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/asan.h> 62 #include <sys/bitset.h> 63 #include <sys/domainset.h> 64 #include <sys/eventhandler.h> 65 #include <sys/kernel.h> 66 #include <sys/types.h> 67 #include <sys/limits.h> 68 #include <sys/queue.h> 69 #include <sys/malloc.h> 70 #include <sys/ktr.h> 71 #include <sys/lock.h> 72 #include <sys/sysctl.h> 73 #include <sys/mutex.h> 74 #include <sys/proc.h> 75 #include <sys/random.h> 76 #include <sys/rwlock.h> 77 #include <sys/sbuf.h> 78 #include <sys/sched.h> 79 #include <sys/sleepqueue.h> 80 #include <sys/smp.h> 81 #include <sys/smr.h> 82 #include <sys/taskqueue.h> 83 #include <sys/vmmeter.h> 84 85 #include <vm/vm.h> 86 #include <vm/vm_param.h> 87 #include <vm/vm_domainset.h> 88 #include <vm/vm_object.h> 89 #include <vm/vm_page.h> 90 #include <vm/vm_pageout.h> 91 #include <vm/vm_phys.h> 92 #include <vm/vm_pagequeue.h> 93 #include <vm/vm_map.h> 94 #include <vm/vm_kern.h> 95 #include <vm/vm_extern.h> 96 #include <vm/vm_dumpset.h> 97 #include <vm/uma.h> 98 #include <vm/uma_int.h> 99 #include <vm/uma_dbg.h> 100 101 #include <ddb/ddb.h> 102 103 #ifdef DEBUG_MEMGUARD 104 #include <vm/memguard.h> 105 #endif 106 107 #include <machine/md_var.h> 108 109 #ifdef INVARIANTS 110 #define UMA_ALWAYS_CTORDTOR 1 111 #else 112 #define UMA_ALWAYS_CTORDTOR 0 113 #endif 114 115 /* 116 * This is the zone and keg from which all zones are spawned. 117 */ 118 static uma_zone_t kegs; 119 static uma_zone_t zones; 120 121 /* 122 * On INVARIANTS builds, the slab contains a second bitset of the same size, 123 * "dbg_bits", which is laid out immediately after us_free. 124 */ 125 #ifdef INVARIANTS 126 #define SLAB_BITSETS 2 127 #else 128 #define SLAB_BITSETS 1 129 #endif 130 131 /* 132 * These are the two zones from which all offpage uma_slab_ts are allocated. 133 * 134 * One zone is for slab headers that can represent a larger number of items, 135 * making the slabs themselves more efficient, and the other zone is for 136 * headers that are smaller and represent fewer items, making the headers more 137 * efficient. 138 */ 139 #define SLABZONE_SIZE(setsize) \ 140 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS) 141 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16) 142 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE 143 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE) 144 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE) 145 static uma_zone_t slabzones[2]; 146 147 /* 148 * The initial hash tables come out of this zone so they can be allocated 149 * prior to malloc coming up. 150 */ 151 static uma_zone_t hashzone; 152 153 /* The boot-time adjusted value for cache line alignment. */ 154 int uma_align_cache = 64 - 1; 155 156 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 157 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc"); 158 159 /* 160 * Are we allowed to allocate buckets? 161 */ 162 static int bucketdisable = 1; 163 164 /* Linked list of all kegs in the system */ 165 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 166 167 /* Linked list of all cache-only zones in the system */ 168 static LIST_HEAD(,uma_zone) uma_cachezones = 169 LIST_HEAD_INITIALIZER(uma_cachezones); 170 171 /* 172 * Mutex for global lists: uma_kegs, uma_cachezones, and the per-keg list of 173 * zones. 174 */ 175 static struct rwlock_padalign __exclusive_cache_line uma_rwlock; 176 177 static struct sx uma_reclaim_lock; 178 179 /* 180 * First available virual address for boot time allocations. 181 */ 182 static vm_offset_t bootstart; 183 static vm_offset_t bootmem; 184 185 /* 186 * kmem soft limit, initialized by uma_set_limit(). Ensure that early 187 * allocations don't trigger a wakeup of the reclaim thread. 188 */ 189 unsigned long uma_kmem_limit = LONG_MAX; 190 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0, 191 "UMA kernel memory soft limit"); 192 unsigned long uma_kmem_total; 193 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0, 194 "UMA kernel memory usage"); 195 196 /* Is the VM done starting up? */ 197 static enum { 198 BOOT_COLD, 199 BOOT_KVA, 200 BOOT_PCPU, 201 BOOT_RUNNING, 202 BOOT_SHUTDOWN, 203 } booted = BOOT_COLD; 204 205 /* 206 * This is the handle used to schedule events that need to happen 207 * outside of the allocation fast path. 208 */ 209 static struct callout uma_callout; 210 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 211 212 /* 213 * This structure is passed as the zone ctor arg so that I don't have to create 214 * a special allocation function just for zones. 215 */ 216 struct uma_zctor_args { 217 const char *name; 218 size_t size; 219 uma_ctor ctor; 220 uma_dtor dtor; 221 uma_init uminit; 222 uma_fini fini; 223 uma_import import; 224 uma_release release; 225 void *arg; 226 uma_keg_t keg; 227 int align; 228 uint32_t flags; 229 }; 230 231 struct uma_kctor_args { 232 uma_zone_t zone; 233 size_t size; 234 uma_init uminit; 235 uma_fini fini; 236 int align; 237 uint32_t flags; 238 }; 239 240 struct uma_bucket_zone { 241 uma_zone_t ubz_zone; 242 const char *ubz_name; 243 int ubz_entries; /* Number of items it can hold. */ 244 int ubz_maxsize; /* Maximum allocation size per-item. */ 245 }; 246 247 /* 248 * Compute the actual number of bucket entries to pack them in power 249 * of two sizes for more efficient space utilization. 250 */ 251 #define BUCKET_SIZE(n) \ 252 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 253 254 #define BUCKET_MAX BUCKET_SIZE(256) 255 256 struct uma_bucket_zone bucket_zones[] = { 257 /* Literal bucket sizes. */ 258 { NULL, "2 Bucket", 2, 4096 }, 259 { NULL, "4 Bucket", 4, 3072 }, 260 { NULL, "8 Bucket", 8, 2048 }, 261 { NULL, "16 Bucket", 16, 1024 }, 262 /* Rounded down power of 2 sizes for efficiency. */ 263 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 264 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 265 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 266 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 267 { NULL, NULL, 0} 268 }; 269 270 /* 271 * Flags and enumerations to be passed to internal functions. 272 */ 273 enum zfreeskip { 274 SKIP_NONE = 0, 275 SKIP_CNT = 0x00000001, 276 SKIP_DTOR = 0x00010000, 277 SKIP_FINI = 0x00020000, 278 }; 279 280 /* Prototypes.. */ 281 282 void uma_startup1(vm_offset_t); 283 void uma_startup2(void); 284 285 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 286 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 287 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 288 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 289 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 290 static void page_free(void *, vm_size_t, uint8_t); 291 static void pcpu_page_free(void *, vm_size_t, uint8_t); 292 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int); 293 static void cache_drain(uma_zone_t); 294 static void bucket_drain(uma_zone_t, uma_bucket_t); 295 static void bucket_cache_reclaim(uma_zone_t zone, bool, int); 296 static bool bucket_cache_reclaim_domain(uma_zone_t, bool, bool, int); 297 static int keg_ctor(void *, int, void *, int); 298 static void keg_dtor(void *, int, void *); 299 static void keg_drain(uma_keg_t keg, int domain); 300 static int zone_ctor(void *, int, void *, int); 301 static void zone_dtor(void *, int, void *); 302 static inline void item_dtor(uma_zone_t zone, void *item, int size, 303 void *udata, enum zfreeskip skip); 304 static int zero_init(void *, int, int); 305 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata, 306 int itemdomain, bool ws); 307 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *); 308 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *); 309 static void zone_timeout(uma_zone_t zone, void *); 310 static int hash_alloc(struct uma_hash *, u_int); 311 static int hash_expand(struct uma_hash *, struct uma_hash *); 312 static void hash_free(struct uma_hash *hash); 313 static void uma_timeout(void *); 314 static void uma_shutdown(void); 315 static void *zone_alloc_item(uma_zone_t, void *, int, int); 316 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 317 static int zone_alloc_limit(uma_zone_t zone, int count, int flags); 318 static void zone_free_limit(uma_zone_t zone, int count); 319 static void bucket_enable(void); 320 static void bucket_init(void); 321 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 322 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 323 static void bucket_zone_drain(int domain); 324 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int); 325 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 326 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item); 327 static size_t slab_sizeof(int nitems); 328 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 329 uma_fini fini, int align, uint32_t flags); 330 static int zone_import(void *, void **, int, int, int); 331 static void zone_release(void *, void **, int); 332 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int); 333 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int); 334 335 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 336 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 337 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS); 338 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS); 339 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS); 340 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS); 341 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS); 342 343 static uint64_t uma_zone_get_allocs(uma_zone_t zone); 344 345 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 346 "Memory allocation debugging"); 347 348 #ifdef INVARIANTS 349 static uint64_t uma_keg_get_allocs(uma_keg_t zone); 350 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg); 351 352 static bool uma_dbg_kskip(uma_keg_t keg, void *mem); 353 static bool uma_dbg_zskip(uma_zone_t zone, void *mem); 354 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item); 355 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item); 356 357 static u_int dbg_divisor = 1; 358 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor, 359 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0, 360 "Debug & thrash every this item in memory allocator"); 361 362 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER; 363 static counter_u64_t uma_skip_cnt = EARLY_COUNTER; 364 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD, 365 &uma_dbg_cnt, "memory items debugged"); 366 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD, 367 &uma_skip_cnt, "memory items skipped, not debugged"); 368 #endif 369 370 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 371 "Universal Memory Allocator"); 372 373 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT, 374 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 375 376 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT, 377 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 378 379 static int zone_warnings = 1; 380 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, 381 "Warn when UMA zones becomes full"); 382 383 static int multipage_slabs = 1; 384 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs); 385 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs, 386 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0, 387 "UMA may choose larger slab sizes for better efficiency"); 388 389 /* 390 * Select the slab zone for an offpage slab with the given maximum item count. 391 */ 392 static inline uma_zone_t 393 slabzone(int ipers) 394 { 395 396 return (slabzones[ipers > SLABZONE0_SETSIZE]); 397 } 398 399 /* 400 * This routine checks to see whether or not it's safe to enable buckets. 401 */ 402 static void 403 bucket_enable(void) 404 { 405 406 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init")); 407 bucketdisable = vm_page_count_min(); 408 } 409 410 /* 411 * Initialize bucket_zones, the array of zones of buckets of various sizes. 412 * 413 * For each zone, calculate the memory required for each bucket, consisting 414 * of the header and an array of pointers. 415 */ 416 static void 417 bucket_init(void) 418 { 419 struct uma_bucket_zone *ubz; 420 int size; 421 422 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 423 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 424 size += sizeof(void *) * ubz->ubz_entries; 425 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 426 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 427 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET | 428 UMA_ZONE_FIRSTTOUCH); 429 } 430 } 431 432 /* 433 * Given a desired number of entries for a bucket, return the zone from which 434 * to allocate the bucket. 435 */ 436 static struct uma_bucket_zone * 437 bucket_zone_lookup(int entries) 438 { 439 struct uma_bucket_zone *ubz; 440 441 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 442 if (ubz->ubz_entries >= entries) 443 return (ubz); 444 ubz--; 445 return (ubz); 446 } 447 448 static int 449 bucket_select(int size) 450 { 451 struct uma_bucket_zone *ubz; 452 453 ubz = &bucket_zones[0]; 454 if (size > ubz->ubz_maxsize) 455 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 456 457 for (; ubz->ubz_entries != 0; ubz++) 458 if (ubz->ubz_maxsize < size) 459 break; 460 ubz--; 461 return (ubz->ubz_entries); 462 } 463 464 static uma_bucket_t 465 bucket_alloc(uma_zone_t zone, void *udata, int flags) 466 { 467 struct uma_bucket_zone *ubz; 468 uma_bucket_t bucket; 469 470 /* 471 * Don't allocate buckets early in boot. 472 */ 473 if (__predict_false(booted < BOOT_KVA)) 474 return (NULL); 475 476 /* 477 * To limit bucket recursion we store the original zone flags 478 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 479 * NOVM flag to persist even through deep recursions. We also 480 * store ZFLAG_BUCKET once we have recursed attempting to allocate 481 * a bucket for a bucket zone so we do not allow infinite bucket 482 * recursion. This cookie will even persist to frees of unused 483 * buckets via the allocation path or bucket allocations in the 484 * free path. 485 */ 486 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 487 udata = (void *)(uintptr_t)zone->uz_flags; 488 else { 489 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 490 return (NULL); 491 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 492 } 493 if (((uintptr_t)udata & UMA_ZONE_VM) != 0) 494 flags |= M_NOVM; 495 ubz = bucket_zone_lookup(atomic_load_16(&zone->uz_bucket_size)); 496 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 497 ubz++; 498 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 499 if (bucket) { 500 #ifdef INVARIANTS 501 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 502 #endif 503 bucket->ub_cnt = 0; 504 bucket->ub_entries = min(ubz->ubz_entries, 505 zone->uz_bucket_size_max); 506 bucket->ub_seq = SMR_SEQ_INVALID; 507 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p", 508 zone->uz_name, zone, bucket); 509 } 510 511 return (bucket); 512 } 513 514 static void 515 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 516 { 517 struct uma_bucket_zone *ubz; 518 519 if (bucket->ub_cnt != 0) 520 bucket_drain(zone, bucket); 521 522 KASSERT(bucket->ub_cnt == 0, 523 ("bucket_free: Freeing a non free bucket.")); 524 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID, 525 ("bucket_free: Freeing an SMR bucket.")); 526 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 527 udata = (void *)(uintptr_t)zone->uz_flags; 528 ubz = bucket_zone_lookup(bucket->ub_entries); 529 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 530 } 531 532 static void 533 bucket_zone_drain(int domain) 534 { 535 struct uma_bucket_zone *ubz; 536 537 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 538 uma_zone_reclaim_domain(ubz->ubz_zone, UMA_RECLAIM_DRAIN, 539 domain); 540 } 541 542 #ifdef KASAN 543 _Static_assert(UMA_SMALLEST_UNIT % KASAN_SHADOW_SCALE == 0, 544 "Base UMA allocation size not a multiple of the KASAN scale factor"); 545 546 static void 547 kasan_mark_item_valid(uma_zone_t zone, void *item) 548 { 549 void *pcpu_item; 550 size_t sz, rsz; 551 int i; 552 553 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0) 554 return; 555 556 sz = zone->uz_size; 557 rsz = roundup2(sz, KASAN_SHADOW_SCALE); 558 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) { 559 kasan_mark(item, sz, rsz, 0); 560 } else { 561 pcpu_item = zpcpu_base_to_offset(item); 562 for (i = 0; i <= mp_maxid; i++) 563 kasan_mark(zpcpu_get_cpu(pcpu_item, i), sz, rsz, 0); 564 } 565 } 566 567 static void 568 kasan_mark_item_invalid(uma_zone_t zone, void *item) 569 { 570 void *pcpu_item; 571 size_t sz; 572 int i; 573 574 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0) 575 return; 576 577 sz = roundup2(zone->uz_size, KASAN_SHADOW_SCALE); 578 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) { 579 kasan_mark(item, 0, sz, KASAN_UMA_FREED); 580 } else { 581 pcpu_item = zpcpu_base_to_offset(item); 582 for (i = 0; i <= mp_maxid; i++) 583 kasan_mark(zpcpu_get_cpu(pcpu_item, i), 0, sz, 0); 584 } 585 } 586 587 static void 588 kasan_mark_slab_valid(uma_keg_t keg, void *mem) 589 { 590 size_t sz; 591 592 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) { 593 sz = keg->uk_ppera * PAGE_SIZE; 594 kasan_mark(mem, sz, sz, 0); 595 } 596 } 597 598 static void 599 kasan_mark_slab_invalid(uma_keg_t keg, void *mem) 600 { 601 size_t sz; 602 603 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) { 604 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0) 605 sz = keg->uk_ppera * PAGE_SIZE; 606 else 607 sz = keg->uk_pgoff; 608 kasan_mark(mem, 0, sz, KASAN_UMA_FREED); 609 } 610 } 611 #else /* !KASAN */ 612 static void 613 kasan_mark_item_valid(uma_zone_t zone __unused, void *item __unused) 614 { 615 } 616 617 static void 618 kasan_mark_item_invalid(uma_zone_t zone __unused, void *item __unused) 619 { 620 } 621 622 static void 623 kasan_mark_slab_valid(uma_keg_t keg __unused, void *mem __unused) 624 { 625 } 626 627 static void 628 kasan_mark_slab_invalid(uma_keg_t keg __unused, void *mem __unused) 629 { 630 } 631 #endif /* KASAN */ 632 633 /* 634 * Acquire the domain lock and record contention. 635 */ 636 static uma_zone_domain_t 637 zone_domain_lock(uma_zone_t zone, int domain) 638 { 639 uma_zone_domain_t zdom; 640 bool lockfail; 641 642 zdom = ZDOM_GET(zone, domain); 643 lockfail = false; 644 if (ZDOM_OWNED(zdom)) 645 lockfail = true; 646 ZDOM_LOCK(zdom); 647 /* This is unsynchronized. The counter does not need to be precise. */ 648 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max) 649 zone->uz_bucket_size++; 650 return (zdom); 651 } 652 653 /* 654 * Search for the domain with the least cached items and return it if it 655 * is out of balance with the preferred domain. 656 */ 657 static __noinline int 658 zone_domain_lowest(uma_zone_t zone, int pref) 659 { 660 long least, nitems, prefitems; 661 int domain; 662 int i; 663 664 prefitems = least = LONG_MAX; 665 domain = 0; 666 for (i = 0; i < vm_ndomains; i++) { 667 nitems = ZDOM_GET(zone, i)->uzd_nitems; 668 if (nitems < least) { 669 domain = i; 670 least = nitems; 671 } 672 if (domain == pref) 673 prefitems = nitems; 674 } 675 if (prefitems < least * 2) 676 return (pref); 677 678 return (domain); 679 } 680 681 /* 682 * Search for the domain with the most cached items and return it or the 683 * preferred domain if it has enough to proceed. 684 */ 685 static __noinline int 686 zone_domain_highest(uma_zone_t zone, int pref) 687 { 688 long most, nitems; 689 int domain; 690 int i; 691 692 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX) 693 return (pref); 694 695 most = 0; 696 domain = 0; 697 for (i = 0; i < vm_ndomains; i++) { 698 nitems = ZDOM_GET(zone, i)->uzd_nitems; 699 if (nitems > most) { 700 domain = i; 701 most = nitems; 702 } 703 } 704 705 return (domain); 706 } 707 708 /* 709 * Set the maximum imax value. 710 */ 711 static void 712 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems) 713 { 714 long old; 715 716 old = zdom->uzd_imax; 717 do { 718 if (old >= nitems) 719 return; 720 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0); 721 722 /* 723 * We are at new maximum, so do the last WSS update for the old 724 * bimin and prepare to measure next allocation batch. 725 */ 726 if (zdom->uzd_wss < old - zdom->uzd_bimin) 727 zdom->uzd_wss = old - zdom->uzd_bimin; 728 zdom->uzd_bimin = nitems; 729 } 730 731 /* 732 * Attempt to satisfy an allocation by retrieving a full bucket from one of the 733 * zone's caches. If a bucket is found the zone is not locked on return. 734 */ 735 static uma_bucket_t 736 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim) 737 { 738 uma_bucket_t bucket; 739 long cnt; 740 int i; 741 bool dtor = false; 742 743 ZDOM_LOCK_ASSERT(zdom); 744 745 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL) 746 return (NULL); 747 748 /* SMR Buckets can not be re-used until readers expire. */ 749 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && 750 bucket->ub_seq != SMR_SEQ_INVALID) { 751 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false)) 752 return (NULL); 753 bucket->ub_seq = SMR_SEQ_INVALID; 754 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR; 755 if (STAILQ_NEXT(bucket, ub_link) != NULL) 756 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq; 757 } 758 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link); 759 760 KASSERT(zdom->uzd_nitems >= bucket->ub_cnt, 761 ("%s: item count underflow (%ld, %d)", 762 __func__, zdom->uzd_nitems, bucket->ub_cnt)); 763 KASSERT(bucket->ub_cnt > 0, 764 ("%s: empty bucket in bucket cache", __func__)); 765 zdom->uzd_nitems -= bucket->ub_cnt; 766 767 if (reclaim) { 768 /* 769 * Shift the bounds of the current WSS interval to avoid 770 * perturbing the estimates. 771 */ 772 cnt = lmin(zdom->uzd_bimin, bucket->ub_cnt); 773 atomic_subtract_long(&zdom->uzd_imax, cnt); 774 zdom->uzd_bimin -= cnt; 775 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt); 776 if (zdom->uzd_limin >= bucket->ub_cnt) { 777 zdom->uzd_limin -= bucket->ub_cnt; 778 } else { 779 zdom->uzd_limin = 0; 780 zdom->uzd_timin = 0; 781 } 782 } else if (zdom->uzd_bimin > zdom->uzd_nitems) { 783 zdom->uzd_bimin = zdom->uzd_nitems; 784 if (zdom->uzd_imin > zdom->uzd_nitems) 785 zdom->uzd_imin = zdom->uzd_nitems; 786 } 787 788 ZDOM_UNLOCK(zdom); 789 if (dtor) 790 for (i = 0; i < bucket->ub_cnt; i++) 791 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size, 792 NULL, SKIP_NONE); 793 794 return (bucket); 795 } 796 797 /* 798 * Insert a full bucket into the specified cache. The "ws" parameter indicates 799 * whether the bucket's contents should be counted as part of the zone's working 800 * set. The bucket may be freed if it exceeds the bucket limit. 801 */ 802 static void 803 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata, 804 const bool ws) 805 { 806 uma_zone_domain_t zdom; 807 808 /* We don't cache empty buckets. This can happen after a reclaim. */ 809 if (bucket->ub_cnt == 0) 810 goto out; 811 zdom = zone_domain_lock(zone, domain); 812 813 /* 814 * Conditionally set the maximum number of items. 815 */ 816 zdom->uzd_nitems += bucket->ub_cnt; 817 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) { 818 if (ws) { 819 zone_domain_imax_set(zdom, zdom->uzd_nitems); 820 } else { 821 /* 822 * Shift the bounds of the current WSS interval to 823 * avoid perturbing the estimates. 824 */ 825 atomic_add_long(&zdom->uzd_imax, bucket->ub_cnt); 826 zdom->uzd_imin += bucket->ub_cnt; 827 zdom->uzd_bimin += bucket->ub_cnt; 828 zdom->uzd_limin += bucket->ub_cnt; 829 } 830 if (STAILQ_EMPTY(&zdom->uzd_buckets)) 831 zdom->uzd_seq = bucket->ub_seq; 832 833 /* 834 * Try to promote reuse of recently used items. For items 835 * protected by SMR, try to defer reuse to minimize polling. 836 */ 837 if (bucket->ub_seq == SMR_SEQ_INVALID) 838 STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link); 839 else 840 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link); 841 ZDOM_UNLOCK(zdom); 842 return; 843 } 844 zdom->uzd_nitems -= bucket->ub_cnt; 845 ZDOM_UNLOCK(zdom); 846 out: 847 bucket_free(zone, bucket, udata); 848 } 849 850 /* Pops an item out of a per-cpu cache bucket. */ 851 static inline void * 852 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket) 853 { 854 void *item; 855 856 CRITICAL_ASSERT(curthread); 857 858 bucket->ucb_cnt--; 859 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt]; 860 #ifdef INVARIANTS 861 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL; 862 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 863 #endif 864 cache->uc_allocs++; 865 866 return (item); 867 } 868 869 /* Pushes an item into a per-cpu cache bucket. */ 870 static inline void 871 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item) 872 { 873 874 CRITICAL_ASSERT(curthread); 875 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL, 876 ("uma_zfree: Freeing to non free bucket index.")); 877 878 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item; 879 bucket->ucb_cnt++; 880 cache->uc_frees++; 881 } 882 883 /* 884 * Unload a UMA bucket from a per-cpu cache. 885 */ 886 static inline uma_bucket_t 887 cache_bucket_unload(uma_cache_bucket_t bucket) 888 { 889 uma_bucket_t b; 890 891 b = bucket->ucb_bucket; 892 if (b != NULL) { 893 MPASS(b->ub_entries == bucket->ucb_entries); 894 b->ub_cnt = bucket->ucb_cnt; 895 bucket->ucb_bucket = NULL; 896 bucket->ucb_entries = bucket->ucb_cnt = 0; 897 } 898 899 return (b); 900 } 901 902 static inline uma_bucket_t 903 cache_bucket_unload_alloc(uma_cache_t cache) 904 { 905 906 return (cache_bucket_unload(&cache->uc_allocbucket)); 907 } 908 909 static inline uma_bucket_t 910 cache_bucket_unload_free(uma_cache_t cache) 911 { 912 913 return (cache_bucket_unload(&cache->uc_freebucket)); 914 } 915 916 static inline uma_bucket_t 917 cache_bucket_unload_cross(uma_cache_t cache) 918 { 919 920 return (cache_bucket_unload(&cache->uc_crossbucket)); 921 } 922 923 /* 924 * Load a bucket into a per-cpu cache bucket. 925 */ 926 static inline void 927 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b) 928 { 929 930 CRITICAL_ASSERT(curthread); 931 MPASS(bucket->ucb_bucket == NULL); 932 MPASS(b->ub_seq == SMR_SEQ_INVALID); 933 934 bucket->ucb_bucket = b; 935 bucket->ucb_cnt = b->ub_cnt; 936 bucket->ucb_entries = b->ub_entries; 937 } 938 939 static inline void 940 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b) 941 { 942 943 cache_bucket_load(&cache->uc_allocbucket, b); 944 } 945 946 static inline void 947 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b) 948 { 949 950 cache_bucket_load(&cache->uc_freebucket, b); 951 } 952 953 #ifdef NUMA 954 static inline void 955 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b) 956 { 957 958 cache_bucket_load(&cache->uc_crossbucket, b); 959 } 960 #endif 961 962 /* 963 * Copy and preserve ucb_spare. 964 */ 965 static inline void 966 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2) 967 { 968 969 b1->ucb_bucket = b2->ucb_bucket; 970 b1->ucb_entries = b2->ucb_entries; 971 b1->ucb_cnt = b2->ucb_cnt; 972 } 973 974 /* 975 * Swap two cache buckets. 976 */ 977 static inline void 978 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2) 979 { 980 struct uma_cache_bucket b3; 981 982 CRITICAL_ASSERT(curthread); 983 984 cache_bucket_copy(&b3, b1); 985 cache_bucket_copy(b1, b2); 986 cache_bucket_copy(b2, &b3); 987 } 988 989 /* 990 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache. 991 */ 992 static uma_bucket_t 993 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain) 994 { 995 uma_zone_domain_t zdom; 996 uma_bucket_t bucket; 997 998 /* 999 * Avoid the lock if possible. 1000 */ 1001 zdom = ZDOM_GET(zone, domain); 1002 if (zdom->uzd_nitems == 0) 1003 return (NULL); 1004 1005 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 && 1006 !smr_poll(zone->uz_smr, zdom->uzd_seq, false)) 1007 return (NULL); 1008 1009 /* 1010 * Check the zone's cache of buckets. 1011 */ 1012 zdom = zone_domain_lock(zone, domain); 1013 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) 1014 return (bucket); 1015 ZDOM_UNLOCK(zdom); 1016 1017 return (NULL); 1018 } 1019 1020 static void 1021 zone_log_warning(uma_zone_t zone) 1022 { 1023 static const struct timeval warninterval = { 300, 0 }; 1024 1025 if (!zone_warnings || zone->uz_warning == NULL) 1026 return; 1027 1028 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 1029 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 1030 } 1031 1032 static inline void 1033 zone_maxaction(uma_zone_t zone) 1034 { 1035 1036 if (zone->uz_maxaction.ta_func != NULL) 1037 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); 1038 } 1039 1040 /* 1041 * Routine called by timeout which is used to fire off some time interval 1042 * based calculations. (stats, hash size, etc.) 1043 * 1044 * Arguments: 1045 * arg Unused 1046 * 1047 * Returns: 1048 * Nothing 1049 */ 1050 static void 1051 uma_timeout(void *unused) 1052 { 1053 bucket_enable(); 1054 zone_foreach(zone_timeout, NULL); 1055 1056 /* Reschedule this event */ 1057 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1058 } 1059 1060 /* 1061 * Update the working set size estimates for the zone's bucket cache. 1062 * The constants chosen here are somewhat arbitrary. 1063 */ 1064 static void 1065 zone_domain_update_wss(uma_zone_domain_t zdom) 1066 { 1067 long m; 1068 1069 ZDOM_LOCK_ASSERT(zdom); 1070 MPASS(zdom->uzd_imax >= zdom->uzd_nitems); 1071 MPASS(zdom->uzd_nitems >= zdom->uzd_bimin); 1072 MPASS(zdom->uzd_bimin >= zdom->uzd_imin); 1073 1074 /* 1075 * Estimate WSS as modified moving average of biggest allocation 1076 * batches for each period over few minutes (UMA_TIMEOUT of 20s). 1077 */ 1078 zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4, 1079 zdom->uzd_imax - zdom->uzd_bimin); 1080 1081 /* 1082 * Estimate longtime minimum item count as a combination of recent 1083 * minimum item count, adjusted by WSS for safety, and the modified 1084 * moving average over the last several hours (UMA_TIMEOUT of 20s). 1085 * timin measures time since limin tried to go negative, that means 1086 * we were dangerously close to or got out of cache. 1087 */ 1088 m = zdom->uzd_imin - zdom->uzd_wss; 1089 if (m >= 0) { 1090 if (zdom->uzd_limin >= m) 1091 zdom->uzd_limin = m; 1092 else 1093 zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256; 1094 zdom->uzd_timin++; 1095 } else { 1096 zdom->uzd_limin = 0; 1097 zdom->uzd_timin = 0; 1098 } 1099 1100 /* To reduce period edge effects on WSS keep half of the imax. */ 1101 atomic_subtract_long(&zdom->uzd_imax, 1102 (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2); 1103 zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems; 1104 } 1105 1106 /* 1107 * Routine to perform timeout driven calculations. This expands the 1108 * hashes and does per cpu statistics aggregation. 1109 * 1110 * Returns nothing. 1111 */ 1112 static void 1113 zone_timeout(uma_zone_t zone, void *unused) 1114 { 1115 uma_keg_t keg; 1116 u_int slabs, pages; 1117 1118 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0) 1119 goto trim; 1120 1121 keg = zone->uz_keg; 1122 1123 /* 1124 * Hash zones are non-numa by definition so the first domain 1125 * is the only one present. 1126 */ 1127 KEG_LOCK(keg, 0); 1128 pages = keg->uk_domain[0].ud_pages; 1129 1130 /* 1131 * Expand the keg hash table. 1132 * 1133 * This is done if the number of slabs is larger than the hash size. 1134 * What I'm trying to do here is completely reduce collisions. This 1135 * may be a little aggressive. Should I allow for two collisions max? 1136 */ 1137 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) { 1138 struct uma_hash newhash; 1139 struct uma_hash oldhash; 1140 int ret; 1141 1142 /* 1143 * This is so involved because allocating and freeing 1144 * while the keg lock is held will lead to deadlock. 1145 * I have to do everything in stages and check for 1146 * races. 1147 */ 1148 KEG_UNLOCK(keg, 0); 1149 ret = hash_alloc(&newhash, 1 << fls(slabs)); 1150 KEG_LOCK(keg, 0); 1151 if (ret) { 1152 if (hash_expand(&keg->uk_hash, &newhash)) { 1153 oldhash = keg->uk_hash; 1154 keg->uk_hash = newhash; 1155 } else 1156 oldhash = newhash; 1157 1158 KEG_UNLOCK(keg, 0); 1159 hash_free(&oldhash); 1160 goto trim; 1161 } 1162 } 1163 KEG_UNLOCK(keg, 0); 1164 1165 trim: 1166 /* Trim caches not used for a long time. */ 1167 for (int i = 0; i < vm_ndomains; i++) { 1168 if (bucket_cache_reclaim_domain(zone, false, false, i) && 1169 (zone->uz_flags & UMA_ZFLAG_CACHE) == 0) 1170 keg_drain(zone->uz_keg, i); 1171 } 1172 } 1173 1174 /* 1175 * Allocate and zero fill the next sized hash table from the appropriate 1176 * backing store. 1177 * 1178 * Arguments: 1179 * hash A new hash structure with the old hash size in uh_hashsize 1180 * 1181 * Returns: 1182 * 1 on success and 0 on failure. 1183 */ 1184 static int 1185 hash_alloc(struct uma_hash *hash, u_int size) 1186 { 1187 size_t alloc; 1188 1189 KASSERT(powerof2(size), ("hash size must be power of 2")); 1190 if (size > UMA_HASH_SIZE_INIT) { 1191 hash->uh_hashsize = size; 1192 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 1193 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT); 1194 } else { 1195 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 1196 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 1197 UMA_ANYDOMAIN, M_WAITOK); 1198 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 1199 } 1200 if (hash->uh_slab_hash) { 1201 bzero(hash->uh_slab_hash, alloc); 1202 hash->uh_hashmask = hash->uh_hashsize - 1; 1203 return (1); 1204 } 1205 1206 return (0); 1207 } 1208 1209 /* 1210 * Expands the hash table for HASH zones. This is done from zone_timeout 1211 * to reduce collisions. This must not be done in the regular allocation 1212 * path, otherwise, we can recurse on the vm while allocating pages. 1213 * 1214 * Arguments: 1215 * oldhash The hash you want to expand 1216 * newhash The hash structure for the new table 1217 * 1218 * Returns: 1219 * Nothing 1220 * 1221 * Discussion: 1222 */ 1223 static int 1224 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 1225 { 1226 uma_hash_slab_t slab; 1227 u_int hval; 1228 u_int idx; 1229 1230 if (!newhash->uh_slab_hash) 1231 return (0); 1232 1233 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 1234 return (0); 1235 1236 /* 1237 * I need to investigate hash algorithms for resizing without a 1238 * full rehash. 1239 */ 1240 1241 for (idx = 0; idx < oldhash->uh_hashsize; idx++) 1242 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) { 1243 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]); 1244 LIST_REMOVE(slab, uhs_hlink); 1245 hval = UMA_HASH(newhash, slab->uhs_data); 1246 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 1247 slab, uhs_hlink); 1248 } 1249 1250 return (1); 1251 } 1252 1253 /* 1254 * Free the hash bucket to the appropriate backing store. 1255 * 1256 * Arguments: 1257 * slab_hash The hash bucket we're freeing 1258 * hashsize The number of entries in that hash bucket 1259 * 1260 * Returns: 1261 * Nothing 1262 */ 1263 static void 1264 hash_free(struct uma_hash *hash) 1265 { 1266 if (hash->uh_slab_hash == NULL) 1267 return; 1268 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 1269 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 1270 else 1271 free(hash->uh_slab_hash, M_UMAHASH); 1272 } 1273 1274 /* 1275 * Frees all outstanding items in a bucket 1276 * 1277 * Arguments: 1278 * zone The zone to free to, must be unlocked. 1279 * bucket The free/alloc bucket with items. 1280 * 1281 * Returns: 1282 * Nothing 1283 */ 1284 static void 1285 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 1286 { 1287 int i; 1288 1289 if (bucket->ub_cnt == 0) 1290 return; 1291 1292 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && 1293 bucket->ub_seq != SMR_SEQ_INVALID) { 1294 smr_wait(zone->uz_smr, bucket->ub_seq); 1295 bucket->ub_seq = SMR_SEQ_INVALID; 1296 for (i = 0; i < bucket->ub_cnt; i++) 1297 item_dtor(zone, bucket->ub_bucket[i], 1298 zone->uz_size, NULL, SKIP_NONE); 1299 } 1300 if (zone->uz_fini) 1301 for (i = 0; i < bucket->ub_cnt; i++) { 1302 kasan_mark_item_valid(zone, bucket->ub_bucket[i]); 1303 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 1304 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]); 1305 } 1306 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 1307 if (zone->uz_max_items > 0) 1308 zone_free_limit(zone, bucket->ub_cnt); 1309 #ifdef INVARIANTS 1310 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt); 1311 #endif 1312 bucket->ub_cnt = 0; 1313 } 1314 1315 /* 1316 * Drains the per cpu caches for a zone. 1317 * 1318 * NOTE: This may only be called while the zone is being torn down, and not 1319 * during normal operation. This is necessary in order that we do not have 1320 * to migrate CPUs to drain the per-CPU caches. 1321 * 1322 * Arguments: 1323 * zone The zone to drain, must be unlocked. 1324 * 1325 * Returns: 1326 * Nothing 1327 */ 1328 static void 1329 cache_drain(uma_zone_t zone) 1330 { 1331 uma_cache_t cache; 1332 uma_bucket_t bucket; 1333 smr_seq_t seq; 1334 int cpu; 1335 1336 /* 1337 * XXX: It is safe to not lock the per-CPU caches, because we're 1338 * tearing down the zone anyway. I.e., there will be no further use 1339 * of the caches at this point. 1340 * 1341 * XXX: It would good to be able to assert that the zone is being 1342 * torn down to prevent improper use of cache_drain(). 1343 */ 1344 seq = SMR_SEQ_INVALID; 1345 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 1346 seq = smr_advance(zone->uz_smr); 1347 CPU_FOREACH(cpu) { 1348 cache = &zone->uz_cpu[cpu]; 1349 bucket = cache_bucket_unload_alloc(cache); 1350 if (bucket != NULL) 1351 bucket_free(zone, bucket, NULL); 1352 bucket = cache_bucket_unload_free(cache); 1353 if (bucket != NULL) { 1354 bucket->ub_seq = seq; 1355 bucket_free(zone, bucket, NULL); 1356 } 1357 bucket = cache_bucket_unload_cross(cache); 1358 if (bucket != NULL) { 1359 bucket->ub_seq = seq; 1360 bucket_free(zone, bucket, NULL); 1361 } 1362 } 1363 bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN); 1364 } 1365 1366 static void 1367 cache_shrink(uma_zone_t zone, void *unused) 1368 { 1369 1370 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 1371 return; 1372 1373 ZONE_LOCK(zone); 1374 zone->uz_bucket_size = 1375 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2; 1376 ZONE_UNLOCK(zone); 1377 } 1378 1379 static void 1380 cache_drain_safe_cpu(uma_zone_t zone, void *unused) 1381 { 1382 uma_cache_t cache; 1383 uma_bucket_t b1, b2, b3; 1384 int domain; 1385 1386 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 1387 return; 1388 1389 b1 = b2 = b3 = NULL; 1390 critical_enter(); 1391 cache = &zone->uz_cpu[curcpu]; 1392 domain = PCPU_GET(domain); 1393 b1 = cache_bucket_unload_alloc(cache); 1394 1395 /* 1396 * Don't flush SMR zone buckets. This leaves the zone without a 1397 * bucket and forces every free to synchronize(). 1398 */ 1399 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) { 1400 b2 = cache_bucket_unload_free(cache); 1401 b3 = cache_bucket_unload_cross(cache); 1402 } 1403 critical_exit(); 1404 1405 if (b1 != NULL) 1406 zone_free_bucket(zone, b1, NULL, domain, false); 1407 if (b2 != NULL) 1408 zone_free_bucket(zone, b2, NULL, domain, false); 1409 if (b3 != NULL) { 1410 /* Adjust the domain so it goes to zone_free_cross. */ 1411 domain = (domain + 1) % vm_ndomains; 1412 zone_free_bucket(zone, b3, NULL, domain, false); 1413 } 1414 } 1415 1416 /* 1417 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 1418 * This is an expensive call because it needs to bind to all CPUs 1419 * one by one and enter a critical section on each of them in order 1420 * to safely access their cache buckets. 1421 * Zone lock must not be held on call this function. 1422 */ 1423 static void 1424 pcpu_cache_drain_safe(uma_zone_t zone) 1425 { 1426 int cpu; 1427 1428 /* 1429 * Polite bucket sizes shrinking was not enough, shrink aggressively. 1430 */ 1431 if (zone) 1432 cache_shrink(zone, NULL); 1433 else 1434 zone_foreach(cache_shrink, NULL); 1435 1436 CPU_FOREACH(cpu) { 1437 thread_lock(curthread); 1438 sched_bind(curthread, cpu); 1439 thread_unlock(curthread); 1440 1441 if (zone) 1442 cache_drain_safe_cpu(zone, NULL); 1443 else 1444 zone_foreach(cache_drain_safe_cpu, NULL); 1445 } 1446 thread_lock(curthread); 1447 sched_unbind(curthread); 1448 thread_unlock(curthread); 1449 } 1450 1451 /* 1452 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller 1453 * requested a drain, otherwise the per-domain caches are trimmed to either 1454 * estimated working set size. 1455 */ 1456 static bool 1457 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain) 1458 { 1459 uma_zone_domain_t zdom; 1460 uma_bucket_t bucket; 1461 long target; 1462 bool done = false; 1463 1464 /* 1465 * The cross bucket is partially filled and not part of 1466 * the item count. Reclaim it individually here. 1467 */ 1468 zdom = ZDOM_GET(zone, domain); 1469 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) { 1470 ZONE_CROSS_LOCK(zone); 1471 bucket = zdom->uzd_cross; 1472 zdom->uzd_cross = NULL; 1473 ZONE_CROSS_UNLOCK(zone); 1474 if (bucket != NULL) 1475 bucket_free(zone, bucket, NULL); 1476 } 1477 1478 /* 1479 * If we were asked to drain the zone, we are done only once 1480 * this bucket cache is empty. If trim, we reclaim items in 1481 * excess of the zone's estimated working set size. Multiple 1482 * consecutive calls will shrink the WSS and so reclaim more. 1483 * If neither drain nor trim, then voluntarily reclaim 1/4 1484 * (to reduce first spike) of items not used for a long time. 1485 */ 1486 ZDOM_LOCK(zdom); 1487 zone_domain_update_wss(zdom); 1488 if (drain) 1489 target = 0; 1490 else if (trim) 1491 target = zdom->uzd_wss; 1492 else if (zdom->uzd_timin > 900 / UMA_TIMEOUT) 1493 target = zdom->uzd_nitems - zdom->uzd_limin / 4; 1494 else { 1495 ZDOM_UNLOCK(zdom); 1496 return (done); 1497 } 1498 while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL && 1499 zdom->uzd_nitems >= target + bucket->ub_cnt) { 1500 bucket = zone_fetch_bucket(zone, zdom, true); 1501 if (bucket == NULL) 1502 break; 1503 bucket_free(zone, bucket, NULL); 1504 done = true; 1505 ZDOM_LOCK(zdom); 1506 } 1507 ZDOM_UNLOCK(zdom); 1508 return (done); 1509 } 1510 1511 static void 1512 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain) 1513 { 1514 int i; 1515 1516 /* 1517 * Shrink the zone bucket size to ensure that the per-CPU caches 1518 * don't grow too large. 1519 */ 1520 if (zone->uz_bucket_size > zone->uz_bucket_size_min) 1521 zone->uz_bucket_size--; 1522 1523 if (domain != UMA_ANYDOMAIN && 1524 (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) { 1525 bucket_cache_reclaim_domain(zone, drain, true, domain); 1526 } else { 1527 for (i = 0; i < vm_ndomains; i++) 1528 bucket_cache_reclaim_domain(zone, drain, true, i); 1529 } 1530 } 1531 1532 static void 1533 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 1534 { 1535 uint8_t *mem; 1536 size_t size; 1537 int i; 1538 uint8_t flags; 1539 1540 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes", 1541 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera); 1542 1543 mem = slab_data(slab, keg); 1544 size = PAGE_SIZE * keg->uk_ppera; 1545 1546 kasan_mark_slab_valid(keg, mem); 1547 if (keg->uk_fini != NULL) { 1548 for (i = start - 1; i > -1; i--) 1549 #ifdef INVARIANTS 1550 /* 1551 * trash_fini implies that dtor was trash_dtor. trash_fini 1552 * would check that memory hasn't been modified since free, 1553 * which executed trash_dtor. 1554 * That's why we need to run uma_dbg_kskip() check here, 1555 * albeit we don't make skip check for other init/fini 1556 * invocations. 1557 */ 1558 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) || 1559 keg->uk_fini != trash_fini) 1560 #endif 1561 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size); 1562 } 1563 flags = slab->us_flags; 1564 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) { 1565 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab), 1566 NULL, SKIP_NONE); 1567 } 1568 keg->uk_freef(mem, size, flags); 1569 uma_total_dec(size); 1570 } 1571 1572 static void 1573 keg_drain_domain(uma_keg_t keg, int domain) 1574 { 1575 struct slabhead freeslabs; 1576 uma_domain_t dom; 1577 uma_slab_t slab, tmp; 1578 uint32_t i, stofree, stokeep, partial; 1579 1580 dom = &keg->uk_domain[domain]; 1581 LIST_INIT(&freeslabs); 1582 1583 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u", 1584 keg->uk_name, keg, domain, dom->ud_free_items); 1585 1586 KEG_LOCK(keg, domain); 1587 1588 /* 1589 * Are the free items in partially allocated slabs sufficient to meet 1590 * the reserve? If not, compute the number of fully free slabs that must 1591 * be kept. 1592 */ 1593 partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers; 1594 if (partial < keg->uk_reserve) { 1595 stokeep = min(dom->ud_free_slabs, 1596 howmany(keg->uk_reserve - partial, keg->uk_ipers)); 1597 } else { 1598 stokeep = 0; 1599 } 1600 stofree = dom->ud_free_slabs - stokeep; 1601 1602 /* 1603 * Partition the free slabs into two sets: those that must be kept in 1604 * order to maintain the reserve, and those that may be released back to 1605 * the system. Since one set may be much larger than the other, 1606 * populate the smaller of the two sets and swap them if necessary. 1607 */ 1608 for (i = min(stofree, stokeep); i > 0; i--) { 1609 slab = LIST_FIRST(&dom->ud_free_slab); 1610 LIST_REMOVE(slab, us_link); 1611 LIST_INSERT_HEAD(&freeslabs, slab, us_link); 1612 } 1613 if (stofree > stokeep) 1614 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link); 1615 1616 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) { 1617 LIST_FOREACH(slab, &freeslabs, us_link) 1618 UMA_HASH_REMOVE(&keg->uk_hash, slab); 1619 } 1620 dom->ud_free_items -= stofree * keg->uk_ipers; 1621 dom->ud_free_slabs -= stofree; 1622 dom->ud_pages -= stofree * keg->uk_ppera; 1623 KEG_UNLOCK(keg, domain); 1624 1625 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp) 1626 keg_free_slab(keg, slab, keg->uk_ipers); 1627 } 1628 1629 /* 1630 * Frees pages from a keg back to the system. This is done on demand from 1631 * the pageout daemon. 1632 * 1633 * Returns nothing. 1634 */ 1635 static void 1636 keg_drain(uma_keg_t keg, int domain) 1637 { 1638 int i; 1639 1640 if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0) 1641 return; 1642 if (domain != UMA_ANYDOMAIN) { 1643 keg_drain_domain(keg, domain); 1644 } else { 1645 for (i = 0; i < vm_ndomains; i++) 1646 keg_drain_domain(keg, i); 1647 } 1648 } 1649 1650 static void 1651 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain) 1652 { 1653 /* 1654 * Count active reclaim operations in order to interlock with 1655 * zone_dtor(), which removes the zone from global lists before 1656 * attempting to reclaim items itself. 1657 * 1658 * The zone may be destroyed while sleeping, so only zone_dtor() should 1659 * specify M_WAITOK. 1660 */ 1661 ZONE_LOCK(zone); 1662 if (waitok == M_WAITOK) { 1663 while (zone->uz_reclaimers > 0) 1664 msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1); 1665 } 1666 zone->uz_reclaimers++; 1667 ZONE_UNLOCK(zone); 1668 bucket_cache_reclaim(zone, drain, domain); 1669 1670 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) 1671 keg_drain(zone->uz_keg, domain); 1672 ZONE_LOCK(zone); 1673 zone->uz_reclaimers--; 1674 if (zone->uz_reclaimers == 0) 1675 wakeup(zone); 1676 ZONE_UNLOCK(zone); 1677 } 1678 1679 static void 1680 zone_drain(uma_zone_t zone, void *arg) 1681 { 1682 int domain; 1683 1684 domain = (int)(uintptr_t)arg; 1685 zone_reclaim(zone, domain, M_NOWAIT, true); 1686 } 1687 1688 static void 1689 zone_trim(uma_zone_t zone, void *arg) 1690 { 1691 int domain; 1692 1693 domain = (int)(uintptr_t)arg; 1694 zone_reclaim(zone, domain, M_NOWAIT, false); 1695 } 1696 1697 /* 1698 * Allocate a new slab for a keg and inserts it into the partial slab list. 1699 * The keg should be unlocked on entry. If the allocation succeeds it will 1700 * be locked on return. 1701 * 1702 * Arguments: 1703 * flags Wait flags for the item initialization routine 1704 * aflags Wait flags for the slab allocation 1705 * 1706 * Returns: 1707 * The slab that was allocated or NULL if there is no memory and the 1708 * caller specified M_NOWAIT. 1709 */ 1710 static uma_slab_t 1711 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags, 1712 int aflags) 1713 { 1714 uma_domain_t dom; 1715 uma_slab_t slab; 1716 unsigned long size; 1717 uint8_t *mem; 1718 uint8_t sflags; 1719 int i; 1720 1721 KASSERT(domain >= 0 && domain < vm_ndomains, 1722 ("keg_alloc_slab: domain %d out of range", domain)); 1723 1724 slab = NULL; 1725 mem = NULL; 1726 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) { 1727 uma_hash_slab_t hslab; 1728 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL, 1729 domain, aflags); 1730 if (hslab == NULL) 1731 goto fail; 1732 slab = &hslab->uhs_slab; 1733 } 1734 1735 /* 1736 * This reproduces the old vm_zone behavior of zero filling pages the 1737 * first time they are added to a zone. 1738 * 1739 * Malloced items are zeroed in uma_zalloc. 1740 */ 1741 1742 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 1743 aflags |= M_ZERO; 1744 else 1745 aflags &= ~M_ZERO; 1746 1747 if (keg->uk_flags & UMA_ZONE_NODUMP) 1748 aflags |= M_NODUMP; 1749 1750 /* zone is passed for legacy reasons. */ 1751 size = keg->uk_ppera * PAGE_SIZE; 1752 mem = keg->uk_allocf(zone, size, domain, &sflags, aflags); 1753 if (mem == NULL) { 1754 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) 1755 zone_free_item(slabzone(keg->uk_ipers), 1756 slab_tohashslab(slab), NULL, SKIP_NONE); 1757 goto fail; 1758 } 1759 uma_total_inc(size); 1760 1761 /* For HASH zones all pages go to the same uma_domain. */ 1762 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) 1763 domain = 0; 1764 1765 /* Point the slab into the allocated memory */ 1766 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) 1767 slab = (uma_slab_t)(mem + keg->uk_pgoff); 1768 else 1769 slab_tohashslab(slab)->uhs_data = mem; 1770 1771 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB) 1772 for (i = 0; i < keg->uk_ppera; i++) 1773 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE), 1774 zone, slab); 1775 1776 slab->us_freecount = keg->uk_ipers; 1777 slab->us_flags = sflags; 1778 slab->us_domain = domain; 1779 1780 BIT_FILL(keg->uk_ipers, &slab->us_free); 1781 #ifdef INVARIANTS 1782 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg)); 1783 #endif 1784 1785 if (keg->uk_init != NULL) { 1786 for (i = 0; i < keg->uk_ipers; i++) 1787 if (keg->uk_init(slab_item(slab, keg, i), 1788 keg->uk_size, flags) != 0) 1789 break; 1790 if (i != keg->uk_ipers) { 1791 keg_free_slab(keg, slab, i); 1792 goto fail; 1793 } 1794 } 1795 kasan_mark_slab_invalid(keg, mem); 1796 KEG_LOCK(keg, domain); 1797 1798 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)", 1799 slab, keg->uk_name, keg); 1800 1801 if (keg->uk_flags & UMA_ZFLAG_HASH) 1802 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1803 1804 /* 1805 * If we got a slab here it's safe to mark it partially used 1806 * and return. We assume that the caller is going to remove 1807 * at least one item. 1808 */ 1809 dom = &keg->uk_domain[domain]; 1810 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 1811 dom->ud_pages += keg->uk_ppera; 1812 dom->ud_free_items += keg->uk_ipers; 1813 1814 return (slab); 1815 1816 fail: 1817 return (NULL); 1818 } 1819 1820 /* 1821 * This function is intended to be used early on in place of page_alloc(). It 1822 * performs contiguous physical memory allocations and uses a bump allocator for 1823 * KVA, so is usable before the kernel map is initialized. 1824 */ 1825 static void * 1826 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1827 int wait) 1828 { 1829 vm_paddr_t pa; 1830 vm_page_t m; 1831 void *mem; 1832 int pages; 1833 int i; 1834 1835 pages = howmany(bytes, PAGE_SIZE); 1836 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__)); 1837 1838 *pflag = UMA_SLAB_BOOT; 1839 m = vm_page_alloc_contig_domain(NULL, 0, domain, 1840 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages, 1841 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT); 1842 if (m == NULL) 1843 return (NULL); 1844 1845 pa = VM_PAGE_TO_PHYS(m); 1846 for (i = 0; i < pages; i++, pa += PAGE_SIZE) { 1847 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 1848 defined(__riscv) || defined(__powerpc64__) 1849 if ((wait & M_NODUMP) == 0) 1850 dump_add_page(pa); 1851 #endif 1852 } 1853 /* Allocate KVA and indirectly advance bootmem. */ 1854 mem = (void *)pmap_map(&bootmem, m->phys_addr, 1855 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE); 1856 if ((wait & M_ZERO) != 0) 1857 bzero(mem, pages * PAGE_SIZE); 1858 1859 return (mem); 1860 } 1861 1862 static void 1863 startup_free(void *mem, vm_size_t bytes) 1864 { 1865 vm_offset_t va; 1866 vm_page_t m; 1867 1868 va = (vm_offset_t)mem; 1869 m = PHYS_TO_VM_PAGE(pmap_kextract(va)); 1870 1871 /* 1872 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid 1873 * unmapping ranges of the direct map. 1874 */ 1875 if (va >= bootstart && va + bytes <= bootmem) 1876 pmap_remove(kernel_pmap, va, va + bytes); 1877 for (; bytes != 0; bytes -= PAGE_SIZE, m++) { 1878 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 1879 defined(__riscv) || defined(__powerpc64__) 1880 dump_drop_page(VM_PAGE_TO_PHYS(m)); 1881 #endif 1882 vm_page_unwire_noq(m); 1883 vm_page_free(m); 1884 } 1885 } 1886 1887 /* 1888 * Allocates a number of pages from the system 1889 * 1890 * Arguments: 1891 * bytes The number of bytes requested 1892 * wait Shall we wait? 1893 * 1894 * Returns: 1895 * A pointer to the alloced memory or possibly 1896 * NULL if M_NOWAIT is set. 1897 */ 1898 static void * 1899 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1900 int wait) 1901 { 1902 void *p; /* Returned page */ 1903 1904 *pflag = UMA_SLAB_KERNEL; 1905 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait); 1906 1907 return (p); 1908 } 1909 1910 static void * 1911 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1912 int wait) 1913 { 1914 struct pglist alloctail; 1915 vm_offset_t addr, zkva; 1916 int cpu, flags; 1917 vm_page_t p, p_next; 1918 #ifdef NUMA 1919 struct pcpu *pc; 1920 #endif 1921 1922 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE); 1923 1924 TAILQ_INIT(&alloctail); 1925 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1926 malloc2vm_flags(wait); 1927 *pflag = UMA_SLAB_KERNEL; 1928 for (cpu = 0; cpu <= mp_maxid; cpu++) { 1929 if (CPU_ABSENT(cpu)) { 1930 p = vm_page_alloc(NULL, 0, flags); 1931 } else { 1932 #ifndef NUMA 1933 p = vm_page_alloc(NULL, 0, flags); 1934 #else 1935 pc = pcpu_find(cpu); 1936 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain))) 1937 p = NULL; 1938 else 1939 p = vm_page_alloc_domain(NULL, 0, 1940 pc->pc_domain, flags); 1941 if (__predict_false(p == NULL)) 1942 p = vm_page_alloc(NULL, 0, flags); 1943 #endif 1944 } 1945 if (__predict_false(p == NULL)) 1946 goto fail; 1947 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1948 } 1949 if ((addr = kva_alloc(bytes)) == 0) 1950 goto fail; 1951 zkva = addr; 1952 TAILQ_FOREACH(p, &alloctail, listq) { 1953 pmap_qenter(zkva, &p, 1); 1954 zkva += PAGE_SIZE; 1955 } 1956 return ((void*)addr); 1957 fail: 1958 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1959 vm_page_unwire_noq(p); 1960 vm_page_free(p); 1961 } 1962 return (NULL); 1963 } 1964 1965 /* 1966 * Allocates a number of pages from within an object 1967 * 1968 * Arguments: 1969 * bytes The number of bytes requested 1970 * wait Shall we wait? 1971 * 1972 * Returns: 1973 * A pointer to the alloced memory or possibly 1974 * NULL if M_NOWAIT is set. 1975 */ 1976 static void * 1977 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, 1978 int wait) 1979 { 1980 TAILQ_HEAD(, vm_page) alloctail; 1981 u_long npages; 1982 vm_offset_t retkva, zkva; 1983 vm_page_t p, p_next; 1984 uma_keg_t keg; 1985 1986 TAILQ_INIT(&alloctail); 1987 keg = zone->uz_keg; 1988 1989 npages = howmany(bytes, PAGE_SIZE); 1990 while (npages > 0) { 1991 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT | 1992 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1993 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK : 1994 VM_ALLOC_NOWAIT)); 1995 if (p != NULL) { 1996 /* 1997 * Since the page does not belong to an object, its 1998 * listq is unused. 1999 */ 2000 TAILQ_INSERT_TAIL(&alloctail, p, listq); 2001 npages--; 2002 continue; 2003 } 2004 /* 2005 * Page allocation failed, free intermediate pages and 2006 * exit. 2007 */ 2008 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 2009 vm_page_unwire_noq(p); 2010 vm_page_free(p); 2011 } 2012 return (NULL); 2013 } 2014 *flags = UMA_SLAB_PRIV; 2015 zkva = keg->uk_kva + 2016 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 2017 retkva = zkva; 2018 TAILQ_FOREACH(p, &alloctail, listq) { 2019 pmap_qenter(zkva, &p, 1); 2020 zkva += PAGE_SIZE; 2021 } 2022 2023 return ((void *)retkva); 2024 } 2025 2026 /* 2027 * Allocate physically contiguous pages. 2028 */ 2029 static void * 2030 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 2031 int wait) 2032 { 2033 2034 *pflag = UMA_SLAB_KERNEL; 2035 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain), 2036 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT)); 2037 } 2038 2039 /* 2040 * Frees a number of pages to the system 2041 * 2042 * Arguments: 2043 * mem A pointer to the memory to be freed 2044 * size The size of the memory being freed 2045 * flags The original p->us_flags field 2046 * 2047 * Returns: 2048 * Nothing 2049 */ 2050 static void 2051 page_free(void *mem, vm_size_t size, uint8_t flags) 2052 { 2053 2054 if ((flags & UMA_SLAB_BOOT) != 0) { 2055 startup_free(mem, size); 2056 return; 2057 } 2058 2059 KASSERT((flags & UMA_SLAB_KERNEL) != 0, 2060 ("UMA: page_free used with invalid flags %x", flags)); 2061 2062 kmem_free((vm_offset_t)mem, size); 2063 } 2064 2065 /* 2066 * Frees pcpu zone allocations 2067 * 2068 * Arguments: 2069 * mem A pointer to the memory to be freed 2070 * size The size of the memory being freed 2071 * flags The original p->us_flags field 2072 * 2073 * Returns: 2074 * Nothing 2075 */ 2076 static void 2077 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags) 2078 { 2079 vm_offset_t sva, curva; 2080 vm_paddr_t paddr; 2081 vm_page_t m; 2082 2083 MPASS(size == (mp_maxid+1)*PAGE_SIZE); 2084 2085 if ((flags & UMA_SLAB_BOOT) != 0) { 2086 startup_free(mem, size); 2087 return; 2088 } 2089 2090 sva = (vm_offset_t)mem; 2091 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) { 2092 paddr = pmap_kextract(curva); 2093 m = PHYS_TO_VM_PAGE(paddr); 2094 vm_page_unwire_noq(m); 2095 vm_page_free(m); 2096 } 2097 pmap_qremove(sva, size >> PAGE_SHIFT); 2098 kva_free(sva, size); 2099 } 2100 2101 /* 2102 * Zero fill initializer 2103 * 2104 * Arguments/Returns follow uma_init specifications 2105 */ 2106 static int 2107 zero_init(void *mem, int size, int flags) 2108 { 2109 bzero(mem, size); 2110 return (0); 2111 } 2112 2113 #ifdef INVARIANTS 2114 static struct noslabbits * 2115 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg) 2116 { 2117 2118 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers))); 2119 } 2120 #endif 2121 2122 /* 2123 * Actual size of embedded struct slab (!OFFPAGE). 2124 */ 2125 static size_t 2126 slab_sizeof(int nitems) 2127 { 2128 size_t s; 2129 2130 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS; 2131 return (roundup(s, UMA_ALIGN_PTR + 1)); 2132 } 2133 2134 #define UMA_FIXPT_SHIFT 31 2135 #define UMA_FRAC_FIXPT(n, d) \ 2136 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d))) 2137 #define UMA_FIXPT_PCT(f) \ 2138 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT)) 2139 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100) 2140 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE) 2141 2142 /* 2143 * Compute the number of items that will fit in a slab. If hdr is true, the 2144 * item count may be limited to provide space in the slab for an inline slab 2145 * header. Otherwise, all slab space will be provided for item storage. 2146 */ 2147 static u_int 2148 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr) 2149 { 2150 u_int ipers; 2151 u_int padpi; 2152 2153 /* The padding between items is not needed after the last item. */ 2154 padpi = rsize - size; 2155 2156 if (hdr) { 2157 /* 2158 * Start with the maximum item count and remove items until 2159 * the slab header first alongside the allocatable memory. 2160 */ 2161 for (ipers = MIN(SLAB_MAX_SETSIZE, 2162 (slabsize + padpi - slab_sizeof(1)) / rsize); 2163 ipers > 0 && 2164 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize; 2165 ipers--) 2166 continue; 2167 } else { 2168 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE); 2169 } 2170 2171 return (ipers); 2172 } 2173 2174 struct keg_layout_result { 2175 u_int format; 2176 u_int slabsize; 2177 u_int ipers; 2178 u_int eff; 2179 }; 2180 2181 static void 2182 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt, 2183 struct keg_layout_result *kl) 2184 { 2185 u_int total; 2186 2187 kl->format = fmt; 2188 kl->slabsize = slabsize; 2189 2190 /* Handle INTERNAL as inline with an extra page. */ 2191 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) { 2192 kl->format &= ~UMA_ZFLAG_INTERNAL; 2193 kl->slabsize += PAGE_SIZE; 2194 } 2195 2196 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize, 2197 (fmt & UMA_ZFLAG_OFFPAGE) == 0); 2198 2199 /* Account for memory used by an offpage slab header. */ 2200 total = kl->slabsize; 2201 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0) 2202 total += slabzone(kl->ipers)->uz_keg->uk_rsize; 2203 2204 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total); 2205 } 2206 2207 /* 2208 * Determine the format of a uma keg. This determines where the slab header 2209 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera. 2210 * 2211 * Arguments 2212 * keg The zone we should initialize 2213 * 2214 * Returns 2215 * Nothing 2216 */ 2217 static void 2218 keg_layout(uma_keg_t keg) 2219 { 2220 struct keg_layout_result kl = {}, kl_tmp; 2221 u_int fmts[2]; 2222 u_int alignsize; 2223 u_int nfmt; 2224 u_int pages; 2225 u_int rsize; 2226 u_int slabsize; 2227 u_int i, j; 2228 2229 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 2230 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE && 2231 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0), 2232 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b", 2233 __func__, keg->uk_name, keg->uk_size, keg->uk_flags, 2234 PRINT_UMA_ZFLAGS)); 2235 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 || 2236 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0, 2237 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags, 2238 PRINT_UMA_ZFLAGS)); 2239 2240 alignsize = keg->uk_align + 1; 2241 2242 /* 2243 * Calculate the size of each allocation (rsize) according to 2244 * alignment. If the requested size is smaller than we have 2245 * allocation bits for we round it up. 2246 */ 2247 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT); 2248 rsize = roundup2(rsize, alignsize); 2249 2250 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) { 2251 /* 2252 * We want one item to start on every align boundary in a page. 2253 * To do this we will span pages. We will also extend the item 2254 * by the size of align if it is an even multiple of align. 2255 * Otherwise, it would fall on the same boundary every time. 2256 */ 2257 if ((rsize & alignsize) == 0) 2258 rsize += alignsize; 2259 slabsize = rsize * (PAGE_SIZE / alignsize); 2260 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE); 2261 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE); 2262 slabsize = round_page(slabsize); 2263 } else { 2264 /* 2265 * Start with a slab size of as many pages as it takes to 2266 * represent a single item. We will try to fit as many 2267 * additional items into the slab as possible. 2268 */ 2269 slabsize = round_page(keg->uk_size); 2270 } 2271 2272 /* Build a list of all of the available formats for this keg. */ 2273 nfmt = 0; 2274 2275 /* Evaluate an inline slab layout. */ 2276 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0) 2277 fmts[nfmt++] = 0; 2278 2279 /* TODO: vm_page-embedded slab. */ 2280 2281 /* 2282 * We can't do OFFPAGE if we're internal or if we've been 2283 * asked to not go to the VM for buckets. If we do this we 2284 * may end up going to the VM for slabs which we do not want 2285 * to do if we're UMA_ZONE_VM, which clearly forbids it. 2286 * In those cases, evaluate a pseudo-format called INTERNAL 2287 * which has an inline slab header and one extra page to 2288 * guarantee that it fits. 2289 * 2290 * Otherwise, see if using an OFFPAGE slab will improve our 2291 * efficiency. 2292 */ 2293 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0) 2294 fmts[nfmt++] = UMA_ZFLAG_INTERNAL; 2295 else 2296 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE; 2297 2298 /* 2299 * Choose a slab size and format which satisfy the minimum efficiency. 2300 * Prefer the smallest slab size that meets the constraints. 2301 * 2302 * Start with a minimum slab size, to accommodate CACHESPREAD. Then, 2303 * for small items (up to PAGE_SIZE), the iteration increment is one 2304 * page; and for large items, the increment is one item. 2305 */ 2306 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize); 2307 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u", 2308 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize, 2309 rsize, i)); 2310 for ( ; ; i++) { 2311 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) : 2312 round_page(rsize * (i - 1) + keg->uk_size); 2313 2314 for (j = 0; j < nfmt; j++) { 2315 /* Only if we have no viable format yet. */ 2316 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 && 2317 kl.ipers > 0) 2318 continue; 2319 2320 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp); 2321 if (kl_tmp.eff <= kl.eff) 2322 continue; 2323 2324 kl = kl_tmp; 2325 2326 CTR6(KTR_UMA, "keg %s layout: format %#x " 2327 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff", 2328 keg->uk_name, kl.format, kl.ipers, rsize, 2329 kl.slabsize, UMA_FIXPT_PCT(kl.eff)); 2330 2331 /* Stop when we reach the minimum efficiency. */ 2332 if (kl.eff >= UMA_MIN_EFF) 2333 break; 2334 } 2335 2336 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs || 2337 slabsize >= SLAB_MAX_SETSIZE * rsize || 2338 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0) 2339 break; 2340 } 2341 2342 pages = atop(kl.slabsize); 2343 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) 2344 pages *= mp_maxid + 1; 2345 2346 keg->uk_rsize = rsize; 2347 keg->uk_ipers = kl.ipers; 2348 keg->uk_ppera = pages; 2349 keg->uk_flags |= kl.format; 2350 2351 /* 2352 * How do we find the slab header if it is offpage or if not all item 2353 * start addresses are in the same page? We could solve the latter 2354 * case with vaddr alignment, but we don't. 2355 */ 2356 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 || 2357 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) { 2358 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0) 2359 keg->uk_flags |= UMA_ZFLAG_HASH; 2360 else 2361 keg->uk_flags |= UMA_ZFLAG_VTOSLAB; 2362 } 2363 2364 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u", 2365 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers, 2366 pages); 2367 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE, 2368 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__, 2369 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize, 2370 keg->uk_ipers, pages)); 2371 } 2372 2373 /* 2374 * Keg header ctor. This initializes all fields, locks, etc. And inserts 2375 * the keg onto the global keg list. 2376 * 2377 * Arguments/Returns follow uma_ctor specifications 2378 * udata Actually uma_kctor_args 2379 */ 2380 static int 2381 keg_ctor(void *mem, int size, void *udata, int flags) 2382 { 2383 struct uma_kctor_args *arg = udata; 2384 uma_keg_t keg = mem; 2385 uma_zone_t zone; 2386 int i; 2387 2388 bzero(keg, size); 2389 keg->uk_size = arg->size; 2390 keg->uk_init = arg->uminit; 2391 keg->uk_fini = arg->fini; 2392 keg->uk_align = arg->align; 2393 keg->uk_reserve = 0; 2394 keg->uk_flags = arg->flags; 2395 2396 /* 2397 * We use a global round-robin policy by default. Zones with 2398 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which 2399 * case the iterator is never run. 2400 */ 2401 keg->uk_dr.dr_policy = DOMAINSET_RR(); 2402 keg->uk_dr.dr_iter = 0; 2403 2404 /* 2405 * The primary zone is passed to us at keg-creation time. 2406 */ 2407 zone = arg->zone; 2408 keg->uk_name = zone->uz_name; 2409 2410 if (arg->flags & UMA_ZONE_ZINIT) 2411 keg->uk_init = zero_init; 2412 2413 if (arg->flags & UMA_ZONE_MALLOC) 2414 keg->uk_flags |= UMA_ZFLAG_VTOSLAB; 2415 2416 #ifndef SMP 2417 keg->uk_flags &= ~UMA_ZONE_PCPU; 2418 #endif 2419 2420 keg_layout(keg); 2421 2422 /* 2423 * Use a first-touch NUMA policy for kegs that pmap_extract() will 2424 * work on. Use round-robin for everything else. 2425 * 2426 * Zones may override the default by specifying either. 2427 */ 2428 #ifdef NUMA 2429 if ((keg->uk_flags & 2430 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0) 2431 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH; 2432 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0) 2433 keg->uk_flags |= UMA_ZONE_ROUNDROBIN; 2434 #endif 2435 2436 /* 2437 * If we haven't booted yet we need allocations to go through the 2438 * startup cache until the vm is ready. 2439 */ 2440 #ifdef UMA_MD_SMALL_ALLOC 2441 if (keg->uk_ppera == 1) 2442 keg->uk_allocf = uma_small_alloc; 2443 else 2444 #endif 2445 if (booted < BOOT_KVA) 2446 keg->uk_allocf = startup_alloc; 2447 else if (keg->uk_flags & UMA_ZONE_PCPU) 2448 keg->uk_allocf = pcpu_page_alloc; 2449 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1) 2450 keg->uk_allocf = contig_alloc; 2451 else 2452 keg->uk_allocf = page_alloc; 2453 #ifdef UMA_MD_SMALL_ALLOC 2454 if (keg->uk_ppera == 1) 2455 keg->uk_freef = uma_small_free; 2456 else 2457 #endif 2458 if (keg->uk_flags & UMA_ZONE_PCPU) 2459 keg->uk_freef = pcpu_page_free; 2460 else 2461 keg->uk_freef = page_free; 2462 2463 /* 2464 * Initialize keg's locks. 2465 */ 2466 for (i = 0; i < vm_ndomains; i++) 2467 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS)); 2468 2469 /* 2470 * If we're putting the slab header in the actual page we need to 2471 * figure out where in each page it goes. See slab_sizeof 2472 * definition. 2473 */ 2474 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) { 2475 size_t shsize; 2476 2477 shsize = slab_sizeof(keg->uk_ipers); 2478 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize; 2479 /* 2480 * The only way the following is possible is if with our 2481 * UMA_ALIGN_PTR adjustments we are now bigger than 2482 * UMA_SLAB_SIZE. I haven't checked whether this is 2483 * mathematically possible for all cases, so we make 2484 * sure here anyway. 2485 */ 2486 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera, 2487 ("zone %s ipers %d rsize %d size %d slab won't fit", 2488 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size)); 2489 } 2490 2491 if (keg->uk_flags & UMA_ZFLAG_HASH) 2492 hash_alloc(&keg->uk_hash, 0); 2493 2494 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone); 2495 2496 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 2497 2498 rw_wlock(&uma_rwlock); 2499 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 2500 rw_wunlock(&uma_rwlock); 2501 return (0); 2502 } 2503 2504 static void 2505 zone_kva_available(uma_zone_t zone, void *unused) 2506 { 2507 uma_keg_t keg; 2508 2509 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 2510 return; 2511 KEG_GET(zone, keg); 2512 2513 if (keg->uk_allocf == startup_alloc) { 2514 /* Switch to the real allocator. */ 2515 if (keg->uk_flags & UMA_ZONE_PCPU) 2516 keg->uk_allocf = pcpu_page_alloc; 2517 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && 2518 keg->uk_ppera > 1) 2519 keg->uk_allocf = contig_alloc; 2520 else 2521 keg->uk_allocf = page_alloc; 2522 } 2523 } 2524 2525 static void 2526 zone_alloc_counters(uma_zone_t zone, void *unused) 2527 { 2528 2529 zone->uz_allocs = counter_u64_alloc(M_WAITOK); 2530 zone->uz_frees = counter_u64_alloc(M_WAITOK); 2531 zone->uz_fails = counter_u64_alloc(M_WAITOK); 2532 zone->uz_xdomain = counter_u64_alloc(M_WAITOK); 2533 } 2534 2535 static void 2536 zone_alloc_sysctl(uma_zone_t zone, void *unused) 2537 { 2538 uma_zone_domain_t zdom; 2539 uma_domain_t dom; 2540 uma_keg_t keg; 2541 struct sysctl_oid *oid, *domainoid; 2542 int domains, i, cnt; 2543 static const char *nokeg = "cache zone"; 2544 char *c; 2545 2546 /* 2547 * Make a sysctl safe copy of the zone name by removing 2548 * any special characters and handling dups by appending 2549 * an index. 2550 */ 2551 if (zone->uz_namecnt != 0) { 2552 /* Count the number of decimal digits and '_' separator. */ 2553 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++) 2554 cnt /= 10; 2555 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1, 2556 M_UMA, M_WAITOK); 2557 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name, 2558 zone->uz_namecnt); 2559 } else 2560 zone->uz_ctlname = strdup(zone->uz_name, M_UMA); 2561 for (c = zone->uz_ctlname; *c != '\0'; c++) 2562 if (strchr("./\\ -", *c) != NULL) 2563 *c = '_'; 2564 2565 /* 2566 * Basic parameters at the root. 2567 */ 2568 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma), 2569 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2570 oid = zone->uz_oid; 2571 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2572 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size"); 2573 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2574 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE, 2575 zone, 0, sysctl_handle_uma_zone_flags, "A", 2576 "Allocator configuration flags"); 2577 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2578 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0, 2579 "Desired per-cpu cache size"); 2580 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2581 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0, 2582 "Maximum allowed per-cpu cache size"); 2583 2584 /* 2585 * keg if present. 2586 */ 2587 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0) 2588 domains = vm_ndomains; 2589 else 2590 domains = 1; 2591 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 2592 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2593 keg = zone->uz_keg; 2594 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) { 2595 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2596 "name", CTLFLAG_RD, keg->uk_name, "Keg name"); 2597 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2598 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0, 2599 "Real object size with alignment"); 2600 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2601 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0, 2602 "pages per-slab allocation"); 2603 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2604 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0, 2605 "items available per-slab"); 2606 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2607 "align", CTLFLAG_RD, &keg->uk_align, 0, 2608 "item alignment mask"); 2609 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2610 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0, 2611 "number of reserved items"); 2612 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2613 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE, 2614 keg, 0, sysctl_handle_uma_slab_efficiency, "I", 2615 "Slab utilization (100 - internal fragmentation %)"); 2616 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid), 2617 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2618 for (i = 0; i < domains; i++) { 2619 dom = &keg->uk_domain[i]; 2620 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid), 2621 OID_AUTO, VM_DOMAIN(i)->vmd_name, 2622 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2623 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2624 "pages", CTLFLAG_RD, &dom->ud_pages, 0, 2625 "Total pages currently allocated from VM"); 2626 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2627 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0, 2628 "items free in the slab layer"); 2629 } 2630 } else 2631 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2632 "name", CTLFLAG_RD, nokeg, "Keg name"); 2633 2634 /* 2635 * Information about zone limits. 2636 */ 2637 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 2638 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2639 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2640 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2641 zone, 0, sysctl_handle_uma_zone_items, "QU", 2642 "Current number of allocated items if limit is set"); 2643 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2644 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0, 2645 "Maximum number of allocated and cached items"); 2646 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2647 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0, 2648 "Number of threads sleeping at limit"); 2649 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2650 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0, 2651 "Total zone limit sleeps"); 2652 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2653 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0, 2654 "Maximum number of items in each domain's bucket cache"); 2655 2656 /* 2657 * Per-domain zone information. 2658 */ 2659 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), 2660 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2661 for (i = 0; i < domains; i++) { 2662 zdom = ZDOM_GET(zone, i); 2663 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid), 2664 OID_AUTO, VM_DOMAIN(i)->vmd_name, 2665 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2666 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2667 "nitems", CTLFLAG_RD, &zdom->uzd_nitems, 2668 "number of items in this domain"); 2669 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2670 "imax", CTLFLAG_RD, &zdom->uzd_imax, 2671 "maximum item count in this period"); 2672 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2673 "imin", CTLFLAG_RD, &zdom->uzd_imin, 2674 "minimum item count in this period"); 2675 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2676 "bimin", CTLFLAG_RD, &zdom->uzd_bimin, 2677 "Minimum item count in this batch"); 2678 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2679 "wss", CTLFLAG_RD, &zdom->uzd_wss, 2680 "Working set size"); 2681 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2682 "limin", CTLFLAG_RD, &zdom->uzd_limin, 2683 "Long time minimum item count"); 2684 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2685 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0, 2686 "Time since zero long time minimum item count"); 2687 } 2688 2689 /* 2690 * General statistics. 2691 */ 2692 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 2693 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2694 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2695 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE, 2696 zone, 1, sysctl_handle_uma_zone_cur, "I", 2697 "Current number of allocated items"); 2698 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2699 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2700 zone, 0, sysctl_handle_uma_zone_allocs, "QU", 2701 "Total allocation calls"); 2702 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2703 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2704 zone, 0, sysctl_handle_uma_zone_frees, "QU", 2705 "Total free calls"); 2706 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2707 "fails", CTLFLAG_RD, &zone->uz_fails, 2708 "Number of allocation failures"); 2709 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2710 "xdomain", CTLFLAG_RD, &zone->uz_xdomain, 2711 "Free calls from the wrong domain"); 2712 } 2713 2714 struct uma_zone_count { 2715 const char *name; 2716 int count; 2717 }; 2718 2719 static void 2720 zone_count(uma_zone_t zone, void *arg) 2721 { 2722 struct uma_zone_count *cnt; 2723 2724 cnt = arg; 2725 /* 2726 * Some zones are rapidly created with identical names and 2727 * destroyed out of order. This can lead to gaps in the count. 2728 * Use one greater than the maximum observed for this name. 2729 */ 2730 if (strcmp(zone->uz_name, cnt->name) == 0) 2731 cnt->count = MAX(cnt->count, 2732 zone->uz_namecnt + 1); 2733 } 2734 2735 static void 2736 zone_update_caches(uma_zone_t zone) 2737 { 2738 int i; 2739 2740 for (i = 0; i <= mp_maxid; i++) { 2741 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size); 2742 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags); 2743 } 2744 } 2745 2746 /* 2747 * Zone header ctor. This initializes all fields, locks, etc. 2748 * 2749 * Arguments/Returns follow uma_ctor specifications 2750 * udata Actually uma_zctor_args 2751 */ 2752 static int 2753 zone_ctor(void *mem, int size, void *udata, int flags) 2754 { 2755 struct uma_zone_count cnt; 2756 struct uma_zctor_args *arg = udata; 2757 uma_zone_domain_t zdom; 2758 uma_zone_t zone = mem; 2759 uma_zone_t z; 2760 uma_keg_t keg; 2761 int i; 2762 2763 bzero(zone, size); 2764 zone->uz_name = arg->name; 2765 zone->uz_ctor = arg->ctor; 2766 zone->uz_dtor = arg->dtor; 2767 zone->uz_init = NULL; 2768 zone->uz_fini = NULL; 2769 zone->uz_sleeps = 0; 2770 zone->uz_bucket_size = 0; 2771 zone->uz_bucket_size_min = 0; 2772 zone->uz_bucket_size_max = BUCKET_MAX; 2773 zone->uz_flags = (arg->flags & UMA_ZONE_SMR); 2774 zone->uz_warning = NULL; 2775 /* The domain structures follow the cpu structures. */ 2776 zone->uz_bucket_max = ULONG_MAX; 2777 timevalclear(&zone->uz_ratecheck); 2778 2779 /* Count the number of duplicate names. */ 2780 cnt.name = arg->name; 2781 cnt.count = 0; 2782 zone_foreach(zone_count, &cnt); 2783 zone->uz_namecnt = cnt.count; 2784 ZONE_CROSS_LOCK_INIT(zone); 2785 2786 for (i = 0; i < vm_ndomains; i++) { 2787 zdom = ZDOM_GET(zone, i); 2788 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS)); 2789 STAILQ_INIT(&zdom->uzd_buckets); 2790 } 2791 2792 #if defined(INVARIANTS) && !defined(KASAN) 2793 if (arg->uminit == trash_init && arg->fini == trash_fini) 2794 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR; 2795 #elif defined(KASAN) 2796 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0) 2797 arg->flags |= UMA_ZONE_NOKASAN; 2798 #endif 2799 2800 /* 2801 * This is a pure cache zone, no kegs. 2802 */ 2803 if (arg->import) { 2804 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0, 2805 ("zone_ctor: Import specified for non-cache zone.")); 2806 zone->uz_flags = arg->flags; 2807 zone->uz_size = arg->size; 2808 zone->uz_import = arg->import; 2809 zone->uz_release = arg->release; 2810 zone->uz_arg = arg->arg; 2811 #ifdef NUMA 2812 /* 2813 * Cache zones are round-robin unless a policy is 2814 * specified because they may have incompatible 2815 * constraints. 2816 */ 2817 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0) 2818 zone->uz_flags |= UMA_ZONE_ROUNDROBIN; 2819 #endif 2820 rw_wlock(&uma_rwlock); 2821 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 2822 rw_wunlock(&uma_rwlock); 2823 goto out; 2824 } 2825 2826 /* 2827 * Use the regular zone/keg/slab allocator. 2828 */ 2829 zone->uz_import = zone_import; 2830 zone->uz_release = zone_release; 2831 zone->uz_arg = zone; 2832 keg = arg->keg; 2833 2834 if (arg->flags & UMA_ZONE_SECONDARY) { 2835 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 2836 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 2837 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 2838 zone->uz_init = arg->uminit; 2839 zone->uz_fini = arg->fini; 2840 zone->uz_flags |= UMA_ZONE_SECONDARY; 2841 rw_wlock(&uma_rwlock); 2842 ZONE_LOCK(zone); 2843 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 2844 if (LIST_NEXT(z, uz_link) == NULL) { 2845 LIST_INSERT_AFTER(z, zone, uz_link); 2846 break; 2847 } 2848 } 2849 ZONE_UNLOCK(zone); 2850 rw_wunlock(&uma_rwlock); 2851 } else if (keg == NULL) { 2852 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 2853 arg->align, arg->flags)) == NULL) 2854 return (ENOMEM); 2855 } else { 2856 struct uma_kctor_args karg; 2857 int error; 2858 2859 /* We should only be here from uma_startup() */ 2860 karg.size = arg->size; 2861 karg.uminit = arg->uminit; 2862 karg.fini = arg->fini; 2863 karg.align = arg->align; 2864 karg.flags = (arg->flags & ~UMA_ZONE_SMR); 2865 karg.zone = zone; 2866 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 2867 flags); 2868 if (error) 2869 return (error); 2870 } 2871 2872 /* Inherit properties from the keg. */ 2873 zone->uz_keg = keg; 2874 zone->uz_size = keg->uk_size; 2875 zone->uz_flags |= (keg->uk_flags & 2876 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 2877 2878 out: 2879 if (booted >= BOOT_PCPU) { 2880 zone_alloc_counters(zone, NULL); 2881 if (booted >= BOOT_RUNNING) 2882 zone_alloc_sysctl(zone, NULL); 2883 } else { 2884 zone->uz_allocs = EARLY_COUNTER; 2885 zone->uz_frees = EARLY_COUNTER; 2886 zone->uz_fails = EARLY_COUNTER; 2887 } 2888 2889 /* Caller requests a private SMR context. */ 2890 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 2891 zone->uz_smr = smr_create(zone->uz_name, 0, 0); 2892 2893 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) != 2894 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET), 2895 ("Invalid zone flag combination")); 2896 if (arg->flags & UMA_ZFLAG_INTERNAL) 2897 zone->uz_bucket_size_max = zone->uz_bucket_size = 0; 2898 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0) 2899 zone->uz_bucket_size = BUCKET_MAX; 2900 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0) 2901 zone->uz_bucket_size = 0; 2902 else 2903 zone->uz_bucket_size = bucket_select(zone->uz_size); 2904 zone->uz_bucket_size_min = zone->uz_bucket_size; 2905 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL) 2906 zone->uz_flags |= UMA_ZFLAG_CTORDTOR; 2907 zone_update_caches(zone); 2908 2909 return (0); 2910 } 2911 2912 /* 2913 * Keg header dtor. This frees all data, destroys locks, frees the hash 2914 * table and removes the keg from the global list. 2915 * 2916 * Arguments/Returns follow uma_dtor specifications 2917 * udata unused 2918 */ 2919 static void 2920 keg_dtor(void *arg, int size, void *udata) 2921 { 2922 uma_keg_t keg; 2923 uint32_t free, pages; 2924 int i; 2925 2926 keg = (uma_keg_t)arg; 2927 free = pages = 0; 2928 for (i = 0; i < vm_ndomains; i++) { 2929 free += keg->uk_domain[i].ud_free_items; 2930 pages += keg->uk_domain[i].ud_pages; 2931 KEG_LOCK_FINI(keg, i); 2932 } 2933 if (pages != 0) 2934 printf("Freed UMA keg (%s) was not empty (%u items). " 2935 " Lost %u pages of memory.\n", 2936 keg->uk_name ? keg->uk_name : "", 2937 pages / keg->uk_ppera * keg->uk_ipers - free, pages); 2938 2939 hash_free(&keg->uk_hash); 2940 } 2941 2942 /* 2943 * Zone header dtor. 2944 * 2945 * Arguments/Returns follow uma_dtor specifications 2946 * udata unused 2947 */ 2948 static void 2949 zone_dtor(void *arg, int size, void *udata) 2950 { 2951 uma_zone_t zone; 2952 uma_keg_t keg; 2953 int i; 2954 2955 zone = (uma_zone_t)arg; 2956 2957 sysctl_remove_oid(zone->uz_oid, 1, 1); 2958 2959 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 2960 cache_drain(zone); 2961 2962 rw_wlock(&uma_rwlock); 2963 LIST_REMOVE(zone, uz_link); 2964 rw_wunlock(&uma_rwlock); 2965 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) { 2966 keg = zone->uz_keg; 2967 keg->uk_reserve = 0; 2968 } 2969 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true); 2970 2971 /* 2972 * We only destroy kegs from non secondary/non cache zones. 2973 */ 2974 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) { 2975 keg = zone->uz_keg; 2976 rw_wlock(&uma_rwlock); 2977 LIST_REMOVE(keg, uk_link); 2978 rw_wunlock(&uma_rwlock); 2979 zone_free_item(kegs, keg, NULL, SKIP_NONE); 2980 } 2981 counter_u64_free(zone->uz_allocs); 2982 counter_u64_free(zone->uz_frees); 2983 counter_u64_free(zone->uz_fails); 2984 counter_u64_free(zone->uz_xdomain); 2985 free(zone->uz_ctlname, M_UMA); 2986 for (i = 0; i < vm_ndomains; i++) 2987 ZDOM_LOCK_FINI(ZDOM_GET(zone, i)); 2988 ZONE_CROSS_LOCK_FINI(zone); 2989 } 2990 2991 static void 2992 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg) 2993 { 2994 uma_keg_t keg; 2995 uma_zone_t zone; 2996 2997 LIST_FOREACH(keg, &uma_kegs, uk_link) { 2998 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 2999 zfunc(zone, arg); 3000 } 3001 LIST_FOREACH(zone, &uma_cachezones, uz_link) 3002 zfunc(zone, arg); 3003 } 3004 3005 /* 3006 * Traverses every zone in the system and calls a callback 3007 * 3008 * Arguments: 3009 * zfunc A pointer to a function which accepts a zone 3010 * as an argument. 3011 * 3012 * Returns: 3013 * Nothing 3014 */ 3015 static void 3016 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg) 3017 { 3018 3019 rw_rlock(&uma_rwlock); 3020 zone_foreach_unlocked(zfunc, arg); 3021 rw_runlock(&uma_rwlock); 3022 } 3023 3024 /* 3025 * Initialize the kernel memory allocator. This is done after pages can be 3026 * allocated but before general KVA is available. 3027 */ 3028 void 3029 uma_startup1(vm_offset_t virtual_avail) 3030 { 3031 struct uma_zctor_args args; 3032 size_t ksize, zsize, size; 3033 uma_keg_t primarykeg; 3034 uintptr_t m; 3035 int domain; 3036 uint8_t pflag; 3037 3038 bootstart = bootmem = virtual_avail; 3039 3040 rw_init(&uma_rwlock, "UMA lock"); 3041 sx_init(&uma_reclaim_lock, "umareclaim"); 3042 3043 ksize = sizeof(struct uma_keg) + 3044 (sizeof(struct uma_domain) * vm_ndomains); 3045 ksize = roundup(ksize, UMA_SUPER_ALIGN); 3046 zsize = sizeof(struct uma_zone) + 3047 (sizeof(struct uma_cache) * (mp_maxid + 1)) + 3048 (sizeof(struct uma_zone_domain) * vm_ndomains); 3049 zsize = roundup(zsize, UMA_SUPER_ALIGN); 3050 3051 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */ 3052 size = (zsize * 2) + ksize; 3053 for (domain = 0; domain < vm_ndomains; domain++) { 3054 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag, 3055 M_NOWAIT | M_ZERO); 3056 if (m != 0) 3057 break; 3058 } 3059 zones = (uma_zone_t)m; 3060 m += zsize; 3061 kegs = (uma_zone_t)m; 3062 m += zsize; 3063 primarykeg = (uma_keg_t)m; 3064 3065 /* "manually" create the initial zone */ 3066 memset(&args, 0, sizeof(args)); 3067 args.name = "UMA Kegs"; 3068 args.size = ksize; 3069 args.ctor = keg_ctor; 3070 args.dtor = keg_dtor; 3071 args.uminit = zero_init; 3072 args.fini = NULL; 3073 args.keg = primarykeg; 3074 args.align = UMA_SUPER_ALIGN - 1; 3075 args.flags = UMA_ZFLAG_INTERNAL; 3076 zone_ctor(kegs, zsize, &args, M_WAITOK); 3077 3078 args.name = "UMA Zones"; 3079 args.size = zsize; 3080 args.ctor = zone_ctor; 3081 args.dtor = zone_dtor; 3082 args.uminit = zero_init; 3083 args.fini = NULL; 3084 args.keg = NULL; 3085 args.align = UMA_SUPER_ALIGN - 1; 3086 args.flags = UMA_ZFLAG_INTERNAL; 3087 zone_ctor(zones, zsize, &args, M_WAITOK); 3088 3089 /* Now make zones for slab headers */ 3090 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE, 3091 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 3092 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE, 3093 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 3094 3095 hashzone = uma_zcreate("UMA Hash", 3096 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 3097 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 3098 3099 bucket_init(); 3100 smr_init(); 3101 } 3102 3103 #ifndef UMA_MD_SMALL_ALLOC 3104 extern void vm_radix_reserve_kva(void); 3105 #endif 3106 3107 /* 3108 * Advertise the availability of normal kva allocations and switch to 3109 * the default back-end allocator. Marks the KVA we consumed on startup 3110 * as used in the map. 3111 */ 3112 void 3113 uma_startup2(void) 3114 { 3115 3116 if (bootstart != bootmem) { 3117 vm_map_lock(kernel_map); 3118 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem, 3119 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 3120 vm_map_unlock(kernel_map); 3121 } 3122 3123 #ifndef UMA_MD_SMALL_ALLOC 3124 /* Set up radix zone to use noobj_alloc. */ 3125 vm_radix_reserve_kva(); 3126 #endif 3127 3128 booted = BOOT_KVA; 3129 zone_foreach_unlocked(zone_kva_available, NULL); 3130 bucket_enable(); 3131 } 3132 3133 /* 3134 * Allocate counters as early as possible so that boot-time allocations are 3135 * accounted more precisely. 3136 */ 3137 static void 3138 uma_startup_pcpu(void *arg __unused) 3139 { 3140 3141 zone_foreach_unlocked(zone_alloc_counters, NULL); 3142 booted = BOOT_PCPU; 3143 } 3144 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL); 3145 3146 /* 3147 * Finish our initialization steps. 3148 */ 3149 static void 3150 uma_startup3(void *arg __unused) 3151 { 3152 3153 #ifdef INVARIANTS 3154 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor); 3155 uma_dbg_cnt = counter_u64_alloc(M_WAITOK); 3156 uma_skip_cnt = counter_u64_alloc(M_WAITOK); 3157 #endif 3158 zone_foreach_unlocked(zone_alloc_sysctl, NULL); 3159 callout_init(&uma_callout, 1); 3160 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 3161 booted = BOOT_RUNNING; 3162 3163 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL, 3164 EVENTHANDLER_PRI_FIRST); 3165 } 3166 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 3167 3168 static void 3169 uma_shutdown(void) 3170 { 3171 3172 booted = BOOT_SHUTDOWN; 3173 } 3174 3175 static uma_keg_t 3176 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 3177 int align, uint32_t flags) 3178 { 3179 struct uma_kctor_args args; 3180 3181 args.size = size; 3182 args.uminit = uminit; 3183 args.fini = fini; 3184 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 3185 args.flags = flags; 3186 args.zone = zone; 3187 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK)); 3188 } 3189 3190 /* Public functions */ 3191 /* See uma.h */ 3192 void 3193 uma_set_align(int align) 3194 { 3195 3196 if (align != UMA_ALIGN_CACHE) 3197 uma_align_cache = align; 3198 } 3199 3200 /* See uma.h */ 3201 uma_zone_t 3202 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 3203 uma_init uminit, uma_fini fini, int align, uint32_t flags) 3204 3205 { 3206 struct uma_zctor_args args; 3207 uma_zone_t res; 3208 3209 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"", 3210 align, name)); 3211 3212 /* This stuff is essential for the zone ctor */ 3213 memset(&args, 0, sizeof(args)); 3214 args.name = name; 3215 args.size = size; 3216 args.ctor = ctor; 3217 args.dtor = dtor; 3218 args.uminit = uminit; 3219 args.fini = fini; 3220 #if defined(INVARIANTS) && !defined(KASAN) 3221 /* 3222 * Inject procedures which check for memory use after free if we are 3223 * allowed to scramble the memory while it is not allocated. This 3224 * requires that: UMA is actually able to access the memory, no init 3225 * or fini procedures, no dependency on the initial value of the 3226 * memory, and no (legitimate) use of the memory after free. Note, 3227 * the ctor and dtor do not need to be empty. 3228 */ 3229 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH | 3230 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) { 3231 args.uminit = trash_init; 3232 args.fini = trash_fini; 3233 } 3234 #endif 3235 args.align = align; 3236 args.flags = flags; 3237 args.keg = NULL; 3238 3239 sx_xlock(&uma_reclaim_lock); 3240 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 3241 sx_xunlock(&uma_reclaim_lock); 3242 3243 return (res); 3244 } 3245 3246 /* See uma.h */ 3247 uma_zone_t 3248 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor, 3249 uma_init zinit, uma_fini zfini, uma_zone_t primary) 3250 { 3251 struct uma_zctor_args args; 3252 uma_keg_t keg; 3253 uma_zone_t res; 3254 3255 keg = primary->uz_keg; 3256 memset(&args, 0, sizeof(args)); 3257 args.name = name; 3258 args.size = keg->uk_size; 3259 args.ctor = ctor; 3260 args.dtor = dtor; 3261 args.uminit = zinit; 3262 args.fini = zfini; 3263 args.align = keg->uk_align; 3264 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 3265 args.keg = keg; 3266 3267 sx_xlock(&uma_reclaim_lock); 3268 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 3269 sx_xunlock(&uma_reclaim_lock); 3270 3271 return (res); 3272 } 3273 3274 /* See uma.h */ 3275 uma_zone_t 3276 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor, 3277 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease, 3278 void *arg, int flags) 3279 { 3280 struct uma_zctor_args args; 3281 3282 memset(&args, 0, sizeof(args)); 3283 args.name = name; 3284 args.size = size; 3285 args.ctor = ctor; 3286 args.dtor = dtor; 3287 args.uminit = zinit; 3288 args.fini = zfini; 3289 args.import = zimport; 3290 args.release = zrelease; 3291 args.arg = arg; 3292 args.align = 0; 3293 args.flags = flags | UMA_ZFLAG_CACHE; 3294 3295 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK)); 3296 } 3297 3298 /* See uma.h */ 3299 void 3300 uma_zdestroy(uma_zone_t zone) 3301 { 3302 3303 /* 3304 * Large slabs are expensive to reclaim, so don't bother doing 3305 * unnecessary work if we're shutting down. 3306 */ 3307 if (booted == BOOT_SHUTDOWN && 3308 zone->uz_fini == NULL && zone->uz_release == zone_release) 3309 return; 3310 sx_xlock(&uma_reclaim_lock); 3311 zone_free_item(zones, zone, NULL, SKIP_NONE); 3312 sx_xunlock(&uma_reclaim_lock); 3313 } 3314 3315 void 3316 uma_zwait(uma_zone_t zone) 3317 { 3318 3319 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 3320 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK)); 3321 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0) 3322 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK)); 3323 else 3324 uma_zfree(zone, uma_zalloc(zone, M_WAITOK)); 3325 } 3326 3327 void * 3328 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags) 3329 { 3330 void *item, *pcpu_item; 3331 #ifdef SMP 3332 int i; 3333 3334 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 3335 #endif 3336 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO); 3337 if (item == NULL) 3338 return (NULL); 3339 pcpu_item = zpcpu_base_to_offset(item); 3340 if (flags & M_ZERO) { 3341 #ifdef SMP 3342 for (i = 0; i <= mp_maxid; i++) 3343 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size); 3344 #else 3345 bzero(item, zone->uz_size); 3346 #endif 3347 } 3348 return (pcpu_item); 3349 } 3350 3351 /* 3352 * A stub while both regular and pcpu cases are identical. 3353 */ 3354 void 3355 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata) 3356 { 3357 void *item; 3358 3359 #ifdef SMP 3360 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 3361 #endif 3362 3363 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */ 3364 if (pcpu_item == NULL) 3365 return; 3366 3367 item = zpcpu_offset_to_base(pcpu_item); 3368 uma_zfree_arg(zone, item, udata); 3369 } 3370 3371 static inline void * 3372 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags, 3373 void *item) 3374 { 3375 #ifdef INVARIANTS 3376 bool skipdbg; 3377 #endif 3378 3379 kasan_mark_item_valid(zone, item); 3380 3381 #ifdef INVARIANTS 3382 skipdbg = uma_dbg_zskip(zone, item); 3383 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 && 3384 zone->uz_ctor != trash_ctor) 3385 trash_ctor(item, size, udata, flags); 3386 #endif 3387 3388 /* Check flags before loading ctor pointer. */ 3389 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) && 3390 __predict_false(zone->uz_ctor != NULL) && 3391 zone->uz_ctor(item, size, udata, flags) != 0) { 3392 counter_u64_add(zone->uz_fails, 1); 3393 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT); 3394 return (NULL); 3395 } 3396 #ifdef INVARIANTS 3397 if (!skipdbg) 3398 uma_dbg_alloc(zone, NULL, item); 3399 #endif 3400 if (__predict_false(flags & M_ZERO)) 3401 return (memset(item, 0, size)); 3402 3403 return (item); 3404 } 3405 3406 static inline void 3407 item_dtor(uma_zone_t zone, void *item, int size, void *udata, 3408 enum zfreeskip skip) 3409 { 3410 #ifdef INVARIANTS 3411 bool skipdbg; 3412 3413 skipdbg = uma_dbg_zskip(zone, item); 3414 if (skip == SKIP_NONE && !skipdbg) { 3415 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0) 3416 uma_dbg_free(zone, udata, item); 3417 else 3418 uma_dbg_free(zone, NULL, item); 3419 } 3420 #endif 3421 if (__predict_true(skip < SKIP_DTOR)) { 3422 if (zone->uz_dtor != NULL) 3423 zone->uz_dtor(item, size, udata); 3424 #ifdef INVARIANTS 3425 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 && 3426 zone->uz_dtor != trash_dtor) 3427 trash_dtor(item, size, udata); 3428 #endif 3429 } 3430 kasan_mark_item_invalid(zone, item); 3431 } 3432 3433 #ifdef NUMA 3434 static int 3435 item_domain(void *item) 3436 { 3437 int domain; 3438 3439 domain = vm_phys_domain(vtophys(item)); 3440 KASSERT(domain >= 0 && domain < vm_ndomains, 3441 ("%s: unknown domain for item %p", __func__, item)); 3442 return (domain); 3443 } 3444 #endif 3445 3446 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS) 3447 #define UMA_ZALLOC_DEBUG 3448 static int 3449 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags) 3450 { 3451 int error; 3452 3453 error = 0; 3454 #ifdef WITNESS 3455 if (flags & M_WAITOK) { 3456 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 3457 "uma_zalloc_debug: zone \"%s\"", zone->uz_name); 3458 } 3459 #endif 3460 3461 #ifdef INVARIANTS 3462 KASSERT((flags & M_EXEC) == 0, 3463 ("uma_zalloc_debug: called with M_EXEC")); 3464 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3465 ("uma_zalloc_debug: called within spinlock or critical section")); 3466 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0, 3467 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO")); 3468 #endif 3469 3470 #ifdef DEBUG_MEMGUARD 3471 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) { 3472 void *item; 3473 item = memguard_alloc(zone->uz_size, flags); 3474 if (item != NULL) { 3475 error = EJUSTRETURN; 3476 if (zone->uz_init != NULL && 3477 zone->uz_init(item, zone->uz_size, flags) != 0) { 3478 *itemp = NULL; 3479 return (error); 3480 } 3481 if (zone->uz_ctor != NULL && 3482 zone->uz_ctor(item, zone->uz_size, udata, 3483 flags) != 0) { 3484 counter_u64_add(zone->uz_fails, 1); 3485 zone->uz_fini(item, zone->uz_size); 3486 *itemp = NULL; 3487 return (error); 3488 } 3489 *itemp = item; 3490 return (error); 3491 } 3492 /* This is unfortunate but should not be fatal. */ 3493 } 3494 #endif 3495 return (error); 3496 } 3497 3498 static int 3499 uma_zfree_debug(uma_zone_t zone, void *item, void *udata) 3500 { 3501 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3502 ("uma_zfree_debug: called with spinlock or critical section held")); 3503 3504 #ifdef DEBUG_MEMGUARD 3505 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) { 3506 if (zone->uz_dtor != NULL) 3507 zone->uz_dtor(item, zone->uz_size, udata); 3508 if (zone->uz_fini != NULL) 3509 zone->uz_fini(item, zone->uz_size); 3510 memguard_free(item); 3511 return (EJUSTRETURN); 3512 } 3513 #endif 3514 return (0); 3515 } 3516 #endif 3517 3518 static inline void * 3519 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket, 3520 void *udata, int flags) 3521 { 3522 void *item; 3523 int size, uz_flags; 3524 3525 item = cache_bucket_pop(cache, bucket); 3526 size = cache_uz_size(cache); 3527 uz_flags = cache_uz_flags(cache); 3528 critical_exit(); 3529 return (item_ctor(zone, uz_flags, size, udata, flags, item)); 3530 } 3531 3532 static __noinline void * 3533 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags) 3534 { 3535 uma_cache_bucket_t bucket; 3536 int domain; 3537 3538 while (cache_alloc(zone, cache, udata, flags)) { 3539 cache = &zone->uz_cpu[curcpu]; 3540 bucket = &cache->uc_allocbucket; 3541 if (__predict_false(bucket->ucb_cnt == 0)) 3542 continue; 3543 return (cache_alloc_item(zone, cache, bucket, udata, flags)); 3544 } 3545 critical_exit(); 3546 3547 /* 3548 * We can not get a bucket so try to return a single item. 3549 */ 3550 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) 3551 domain = PCPU_GET(domain); 3552 else 3553 domain = UMA_ANYDOMAIN; 3554 return (zone_alloc_item(zone, udata, domain, flags)); 3555 } 3556 3557 /* See uma.h */ 3558 void * 3559 uma_zalloc_smr(uma_zone_t zone, int flags) 3560 { 3561 uma_cache_bucket_t bucket; 3562 uma_cache_t cache; 3563 3564 #ifdef UMA_ZALLOC_DEBUG 3565 void *item; 3566 3567 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0, 3568 ("uma_zalloc_arg: called with non-SMR zone.")); 3569 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN) 3570 return (item); 3571 #endif 3572 3573 critical_enter(); 3574 cache = &zone->uz_cpu[curcpu]; 3575 bucket = &cache->uc_allocbucket; 3576 if (__predict_false(bucket->ucb_cnt == 0)) 3577 return (cache_alloc_retry(zone, cache, NULL, flags)); 3578 return (cache_alloc_item(zone, cache, bucket, NULL, flags)); 3579 } 3580 3581 /* See uma.h */ 3582 void * 3583 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 3584 { 3585 uma_cache_bucket_t bucket; 3586 uma_cache_t cache; 3587 3588 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3589 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3590 3591 /* This is the fast path allocation */ 3592 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name, 3593 zone, flags); 3594 3595 #ifdef UMA_ZALLOC_DEBUG 3596 void *item; 3597 3598 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0, 3599 ("uma_zalloc_arg: called with SMR zone.")); 3600 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN) 3601 return (item); 3602 #endif 3603 3604 /* 3605 * If possible, allocate from the per-CPU cache. There are two 3606 * requirements for safe access to the per-CPU cache: (1) the thread 3607 * accessing the cache must not be preempted or yield during access, 3608 * and (2) the thread must not migrate CPUs without switching which 3609 * cache it accesses. We rely on a critical section to prevent 3610 * preemption and migration. We release the critical section in 3611 * order to acquire the zone mutex if we are unable to allocate from 3612 * the current cache; when we re-acquire the critical section, we 3613 * must detect and handle migration if it has occurred. 3614 */ 3615 critical_enter(); 3616 cache = &zone->uz_cpu[curcpu]; 3617 bucket = &cache->uc_allocbucket; 3618 if (__predict_false(bucket->ucb_cnt == 0)) 3619 return (cache_alloc_retry(zone, cache, udata, flags)); 3620 return (cache_alloc_item(zone, cache, bucket, udata, flags)); 3621 } 3622 3623 /* 3624 * Replenish an alloc bucket and possibly restore an old one. Called in 3625 * a critical section. Returns in a critical section. 3626 * 3627 * A false return value indicates an allocation failure. 3628 * A true return value indicates success and the caller should retry. 3629 */ 3630 static __noinline bool 3631 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags) 3632 { 3633 uma_bucket_t bucket; 3634 int curdomain, domain; 3635 bool new; 3636 3637 CRITICAL_ASSERT(curthread); 3638 3639 /* 3640 * If we have run out of items in our alloc bucket see 3641 * if we can switch with the free bucket. 3642 * 3643 * SMR Zones can't re-use the free bucket until the sequence has 3644 * expired. 3645 */ 3646 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 && 3647 cache->uc_freebucket.ucb_cnt != 0) { 3648 cache_bucket_swap(&cache->uc_freebucket, 3649 &cache->uc_allocbucket); 3650 return (true); 3651 } 3652 3653 /* 3654 * Discard any empty allocation bucket while we hold no locks. 3655 */ 3656 bucket = cache_bucket_unload_alloc(cache); 3657 critical_exit(); 3658 3659 if (bucket != NULL) { 3660 KASSERT(bucket->ub_cnt == 0, 3661 ("cache_alloc: Entered with non-empty alloc bucket.")); 3662 bucket_free(zone, bucket, udata); 3663 } 3664 3665 /* 3666 * Attempt to retrieve the item from the per-CPU cache has failed, so 3667 * we must go back to the zone. This requires the zdom lock, so we 3668 * must drop the critical section, then re-acquire it when we go back 3669 * to the cache. Since the critical section is released, we may be 3670 * preempted or migrate. As such, make sure not to maintain any 3671 * thread-local state specific to the cache from prior to releasing 3672 * the critical section. 3673 */ 3674 domain = PCPU_GET(domain); 3675 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 || 3676 VM_DOMAIN_EMPTY(domain)) 3677 domain = zone_domain_highest(zone, domain); 3678 bucket = cache_fetch_bucket(zone, cache, domain); 3679 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) { 3680 bucket = zone_alloc_bucket(zone, udata, domain, flags); 3681 new = true; 3682 } else { 3683 new = false; 3684 } 3685 3686 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p", 3687 zone->uz_name, zone, bucket); 3688 if (bucket == NULL) { 3689 critical_enter(); 3690 return (false); 3691 } 3692 3693 /* 3694 * See if we lost the race or were migrated. Cache the 3695 * initialized bucket to make this less likely or claim 3696 * the memory directly. 3697 */ 3698 critical_enter(); 3699 cache = &zone->uz_cpu[curcpu]; 3700 if (cache->uc_allocbucket.ucb_bucket == NULL && 3701 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 || 3702 (curdomain = PCPU_GET(domain)) == domain || 3703 VM_DOMAIN_EMPTY(curdomain))) { 3704 if (new) 3705 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax, 3706 bucket->ub_cnt); 3707 cache_bucket_load_alloc(cache, bucket); 3708 return (true); 3709 } 3710 3711 /* 3712 * We lost the race, release this bucket and start over. 3713 */ 3714 critical_exit(); 3715 zone_put_bucket(zone, domain, bucket, udata, !new); 3716 critical_enter(); 3717 3718 return (true); 3719 } 3720 3721 void * 3722 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags) 3723 { 3724 #ifdef NUMA 3725 uma_bucket_t bucket; 3726 uma_zone_domain_t zdom; 3727 void *item; 3728 #endif 3729 3730 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3731 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3732 3733 /* This is the fast path allocation */ 3734 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d", 3735 zone->uz_name, zone, domain, flags); 3736 3737 if (flags & M_WAITOK) { 3738 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 3739 "uma_zalloc_domain: zone \"%s\"", zone->uz_name); 3740 } 3741 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3742 ("uma_zalloc_domain: called with spinlock or critical section held")); 3743 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0, 3744 ("uma_zalloc_domain: called with SMR zone.")); 3745 #ifdef NUMA 3746 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0, 3747 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone.")); 3748 3749 if (vm_ndomains == 1) 3750 return (uma_zalloc_arg(zone, udata, flags)); 3751 3752 /* 3753 * Try to allocate from the bucket cache before falling back to the keg. 3754 * We could try harder and attempt to allocate from per-CPU caches or 3755 * the per-domain cross-domain buckets, but the complexity is probably 3756 * not worth it. It is more important that frees of previous 3757 * cross-domain allocations do not blow up the cache. 3758 */ 3759 zdom = zone_domain_lock(zone, domain); 3760 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) { 3761 item = bucket->ub_bucket[bucket->ub_cnt - 1]; 3762 #ifdef INVARIANTS 3763 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL; 3764 #endif 3765 bucket->ub_cnt--; 3766 zone_put_bucket(zone, domain, bucket, udata, true); 3767 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, 3768 flags, item); 3769 if (item != NULL) { 3770 KASSERT(item_domain(item) == domain, 3771 ("%s: bucket cache item %p from wrong domain", 3772 __func__, item)); 3773 counter_u64_add(zone->uz_allocs, 1); 3774 } 3775 return (item); 3776 } 3777 ZDOM_UNLOCK(zdom); 3778 return (zone_alloc_item(zone, udata, domain, flags)); 3779 #else 3780 return (uma_zalloc_arg(zone, udata, flags)); 3781 #endif 3782 } 3783 3784 /* 3785 * Find a slab with some space. Prefer slabs that are partially used over those 3786 * that are totally full. This helps to reduce fragmentation. 3787 * 3788 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check 3789 * only 'domain'. 3790 */ 3791 static uma_slab_t 3792 keg_first_slab(uma_keg_t keg, int domain, bool rr) 3793 { 3794 uma_domain_t dom; 3795 uma_slab_t slab; 3796 int start; 3797 3798 KASSERT(domain >= 0 && domain < vm_ndomains, 3799 ("keg_first_slab: domain %d out of range", domain)); 3800 KEG_LOCK_ASSERT(keg, domain); 3801 3802 slab = NULL; 3803 start = domain; 3804 do { 3805 dom = &keg->uk_domain[domain]; 3806 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL) 3807 return (slab); 3808 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) { 3809 LIST_REMOVE(slab, us_link); 3810 dom->ud_free_slabs--; 3811 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 3812 return (slab); 3813 } 3814 if (rr) 3815 domain = (domain + 1) % vm_ndomains; 3816 } while (domain != start); 3817 3818 return (NULL); 3819 } 3820 3821 /* 3822 * Fetch an existing slab from a free or partial list. Returns with the 3823 * keg domain lock held if a slab was found or unlocked if not. 3824 */ 3825 static uma_slab_t 3826 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags) 3827 { 3828 uma_slab_t slab; 3829 uint32_t reserve; 3830 3831 /* HASH has a single free list. */ 3832 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) 3833 domain = 0; 3834 3835 KEG_LOCK(keg, domain); 3836 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve; 3837 if (keg->uk_domain[domain].ud_free_items <= reserve || 3838 (slab = keg_first_slab(keg, domain, rr)) == NULL) { 3839 KEG_UNLOCK(keg, domain); 3840 return (NULL); 3841 } 3842 return (slab); 3843 } 3844 3845 static uma_slab_t 3846 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags) 3847 { 3848 struct vm_domainset_iter di; 3849 uma_slab_t slab; 3850 int aflags, domain; 3851 bool rr; 3852 3853 restart: 3854 /* 3855 * Use the keg's policy if upper layers haven't already specified a 3856 * domain (as happens with first-touch zones). 3857 * 3858 * To avoid races we run the iterator with the keg lock held, but that 3859 * means that we cannot allow the vm_domainset layer to sleep. Thus, 3860 * clear M_WAITOK and handle low memory conditions locally. 3861 */ 3862 rr = rdomain == UMA_ANYDOMAIN; 3863 if (rr) { 3864 aflags = (flags & ~M_WAITOK) | M_NOWAIT; 3865 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 3866 &aflags); 3867 } else { 3868 aflags = flags; 3869 domain = rdomain; 3870 } 3871 3872 for (;;) { 3873 slab = keg_fetch_free_slab(keg, domain, rr, flags); 3874 if (slab != NULL) 3875 return (slab); 3876 3877 /* 3878 * M_NOVM means don't ask at all! 3879 */ 3880 if (flags & M_NOVM) 3881 break; 3882 3883 slab = keg_alloc_slab(keg, zone, domain, flags, aflags); 3884 if (slab != NULL) 3885 return (slab); 3886 if (!rr && (flags & M_WAITOK) == 0) 3887 break; 3888 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) { 3889 if ((flags & M_WAITOK) != 0) { 3890 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0); 3891 goto restart; 3892 } 3893 break; 3894 } 3895 } 3896 3897 /* 3898 * We might not have been able to get a slab but another cpu 3899 * could have while we were unlocked. Check again before we 3900 * fail. 3901 */ 3902 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL) 3903 return (slab); 3904 3905 return (NULL); 3906 } 3907 3908 static void * 3909 slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 3910 { 3911 uma_domain_t dom; 3912 void *item; 3913 int freei; 3914 3915 KEG_LOCK_ASSERT(keg, slab->us_domain); 3916 3917 dom = &keg->uk_domain[slab->us_domain]; 3918 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1; 3919 BIT_CLR(keg->uk_ipers, freei, &slab->us_free); 3920 item = slab_item(slab, keg, freei); 3921 slab->us_freecount--; 3922 dom->ud_free_items--; 3923 3924 /* 3925 * Move this slab to the full list. It must be on the partial list, so 3926 * we do not need to update the free slab count. In particular, 3927 * keg_fetch_slab() always returns slabs on the partial list. 3928 */ 3929 if (slab->us_freecount == 0) { 3930 LIST_REMOVE(slab, us_link); 3931 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link); 3932 } 3933 3934 return (item); 3935 } 3936 3937 static int 3938 zone_import(void *arg, void **bucket, int max, int domain, int flags) 3939 { 3940 uma_domain_t dom; 3941 uma_zone_t zone; 3942 uma_slab_t slab; 3943 uma_keg_t keg; 3944 #ifdef NUMA 3945 int stripe; 3946 #endif 3947 int i; 3948 3949 zone = arg; 3950 slab = NULL; 3951 keg = zone->uz_keg; 3952 /* Try to keep the buckets totally full */ 3953 for (i = 0; i < max; ) { 3954 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL) 3955 break; 3956 #ifdef NUMA 3957 stripe = howmany(max, vm_ndomains); 3958 #endif 3959 dom = &keg->uk_domain[slab->us_domain]; 3960 do { 3961 bucket[i++] = slab_alloc_item(keg, slab); 3962 if (dom->ud_free_items <= keg->uk_reserve) { 3963 /* 3964 * Avoid depleting the reserve after a 3965 * successful item allocation, even if 3966 * M_USE_RESERVE is specified. 3967 */ 3968 KEG_UNLOCK(keg, slab->us_domain); 3969 goto out; 3970 } 3971 #ifdef NUMA 3972 /* 3973 * If the zone is striped we pick a new slab for every 3974 * N allocations. Eliminating this conditional will 3975 * instead pick a new domain for each bucket rather 3976 * than stripe within each bucket. The current option 3977 * produces more fragmentation and requires more cpu 3978 * time but yields better distribution. 3979 */ 3980 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 && 3981 vm_ndomains > 1 && --stripe == 0) 3982 break; 3983 #endif 3984 } while (slab->us_freecount != 0 && i < max); 3985 KEG_UNLOCK(keg, slab->us_domain); 3986 3987 /* Don't block if we allocated any successfully. */ 3988 flags &= ~M_WAITOK; 3989 flags |= M_NOWAIT; 3990 } 3991 out: 3992 return i; 3993 } 3994 3995 static int 3996 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags) 3997 { 3998 uint64_t old, new, total, max; 3999 4000 /* 4001 * The hard case. We're going to sleep because there were existing 4002 * sleepers or because we ran out of items. This routine enforces 4003 * fairness by keeping fifo order. 4004 * 4005 * First release our ill gotten gains and make some noise. 4006 */ 4007 for (;;) { 4008 zone_free_limit(zone, count); 4009 zone_log_warning(zone); 4010 zone_maxaction(zone); 4011 if (flags & M_NOWAIT) 4012 return (0); 4013 4014 /* 4015 * We need to allocate an item or set ourself as a sleeper 4016 * while the sleepq lock is held to avoid wakeup races. This 4017 * is essentially a home rolled semaphore. 4018 */ 4019 sleepq_lock(&zone->uz_max_items); 4020 old = zone->uz_items; 4021 do { 4022 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX); 4023 /* Cache the max since we will evaluate twice. */ 4024 max = zone->uz_max_items; 4025 if (UZ_ITEMS_SLEEPERS(old) != 0 || 4026 UZ_ITEMS_COUNT(old) >= max) 4027 new = old + UZ_ITEMS_SLEEPER; 4028 else 4029 new = old + MIN(count, max - old); 4030 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0); 4031 4032 /* We may have successfully allocated under the sleepq lock. */ 4033 if (UZ_ITEMS_SLEEPERS(new) == 0) { 4034 sleepq_release(&zone->uz_max_items); 4035 return (new - old); 4036 } 4037 4038 /* 4039 * This is in a different cacheline from uz_items so that we 4040 * don't constantly invalidate the fastpath cacheline when we 4041 * adjust item counts. This could be limited to toggling on 4042 * transitions. 4043 */ 4044 atomic_add_32(&zone->uz_sleepers, 1); 4045 atomic_add_64(&zone->uz_sleeps, 1); 4046 4047 /* 4048 * We have added ourselves as a sleeper. The sleepq lock 4049 * protects us from wakeup races. Sleep now and then retry. 4050 */ 4051 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0); 4052 sleepq_wait(&zone->uz_max_items, PVM); 4053 4054 /* 4055 * After wakeup, remove ourselves as a sleeper and try 4056 * again. We no longer have the sleepq lock for protection. 4057 * 4058 * Subract ourselves as a sleeper while attempting to add 4059 * our count. 4060 */ 4061 atomic_subtract_32(&zone->uz_sleepers, 1); 4062 old = atomic_fetchadd_64(&zone->uz_items, 4063 -(UZ_ITEMS_SLEEPER - count)); 4064 /* We're no longer a sleeper. */ 4065 old -= UZ_ITEMS_SLEEPER; 4066 4067 /* 4068 * If we're still at the limit, restart. Notably do not 4069 * block on other sleepers. Cache the max value to protect 4070 * against changes via sysctl. 4071 */ 4072 total = UZ_ITEMS_COUNT(old); 4073 max = zone->uz_max_items; 4074 if (total >= max) 4075 continue; 4076 /* Truncate if necessary, otherwise wake other sleepers. */ 4077 if (total + count > max) { 4078 zone_free_limit(zone, total + count - max); 4079 count = max - total; 4080 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0) 4081 wakeup_one(&zone->uz_max_items); 4082 4083 return (count); 4084 } 4085 } 4086 4087 /* 4088 * Allocate 'count' items from our max_items limit. Returns the number 4089 * available. If M_NOWAIT is not specified it will sleep until at least 4090 * one item can be allocated. 4091 */ 4092 static int 4093 zone_alloc_limit(uma_zone_t zone, int count, int flags) 4094 { 4095 uint64_t old; 4096 uint64_t max; 4097 4098 max = zone->uz_max_items; 4099 MPASS(max > 0); 4100 4101 /* 4102 * We expect normal allocations to succeed with a simple 4103 * fetchadd. 4104 */ 4105 old = atomic_fetchadd_64(&zone->uz_items, count); 4106 if (__predict_true(old + count <= max)) 4107 return (count); 4108 4109 /* 4110 * If we had some items and no sleepers just return the 4111 * truncated value. We have to release the excess space 4112 * though because that may wake sleepers who weren't woken 4113 * because we were temporarily over the limit. 4114 */ 4115 if (old < max) { 4116 zone_free_limit(zone, (old + count) - max); 4117 return (max - old); 4118 } 4119 return (zone_alloc_limit_hard(zone, count, flags)); 4120 } 4121 4122 /* 4123 * Free a number of items back to the limit. 4124 */ 4125 static void 4126 zone_free_limit(uma_zone_t zone, int count) 4127 { 4128 uint64_t old; 4129 4130 MPASS(count > 0); 4131 4132 /* 4133 * In the common case we either have no sleepers or 4134 * are still over the limit and can just return. 4135 */ 4136 old = atomic_fetchadd_64(&zone->uz_items, -count); 4137 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 || 4138 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items)) 4139 return; 4140 4141 /* 4142 * Moderate the rate of wakeups. Sleepers will continue 4143 * to generate wakeups if necessary. 4144 */ 4145 wakeup_one(&zone->uz_max_items); 4146 } 4147 4148 static uma_bucket_t 4149 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags) 4150 { 4151 uma_bucket_t bucket; 4152 int error, maxbucket, cnt; 4153 4154 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name, 4155 zone, domain); 4156 4157 /* Avoid allocs targeting empty domains. */ 4158 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain)) 4159 domain = UMA_ANYDOMAIN; 4160 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0) 4161 domain = UMA_ANYDOMAIN; 4162 4163 if (zone->uz_max_items > 0) 4164 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size, 4165 M_NOWAIT); 4166 else 4167 maxbucket = zone->uz_bucket_size; 4168 if (maxbucket == 0) 4169 return (false); 4170 4171 /* Don't wait for buckets, preserve caller's NOVM setting. */ 4172 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 4173 if (bucket == NULL) { 4174 cnt = 0; 4175 goto out; 4176 } 4177 4178 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 4179 MIN(maxbucket, bucket->ub_entries), domain, flags); 4180 4181 /* 4182 * Initialize the memory if necessary. 4183 */ 4184 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 4185 int i; 4186 4187 for (i = 0; i < bucket->ub_cnt; i++) { 4188 kasan_mark_item_valid(zone, bucket->ub_bucket[i]); 4189 error = zone->uz_init(bucket->ub_bucket[i], 4190 zone->uz_size, flags); 4191 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]); 4192 if (error != 0) 4193 break; 4194 } 4195 4196 /* 4197 * If we couldn't initialize the whole bucket, put the 4198 * rest back onto the freelist. 4199 */ 4200 if (i != bucket->ub_cnt) { 4201 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 4202 bucket->ub_cnt - i); 4203 #ifdef INVARIANTS 4204 bzero(&bucket->ub_bucket[i], 4205 sizeof(void *) * (bucket->ub_cnt - i)); 4206 #endif 4207 bucket->ub_cnt = i; 4208 } 4209 } 4210 4211 cnt = bucket->ub_cnt; 4212 if (bucket->ub_cnt == 0) { 4213 bucket_free(zone, bucket, udata); 4214 counter_u64_add(zone->uz_fails, 1); 4215 bucket = NULL; 4216 } 4217 out: 4218 if (zone->uz_max_items > 0 && cnt < maxbucket) 4219 zone_free_limit(zone, maxbucket - cnt); 4220 4221 return (bucket); 4222 } 4223 4224 /* 4225 * Allocates a single item from a zone. 4226 * 4227 * Arguments 4228 * zone The zone to alloc for. 4229 * udata The data to be passed to the constructor. 4230 * domain The domain to allocate from or UMA_ANYDOMAIN. 4231 * flags M_WAITOK, M_NOWAIT, M_ZERO. 4232 * 4233 * Returns 4234 * NULL if there is no memory and M_NOWAIT is set 4235 * An item if successful 4236 */ 4237 4238 static void * 4239 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags) 4240 { 4241 void *item; 4242 4243 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) { 4244 counter_u64_add(zone->uz_fails, 1); 4245 return (NULL); 4246 } 4247 4248 /* Avoid allocs targeting empty domains. */ 4249 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain)) 4250 domain = UMA_ANYDOMAIN; 4251 4252 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1) 4253 goto fail_cnt; 4254 4255 /* 4256 * We have to call both the zone's init (not the keg's init) 4257 * and the zone's ctor. This is because the item is going from 4258 * a keg slab directly to the user, and the user is expecting it 4259 * to be both zone-init'd as well as zone-ctor'd. 4260 */ 4261 if (zone->uz_init != NULL) { 4262 int error; 4263 4264 kasan_mark_item_valid(zone, item); 4265 error = zone->uz_init(item, zone->uz_size, flags); 4266 kasan_mark_item_invalid(zone, item); 4267 if (error != 0) { 4268 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT); 4269 goto fail_cnt; 4270 } 4271 } 4272 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags, 4273 item); 4274 if (item == NULL) 4275 goto fail; 4276 4277 counter_u64_add(zone->uz_allocs, 1); 4278 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item, 4279 zone->uz_name, zone); 4280 4281 return (item); 4282 4283 fail_cnt: 4284 counter_u64_add(zone->uz_fails, 1); 4285 fail: 4286 if (zone->uz_max_items > 0) 4287 zone_free_limit(zone, 1); 4288 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)", 4289 zone->uz_name, zone); 4290 4291 return (NULL); 4292 } 4293 4294 /* See uma.h */ 4295 void 4296 uma_zfree_smr(uma_zone_t zone, void *item) 4297 { 4298 uma_cache_t cache; 4299 uma_cache_bucket_t bucket; 4300 int itemdomain, uz_flags; 4301 4302 #ifdef UMA_ZALLOC_DEBUG 4303 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0, 4304 ("uma_zfree_smr: called with non-SMR zone.")); 4305 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer.")); 4306 SMR_ASSERT_NOT_ENTERED(zone->uz_smr); 4307 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN) 4308 return; 4309 #endif 4310 cache = &zone->uz_cpu[curcpu]; 4311 uz_flags = cache_uz_flags(cache); 4312 itemdomain = 0; 4313 #ifdef NUMA 4314 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 4315 itemdomain = item_domain(item); 4316 #endif 4317 critical_enter(); 4318 do { 4319 cache = &zone->uz_cpu[curcpu]; 4320 /* SMR Zones must free to the free bucket. */ 4321 bucket = &cache->uc_freebucket; 4322 #ifdef NUMA 4323 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && 4324 PCPU_GET(domain) != itemdomain) { 4325 bucket = &cache->uc_crossbucket; 4326 } 4327 #endif 4328 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) { 4329 cache_bucket_push(cache, bucket, item); 4330 critical_exit(); 4331 return; 4332 } 4333 } while (cache_free(zone, cache, NULL, item, itemdomain)); 4334 critical_exit(); 4335 4336 /* 4337 * If nothing else caught this, we'll just do an internal free. 4338 */ 4339 zone_free_item(zone, item, NULL, SKIP_NONE); 4340 } 4341 4342 /* See uma.h */ 4343 void 4344 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 4345 { 4346 uma_cache_t cache; 4347 uma_cache_bucket_t bucket; 4348 int itemdomain, uz_flags; 4349 4350 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 4351 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 4352 4353 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone); 4354 4355 #ifdef UMA_ZALLOC_DEBUG 4356 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0, 4357 ("uma_zfree_arg: called with SMR zone.")); 4358 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN) 4359 return; 4360 #endif 4361 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 4362 if (item == NULL) 4363 return; 4364 4365 /* 4366 * We are accessing the per-cpu cache without a critical section to 4367 * fetch size and flags. This is acceptable, if we are preempted we 4368 * will simply read another cpu's line. 4369 */ 4370 cache = &zone->uz_cpu[curcpu]; 4371 uz_flags = cache_uz_flags(cache); 4372 if (UMA_ALWAYS_CTORDTOR || 4373 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0)) 4374 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE); 4375 4376 /* 4377 * The race here is acceptable. If we miss it we'll just have to wait 4378 * a little longer for the limits to be reset. 4379 */ 4380 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) { 4381 if (atomic_load_32(&zone->uz_sleepers) > 0) 4382 goto zfree_item; 4383 } 4384 4385 /* 4386 * If possible, free to the per-CPU cache. There are two 4387 * requirements for safe access to the per-CPU cache: (1) the thread 4388 * accessing the cache must not be preempted or yield during access, 4389 * and (2) the thread must not migrate CPUs without switching which 4390 * cache it accesses. We rely on a critical section to prevent 4391 * preemption and migration. We release the critical section in 4392 * order to acquire the zone mutex if we are unable to free to the 4393 * current cache; when we re-acquire the critical section, we must 4394 * detect and handle migration if it has occurred. 4395 */ 4396 itemdomain = 0; 4397 #ifdef NUMA 4398 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 4399 itemdomain = item_domain(item); 4400 #endif 4401 critical_enter(); 4402 do { 4403 cache = &zone->uz_cpu[curcpu]; 4404 /* 4405 * Try to free into the allocbucket first to give LIFO 4406 * ordering for cache-hot datastructures. Spill over 4407 * into the freebucket if necessary. Alloc will swap 4408 * them if one runs dry. 4409 */ 4410 bucket = &cache->uc_allocbucket; 4411 #ifdef NUMA 4412 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && 4413 PCPU_GET(domain) != itemdomain) { 4414 bucket = &cache->uc_crossbucket; 4415 } else 4416 #endif 4417 if (bucket->ucb_cnt == bucket->ucb_entries && 4418 cache->uc_freebucket.ucb_cnt < 4419 cache->uc_freebucket.ucb_entries) 4420 cache_bucket_swap(&cache->uc_freebucket, 4421 &cache->uc_allocbucket); 4422 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) { 4423 cache_bucket_push(cache, bucket, item); 4424 critical_exit(); 4425 return; 4426 } 4427 } while (cache_free(zone, cache, udata, item, itemdomain)); 4428 critical_exit(); 4429 4430 /* 4431 * If nothing else caught this, we'll just do an internal free. 4432 */ 4433 zfree_item: 4434 zone_free_item(zone, item, udata, SKIP_DTOR); 4435 } 4436 4437 #ifdef NUMA 4438 /* 4439 * sort crossdomain free buckets to domain correct buckets and cache 4440 * them. 4441 */ 4442 static void 4443 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata) 4444 { 4445 struct uma_bucketlist emptybuckets, fullbuckets; 4446 uma_zone_domain_t zdom; 4447 uma_bucket_t b; 4448 smr_seq_t seq; 4449 void *item; 4450 int domain; 4451 4452 CTR3(KTR_UMA, 4453 "uma_zfree: zone %s(%p) draining cross bucket %p", 4454 zone->uz_name, zone, bucket); 4455 4456 /* 4457 * It is possible for buckets to arrive here out of order so we fetch 4458 * the current smr seq rather than accepting the bucket's. 4459 */ 4460 seq = SMR_SEQ_INVALID; 4461 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 4462 seq = smr_advance(zone->uz_smr); 4463 4464 /* 4465 * To avoid having ndomain * ndomain buckets for sorting we have a 4466 * lock on the current crossfree bucket. A full matrix with 4467 * per-domain locking could be used if necessary. 4468 */ 4469 STAILQ_INIT(&emptybuckets); 4470 STAILQ_INIT(&fullbuckets); 4471 ZONE_CROSS_LOCK(zone); 4472 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) { 4473 item = bucket->ub_bucket[bucket->ub_cnt - 1]; 4474 domain = item_domain(item); 4475 zdom = ZDOM_GET(zone, domain); 4476 if (zdom->uzd_cross == NULL) { 4477 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) { 4478 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link); 4479 zdom->uzd_cross = b; 4480 } else { 4481 /* 4482 * Avoid allocating a bucket with the cross lock 4483 * held, since allocation can trigger a 4484 * cross-domain free and bucket zones may 4485 * allocate from each other. 4486 */ 4487 ZONE_CROSS_UNLOCK(zone); 4488 b = bucket_alloc(zone, udata, M_NOWAIT); 4489 if (b == NULL) 4490 goto out; 4491 ZONE_CROSS_LOCK(zone); 4492 if (zdom->uzd_cross != NULL) { 4493 STAILQ_INSERT_HEAD(&emptybuckets, b, 4494 ub_link); 4495 } else { 4496 zdom->uzd_cross = b; 4497 } 4498 } 4499 } 4500 b = zdom->uzd_cross; 4501 b->ub_bucket[b->ub_cnt++] = item; 4502 b->ub_seq = seq; 4503 if (b->ub_cnt == b->ub_entries) { 4504 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link); 4505 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) 4506 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link); 4507 zdom->uzd_cross = b; 4508 } 4509 } 4510 ZONE_CROSS_UNLOCK(zone); 4511 out: 4512 if (bucket->ub_cnt == 0) 4513 bucket->ub_seq = SMR_SEQ_INVALID; 4514 bucket_free(zone, bucket, udata); 4515 4516 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) { 4517 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link); 4518 bucket_free(zone, b, udata); 4519 } 4520 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) { 4521 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link); 4522 domain = item_domain(b->ub_bucket[0]); 4523 zone_put_bucket(zone, domain, b, udata, true); 4524 } 4525 } 4526 #endif 4527 4528 static void 4529 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata, 4530 int itemdomain, bool ws) 4531 { 4532 4533 #ifdef NUMA 4534 /* 4535 * Buckets coming from the wrong domain will be entirely for the 4536 * only other domain on two domain systems. In this case we can 4537 * simply cache them. Otherwise we need to sort them back to 4538 * correct domains. 4539 */ 4540 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && 4541 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) { 4542 zone_free_cross(zone, bucket, udata); 4543 return; 4544 } 4545 #endif 4546 4547 /* 4548 * Attempt to save the bucket in the zone's domain bucket cache. 4549 */ 4550 CTR3(KTR_UMA, 4551 "uma_zfree: zone %s(%p) putting bucket %p on free list", 4552 zone->uz_name, zone, bucket); 4553 /* ub_cnt is pointing to the last free item */ 4554 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0) 4555 itemdomain = zone_domain_lowest(zone, itemdomain); 4556 zone_put_bucket(zone, itemdomain, bucket, udata, ws); 4557 } 4558 4559 /* 4560 * Populate a free or cross bucket for the current cpu cache. Free any 4561 * existing full bucket either to the zone cache or back to the slab layer. 4562 * 4563 * Enters and returns in a critical section. false return indicates that 4564 * we can not satisfy this free in the cache layer. true indicates that 4565 * the caller should retry. 4566 */ 4567 static __noinline bool 4568 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item, 4569 int itemdomain) 4570 { 4571 uma_cache_bucket_t cbucket; 4572 uma_bucket_t newbucket, bucket; 4573 4574 CRITICAL_ASSERT(curthread); 4575 4576 if (zone->uz_bucket_size == 0) 4577 return false; 4578 4579 cache = &zone->uz_cpu[curcpu]; 4580 newbucket = NULL; 4581 4582 /* 4583 * FIRSTTOUCH domains need to free to the correct zdom. When 4584 * enabled this is the zdom of the item. The bucket is the 4585 * cross bucket if the current domain and itemdomain do not match. 4586 */ 4587 cbucket = &cache->uc_freebucket; 4588 #ifdef NUMA 4589 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) { 4590 if (PCPU_GET(domain) != itemdomain) { 4591 cbucket = &cache->uc_crossbucket; 4592 if (cbucket->ucb_cnt != 0) 4593 counter_u64_add(zone->uz_xdomain, 4594 cbucket->ucb_cnt); 4595 } 4596 } 4597 #endif 4598 bucket = cache_bucket_unload(cbucket); 4599 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries, 4600 ("cache_free: Entered with non-full free bucket.")); 4601 4602 /* We are no longer associated with this CPU. */ 4603 critical_exit(); 4604 4605 /* 4606 * Don't let SMR zones operate without a free bucket. Force 4607 * a synchronize and re-use this one. We will only degrade 4608 * to a synchronize every bucket_size items rather than every 4609 * item if we fail to allocate a bucket. 4610 */ 4611 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) { 4612 if (bucket != NULL) 4613 bucket->ub_seq = smr_advance(zone->uz_smr); 4614 newbucket = bucket_alloc(zone, udata, M_NOWAIT); 4615 if (newbucket == NULL && bucket != NULL) { 4616 bucket_drain(zone, bucket); 4617 newbucket = bucket; 4618 bucket = NULL; 4619 } 4620 } else if (!bucketdisable) 4621 newbucket = bucket_alloc(zone, udata, M_NOWAIT); 4622 4623 if (bucket != NULL) 4624 zone_free_bucket(zone, bucket, udata, itemdomain, true); 4625 4626 critical_enter(); 4627 if ((bucket = newbucket) == NULL) 4628 return (false); 4629 cache = &zone->uz_cpu[curcpu]; 4630 #ifdef NUMA 4631 /* 4632 * Check to see if we should be populating the cross bucket. If it 4633 * is already populated we will fall through and attempt to populate 4634 * the free bucket. 4635 */ 4636 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) { 4637 if (PCPU_GET(domain) != itemdomain && 4638 cache->uc_crossbucket.ucb_bucket == NULL) { 4639 cache_bucket_load_cross(cache, bucket); 4640 return (true); 4641 } 4642 } 4643 #endif 4644 /* 4645 * We may have lost the race to fill the bucket or switched CPUs. 4646 */ 4647 if (cache->uc_freebucket.ucb_bucket != NULL) { 4648 critical_exit(); 4649 bucket_free(zone, bucket, udata); 4650 critical_enter(); 4651 } else 4652 cache_bucket_load_free(cache, bucket); 4653 4654 return (true); 4655 } 4656 4657 static void 4658 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item) 4659 { 4660 uma_keg_t keg; 4661 uma_domain_t dom; 4662 int freei; 4663 4664 keg = zone->uz_keg; 4665 KEG_LOCK_ASSERT(keg, slab->us_domain); 4666 4667 /* Do we need to remove from any lists? */ 4668 dom = &keg->uk_domain[slab->us_domain]; 4669 if (slab->us_freecount + 1 == keg->uk_ipers) { 4670 LIST_REMOVE(slab, us_link); 4671 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); 4672 dom->ud_free_slabs++; 4673 } else if (slab->us_freecount == 0) { 4674 LIST_REMOVE(slab, us_link); 4675 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 4676 } 4677 4678 /* Slab management. */ 4679 freei = slab_item_index(slab, keg, item); 4680 BIT_SET(keg->uk_ipers, freei, &slab->us_free); 4681 slab->us_freecount++; 4682 4683 /* Keg statistics. */ 4684 dom->ud_free_items++; 4685 } 4686 4687 static void 4688 zone_release(void *arg, void **bucket, int cnt) 4689 { 4690 struct mtx *lock; 4691 uma_zone_t zone; 4692 uma_slab_t slab; 4693 uma_keg_t keg; 4694 uint8_t *mem; 4695 void *item; 4696 int i; 4697 4698 zone = arg; 4699 keg = zone->uz_keg; 4700 lock = NULL; 4701 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0)) 4702 lock = KEG_LOCK(keg, 0); 4703 for (i = 0; i < cnt; i++) { 4704 item = bucket[i]; 4705 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) { 4706 slab = vtoslab((vm_offset_t)item); 4707 } else { 4708 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 4709 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0) 4710 slab = hash_sfind(&keg->uk_hash, mem); 4711 else 4712 slab = (uma_slab_t)(mem + keg->uk_pgoff); 4713 } 4714 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) { 4715 if (lock != NULL) 4716 mtx_unlock(lock); 4717 lock = KEG_LOCK(keg, slab->us_domain); 4718 } 4719 slab_free_item(zone, slab, item); 4720 } 4721 if (lock != NULL) 4722 mtx_unlock(lock); 4723 } 4724 4725 /* 4726 * Frees a single item to any zone. 4727 * 4728 * Arguments: 4729 * zone The zone to free to 4730 * item The item we're freeing 4731 * udata User supplied data for the dtor 4732 * skip Skip dtors and finis 4733 */ 4734 static __noinline void 4735 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 4736 { 4737 4738 /* 4739 * If a free is sent directly to an SMR zone we have to 4740 * synchronize immediately because the item can instantly 4741 * be reallocated. This should only happen in degenerate 4742 * cases when no memory is available for per-cpu caches. 4743 */ 4744 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE) 4745 smr_synchronize(zone->uz_smr); 4746 4747 item_dtor(zone, item, zone->uz_size, udata, skip); 4748 4749 if (skip < SKIP_FINI && zone->uz_fini) { 4750 kasan_mark_item_valid(zone, item); 4751 zone->uz_fini(item, zone->uz_size); 4752 kasan_mark_item_invalid(zone, item); 4753 } 4754 4755 zone->uz_release(zone->uz_arg, &item, 1); 4756 4757 if (skip & SKIP_CNT) 4758 return; 4759 4760 counter_u64_add(zone->uz_frees, 1); 4761 4762 if (zone->uz_max_items > 0) 4763 zone_free_limit(zone, 1); 4764 } 4765 4766 /* See uma.h */ 4767 int 4768 uma_zone_set_max(uma_zone_t zone, int nitems) 4769 { 4770 4771 /* 4772 * If the limit is small, we may need to constrain the maximum per-CPU 4773 * cache size, or disable caching entirely. 4774 */ 4775 uma_zone_set_maxcache(zone, nitems); 4776 4777 /* 4778 * XXX This can misbehave if the zone has any allocations with 4779 * no limit and a limit is imposed. There is currently no 4780 * way to clear a limit. 4781 */ 4782 ZONE_LOCK(zone); 4783 zone->uz_max_items = nitems; 4784 zone->uz_flags |= UMA_ZFLAG_LIMIT; 4785 zone_update_caches(zone); 4786 /* We may need to wake waiters. */ 4787 wakeup(&zone->uz_max_items); 4788 ZONE_UNLOCK(zone); 4789 4790 return (nitems); 4791 } 4792 4793 /* See uma.h */ 4794 void 4795 uma_zone_set_maxcache(uma_zone_t zone, int nitems) 4796 { 4797 int bpcpu, bpdom, bsize, nb; 4798 4799 ZONE_LOCK(zone); 4800 4801 /* 4802 * Compute a lower bound on the number of items that may be cached in 4803 * the zone. Each CPU gets at least two buckets, and for cross-domain 4804 * frees we use an additional bucket per CPU and per domain. Select the 4805 * largest bucket size that does not exceed half of the requested limit, 4806 * with the left over space given to the full bucket cache. 4807 */ 4808 bpdom = 0; 4809 bpcpu = 2; 4810 #ifdef NUMA 4811 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) { 4812 bpcpu++; 4813 bpdom++; 4814 } 4815 #endif 4816 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains; 4817 bsize = nitems / nb / 2; 4818 if (bsize > BUCKET_MAX) 4819 bsize = BUCKET_MAX; 4820 else if (bsize == 0 && nitems / nb > 0) 4821 bsize = 1; 4822 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize; 4823 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max) 4824 zone->uz_bucket_size_min = zone->uz_bucket_size_max; 4825 zone->uz_bucket_max = nitems - nb * bsize; 4826 ZONE_UNLOCK(zone); 4827 } 4828 4829 /* See uma.h */ 4830 int 4831 uma_zone_get_max(uma_zone_t zone) 4832 { 4833 int nitems; 4834 4835 nitems = atomic_load_64(&zone->uz_max_items); 4836 4837 return (nitems); 4838 } 4839 4840 /* See uma.h */ 4841 void 4842 uma_zone_set_warning(uma_zone_t zone, const char *warning) 4843 { 4844 4845 ZONE_ASSERT_COLD(zone); 4846 zone->uz_warning = warning; 4847 } 4848 4849 /* See uma.h */ 4850 void 4851 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) 4852 { 4853 4854 ZONE_ASSERT_COLD(zone); 4855 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); 4856 } 4857 4858 /* See uma.h */ 4859 int 4860 uma_zone_get_cur(uma_zone_t zone) 4861 { 4862 int64_t nitems; 4863 u_int i; 4864 4865 nitems = 0; 4866 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER) 4867 nitems = counter_u64_fetch(zone->uz_allocs) - 4868 counter_u64_fetch(zone->uz_frees); 4869 CPU_FOREACH(i) 4870 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) - 4871 atomic_load_64(&zone->uz_cpu[i].uc_frees); 4872 4873 return (nitems < 0 ? 0 : nitems); 4874 } 4875 4876 static uint64_t 4877 uma_zone_get_allocs(uma_zone_t zone) 4878 { 4879 uint64_t nitems; 4880 u_int i; 4881 4882 nitems = 0; 4883 if (zone->uz_allocs != EARLY_COUNTER) 4884 nitems = counter_u64_fetch(zone->uz_allocs); 4885 CPU_FOREACH(i) 4886 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs); 4887 4888 return (nitems); 4889 } 4890 4891 static uint64_t 4892 uma_zone_get_frees(uma_zone_t zone) 4893 { 4894 uint64_t nitems; 4895 u_int i; 4896 4897 nitems = 0; 4898 if (zone->uz_frees != EARLY_COUNTER) 4899 nitems = counter_u64_fetch(zone->uz_frees); 4900 CPU_FOREACH(i) 4901 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees); 4902 4903 return (nitems); 4904 } 4905 4906 #ifdef INVARIANTS 4907 /* Used only for KEG_ASSERT_COLD(). */ 4908 static uint64_t 4909 uma_keg_get_allocs(uma_keg_t keg) 4910 { 4911 uma_zone_t z; 4912 uint64_t nitems; 4913 4914 nitems = 0; 4915 LIST_FOREACH(z, &keg->uk_zones, uz_link) 4916 nitems += uma_zone_get_allocs(z); 4917 4918 return (nitems); 4919 } 4920 #endif 4921 4922 /* See uma.h */ 4923 void 4924 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 4925 { 4926 uma_keg_t keg; 4927 4928 KEG_GET(zone, keg); 4929 KEG_ASSERT_COLD(keg); 4930 keg->uk_init = uminit; 4931 } 4932 4933 /* See uma.h */ 4934 void 4935 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 4936 { 4937 uma_keg_t keg; 4938 4939 KEG_GET(zone, keg); 4940 KEG_ASSERT_COLD(keg); 4941 keg->uk_fini = fini; 4942 } 4943 4944 /* See uma.h */ 4945 void 4946 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 4947 { 4948 4949 ZONE_ASSERT_COLD(zone); 4950 zone->uz_init = zinit; 4951 } 4952 4953 /* See uma.h */ 4954 void 4955 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 4956 { 4957 4958 ZONE_ASSERT_COLD(zone); 4959 zone->uz_fini = zfini; 4960 } 4961 4962 /* See uma.h */ 4963 void 4964 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 4965 { 4966 uma_keg_t keg; 4967 4968 KEG_GET(zone, keg); 4969 KEG_ASSERT_COLD(keg); 4970 keg->uk_freef = freef; 4971 } 4972 4973 /* See uma.h */ 4974 void 4975 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 4976 { 4977 uma_keg_t keg; 4978 4979 KEG_GET(zone, keg); 4980 KEG_ASSERT_COLD(keg); 4981 keg->uk_allocf = allocf; 4982 } 4983 4984 /* See uma.h */ 4985 void 4986 uma_zone_set_smr(uma_zone_t zone, smr_t smr) 4987 { 4988 4989 ZONE_ASSERT_COLD(zone); 4990 4991 KASSERT(smr != NULL, ("Got NULL smr")); 4992 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0, 4993 ("zone %p (%s) already uses SMR", zone, zone->uz_name)); 4994 zone->uz_flags |= UMA_ZONE_SMR; 4995 zone->uz_smr = smr; 4996 zone_update_caches(zone); 4997 } 4998 4999 smr_t 5000 uma_zone_get_smr(uma_zone_t zone) 5001 { 5002 5003 return (zone->uz_smr); 5004 } 5005 5006 /* See uma.h */ 5007 void 5008 uma_zone_reserve(uma_zone_t zone, int items) 5009 { 5010 uma_keg_t keg; 5011 5012 KEG_GET(zone, keg); 5013 KEG_ASSERT_COLD(keg); 5014 keg->uk_reserve = items; 5015 } 5016 5017 /* See uma.h */ 5018 int 5019 uma_zone_reserve_kva(uma_zone_t zone, int count) 5020 { 5021 uma_keg_t keg; 5022 vm_offset_t kva; 5023 u_int pages; 5024 5025 KEG_GET(zone, keg); 5026 KEG_ASSERT_COLD(keg); 5027 ZONE_ASSERT_COLD(zone); 5028 5029 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera; 5030 5031 #ifdef UMA_MD_SMALL_ALLOC 5032 if (keg->uk_ppera > 1) { 5033 #else 5034 if (1) { 5035 #endif 5036 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE); 5037 if (kva == 0) 5038 return (0); 5039 } else 5040 kva = 0; 5041 5042 MPASS(keg->uk_kva == 0); 5043 keg->uk_kva = kva; 5044 keg->uk_offset = 0; 5045 zone->uz_max_items = pages * keg->uk_ipers; 5046 #ifdef UMA_MD_SMALL_ALLOC 5047 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 5048 #else 5049 keg->uk_allocf = noobj_alloc; 5050 #endif 5051 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE; 5052 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE; 5053 zone_update_caches(zone); 5054 5055 return (1); 5056 } 5057 5058 /* See uma.h */ 5059 void 5060 uma_prealloc(uma_zone_t zone, int items) 5061 { 5062 struct vm_domainset_iter di; 5063 uma_domain_t dom; 5064 uma_slab_t slab; 5065 uma_keg_t keg; 5066 int aflags, domain, slabs; 5067 5068 KEG_GET(zone, keg); 5069 slabs = howmany(items, keg->uk_ipers); 5070 while (slabs-- > 0) { 5071 aflags = M_NOWAIT; 5072 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 5073 &aflags); 5074 for (;;) { 5075 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK, 5076 aflags); 5077 if (slab != NULL) { 5078 dom = &keg->uk_domain[slab->us_domain]; 5079 /* 5080 * keg_alloc_slab() always returns a slab on the 5081 * partial list. 5082 */ 5083 LIST_REMOVE(slab, us_link); 5084 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, 5085 us_link); 5086 dom->ud_free_slabs++; 5087 KEG_UNLOCK(keg, slab->us_domain); 5088 break; 5089 } 5090 if (vm_domainset_iter_policy(&di, &domain) != 0) 5091 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0); 5092 } 5093 } 5094 } 5095 5096 /* 5097 * Returns a snapshot of memory consumption in bytes. 5098 */ 5099 size_t 5100 uma_zone_memory(uma_zone_t zone) 5101 { 5102 size_t sz; 5103 int i; 5104 5105 sz = 0; 5106 if (zone->uz_flags & UMA_ZFLAG_CACHE) { 5107 for (i = 0; i < vm_ndomains; i++) 5108 sz += ZDOM_GET(zone, i)->uzd_nitems; 5109 return (sz * zone->uz_size); 5110 } 5111 for (i = 0; i < vm_ndomains; i++) 5112 sz += zone->uz_keg->uk_domain[i].ud_pages; 5113 5114 return (sz * PAGE_SIZE); 5115 } 5116 5117 /* See uma.h */ 5118 void 5119 uma_reclaim(int req) 5120 { 5121 uma_reclaim_domain(req, UMA_ANYDOMAIN); 5122 } 5123 5124 void 5125 uma_reclaim_domain(int req, int domain) 5126 { 5127 void *arg; 5128 5129 bucket_enable(); 5130 5131 arg = (void *)(uintptr_t)domain; 5132 sx_slock(&uma_reclaim_lock); 5133 switch (req) { 5134 case UMA_RECLAIM_TRIM: 5135 zone_foreach(zone_trim, arg); 5136 break; 5137 case UMA_RECLAIM_DRAIN: 5138 zone_foreach(zone_drain, arg); 5139 break; 5140 case UMA_RECLAIM_DRAIN_CPU: 5141 zone_foreach(zone_drain, arg); 5142 pcpu_cache_drain_safe(NULL); 5143 zone_foreach(zone_drain, arg); 5144 break; 5145 default: 5146 panic("unhandled reclamation request %d", req); 5147 } 5148 5149 /* 5150 * Some slabs may have been freed but this zone will be visited early 5151 * we visit again so that we can free pages that are empty once other 5152 * zones are drained. We have to do the same for buckets. 5153 */ 5154 zone_drain(slabzones[0], arg); 5155 zone_drain(slabzones[1], arg); 5156 bucket_zone_drain(domain); 5157 sx_sunlock(&uma_reclaim_lock); 5158 } 5159 5160 static volatile int uma_reclaim_needed; 5161 5162 void 5163 uma_reclaim_wakeup(void) 5164 { 5165 5166 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0) 5167 wakeup(uma_reclaim); 5168 } 5169 5170 void 5171 uma_reclaim_worker(void *arg __unused) 5172 { 5173 5174 for (;;) { 5175 sx_xlock(&uma_reclaim_lock); 5176 while (atomic_load_int(&uma_reclaim_needed) == 0) 5177 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl", 5178 hz); 5179 sx_xunlock(&uma_reclaim_lock); 5180 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); 5181 uma_reclaim(UMA_RECLAIM_DRAIN_CPU); 5182 atomic_store_int(&uma_reclaim_needed, 0); 5183 /* Don't fire more than once per-second. */ 5184 pause("umarclslp", hz); 5185 } 5186 } 5187 5188 /* See uma.h */ 5189 void 5190 uma_zone_reclaim(uma_zone_t zone, int req) 5191 { 5192 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN); 5193 } 5194 5195 void 5196 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain) 5197 { 5198 void *arg; 5199 5200 arg = (void *)(uintptr_t)domain; 5201 switch (req) { 5202 case UMA_RECLAIM_TRIM: 5203 zone_trim(zone, arg); 5204 break; 5205 case UMA_RECLAIM_DRAIN: 5206 zone_drain(zone, arg); 5207 break; 5208 case UMA_RECLAIM_DRAIN_CPU: 5209 pcpu_cache_drain_safe(zone); 5210 zone_drain(zone, arg); 5211 break; 5212 default: 5213 panic("unhandled reclamation request %d", req); 5214 } 5215 } 5216 5217 /* See uma.h */ 5218 int 5219 uma_zone_exhausted(uma_zone_t zone) 5220 { 5221 5222 return (atomic_load_32(&zone->uz_sleepers) > 0); 5223 } 5224 5225 unsigned long 5226 uma_limit(void) 5227 { 5228 5229 return (uma_kmem_limit); 5230 } 5231 5232 void 5233 uma_set_limit(unsigned long limit) 5234 { 5235 5236 uma_kmem_limit = limit; 5237 } 5238 5239 unsigned long 5240 uma_size(void) 5241 { 5242 5243 return (atomic_load_long(&uma_kmem_total)); 5244 } 5245 5246 long 5247 uma_avail(void) 5248 { 5249 5250 return (uma_kmem_limit - uma_size()); 5251 } 5252 5253 #ifdef DDB 5254 /* 5255 * Generate statistics across both the zone and its per-cpu cache's. Return 5256 * desired statistics if the pointer is non-NULL for that statistic. 5257 * 5258 * Note: does not update the zone statistics, as it can't safely clear the 5259 * per-CPU cache statistic. 5260 * 5261 */ 5262 static void 5263 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp, 5264 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp) 5265 { 5266 uma_cache_t cache; 5267 uint64_t allocs, frees, sleeps, xdomain; 5268 int cachefree, cpu; 5269 5270 allocs = frees = sleeps = xdomain = 0; 5271 cachefree = 0; 5272 CPU_FOREACH(cpu) { 5273 cache = &z->uz_cpu[cpu]; 5274 cachefree += cache->uc_allocbucket.ucb_cnt; 5275 cachefree += cache->uc_freebucket.ucb_cnt; 5276 xdomain += cache->uc_crossbucket.ucb_cnt; 5277 cachefree += cache->uc_crossbucket.ucb_cnt; 5278 allocs += cache->uc_allocs; 5279 frees += cache->uc_frees; 5280 } 5281 allocs += counter_u64_fetch(z->uz_allocs); 5282 frees += counter_u64_fetch(z->uz_frees); 5283 xdomain += counter_u64_fetch(z->uz_xdomain); 5284 sleeps += z->uz_sleeps; 5285 if (cachefreep != NULL) 5286 *cachefreep = cachefree; 5287 if (allocsp != NULL) 5288 *allocsp = allocs; 5289 if (freesp != NULL) 5290 *freesp = frees; 5291 if (sleepsp != NULL) 5292 *sleepsp = sleeps; 5293 if (xdomainp != NULL) 5294 *xdomainp = xdomain; 5295 } 5296 #endif /* DDB */ 5297 5298 static int 5299 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 5300 { 5301 uma_keg_t kz; 5302 uma_zone_t z; 5303 int count; 5304 5305 count = 0; 5306 rw_rlock(&uma_rwlock); 5307 LIST_FOREACH(kz, &uma_kegs, uk_link) { 5308 LIST_FOREACH(z, &kz->uk_zones, uz_link) 5309 count++; 5310 } 5311 LIST_FOREACH(z, &uma_cachezones, uz_link) 5312 count++; 5313 5314 rw_runlock(&uma_rwlock); 5315 return (sysctl_handle_int(oidp, &count, 0, req)); 5316 } 5317 5318 static void 5319 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf, 5320 struct uma_percpu_stat *ups, bool internal) 5321 { 5322 uma_zone_domain_t zdom; 5323 uma_cache_t cache; 5324 int i; 5325 5326 for (i = 0; i < vm_ndomains; i++) { 5327 zdom = ZDOM_GET(z, i); 5328 uth->uth_zone_free += zdom->uzd_nitems; 5329 } 5330 uth->uth_allocs = counter_u64_fetch(z->uz_allocs); 5331 uth->uth_frees = counter_u64_fetch(z->uz_frees); 5332 uth->uth_fails = counter_u64_fetch(z->uz_fails); 5333 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain); 5334 uth->uth_sleeps = z->uz_sleeps; 5335 5336 for (i = 0; i < mp_maxid + 1; i++) { 5337 bzero(&ups[i], sizeof(*ups)); 5338 if (internal || CPU_ABSENT(i)) 5339 continue; 5340 cache = &z->uz_cpu[i]; 5341 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt; 5342 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt; 5343 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt; 5344 ups[i].ups_allocs = cache->uc_allocs; 5345 ups[i].ups_frees = cache->uc_frees; 5346 } 5347 } 5348 5349 static int 5350 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 5351 { 5352 struct uma_stream_header ush; 5353 struct uma_type_header uth; 5354 struct uma_percpu_stat *ups; 5355 struct sbuf sbuf; 5356 uma_keg_t kz; 5357 uma_zone_t z; 5358 uint64_t items; 5359 uint32_t kfree, pages; 5360 int count, error, i; 5361 5362 error = sysctl_wire_old_buffer(req, 0); 5363 if (error != 0) 5364 return (error); 5365 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 5366 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); 5367 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK); 5368 5369 count = 0; 5370 rw_rlock(&uma_rwlock); 5371 LIST_FOREACH(kz, &uma_kegs, uk_link) { 5372 LIST_FOREACH(z, &kz->uk_zones, uz_link) 5373 count++; 5374 } 5375 5376 LIST_FOREACH(z, &uma_cachezones, uz_link) 5377 count++; 5378 5379 /* 5380 * Insert stream header. 5381 */ 5382 bzero(&ush, sizeof(ush)); 5383 ush.ush_version = UMA_STREAM_VERSION; 5384 ush.ush_maxcpus = (mp_maxid + 1); 5385 ush.ush_count = count; 5386 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 5387 5388 LIST_FOREACH(kz, &uma_kegs, uk_link) { 5389 kfree = pages = 0; 5390 for (i = 0; i < vm_ndomains; i++) { 5391 kfree += kz->uk_domain[i].ud_free_items; 5392 pages += kz->uk_domain[i].ud_pages; 5393 } 5394 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 5395 bzero(&uth, sizeof(uth)); 5396 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 5397 uth.uth_align = kz->uk_align; 5398 uth.uth_size = kz->uk_size; 5399 uth.uth_rsize = kz->uk_rsize; 5400 if (z->uz_max_items > 0) { 5401 items = UZ_ITEMS_COUNT(z->uz_items); 5402 uth.uth_pages = (items / kz->uk_ipers) * 5403 kz->uk_ppera; 5404 } else 5405 uth.uth_pages = pages; 5406 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) * 5407 kz->uk_ppera; 5408 uth.uth_limit = z->uz_max_items; 5409 uth.uth_keg_free = kfree; 5410 5411 /* 5412 * A zone is secondary is it is not the first entry 5413 * on the keg's zone list. 5414 */ 5415 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 5416 (LIST_FIRST(&kz->uk_zones) != z)) 5417 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 5418 uma_vm_zone_stats(&uth, z, &sbuf, ups, 5419 kz->uk_flags & UMA_ZFLAG_INTERNAL); 5420 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 5421 for (i = 0; i < mp_maxid + 1; i++) 5422 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 5423 } 5424 } 5425 LIST_FOREACH(z, &uma_cachezones, uz_link) { 5426 bzero(&uth, sizeof(uth)); 5427 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 5428 uth.uth_size = z->uz_size; 5429 uma_vm_zone_stats(&uth, z, &sbuf, ups, false); 5430 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 5431 for (i = 0; i < mp_maxid + 1; i++) 5432 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 5433 } 5434 5435 rw_runlock(&uma_rwlock); 5436 error = sbuf_finish(&sbuf); 5437 sbuf_delete(&sbuf); 5438 free(ups, M_TEMP); 5439 return (error); 5440 } 5441 5442 int 5443 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 5444 { 5445 uma_zone_t zone = *(uma_zone_t *)arg1; 5446 int error, max; 5447 5448 max = uma_zone_get_max(zone); 5449 error = sysctl_handle_int(oidp, &max, 0, req); 5450 if (error || !req->newptr) 5451 return (error); 5452 5453 uma_zone_set_max(zone, max); 5454 5455 return (0); 5456 } 5457 5458 int 5459 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 5460 { 5461 uma_zone_t zone; 5462 int cur; 5463 5464 /* 5465 * Some callers want to add sysctls for global zones that 5466 * may not yet exist so they pass a pointer to a pointer. 5467 */ 5468 if (arg2 == 0) 5469 zone = *(uma_zone_t *)arg1; 5470 else 5471 zone = arg1; 5472 cur = uma_zone_get_cur(zone); 5473 return (sysctl_handle_int(oidp, &cur, 0, req)); 5474 } 5475 5476 static int 5477 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS) 5478 { 5479 uma_zone_t zone = arg1; 5480 uint64_t cur; 5481 5482 cur = uma_zone_get_allocs(zone); 5483 return (sysctl_handle_64(oidp, &cur, 0, req)); 5484 } 5485 5486 static int 5487 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS) 5488 { 5489 uma_zone_t zone = arg1; 5490 uint64_t cur; 5491 5492 cur = uma_zone_get_frees(zone); 5493 return (sysctl_handle_64(oidp, &cur, 0, req)); 5494 } 5495 5496 static int 5497 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS) 5498 { 5499 struct sbuf sbuf; 5500 uma_zone_t zone = arg1; 5501 int error; 5502 5503 sbuf_new_for_sysctl(&sbuf, NULL, 0, req); 5504 if (zone->uz_flags != 0) 5505 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS); 5506 else 5507 sbuf_printf(&sbuf, "0"); 5508 error = sbuf_finish(&sbuf); 5509 sbuf_delete(&sbuf); 5510 5511 return (error); 5512 } 5513 5514 static int 5515 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS) 5516 { 5517 uma_keg_t keg = arg1; 5518 int avail, effpct, total; 5519 5520 total = keg->uk_ppera * PAGE_SIZE; 5521 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0) 5522 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize; 5523 /* 5524 * We consider the client's requested size and alignment here, not the 5525 * real size determination uk_rsize, because we also adjust the real 5526 * size for internal implementation reasons (max bitset size). 5527 */ 5528 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1); 5529 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) 5530 avail *= mp_maxid + 1; 5531 effpct = 100 * avail / total; 5532 return (sysctl_handle_int(oidp, &effpct, 0, req)); 5533 } 5534 5535 static int 5536 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS) 5537 { 5538 uma_zone_t zone = arg1; 5539 uint64_t cur; 5540 5541 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items)); 5542 return (sysctl_handle_64(oidp, &cur, 0, req)); 5543 } 5544 5545 #ifdef INVARIANTS 5546 static uma_slab_t 5547 uma_dbg_getslab(uma_zone_t zone, void *item) 5548 { 5549 uma_slab_t slab; 5550 uma_keg_t keg; 5551 uint8_t *mem; 5552 5553 /* 5554 * It is safe to return the slab here even though the 5555 * zone is unlocked because the item's allocation state 5556 * essentially holds a reference. 5557 */ 5558 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 5559 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 5560 return (NULL); 5561 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB) 5562 return (vtoslab((vm_offset_t)mem)); 5563 keg = zone->uz_keg; 5564 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0) 5565 return ((uma_slab_t)(mem + keg->uk_pgoff)); 5566 KEG_LOCK(keg, 0); 5567 slab = hash_sfind(&keg->uk_hash, mem); 5568 KEG_UNLOCK(keg, 0); 5569 5570 return (slab); 5571 } 5572 5573 static bool 5574 uma_dbg_zskip(uma_zone_t zone, void *mem) 5575 { 5576 5577 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 5578 return (true); 5579 5580 return (uma_dbg_kskip(zone->uz_keg, mem)); 5581 } 5582 5583 static bool 5584 uma_dbg_kskip(uma_keg_t keg, void *mem) 5585 { 5586 uintptr_t idx; 5587 5588 if (dbg_divisor == 0) 5589 return (true); 5590 5591 if (dbg_divisor == 1) 5592 return (false); 5593 5594 idx = (uintptr_t)mem >> PAGE_SHIFT; 5595 if (keg->uk_ipers > 1) { 5596 idx *= keg->uk_ipers; 5597 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize; 5598 } 5599 5600 if ((idx / dbg_divisor) * dbg_divisor != idx) { 5601 counter_u64_add(uma_skip_cnt, 1); 5602 return (true); 5603 } 5604 counter_u64_add(uma_dbg_cnt, 1); 5605 5606 return (false); 5607 } 5608 5609 /* 5610 * Set up the slab's freei data such that uma_dbg_free can function. 5611 * 5612 */ 5613 static void 5614 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) 5615 { 5616 uma_keg_t keg; 5617 int freei; 5618 5619 if (slab == NULL) { 5620 slab = uma_dbg_getslab(zone, item); 5621 if (slab == NULL) 5622 panic("uma: item %p did not belong to zone %s", 5623 item, zone->uz_name); 5624 } 5625 keg = zone->uz_keg; 5626 freei = slab_item_index(slab, keg, item); 5627 5628 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei, 5629 slab_dbg_bits(slab, keg))) 5630 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)", 5631 item, zone, zone->uz_name, slab, freei); 5632 } 5633 5634 /* 5635 * Verifies freed addresses. Checks for alignment, valid slab membership 5636 * and duplicate frees. 5637 * 5638 */ 5639 static void 5640 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) 5641 { 5642 uma_keg_t keg; 5643 int freei; 5644 5645 if (slab == NULL) { 5646 slab = uma_dbg_getslab(zone, item); 5647 if (slab == NULL) 5648 panic("uma: Freed item %p did not belong to zone %s", 5649 item, zone->uz_name); 5650 } 5651 keg = zone->uz_keg; 5652 freei = slab_item_index(slab, keg, item); 5653 5654 if (freei >= keg->uk_ipers) 5655 panic("Invalid free of %p from zone %p(%s) slab %p(%d)", 5656 item, zone, zone->uz_name, slab, freei); 5657 5658 if (slab_item(slab, keg, freei) != item) 5659 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)", 5660 item, zone, zone->uz_name, slab, freei); 5661 5662 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei, 5663 slab_dbg_bits(slab, keg))) 5664 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)", 5665 item, zone, zone->uz_name, slab, freei); 5666 } 5667 #endif /* INVARIANTS */ 5668 5669 #ifdef DDB 5670 static int64_t 5671 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used, 5672 uint64_t *sleeps, long *cachefree, uint64_t *xdomain) 5673 { 5674 uint64_t frees; 5675 int i; 5676 5677 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 5678 *allocs = counter_u64_fetch(z->uz_allocs); 5679 frees = counter_u64_fetch(z->uz_frees); 5680 *sleeps = z->uz_sleeps; 5681 *cachefree = 0; 5682 *xdomain = 0; 5683 } else 5684 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps, 5685 xdomain); 5686 for (i = 0; i < vm_ndomains; i++) { 5687 *cachefree += ZDOM_GET(z, i)->uzd_nitems; 5688 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 5689 (LIST_FIRST(&kz->uk_zones) != z))) 5690 *cachefree += kz->uk_domain[i].ud_free_items; 5691 } 5692 *used = *allocs - frees; 5693 return (((int64_t)*used + *cachefree) * kz->uk_size); 5694 } 5695 5696 DB_SHOW_COMMAND(uma, db_show_uma) 5697 { 5698 const char *fmt_hdr, *fmt_entry; 5699 uma_keg_t kz; 5700 uma_zone_t z; 5701 uint64_t allocs, used, sleeps, xdomain; 5702 long cachefree; 5703 /* variables for sorting */ 5704 uma_keg_t cur_keg; 5705 uma_zone_t cur_zone, last_zone; 5706 int64_t cur_size, last_size, size; 5707 int ties; 5708 5709 /* /i option produces machine-parseable CSV output */ 5710 if (modif[0] == 'i') { 5711 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n"; 5712 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n"; 5713 } else { 5714 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n"; 5715 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n"; 5716 } 5717 5718 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests", 5719 "Sleeps", "Bucket", "Total Mem", "XFree"); 5720 5721 /* Sort the zones with largest size first. */ 5722 last_zone = NULL; 5723 last_size = INT64_MAX; 5724 for (;;) { 5725 cur_zone = NULL; 5726 cur_size = -1; 5727 ties = 0; 5728 LIST_FOREACH(kz, &uma_kegs, uk_link) { 5729 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 5730 /* 5731 * In the case of size ties, print out zones 5732 * in the order they are encountered. That is, 5733 * when we encounter the most recently output 5734 * zone, we have already printed all preceding 5735 * ties, and we must print all following ties. 5736 */ 5737 if (z == last_zone) { 5738 ties = 1; 5739 continue; 5740 } 5741 size = get_uma_stats(kz, z, &allocs, &used, 5742 &sleeps, &cachefree, &xdomain); 5743 if (size > cur_size && size < last_size + ties) 5744 { 5745 cur_size = size; 5746 cur_zone = z; 5747 cur_keg = kz; 5748 } 5749 } 5750 } 5751 if (cur_zone == NULL) 5752 break; 5753 5754 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used, 5755 &sleeps, &cachefree, &xdomain); 5756 db_printf(fmt_entry, cur_zone->uz_name, 5757 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree, 5758 (uintmax_t)allocs, (uintmax_t)sleeps, 5759 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size, 5760 xdomain); 5761 5762 if (db_pager_quit) 5763 return; 5764 last_zone = cur_zone; 5765 last_size = cur_size; 5766 } 5767 } 5768 5769 DB_SHOW_COMMAND(umacache, db_show_umacache) 5770 { 5771 uma_zone_t z; 5772 uint64_t allocs, frees; 5773 long cachefree; 5774 int i; 5775 5776 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 5777 "Requests", "Bucket"); 5778 LIST_FOREACH(z, &uma_cachezones, uz_link) { 5779 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL); 5780 for (i = 0; i < vm_ndomains; i++) 5781 cachefree += ZDOM_GET(z, i)->uzd_nitems; 5782 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n", 5783 z->uz_name, (uintmax_t)z->uz_size, 5784 (intmax_t)(allocs - frees), cachefree, 5785 (uintmax_t)allocs, z->uz_bucket_size); 5786 if (db_pager_quit) 5787 return; 5788 } 5789 } 5790 #endif /* DDB */ 5791