1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 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 #include "opt_ddb.h" 54 #include "opt_param.h" 55 #include "opt_vm.h" 56 57 #include <sys/param.h> 58 #include <sys/systm.h> 59 #include <sys/asan.h> 60 #include <sys/bitset.h> 61 #include <sys/domainset.h> 62 #include <sys/eventhandler.h> 63 #include <sys/kernel.h> 64 #include <sys/types.h> 65 #include <sys/limits.h> 66 #include <sys/queue.h> 67 #include <sys/malloc.h> 68 #include <sys/ktr.h> 69 #include <sys/lock.h> 70 #include <sys/msan.h> 71 #include <sys/mutex.h> 72 #include <sys/proc.h> 73 #include <sys/random.h> 74 #include <sys/rwlock.h> 75 #include <sys/sbuf.h> 76 #include <sys/sched.h> 77 #include <sys/sleepqueue.h> 78 #include <sys/smp.h> 79 #include <sys/smr.h> 80 #include <sys/sysctl.h> 81 #include <sys/taskqueue.h> 82 #include <sys/vmmeter.h> 83 84 #include <vm/vm.h> 85 #include <vm/vm_param.h> 86 #include <vm/vm_domainset.h> 87 #include <vm/vm_object.h> 88 #include <vm/vm_page.h> 89 #include <vm/vm_pageout.h> 90 #include <vm/vm_phys.h> 91 #include <vm/vm_pagequeue.h> 92 #include <vm/vm_map.h> 93 #include <vm/vm_kern.h> 94 #include <vm/vm_extern.h> 95 #include <vm/vm_dumpset.h> 96 #include <vm/uma.h> 97 #include <vm/uma_int.h> 98 #include <vm/uma_dbg.h> 99 100 #include <ddb/ddb.h> 101 102 #ifdef DEBUG_MEMGUARD 103 #include <vm/memguard.h> 104 #endif 105 106 #include <machine/md_var.h> 107 108 #ifdef INVARIANTS 109 #define UMA_ALWAYS_CTORDTOR 1 110 #else 111 #define UMA_ALWAYS_CTORDTOR 0 112 #endif 113 114 /* 115 * This is the zone and keg from which all zones are spawned. 116 */ 117 static uma_zone_t kegs; 118 static uma_zone_t zones; 119 120 /* 121 * On INVARIANTS builds, the slab contains a second bitset of the same size, 122 * "dbg_bits", which is laid out immediately after us_free. 123 */ 124 #ifdef INVARIANTS 125 #define SLAB_BITSETS 2 126 #else 127 #define SLAB_BITSETS 1 128 #endif 129 130 /* 131 * These are the two zones from which all offpage uma_slab_ts are allocated. 132 * 133 * One zone is for slab headers that can represent a larger number of items, 134 * making the slabs themselves more efficient, and the other zone is for 135 * headers that are smaller and represent fewer items, making the headers more 136 * efficient. 137 */ 138 #define SLABZONE_SIZE(setsize) \ 139 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS) 140 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16) 141 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE 142 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE) 143 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE) 144 static uma_zone_t slabzones[2]; 145 146 /* 147 * The initial hash tables come out of this zone so they can be allocated 148 * prior to malloc coming up. 149 */ 150 static uma_zone_t hashzone; 151 152 /* The boot-time adjusted value for cache line alignment. */ 153 static unsigned int uma_cache_align_mask = 64 - 1; 154 155 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 156 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc"); 157 158 /* 159 * Are we allowed to allocate buckets? 160 */ 161 static int bucketdisable = 1; 162 163 /* Linked list of all kegs in the system */ 164 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 165 166 /* Linked list of all cache-only zones in the system */ 167 static LIST_HEAD(,uma_zone) uma_cachezones = 168 LIST_HEAD_INITIALIZER(uma_cachezones); 169 170 /* 171 * Mutex for global lists: uma_kegs, uma_cachezones, and the per-keg list of 172 * zones. 173 */ 174 static struct rwlock_padalign __exclusive_cache_line uma_rwlock; 175 176 static struct sx uma_reclaim_lock; 177 178 /* 179 * First available virual address for boot time allocations. 180 */ 181 static vm_offset_t bootstart; 182 static vm_offset_t bootmem; 183 184 /* 185 * kmem soft limit, initialized by uma_set_limit(). Ensure that early 186 * allocations don't trigger a wakeup of the reclaim thread. 187 */ 188 unsigned long uma_kmem_limit = LONG_MAX; 189 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0, 190 "UMA kernel memory soft limit"); 191 unsigned long uma_kmem_total; 192 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0, 193 "UMA kernel memory usage"); 194 195 /* Is the VM done starting up? */ 196 static enum { 197 BOOT_COLD, 198 BOOT_KVA, 199 BOOT_PCPU, 200 BOOT_RUNNING, 201 BOOT_SHUTDOWN, 202 } booted = BOOT_COLD; 203 204 /* 205 * This is the handle used to schedule events that need to happen 206 * outside of the allocation fast path. 207 */ 208 static struct timeout_task uma_timeout_task; 209 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 210 211 /* 212 * This structure is passed as the zone ctor arg so that I don't have to create 213 * a special allocation function just for zones. 214 */ 215 struct uma_zctor_args { 216 const char *name; 217 size_t size; 218 uma_ctor ctor; 219 uma_dtor dtor; 220 uma_init uminit; 221 uma_fini fini; 222 uma_import import; 223 uma_release release; 224 void *arg; 225 uma_keg_t keg; 226 int align; 227 uint32_t flags; 228 }; 229 230 struct uma_kctor_args { 231 uma_zone_t zone; 232 size_t size; 233 uma_init uminit; 234 uma_fini fini; 235 int align; 236 uint32_t flags; 237 }; 238 239 struct uma_bucket_zone { 240 uma_zone_t ubz_zone; 241 const char *ubz_name; 242 int ubz_entries; /* Number of items it can hold. */ 243 int ubz_maxsize; /* Maximum allocation size per-item. */ 244 }; 245 246 /* 247 * Compute the actual number of bucket entries to pack them in power 248 * of two sizes for more efficient space utilization. 249 */ 250 #define BUCKET_SIZE(n) \ 251 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 252 253 #define BUCKET_MAX BUCKET_SIZE(256) 254 255 struct uma_bucket_zone bucket_zones[] = { 256 /* Literal bucket sizes. */ 257 { NULL, "2 Bucket", 2, 4096 }, 258 { NULL, "4 Bucket", 4, 3072 }, 259 { NULL, "8 Bucket", 8, 2048 }, 260 { NULL, "16 Bucket", 16, 1024 }, 261 /* Rounded down power of 2 sizes for efficiency. */ 262 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 263 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 264 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 265 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 266 { NULL, NULL, 0} 267 }; 268 269 /* 270 * Flags and enumerations to be passed to internal functions. 271 */ 272 enum zfreeskip { 273 SKIP_NONE = 0, 274 SKIP_CNT = 0x00000001, 275 SKIP_DTOR = 0x00010000, 276 SKIP_FINI = 0x00020000, 277 }; 278 279 /* Prototypes.. */ 280 281 void uma_startup1(vm_offset_t); 282 void uma_startup2(void); 283 284 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 285 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 286 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 287 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 288 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 289 static void page_free(void *, vm_size_t, uint8_t); 290 static void pcpu_page_free(void *, vm_size_t, uint8_t); 291 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int); 292 static void cache_drain(uma_zone_t); 293 static void bucket_drain(uma_zone_t, uma_bucket_t); 294 static void bucket_cache_reclaim(uma_zone_t zone, bool, int); 295 static bool bucket_cache_reclaim_domain(uma_zone_t, bool, bool, int); 296 static int keg_ctor(void *, int, void *, int); 297 static void keg_dtor(void *, int, void *); 298 static void keg_drain(uma_keg_t keg, int domain); 299 static int zone_ctor(void *, int, void *, int); 300 static void zone_dtor(void *, int, void *); 301 static inline void item_dtor(uma_zone_t zone, void *item, int size, 302 void *udata, enum zfreeskip skip); 303 static int zero_init(void *, int, int); 304 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata, 305 int itemdomain, bool ws); 306 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *); 307 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *); 308 static void zone_timeout(uma_zone_t zone, void *); 309 static int hash_alloc(struct uma_hash *, u_int); 310 static int hash_expand(struct uma_hash *, struct uma_hash *); 311 static void hash_free(struct uma_hash *hash); 312 static void uma_timeout(void *, int); 313 static void uma_shutdown(void); 314 static void *zone_alloc_item(uma_zone_t, void *, int, int); 315 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 316 static int zone_alloc_limit(uma_zone_t zone, int count, int flags); 317 static void zone_free_limit(uma_zone_t zone, int count); 318 static void bucket_enable(void); 319 static void bucket_init(void); 320 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 321 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 322 static void bucket_zone_drain(int domain); 323 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int); 324 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 325 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item); 326 static size_t slab_sizeof(int nitems); 327 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 328 uma_fini fini, int align, uint32_t flags); 329 static int zone_import(void *, void **, int, int, int); 330 static void zone_release(void *, void **, int); 331 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int); 332 static bool cache_free(uma_zone_t, uma_cache_t, void *, int); 333 334 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 335 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 336 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS); 337 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS); 338 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS); 339 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS); 340 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS); 341 342 static uint64_t uma_zone_get_allocs(uma_zone_t zone); 343 344 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 345 "Memory allocation debugging"); 346 347 #ifdef INVARIANTS 348 static uint64_t uma_keg_get_allocs(uma_keg_t zone); 349 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg); 350 351 static bool uma_dbg_kskip(uma_keg_t keg, void *mem); 352 static bool uma_dbg_zskip(uma_zone_t zone, void *mem); 353 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item); 354 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item); 355 356 static u_int dbg_divisor = 1; 357 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor, 358 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0, 359 "Debug & thrash every this item in memory allocator"); 360 361 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER; 362 static counter_u64_t uma_skip_cnt = EARLY_COUNTER; 363 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD, 364 &uma_dbg_cnt, "memory items debugged"); 365 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD, 366 &uma_skip_cnt, "memory items skipped, not debugged"); 367 #endif 368 369 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 370 "Universal Memory Allocator"); 371 372 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT, 373 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 374 375 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT, 376 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 377 378 static int zone_warnings = 1; 379 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, 380 "Warn when UMA zones becomes full"); 381 382 static int multipage_slabs = 1; 383 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs); 384 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs, 385 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0, 386 "UMA may choose larger slab sizes for better efficiency"); 387 388 /* 389 * Select the slab zone for an offpage slab with the given maximum item count. 390 */ 391 static inline uma_zone_t 392 slabzone(int ipers) 393 { 394 395 return (slabzones[ipers > SLABZONE0_SETSIZE]); 396 } 397 398 /* 399 * This routine checks to see whether or not it's safe to enable buckets. 400 */ 401 static void 402 bucket_enable(void) 403 { 404 405 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init")); 406 bucketdisable = vm_page_count_min(); 407 } 408 409 /* 410 * Initialize bucket_zones, the array of zones of buckets of various sizes. 411 * 412 * For each zone, calculate the memory required for each bucket, consisting 413 * of the header and an array of pointers. 414 */ 415 static void 416 bucket_init(void) 417 { 418 struct uma_bucket_zone *ubz; 419 int size; 420 421 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 422 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 423 size += sizeof(void *) * ubz->ubz_entries; 424 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 425 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 426 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET | 427 UMA_ZONE_FIRSTTOUCH); 428 } 429 } 430 431 /* 432 * Given a desired number of entries for a bucket, return the zone from which 433 * to allocate the bucket. 434 */ 435 static struct uma_bucket_zone * 436 bucket_zone_lookup(int entries) 437 { 438 struct uma_bucket_zone *ubz; 439 440 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 441 if (ubz->ubz_entries >= entries) 442 return (ubz); 443 ubz--; 444 return (ubz); 445 } 446 447 static int 448 bucket_select(int size) 449 { 450 struct uma_bucket_zone *ubz; 451 452 ubz = &bucket_zones[0]; 453 if (size > ubz->ubz_maxsize) 454 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 455 456 for (; ubz->ubz_entries != 0; ubz++) 457 if (ubz->ubz_maxsize < size) 458 break; 459 ubz--; 460 return (ubz->ubz_entries); 461 } 462 463 static uma_bucket_t 464 bucket_alloc(uma_zone_t zone, void *udata, int flags) 465 { 466 struct uma_bucket_zone *ubz; 467 uma_bucket_t bucket; 468 469 /* 470 * Don't allocate buckets early in boot. 471 */ 472 if (__predict_false(booted < BOOT_KVA)) 473 return (NULL); 474 475 /* 476 * To limit bucket recursion we store the original zone flags 477 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 478 * NOVM flag to persist even through deep recursions. We also 479 * store ZFLAG_BUCKET once we have recursed attempting to allocate 480 * a bucket for a bucket zone so we do not allow infinite bucket 481 * recursion. This cookie will even persist to frees of unused 482 * buckets via the allocation path or bucket allocations in the 483 * free path. 484 */ 485 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 486 udata = (void *)(uintptr_t)zone->uz_flags; 487 else { 488 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 489 return (NULL); 490 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 491 } 492 if (((uintptr_t)udata & UMA_ZONE_VM) != 0) 493 flags |= M_NOVM; 494 ubz = bucket_zone_lookup(atomic_load_16(&zone->uz_bucket_size)); 495 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 496 ubz++; 497 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 498 if (bucket) { 499 #ifdef INVARIANTS 500 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 501 #endif 502 bucket->ub_cnt = 0; 503 bucket->ub_entries = min(ubz->ubz_entries, 504 zone->uz_bucket_size_max); 505 bucket->ub_seq = SMR_SEQ_INVALID; 506 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p", 507 zone->uz_name, zone, bucket); 508 } 509 510 return (bucket); 511 } 512 513 static void 514 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 515 { 516 struct uma_bucket_zone *ubz; 517 518 if (bucket->ub_cnt != 0) 519 bucket_drain(zone, bucket); 520 521 KASSERT(bucket->ub_cnt == 0, 522 ("bucket_free: Freeing a non free bucket.")); 523 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID, 524 ("bucket_free: Freeing an SMR bucket.")); 525 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 526 udata = (void *)(uintptr_t)zone->uz_flags; 527 ubz = bucket_zone_lookup(bucket->ub_entries); 528 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 529 } 530 531 static void 532 bucket_zone_drain(int domain) 533 { 534 struct uma_bucket_zone *ubz; 535 536 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 537 uma_zone_reclaim_domain(ubz->ubz_zone, UMA_RECLAIM_DRAIN, 538 domain); 539 } 540 541 #ifdef KASAN 542 _Static_assert(UMA_SMALLEST_UNIT % KASAN_SHADOW_SCALE == 0, 543 "Base UMA allocation size not a multiple of the KASAN scale factor"); 544 545 static void 546 kasan_mark_item_valid(uma_zone_t zone, void *item) 547 { 548 void *pcpu_item; 549 size_t sz, rsz; 550 int i; 551 552 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0) 553 return; 554 555 sz = zone->uz_size; 556 rsz = roundup2(sz, KASAN_SHADOW_SCALE); 557 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) { 558 kasan_mark(item, sz, rsz, KASAN_GENERIC_REDZONE); 559 } else { 560 pcpu_item = zpcpu_base_to_offset(item); 561 for (i = 0; i <= mp_maxid; i++) 562 kasan_mark(zpcpu_get_cpu(pcpu_item, i), sz, rsz, 563 KASAN_GENERIC_REDZONE); 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, 584 KASAN_UMA_FREED); 585 } 586 } 587 588 static void 589 kasan_mark_slab_valid(uma_keg_t keg, void *mem) 590 { 591 size_t sz; 592 593 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) { 594 sz = keg->uk_ppera * PAGE_SIZE; 595 kasan_mark(mem, sz, sz, 0); 596 } 597 } 598 599 static void 600 kasan_mark_slab_invalid(uma_keg_t keg, void *mem) 601 { 602 size_t sz; 603 604 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) { 605 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0) 606 sz = keg->uk_ppera * PAGE_SIZE; 607 else 608 sz = keg->uk_pgoff; 609 kasan_mark(mem, 0, sz, KASAN_UMA_FREED); 610 } 611 } 612 #else /* !KASAN */ 613 static void 614 kasan_mark_item_valid(uma_zone_t zone __unused, void *item __unused) 615 { 616 } 617 618 static void 619 kasan_mark_item_invalid(uma_zone_t zone __unused, void *item __unused) 620 { 621 } 622 623 static void 624 kasan_mark_slab_valid(uma_keg_t keg __unused, void *mem __unused) 625 { 626 } 627 628 static void 629 kasan_mark_slab_invalid(uma_keg_t keg __unused, void *mem __unused) 630 { 631 } 632 #endif /* KASAN */ 633 634 #ifdef KMSAN 635 static inline void 636 kmsan_mark_item_uninitialized(uma_zone_t zone, void *item) 637 { 638 void *pcpu_item; 639 size_t sz; 640 int i; 641 642 if ((zone->uz_flags & 643 (UMA_ZFLAG_CACHE | UMA_ZONE_SECONDARY | UMA_ZONE_MALLOC)) != 0) { 644 /* 645 * Cache zones should not be instrumented by default, as UMA 646 * does not have enough information to do so correctly. 647 * Consumers can mark items themselves if it makes sense to do 648 * so. 649 * 650 * Items from secondary zones are initialized by the parent 651 * zone and thus cannot safely be marked by UMA. 652 * 653 * malloc zones are handled directly by malloc(9) and friends, 654 * since they can provide more precise origin tracking. 655 */ 656 return; 657 } 658 if (zone->uz_keg->uk_init != NULL) { 659 /* 660 * By definition, initialized items cannot be marked. The 661 * best we can do is mark items from these zones after they 662 * are freed to the keg. 663 */ 664 return; 665 } 666 667 sz = zone->uz_size; 668 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) { 669 kmsan_orig(item, sz, KMSAN_TYPE_UMA, KMSAN_RET_ADDR); 670 kmsan_mark(item, sz, KMSAN_STATE_UNINIT); 671 } else { 672 pcpu_item = zpcpu_base_to_offset(item); 673 for (i = 0; i <= mp_maxid; i++) { 674 kmsan_orig(zpcpu_get_cpu(pcpu_item, i), sz, 675 KMSAN_TYPE_UMA, KMSAN_RET_ADDR); 676 kmsan_mark(zpcpu_get_cpu(pcpu_item, i), sz, 677 KMSAN_STATE_INITED); 678 } 679 } 680 } 681 #else /* !KMSAN */ 682 static inline void 683 kmsan_mark_item_uninitialized(uma_zone_t zone __unused, void *item __unused) 684 { 685 } 686 #endif /* KMSAN */ 687 688 /* 689 * Acquire the domain lock and record contention. 690 */ 691 static uma_zone_domain_t 692 zone_domain_lock(uma_zone_t zone, int domain) 693 { 694 uma_zone_domain_t zdom; 695 bool lockfail; 696 697 zdom = ZDOM_GET(zone, domain); 698 lockfail = false; 699 if (ZDOM_OWNED(zdom)) 700 lockfail = true; 701 ZDOM_LOCK(zdom); 702 /* This is unsynchronized. The counter does not need to be precise. */ 703 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max) 704 zone->uz_bucket_size++; 705 return (zdom); 706 } 707 708 /* 709 * Search for the domain with the least cached items and return it if it 710 * is out of balance with the preferred domain. 711 */ 712 static __noinline int 713 zone_domain_lowest(uma_zone_t zone, int pref) 714 { 715 long least, nitems, prefitems; 716 int domain; 717 int i; 718 719 prefitems = least = LONG_MAX; 720 domain = 0; 721 for (i = 0; i < vm_ndomains; i++) { 722 nitems = ZDOM_GET(zone, i)->uzd_nitems; 723 if (nitems < least) { 724 domain = i; 725 least = nitems; 726 } 727 if (domain == pref) 728 prefitems = nitems; 729 } 730 if (prefitems < least * 2) 731 return (pref); 732 733 return (domain); 734 } 735 736 /* 737 * Search for the domain with the most cached items and return it or the 738 * preferred domain if it has enough to proceed. 739 */ 740 static __noinline int 741 zone_domain_highest(uma_zone_t zone, int pref) 742 { 743 long most, nitems; 744 int domain; 745 int i; 746 747 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX) 748 return (pref); 749 750 most = 0; 751 domain = 0; 752 for (i = 0; i < vm_ndomains; i++) { 753 nitems = ZDOM_GET(zone, i)->uzd_nitems; 754 if (nitems > most) { 755 domain = i; 756 most = nitems; 757 } 758 } 759 760 return (domain); 761 } 762 763 /* 764 * Set the maximum imax value. 765 */ 766 static void 767 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems) 768 { 769 long old; 770 771 old = zdom->uzd_imax; 772 do { 773 if (old >= nitems) 774 return; 775 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0); 776 777 /* 778 * We are at new maximum, so do the last WSS update for the old 779 * bimin and prepare to measure next allocation batch. 780 */ 781 if (zdom->uzd_wss < old - zdom->uzd_bimin) 782 zdom->uzd_wss = old - zdom->uzd_bimin; 783 zdom->uzd_bimin = nitems; 784 } 785 786 /* 787 * Attempt to satisfy an allocation by retrieving a full bucket from one of the 788 * zone's caches. If a bucket is found the zone is not locked on return. 789 */ 790 static uma_bucket_t 791 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim) 792 { 793 uma_bucket_t bucket; 794 long cnt; 795 int i; 796 bool dtor = false; 797 798 ZDOM_LOCK_ASSERT(zdom); 799 800 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL) 801 return (NULL); 802 803 /* SMR Buckets can not be re-used until readers expire. */ 804 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && 805 bucket->ub_seq != SMR_SEQ_INVALID) { 806 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false)) 807 return (NULL); 808 bucket->ub_seq = SMR_SEQ_INVALID; 809 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR; 810 if (STAILQ_NEXT(bucket, ub_link) != NULL) 811 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq; 812 } 813 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link); 814 815 KASSERT(zdom->uzd_nitems >= bucket->ub_cnt, 816 ("%s: item count underflow (%ld, %d)", 817 __func__, zdom->uzd_nitems, bucket->ub_cnt)); 818 KASSERT(bucket->ub_cnt > 0, 819 ("%s: empty bucket in bucket cache", __func__)); 820 zdom->uzd_nitems -= bucket->ub_cnt; 821 822 if (reclaim) { 823 /* 824 * Shift the bounds of the current WSS interval to avoid 825 * perturbing the estimates. 826 */ 827 cnt = lmin(zdom->uzd_bimin, bucket->ub_cnt); 828 atomic_subtract_long(&zdom->uzd_imax, cnt); 829 zdom->uzd_bimin -= cnt; 830 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt); 831 if (zdom->uzd_limin >= bucket->ub_cnt) { 832 zdom->uzd_limin -= bucket->ub_cnt; 833 } else { 834 zdom->uzd_limin = 0; 835 zdom->uzd_timin = 0; 836 } 837 } else if (zdom->uzd_bimin > zdom->uzd_nitems) { 838 zdom->uzd_bimin = zdom->uzd_nitems; 839 if (zdom->uzd_imin > zdom->uzd_nitems) 840 zdom->uzd_imin = zdom->uzd_nitems; 841 } 842 843 ZDOM_UNLOCK(zdom); 844 if (dtor) 845 for (i = 0; i < bucket->ub_cnt; i++) 846 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size, 847 NULL, SKIP_NONE); 848 849 return (bucket); 850 } 851 852 /* 853 * Insert a full bucket into the specified cache. The "ws" parameter indicates 854 * whether the bucket's contents should be counted as part of the zone's working 855 * set. The bucket may be freed if it exceeds the bucket limit. 856 */ 857 static void 858 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata, 859 const bool ws) 860 { 861 uma_zone_domain_t zdom; 862 863 /* We don't cache empty buckets. This can happen after a reclaim. */ 864 if (bucket->ub_cnt == 0) 865 goto out; 866 zdom = zone_domain_lock(zone, domain); 867 868 /* 869 * Conditionally set the maximum number of items. 870 */ 871 zdom->uzd_nitems += bucket->ub_cnt; 872 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) { 873 if (ws) { 874 zone_domain_imax_set(zdom, zdom->uzd_nitems); 875 } else { 876 /* 877 * Shift the bounds of the current WSS interval to 878 * avoid perturbing the estimates. 879 */ 880 atomic_add_long(&zdom->uzd_imax, bucket->ub_cnt); 881 zdom->uzd_imin += bucket->ub_cnt; 882 zdom->uzd_bimin += bucket->ub_cnt; 883 zdom->uzd_limin += bucket->ub_cnt; 884 } 885 if (STAILQ_EMPTY(&zdom->uzd_buckets)) 886 zdom->uzd_seq = bucket->ub_seq; 887 888 /* 889 * Try to promote reuse of recently used items. For items 890 * protected by SMR, try to defer reuse to minimize polling. 891 */ 892 if (bucket->ub_seq == SMR_SEQ_INVALID) 893 STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link); 894 else 895 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link); 896 ZDOM_UNLOCK(zdom); 897 return; 898 } 899 zdom->uzd_nitems -= bucket->ub_cnt; 900 ZDOM_UNLOCK(zdom); 901 out: 902 bucket_free(zone, bucket, udata); 903 } 904 905 /* Pops an item out of a per-cpu cache bucket. */ 906 static inline void * 907 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket) 908 { 909 void *item; 910 911 CRITICAL_ASSERT(curthread); 912 913 bucket->ucb_cnt--; 914 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt]; 915 #ifdef INVARIANTS 916 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL; 917 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 918 #endif 919 cache->uc_allocs++; 920 921 return (item); 922 } 923 924 /* Pushes an item into a per-cpu cache bucket. */ 925 static inline void 926 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item) 927 { 928 929 CRITICAL_ASSERT(curthread); 930 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL, 931 ("uma_zfree: Freeing to non free bucket index.")); 932 933 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item; 934 bucket->ucb_cnt++; 935 cache->uc_frees++; 936 } 937 938 /* 939 * Unload a UMA bucket from a per-cpu cache. 940 */ 941 static inline uma_bucket_t 942 cache_bucket_unload(uma_cache_bucket_t bucket) 943 { 944 uma_bucket_t b; 945 946 b = bucket->ucb_bucket; 947 if (b != NULL) { 948 MPASS(b->ub_entries == bucket->ucb_entries); 949 b->ub_cnt = bucket->ucb_cnt; 950 bucket->ucb_bucket = NULL; 951 bucket->ucb_entries = bucket->ucb_cnt = 0; 952 } 953 954 return (b); 955 } 956 957 static inline uma_bucket_t 958 cache_bucket_unload_alloc(uma_cache_t cache) 959 { 960 961 return (cache_bucket_unload(&cache->uc_allocbucket)); 962 } 963 964 static inline uma_bucket_t 965 cache_bucket_unload_free(uma_cache_t cache) 966 { 967 968 return (cache_bucket_unload(&cache->uc_freebucket)); 969 } 970 971 static inline uma_bucket_t 972 cache_bucket_unload_cross(uma_cache_t cache) 973 { 974 975 return (cache_bucket_unload(&cache->uc_crossbucket)); 976 } 977 978 /* 979 * Load a bucket into a per-cpu cache bucket. 980 */ 981 static inline void 982 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b) 983 { 984 985 CRITICAL_ASSERT(curthread); 986 MPASS(bucket->ucb_bucket == NULL); 987 MPASS(b->ub_seq == SMR_SEQ_INVALID); 988 989 bucket->ucb_bucket = b; 990 bucket->ucb_cnt = b->ub_cnt; 991 bucket->ucb_entries = b->ub_entries; 992 } 993 994 static inline void 995 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b) 996 { 997 998 cache_bucket_load(&cache->uc_allocbucket, b); 999 } 1000 1001 static inline void 1002 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b) 1003 { 1004 1005 cache_bucket_load(&cache->uc_freebucket, b); 1006 } 1007 1008 #ifdef NUMA 1009 static inline void 1010 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b) 1011 { 1012 1013 cache_bucket_load(&cache->uc_crossbucket, b); 1014 } 1015 #endif 1016 1017 /* 1018 * Copy and preserve ucb_spare. 1019 */ 1020 static inline void 1021 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2) 1022 { 1023 1024 b1->ucb_bucket = b2->ucb_bucket; 1025 b1->ucb_entries = b2->ucb_entries; 1026 b1->ucb_cnt = b2->ucb_cnt; 1027 } 1028 1029 /* 1030 * Swap two cache buckets. 1031 */ 1032 static inline void 1033 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2) 1034 { 1035 struct uma_cache_bucket b3; 1036 1037 CRITICAL_ASSERT(curthread); 1038 1039 cache_bucket_copy(&b3, b1); 1040 cache_bucket_copy(b1, b2); 1041 cache_bucket_copy(b2, &b3); 1042 } 1043 1044 /* 1045 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache. 1046 */ 1047 static uma_bucket_t 1048 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain) 1049 { 1050 uma_zone_domain_t zdom; 1051 uma_bucket_t bucket; 1052 smr_seq_t seq; 1053 1054 /* 1055 * Avoid the lock if possible. 1056 */ 1057 zdom = ZDOM_GET(zone, domain); 1058 if (zdom->uzd_nitems == 0) 1059 return (NULL); 1060 1061 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 && 1062 (seq = atomic_load_32(&zdom->uzd_seq)) != SMR_SEQ_INVALID && 1063 !smr_poll(zone->uz_smr, seq, false)) 1064 return (NULL); 1065 1066 /* 1067 * Check the zone's cache of buckets. 1068 */ 1069 zdom = zone_domain_lock(zone, domain); 1070 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) 1071 return (bucket); 1072 ZDOM_UNLOCK(zdom); 1073 1074 return (NULL); 1075 } 1076 1077 static void 1078 zone_log_warning(uma_zone_t zone) 1079 { 1080 static const struct timeval warninterval = { 300, 0 }; 1081 1082 if (!zone_warnings || zone->uz_warning == NULL) 1083 return; 1084 1085 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 1086 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 1087 } 1088 1089 static inline void 1090 zone_maxaction(uma_zone_t zone) 1091 { 1092 1093 if (zone->uz_maxaction.ta_func != NULL) 1094 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); 1095 } 1096 1097 /* 1098 * Routine called by timeout which is used to fire off some time interval 1099 * based calculations. (stats, hash size, etc.) 1100 * 1101 * Arguments: 1102 * arg Unused 1103 * 1104 * Returns: 1105 * Nothing 1106 */ 1107 static void 1108 uma_timeout(void *context __unused, int pending __unused) 1109 { 1110 bucket_enable(); 1111 zone_foreach(zone_timeout, NULL); 1112 1113 /* Reschedule this event */ 1114 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task, 1115 UMA_TIMEOUT * hz); 1116 } 1117 1118 /* 1119 * Update the working set size estimates for the zone's bucket cache. 1120 * The constants chosen here are somewhat arbitrary. 1121 */ 1122 static void 1123 zone_domain_update_wss(uma_zone_domain_t zdom) 1124 { 1125 long m; 1126 1127 ZDOM_LOCK_ASSERT(zdom); 1128 MPASS(zdom->uzd_imax >= zdom->uzd_nitems); 1129 MPASS(zdom->uzd_nitems >= zdom->uzd_bimin); 1130 MPASS(zdom->uzd_bimin >= zdom->uzd_imin); 1131 1132 /* 1133 * Estimate WSS as modified moving average of biggest allocation 1134 * batches for each period over few minutes (UMA_TIMEOUT of 20s). 1135 */ 1136 zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4, 1137 zdom->uzd_imax - zdom->uzd_bimin); 1138 1139 /* 1140 * Estimate longtime minimum item count as a combination of recent 1141 * minimum item count, adjusted by WSS for safety, and the modified 1142 * moving average over the last several hours (UMA_TIMEOUT of 20s). 1143 * timin measures time since limin tried to go negative, that means 1144 * we were dangerously close to or got out of cache. 1145 */ 1146 m = zdom->uzd_imin - zdom->uzd_wss; 1147 if (m >= 0) { 1148 if (zdom->uzd_limin >= m) 1149 zdom->uzd_limin = m; 1150 else 1151 zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256; 1152 zdom->uzd_timin++; 1153 } else { 1154 zdom->uzd_limin = 0; 1155 zdom->uzd_timin = 0; 1156 } 1157 1158 /* To reduce period edge effects on WSS keep half of the imax. */ 1159 atomic_subtract_long(&zdom->uzd_imax, 1160 (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2); 1161 zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems; 1162 } 1163 1164 /* 1165 * Routine to perform timeout driven calculations. This expands the 1166 * hashes and does per cpu statistics aggregation. 1167 * 1168 * Returns nothing. 1169 */ 1170 static void 1171 zone_timeout(uma_zone_t zone, void *unused) 1172 { 1173 uma_keg_t keg; 1174 u_int slabs, pages; 1175 1176 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0) 1177 goto trim; 1178 1179 keg = zone->uz_keg; 1180 1181 /* 1182 * Hash zones are non-numa by definition so the first domain 1183 * is the only one present. 1184 */ 1185 KEG_LOCK(keg, 0); 1186 pages = keg->uk_domain[0].ud_pages; 1187 1188 /* 1189 * Expand the keg hash table. 1190 * 1191 * This is done if the number of slabs is larger than the hash size. 1192 * What I'm trying to do here is completely reduce collisions. This 1193 * may be a little aggressive. Should I allow for two collisions max? 1194 */ 1195 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) { 1196 struct uma_hash newhash; 1197 struct uma_hash oldhash; 1198 int ret; 1199 1200 /* 1201 * This is so involved because allocating and freeing 1202 * while the keg lock is held will lead to deadlock. 1203 * I have to do everything in stages and check for 1204 * races. 1205 */ 1206 KEG_UNLOCK(keg, 0); 1207 ret = hash_alloc(&newhash, 1 << fls(slabs)); 1208 KEG_LOCK(keg, 0); 1209 if (ret) { 1210 if (hash_expand(&keg->uk_hash, &newhash)) { 1211 oldhash = keg->uk_hash; 1212 keg->uk_hash = newhash; 1213 } else 1214 oldhash = newhash; 1215 1216 KEG_UNLOCK(keg, 0); 1217 hash_free(&oldhash); 1218 goto trim; 1219 } 1220 } 1221 KEG_UNLOCK(keg, 0); 1222 1223 trim: 1224 /* Trim caches not used for a long time. */ 1225 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0) { 1226 for (int i = 0; i < vm_ndomains; i++) { 1227 if (bucket_cache_reclaim_domain(zone, false, false, i) && 1228 (zone->uz_flags & UMA_ZFLAG_CACHE) == 0) 1229 keg_drain(zone->uz_keg, i); 1230 } 1231 } 1232 } 1233 1234 /* 1235 * Allocate and zero fill the next sized hash table from the appropriate 1236 * backing store. 1237 * 1238 * Arguments: 1239 * hash A new hash structure with the old hash size in uh_hashsize 1240 * 1241 * Returns: 1242 * 1 on success and 0 on failure. 1243 */ 1244 static int 1245 hash_alloc(struct uma_hash *hash, u_int size) 1246 { 1247 size_t alloc; 1248 1249 KASSERT(powerof2(size), ("hash size must be power of 2")); 1250 if (size > UMA_HASH_SIZE_INIT) { 1251 hash->uh_hashsize = size; 1252 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 1253 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT); 1254 } else { 1255 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 1256 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 1257 UMA_ANYDOMAIN, M_WAITOK); 1258 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 1259 } 1260 if (hash->uh_slab_hash) { 1261 bzero(hash->uh_slab_hash, alloc); 1262 hash->uh_hashmask = hash->uh_hashsize - 1; 1263 return (1); 1264 } 1265 1266 return (0); 1267 } 1268 1269 /* 1270 * Expands the hash table for HASH zones. This is done from zone_timeout 1271 * to reduce collisions. This must not be done in the regular allocation 1272 * path, otherwise, we can recurse on the vm while allocating pages. 1273 * 1274 * Arguments: 1275 * oldhash The hash you want to expand 1276 * newhash The hash structure for the new table 1277 * 1278 * Returns: 1279 * Nothing 1280 * 1281 * Discussion: 1282 */ 1283 static int 1284 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 1285 { 1286 uma_hash_slab_t slab; 1287 u_int hval; 1288 u_int idx; 1289 1290 if (!newhash->uh_slab_hash) 1291 return (0); 1292 1293 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 1294 return (0); 1295 1296 /* 1297 * I need to investigate hash algorithms for resizing without a 1298 * full rehash. 1299 */ 1300 1301 for (idx = 0; idx < oldhash->uh_hashsize; idx++) 1302 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) { 1303 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]); 1304 LIST_REMOVE(slab, uhs_hlink); 1305 hval = UMA_HASH(newhash, slab->uhs_data); 1306 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 1307 slab, uhs_hlink); 1308 } 1309 1310 return (1); 1311 } 1312 1313 /* 1314 * Free the hash bucket to the appropriate backing store. 1315 * 1316 * Arguments: 1317 * slab_hash The hash bucket we're freeing 1318 * hashsize The number of entries in that hash bucket 1319 * 1320 * Returns: 1321 * Nothing 1322 */ 1323 static void 1324 hash_free(struct uma_hash *hash) 1325 { 1326 if (hash->uh_slab_hash == NULL) 1327 return; 1328 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 1329 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 1330 else 1331 free(hash->uh_slab_hash, M_UMAHASH); 1332 } 1333 1334 /* 1335 * Frees all outstanding items in a bucket 1336 * 1337 * Arguments: 1338 * zone The zone to free to, must be unlocked. 1339 * bucket The free/alloc bucket with items. 1340 * 1341 * Returns: 1342 * Nothing 1343 */ 1344 static void 1345 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 1346 { 1347 int i; 1348 1349 if (bucket->ub_cnt == 0) 1350 return; 1351 1352 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && 1353 bucket->ub_seq != SMR_SEQ_INVALID) { 1354 smr_wait(zone->uz_smr, bucket->ub_seq); 1355 bucket->ub_seq = SMR_SEQ_INVALID; 1356 for (i = 0; i < bucket->ub_cnt; i++) 1357 item_dtor(zone, bucket->ub_bucket[i], 1358 zone->uz_size, NULL, SKIP_NONE); 1359 } 1360 if (zone->uz_fini) 1361 for (i = 0; i < bucket->ub_cnt; i++) { 1362 kasan_mark_item_valid(zone, bucket->ub_bucket[i]); 1363 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 1364 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]); 1365 } 1366 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 1367 if (zone->uz_max_items > 0) 1368 zone_free_limit(zone, bucket->ub_cnt); 1369 #ifdef INVARIANTS 1370 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt); 1371 #endif 1372 bucket->ub_cnt = 0; 1373 } 1374 1375 /* 1376 * Drains the per cpu caches for a zone. 1377 * 1378 * NOTE: This may only be called while the zone is being torn down, and not 1379 * during normal operation. This is necessary in order that we do not have 1380 * to migrate CPUs to drain the per-CPU caches. 1381 * 1382 * Arguments: 1383 * zone The zone to drain, must be unlocked. 1384 * 1385 * Returns: 1386 * Nothing 1387 */ 1388 static void 1389 cache_drain(uma_zone_t zone) 1390 { 1391 uma_cache_t cache; 1392 uma_bucket_t bucket; 1393 smr_seq_t seq; 1394 int cpu; 1395 1396 /* 1397 * XXX: It is safe to not lock the per-CPU caches, because we're 1398 * tearing down the zone anyway. I.e., there will be no further use 1399 * of the caches at this point. 1400 * 1401 * XXX: It would good to be able to assert that the zone is being 1402 * torn down to prevent improper use of cache_drain(). 1403 */ 1404 seq = SMR_SEQ_INVALID; 1405 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 1406 seq = smr_advance(zone->uz_smr); 1407 CPU_FOREACH(cpu) { 1408 cache = &zone->uz_cpu[cpu]; 1409 bucket = cache_bucket_unload_alloc(cache); 1410 if (bucket != NULL) 1411 bucket_free(zone, bucket, NULL); 1412 bucket = cache_bucket_unload_free(cache); 1413 if (bucket != NULL) { 1414 bucket->ub_seq = seq; 1415 bucket_free(zone, bucket, NULL); 1416 } 1417 bucket = cache_bucket_unload_cross(cache); 1418 if (bucket != NULL) { 1419 bucket->ub_seq = seq; 1420 bucket_free(zone, bucket, NULL); 1421 } 1422 } 1423 bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN); 1424 } 1425 1426 static void 1427 cache_shrink(uma_zone_t zone, void *unused) 1428 { 1429 1430 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 1431 return; 1432 1433 ZONE_LOCK(zone); 1434 zone->uz_bucket_size = 1435 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2; 1436 ZONE_UNLOCK(zone); 1437 } 1438 1439 static void 1440 cache_drain_safe_cpu(uma_zone_t zone, void *unused) 1441 { 1442 uma_cache_t cache; 1443 uma_bucket_t b1, b2, b3; 1444 int domain; 1445 1446 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 1447 return; 1448 1449 b1 = b2 = b3 = NULL; 1450 critical_enter(); 1451 cache = &zone->uz_cpu[curcpu]; 1452 domain = PCPU_GET(domain); 1453 b1 = cache_bucket_unload_alloc(cache); 1454 1455 /* 1456 * Don't flush SMR zone buckets. This leaves the zone without a 1457 * bucket and forces every free to synchronize(). 1458 */ 1459 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) { 1460 b2 = cache_bucket_unload_free(cache); 1461 b3 = cache_bucket_unload_cross(cache); 1462 } 1463 critical_exit(); 1464 1465 if (b1 != NULL) 1466 zone_free_bucket(zone, b1, NULL, domain, false); 1467 if (b2 != NULL) 1468 zone_free_bucket(zone, b2, NULL, domain, false); 1469 if (b3 != NULL) { 1470 /* Adjust the domain so it goes to zone_free_cross. */ 1471 domain = (domain + 1) % vm_ndomains; 1472 zone_free_bucket(zone, b3, NULL, domain, false); 1473 } 1474 } 1475 1476 /* 1477 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 1478 * This is an expensive call because it needs to bind to all CPUs 1479 * one by one and enter a critical section on each of them in order 1480 * to safely access their cache buckets. 1481 * Zone lock must not be held on call this function. 1482 */ 1483 static void 1484 pcpu_cache_drain_safe(uma_zone_t zone) 1485 { 1486 int cpu; 1487 1488 /* 1489 * Polite bucket sizes shrinking was not enough, shrink aggressively. 1490 */ 1491 if (zone) 1492 cache_shrink(zone, NULL); 1493 else 1494 zone_foreach(cache_shrink, NULL); 1495 1496 CPU_FOREACH(cpu) { 1497 thread_lock(curthread); 1498 sched_bind(curthread, cpu); 1499 thread_unlock(curthread); 1500 1501 if (zone) 1502 cache_drain_safe_cpu(zone, NULL); 1503 else 1504 zone_foreach(cache_drain_safe_cpu, NULL); 1505 } 1506 thread_lock(curthread); 1507 sched_unbind(curthread); 1508 thread_unlock(curthread); 1509 } 1510 1511 /* 1512 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller 1513 * requested a drain, otherwise the per-domain caches are trimmed to either 1514 * estimated working set size. 1515 */ 1516 static bool 1517 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain) 1518 { 1519 uma_zone_domain_t zdom; 1520 uma_bucket_t bucket; 1521 long target; 1522 bool done = false; 1523 1524 /* 1525 * The cross bucket is partially filled and not part of 1526 * the item count. Reclaim it individually here. 1527 */ 1528 zdom = ZDOM_GET(zone, domain); 1529 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) { 1530 ZONE_CROSS_LOCK(zone); 1531 bucket = zdom->uzd_cross; 1532 zdom->uzd_cross = NULL; 1533 ZONE_CROSS_UNLOCK(zone); 1534 if (bucket != NULL) 1535 bucket_free(zone, bucket, NULL); 1536 } 1537 1538 /* 1539 * If we were asked to drain the zone, we are done only once 1540 * this bucket cache is empty. If trim, we reclaim items in 1541 * excess of the zone's estimated working set size. Multiple 1542 * consecutive calls will shrink the WSS and so reclaim more. 1543 * If neither drain nor trim, then voluntarily reclaim 1/4 1544 * (to reduce first spike) of items not used for a long time. 1545 */ 1546 ZDOM_LOCK(zdom); 1547 zone_domain_update_wss(zdom); 1548 if (drain) 1549 target = 0; 1550 else if (trim) 1551 target = zdom->uzd_wss; 1552 else if (zdom->uzd_timin > 900 / UMA_TIMEOUT) 1553 target = zdom->uzd_nitems - zdom->uzd_limin / 4; 1554 else { 1555 ZDOM_UNLOCK(zdom); 1556 return (done); 1557 } 1558 while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL && 1559 zdom->uzd_nitems >= target + bucket->ub_cnt) { 1560 bucket = zone_fetch_bucket(zone, zdom, true); 1561 if (bucket == NULL) 1562 break; 1563 bucket_free(zone, bucket, NULL); 1564 done = true; 1565 ZDOM_LOCK(zdom); 1566 } 1567 ZDOM_UNLOCK(zdom); 1568 return (done); 1569 } 1570 1571 static void 1572 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain) 1573 { 1574 int i; 1575 1576 /* 1577 * Shrink the zone bucket size to ensure that the per-CPU caches 1578 * don't grow too large. 1579 */ 1580 if (zone->uz_bucket_size > zone->uz_bucket_size_min) 1581 zone->uz_bucket_size--; 1582 1583 if (domain != UMA_ANYDOMAIN && 1584 (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) { 1585 bucket_cache_reclaim_domain(zone, drain, true, domain); 1586 } else { 1587 for (i = 0; i < vm_ndomains; i++) 1588 bucket_cache_reclaim_domain(zone, drain, true, i); 1589 } 1590 } 1591 1592 static void 1593 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 1594 { 1595 uint8_t *mem; 1596 size_t size; 1597 int i; 1598 uint8_t flags; 1599 1600 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes", 1601 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera); 1602 1603 mem = slab_data(slab, keg); 1604 size = PAGE_SIZE * keg->uk_ppera; 1605 1606 kasan_mark_slab_valid(keg, mem); 1607 if (keg->uk_fini != NULL) { 1608 for (i = start - 1; i > -1; i--) 1609 #ifdef INVARIANTS 1610 /* 1611 * trash_fini implies that dtor was trash_dtor. trash_fini 1612 * would check that memory hasn't been modified since free, 1613 * which executed trash_dtor. 1614 * That's why we need to run uma_dbg_kskip() check here, 1615 * albeit we don't make skip check for other init/fini 1616 * invocations. 1617 */ 1618 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) || 1619 keg->uk_fini != trash_fini) 1620 #endif 1621 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size); 1622 } 1623 flags = slab->us_flags; 1624 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) { 1625 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab), 1626 NULL, SKIP_NONE); 1627 } 1628 keg->uk_freef(mem, size, flags); 1629 uma_total_dec(size); 1630 } 1631 1632 static void 1633 keg_drain_domain(uma_keg_t keg, int domain) 1634 { 1635 struct slabhead freeslabs; 1636 uma_domain_t dom; 1637 uma_slab_t slab, tmp; 1638 uint32_t i, stofree, stokeep, partial; 1639 1640 dom = &keg->uk_domain[domain]; 1641 LIST_INIT(&freeslabs); 1642 1643 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u", 1644 keg->uk_name, keg, domain, dom->ud_free_items); 1645 1646 KEG_LOCK(keg, domain); 1647 1648 /* 1649 * Are the free items in partially allocated slabs sufficient to meet 1650 * the reserve? If not, compute the number of fully free slabs that must 1651 * be kept. 1652 */ 1653 partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers; 1654 if (partial < keg->uk_reserve) { 1655 stokeep = min(dom->ud_free_slabs, 1656 howmany(keg->uk_reserve - partial, keg->uk_ipers)); 1657 } else { 1658 stokeep = 0; 1659 } 1660 stofree = dom->ud_free_slabs - stokeep; 1661 1662 /* 1663 * Partition the free slabs into two sets: those that must be kept in 1664 * order to maintain the reserve, and those that may be released back to 1665 * the system. Since one set may be much larger than the other, 1666 * populate the smaller of the two sets and swap them if necessary. 1667 */ 1668 for (i = min(stofree, stokeep); i > 0; i--) { 1669 slab = LIST_FIRST(&dom->ud_free_slab); 1670 LIST_REMOVE(slab, us_link); 1671 LIST_INSERT_HEAD(&freeslabs, slab, us_link); 1672 } 1673 if (stofree > stokeep) 1674 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link); 1675 1676 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) { 1677 LIST_FOREACH(slab, &freeslabs, us_link) 1678 UMA_HASH_REMOVE(&keg->uk_hash, slab); 1679 } 1680 dom->ud_free_items -= stofree * keg->uk_ipers; 1681 dom->ud_free_slabs -= stofree; 1682 dom->ud_pages -= stofree * keg->uk_ppera; 1683 KEG_UNLOCK(keg, domain); 1684 1685 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp) 1686 keg_free_slab(keg, slab, keg->uk_ipers); 1687 } 1688 1689 /* 1690 * Frees pages from a keg back to the system. This is done on demand from 1691 * the pageout daemon. 1692 * 1693 * Returns nothing. 1694 */ 1695 static void 1696 keg_drain(uma_keg_t keg, int domain) 1697 { 1698 int i; 1699 1700 if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0) 1701 return; 1702 if (domain != UMA_ANYDOMAIN) { 1703 keg_drain_domain(keg, domain); 1704 } else { 1705 for (i = 0; i < vm_ndomains; i++) 1706 keg_drain_domain(keg, i); 1707 } 1708 } 1709 1710 static void 1711 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain) 1712 { 1713 /* 1714 * Count active reclaim operations in order to interlock with 1715 * zone_dtor(), which removes the zone from global lists before 1716 * attempting to reclaim items itself. 1717 * 1718 * The zone may be destroyed while sleeping, so only zone_dtor() should 1719 * specify M_WAITOK. 1720 */ 1721 ZONE_LOCK(zone); 1722 if (waitok == M_WAITOK) { 1723 while (zone->uz_reclaimers > 0) 1724 msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1); 1725 } 1726 zone->uz_reclaimers++; 1727 ZONE_UNLOCK(zone); 1728 bucket_cache_reclaim(zone, drain, domain); 1729 1730 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) 1731 keg_drain(zone->uz_keg, domain); 1732 ZONE_LOCK(zone); 1733 zone->uz_reclaimers--; 1734 if (zone->uz_reclaimers == 0) 1735 wakeup(zone); 1736 ZONE_UNLOCK(zone); 1737 } 1738 1739 /* 1740 * Allocate a new slab for a keg and inserts it into the partial slab list. 1741 * The keg should be unlocked on entry. If the allocation succeeds it will 1742 * be locked on return. 1743 * 1744 * Arguments: 1745 * flags Wait flags for the item initialization routine 1746 * aflags Wait flags for the slab allocation 1747 * 1748 * Returns: 1749 * The slab that was allocated or NULL if there is no memory and the 1750 * caller specified M_NOWAIT. 1751 */ 1752 static uma_slab_t 1753 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags, 1754 int aflags) 1755 { 1756 uma_domain_t dom; 1757 uma_slab_t slab; 1758 unsigned long size; 1759 uint8_t *mem; 1760 uint8_t sflags; 1761 int i; 1762 1763 TSENTER(); 1764 1765 KASSERT(domain >= 0 && domain < vm_ndomains, 1766 ("keg_alloc_slab: domain %d out of range", domain)); 1767 1768 slab = NULL; 1769 mem = NULL; 1770 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) { 1771 uma_hash_slab_t hslab; 1772 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL, 1773 domain, aflags); 1774 if (hslab == NULL) 1775 goto fail; 1776 slab = &hslab->uhs_slab; 1777 } 1778 1779 /* 1780 * This reproduces the old vm_zone behavior of zero filling pages the 1781 * first time they are added to a zone. 1782 * 1783 * Malloced items are zeroed in uma_zalloc. 1784 */ 1785 1786 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 1787 aflags |= M_ZERO; 1788 else 1789 aflags &= ~M_ZERO; 1790 1791 if (keg->uk_flags & UMA_ZONE_NODUMP) 1792 aflags |= M_NODUMP; 1793 1794 /* zone is passed for legacy reasons. */ 1795 size = keg->uk_ppera * PAGE_SIZE; 1796 mem = keg->uk_allocf(zone, size, domain, &sflags, aflags); 1797 if (mem == NULL) { 1798 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) 1799 zone_free_item(slabzone(keg->uk_ipers), 1800 slab_tohashslab(slab), NULL, SKIP_NONE); 1801 goto fail; 1802 } 1803 uma_total_inc(size); 1804 1805 /* For HASH zones all pages go to the same uma_domain. */ 1806 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) 1807 domain = 0; 1808 1809 kmsan_mark(mem, size, 1810 (aflags & M_ZERO) != 0 ? KMSAN_STATE_INITED : KMSAN_STATE_UNINIT); 1811 1812 /* Point the slab into the allocated memory */ 1813 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) 1814 slab = (uma_slab_t)(mem + keg->uk_pgoff); 1815 else 1816 slab_tohashslab(slab)->uhs_data = mem; 1817 1818 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB) 1819 for (i = 0; i < keg->uk_ppera; i++) 1820 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE), 1821 zone, slab); 1822 1823 slab->us_freecount = keg->uk_ipers; 1824 slab->us_flags = sflags; 1825 slab->us_domain = domain; 1826 1827 BIT_FILL(keg->uk_ipers, &slab->us_free); 1828 #ifdef INVARIANTS 1829 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg)); 1830 #endif 1831 1832 if (keg->uk_init != NULL) { 1833 for (i = 0; i < keg->uk_ipers; i++) 1834 if (keg->uk_init(slab_item(slab, keg, i), 1835 keg->uk_size, flags) != 0) 1836 break; 1837 if (i != keg->uk_ipers) { 1838 keg_free_slab(keg, slab, i); 1839 goto fail; 1840 } 1841 } 1842 kasan_mark_slab_invalid(keg, mem); 1843 KEG_LOCK(keg, domain); 1844 1845 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)", 1846 slab, keg->uk_name, keg); 1847 1848 if (keg->uk_flags & UMA_ZFLAG_HASH) 1849 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1850 1851 /* 1852 * If we got a slab here it's safe to mark it partially used 1853 * and return. We assume that the caller is going to remove 1854 * at least one item. 1855 */ 1856 dom = &keg->uk_domain[domain]; 1857 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 1858 dom->ud_pages += keg->uk_ppera; 1859 dom->ud_free_items += keg->uk_ipers; 1860 1861 TSEXIT(); 1862 return (slab); 1863 1864 fail: 1865 return (NULL); 1866 } 1867 1868 /* 1869 * This function is intended to be used early on in place of page_alloc(). It 1870 * performs contiguous physical memory allocations and uses a bump allocator for 1871 * KVA, so is usable before the kernel map is initialized. 1872 */ 1873 static void * 1874 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1875 int wait) 1876 { 1877 vm_paddr_t pa; 1878 vm_page_t m; 1879 int i, pages; 1880 1881 pages = howmany(bytes, PAGE_SIZE); 1882 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__)); 1883 1884 *pflag = UMA_SLAB_BOOT; 1885 m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) | 1886 VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, 1887 VM_MEMATTR_DEFAULT); 1888 if (m == NULL) 1889 return (NULL); 1890 1891 pa = VM_PAGE_TO_PHYS(m); 1892 for (i = 0; i < pages; i++, pa += PAGE_SIZE) { 1893 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING 1894 if ((wait & M_NODUMP) == 0) 1895 dump_add_page(pa); 1896 #endif 1897 } 1898 1899 /* Allocate KVA and indirectly advance bootmem. */ 1900 return ((void *)pmap_map(&bootmem, m->phys_addr, 1901 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE)); 1902 } 1903 1904 static void 1905 startup_free(void *mem, vm_size_t bytes) 1906 { 1907 vm_offset_t va; 1908 vm_page_t m; 1909 1910 va = (vm_offset_t)mem; 1911 m = PHYS_TO_VM_PAGE(pmap_kextract(va)); 1912 1913 /* 1914 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid 1915 * unmapping ranges of the direct map. 1916 */ 1917 if (va >= bootstart && va + bytes <= bootmem) 1918 pmap_remove(kernel_pmap, va, va + bytes); 1919 for (; bytes != 0; bytes -= PAGE_SIZE, m++) { 1920 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING 1921 dump_drop_page(VM_PAGE_TO_PHYS(m)); 1922 #endif 1923 vm_page_unwire_noq(m); 1924 vm_page_free(m); 1925 } 1926 } 1927 1928 /* 1929 * Allocates a number of pages from the system 1930 * 1931 * Arguments: 1932 * bytes The number of bytes requested 1933 * wait Shall we wait? 1934 * 1935 * Returns: 1936 * A pointer to the alloced memory or possibly 1937 * NULL if M_NOWAIT is set. 1938 */ 1939 static void * 1940 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1941 int wait) 1942 { 1943 void *p; /* Returned page */ 1944 1945 *pflag = UMA_SLAB_KERNEL; 1946 p = kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait); 1947 1948 return (p); 1949 } 1950 1951 static void * 1952 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1953 int wait) 1954 { 1955 struct pglist alloctail; 1956 vm_offset_t addr, zkva; 1957 int cpu, flags; 1958 vm_page_t p, p_next; 1959 #ifdef NUMA 1960 struct pcpu *pc; 1961 #endif 1962 1963 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE); 1964 1965 TAILQ_INIT(&alloctail); 1966 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait); 1967 *pflag = UMA_SLAB_KERNEL; 1968 for (cpu = 0; cpu <= mp_maxid; cpu++) { 1969 if (CPU_ABSENT(cpu)) { 1970 p = vm_page_alloc_noobj(flags); 1971 } else { 1972 #ifndef NUMA 1973 p = vm_page_alloc_noobj(flags); 1974 #else 1975 pc = pcpu_find(cpu); 1976 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain))) 1977 p = NULL; 1978 else 1979 p = vm_page_alloc_noobj_domain(pc->pc_domain, 1980 flags); 1981 if (__predict_false(p == NULL)) 1982 p = vm_page_alloc_noobj(flags); 1983 #endif 1984 } 1985 if (__predict_false(p == NULL)) 1986 goto fail; 1987 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1988 } 1989 if ((addr = kva_alloc(bytes)) == 0) 1990 goto fail; 1991 zkva = addr; 1992 TAILQ_FOREACH(p, &alloctail, listq) { 1993 pmap_qenter(zkva, &p, 1); 1994 zkva += PAGE_SIZE; 1995 } 1996 return ((void*)addr); 1997 fail: 1998 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1999 vm_page_unwire_noq(p); 2000 vm_page_free(p); 2001 } 2002 return (NULL); 2003 } 2004 2005 /* 2006 * Allocates a number of pages not belonging to a VM object 2007 * 2008 * Arguments: 2009 * bytes The number of bytes requested 2010 * wait Shall we wait? 2011 * 2012 * Returns: 2013 * A pointer to the alloced memory or possibly 2014 * NULL if M_NOWAIT is set. 2015 */ 2016 static void * 2017 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, 2018 int wait) 2019 { 2020 TAILQ_HEAD(, vm_page) alloctail; 2021 u_long npages; 2022 vm_offset_t retkva, zkva; 2023 vm_page_t p, p_next; 2024 uma_keg_t keg; 2025 int req; 2026 2027 TAILQ_INIT(&alloctail); 2028 keg = zone->uz_keg; 2029 req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED; 2030 if ((wait & M_WAITOK) != 0) 2031 req |= VM_ALLOC_WAITOK; 2032 2033 npages = howmany(bytes, PAGE_SIZE); 2034 while (npages > 0) { 2035 p = vm_page_alloc_noobj_domain(domain, req); 2036 if (p != NULL) { 2037 /* 2038 * Since the page does not belong to an object, its 2039 * listq is unused. 2040 */ 2041 TAILQ_INSERT_TAIL(&alloctail, p, listq); 2042 npages--; 2043 continue; 2044 } 2045 /* 2046 * Page allocation failed, free intermediate pages and 2047 * exit. 2048 */ 2049 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 2050 vm_page_unwire_noq(p); 2051 vm_page_free(p); 2052 } 2053 return (NULL); 2054 } 2055 *flags = UMA_SLAB_PRIV; 2056 zkva = keg->uk_kva + 2057 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 2058 retkva = zkva; 2059 TAILQ_FOREACH(p, &alloctail, listq) { 2060 pmap_qenter(zkva, &p, 1); 2061 zkva += PAGE_SIZE; 2062 } 2063 2064 return ((void *)retkva); 2065 } 2066 2067 /* 2068 * Allocate physically contiguous pages. 2069 */ 2070 static void * 2071 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 2072 int wait) 2073 { 2074 2075 *pflag = UMA_SLAB_KERNEL; 2076 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain), 2077 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT)); 2078 } 2079 2080 #if defined(UMA_USE_DMAP) && !defined(UMA_MD_SMALL_ALLOC) 2081 void * 2082 uma_small_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, 2083 int wait) 2084 { 2085 vm_page_t m; 2086 vm_paddr_t pa; 2087 void *va; 2088 2089 *flags = UMA_SLAB_PRIV; 2090 m = vm_page_alloc_noobj_domain(domain, 2091 malloc2vm_flags(wait) | VM_ALLOC_WIRED); 2092 if (m == NULL) 2093 return (NULL); 2094 pa = m->phys_addr; 2095 if ((wait & M_NODUMP) == 0) 2096 dump_add_page(pa); 2097 va = (void *)PHYS_TO_DMAP(pa); 2098 return (va); 2099 } 2100 #endif 2101 2102 /* 2103 * Frees a number of pages to the system 2104 * 2105 * Arguments: 2106 * mem A pointer to the memory to be freed 2107 * size The size of the memory being freed 2108 * flags The original p->us_flags field 2109 * 2110 * Returns: 2111 * Nothing 2112 */ 2113 static void 2114 page_free(void *mem, vm_size_t size, uint8_t flags) 2115 { 2116 2117 if ((flags & UMA_SLAB_BOOT) != 0) { 2118 startup_free(mem, size); 2119 return; 2120 } 2121 2122 KASSERT((flags & UMA_SLAB_KERNEL) != 0, 2123 ("UMA: page_free used with invalid flags %x", flags)); 2124 2125 kmem_free(mem, size); 2126 } 2127 2128 /* 2129 * Frees pcpu zone allocations 2130 * 2131 * Arguments: 2132 * mem A pointer to the memory to be freed 2133 * size The size of the memory being freed 2134 * flags The original p->us_flags field 2135 * 2136 * Returns: 2137 * Nothing 2138 */ 2139 static void 2140 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags) 2141 { 2142 vm_offset_t sva, curva; 2143 vm_paddr_t paddr; 2144 vm_page_t m; 2145 2146 MPASS(size == (mp_maxid+1)*PAGE_SIZE); 2147 2148 if ((flags & UMA_SLAB_BOOT) != 0) { 2149 startup_free(mem, size); 2150 return; 2151 } 2152 2153 sva = (vm_offset_t)mem; 2154 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) { 2155 paddr = pmap_kextract(curva); 2156 m = PHYS_TO_VM_PAGE(paddr); 2157 vm_page_unwire_noq(m); 2158 vm_page_free(m); 2159 } 2160 pmap_qremove(sva, size >> PAGE_SHIFT); 2161 kva_free(sva, size); 2162 } 2163 2164 #if defined(UMA_USE_DMAP) && !defined(UMA_MD_SMALL_ALLOC) 2165 void 2166 uma_small_free(void *mem, vm_size_t size, uint8_t flags) 2167 { 2168 vm_page_t m; 2169 vm_paddr_t pa; 2170 2171 pa = DMAP_TO_PHYS((vm_offset_t)mem); 2172 dump_drop_page(pa); 2173 m = PHYS_TO_VM_PAGE(pa); 2174 vm_page_unwire_noq(m); 2175 vm_page_free(m); 2176 } 2177 #endif 2178 2179 /* 2180 * Zero fill initializer 2181 * 2182 * Arguments/Returns follow uma_init specifications 2183 */ 2184 static int 2185 zero_init(void *mem, int size, int flags) 2186 { 2187 bzero(mem, size); 2188 return (0); 2189 } 2190 2191 #ifdef INVARIANTS 2192 static struct noslabbits * 2193 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg) 2194 { 2195 2196 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers))); 2197 } 2198 #endif 2199 2200 /* 2201 * Actual size of embedded struct slab (!OFFPAGE). 2202 */ 2203 static size_t 2204 slab_sizeof(int nitems) 2205 { 2206 size_t s; 2207 2208 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS; 2209 return (roundup(s, UMA_ALIGN_PTR + 1)); 2210 } 2211 2212 #define UMA_FIXPT_SHIFT 31 2213 #define UMA_FRAC_FIXPT(n, d) \ 2214 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d))) 2215 #define UMA_FIXPT_PCT(f) \ 2216 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT)) 2217 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100) 2218 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE) 2219 2220 /* 2221 * Compute the number of items that will fit in a slab. If hdr is true, the 2222 * item count may be limited to provide space in the slab for an inline slab 2223 * header. Otherwise, all slab space will be provided for item storage. 2224 */ 2225 static u_int 2226 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr) 2227 { 2228 u_int ipers; 2229 u_int padpi; 2230 2231 /* The padding between items is not needed after the last item. */ 2232 padpi = rsize - size; 2233 2234 if (hdr) { 2235 /* 2236 * Start with the maximum item count and remove items until 2237 * the slab header first alongside the allocatable memory. 2238 */ 2239 for (ipers = MIN(SLAB_MAX_SETSIZE, 2240 (slabsize + padpi - slab_sizeof(1)) / rsize); 2241 ipers > 0 && 2242 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize; 2243 ipers--) 2244 continue; 2245 } else { 2246 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE); 2247 } 2248 2249 return (ipers); 2250 } 2251 2252 struct keg_layout_result { 2253 u_int format; 2254 u_int slabsize; 2255 u_int ipers; 2256 u_int eff; 2257 }; 2258 2259 static void 2260 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt, 2261 struct keg_layout_result *kl) 2262 { 2263 u_int total; 2264 2265 kl->format = fmt; 2266 kl->slabsize = slabsize; 2267 2268 /* Handle INTERNAL as inline with an extra page. */ 2269 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) { 2270 kl->format &= ~UMA_ZFLAG_INTERNAL; 2271 kl->slabsize += PAGE_SIZE; 2272 } 2273 2274 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize, 2275 (fmt & UMA_ZFLAG_OFFPAGE) == 0); 2276 2277 /* Account for memory used by an offpage slab header. */ 2278 total = kl->slabsize; 2279 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0) 2280 total += slabzone(kl->ipers)->uz_keg->uk_rsize; 2281 2282 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total); 2283 } 2284 2285 /* 2286 * Determine the format of a uma keg. This determines where the slab header 2287 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera. 2288 * 2289 * Arguments 2290 * keg The zone we should initialize 2291 * 2292 * Returns 2293 * Nothing 2294 */ 2295 static void 2296 keg_layout(uma_keg_t keg) 2297 { 2298 struct keg_layout_result kl = {}, kl_tmp; 2299 u_int fmts[2]; 2300 u_int alignsize; 2301 u_int nfmt; 2302 u_int pages; 2303 u_int rsize; 2304 u_int slabsize; 2305 u_int i, j; 2306 2307 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 2308 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE && 2309 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0), 2310 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b", 2311 __func__, keg->uk_name, keg->uk_size, keg->uk_flags, 2312 PRINT_UMA_ZFLAGS)); 2313 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 || 2314 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0, 2315 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags, 2316 PRINT_UMA_ZFLAGS)); 2317 2318 alignsize = keg->uk_align + 1; 2319 #ifdef KASAN 2320 /* 2321 * ASAN requires that each allocation be aligned to the shadow map 2322 * scale factor. 2323 */ 2324 if (alignsize < KASAN_SHADOW_SCALE) 2325 alignsize = KASAN_SHADOW_SCALE; 2326 #endif 2327 2328 /* 2329 * Calculate the size of each allocation (rsize) according to 2330 * alignment. If the requested size is smaller than we have 2331 * allocation bits for we round it up. 2332 */ 2333 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT); 2334 rsize = roundup2(rsize, alignsize); 2335 2336 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) { 2337 /* 2338 * We want one item to start on every align boundary in a page. 2339 * To do this we will span pages. We will also extend the item 2340 * by the size of align if it is an even multiple of align. 2341 * Otherwise, it would fall on the same boundary every time. 2342 */ 2343 if ((rsize & alignsize) == 0) 2344 rsize += alignsize; 2345 slabsize = rsize * (PAGE_SIZE / alignsize); 2346 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE); 2347 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE); 2348 slabsize = round_page(slabsize); 2349 } else { 2350 /* 2351 * Start with a slab size of as many pages as it takes to 2352 * represent a single item. We will try to fit as many 2353 * additional items into the slab as possible. 2354 */ 2355 slabsize = round_page(keg->uk_size); 2356 } 2357 2358 /* Build a list of all of the available formats for this keg. */ 2359 nfmt = 0; 2360 2361 /* Evaluate an inline slab layout. */ 2362 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0) 2363 fmts[nfmt++] = 0; 2364 2365 /* TODO: vm_page-embedded slab. */ 2366 2367 /* 2368 * We can't do OFFPAGE if we're internal or if we've been 2369 * asked to not go to the VM for buckets. If we do this we 2370 * may end up going to the VM for slabs which we do not want 2371 * to do if we're UMA_ZONE_VM, which clearly forbids it. 2372 * In those cases, evaluate a pseudo-format called INTERNAL 2373 * which has an inline slab header and one extra page to 2374 * guarantee that it fits. 2375 * 2376 * Otherwise, see if using an OFFPAGE slab will improve our 2377 * efficiency. 2378 */ 2379 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0) 2380 fmts[nfmt++] = UMA_ZFLAG_INTERNAL; 2381 else 2382 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE; 2383 2384 /* 2385 * Choose a slab size and format which satisfy the minimum efficiency. 2386 * Prefer the smallest slab size that meets the constraints. 2387 * 2388 * Start with a minimum slab size, to accommodate CACHESPREAD. Then, 2389 * for small items (up to PAGE_SIZE), the iteration increment is one 2390 * page; and for large items, the increment is one item. 2391 */ 2392 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize); 2393 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u", 2394 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize, 2395 rsize, i)); 2396 for ( ; ; i++) { 2397 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) : 2398 round_page(rsize * (i - 1) + keg->uk_size); 2399 2400 for (j = 0; j < nfmt; j++) { 2401 /* Only if we have no viable format yet. */ 2402 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 && 2403 kl.ipers > 0) 2404 continue; 2405 2406 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp); 2407 if (kl_tmp.eff <= kl.eff) 2408 continue; 2409 2410 kl = kl_tmp; 2411 2412 CTR6(KTR_UMA, "keg %s layout: format %#x " 2413 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff", 2414 keg->uk_name, kl.format, kl.ipers, rsize, 2415 kl.slabsize, UMA_FIXPT_PCT(kl.eff)); 2416 2417 /* Stop when we reach the minimum efficiency. */ 2418 if (kl.eff >= UMA_MIN_EFF) 2419 break; 2420 } 2421 2422 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs || 2423 slabsize >= SLAB_MAX_SETSIZE * rsize || 2424 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0) 2425 break; 2426 } 2427 2428 pages = atop(kl.slabsize); 2429 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) 2430 pages *= mp_maxid + 1; 2431 2432 keg->uk_rsize = rsize; 2433 keg->uk_ipers = kl.ipers; 2434 keg->uk_ppera = pages; 2435 keg->uk_flags |= kl.format; 2436 2437 /* 2438 * How do we find the slab header if it is offpage or if not all item 2439 * start addresses are in the same page? We could solve the latter 2440 * case with vaddr alignment, but we don't. 2441 */ 2442 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 || 2443 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) { 2444 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0) 2445 keg->uk_flags |= UMA_ZFLAG_HASH; 2446 else 2447 keg->uk_flags |= UMA_ZFLAG_VTOSLAB; 2448 } 2449 2450 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u", 2451 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers, 2452 pages); 2453 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE, 2454 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__, 2455 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize, 2456 keg->uk_ipers, pages)); 2457 } 2458 2459 /* 2460 * Keg header ctor. This initializes all fields, locks, etc. And inserts 2461 * the keg onto the global keg list. 2462 * 2463 * Arguments/Returns follow uma_ctor specifications 2464 * udata Actually uma_kctor_args 2465 */ 2466 static int 2467 keg_ctor(void *mem, int size, void *udata, int flags) 2468 { 2469 struct uma_kctor_args *arg = udata; 2470 uma_keg_t keg = mem; 2471 uma_zone_t zone; 2472 int i; 2473 2474 bzero(keg, size); 2475 keg->uk_size = arg->size; 2476 keg->uk_init = arg->uminit; 2477 keg->uk_fini = arg->fini; 2478 keg->uk_align = arg->align; 2479 keg->uk_reserve = 0; 2480 keg->uk_flags = arg->flags; 2481 2482 /* 2483 * We use a global round-robin policy by default. Zones with 2484 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which 2485 * case the iterator is never run. 2486 */ 2487 keg->uk_dr.dr_policy = DOMAINSET_RR(); 2488 keg->uk_dr.dr_iter = 0; 2489 2490 /* 2491 * The primary zone is passed to us at keg-creation time. 2492 */ 2493 zone = arg->zone; 2494 keg->uk_name = zone->uz_name; 2495 2496 if (arg->flags & UMA_ZONE_ZINIT) 2497 keg->uk_init = zero_init; 2498 2499 if (arg->flags & UMA_ZONE_MALLOC) 2500 keg->uk_flags |= UMA_ZFLAG_VTOSLAB; 2501 2502 #ifndef SMP 2503 keg->uk_flags &= ~UMA_ZONE_PCPU; 2504 #endif 2505 2506 keg_layout(keg); 2507 2508 /* 2509 * Use a first-touch NUMA policy for kegs that pmap_extract() will 2510 * work on. Use round-robin for everything else. 2511 * 2512 * Zones may override the default by specifying either. 2513 */ 2514 #ifdef NUMA 2515 if ((keg->uk_flags & 2516 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0) 2517 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH; 2518 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0) 2519 keg->uk_flags |= UMA_ZONE_ROUNDROBIN; 2520 #endif 2521 2522 /* 2523 * If we haven't booted yet we need allocations to go through the 2524 * startup cache until the vm is ready. 2525 */ 2526 #ifdef UMA_USE_DMAP 2527 if (keg->uk_ppera == 1) 2528 keg->uk_allocf = uma_small_alloc; 2529 else 2530 #endif 2531 if (booted < BOOT_KVA) 2532 keg->uk_allocf = startup_alloc; 2533 else if (keg->uk_flags & UMA_ZONE_PCPU) 2534 keg->uk_allocf = pcpu_page_alloc; 2535 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1) 2536 keg->uk_allocf = contig_alloc; 2537 else 2538 keg->uk_allocf = page_alloc; 2539 #ifdef UMA_USE_DMAP 2540 if (keg->uk_ppera == 1) 2541 keg->uk_freef = uma_small_free; 2542 else 2543 #endif 2544 if (keg->uk_flags & UMA_ZONE_PCPU) 2545 keg->uk_freef = pcpu_page_free; 2546 else 2547 keg->uk_freef = page_free; 2548 2549 /* 2550 * Initialize keg's locks. 2551 */ 2552 for (i = 0; i < vm_ndomains; i++) 2553 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS)); 2554 2555 /* 2556 * If we're putting the slab header in the actual page we need to 2557 * figure out where in each page it goes. See slab_sizeof 2558 * definition. 2559 */ 2560 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) { 2561 size_t shsize; 2562 2563 shsize = slab_sizeof(keg->uk_ipers); 2564 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize; 2565 /* 2566 * The only way the following is possible is if with our 2567 * UMA_ALIGN_PTR adjustments we are now bigger than 2568 * UMA_SLAB_SIZE. I haven't checked whether this is 2569 * mathematically possible for all cases, so we make 2570 * sure here anyway. 2571 */ 2572 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera, 2573 ("zone %s ipers %d rsize %d size %d slab won't fit", 2574 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size)); 2575 } 2576 2577 if (keg->uk_flags & UMA_ZFLAG_HASH) 2578 hash_alloc(&keg->uk_hash, 0); 2579 2580 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone); 2581 2582 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 2583 2584 rw_wlock(&uma_rwlock); 2585 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 2586 rw_wunlock(&uma_rwlock); 2587 return (0); 2588 } 2589 2590 static void 2591 zone_kva_available(uma_zone_t zone, void *unused) 2592 { 2593 uma_keg_t keg; 2594 2595 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 2596 return; 2597 KEG_GET(zone, keg); 2598 2599 if (keg->uk_allocf == startup_alloc) { 2600 /* Switch to the real allocator. */ 2601 if (keg->uk_flags & UMA_ZONE_PCPU) 2602 keg->uk_allocf = pcpu_page_alloc; 2603 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && 2604 keg->uk_ppera > 1) 2605 keg->uk_allocf = contig_alloc; 2606 else 2607 keg->uk_allocf = page_alloc; 2608 } 2609 } 2610 2611 static void 2612 zone_alloc_counters(uma_zone_t zone, void *unused) 2613 { 2614 2615 zone->uz_allocs = counter_u64_alloc(M_WAITOK); 2616 zone->uz_frees = counter_u64_alloc(M_WAITOK); 2617 zone->uz_fails = counter_u64_alloc(M_WAITOK); 2618 zone->uz_xdomain = counter_u64_alloc(M_WAITOK); 2619 } 2620 2621 static void 2622 zone_alloc_sysctl(uma_zone_t zone, void *unused) 2623 { 2624 uma_zone_domain_t zdom; 2625 uma_domain_t dom; 2626 uma_keg_t keg; 2627 struct sysctl_oid *oid, *domainoid; 2628 int domains, i, cnt; 2629 static const char *nokeg = "cache zone"; 2630 char *c; 2631 2632 /* 2633 * Make a sysctl safe copy of the zone name by removing 2634 * any special characters and handling dups by appending 2635 * an index. 2636 */ 2637 if (zone->uz_namecnt != 0) { 2638 /* Count the number of decimal digits and '_' separator. */ 2639 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++) 2640 cnt /= 10; 2641 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1, 2642 M_UMA, M_WAITOK); 2643 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name, 2644 zone->uz_namecnt); 2645 } else 2646 zone->uz_ctlname = strdup(zone->uz_name, M_UMA); 2647 for (c = zone->uz_ctlname; *c != '\0'; c++) 2648 if (strchr("./\\ -", *c) != NULL) 2649 *c = '_'; 2650 2651 /* 2652 * Basic parameters at the root. 2653 */ 2654 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma), 2655 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2656 oid = zone->uz_oid; 2657 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2658 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size"); 2659 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2660 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE, 2661 zone, 0, sysctl_handle_uma_zone_flags, "A", 2662 "Allocator configuration flags"); 2663 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2664 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0, 2665 "Desired per-cpu cache size"); 2666 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2667 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0, 2668 "Maximum allowed per-cpu cache size"); 2669 2670 /* 2671 * keg if present. 2672 */ 2673 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0) 2674 domains = vm_ndomains; 2675 else 2676 domains = 1; 2677 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 2678 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2679 keg = zone->uz_keg; 2680 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) { 2681 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2682 "name", CTLFLAG_RD, keg->uk_name, "Keg name"); 2683 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2684 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0, 2685 "Real object size with alignment"); 2686 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2687 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0, 2688 "pages per-slab allocation"); 2689 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2690 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0, 2691 "items available per-slab"); 2692 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2693 "align", CTLFLAG_RD, &keg->uk_align, 0, 2694 "item alignment mask"); 2695 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2696 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0, 2697 "number of reserved items"); 2698 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2699 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE, 2700 keg, 0, sysctl_handle_uma_slab_efficiency, "I", 2701 "Slab utilization (100 - internal fragmentation %)"); 2702 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid), 2703 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2704 for (i = 0; i < domains; i++) { 2705 dom = &keg->uk_domain[i]; 2706 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid), 2707 OID_AUTO, VM_DOMAIN(i)->vmd_name, 2708 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2709 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2710 "pages", CTLFLAG_RD, &dom->ud_pages, 0, 2711 "Total pages currently allocated from VM"); 2712 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2713 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0, 2714 "Items free in the slab layer"); 2715 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2716 "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0, 2717 "Unused slabs"); 2718 } 2719 } else 2720 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2721 "name", CTLFLAG_RD, nokeg, "Keg name"); 2722 2723 /* 2724 * Information about zone limits. 2725 */ 2726 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 2727 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2728 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2729 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2730 zone, 0, sysctl_handle_uma_zone_items, "QU", 2731 "Current number of allocated items if limit is set"); 2732 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2733 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0, 2734 "Maximum number of allocated and cached items"); 2735 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2736 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0, 2737 "Number of threads sleeping at limit"); 2738 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2739 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0, 2740 "Total zone limit sleeps"); 2741 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2742 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0, 2743 "Maximum number of items in each domain's bucket cache"); 2744 2745 /* 2746 * Per-domain zone information. 2747 */ 2748 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), 2749 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2750 for (i = 0; i < domains; i++) { 2751 zdom = ZDOM_GET(zone, i); 2752 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid), 2753 OID_AUTO, VM_DOMAIN(i)->vmd_name, 2754 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2755 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2756 "nitems", CTLFLAG_RD, &zdom->uzd_nitems, 2757 "number of items in this domain"); 2758 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2759 "imax", CTLFLAG_RD, &zdom->uzd_imax, 2760 "maximum item count in this period"); 2761 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2762 "imin", CTLFLAG_RD, &zdom->uzd_imin, 2763 "minimum item count in this period"); 2764 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2765 "bimin", CTLFLAG_RD, &zdom->uzd_bimin, 2766 "Minimum item count in this batch"); 2767 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2768 "wss", CTLFLAG_RD, &zdom->uzd_wss, 2769 "Working set size"); 2770 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2771 "limin", CTLFLAG_RD, &zdom->uzd_limin, 2772 "Long time minimum item count"); 2773 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2774 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0, 2775 "Time since zero long time minimum item count"); 2776 } 2777 2778 /* 2779 * General statistics. 2780 */ 2781 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 2782 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); 2783 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2784 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE, 2785 zone, 1, sysctl_handle_uma_zone_cur, "I", 2786 "Current number of allocated items"); 2787 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2788 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2789 zone, 0, sysctl_handle_uma_zone_allocs, "QU", 2790 "Total allocation calls"); 2791 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2792 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2793 zone, 0, sysctl_handle_uma_zone_frees, "QU", 2794 "Total free calls"); 2795 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2796 "fails", CTLFLAG_RD, &zone->uz_fails, 2797 "Number of allocation failures"); 2798 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2799 "xdomain", CTLFLAG_RD, &zone->uz_xdomain, 2800 "Free calls from the wrong domain"); 2801 } 2802 2803 struct uma_zone_count { 2804 const char *name; 2805 int count; 2806 }; 2807 2808 static void 2809 zone_count(uma_zone_t zone, void *arg) 2810 { 2811 struct uma_zone_count *cnt; 2812 2813 cnt = arg; 2814 /* 2815 * Some zones are rapidly created with identical names and 2816 * destroyed out of order. This can lead to gaps in the count. 2817 * Use one greater than the maximum observed for this name. 2818 */ 2819 if (strcmp(zone->uz_name, cnt->name) == 0) 2820 cnt->count = MAX(cnt->count, 2821 zone->uz_namecnt + 1); 2822 } 2823 2824 static void 2825 zone_update_caches(uma_zone_t zone) 2826 { 2827 int i; 2828 2829 for (i = 0; i <= mp_maxid; i++) { 2830 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size); 2831 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags); 2832 } 2833 } 2834 2835 /* 2836 * Zone header ctor. This initializes all fields, locks, etc. 2837 * 2838 * Arguments/Returns follow uma_ctor specifications 2839 * udata Actually uma_zctor_args 2840 */ 2841 static int 2842 zone_ctor(void *mem, int size, void *udata, int flags) 2843 { 2844 struct uma_zone_count cnt; 2845 struct uma_zctor_args *arg = udata; 2846 uma_zone_domain_t zdom; 2847 uma_zone_t zone = mem; 2848 uma_zone_t z; 2849 uma_keg_t keg; 2850 int i; 2851 2852 bzero(zone, size); 2853 zone->uz_name = arg->name; 2854 zone->uz_ctor = arg->ctor; 2855 zone->uz_dtor = arg->dtor; 2856 zone->uz_init = NULL; 2857 zone->uz_fini = NULL; 2858 zone->uz_sleeps = 0; 2859 zone->uz_bucket_size = 0; 2860 zone->uz_bucket_size_min = 0; 2861 zone->uz_bucket_size_max = BUCKET_MAX; 2862 zone->uz_flags = (arg->flags & UMA_ZONE_SMR); 2863 zone->uz_warning = NULL; 2864 /* The domain structures follow the cpu structures. */ 2865 zone->uz_bucket_max = ULONG_MAX; 2866 timevalclear(&zone->uz_ratecheck); 2867 2868 /* Count the number of duplicate names. */ 2869 cnt.name = arg->name; 2870 cnt.count = 0; 2871 zone_foreach(zone_count, &cnt); 2872 zone->uz_namecnt = cnt.count; 2873 ZONE_CROSS_LOCK_INIT(zone); 2874 2875 for (i = 0; i < vm_ndomains; i++) { 2876 zdom = ZDOM_GET(zone, i); 2877 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS)); 2878 STAILQ_INIT(&zdom->uzd_buckets); 2879 } 2880 2881 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN) 2882 if (arg->uminit == trash_init && arg->fini == trash_fini) 2883 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR; 2884 #elif defined(KASAN) 2885 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0) 2886 arg->flags |= UMA_ZONE_NOKASAN; 2887 #endif 2888 2889 /* 2890 * This is a pure cache zone, no kegs. 2891 */ 2892 if (arg->import) { 2893 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0, 2894 ("zone_ctor: Import specified for non-cache zone.")); 2895 zone->uz_flags = arg->flags; 2896 zone->uz_size = arg->size; 2897 zone->uz_import = arg->import; 2898 zone->uz_release = arg->release; 2899 zone->uz_arg = arg->arg; 2900 #ifdef NUMA 2901 /* 2902 * Cache zones are round-robin unless a policy is 2903 * specified because they may have incompatible 2904 * constraints. 2905 */ 2906 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0) 2907 zone->uz_flags |= UMA_ZONE_ROUNDROBIN; 2908 #endif 2909 rw_wlock(&uma_rwlock); 2910 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 2911 rw_wunlock(&uma_rwlock); 2912 goto out; 2913 } 2914 2915 /* 2916 * Use the regular zone/keg/slab allocator. 2917 */ 2918 zone->uz_import = zone_import; 2919 zone->uz_release = zone_release; 2920 zone->uz_arg = zone; 2921 keg = arg->keg; 2922 2923 if (arg->flags & UMA_ZONE_SECONDARY) { 2924 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 2925 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 2926 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 2927 zone->uz_init = arg->uminit; 2928 zone->uz_fini = arg->fini; 2929 zone->uz_flags |= UMA_ZONE_SECONDARY; 2930 rw_wlock(&uma_rwlock); 2931 ZONE_LOCK(zone); 2932 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 2933 if (LIST_NEXT(z, uz_link) == NULL) { 2934 LIST_INSERT_AFTER(z, zone, uz_link); 2935 break; 2936 } 2937 } 2938 ZONE_UNLOCK(zone); 2939 rw_wunlock(&uma_rwlock); 2940 } else if (keg == NULL) { 2941 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 2942 arg->align, arg->flags)) == NULL) 2943 return (ENOMEM); 2944 } else { 2945 struct uma_kctor_args karg; 2946 int error; 2947 2948 /* We should only be here from uma_startup() */ 2949 karg.size = arg->size; 2950 karg.uminit = arg->uminit; 2951 karg.fini = arg->fini; 2952 karg.align = arg->align; 2953 karg.flags = (arg->flags & ~UMA_ZONE_SMR); 2954 karg.zone = zone; 2955 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 2956 flags); 2957 if (error) 2958 return (error); 2959 } 2960 2961 /* Inherit properties from the keg. */ 2962 zone->uz_keg = keg; 2963 zone->uz_size = keg->uk_size; 2964 zone->uz_flags |= (keg->uk_flags & 2965 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 2966 2967 out: 2968 if (booted >= BOOT_PCPU) { 2969 zone_alloc_counters(zone, NULL); 2970 if (booted >= BOOT_RUNNING) 2971 zone_alloc_sysctl(zone, NULL); 2972 } else { 2973 zone->uz_allocs = EARLY_COUNTER; 2974 zone->uz_frees = EARLY_COUNTER; 2975 zone->uz_fails = EARLY_COUNTER; 2976 } 2977 2978 /* Caller requests a private SMR context. */ 2979 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 2980 zone->uz_smr = smr_create(zone->uz_name, 0, 0); 2981 2982 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) != 2983 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET), 2984 ("Invalid zone flag combination")); 2985 if (arg->flags & UMA_ZFLAG_INTERNAL) 2986 zone->uz_bucket_size_max = zone->uz_bucket_size = 0; 2987 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0) 2988 zone->uz_bucket_size = BUCKET_MAX; 2989 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0) 2990 zone->uz_bucket_size = 0; 2991 else 2992 zone->uz_bucket_size = bucket_select(zone->uz_size); 2993 zone->uz_bucket_size_min = zone->uz_bucket_size; 2994 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL) 2995 zone->uz_flags |= UMA_ZFLAG_CTORDTOR; 2996 zone_update_caches(zone); 2997 2998 return (0); 2999 } 3000 3001 /* 3002 * Keg header dtor. This frees all data, destroys locks, frees the hash 3003 * table and removes the keg from the global list. 3004 * 3005 * Arguments/Returns follow uma_dtor specifications 3006 * udata unused 3007 */ 3008 static void 3009 keg_dtor(void *arg, int size, void *udata) 3010 { 3011 uma_keg_t keg; 3012 uint32_t free, pages; 3013 int i; 3014 3015 keg = (uma_keg_t)arg; 3016 free = pages = 0; 3017 for (i = 0; i < vm_ndomains; i++) { 3018 free += keg->uk_domain[i].ud_free_items; 3019 pages += keg->uk_domain[i].ud_pages; 3020 KEG_LOCK_FINI(keg, i); 3021 } 3022 if (pages != 0) 3023 printf("Freed UMA keg (%s) was not empty (%u items). " 3024 " Lost %u pages of memory.\n", 3025 keg->uk_name ? keg->uk_name : "", 3026 pages / keg->uk_ppera * keg->uk_ipers - free, pages); 3027 3028 hash_free(&keg->uk_hash); 3029 } 3030 3031 /* 3032 * Zone header dtor. 3033 * 3034 * Arguments/Returns follow uma_dtor specifications 3035 * udata unused 3036 */ 3037 static void 3038 zone_dtor(void *arg, int size, void *udata) 3039 { 3040 uma_zone_t zone; 3041 uma_keg_t keg; 3042 int i; 3043 3044 zone = (uma_zone_t)arg; 3045 3046 sysctl_remove_oid(zone->uz_oid, 1, 1); 3047 3048 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 3049 cache_drain(zone); 3050 3051 rw_wlock(&uma_rwlock); 3052 LIST_REMOVE(zone, uz_link); 3053 rw_wunlock(&uma_rwlock); 3054 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) { 3055 keg = zone->uz_keg; 3056 keg->uk_reserve = 0; 3057 } 3058 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true); 3059 3060 /* 3061 * We only destroy kegs from non secondary/non cache zones. 3062 */ 3063 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) { 3064 keg = zone->uz_keg; 3065 rw_wlock(&uma_rwlock); 3066 LIST_REMOVE(keg, uk_link); 3067 rw_wunlock(&uma_rwlock); 3068 zone_free_item(kegs, keg, NULL, SKIP_NONE); 3069 } 3070 counter_u64_free(zone->uz_allocs); 3071 counter_u64_free(zone->uz_frees); 3072 counter_u64_free(zone->uz_fails); 3073 counter_u64_free(zone->uz_xdomain); 3074 free(zone->uz_ctlname, M_UMA); 3075 for (i = 0; i < vm_ndomains; i++) 3076 ZDOM_LOCK_FINI(ZDOM_GET(zone, i)); 3077 ZONE_CROSS_LOCK_FINI(zone); 3078 } 3079 3080 static void 3081 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg) 3082 { 3083 uma_keg_t keg; 3084 uma_zone_t zone; 3085 3086 LIST_FOREACH(keg, &uma_kegs, uk_link) { 3087 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 3088 zfunc(zone, arg); 3089 } 3090 LIST_FOREACH(zone, &uma_cachezones, uz_link) 3091 zfunc(zone, arg); 3092 } 3093 3094 /* 3095 * Traverses every zone in the system and calls a callback 3096 * 3097 * Arguments: 3098 * zfunc A pointer to a function which accepts a zone 3099 * as an argument. 3100 * 3101 * Returns: 3102 * Nothing 3103 */ 3104 static void 3105 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg) 3106 { 3107 3108 rw_rlock(&uma_rwlock); 3109 zone_foreach_unlocked(zfunc, arg); 3110 rw_runlock(&uma_rwlock); 3111 } 3112 3113 /* 3114 * Initialize the kernel memory allocator. This is done after pages can be 3115 * allocated but before general KVA is available. 3116 */ 3117 void 3118 uma_startup1(vm_offset_t virtual_avail) 3119 { 3120 struct uma_zctor_args args; 3121 size_t ksize, zsize, size; 3122 uma_keg_t primarykeg; 3123 uintptr_t m; 3124 int domain; 3125 uint8_t pflag; 3126 3127 bootstart = bootmem = virtual_avail; 3128 3129 rw_init(&uma_rwlock, "UMA lock"); 3130 sx_init(&uma_reclaim_lock, "umareclaim"); 3131 3132 ksize = sizeof(struct uma_keg) + 3133 (sizeof(struct uma_domain) * vm_ndomains); 3134 ksize = roundup(ksize, UMA_SUPER_ALIGN); 3135 zsize = sizeof(struct uma_zone) + 3136 (sizeof(struct uma_cache) * (mp_maxid + 1)) + 3137 (sizeof(struct uma_zone_domain) * vm_ndomains); 3138 zsize = roundup(zsize, UMA_SUPER_ALIGN); 3139 3140 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */ 3141 size = (zsize * 2) + ksize; 3142 for (domain = 0; domain < vm_ndomains; domain++) { 3143 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag, 3144 M_NOWAIT | M_ZERO); 3145 if (m != 0) 3146 break; 3147 } 3148 zones = (uma_zone_t)m; 3149 m += zsize; 3150 kegs = (uma_zone_t)m; 3151 m += zsize; 3152 primarykeg = (uma_keg_t)m; 3153 3154 /* "manually" create the initial zone */ 3155 memset(&args, 0, sizeof(args)); 3156 args.name = "UMA Kegs"; 3157 args.size = ksize; 3158 args.ctor = keg_ctor; 3159 args.dtor = keg_dtor; 3160 args.uminit = zero_init; 3161 args.fini = NULL; 3162 args.keg = primarykeg; 3163 args.align = UMA_SUPER_ALIGN - 1; 3164 args.flags = UMA_ZFLAG_INTERNAL; 3165 zone_ctor(kegs, zsize, &args, M_WAITOK); 3166 3167 args.name = "UMA Zones"; 3168 args.size = zsize; 3169 args.ctor = zone_ctor; 3170 args.dtor = zone_dtor; 3171 args.uminit = zero_init; 3172 args.fini = NULL; 3173 args.keg = NULL; 3174 args.align = UMA_SUPER_ALIGN - 1; 3175 args.flags = UMA_ZFLAG_INTERNAL; 3176 zone_ctor(zones, zsize, &args, M_WAITOK); 3177 3178 /* Now make zones for slab headers */ 3179 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE, 3180 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 3181 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE, 3182 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 3183 3184 hashzone = uma_zcreate("UMA Hash", 3185 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 3186 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 3187 3188 bucket_init(); 3189 smr_init(); 3190 } 3191 3192 #ifndef UMA_USE_DMAP 3193 extern void vm_radix_reserve_kva(void); 3194 #endif 3195 3196 /* 3197 * Advertise the availability of normal kva allocations and switch to 3198 * the default back-end allocator. Marks the KVA we consumed on startup 3199 * as used in the map. 3200 */ 3201 void 3202 uma_startup2(void) 3203 { 3204 3205 if (bootstart != bootmem) { 3206 vm_map_lock(kernel_map); 3207 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem, 3208 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 3209 vm_map_unlock(kernel_map); 3210 } 3211 3212 #ifndef UMA_USE_DMAP 3213 /* Set up radix zone to use noobj_alloc. */ 3214 vm_radix_reserve_kva(); 3215 #endif 3216 3217 booted = BOOT_KVA; 3218 zone_foreach_unlocked(zone_kva_available, NULL); 3219 bucket_enable(); 3220 } 3221 3222 /* 3223 * Allocate counters as early as possible so that boot-time allocations are 3224 * accounted more precisely. 3225 */ 3226 static void 3227 uma_startup_pcpu(void *arg __unused) 3228 { 3229 3230 zone_foreach_unlocked(zone_alloc_counters, NULL); 3231 booted = BOOT_PCPU; 3232 } 3233 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL); 3234 3235 /* 3236 * Finish our initialization steps. 3237 */ 3238 static void 3239 uma_startup3(void *arg __unused) 3240 { 3241 3242 #ifdef INVARIANTS 3243 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor); 3244 uma_dbg_cnt = counter_u64_alloc(M_WAITOK); 3245 uma_skip_cnt = counter_u64_alloc(M_WAITOK); 3246 #endif 3247 zone_foreach_unlocked(zone_alloc_sysctl, NULL); 3248 booted = BOOT_RUNNING; 3249 3250 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL, 3251 EVENTHANDLER_PRI_FIRST); 3252 } 3253 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 3254 3255 static void 3256 uma_startup4(void *arg __unused) 3257 { 3258 TIMEOUT_TASK_INIT(taskqueue_thread, &uma_timeout_task, 0, uma_timeout, 3259 NULL); 3260 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task, 3261 UMA_TIMEOUT * hz); 3262 } 3263 SYSINIT(uma_startup4, SI_SUB_TASKQ, SI_ORDER_ANY, uma_startup4, NULL); 3264 3265 static void 3266 uma_shutdown(void) 3267 { 3268 3269 booted = BOOT_SHUTDOWN; 3270 } 3271 3272 static uma_keg_t 3273 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 3274 int align, uint32_t flags) 3275 { 3276 struct uma_kctor_args args; 3277 3278 args.size = size; 3279 args.uminit = uminit; 3280 args.fini = fini; 3281 args.align = align; 3282 args.flags = flags; 3283 args.zone = zone; 3284 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK)); 3285 } 3286 3287 3288 static void 3289 check_align_mask(unsigned int mask) 3290 { 3291 3292 KASSERT(powerof2(mask + 1), 3293 ("UMA: %s: Not the mask of a power of 2 (%#x)", __func__, mask)); 3294 /* 3295 * Make sure the stored align mask doesn't have its highest bit set, 3296 * which would cause implementation-defined behavior when passing it as 3297 * the 'align' argument of uma_zcreate(). Such very large alignments do 3298 * not make sense anyway. 3299 */ 3300 KASSERT(mask <= INT_MAX, 3301 ("UMA: %s: Mask too big (%#x)", __func__, mask)); 3302 } 3303 3304 /* Public functions */ 3305 /* See uma.h */ 3306 void 3307 uma_set_cache_align_mask(unsigned int mask) 3308 { 3309 3310 check_align_mask(mask); 3311 uma_cache_align_mask = mask; 3312 } 3313 3314 /* Returns the alignment mask to use to request cache alignment. */ 3315 unsigned int 3316 uma_get_cache_align_mask(void) 3317 { 3318 return (uma_cache_align_mask); 3319 } 3320 3321 /* See uma.h */ 3322 uma_zone_t 3323 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 3324 uma_init uminit, uma_fini fini, int align, uint32_t flags) 3325 3326 { 3327 struct uma_zctor_args args; 3328 uma_zone_t res; 3329 3330 check_align_mask(align); 3331 3332 /* This stuff is essential for the zone ctor */ 3333 memset(&args, 0, sizeof(args)); 3334 args.name = name; 3335 args.size = size; 3336 args.ctor = ctor; 3337 args.dtor = dtor; 3338 args.uminit = uminit; 3339 args.fini = fini; 3340 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN) 3341 /* 3342 * Inject procedures which check for memory use after free if we are 3343 * allowed to scramble the memory while it is not allocated. This 3344 * requires that: UMA is actually able to access the memory, no init 3345 * or fini procedures, no dependency on the initial value of the 3346 * memory, and no (legitimate) use of the memory after free. Note, 3347 * the ctor and dtor do not need to be empty. 3348 */ 3349 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH | 3350 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) { 3351 args.uminit = trash_init; 3352 args.fini = trash_fini; 3353 } 3354 #endif 3355 args.align = align; 3356 args.flags = flags; 3357 args.keg = NULL; 3358 3359 sx_xlock(&uma_reclaim_lock); 3360 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 3361 sx_xunlock(&uma_reclaim_lock); 3362 3363 return (res); 3364 } 3365 3366 /* See uma.h */ 3367 uma_zone_t 3368 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor, 3369 uma_init zinit, uma_fini zfini, uma_zone_t primary) 3370 { 3371 struct uma_zctor_args args; 3372 uma_keg_t keg; 3373 uma_zone_t res; 3374 3375 keg = primary->uz_keg; 3376 memset(&args, 0, sizeof(args)); 3377 args.name = name; 3378 args.size = keg->uk_size; 3379 args.ctor = ctor; 3380 args.dtor = dtor; 3381 args.uminit = zinit; 3382 args.fini = zfini; 3383 args.align = keg->uk_align; 3384 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 3385 args.keg = keg; 3386 3387 sx_xlock(&uma_reclaim_lock); 3388 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 3389 sx_xunlock(&uma_reclaim_lock); 3390 3391 return (res); 3392 } 3393 3394 /* See uma.h */ 3395 uma_zone_t 3396 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor, 3397 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease, 3398 void *arg, int flags) 3399 { 3400 struct uma_zctor_args args; 3401 3402 memset(&args, 0, sizeof(args)); 3403 args.name = name; 3404 args.size = size; 3405 args.ctor = ctor; 3406 args.dtor = dtor; 3407 args.uminit = zinit; 3408 args.fini = zfini; 3409 args.import = zimport; 3410 args.release = zrelease; 3411 args.arg = arg; 3412 args.align = 0; 3413 args.flags = flags | UMA_ZFLAG_CACHE; 3414 3415 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK)); 3416 } 3417 3418 /* See uma.h */ 3419 void 3420 uma_zdestroy(uma_zone_t zone) 3421 { 3422 3423 /* 3424 * Large slabs are expensive to reclaim, so don't bother doing 3425 * unnecessary work if we're shutting down. 3426 */ 3427 if (booted == BOOT_SHUTDOWN && 3428 zone->uz_fini == NULL && zone->uz_release == zone_release) 3429 return; 3430 sx_xlock(&uma_reclaim_lock); 3431 zone_free_item(zones, zone, NULL, SKIP_NONE); 3432 sx_xunlock(&uma_reclaim_lock); 3433 } 3434 3435 void 3436 uma_zwait(uma_zone_t zone) 3437 { 3438 3439 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 3440 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK)); 3441 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0) 3442 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK)); 3443 else 3444 uma_zfree(zone, uma_zalloc(zone, M_WAITOK)); 3445 } 3446 3447 void * 3448 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags) 3449 { 3450 void *item, *pcpu_item; 3451 #ifdef SMP 3452 int i; 3453 3454 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 3455 #endif 3456 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO); 3457 if (item == NULL) 3458 return (NULL); 3459 pcpu_item = zpcpu_base_to_offset(item); 3460 if (flags & M_ZERO) { 3461 #ifdef SMP 3462 for (i = 0; i <= mp_maxid; i++) 3463 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size); 3464 #else 3465 bzero(item, zone->uz_size); 3466 #endif 3467 } 3468 return (pcpu_item); 3469 } 3470 3471 /* 3472 * A stub while both regular and pcpu cases are identical. 3473 */ 3474 void 3475 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata) 3476 { 3477 void *item; 3478 3479 #ifdef SMP 3480 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 3481 #endif 3482 3483 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */ 3484 if (pcpu_item == NULL) 3485 return; 3486 3487 item = zpcpu_offset_to_base(pcpu_item); 3488 uma_zfree_arg(zone, item, udata); 3489 } 3490 3491 static inline void * 3492 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags, 3493 void *item) 3494 { 3495 #ifdef INVARIANTS 3496 bool skipdbg; 3497 #endif 3498 3499 kasan_mark_item_valid(zone, item); 3500 kmsan_mark_item_uninitialized(zone, item); 3501 3502 #ifdef INVARIANTS 3503 skipdbg = uma_dbg_zskip(zone, item); 3504 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 && 3505 zone->uz_ctor != trash_ctor) 3506 trash_ctor(item, size, zone, flags); 3507 #endif 3508 3509 /* Check flags before loading ctor pointer. */ 3510 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) && 3511 __predict_false(zone->uz_ctor != NULL) && 3512 zone->uz_ctor(item, size, udata, flags) != 0) { 3513 counter_u64_add(zone->uz_fails, 1); 3514 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT); 3515 return (NULL); 3516 } 3517 #ifdef INVARIANTS 3518 if (!skipdbg) 3519 uma_dbg_alloc(zone, NULL, item); 3520 #endif 3521 if (__predict_false(flags & M_ZERO)) 3522 return (memset(item, 0, size)); 3523 3524 return (item); 3525 } 3526 3527 static inline void 3528 item_dtor(uma_zone_t zone, void *item, int size, void *udata, 3529 enum zfreeskip skip) 3530 { 3531 #ifdef INVARIANTS 3532 bool skipdbg; 3533 3534 skipdbg = uma_dbg_zskip(zone, item); 3535 if (skip == SKIP_NONE && !skipdbg) { 3536 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0) 3537 uma_dbg_free(zone, udata, item); 3538 else 3539 uma_dbg_free(zone, NULL, item); 3540 } 3541 #endif 3542 if (__predict_true(skip < SKIP_DTOR)) { 3543 if (zone->uz_dtor != NULL) 3544 zone->uz_dtor(item, size, udata); 3545 #ifdef INVARIANTS 3546 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 && 3547 zone->uz_dtor != trash_dtor) 3548 trash_dtor(item, size, zone); 3549 #endif 3550 } 3551 kasan_mark_item_invalid(zone, item); 3552 } 3553 3554 #ifdef NUMA 3555 static int 3556 item_domain(void *item) 3557 { 3558 int domain; 3559 3560 domain = vm_phys_domain(vtophys(item)); 3561 KASSERT(domain >= 0 && domain < vm_ndomains, 3562 ("%s: unknown domain for item %p", __func__, item)); 3563 return (domain); 3564 } 3565 #endif 3566 3567 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS) 3568 #if defined(INVARIANTS) && (defined(DDB) || defined(STACK)) 3569 #include <sys/stack.h> 3570 #endif 3571 #define UMA_ZALLOC_DEBUG 3572 static int 3573 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags) 3574 { 3575 int error; 3576 3577 error = 0; 3578 #ifdef WITNESS 3579 if (flags & M_WAITOK) { 3580 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 3581 "uma_zalloc_debug: zone \"%s\"", zone->uz_name); 3582 } 3583 #endif 3584 3585 #ifdef INVARIANTS 3586 KASSERT((flags & M_EXEC) == 0, 3587 ("uma_zalloc_debug: called with M_EXEC")); 3588 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3589 ("uma_zalloc_debug: called within spinlock or critical section")); 3590 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0, 3591 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO")); 3592 3593 _Static_assert(M_NOWAIT != 0 && M_WAITOK != 0, 3594 "M_NOWAIT and M_WAITOK must be non-zero for this assertion:"); 3595 #if 0 3596 /* 3597 * Give the #elif clause time to find problems, then remove it 3598 * and enable this. (Remove <sys/stack.h> above, too.) 3599 */ 3600 KASSERT((flags & (M_NOWAIT|M_WAITOK)) == M_NOWAIT || 3601 (flags & (M_NOWAIT|M_WAITOK)) == M_WAITOK, 3602 ("uma_zalloc_debug: must pass one of M_NOWAIT or M_WAITOK")); 3603 #elif defined(DDB) || defined(STACK) 3604 if (__predict_false((flags & (M_NOWAIT|M_WAITOK)) != M_NOWAIT && 3605 (flags & (M_NOWAIT|M_WAITOK)) != M_WAITOK)) { 3606 static int stack_count; 3607 struct stack st; 3608 3609 if (stack_count < 10) { 3610 ++stack_count; 3611 printf("uma_zalloc* called with bad WAIT flags:\n"); 3612 stack_save(&st); 3613 stack_print(&st); 3614 } 3615 } 3616 #endif 3617 #endif 3618 3619 #ifdef DEBUG_MEMGUARD 3620 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 && 3621 memguard_cmp_zone(zone)) { 3622 void *item; 3623 item = memguard_alloc(zone->uz_size, flags); 3624 if (item != NULL) { 3625 error = EJUSTRETURN; 3626 if (zone->uz_init != NULL && 3627 zone->uz_init(item, zone->uz_size, flags) != 0) { 3628 *itemp = NULL; 3629 return (error); 3630 } 3631 if (zone->uz_ctor != NULL && 3632 zone->uz_ctor(item, zone->uz_size, udata, 3633 flags) != 0) { 3634 counter_u64_add(zone->uz_fails, 1); 3635 if (zone->uz_fini != NULL) 3636 zone->uz_fini(item, zone->uz_size); 3637 *itemp = NULL; 3638 return (error); 3639 } 3640 *itemp = item; 3641 return (error); 3642 } 3643 /* This is unfortunate but should not be fatal. */ 3644 } 3645 #endif 3646 return (error); 3647 } 3648 3649 static int 3650 uma_zfree_debug(uma_zone_t zone, void *item, void *udata) 3651 { 3652 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3653 ("uma_zfree_debug: called with spinlock or critical section held")); 3654 3655 #ifdef DEBUG_MEMGUARD 3656 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 && 3657 is_memguard_addr(item)) { 3658 if (zone->uz_dtor != NULL) 3659 zone->uz_dtor(item, zone->uz_size, udata); 3660 if (zone->uz_fini != NULL) 3661 zone->uz_fini(item, zone->uz_size); 3662 memguard_free(item); 3663 return (EJUSTRETURN); 3664 } 3665 #endif 3666 return (0); 3667 } 3668 #endif 3669 3670 static inline void * 3671 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket, 3672 void *udata, int flags) 3673 { 3674 void *item; 3675 int size, uz_flags; 3676 3677 item = cache_bucket_pop(cache, bucket); 3678 size = cache_uz_size(cache); 3679 uz_flags = cache_uz_flags(cache); 3680 critical_exit(); 3681 return (item_ctor(zone, uz_flags, size, udata, flags, item)); 3682 } 3683 3684 static __noinline void * 3685 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags) 3686 { 3687 uma_cache_bucket_t bucket; 3688 int domain; 3689 3690 while (cache_alloc(zone, cache, udata, flags)) { 3691 cache = &zone->uz_cpu[curcpu]; 3692 bucket = &cache->uc_allocbucket; 3693 if (__predict_false(bucket->ucb_cnt == 0)) 3694 continue; 3695 return (cache_alloc_item(zone, cache, bucket, udata, flags)); 3696 } 3697 critical_exit(); 3698 3699 /* 3700 * We can not get a bucket so try to return a single item. 3701 */ 3702 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) 3703 domain = PCPU_GET(domain); 3704 else 3705 domain = UMA_ANYDOMAIN; 3706 return (zone_alloc_item(zone, udata, domain, flags)); 3707 } 3708 3709 /* See uma.h */ 3710 void * 3711 uma_zalloc_smr(uma_zone_t zone, int flags) 3712 { 3713 uma_cache_bucket_t bucket; 3714 uma_cache_t cache; 3715 3716 CTR3(KTR_UMA, "uma_zalloc_smr zone %s(%p) flags %d", zone->uz_name, 3717 zone, flags); 3718 3719 #ifdef UMA_ZALLOC_DEBUG 3720 void *item; 3721 3722 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0, 3723 ("uma_zalloc_arg: called with non-SMR zone.")); 3724 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN) 3725 return (item); 3726 #endif 3727 3728 critical_enter(); 3729 cache = &zone->uz_cpu[curcpu]; 3730 bucket = &cache->uc_allocbucket; 3731 if (__predict_false(bucket->ucb_cnt == 0)) 3732 return (cache_alloc_retry(zone, cache, NULL, flags)); 3733 return (cache_alloc_item(zone, cache, bucket, NULL, flags)); 3734 } 3735 3736 /* See uma.h */ 3737 void * 3738 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 3739 { 3740 uma_cache_bucket_t bucket; 3741 uma_cache_t cache; 3742 3743 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3744 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3745 3746 /* This is the fast path allocation */ 3747 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name, 3748 zone, flags); 3749 3750 #ifdef UMA_ZALLOC_DEBUG 3751 void *item; 3752 3753 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0, 3754 ("uma_zalloc_arg: called with SMR zone.")); 3755 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN) 3756 return (item); 3757 #endif 3758 3759 /* 3760 * If possible, allocate from the per-CPU cache. There are two 3761 * requirements for safe access to the per-CPU cache: (1) the thread 3762 * accessing the cache must not be preempted or yield during access, 3763 * and (2) the thread must not migrate CPUs without switching which 3764 * cache it accesses. We rely on a critical section to prevent 3765 * preemption and migration. We release the critical section in 3766 * order to acquire the zone mutex if we are unable to allocate from 3767 * the current cache; when we re-acquire the critical section, we 3768 * must detect and handle migration if it has occurred. 3769 */ 3770 critical_enter(); 3771 cache = &zone->uz_cpu[curcpu]; 3772 bucket = &cache->uc_allocbucket; 3773 if (__predict_false(bucket->ucb_cnt == 0)) 3774 return (cache_alloc_retry(zone, cache, udata, flags)); 3775 return (cache_alloc_item(zone, cache, bucket, udata, flags)); 3776 } 3777 3778 /* 3779 * Replenish an alloc bucket and possibly restore an old one. Called in 3780 * a critical section. Returns in a critical section. 3781 * 3782 * A false return value indicates an allocation failure. 3783 * A true return value indicates success and the caller should retry. 3784 */ 3785 static __noinline bool 3786 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags) 3787 { 3788 uma_bucket_t bucket; 3789 int curdomain, domain; 3790 bool new; 3791 3792 CRITICAL_ASSERT(curthread); 3793 3794 /* 3795 * If we have run out of items in our alloc bucket see 3796 * if we can switch with the free bucket. 3797 * 3798 * SMR Zones can't re-use the free bucket until the sequence has 3799 * expired. 3800 */ 3801 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 && 3802 cache->uc_freebucket.ucb_cnt != 0) { 3803 cache_bucket_swap(&cache->uc_freebucket, 3804 &cache->uc_allocbucket); 3805 return (true); 3806 } 3807 3808 /* 3809 * Discard any empty allocation bucket while we hold no locks. 3810 */ 3811 bucket = cache_bucket_unload_alloc(cache); 3812 critical_exit(); 3813 3814 if (bucket != NULL) { 3815 KASSERT(bucket->ub_cnt == 0, 3816 ("cache_alloc: Entered with non-empty alloc bucket.")); 3817 bucket_free(zone, bucket, udata); 3818 } 3819 3820 /* 3821 * Attempt to retrieve the item from the per-CPU cache has failed, so 3822 * we must go back to the zone. This requires the zdom lock, so we 3823 * must drop the critical section, then re-acquire it when we go back 3824 * to the cache. Since the critical section is released, we may be 3825 * preempted or migrate. As such, make sure not to maintain any 3826 * thread-local state specific to the cache from prior to releasing 3827 * the critical section. 3828 */ 3829 domain = PCPU_GET(domain); 3830 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 || 3831 VM_DOMAIN_EMPTY(domain)) 3832 domain = zone_domain_highest(zone, domain); 3833 bucket = cache_fetch_bucket(zone, cache, domain); 3834 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) { 3835 bucket = zone_alloc_bucket(zone, udata, domain, flags); 3836 new = true; 3837 } else { 3838 new = false; 3839 } 3840 3841 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p", 3842 zone->uz_name, zone, bucket); 3843 if (bucket == NULL) { 3844 critical_enter(); 3845 return (false); 3846 } 3847 3848 /* 3849 * See if we lost the race or were migrated. Cache the 3850 * initialized bucket to make this less likely or claim 3851 * the memory directly. 3852 */ 3853 critical_enter(); 3854 cache = &zone->uz_cpu[curcpu]; 3855 if (cache->uc_allocbucket.ucb_bucket == NULL && 3856 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 || 3857 (curdomain = PCPU_GET(domain)) == domain || 3858 VM_DOMAIN_EMPTY(curdomain))) { 3859 if (new) 3860 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax, 3861 bucket->ub_cnt); 3862 cache_bucket_load_alloc(cache, bucket); 3863 return (true); 3864 } 3865 3866 /* 3867 * We lost the race, release this bucket and start over. 3868 */ 3869 critical_exit(); 3870 zone_put_bucket(zone, domain, bucket, udata, !new); 3871 critical_enter(); 3872 3873 return (true); 3874 } 3875 3876 void * 3877 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags) 3878 { 3879 #ifdef NUMA 3880 uma_bucket_t bucket; 3881 uma_zone_domain_t zdom; 3882 void *item; 3883 #endif 3884 3885 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3886 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3887 3888 /* This is the fast path allocation */ 3889 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d", 3890 zone->uz_name, zone, domain, flags); 3891 3892 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0, 3893 ("uma_zalloc_domain: called with SMR zone.")); 3894 #ifdef NUMA 3895 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0, 3896 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone.")); 3897 3898 if (vm_ndomains == 1) 3899 return (uma_zalloc_arg(zone, udata, flags)); 3900 3901 #ifdef UMA_ZALLOC_DEBUG 3902 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN) 3903 return (item); 3904 #endif 3905 3906 /* 3907 * Try to allocate from the bucket cache before falling back to the keg. 3908 * We could try harder and attempt to allocate from per-CPU caches or 3909 * the per-domain cross-domain buckets, but the complexity is probably 3910 * not worth it. It is more important that frees of previous 3911 * cross-domain allocations do not blow up the cache. 3912 */ 3913 zdom = zone_domain_lock(zone, domain); 3914 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) { 3915 item = bucket->ub_bucket[bucket->ub_cnt - 1]; 3916 #ifdef INVARIANTS 3917 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL; 3918 #endif 3919 bucket->ub_cnt--; 3920 zone_put_bucket(zone, domain, bucket, udata, true); 3921 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, 3922 flags, item); 3923 if (item != NULL) { 3924 KASSERT(item_domain(item) == domain, 3925 ("%s: bucket cache item %p from wrong domain", 3926 __func__, item)); 3927 counter_u64_add(zone->uz_allocs, 1); 3928 } 3929 return (item); 3930 } 3931 ZDOM_UNLOCK(zdom); 3932 return (zone_alloc_item(zone, udata, domain, flags)); 3933 #else 3934 return (uma_zalloc_arg(zone, udata, flags)); 3935 #endif 3936 } 3937 3938 /* 3939 * Find a slab with some space. Prefer slabs that are partially used over those 3940 * that are totally full. This helps to reduce fragmentation. 3941 * 3942 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check 3943 * only 'domain'. 3944 */ 3945 static uma_slab_t 3946 keg_first_slab(uma_keg_t keg, int domain, bool rr) 3947 { 3948 uma_domain_t dom; 3949 uma_slab_t slab; 3950 int start; 3951 3952 KASSERT(domain >= 0 && domain < vm_ndomains, 3953 ("keg_first_slab: domain %d out of range", domain)); 3954 KEG_LOCK_ASSERT(keg, domain); 3955 3956 slab = NULL; 3957 start = domain; 3958 do { 3959 dom = &keg->uk_domain[domain]; 3960 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL) 3961 return (slab); 3962 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) { 3963 LIST_REMOVE(slab, us_link); 3964 dom->ud_free_slabs--; 3965 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 3966 return (slab); 3967 } 3968 if (rr) 3969 domain = (domain + 1) % vm_ndomains; 3970 } while (domain != start); 3971 3972 return (NULL); 3973 } 3974 3975 /* 3976 * Fetch an existing slab from a free or partial list. Returns with the 3977 * keg domain lock held if a slab was found or unlocked if not. 3978 */ 3979 static uma_slab_t 3980 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags) 3981 { 3982 uma_slab_t slab; 3983 uint32_t reserve; 3984 3985 /* HASH has a single free list. */ 3986 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) 3987 domain = 0; 3988 3989 KEG_LOCK(keg, domain); 3990 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve; 3991 if (keg->uk_domain[domain].ud_free_items <= reserve || 3992 (slab = keg_first_slab(keg, domain, rr)) == NULL) { 3993 KEG_UNLOCK(keg, domain); 3994 return (NULL); 3995 } 3996 return (slab); 3997 } 3998 3999 static uma_slab_t 4000 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags) 4001 { 4002 struct vm_domainset_iter di; 4003 uma_slab_t slab; 4004 int aflags, domain; 4005 bool rr; 4006 4007 KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM), 4008 ("%s: invalid flags %#x", __func__, flags)); 4009 4010 restart: 4011 /* 4012 * Use the keg's policy if upper layers haven't already specified a 4013 * domain (as happens with first-touch zones). 4014 * 4015 * To avoid races we run the iterator with the keg lock held, but that 4016 * means that we cannot allow the vm_domainset layer to sleep. Thus, 4017 * clear M_WAITOK and handle low memory conditions locally. 4018 */ 4019 rr = rdomain == UMA_ANYDOMAIN; 4020 if (rr) { 4021 aflags = (flags & ~M_WAITOK) | M_NOWAIT; 4022 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 4023 &aflags); 4024 } else { 4025 aflags = flags; 4026 domain = rdomain; 4027 } 4028 4029 for (;;) { 4030 slab = keg_fetch_free_slab(keg, domain, rr, flags); 4031 if (slab != NULL) 4032 return (slab); 4033 4034 /* 4035 * M_NOVM is used to break the recursion that can otherwise 4036 * occur if low-level memory management routines use UMA. 4037 */ 4038 if ((flags & M_NOVM) == 0) { 4039 slab = keg_alloc_slab(keg, zone, domain, flags, aflags); 4040 if (slab != NULL) 4041 return (slab); 4042 } 4043 4044 if (!rr) { 4045 if ((flags & M_USE_RESERVE) != 0) { 4046 /* 4047 * Drain reserves from other domains before 4048 * giving up or sleeping. It may be useful to 4049 * support per-domain reserves eventually. 4050 */ 4051 rdomain = UMA_ANYDOMAIN; 4052 goto restart; 4053 } 4054 if ((flags & M_WAITOK) == 0) 4055 break; 4056 vm_wait_domain(domain); 4057 } else if (vm_domainset_iter_policy(&di, &domain) != 0) { 4058 if ((flags & M_WAITOK) != 0) { 4059 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0); 4060 goto restart; 4061 } 4062 break; 4063 } 4064 } 4065 4066 /* 4067 * We might not have been able to get a slab but another cpu 4068 * could have while we were unlocked. Check again before we 4069 * fail. 4070 */ 4071 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL) 4072 return (slab); 4073 4074 return (NULL); 4075 } 4076 4077 static void * 4078 slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 4079 { 4080 uma_domain_t dom; 4081 void *item; 4082 int freei; 4083 4084 KEG_LOCK_ASSERT(keg, slab->us_domain); 4085 4086 dom = &keg->uk_domain[slab->us_domain]; 4087 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1; 4088 BIT_CLR(keg->uk_ipers, freei, &slab->us_free); 4089 item = slab_item(slab, keg, freei); 4090 slab->us_freecount--; 4091 dom->ud_free_items--; 4092 4093 /* 4094 * Move this slab to the full list. It must be on the partial list, so 4095 * we do not need to update the free slab count. In particular, 4096 * keg_fetch_slab() always returns slabs on the partial list. 4097 */ 4098 if (slab->us_freecount == 0) { 4099 LIST_REMOVE(slab, us_link); 4100 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link); 4101 } 4102 4103 return (item); 4104 } 4105 4106 static int 4107 zone_import(void *arg, void **bucket, int max, int domain, int flags) 4108 { 4109 uma_domain_t dom; 4110 uma_zone_t zone; 4111 uma_slab_t slab; 4112 uma_keg_t keg; 4113 #ifdef NUMA 4114 int stripe; 4115 #endif 4116 int i; 4117 4118 zone = arg; 4119 slab = NULL; 4120 keg = zone->uz_keg; 4121 /* Try to keep the buckets totally full */ 4122 for (i = 0; i < max; ) { 4123 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL) 4124 break; 4125 #ifdef NUMA 4126 stripe = howmany(max, vm_ndomains); 4127 #endif 4128 dom = &keg->uk_domain[slab->us_domain]; 4129 do { 4130 bucket[i++] = slab_alloc_item(keg, slab); 4131 if (keg->uk_reserve > 0 && 4132 dom->ud_free_items <= keg->uk_reserve) { 4133 /* 4134 * Avoid depleting the reserve after a 4135 * successful item allocation, even if 4136 * M_USE_RESERVE is specified. 4137 */ 4138 KEG_UNLOCK(keg, slab->us_domain); 4139 goto out; 4140 } 4141 #ifdef NUMA 4142 /* 4143 * If the zone is striped we pick a new slab for every 4144 * N allocations. Eliminating this conditional will 4145 * instead pick a new domain for each bucket rather 4146 * than stripe within each bucket. The current option 4147 * produces more fragmentation and requires more cpu 4148 * time but yields better distribution. 4149 */ 4150 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 && 4151 vm_ndomains > 1 && --stripe == 0) 4152 break; 4153 #endif 4154 } while (slab->us_freecount != 0 && i < max); 4155 KEG_UNLOCK(keg, slab->us_domain); 4156 4157 /* Don't block if we allocated any successfully. */ 4158 flags &= ~M_WAITOK; 4159 flags |= M_NOWAIT; 4160 } 4161 out: 4162 return i; 4163 } 4164 4165 static int 4166 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags) 4167 { 4168 uint64_t old, new, total, max; 4169 4170 /* 4171 * The hard case. We're going to sleep because there were existing 4172 * sleepers or because we ran out of items. This routine enforces 4173 * fairness by keeping fifo order. 4174 * 4175 * First release our ill gotten gains and make some noise. 4176 */ 4177 for (;;) { 4178 zone_free_limit(zone, count); 4179 zone_log_warning(zone); 4180 zone_maxaction(zone); 4181 if (flags & M_NOWAIT) 4182 return (0); 4183 4184 /* 4185 * We need to allocate an item or set ourself as a sleeper 4186 * while the sleepq lock is held to avoid wakeup races. This 4187 * is essentially a home rolled semaphore. 4188 */ 4189 sleepq_lock(&zone->uz_max_items); 4190 old = zone->uz_items; 4191 do { 4192 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX); 4193 /* Cache the max since we will evaluate twice. */ 4194 max = zone->uz_max_items; 4195 if (UZ_ITEMS_SLEEPERS(old) != 0 || 4196 UZ_ITEMS_COUNT(old) >= max) 4197 new = old + UZ_ITEMS_SLEEPER; 4198 else 4199 new = old + MIN(count, max - old); 4200 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0); 4201 4202 /* We may have successfully allocated under the sleepq lock. */ 4203 if (UZ_ITEMS_SLEEPERS(new) == 0) { 4204 sleepq_release(&zone->uz_max_items); 4205 return (new - old); 4206 } 4207 4208 /* 4209 * This is in a different cacheline from uz_items so that we 4210 * don't constantly invalidate the fastpath cacheline when we 4211 * adjust item counts. This could be limited to toggling on 4212 * transitions. 4213 */ 4214 atomic_add_32(&zone->uz_sleepers, 1); 4215 atomic_add_64(&zone->uz_sleeps, 1); 4216 4217 /* 4218 * We have added ourselves as a sleeper. The sleepq lock 4219 * protects us from wakeup races. Sleep now and then retry. 4220 */ 4221 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0); 4222 sleepq_wait(&zone->uz_max_items, PVM); 4223 4224 /* 4225 * After wakeup, remove ourselves as a sleeper and try 4226 * again. We no longer have the sleepq lock for protection. 4227 * 4228 * Subract ourselves as a sleeper while attempting to add 4229 * our count. 4230 */ 4231 atomic_subtract_32(&zone->uz_sleepers, 1); 4232 old = atomic_fetchadd_64(&zone->uz_items, 4233 -(UZ_ITEMS_SLEEPER - count)); 4234 /* We're no longer a sleeper. */ 4235 old -= UZ_ITEMS_SLEEPER; 4236 4237 /* 4238 * If we're still at the limit, restart. Notably do not 4239 * block on other sleepers. Cache the max value to protect 4240 * against changes via sysctl. 4241 */ 4242 total = UZ_ITEMS_COUNT(old); 4243 max = zone->uz_max_items; 4244 if (total >= max) 4245 continue; 4246 /* Truncate if necessary, otherwise wake other sleepers. */ 4247 if (total + count > max) { 4248 zone_free_limit(zone, total + count - max); 4249 count = max - total; 4250 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0) 4251 wakeup_one(&zone->uz_max_items); 4252 4253 return (count); 4254 } 4255 } 4256 4257 /* 4258 * Allocate 'count' items from our max_items limit. Returns the number 4259 * available. If M_NOWAIT is not specified it will sleep until at least 4260 * one item can be allocated. 4261 */ 4262 static int 4263 zone_alloc_limit(uma_zone_t zone, int count, int flags) 4264 { 4265 uint64_t old; 4266 uint64_t max; 4267 4268 max = zone->uz_max_items; 4269 MPASS(max > 0); 4270 4271 /* 4272 * We expect normal allocations to succeed with a simple 4273 * fetchadd. 4274 */ 4275 old = atomic_fetchadd_64(&zone->uz_items, count); 4276 if (__predict_true(old + count <= max)) 4277 return (count); 4278 4279 /* 4280 * If we had some items and no sleepers just return the 4281 * truncated value. We have to release the excess space 4282 * though because that may wake sleepers who weren't woken 4283 * because we were temporarily over the limit. 4284 */ 4285 if (old < max) { 4286 zone_free_limit(zone, (old + count) - max); 4287 return (max - old); 4288 } 4289 return (zone_alloc_limit_hard(zone, count, flags)); 4290 } 4291 4292 /* 4293 * Free a number of items back to the limit. 4294 */ 4295 static void 4296 zone_free_limit(uma_zone_t zone, int count) 4297 { 4298 uint64_t old; 4299 4300 MPASS(count > 0); 4301 4302 /* 4303 * In the common case we either have no sleepers or 4304 * are still over the limit and can just return. 4305 */ 4306 old = atomic_fetchadd_64(&zone->uz_items, -count); 4307 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 || 4308 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items)) 4309 return; 4310 4311 /* 4312 * Moderate the rate of wakeups. Sleepers will continue 4313 * to generate wakeups if necessary. 4314 */ 4315 wakeup_one(&zone->uz_max_items); 4316 } 4317 4318 static uma_bucket_t 4319 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags) 4320 { 4321 uma_bucket_t bucket; 4322 int error, maxbucket, cnt; 4323 4324 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name, 4325 zone, domain); 4326 4327 /* Avoid allocs targeting empty domains. */ 4328 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain)) 4329 domain = UMA_ANYDOMAIN; 4330 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0) 4331 domain = UMA_ANYDOMAIN; 4332 4333 if (zone->uz_max_items > 0) 4334 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size, 4335 M_NOWAIT); 4336 else 4337 maxbucket = zone->uz_bucket_size; 4338 if (maxbucket == 0) 4339 return (NULL); 4340 4341 /* Don't wait for buckets, preserve caller's NOVM setting. */ 4342 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 4343 if (bucket == NULL) { 4344 cnt = 0; 4345 goto out; 4346 } 4347 4348 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 4349 MIN(maxbucket, bucket->ub_entries), domain, flags); 4350 4351 /* 4352 * Initialize the memory if necessary. 4353 */ 4354 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 4355 int i; 4356 4357 for (i = 0; i < bucket->ub_cnt; i++) { 4358 kasan_mark_item_valid(zone, bucket->ub_bucket[i]); 4359 error = zone->uz_init(bucket->ub_bucket[i], 4360 zone->uz_size, flags); 4361 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]); 4362 if (error != 0) 4363 break; 4364 } 4365 4366 /* 4367 * If we couldn't initialize the whole bucket, put the 4368 * rest back onto the freelist. 4369 */ 4370 if (i != bucket->ub_cnt) { 4371 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 4372 bucket->ub_cnt - i); 4373 #ifdef INVARIANTS 4374 bzero(&bucket->ub_bucket[i], 4375 sizeof(void *) * (bucket->ub_cnt - i)); 4376 #endif 4377 bucket->ub_cnt = i; 4378 } 4379 } 4380 4381 cnt = bucket->ub_cnt; 4382 if (bucket->ub_cnt == 0) { 4383 bucket_free(zone, bucket, udata); 4384 counter_u64_add(zone->uz_fails, 1); 4385 bucket = NULL; 4386 } 4387 out: 4388 if (zone->uz_max_items > 0 && cnt < maxbucket) 4389 zone_free_limit(zone, maxbucket - cnt); 4390 4391 return (bucket); 4392 } 4393 4394 /* 4395 * Allocates a single item from a zone. 4396 * 4397 * Arguments 4398 * zone The zone to alloc for. 4399 * udata The data to be passed to the constructor. 4400 * domain The domain to allocate from or UMA_ANYDOMAIN. 4401 * flags M_WAITOK, M_NOWAIT, M_ZERO. 4402 * 4403 * Returns 4404 * NULL if there is no memory and M_NOWAIT is set 4405 * An item if successful 4406 */ 4407 4408 static void * 4409 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags) 4410 { 4411 void *item; 4412 4413 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) { 4414 counter_u64_add(zone->uz_fails, 1); 4415 return (NULL); 4416 } 4417 4418 /* Avoid allocs targeting empty domains. */ 4419 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain)) 4420 domain = UMA_ANYDOMAIN; 4421 4422 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1) 4423 goto fail_cnt; 4424 4425 /* 4426 * We have to call both the zone's init (not the keg's init) 4427 * and the zone's ctor. This is because the item is going from 4428 * a keg slab directly to the user, and the user is expecting it 4429 * to be both zone-init'd as well as zone-ctor'd. 4430 */ 4431 if (zone->uz_init != NULL) { 4432 int error; 4433 4434 kasan_mark_item_valid(zone, item); 4435 error = zone->uz_init(item, zone->uz_size, flags); 4436 kasan_mark_item_invalid(zone, item); 4437 if (error != 0) { 4438 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT); 4439 goto fail_cnt; 4440 } 4441 } 4442 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags, 4443 item); 4444 if (item == NULL) 4445 goto fail; 4446 4447 counter_u64_add(zone->uz_allocs, 1); 4448 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item, 4449 zone->uz_name, zone); 4450 4451 return (item); 4452 4453 fail_cnt: 4454 counter_u64_add(zone->uz_fails, 1); 4455 fail: 4456 if (zone->uz_max_items > 0) 4457 zone_free_limit(zone, 1); 4458 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)", 4459 zone->uz_name, zone); 4460 4461 return (NULL); 4462 } 4463 4464 /* See uma.h */ 4465 void 4466 uma_zfree_smr(uma_zone_t zone, void *item) 4467 { 4468 uma_cache_t cache; 4469 uma_cache_bucket_t bucket; 4470 int itemdomain; 4471 #ifdef NUMA 4472 int uz_flags; 4473 #endif 4474 4475 CTR3(KTR_UMA, "uma_zfree_smr zone %s(%p) item %p", 4476 zone->uz_name, zone, item); 4477 4478 #ifdef UMA_ZALLOC_DEBUG 4479 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0, 4480 ("uma_zfree_smr: called with non-SMR zone.")); 4481 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer.")); 4482 SMR_ASSERT_NOT_ENTERED(zone->uz_smr); 4483 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN) 4484 return; 4485 #endif 4486 cache = &zone->uz_cpu[curcpu]; 4487 itemdomain = 0; 4488 #ifdef NUMA 4489 uz_flags = cache_uz_flags(cache); 4490 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 4491 itemdomain = item_domain(item); 4492 #endif 4493 critical_enter(); 4494 do { 4495 cache = &zone->uz_cpu[curcpu]; 4496 /* SMR Zones must free to the free bucket. */ 4497 bucket = &cache->uc_freebucket; 4498 #ifdef NUMA 4499 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && 4500 PCPU_GET(domain) != itemdomain) { 4501 bucket = &cache->uc_crossbucket; 4502 } 4503 #endif 4504 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) { 4505 cache_bucket_push(cache, bucket, item); 4506 critical_exit(); 4507 return; 4508 } 4509 } while (cache_free(zone, cache, NULL, itemdomain)); 4510 critical_exit(); 4511 4512 /* 4513 * If nothing else caught this, we'll just do an internal free. 4514 */ 4515 zone_free_item(zone, item, NULL, SKIP_NONE); 4516 } 4517 4518 /* See uma.h */ 4519 void 4520 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 4521 { 4522 uma_cache_t cache; 4523 uma_cache_bucket_t bucket; 4524 int itemdomain, uz_flags; 4525 4526 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 4527 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 4528 4529 CTR3(KTR_UMA, "uma_zfree_arg zone %s(%p) item %p", 4530 zone->uz_name, zone, item); 4531 4532 #ifdef UMA_ZALLOC_DEBUG 4533 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0, 4534 ("uma_zfree_arg: called with SMR zone.")); 4535 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN) 4536 return; 4537 #endif 4538 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 4539 if (item == NULL) 4540 return; 4541 4542 /* 4543 * We are accessing the per-cpu cache without a critical section to 4544 * fetch size and flags. This is acceptable, if we are preempted we 4545 * will simply read another cpu's line. 4546 */ 4547 cache = &zone->uz_cpu[curcpu]; 4548 uz_flags = cache_uz_flags(cache); 4549 if (UMA_ALWAYS_CTORDTOR || 4550 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0)) 4551 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE); 4552 4553 /* 4554 * The race here is acceptable. If we miss it we'll just have to wait 4555 * a little longer for the limits to be reset. 4556 */ 4557 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) { 4558 if (atomic_load_32(&zone->uz_sleepers) > 0) 4559 goto zfree_item; 4560 } 4561 4562 /* 4563 * If possible, free to the per-CPU cache. There are two 4564 * requirements for safe access to the per-CPU cache: (1) the thread 4565 * accessing the cache must not be preempted or yield during access, 4566 * and (2) the thread must not migrate CPUs without switching which 4567 * cache it accesses. We rely on a critical section to prevent 4568 * preemption and migration. We release the critical section in 4569 * order to acquire the zone mutex if we are unable to free to the 4570 * current cache; when we re-acquire the critical section, we must 4571 * detect and handle migration if it has occurred. 4572 */ 4573 itemdomain = 0; 4574 #ifdef NUMA 4575 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 4576 itemdomain = item_domain(item); 4577 #endif 4578 critical_enter(); 4579 do { 4580 cache = &zone->uz_cpu[curcpu]; 4581 /* 4582 * Try to free into the allocbucket first to give LIFO 4583 * ordering for cache-hot datastructures. Spill over 4584 * into the freebucket if necessary. Alloc will swap 4585 * them if one runs dry. 4586 */ 4587 bucket = &cache->uc_allocbucket; 4588 #ifdef NUMA 4589 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && 4590 PCPU_GET(domain) != itemdomain) { 4591 bucket = &cache->uc_crossbucket; 4592 } else 4593 #endif 4594 if (bucket->ucb_cnt == bucket->ucb_entries && 4595 cache->uc_freebucket.ucb_cnt < 4596 cache->uc_freebucket.ucb_entries) 4597 cache_bucket_swap(&cache->uc_freebucket, 4598 &cache->uc_allocbucket); 4599 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) { 4600 cache_bucket_push(cache, bucket, item); 4601 critical_exit(); 4602 return; 4603 } 4604 } while (cache_free(zone, cache, udata, itemdomain)); 4605 critical_exit(); 4606 4607 /* 4608 * If nothing else caught this, we'll just do an internal free. 4609 */ 4610 zfree_item: 4611 zone_free_item(zone, item, udata, SKIP_DTOR); 4612 } 4613 4614 #ifdef NUMA 4615 /* 4616 * sort crossdomain free buckets to domain correct buckets and cache 4617 * them. 4618 */ 4619 static void 4620 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata) 4621 { 4622 struct uma_bucketlist emptybuckets, fullbuckets; 4623 uma_zone_domain_t zdom; 4624 uma_bucket_t b; 4625 smr_seq_t seq; 4626 void *item; 4627 int domain; 4628 4629 CTR3(KTR_UMA, 4630 "uma_zfree: zone %s(%p) draining cross bucket %p", 4631 zone->uz_name, zone, bucket); 4632 4633 /* 4634 * It is possible for buckets to arrive here out of order so we fetch 4635 * the current smr seq rather than accepting the bucket's. 4636 */ 4637 seq = SMR_SEQ_INVALID; 4638 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 4639 seq = smr_advance(zone->uz_smr); 4640 4641 /* 4642 * To avoid having ndomain * ndomain buckets for sorting we have a 4643 * lock on the current crossfree bucket. A full matrix with 4644 * per-domain locking could be used if necessary. 4645 */ 4646 STAILQ_INIT(&emptybuckets); 4647 STAILQ_INIT(&fullbuckets); 4648 ZONE_CROSS_LOCK(zone); 4649 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) { 4650 item = bucket->ub_bucket[bucket->ub_cnt - 1]; 4651 domain = item_domain(item); 4652 zdom = ZDOM_GET(zone, domain); 4653 if (zdom->uzd_cross == NULL) { 4654 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) { 4655 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link); 4656 zdom->uzd_cross = b; 4657 } else { 4658 /* 4659 * Avoid allocating a bucket with the cross lock 4660 * held, since allocation can trigger a 4661 * cross-domain free and bucket zones may 4662 * allocate from each other. 4663 */ 4664 ZONE_CROSS_UNLOCK(zone); 4665 b = bucket_alloc(zone, udata, M_NOWAIT); 4666 if (b == NULL) 4667 goto out; 4668 ZONE_CROSS_LOCK(zone); 4669 if (zdom->uzd_cross != NULL) { 4670 STAILQ_INSERT_HEAD(&emptybuckets, b, 4671 ub_link); 4672 } else { 4673 zdom->uzd_cross = b; 4674 } 4675 } 4676 } 4677 b = zdom->uzd_cross; 4678 b->ub_bucket[b->ub_cnt++] = item; 4679 b->ub_seq = seq; 4680 if (b->ub_cnt == b->ub_entries) { 4681 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link); 4682 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) 4683 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link); 4684 zdom->uzd_cross = b; 4685 } 4686 } 4687 ZONE_CROSS_UNLOCK(zone); 4688 out: 4689 if (bucket->ub_cnt == 0) 4690 bucket->ub_seq = SMR_SEQ_INVALID; 4691 bucket_free(zone, bucket, udata); 4692 4693 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) { 4694 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link); 4695 bucket_free(zone, b, udata); 4696 } 4697 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) { 4698 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link); 4699 domain = item_domain(b->ub_bucket[0]); 4700 zone_put_bucket(zone, domain, b, udata, true); 4701 } 4702 } 4703 #endif 4704 4705 static void 4706 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata, 4707 int itemdomain, bool ws) 4708 { 4709 4710 #ifdef NUMA 4711 /* 4712 * Buckets coming from the wrong domain will be entirely for the 4713 * only other domain on two domain systems. In this case we can 4714 * simply cache them. Otherwise we need to sort them back to 4715 * correct domains. 4716 */ 4717 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && 4718 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) { 4719 zone_free_cross(zone, bucket, udata); 4720 return; 4721 } 4722 #endif 4723 4724 /* 4725 * Attempt to save the bucket in the zone's domain bucket cache. 4726 */ 4727 CTR3(KTR_UMA, 4728 "uma_zfree: zone %s(%p) putting bucket %p on free list", 4729 zone->uz_name, zone, bucket); 4730 /* ub_cnt is pointing to the last free item */ 4731 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0) 4732 itemdomain = zone_domain_lowest(zone, itemdomain); 4733 zone_put_bucket(zone, itemdomain, bucket, udata, ws); 4734 } 4735 4736 /* 4737 * Populate a free or cross bucket for the current cpu cache. Free any 4738 * existing full bucket either to the zone cache or back to the slab layer. 4739 * 4740 * Enters and returns in a critical section. false return indicates that 4741 * we can not satisfy this free in the cache layer. true indicates that 4742 * the caller should retry. 4743 */ 4744 static __noinline bool 4745 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, int itemdomain) 4746 { 4747 uma_cache_bucket_t cbucket; 4748 uma_bucket_t newbucket, bucket; 4749 4750 CRITICAL_ASSERT(curthread); 4751 4752 if (zone->uz_bucket_size == 0) 4753 return false; 4754 4755 cache = &zone->uz_cpu[curcpu]; 4756 newbucket = NULL; 4757 4758 /* 4759 * FIRSTTOUCH domains need to free to the correct zdom. When 4760 * enabled this is the zdom of the item. The bucket is the 4761 * cross bucket if the current domain and itemdomain do not match. 4762 */ 4763 cbucket = &cache->uc_freebucket; 4764 #ifdef NUMA 4765 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) { 4766 if (PCPU_GET(domain) != itemdomain) { 4767 cbucket = &cache->uc_crossbucket; 4768 if (cbucket->ucb_cnt != 0) 4769 counter_u64_add(zone->uz_xdomain, 4770 cbucket->ucb_cnt); 4771 } 4772 } 4773 #endif 4774 bucket = cache_bucket_unload(cbucket); 4775 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries, 4776 ("cache_free: Entered with non-full free bucket.")); 4777 4778 /* We are no longer associated with this CPU. */ 4779 critical_exit(); 4780 4781 /* 4782 * Don't let SMR zones operate without a free bucket. Force 4783 * a synchronize and re-use this one. We will only degrade 4784 * to a synchronize every bucket_size items rather than every 4785 * item if we fail to allocate a bucket. 4786 */ 4787 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) { 4788 if (bucket != NULL) 4789 bucket->ub_seq = smr_advance(zone->uz_smr); 4790 newbucket = bucket_alloc(zone, udata, M_NOWAIT); 4791 if (newbucket == NULL && bucket != NULL) { 4792 bucket_drain(zone, bucket); 4793 newbucket = bucket; 4794 bucket = NULL; 4795 } 4796 } else if (!bucketdisable) 4797 newbucket = bucket_alloc(zone, udata, M_NOWAIT); 4798 4799 if (bucket != NULL) 4800 zone_free_bucket(zone, bucket, udata, itemdomain, true); 4801 4802 critical_enter(); 4803 if ((bucket = newbucket) == NULL) 4804 return (false); 4805 cache = &zone->uz_cpu[curcpu]; 4806 #ifdef NUMA 4807 /* 4808 * Check to see if we should be populating the cross bucket. If it 4809 * is already populated we will fall through and attempt to populate 4810 * the free bucket. 4811 */ 4812 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) { 4813 if (PCPU_GET(domain) != itemdomain && 4814 cache->uc_crossbucket.ucb_bucket == NULL) { 4815 cache_bucket_load_cross(cache, bucket); 4816 return (true); 4817 } 4818 } 4819 #endif 4820 /* 4821 * We may have lost the race to fill the bucket or switched CPUs. 4822 */ 4823 if (cache->uc_freebucket.ucb_bucket != NULL) { 4824 critical_exit(); 4825 bucket_free(zone, bucket, udata); 4826 critical_enter(); 4827 } else 4828 cache_bucket_load_free(cache, bucket); 4829 4830 return (true); 4831 } 4832 4833 static void 4834 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item) 4835 { 4836 uma_keg_t keg; 4837 uma_domain_t dom; 4838 int freei; 4839 4840 keg = zone->uz_keg; 4841 KEG_LOCK_ASSERT(keg, slab->us_domain); 4842 4843 /* Do we need to remove from any lists? */ 4844 dom = &keg->uk_domain[slab->us_domain]; 4845 if (slab->us_freecount + 1 == keg->uk_ipers) { 4846 LIST_REMOVE(slab, us_link); 4847 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); 4848 dom->ud_free_slabs++; 4849 } else if (slab->us_freecount == 0) { 4850 LIST_REMOVE(slab, us_link); 4851 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 4852 } 4853 4854 /* Slab management. */ 4855 freei = slab_item_index(slab, keg, item); 4856 BIT_SET(keg->uk_ipers, freei, &slab->us_free); 4857 slab->us_freecount++; 4858 4859 /* Keg statistics. */ 4860 dom->ud_free_items++; 4861 } 4862 4863 static void 4864 zone_release(void *arg, void **bucket, int cnt) 4865 { 4866 struct mtx *lock; 4867 uma_zone_t zone; 4868 uma_slab_t slab; 4869 uma_keg_t keg; 4870 uint8_t *mem; 4871 void *item; 4872 int i; 4873 4874 zone = arg; 4875 keg = zone->uz_keg; 4876 lock = NULL; 4877 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0)) 4878 lock = KEG_LOCK(keg, 0); 4879 for (i = 0; i < cnt; i++) { 4880 item = bucket[i]; 4881 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) { 4882 slab = vtoslab((vm_offset_t)item); 4883 } else { 4884 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 4885 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0) 4886 slab = hash_sfind(&keg->uk_hash, mem); 4887 else 4888 slab = (uma_slab_t)(mem + keg->uk_pgoff); 4889 } 4890 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) { 4891 if (lock != NULL) 4892 mtx_unlock(lock); 4893 lock = KEG_LOCK(keg, slab->us_domain); 4894 } 4895 slab_free_item(zone, slab, item); 4896 } 4897 if (lock != NULL) 4898 mtx_unlock(lock); 4899 } 4900 4901 /* 4902 * Frees a single item to any zone. 4903 * 4904 * Arguments: 4905 * zone The zone to free to 4906 * item The item we're freeing 4907 * udata User supplied data for the dtor 4908 * skip Skip dtors and finis 4909 */ 4910 static __noinline void 4911 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 4912 { 4913 4914 /* 4915 * If a free is sent directly to an SMR zone we have to 4916 * synchronize immediately because the item can instantly 4917 * be reallocated. This should only happen in degenerate 4918 * cases when no memory is available for per-cpu caches. 4919 */ 4920 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE) 4921 smr_synchronize(zone->uz_smr); 4922 4923 item_dtor(zone, item, zone->uz_size, udata, skip); 4924 4925 if (skip < SKIP_FINI && zone->uz_fini) { 4926 kasan_mark_item_valid(zone, item); 4927 zone->uz_fini(item, zone->uz_size); 4928 kasan_mark_item_invalid(zone, item); 4929 } 4930 4931 zone->uz_release(zone->uz_arg, &item, 1); 4932 4933 if (skip & SKIP_CNT) 4934 return; 4935 4936 counter_u64_add(zone->uz_frees, 1); 4937 4938 if (zone->uz_max_items > 0) 4939 zone_free_limit(zone, 1); 4940 } 4941 4942 /* See uma.h */ 4943 int 4944 uma_zone_set_max(uma_zone_t zone, int nitems) 4945 { 4946 4947 /* 4948 * If the limit is small, we may need to constrain the maximum per-CPU 4949 * cache size, or disable caching entirely. 4950 */ 4951 uma_zone_set_maxcache(zone, nitems); 4952 4953 /* 4954 * XXX This can misbehave if the zone has any allocations with 4955 * no limit and a limit is imposed. There is currently no 4956 * way to clear a limit. 4957 */ 4958 ZONE_LOCK(zone); 4959 if (zone->uz_max_items == 0) 4960 ZONE_ASSERT_COLD(zone); 4961 zone->uz_max_items = nitems; 4962 zone->uz_flags |= UMA_ZFLAG_LIMIT; 4963 zone_update_caches(zone); 4964 /* We may need to wake waiters. */ 4965 wakeup(&zone->uz_max_items); 4966 ZONE_UNLOCK(zone); 4967 4968 return (nitems); 4969 } 4970 4971 /* See uma.h */ 4972 void 4973 uma_zone_set_maxcache(uma_zone_t zone, int nitems) 4974 { 4975 int bpcpu, bpdom, bsize, nb; 4976 4977 ZONE_LOCK(zone); 4978 4979 /* 4980 * Compute a lower bound on the number of items that may be cached in 4981 * the zone. Each CPU gets at least two buckets, and for cross-domain 4982 * frees we use an additional bucket per CPU and per domain. Select the 4983 * largest bucket size that does not exceed half of the requested limit, 4984 * with the left over space given to the full bucket cache. 4985 */ 4986 bpdom = 0; 4987 bpcpu = 2; 4988 #ifdef NUMA 4989 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) { 4990 bpcpu++; 4991 bpdom++; 4992 } 4993 #endif 4994 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains; 4995 bsize = nitems / nb / 2; 4996 if (bsize > BUCKET_MAX) 4997 bsize = BUCKET_MAX; 4998 else if (bsize == 0 && nitems / nb > 0) 4999 bsize = 1; 5000 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize; 5001 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max) 5002 zone->uz_bucket_size_min = zone->uz_bucket_size_max; 5003 zone->uz_bucket_max = nitems - nb * bsize; 5004 ZONE_UNLOCK(zone); 5005 } 5006 5007 /* See uma.h */ 5008 int 5009 uma_zone_get_max(uma_zone_t zone) 5010 { 5011 int nitems; 5012 5013 nitems = atomic_load_64(&zone->uz_max_items); 5014 5015 return (nitems); 5016 } 5017 5018 /* See uma.h */ 5019 void 5020 uma_zone_set_warning(uma_zone_t zone, const char *warning) 5021 { 5022 5023 ZONE_ASSERT_COLD(zone); 5024 zone->uz_warning = warning; 5025 } 5026 5027 /* See uma.h */ 5028 void 5029 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) 5030 { 5031 5032 ZONE_ASSERT_COLD(zone); 5033 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); 5034 } 5035 5036 /* See uma.h */ 5037 int 5038 uma_zone_get_cur(uma_zone_t zone) 5039 { 5040 int64_t nitems; 5041 u_int i; 5042 5043 nitems = 0; 5044 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER) 5045 nitems = counter_u64_fetch(zone->uz_allocs) - 5046 counter_u64_fetch(zone->uz_frees); 5047 CPU_FOREACH(i) 5048 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) - 5049 atomic_load_64(&zone->uz_cpu[i].uc_frees); 5050 5051 return (nitems < 0 ? 0 : nitems); 5052 } 5053 5054 static uint64_t 5055 uma_zone_get_allocs(uma_zone_t zone) 5056 { 5057 uint64_t nitems; 5058 u_int i; 5059 5060 nitems = 0; 5061 if (zone->uz_allocs != EARLY_COUNTER) 5062 nitems = counter_u64_fetch(zone->uz_allocs); 5063 CPU_FOREACH(i) 5064 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs); 5065 5066 return (nitems); 5067 } 5068 5069 static uint64_t 5070 uma_zone_get_frees(uma_zone_t zone) 5071 { 5072 uint64_t nitems; 5073 u_int i; 5074 5075 nitems = 0; 5076 if (zone->uz_frees != EARLY_COUNTER) 5077 nitems = counter_u64_fetch(zone->uz_frees); 5078 CPU_FOREACH(i) 5079 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees); 5080 5081 return (nitems); 5082 } 5083 5084 #ifdef INVARIANTS 5085 /* Used only for KEG_ASSERT_COLD(). */ 5086 static uint64_t 5087 uma_keg_get_allocs(uma_keg_t keg) 5088 { 5089 uma_zone_t z; 5090 uint64_t nitems; 5091 5092 nitems = 0; 5093 LIST_FOREACH(z, &keg->uk_zones, uz_link) 5094 nitems += uma_zone_get_allocs(z); 5095 5096 return (nitems); 5097 } 5098 #endif 5099 5100 /* See uma.h */ 5101 void 5102 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 5103 { 5104 uma_keg_t keg; 5105 5106 KEG_GET(zone, keg); 5107 KEG_ASSERT_COLD(keg); 5108 keg->uk_init = uminit; 5109 } 5110 5111 /* See uma.h */ 5112 void 5113 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 5114 { 5115 uma_keg_t keg; 5116 5117 KEG_GET(zone, keg); 5118 KEG_ASSERT_COLD(keg); 5119 keg->uk_fini = fini; 5120 } 5121 5122 /* See uma.h */ 5123 void 5124 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 5125 { 5126 5127 ZONE_ASSERT_COLD(zone); 5128 zone->uz_init = zinit; 5129 } 5130 5131 /* See uma.h */ 5132 void 5133 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 5134 { 5135 5136 ZONE_ASSERT_COLD(zone); 5137 zone->uz_fini = zfini; 5138 } 5139 5140 /* See uma.h */ 5141 void 5142 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 5143 { 5144 uma_keg_t keg; 5145 5146 KEG_GET(zone, keg); 5147 KEG_ASSERT_COLD(keg); 5148 keg->uk_freef = freef; 5149 } 5150 5151 /* See uma.h */ 5152 void 5153 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 5154 { 5155 uma_keg_t keg; 5156 5157 KEG_GET(zone, keg); 5158 KEG_ASSERT_COLD(keg); 5159 keg->uk_allocf = allocf; 5160 } 5161 5162 /* See uma.h */ 5163 void 5164 uma_zone_set_smr(uma_zone_t zone, smr_t smr) 5165 { 5166 5167 ZONE_ASSERT_COLD(zone); 5168 5169 KASSERT(smr != NULL, ("Got NULL smr")); 5170 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0, 5171 ("zone %p (%s) already uses SMR", zone, zone->uz_name)); 5172 zone->uz_flags |= UMA_ZONE_SMR; 5173 zone->uz_smr = smr; 5174 zone_update_caches(zone); 5175 } 5176 5177 smr_t 5178 uma_zone_get_smr(uma_zone_t zone) 5179 { 5180 5181 return (zone->uz_smr); 5182 } 5183 5184 /* See uma.h */ 5185 void 5186 uma_zone_reserve(uma_zone_t zone, int items) 5187 { 5188 uma_keg_t keg; 5189 5190 KEG_GET(zone, keg); 5191 KEG_ASSERT_COLD(keg); 5192 keg->uk_reserve = items; 5193 } 5194 5195 /* See uma.h */ 5196 int 5197 uma_zone_reserve_kva(uma_zone_t zone, int count) 5198 { 5199 uma_keg_t keg; 5200 vm_offset_t kva; 5201 u_int pages; 5202 5203 KEG_GET(zone, keg); 5204 KEG_ASSERT_COLD(keg); 5205 ZONE_ASSERT_COLD(zone); 5206 5207 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera; 5208 5209 #ifdef UMA_USE_DMAP 5210 if (keg->uk_ppera > 1) { 5211 #else 5212 if (1) { 5213 #endif 5214 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE); 5215 if (kva == 0) 5216 return (0); 5217 } else 5218 kva = 0; 5219 5220 MPASS(keg->uk_kva == 0); 5221 keg->uk_kva = kva; 5222 keg->uk_offset = 0; 5223 zone->uz_max_items = pages * keg->uk_ipers; 5224 #ifdef UMA_USE_DMAP 5225 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 5226 #else 5227 keg->uk_allocf = noobj_alloc; 5228 #endif 5229 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE; 5230 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE; 5231 zone_update_caches(zone); 5232 5233 return (1); 5234 } 5235 5236 /* See uma.h */ 5237 void 5238 uma_prealloc(uma_zone_t zone, int items) 5239 { 5240 struct vm_domainset_iter di; 5241 uma_domain_t dom; 5242 uma_slab_t slab; 5243 uma_keg_t keg; 5244 int aflags, domain, slabs; 5245 5246 KEG_GET(zone, keg); 5247 slabs = howmany(items, keg->uk_ipers); 5248 while (slabs-- > 0) { 5249 aflags = M_NOWAIT; 5250 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 5251 &aflags); 5252 for (;;) { 5253 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK, 5254 aflags); 5255 if (slab != NULL) { 5256 dom = &keg->uk_domain[slab->us_domain]; 5257 /* 5258 * keg_alloc_slab() always returns a slab on the 5259 * partial list. 5260 */ 5261 LIST_REMOVE(slab, us_link); 5262 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, 5263 us_link); 5264 dom->ud_free_slabs++; 5265 KEG_UNLOCK(keg, slab->us_domain); 5266 break; 5267 } 5268 if (vm_domainset_iter_policy(&di, &domain) != 0) 5269 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0); 5270 } 5271 } 5272 } 5273 5274 /* 5275 * Returns a snapshot of memory consumption in bytes. 5276 */ 5277 size_t 5278 uma_zone_memory(uma_zone_t zone) 5279 { 5280 size_t sz; 5281 int i; 5282 5283 sz = 0; 5284 if (zone->uz_flags & UMA_ZFLAG_CACHE) { 5285 for (i = 0; i < vm_ndomains; i++) 5286 sz += ZDOM_GET(zone, i)->uzd_nitems; 5287 return (sz * zone->uz_size); 5288 } 5289 for (i = 0; i < vm_ndomains; i++) 5290 sz += zone->uz_keg->uk_domain[i].ud_pages; 5291 5292 return (sz * PAGE_SIZE); 5293 } 5294 5295 struct uma_reclaim_args { 5296 int domain; 5297 int req; 5298 }; 5299 5300 static void 5301 uma_reclaim_domain_cb(uma_zone_t zone, void *arg) 5302 { 5303 struct uma_reclaim_args *args; 5304 5305 args = arg; 5306 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0) 5307 uma_zone_reclaim_domain(zone, args->req, args->domain); 5308 } 5309 5310 /* See uma.h */ 5311 void 5312 uma_reclaim(int req) 5313 { 5314 uma_reclaim_domain(req, UMA_ANYDOMAIN); 5315 } 5316 5317 void 5318 uma_reclaim_domain(int req, int domain) 5319 { 5320 struct uma_reclaim_args args; 5321 5322 bucket_enable(); 5323 5324 args.domain = domain; 5325 args.req = req; 5326 5327 sx_slock(&uma_reclaim_lock); 5328 switch (req) { 5329 case UMA_RECLAIM_TRIM: 5330 case UMA_RECLAIM_DRAIN: 5331 zone_foreach(uma_reclaim_domain_cb, &args); 5332 break; 5333 case UMA_RECLAIM_DRAIN_CPU: 5334 zone_foreach(uma_reclaim_domain_cb, &args); 5335 pcpu_cache_drain_safe(NULL); 5336 zone_foreach(uma_reclaim_domain_cb, &args); 5337 break; 5338 default: 5339 panic("unhandled reclamation request %d", req); 5340 } 5341 5342 /* 5343 * Some slabs may have been freed but this zone will be visited early 5344 * we visit again so that we can free pages that are empty once other 5345 * zones are drained. We have to do the same for buckets. 5346 */ 5347 uma_zone_reclaim_domain(slabzones[0], UMA_RECLAIM_DRAIN, domain); 5348 uma_zone_reclaim_domain(slabzones[1], UMA_RECLAIM_DRAIN, domain); 5349 bucket_zone_drain(domain); 5350 sx_sunlock(&uma_reclaim_lock); 5351 } 5352 5353 static volatile int uma_reclaim_needed; 5354 5355 void 5356 uma_reclaim_wakeup(void) 5357 { 5358 5359 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0) 5360 wakeup(uma_reclaim); 5361 } 5362 5363 void 5364 uma_reclaim_worker(void *arg __unused) 5365 { 5366 5367 for (;;) { 5368 sx_xlock(&uma_reclaim_lock); 5369 while (atomic_load_int(&uma_reclaim_needed) == 0) 5370 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl", 5371 hz); 5372 sx_xunlock(&uma_reclaim_lock); 5373 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); 5374 uma_reclaim(UMA_RECLAIM_DRAIN_CPU); 5375 atomic_store_int(&uma_reclaim_needed, 0); 5376 /* Don't fire more than once per-second. */ 5377 pause("umarclslp", hz); 5378 } 5379 } 5380 5381 /* See uma.h */ 5382 void 5383 uma_zone_reclaim(uma_zone_t zone, int req) 5384 { 5385 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN); 5386 } 5387 5388 void 5389 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain) 5390 { 5391 switch (req) { 5392 case UMA_RECLAIM_TRIM: 5393 zone_reclaim(zone, domain, M_NOWAIT, false); 5394 break; 5395 case UMA_RECLAIM_DRAIN: 5396 zone_reclaim(zone, domain, M_NOWAIT, true); 5397 break; 5398 case UMA_RECLAIM_DRAIN_CPU: 5399 pcpu_cache_drain_safe(zone); 5400 zone_reclaim(zone, domain, M_NOWAIT, true); 5401 break; 5402 default: 5403 panic("unhandled reclamation request %d", req); 5404 } 5405 } 5406 5407 /* See uma.h */ 5408 int 5409 uma_zone_exhausted(uma_zone_t zone) 5410 { 5411 5412 return (atomic_load_32(&zone->uz_sleepers) > 0); 5413 } 5414 5415 unsigned long 5416 uma_limit(void) 5417 { 5418 5419 return (uma_kmem_limit); 5420 } 5421 5422 void 5423 uma_set_limit(unsigned long limit) 5424 { 5425 5426 uma_kmem_limit = limit; 5427 } 5428 5429 unsigned long 5430 uma_size(void) 5431 { 5432 5433 return (atomic_load_long(&uma_kmem_total)); 5434 } 5435 5436 long 5437 uma_avail(void) 5438 { 5439 5440 return (uma_kmem_limit - uma_size()); 5441 } 5442 5443 #ifdef DDB 5444 /* 5445 * Generate statistics across both the zone and its per-cpu cache's. Return 5446 * desired statistics if the pointer is non-NULL for that statistic. 5447 * 5448 * Note: does not update the zone statistics, as it can't safely clear the 5449 * per-CPU cache statistic. 5450 * 5451 */ 5452 static void 5453 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp, 5454 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp) 5455 { 5456 uma_cache_t cache; 5457 uint64_t allocs, frees, sleeps, xdomain; 5458 int cachefree, cpu; 5459 5460 allocs = frees = sleeps = xdomain = 0; 5461 cachefree = 0; 5462 CPU_FOREACH(cpu) { 5463 cache = &z->uz_cpu[cpu]; 5464 cachefree += cache->uc_allocbucket.ucb_cnt; 5465 cachefree += cache->uc_freebucket.ucb_cnt; 5466 xdomain += cache->uc_crossbucket.ucb_cnt; 5467 cachefree += cache->uc_crossbucket.ucb_cnt; 5468 allocs += cache->uc_allocs; 5469 frees += cache->uc_frees; 5470 } 5471 allocs += counter_u64_fetch(z->uz_allocs); 5472 frees += counter_u64_fetch(z->uz_frees); 5473 xdomain += counter_u64_fetch(z->uz_xdomain); 5474 sleeps += z->uz_sleeps; 5475 if (cachefreep != NULL) 5476 *cachefreep = cachefree; 5477 if (allocsp != NULL) 5478 *allocsp = allocs; 5479 if (freesp != NULL) 5480 *freesp = frees; 5481 if (sleepsp != NULL) 5482 *sleepsp = sleeps; 5483 if (xdomainp != NULL) 5484 *xdomainp = xdomain; 5485 } 5486 #endif /* DDB */ 5487 5488 static int 5489 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 5490 { 5491 uma_keg_t kz; 5492 uma_zone_t z; 5493 int count; 5494 5495 count = 0; 5496 rw_rlock(&uma_rwlock); 5497 LIST_FOREACH(kz, &uma_kegs, uk_link) { 5498 LIST_FOREACH(z, &kz->uk_zones, uz_link) 5499 count++; 5500 } 5501 LIST_FOREACH(z, &uma_cachezones, uz_link) 5502 count++; 5503 5504 rw_runlock(&uma_rwlock); 5505 return (sysctl_handle_int(oidp, &count, 0, req)); 5506 } 5507 5508 static void 5509 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf, 5510 struct uma_percpu_stat *ups, bool internal) 5511 { 5512 uma_zone_domain_t zdom; 5513 uma_cache_t cache; 5514 int i; 5515 5516 for (i = 0; i < vm_ndomains; i++) { 5517 zdom = ZDOM_GET(z, i); 5518 uth->uth_zone_free += zdom->uzd_nitems; 5519 } 5520 uth->uth_allocs = counter_u64_fetch(z->uz_allocs); 5521 uth->uth_frees = counter_u64_fetch(z->uz_frees); 5522 uth->uth_fails = counter_u64_fetch(z->uz_fails); 5523 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain); 5524 uth->uth_sleeps = z->uz_sleeps; 5525 5526 for (i = 0; i < mp_maxid + 1; i++) { 5527 bzero(&ups[i], sizeof(*ups)); 5528 if (internal || CPU_ABSENT(i)) 5529 continue; 5530 cache = &z->uz_cpu[i]; 5531 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt; 5532 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt; 5533 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt; 5534 ups[i].ups_allocs = cache->uc_allocs; 5535 ups[i].ups_frees = cache->uc_frees; 5536 } 5537 } 5538 5539 static int 5540 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 5541 { 5542 struct uma_stream_header ush; 5543 struct uma_type_header uth; 5544 struct uma_percpu_stat *ups; 5545 struct sbuf sbuf; 5546 uma_keg_t kz; 5547 uma_zone_t z; 5548 uint64_t items; 5549 uint32_t kfree, pages; 5550 int count, error, i; 5551 5552 error = sysctl_wire_old_buffer(req, 0); 5553 if (error != 0) 5554 return (error); 5555 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 5556 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); 5557 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK); 5558 5559 count = 0; 5560 rw_rlock(&uma_rwlock); 5561 LIST_FOREACH(kz, &uma_kegs, uk_link) { 5562 LIST_FOREACH(z, &kz->uk_zones, uz_link) 5563 count++; 5564 } 5565 5566 LIST_FOREACH(z, &uma_cachezones, uz_link) 5567 count++; 5568 5569 /* 5570 * Insert stream header. 5571 */ 5572 bzero(&ush, sizeof(ush)); 5573 ush.ush_version = UMA_STREAM_VERSION; 5574 ush.ush_maxcpus = (mp_maxid + 1); 5575 ush.ush_count = count; 5576 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 5577 5578 LIST_FOREACH(kz, &uma_kegs, uk_link) { 5579 kfree = pages = 0; 5580 for (i = 0; i < vm_ndomains; i++) { 5581 kfree += kz->uk_domain[i].ud_free_items; 5582 pages += kz->uk_domain[i].ud_pages; 5583 } 5584 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 5585 bzero(&uth, sizeof(uth)); 5586 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 5587 uth.uth_align = kz->uk_align; 5588 uth.uth_size = kz->uk_size; 5589 uth.uth_rsize = kz->uk_rsize; 5590 if (z->uz_max_items > 0) { 5591 items = UZ_ITEMS_COUNT(z->uz_items); 5592 uth.uth_pages = (items / kz->uk_ipers) * 5593 kz->uk_ppera; 5594 } else 5595 uth.uth_pages = pages; 5596 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) * 5597 kz->uk_ppera; 5598 uth.uth_limit = z->uz_max_items; 5599 uth.uth_keg_free = kfree; 5600 5601 /* 5602 * A zone is secondary is it is not the first entry 5603 * on the keg's zone list. 5604 */ 5605 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 5606 (LIST_FIRST(&kz->uk_zones) != z)) 5607 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 5608 uma_vm_zone_stats(&uth, z, &sbuf, ups, 5609 kz->uk_flags & UMA_ZFLAG_INTERNAL); 5610 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 5611 for (i = 0; i < mp_maxid + 1; i++) 5612 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 5613 } 5614 } 5615 LIST_FOREACH(z, &uma_cachezones, uz_link) { 5616 bzero(&uth, sizeof(uth)); 5617 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 5618 uth.uth_size = z->uz_size; 5619 uma_vm_zone_stats(&uth, z, &sbuf, ups, false); 5620 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 5621 for (i = 0; i < mp_maxid + 1; i++) 5622 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 5623 } 5624 5625 rw_runlock(&uma_rwlock); 5626 error = sbuf_finish(&sbuf); 5627 sbuf_delete(&sbuf); 5628 free(ups, M_TEMP); 5629 return (error); 5630 } 5631 5632 int 5633 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 5634 { 5635 uma_zone_t zone = *(uma_zone_t *)arg1; 5636 int error, max; 5637 5638 max = uma_zone_get_max(zone); 5639 error = sysctl_handle_int(oidp, &max, 0, req); 5640 if (error || !req->newptr) 5641 return (error); 5642 5643 uma_zone_set_max(zone, max); 5644 5645 return (0); 5646 } 5647 5648 int 5649 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 5650 { 5651 uma_zone_t zone; 5652 int cur; 5653 5654 /* 5655 * Some callers want to add sysctls for global zones that 5656 * may not yet exist so they pass a pointer to a pointer. 5657 */ 5658 if (arg2 == 0) 5659 zone = *(uma_zone_t *)arg1; 5660 else 5661 zone = arg1; 5662 cur = uma_zone_get_cur(zone); 5663 return (sysctl_handle_int(oidp, &cur, 0, req)); 5664 } 5665 5666 static int 5667 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS) 5668 { 5669 uma_zone_t zone = arg1; 5670 uint64_t cur; 5671 5672 cur = uma_zone_get_allocs(zone); 5673 return (sysctl_handle_64(oidp, &cur, 0, req)); 5674 } 5675 5676 static int 5677 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS) 5678 { 5679 uma_zone_t zone = arg1; 5680 uint64_t cur; 5681 5682 cur = uma_zone_get_frees(zone); 5683 return (sysctl_handle_64(oidp, &cur, 0, req)); 5684 } 5685 5686 static int 5687 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS) 5688 { 5689 struct sbuf sbuf; 5690 uma_zone_t zone = arg1; 5691 int error; 5692 5693 sbuf_new_for_sysctl(&sbuf, NULL, 0, req); 5694 if (zone->uz_flags != 0) 5695 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS); 5696 else 5697 sbuf_printf(&sbuf, "0"); 5698 error = sbuf_finish(&sbuf); 5699 sbuf_delete(&sbuf); 5700 5701 return (error); 5702 } 5703 5704 static int 5705 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS) 5706 { 5707 uma_keg_t keg = arg1; 5708 int avail, effpct, total; 5709 5710 total = keg->uk_ppera * PAGE_SIZE; 5711 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0) 5712 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize; 5713 /* 5714 * We consider the client's requested size and alignment here, not the 5715 * real size determination uk_rsize, because we also adjust the real 5716 * size for internal implementation reasons (max bitset size). 5717 */ 5718 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1); 5719 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) 5720 avail *= mp_maxid + 1; 5721 effpct = 100 * avail / total; 5722 return (sysctl_handle_int(oidp, &effpct, 0, req)); 5723 } 5724 5725 static int 5726 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS) 5727 { 5728 uma_zone_t zone = arg1; 5729 uint64_t cur; 5730 5731 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items)); 5732 return (sysctl_handle_64(oidp, &cur, 0, req)); 5733 } 5734 5735 #ifdef INVARIANTS 5736 static uma_slab_t 5737 uma_dbg_getslab(uma_zone_t zone, void *item) 5738 { 5739 uma_slab_t slab; 5740 uma_keg_t keg; 5741 uint8_t *mem; 5742 5743 /* 5744 * It is safe to return the slab here even though the 5745 * zone is unlocked because the item's allocation state 5746 * essentially holds a reference. 5747 */ 5748 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 5749 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 5750 return (NULL); 5751 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB) 5752 return (vtoslab((vm_offset_t)mem)); 5753 keg = zone->uz_keg; 5754 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0) 5755 return ((uma_slab_t)(mem + keg->uk_pgoff)); 5756 KEG_LOCK(keg, 0); 5757 slab = hash_sfind(&keg->uk_hash, mem); 5758 KEG_UNLOCK(keg, 0); 5759 5760 return (slab); 5761 } 5762 5763 static bool 5764 uma_dbg_zskip(uma_zone_t zone, void *mem) 5765 { 5766 5767 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 5768 return (true); 5769 5770 return (uma_dbg_kskip(zone->uz_keg, mem)); 5771 } 5772 5773 static bool 5774 uma_dbg_kskip(uma_keg_t keg, void *mem) 5775 { 5776 uintptr_t idx; 5777 5778 if (dbg_divisor == 0) 5779 return (true); 5780 5781 if (dbg_divisor == 1) 5782 return (false); 5783 5784 idx = (uintptr_t)mem >> PAGE_SHIFT; 5785 if (keg->uk_ipers > 1) { 5786 idx *= keg->uk_ipers; 5787 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize; 5788 } 5789 5790 if ((idx / dbg_divisor) * dbg_divisor != idx) { 5791 counter_u64_add(uma_skip_cnt, 1); 5792 return (true); 5793 } 5794 counter_u64_add(uma_dbg_cnt, 1); 5795 5796 return (false); 5797 } 5798 5799 /* 5800 * Set up the slab's freei data such that uma_dbg_free can function. 5801 * 5802 */ 5803 static void 5804 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) 5805 { 5806 uma_keg_t keg; 5807 int freei; 5808 5809 if (slab == NULL) { 5810 slab = uma_dbg_getslab(zone, item); 5811 if (slab == NULL) 5812 panic("uma: item %p did not belong to zone %s", 5813 item, zone->uz_name); 5814 } 5815 keg = zone->uz_keg; 5816 freei = slab_item_index(slab, keg, item); 5817 5818 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei, 5819 slab_dbg_bits(slab, keg))) 5820 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)", 5821 item, zone, zone->uz_name, slab, freei); 5822 } 5823 5824 /* 5825 * Verifies freed addresses. Checks for alignment, valid slab membership 5826 * and duplicate frees. 5827 * 5828 */ 5829 static void 5830 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) 5831 { 5832 uma_keg_t keg; 5833 int freei; 5834 5835 if (slab == NULL) { 5836 slab = uma_dbg_getslab(zone, item); 5837 if (slab == NULL) 5838 panic("uma: Freed item %p did not belong to zone %s", 5839 item, zone->uz_name); 5840 } 5841 keg = zone->uz_keg; 5842 freei = slab_item_index(slab, keg, item); 5843 5844 if (freei >= keg->uk_ipers) 5845 panic("Invalid free of %p from zone %p(%s) slab %p(%d)", 5846 item, zone, zone->uz_name, slab, freei); 5847 5848 if (slab_item(slab, keg, freei) != item) 5849 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)", 5850 item, zone, zone->uz_name, slab, freei); 5851 5852 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei, 5853 slab_dbg_bits(slab, keg))) 5854 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)", 5855 item, zone, zone->uz_name, slab, freei); 5856 } 5857 #endif /* INVARIANTS */ 5858 5859 #ifdef DDB 5860 static int64_t 5861 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used, 5862 uint64_t *sleeps, long *cachefree, uint64_t *xdomain) 5863 { 5864 uint64_t frees; 5865 int i; 5866 5867 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 5868 *allocs = counter_u64_fetch(z->uz_allocs); 5869 frees = counter_u64_fetch(z->uz_frees); 5870 *sleeps = z->uz_sleeps; 5871 *cachefree = 0; 5872 *xdomain = 0; 5873 } else 5874 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps, 5875 xdomain); 5876 for (i = 0; i < vm_ndomains; i++) { 5877 *cachefree += ZDOM_GET(z, i)->uzd_nitems; 5878 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 5879 (LIST_FIRST(&kz->uk_zones) != z))) 5880 *cachefree += kz->uk_domain[i].ud_free_items; 5881 } 5882 *used = *allocs - frees; 5883 return (((int64_t)*used + *cachefree) * kz->uk_size); 5884 } 5885 5886 DB_SHOW_COMMAND_FLAGS(uma, db_show_uma, DB_CMD_MEMSAFE) 5887 { 5888 const char *fmt_hdr, *fmt_entry; 5889 uma_keg_t kz; 5890 uma_zone_t z; 5891 uint64_t allocs, used, sleeps, xdomain; 5892 long cachefree; 5893 /* variables for sorting */ 5894 uma_keg_t cur_keg; 5895 uma_zone_t cur_zone, last_zone; 5896 int64_t cur_size, last_size, size; 5897 int ties; 5898 5899 /* /i option produces machine-parseable CSV output */ 5900 if (modif[0] == 'i') { 5901 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n"; 5902 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n"; 5903 } else { 5904 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n"; 5905 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n"; 5906 } 5907 5908 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests", 5909 "Sleeps", "Bucket", "Total Mem", "XFree"); 5910 5911 /* Sort the zones with largest size first. */ 5912 last_zone = NULL; 5913 last_size = INT64_MAX; 5914 for (;;) { 5915 cur_zone = NULL; 5916 cur_size = -1; 5917 ties = 0; 5918 LIST_FOREACH(kz, &uma_kegs, uk_link) { 5919 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 5920 /* 5921 * In the case of size ties, print out zones 5922 * in the order they are encountered. That is, 5923 * when we encounter the most recently output 5924 * zone, we have already printed all preceding 5925 * ties, and we must print all following ties. 5926 */ 5927 if (z == last_zone) { 5928 ties = 1; 5929 continue; 5930 } 5931 size = get_uma_stats(kz, z, &allocs, &used, 5932 &sleeps, &cachefree, &xdomain); 5933 if (size > cur_size && size < last_size + ties) 5934 { 5935 cur_size = size; 5936 cur_zone = z; 5937 cur_keg = kz; 5938 } 5939 } 5940 } 5941 if (cur_zone == NULL) 5942 break; 5943 5944 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used, 5945 &sleeps, &cachefree, &xdomain); 5946 db_printf(fmt_entry, cur_zone->uz_name, 5947 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree, 5948 (uintmax_t)allocs, (uintmax_t)sleeps, 5949 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size, 5950 xdomain); 5951 5952 if (db_pager_quit) 5953 return; 5954 last_zone = cur_zone; 5955 last_size = cur_size; 5956 } 5957 } 5958 5959 DB_SHOW_COMMAND_FLAGS(umacache, db_show_umacache, DB_CMD_MEMSAFE) 5960 { 5961 uma_zone_t z; 5962 uint64_t allocs, frees; 5963 long cachefree; 5964 int i; 5965 5966 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 5967 "Requests", "Bucket"); 5968 LIST_FOREACH(z, &uma_cachezones, uz_link) { 5969 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL); 5970 for (i = 0; i < vm_ndomains; i++) 5971 cachefree += ZDOM_GET(z, i)->uzd_nitems; 5972 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n", 5973 z->uz_name, (uintmax_t)z->uz_size, 5974 (intmax_t)(allocs - frees), cachefree, 5975 (uintmax_t)allocs, z->uz_bucket_size); 5976 if (db_pager_quit) 5977 return; 5978 } 5979 } 5980 #endif /* DDB */ 5981