 * this if it permits more items per-slab. 1432 * 1433 * XXX We could try growing slabsize to limit max waste as well. 1434 * Historically this was not done because the VM could not 1435 * efficiently handle contiguous allocations. 1436 */ 1437 if ((wastedspace >= slabsize / UMA_MAX_WASTE) && 1438 (keg->uk_ipers < (slabsize / keg->uk_rsize))) { 1439 keg->uk_ipers = slabsize / keg->uk_rsize; 1440 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1441 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1442 CTR6(KTR_UMA, "UMA decided we need offpage slab headers for " 1443 "keg: %s(%p), calculated wastedspace = %d, " 1444 "maximum wasted space allowed = %d, " 1445 "calculated ipers = %d, " 1446 "new wasted space = %d\n", keg->uk_name, keg, wastedspace, 1447 slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1448 slabsize - keg->uk_ipers * keg->uk_rsize); 1449 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1450 } 1451 1452 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1453 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1454 keg->uk_flags |= UMA_ZONE_HASH; 1455 } 1456 1457 /* 1458 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1459 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1460 * more complicated. 1461 * 1462 * Arguments 1463 * keg The keg we should initialize 1464 * 1465 * Returns 1466 * Nothing 1467 */ 1468 static void 1469 keg_large_init(uma_keg_t keg) 1470 { 1471 u_int shsize; 1472 1473 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1474 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1475 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1476 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1477 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1478 1479 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1480 keg->uk_ipers = 1; 1481 keg->uk_rsize = keg->uk_size; 1482 1483 /* Check whether we have enough space to not do OFFPAGE. */ 1484 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) { 1485 shsize = sizeof(struct uma_slab); 1486 if (shsize & UMA_ALIGN_PTR) 1487 shsize = (shsize & ~UMA_ALIGN_PTR) + 1488 (UMA_ALIGN_PTR + 1); 1489 1490 if (PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < shsize) { 1491 /* 1492 * We can't do OFFPAGE if we're internal, in which case 1493 * we need an extra page per allocation to contain the 1494 * slab header. 1495 */ 1496 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0) 1497 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1498 else 1499 keg->uk_ppera++; 1500 } 1501 } 1502 1503 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1504 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1505 keg->uk_flags |= UMA_ZONE_HASH; 1506 } 1507 1508 static void 1509 keg_cachespread_init(uma_keg_t keg) 1510 { 1511 int alignsize; 1512 int trailer; 1513 int pages; 1514 int rsize; 1515 1516 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1517 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1518 1519 alignsize = keg->uk_align + 1; 1520 rsize = keg->uk_size; 1521 /* 1522 * We want one item to start on every align boundary in a page. To 1523 * do this we will span pages. We will also extend the item by the 1524 * size of align if it is an even multiple of align. Otherwise, it 1525 * would fall on the same boundary every time. 1526 */ 1527 if (rsize & keg->uk_align) 1528 rsize = (rsize & ~keg->uk_align) + alignsize; 1529 if ((rsize & alignsize) == 0) 1530 rsize += alignsize; 1531 trailer = rsize - keg->uk_size; 1532 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1533 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1534 keg->uk_rsize = rsize; 1535 keg->uk_ppera = pages; 1536 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1537 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1538 KASSERT(keg->uk_ipers <= SLAB_SETSIZE, 1539 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1540 keg->uk_ipers)); 1541 } 1542 1543 /* 1544 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1545 * the keg onto the global keg list. 1546 * 1547 * Arguments/Returns follow uma_ctor specifications 1548 * udata Actually uma_kctor_args 1549 */ 1550 static int 1551 keg_ctor(void *mem, int size, void *udata, int flags) 1552 { 1553 struct uma_kctor_args *arg = udata; 1554 uma_keg_t keg = mem; 1555 uma_zone_t zone; 1556 1557 bzero(keg, size); 1558 keg->uk_size = arg->size; 1559 keg->uk_init = arg->uminit; 1560 keg->uk_fini = arg->fini; 1561 keg->uk_align = arg->align; 1562 keg->uk_cursor = 0; 1563 keg->uk_free = 0; 1564 keg->uk_reserve = 0; 1565 keg->uk_pages = 0; 1566 keg->uk_flags = arg->flags; 1567 keg->uk_slabzone = NULL; 1568 1569 /* 1570 * The master zone is passed to us at keg-creation time. 1571 */ 1572 zone = arg->zone; 1573 keg->uk_name = zone->uz_name; 1574 1575 if (arg->flags & UMA_ZONE_VM) 1576 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1577 1578 if (arg->flags & UMA_ZONE_ZINIT) 1579 keg->uk_init = zero_init; 1580 1581 if (arg->flags & UMA_ZONE_MALLOC) 1582 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1583 1584 if (arg->flags & UMA_ZONE_PCPU) 1585 #ifdef SMP 1586 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1587 #else 1588 keg->uk_flags &= ~UMA_ZONE_PCPU; 1589 #endif 1590 1591 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1592 keg_cachespread_init(keg); 1593 } else { 1594 if (keg->uk_size > UMA_SLAB_SPACE) 1595 keg_large_init(keg); 1596 else 1597 keg_small_init(keg); 1598 } 1599 1600 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1601 keg->uk_slabzone = slabzone; 1602 1603 /* 1604 * If we haven't booted yet we need allocations to go through the 1605 * startup cache until the vm is ready. 1606 */ 1607 if (booted < BOOT_PAGEALLOC) 1608 keg->uk_allocf = startup_alloc; 1609 #ifdef UMA_MD_SMALL_ALLOC 1610 else if (keg->uk_ppera == 1) 1611 keg->uk_allocf = uma_small_alloc; 1612 #endif 1613 else if (keg->uk_flags & UMA_ZONE_PCPU) 1614 keg->uk_allocf = pcpu_page_alloc; 1615 else 1616 keg->uk_allocf = page_alloc; 1617 #ifdef UMA_MD_SMALL_ALLOC 1618 if (keg->uk_ppera == 1) 1619 keg->uk_freef = uma_small_free; 1620 else 1621 #endif 1622 if (keg->uk_flags & UMA_ZONE_PCPU) 1623 keg->uk_freef = pcpu_page_free; 1624 else 1625 keg->uk_freef = page_free; 1626 1627 /* 1628 * Initialize keg's lock 1629 */ 1630 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1631 1632 /* 1633 * If we're putting the slab header in the actual page we need to 1634 * figure out where in each page it goes. This calculates a right 1635 * justified offset into the memory on an ALIGN_PTR boundary. 1636 */ 1637 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1638 u_int totsize; 1639 1640 /* Size of the slab struct and free list */ 1641 totsize = sizeof(struct uma_slab); 1642 1643 if (totsize & UMA_ALIGN_PTR) 1644 totsize = (totsize & ~UMA_ALIGN_PTR) + 1645 (UMA_ALIGN_PTR + 1); 1646 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1647 1648 /* 1649 * The only way the following is possible is if with our 1650 * UMA_ALIGN_PTR adjustments we are now bigger than 1651 * UMA_SLAB_SIZE. I haven't checked whether this is 1652 * mathematically possible for all cases, so we make 1653 * sure here anyway. 1654 */ 1655 totsize = keg->uk_pgoff + sizeof(struct uma_slab); 1656 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1657 printf("zone %s ipers %d rsize %d size %d\n", 1658 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1659 keg->uk_size); 1660 panic("UMA slab won't fit."); 1661 } 1662 } 1663 1664 if (keg->uk_flags & UMA_ZONE_HASH) 1665 hash_alloc(&keg->uk_hash); 1666 1667 CTR5(KTR_UMA, "keg_ctor %p zone %s(%p) out %d free %d\n", 1668 keg, zone->uz_name, zone, 1669 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, 1670 keg->uk_free); 1671 1672 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1673 1674 rw_wlock(&uma_rwlock); 1675 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1676 rw_wunlock(&uma_rwlock); 1677 return (0); 1678 } 1679 1680 /* 1681 * Zone header ctor. This initializes all fields, locks, etc. 1682 * 1683 * Arguments/Returns follow uma_ctor specifications 1684 * udata Actually uma_zctor_args 1685 */ 1686 static int 1687 zone_ctor(void *mem, int size, void *udata, int flags) 1688 { 1689 struct uma_zctor_args *arg = udata; 1690 uma_zone_t zone = mem; 1691 uma_zone_t z; 1692 uma_keg_t keg; 1693 1694 bzero(zone, size); 1695 zone->uz_name = arg->name; 1696 zone->uz_ctor = arg->ctor; 1697 zone->uz_dtor = arg->dtor; 1698 zone->uz_slab = zone_fetch_slab; 1699 zone->uz_init = NULL; 1700 zone->uz_fini = NULL; 1701 zone->uz_allocs = 0; 1702 zone->uz_frees = 0; 1703 zone->uz_fails = 0; 1704 zone->uz_sleeps = 0; 1705 zone->uz_count = 0; 1706 zone->uz_count_min = 0; 1707 zone->uz_flags = 0; 1708 zone->uz_warning = NULL; 1709 /* The domain structures follow the cpu structures. */ 1710 zone->uz_domain = (struct uma_zone_domain *)&zone->uz_cpu[mp_ncpus]; 1711 timevalclear(&zone->uz_ratecheck); 1712 keg = arg->keg; 1713 1714 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 1715 1716 /* 1717 * This is a pure cache zone, no kegs. 1718 */ 1719 if (arg->import) { 1720 if (arg->flags & UMA_ZONE_VM) 1721 arg->flags |= UMA_ZFLAG_CACHEONLY; 1722 zone->uz_flags = arg->flags; 1723 zone->uz_size = arg->size; 1724 zone->uz_import = arg->import; 1725 zone->uz_release = arg->release; 1726 zone->uz_arg = arg->arg; 1727 zone->uz_lockptr = &zone->uz_lock; 1728 rw_wlock(&uma_rwlock); 1729 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 1730 rw_wunlock(&uma_rwlock); 1731 goto out; 1732 } 1733 1734 /* 1735 * Use the regular zone/keg/slab allocator. 1736 */ 1737 zone->uz_import = (uma_import)zone_import; 1738 zone->uz_release = (uma_release)zone_release; 1739 zone->uz_arg = zone; 1740 1741 if (arg->flags & UMA_ZONE_SECONDARY) { 1742 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1743 zone->uz_init = arg->uminit; 1744 zone->uz_fini = arg->fini; 1745 zone->uz_lockptr = &keg->uk_lock; 1746 zone->uz_flags |= UMA_ZONE_SECONDARY; 1747 rw_wlock(&uma_rwlock); 1748 ZONE_LOCK(zone); 1749 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1750 if (LIST_NEXT(z, uz_link) == NULL) { 1751 LIST_INSERT_AFTER(z, zone, uz_link); 1752 break; 1753 } 1754 } 1755 ZONE_UNLOCK(zone); 1756 rw_wunlock(&uma_rwlock); 1757 } else if (keg == NULL) { 1758 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1759 arg->align, arg->flags)) == NULL) 1760 return (ENOMEM); 1761 } else { 1762 struct uma_kctor_args karg; 1763 int error; 1764 1765 /* We should only be here from uma_startup() */ 1766 karg.size = arg->size; 1767 karg.uminit = arg->uminit; 1768 karg.fini = arg->fini; 1769 karg.align = arg->align; 1770 karg.flags = arg->flags; 1771 karg.zone = zone; 1772 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1773 flags); 1774 if (error) 1775 return (error); 1776 } 1777 1778 /* 1779 * Link in the first keg. 1780 */ 1781 zone->uz_klink.kl_keg = keg; 1782 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1783 zone->uz_lockptr = &keg->uk_lock; 1784 zone->uz_size = keg->uk_size; 1785 zone->uz_flags |= (keg->uk_flags & 1786 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1787 1788 /* 1789 * Some internal zones don't have room allocated for the per cpu 1790 * caches. If we're internal, bail out here. 1791 */ 1792 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1793 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1794 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1795 return (0); 1796 } 1797 1798 out: 1799 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) != 1800 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET), 1801 ("Invalid zone flag combination")); 1802 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0) 1803 zone->uz_count = BUCKET_MAX; 1804 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0) 1805 zone->uz_count = 0; 1806 else 1807 zone->uz_count = bucket_select(zone->uz_size); 1808 zone->uz_count_min = zone->uz_count; 1809 1810 return (0); 1811 } 1812 1813 /* 1814 * Keg header dtor. This frees all data, destroys locks, frees the hash 1815 * table and removes the keg from the global list. 1816 * 1817 * Arguments/Returns follow uma_dtor specifications 1818 * udata unused 1819 */ 1820 static void 1821 keg_dtor(void *arg, int size, void *udata) 1822 { 1823 uma_keg_t keg; 1824 1825 keg = (uma_keg_t)arg; 1826 KEG_LOCK(keg); 1827 if (keg->uk_free != 0) { 1828 printf("Freed UMA keg (%s) was not empty (%d items). " 1829 " Lost %d pages of memory.\n", 1830 keg->uk_name ? keg->uk_name : "", 1831 keg->uk_free, keg->uk_pages); 1832 } 1833 KEG_UNLOCK(keg); 1834 1835 hash_free(&keg->uk_hash); 1836 1837 KEG_LOCK_FINI(keg); 1838 } 1839 1840 /* 1841 * Zone header dtor. 1842 * 1843 * Arguments/Returns follow uma_dtor specifications 1844 * udata unused 1845 */ 1846 static void 1847 zone_dtor(void *arg, int size, void *udata) 1848 { 1849 uma_klink_t klink; 1850 uma_zone_t zone; 1851 uma_keg_t keg; 1852 1853 zone = (uma_zone_t)arg; 1854 keg = zone_first_keg(zone); 1855 1856 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1857 cache_drain(zone); 1858 1859 rw_wlock(&uma_rwlock); 1860 LIST_REMOVE(zone, uz_link); 1861 rw_wunlock(&uma_rwlock); 1862 /* 1863 * XXX there are some races here where 1864 * the zone can be drained but zone lock 1865 * released and then refilled before we 1866 * remove it... we dont care for now 1867 */ 1868 zone_drain_wait(zone, M_WAITOK); 1869 /* 1870 * Unlink all of our kegs. 1871 */ 1872 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1873 klink->kl_keg = NULL; 1874 LIST_REMOVE(klink, kl_link); 1875 if (klink == &zone->uz_klink) 1876 continue; 1877 free(klink, M_TEMP); 1878 } 1879 /* 1880 * We only destroy kegs from non secondary zones. 1881 */ 1882 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1883 rw_wlock(&uma_rwlock); 1884 LIST_REMOVE(keg, uk_link); 1885 rw_wunlock(&uma_rwlock); 1886 zone_free_item(kegs, keg, NULL, SKIP_NONE); 1887 } 1888 ZONE_LOCK_FINI(zone); 1889 } 1890 1891 /* 1892 * Traverses every zone in the system and calls a callback 1893 * 1894 * Arguments: 1895 * zfunc A pointer to a function which accepts a zone 1896 * as an argument. 1897 * 1898 * Returns: 1899 * Nothing 1900 */ 1901 static void 1902 zone_foreach(void (*zfunc)(uma_zone_t)) 1903 { 1904 uma_keg_t keg; 1905 uma_zone_t zone; 1906 1907 rw_rlock(&uma_rwlock); 1908 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1909 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1910 zfunc(zone); 1911 } 1912 rw_runlock(&uma_rwlock); 1913 } 1914 1915 /* 1916 * Count how many pages do we need to bootstrap. VM supplies 1917 * its need in early zones in the argument, we add up our zones, 1918 * which consist of: UMA Slabs, UMA Hash and 9 Bucket zones. The 1919 * zone of zones and zone of kegs are accounted separately. 1920 */ 1921 #define UMA_BOOT_ZONES 11 1922 /* Zone of zones and zone of kegs have arbitrary alignment. */ 1923 #define UMA_BOOT_ALIGN 32 1924 static int zsize, ksize; 1925 int 1926 uma_startup_count(int vm_zones) 1927 { 1928 int zones, pages; 1929 1930 ksize = sizeof(struct uma_keg) + 1931 (sizeof(struct uma_domain) * vm_ndomains); 1932 zsize = sizeof(struct uma_zone) + 1933 (sizeof(struct uma_cache) * (mp_maxid + 1)) + 1934 (sizeof(struct uma_zone_domain) * vm_ndomains); 1935 1936 /* 1937 * Memory for the zone of kegs and its keg, 1938 * and for zone of zones. 1939 */ 1940 pages = howmany(roundup(zsize, CACHE_LINE_SIZE) * 2 + 1941 roundup(ksize, CACHE_LINE_SIZE), PAGE_SIZE); 1942 1943 #ifdef UMA_MD_SMALL_ALLOC 1944 zones = UMA_BOOT_ZONES; 1945 #else 1946 zones = UMA_BOOT_ZONES + vm_zones; 1947 vm_zones = 0; 1948 #endif 1949 1950 /* Memory for the rest of startup zones, UMA and VM, ... */ 1951 if (zsize > UMA_SLAB_SPACE) 1952 pages += (zones + vm_zones) * 1953 howmany(roundup2(zsize, UMA_BOOT_ALIGN), UMA_SLAB_SIZE); 1954 else if (roundup2(zsize, UMA_BOOT_ALIGN) > UMA_SLAB_SPACE) 1955 pages += zones; 1956 else 1957 pages += howmany(zones, 1958 UMA_SLAB_SPACE / roundup2(zsize, UMA_BOOT_ALIGN)); 1959 1960 /* ... and their kegs. Note that zone of zones allocates a keg! */ 1961 pages += howmany(zones + 1, 1962 UMA_SLAB_SPACE / roundup2(ksize, UMA_BOOT_ALIGN)); 1963 1964 /* 1965 * Most of startup zones are not going to be offpages, that's 1966 * why we use UMA_SLAB_SPACE instead of UMA_SLAB_SIZE in all 1967 * calculations. Some large bucket zones will be offpage, and 1968 * thus will allocate hashes. We take conservative approach 1969 * and assume that all zones may allocate hash. This may give 1970 * us some positive inaccuracy, usually an extra single page. 1971 */ 1972 pages += howmany(zones, UMA_SLAB_SPACE / 1973 (sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT)); 1974 1975 return (pages); 1976 } 1977 1978 void 1979 uma_startup(void *mem, int npages) 1980 { 1981 struct uma_zctor_args args; 1982 uma_keg_t masterkeg; 1983 uintptr_t m; 1984 1985 #ifdef DIAGNOSTIC 1986 printf("Entering %s with %d boot pages configured\n", __func__, npages); 1987 #endif 1988 1989 rw_init(&uma_rwlock, "UMA lock"); 1990 1991 /* Use bootpages memory for the zone of zones and zone of kegs. */ 1992 m = (uintptr_t)mem; 1993 zones = (uma_zone_t)m; 1994 m += roundup(zsize, CACHE_LINE_SIZE); 1995 kegs = (uma_zone_t)m; 1996 m += roundup(zsize, CACHE_LINE_SIZE); 1997 masterkeg = (uma_keg_t)m; 1998 m += roundup(ksize, CACHE_LINE_SIZE); 1999 m = roundup(m, PAGE_SIZE); 2000 npages -= (m - (uintptr_t)mem) / PAGE_SIZE; 2001 mem = (void *)m; 2002 2003 /* "manually" create the initial zone */ 2004 memset(&args, 0, sizeof(args)); 2005 args.name = "UMA Kegs"; 2006 args.size = ksize; 2007 args.ctor = keg_ctor; 2008 args.dtor = keg_dtor; 2009 args.uminit = zero_init; 2010 args.fini = NULL; 2011 args.keg = masterkeg; 2012 args.align = UMA_BOOT_ALIGN - 1; 2013 args.flags = UMA_ZFLAG_INTERNAL; 2014 zone_ctor(kegs, zsize, &args, M_WAITOK); 2015 2016 bootmem = mem; 2017 boot_pages = npages; 2018 2019 args.name = "UMA Zones"; 2020 args.size = zsize; 2021 args.ctor = zone_ctor; 2022 args.dtor = zone_dtor; 2023 args.uminit = zero_init; 2024 args.fini = NULL; 2025 args.keg = NULL; 2026 args.align = UMA_BOOT_ALIGN - 1; 2027 args.flags = UMA_ZFLAG_INTERNAL; 2028 zone_ctor(zones, zsize, &args, M_WAITOK); 2029 2030 /* Now make a zone for slab headers */ 2031 slabzone = uma_zcreate("UMA Slabs", 2032 sizeof(struct uma_slab), 2033 NULL, NULL, NULL, NULL, 2034 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2035 2036 hashzone = uma_zcreate("UMA Hash", 2037 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 2038 NULL, NULL, NULL, NULL, 2039 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2040 2041 bucket_init(); 2042 2043 booted = BOOT_STRAPPED; 2044 } 2045 2046 void 2047 uma_startup1(void) 2048 { 2049 2050 #ifdef DIAGNOSTIC 2051 printf("Entering %s with %d boot pages left\n", __func__, boot_pages); 2052 #endif 2053 booted = BOOT_PAGEALLOC; 2054 } 2055 2056 void 2057 uma_startup2(void) 2058 { 2059 2060 #ifdef DIAGNOSTIC 2061 printf("Entering %s with %d boot pages left\n", __func__, boot_pages); 2062 #endif 2063 booted = BOOT_BUCKETS; 2064 sx_init(&uma_drain_lock, "umadrain"); 2065 bucket_enable(); 2066 } 2067 2068 /* 2069 * Initialize our callout handle 2070 * 2071 */ 2072 static void 2073 uma_startup3(void) 2074 { 2075 2076 #ifdef INVARIANTS 2077 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor); 2078 uma_dbg_cnt = counter_u64_alloc(M_WAITOK); 2079 uma_skip_cnt = counter_u64_alloc(M_WAITOK); 2080 #endif 2081 callout_init(&uma_callout, 1); 2082 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 2083 booted = BOOT_RUNNING; 2084 } 2085 2086 static uma_keg_t 2087 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 2088 int align, uint32_t flags) 2089 { 2090 struct uma_kctor_args args; 2091 2092 args.size = size; 2093 args.uminit = uminit; 2094 args.fini = fini; 2095 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 2096 args.flags = flags; 2097 args.zone = zone; 2098 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK)); 2099 } 2100 2101 /* Public functions */ 2102 /* See uma.h */ 2103 void 2104 uma_set_align(int align) 2105 { 2106 2107 if (align != UMA_ALIGN_CACHE) 2108 uma_align_cache = align; 2109 } 2110 2111 /* See uma.h */ 2112 uma_zone_t 2113 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 2114 uma_init uminit, uma_fini fini, int align, uint32_t flags) 2115 2116 { 2117 struct uma_zctor_args args; 2118 uma_zone_t res; 2119 bool locked; 2120 2121 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"", 2122 align, name)); 2123 2124 /* This stuff is essential for the zone ctor */ 2125 memset(&args, 0, sizeof(args)); 2126 args.name = name; 2127 args.size = size; 2128 args.ctor = ctor; 2129 args.dtor = dtor; 2130 args.uminit = uminit; 2131 args.fini = fini; 2132 #ifdef INVARIANTS 2133 /* 2134 * If a zone is being created with an empty constructor and 2135 * destructor, pass UMA constructor/destructor which checks for 2136 * memory use after free. 2137 */ 2138 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) && 2139 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) { 2140 args.ctor = trash_ctor; 2141 args.dtor = trash_dtor; 2142 args.uminit = trash_init; 2143 args.fini = trash_fini; 2144 } 2145 #endif 2146 args.align = align; 2147 args.flags = flags; 2148 args.keg = NULL; 2149 2150 if (booted < BOOT_BUCKETS) { 2151 locked = false; 2152 } else { 2153 sx_slock(&uma_drain_lock); 2154 locked = true; 2155 } 2156 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2157 if (locked) 2158 sx_sunlock(&uma_drain_lock); 2159 return (res); 2160 } 2161 2162 /* See uma.h */ 2163 uma_zone_t 2164 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 2165 uma_init zinit, uma_fini zfini, uma_zone_t master) 2166 { 2167 struct uma_zctor_args args; 2168 uma_keg_t keg; 2169 uma_zone_t res; 2170 bool locked; 2171 2172 keg = zone_first_keg(master); 2173 memset(&args, 0, sizeof(args)); 2174 args.name = name; 2175 args.size = keg->uk_size; 2176 args.ctor = ctor; 2177 args.dtor = dtor; 2178 args.uminit = zinit; 2179 args.fini = zfini; 2180 args.align = keg->uk_align; 2181 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 2182 args.keg = keg; 2183 2184 if (booted < BOOT_BUCKETS) { 2185 locked = false; 2186 } else { 2187 sx_slock(&uma_drain_lock); 2188 locked = true; 2189 } 2190 /* XXX Attaches only one keg of potentially many. */ 2191 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2192 if (locked) 2193 sx_sunlock(&uma_drain_lock); 2194 return (res); 2195 } 2196 2197 /* See uma.h */ 2198 uma_zone_t 2199 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 2200 uma_init zinit, uma_fini zfini, uma_import zimport, 2201 uma_release zrelease, void *arg, int flags) 2202 { 2203 struct uma_zctor_args args; 2204 2205 memset(&args, 0, sizeof(args)); 2206 args.name = name; 2207 args.size = size; 2208 args.ctor = ctor; 2209 args.dtor = dtor; 2210 args.uminit = zinit; 2211 args.fini = zfini; 2212 args.import = zimport; 2213 args.release = zrelease; 2214 args.arg = arg; 2215 args.align = 0; 2216 args.flags = flags; 2217 2218 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK)); 2219 } 2220 2221 static void 2222 zone_lock_pair(uma_zone_t a, uma_zone_t b) 2223 { 2224 if (a < b) { 2225 ZONE_LOCK(a); 2226 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); 2227 } else { 2228 ZONE_LOCK(b); 2229 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); 2230 } 2231 } 2232 2233 static void 2234 zone_unlock_pair(uma_zone_t a, uma_zone_t b) 2235 { 2236 2237 ZONE_UNLOCK(a); 2238 ZONE_UNLOCK(b); 2239 } 2240 2241 int 2242 uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 2243 { 2244 uma_klink_t klink; 2245 uma_klink_t kl; 2246 int error; 2247 2248 error = 0; 2249 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 2250 2251 zone_lock_pair(zone, master); 2252 /* 2253 * zone must use vtoslab() to resolve objects and must already be 2254 * a secondary. 2255 */ 2256 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 2257 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 2258 error = EINVAL; 2259 goto out; 2260 } 2261 /* 2262 * The new master must also use vtoslab(). 2263 */ 2264 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 2265 error = EINVAL; 2266 goto out; 2267 } 2268 2269 /* 2270 * The underlying object must be the same size. rsize 2271 * may be different. 2272 */ 2273 if (master->uz_size != zone->uz_size) { 2274 error = E2BIG; 2275 goto out; 2276 } 2277 /* 2278 * Put it at the end of the list. 2279 */ 2280 klink->kl_keg = zone_first_keg(master); 2281 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 2282 if (LIST_NEXT(kl, kl_link) == NULL) { 2283 LIST_INSERT_AFTER(kl, klink, kl_link); 2284 break; 2285 } 2286 } 2287 klink = NULL; 2288 zone->uz_flags |= UMA_ZFLAG_MULTI; 2289 zone->uz_slab = zone_fetch_slab_multi; 2290 2291 out: 2292 zone_unlock_pair(zone, master); 2293 if (klink != NULL) 2294 free(klink, M_TEMP); 2295 2296 return (error); 2297 } 2298 2299 2300 /* See uma.h */ 2301 void 2302 uma_zdestroy(uma_zone_t zone) 2303 { 2304 2305 sx_slock(&uma_drain_lock); 2306 zone_free_item(zones, zone, NULL, SKIP_NONE); 2307 sx_sunlock(&uma_drain_lock); 2308 } 2309 2310 void 2311 uma_zwait(uma_zone_t zone) 2312 { 2313 void *item; 2314 2315 item = uma_zalloc_arg(zone, NULL, M_WAITOK); 2316 uma_zfree(zone, item); 2317 } 2318 2319 void * 2320 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags) 2321 { 2322 void *item; 2323 #ifdef SMP 2324 int i; 2325 2326 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2327 #endif 2328 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO); 2329 if (item != NULL && (flags & M_ZERO)) { 2330 #ifdef SMP 2331 for (i = 0; i <= mp_maxid; i++) 2332 bzero(zpcpu_get_cpu(item, i), zone->uz_size); 2333 #else 2334 bzero(item, zone->uz_size); 2335 #endif 2336 } 2337 return (item); 2338 } 2339 2340 /* 2341 * A stub while both regular and pcpu cases are identical. 2342 */ 2343 void 2344 uma_zfree_pcpu_arg(uma_zone_t zone, void *item, void *udata) 2345 { 2346 2347 #ifdef SMP 2348 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2349 #endif 2350 uma_zfree_arg(zone, item, udata); 2351 } 2352 2353 /* See uma.h */ 2354 void * 2355 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2356 { 2357 uma_zone_domain_t zdom; 2358 uma_bucket_t bucket; 2359 uma_cache_t cache; 2360 void *item; 2361 int cpu, domain, lockfail; 2362 #ifdef INVARIANTS 2363 bool skipdbg; 2364 #endif 2365 2366 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2367 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 2368 2369 /* This is the fast path allocation */ 2370 CTR4(KTR_UMA, "uma_zalloc_arg thread %x zone %s(%p) flags %d", 2371 curthread, zone->uz_name, zone, flags); 2372 2373 if (flags & M_WAITOK) { 2374 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2375 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2376 } 2377 KASSERT((flags & M_EXEC) == 0, ("uma_zalloc_arg: called with M_EXEC")); 2378 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2379 ("uma_zalloc_arg: called with spinlock or critical section held")); 2380 if (zone->uz_flags & UMA_ZONE_PCPU) 2381 KASSERT((flags & M_ZERO) == 0, ("allocating from a pcpu zone " 2382 "with M_ZERO passed")); 2383 2384 #ifdef DEBUG_MEMGUARD 2385 if (memguard_cmp_zone(zone)) { 2386 item = memguard_alloc(zone->uz_size, flags); 2387 if (item != NULL) { 2388 if (zone->uz_init != NULL && 2389 zone->uz_init(item, zone->uz_size, flags) != 0) 2390 return (NULL); 2391 if (zone->uz_ctor != NULL && 2392 zone->uz_ctor(item, zone->uz_size, udata, 2393 flags) != 0) { 2394 zone->uz_fini(item, zone->uz_size); 2395 return (NULL); 2396 } 2397 return (item); 2398 } 2399 /* This is unfortunate but should not be fatal. */ 2400 } 2401 #endif 2402 /* 2403 * If possible, allocate from the per-CPU cache. There are two 2404 * requirements for safe access to the per-CPU cache: (1) the thread 2405 * accessing the cache must not be preempted or yield during access, 2406 * and (2) the thread must not migrate CPUs without switching which 2407 * cache it accesses. We rely on a critical section to prevent 2408 * preemption and migration. We release the critical section in 2409 * order to acquire the zone mutex if we are unable to allocate from 2410 * the current cache; when we re-acquire the critical section, we 2411 * must detect and handle migration if it has occurred. 2412 */ 2413 critical_enter(); 2414 cpu = curcpu; 2415 cache = &zone->uz_cpu[cpu]; 2416 2417 zalloc_start: 2418 bucket = cache->uc_allocbucket; 2419 if (bucket != NULL && bucket->ub_cnt > 0) { 2420 bucket->ub_cnt--; 2421 item = bucket->ub_bucket[bucket->ub_cnt]; 2422 #ifdef INVARIANTS 2423 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2424 #endif 2425 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2426 cache->uc_allocs++; 2427 critical_exit(); 2428 #ifdef INVARIANTS 2429 skipdbg = uma_dbg_zskip(zone, item); 2430 #endif 2431 if (zone->uz_ctor != NULL && 2432 #ifdef INVARIANTS 2433 (!skipdbg || zone->uz_ctor != trash_ctor || 2434 zone->uz_dtor != trash_dtor) && 2435 #endif 2436 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2437 atomic_add_long(&zone->uz_fails, 1); 2438 zone_free_item(zone, item, udata, SKIP_DTOR); 2439 return (NULL); 2440 } 2441 #ifdef INVARIANTS 2442 if (!skipdbg) 2443 uma_dbg_alloc(zone, NULL, item); 2444 #endif 2445 if (flags & M_ZERO) 2446 uma_zero_item(item, zone); 2447 return (item); 2448 } 2449 2450 /* 2451 * We have run out of items in our alloc bucket. 2452 * See if we can switch with our free bucket. 2453 */ 2454 bucket = cache->uc_freebucket; 2455 if (bucket != NULL && bucket->ub_cnt > 0) { 2456 CTR2(KTR_UMA, 2457 "uma_zalloc: zone %s(%p) swapping empty with alloc", 2458 zone->uz_name, zone); 2459 cache->uc_freebucket = cache->uc_allocbucket; 2460 cache->uc_allocbucket = bucket; 2461 goto zalloc_start; 2462 } 2463 2464 /* 2465 * Discard any empty allocation bucket while we hold no locks. 2466 */ 2467 bucket = cache->uc_allocbucket; 2468 cache->uc_allocbucket = NULL; 2469 critical_exit(); 2470 if (bucket != NULL) 2471 bucket_free(zone, bucket, udata); 2472 2473 if (zone->uz_flags & UMA_ZONE_NUMA) { 2474 domain = PCPU_GET(domain); 2475 if (VM_DOMAIN_EMPTY(domain)) 2476 domain = UMA_ANYDOMAIN; 2477 } else 2478 domain = UMA_ANYDOMAIN; 2479 2480 /* Short-circuit for zones without buckets and low memory. */ 2481 if (zone->uz_count == 0 || bucketdisable) 2482 goto zalloc_item; 2483 2484 /* 2485 * Attempt to retrieve the item from the per-CPU cache has failed, so 2486 * we must go back to the zone. This requires the zone lock, so we 2487 * must drop the critical section, then re-acquire it when we go back 2488 * to the cache. Since the critical section is released, we may be 2489 * preempted or migrate. As such, make sure not to maintain any 2490 * thread-local state specific to the cache from prior to releasing 2491 * the critical section. 2492 */ 2493 lockfail = 0; 2494 if (ZONE_TRYLOCK(zone) == 0) { 2495 /* Record contention to size the buckets. */ 2496 ZONE_LOCK(zone); 2497 lockfail = 1; 2498 } 2499 critical_enter(); 2500 cpu = curcpu; 2501 cache = &zone->uz_cpu[cpu]; 2502 2503 /* See if we lost the race to fill the cache. */ 2504 if (cache->uc_allocbucket != NULL) { 2505 ZONE_UNLOCK(zone); 2506 goto zalloc_start; 2507 } 2508 2509 /* 2510 * Check the zone's cache of buckets. 2511 */ 2512 if (domain == UMA_ANYDOMAIN) 2513 zdom = &zone->uz_domain[0]; 2514 else 2515 zdom = &zone->uz_domain[domain]; 2516 if ((bucket = LIST_FIRST(&zdom->uzd_buckets)) != NULL) { 2517 KASSERT(bucket->ub_cnt != 0, 2518 ("uma_zalloc_arg: Returning an empty bucket.")); 2519 2520 LIST_REMOVE(bucket, ub_link); 2521 cache->uc_allocbucket = bucket; 2522 ZONE_UNLOCK(zone); 2523 goto zalloc_start; 2524 } 2525 /* We are no longer associated with this CPU. */ 2526 critical_exit(); 2527 2528 /* 2529 * We bump the uz count when the cache size is insufficient to 2530 * handle the working set. 2531 */ 2532 if (lockfail && zone->uz_count < BUCKET_MAX) 2533 zone->uz_count++; 2534 ZONE_UNLOCK(zone); 2535 2536 /* 2537 * Now lets just fill a bucket and put it on the free list. If that 2538 * works we'll restart the allocation from the beginning and it 2539 * will use the just filled bucket. 2540 */ 2541 bucket = zone_alloc_bucket(zone, udata, domain, flags); 2542 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p", 2543 zone->uz_name, zone, bucket); 2544 if (bucket != NULL) { 2545 ZONE_LOCK(zone); 2546 critical_enter(); 2547 cpu = curcpu; 2548 cache = &zone->uz_cpu[cpu]; 2549 /* 2550 * See if we lost the race or were migrated. Cache the 2551 * initialized bucket to make this less likely or claim 2552 * the memory directly. 2553 */ 2554 if (cache->uc_allocbucket != NULL || 2555 (zone->uz_flags & UMA_ZONE_NUMA && 2556 domain != PCPU_GET(domain))) 2557 LIST_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link); 2558 else 2559 cache->uc_allocbucket = bucket; 2560 ZONE_UNLOCK(zone); 2561 goto zalloc_start; 2562 } 2563 2564 /* 2565 * We may not be able to get a bucket so return an actual item. 2566 */ 2567 zalloc_item: 2568 item = zone_alloc_item(zone, udata, domain, flags); 2569 2570 return (item); 2571 } 2572 2573 void * 2574 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags) 2575 { 2576 2577 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2578 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 2579 2580 /* This is the fast path allocation */ 2581 CTR5(KTR_UMA, 2582 "uma_zalloc_domain thread %x zone %s(%p) domain %d flags %d", 2583 curthread, zone->uz_name, zone, domain, flags); 2584 2585 if (flags & M_WAITOK) { 2586 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2587 "uma_zalloc_domain: zone \"%s\"", zone->uz_name); 2588 } 2589 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2590 ("uma_zalloc_domain: called with spinlock or critical section held")); 2591 2592 return (zone_alloc_item(zone, udata, domain, flags)); 2593 } 2594 2595 /* 2596 * Find a slab with some space. Prefer slabs that are partially used over those 2597 * that are totally full. This helps to reduce fragmentation. 2598 * 2599 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check 2600 * only 'domain'. 2601 */ 2602 static uma_slab_t 2603 keg_first_slab(uma_keg_t keg, int domain, int rr) 2604 { 2605 uma_domain_t dom; 2606 uma_slab_t slab; 2607 int start; 2608 2609 KASSERT(domain >= 0 && domain < vm_ndomains, 2610 ("keg_first_slab: domain %d out of range", domain)); 2611 2612 slab = NULL; 2613 start = domain; 2614 do { 2615 dom = &keg->uk_domain[domain]; 2616 if (!LIST_EMPTY(&dom->ud_part_slab)) 2617 return (LIST_FIRST(&dom->ud_part_slab)); 2618 if (!LIST_EMPTY(&dom->ud_free_slab)) { 2619 slab = LIST_FIRST(&dom->ud_free_slab); 2620 LIST_REMOVE(slab, us_link); 2621 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 2622 return (slab); 2623 } 2624 if (rr) 2625 domain = (domain + 1) % vm_ndomains; 2626 } while (domain != start); 2627 2628 return (NULL); 2629 } 2630 2631 static uma_slab_t 2632 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, int flags) 2633 { 2634 uma_domain_t dom; 2635 uma_slab_t slab; 2636 int allocflags, domain, reserve, rr, start; 2637 2638 mtx_assert(&keg->uk_lock, MA_OWNED); 2639 slab = NULL; 2640 reserve = 0; 2641 allocflags = flags; 2642 if ((flags & M_USE_RESERVE) == 0) 2643 reserve = keg->uk_reserve; 2644 2645 /* 2646 * Round-robin for non first-touch zones when there is more than one 2647 * domain. 2648 */ 2649 if (vm_ndomains == 1) 2650 rdomain = 0; 2651 rr = rdomain == UMA_ANYDOMAIN; 2652 if (rr) { 2653 start = keg->uk_cursor; 2654 do { 2655 keg->uk_cursor = (keg->uk_cursor + 1) % vm_ndomains; 2656 domain = keg->uk_cursor; 2657 } while (VM_DOMAIN_EMPTY(domain) && domain != start); 2658 domain = start = keg->uk_cursor; 2659 /* Only block on the second pass. */ 2660 if ((flags & (M_WAITOK | M_NOVM)) == M_WAITOK) 2661 allocflags = (allocflags & ~M_WAITOK) | M_NOWAIT; 2662 } else 2663 domain = start = rdomain; 2664 2665 again: 2666 do { 2667 if (keg->uk_free > reserve && 2668 (slab = keg_first_slab(keg, domain, rr)) != NULL) { 2669 MPASS(slab->us_keg == keg); 2670 return (slab); 2671 } 2672 2673 /* 2674 * M_NOVM means don't ask at all! 2675 */ 2676 if (flags & M_NOVM) 2677 break; 2678 2679 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2680 keg->uk_flags |= UMA_ZFLAG_FULL; 2681 /* 2682 * If this is not a multi-zone, set the FULL bit. 2683 * Otherwise slab_multi() takes care of it. 2684 */ 2685 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2686 zone->uz_flags |= UMA_ZFLAG_FULL; 2687 zone_log_warning(zone); 2688 zone_maxaction(zone); 2689 } 2690 if (flags & M_NOWAIT) 2691 return (NULL); 2692 zone->uz_sleeps++; 2693 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2694 continue; 2695 } 2696 slab = keg_alloc_slab(keg, zone, domain, allocflags); 2697 /* 2698 * If we got a slab here it's safe to mark it partially used 2699 * and return. We assume that the caller is going to remove 2700 * at least one item. 2701 */ 2702 if (slab) { 2703 MPASS(slab->us_keg == keg); 2704 dom = &keg->uk_domain[slab->us_domain]; 2705 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 2706 return (slab); 2707 } 2708 if (rr) { 2709 do { 2710 domain = (domain + 1) % vm_ndomains; 2711 } while (VM_DOMAIN_EMPTY(domain) && domain != start); 2712 } 2713 } while (domain != start); 2714 2715 /* Retry domain scan with blocking. */ 2716 if (allocflags != flags) { 2717 allocflags = flags; 2718 goto again; 2719 } 2720 2721 /* 2722 * We might not have been able to get a slab but another cpu 2723 * could have while we were unlocked. Check again before we 2724 * fail. 2725 */ 2726 if (keg->uk_free > reserve && 2727 (slab = keg_first_slab(keg, domain, rr)) != NULL) { 2728 MPASS(slab->us_keg == keg); 2729 return (slab); 2730 } 2731 return (NULL); 2732 } 2733 2734 static uma_slab_t 2735 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int domain, int flags) 2736 { 2737 uma_slab_t slab; 2738 2739 if (keg == NULL) { 2740 keg = zone_first_keg(zone); 2741 KEG_LOCK(keg); 2742 } 2743 2744 for (;;) { 2745 slab = keg_fetch_slab(keg, zone, domain, flags); 2746 if (slab) 2747 return (slab); 2748 if (flags & (M_NOWAIT | M_NOVM)) 2749 break; 2750 } 2751 KEG_UNLOCK(keg); 2752 return (NULL); 2753 } 2754 2755 /* 2756 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2757 * with the keg locked. On NULL no lock is held. 2758 * 2759 * The last pointer is used to seed the search. It is not required. 2760 */ 2761 static uma_slab_t 2762 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int domain, int rflags) 2763 { 2764 uma_klink_t klink; 2765 uma_slab_t slab; 2766 uma_keg_t keg; 2767 int flags; 2768 int empty; 2769 int full; 2770 2771 /* 2772 * Don't wait on the first pass. This will skip limit tests 2773 * as well. We don't want to block if we can find a provider 2774 * without blocking. 2775 */ 2776 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2777 /* 2778 * Use the last slab allocated as a hint for where to start 2779 * the search. 2780 */ 2781 if (last != NULL) { 2782 slab = keg_fetch_slab(last, zone, domain, flags); 2783 if (slab) 2784 return (slab); 2785 KEG_UNLOCK(last); 2786 } 2787 /* 2788 * Loop until we have a slab incase of transient failures 2789 * while M_WAITOK is specified. I'm not sure this is 100% 2790 * required but we've done it for so long now. 2791 */ 2792 for (;;) { 2793 empty = 0; 2794 full = 0; 2795 /* 2796 * Search the available kegs for slabs. Be careful to hold the 2797 * correct lock while calling into the keg layer. 2798 */ 2799 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2800 keg = klink->kl_keg; 2801 KEG_LOCK(keg); 2802 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2803 slab = keg_fetch_slab(keg, zone, domain, flags); 2804 if (slab) 2805 return (slab); 2806 } 2807 if (keg->uk_flags & UMA_ZFLAG_FULL) 2808 full++; 2809 else 2810 empty++; 2811 KEG_UNLOCK(keg); 2812 } 2813 if (rflags & (M_NOWAIT | M_NOVM)) 2814 break; 2815 flags = rflags; 2816 /* 2817 * All kegs are full. XXX We can't atomically check all kegs 2818 * and sleep so just sleep for a short period and retry. 2819 */ 2820 if (full && !empty) { 2821 ZONE_LOCK(zone); 2822 zone->uz_flags |= UMA_ZFLAG_FULL; 2823 zone->uz_sleeps++; 2824 zone_log_warning(zone); 2825 zone_maxaction(zone); 2826 msleep(zone, zone->uz_lockptr, PVM, 2827 "zonelimit", hz/100); 2828 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2829 ZONE_UNLOCK(zone); 2830 continue; 2831 } 2832 } 2833 return (NULL); 2834 } 2835 2836 static void * 2837 slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 2838 { 2839 uma_domain_t dom; 2840 void *item; 2841 uint8_t freei; 2842 2843 MPASS(keg == slab->us_keg); 2844 mtx_assert(&keg->uk_lock, MA_OWNED); 2845 2846 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; 2847 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); 2848 item = slab->us_data + (keg->uk_rsize * freei); 2849 slab->us_freecount--; 2850 keg->uk_free--; 2851 2852 /* Move this slab to the full list */ 2853 if (slab->us_freecount == 0) { 2854 LIST_REMOVE(slab, us_link); 2855 dom = &keg->uk_domain[slab->us_domain]; 2856 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link); 2857 } 2858 2859 return (item); 2860 } 2861 2862 static int 2863 zone_import(uma_zone_t zone, void **bucket, int max, int domain, int flags) 2864 { 2865 uma_slab_t slab; 2866 uma_keg_t keg; 2867 #ifdef NUMA 2868 int stripe; 2869 #endif 2870 int i; 2871 2872 slab = NULL; 2873 keg = NULL; 2874 /* Try to keep the buckets totally full */ 2875 for (i = 0; i < max; ) { 2876 if ((slab = zone->uz_slab(zone, keg, domain, flags)) == NULL) 2877 break; 2878 keg = slab->us_keg; 2879 #ifdef NUMA 2880 stripe = howmany(max, vm_ndomains); 2881 #endif 2882 while (slab->us_freecount && i < max) { 2883 bucket[i++] = slab_alloc_item(keg, slab); 2884 if (keg->uk_free <= keg->uk_reserve) 2885 break; 2886 #ifdef NUMA 2887 /* 2888 * If the zone is striped we pick a new slab for every 2889 * N allocations. Eliminating this conditional will 2890 * instead pick a new domain for each bucket rather 2891 * than stripe within each bucket. The current option 2892 * produces more fragmentation and requires more cpu 2893 * time but yields better distribution. 2894 */ 2895 if ((zone->uz_flags & UMA_ZONE_NUMA) == 0 && 2896 vm_ndomains > 1 && --stripe == 0) 2897 break; 2898 #endif 2899 } 2900 /* Don't block if we allocated any successfully. */ 2901 flags &= ~M_WAITOK; 2902 flags |= M_NOWAIT; 2903 } 2904 if (slab != NULL) 2905 KEG_UNLOCK(keg); 2906 2907 return i; 2908 } 2909 2910 static uma_bucket_t 2911 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags) 2912 { 2913 uma_bucket_t bucket; 2914 int max; 2915 2916 CTR1(KTR_UMA, "zone_alloc:_bucket domain %d)", domain); 2917 2918 /* Don't wait for buckets, preserve caller's NOVM setting. */ 2919 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 2920 if (bucket == NULL) 2921 return (NULL); 2922 2923 max = MIN(bucket->ub_entries, zone->uz_count); 2924 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 2925 max, domain, flags); 2926 2927 /* 2928 * Initialize the memory if necessary. 2929 */ 2930 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 2931 int i; 2932 2933 for (i = 0; i < bucket->ub_cnt; i++) 2934 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2935 flags) != 0) 2936 break; 2937 /* 2938 * If we couldn't initialize the whole bucket, put the 2939 * rest back onto the freelist. 2940 */ 2941 if (i != bucket->ub_cnt) { 2942 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 2943 bucket->ub_cnt - i); 2944 #ifdef INVARIANTS 2945 bzero(&bucket->ub_bucket[i], 2946 sizeof(void *) * (bucket->ub_cnt - i)); 2947 #endif 2948 bucket->ub_cnt = i; 2949 } 2950 } 2951 2952 if (bucket->ub_cnt == 0) { 2953 bucket_free(zone, bucket, udata); 2954 atomic_add_long(&zone->uz_fails, 1); 2955 return (NULL); 2956 } 2957 2958 return (bucket); 2959 } 2960 2961 /* 2962 * Allocates a single item from a zone. 2963 * 2964 * Arguments 2965 * zone The zone to alloc for. 2966 * udata The data to be passed to the constructor. 2967 * domain The domain to allocate from or UMA_ANYDOMAIN. 2968 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2969 * 2970 * Returns 2971 * NULL if there is no memory and M_NOWAIT is set 2972 * An item if successful 2973 */ 2974 2975 static void * 2976 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags) 2977 { 2978 void *item; 2979 #ifdef INVARIANTS 2980 bool skipdbg; 2981 #endif 2982 2983 item = NULL; 2984 2985 if (domain != UMA_ANYDOMAIN) { 2986 /* avoid allocs targeting empty domains */ 2987 if (VM_DOMAIN_EMPTY(domain)) 2988 domain = UMA_ANYDOMAIN; 2989 } 2990 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1) 2991 goto fail; 2992 atomic_add_long(&zone->uz_allocs, 1); 2993 2994 #ifdef INVARIANTS 2995 skipdbg = uma_dbg_zskip(zone, item); 2996 #endif 2997 /* 2998 * We have to call both the zone's init (not the keg's init) 2999 * and the zone's ctor. This is because the item is going from 3000 * a keg slab directly to the user, and the user is expecting it 3001 * to be both zone-init'd as well as zone-ctor'd. 3002 */ 3003 if (zone->uz_init != NULL) { 3004 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 3005 zone_free_item(zone, item, udata, SKIP_FINI); 3006 goto fail; 3007 } 3008 } 3009 if (zone->uz_ctor != NULL && 3010 #ifdef INVARIANTS 3011 (!skipdbg || zone->uz_ctor != trash_ctor || 3012 zone->uz_dtor != trash_dtor) && 3013 #endif 3014 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 3015 zone_free_item(zone, item, udata, SKIP_DTOR); 3016 goto fail; 3017 } 3018 #ifdef INVARIANTS 3019 if (!skipdbg) 3020 uma_dbg_alloc(zone, NULL, item); 3021 #endif 3022 if (flags & M_ZERO) 3023 uma_zero_item(item, zone); 3024 3025 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item, 3026 zone->uz_name, zone); 3027 3028 return (item); 3029 3030 fail: 3031 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)", 3032 zone->uz_name, zone); 3033 atomic_add_long(&zone->uz_fails, 1); 3034 return (NULL); 3035 } 3036 3037 /* See uma.h */ 3038 void 3039 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 3040 { 3041 uma_cache_t cache; 3042 uma_bucket_t bucket; 3043 uma_zone_domain_t zdom; 3044 int cpu, domain, lockfail; 3045 #ifdef INVARIANTS 3046 bool skipdbg; 3047 #endif 3048 3049 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3050 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3051 3052 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 3053 zone->uz_name); 3054 3055 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3056 ("uma_zfree_arg: called with spinlock or critical section held")); 3057 3058 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 3059 if (item == NULL) 3060 return; 3061 #ifdef DEBUG_MEMGUARD 3062 if (is_memguard_addr(item)) { 3063 if (zone->uz_dtor != NULL) 3064 zone->uz_dtor(item, zone->uz_size, udata); 3065 if (zone->uz_fini != NULL) 3066 zone->uz_fini(item, zone->uz_size); 3067 memguard_free(item); 3068 return; 3069 } 3070 #endif 3071 #ifdef INVARIANTS 3072 skipdbg = uma_dbg_zskip(zone, item); 3073 if (skipdbg == false) { 3074 if (zone->uz_flags & UMA_ZONE_MALLOC) 3075 uma_dbg_free(zone, udata, item); 3076 else 3077 uma_dbg_free(zone, NULL, item); 3078 } 3079 if (zone->uz_dtor != NULL && (!skipdbg || 3080 zone->uz_dtor != trash_dtor || zone->uz_ctor != trash_ctor)) 3081 #else 3082 if (zone->uz_dtor != NULL) 3083 #endif 3084 zone->uz_dtor(item, zone->uz_size, udata); 3085 3086 /* 3087 * The race here is acceptable. If we miss it we'll just have to wait 3088 * a little longer for the limits to be reset. 3089 */ 3090 if (zone->uz_flags & UMA_ZFLAG_FULL) 3091 goto zfree_item; 3092 3093 /* 3094 * If possible, free to the per-CPU cache. There are two 3095 * requirements for safe access to the per-CPU cache: (1) the thread 3096 * accessing the cache must not be preempted or yield during access, 3097 * and (2) the thread must not migrate CPUs without switching which 3098 * cache it accesses. We rely on a critical section to prevent 3099 * preemption and migration. We release the critical section in 3100 * order to acquire the zone mutex if we are unable to free to the 3101 * current cache; when we re-acquire the critical section, we must 3102 * detect and handle migration if it has occurred. 3103 */ 3104 zfree_restart: 3105 critical_enter(); 3106 cpu = curcpu; 3107 cache = &zone->uz_cpu[cpu]; 3108 3109 zfree_start: 3110 /* 3111 * Try to free into the allocbucket first to give LIFO ordering 3112 * for cache-hot datastructures. Spill over into the freebucket 3113 * if necessary. Alloc will swap them if one runs dry. 3114 */ 3115 bucket = cache->uc_allocbucket; 3116 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 3117 bucket = cache->uc_freebucket; 3118 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 3119 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 3120 ("uma_zfree: Freeing to non free bucket index.")); 3121 bucket->ub_bucket[bucket->ub_cnt] = item; 3122 bucket->ub_cnt++; 3123 cache->uc_frees++; 3124 critical_exit(); 3125 return; 3126 } 3127 3128 /* 3129 * We must go back the zone, which requires acquiring the zone lock, 3130 * which in turn means we must release and re-acquire the critical 3131 * section. Since the critical section is released, we may be 3132 * preempted or migrate. As such, make sure not to maintain any 3133 * thread-local state specific to the cache from prior to releasing 3134 * the critical section. 3135 */ 3136 critical_exit(); 3137 if (zone->uz_count == 0 || bucketdisable) 3138 goto zfree_item; 3139 3140 lockfail = 0; 3141 if (ZONE_TRYLOCK(zone) == 0) { 3142 /* Record contention to size the buckets. */ 3143 ZONE_LOCK(zone); 3144 lockfail = 1; 3145 } 3146 critical_enter(); 3147 cpu = curcpu; 3148 cache = &zone->uz_cpu[cpu]; 3149 3150 bucket = cache->uc_freebucket; 3151 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 3152 ZONE_UNLOCK(zone); 3153 goto zfree_start; 3154 } 3155 cache->uc_freebucket = NULL; 3156 /* We are no longer associated with this CPU. */ 3157 critical_exit(); 3158 3159 if ((zone->uz_flags & UMA_ZONE_NUMA) != 0) { 3160 domain = PCPU_GET(domain); 3161 if (VM_DOMAIN_EMPTY(domain)) 3162 domain = UMA_ANYDOMAIN; 3163 } else 3164 domain = 0; 3165 zdom = &zone->uz_domain[0]; 3166 3167 /* Can we throw this on the zone full list? */ 3168 if (bucket != NULL) { 3169 CTR3(KTR_UMA, 3170 "uma_zfree: zone %s(%p) putting bucket %p on free list", 3171 zone->uz_name, zone, bucket); 3172 /* ub_cnt is pointing to the last free item */ 3173 KASSERT(bucket->ub_cnt != 0, 3174 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 3175 if ((zone->uz_flags & UMA_ZONE_NOBUCKETCACHE) != 0) { 3176 ZONE_UNLOCK(zone); 3177 bucket_drain(zone, bucket); 3178 bucket_free(zone, bucket, udata); 3179 goto zfree_restart; 3180 } else 3181 LIST_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link); 3182 } 3183 3184 /* 3185 * We bump the uz count when the cache size is insufficient to 3186 * handle the working set. 3187 */ 3188 if (lockfail && zone->uz_count < BUCKET_MAX) 3189 zone->uz_count++; 3190 ZONE_UNLOCK(zone); 3191 3192 bucket = bucket_alloc(zone, udata, M_NOWAIT); 3193 CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p", 3194 zone->uz_name, zone, bucket); 3195 if (bucket) { 3196 critical_enter(); 3197 cpu = curcpu; 3198 cache = &zone->uz_cpu[cpu]; 3199 if (cache->uc_freebucket == NULL && 3200 ((zone->uz_flags & UMA_ZONE_NUMA) == 0 || 3201 domain == PCPU_GET(domain))) { 3202 cache->uc_freebucket = bucket; 3203 goto zfree_start; 3204 } 3205 /* 3206 * We lost the race, start over. We have to drop our 3207 * critical section to free the bucket. 3208 */ 3209 critical_exit(); 3210 bucket_free(zone, bucket, udata); 3211 goto zfree_restart; 3212 } 3213 3214 /* 3215 * If nothing else caught this, we'll just do an internal free. 3216 */ 3217 zfree_item: 3218 zone_free_item(zone, item, udata, SKIP_DTOR); 3219 3220 return; 3221 } 3222 3223 void 3224 uma_zfree_domain(uma_zone_t zone, void *item, void *udata) 3225 { 3226 3227 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3228 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3229 3230 CTR2(KTR_UMA, "uma_zfree_domain thread %x zone %s", curthread, 3231 zone->uz_name); 3232 3233 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3234 ("uma_zfree_domain: called with spinlock or critical section held")); 3235 3236 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 3237 if (item == NULL) 3238 return; 3239 zone_free_item(zone, item, udata, SKIP_NONE); 3240 } 3241 3242 static void 3243 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) 3244 { 3245 uma_domain_t dom; 3246 uint8_t freei; 3247 3248 mtx_assert(&keg->uk_lock, MA_OWNED); 3249 MPASS(keg == slab->us_keg); 3250 3251 dom = &keg->uk_domain[slab->us_domain]; 3252 3253 /* Do we need to remove from any lists? */ 3254 if (slab->us_freecount+1 == keg->uk_ipers) { 3255 LIST_REMOVE(slab, us_link); 3256 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); 3257 } else if (slab->us_freecount == 0) { 3258 LIST_REMOVE(slab, us_link); 3259 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 3260 } 3261 3262 /* Slab management. */ 3263 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3264 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); 3265 slab->us_freecount++; 3266 3267 /* Keg statistics. */ 3268 keg->uk_free++; 3269 } 3270 3271 static void 3272 zone_release(uma_zone_t zone, void **bucket, int cnt) 3273 { 3274 void *item; 3275 uma_slab_t slab; 3276 uma_keg_t keg; 3277 uint8_t *mem; 3278 int clearfull; 3279 int i; 3280 3281 clearfull = 0; 3282 keg = zone_first_keg(zone); 3283 KEG_LOCK(keg); 3284 for (i = 0; i < cnt; i++) { 3285 item = bucket[i]; 3286 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 3287 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 3288 if (zone->uz_flags & UMA_ZONE_HASH) { 3289 slab = hash_sfind(&keg->uk_hash, mem); 3290 } else { 3291 mem += keg->uk_pgoff; 3292 slab = (uma_slab_t)mem; 3293 } 3294 } else { 3295 slab = vtoslab((vm_offset_t)item); 3296 if (slab->us_keg != keg) { 3297 KEG_UNLOCK(keg); 3298 keg = slab->us_keg; 3299 KEG_LOCK(keg); 3300 } 3301 } 3302 slab_free_item(keg, slab, item); 3303 if (keg->uk_flags & UMA_ZFLAG_FULL) { 3304 if (keg->uk_pages < keg->uk_maxpages) { 3305 keg->uk_flags &= ~UMA_ZFLAG_FULL; 3306 clearfull = 1; 3307 } 3308 3309 /* 3310 * We can handle one more allocation. Since we're 3311 * clearing ZFLAG_FULL, wake up all procs blocked 3312 * on pages. This should be uncommon, so keeping this 3313 * simple for now (rather than adding count of blocked 3314 * threads etc). 3315 */ 3316 wakeup(keg); 3317 } 3318 } 3319 KEG_UNLOCK(keg); 3320 if (clearfull) { 3321 ZONE_LOCK(zone); 3322 zone->uz_flags &= ~UMA_ZFLAG_FULL; 3323 wakeup(zone); 3324 ZONE_UNLOCK(zone); 3325 } 3326 3327 } 3328 3329 /* 3330 * Frees a single item to any zone. 3331 * 3332 * Arguments: 3333 * zone The zone to free to 3334 * item The item we're freeing 3335 * udata User supplied data for the dtor 3336 * skip Skip dtors and finis 3337 */ 3338 static void 3339 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 3340 { 3341 #ifdef INVARIANTS 3342 bool skipdbg; 3343 3344 skipdbg = uma_dbg_zskip(zone, item); 3345 if (skip == SKIP_NONE && !skipdbg) { 3346 if (zone->uz_flags & UMA_ZONE_MALLOC) 3347 uma_dbg_free(zone, udata, item); 3348 else 3349 uma_dbg_free(zone, NULL, item); 3350 } 3351 3352 if (skip < SKIP_DTOR && zone->uz_dtor != NULL && 3353 (!skipdbg || zone->uz_dtor != trash_dtor || 3354 zone->uz_ctor != trash_ctor)) 3355 #else 3356 if (skip < SKIP_DTOR && zone->uz_dtor != NULL) 3357 #endif 3358 zone->uz_dtor(item, zone->uz_size, udata); 3359 3360 if (skip < SKIP_FINI && zone->uz_fini) 3361 zone->uz_fini(item, zone->uz_size); 3362 3363 atomic_add_long(&zone->uz_frees, 1); 3364 zone->uz_release(zone->uz_arg, &item, 1); 3365 } 3366 3367 /* See uma.h */ 3368 int 3369 uma_zone_set_max(uma_zone_t zone, int nitems) 3370 { 3371 uma_keg_t keg; 3372 3373 keg = zone_first_keg(zone); 3374 if (keg == NULL) 3375 return (0); 3376 KEG_LOCK(keg); 3377 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 3378 if (keg->uk_maxpages * keg->uk_ipers < nitems) 3379 keg->uk_maxpages += keg->uk_ppera; 3380 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers; 3381 KEG_UNLOCK(keg); 3382 3383 return (nitems); 3384 } 3385 3386 /* See uma.h */ 3387 int 3388 uma_zone_get_max(uma_zone_t zone) 3389 { 3390 int nitems; 3391 uma_keg_t keg; 3392 3393 keg = zone_first_keg(zone); 3394 if (keg == NULL) 3395 return (0); 3396 KEG_LOCK(keg); 3397 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers; 3398 KEG_UNLOCK(keg); 3399 3400 return (nitems); 3401 } 3402 3403 /* See uma.h */ 3404 void 3405 uma_zone_set_warning(uma_zone_t zone, const char *warning) 3406 { 3407 3408 ZONE_LOCK(zone); 3409 zone->uz_warning = warning; 3410 ZONE_UNLOCK(zone); 3411 } 3412 3413 /* See uma.h */ 3414 void 3415 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) 3416 { 3417 3418 ZONE_LOCK(zone); 3419 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); 3420 ZONE_UNLOCK(zone); 3421 } 3422 3423 /* See uma.h */ 3424 int 3425 uma_zone_get_cur(uma_zone_t zone) 3426 { 3427 int64_t nitems; 3428 u_int i; 3429 3430 ZONE_LOCK(zone); 3431 nitems = zone->uz_allocs - zone->uz_frees; 3432 CPU_FOREACH(i) { 3433 /* 3434 * See the comment in sysctl_vm_zone_stats() regarding the 3435 * safety of accessing the per-cpu caches. With the zone lock 3436 * held, it is safe, but can potentially result in stale data. 3437 */ 3438 nitems += zone->uz_cpu[i].uc_allocs - 3439 zone->uz_cpu[i].uc_frees; 3440 } 3441 ZONE_UNLOCK(zone); 3442 3443 return (nitems < 0 ? 0 : nitems); 3444 } 3445 3446 /* See uma.h */ 3447 void 3448 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 3449 { 3450 uma_keg_t keg; 3451 3452 keg = zone_first_keg(zone); 3453 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 3454 KEG_LOCK(keg); 3455 KASSERT(keg->uk_pages == 0, 3456 ("uma_zone_set_init on non-empty keg")); 3457 keg->uk_init = uminit; 3458 KEG_UNLOCK(keg); 3459 } 3460 3461 /* See uma.h */ 3462 void 3463 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 3464 { 3465 uma_keg_t keg; 3466 3467 keg = zone_first_keg(zone); 3468 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type")); 3469 KEG_LOCK(keg); 3470 KASSERT(keg->uk_pages == 0, 3471 ("uma_zone_set_fini on non-empty keg")); 3472 keg->uk_fini = fini; 3473 KEG_UNLOCK(keg); 3474 } 3475 3476 /* See uma.h */ 3477 void 3478 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 3479 { 3480 3481 ZONE_LOCK(zone); 3482 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3483 ("uma_zone_set_zinit on non-empty keg")); 3484 zone->uz_init = zinit; 3485 ZONE_UNLOCK(zone); 3486 } 3487 3488 /* See uma.h */ 3489 void 3490 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 3491 { 3492 3493 ZONE_LOCK(zone); 3494 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3495 ("uma_zone_set_zfini on non-empty keg")); 3496 zone->uz_fini = zfini; 3497 ZONE_UNLOCK(zone); 3498 } 3499 3500 /* See uma.h */ 3501 /* XXX uk_freef is not actually used with the zone locked */ 3502 void 3503 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 3504 { 3505 uma_keg_t keg; 3506 3507 keg = zone_first_keg(zone); 3508 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type")); 3509 KEG_LOCK(keg); 3510 keg->uk_freef = freef; 3511 KEG_UNLOCK(keg); 3512 } 3513 3514 /* See uma.h */ 3515 /* XXX uk_allocf is not actually used with the zone locked */ 3516 void 3517 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3518 { 3519 uma_keg_t keg; 3520 3521 keg = zone_first_keg(zone); 3522 KEG_LOCK(keg); 3523 keg->uk_allocf = allocf; 3524 KEG_UNLOCK(keg); 3525 } 3526 3527 /* See uma.h */ 3528 void 3529 uma_zone_reserve(uma_zone_t zone, int items) 3530 { 3531 uma_keg_t keg; 3532 3533 keg = zone_first_keg(zone); 3534 if (keg == NULL) 3535 return; 3536 KEG_LOCK(keg); 3537 keg->uk_reserve = items; 3538 KEG_UNLOCK(keg); 3539 3540 return; 3541 } 3542 3543 /* See uma.h */ 3544 int 3545 uma_zone_reserve_kva(uma_zone_t zone, int count) 3546 { 3547 uma_keg_t keg; 3548 vm_offset_t kva; 3549 u_int pages; 3550 3551 keg = zone_first_keg(zone); 3552 if (keg == NULL) 3553 return (0); 3554 pages = count / keg->uk_ipers; 3555 3556 if (pages * keg->uk_ipers < count) 3557 pages++; 3558 pages *= keg->uk_ppera; 3559 3560 #ifdef UMA_MD_SMALL_ALLOC 3561 if (keg->uk_ppera > 1) { 3562 #else 3563 if (1) { 3564 #endif 3565 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE); 3566 if (kva == 0) 3567 return (0); 3568 } else 3569 kva = 0; 3570 KEG_LOCK(keg); 3571 keg->uk_kva = kva; 3572 keg->uk_offset = 0; 3573 keg->uk_maxpages = pages; 3574 #ifdef UMA_MD_SMALL_ALLOC 3575 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3576 #else 3577 keg->uk_allocf = noobj_alloc; 3578 #endif 3579 keg->uk_flags |= UMA_ZONE_NOFREE; 3580 KEG_UNLOCK(keg); 3581 3582 return (1); 3583 } 3584 3585 /* See uma.h */ 3586 void 3587 uma_prealloc(uma_zone_t zone, int items) 3588 { 3589 uma_domain_t dom; 3590 uma_slab_t slab; 3591 uma_keg_t keg; 3592 int domain, slabs; 3593 3594 keg = zone_first_keg(zone); 3595 if (keg == NULL) 3596 return; 3597 KEG_LOCK(keg); 3598 slabs = items / keg->uk_ipers; 3599 domain = 0; 3600 if (slabs * keg->uk_ipers < items) 3601 slabs++; 3602 while (slabs > 0) { 3603 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK); 3604 if (slab == NULL) 3605 break; 3606 MPASS(slab->us_keg == keg); 3607 dom = &keg->uk_domain[slab->us_domain]; 3608 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); 3609 slabs--; 3610 do { 3611 domain = (domain + 1) % vm_ndomains; 3612 } while (VM_DOMAIN_EMPTY(domain)); 3613 } 3614 KEG_UNLOCK(keg); 3615 } 3616 3617 /* See uma.h */ 3618 static void 3619 uma_reclaim_locked(bool kmem_danger) 3620 { 3621 3622 CTR0(KTR_UMA, "UMA: vm asked us to release pages!"); 3623 sx_assert(&uma_drain_lock, SA_XLOCKED); 3624 bucket_enable(); 3625 zone_foreach(zone_drain); 3626 if (vm_page_count_min() || kmem_danger) { 3627 cache_drain_safe(NULL); 3628 zone_foreach(zone_drain); 3629 } 3630 /* 3631 * Some slabs may have been freed but this zone will be visited early 3632 * we visit again so that we can free pages that are empty once other 3633 * zones are drained. We have to do the same for buckets. 3634 */ 3635 zone_drain(slabzone); 3636 bucket_zone_drain(); 3637 } 3638 3639 void 3640 uma_reclaim(void) 3641 { 3642 3643 sx_xlock(&uma_drain_lock); 3644 uma_reclaim_locked(false); 3645 sx_xunlock(&uma_drain_lock); 3646 } 3647 3648 static volatile int uma_reclaim_needed; 3649 3650 void 3651 uma_reclaim_wakeup(void) 3652 { 3653 3654 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0) 3655 wakeup(uma_reclaim); 3656 } 3657 3658 void 3659 uma_reclaim_worker(void *arg __unused) 3660 { 3661 3662 for (;;) { 3663 sx_xlock(&uma_drain_lock); 3664 while (atomic_load_int(&uma_reclaim_needed) == 0) 3665 sx_sleep(uma_reclaim, &uma_drain_lock, PVM, "umarcl", 3666 hz); 3667 sx_xunlock(&uma_drain_lock); 3668 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); 3669 sx_xlock(&uma_drain_lock); 3670 uma_reclaim_locked(true); 3671 atomic_store_int(&uma_reclaim_needed, 0); 3672 sx_xunlock(&uma_drain_lock); 3673 /* Don't fire more than once per-second. */ 3674 pause("umarclslp", hz); 3675 } 3676 } 3677 3678 /* See uma.h */ 3679 int 3680 uma_zone_exhausted(uma_zone_t zone) 3681 { 3682 int full; 3683 3684 ZONE_LOCK(zone); 3685 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3686 ZONE_UNLOCK(zone); 3687 return (full); 3688 } 3689 3690 int 3691 uma_zone_exhausted_nolock(uma_zone_t zone) 3692 { 3693 return (zone->uz_flags & UMA_ZFLAG_FULL); 3694 } 3695 3696 void * 3697 uma_large_malloc_domain(vm_size_t size, int domain, int wait) 3698 { 3699 vm_offset_t addr; 3700 uma_slab_t slab; 3701 3702 if (domain != UMA_ANYDOMAIN) { 3703 /* avoid allocs targeting empty domains */ 3704 if (VM_DOMAIN_EMPTY(domain)) 3705 domain = UMA_ANYDOMAIN; 3706 } 3707 slab = zone_alloc_item(slabzone, NULL, domain, wait); 3708 if (slab == NULL) 3709 return (NULL); 3710 if (domain == UMA_ANYDOMAIN) 3711 addr = kmem_malloc(size, wait); 3712 else 3713 addr = kmem_malloc_domain(domain, size, wait); 3714 if (addr != 0) { 3715 vsetslab(addr, slab); 3716 slab->us_data = (void *)addr; 3717 slab->us_flags = UMA_SLAB_KERNEL | UMA_SLAB_MALLOC; 3718 slab->us_size = size; 3719 slab->us_domain = vm_phys_domain(PHYS_TO_VM_PAGE( 3720 pmap_kextract(addr))); 3721 uma_total_inc(size); 3722 } else { 3723 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3724 } 3725 3726 return ((void *)addr); 3727 } 3728 3729 void * 3730 uma_large_malloc(vm_size_t size, int wait) 3731 { 3732 3733 return uma_large_malloc_domain(size, UMA_ANYDOMAIN, wait); 3734 } 3735 3736 void 3737 uma_large_free(uma_slab_t slab) 3738 { 3739 3740 KASSERT((slab->us_flags & UMA_SLAB_KERNEL) != 0, 3741 ("uma_large_free: Memory not allocated with uma_large_malloc.")); 3742 kmem_free((vm_offset_t)slab->us_data, slab->us_size); 3743 uma_total_dec(slab->us_size); 3744 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3745 } 3746 3747 static void 3748 uma_zero_item(void *item, uma_zone_t zone) 3749 { 3750 3751 bzero(item, zone->uz_size); 3752 } 3753 3754 unsigned long 3755 uma_limit(void) 3756 { 3757 3758 return (uma_kmem_limit); 3759 } 3760 3761 void 3762 uma_set_limit(unsigned long limit) 3763 { 3764 3765 uma_kmem_limit = limit; 3766 } 3767 3768 unsigned long 3769 uma_size(void) 3770 { 3771 3772 return (uma_kmem_total); 3773 } 3774 3775 long 3776 uma_avail(void) 3777 { 3778 3779 return (uma_kmem_limit - uma_kmem_total); 3780 } 3781 3782 void 3783 uma_print_stats(void) 3784 { 3785 zone_foreach(uma_print_zone); 3786 } 3787 3788 static void 3789 slab_print(uma_slab_t slab) 3790 { 3791 printf("slab: keg %p, data %p, freecount %d\n", 3792 slab->us_keg, slab->us_data, slab->us_freecount); 3793 } 3794 3795 static void 3796 cache_print(uma_cache_t cache) 3797 { 3798 printf("alloc: %p(%d), free: %p(%d)\n", 3799 cache->uc_allocbucket, 3800 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3801 cache->uc_freebucket, 3802 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3803 } 3804 3805 static void 3806 uma_print_keg(uma_keg_t keg) 3807 { 3808 uma_domain_t dom; 3809 uma_slab_t slab; 3810 int i; 3811 3812 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3813 "out %d free %d limit %d\n", 3814 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3815 keg->uk_ipers, keg->uk_ppera, 3816 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, 3817 keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3818 for (i = 0; i < vm_ndomains; i++) { 3819 dom = &keg->uk_domain[i]; 3820 printf("Part slabs:\n"); 3821 LIST_FOREACH(slab, &dom->ud_part_slab, us_link) 3822 slab_print(slab); 3823 printf("Free slabs:\n"); 3824 LIST_FOREACH(slab, &dom->ud_free_slab, us_link) 3825 slab_print(slab); 3826 printf("Full slabs:\n"); 3827 LIST_FOREACH(slab, &dom->ud_full_slab, us_link) 3828 slab_print(slab); 3829 } 3830 } 3831 3832 void 3833 uma_print_zone(uma_zone_t zone) 3834 { 3835 uma_cache_t cache; 3836 uma_klink_t kl; 3837 int i; 3838 3839 printf("zone: %s(%p) size %d flags %#x\n", 3840 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3841 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3842 uma_print_keg(kl->kl_keg); 3843 CPU_FOREACH(i) { 3844 cache = &zone->uz_cpu[i]; 3845 printf("CPU %d Cache:\n", i); 3846 cache_print(cache); 3847 } 3848 } 3849 3850 #ifdef DDB 3851 /* 3852 * Generate statistics across both the zone and its per-cpu cache's. Return 3853 * desired statistics if the pointer is non-NULL for that statistic. 3854 * 3855 * Note: does not update the zone statistics, as it can't safely clear the 3856 * per-CPU cache statistic. 3857 * 3858 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3859 * safe from off-CPU; we should modify the caches to track this information 3860 * directly so that we don't have to. 3861 */ 3862 static void 3863 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3864 uint64_t *freesp, uint64_t *sleepsp) 3865 { 3866 uma_cache_t cache; 3867 uint64_t allocs, frees, sleeps; 3868 int cachefree, cpu; 3869 3870 allocs = frees = sleeps = 0; 3871 cachefree = 0; 3872 CPU_FOREACH(cpu) { 3873 cache = &z->uz_cpu[cpu]; 3874 if (cache->uc_allocbucket != NULL) 3875 cachefree += cache->uc_allocbucket->ub_cnt; 3876 if (cache->uc_freebucket != NULL) 3877 cachefree += cache->uc_freebucket->ub_cnt; 3878 allocs += cache->uc_allocs; 3879 frees += cache->uc_frees; 3880 } 3881 allocs += z->uz_allocs; 3882 frees += z->uz_frees; 3883 sleeps += z->uz_sleeps; 3884 if (cachefreep != NULL) 3885 *cachefreep = cachefree; 3886 if (allocsp != NULL) 3887 *allocsp = allocs; 3888 if (freesp != NULL) 3889 *freesp = frees; 3890 if (sleepsp != NULL) 3891 *sleepsp = sleeps; 3892 } 3893 #endif /* DDB */ 3894 3895 static int 3896 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3897 { 3898 uma_keg_t kz; 3899 uma_zone_t z; 3900 int count; 3901 3902 count = 0; 3903 rw_rlock(&uma_rwlock); 3904 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3905 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3906 count++; 3907 } 3908 rw_runlock(&uma_rwlock); 3909 return (sysctl_handle_int(oidp, &count, 0, req)); 3910 } 3911 3912 static int 3913 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3914 { 3915 struct uma_stream_header ush; 3916 struct uma_type_header uth; 3917 struct uma_percpu_stat *ups; 3918 uma_bucket_t bucket; 3919 uma_zone_domain_t zdom; 3920 struct sbuf sbuf; 3921 uma_cache_t cache; 3922 uma_klink_t kl; 3923 uma_keg_t kz; 3924 uma_zone_t z; 3925 uma_keg_t k; 3926 int count, error, i; 3927 3928 error = sysctl_wire_old_buffer(req, 0); 3929 if (error != 0) 3930 return (error); 3931 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3932 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); 3933 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK); 3934 3935 count = 0; 3936 rw_rlock(&uma_rwlock); 3937 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3938 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3939 count++; 3940 } 3941 3942 /* 3943 * Insert stream header. 3944 */ 3945 bzero(&ush, sizeof(ush)); 3946 ush.ush_version = UMA_STREAM_VERSION; 3947 ush.ush_maxcpus = (mp_maxid + 1); 3948 ush.ush_count = count; 3949 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3950 3951 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3952 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3953 bzero(&uth, sizeof(uth)); 3954 ZONE_LOCK(z); 3955 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3956 uth.uth_align = kz->uk_align; 3957 uth.uth_size = kz->uk_size; 3958 uth.uth_rsize = kz->uk_rsize; 3959 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3960 k = kl->kl_keg; 3961 uth.uth_maxpages += k->uk_maxpages; 3962 uth.uth_pages += k->uk_pages; 3963 uth.uth_keg_free += k->uk_free; 3964 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3965 * k->uk_ipers; 3966 } 3967 3968 /* 3969 * A zone is secondary is it is not the first entry 3970 * on the keg's zone list. 3971 */ 3972 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3973 (LIST_FIRST(&kz->uk_zones) != z)) 3974 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3975 3976 for (i = 0; i < vm_ndomains; i++) { 3977 zdom = &z->uz_domain[i]; 3978 LIST_FOREACH(bucket, &zdom->uzd_buckets, 3979 ub_link) 3980 uth.uth_zone_free += bucket->ub_cnt; 3981 } 3982 uth.uth_allocs = z->uz_allocs; 3983 uth.uth_frees = z->uz_frees; 3984 uth.uth_fails = z->uz_fails; 3985 uth.uth_sleeps = z->uz_sleeps; 3986 /* 3987 * While it is not normally safe to access the cache 3988 * bucket pointers while not on the CPU that owns the 3989 * cache, we only allow the pointers to be exchanged 3990 * without the zone lock held, not invalidated, so 3991 * accept the possible race associated with bucket 3992 * exchange during monitoring. 3993 */ 3994 for (i = 0; i < mp_maxid + 1; i++) { 3995 bzero(&ups[i], sizeof(*ups)); 3996 if (kz->uk_flags & UMA_ZFLAG_INTERNAL || 3997 CPU_ABSENT(i)) 3998 continue; 3999 cache = &z->uz_cpu[i]; 4000 if (cache->uc_allocbucket != NULL) 4001 ups[i].ups_cache_free += 4002 cache->uc_allocbucket->ub_cnt; 4003 if (cache->uc_freebucket != NULL) 4004 ups[i].ups_cache_free += 4005 cache->uc_freebucket->ub_cnt; 4006 ups[i].ups_allocs = cache->uc_allocs; 4007 ups[i].ups_frees = cache->uc_frees; 4008 } 4009 ZONE_UNLOCK(z); 4010 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 4011 for (i = 0; i < mp_maxid + 1; i++) 4012 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 4013 } 4014 } 4015 rw_runlock(&uma_rwlock); 4016 error = sbuf_finish(&sbuf); 4017 sbuf_delete(&sbuf); 4018 free(ups, M_TEMP); 4019 return (error); 4020 } 4021 4022 int 4023 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 4024 { 4025 uma_zone_t zone = *(uma_zone_t *)arg1; 4026 int error, max; 4027 4028 max = uma_zone_get_max(zone); 4029 error = sysctl_handle_int(oidp, &max, 0, req); 4030 if (error || !req->newptr) 4031 return (error); 4032 4033 uma_zone_set_max(zone, max); 4034 4035 return (0); 4036 } 4037 4038 int 4039 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 4040 { 4041 uma_zone_t zone = *(uma_zone_t *)arg1; 4042 int cur; 4043 4044 cur = uma_zone_get_cur(zone); 4045 return (sysctl_handle_int(oidp, &cur, 0, req)); 4046 } 4047 4048 #ifdef INVARIANTS 4049 static uma_slab_t 4050 uma_dbg_getslab(uma_zone_t zone, void *item) 4051 { 4052 uma_slab_t slab; 4053 uma_keg_t keg; 4054 uint8_t *mem; 4055 4056 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 4057 if (zone->uz_flags & UMA_ZONE_VTOSLAB) { 4058 slab = vtoslab((vm_offset_t)mem); 4059 } else { 4060 /* 4061 * It is safe to return the slab here even though the 4062 * zone is unlocked because the item's allocation state 4063 * essentially holds a reference. 4064 */ 4065 ZONE_LOCK(zone); 4066 keg = LIST_FIRST(&zone->uz_kegs)->kl_keg; 4067 if (keg->uk_flags & UMA_ZONE_HASH) 4068 slab = hash_sfind(&keg->uk_hash, mem); 4069 else 4070 slab = (uma_slab_t)(mem + keg->uk_pgoff); 4071 ZONE_UNLOCK(zone); 4072 } 4073 4074 return (slab); 4075 } 4076 4077 static bool 4078 uma_dbg_zskip(uma_zone_t zone, void *mem) 4079 { 4080 uma_keg_t keg; 4081 4082 if ((keg = zone_first_keg(zone)) == NULL) 4083 return (true); 4084 4085 return (uma_dbg_kskip(keg, mem)); 4086 } 4087 4088 static bool 4089 uma_dbg_kskip(uma_keg_t keg, void *mem) 4090 { 4091 uintptr_t idx; 4092 4093 if (dbg_divisor == 0) 4094 return (true); 4095 4096 if (dbg_divisor == 1) 4097 return (false); 4098 4099 idx = (uintptr_t)mem >> PAGE_SHIFT; 4100 if (keg->uk_ipers > 1) { 4101 idx *= keg->uk_ipers; 4102 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize; 4103 } 4104 4105 if ((idx / dbg_divisor) * dbg_divisor != idx) { 4106 counter_u64_add(uma_skip_cnt, 1); 4107 return (true); 4108 } 4109 counter_u64_add(uma_dbg_cnt, 1); 4110 4111 return (false); 4112 } 4113 4114 /* 4115 * Set up the slab's freei data such that uma_dbg_free can function. 4116 * 4117 */ 4118 static void 4119 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) 4120 { 4121 uma_keg_t keg; 4122 int freei; 4123 4124 if (slab == NULL) { 4125 slab = uma_dbg_getslab(zone, item); 4126 if (slab == NULL) 4127 panic("uma: item %p did not belong to zone %s\n", 4128 item, zone->uz_name); 4129 } 4130 keg = slab->us_keg; 4131 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 4132 4133 if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) 4134 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n", 4135 item, zone, zone->uz_name, slab, freei); 4136 BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); 4137 4138 return; 4139 } 4140 4141 /* 4142 * Verifies freed addresses. Checks for alignment, valid slab membership 4143 * and duplicate frees. 4144 * 4145 */ 4146 static void 4147 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) 4148 { 4149 uma_keg_t keg; 4150 int freei; 4151 4152 if (slab == NULL) { 4153 slab = uma_dbg_getslab(zone, item); 4154 if (slab == NULL) 4155 panic("uma: Freed item %p did not belong to zone %s\n", 4156 item, zone->uz_name); 4157 } 4158 keg = slab->us_keg; 4159 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 4160 4161 if (freei >= keg->uk_ipers) 4162 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n", 4163 item, zone, zone->uz_name, slab, freei); 4164 4165 if (((freei * keg->uk_rsize) + slab->us_data) != item) 4166 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n", 4167 item, zone, zone->uz_name, slab, freei); 4168 4169 if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) 4170 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n", 4171 item, zone, zone->uz_name, slab, freei); 4172 4173 BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); 4174 } 4175 #endif /* INVARIANTS */ 4176 4177 #ifdef DDB 4178 DB_SHOW_COMMAND(uma, db_show_uma) 4179 { 4180 uma_bucket_t bucket; 4181 uma_keg_t kz; 4182 uma_zone_t z; 4183 uma_zone_domain_t zdom; 4184 uint64_t allocs, frees, sleeps; 4185 int cachefree, i; 4186 4187 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used", 4188 "Free", "Requests", "Sleeps", "Bucket"); 4189 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4190 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 4191 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 4192 allocs = z->uz_allocs; 4193 frees = z->uz_frees; 4194 sleeps = z->uz_sleeps; 4195 cachefree = 0; 4196 } else 4197 uma_zone_sumstat(z, &cachefree, &allocs, 4198 &frees, &sleeps); 4199 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 4200 (LIST_FIRST(&kz->uk_zones) != z))) 4201 cachefree += kz->uk_free; 4202 for (i = 0; i < vm_ndomains; i++) { 4203 zdom = &z->uz_domain[i]; 4204 LIST_FOREACH(bucket, &zdom->uzd_buckets, 4205 ub_link) 4206 cachefree += bucket->ub_cnt; 4207 } 4208 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n", 4209 z->uz_name, (uintmax_t)kz->uk_size, 4210 (intmax_t)(allocs - frees), cachefree, 4211 (uintmax_t)allocs, sleeps, z->uz_count); 4212 if (db_pager_quit) 4213 return; 4214 } 4215 } 4216 } 4217 4218 DB_SHOW_COMMAND(umacache, db_show_umacache) 4219 { 4220 uma_bucket_t bucket; 4221 uma_zone_t z; 4222 uma_zone_domain_t zdom; 4223 uint64_t allocs, frees; 4224 int cachefree, i; 4225 4226 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 4227 "Requests", "Bucket"); 4228 LIST_FOREACH(z, &uma_cachezones, uz_link) { 4229 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL); 4230 for (i = 0; i < vm_ndomains; i++) { 4231 zdom = &z->uz_domain[i]; 4232 LIST_FOREACH(bucket, &zdom->uzd_buckets, ub_link) 4233 cachefree += bucket->ub_cnt; 4234 } 4235 db_printf("%18s %8ju %8jd %8d %12ju %8u\n", 4236 z->uz_name, (uintmax_t)z->uz_size, 4237 (intmax_t)(allocs - frees), cachefree, 4238 (uintmax_t)allocs, z->uz_count); 4239 if (db_pager_quit) 4240 return; 4241 } 4242 } 4243 #endif /* DDB */ 4244