1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2002-2019 Jeffrey Roberson <jeff@FreeBSD.org> 5 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org> 6 * Copyright (c) 2004-2006 Robert N. M. Watson 7 * All rights reserved. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice unmodified, this list of conditions, and the following 14 * disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 29 */ 30 31 /* 32 * uma_core.c Implementation of the Universal Memory allocator 33 * 34 * This allocator is intended to replace the multitude of similar object caches 35 * in the standard FreeBSD kernel. The intent is to be flexible as well as 36 * efficient. A primary design goal is to return unused memory to the rest of 37 * the system. This will make the system as a whole more flexible due to the 38 * ability to move memory to subsystems which most need it instead of leaving 39 * pools of reserved memory unused. 40 * 41 * The basic ideas stem from similar slab/zone based allocators whose algorithms 42 * are well known. 43 * 44 */ 45 46 /* 47 * TODO: 48 * - Improve memory usage for large allocations 49 * - Investigate cache size adjustments 50 */ 51 52 #include <sys/cdefs.h> 53 __FBSDID("$FreeBSD$"); 54 55 #include "opt_ddb.h" 56 #include "opt_param.h" 57 #include "opt_vm.h" 58 59 #include <sys/param.h> 60 #include <sys/systm.h> 61 #include <sys/bitset.h> 62 #include <sys/domainset.h> 63 #include <sys/eventhandler.h> 64 #include <sys/kernel.h> 65 #include <sys/types.h> 66 #include <sys/limits.h> 67 #include <sys/queue.h> 68 #include <sys/malloc.h> 69 #include <sys/ktr.h> 70 #include <sys/lock.h> 71 #include <sys/sysctl.h> 72 #include <sys/mutex.h> 73 #include <sys/proc.h> 74 #include <sys/random.h> 75 #include <sys/rwlock.h> 76 #include <sys/sbuf.h> 77 #include <sys/sched.h> 78 #include <sys/sleepqueue.h> 79 #include <sys/smp.h> 80 #include <sys/taskqueue.h> 81 #include <sys/vmmeter.h> 82 83 #include <vm/vm.h> 84 #include <vm/vm_domainset.h> 85 #include <vm/vm_object.h> 86 #include <vm/vm_page.h> 87 #include <vm/vm_pageout.h> 88 #include <vm/vm_param.h> 89 #include <vm/vm_phys.h> 90 #include <vm/vm_pagequeue.h> 91 #include <vm/vm_map.h> 92 #include <vm/vm_kern.h> 93 #include <vm/vm_extern.h> 94 #include <vm/uma.h> 95 #include <vm/uma_int.h> 96 #include <vm/uma_dbg.h> 97 98 #include <ddb/ddb.h> 99 100 #ifdef DEBUG_MEMGUARD 101 #include <vm/memguard.h> 102 #endif 103 104 #include <machine/md_var.h> 105 106 /* 107 * This is the zone and keg from which all zones are spawned. 108 */ 109 static uma_zone_t kegs; 110 static uma_zone_t zones; 111 112 /* 113 * These are the two zones from which all offpage uma_slab_ts are allocated. 114 * 115 * One zone is for slab headers that can represent a larger number of items, 116 * making the slabs themselves more efficient, and the other zone is for 117 * headers that are smaller and represent fewer items, making the headers more 118 * efficient. 119 */ 120 #define SLABZONE_SIZE(setsize) \ 121 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS) 122 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16) 123 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE 124 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE) 125 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE) 126 static uma_zone_t slabzones[2]; 127 128 /* 129 * The initial hash tables come out of this zone so they can be allocated 130 * prior to malloc coming up. 131 */ 132 static uma_zone_t hashzone; 133 134 /* The boot-time adjusted value for cache line alignment. */ 135 int uma_align_cache = 64 - 1; 136 137 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 138 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc"); 139 140 /* 141 * Are we allowed to allocate buckets? 142 */ 143 static int bucketdisable = 1; 144 145 /* Linked list of all kegs in the system */ 146 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 147 148 /* Linked list of all cache-only zones in the system */ 149 static LIST_HEAD(,uma_zone) uma_cachezones = 150 LIST_HEAD_INITIALIZER(uma_cachezones); 151 152 /* This RW lock protects the keg list */ 153 static struct rwlock_padalign __exclusive_cache_line uma_rwlock; 154 155 /* 156 * First available virual address for boot time allocations. 157 */ 158 static vm_offset_t bootstart; 159 static vm_offset_t bootmem; 160 161 static struct sx uma_reclaim_lock; 162 163 /* 164 * kmem soft limit, initialized by uma_set_limit(). Ensure that early 165 * allocations don't trigger a wakeup of the reclaim thread. 166 */ 167 unsigned long uma_kmem_limit = LONG_MAX; 168 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0, 169 "UMA kernel memory soft limit"); 170 unsigned long uma_kmem_total; 171 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0, 172 "UMA kernel memory usage"); 173 174 /* Is the VM done starting up? */ 175 static enum { 176 BOOT_COLD, 177 BOOT_KVA, 178 BOOT_RUNNING, 179 BOOT_SHUTDOWN, 180 } booted = BOOT_COLD; 181 182 /* 183 * This is the handle used to schedule events that need to happen 184 * outside of the allocation fast path. 185 */ 186 static struct callout uma_callout; 187 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 188 189 /* 190 * This structure is passed as the zone ctor arg so that I don't have to create 191 * a special allocation function just for zones. 192 */ 193 struct uma_zctor_args { 194 const char *name; 195 size_t size; 196 uma_ctor ctor; 197 uma_dtor dtor; 198 uma_init uminit; 199 uma_fini fini; 200 uma_import import; 201 uma_release release; 202 void *arg; 203 uma_keg_t keg; 204 int align; 205 uint32_t flags; 206 }; 207 208 struct uma_kctor_args { 209 uma_zone_t zone; 210 size_t size; 211 uma_init uminit; 212 uma_fini fini; 213 int align; 214 uint32_t flags; 215 }; 216 217 struct uma_bucket_zone { 218 uma_zone_t ubz_zone; 219 char *ubz_name; 220 int ubz_entries; /* Number of items it can hold. */ 221 int ubz_maxsize; /* Maximum allocation size per-item. */ 222 }; 223 224 /* 225 * Compute the actual number of bucket entries to pack them in power 226 * of two sizes for more efficient space utilization. 227 */ 228 #define BUCKET_SIZE(n) \ 229 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 230 231 #define BUCKET_MAX BUCKET_SIZE(256) 232 #define BUCKET_MIN BUCKET_SIZE(4) 233 234 struct uma_bucket_zone bucket_zones[] = { 235 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 236 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 237 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 238 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 239 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 240 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 241 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 242 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 243 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 244 { NULL, NULL, 0} 245 }; 246 247 /* 248 * Flags and enumerations to be passed to internal functions. 249 */ 250 enum zfreeskip { 251 SKIP_NONE = 0, 252 SKIP_CNT = 0x00000001, 253 SKIP_DTOR = 0x00010000, 254 SKIP_FINI = 0x00020000, 255 }; 256 257 /* Prototypes.. */ 258 259 void uma_startup1(vm_offset_t); 260 void uma_startup2(void); 261 262 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 263 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 264 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 265 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 266 static void page_free(void *, vm_size_t, uint8_t); 267 static void pcpu_page_free(void *, vm_size_t, uint8_t); 268 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int); 269 static void cache_drain(uma_zone_t); 270 static void bucket_drain(uma_zone_t, uma_bucket_t); 271 static void bucket_cache_reclaim(uma_zone_t zone, bool); 272 static int keg_ctor(void *, int, void *, int); 273 static void keg_dtor(void *, int, void *); 274 static int zone_ctor(void *, int, void *, int); 275 static void zone_dtor(void *, int, void *); 276 static int zero_init(void *, int, int); 277 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *); 278 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *); 279 static void zone_timeout(uma_zone_t zone, void *); 280 static int hash_alloc(struct uma_hash *, u_int); 281 static int hash_expand(struct uma_hash *, struct uma_hash *); 282 static void hash_free(struct uma_hash *hash); 283 static void uma_timeout(void *); 284 static void uma_startup3(void); 285 static void uma_shutdown(void); 286 static void *zone_alloc_item(uma_zone_t, void *, int, int); 287 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 288 static int zone_alloc_limit(uma_zone_t zone, int count, int flags); 289 static void zone_free_limit(uma_zone_t zone, int count); 290 static void bucket_enable(void); 291 static void bucket_init(void); 292 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 293 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 294 static void bucket_zone_drain(void); 295 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int); 296 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 297 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item); 298 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 299 uma_fini fini, int align, uint32_t flags); 300 static int zone_import(void *, void **, int, int, int); 301 static void zone_release(void *, void **, int); 302 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int); 303 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int); 304 305 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 306 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 307 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS); 308 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS); 309 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS); 310 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS); 311 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS); 312 313 static uint64_t uma_zone_get_allocs(uma_zone_t zone); 314 315 #ifdef INVARIANTS 316 static uint64_t uma_keg_get_allocs(uma_keg_t zone); 317 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg); 318 319 static bool uma_dbg_kskip(uma_keg_t keg, void *mem); 320 static bool uma_dbg_zskip(uma_zone_t zone, void *mem); 321 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item); 322 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item); 323 324 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0, 325 "Memory allocation debugging"); 326 327 static u_int dbg_divisor = 1; 328 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor, 329 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0, 330 "Debug & thrash every this item in memory allocator"); 331 332 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER; 333 static counter_u64_t uma_skip_cnt = EARLY_COUNTER; 334 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD, 335 &uma_dbg_cnt, "memory items debugged"); 336 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD, 337 &uma_skip_cnt, "memory items skipped, not debugged"); 338 #endif 339 340 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 341 342 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW, 0, "Universal Memory Allocator"); 343 344 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT, 345 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 346 347 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT, 348 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 349 350 static int zone_warnings = 1; 351 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, 352 "Warn when UMA zones becomes full"); 353 354 /* 355 * Select the slab zone for an offpage slab with the given maximum item count. 356 */ 357 static inline uma_zone_t 358 slabzone(int ipers) 359 { 360 361 return (slabzones[ipers > SLABZONE0_SETSIZE]); 362 } 363 364 /* 365 * This routine checks to see whether or not it's safe to enable buckets. 366 */ 367 static void 368 bucket_enable(void) 369 { 370 371 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init")); 372 bucketdisable = vm_page_count_min(); 373 } 374 375 /* 376 * Initialize bucket_zones, the array of zones of buckets of various sizes. 377 * 378 * For each zone, calculate the memory required for each bucket, consisting 379 * of the header and an array of pointers. 380 */ 381 static void 382 bucket_init(void) 383 { 384 struct uma_bucket_zone *ubz; 385 int size; 386 387 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 388 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 389 size += sizeof(void *) * ubz->ubz_entries; 390 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 391 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 392 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET | 393 UMA_ZONE_FIRSTTOUCH); 394 } 395 } 396 397 /* 398 * Given a desired number of entries for a bucket, return the zone from which 399 * to allocate the bucket. 400 */ 401 static struct uma_bucket_zone * 402 bucket_zone_lookup(int entries) 403 { 404 struct uma_bucket_zone *ubz; 405 406 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 407 if (ubz->ubz_entries >= entries) 408 return (ubz); 409 ubz--; 410 return (ubz); 411 } 412 413 static struct uma_bucket_zone * 414 bucket_zone_max(uma_zone_t zone, int nitems) 415 { 416 struct uma_bucket_zone *ubz; 417 int bpcpu; 418 419 bpcpu = 2; 420 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 421 /* Count the cross-domain bucket. */ 422 bpcpu++; 423 424 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 425 if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems) 426 break; 427 if (ubz == &bucket_zones[0]) 428 ubz = NULL; 429 else 430 ubz--; 431 return (ubz); 432 } 433 434 static int 435 bucket_select(int size) 436 { 437 struct uma_bucket_zone *ubz; 438 439 ubz = &bucket_zones[0]; 440 if (size > ubz->ubz_maxsize) 441 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 442 443 for (; ubz->ubz_entries != 0; ubz++) 444 if (ubz->ubz_maxsize < size) 445 break; 446 ubz--; 447 return (ubz->ubz_entries); 448 } 449 450 static uma_bucket_t 451 bucket_alloc(uma_zone_t zone, void *udata, int flags) 452 { 453 struct uma_bucket_zone *ubz; 454 uma_bucket_t bucket; 455 456 /* 457 * Don't allocate buckets in low memory situations. 458 */ 459 if (bucketdisable) 460 return (NULL); 461 462 /* 463 * To limit bucket recursion we store the original zone flags 464 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 465 * NOVM flag to persist even through deep recursions. We also 466 * store ZFLAG_BUCKET once we have recursed attempting to allocate 467 * a bucket for a bucket zone so we do not allow infinite bucket 468 * recursion. This cookie will even persist to frees of unused 469 * buckets via the allocation path or bucket allocations in the 470 * free path. 471 */ 472 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 473 udata = (void *)(uintptr_t)zone->uz_flags; 474 else { 475 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 476 return (NULL); 477 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 478 } 479 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 480 flags |= M_NOVM; 481 ubz = bucket_zone_lookup(zone->uz_bucket_size); 482 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 483 ubz++; 484 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 485 if (bucket) { 486 #ifdef INVARIANTS 487 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 488 #endif 489 bucket->ub_cnt = 0; 490 bucket->ub_entries = ubz->ubz_entries; 491 } 492 493 return (bucket); 494 } 495 496 static void 497 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 498 { 499 struct uma_bucket_zone *ubz; 500 501 KASSERT(bucket->ub_cnt == 0, 502 ("bucket_free: Freeing a non free bucket.")); 503 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 504 udata = (void *)(uintptr_t)zone->uz_flags; 505 ubz = bucket_zone_lookup(bucket->ub_entries); 506 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 507 } 508 509 static void 510 bucket_zone_drain(void) 511 { 512 struct uma_bucket_zone *ubz; 513 514 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 515 uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN); 516 } 517 518 /* 519 * Attempt to satisfy an allocation by retrieving a full bucket from one of the 520 * zone's caches. 521 */ 522 static uma_bucket_t 523 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom) 524 { 525 uma_bucket_t bucket; 526 527 ZONE_LOCK_ASSERT(zone); 528 529 if ((bucket = TAILQ_FIRST(&zdom->uzd_buckets)) != NULL) { 530 MPASS(zdom->uzd_nitems >= bucket->ub_cnt); 531 TAILQ_REMOVE(&zdom->uzd_buckets, bucket, ub_link); 532 zdom->uzd_nitems -= bucket->ub_cnt; 533 if (zdom->uzd_imin > zdom->uzd_nitems) 534 zdom->uzd_imin = zdom->uzd_nitems; 535 zone->uz_bkt_count -= bucket->ub_cnt; 536 } 537 return (bucket); 538 } 539 540 /* 541 * Insert a full bucket into the specified cache. The "ws" parameter indicates 542 * whether the bucket's contents should be counted as part of the zone's working 543 * set. 544 */ 545 static void 546 zone_put_bucket(uma_zone_t zone, uma_zone_domain_t zdom, uma_bucket_t bucket, 547 const bool ws) 548 { 549 550 ZONE_LOCK_ASSERT(zone); 551 KASSERT(!ws || zone->uz_bkt_count < zone->uz_bkt_max, 552 ("%s: zone %p overflow", __func__, zone)); 553 554 if (ws) 555 TAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link); 556 else 557 TAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link); 558 zdom->uzd_nitems += bucket->ub_cnt; 559 if (ws && zdom->uzd_imax < zdom->uzd_nitems) 560 zdom->uzd_imax = zdom->uzd_nitems; 561 zone->uz_bkt_count += bucket->ub_cnt; 562 } 563 564 /* Pops an item out of a per-cpu cache bucket. */ 565 static inline void * 566 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket) 567 { 568 void *item; 569 570 CRITICAL_ASSERT(curthread); 571 572 bucket->ucb_cnt--; 573 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt]; 574 #ifdef INVARIANTS 575 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL; 576 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 577 #endif 578 cache->uc_allocs++; 579 580 return (item); 581 } 582 583 /* Pushes an item into a per-cpu cache bucket. */ 584 static inline void 585 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item) 586 { 587 588 CRITICAL_ASSERT(curthread); 589 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL, 590 ("uma_zfree: Freeing to non free bucket index.")); 591 592 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item; 593 bucket->ucb_cnt++; 594 cache->uc_frees++; 595 } 596 597 /* 598 * Unload a UMA bucket from a per-cpu cache. 599 */ 600 static inline uma_bucket_t 601 cache_bucket_unload(uma_cache_bucket_t bucket) 602 { 603 uma_bucket_t b; 604 605 b = bucket->ucb_bucket; 606 if (b != NULL) { 607 MPASS(b->ub_entries == bucket->ucb_entries); 608 b->ub_cnt = bucket->ucb_cnt; 609 bucket->ucb_bucket = NULL; 610 bucket->ucb_entries = bucket->ucb_cnt = 0; 611 } 612 613 return (b); 614 } 615 616 static inline uma_bucket_t 617 cache_bucket_unload_alloc(uma_cache_t cache) 618 { 619 620 return (cache_bucket_unload(&cache->uc_allocbucket)); 621 } 622 623 static inline uma_bucket_t 624 cache_bucket_unload_free(uma_cache_t cache) 625 { 626 627 return (cache_bucket_unload(&cache->uc_freebucket)); 628 } 629 630 static inline uma_bucket_t 631 cache_bucket_unload_cross(uma_cache_t cache) 632 { 633 634 return (cache_bucket_unload(&cache->uc_crossbucket)); 635 } 636 637 /* 638 * Load a bucket into a per-cpu cache bucket. 639 */ 640 static inline void 641 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b) 642 { 643 644 CRITICAL_ASSERT(curthread); 645 MPASS(bucket->ucb_bucket == NULL); 646 647 bucket->ucb_bucket = b; 648 bucket->ucb_cnt = b->ub_cnt; 649 bucket->ucb_entries = b->ub_entries; 650 } 651 652 static inline void 653 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b) 654 { 655 656 cache_bucket_load(&cache->uc_allocbucket, b); 657 } 658 659 static inline void 660 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b) 661 { 662 663 cache_bucket_load(&cache->uc_freebucket, b); 664 } 665 666 #ifdef NUMA 667 static inline void 668 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b) 669 { 670 671 cache_bucket_load(&cache->uc_crossbucket, b); 672 } 673 #endif 674 675 /* 676 * Copy and preserve ucb_spare. 677 */ 678 static inline void 679 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2) 680 { 681 682 b1->ucb_bucket = b2->ucb_bucket; 683 b1->ucb_entries = b2->ucb_entries; 684 b1->ucb_cnt = b2->ucb_cnt; 685 } 686 687 /* 688 * Swap two cache buckets. 689 */ 690 static inline void 691 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2) 692 { 693 struct uma_cache_bucket b3; 694 695 CRITICAL_ASSERT(curthread); 696 697 cache_bucket_copy(&b3, b1); 698 cache_bucket_copy(b1, b2); 699 cache_bucket_copy(b2, &b3); 700 } 701 702 static void 703 zone_log_warning(uma_zone_t zone) 704 { 705 static const struct timeval warninterval = { 300, 0 }; 706 707 if (!zone_warnings || zone->uz_warning == NULL) 708 return; 709 710 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 711 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 712 } 713 714 static inline void 715 zone_maxaction(uma_zone_t zone) 716 { 717 718 if (zone->uz_maxaction.ta_func != NULL) 719 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); 720 } 721 722 /* 723 * Routine called by timeout which is used to fire off some time interval 724 * based calculations. (stats, hash size, etc.) 725 * 726 * Arguments: 727 * arg Unused 728 * 729 * Returns: 730 * Nothing 731 */ 732 static void 733 uma_timeout(void *unused) 734 { 735 bucket_enable(); 736 zone_foreach(zone_timeout, NULL); 737 738 /* Reschedule this event */ 739 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 740 } 741 742 /* 743 * Update the working set size estimate for the zone's bucket cache. 744 * The constants chosen here are somewhat arbitrary. With an update period of 745 * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the 746 * last 100s. 747 */ 748 static void 749 zone_domain_update_wss(uma_zone_domain_t zdom) 750 { 751 long wss; 752 753 MPASS(zdom->uzd_imax >= zdom->uzd_imin); 754 wss = zdom->uzd_imax - zdom->uzd_imin; 755 zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems; 756 zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5; 757 } 758 759 /* 760 * Routine to perform timeout driven calculations. This expands the 761 * hashes and does per cpu statistics aggregation. 762 * 763 * Returns nothing. 764 */ 765 static void 766 zone_timeout(uma_zone_t zone, void *unused) 767 { 768 uma_keg_t keg; 769 u_int slabs, pages; 770 771 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0) 772 goto update_wss; 773 774 keg = zone->uz_keg; 775 776 /* 777 * Hash zones are non-numa by definition so the first domain 778 * is the only one present. 779 */ 780 KEG_LOCK(keg, 0); 781 pages = keg->uk_domain[0].ud_pages; 782 783 /* 784 * Expand the keg hash table. 785 * 786 * This is done if the number of slabs is larger than the hash size. 787 * What I'm trying to do here is completely reduce collisions. This 788 * may be a little aggressive. Should I allow for two collisions max? 789 */ 790 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) { 791 struct uma_hash newhash; 792 struct uma_hash oldhash; 793 int ret; 794 795 /* 796 * This is so involved because allocating and freeing 797 * while the keg lock is held will lead to deadlock. 798 * I have to do everything in stages and check for 799 * races. 800 */ 801 KEG_UNLOCK(keg, 0); 802 ret = hash_alloc(&newhash, 1 << fls(slabs)); 803 KEG_LOCK(keg, 0); 804 if (ret) { 805 if (hash_expand(&keg->uk_hash, &newhash)) { 806 oldhash = keg->uk_hash; 807 keg->uk_hash = newhash; 808 } else 809 oldhash = newhash; 810 811 KEG_UNLOCK(keg, 0); 812 hash_free(&oldhash); 813 goto update_wss; 814 } 815 } 816 KEG_UNLOCK(keg, 0); 817 818 update_wss: 819 ZONE_LOCK(zone); 820 for (int i = 0; i < vm_ndomains; i++) 821 zone_domain_update_wss(&zone->uz_domain[i]); 822 ZONE_UNLOCK(zone); 823 } 824 825 /* 826 * Allocate and zero fill the next sized hash table from the appropriate 827 * backing store. 828 * 829 * Arguments: 830 * hash A new hash structure with the old hash size in uh_hashsize 831 * 832 * Returns: 833 * 1 on success and 0 on failure. 834 */ 835 static int 836 hash_alloc(struct uma_hash *hash, u_int size) 837 { 838 size_t alloc; 839 840 KASSERT(powerof2(size), ("hash size must be power of 2")); 841 if (size > UMA_HASH_SIZE_INIT) { 842 hash->uh_hashsize = size; 843 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 844 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT); 845 } else { 846 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 847 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 848 UMA_ANYDOMAIN, M_WAITOK); 849 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 850 } 851 if (hash->uh_slab_hash) { 852 bzero(hash->uh_slab_hash, alloc); 853 hash->uh_hashmask = hash->uh_hashsize - 1; 854 return (1); 855 } 856 857 return (0); 858 } 859 860 /* 861 * Expands the hash table for HASH zones. This is done from zone_timeout 862 * to reduce collisions. This must not be done in the regular allocation 863 * path, otherwise, we can recurse on the vm while allocating pages. 864 * 865 * Arguments: 866 * oldhash The hash you want to expand 867 * newhash The hash structure for the new table 868 * 869 * Returns: 870 * Nothing 871 * 872 * Discussion: 873 */ 874 static int 875 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 876 { 877 uma_hash_slab_t slab; 878 u_int hval; 879 u_int idx; 880 881 if (!newhash->uh_slab_hash) 882 return (0); 883 884 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 885 return (0); 886 887 /* 888 * I need to investigate hash algorithms for resizing without a 889 * full rehash. 890 */ 891 892 for (idx = 0; idx < oldhash->uh_hashsize; idx++) 893 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) { 894 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]); 895 LIST_REMOVE(slab, uhs_hlink); 896 hval = UMA_HASH(newhash, slab->uhs_data); 897 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 898 slab, uhs_hlink); 899 } 900 901 return (1); 902 } 903 904 /* 905 * Free the hash bucket to the appropriate backing store. 906 * 907 * Arguments: 908 * slab_hash The hash bucket we're freeing 909 * hashsize The number of entries in that hash bucket 910 * 911 * Returns: 912 * Nothing 913 */ 914 static void 915 hash_free(struct uma_hash *hash) 916 { 917 if (hash->uh_slab_hash == NULL) 918 return; 919 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 920 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 921 else 922 free(hash->uh_slab_hash, M_UMAHASH); 923 } 924 925 /* 926 * Frees all outstanding items in a bucket 927 * 928 * Arguments: 929 * zone The zone to free to, must be unlocked. 930 * bucket The free/alloc bucket with items. 931 * 932 * Returns: 933 * Nothing 934 */ 935 936 static void 937 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 938 { 939 int i; 940 941 if (bucket == NULL || bucket->ub_cnt == 0) 942 return; 943 944 if (zone->uz_fini) 945 for (i = 0; i < bucket->ub_cnt; i++) 946 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 947 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 948 if (zone->uz_max_items > 0) 949 zone_free_limit(zone, bucket->ub_cnt); 950 bucket->ub_cnt = 0; 951 } 952 953 /* 954 * Drains the per cpu caches for a zone. 955 * 956 * NOTE: This may only be called while the zone is being torn down, and not 957 * during normal operation. This is necessary in order that we do not have 958 * to migrate CPUs to drain the per-CPU caches. 959 * 960 * Arguments: 961 * zone The zone to drain, must be unlocked. 962 * 963 * Returns: 964 * Nothing 965 */ 966 static void 967 cache_drain(uma_zone_t zone) 968 { 969 uma_cache_t cache; 970 uma_bucket_t bucket; 971 int cpu; 972 973 /* 974 * XXX: It is safe to not lock the per-CPU caches, because we're 975 * tearing down the zone anyway. I.e., there will be no further use 976 * of the caches at this point. 977 * 978 * XXX: It would good to be able to assert that the zone is being 979 * torn down to prevent improper use of cache_drain(). 980 */ 981 CPU_FOREACH(cpu) { 982 cache = &zone->uz_cpu[cpu]; 983 bucket = cache_bucket_unload_alloc(cache); 984 if (bucket != NULL) { 985 bucket_drain(zone, bucket); 986 bucket_free(zone, bucket, NULL); 987 } 988 bucket = cache_bucket_unload_free(cache); 989 if (bucket != NULL) { 990 bucket_drain(zone, bucket); 991 bucket_free(zone, bucket, NULL); 992 } 993 bucket = cache_bucket_unload_cross(cache); 994 if (bucket != NULL) { 995 bucket_drain(zone, bucket); 996 bucket_free(zone, bucket, NULL); 997 } 998 } 999 bucket_cache_reclaim(zone, true); 1000 } 1001 1002 static void 1003 cache_shrink(uma_zone_t zone, void *unused) 1004 { 1005 1006 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 1007 return; 1008 1009 ZONE_LOCK(zone); 1010 zone->uz_bucket_size = 1011 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2; 1012 ZONE_UNLOCK(zone); 1013 } 1014 1015 static void 1016 cache_drain_safe_cpu(uma_zone_t zone, void *unused) 1017 { 1018 uma_cache_t cache; 1019 uma_bucket_t b1, b2, b3; 1020 int domain; 1021 1022 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 1023 return; 1024 1025 b1 = b2 = b3 = NULL; 1026 ZONE_LOCK(zone); 1027 critical_enter(); 1028 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) 1029 domain = PCPU_GET(domain); 1030 else 1031 domain = 0; 1032 cache = &zone->uz_cpu[curcpu]; 1033 b1 = cache_bucket_unload_alloc(cache); 1034 if (b1 != NULL && b1->ub_cnt != 0) { 1035 zone_put_bucket(zone, &zone->uz_domain[domain], b1, false); 1036 b1 = NULL; 1037 } 1038 b2 = cache_bucket_unload_free(cache); 1039 if (b2 != NULL && b2->ub_cnt != 0) { 1040 zone_put_bucket(zone, &zone->uz_domain[domain], b2, false); 1041 b2 = NULL; 1042 } 1043 b3 = cache_bucket_unload_cross(cache); 1044 critical_exit(); 1045 ZONE_UNLOCK(zone); 1046 if (b1) 1047 bucket_free(zone, b1, NULL); 1048 if (b2) 1049 bucket_free(zone, b2, NULL); 1050 if (b3) { 1051 bucket_drain(zone, b3); 1052 bucket_free(zone, b3, NULL); 1053 } 1054 } 1055 1056 /* 1057 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 1058 * This is an expensive call because it needs to bind to all CPUs 1059 * one by one and enter a critical section on each of them in order 1060 * to safely access their cache buckets. 1061 * Zone lock must not be held on call this function. 1062 */ 1063 static void 1064 pcpu_cache_drain_safe(uma_zone_t zone) 1065 { 1066 int cpu; 1067 1068 /* 1069 * Polite bucket sizes shrinking was not enough, shrink aggressively. 1070 */ 1071 if (zone) 1072 cache_shrink(zone, NULL); 1073 else 1074 zone_foreach(cache_shrink, NULL); 1075 1076 CPU_FOREACH(cpu) { 1077 thread_lock(curthread); 1078 sched_bind(curthread, cpu); 1079 thread_unlock(curthread); 1080 1081 if (zone) 1082 cache_drain_safe_cpu(zone, NULL); 1083 else 1084 zone_foreach(cache_drain_safe_cpu, NULL); 1085 } 1086 thread_lock(curthread); 1087 sched_unbind(curthread); 1088 thread_unlock(curthread); 1089 } 1090 1091 /* 1092 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller 1093 * requested a drain, otherwise the per-domain caches are trimmed to either 1094 * estimated working set size. 1095 */ 1096 static void 1097 bucket_cache_reclaim(uma_zone_t zone, bool drain) 1098 { 1099 uma_zone_domain_t zdom; 1100 uma_bucket_t bucket; 1101 long target, tofree; 1102 int i; 1103 1104 for (i = 0; i < vm_ndomains; i++) { 1105 /* 1106 * The cross bucket is partially filled and not part of 1107 * the item count. Reclaim it individually here. 1108 */ 1109 zdom = &zone->uz_domain[i]; 1110 ZONE_CROSS_LOCK(zone); 1111 bucket = zdom->uzd_cross; 1112 zdom->uzd_cross = NULL; 1113 ZONE_CROSS_UNLOCK(zone); 1114 if (bucket != NULL) { 1115 bucket_drain(zone, bucket); 1116 bucket_free(zone, bucket, NULL); 1117 } 1118 1119 /* 1120 * Shrink the zone bucket size to ensure that the per-CPU caches 1121 * don't grow too large. 1122 */ 1123 ZONE_LOCK(zone); 1124 if (i == 0 && zone->uz_bucket_size > zone->uz_bucket_size_min) 1125 zone->uz_bucket_size--; 1126 1127 /* 1128 * If we were asked to drain the zone, we are done only once 1129 * this bucket cache is empty. Otherwise, we reclaim items in 1130 * excess of the zone's estimated working set size. If the 1131 * difference nitems - imin is larger than the WSS estimate, 1132 * then the estimate will grow at the end of this interval and 1133 * we ignore the historical average. 1134 */ 1135 target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems - 1136 zdom->uzd_imin); 1137 while (zdom->uzd_nitems > target) { 1138 bucket = TAILQ_LAST(&zdom->uzd_buckets, uma_bucketlist); 1139 if (bucket == NULL) 1140 break; 1141 tofree = bucket->ub_cnt; 1142 TAILQ_REMOVE(&zdom->uzd_buckets, bucket, ub_link); 1143 zdom->uzd_nitems -= tofree; 1144 1145 /* 1146 * Shift the bounds of the current WSS interval to avoid 1147 * perturbing the estimate. 1148 */ 1149 zdom->uzd_imax -= lmin(zdom->uzd_imax, tofree); 1150 zdom->uzd_imin -= lmin(zdom->uzd_imin, tofree); 1151 1152 ZONE_UNLOCK(zone); 1153 bucket_drain(zone, bucket); 1154 bucket_free(zone, bucket, NULL); 1155 ZONE_LOCK(zone); 1156 } 1157 ZONE_UNLOCK(zone); 1158 } 1159 } 1160 1161 static void 1162 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 1163 { 1164 uint8_t *mem; 1165 int i; 1166 uint8_t flags; 1167 1168 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes", 1169 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera); 1170 1171 mem = slab_data(slab, keg); 1172 flags = slab->us_flags; 1173 i = start; 1174 if (keg->uk_fini != NULL) { 1175 for (i--; i > -1; i--) 1176 #ifdef INVARIANTS 1177 /* 1178 * trash_fini implies that dtor was trash_dtor. trash_fini 1179 * would check that memory hasn't been modified since free, 1180 * which executed trash_dtor. 1181 * That's why we need to run uma_dbg_kskip() check here, 1182 * albeit we don't make skip check for other init/fini 1183 * invocations. 1184 */ 1185 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) || 1186 keg->uk_fini != trash_fini) 1187 #endif 1188 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size); 1189 } 1190 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) 1191 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab), 1192 NULL, SKIP_NONE); 1193 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 1194 uma_total_dec(PAGE_SIZE * keg->uk_ppera); 1195 } 1196 1197 /* 1198 * Frees pages from a keg back to the system. This is done on demand from 1199 * the pageout daemon. 1200 * 1201 * Returns nothing. 1202 */ 1203 static void 1204 keg_drain(uma_keg_t keg) 1205 { 1206 struct slabhead freeslabs = { 0 }; 1207 uma_domain_t dom; 1208 uma_slab_t slab, tmp; 1209 int i, n; 1210 1211 /* 1212 * We don't want to take pages from statically allocated kegs at this 1213 * time 1214 */ 1215 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 1216 return; 1217 1218 for (i = 0; i < vm_ndomains; i++) { 1219 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u", 1220 keg->uk_name, keg, i, dom->ud_free); 1221 n = 0; 1222 dom = &keg->uk_domain[i]; 1223 KEG_LOCK(keg, i); 1224 LIST_FOREACH_SAFE(slab, &dom->ud_free_slab, us_link, tmp) { 1225 if (keg->uk_flags & UMA_ZFLAG_HASH) 1226 UMA_HASH_REMOVE(&keg->uk_hash, slab); 1227 n++; 1228 LIST_REMOVE(slab, us_link); 1229 LIST_INSERT_HEAD(&freeslabs, slab, us_link); 1230 } 1231 dom->ud_pages -= n * keg->uk_ppera; 1232 dom->ud_free -= n * keg->uk_ipers; 1233 KEG_UNLOCK(keg, i); 1234 } 1235 1236 while ((slab = LIST_FIRST(&freeslabs)) != NULL) { 1237 LIST_REMOVE(slab, us_link); 1238 keg_free_slab(keg, slab, keg->uk_ipers); 1239 } 1240 } 1241 1242 static void 1243 zone_reclaim(uma_zone_t zone, int waitok, bool drain) 1244 { 1245 1246 /* 1247 * Set draining to interlock with zone_dtor() so we can release our 1248 * locks as we go. Only dtor() should do a WAITOK call since it 1249 * is the only call that knows the structure will still be available 1250 * when it wakes up. 1251 */ 1252 ZONE_LOCK(zone); 1253 while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) { 1254 if (waitok == M_NOWAIT) 1255 goto out; 1256 msleep(zone, &zone->uz_lock, PVM, "zonedrain", 1); 1257 } 1258 zone->uz_flags |= UMA_ZFLAG_RECLAIMING; 1259 ZONE_UNLOCK(zone); 1260 bucket_cache_reclaim(zone, drain); 1261 1262 /* 1263 * The DRAINING flag protects us from being freed while 1264 * we're running. Normally the uma_rwlock would protect us but we 1265 * must be able to release and acquire the right lock for each keg. 1266 */ 1267 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) 1268 keg_drain(zone->uz_keg); 1269 ZONE_LOCK(zone); 1270 zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING; 1271 wakeup(zone); 1272 out: 1273 ZONE_UNLOCK(zone); 1274 } 1275 1276 static void 1277 zone_drain(uma_zone_t zone, void *unused) 1278 { 1279 1280 zone_reclaim(zone, M_NOWAIT, true); 1281 } 1282 1283 static void 1284 zone_trim(uma_zone_t zone, void *unused) 1285 { 1286 1287 zone_reclaim(zone, M_NOWAIT, false); 1288 } 1289 1290 /* 1291 * Allocate a new slab for a keg and inserts it into the partial slab list. 1292 * The keg should be unlocked on entry. If the allocation succeeds it will 1293 * be locked on return. 1294 * 1295 * Arguments: 1296 * flags Wait flags for the item initialization routine 1297 * aflags Wait flags for the slab allocation 1298 * 1299 * Returns: 1300 * The slab that was allocated or NULL if there is no memory and the 1301 * caller specified M_NOWAIT. 1302 */ 1303 static uma_slab_t 1304 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags, 1305 int aflags) 1306 { 1307 uma_domain_t dom; 1308 uma_alloc allocf; 1309 uma_slab_t slab; 1310 unsigned long size; 1311 uint8_t *mem; 1312 uint8_t sflags; 1313 int i; 1314 1315 KASSERT(domain >= 0 && domain < vm_ndomains, 1316 ("keg_alloc_slab: domain %d out of range", domain)); 1317 1318 allocf = keg->uk_allocf; 1319 slab = NULL; 1320 mem = NULL; 1321 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) { 1322 uma_hash_slab_t hslab; 1323 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL, 1324 domain, aflags); 1325 if (hslab == NULL) 1326 goto fail; 1327 slab = &hslab->uhs_slab; 1328 } 1329 1330 /* 1331 * This reproduces the old vm_zone behavior of zero filling pages the 1332 * first time they are added to a zone. 1333 * 1334 * Malloced items are zeroed in uma_zalloc. 1335 */ 1336 1337 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 1338 aflags |= M_ZERO; 1339 else 1340 aflags &= ~M_ZERO; 1341 1342 if (keg->uk_flags & UMA_ZONE_NODUMP) 1343 aflags |= M_NODUMP; 1344 1345 /* zone is passed for legacy reasons. */ 1346 size = keg->uk_ppera * PAGE_SIZE; 1347 mem = allocf(zone, size, domain, &sflags, aflags); 1348 if (mem == NULL) { 1349 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) 1350 zone_free_item(slabzone(keg->uk_ipers), 1351 slab_tohashslab(slab), NULL, SKIP_NONE); 1352 goto fail; 1353 } 1354 uma_total_inc(size); 1355 1356 /* For HASH zones all pages go to the same uma_domain. */ 1357 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) 1358 domain = 0; 1359 1360 /* Point the slab into the allocated memory */ 1361 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) 1362 slab = (uma_slab_t )(mem + keg->uk_pgoff); 1363 else 1364 slab_tohashslab(slab)->uhs_data = mem; 1365 1366 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB) 1367 for (i = 0; i < keg->uk_ppera; i++) 1368 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE), 1369 zone, slab); 1370 1371 slab->us_freecount = keg->uk_ipers; 1372 slab->us_flags = sflags; 1373 slab->us_domain = domain; 1374 1375 BIT_FILL(keg->uk_ipers, &slab->us_free); 1376 #ifdef INVARIANTS 1377 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg)); 1378 #endif 1379 1380 if (keg->uk_init != NULL) { 1381 for (i = 0; i < keg->uk_ipers; i++) 1382 if (keg->uk_init(slab_item(slab, keg, i), 1383 keg->uk_size, flags) != 0) 1384 break; 1385 if (i != keg->uk_ipers) { 1386 keg_free_slab(keg, slab, i); 1387 goto fail; 1388 } 1389 } 1390 KEG_LOCK(keg, domain); 1391 1392 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)", 1393 slab, keg->uk_name, keg); 1394 1395 if (keg->uk_flags & UMA_ZFLAG_HASH) 1396 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1397 1398 /* 1399 * If we got a slab here it's safe to mark it partially used 1400 * and return. We assume that the caller is going to remove 1401 * at least one item. 1402 */ 1403 dom = &keg->uk_domain[domain]; 1404 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 1405 dom->ud_pages += keg->uk_ppera; 1406 dom->ud_free += keg->uk_ipers; 1407 1408 return (slab); 1409 1410 fail: 1411 return (NULL); 1412 } 1413 1414 /* 1415 * This function is intended to be used early on in place of page_alloc() so 1416 * that we may use the boot time page cache to satisfy allocations before 1417 * the VM is ready. 1418 */ 1419 static void * 1420 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1421 int wait) 1422 { 1423 vm_paddr_t pa; 1424 vm_page_t m; 1425 void *mem; 1426 int pages; 1427 int i; 1428 1429 pages = howmany(bytes, PAGE_SIZE); 1430 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__)); 1431 1432 *pflag = UMA_SLAB_BOOT; 1433 m = vm_page_alloc_contig_domain(NULL, 0, domain, 1434 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages, 1435 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT); 1436 if (m == NULL) 1437 return (NULL); 1438 1439 pa = VM_PAGE_TO_PHYS(m); 1440 for (i = 0; i < pages; i++, pa += PAGE_SIZE) { 1441 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 1442 defined(__riscv) || defined(__powerpc64__) 1443 if ((wait & M_NODUMP) == 0) 1444 dump_add_page(pa); 1445 #endif 1446 } 1447 /* Allocate KVA and indirectly advance bootmem. */ 1448 mem = (void *)pmap_map(&bootmem, m->phys_addr, 1449 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE); 1450 if ((wait & M_ZERO) != 0) 1451 bzero(mem, pages * PAGE_SIZE); 1452 1453 return (mem); 1454 } 1455 1456 static void 1457 startup_free(void *mem, vm_size_t bytes) 1458 { 1459 vm_offset_t va; 1460 vm_page_t m; 1461 1462 va = (vm_offset_t)mem; 1463 m = PHYS_TO_VM_PAGE(pmap_kextract(va)); 1464 pmap_remove(kernel_pmap, va, va + bytes); 1465 for (; bytes != 0; bytes -= PAGE_SIZE, m++) { 1466 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 1467 defined(__riscv) || defined(__powerpc64__) 1468 dump_drop_page(VM_PAGE_TO_PHYS(m)); 1469 #endif 1470 vm_page_unwire_noq(m); 1471 vm_page_free(m); 1472 } 1473 } 1474 1475 /* 1476 * Allocates a number of pages from the system 1477 * 1478 * Arguments: 1479 * bytes The number of bytes requested 1480 * wait Shall we wait? 1481 * 1482 * Returns: 1483 * A pointer to the alloced memory or possibly 1484 * NULL if M_NOWAIT is set. 1485 */ 1486 static void * 1487 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1488 int wait) 1489 { 1490 void *p; /* Returned page */ 1491 1492 *pflag = UMA_SLAB_KERNEL; 1493 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait); 1494 1495 return (p); 1496 } 1497 1498 static void * 1499 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1500 int wait) 1501 { 1502 struct pglist alloctail; 1503 vm_offset_t addr, zkva; 1504 int cpu, flags; 1505 vm_page_t p, p_next; 1506 #ifdef NUMA 1507 struct pcpu *pc; 1508 #endif 1509 1510 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE); 1511 1512 TAILQ_INIT(&alloctail); 1513 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1514 malloc2vm_flags(wait); 1515 *pflag = UMA_SLAB_KERNEL; 1516 for (cpu = 0; cpu <= mp_maxid; cpu++) { 1517 if (CPU_ABSENT(cpu)) { 1518 p = vm_page_alloc(NULL, 0, flags); 1519 } else { 1520 #ifndef NUMA 1521 p = vm_page_alloc(NULL, 0, flags); 1522 #else 1523 pc = pcpu_find(cpu); 1524 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain))) 1525 p = NULL; 1526 else 1527 p = vm_page_alloc_domain(NULL, 0, 1528 pc->pc_domain, flags); 1529 if (__predict_false(p == NULL)) 1530 p = vm_page_alloc(NULL, 0, flags); 1531 #endif 1532 } 1533 if (__predict_false(p == NULL)) 1534 goto fail; 1535 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1536 } 1537 if ((addr = kva_alloc(bytes)) == 0) 1538 goto fail; 1539 zkva = addr; 1540 TAILQ_FOREACH(p, &alloctail, listq) { 1541 pmap_qenter(zkva, &p, 1); 1542 zkva += PAGE_SIZE; 1543 } 1544 return ((void*)addr); 1545 fail: 1546 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1547 vm_page_unwire_noq(p); 1548 vm_page_free(p); 1549 } 1550 return (NULL); 1551 } 1552 1553 /* 1554 * Allocates a number of pages from within an object 1555 * 1556 * Arguments: 1557 * bytes The number of bytes requested 1558 * wait Shall we wait? 1559 * 1560 * Returns: 1561 * A pointer to the alloced memory or possibly 1562 * NULL if M_NOWAIT is set. 1563 */ 1564 static void * 1565 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, 1566 int wait) 1567 { 1568 TAILQ_HEAD(, vm_page) alloctail; 1569 u_long npages; 1570 vm_offset_t retkva, zkva; 1571 vm_page_t p, p_next; 1572 uma_keg_t keg; 1573 1574 TAILQ_INIT(&alloctail); 1575 keg = zone->uz_keg; 1576 1577 npages = howmany(bytes, PAGE_SIZE); 1578 while (npages > 0) { 1579 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT | 1580 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1581 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK : 1582 VM_ALLOC_NOWAIT)); 1583 if (p != NULL) { 1584 /* 1585 * Since the page does not belong to an object, its 1586 * listq is unused. 1587 */ 1588 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1589 npages--; 1590 continue; 1591 } 1592 /* 1593 * Page allocation failed, free intermediate pages and 1594 * exit. 1595 */ 1596 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1597 vm_page_unwire_noq(p); 1598 vm_page_free(p); 1599 } 1600 return (NULL); 1601 } 1602 *flags = UMA_SLAB_PRIV; 1603 zkva = keg->uk_kva + 1604 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1605 retkva = zkva; 1606 TAILQ_FOREACH(p, &alloctail, listq) { 1607 pmap_qenter(zkva, &p, 1); 1608 zkva += PAGE_SIZE; 1609 } 1610 1611 return ((void *)retkva); 1612 } 1613 1614 /* 1615 * Frees a number of pages to the system 1616 * 1617 * Arguments: 1618 * mem A pointer to the memory to be freed 1619 * size The size of the memory being freed 1620 * flags The original p->us_flags field 1621 * 1622 * Returns: 1623 * Nothing 1624 */ 1625 static void 1626 page_free(void *mem, vm_size_t size, uint8_t flags) 1627 { 1628 1629 if ((flags & UMA_SLAB_BOOT) != 0) { 1630 startup_free(mem, size); 1631 return; 1632 } 1633 1634 if ((flags & UMA_SLAB_KERNEL) == 0) 1635 panic("UMA: page_free used with invalid flags %x", flags); 1636 1637 kmem_free((vm_offset_t)mem, size); 1638 } 1639 1640 /* 1641 * Frees pcpu zone allocations 1642 * 1643 * Arguments: 1644 * mem A pointer to the memory to be freed 1645 * size The size of the memory being freed 1646 * flags The original p->us_flags field 1647 * 1648 * Returns: 1649 * Nothing 1650 */ 1651 static void 1652 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags) 1653 { 1654 vm_offset_t sva, curva; 1655 vm_paddr_t paddr; 1656 vm_page_t m; 1657 1658 MPASS(size == (mp_maxid+1)*PAGE_SIZE); 1659 sva = (vm_offset_t)mem; 1660 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) { 1661 paddr = pmap_kextract(curva); 1662 m = PHYS_TO_VM_PAGE(paddr); 1663 vm_page_unwire_noq(m); 1664 vm_page_free(m); 1665 } 1666 pmap_qremove(sva, size >> PAGE_SHIFT); 1667 kva_free(sva, size); 1668 } 1669 1670 1671 /* 1672 * Zero fill initializer 1673 * 1674 * Arguments/Returns follow uma_init specifications 1675 */ 1676 static int 1677 zero_init(void *mem, int size, int flags) 1678 { 1679 bzero(mem, size); 1680 return (0); 1681 } 1682 1683 #ifdef INVARIANTS 1684 struct noslabbits * 1685 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg) 1686 { 1687 1688 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers))); 1689 } 1690 #endif 1691 1692 /* 1693 * Actual size of embedded struct slab (!OFFPAGE). 1694 */ 1695 size_t 1696 slab_sizeof(int nitems) 1697 { 1698 size_t s; 1699 1700 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS; 1701 return (roundup(s, UMA_ALIGN_PTR + 1)); 1702 } 1703 1704 /* 1705 * Size of memory for embedded slabs (!OFFPAGE). 1706 */ 1707 size_t 1708 slab_space(int nitems) 1709 { 1710 return (UMA_SLAB_SIZE - slab_sizeof(nitems)); 1711 } 1712 1713 #define UMA_FIXPT_SHIFT 31 1714 #define UMA_FRAC_FIXPT(n, d) \ 1715 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d))) 1716 #define UMA_FIXPT_PCT(f) \ 1717 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT)) 1718 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100) 1719 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE) 1720 1721 /* 1722 * Compute the number of items that will fit in a slab. If hdr is true, the 1723 * item count may be limited to provide space in the slab for an inline slab 1724 * header. Otherwise, all slab space will be provided for item storage. 1725 */ 1726 static u_int 1727 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr) 1728 { 1729 u_int ipers; 1730 u_int padpi; 1731 1732 /* The padding between items is not needed after the last item. */ 1733 padpi = rsize - size; 1734 1735 if (hdr) { 1736 /* 1737 * Start with the maximum item count and remove items until 1738 * the slab header first alongside the allocatable memory. 1739 */ 1740 for (ipers = MIN(SLAB_MAX_SETSIZE, 1741 (slabsize + padpi - slab_sizeof(1)) / rsize); 1742 ipers > 0 && 1743 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize; 1744 ipers--) 1745 continue; 1746 } else { 1747 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE); 1748 } 1749 1750 return (ipers); 1751 } 1752 1753 /* 1754 * Compute the number of items that will fit in a slab for a startup zone. 1755 */ 1756 int 1757 slab_ipers(size_t size, int align) 1758 { 1759 int rsize; 1760 1761 rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */ 1762 return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true)); 1763 } 1764 1765 /* 1766 * Determine the format of a uma keg. This determines where the slab header 1767 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera. 1768 * 1769 * Arguments 1770 * keg The zone we should initialize 1771 * 1772 * Returns 1773 * Nothing 1774 */ 1775 static void 1776 keg_layout(uma_keg_t keg) 1777 { 1778 u_int alignsize; 1779 u_int eff; 1780 u_int eff_offpage; 1781 u_int format; 1782 u_int ipers; 1783 u_int ipers_offpage; 1784 u_int pages; 1785 u_int rsize; 1786 u_int slabsize; 1787 1788 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1789 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE && 1790 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0), 1791 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b", 1792 __func__, keg->uk_name, keg->uk_size, keg->uk_flags, 1793 PRINT_UMA_ZFLAGS)); 1794 KASSERT((keg->uk_flags & 1795 (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY)) == 0 || 1796 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0, 1797 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags, 1798 PRINT_UMA_ZFLAGS)); 1799 1800 alignsize = keg->uk_align + 1; 1801 format = 0; 1802 ipers = 0; 1803 1804 /* 1805 * Calculate the size of each allocation (rsize) according to 1806 * alignment. If the requested size is smaller than we have 1807 * allocation bits for we round it up. 1808 */ 1809 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT); 1810 rsize = roundup2(rsize, alignsize); 1811 1812 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) { 1813 slabsize = UMA_PCPU_ALLOC_SIZE; 1814 pages = mp_maxid + 1; 1815 } else if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) { 1816 /* 1817 * We want one item to start on every align boundary in a page. 1818 * To do this we will span pages. We will also extend the item 1819 * by the size of align if it is an even multiple of align. 1820 * Otherwise, it would fall on the same boundary every time. 1821 */ 1822 if ((rsize & alignsize) == 0) 1823 rsize += alignsize; 1824 slabsize = rsize * (PAGE_SIZE / alignsize); 1825 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE); 1826 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE); 1827 pages = howmany(slabsize, PAGE_SIZE); 1828 slabsize = ptoa(pages); 1829 } else { 1830 /* 1831 * Choose a slab size of as many pages as it takes to represent 1832 * a single item. We will then try to fit as many additional 1833 * items into the slab as possible. At some point, we may want 1834 * to increase the slab size for awkward item sizes in order to 1835 * increase efficiency. 1836 */ 1837 pages = howmany(keg->uk_size, PAGE_SIZE); 1838 slabsize = ptoa(pages); 1839 } 1840 1841 /* Evaluate an inline slab layout. */ 1842 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0) 1843 ipers = slab_ipers_hdr(keg->uk_size, rsize, slabsize, true); 1844 1845 /* TODO: vm_page-embedded slab. */ 1846 1847 /* 1848 * We can't do OFFPAGE if we're internal or if we've been 1849 * asked to not go to the VM for buckets. If we do this we 1850 * may end up going to the VM for slabs which we do not 1851 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1852 * of UMA_ZONE_VM, which clearly forbids it. 1853 */ 1854 if ((keg->uk_flags & 1855 (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY)) != 0) { 1856 if (ipers == 0) { 1857 /* We need an extra page for the slab header. */ 1858 pages++; 1859 slabsize = ptoa(pages); 1860 ipers = slab_ipers_hdr(keg->uk_size, rsize, slabsize, 1861 true); 1862 } 1863 goto out; 1864 } 1865 1866 /* 1867 * See if using an OFFPAGE slab will improve our efficiency. 1868 * Only do this if we are below our efficiency threshold. 1869 * 1870 * XXX We could try growing slabsize to limit max waste as well. 1871 * Historically this was not done because the VM could not 1872 * efficiently handle contiguous allocations. 1873 */ 1874 eff = UMA_FRAC_FIXPT(ipers * rsize, slabsize); 1875 ipers_offpage = slab_ipers_hdr(keg->uk_size, rsize, slabsize, false); 1876 eff_offpage = UMA_FRAC_FIXPT(ipers_offpage * rsize, 1877 slabsize + slabzone(ipers_offpage)->uz_keg->uk_rsize); 1878 if (ipers == 0 || (eff < UMA_MIN_EFF && eff < eff_offpage)) { 1879 CTR5(KTR_UMA, "UMA decided we need offpage slab headers for " 1880 "keg: %s(%p), minimum efficiency allowed = %u%%, " 1881 "old efficiency = %u%%, offpage efficiency = %u%%", 1882 keg->uk_name, keg, UMA_FIXPT_PCT(UMA_MIN_EFF), 1883 UMA_FIXPT_PCT(eff), UMA_FIXPT_PCT(eff_offpage)); 1884 format = UMA_ZFLAG_OFFPAGE; 1885 ipers = ipers_offpage; 1886 } 1887 1888 out: 1889 /* 1890 * How do we find the slab header if it is offpage or if not all item 1891 * start addresses are in the same page? We could solve the latter 1892 * case with vaddr alignment, but we don't. 1893 */ 1894 if ((format & UMA_ZFLAG_OFFPAGE) != 0 || 1895 (ipers - 1) * rsize >= PAGE_SIZE) { 1896 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0) 1897 format |= UMA_ZFLAG_HASH; 1898 else 1899 format |= UMA_ZFLAG_VTOSLAB; 1900 } 1901 keg->uk_ipers = ipers; 1902 keg->uk_rsize = rsize; 1903 keg->uk_flags |= format; 1904 keg->uk_ppera = pages; 1905 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u", 1906 __func__, keg->uk_name, keg->uk_flags, rsize, ipers, pages); 1907 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE, 1908 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__, 1909 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize, ipers, 1910 pages)); 1911 } 1912 1913 /* 1914 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1915 * the keg onto the global keg list. 1916 * 1917 * Arguments/Returns follow uma_ctor specifications 1918 * udata Actually uma_kctor_args 1919 */ 1920 static int 1921 keg_ctor(void *mem, int size, void *udata, int flags) 1922 { 1923 struct uma_kctor_args *arg = udata; 1924 uma_keg_t keg = mem; 1925 uma_zone_t zone; 1926 int i; 1927 1928 bzero(keg, size); 1929 keg->uk_size = arg->size; 1930 keg->uk_init = arg->uminit; 1931 keg->uk_fini = arg->fini; 1932 keg->uk_align = arg->align; 1933 keg->uk_reserve = 0; 1934 keg->uk_flags = arg->flags; 1935 1936 /* 1937 * We use a global round-robin policy by default. Zones with 1938 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which 1939 * case the iterator is never run. 1940 */ 1941 keg->uk_dr.dr_policy = DOMAINSET_RR(); 1942 keg->uk_dr.dr_iter = 0; 1943 1944 /* 1945 * The master zone is passed to us at keg-creation time. 1946 */ 1947 zone = arg->zone; 1948 keg->uk_name = zone->uz_name; 1949 1950 if (arg->flags & UMA_ZONE_VM) 1951 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1952 1953 if (arg->flags & UMA_ZONE_ZINIT) 1954 keg->uk_init = zero_init; 1955 1956 if (arg->flags & UMA_ZONE_MALLOC) 1957 keg->uk_flags |= UMA_ZFLAG_VTOSLAB; 1958 1959 #ifndef SMP 1960 keg->uk_flags &= ~UMA_ZONE_PCPU; 1961 #endif 1962 1963 keg_layout(keg); 1964 1965 /* 1966 * Use a first-touch NUMA policy for all kegs that pmap_extract() 1967 * will work on with the exception of critical VM structures 1968 * necessary for paging. 1969 * 1970 * Zones may override the default by specifying either. 1971 */ 1972 #ifdef NUMA 1973 if ((keg->uk_flags & 1974 (UMA_ZFLAG_HASH | UMA_ZONE_VM | UMA_ZONE_ROUNDROBIN)) == 0) 1975 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH; 1976 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0) 1977 keg->uk_flags |= UMA_ZONE_ROUNDROBIN; 1978 #endif 1979 1980 /* 1981 * If we haven't booted yet we need allocations to go through the 1982 * startup cache until the vm is ready. 1983 */ 1984 #ifdef UMA_MD_SMALL_ALLOC 1985 if (keg->uk_ppera == 1) 1986 keg->uk_allocf = uma_small_alloc; 1987 else 1988 #endif 1989 if (booted < BOOT_KVA) 1990 keg->uk_allocf = startup_alloc; 1991 else if (keg->uk_flags & UMA_ZONE_PCPU) 1992 keg->uk_allocf = pcpu_page_alloc; 1993 else 1994 keg->uk_allocf = page_alloc; 1995 #ifdef UMA_MD_SMALL_ALLOC 1996 if (keg->uk_ppera == 1) 1997 keg->uk_freef = uma_small_free; 1998 else 1999 #endif 2000 if (keg->uk_flags & UMA_ZONE_PCPU) 2001 keg->uk_freef = pcpu_page_free; 2002 else 2003 keg->uk_freef = page_free; 2004 2005 /* 2006 * Initialize keg's locks. 2007 */ 2008 for (i = 0; i < vm_ndomains; i++) 2009 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS)); 2010 2011 /* 2012 * If we're putting the slab header in the actual page we need to 2013 * figure out where in each page it goes. See slab_sizeof 2014 * definition. 2015 */ 2016 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) { 2017 size_t shsize; 2018 2019 shsize = slab_sizeof(keg->uk_ipers); 2020 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize; 2021 /* 2022 * The only way the following is possible is if with our 2023 * UMA_ALIGN_PTR adjustments we are now bigger than 2024 * UMA_SLAB_SIZE. I haven't checked whether this is 2025 * mathematically possible for all cases, so we make 2026 * sure here anyway. 2027 */ 2028 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera, 2029 ("zone %s ipers %d rsize %d size %d slab won't fit", 2030 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size)); 2031 } 2032 2033 if (keg->uk_flags & UMA_ZFLAG_HASH) 2034 hash_alloc(&keg->uk_hash, 0); 2035 2036 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone); 2037 2038 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 2039 2040 rw_wlock(&uma_rwlock); 2041 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 2042 rw_wunlock(&uma_rwlock); 2043 return (0); 2044 } 2045 2046 static void 2047 zone_kva_available(uma_zone_t zone, void *unused) 2048 { 2049 uma_keg_t keg; 2050 2051 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 2052 return; 2053 KEG_GET(zone, keg); 2054 if (keg->uk_allocf == startup_alloc) 2055 keg->uk_allocf = page_alloc; 2056 } 2057 2058 static void 2059 zone_alloc_counters(uma_zone_t zone, void *unused) 2060 { 2061 2062 zone->uz_allocs = counter_u64_alloc(M_WAITOK); 2063 zone->uz_frees = counter_u64_alloc(M_WAITOK); 2064 zone->uz_fails = counter_u64_alloc(M_WAITOK); 2065 } 2066 2067 static void 2068 zone_alloc_sysctl(uma_zone_t zone, void *unused) 2069 { 2070 uma_zone_domain_t zdom; 2071 uma_domain_t dom; 2072 uma_keg_t keg; 2073 struct sysctl_oid *oid, *domainoid; 2074 int domains, i, cnt; 2075 static const char *nokeg = "cache zone"; 2076 char *c; 2077 2078 /* 2079 * Make a sysctl safe copy of the zone name by removing 2080 * any special characters and handling dups by appending 2081 * an index. 2082 */ 2083 if (zone->uz_namecnt != 0) { 2084 /* Count the number of decimal digits and '_' separator. */ 2085 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++) 2086 cnt /= 10; 2087 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1, 2088 M_UMA, M_WAITOK); 2089 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name, 2090 zone->uz_namecnt); 2091 } else 2092 zone->uz_ctlname = strdup(zone->uz_name, M_UMA); 2093 for (c = zone->uz_ctlname; *c != '\0'; c++) 2094 if (strchr("./\\ -", *c) != NULL) 2095 *c = '_'; 2096 2097 /* 2098 * Basic parameters at the root. 2099 */ 2100 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma), 2101 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD, NULL, ""); 2102 oid = zone->uz_oid; 2103 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2104 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size"); 2105 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2106 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE, 2107 zone, 0, sysctl_handle_uma_zone_flags, "A", 2108 "Allocator configuration flags"); 2109 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2110 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0, 2111 "Desired per-cpu cache size"); 2112 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2113 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0, 2114 "Maximum allowed per-cpu cache size"); 2115 2116 /* 2117 * keg if present. 2118 */ 2119 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0) 2120 domains = vm_ndomains; 2121 else 2122 domains = 1; 2123 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 2124 "keg", CTLFLAG_RD, NULL, ""); 2125 keg = zone->uz_keg; 2126 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) { 2127 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2128 "name", CTLFLAG_RD, keg->uk_name, "Keg name"); 2129 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2130 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0, 2131 "Real object size with alignment"); 2132 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2133 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0, 2134 "pages per-slab allocation"); 2135 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2136 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0, 2137 "items available per-slab"); 2138 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2139 "align", CTLFLAG_RD, &keg->uk_align, 0, 2140 "item alignment mask"); 2141 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2142 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE, 2143 keg, 0, sysctl_handle_uma_slab_efficiency, "I", 2144 "Slab utilization (100 - internal fragmentation %)"); 2145 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid), 2146 OID_AUTO, "domain", CTLFLAG_RD, NULL, ""); 2147 for (i = 0; i < domains; i++) { 2148 dom = &keg->uk_domain[i]; 2149 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid), 2150 OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD, 2151 NULL, ""); 2152 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2153 "pages", CTLFLAG_RD, &dom->ud_pages, 0, 2154 "Total pages currently allocated from VM"); 2155 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2156 "free", CTLFLAG_RD, &dom->ud_free, 0, 2157 "items free in the slab layer"); 2158 } 2159 } else 2160 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2161 "name", CTLFLAG_RD, nokeg, "Keg name"); 2162 2163 /* 2164 * Information about zone limits. 2165 */ 2166 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 2167 "limit", CTLFLAG_RD, NULL, ""); 2168 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2169 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2170 zone, 0, sysctl_handle_uma_zone_items, "QU", 2171 "current number of allocated items if limit is set"); 2172 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2173 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0, 2174 "Maximum number of cached items"); 2175 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2176 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0, 2177 "Number of threads sleeping at limit"); 2178 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2179 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0, 2180 "Total zone limit sleeps"); 2181 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2182 "bucket_max", CTLFLAG_RD, &zone->uz_bkt_max, 0, 2183 "Maximum number of items in the bucket cache"); 2184 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2185 "bucket_cnt", CTLFLAG_RD, &zone->uz_bkt_count, 0, 2186 "Number of items in the bucket cache"); 2187 2188 /* 2189 * Per-domain zone information. 2190 */ 2191 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), 2192 OID_AUTO, "domain", CTLFLAG_RD, NULL, ""); 2193 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0) 2194 domains = 1; 2195 for (i = 0; i < domains; i++) { 2196 zdom = &zone->uz_domain[i]; 2197 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid), 2198 OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD, NULL, ""); 2199 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2200 "nitems", CTLFLAG_RD, &zdom->uzd_nitems, 2201 "number of items in this domain"); 2202 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2203 "imax", CTLFLAG_RD, &zdom->uzd_imax, 2204 "maximum item count in this period"); 2205 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2206 "imin", CTLFLAG_RD, &zdom->uzd_imin, 2207 "minimum item count in this period"); 2208 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2209 "wss", CTLFLAG_RD, &zdom->uzd_wss, 2210 "Working set size"); 2211 } 2212 2213 /* 2214 * General statistics. 2215 */ 2216 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 2217 "stats", CTLFLAG_RD, NULL, ""); 2218 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2219 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE, 2220 zone, 1, sysctl_handle_uma_zone_cur, "I", 2221 "Current number of allocated items"); 2222 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2223 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2224 zone, 0, sysctl_handle_uma_zone_allocs, "QU", 2225 "Total allocation calls"); 2226 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2227 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2228 zone, 0, sysctl_handle_uma_zone_frees, "QU", 2229 "Total free calls"); 2230 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2231 "fails", CTLFLAG_RD, &zone->uz_fails, 2232 "Number of allocation failures"); 2233 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2234 "xdomain", CTLFLAG_RD, &zone->uz_xdomain, 0, 2235 "Free calls from the wrong domain"); 2236 } 2237 2238 struct uma_zone_count { 2239 const char *name; 2240 int count; 2241 }; 2242 2243 static void 2244 zone_count(uma_zone_t zone, void *arg) 2245 { 2246 struct uma_zone_count *cnt; 2247 2248 cnt = arg; 2249 /* 2250 * Some zones are rapidly created with identical names and 2251 * destroyed out of order. This can lead to gaps in the count. 2252 * Use one greater than the maximum observed for this name. 2253 */ 2254 if (strcmp(zone->uz_name, cnt->name) == 0) 2255 cnt->count = MAX(cnt->count, 2256 zone->uz_namecnt + 1); 2257 } 2258 2259 static void 2260 zone_update_caches(uma_zone_t zone) 2261 { 2262 int i; 2263 2264 for (i = 0; i <= mp_maxid; i++) { 2265 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size); 2266 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags); 2267 } 2268 } 2269 2270 /* 2271 * Zone header ctor. This initializes all fields, locks, etc. 2272 * 2273 * Arguments/Returns follow uma_ctor specifications 2274 * udata Actually uma_zctor_args 2275 */ 2276 static int 2277 zone_ctor(void *mem, int size, void *udata, int flags) 2278 { 2279 struct uma_zone_count cnt; 2280 struct uma_zctor_args *arg = udata; 2281 uma_zone_t zone = mem; 2282 uma_zone_t z; 2283 uma_keg_t keg; 2284 int i; 2285 2286 bzero(zone, size); 2287 zone->uz_name = arg->name; 2288 zone->uz_ctor = arg->ctor; 2289 zone->uz_dtor = arg->dtor; 2290 zone->uz_init = NULL; 2291 zone->uz_fini = NULL; 2292 zone->uz_sleeps = 0; 2293 zone->uz_xdomain = 0; 2294 zone->uz_bucket_size = 0; 2295 zone->uz_bucket_size_min = 0; 2296 zone->uz_bucket_size_max = BUCKET_MAX; 2297 zone->uz_flags = 0; 2298 zone->uz_warning = NULL; 2299 /* The domain structures follow the cpu structures. */ 2300 zone->uz_domain = 2301 (struct uma_zone_domain *)&zone->uz_cpu[mp_maxid + 1]; 2302 zone->uz_bkt_max = ULONG_MAX; 2303 timevalclear(&zone->uz_ratecheck); 2304 2305 /* Count the number of duplicate names. */ 2306 cnt.name = arg->name; 2307 cnt.count = 0; 2308 zone_foreach(zone_count, &cnt); 2309 zone->uz_namecnt = cnt.count; 2310 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 2311 ZONE_CROSS_LOCK_INIT(zone); 2312 2313 for (i = 0; i < vm_ndomains; i++) 2314 TAILQ_INIT(&zone->uz_domain[i].uzd_buckets); 2315 2316 #ifdef INVARIANTS 2317 if (arg->uminit == trash_init && arg->fini == trash_fini) 2318 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR; 2319 #endif 2320 2321 /* 2322 * This is a pure cache zone, no kegs. 2323 */ 2324 if (arg->import) { 2325 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0, 2326 ("zone_ctor: Import specified for non-cache zone.")); 2327 if (arg->flags & UMA_ZONE_VM) 2328 arg->flags |= UMA_ZFLAG_CACHEONLY; 2329 zone->uz_flags = arg->flags; 2330 zone->uz_size = arg->size; 2331 zone->uz_import = arg->import; 2332 zone->uz_release = arg->release; 2333 zone->uz_arg = arg->arg; 2334 rw_wlock(&uma_rwlock); 2335 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 2336 rw_wunlock(&uma_rwlock); 2337 goto out; 2338 } 2339 2340 /* 2341 * Use the regular zone/keg/slab allocator. 2342 */ 2343 zone->uz_import = zone_import; 2344 zone->uz_release = zone_release; 2345 zone->uz_arg = zone; 2346 keg = arg->keg; 2347 2348 if (arg->flags & UMA_ZONE_SECONDARY) { 2349 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 2350 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 2351 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 2352 zone->uz_init = arg->uminit; 2353 zone->uz_fini = arg->fini; 2354 zone->uz_flags |= UMA_ZONE_SECONDARY; 2355 rw_wlock(&uma_rwlock); 2356 ZONE_LOCK(zone); 2357 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 2358 if (LIST_NEXT(z, uz_link) == NULL) { 2359 LIST_INSERT_AFTER(z, zone, uz_link); 2360 break; 2361 } 2362 } 2363 ZONE_UNLOCK(zone); 2364 rw_wunlock(&uma_rwlock); 2365 } else if (keg == NULL) { 2366 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 2367 arg->align, arg->flags)) == NULL) 2368 return (ENOMEM); 2369 } else { 2370 struct uma_kctor_args karg; 2371 int error; 2372 2373 /* We should only be here from uma_startup() */ 2374 karg.size = arg->size; 2375 karg.uminit = arg->uminit; 2376 karg.fini = arg->fini; 2377 karg.align = arg->align; 2378 karg.flags = arg->flags; 2379 karg.zone = zone; 2380 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 2381 flags); 2382 if (error) 2383 return (error); 2384 } 2385 2386 /* Inherit properties from the keg. */ 2387 zone->uz_keg = keg; 2388 zone->uz_size = keg->uk_size; 2389 zone->uz_flags |= (keg->uk_flags & 2390 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 2391 2392 out: 2393 if (__predict_true(booted >= BOOT_RUNNING)) { 2394 zone_alloc_counters(zone, NULL); 2395 zone_alloc_sysctl(zone, NULL); 2396 } else { 2397 zone->uz_allocs = EARLY_COUNTER; 2398 zone->uz_frees = EARLY_COUNTER; 2399 zone->uz_fails = EARLY_COUNTER; 2400 } 2401 2402 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) != 2403 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET), 2404 ("Invalid zone flag combination")); 2405 if (arg->flags & UMA_ZFLAG_INTERNAL) 2406 zone->uz_bucket_size_max = zone->uz_bucket_size = 0; 2407 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0) 2408 zone->uz_bucket_size = BUCKET_MAX; 2409 else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0) 2410 zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN; 2411 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0) 2412 zone->uz_bucket_size = 0; 2413 else 2414 zone->uz_bucket_size = bucket_select(zone->uz_size); 2415 zone->uz_bucket_size_min = zone->uz_bucket_size; 2416 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL) 2417 zone->uz_flags |= UMA_ZFLAG_CTORDTOR; 2418 zone_update_caches(zone); 2419 2420 return (0); 2421 } 2422 2423 /* 2424 * Keg header dtor. This frees all data, destroys locks, frees the hash 2425 * table and removes the keg from the global list. 2426 * 2427 * Arguments/Returns follow uma_dtor specifications 2428 * udata unused 2429 */ 2430 static void 2431 keg_dtor(void *arg, int size, void *udata) 2432 { 2433 uma_keg_t keg; 2434 uint32_t free, pages; 2435 int i; 2436 2437 keg = (uma_keg_t)arg; 2438 free = pages = 0; 2439 for (i = 0; i < vm_ndomains; i++) { 2440 free += keg->uk_domain[i].ud_free; 2441 pages += keg->uk_domain[i].ud_pages; 2442 KEG_LOCK_FINI(keg, i); 2443 } 2444 if (pages != 0) 2445 printf("Freed UMA keg (%s) was not empty (%u items). " 2446 " Lost %u pages of memory.\n", 2447 keg->uk_name ? keg->uk_name : "", 2448 pages / keg->uk_ppera * keg->uk_ipers - free, pages); 2449 2450 hash_free(&keg->uk_hash); 2451 } 2452 2453 /* 2454 * Zone header dtor. 2455 * 2456 * Arguments/Returns follow uma_dtor specifications 2457 * udata unused 2458 */ 2459 static void 2460 zone_dtor(void *arg, int size, void *udata) 2461 { 2462 uma_zone_t zone; 2463 uma_keg_t keg; 2464 2465 zone = (uma_zone_t)arg; 2466 2467 sysctl_remove_oid(zone->uz_oid, 1, 1); 2468 2469 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 2470 cache_drain(zone); 2471 2472 rw_wlock(&uma_rwlock); 2473 LIST_REMOVE(zone, uz_link); 2474 rw_wunlock(&uma_rwlock); 2475 /* 2476 * XXX there are some races here where 2477 * the zone can be drained but zone lock 2478 * released and then refilled before we 2479 * remove it... we dont care for now 2480 */ 2481 zone_reclaim(zone, M_WAITOK, true); 2482 /* 2483 * We only destroy kegs from non secondary/non cache zones. 2484 */ 2485 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) { 2486 keg = zone->uz_keg; 2487 rw_wlock(&uma_rwlock); 2488 LIST_REMOVE(keg, uk_link); 2489 rw_wunlock(&uma_rwlock); 2490 zone_free_item(kegs, keg, NULL, SKIP_NONE); 2491 } 2492 counter_u64_free(zone->uz_allocs); 2493 counter_u64_free(zone->uz_frees); 2494 counter_u64_free(zone->uz_fails); 2495 free(zone->uz_ctlname, M_UMA); 2496 ZONE_LOCK_FINI(zone); 2497 ZONE_CROSS_LOCK_FINI(zone); 2498 } 2499 2500 static void 2501 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg) 2502 { 2503 uma_keg_t keg; 2504 uma_zone_t zone; 2505 2506 LIST_FOREACH(keg, &uma_kegs, uk_link) { 2507 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 2508 zfunc(zone, arg); 2509 } 2510 LIST_FOREACH(zone, &uma_cachezones, uz_link) 2511 zfunc(zone, arg); 2512 } 2513 2514 /* 2515 * Traverses every zone in the system and calls a callback 2516 * 2517 * Arguments: 2518 * zfunc A pointer to a function which accepts a zone 2519 * as an argument. 2520 * 2521 * Returns: 2522 * Nothing 2523 */ 2524 static void 2525 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg) 2526 { 2527 2528 rw_rlock(&uma_rwlock); 2529 zone_foreach_unlocked(zfunc, arg); 2530 rw_runlock(&uma_rwlock); 2531 } 2532 2533 /* 2534 * Initialize the kernel memory allocator. This is done after pages can be 2535 * allocated but before general KVA is available. 2536 */ 2537 void 2538 uma_startup1(vm_offset_t virtual_avail) 2539 { 2540 struct uma_zctor_args args; 2541 size_t ksize, zsize, size; 2542 uma_keg_t masterkeg; 2543 uintptr_t m; 2544 uint8_t pflag; 2545 2546 bootstart = bootmem = virtual_avail; 2547 2548 rw_init(&uma_rwlock, "UMA lock"); 2549 sx_init(&uma_reclaim_lock, "umareclaim"); 2550 2551 ksize = sizeof(struct uma_keg) + 2552 (sizeof(struct uma_domain) * vm_ndomains); 2553 ksize = roundup(ksize, UMA_SUPER_ALIGN); 2554 zsize = sizeof(struct uma_zone) + 2555 (sizeof(struct uma_cache) * (mp_maxid + 1)) + 2556 (sizeof(struct uma_zone_domain) * vm_ndomains); 2557 zsize = roundup(zsize, UMA_SUPER_ALIGN); 2558 2559 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */ 2560 size = (zsize * 2) + ksize; 2561 m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO); 2562 zones = (uma_zone_t)m; 2563 m += zsize; 2564 kegs = (uma_zone_t)m; 2565 m += zsize; 2566 masterkeg = (uma_keg_t)m; 2567 2568 /* "manually" create the initial zone */ 2569 memset(&args, 0, sizeof(args)); 2570 args.name = "UMA Kegs"; 2571 args.size = ksize; 2572 args.ctor = keg_ctor; 2573 args.dtor = keg_dtor; 2574 args.uminit = zero_init; 2575 args.fini = NULL; 2576 args.keg = masterkeg; 2577 args.align = UMA_SUPER_ALIGN - 1; 2578 args.flags = UMA_ZFLAG_INTERNAL; 2579 zone_ctor(kegs, zsize, &args, M_WAITOK); 2580 2581 args.name = "UMA Zones"; 2582 args.size = zsize; 2583 args.ctor = zone_ctor; 2584 args.dtor = zone_dtor; 2585 args.uminit = zero_init; 2586 args.fini = NULL; 2587 args.keg = NULL; 2588 args.align = UMA_SUPER_ALIGN - 1; 2589 args.flags = UMA_ZFLAG_INTERNAL; 2590 zone_ctor(zones, zsize, &args, M_WAITOK); 2591 2592 /* Now make zones for slab headers */ 2593 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE, 2594 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2595 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE, 2596 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2597 2598 hashzone = uma_zcreate("UMA Hash", 2599 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 2600 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2601 2602 bucket_init(); 2603 } 2604 2605 #ifndef UMA_MD_SMALL_ALLOC 2606 extern void vm_radix_reserve_kva(void); 2607 #endif 2608 2609 /* 2610 * Advertise the availability of normal kva allocations and switch to 2611 * the default back-end allocator. Marks the KVA we consumed on startup 2612 * as used in the map. 2613 */ 2614 void 2615 uma_startup2(void) 2616 { 2617 2618 if (bootstart != bootmem) { 2619 vm_map_lock(kernel_map); 2620 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem, 2621 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 2622 vm_map_unlock(kernel_map); 2623 } 2624 2625 #ifndef UMA_MD_SMALL_ALLOC 2626 /* Set up radix zone to use noobj_alloc. */ 2627 vm_radix_reserve_kva(); 2628 #endif 2629 2630 booted = BOOT_KVA; 2631 zone_foreach_unlocked(zone_kva_available, NULL); 2632 bucket_enable(); 2633 } 2634 2635 /* 2636 * Finish our initialization steps. 2637 */ 2638 static void 2639 uma_startup3(void) 2640 { 2641 2642 #ifdef INVARIANTS 2643 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor); 2644 uma_dbg_cnt = counter_u64_alloc(M_WAITOK); 2645 uma_skip_cnt = counter_u64_alloc(M_WAITOK); 2646 #endif 2647 zone_foreach_unlocked(zone_alloc_counters, NULL); 2648 zone_foreach_unlocked(zone_alloc_sysctl, NULL); 2649 callout_init(&uma_callout, 1); 2650 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 2651 booted = BOOT_RUNNING; 2652 2653 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL, 2654 EVENTHANDLER_PRI_FIRST); 2655 } 2656 2657 static void 2658 uma_shutdown(void) 2659 { 2660 2661 booted = BOOT_SHUTDOWN; 2662 } 2663 2664 static uma_keg_t 2665 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 2666 int align, uint32_t flags) 2667 { 2668 struct uma_kctor_args args; 2669 2670 args.size = size; 2671 args.uminit = uminit; 2672 args.fini = fini; 2673 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 2674 args.flags = flags; 2675 args.zone = zone; 2676 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK)); 2677 } 2678 2679 /* Public functions */ 2680 /* See uma.h */ 2681 void 2682 uma_set_align(int align) 2683 { 2684 2685 if (align != UMA_ALIGN_CACHE) 2686 uma_align_cache = align; 2687 } 2688 2689 /* See uma.h */ 2690 uma_zone_t 2691 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 2692 uma_init uminit, uma_fini fini, int align, uint32_t flags) 2693 2694 { 2695 struct uma_zctor_args args; 2696 uma_zone_t res; 2697 2698 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"", 2699 align, name)); 2700 2701 /* This stuff is essential for the zone ctor */ 2702 memset(&args, 0, sizeof(args)); 2703 args.name = name; 2704 args.size = size; 2705 args.ctor = ctor; 2706 args.dtor = dtor; 2707 args.uminit = uminit; 2708 args.fini = fini; 2709 #ifdef INVARIANTS 2710 /* 2711 * Inject procedures which check for memory use after free if we are 2712 * allowed to scramble the memory while it is not allocated. This 2713 * requires that: UMA is actually able to access the memory, no init 2714 * or fini procedures, no dependency on the initial value of the 2715 * memory, and no (legitimate) use of the memory after free. Note, 2716 * the ctor and dtor do not need to be empty. 2717 */ 2718 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH | 2719 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) { 2720 args.uminit = trash_init; 2721 args.fini = trash_fini; 2722 } 2723 #endif 2724 args.align = align; 2725 args.flags = flags; 2726 args.keg = NULL; 2727 2728 sx_slock(&uma_reclaim_lock); 2729 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2730 sx_sunlock(&uma_reclaim_lock); 2731 2732 return (res); 2733 } 2734 2735 /* See uma.h */ 2736 uma_zone_t 2737 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 2738 uma_init zinit, uma_fini zfini, uma_zone_t master) 2739 { 2740 struct uma_zctor_args args; 2741 uma_keg_t keg; 2742 uma_zone_t res; 2743 2744 keg = master->uz_keg; 2745 memset(&args, 0, sizeof(args)); 2746 args.name = name; 2747 args.size = keg->uk_size; 2748 args.ctor = ctor; 2749 args.dtor = dtor; 2750 args.uminit = zinit; 2751 args.fini = zfini; 2752 args.align = keg->uk_align; 2753 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 2754 args.keg = keg; 2755 2756 sx_slock(&uma_reclaim_lock); 2757 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2758 sx_sunlock(&uma_reclaim_lock); 2759 2760 return (res); 2761 } 2762 2763 /* See uma.h */ 2764 uma_zone_t 2765 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 2766 uma_init zinit, uma_fini zfini, uma_import zimport, 2767 uma_release zrelease, void *arg, int flags) 2768 { 2769 struct uma_zctor_args args; 2770 2771 memset(&args, 0, sizeof(args)); 2772 args.name = name; 2773 args.size = size; 2774 args.ctor = ctor; 2775 args.dtor = dtor; 2776 args.uminit = zinit; 2777 args.fini = zfini; 2778 args.import = zimport; 2779 args.release = zrelease; 2780 args.arg = arg; 2781 args.align = 0; 2782 args.flags = flags | UMA_ZFLAG_CACHE; 2783 2784 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK)); 2785 } 2786 2787 /* See uma.h */ 2788 void 2789 uma_zdestroy(uma_zone_t zone) 2790 { 2791 2792 /* 2793 * Large slabs are expensive to reclaim, so don't bother doing 2794 * unnecessary work if we're shutting down. 2795 */ 2796 if (booted == BOOT_SHUTDOWN && 2797 zone->uz_fini == NULL && zone->uz_release == zone_release) 2798 return; 2799 sx_slock(&uma_reclaim_lock); 2800 zone_free_item(zones, zone, NULL, SKIP_NONE); 2801 sx_sunlock(&uma_reclaim_lock); 2802 } 2803 2804 void 2805 uma_zwait(uma_zone_t zone) 2806 { 2807 void *item; 2808 2809 item = uma_zalloc_arg(zone, NULL, M_WAITOK); 2810 uma_zfree(zone, item); 2811 } 2812 2813 void * 2814 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags) 2815 { 2816 void *item; 2817 #ifdef SMP 2818 int i; 2819 2820 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2821 #endif 2822 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO); 2823 if (item != NULL && (flags & M_ZERO)) { 2824 #ifdef SMP 2825 for (i = 0; i <= mp_maxid; i++) 2826 bzero(zpcpu_get_cpu(item, i), zone->uz_size); 2827 #else 2828 bzero(item, zone->uz_size); 2829 #endif 2830 } 2831 return (item); 2832 } 2833 2834 /* 2835 * A stub while both regular and pcpu cases are identical. 2836 */ 2837 void 2838 uma_zfree_pcpu_arg(uma_zone_t zone, void *item, void *udata) 2839 { 2840 2841 #ifdef SMP 2842 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2843 #endif 2844 uma_zfree_arg(zone, item, udata); 2845 } 2846 2847 #ifdef INVARIANTS 2848 #define UMA_ALWAYS_CTORDTOR 1 2849 #else 2850 #define UMA_ALWAYS_CTORDTOR 0 2851 #endif 2852 2853 static void * 2854 item_ctor(uma_zone_t zone, int size, void *udata, int flags, void *item) 2855 { 2856 #ifdef INVARIANTS 2857 bool skipdbg; 2858 2859 skipdbg = uma_dbg_zskip(zone, item); 2860 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 && 2861 zone->uz_ctor != trash_ctor) 2862 trash_ctor(item, size, udata, flags); 2863 #endif 2864 if (__predict_false(zone->uz_ctor != NULL) && 2865 zone->uz_ctor(item, size, udata, flags) != 0) { 2866 counter_u64_add(zone->uz_fails, 1); 2867 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT); 2868 return (NULL); 2869 } 2870 #ifdef INVARIANTS 2871 if (!skipdbg) 2872 uma_dbg_alloc(zone, NULL, item); 2873 #endif 2874 if (flags & M_ZERO) 2875 bzero(item, size); 2876 2877 return (item); 2878 } 2879 2880 static inline void 2881 item_dtor(uma_zone_t zone, void *item, int size, void *udata, 2882 enum zfreeskip skip) 2883 { 2884 #ifdef INVARIANTS 2885 bool skipdbg; 2886 2887 skipdbg = uma_dbg_zskip(zone, item); 2888 if (skip == SKIP_NONE && !skipdbg) { 2889 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0) 2890 uma_dbg_free(zone, udata, item); 2891 else 2892 uma_dbg_free(zone, NULL, item); 2893 } 2894 #endif 2895 if (__predict_true(skip < SKIP_DTOR)) { 2896 if (zone->uz_dtor != NULL) 2897 zone->uz_dtor(item, size, udata); 2898 #ifdef INVARIANTS 2899 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 && 2900 zone->uz_dtor != trash_dtor) 2901 trash_dtor(item, size, udata); 2902 #endif 2903 } 2904 } 2905 2906 /* See uma.h */ 2907 void * 2908 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2909 { 2910 uma_cache_bucket_t bucket; 2911 uma_cache_t cache; 2912 void *item; 2913 int domain, size, uz_flags; 2914 2915 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2916 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 2917 2918 /* This is the fast path allocation */ 2919 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name, 2920 zone, flags); 2921 2922 #ifdef WITNESS 2923 if (flags & M_WAITOK) { 2924 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2925 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2926 } 2927 #endif 2928 2929 #ifdef INVARIANTS 2930 KASSERT((flags & M_EXEC) == 0, ("uma_zalloc_arg: called with M_EXEC")); 2931 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2932 ("uma_zalloc_arg: called with spinlock or critical section held")); 2933 if (zone->uz_flags & UMA_ZONE_PCPU) 2934 KASSERT((flags & M_ZERO) == 0, ("allocating from a pcpu zone " 2935 "with M_ZERO passed")); 2936 #endif 2937 2938 #ifdef DEBUG_MEMGUARD 2939 if (memguard_cmp_zone(zone)) { 2940 item = memguard_alloc(zone->uz_size, flags); 2941 if (item != NULL) { 2942 if (zone->uz_init != NULL && 2943 zone->uz_init(item, zone->uz_size, flags) != 0) 2944 return (NULL); 2945 if (zone->uz_ctor != NULL && 2946 zone->uz_ctor(item, zone->uz_size, udata, 2947 flags) != 0) { 2948 counter_u64_add(zone->uz_fails, 1); 2949 zone->uz_fini(item, zone->uz_size); 2950 return (NULL); 2951 } 2952 return (item); 2953 } 2954 /* This is unfortunate but should not be fatal. */ 2955 } 2956 #endif 2957 /* 2958 * If possible, allocate from the per-CPU cache. There are two 2959 * requirements for safe access to the per-CPU cache: (1) the thread 2960 * accessing the cache must not be preempted or yield during access, 2961 * and (2) the thread must not migrate CPUs without switching which 2962 * cache it accesses. We rely on a critical section to prevent 2963 * preemption and migration. We release the critical section in 2964 * order to acquire the zone mutex if we are unable to allocate from 2965 * the current cache; when we re-acquire the critical section, we 2966 * must detect and handle migration if it has occurred. 2967 */ 2968 critical_enter(); 2969 do { 2970 cache = &zone->uz_cpu[curcpu]; 2971 bucket = &cache->uc_allocbucket; 2972 size = cache_uz_size(cache); 2973 uz_flags = cache_uz_flags(cache); 2974 if (__predict_true(bucket->ucb_cnt != 0)) { 2975 item = cache_bucket_pop(cache, bucket); 2976 critical_exit(); 2977 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0 || 2978 UMA_ALWAYS_CTORDTOR)) 2979 return (item_ctor(zone, size, udata, flags, item)); 2980 if (flags & M_ZERO) 2981 bzero(item, size); 2982 return (item); 2983 } 2984 } while (cache_alloc(zone, cache, udata, flags)); 2985 critical_exit(); 2986 2987 /* 2988 * We can not get a bucket so try to return a single item. 2989 */ 2990 if (uz_flags & UMA_ZONE_FIRSTTOUCH) 2991 domain = PCPU_GET(domain); 2992 else 2993 domain = UMA_ANYDOMAIN; 2994 return (zone_alloc_item(zone, udata, domain, flags)); 2995 } 2996 2997 /* 2998 * Replenish an alloc bucket and possibly restore an old one. Called in 2999 * a critical section. Returns in a critical section. 3000 * 3001 * A false return value indicates an allocation failure. 3002 * A true return value indicates success and the caller should retry. 3003 */ 3004 static __noinline bool 3005 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags) 3006 { 3007 uma_zone_domain_t zdom; 3008 uma_bucket_t bucket; 3009 int domain; 3010 bool lockfail; 3011 3012 CRITICAL_ASSERT(curthread); 3013 3014 /* 3015 * If we have run out of items in our alloc bucket see 3016 * if we can switch with the free bucket. 3017 */ 3018 if (cache->uc_freebucket.ucb_cnt != 0) { 3019 cache_bucket_swap(&cache->uc_freebucket, &cache->uc_allocbucket); 3020 return (true); 3021 } 3022 3023 /* 3024 * Discard any empty allocation bucket while we hold no locks. 3025 */ 3026 bucket = cache_bucket_unload_alloc(cache); 3027 critical_exit(); 3028 if (bucket != NULL) 3029 bucket_free(zone, bucket, udata); 3030 3031 /* Short-circuit for zones without buckets and low memory. */ 3032 if (zone->uz_bucket_size == 0 || bucketdisable) { 3033 critical_enter(); 3034 return (false); 3035 } 3036 3037 /* 3038 * Attempt to retrieve the item from the per-CPU cache has failed, so 3039 * we must go back to the zone. This requires the zone lock, so we 3040 * must drop the critical section, then re-acquire it when we go back 3041 * to the cache. Since the critical section is released, we may be 3042 * preempted or migrate. As such, make sure not to maintain any 3043 * thread-local state specific to the cache from prior to releasing 3044 * the critical section. 3045 */ 3046 lockfail = 0; 3047 if (ZONE_TRYLOCK(zone) == 0) { 3048 /* Record contention to size the buckets. */ 3049 ZONE_LOCK(zone); 3050 lockfail = 1; 3051 } 3052 3053 /* See if we lost the race to fill the cache. */ 3054 critical_enter(); 3055 cache = &zone->uz_cpu[curcpu]; 3056 if (cache->uc_allocbucket.ucb_bucket != NULL) { 3057 ZONE_UNLOCK(zone); 3058 return (true); 3059 } 3060 3061 /* 3062 * Check the zone's cache of buckets. 3063 */ 3064 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) { 3065 domain = PCPU_GET(domain); 3066 zdom = &zone->uz_domain[domain]; 3067 } else { 3068 domain = UMA_ANYDOMAIN; 3069 zdom = &zone->uz_domain[0]; 3070 } 3071 3072 if ((bucket = zone_fetch_bucket(zone, zdom)) != NULL) { 3073 ZONE_UNLOCK(zone); 3074 KASSERT(bucket->ub_cnt != 0, 3075 ("uma_zalloc_arg: Returning an empty bucket.")); 3076 cache_bucket_load_alloc(cache, bucket); 3077 return (true); 3078 } 3079 /* We are no longer associated with this CPU. */ 3080 critical_exit(); 3081 3082 /* 3083 * We bump the uz count when the cache size is insufficient to 3084 * handle the working set. 3085 */ 3086 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max) 3087 zone->uz_bucket_size++; 3088 ZONE_UNLOCK(zone); 3089 3090 /* 3091 * Fill a bucket and attempt to use it as the alloc bucket. 3092 */ 3093 bucket = zone_alloc_bucket(zone, udata, domain, flags); 3094 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p", 3095 zone->uz_name, zone, bucket); 3096 if (bucket == NULL) { 3097 critical_enter(); 3098 return (false); 3099 } 3100 3101 /* 3102 * See if we lost the race or were migrated. Cache the 3103 * initialized bucket to make this less likely or claim 3104 * the memory directly. 3105 */ 3106 ZONE_LOCK(zone); 3107 critical_enter(); 3108 cache = &zone->uz_cpu[curcpu]; 3109 if (cache->uc_allocbucket.ucb_bucket == NULL && 3110 ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0 || 3111 domain == PCPU_GET(domain))) { 3112 cache_bucket_load_alloc(cache, bucket); 3113 zdom->uzd_imax += bucket->ub_cnt; 3114 } else if (zone->uz_bkt_count >= zone->uz_bkt_max) { 3115 critical_exit(); 3116 ZONE_UNLOCK(zone); 3117 bucket_drain(zone, bucket); 3118 bucket_free(zone, bucket, udata); 3119 critical_enter(); 3120 return (true); 3121 } else 3122 zone_put_bucket(zone, zdom, bucket, false); 3123 ZONE_UNLOCK(zone); 3124 return (true); 3125 } 3126 3127 void * 3128 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags) 3129 { 3130 3131 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3132 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3133 3134 /* This is the fast path allocation */ 3135 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d", 3136 zone->uz_name, zone, domain, flags); 3137 3138 if (flags & M_WAITOK) { 3139 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 3140 "uma_zalloc_domain: zone \"%s\"", zone->uz_name); 3141 } 3142 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3143 ("uma_zalloc_domain: called with spinlock or critical section held")); 3144 3145 return (zone_alloc_item(zone, udata, domain, flags)); 3146 } 3147 3148 /* 3149 * Find a slab with some space. Prefer slabs that are partially used over those 3150 * that are totally full. This helps to reduce fragmentation. 3151 * 3152 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check 3153 * only 'domain'. 3154 */ 3155 static uma_slab_t 3156 keg_first_slab(uma_keg_t keg, int domain, bool rr) 3157 { 3158 uma_domain_t dom; 3159 uma_slab_t slab; 3160 int start; 3161 3162 KASSERT(domain >= 0 && domain < vm_ndomains, 3163 ("keg_first_slab: domain %d out of range", domain)); 3164 KEG_LOCK_ASSERT(keg, domain); 3165 3166 slab = NULL; 3167 start = domain; 3168 do { 3169 dom = &keg->uk_domain[domain]; 3170 if (!LIST_EMPTY(&dom->ud_part_slab)) 3171 return (LIST_FIRST(&dom->ud_part_slab)); 3172 if (!LIST_EMPTY(&dom->ud_free_slab)) { 3173 slab = LIST_FIRST(&dom->ud_free_slab); 3174 LIST_REMOVE(slab, us_link); 3175 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 3176 return (slab); 3177 } 3178 if (rr) 3179 domain = (domain + 1) % vm_ndomains; 3180 } while (domain != start); 3181 3182 return (NULL); 3183 } 3184 3185 /* 3186 * Fetch an existing slab from a free or partial list. Returns with the 3187 * keg domain lock held if a slab was found or unlocked if not. 3188 */ 3189 static uma_slab_t 3190 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags) 3191 { 3192 uma_slab_t slab; 3193 uint32_t reserve; 3194 3195 /* HASH has a single free list. */ 3196 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) 3197 domain = 0; 3198 3199 KEG_LOCK(keg, domain); 3200 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve; 3201 if (keg->uk_domain[domain].ud_free <= reserve || 3202 (slab = keg_first_slab(keg, domain, rr)) == NULL) { 3203 KEG_UNLOCK(keg, domain); 3204 return (NULL); 3205 } 3206 return (slab); 3207 } 3208 3209 static uma_slab_t 3210 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags) 3211 { 3212 struct vm_domainset_iter di; 3213 uma_slab_t slab; 3214 int aflags, domain; 3215 bool rr; 3216 3217 restart: 3218 /* 3219 * Use the keg's policy if upper layers haven't already specified a 3220 * domain (as happens with first-touch zones). 3221 * 3222 * To avoid races we run the iterator with the keg lock held, but that 3223 * means that we cannot allow the vm_domainset layer to sleep. Thus, 3224 * clear M_WAITOK and handle low memory conditions locally. 3225 */ 3226 rr = rdomain == UMA_ANYDOMAIN; 3227 if (rr) { 3228 aflags = (flags & ~M_WAITOK) | M_NOWAIT; 3229 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 3230 &aflags); 3231 } else { 3232 aflags = flags; 3233 domain = rdomain; 3234 } 3235 3236 for (;;) { 3237 slab = keg_fetch_free_slab(keg, domain, rr, flags); 3238 if (slab != NULL) 3239 return (slab); 3240 3241 /* 3242 * M_NOVM means don't ask at all! 3243 */ 3244 if (flags & M_NOVM) 3245 break; 3246 3247 slab = keg_alloc_slab(keg, zone, domain, flags, aflags); 3248 if (slab != NULL) 3249 return (slab); 3250 if (!rr && (flags & M_WAITOK) == 0) 3251 break; 3252 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) { 3253 if ((flags & M_WAITOK) != 0) { 3254 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask); 3255 goto restart; 3256 } 3257 break; 3258 } 3259 } 3260 3261 /* 3262 * We might not have been able to get a slab but another cpu 3263 * could have while we were unlocked. Check again before we 3264 * fail. 3265 */ 3266 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL) 3267 return (slab); 3268 3269 return (NULL); 3270 } 3271 3272 static void * 3273 slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 3274 { 3275 uma_domain_t dom; 3276 void *item; 3277 int freei; 3278 3279 KEG_LOCK_ASSERT(keg, slab->us_domain); 3280 3281 dom = &keg->uk_domain[slab->us_domain]; 3282 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1; 3283 BIT_CLR(keg->uk_ipers, freei, &slab->us_free); 3284 item = slab_item(slab, keg, freei); 3285 slab->us_freecount--; 3286 dom->ud_free--; 3287 3288 /* Move this slab to the full list */ 3289 if (slab->us_freecount == 0) { 3290 LIST_REMOVE(slab, us_link); 3291 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link); 3292 } 3293 3294 return (item); 3295 } 3296 3297 static int 3298 zone_import(void *arg, void **bucket, int max, int domain, int flags) 3299 { 3300 uma_domain_t dom; 3301 uma_zone_t zone; 3302 uma_slab_t slab; 3303 uma_keg_t keg; 3304 #ifdef NUMA 3305 int stripe; 3306 #endif 3307 int i; 3308 3309 zone = arg; 3310 slab = NULL; 3311 keg = zone->uz_keg; 3312 /* Try to keep the buckets totally full */ 3313 for (i = 0; i < max; ) { 3314 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL) 3315 break; 3316 #ifdef NUMA 3317 stripe = howmany(max, vm_ndomains); 3318 #endif 3319 dom = &keg->uk_domain[slab->us_domain]; 3320 while (slab->us_freecount && i < max) { 3321 bucket[i++] = slab_alloc_item(keg, slab); 3322 if (dom->ud_free <= keg->uk_reserve) 3323 break; 3324 #ifdef NUMA 3325 /* 3326 * If the zone is striped we pick a new slab for every 3327 * N allocations. Eliminating this conditional will 3328 * instead pick a new domain for each bucket rather 3329 * than stripe within each bucket. The current option 3330 * produces more fragmentation and requires more cpu 3331 * time but yields better distribution. 3332 */ 3333 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 && 3334 vm_ndomains > 1 && --stripe == 0) 3335 break; 3336 #endif 3337 } 3338 KEG_UNLOCK(keg, slab->us_domain); 3339 /* Don't block if we allocated any successfully. */ 3340 flags &= ~M_WAITOK; 3341 flags |= M_NOWAIT; 3342 } 3343 3344 return i; 3345 } 3346 3347 static int 3348 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags) 3349 { 3350 uint64_t old, new, total, max; 3351 3352 /* 3353 * The hard case. We're going to sleep because there were existing 3354 * sleepers or because we ran out of items. This routine enforces 3355 * fairness by keeping fifo order. 3356 * 3357 * First release our ill gotten gains and make some noise. 3358 */ 3359 for (;;) { 3360 zone_free_limit(zone, count); 3361 zone_log_warning(zone); 3362 zone_maxaction(zone); 3363 if (flags & M_NOWAIT) 3364 return (0); 3365 3366 /* 3367 * We need to allocate an item or set ourself as a sleeper 3368 * while the sleepq lock is held to avoid wakeup races. This 3369 * is essentially a home rolled semaphore. 3370 */ 3371 sleepq_lock(&zone->uz_max_items); 3372 old = zone->uz_items; 3373 do { 3374 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX); 3375 /* Cache the max since we will evaluate twice. */ 3376 max = zone->uz_max_items; 3377 if (UZ_ITEMS_SLEEPERS(old) != 0 || 3378 UZ_ITEMS_COUNT(old) >= max) 3379 new = old + UZ_ITEMS_SLEEPER; 3380 else 3381 new = old + MIN(count, max - old); 3382 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0); 3383 3384 /* We may have successfully allocated under the sleepq lock. */ 3385 if (UZ_ITEMS_SLEEPERS(new) == 0) { 3386 sleepq_release(&zone->uz_max_items); 3387 return (new - old); 3388 } 3389 3390 /* 3391 * This is in a different cacheline from uz_items so that we 3392 * don't constantly invalidate the fastpath cacheline when we 3393 * adjust item counts. This could be limited to toggling on 3394 * transitions. 3395 */ 3396 atomic_add_32(&zone->uz_sleepers, 1); 3397 atomic_add_64(&zone->uz_sleeps, 1); 3398 3399 /* 3400 * We have added ourselves as a sleeper. The sleepq lock 3401 * protects us from wakeup races. Sleep now and then retry. 3402 */ 3403 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0); 3404 sleepq_wait(&zone->uz_max_items, PVM); 3405 3406 /* 3407 * After wakeup, remove ourselves as a sleeper and try 3408 * again. We no longer have the sleepq lock for protection. 3409 * 3410 * Subract ourselves as a sleeper while attempting to add 3411 * our count. 3412 */ 3413 atomic_subtract_32(&zone->uz_sleepers, 1); 3414 old = atomic_fetchadd_64(&zone->uz_items, 3415 -(UZ_ITEMS_SLEEPER - count)); 3416 /* We're no longer a sleeper. */ 3417 old -= UZ_ITEMS_SLEEPER; 3418 3419 /* 3420 * If we're still at the limit, restart. Notably do not 3421 * block on other sleepers. Cache the max value to protect 3422 * against changes via sysctl. 3423 */ 3424 total = UZ_ITEMS_COUNT(old); 3425 max = zone->uz_max_items; 3426 if (total >= max) 3427 continue; 3428 /* Truncate if necessary, otherwise wake other sleepers. */ 3429 if (total + count > max) { 3430 zone_free_limit(zone, total + count - max); 3431 count = max - total; 3432 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0) 3433 wakeup_one(&zone->uz_max_items); 3434 3435 return (count); 3436 } 3437 } 3438 3439 /* 3440 * Allocate 'count' items from our max_items limit. Returns the number 3441 * available. If M_NOWAIT is not specified it will sleep until at least 3442 * one item can be allocated. 3443 */ 3444 static int 3445 zone_alloc_limit(uma_zone_t zone, int count, int flags) 3446 { 3447 uint64_t old; 3448 uint64_t max; 3449 3450 max = zone->uz_max_items; 3451 MPASS(max > 0); 3452 3453 /* 3454 * We expect normal allocations to succeed with a simple 3455 * fetchadd. 3456 */ 3457 old = atomic_fetchadd_64(&zone->uz_items, count); 3458 if (__predict_true(old + count <= max)) 3459 return (count); 3460 3461 /* 3462 * If we had some items and no sleepers just return the 3463 * truncated value. We have to release the excess space 3464 * though because that may wake sleepers who weren't woken 3465 * because we were temporarily over the limit. 3466 */ 3467 if (old < max) { 3468 zone_free_limit(zone, (old + count) - max); 3469 return (max - old); 3470 } 3471 return (zone_alloc_limit_hard(zone, count, flags)); 3472 } 3473 3474 /* 3475 * Free a number of items back to the limit. 3476 */ 3477 static void 3478 zone_free_limit(uma_zone_t zone, int count) 3479 { 3480 uint64_t old; 3481 3482 MPASS(count > 0); 3483 3484 /* 3485 * In the common case we either have no sleepers or 3486 * are still over the limit and can just return. 3487 */ 3488 old = atomic_fetchadd_64(&zone->uz_items, -count); 3489 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 || 3490 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items)) 3491 return; 3492 3493 /* 3494 * Moderate the rate of wakeups. Sleepers will continue 3495 * to generate wakeups if necessary. 3496 */ 3497 wakeup_one(&zone->uz_max_items); 3498 } 3499 3500 static uma_bucket_t 3501 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags) 3502 { 3503 uma_bucket_t bucket; 3504 int maxbucket, cnt; 3505 3506 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name, 3507 zone, domain); 3508 3509 /* Avoid allocs targeting empty domains. */ 3510 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain)) 3511 domain = UMA_ANYDOMAIN; 3512 3513 if (zone->uz_max_items > 0) 3514 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size, 3515 M_NOWAIT); 3516 else 3517 maxbucket = zone->uz_bucket_size; 3518 if (maxbucket == 0) 3519 return (false); 3520 3521 /* Don't wait for buckets, preserve caller's NOVM setting. */ 3522 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 3523 if (bucket == NULL) { 3524 cnt = 0; 3525 goto out; 3526 } 3527 3528 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 3529 MIN(maxbucket, bucket->ub_entries), domain, flags); 3530 3531 /* 3532 * Initialize the memory if necessary. 3533 */ 3534 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 3535 int i; 3536 3537 for (i = 0; i < bucket->ub_cnt; i++) 3538 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 3539 flags) != 0) 3540 break; 3541 /* 3542 * If we couldn't initialize the whole bucket, put the 3543 * rest back onto the freelist. 3544 */ 3545 if (i != bucket->ub_cnt) { 3546 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 3547 bucket->ub_cnt - i); 3548 #ifdef INVARIANTS 3549 bzero(&bucket->ub_bucket[i], 3550 sizeof(void *) * (bucket->ub_cnt - i)); 3551 #endif 3552 bucket->ub_cnt = i; 3553 } 3554 } 3555 3556 cnt = bucket->ub_cnt; 3557 if (bucket->ub_cnt == 0) { 3558 bucket_free(zone, bucket, udata); 3559 counter_u64_add(zone->uz_fails, 1); 3560 bucket = NULL; 3561 } 3562 out: 3563 if (zone->uz_max_items > 0 && cnt < maxbucket) 3564 zone_free_limit(zone, maxbucket - cnt); 3565 3566 return (bucket); 3567 } 3568 3569 /* 3570 * Allocates a single item from a zone. 3571 * 3572 * Arguments 3573 * zone The zone to alloc for. 3574 * udata The data to be passed to the constructor. 3575 * domain The domain to allocate from or UMA_ANYDOMAIN. 3576 * flags M_WAITOK, M_NOWAIT, M_ZERO. 3577 * 3578 * Returns 3579 * NULL if there is no memory and M_NOWAIT is set 3580 * An item if successful 3581 */ 3582 3583 static void * 3584 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags) 3585 { 3586 void *item; 3587 3588 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) 3589 return (NULL); 3590 3591 /* Avoid allocs targeting empty domains. */ 3592 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain)) 3593 domain = UMA_ANYDOMAIN; 3594 3595 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1) 3596 goto fail_cnt; 3597 3598 /* 3599 * We have to call both the zone's init (not the keg's init) 3600 * and the zone's ctor. This is because the item is going from 3601 * a keg slab directly to the user, and the user is expecting it 3602 * to be both zone-init'd as well as zone-ctor'd. 3603 */ 3604 if (zone->uz_init != NULL) { 3605 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 3606 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT); 3607 goto fail_cnt; 3608 } 3609 } 3610 item = item_ctor(zone, zone->uz_size, udata, flags, item); 3611 if (item == NULL) 3612 goto fail; 3613 3614 counter_u64_add(zone->uz_allocs, 1); 3615 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item, 3616 zone->uz_name, zone); 3617 3618 return (item); 3619 3620 fail_cnt: 3621 counter_u64_add(zone->uz_fails, 1); 3622 fail: 3623 if (zone->uz_max_items > 0) 3624 zone_free_limit(zone, 1); 3625 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)", 3626 zone->uz_name, zone); 3627 3628 return (NULL); 3629 } 3630 3631 /* See uma.h */ 3632 void 3633 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 3634 { 3635 uma_cache_t cache; 3636 uma_cache_bucket_t bucket; 3637 int domain, itemdomain, uz_flags; 3638 3639 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3640 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3641 3642 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone); 3643 3644 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3645 ("uma_zfree_arg: called with spinlock or critical section held")); 3646 3647 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 3648 if (item == NULL) 3649 return; 3650 #ifdef DEBUG_MEMGUARD 3651 if (is_memguard_addr(item)) { 3652 if (zone->uz_dtor != NULL) 3653 zone->uz_dtor(item, zone->uz_size, udata); 3654 if (zone->uz_fini != NULL) 3655 zone->uz_fini(item, zone->uz_size); 3656 memguard_free(item); 3657 return; 3658 } 3659 #endif 3660 3661 /* 3662 * We are accessing the per-cpu cache without a critical section to 3663 * fetch size and flags. This is acceptable, if we are preempted we 3664 * will simply read another cpu's line. 3665 */ 3666 cache = &zone->uz_cpu[curcpu]; 3667 uz_flags = cache_uz_flags(cache); 3668 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0 || 3669 UMA_ALWAYS_CTORDTOR)) 3670 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE); 3671 3672 /* 3673 * The race here is acceptable. If we miss it we'll just have to wait 3674 * a little longer for the limits to be reset. 3675 */ 3676 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) { 3677 if (zone->uz_sleepers > 0) 3678 goto zfree_item; 3679 } 3680 3681 /* 3682 * If possible, free to the per-CPU cache. There are two 3683 * requirements for safe access to the per-CPU cache: (1) the thread 3684 * accessing the cache must not be preempted or yield during access, 3685 * and (2) the thread must not migrate CPUs without switching which 3686 * cache it accesses. We rely on a critical section to prevent 3687 * preemption and migration. We release the critical section in 3688 * order to acquire the zone mutex if we are unable to free to the 3689 * current cache; when we re-acquire the critical section, we must 3690 * detect and handle migration if it has occurred. 3691 */ 3692 domain = itemdomain = 0; 3693 #ifdef NUMA 3694 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 3695 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item)); 3696 #endif 3697 critical_enter(); 3698 do { 3699 cache = &zone->uz_cpu[curcpu]; 3700 #ifdef NUMA 3701 domain = PCPU_GET(domain); 3702 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && 3703 domain != itemdomain) { 3704 bucket = &cache->uc_crossbucket; 3705 } else 3706 #endif 3707 { 3708 /* 3709 * Try to free into the allocbucket first to give LIFO 3710 * ordering for cache-hot datastructures. Spill over 3711 * into the freebucket if necessary. Alloc will swap 3712 * them if one runs dry. 3713 */ 3714 bucket = &cache->uc_allocbucket; 3715 if (__predict_false(bucket->ucb_cnt >= 3716 bucket->ucb_entries)) 3717 bucket = &cache->uc_freebucket; 3718 } 3719 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) { 3720 cache_bucket_push(cache, bucket, item); 3721 critical_exit(); 3722 return; 3723 } 3724 } while (cache_free(zone, cache, udata, item, itemdomain)); 3725 critical_exit(); 3726 3727 /* 3728 * If nothing else caught this, we'll just do an internal free. 3729 */ 3730 zfree_item: 3731 zone_free_item(zone, item, udata, SKIP_DTOR); 3732 } 3733 3734 #ifdef NUMA 3735 /* 3736 * sort crossdomain free buckets to domain correct buckets and cache 3737 * them. 3738 */ 3739 static void 3740 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata) 3741 { 3742 struct uma_bucketlist fullbuckets; 3743 uma_zone_domain_t zdom; 3744 uma_bucket_t b; 3745 void *item; 3746 int domain; 3747 3748 CTR3(KTR_UMA, 3749 "uma_zfree: zone %s(%p) draining cross bucket %p", 3750 zone->uz_name, zone, bucket); 3751 3752 TAILQ_INIT(&fullbuckets); 3753 3754 /* 3755 * To avoid having ndomain * ndomain buckets for sorting we have a 3756 * lock on the current crossfree bucket. A full matrix with 3757 * per-domain locking could be used if necessary. 3758 */ 3759 ZONE_CROSS_LOCK(zone); 3760 while (bucket->ub_cnt > 0) { 3761 item = bucket->ub_bucket[bucket->ub_cnt - 1]; 3762 domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item)); 3763 zdom = &zone->uz_domain[domain]; 3764 if (zdom->uzd_cross == NULL) { 3765 zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT); 3766 if (zdom->uzd_cross == NULL) 3767 break; 3768 } 3769 zdom->uzd_cross->ub_bucket[zdom->uzd_cross->ub_cnt++] = item; 3770 if (zdom->uzd_cross->ub_cnt == zdom->uzd_cross->ub_entries) { 3771 TAILQ_INSERT_HEAD(&fullbuckets, zdom->uzd_cross, 3772 ub_link); 3773 zdom->uzd_cross = NULL; 3774 } 3775 bucket->ub_cnt--; 3776 } 3777 ZONE_CROSS_UNLOCK(zone); 3778 if (!TAILQ_EMPTY(&fullbuckets)) { 3779 ZONE_LOCK(zone); 3780 while ((b = TAILQ_FIRST(&fullbuckets)) != NULL) { 3781 TAILQ_REMOVE(&fullbuckets, b, ub_link); 3782 if (zone->uz_bkt_count >= zone->uz_bkt_max) { 3783 ZONE_UNLOCK(zone); 3784 bucket_drain(zone, b); 3785 bucket_free(zone, b, udata); 3786 ZONE_LOCK(zone); 3787 } else { 3788 domain = _vm_phys_domain( 3789 pmap_kextract( 3790 (vm_offset_t)b->ub_bucket[0])); 3791 zdom = &zone->uz_domain[domain]; 3792 zone_put_bucket(zone, zdom, b, true); 3793 } 3794 } 3795 ZONE_UNLOCK(zone); 3796 } 3797 if (bucket->ub_cnt != 0) 3798 bucket_drain(zone, bucket); 3799 bucket_free(zone, bucket, udata); 3800 } 3801 #endif 3802 3803 static void 3804 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata, 3805 int domain, int itemdomain) 3806 { 3807 uma_zone_domain_t zdom; 3808 3809 #ifdef NUMA 3810 /* 3811 * Buckets coming from the wrong domain will be entirely for the 3812 * only other domain on two domain systems. In this case we can 3813 * simply cache them. Otherwise we need to sort them back to 3814 * correct domains. 3815 */ 3816 if (domain != itemdomain && vm_ndomains > 2) { 3817 zone_free_cross(zone, bucket, udata); 3818 return; 3819 } 3820 #endif 3821 3822 /* 3823 * Attempt to save the bucket in the zone's domain bucket cache. 3824 * 3825 * We bump the uz count when the cache size is insufficient to 3826 * handle the working set. 3827 */ 3828 if (ZONE_TRYLOCK(zone) == 0) { 3829 /* Record contention to size the buckets. */ 3830 ZONE_LOCK(zone); 3831 if (zone->uz_bucket_size < zone->uz_bucket_size_max) 3832 zone->uz_bucket_size++; 3833 } 3834 3835 CTR3(KTR_UMA, 3836 "uma_zfree: zone %s(%p) putting bucket %p on free list", 3837 zone->uz_name, zone, bucket); 3838 /* ub_cnt is pointing to the last free item */ 3839 KASSERT(bucket->ub_cnt == bucket->ub_entries, 3840 ("uma_zfree: Attempting to insert partial bucket onto the full list.\n")); 3841 if (zone->uz_bkt_count >= zone->uz_bkt_max) { 3842 ZONE_UNLOCK(zone); 3843 bucket_drain(zone, bucket); 3844 bucket_free(zone, bucket, udata); 3845 } else { 3846 zdom = &zone->uz_domain[itemdomain]; 3847 zone_put_bucket(zone, zdom, bucket, true); 3848 ZONE_UNLOCK(zone); 3849 } 3850 } 3851 3852 /* 3853 * Populate a free or cross bucket for the current cpu cache. Free any 3854 * existing full bucket either to the zone cache or back to the slab layer. 3855 * 3856 * Enters and returns in a critical section. false return indicates that 3857 * we can not satisfy this free in the cache layer. true indicates that 3858 * the caller should retry. 3859 */ 3860 static __noinline bool 3861 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item, 3862 int itemdomain) 3863 { 3864 uma_cache_bucket_t cbucket; 3865 uma_bucket_t bucket; 3866 int domain; 3867 3868 CRITICAL_ASSERT(curthread); 3869 3870 if (zone->uz_bucket_size == 0 || bucketdisable) 3871 return false; 3872 3873 cache = &zone->uz_cpu[curcpu]; 3874 3875 /* 3876 * FIRSTTOUCH domains need to free to the correct zdom. When 3877 * enabled this is the zdom of the item. The bucket is the 3878 * cross bucket if the current domain and itemdomain do not match. 3879 */ 3880 cbucket = &cache->uc_freebucket; 3881 #ifdef NUMA 3882 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) { 3883 domain = PCPU_GET(domain); 3884 if (domain != itemdomain) { 3885 cbucket = &cache->uc_crossbucket; 3886 if (cbucket->ucb_cnt != 0) 3887 atomic_add_64(&zone->uz_xdomain, 3888 cbucket->ucb_cnt); 3889 } 3890 } else 3891 #endif 3892 itemdomain = domain = 0; 3893 bucket = cache_bucket_unload(cbucket); 3894 3895 /* We are no longer associated with this CPU. */ 3896 critical_exit(); 3897 3898 if (bucket != NULL) 3899 zone_free_bucket(zone, bucket, udata, domain, itemdomain); 3900 3901 bucket = bucket_alloc(zone, udata, M_NOWAIT); 3902 CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p", 3903 zone->uz_name, zone, bucket); 3904 critical_enter(); 3905 if (bucket == NULL) 3906 return (false); 3907 cache = &zone->uz_cpu[curcpu]; 3908 #ifdef NUMA 3909 /* 3910 * Check to see if we should be populating the cross bucket. If it 3911 * is already populated we will fall through and attempt to populate 3912 * the free bucket. 3913 */ 3914 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) { 3915 domain = PCPU_GET(domain); 3916 if (domain != itemdomain && 3917 cache->uc_crossbucket.ucb_bucket == NULL) { 3918 cache_bucket_load_cross(cache, bucket); 3919 return (true); 3920 } 3921 } 3922 #endif 3923 /* 3924 * We may have lost the race to fill the bucket or switched CPUs. 3925 */ 3926 if (cache->uc_freebucket.ucb_bucket != NULL) { 3927 critical_exit(); 3928 bucket_free(zone, bucket, udata); 3929 critical_enter(); 3930 } else 3931 cache_bucket_load_free(cache, bucket); 3932 3933 return (true); 3934 } 3935 3936 void 3937 uma_zfree_domain(uma_zone_t zone, void *item, void *udata) 3938 { 3939 3940 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3941 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3942 3943 CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone); 3944 3945 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3946 ("uma_zfree_domain: called with spinlock or critical section held")); 3947 3948 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 3949 if (item == NULL) 3950 return; 3951 zone_free_item(zone, item, udata, SKIP_NONE); 3952 } 3953 3954 static void 3955 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item) 3956 { 3957 uma_keg_t keg; 3958 uma_domain_t dom; 3959 int freei; 3960 3961 keg = zone->uz_keg; 3962 KEG_LOCK_ASSERT(keg, slab->us_domain); 3963 3964 /* Do we need to remove from any lists? */ 3965 dom = &keg->uk_domain[slab->us_domain]; 3966 if (slab->us_freecount+1 == keg->uk_ipers) { 3967 LIST_REMOVE(slab, us_link); 3968 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); 3969 } else if (slab->us_freecount == 0) { 3970 LIST_REMOVE(slab, us_link); 3971 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 3972 } 3973 3974 /* Slab management. */ 3975 freei = slab_item_index(slab, keg, item); 3976 BIT_SET(keg->uk_ipers, freei, &slab->us_free); 3977 slab->us_freecount++; 3978 3979 /* Keg statistics. */ 3980 dom->ud_free++; 3981 } 3982 3983 static void 3984 zone_release(void *arg, void **bucket, int cnt) 3985 { 3986 struct mtx *lock; 3987 uma_zone_t zone; 3988 uma_slab_t slab; 3989 uma_keg_t keg; 3990 uint8_t *mem; 3991 void *item; 3992 int i; 3993 3994 zone = arg; 3995 keg = zone->uz_keg; 3996 lock = NULL; 3997 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0)) 3998 lock = KEG_LOCK(keg, 0); 3999 for (i = 0; i < cnt; i++) { 4000 item = bucket[i]; 4001 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) { 4002 slab = vtoslab((vm_offset_t)item); 4003 } else { 4004 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 4005 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0) 4006 slab = hash_sfind(&keg->uk_hash, mem); 4007 else 4008 slab = (uma_slab_t)(mem + keg->uk_pgoff); 4009 } 4010 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) { 4011 if (lock != NULL) 4012 mtx_unlock(lock); 4013 lock = KEG_LOCK(keg, slab->us_domain); 4014 } 4015 slab_free_item(zone, slab, item); 4016 } 4017 if (lock != NULL) 4018 mtx_unlock(lock); 4019 } 4020 4021 /* 4022 * Frees a single item to any zone. 4023 * 4024 * Arguments: 4025 * zone The zone to free to 4026 * item The item we're freeing 4027 * udata User supplied data for the dtor 4028 * skip Skip dtors and finis 4029 */ 4030 static void 4031 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 4032 { 4033 4034 item_dtor(zone, item, zone->uz_size, udata, skip); 4035 4036 if (skip < SKIP_FINI && zone->uz_fini) 4037 zone->uz_fini(item, zone->uz_size); 4038 4039 zone->uz_release(zone->uz_arg, &item, 1); 4040 4041 if (skip & SKIP_CNT) 4042 return; 4043 4044 counter_u64_add(zone->uz_frees, 1); 4045 4046 if (zone->uz_max_items > 0) 4047 zone_free_limit(zone, 1); 4048 } 4049 4050 /* See uma.h */ 4051 int 4052 uma_zone_set_max(uma_zone_t zone, int nitems) 4053 { 4054 struct uma_bucket_zone *ubz; 4055 int count; 4056 4057 /* 4058 * XXX This can misbehave if the zone has any allocations with 4059 * no limit and a limit is imposed. There is currently no 4060 * way to clear a limit. 4061 */ 4062 ZONE_LOCK(zone); 4063 ubz = bucket_zone_max(zone, nitems); 4064 count = ubz != NULL ? ubz->ubz_entries : 0; 4065 zone->uz_bucket_size_max = zone->uz_bucket_size = count; 4066 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max) 4067 zone->uz_bucket_size_min = zone->uz_bucket_size_max; 4068 zone->uz_max_items = nitems; 4069 zone->uz_flags |= UMA_ZFLAG_LIMIT; 4070 zone_update_caches(zone); 4071 /* We may need to wake waiters. */ 4072 wakeup(&zone->uz_max_items); 4073 ZONE_UNLOCK(zone); 4074 4075 return (nitems); 4076 } 4077 4078 /* See uma.h */ 4079 void 4080 uma_zone_set_maxcache(uma_zone_t zone, int nitems) 4081 { 4082 struct uma_bucket_zone *ubz; 4083 int bpcpu; 4084 4085 ZONE_LOCK(zone); 4086 ubz = bucket_zone_max(zone, nitems); 4087 if (ubz != NULL) { 4088 bpcpu = 2; 4089 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 4090 /* Count the cross-domain bucket. */ 4091 bpcpu++; 4092 nitems -= ubz->ubz_entries * bpcpu * mp_ncpus; 4093 zone->uz_bucket_size_max = ubz->ubz_entries; 4094 } else { 4095 zone->uz_bucket_size_max = zone->uz_bucket_size = 0; 4096 } 4097 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max) 4098 zone->uz_bucket_size_min = zone->uz_bucket_size_max; 4099 zone->uz_bkt_max = nitems; 4100 ZONE_UNLOCK(zone); 4101 } 4102 4103 /* See uma.h */ 4104 int 4105 uma_zone_get_max(uma_zone_t zone) 4106 { 4107 int nitems; 4108 4109 nitems = atomic_load_64(&zone->uz_max_items); 4110 4111 return (nitems); 4112 } 4113 4114 /* See uma.h */ 4115 void 4116 uma_zone_set_warning(uma_zone_t zone, const char *warning) 4117 { 4118 4119 ZONE_ASSERT_COLD(zone); 4120 zone->uz_warning = warning; 4121 } 4122 4123 /* See uma.h */ 4124 void 4125 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) 4126 { 4127 4128 ZONE_ASSERT_COLD(zone); 4129 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); 4130 } 4131 4132 /* See uma.h */ 4133 int 4134 uma_zone_get_cur(uma_zone_t zone) 4135 { 4136 int64_t nitems; 4137 u_int i; 4138 4139 nitems = 0; 4140 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER) 4141 nitems = counter_u64_fetch(zone->uz_allocs) - 4142 counter_u64_fetch(zone->uz_frees); 4143 CPU_FOREACH(i) 4144 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) - 4145 atomic_load_64(&zone->uz_cpu[i].uc_frees); 4146 4147 return (nitems < 0 ? 0 : nitems); 4148 } 4149 4150 static uint64_t 4151 uma_zone_get_allocs(uma_zone_t zone) 4152 { 4153 uint64_t nitems; 4154 u_int i; 4155 4156 nitems = 0; 4157 if (zone->uz_allocs != EARLY_COUNTER) 4158 nitems = counter_u64_fetch(zone->uz_allocs); 4159 CPU_FOREACH(i) 4160 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs); 4161 4162 return (nitems); 4163 } 4164 4165 static uint64_t 4166 uma_zone_get_frees(uma_zone_t zone) 4167 { 4168 uint64_t nitems; 4169 u_int i; 4170 4171 nitems = 0; 4172 if (zone->uz_frees != EARLY_COUNTER) 4173 nitems = counter_u64_fetch(zone->uz_frees); 4174 CPU_FOREACH(i) 4175 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees); 4176 4177 return (nitems); 4178 } 4179 4180 #ifdef INVARIANTS 4181 /* Used only for KEG_ASSERT_COLD(). */ 4182 static uint64_t 4183 uma_keg_get_allocs(uma_keg_t keg) 4184 { 4185 uma_zone_t z; 4186 uint64_t nitems; 4187 4188 nitems = 0; 4189 LIST_FOREACH(z, &keg->uk_zones, uz_link) 4190 nitems += uma_zone_get_allocs(z); 4191 4192 return (nitems); 4193 } 4194 #endif 4195 4196 /* See uma.h */ 4197 void 4198 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 4199 { 4200 uma_keg_t keg; 4201 4202 KEG_GET(zone, keg); 4203 KEG_ASSERT_COLD(keg); 4204 keg->uk_init = uminit; 4205 } 4206 4207 /* See uma.h */ 4208 void 4209 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 4210 { 4211 uma_keg_t keg; 4212 4213 KEG_GET(zone, keg); 4214 KEG_ASSERT_COLD(keg); 4215 keg->uk_fini = fini; 4216 } 4217 4218 /* See uma.h */ 4219 void 4220 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 4221 { 4222 4223 ZONE_ASSERT_COLD(zone); 4224 zone->uz_init = zinit; 4225 } 4226 4227 /* See uma.h */ 4228 void 4229 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 4230 { 4231 4232 ZONE_ASSERT_COLD(zone); 4233 zone->uz_fini = zfini; 4234 } 4235 4236 /* See uma.h */ 4237 void 4238 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 4239 { 4240 uma_keg_t keg; 4241 4242 KEG_GET(zone, keg); 4243 KEG_ASSERT_COLD(keg); 4244 keg->uk_freef = freef; 4245 } 4246 4247 /* See uma.h */ 4248 void 4249 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 4250 { 4251 uma_keg_t keg; 4252 4253 KEG_GET(zone, keg); 4254 KEG_ASSERT_COLD(keg); 4255 keg->uk_allocf = allocf; 4256 } 4257 4258 /* See uma.h */ 4259 void 4260 uma_zone_reserve(uma_zone_t zone, int items) 4261 { 4262 uma_keg_t keg; 4263 4264 KEG_GET(zone, keg); 4265 KEG_ASSERT_COLD(keg); 4266 keg->uk_reserve = items; 4267 } 4268 4269 /* See uma.h */ 4270 int 4271 uma_zone_reserve_kva(uma_zone_t zone, int count) 4272 { 4273 uma_keg_t keg; 4274 vm_offset_t kva; 4275 u_int pages; 4276 4277 KEG_GET(zone, keg); 4278 KEG_ASSERT_COLD(keg); 4279 ZONE_ASSERT_COLD(zone); 4280 4281 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera; 4282 4283 #ifdef UMA_MD_SMALL_ALLOC 4284 if (keg->uk_ppera > 1) { 4285 #else 4286 if (1) { 4287 #endif 4288 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE); 4289 if (kva == 0) 4290 return (0); 4291 } else 4292 kva = 0; 4293 4294 ZONE_LOCK(zone); 4295 MPASS(keg->uk_kva == 0); 4296 keg->uk_kva = kva; 4297 keg->uk_offset = 0; 4298 zone->uz_max_items = pages * keg->uk_ipers; 4299 #ifdef UMA_MD_SMALL_ALLOC 4300 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 4301 #else 4302 keg->uk_allocf = noobj_alloc; 4303 #endif 4304 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE; 4305 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE; 4306 zone_update_caches(zone); 4307 ZONE_UNLOCK(zone); 4308 4309 return (1); 4310 } 4311 4312 /* See uma.h */ 4313 void 4314 uma_prealloc(uma_zone_t zone, int items) 4315 { 4316 struct vm_domainset_iter di; 4317 uma_domain_t dom; 4318 uma_slab_t slab; 4319 uma_keg_t keg; 4320 int aflags, domain, slabs; 4321 4322 KEG_GET(zone, keg); 4323 slabs = howmany(items, keg->uk_ipers); 4324 while (slabs-- > 0) { 4325 aflags = M_NOWAIT; 4326 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 4327 &aflags); 4328 for (;;) { 4329 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK, 4330 aflags); 4331 if (slab != NULL) { 4332 dom = &keg->uk_domain[slab->us_domain]; 4333 LIST_REMOVE(slab, us_link); 4334 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, 4335 us_link); 4336 KEG_UNLOCK(keg, slab->us_domain); 4337 break; 4338 } 4339 if (vm_domainset_iter_policy(&di, &domain) != 0) 4340 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask); 4341 } 4342 } 4343 } 4344 4345 /* See uma.h */ 4346 void 4347 uma_reclaim(int req) 4348 { 4349 4350 CTR0(KTR_UMA, "UMA: vm asked us to release pages!"); 4351 sx_xlock(&uma_reclaim_lock); 4352 bucket_enable(); 4353 4354 switch (req) { 4355 case UMA_RECLAIM_TRIM: 4356 zone_foreach(zone_trim, NULL); 4357 break; 4358 case UMA_RECLAIM_DRAIN: 4359 case UMA_RECLAIM_DRAIN_CPU: 4360 zone_foreach(zone_drain, NULL); 4361 if (req == UMA_RECLAIM_DRAIN_CPU) { 4362 pcpu_cache_drain_safe(NULL); 4363 zone_foreach(zone_drain, NULL); 4364 } 4365 break; 4366 default: 4367 panic("unhandled reclamation request %d", req); 4368 } 4369 4370 /* 4371 * Some slabs may have been freed but this zone will be visited early 4372 * we visit again so that we can free pages that are empty once other 4373 * zones are drained. We have to do the same for buckets. 4374 */ 4375 zone_drain(slabzones[0], NULL); 4376 zone_drain(slabzones[1], NULL); 4377 bucket_zone_drain(); 4378 sx_xunlock(&uma_reclaim_lock); 4379 } 4380 4381 static volatile int uma_reclaim_needed; 4382 4383 void 4384 uma_reclaim_wakeup(void) 4385 { 4386 4387 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0) 4388 wakeup(uma_reclaim); 4389 } 4390 4391 void 4392 uma_reclaim_worker(void *arg __unused) 4393 { 4394 4395 for (;;) { 4396 sx_xlock(&uma_reclaim_lock); 4397 while (atomic_load_int(&uma_reclaim_needed) == 0) 4398 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl", 4399 hz); 4400 sx_xunlock(&uma_reclaim_lock); 4401 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); 4402 uma_reclaim(UMA_RECLAIM_DRAIN_CPU); 4403 atomic_store_int(&uma_reclaim_needed, 0); 4404 /* Don't fire more than once per-second. */ 4405 pause("umarclslp", hz); 4406 } 4407 } 4408 4409 /* See uma.h */ 4410 void 4411 uma_zone_reclaim(uma_zone_t zone, int req) 4412 { 4413 4414 switch (req) { 4415 case UMA_RECLAIM_TRIM: 4416 zone_trim(zone, NULL); 4417 break; 4418 case UMA_RECLAIM_DRAIN: 4419 zone_drain(zone, NULL); 4420 break; 4421 case UMA_RECLAIM_DRAIN_CPU: 4422 pcpu_cache_drain_safe(zone); 4423 zone_drain(zone, NULL); 4424 break; 4425 default: 4426 panic("unhandled reclamation request %d", req); 4427 } 4428 } 4429 4430 /* See uma.h */ 4431 int 4432 uma_zone_exhausted(uma_zone_t zone) 4433 { 4434 4435 return (atomic_load_32(&zone->uz_sleepers) > 0); 4436 } 4437 4438 unsigned long 4439 uma_limit(void) 4440 { 4441 4442 return (uma_kmem_limit); 4443 } 4444 4445 void 4446 uma_set_limit(unsigned long limit) 4447 { 4448 4449 uma_kmem_limit = limit; 4450 } 4451 4452 unsigned long 4453 uma_size(void) 4454 { 4455 4456 return (atomic_load_long(&uma_kmem_total)); 4457 } 4458 4459 long 4460 uma_avail(void) 4461 { 4462 4463 return (uma_kmem_limit - uma_size()); 4464 } 4465 4466 #ifdef DDB 4467 /* 4468 * Generate statistics across both the zone and its per-cpu cache's. Return 4469 * desired statistics if the pointer is non-NULL for that statistic. 4470 * 4471 * Note: does not update the zone statistics, as it can't safely clear the 4472 * per-CPU cache statistic. 4473 * 4474 */ 4475 static void 4476 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp, 4477 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp) 4478 { 4479 uma_cache_t cache; 4480 uint64_t allocs, frees, sleeps, xdomain; 4481 int cachefree, cpu; 4482 4483 allocs = frees = sleeps = xdomain = 0; 4484 cachefree = 0; 4485 CPU_FOREACH(cpu) { 4486 cache = &z->uz_cpu[cpu]; 4487 cachefree += cache->uc_allocbucket.ucb_cnt; 4488 cachefree += cache->uc_freebucket.ucb_cnt; 4489 xdomain += cache->uc_crossbucket.ucb_cnt; 4490 cachefree += cache->uc_crossbucket.ucb_cnt; 4491 allocs += cache->uc_allocs; 4492 frees += cache->uc_frees; 4493 } 4494 allocs += counter_u64_fetch(z->uz_allocs); 4495 frees += counter_u64_fetch(z->uz_frees); 4496 sleeps += z->uz_sleeps; 4497 xdomain += z->uz_xdomain; 4498 if (cachefreep != NULL) 4499 *cachefreep = cachefree; 4500 if (allocsp != NULL) 4501 *allocsp = allocs; 4502 if (freesp != NULL) 4503 *freesp = frees; 4504 if (sleepsp != NULL) 4505 *sleepsp = sleeps; 4506 if (xdomainp != NULL) 4507 *xdomainp = xdomain; 4508 } 4509 #endif /* DDB */ 4510 4511 static int 4512 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 4513 { 4514 uma_keg_t kz; 4515 uma_zone_t z; 4516 int count; 4517 4518 count = 0; 4519 rw_rlock(&uma_rwlock); 4520 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4521 LIST_FOREACH(z, &kz->uk_zones, uz_link) 4522 count++; 4523 } 4524 LIST_FOREACH(z, &uma_cachezones, uz_link) 4525 count++; 4526 4527 rw_runlock(&uma_rwlock); 4528 return (sysctl_handle_int(oidp, &count, 0, req)); 4529 } 4530 4531 static void 4532 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf, 4533 struct uma_percpu_stat *ups, bool internal) 4534 { 4535 uma_zone_domain_t zdom; 4536 uma_cache_t cache; 4537 int i; 4538 4539 4540 for (i = 0; i < vm_ndomains; i++) { 4541 zdom = &z->uz_domain[i]; 4542 uth->uth_zone_free += zdom->uzd_nitems; 4543 } 4544 uth->uth_allocs = counter_u64_fetch(z->uz_allocs); 4545 uth->uth_frees = counter_u64_fetch(z->uz_frees); 4546 uth->uth_fails = counter_u64_fetch(z->uz_fails); 4547 uth->uth_sleeps = z->uz_sleeps; 4548 uth->uth_xdomain = z->uz_xdomain; 4549 4550 /* 4551 * While it is not normally safe to access the cache bucket pointers 4552 * while not on the CPU that owns the cache, we only allow the pointers 4553 * to be exchanged without the zone lock held, not invalidated, so 4554 * accept the possible race associated with bucket exchange during 4555 * monitoring. Use atomic_load_ptr() to ensure that the bucket pointers 4556 * are loaded only once. 4557 */ 4558 for (i = 0; i < mp_maxid + 1; i++) { 4559 bzero(&ups[i], sizeof(*ups)); 4560 if (internal || CPU_ABSENT(i)) 4561 continue; 4562 cache = &z->uz_cpu[i]; 4563 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt; 4564 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt; 4565 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt; 4566 ups[i].ups_allocs = cache->uc_allocs; 4567 ups[i].ups_frees = cache->uc_frees; 4568 } 4569 } 4570 4571 static int 4572 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 4573 { 4574 struct uma_stream_header ush; 4575 struct uma_type_header uth; 4576 struct uma_percpu_stat *ups; 4577 struct sbuf sbuf; 4578 uma_keg_t kz; 4579 uma_zone_t z; 4580 uint64_t items; 4581 uint32_t kfree, pages; 4582 int count, error, i; 4583 4584 error = sysctl_wire_old_buffer(req, 0); 4585 if (error != 0) 4586 return (error); 4587 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 4588 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); 4589 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK); 4590 4591 count = 0; 4592 rw_rlock(&uma_rwlock); 4593 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4594 LIST_FOREACH(z, &kz->uk_zones, uz_link) 4595 count++; 4596 } 4597 4598 LIST_FOREACH(z, &uma_cachezones, uz_link) 4599 count++; 4600 4601 /* 4602 * Insert stream header. 4603 */ 4604 bzero(&ush, sizeof(ush)); 4605 ush.ush_version = UMA_STREAM_VERSION; 4606 ush.ush_maxcpus = (mp_maxid + 1); 4607 ush.ush_count = count; 4608 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 4609 4610 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4611 kfree = pages = 0; 4612 for (i = 0; i < vm_ndomains; i++) { 4613 kfree += kz->uk_domain[i].ud_free; 4614 pages += kz->uk_domain[i].ud_pages; 4615 } 4616 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 4617 bzero(&uth, sizeof(uth)); 4618 ZONE_LOCK(z); 4619 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 4620 uth.uth_align = kz->uk_align; 4621 uth.uth_size = kz->uk_size; 4622 uth.uth_rsize = kz->uk_rsize; 4623 if (z->uz_max_items > 0) { 4624 items = UZ_ITEMS_COUNT(z->uz_items); 4625 uth.uth_pages = (items / kz->uk_ipers) * 4626 kz->uk_ppera; 4627 } else 4628 uth.uth_pages = pages; 4629 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) * 4630 kz->uk_ppera; 4631 uth.uth_limit = z->uz_max_items; 4632 uth.uth_keg_free = kfree; 4633 4634 /* 4635 * A zone is secondary is it is not the first entry 4636 * on the keg's zone list. 4637 */ 4638 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 4639 (LIST_FIRST(&kz->uk_zones) != z)) 4640 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 4641 uma_vm_zone_stats(&uth, z, &sbuf, ups, 4642 kz->uk_flags & UMA_ZFLAG_INTERNAL); 4643 ZONE_UNLOCK(z); 4644 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 4645 for (i = 0; i < mp_maxid + 1; i++) 4646 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 4647 } 4648 } 4649 LIST_FOREACH(z, &uma_cachezones, uz_link) { 4650 bzero(&uth, sizeof(uth)); 4651 ZONE_LOCK(z); 4652 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 4653 uth.uth_size = z->uz_size; 4654 uma_vm_zone_stats(&uth, z, &sbuf, ups, false); 4655 ZONE_UNLOCK(z); 4656 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 4657 for (i = 0; i < mp_maxid + 1; i++) 4658 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 4659 } 4660 4661 rw_runlock(&uma_rwlock); 4662 error = sbuf_finish(&sbuf); 4663 sbuf_delete(&sbuf); 4664 free(ups, M_TEMP); 4665 return (error); 4666 } 4667 4668 int 4669 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 4670 { 4671 uma_zone_t zone = *(uma_zone_t *)arg1; 4672 int error, max; 4673 4674 max = uma_zone_get_max(zone); 4675 error = sysctl_handle_int(oidp, &max, 0, req); 4676 if (error || !req->newptr) 4677 return (error); 4678 4679 uma_zone_set_max(zone, max); 4680 4681 return (0); 4682 } 4683 4684 int 4685 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 4686 { 4687 uma_zone_t zone; 4688 int cur; 4689 4690 /* 4691 * Some callers want to add sysctls for global zones that 4692 * may not yet exist so they pass a pointer to a pointer. 4693 */ 4694 if (arg2 == 0) 4695 zone = *(uma_zone_t *)arg1; 4696 else 4697 zone = arg1; 4698 cur = uma_zone_get_cur(zone); 4699 return (sysctl_handle_int(oidp, &cur, 0, req)); 4700 } 4701 4702 static int 4703 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS) 4704 { 4705 uma_zone_t zone = arg1; 4706 uint64_t cur; 4707 4708 cur = uma_zone_get_allocs(zone); 4709 return (sysctl_handle_64(oidp, &cur, 0, req)); 4710 } 4711 4712 static int 4713 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS) 4714 { 4715 uma_zone_t zone = arg1; 4716 uint64_t cur; 4717 4718 cur = uma_zone_get_frees(zone); 4719 return (sysctl_handle_64(oidp, &cur, 0, req)); 4720 } 4721 4722 static int 4723 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS) 4724 { 4725 struct sbuf sbuf; 4726 uma_zone_t zone = arg1; 4727 int error; 4728 4729 sbuf_new_for_sysctl(&sbuf, NULL, 0, req); 4730 if (zone->uz_flags != 0) 4731 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS); 4732 else 4733 sbuf_printf(&sbuf, "0"); 4734 error = sbuf_finish(&sbuf); 4735 sbuf_delete(&sbuf); 4736 4737 return (error); 4738 } 4739 4740 static int 4741 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS) 4742 { 4743 uma_keg_t keg = arg1; 4744 int avail, effpct, total; 4745 4746 total = keg->uk_ppera * PAGE_SIZE; 4747 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0) 4748 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize; 4749 /* 4750 * We consider the client's requested size and alignment here, not the 4751 * real size determination uk_rsize, because we also adjust the real 4752 * size for internal implementation reasons (max bitset size). 4753 */ 4754 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1); 4755 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) 4756 avail *= mp_maxid + 1; 4757 effpct = 100 * avail / total; 4758 return (sysctl_handle_int(oidp, &effpct, 0, req)); 4759 } 4760 4761 static int 4762 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS) 4763 { 4764 uma_zone_t zone = arg1; 4765 uint64_t cur; 4766 4767 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items)); 4768 return (sysctl_handle_64(oidp, &cur, 0, req)); 4769 } 4770 4771 #ifdef INVARIANTS 4772 static uma_slab_t 4773 uma_dbg_getslab(uma_zone_t zone, void *item) 4774 { 4775 uma_slab_t slab; 4776 uma_keg_t keg; 4777 uint8_t *mem; 4778 4779 /* 4780 * It is safe to return the slab here even though the 4781 * zone is unlocked because the item's allocation state 4782 * essentially holds a reference. 4783 */ 4784 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 4785 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 4786 return (NULL); 4787 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB) 4788 return (vtoslab((vm_offset_t)mem)); 4789 keg = zone->uz_keg; 4790 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0) 4791 return ((uma_slab_t)(mem + keg->uk_pgoff)); 4792 KEG_LOCK(keg, 0); 4793 slab = hash_sfind(&keg->uk_hash, mem); 4794 KEG_UNLOCK(keg, 0); 4795 4796 return (slab); 4797 } 4798 4799 static bool 4800 uma_dbg_zskip(uma_zone_t zone, void *mem) 4801 { 4802 4803 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 4804 return (true); 4805 4806 return (uma_dbg_kskip(zone->uz_keg, mem)); 4807 } 4808 4809 static bool 4810 uma_dbg_kskip(uma_keg_t keg, void *mem) 4811 { 4812 uintptr_t idx; 4813 4814 if (dbg_divisor == 0) 4815 return (true); 4816 4817 if (dbg_divisor == 1) 4818 return (false); 4819 4820 idx = (uintptr_t)mem >> PAGE_SHIFT; 4821 if (keg->uk_ipers > 1) { 4822 idx *= keg->uk_ipers; 4823 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize; 4824 } 4825 4826 if ((idx / dbg_divisor) * dbg_divisor != idx) { 4827 counter_u64_add(uma_skip_cnt, 1); 4828 return (true); 4829 } 4830 counter_u64_add(uma_dbg_cnt, 1); 4831 4832 return (false); 4833 } 4834 4835 /* 4836 * Set up the slab's freei data such that uma_dbg_free can function. 4837 * 4838 */ 4839 static void 4840 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) 4841 { 4842 uma_keg_t keg; 4843 int freei; 4844 4845 if (slab == NULL) { 4846 slab = uma_dbg_getslab(zone, item); 4847 if (slab == NULL) 4848 panic("uma: item %p did not belong to zone %s\n", 4849 item, zone->uz_name); 4850 } 4851 keg = zone->uz_keg; 4852 freei = slab_item_index(slab, keg, item); 4853 4854 if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg))) 4855 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n", 4856 item, zone, zone->uz_name, slab, freei); 4857 BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)); 4858 } 4859 4860 /* 4861 * Verifies freed addresses. Checks for alignment, valid slab membership 4862 * and duplicate frees. 4863 * 4864 */ 4865 static void 4866 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) 4867 { 4868 uma_keg_t keg; 4869 int freei; 4870 4871 if (slab == NULL) { 4872 slab = uma_dbg_getslab(zone, item); 4873 if (slab == NULL) 4874 panic("uma: Freed item %p did not belong to zone %s\n", 4875 item, zone->uz_name); 4876 } 4877 keg = zone->uz_keg; 4878 freei = slab_item_index(slab, keg, item); 4879 4880 if (freei >= keg->uk_ipers) 4881 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n", 4882 item, zone, zone->uz_name, slab, freei); 4883 4884 if (slab_item(slab, keg, freei) != item) 4885 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n", 4886 item, zone, zone->uz_name, slab, freei); 4887 4888 if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg))) 4889 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n", 4890 item, zone, zone->uz_name, slab, freei); 4891 4892 BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)); 4893 } 4894 #endif /* INVARIANTS */ 4895 4896 #ifdef DDB 4897 static int64_t 4898 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used, 4899 uint64_t *sleeps, long *cachefree, uint64_t *xdomain) 4900 { 4901 uint64_t frees; 4902 int i; 4903 4904 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 4905 *allocs = counter_u64_fetch(z->uz_allocs); 4906 frees = counter_u64_fetch(z->uz_frees); 4907 *sleeps = z->uz_sleeps; 4908 *cachefree = 0; 4909 *xdomain = 0; 4910 } else 4911 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps, 4912 xdomain); 4913 for (i = 0; i < vm_ndomains; i++) { 4914 *cachefree += z->uz_domain[i].uzd_nitems; 4915 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 4916 (LIST_FIRST(&kz->uk_zones) != z))) 4917 *cachefree += kz->uk_domain[i].ud_free; 4918 } 4919 *used = *allocs - frees; 4920 return (((int64_t)*used + *cachefree) * kz->uk_size); 4921 } 4922 4923 DB_SHOW_COMMAND(uma, db_show_uma) 4924 { 4925 const char *fmt_hdr, *fmt_entry; 4926 uma_keg_t kz; 4927 uma_zone_t z; 4928 uint64_t allocs, used, sleeps, xdomain; 4929 long cachefree; 4930 /* variables for sorting */ 4931 uma_keg_t cur_keg; 4932 uma_zone_t cur_zone, last_zone; 4933 int64_t cur_size, last_size, size; 4934 int ties; 4935 4936 /* /i option produces machine-parseable CSV output */ 4937 if (modif[0] == 'i') { 4938 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n"; 4939 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n"; 4940 } else { 4941 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n"; 4942 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n"; 4943 } 4944 4945 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests", 4946 "Sleeps", "Bucket", "Total Mem", "XFree"); 4947 4948 /* Sort the zones with largest size first. */ 4949 last_zone = NULL; 4950 last_size = INT64_MAX; 4951 for (;;) { 4952 cur_zone = NULL; 4953 cur_size = -1; 4954 ties = 0; 4955 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4956 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 4957 /* 4958 * In the case of size ties, print out zones 4959 * in the order they are encountered. That is, 4960 * when we encounter the most recently output 4961 * zone, we have already printed all preceding 4962 * ties, and we must print all following ties. 4963 */ 4964 if (z == last_zone) { 4965 ties = 1; 4966 continue; 4967 } 4968 size = get_uma_stats(kz, z, &allocs, &used, 4969 &sleeps, &cachefree, &xdomain); 4970 if (size > cur_size && size < last_size + ties) 4971 { 4972 cur_size = size; 4973 cur_zone = z; 4974 cur_keg = kz; 4975 } 4976 } 4977 } 4978 if (cur_zone == NULL) 4979 break; 4980 4981 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used, 4982 &sleeps, &cachefree, &xdomain); 4983 db_printf(fmt_entry, cur_zone->uz_name, 4984 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree, 4985 (uintmax_t)allocs, (uintmax_t)sleeps, 4986 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size, 4987 xdomain); 4988 4989 if (db_pager_quit) 4990 return; 4991 last_zone = cur_zone; 4992 last_size = cur_size; 4993 } 4994 } 4995 4996 DB_SHOW_COMMAND(umacache, db_show_umacache) 4997 { 4998 uma_zone_t z; 4999 uint64_t allocs, frees; 5000 long cachefree; 5001 int i; 5002 5003 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 5004 "Requests", "Bucket"); 5005 LIST_FOREACH(z, &uma_cachezones, uz_link) { 5006 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL); 5007 for (i = 0; i < vm_ndomains; i++) 5008 cachefree += z->uz_domain[i].uzd_nitems; 5009 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n", 5010 z->uz_name, (uintmax_t)z->uz_size, 5011 (intmax_t)(allocs - frees), cachefree, 5012 (uintmax_t)allocs, z->uz_bucket_size); 5013 if (db_pager_quit) 5014 return; 5015 } 5016 } 5017 #endif /* DDB */ 5018