1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2002-2019 Jeffrey Roberson <jeff@FreeBSD.org> 5 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org> 6 * Copyright (c) 2004-2006 Robert N. M. Watson 7 * All rights reserved. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice unmodified, this list of conditions, and the following 14 * disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 29 */ 30 31 /* 32 * uma_core.c Implementation of the Universal Memory allocator 33 * 34 * This allocator is intended to replace the multitude of similar object caches 35 * in the standard FreeBSD kernel. The intent is to be flexible as well as 36 * efficient. A primary design goal is to return unused memory to the rest of 37 * the system. This will make the system as a whole more flexible due to the 38 * ability to move memory to subsystems which most need it instead of leaving 39 * pools of reserved memory unused. 40 * 41 * The basic ideas stem from similar slab/zone based allocators whose algorithms 42 * are well known. 43 * 44 */ 45 46 /* 47 * TODO: 48 * - Improve memory usage for large allocations 49 * - Investigate cache size adjustments 50 */ 51 52 #include <sys/cdefs.h> 53 __FBSDID("$FreeBSD$"); 54 55 #include "opt_ddb.h" 56 #include "opt_param.h" 57 #include "opt_vm.h" 58 59 #include <sys/param.h> 60 #include <sys/systm.h> 61 #include <sys/bitset.h> 62 #include <sys/domainset.h> 63 #include <sys/eventhandler.h> 64 #include <sys/kernel.h> 65 #include <sys/types.h> 66 #include <sys/limits.h> 67 #include <sys/queue.h> 68 #include <sys/malloc.h> 69 #include <sys/ktr.h> 70 #include <sys/lock.h> 71 #include <sys/sysctl.h> 72 #include <sys/mutex.h> 73 #include <sys/proc.h> 74 #include <sys/random.h> 75 #include <sys/rwlock.h> 76 #include <sys/sbuf.h> 77 #include <sys/sched.h> 78 #include <sys/sleepqueue.h> 79 #include <sys/smp.h> 80 #include <sys/smr.h> 81 #include <sys/taskqueue.h> 82 #include <sys/vmmeter.h> 83 84 #include <vm/vm.h> 85 #include <vm/vm_domainset.h> 86 #include <vm/vm_object.h> 87 #include <vm/vm_page.h> 88 #include <vm/vm_pageout.h> 89 #include <vm/vm_param.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/uma.h> 96 #include <vm/uma_int.h> 97 #include <vm/uma_dbg.h> 98 99 #include <ddb/ddb.h> 100 101 #ifdef DEBUG_MEMGUARD 102 #include <vm/memguard.h> 103 #endif 104 105 #include <machine/md_var.h> 106 107 #ifdef INVARIANTS 108 #define UMA_ALWAYS_CTORDTOR 1 109 #else 110 #define UMA_ALWAYS_CTORDTOR 0 111 #endif 112 113 /* 114 * This is the zone and keg from which all zones are spawned. 115 */ 116 static uma_zone_t kegs; 117 static uma_zone_t zones; 118 119 /* 120 * These are the two zones from which all offpage uma_slab_ts are allocated. 121 * 122 * One zone is for slab headers that can represent a larger number of items, 123 * making the slabs themselves more efficient, and the other zone is for 124 * headers that are smaller and represent fewer items, making the headers more 125 * efficient. 126 */ 127 #define SLABZONE_SIZE(setsize) \ 128 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS) 129 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16) 130 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE 131 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE) 132 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE) 133 static uma_zone_t slabzones[2]; 134 135 /* 136 * The initial hash tables come out of this zone so they can be allocated 137 * prior to malloc coming up. 138 */ 139 static uma_zone_t hashzone; 140 141 /* The boot-time adjusted value for cache line alignment. */ 142 int uma_align_cache = 64 - 1; 143 144 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 145 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc"); 146 147 /* 148 * Are we allowed to allocate buckets? 149 */ 150 static int bucketdisable = 1; 151 152 /* Linked list of all kegs in the system */ 153 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 154 155 /* Linked list of all cache-only zones in the system */ 156 static LIST_HEAD(,uma_zone) uma_cachezones = 157 LIST_HEAD_INITIALIZER(uma_cachezones); 158 159 /* This RW lock protects the keg list */ 160 static struct rwlock_padalign __exclusive_cache_line uma_rwlock; 161 162 /* 163 * First available virual address for boot time allocations. 164 */ 165 static vm_offset_t bootstart; 166 static vm_offset_t bootmem; 167 168 static struct sx uma_reclaim_lock; 169 170 /* 171 * kmem soft limit, initialized by uma_set_limit(). Ensure that early 172 * allocations don't trigger a wakeup of the reclaim thread. 173 */ 174 unsigned long uma_kmem_limit = LONG_MAX; 175 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0, 176 "UMA kernel memory soft limit"); 177 unsigned long uma_kmem_total; 178 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0, 179 "UMA kernel memory usage"); 180 181 /* Is the VM done starting up? */ 182 static enum { 183 BOOT_COLD, 184 BOOT_KVA, 185 BOOT_RUNNING, 186 BOOT_SHUTDOWN, 187 } booted = BOOT_COLD; 188 189 /* 190 * This is the handle used to schedule events that need to happen 191 * outside of the allocation fast path. 192 */ 193 static struct callout uma_callout; 194 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 195 196 /* 197 * This structure is passed as the zone ctor arg so that I don't have to create 198 * a special allocation function just for zones. 199 */ 200 struct uma_zctor_args { 201 const char *name; 202 size_t size; 203 uma_ctor ctor; 204 uma_dtor dtor; 205 uma_init uminit; 206 uma_fini fini; 207 uma_import import; 208 uma_release release; 209 void *arg; 210 uma_keg_t keg; 211 int align; 212 uint32_t flags; 213 }; 214 215 struct uma_kctor_args { 216 uma_zone_t zone; 217 size_t size; 218 uma_init uminit; 219 uma_fini fini; 220 int align; 221 uint32_t flags; 222 }; 223 224 struct uma_bucket_zone { 225 uma_zone_t ubz_zone; 226 char *ubz_name; 227 int ubz_entries; /* Number of items it can hold. */ 228 int ubz_maxsize; /* Maximum allocation size per-item. */ 229 }; 230 231 /* 232 * Compute the actual number of bucket entries to pack them in power 233 * of two sizes for more efficient space utilization. 234 */ 235 #define BUCKET_SIZE(n) \ 236 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 237 238 #define BUCKET_MAX BUCKET_SIZE(256) 239 #define BUCKET_MIN 2 240 241 struct uma_bucket_zone bucket_zones[] = { 242 /* Literal bucket sizes. */ 243 { NULL, "2 Bucket", 2, 4096 }, 244 { NULL, "4 Bucket", 4, 3072 }, 245 { NULL, "8 Bucket", 8, 2048 }, 246 { NULL, "16 Bucket", 16, 1024 }, 247 /* Rounded down power of 2 sizes for efficiency. */ 248 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 249 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 250 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 251 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 252 { NULL, NULL, 0} 253 }; 254 255 /* 256 * Flags and enumerations to be passed to internal functions. 257 */ 258 enum zfreeskip { 259 SKIP_NONE = 0, 260 SKIP_CNT = 0x00000001, 261 SKIP_DTOR = 0x00010000, 262 SKIP_FINI = 0x00020000, 263 }; 264 265 /* Prototypes.. */ 266 267 void uma_startup1(vm_offset_t); 268 void uma_startup2(void); 269 270 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 271 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 272 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 273 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 274 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 275 static void page_free(void *, vm_size_t, uint8_t); 276 static void pcpu_page_free(void *, vm_size_t, uint8_t); 277 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int); 278 static void cache_drain(uma_zone_t); 279 static void bucket_drain(uma_zone_t, uma_bucket_t); 280 static void bucket_cache_reclaim(uma_zone_t zone, bool); 281 static int keg_ctor(void *, int, void *, int); 282 static void keg_dtor(void *, int, void *); 283 static int zone_ctor(void *, int, void *, int); 284 static void zone_dtor(void *, int, void *); 285 static inline void item_dtor(uma_zone_t zone, void *item, int size, 286 void *udata, enum zfreeskip skip); 287 static int zero_init(void *, int, int); 288 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *); 289 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *); 290 static void zone_timeout(uma_zone_t zone, void *); 291 static int hash_alloc(struct uma_hash *, u_int); 292 static int hash_expand(struct uma_hash *, struct uma_hash *); 293 static void hash_free(struct uma_hash *hash); 294 static void uma_timeout(void *); 295 static void uma_startup3(void); 296 static void uma_shutdown(void); 297 static void *zone_alloc_item(uma_zone_t, void *, int, int); 298 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 299 static int zone_alloc_limit(uma_zone_t zone, int count, int flags); 300 static void zone_free_limit(uma_zone_t zone, int count); 301 static void bucket_enable(void); 302 static void bucket_init(void); 303 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 304 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 305 static void bucket_zone_drain(void); 306 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int); 307 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 308 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item); 309 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 310 uma_fini fini, int align, uint32_t flags); 311 static int zone_import(void *, void **, int, int, int); 312 static void zone_release(void *, void **, int); 313 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int); 314 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int); 315 316 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 317 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 318 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS); 319 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS); 320 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS); 321 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS); 322 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS); 323 324 static uint64_t uma_zone_get_allocs(uma_zone_t zone); 325 326 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0, 327 "Memory allocation debugging"); 328 329 #ifdef INVARIANTS 330 static uint64_t uma_keg_get_allocs(uma_keg_t zone); 331 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg); 332 333 static bool uma_dbg_kskip(uma_keg_t keg, void *mem); 334 static bool uma_dbg_zskip(uma_zone_t zone, void *mem); 335 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item); 336 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item); 337 338 static u_int dbg_divisor = 1; 339 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor, 340 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0, 341 "Debug & thrash every this item in memory allocator"); 342 343 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER; 344 static counter_u64_t uma_skip_cnt = EARLY_COUNTER; 345 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD, 346 &uma_dbg_cnt, "memory items debugged"); 347 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD, 348 &uma_skip_cnt, "memory items skipped, not debugged"); 349 #endif 350 351 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 352 353 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW, 0, "Universal Memory Allocator"); 354 355 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT, 356 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 357 358 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT, 359 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 360 361 static int zone_warnings = 1; 362 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, 363 "Warn when UMA zones becomes full"); 364 365 static int multipage_slabs = 1; 366 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs); 367 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs, 368 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0, 369 "UMA may choose larger slab sizes for better efficiency"); 370 371 /* 372 * Select the slab zone for an offpage slab with the given maximum item count. 373 */ 374 static inline uma_zone_t 375 slabzone(int ipers) 376 { 377 378 return (slabzones[ipers > SLABZONE0_SETSIZE]); 379 } 380 381 /* 382 * This routine checks to see whether or not it's safe to enable buckets. 383 */ 384 static void 385 bucket_enable(void) 386 { 387 388 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init")); 389 bucketdisable = vm_page_count_min(); 390 } 391 392 /* 393 * Initialize bucket_zones, the array of zones of buckets of various sizes. 394 * 395 * For each zone, calculate the memory required for each bucket, consisting 396 * of the header and an array of pointers. 397 */ 398 static void 399 bucket_init(void) 400 { 401 struct uma_bucket_zone *ubz; 402 int size; 403 404 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 405 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 406 size += sizeof(void *) * ubz->ubz_entries; 407 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 408 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 409 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET | 410 UMA_ZONE_FIRSTTOUCH); 411 } 412 } 413 414 /* 415 * Given a desired number of entries for a bucket, return the zone from which 416 * to allocate the bucket. 417 */ 418 static struct uma_bucket_zone * 419 bucket_zone_lookup(int entries) 420 { 421 struct uma_bucket_zone *ubz; 422 423 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 424 if (ubz->ubz_entries >= entries) 425 return (ubz); 426 ubz--; 427 return (ubz); 428 } 429 430 static struct uma_bucket_zone * 431 bucket_zone_max(uma_zone_t zone, int nitems) 432 { 433 struct uma_bucket_zone *ubz; 434 int bpcpu; 435 436 bpcpu = 2; 437 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 438 /* Count the cross-domain bucket. */ 439 bpcpu++; 440 441 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 442 if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems) 443 break; 444 if (ubz == &bucket_zones[0]) 445 ubz = NULL; 446 else 447 ubz--; 448 return (ubz); 449 } 450 451 static int 452 bucket_select(int size) 453 { 454 struct uma_bucket_zone *ubz; 455 456 ubz = &bucket_zones[0]; 457 if (size > ubz->ubz_maxsize) 458 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 459 460 for (; ubz->ubz_entries != 0; ubz++) 461 if (ubz->ubz_maxsize < size) 462 break; 463 ubz--; 464 return (ubz->ubz_entries); 465 } 466 467 static uma_bucket_t 468 bucket_alloc(uma_zone_t zone, void *udata, int flags) 469 { 470 struct uma_bucket_zone *ubz; 471 uma_bucket_t bucket; 472 473 /* 474 * Don't allocate buckets early in boot. 475 */ 476 if (__predict_false(booted < BOOT_KVA)) 477 return (NULL); 478 479 /* 480 * To limit bucket recursion we store the original zone flags 481 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 482 * NOVM flag to persist even through deep recursions. We also 483 * store ZFLAG_BUCKET once we have recursed attempting to allocate 484 * a bucket for a bucket zone so we do not allow infinite bucket 485 * recursion. This cookie will even persist to frees of unused 486 * buckets via the allocation path or bucket allocations in the 487 * free path. 488 */ 489 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 490 udata = (void *)(uintptr_t)zone->uz_flags; 491 else { 492 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 493 return (NULL); 494 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 495 } 496 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 497 flags |= M_NOVM; 498 ubz = bucket_zone_lookup(zone->uz_bucket_size); 499 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 500 ubz++; 501 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 502 if (bucket) { 503 #ifdef INVARIANTS 504 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 505 #endif 506 bucket->ub_cnt = 0; 507 bucket->ub_entries = ubz->ubz_entries; 508 bucket->ub_seq = SMR_SEQ_INVALID; 509 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p", 510 zone->uz_name, zone, bucket); 511 } 512 513 return (bucket); 514 } 515 516 static void 517 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 518 { 519 struct uma_bucket_zone *ubz; 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(void) 533 { 534 struct uma_bucket_zone *ubz; 535 536 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 537 uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN); 538 } 539 540 /* 541 * Attempt to satisfy an allocation by retrieving a full bucket from one of the 542 * zone's caches. If a bucket is found the zone is not locked on return. 543 */ 544 static uma_bucket_t 545 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom) 546 { 547 uma_bucket_t bucket; 548 int i; 549 bool dtor = false; 550 551 ZONE_LOCK_ASSERT(zone); 552 553 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL) 554 return (NULL); 555 556 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && 557 bucket->ub_seq != SMR_SEQ_INVALID) { 558 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false)) 559 return (NULL); 560 bucket->ub_seq = SMR_SEQ_INVALID; 561 dtor = (zone->uz_dtor != NULL) | UMA_ALWAYS_CTORDTOR; 562 } 563 MPASS(zdom->uzd_nitems >= bucket->ub_cnt); 564 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link); 565 zdom->uzd_nitems -= bucket->ub_cnt; 566 if (zdom->uzd_imin > zdom->uzd_nitems) 567 zdom->uzd_imin = zdom->uzd_nitems; 568 zone->uz_bkt_count -= bucket->ub_cnt; 569 ZONE_UNLOCK(zone); 570 if (dtor) 571 for (i = 0; i < bucket->ub_cnt; i++) 572 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size, 573 NULL, SKIP_NONE); 574 575 return (bucket); 576 } 577 578 /* 579 * Insert a full bucket into the specified cache. The "ws" parameter indicates 580 * whether the bucket's contents should be counted as part of the zone's working 581 * set. 582 */ 583 static void 584 zone_put_bucket(uma_zone_t zone, uma_zone_domain_t zdom, uma_bucket_t bucket, 585 const bool ws) 586 { 587 588 ZONE_LOCK_ASSERT(zone); 589 KASSERT(!ws || zone->uz_bkt_count < zone->uz_bkt_max, 590 ("%s: zone %p overflow", __func__, zone)); 591 592 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link); 593 zdom->uzd_nitems += bucket->ub_cnt; 594 if (ws && zdom->uzd_imax < zdom->uzd_nitems) 595 zdom->uzd_imax = zdom->uzd_nitems; 596 zone->uz_bkt_count += bucket->ub_cnt; 597 } 598 599 /* Pops an item out of a per-cpu cache bucket. */ 600 static inline void * 601 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket) 602 { 603 void *item; 604 605 CRITICAL_ASSERT(curthread); 606 607 bucket->ucb_cnt--; 608 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt]; 609 #ifdef INVARIANTS 610 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL; 611 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 612 #endif 613 cache->uc_allocs++; 614 615 return (item); 616 } 617 618 /* Pushes an item into a per-cpu cache bucket. */ 619 static inline void 620 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item) 621 { 622 623 CRITICAL_ASSERT(curthread); 624 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL, 625 ("uma_zfree: Freeing to non free bucket index.")); 626 627 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item; 628 bucket->ucb_cnt++; 629 cache->uc_frees++; 630 } 631 632 /* 633 * Unload a UMA bucket from a per-cpu cache. 634 */ 635 static inline uma_bucket_t 636 cache_bucket_unload(uma_cache_bucket_t bucket) 637 { 638 uma_bucket_t b; 639 640 b = bucket->ucb_bucket; 641 if (b != NULL) { 642 MPASS(b->ub_entries == bucket->ucb_entries); 643 b->ub_cnt = bucket->ucb_cnt; 644 bucket->ucb_bucket = NULL; 645 bucket->ucb_entries = bucket->ucb_cnt = 0; 646 } 647 648 return (b); 649 } 650 651 static inline uma_bucket_t 652 cache_bucket_unload_alloc(uma_cache_t cache) 653 { 654 655 return (cache_bucket_unload(&cache->uc_allocbucket)); 656 } 657 658 static inline uma_bucket_t 659 cache_bucket_unload_free(uma_cache_t cache) 660 { 661 662 return (cache_bucket_unload(&cache->uc_freebucket)); 663 } 664 665 static inline uma_bucket_t 666 cache_bucket_unload_cross(uma_cache_t cache) 667 { 668 669 return (cache_bucket_unload(&cache->uc_crossbucket)); 670 } 671 672 /* 673 * Load a bucket into a per-cpu cache bucket. 674 */ 675 static inline void 676 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b) 677 { 678 679 CRITICAL_ASSERT(curthread); 680 MPASS(bucket->ucb_bucket == NULL); 681 682 bucket->ucb_bucket = b; 683 bucket->ucb_cnt = b->ub_cnt; 684 bucket->ucb_entries = b->ub_entries; 685 } 686 687 static inline void 688 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b) 689 { 690 691 cache_bucket_load(&cache->uc_allocbucket, b); 692 } 693 694 static inline void 695 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b) 696 { 697 698 cache_bucket_load(&cache->uc_freebucket, b); 699 } 700 701 #ifdef NUMA 702 static inline void 703 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b) 704 { 705 706 cache_bucket_load(&cache->uc_crossbucket, b); 707 } 708 #endif 709 710 /* 711 * Copy and preserve ucb_spare. 712 */ 713 static inline void 714 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2) 715 { 716 717 b1->ucb_bucket = b2->ucb_bucket; 718 b1->ucb_entries = b2->ucb_entries; 719 b1->ucb_cnt = b2->ucb_cnt; 720 } 721 722 /* 723 * Swap two cache buckets. 724 */ 725 static inline void 726 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2) 727 { 728 struct uma_cache_bucket b3; 729 730 CRITICAL_ASSERT(curthread); 731 732 cache_bucket_copy(&b3, b1); 733 cache_bucket_copy(b1, b2); 734 cache_bucket_copy(b2, &b3); 735 } 736 737 static void 738 zone_log_warning(uma_zone_t zone) 739 { 740 static const struct timeval warninterval = { 300, 0 }; 741 742 if (!zone_warnings || zone->uz_warning == NULL) 743 return; 744 745 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 746 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 747 } 748 749 static inline void 750 zone_maxaction(uma_zone_t zone) 751 { 752 753 if (zone->uz_maxaction.ta_func != NULL) 754 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); 755 } 756 757 /* 758 * Routine called by timeout which is used to fire off some time interval 759 * based calculations. (stats, hash size, etc.) 760 * 761 * Arguments: 762 * arg Unused 763 * 764 * Returns: 765 * Nothing 766 */ 767 static void 768 uma_timeout(void *unused) 769 { 770 bucket_enable(); 771 zone_foreach(zone_timeout, NULL); 772 773 /* Reschedule this event */ 774 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 775 } 776 777 /* 778 * Update the working set size estimate for the zone's bucket cache. 779 * The constants chosen here are somewhat arbitrary. With an update period of 780 * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the 781 * last 100s. 782 */ 783 static void 784 zone_domain_update_wss(uma_zone_domain_t zdom) 785 { 786 long wss; 787 788 MPASS(zdom->uzd_imax >= zdom->uzd_imin); 789 wss = zdom->uzd_imax - zdom->uzd_imin; 790 zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems; 791 zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5; 792 } 793 794 /* 795 * Routine to perform timeout driven calculations. This expands the 796 * hashes and does per cpu statistics aggregation. 797 * 798 * Returns nothing. 799 */ 800 static void 801 zone_timeout(uma_zone_t zone, void *unused) 802 { 803 uma_keg_t keg; 804 u_int slabs, pages; 805 806 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0) 807 goto update_wss; 808 809 keg = zone->uz_keg; 810 811 /* 812 * Hash zones are non-numa by definition so the first domain 813 * is the only one present. 814 */ 815 KEG_LOCK(keg, 0); 816 pages = keg->uk_domain[0].ud_pages; 817 818 /* 819 * Expand the keg hash table. 820 * 821 * This is done if the number of slabs is larger than the hash size. 822 * What I'm trying to do here is completely reduce collisions. This 823 * may be a little aggressive. Should I allow for two collisions max? 824 */ 825 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) { 826 struct uma_hash newhash; 827 struct uma_hash oldhash; 828 int ret; 829 830 /* 831 * This is so involved because allocating and freeing 832 * while the keg lock is held will lead to deadlock. 833 * I have to do everything in stages and check for 834 * races. 835 */ 836 KEG_UNLOCK(keg, 0); 837 ret = hash_alloc(&newhash, 1 << fls(slabs)); 838 KEG_LOCK(keg, 0); 839 if (ret) { 840 if (hash_expand(&keg->uk_hash, &newhash)) { 841 oldhash = keg->uk_hash; 842 keg->uk_hash = newhash; 843 } else 844 oldhash = newhash; 845 846 KEG_UNLOCK(keg, 0); 847 hash_free(&oldhash); 848 goto update_wss; 849 } 850 } 851 KEG_UNLOCK(keg, 0); 852 853 update_wss: 854 ZONE_LOCK(zone); 855 for (int i = 0; i < vm_ndomains; i++) 856 zone_domain_update_wss(&zone->uz_domain[i]); 857 ZONE_UNLOCK(zone); 858 } 859 860 /* 861 * Allocate and zero fill the next sized hash table from the appropriate 862 * backing store. 863 * 864 * Arguments: 865 * hash A new hash structure with the old hash size in uh_hashsize 866 * 867 * Returns: 868 * 1 on success and 0 on failure. 869 */ 870 static int 871 hash_alloc(struct uma_hash *hash, u_int size) 872 { 873 size_t alloc; 874 875 KASSERT(powerof2(size), ("hash size must be power of 2")); 876 if (size > UMA_HASH_SIZE_INIT) { 877 hash->uh_hashsize = size; 878 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 879 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT); 880 } else { 881 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 882 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 883 UMA_ANYDOMAIN, M_WAITOK); 884 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 885 } 886 if (hash->uh_slab_hash) { 887 bzero(hash->uh_slab_hash, alloc); 888 hash->uh_hashmask = hash->uh_hashsize - 1; 889 return (1); 890 } 891 892 return (0); 893 } 894 895 /* 896 * Expands the hash table for HASH zones. This is done from zone_timeout 897 * to reduce collisions. This must not be done in the regular allocation 898 * path, otherwise, we can recurse on the vm while allocating pages. 899 * 900 * Arguments: 901 * oldhash The hash you want to expand 902 * newhash The hash structure for the new table 903 * 904 * Returns: 905 * Nothing 906 * 907 * Discussion: 908 */ 909 static int 910 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 911 { 912 uma_hash_slab_t slab; 913 u_int hval; 914 u_int idx; 915 916 if (!newhash->uh_slab_hash) 917 return (0); 918 919 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 920 return (0); 921 922 /* 923 * I need to investigate hash algorithms for resizing without a 924 * full rehash. 925 */ 926 927 for (idx = 0; idx < oldhash->uh_hashsize; idx++) 928 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) { 929 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]); 930 LIST_REMOVE(slab, uhs_hlink); 931 hval = UMA_HASH(newhash, slab->uhs_data); 932 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 933 slab, uhs_hlink); 934 } 935 936 return (1); 937 } 938 939 /* 940 * Free the hash bucket to the appropriate backing store. 941 * 942 * Arguments: 943 * slab_hash The hash bucket we're freeing 944 * hashsize The number of entries in that hash bucket 945 * 946 * Returns: 947 * Nothing 948 */ 949 static void 950 hash_free(struct uma_hash *hash) 951 { 952 if (hash->uh_slab_hash == NULL) 953 return; 954 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 955 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 956 else 957 free(hash->uh_slab_hash, M_UMAHASH); 958 } 959 960 /* 961 * Frees all outstanding items in a bucket 962 * 963 * Arguments: 964 * zone The zone to free to, must be unlocked. 965 * bucket The free/alloc bucket with items. 966 * 967 * Returns: 968 * Nothing 969 */ 970 971 static void 972 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 973 { 974 int i; 975 976 if (bucket == NULL || bucket->ub_cnt == 0) 977 return; 978 979 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && 980 bucket->ub_seq != SMR_SEQ_INVALID) { 981 smr_wait(zone->uz_smr, bucket->ub_seq); 982 for (i = 0; i < bucket->ub_cnt; i++) 983 item_dtor(zone, bucket->ub_bucket[i], 984 zone->uz_size, NULL, SKIP_NONE); 985 bucket->ub_seq = SMR_SEQ_INVALID; 986 } 987 if (zone->uz_fini) 988 for (i = 0; i < bucket->ub_cnt; i++) 989 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 990 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 991 if (zone->uz_max_items > 0) 992 zone_free_limit(zone, bucket->ub_cnt); 993 #ifdef INVARIANTS 994 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt); 995 #endif 996 bucket->ub_cnt = 0; 997 } 998 999 /* 1000 * Drains the per cpu caches for a zone. 1001 * 1002 * NOTE: This may only be called while the zone is being torn down, and not 1003 * during normal operation. This is necessary in order that we do not have 1004 * to migrate CPUs to drain the per-CPU caches. 1005 * 1006 * Arguments: 1007 * zone The zone to drain, must be unlocked. 1008 * 1009 * Returns: 1010 * Nothing 1011 */ 1012 static void 1013 cache_drain(uma_zone_t zone) 1014 { 1015 uma_cache_t cache; 1016 uma_bucket_t bucket; 1017 int cpu; 1018 1019 /* 1020 * XXX: It is safe to not lock the per-CPU caches, because we're 1021 * tearing down the zone anyway. I.e., there will be no further use 1022 * of the caches at this point. 1023 * 1024 * XXX: It would good to be able to assert that the zone is being 1025 * torn down to prevent improper use of cache_drain(). 1026 */ 1027 CPU_FOREACH(cpu) { 1028 cache = &zone->uz_cpu[cpu]; 1029 bucket = cache_bucket_unload_alloc(cache); 1030 if (bucket != NULL) { 1031 bucket_drain(zone, bucket); 1032 bucket_free(zone, bucket, NULL); 1033 } 1034 bucket = cache_bucket_unload_free(cache); 1035 if (bucket != NULL) { 1036 bucket_drain(zone, bucket); 1037 bucket_free(zone, bucket, NULL); 1038 } 1039 bucket = cache_bucket_unload_cross(cache); 1040 if (bucket != NULL) { 1041 bucket_drain(zone, bucket); 1042 bucket_free(zone, bucket, NULL); 1043 } 1044 } 1045 bucket_cache_reclaim(zone, true); 1046 } 1047 1048 static void 1049 cache_shrink(uma_zone_t zone, void *unused) 1050 { 1051 1052 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 1053 return; 1054 1055 ZONE_LOCK(zone); 1056 zone->uz_bucket_size = 1057 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2; 1058 ZONE_UNLOCK(zone); 1059 } 1060 1061 static void 1062 cache_drain_safe_cpu(uma_zone_t zone, void *unused) 1063 { 1064 uma_cache_t cache; 1065 uma_bucket_t b1, b2, b3; 1066 int domain; 1067 1068 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 1069 return; 1070 1071 b1 = b2 = b3 = NULL; 1072 ZONE_LOCK(zone); 1073 critical_enter(); 1074 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) 1075 domain = PCPU_GET(domain); 1076 else 1077 domain = 0; 1078 cache = &zone->uz_cpu[curcpu]; 1079 b1 = cache_bucket_unload_alloc(cache); 1080 if (b1 != NULL && b1->ub_cnt != 0) { 1081 zone_put_bucket(zone, &zone->uz_domain[domain], b1, false); 1082 b1 = NULL; 1083 } 1084 1085 /* 1086 * Don't flush SMR zone buckets. This leaves the zone without a 1087 * bucket and forces every free to synchronize(). 1088 */ 1089 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 1090 goto out; 1091 b2 = cache_bucket_unload_free(cache); 1092 if (b2 != NULL && b2->ub_cnt != 0) { 1093 zone_put_bucket(zone, &zone->uz_domain[domain], b2, false); 1094 b2 = NULL; 1095 } 1096 b3 = cache_bucket_unload_cross(cache); 1097 1098 out: 1099 critical_exit(); 1100 ZONE_UNLOCK(zone); 1101 if (b1) 1102 bucket_free(zone, b1, NULL); 1103 if (b2) 1104 bucket_free(zone, b2, NULL); 1105 if (b3) { 1106 bucket_drain(zone, b3); 1107 bucket_free(zone, b3, NULL); 1108 } 1109 } 1110 1111 /* 1112 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 1113 * This is an expensive call because it needs to bind to all CPUs 1114 * one by one and enter a critical section on each of them in order 1115 * to safely access their cache buckets. 1116 * Zone lock must not be held on call this function. 1117 */ 1118 static void 1119 pcpu_cache_drain_safe(uma_zone_t zone) 1120 { 1121 int cpu; 1122 1123 /* 1124 * Polite bucket sizes shrinking was not enough, shrink aggressively. 1125 */ 1126 if (zone) 1127 cache_shrink(zone, NULL); 1128 else 1129 zone_foreach(cache_shrink, NULL); 1130 1131 CPU_FOREACH(cpu) { 1132 thread_lock(curthread); 1133 sched_bind(curthread, cpu); 1134 thread_unlock(curthread); 1135 1136 if (zone) 1137 cache_drain_safe_cpu(zone, NULL); 1138 else 1139 zone_foreach(cache_drain_safe_cpu, NULL); 1140 } 1141 thread_lock(curthread); 1142 sched_unbind(curthread); 1143 thread_unlock(curthread); 1144 } 1145 1146 /* 1147 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller 1148 * requested a drain, otherwise the per-domain caches are trimmed to either 1149 * estimated working set size. 1150 */ 1151 static void 1152 bucket_cache_reclaim(uma_zone_t zone, bool drain) 1153 { 1154 uma_zone_domain_t zdom; 1155 uma_bucket_t bucket; 1156 long target, tofree; 1157 int i; 1158 1159 for (i = 0; i < vm_ndomains; i++) { 1160 /* 1161 * The cross bucket is partially filled and not part of 1162 * the item count. Reclaim it individually here. 1163 */ 1164 zdom = &zone->uz_domain[i]; 1165 ZONE_CROSS_LOCK(zone); 1166 bucket = zdom->uzd_cross; 1167 zdom->uzd_cross = NULL; 1168 ZONE_CROSS_UNLOCK(zone); 1169 if (bucket != NULL) { 1170 bucket_drain(zone, bucket); 1171 bucket_free(zone, bucket, NULL); 1172 } 1173 1174 /* 1175 * Shrink the zone bucket size to ensure that the per-CPU caches 1176 * don't grow too large. 1177 */ 1178 ZONE_LOCK(zone); 1179 if (i == 0 && zone->uz_bucket_size > zone->uz_bucket_size_min) 1180 zone->uz_bucket_size--; 1181 1182 /* 1183 * If we were asked to drain the zone, we are done only once 1184 * this bucket cache is empty. Otherwise, we reclaim items in 1185 * excess of the zone's estimated working set size. If the 1186 * difference nitems - imin is larger than the WSS estimate, 1187 * then the estimate will grow at the end of this interval and 1188 * we ignore the historical average. 1189 */ 1190 target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems - 1191 zdom->uzd_imin); 1192 while (zdom->uzd_nitems > target) { 1193 bucket = STAILQ_FIRST(&zdom->uzd_buckets); 1194 if (bucket == NULL) 1195 break; 1196 tofree = bucket->ub_cnt; 1197 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link); 1198 zdom->uzd_nitems -= tofree; 1199 1200 /* 1201 * Shift the bounds of the current WSS interval to avoid 1202 * perturbing the estimate. 1203 */ 1204 zdom->uzd_imax -= lmin(zdom->uzd_imax, tofree); 1205 zdom->uzd_imin -= lmin(zdom->uzd_imin, tofree); 1206 1207 ZONE_UNLOCK(zone); 1208 bucket_drain(zone, bucket); 1209 bucket_free(zone, bucket, NULL); 1210 ZONE_LOCK(zone); 1211 } 1212 ZONE_UNLOCK(zone); 1213 } 1214 } 1215 1216 static void 1217 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 1218 { 1219 uint8_t *mem; 1220 int i; 1221 uint8_t flags; 1222 1223 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes", 1224 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera); 1225 1226 mem = slab_data(slab, keg); 1227 flags = slab->us_flags; 1228 i = start; 1229 if (keg->uk_fini != NULL) { 1230 for (i--; i > -1; i--) 1231 #ifdef INVARIANTS 1232 /* 1233 * trash_fini implies that dtor was trash_dtor. trash_fini 1234 * would check that memory hasn't been modified since free, 1235 * which executed trash_dtor. 1236 * That's why we need to run uma_dbg_kskip() check here, 1237 * albeit we don't make skip check for other init/fini 1238 * invocations. 1239 */ 1240 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) || 1241 keg->uk_fini != trash_fini) 1242 #endif 1243 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size); 1244 } 1245 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) 1246 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab), 1247 NULL, SKIP_NONE); 1248 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 1249 uma_total_dec(PAGE_SIZE * keg->uk_ppera); 1250 } 1251 1252 /* 1253 * Frees pages from a keg back to the system. This is done on demand from 1254 * the pageout daemon. 1255 * 1256 * Returns nothing. 1257 */ 1258 static void 1259 keg_drain(uma_keg_t keg) 1260 { 1261 struct slabhead freeslabs = { 0 }; 1262 uma_domain_t dom; 1263 uma_slab_t slab, tmp; 1264 int i, n; 1265 1266 /* 1267 * We don't want to take pages from statically allocated kegs at this 1268 * time 1269 */ 1270 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 1271 return; 1272 1273 for (i = 0; i < vm_ndomains; i++) { 1274 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u", 1275 keg->uk_name, keg, i, dom->ud_free); 1276 n = 0; 1277 dom = &keg->uk_domain[i]; 1278 KEG_LOCK(keg, i); 1279 LIST_FOREACH_SAFE(slab, &dom->ud_free_slab, us_link, tmp) { 1280 if (keg->uk_flags & UMA_ZFLAG_HASH) 1281 UMA_HASH_REMOVE(&keg->uk_hash, slab); 1282 n++; 1283 LIST_REMOVE(slab, us_link); 1284 LIST_INSERT_HEAD(&freeslabs, slab, us_link); 1285 } 1286 dom->ud_pages -= n * keg->uk_ppera; 1287 dom->ud_free -= n * keg->uk_ipers; 1288 KEG_UNLOCK(keg, i); 1289 } 1290 1291 while ((slab = LIST_FIRST(&freeslabs)) != NULL) { 1292 LIST_REMOVE(slab, us_link); 1293 keg_free_slab(keg, slab, keg->uk_ipers); 1294 } 1295 } 1296 1297 static void 1298 zone_reclaim(uma_zone_t zone, int waitok, bool drain) 1299 { 1300 1301 /* 1302 * Set draining to interlock with zone_dtor() so we can release our 1303 * locks as we go. Only dtor() should do a WAITOK call since it 1304 * is the only call that knows the structure will still be available 1305 * when it wakes up. 1306 */ 1307 ZONE_LOCK(zone); 1308 while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) { 1309 if (waitok == M_NOWAIT) 1310 goto out; 1311 msleep(zone, &zone->uz_lock, PVM, "zonedrain", 1); 1312 } 1313 zone->uz_flags |= UMA_ZFLAG_RECLAIMING; 1314 ZONE_UNLOCK(zone); 1315 bucket_cache_reclaim(zone, drain); 1316 1317 /* 1318 * The DRAINING flag protects us from being freed while 1319 * we're running. Normally the uma_rwlock would protect us but we 1320 * must be able to release and acquire the right lock for each keg. 1321 */ 1322 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) 1323 keg_drain(zone->uz_keg); 1324 ZONE_LOCK(zone); 1325 zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING; 1326 wakeup(zone); 1327 out: 1328 ZONE_UNLOCK(zone); 1329 } 1330 1331 static void 1332 zone_drain(uma_zone_t zone, void *unused) 1333 { 1334 1335 zone_reclaim(zone, M_NOWAIT, true); 1336 } 1337 1338 static void 1339 zone_trim(uma_zone_t zone, void *unused) 1340 { 1341 1342 zone_reclaim(zone, M_NOWAIT, false); 1343 } 1344 1345 /* 1346 * Allocate a new slab for a keg and inserts it into the partial slab list. 1347 * The keg should be unlocked on entry. If the allocation succeeds it will 1348 * be locked on return. 1349 * 1350 * Arguments: 1351 * flags Wait flags for the item initialization routine 1352 * aflags Wait flags for the slab allocation 1353 * 1354 * Returns: 1355 * The slab that was allocated or NULL if there is no memory and the 1356 * caller specified M_NOWAIT. 1357 */ 1358 static uma_slab_t 1359 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags, 1360 int aflags) 1361 { 1362 uma_domain_t dom; 1363 uma_alloc allocf; 1364 uma_slab_t slab; 1365 unsigned long size; 1366 uint8_t *mem; 1367 uint8_t sflags; 1368 int i; 1369 1370 KASSERT(domain >= 0 && domain < vm_ndomains, 1371 ("keg_alloc_slab: domain %d out of range", domain)); 1372 1373 allocf = keg->uk_allocf; 1374 slab = NULL; 1375 mem = NULL; 1376 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) { 1377 uma_hash_slab_t hslab; 1378 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL, 1379 domain, aflags); 1380 if (hslab == NULL) 1381 goto fail; 1382 slab = &hslab->uhs_slab; 1383 } 1384 1385 /* 1386 * This reproduces the old vm_zone behavior of zero filling pages the 1387 * first time they are added to a zone. 1388 * 1389 * Malloced items are zeroed in uma_zalloc. 1390 */ 1391 1392 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 1393 aflags |= M_ZERO; 1394 else 1395 aflags &= ~M_ZERO; 1396 1397 if (keg->uk_flags & UMA_ZONE_NODUMP) 1398 aflags |= M_NODUMP; 1399 1400 /* zone is passed for legacy reasons. */ 1401 size = keg->uk_ppera * PAGE_SIZE; 1402 mem = allocf(zone, size, domain, &sflags, aflags); 1403 if (mem == NULL) { 1404 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) 1405 zone_free_item(slabzone(keg->uk_ipers), 1406 slab_tohashslab(slab), NULL, SKIP_NONE); 1407 goto fail; 1408 } 1409 uma_total_inc(size); 1410 1411 /* For HASH zones all pages go to the same uma_domain. */ 1412 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) 1413 domain = 0; 1414 1415 /* Point the slab into the allocated memory */ 1416 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) 1417 slab = (uma_slab_t )(mem + keg->uk_pgoff); 1418 else 1419 slab_tohashslab(slab)->uhs_data = mem; 1420 1421 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB) 1422 for (i = 0; i < keg->uk_ppera; i++) 1423 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE), 1424 zone, slab); 1425 1426 slab->us_freecount = keg->uk_ipers; 1427 slab->us_flags = sflags; 1428 slab->us_domain = domain; 1429 1430 BIT_FILL(keg->uk_ipers, &slab->us_free); 1431 #ifdef INVARIANTS 1432 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg)); 1433 #endif 1434 1435 if (keg->uk_init != NULL) { 1436 for (i = 0; i < keg->uk_ipers; i++) 1437 if (keg->uk_init(slab_item(slab, keg, i), 1438 keg->uk_size, flags) != 0) 1439 break; 1440 if (i != keg->uk_ipers) { 1441 keg_free_slab(keg, slab, i); 1442 goto fail; 1443 } 1444 } 1445 KEG_LOCK(keg, domain); 1446 1447 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)", 1448 slab, keg->uk_name, keg); 1449 1450 if (keg->uk_flags & UMA_ZFLAG_HASH) 1451 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1452 1453 /* 1454 * If we got a slab here it's safe to mark it partially used 1455 * and return. We assume that the caller is going to remove 1456 * at least one item. 1457 */ 1458 dom = &keg->uk_domain[domain]; 1459 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 1460 dom->ud_pages += keg->uk_ppera; 1461 dom->ud_free += keg->uk_ipers; 1462 1463 return (slab); 1464 1465 fail: 1466 return (NULL); 1467 } 1468 1469 /* 1470 * This function is intended to be used early on in place of page_alloc() so 1471 * that we may use the boot time page cache to satisfy allocations before 1472 * the VM is ready. 1473 */ 1474 static void * 1475 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1476 int wait) 1477 { 1478 vm_paddr_t pa; 1479 vm_page_t m; 1480 void *mem; 1481 int pages; 1482 int i; 1483 1484 pages = howmany(bytes, PAGE_SIZE); 1485 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__)); 1486 1487 *pflag = UMA_SLAB_BOOT; 1488 m = vm_page_alloc_contig_domain(NULL, 0, domain, 1489 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages, 1490 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT); 1491 if (m == NULL) 1492 return (NULL); 1493 1494 pa = VM_PAGE_TO_PHYS(m); 1495 for (i = 0; i < pages; i++, pa += PAGE_SIZE) { 1496 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 1497 defined(__riscv) || defined(__powerpc64__) 1498 if ((wait & M_NODUMP) == 0) 1499 dump_add_page(pa); 1500 #endif 1501 } 1502 /* Allocate KVA and indirectly advance bootmem. */ 1503 mem = (void *)pmap_map(&bootmem, m->phys_addr, 1504 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE); 1505 if ((wait & M_ZERO) != 0) 1506 bzero(mem, pages * PAGE_SIZE); 1507 1508 return (mem); 1509 } 1510 1511 static void 1512 startup_free(void *mem, vm_size_t bytes) 1513 { 1514 vm_offset_t va; 1515 vm_page_t m; 1516 1517 va = (vm_offset_t)mem; 1518 m = PHYS_TO_VM_PAGE(pmap_kextract(va)); 1519 pmap_remove(kernel_pmap, va, va + bytes); 1520 for (; bytes != 0; bytes -= PAGE_SIZE, m++) { 1521 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 1522 defined(__riscv) || defined(__powerpc64__) 1523 dump_drop_page(VM_PAGE_TO_PHYS(m)); 1524 #endif 1525 vm_page_unwire_noq(m); 1526 vm_page_free(m); 1527 } 1528 } 1529 1530 /* 1531 * Allocates a number of pages from the system 1532 * 1533 * Arguments: 1534 * bytes The number of bytes requested 1535 * wait Shall we wait? 1536 * 1537 * Returns: 1538 * A pointer to the alloced memory or possibly 1539 * NULL if M_NOWAIT is set. 1540 */ 1541 static void * 1542 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1543 int wait) 1544 { 1545 void *p; /* Returned page */ 1546 1547 *pflag = UMA_SLAB_KERNEL; 1548 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait); 1549 1550 return (p); 1551 } 1552 1553 static void * 1554 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1555 int wait) 1556 { 1557 struct pglist alloctail; 1558 vm_offset_t addr, zkva; 1559 int cpu, flags; 1560 vm_page_t p, p_next; 1561 #ifdef NUMA 1562 struct pcpu *pc; 1563 #endif 1564 1565 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE); 1566 1567 TAILQ_INIT(&alloctail); 1568 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1569 malloc2vm_flags(wait); 1570 *pflag = UMA_SLAB_KERNEL; 1571 for (cpu = 0; cpu <= mp_maxid; cpu++) { 1572 if (CPU_ABSENT(cpu)) { 1573 p = vm_page_alloc(NULL, 0, flags); 1574 } else { 1575 #ifndef NUMA 1576 p = vm_page_alloc(NULL, 0, flags); 1577 #else 1578 pc = pcpu_find(cpu); 1579 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain))) 1580 p = NULL; 1581 else 1582 p = vm_page_alloc_domain(NULL, 0, 1583 pc->pc_domain, flags); 1584 if (__predict_false(p == NULL)) 1585 p = vm_page_alloc(NULL, 0, flags); 1586 #endif 1587 } 1588 if (__predict_false(p == NULL)) 1589 goto fail; 1590 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1591 } 1592 if ((addr = kva_alloc(bytes)) == 0) 1593 goto fail; 1594 zkva = addr; 1595 TAILQ_FOREACH(p, &alloctail, listq) { 1596 pmap_qenter(zkva, &p, 1); 1597 zkva += PAGE_SIZE; 1598 } 1599 return ((void*)addr); 1600 fail: 1601 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1602 vm_page_unwire_noq(p); 1603 vm_page_free(p); 1604 } 1605 return (NULL); 1606 } 1607 1608 /* 1609 * Allocates a number of pages from within an object 1610 * 1611 * Arguments: 1612 * bytes The number of bytes requested 1613 * wait Shall we wait? 1614 * 1615 * Returns: 1616 * A pointer to the alloced memory or possibly 1617 * NULL if M_NOWAIT is set. 1618 */ 1619 static void * 1620 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, 1621 int wait) 1622 { 1623 TAILQ_HEAD(, vm_page) alloctail; 1624 u_long npages; 1625 vm_offset_t retkva, zkva; 1626 vm_page_t p, p_next; 1627 uma_keg_t keg; 1628 1629 TAILQ_INIT(&alloctail); 1630 keg = zone->uz_keg; 1631 1632 npages = howmany(bytes, PAGE_SIZE); 1633 while (npages > 0) { 1634 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT | 1635 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1636 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK : 1637 VM_ALLOC_NOWAIT)); 1638 if (p != NULL) { 1639 /* 1640 * Since the page does not belong to an object, its 1641 * listq is unused. 1642 */ 1643 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1644 npages--; 1645 continue; 1646 } 1647 /* 1648 * Page allocation failed, free intermediate pages and 1649 * exit. 1650 */ 1651 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1652 vm_page_unwire_noq(p); 1653 vm_page_free(p); 1654 } 1655 return (NULL); 1656 } 1657 *flags = UMA_SLAB_PRIV; 1658 zkva = keg->uk_kva + 1659 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1660 retkva = zkva; 1661 TAILQ_FOREACH(p, &alloctail, listq) { 1662 pmap_qenter(zkva, &p, 1); 1663 zkva += PAGE_SIZE; 1664 } 1665 1666 return ((void *)retkva); 1667 } 1668 1669 /* 1670 * Allocate physically contiguous pages. 1671 */ 1672 static void * 1673 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1674 int wait) 1675 { 1676 1677 *pflag = UMA_SLAB_KERNEL; 1678 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain), 1679 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT)); 1680 } 1681 1682 /* 1683 * Frees a number of pages to the system 1684 * 1685 * Arguments: 1686 * mem A pointer to the memory to be freed 1687 * size The size of the memory being freed 1688 * flags The original p->us_flags field 1689 * 1690 * Returns: 1691 * Nothing 1692 */ 1693 static void 1694 page_free(void *mem, vm_size_t size, uint8_t flags) 1695 { 1696 1697 if ((flags & UMA_SLAB_BOOT) != 0) { 1698 startup_free(mem, size); 1699 return; 1700 } 1701 1702 KASSERT((flags & UMA_SLAB_KERNEL) != 0, 1703 ("UMA: page_free used with invalid flags %x", flags)); 1704 1705 kmem_free((vm_offset_t)mem, size); 1706 } 1707 1708 /* 1709 * Frees pcpu zone allocations 1710 * 1711 * Arguments: 1712 * mem A pointer to the memory to be freed 1713 * size The size of the memory being freed 1714 * flags The original p->us_flags field 1715 * 1716 * Returns: 1717 * Nothing 1718 */ 1719 static void 1720 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags) 1721 { 1722 vm_offset_t sva, curva; 1723 vm_paddr_t paddr; 1724 vm_page_t m; 1725 1726 MPASS(size == (mp_maxid+1)*PAGE_SIZE); 1727 1728 if ((flags & UMA_SLAB_BOOT) != 0) { 1729 startup_free(mem, size); 1730 return; 1731 } 1732 1733 sva = (vm_offset_t)mem; 1734 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) { 1735 paddr = pmap_kextract(curva); 1736 m = PHYS_TO_VM_PAGE(paddr); 1737 vm_page_unwire_noq(m); 1738 vm_page_free(m); 1739 } 1740 pmap_qremove(sva, size >> PAGE_SHIFT); 1741 kva_free(sva, size); 1742 } 1743 1744 1745 /* 1746 * Zero fill initializer 1747 * 1748 * Arguments/Returns follow uma_init specifications 1749 */ 1750 static int 1751 zero_init(void *mem, int size, int flags) 1752 { 1753 bzero(mem, size); 1754 return (0); 1755 } 1756 1757 #ifdef INVARIANTS 1758 struct noslabbits * 1759 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg) 1760 { 1761 1762 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers))); 1763 } 1764 #endif 1765 1766 /* 1767 * Actual size of embedded struct slab (!OFFPAGE). 1768 */ 1769 size_t 1770 slab_sizeof(int nitems) 1771 { 1772 size_t s; 1773 1774 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS; 1775 return (roundup(s, UMA_ALIGN_PTR + 1)); 1776 } 1777 1778 /* 1779 * Size of memory for embedded slabs (!OFFPAGE). 1780 */ 1781 size_t 1782 slab_space(int nitems) 1783 { 1784 return (UMA_SLAB_SIZE - slab_sizeof(nitems)); 1785 } 1786 1787 #define UMA_FIXPT_SHIFT 31 1788 #define UMA_FRAC_FIXPT(n, d) \ 1789 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d))) 1790 #define UMA_FIXPT_PCT(f) \ 1791 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT)) 1792 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100) 1793 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE) 1794 1795 /* 1796 * Compute the number of items that will fit in a slab. If hdr is true, the 1797 * item count may be limited to provide space in the slab for an inline slab 1798 * header. Otherwise, all slab space will be provided for item storage. 1799 */ 1800 static u_int 1801 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr) 1802 { 1803 u_int ipers; 1804 u_int padpi; 1805 1806 /* The padding between items is not needed after the last item. */ 1807 padpi = rsize - size; 1808 1809 if (hdr) { 1810 /* 1811 * Start with the maximum item count and remove items until 1812 * the slab header first alongside the allocatable memory. 1813 */ 1814 for (ipers = MIN(SLAB_MAX_SETSIZE, 1815 (slabsize + padpi - slab_sizeof(1)) / rsize); 1816 ipers > 0 && 1817 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize; 1818 ipers--) 1819 continue; 1820 } else { 1821 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE); 1822 } 1823 1824 return (ipers); 1825 } 1826 1827 /* 1828 * Compute the number of items that will fit in a slab for a startup zone. 1829 */ 1830 int 1831 slab_ipers(size_t size, int align) 1832 { 1833 int rsize; 1834 1835 rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */ 1836 return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true)); 1837 } 1838 1839 struct keg_layout_result { 1840 u_int format; 1841 u_int slabsize; 1842 u_int ipers; 1843 u_int eff; 1844 }; 1845 1846 static void 1847 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt, 1848 struct keg_layout_result *kl) 1849 { 1850 u_int total; 1851 1852 kl->format = fmt; 1853 kl->slabsize = slabsize; 1854 1855 /* Handle INTERNAL as inline with an extra page. */ 1856 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) { 1857 kl->format &= ~UMA_ZFLAG_INTERNAL; 1858 kl->slabsize += PAGE_SIZE; 1859 } 1860 1861 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize, 1862 (fmt & UMA_ZFLAG_OFFPAGE) == 0); 1863 1864 /* Account for memory used by an offpage slab header. */ 1865 total = kl->slabsize; 1866 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0) 1867 total += slabzone(kl->ipers)->uz_keg->uk_rsize; 1868 1869 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total); 1870 } 1871 1872 /* 1873 * Determine the format of a uma keg. This determines where the slab header 1874 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera. 1875 * 1876 * Arguments 1877 * keg The zone we should initialize 1878 * 1879 * Returns 1880 * Nothing 1881 */ 1882 static void 1883 keg_layout(uma_keg_t keg) 1884 { 1885 struct keg_layout_result kl = {}, kl_tmp; 1886 u_int fmts[2]; 1887 u_int alignsize; 1888 u_int nfmt; 1889 u_int pages; 1890 u_int rsize; 1891 u_int slabsize; 1892 u_int i, j; 1893 1894 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1895 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE && 1896 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0), 1897 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b", 1898 __func__, keg->uk_name, keg->uk_size, keg->uk_flags, 1899 PRINT_UMA_ZFLAGS)); 1900 KASSERT((keg->uk_flags & 1901 (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY)) == 0 || 1902 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0, 1903 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags, 1904 PRINT_UMA_ZFLAGS)); 1905 1906 alignsize = keg->uk_align + 1; 1907 1908 /* 1909 * Calculate the size of each allocation (rsize) according to 1910 * alignment. If the requested size is smaller than we have 1911 * allocation bits for we round it up. 1912 */ 1913 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT); 1914 rsize = roundup2(rsize, alignsize); 1915 1916 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) { 1917 /* 1918 * We want one item to start on every align boundary in a page. 1919 * To do this we will span pages. We will also extend the item 1920 * by the size of align if it is an even multiple of align. 1921 * Otherwise, it would fall on the same boundary every time. 1922 */ 1923 if ((rsize & alignsize) == 0) 1924 rsize += alignsize; 1925 slabsize = rsize * (PAGE_SIZE / alignsize); 1926 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE); 1927 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE); 1928 slabsize = round_page(slabsize); 1929 } else { 1930 /* 1931 * Start with a slab size of as many pages as it takes to 1932 * represent a single item. We will try to fit as many 1933 * additional items into the slab as possible. 1934 */ 1935 slabsize = round_page(keg->uk_size); 1936 } 1937 1938 /* Build a list of all of the available formats for this keg. */ 1939 nfmt = 0; 1940 1941 /* Evaluate an inline slab layout. */ 1942 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0) 1943 fmts[nfmt++] = 0; 1944 1945 /* TODO: vm_page-embedded slab. */ 1946 1947 /* 1948 * We can't do OFFPAGE if we're internal or if we've been 1949 * asked to not go to the VM for buckets. If we do this we 1950 * may end up going to the VM for slabs which we do not 1951 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1952 * of UMA_ZONE_VM, which clearly forbids it. In those cases, 1953 * evaluate a pseudo-format called INTERNAL which has an inline 1954 * slab header and one extra page to guarantee that it fits. 1955 * 1956 * Otherwise, see if using an OFFPAGE slab will improve our 1957 * efficiency. 1958 */ 1959 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZFLAG_CACHEONLY)) != 0) 1960 fmts[nfmt++] = UMA_ZFLAG_INTERNAL; 1961 else 1962 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE; 1963 1964 /* 1965 * Choose a slab size and format which satisfy the minimum efficiency. 1966 * Prefer the smallest slab size that meets the constraints. 1967 * 1968 * Start with a minimum slab size, to accommodate CACHESPREAD. Then, 1969 * for small items (up to PAGE_SIZE), the iteration increment is one 1970 * page; and for large items, the increment is one item. 1971 */ 1972 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize); 1973 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u", 1974 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize, 1975 rsize, i)); 1976 for ( ; ; i++) { 1977 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) : 1978 round_page(rsize * (i - 1) + keg->uk_size); 1979 1980 for (j = 0; j < nfmt; j++) { 1981 /* Only if we have no viable format yet. */ 1982 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 && 1983 kl.ipers > 0) 1984 continue; 1985 1986 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp); 1987 if (kl_tmp.eff <= kl.eff) 1988 continue; 1989 1990 kl = kl_tmp; 1991 1992 CTR6(KTR_UMA, "keg %s layout: format %#x " 1993 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff", 1994 keg->uk_name, kl.format, kl.ipers, rsize, 1995 kl.slabsize, UMA_FIXPT_PCT(kl.eff)); 1996 1997 /* Stop when we reach the minimum efficiency. */ 1998 if (kl.eff >= UMA_MIN_EFF) 1999 break; 2000 } 2001 2002 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs || 2003 slabsize >= SLAB_MAX_SETSIZE * rsize || 2004 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0) 2005 break; 2006 } 2007 2008 pages = atop(kl.slabsize); 2009 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) 2010 pages *= mp_maxid + 1; 2011 2012 keg->uk_rsize = rsize; 2013 keg->uk_ipers = kl.ipers; 2014 keg->uk_ppera = pages; 2015 keg->uk_flags |= kl.format; 2016 2017 /* 2018 * How do we find the slab header if it is offpage or if not all item 2019 * start addresses are in the same page? We could solve the latter 2020 * case with vaddr alignment, but we don't. 2021 */ 2022 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 || 2023 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) { 2024 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0) 2025 keg->uk_flags |= UMA_ZFLAG_HASH; 2026 else 2027 keg->uk_flags |= UMA_ZFLAG_VTOSLAB; 2028 } 2029 2030 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u", 2031 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers, 2032 pages); 2033 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE, 2034 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__, 2035 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize, 2036 keg->uk_ipers, pages)); 2037 } 2038 2039 /* 2040 * Keg header ctor. This initializes all fields, locks, etc. And inserts 2041 * the keg onto the global keg list. 2042 * 2043 * Arguments/Returns follow uma_ctor specifications 2044 * udata Actually uma_kctor_args 2045 */ 2046 static int 2047 keg_ctor(void *mem, int size, void *udata, int flags) 2048 { 2049 struct uma_kctor_args *arg = udata; 2050 uma_keg_t keg = mem; 2051 uma_zone_t zone; 2052 int i; 2053 2054 bzero(keg, size); 2055 keg->uk_size = arg->size; 2056 keg->uk_init = arg->uminit; 2057 keg->uk_fini = arg->fini; 2058 keg->uk_align = arg->align; 2059 keg->uk_reserve = 0; 2060 keg->uk_flags = arg->flags; 2061 2062 /* 2063 * We use a global round-robin policy by default. Zones with 2064 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which 2065 * case the iterator is never run. 2066 */ 2067 keg->uk_dr.dr_policy = DOMAINSET_RR(); 2068 keg->uk_dr.dr_iter = 0; 2069 2070 /* 2071 * The master zone is passed to us at keg-creation time. 2072 */ 2073 zone = arg->zone; 2074 keg->uk_name = zone->uz_name; 2075 2076 if (arg->flags & UMA_ZONE_VM) 2077 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 2078 2079 if (arg->flags & UMA_ZONE_ZINIT) 2080 keg->uk_init = zero_init; 2081 2082 if (arg->flags & UMA_ZONE_MALLOC) 2083 keg->uk_flags |= UMA_ZFLAG_VTOSLAB; 2084 2085 #ifndef SMP 2086 keg->uk_flags &= ~UMA_ZONE_PCPU; 2087 #endif 2088 2089 keg_layout(keg); 2090 2091 /* 2092 * Use a first-touch NUMA policy for all kegs that pmap_extract() 2093 * will work on with the exception of critical VM structures 2094 * necessary for paging. 2095 * 2096 * Zones may override the default by specifying either. 2097 */ 2098 #ifdef NUMA 2099 if ((keg->uk_flags & 2100 (UMA_ZFLAG_HASH | UMA_ZONE_VM | UMA_ZONE_ROUNDROBIN)) == 0) 2101 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH; 2102 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0) 2103 keg->uk_flags |= UMA_ZONE_ROUNDROBIN; 2104 #endif 2105 2106 /* 2107 * If we haven't booted yet we need allocations to go through the 2108 * startup cache until the vm is ready. 2109 */ 2110 #ifdef UMA_MD_SMALL_ALLOC 2111 if (keg->uk_ppera == 1) 2112 keg->uk_allocf = uma_small_alloc; 2113 else 2114 #endif 2115 if (booted < BOOT_KVA) 2116 keg->uk_allocf = startup_alloc; 2117 else if (keg->uk_flags & UMA_ZONE_PCPU) 2118 keg->uk_allocf = pcpu_page_alloc; 2119 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1) 2120 keg->uk_allocf = contig_alloc; 2121 else 2122 keg->uk_allocf = page_alloc; 2123 #ifdef UMA_MD_SMALL_ALLOC 2124 if (keg->uk_ppera == 1) 2125 keg->uk_freef = uma_small_free; 2126 else 2127 #endif 2128 if (keg->uk_flags & UMA_ZONE_PCPU) 2129 keg->uk_freef = pcpu_page_free; 2130 else 2131 keg->uk_freef = page_free; 2132 2133 /* 2134 * Initialize keg's locks. 2135 */ 2136 for (i = 0; i < vm_ndomains; i++) 2137 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS)); 2138 2139 /* 2140 * If we're putting the slab header in the actual page we need to 2141 * figure out where in each page it goes. See slab_sizeof 2142 * definition. 2143 */ 2144 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) { 2145 size_t shsize; 2146 2147 shsize = slab_sizeof(keg->uk_ipers); 2148 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize; 2149 /* 2150 * The only way the following is possible is if with our 2151 * UMA_ALIGN_PTR adjustments we are now bigger than 2152 * UMA_SLAB_SIZE. I haven't checked whether this is 2153 * mathematically possible for all cases, so we make 2154 * sure here anyway. 2155 */ 2156 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera, 2157 ("zone %s ipers %d rsize %d size %d slab won't fit", 2158 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size)); 2159 } 2160 2161 if (keg->uk_flags & UMA_ZFLAG_HASH) 2162 hash_alloc(&keg->uk_hash, 0); 2163 2164 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone); 2165 2166 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 2167 2168 rw_wlock(&uma_rwlock); 2169 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 2170 rw_wunlock(&uma_rwlock); 2171 return (0); 2172 } 2173 2174 static void 2175 zone_kva_available(uma_zone_t zone, void *unused) 2176 { 2177 uma_keg_t keg; 2178 2179 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 2180 return; 2181 KEG_GET(zone, keg); 2182 2183 if (keg->uk_allocf == startup_alloc) { 2184 /* Switch to the real allocator. */ 2185 if (keg->uk_flags & UMA_ZONE_PCPU) 2186 keg->uk_allocf = pcpu_page_alloc; 2187 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && 2188 keg->uk_ppera > 1) 2189 keg->uk_allocf = contig_alloc; 2190 else 2191 keg->uk_allocf = page_alloc; 2192 } 2193 } 2194 2195 static void 2196 zone_alloc_counters(uma_zone_t zone, void *unused) 2197 { 2198 2199 zone->uz_allocs = counter_u64_alloc(M_WAITOK); 2200 zone->uz_frees = counter_u64_alloc(M_WAITOK); 2201 zone->uz_fails = counter_u64_alloc(M_WAITOK); 2202 } 2203 2204 static void 2205 zone_alloc_sysctl(uma_zone_t zone, void *unused) 2206 { 2207 uma_zone_domain_t zdom; 2208 uma_domain_t dom; 2209 uma_keg_t keg; 2210 struct sysctl_oid *oid, *domainoid; 2211 int domains, i, cnt; 2212 static const char *nokeg = "cache zone"; 2213 char *c; 2214 2215 /* 2216 * Make a sysctl safe copy of the zone name by removing 2217 * any special characters and handling dups by appending 2218 * an index. 2219 */ 2220 if (zone->uz_namecnt != 0) { 2221 /* Count the number of decimal digits and '_' separator. */ 2222 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++) 2223 cnt /= 10; 2224 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1, 2225 M_UMA, M_WAITOK); 2226 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name, 2227 zone->uz_namecnt); 2228 } else 2229 zone->uz_ctlname = strdup(zone->uz_name, M_UMA); 2230 for (c = zone->uz_ctlname; *c != '\0'; c++) 2231 if (strchr("./\\ -", *c) != NULL) 2232 *c = '_'; 2233 2234 /* 2235 * Basic parameters at the root. 2236 */ 2237 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma), 2238 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD, NULL, ""); 2239 oid = zone->uz_oid; 2240 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2241 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size"); 2242 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2243 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE, 2244 zone, 0, sysctl_handle_uma_zone_flags, "A", 2245 "Allocator configuration flags"); 2246 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2247 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0, 2248 "Desired per-cpu cache size"); 2249 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2250 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0, 2251 "Maximum allowed per-cpu cache size"); 2252 2253 /* 2254 * keg if present. 2255 */ 2256 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0) 2257 domains = vm_ndomains; 2258 else 2259 domains = 1; 2260 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 2261 "keg", CTLFLAG_RD, NULL, ""); 2262 keg = zone->uz_keg; 2263 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) { 2264 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2265 "name", CTLFLAG_RD, keg->uk_name, "Keg name"); 2266 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2267 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0, 2268 "Real object size with alignment"); 2269 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2270 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0, 2271 "pages per-slab allocation"); 2272 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2273 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0, 2274 "items available per-slab"); 2275 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2276 "align", CTLFLAG_RD, &keg->uk_align, 0, 2277 "item alignment mask"); 2278 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2279 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE, 2280 keg, 0, sysctl_handle_uma_slab_efficiency, "I", 2281 "Slab utilization (100 - internal fragmentation %)"); 2282 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid), 2283 OID_AUTO, "domain", CTLFLAG_RD, NULL, ""); 2284 for (i = 0; i < domains; i++) { 2285 dom = &keg->uk_domain[i]; 2286 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid), 2287 OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD, 2288 NULL, ""); 2289 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2290 "pages", CTLFLAG_RD, &dom->ud_pages, 0, 2291 "Total pages currently allocated from VM"); 2292 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2293 "free", CTLFLAG_RD, &dom->ud_free, 0, 2294 "items free in the slab layer"); 2295 } 2296 } else 2297 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2298 "name", CTLFLAG_RD, nokeg, "Keg name"); 2299 2300 /* 2301 * Information about zone limits. 2302 */ 2303 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 2304 "limit", CTLFLAG_RD, NULL, ""); 2305 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2306 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2307 zone, 0, sysctl_handle_uma_zone_items, "QU", 2308 "current number of allocated items if limit is set"); 2309 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2310 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0, 2311 "Maximum number of cached items"); 2312 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2313 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0, 2314 "Number of threads sleeping at limit"); 2315 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2316 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0, 2317 "Total zone limit sleeps"); 2318 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2319 "bucket_max", CTLFLAG_RD, &zone->uz_bkt_max, 0, 2320 "Maximum number of items in the bucket cache"); 2321 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2322 "bucket_cnt", CTLFLAG_RD, &zone->uz_bkt_count, 0, 2323 "Number of items in the bucket cache"); 2324 2325 /* 2326 * Per-domain zone information. 2327 */ 2328 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), 2329 OID_AUTO, "domain", CTLFLAG_RD, NULL, ""); 2330 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0) 2331 domains = 1; 2332 for (i = 0; i < domains; i++) { 2333 zdom = &zone->uz_domain[i]; 2334 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid), 2335 OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD, NULL, ""); 2336 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2337 "nitems", CTLFLAG_RD, &zdom->uzd_nitems, 2338 "number of items in this domain"); 2339 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2340 "imax", CTLFLAG_RD, &zdom->uzd_imax, 2341 "maximum item count in this period"); 2342 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2343 "imin", CTLFLAG_RD, &zdom->uzd_imin, 2344 "minimum item count in this period"); 2345 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2346 "wss", CTLFLAG_RD, &zdom->uzd_wss, 2347 "Working set size"); 2348 } 2349 2350 /* 2351 * General statistics. 2352 */ 2353 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 2354 "stats", CTLFLAG_RD, NULL, ""); 2355 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2356 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE, 2357 zone, 1, sysctl_handle_uma_zone_cur, "I", 2358 "Current number of allocated items"); 2359 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2360 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2361 zone, 0, sysctl_handle_uma_zone_allocs, "QU", 2362 "Total allocation calls"); 2363 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2364 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2365 zone, 0, sysctl_handle_uma_zone_frees, "QU", 2366 "Total free calls"); 2367 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2368 "fails", CTLFLAG_RD, &zone->uz_fails, 2369 "Number of allocation failures"); 2370 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2371 "xdomain", CTLFLAG_RD, &zone->uz_xdomain, 0, 2372 "Free calls from the wrong domain"); 2373 } 2374 2375 struct uma_zone_count { 2376 const char *name; 2377 int count; 2378 }; 2379 2380 static void 2381 zone_count(uma_zone_t zone, void *arg) 2382 { 2383 struct uma_zone_count *cnt; 2384 2385 cnt = arg; 2386 /* 2387 * Some zones are rapidly created with identical names and 2388 * destroyed out of order. This can lead to gaps in the count. 2389 * Use one greater than the maximum observed for this name. 2390 */ 2391 if (strcmp(zone->uz_name, cnt->name) == 0) 2392 cnt->count = MAX(cnt->count, 2393 zone->uz_namecnt + 1); 2394 } 2395 2396 static void 2397 zone_update_caches(uma_zone_t zone) 2398 { 2399 int i; 2400 2401 for (i = 0; i <= mp_maxid; i++) { 2402 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size); 2403 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags); 2404 } 2405 } 2406 2407 /* 2408 * Zone header ctor. This initializes all fields, locks, etc. 2409 * 2410 * Arguments/Returns follow uma_ctor specifications 2411 * udata Actually uma_zctor_args 2412 */ 2413 static int 2414 zone_ctor(void *mem, int size, void *udata, int flags) 2415 { 2416 struct uma_zone_count cnt; 2417 struct uma_zctor_args *arg = udata; 2418 uma_zone_t zone = mem; 2419 uma_zone_t z; 2420 uma_keg_t keg; 2421 int i; 2422 2423 bzero(zone, size); 2424 zone->uz_name = arg->name; 2425 zone->uz_ctor = arg->ctor; 2426 zone->uz_dtor = arg->dtor; 2427 zone->uz_init = NULL; 2428 zone->uz_fini = NULL; 2429 zone->uz_sleeps = 0; 2430 zone->uz_xdomain = 0; 2431 zone->uz_bucket_size = 0; 2432 zone->uz_bucket_size_min = 0; 2433 zone->uz_bucket_size_max = BUCKET_MAX; 2434 zone->uz_flags = (arg->flags & UMA_ZONE_SMR); 2435 zone->uz_warning = NULL; 2436 /* The domain structures follow the cpu structures. */ 2437 zone->uz_domain = 2438 (struct uma_zone_domain *)&zone->uz_cpu[mp_maxid + 1]; 2439 zone->uz_bkt_max = ULONG_MAX; 2440 timevalclear(&zone->uz_ratecheck); 2441 2442 /* Count the number of duplicate names. */ 2443 cnt.name = arg->name; 2444 cnt.count = 0; 2445 zone_foreach(zone_count, &cnt); 2446 zone->uz_namecnt = cnt.count; 2447 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 2448 ZONE_CROSS_LOCK_INIT(zone); 2449 2450 for (i = 0; i < vm_ndomains; i++) 2451 STAILQ_INIT(&zone->uz_domain[i].uzd_buckets); 2452 2453 #ifdef INVARIANTS 2454 if (arg->uminit == trash_init && arg->fini == trash_fini) 2455 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR; 2456 #endif 2457 2458 /* 2459 * This is a pure cache zone, no kegs. 2460 */ 2461 if (arg->import) { 2462 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0, 2463 ("zone_ctor: Import specified for non-cache zone.")); 2464 if (arg->flags & UMA_ZONE_VM) 2465 arg->flags |= UMA_ZFLAG_CACHEONLY; 2466 zone->uz_flags = arg->flags; 2467 zone->uz_size = arg->size; 2468 zone->uz_import = arg->import; 2469 zone->uz_release = arg->release; 2470 zone->uz_arg = arg->arg; 2471 rw_wlock(&uma_rwlock); 2472 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 2473 rw_wunlock(&uma_rwlock); 2474 goto out; 2475 } 2476 2477 /* 2478 * Use the regular zone/keg/slab allocator. 2479 */ 2480 zone->uz_import = zone_import; 2481 zone->uz_release = zone_release; 2482 zone->uz_arg = zone; 2483 keg = arg->keg; 2484 2485 if (arg->flags & UMA_ZONE_SECONDARY) { 2486 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 2487 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 2488 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 2489 zone->uz_init = arg->uminit; 2490 zone->uz_fini = arg->fini; 2491 zone->uz_flags |= UMA_ZONE_SECONDARY; 2492 rw_wlock(&uma_rwlock); 2493 ZONE_LOCK(zone); 2494 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 2495 if (LIST_NEXT(z, uz_link) == NULL) { 2496 LIST_INSERT_AFTER(z, zone, uz_link); 2497 break; 2498 } 2499 } 2500 ZONE_UNLOCK(zone); 2501 rw_wunlock(&uma_rwlock); 2502 } else if (keg == NULL) { 2503 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 2504 arg->align, arg->flags)) == NULL) 2505 return (ENOMEM); 2506 } else { 2507 struct uma_kctor_args karg; 2508 int error; 2509 2510 /* We should only be here from uma_startup() */ 2511 karg.size = arg->size; 2512 karg.uminit = arg->uminit; 2513 karg.fini = arg->fini; 2514 karg.align = arg->align; 2515 karg.flags = (arg->flags & ~UMA_ZONE_SMR); 2516 karg.zone = zone; 2517 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 2518 flags); 2519 if (error) 2520 return (error); 2521 } 2522 2523 /* Inherit properties from the keg. */ 2524 zone->uz_keg = keg; 2525 zone->uz_size = keg->uk_size; 2526 zone->uz_flags |= (keg->uk_flags & 2527 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 2528 2529 out: 2530 if (__predict_true(booted >= BOOT_RUNNING)) { 2531 zone_alloc_counters(zone, NULL); 2532 zone_alloc_sysctl(zone, NULL); 2533 } else { 2534 zone->uz_allocs = EARLY_COUNTER; 2535 zone->uz_frees = EARLY_COUNTER; 2536 zone->uz_fails = EARLY_COUNTER; 2537 } 2538 2539 /* Caller requests a private SMR context. */ 2540 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 2541 zone->uz_smr = smr_create(zone->uz_name); 2542 2543 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) != 2544 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET), 2545 ("Invalid zone flag combination")); 2546 if (arg->flags & UMA_ZFLAG_INTERNAL) 2547 zone->uz_bucket_size_max = zone->uz_bucket_size = 0; 2548 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0) 2549 zone->uz_bucket_size = BUCKET_MAX; 2550 else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0) 2551 zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN; 2552 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0) 2553 zone->uz_bucket_size = 0; 2554 else 2555 zone->uz_bucket_size = bucket_select(zone->uz_size); 2556 zone->uz_bucket_size_min = zone->uz_bucket_size; 2557 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL) 2558 zone->uz_flags |= UMA_ZFLAG_CTORDTOR; 2559 zone_update_caches(zone); 2560 2561 return (0); 2562 } 2563 2564 /* 2565 * Keg header dtor. This frees all data, destroys locks, frees the hash 2566 * table and removes the keg from the global list. 2567 * 2568 * Arguments/Returns follow uma_dtor specifications 2569 * udata unused 2570 */ 2571 static void 2572 keg_dtor(void *arg, int size, void *udata) 2573 { 2574 uma_keg_t keg; 2575 uint32_t free, pages; 2576 int i; 2577 2578 keg = (uma_keg_t)arg; 2579 free = pages = 0; 2580 for (i = 0; i < vm_ndomains; i++) { 2581 free += keg->uk_domain[i].ud_free; 2582 pages += keg->uk_domain[i].ud_pages; 2583 KEG_LOCK_FINI(keg, i); 2584 } 2585 if (pages != 0) 2586 printf("Freed UMA keg (%s) was not empty (%u items). " 2587 " Lost %u pages of memory.\n", 2588 keg->uk_name ? keg->uk_name : "", 2589 pages / keg->uk_ppera * keg->uk_ipers - free, pages); 2590 2591 hash_free(&keg->uk_hash); 2592 } 2593 2594 /* 2595 * Zone header dtor. 2596 * 2597 * Arguments/Returns follow uma_dtor specifications 2598 * udata unused 2599 */ 2600 static void 2601 zone_dtor(void *arg, int size, void *udata) 2602 { 2603 uma_zone_t zone; 2604 uma_keg_t keg; 2605 2606 zone = (uma_zone_t)arg; 2607 2608 sysctl_remove_oid(zone->uz_oid, 1, 1); 2609 2610 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 2611 cache_drain(zone); 2612 2613 rw_wlock(&uma_rwlock); 2614 LIST_REMOVE(zone, uz_link); 2615 rw_wunlock(&uma_rwlock); 2616 /* 2617 * XXX there are some races here where 2618 * the zone can be drained but zone lock 2619 * released and then refilled before we 2620 * remove it... we dont care for now 2621 */ 2622 zone_reclaim(zone, M_WAITOK, true); 2623 /* 2624 * We only destroy kegs from non secondary/non cache zones. 2625 */ 2626 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) { 2627 keg = zone->uz_keg; 2628 rw_wlock(&uma_rwlock); 2629 LIST_REMOVE(keg, uk_link); 2630 rw_wunlock(&uma_rwlock); 2631 zone_free_item(kegs, keg, NULL, SKIP_NONE); 2632 } 2633 counter_u64_free(zone->uz_allocs); 2634 counter_u64_free(zone->uz_frees); 2635 counter_u64_free(zone->uz_fails); 2636 free(zone->uz_ctlname, M_UMA); 2637 ZONE_LOCK_FINI(zone); 2638 ZONE_CROSS_LOCK_FINI(zone); 2639 } 2640 2641 static void 2642 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg) 2643 { 2644 uma_keg_t keg; 2645 uma_zone_t zone; 2646 2647 LIST_FOREACH(keg, &uma_kegs, uk_link) { 2648 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 2649 zfunc(zone, arg); 2650 } 2651 LIST_FOREACH(zone, &uma_cachezones, uz_link) 2652 zfunc(zone, arg); 2653 } 2654 2655 /* 2656 * Traverses every zone in the system and calls a callback 2657 * 2658 * Arguments: 2659 * zfunc A pointer to a function which accepts a zone 2660 * as an argument. 2661 * 2662 * Returns: 2663 * Nothing 2664 */ 2665 static void 2666 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg) 2667 { 2668 2669 rw_rlock(&uma_rwlock); 2670 zone_foreach_unlocked(zfunc, arg); 2671 rw_runlock(&uma_rwlock); 2672 } 2673 2674 /* 2675 * Initialize the kernel memory allocator. This is done after pages can be 2676 * allocated but before general KVA is available. 2677 */ 2678 void 2679 uma_startup1(vm_offset_t virtual_avail) 2680 { 2681 struct uma_zctor_args args; 2682 size_t ksize, zsize, size; 2683 uma_keg_t masterkeg; 2684 uintptr_t m; 2685 uint8_t pflag; 2686 2687 bootstart = bootmem = virtual_avail; 2688 2689 rw_init(&uma_rwlock, "UMA lock"); 2690 sx_init(&uma_reclaim_lock, "umareclaim"); 2691 2692 ksize = sizeof(struct uma_keg) + 2693 (sizeof(struct uma_domain) * vm_ndomains); 2694 ksize = roundup(ksize, UMA_SUPER_ALIGN); 2695 zsize = sizeof(struct uma_zone) + 2696 (sizeof(struct uma_cache) * (mp_maxid + 1)) + 2697 (sizeof(struct uma_zone_domain) * vm_ndomains); 2698 zsize = roundup(zsize, UMA_SUPER_ALIGN); 2699 2700 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */ 2701 size = (zsize * 2) + ksize; 2702 m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO); 2703 zones = (uma_zone_t)m; 2704 m += zsize; 2705 kegs = (uma_zone_t)m; 2706 m += zsize; 2707 masterkeg = (uma_keg_t)m; 2708 2709 /* "manually" create the initial zone */ 2710 memset(&args, 0, sizeof(args)); 2711 args.name = "UMA Kegs"; 2712 args.size = ksize; 2713 args.ctor = keg_ctor; 2714 args.dtor = keg_dtor; 2715 args.uminit = zero_init; 2716 args.fini = NULL; 2717 args.keg = masterkeg; 2718 args.align = UMA_SUPER_ALIGN - 1; 2719 args.flags = UMA_ZFLAG_INTERNAL; 2720 zone_ctor(kegs, zsize, &args, M_WAITOK); 2721 2722 args.name = "UMA Zones"; 2723 args.size = zsize; 2724 args.ctor = zone_ctor; 2725 args.dtor = zone_dtor; 2726 args.uminit = zero_init; 2727 args.fini = NULL; 2728 args.keg = NULL; 2729 args.align = UMA_SUPER_ALIGN - 1; 2730 args.flags = UMA_ZFLAG_INTERNAL; 2731 zone_ctor(zones, zsize, &args, M_WAITOK); 2732 2733 /* Now make zones for slab headers */ 2734 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE, 2735 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2736 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE, 2737 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2738 2739 hashzone = uma_zcreate("UMA Hash", 2740 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 2741 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2742 2743 bucket_init(); 2744 smr_init(); 2745 } 2746 2747 #ifndef UMA_MD_SMALL_ALLOC 2748 extern void vm_radix_reserve_kva(void); 2749 #endif 2750 2751 /* 2752 * Advertise the availability of normal kva allocations and switch to 2753 * the default back-end allocator. Marks the KVA we consumed on startup 2754 * as used in the map. 2755 */ 2756 void 2757 uma_startup2(void) 2758 { 2759 2760 if (bootstart != bootmem) { 2761 vm_map_lock(kernel_map); 2762 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem, 2763 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 2764 vm_map_unlock(kernel_map); 2765 } 2766 2767 #ifndef UMA_MD_SMALL_ALLOC 2768 /* Set up radix zone to use noobj_alloc. */ 2769 vm_radix_reserve_kva(); 2770 #endif 2771 2772 booted = BOOT_KVA; 2773 zone_foreach_unlocked(zone_kva_available, NULL); 2774 bucket_enable(); 2775 } 2776 2777 /* 2778 * Finish our initialization steps. 2779 */ 2780 static void 2781 uma_startup3(void) 2782 { 2783 2784 #ifdef INVARIANTS 2785 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor); 2786 uma_dbg_cnt = counter_u64_alloc(M_WAITOK); 2787 uma_skip_cnt = counter_u64_alloc(M_WAITOK); 2788 #endif 2789 zone_foreach_unlocked(zone_alloc_counters, NULL); 2790 zone_foreach_unlocked(zone_alloc_sysctl, NULL); 2791 callout_init(&uma_callout, 1); 2792 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 2793 booted = BOOT_RUNNING; 2794 2795 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL, 2796 EVENTHANDLER_PRI_FIRST); 2797 } 2798 2799 static void 2800 uma_shutdown(void) 2801 { 2802 2803 booted = BOOT_SHUTDOWN; 2804 } 2805 2806 static uma_keg_t 2807 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 2808 int align, uint32_t flags) 2809 { 2810 struct uma_kctor_args args; 2811 2812 args.size = size; 2813 args.uminit = uminit; 2814 args.fini = fini; 2815 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 2816 args.flags = flags; 2817 args.zone = zone; 2818 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK)); 2819 } 2820 2821 /* Public functions */ 2822 /* See uma.h */ 2823 void 2824 uma_set_align(int align) 2825 { 2826 2827 if (align != UMA_ALIGN_CACHE) 2828 uma_align_cache = align; 2829 } 2830 2831 /* See uma.h */ 2832 uma_zone_t 2833 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 2834 uma_init uminit, uma_fini fini, int align, uint32_t flags) 2835 2836 { 2837 struct uma_zctor_args args; 2838 uma_zone_t res; 2839 2840 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"", 2841 align, name)); 2842 2843 /* This stuff is essential for the zone ctor */ 2844 memset(&args, 0, sizeof(args)); 2845 args.name = name; 2846 args.size = size; 2847 args.ctor = ctor; 2848 args.dtor = dtor; 2849 args.uminit = uminit; 2850 args.fini = fini; 2851 #ifdef INVARIANTS 2852 /* 2853 * Inject procedures which check for memory use after free if we are 2854 * allowed to scramble the memory while it is not allocated. This 2855 * requires that: UMA is actually able to access the memory, no init 2856 * or fini procedures, no dependency on the initial value of the 2857 * memory, and no (legitimate) use of the memory after free. Note, 2858 * the ctor and dtor do not need to be empty. 2859 */ 2860 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH | 2861 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) { 2862 args.uminit = trash_init; 2863 args.fini = trash_fini; 2864 } 2865 #endif 2866 args.align = align; 2867 args.flags = flags; 2868 args.keg = NULL; 2869 2870 sx_slock(&uma_reclaim_lock); 2871 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2872 sx_sunlock(&uma_reclaim_lock); 2873 2874 return (res); 2875 } 2876 2877 /* See uma.h */ 2878 uma_zone_t 2879 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 2880 uma_init zinit, uma_fini zfini, uma_zone_t master) 2881 { 2882 struct uma_zctor_args args; 2883 uma_keg_t keg; 2884 uma_zone_t res; 2885 2886 keg = master->uz_keg; 2887 memset(&args, 0, sizeof(args)); 2888 args.name = name; 2889 args.size = keg->uk_size; 2890 args.ctor = ctor; 2891 args.dtor = dtor; 2892 args.uminit = zinit; 2893 args.fini = zfini; 2894 args.align = keg->uk_align; 2895 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 2896 args.keg = keg; 2897 2898 sx_slock(&uma_reclaim_lock); 2899 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2900 sx_sunlock(&uma_reclaim_lock); 2901 2902 return (res); 2903 } 2904 2905 /* See uma.h */ 2906 uma_zone_t 2907 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 2908 uma_init zinit, uma_fini zfini, uma_import zimport, 2909 uma_release zrelease, void *arg, int flags) 2910 { 2911 struct uma_zctor_args args; 2912 2913 memset(&args, 0, sizeof(args)); 2914 args.name = name; 2915 args.size = size; 2916 args.ctor = ctor; 2917 args.dtor = dtor; 2918 args.uminit = zinit; 2919 args.fini = zfini; 2920 args.import = zimport; 2921 args.release = zrelease; 2922 args.arg = arg; 2923 args.align = 0; 2924 args.flags = flags | UMA_ZFLAG_CACHE; 2925 2926 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK)); 2927 } 2928 2929 /* See uma.h */ 2930 void 2931 uma_zdestroy(uma_zone_t zone) 2932 { 2933 2934 /* 2935 * Large slabs are expensive to reclaim, so don't bother doing 2936 * unnecessary work if we're shutting down. 2937 */ 2938 if (booted == BOOT_SHUTDOWN && 2939 zone->uz_fini == NULL && zone->uz_release == zone_release) 2940 return; 2941 sx_slock(&uma_reclaim_lock); 2942 zone_free_item(zones, zone, NULL, SKIP_NONE); 2943 sx_sunlock(&uma_reclaim_lock); 2944 } 2945 2946 void 2947 uma_zwait(uma_zone_t zone) 2948 { 2949 void *item; 2950 2951 item = uma_zalloc_arg(zone, NULL, M_WAITOK); 2952 uma_zfree(zone, item); 2953 } 2954 2955 void * 2956 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags) 2957 { 2958 void *item; 2959 #ifdef SMP 2960 int i; 2961 2962 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2963 #endif 2964 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO); 2965 if (item != NULL && (flags & M_ZERO)) { 2966 #ifdef SMP 2967 for (i = 0; i <= mp_maxid; i++) 2968 bzero(zpcpu_get_cpu(item, i), zone->uz_size); 2969 #else 2970 bzero(item, zone->uz_size); 2971 #endif 2972 } 2973 return (item); 2974 } 2975 2976 /* 2977 * A stub while both regular and pcpu cases are identical. 2978 */ 2979 void 2980 uma_zfree_pcpu_arg(uma_zone_t zone, void *item, void *udata) 2981 { 2982 2983 #ifdef SMP 2984 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2985 #endif 2986 uma_zfree_arg(zone, item, udata); 2987 } 2988 2989 static inline void * 2990 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags, 2991 void *item) 2992 { 2993 #ifdef INVARIANTS 2994 bool skipdbg; 2995 2996 skipdbg = uma_dbg_zskip(zone, item); 2997 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 && 2998 zone->uz_ctor != trash_ctor) 2999 trash_ctor(item, size, udata, flags); 3000 #endif 3001 /* Check flags before loading ctor pointer. */ 3002 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) && 3003 __predict_false(zone->uz_ctor != NULL) && 3004 zone->uz_ctor(item, size, udata, flags) != 0) { 3005 counter_u64_add(zone->uz_fails, 1); 3006 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT); 3007 return (NULL); 3008 } 3009 #ifdef INVARIANTS 3010 if (!skipdbg) 3011 uma_dbg_alloc(zone, NULL, item); 3012 #endif 3013 if (flags & M_ZERO) 3014 bzero(item, size); 3015 3016 return (item); 3017 } 3018 3019 static inline void 3020 item_dtor(uma_zone_t zone, void *item, int size, void *udata, 3021 enum zfreeskip skip) 3022 { 3023 #ifdef INVARIANTS 3024 bool skipdbg; 3025 3026 skipdbg = uma_dbg_zskip(zone, item); 3027 if (skip == SKIP_NONE && !skipdbg) { 3028 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0) 3029 uma_dbg_free(zone, udata, item); 3030 else 3031 uma_dbg_free(zone, NULL, item); 3032 } 3033 #endif 3034 if (__predict_true(skip < SKIP_DTOR)) { 3035 if (zone->uz_dtor != NULL) 3036 zone->uz_dtor(item, size, udata); 3037 #ifdef INVARIANTS 3038 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 && 3039 zone->uz_dtor != trash_dtor) 3040 trash_dtor(item, size, udata); 3041 #endif 3042 } 3043 } 3044 3045 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS) 3046 #define UMA_ZALLOC_DEBUG 3047 static int 3048 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags) 3049 { 3050 int error; 3051 3052 error = 0; 3053 #ifdef WITNESS 3054 if (flags & M_WAITOK) { 3055 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 3056 "uma_zalloc_debug: zone \"%s\"", zone->uz_name); 3057 } 3058 #endif 3059 3060 #ifdef INVARIANTS 3061 KASSERT((flags & M_EXEC) == 0, 3062 ("uma_zalloc_debug: called with M_EXEC")); 3063 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3064 ("uma_zalloc_debug: called within spinlock or critical section")); 3065 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0, 3066 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO")); 3067 #endif 3068 3069 #ifdef DEBUG_MEMGUARD 3070 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) { 3071 void *item; 3072 item = memguard_alloc(zone->uz_size, flags); 3073 if (item != NULL) { 3074 error = EJUSTRETURN; 3075 if (zone->uz_init != NULL && 3076 zone->uz_init(item, zone->uz_size, flags) != 0) { 3077 *itemp = NULL; 3078 return (error); 3079 } 3080 if (zone->uz_ctor != NULL && 3081 zone->uz_ctor(item, zone->uz_size, udata, 3082 flags) != 0) { 3083 counter_u64_add(zone->uz_fails, 1); 3084 zone->uz_fini(item, zone->uz_size); 3085 *itemp = NULL; 3086 return (error); 3087 } 3088 *itemp = item; 3089 return (error); 3090 } 3091 /* This is unfortunate but should not be fatal. */ 3092 } 3093 #endif 3094 return (error); 3095 } 3096 3097 static int 3098 uma_zfree_debug(uma_zone_t zone, void *item, void *udata) 3099 { 3100 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3101 ("uma_zfree_debug: called with spinlock or critical section held")); 3102 3103 #ifdef DEBUG_MEMGUARD 3104 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) { 3105 if (zone->uz_dtor != NULL) 3106 zone->uz_dtor(item, zone->uz_size, udata); 3107 if (zone->uz_fini != NULL) 3108 zone->uz_fini(item, zone->uz_size); 3109 memguard_free(item); 3110 return (EJUSTRETURN); 3111 } 3112 #endif 3113 return (0); 3114 } 3115 #endif 3116 3117 static __noinline void * 3118 uma_zalloc_single(uma_zone_t zone, void *udata, int flags) 3119 { 3120 int domain; 3121 3122 /* 3123 * We can not get a bucket so try to return a single item. 3124 */ 3125 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) 3126 domain = PCPU_GET(domain); 3127 else 3128 domain = UMA_ANYDOMAIN; 3129 return (zone_alloc_item(zone, udata, domain, flags)); 3130 } 3131 3132 /* See uma.h */ 3133 void * 3134 uma_zalloc_smr(uma_zone_t zone, int flags) 3135 { 3136 uma_cache_bucket_t bucket; 3137 uma_cache_t cache; 3138 void *item; 3139 int size, uz_flags; 3140 3141 #ifdef UMA_ZALLOC_DEBUG 3142 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0, 3143 ("uma_zalloc_arg: called with non-SMR zone.\n")); 3144 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN) 3145 return (item); 3146 #endif 3147 3148 critical_enter(); 3149 do { 3150 cache = &zone->uz_cpu[curcpu]; 3151 bucket = &cache->uc_allocbucket; 3152 size = cache_uz_size(cache); 3153 uz_flags = cache_uz_flags(cache); 3154 if (__predict_true(bucket->ucb_cnt != 0)) { 3155 item = cache_bucket_pop(cache, bucket); 3156 critical_exit(); 3157 return (item_ctor(zone, uz_flags, size, NULL, flags, 3158 item)); 3159 } 3160 } while (cache_alloc(zone, cache, NULL, flags)); 3161 critical_exit(); 3162 3163 return (uma_zalloc_single(zone, NULL, flags)); 3164 } 3165 3166 /* See uma.h */ 3167 void * 3168 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 3169 { 3170 uma_cache_bucket_t bucket; 3171 uma_cache_t cache; 3172 void *item; 3173 int size, uz_flags; 3174 3175 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3176 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3177 3178 /* This is the fast path allocation */ 3179 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name, 3180 zone, flags); 3181 3182 #ifdef UMA_ZALLOC_DEBUG 3183 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0, 3184 ("uma_zalloc_arg: called with SMR zone.\n")); 3185 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN) 3186 return (item); 3187 #endif 3188 3189 /* 3190 * If possible, allocate from the per-CPU cache. There are two 3191 * requirements for safe access to the per-CPU cache: (1) the thread 3192 * accessing the cache must not be preempted or yield during access, 3193 * and (2) the thread must not migrate CPUs without switching which 3194 * cache it accesses. We rely on a critical section to prevent 3195 * preemption and migration. We release the critical section in 3196 * order to acquire the zone mutex if we are unable to allocate from 3197 * the current cache; when we re-acquire the critical section, we 3198 * must detect and handle migration if it has occurred. 3199 */ 3200 critical_enter(); 3201 do { 3202 cache = &zone->uz_cpu[curcpu]; 3203 bucket = &cache->uc_allocbucket; 3204 size = cache_uz_size(cache); 3205 uz_flags = cache_uz_flags(cache); 3206 if (__predict_true(bucket->ucb_cnt != 0)) { 3207 item = cache_bucket_pop(cache, bucket); 3208 critical_exit(); 3209 return (item_ctor(zone, uz_flags, size, udata, flags, 3210 item)); 3211 } 3212 } while (cache_alloc(zone, cache, udata, flags)); 3213 critical_exit(); 3214 3215 return (uma_zalloc_single(zone, udata, flags)); 3216 } 3217 3218 /* 3219 * Replenish an alloc bucket and possibly restore an old one. Called in 3220 * a critical section. Returns in a critical section. 3221 * 3222 * A false return value indicates an allocation failure. 3223 * A true return value indicates success and the caller should retry. 3224 */ 3225 static __noinline bool 3226 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags) 3227 { 3228 uma_zone_domain_t zdom; 3229 uma_bucket_t bucket; 3230 int domain; 3231 bool lockfail; 3232 3233 CRITICAL_ASSERT(curthread); 3234 3235 /* 3236 * If we have run out of items in our alloc bucket see 3237 * if we can switch with the free bucket. 3238 * 3239 * SMR Zones can't re-use the free bucket until the sequence has 3240 * expired. 3241 */ 3242 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && 3243 cache->uc_freebucket.ucb_cnt != 0) { 3244 cache_bucket_swap(&cache->uc_freebucket, 3245 &cache->uc_allocbucket); 3246 return (true); 3247 } 3248 3249 /* 3250 * Discard any empty allocation bucket while we hold no locks. 3251 */ 3252 bucket = cache_bucket_unload_alloc(cache); 3253 critical_exit(); 3254 if (bucket != NULL) 3255 bucket_free(zone, bucket, udata); 3256 3257 /* Short-circuit for zones without buckets and low memory. */ 3258 if (zone->uz_bucket_size == 0 || bucketdisable) { 3259 critical_enter(); 3260 return (false); 3261 } 3262 3263 /* 3264 * Attempt to retrieve the item from the per-CPU cache has failed, so 3265 * we must go back to the zone. This requires the zone lock, so we 3266 * must drop the critical section, then re-acquire it when we go back 3267 * to the cache. Since the critical section is released, we may be 3268 * preempted or migrate. As such, make sure not to maintain any 3269 * thread-local state specific to the cache from prior to releasing 3270 * the critical section. 3271 */ 3272 lockfail = 0; 3273 if (ZONE_TRYLOCK(zone) == 0) { 3274 /* Record contention to size the buckets. */ 3275 ZONE_LOCK(zone); 3276 lockfail = 1; 3277 } 3278 3279 /* See if we lost the race to fill the cache. */ 3280 critical_enter(); 3281 cache = &zone->uz_cpu[curcpu]; 3282 if (cache->uc_allocbucket.ucb_bucket != NULL) { 3283 ZONE_UNLOCK(zone); 3284 return (true); 3285 } 3286 3287 /* 3288 * Check the zone's cache of buckets. 3289 */ 3290 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) { 3291 domain = PCPU_GET(domain); 3292 zdom = &zone->uz_domain[domain]; 3293 } else { 3294 domain = UMA_ANYDOMAIN; 3295 zdom = &zone->uz_domain[0]; 3296 } 3297 3298 if ((bucket = zone_fetch_bucket(zone, zdom)) != NULL) { 3299 KASSERT(bucket->ub_cnt != 0, 3300 ("uma_zalloc_arg: Returning an empty bucket.")); 3301 cache_bucket_load_alloc(cache, bucket); 3302 return (true); 3303 } 3304 /* We are no longer associated with this CPU. */ 3305 critical_exit(); 3306 3307 /* 3308 * We bump the uz count when the cache size is insufficient to 3309 * handle the working set. 3310 */ 3311 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max) 3312 zone->uz_bucket_size++; 3313 ZONE_UNLOCK(zone); 3314 3315 /* 3316 * Fill a bucket and attempt to use it as the alloc bucket. 3317 */ 3318 bucket = zone_alloc_bucket(zone, udata, domain, flags); 3319 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p", 3320 zone->uz_name, zone, bucket); 3321 if (bucket == NULL) { 3322 critical_enter(); 3323 return (false); 3324 } 3325 3326 /* 3327 * See if we lost the race or were migrated. Cache the 3328 * initialized bucket to make this less likely or claim 3329 * the memory directly. 3330 */ 3331 ZONE_LOCK(zone); 3332 critical_enter(); 3333 cache = &zone->uz_cpu[curcpu]; 3334 if (cache->uc_allocbucket.ucb_bucket == NULL && 3335 ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0 || 3336 domain == PCPU_GET(domain))) { 3337 cache_bucket_load_alloc(cache, bucket); 3338 zdom->uzd_imax += bucket->ub_cnt; 3339 } else if (zone->uz_bkt_count >= zone->uz_bkt_max) { 3340 critical_exit(); 3341 ZONE_UNLOCK(zone); 3342 bucket_drain(zone, bucket); 3343 bucket_free(zone, bucket, udata); 3344 critical_enter(); 3345 return (true); 3346 } else 3347 zone_put_bucket(zone, zdom, bucket, false); 3348 ZONE_UNLOCK(zone); 3349 return (true); 3350 } 3351 3352 void * 3353 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags) 3354 { 3355 3356 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3357 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3358 3359 /* This is the fast path allocation */ 3360 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d", 3361 zone->uz_name, zone, domain, flags); 3362 3363 if (flags & M_WAITOK) { 3364 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 3365 "uma_zalloc_domain: zone \"%s\"", zone->uz_name); 3366 } 3367 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3368 ("uma_zalloc_domain: called with spinlock or critical section held")); 3369 3370 return (zone_alloc_item(zone, udata, domain, flags)); 3371 } 3372 3373 /* 3374 * Find a slab with some space. Prefer slabs that are partially used over those 3375 * that are totally full. This helps to reduce fragmentation. 3376 * 3377 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check 3378 * only 'domain'. 3379 */ 3380 static uma_slab_t 3381 keg_first_slab(uma_keg_t keg, int domain, bool rr) 3382 { 3383 uma_domain_t dom; 3384 uma_slab_t slab; 3385 int start; 3386 3387 KASSERT(domain >= 0 && domain < vm_ndomains, 3388 ("keg_first_slab: domain %d out of range", domain)); 3389 KEG_LOCK_ASSERT(keg, domain); 3390 3391 slab = NULL; 3392 start = domain; 3393 do { 3394 dom = &keg->uk_domain[domain]; 3395 if (!LIST_EMPTY(&dom->ud_part_slab)) 3396 return (LIST_FIRST(&dom->ud_part_slab)); 3397 if (!LIST_EMPTY(&dom->ud_free_slab)) { 3398 slab = LIST_FIRST(&dom->ud_free_slab); 3399 LIST_REMOVE(slab, us_link); 3400 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 3401 return (slab); 3402 } 3403 if (rr) 3404 domain = (domain + 1) % vm_ndomains; 3405 } while (domain != start); 3406 3407 return (NULL); 3408 } 3409 3410 /* 3411 * Fetch an existing slab from a free or partial list. Returns with the 3412 * keg domain lock held if a slab was found or unlocked if not. 3413 */ 3414 static uma_slab_t 3415 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags) 3416 { 3417 uma_slab_t slab; 3418 uint32_t reserve; 3419 3420 /* HASH has a single free list. */ 3421 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) 3422 domain = 0; 3423 3424 KEG_LOCK(keg, domain); 3425 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve; 3426 if (keg->uk_domain[domain].ud_free <= reserve || 3427 (slab = keg_first_slab(keg, domain, rr)) == NULL) { 3428 KEG_UNLOCK(keg, domain); 3429 return (NULL); 3430 } 3431 return (slab); 3432 } 3433 3434 static uma_slab_t 3435 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags) 3436 { 3437 struct vm_domainset_iter di; 3438 uma_slab_t slab; 3439 int aflags, domain; 3440 bool rr; 3441 3442 restart: 3443 /* 3444 * Use the keg's policy if upper layers haven't already specified a 3445 * domain (as happens with first-touch zones). 3446 * 3447 * To avoid races we run the iterator with the keg lock held, but that 3448 * means that we cannot allow the vm_domainset layer to sleep. Thus, 3449 * clear M_WAITOK and handle low memory conditions locally. 3450 */ 3451 rr = rdomain == UMA_ANYDOMAIN; 3452 if (rr) { 3453 aflags = (flags & ~M_WAITOK) | M_NOWAIT; 3454 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 3455 &aflags); 3456 } else { 3457 aflags = flags; 3458 domain = rdomain; 3459 } 3460 3461 for (;;) { 3462 slab = keg_fetch_free_slab(keg, domain, rr, flags); 3463 if (slab != NULL) 3464 return (slab); 3465 3466 /* 3467 * M_NOVM means don't ask at all! 3468 */ 3469 if (flags & M_NOVM) 3470 break; 3471 3472 slab = keg_alloc_slab(keg, zone, domain, flags, aflags); 3473 if (slab != NULL) 3474 return (slab); 3475 if (!rr && (flags & M_WAITOK) == 0) 3476 break; 3477 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) { 3478 if ((flags & M_WAITOK) != 0) { 3479 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask); 3480 goto restart; 3481 } 3482 break; 3483 } 3484 } 3485 3486 /* 3487 * We might not have been able to get a slab but another cpu 3488 * could have while we were unlocked. Check again before we 3489 * fail. 3490 */ 3491 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL) 3492 return (slab); 3493 3494 return (NULL); 3495 } 3496 3497 static void * 3498 slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 3499 { 3500 uma_domain_t dom; 3501 void *item; 3502 int freei; 3503 3504 KEG_LOCK_ASSERT(keg, slab->us_domain); 3505 3506 dom = &keg->uk_domain[slab->us_domain]; 3507 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1; 3508 BIT_CLR(keg->uk_ipers, freei, &slab->us_free); 3509 item = slab_item(slab, keg, freei); 3510 slab->us_freecount--; 3511 dom->ud_free--; 3512 3513 /* Move this slab to the full list */ 3514 if (slab->us_freecount == 0) { 3515 LIST_REMOVE(slab, us_link); 3516 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link); 3517 } 3518 3519 return (item); 3520 } 3521 3522 static int 3523 zone_import(void *arg, void **bucket, int max, int domain, int flags) 3524 { 3525 uma_domain_t dom; 3526 uma_zone_t zone; 3527 uma_slab_t slab; 3528 uma_keg_t keg; 3529 #ifdef NUMA 3530 int stripe; 3531 #endif 3532 int i; 3533 3534 zone = arg; 3535 slab = NULL; 3536 keg = zone->uz_keg; 3537 /* Try to keep the buckets totally full */ 3538 for (i = 0; i < max; ) { 3539 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL) 3540 break; 3541 #ifdef NUMA 3542 stripe = howmany(max, vm_ndomains); 3543 #endif 3544 dom = &keg->uk_domain[slab->us_domain]; 3545 while (slab->us_freecount && i < max) { 3546 bucket[i++] = slab_alloc_item(keg, slab); 3547 if (dom->ud_free <= keg->uk_reserve) 3548 break; 3549 #ifdef NUMA 3550 /* 3551 * If the zone is striped we pick a new slab for every 3552 * N allocations. Eliminating this conditional will 3553 * instead pick a new domain for each bucket rather 3554 * than stripe within each bucket. The current option 3555 * produces more fragmentation and requires more cpu 3556 * time but yields better distribution. 3557 */ 3558 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 && 3559 vm_ndomains > 1 && --stripe == 0) 3560 break; 3561 #endif 3562 } 3563 KEG_UNLOCK(keg, slab->us_domain); 3564 /* Don't block if we allocated any successfully. */ 3565 flags &= ~M_WAITOK; 3566 flags |= M_NOWAIT; 3567 } 3568 3569 return i; 3570 } 3571 3572 static int 3573 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags) 3574 { 3575 uint64_t old, new, total, max; 3576 3577 /* 3578 * The hard case. We're going to sleep because there were existing 3579 * sleepers or because we ran out of items. This routine enforces 3580 * fairness by keeping fifo order. 3581 * 3582 * First release our ill gotten gains and make some noise. 3583 */ 3584 for (;;) { 3585 zone_free_limit(zone, count); 3586 zone_log_warning(zone); 3587 zone_maxaction(zone); 3588 if (flags & M_NOWAIT) 3589 return (0); 3590 3591 /* 3592 * We need to allocate an item or set ourself as a sleeper 3593 * while the sleepq lock is held to avoid wakeup races. This 3594 * is essentially a home rolled semaphore. 3595 */ 3596 sleepq_lock(&zone->uz_max_items); 3597 old = zone->uz_items; 3598 do { 3599 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX); 3600 /* Cache the max since we will evaluate twice. */ 3601 max = zone->uz_max_items; 3602 if (UZ_ITEMS_SLEEPERS(old) != 0 || 3603 UZ_ITEMS_COUNT(old) >= max) 3604 new = old + UZ_ITEMS_SLEEPER; 3605 else 3606 new = old + MIN(count, max - old); 3607 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0); 3608 3609 /* We may have successfully allocated under the sleepq lock. */ 3610 if (UZ_ITEMS_SLEEPERS(new) == 0) { 3611 sleepq_release(&zone->uz_max_items); 3612 return (new - old); 3613 } 3614 3615 /* 3616 * This is in a different cacheline from uz_items so that we 3617 * don't constantly invalidate the fastpath cacheline when we 3618 * adjust item counts. This could be limited to toggling on 3619 * transitions. 3620 */ 3621 atomic_add_32(&zone->uz_sleepers, 1); 3622 atomic_add_64(&zone->uz_sleeps, 1); 3623 3624 /* 3625 * We have added ourselves as a sleeper. The sleepq lock 3626 * protects us from wakeup races. Sleep now and then retry. 3627 */ 3628 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0); 3629 sleepq_wait(&zone->uz_max_items, PVM); 3630 3631 /* 3632 * After wakeup, remove ourselves as a sleeper and try 3633 * again. We no longer have the sleepq lock for protection. 3634 * 3635 * Subract ourselves as a sleeper while attempting to add 3636 * our count. 3637 */ 3638 atomic_subtract_32(&zone->uz_sleepers, 1); 3639 old = atomic_fetchadd_64(&zone->uz_items, 3640 -(UZ_ITEMS_SLEEPER - count)); 3641 /* We're no longer a sleeper. */ 3642 old -= UZ_ITEMS_SLEEPER; 3643 3644 /* 3645 * If we're still at the limit, restart. Notably do not 3646 * block on other sleepers. Cache the max value to protect 3647 * against changes via sysctl. 3648 */ 3649 total = UZ_ITEMS_COUNT(old); 3650 max = zone->uz_max_items; 3651 if (total >= max) 3652 continue; 3653 /* Truncate if necessary, otherwise wake other sleepers. */ 3654 if (total + count > max) { 3655 zone_free_limit(zone, total + count - max); 3656 count = max - total; 3657 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0) 3658 wakeup_one(&zone->uz_max_items); 3659 3660 return (count); 3661 } 3662 } 3663 3664 /* 3665 * Allocate 'count' items from our max_items limit. Returns the number 3666 * available. If M_NOWAIT is not specified it will sleep until at least 3667 * one item can be allocated. 3668 */ 3669 static int 3670 zone_alloc_limit(uma_zone_t zone, int count, int flags) 3671 { 3672 uint64_t old; 3673 uint64_t max; 3674 3675 max = zone->uz_max_items; 3676 MPASS(max > 0); 3677 3678 /* 3679 * We expect normal allocations to succeed with a simple 3680 * fetchadd. 3681 */ 3682 old = atomic_fetchadd_64(&zone->uz_items, count); 3683 if (__predict_true(old + count <= max)) 3684 return (count); 3685 3686 /* 3687 * If we had some items and no sleepers just return the 3688 * truncated value. We have to release the excess space 3689 * though because that may wake sleepers who weren't woken 3690 * because we were temporarily over the limit. 3691 */ 3692 if (old < max) { 3693 zone_free_limit(zone, (old + count) - max); 3694 return (max - old); 3695 } 3696 return (zone_alloc_limit_hard(zone, count, flags)); 3697 } 3698 3699 /* 3700 * Free a number of items back to the limit. 3701 */ 3702 static void 3703 zone_free_limit(uma_zone_t zone, int count) 3704 { 3705 uint64_t old; 3706 3707 MPASS(count > 0); 3708 3709 /* 3710 * In the common case we either have no sleepers or 3711 * are still over the limit and can just return. 3712 */ 3713 old = atomic_fetchadd_64(&zone->uz_items, -count); 3714 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 || 3715 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items)) 3716 return; 3717 3718 /* 3719 * Moderate the rate of wakeups. Sleepers will continue 3720 * to generate wakeups if necessary. 3721 */ 3722 wakeup_one(&zone->uz_max_items); 3723 } 3724 3725 static uma_bucket_t 3726 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags) 3727 { 3728 uma_bucket_t bucket; 3729 int maxbucket, cnt; 3730 3731 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name, 3732 zone, domain); 3733 3734 /* Avoid allocs targeting empty domains. */ 3735 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain)) 3736 domain = UMA_ANYDOMAIN; 3737 3738 if (zone->uz_max_items > 0) 3739 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size, 3740 M_NOWAIT); 3741 else 3742 maxbucket = zone->uz_bucket_size; 3743 if (maxbucket == 0) 3744 return (false); 3745 3746 /* Don't wait for buckets, preserve caller's NOVM setting. */ 3747 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 3748 if (bucket == NULL) { 3749 cnt = 0; 3750 goto out; 3751 } 3752 3753 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 3754 MIN(maxbucket, bucket->ub_entries), domain, flags); 3755 3756 /* 3757 * Initialize the memory if necessary. 3758 */ 3759 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 3760 int i; 3761 3762 for (i = 0; i < bucket->ub_cnt; i++) 3763 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 3764 flags) != 0) 3765 break; 3766 /* 3767 * If we couldn't initialize the whole bucket, put the 3768 * rest back onto the freelist. 3769 */ 3770 if (i != bucket->ub_cnt) { 3771 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 3772 bucket->ub_cnt - i); 3773 #ifdef INVARIANTS 3774 bzero(&bucket->ub_bucket[i], 3775 sizeof(void *) * (bucket->ub_cnt - i)); 3776 #endif 3777 bucket->ub_cnt = i; 3778 } 3779 } 3780 3781 cnt = bucket->ub_cnt; 3782 if (bucket->ub_cnt == 0) { 3783 bucket_free(zone, bucket, udata); 3784 counter_u64_add(zone->uz_fails, 1); 3785 bucket = NULL; 3786 } 3787 out: 3788 if (zone->uz_max_items > 0 && cnt < maxbucket) 3789 zone_free_limit(zone, maxbucket - cnt); 3790 3791 return (bucket); 3792 } 3793 3794 /* 3795 * Allocates a single item from a zone. 3796 * 3797 * Arguments 3798 * zone The zone to alloc for. 3799 * udata The data to be passed to the constructor. 3800 * domain The domain to allocate from or UMA_ANYDOMAIN. 3801 * flags M_WAITOK, M_NOWAIT, M_ZERO. 3802 * 3803 * Returns 3804 * NULL if there is no memory and M_NOWAIT is set 3805 * An item if successful 3806 */ 3807 3808 static void * 3809 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags) 3810 { 3811 void *item; 3812 3813 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) 3814 return (NULL); 3815 3816 /* Avoid allocs targeting empty domains. */ 3817 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain)) 3818 domain = UMA_ANYDOMAIN; 3819 3820 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1) 3821 goto fail_cnt; 3822 3823 /* 3824 * We have to call both the zone's init (not the keg's init) 3825 * and the zone's ctor. This is because the item is going from 3826 * a keg slab directly to the user, and the user is expecting it 3827 * to be both zone-init'd as well as zone-ctor'd. 3828 */ 3829 if (zone->uz_init != NULL) { 3830 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 3831 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT); 3832 goto fail_cnt; 3833 } 3834 } 3835 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags, 3836 item); 3837 if (item == NULL) 3838 goto fail; 3839 3840 counter_u64_add(zone->uz_allocs, 1); 3841 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item, 3842 zone->uz_name, zone); 3843 3844 return (item); 3845 3846 fail_cnt: 3847 counter_u64_add(zone->uz_fails, 1); 3848 fail: 3849 if (zone->uz_max_items > 0) 3850 zone_free_limit(zone, 1); 3851 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)", 3852 zone->uz_name, zone); 3853 3854 return (NULL); 3855 } 3856 3857 /* See uma.h */ 3858 void 3859 uma_zfree_smr(uma_zone_t zone, void *item) 3860 { 3861 uma_cache_t cache; 3862 uma_cache_bucket_t bucket; 3863 int domain, itemdomain, uz_flags; 3864 3865 #ifdef UMA_ZALLOC_DEBUG 3866 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0, 3867 ("uma_zfree_smr: called with non-SMR zone.\n")); 3868 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer.")); 3869 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN) 3870 return; 3871 #endif 3872 cache = &zone->uz_cpu[curcpu]; 3873 uz_flags = cache_uz_flags(cache); 3874 domain = itemdomain = 0; 3875 #ifdef NUMA 3876 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 3877 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item)); 3878 #endif 3879 critical_enter(); 3880 do { 3881 cache = &zone->uz_cpu[curcpu]; 3882 /* SMR Zones must free to the free bucket. */ 3883 bucket = &cache->uc_freebucket; 3884 #ifdef NUMA 3885 domain = PCPU_GET(domain); 3886 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && 3887 domain != itemdomain) { 3888 bucket = &cache->uc_crossbucket; 3889 } 3890 #endif 3891 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) { 3892 cache_bucket_push(cache, bucket, item); 3893 critical_exit(); 3894 return; 3895 } 3896 } while (cache_free(zone, cache, NULL, item, itemdomain)); 3897 critical_exit(); 3898 3899 /* 3900 * If nothing else caught this, we'll just do an internal free. 3901 */ 3902 zone_free_item(zone, item, NULL, SKIP_NONE); 3903 } 3904 3905 /* See uma.h */ 3906 void 3907 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 3908 { 3909 uma_cache_t cache; 3910 uma_cache_bucket_t bucket; 3911 int domain, itemdomain, uz_flags; 3912 3913 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3914 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3915 3916 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone); 3917 3918 #ifdef UMA_ZALLOC_DEBUG 3919 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0, 3920 ("uma_zfree_arg: called with SMR zone.\n")); 3921 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN) 3922 return; 3923 #endif 3924 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 3925 if (item == NULL) 3926 return; 3927 3928 /* 3929 * We are accessing the per-cpu cache without a critical section to 3930 * fetch size and flags. This is acceptable, if we are preempted we 3931 * will simply read another cpu's line. 3932 */ 3933 cache = &zone->uz_cpu[curcpu]; 3934 uz_flags = cache_uz_flags(cache); 3935 if (UMA_ALWAYS_CTORDTOR || 3936 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0)) 3937 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE); 3938 3939 /* 3940 * The race here is acceptable. If we miss it we'll just have to wait 3941 * a little longer for the limits to be reset. 3942 */ 3943 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) { 3944 if (zone->uz_sleepers > 0) 3945 goto zfree_item; 3946 } 3947 3948 /* 3949 * If possible, free to the per-CPU cache. There are two 3950 * requirements for safe access to the per-CPU cache: (1) the thread 3951 * accessing the cache must not be preempted or yield during access, 3952 * and (2) the thread must not migrate CPUs without switching which 3953 * cache it accesses. We rely on a critical section to prevent 3954 * preemption and migration. We release the critical section in 3955 * order to acquire the zone mutex if we are unable to free to the 3956 * current cache; when we re-acquire the critical section, we must 3957 * detect and handle migration if it has occurred. 3958 */ 3959 domain = itemdomain = 0; 3960 #ifdef NUMA 3961 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 3962 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item)); 3963 #endif 3964 critical_enter(); 3965 do { 3966 cache = &zone->uz_cpu[curcpu]; 3967 /* 3968 * Try to free into the allocbucket first to give LIFO 3969 * ordering for cache-hot datastructures. Spill over 3970 * into the freebucket if necessary. Alloc will swap 3971 * them if one runs dry. 3972 */ 3973 bucket = &cache->uc_allocbucket; 3974 #ifdef NUMA 3975 domain = PCPU_GET(domain); 3976 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && 3977 domain != itemdomain) { 3978 bucket = &cache->uc_crossbucket; 3979 } else 3980 #endif 3981 if (bucket->ucb_cnt >= bucket->ucb_entries) 3982 bucket = &cache->uc_freebucket; 3983 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) { 3984 cache_bucket_push(cache, bucket, item); 3985 critical_exit(); 3986 return; 3987 } 3988 } while (cache_free(zone, cache, udata, item, itemdomain)); 3989 critical_exit(); 3990 3991 /* 3992 * If nothing else caught this, we'll just do an internal free. 3993 */ 3994 zfree_item: 3995 zone_free_item(zone, item, udata, SKIP_DTOR); 3996 } 3997 3998 #ifdef NUMA 3999 /* 4000 * sort crossdomain free buckets to domain correct buckets and cache 4001 * them. 4002 */ 4003 static void 4004 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata) 4005 { 4006 struct uma_bucketlist fullbuckets; 4007 uma_zone_domain_t zdom; 4008 uma_bucket_t b; 4009 void *item; 4010 int domain; 4011 4012 CTR3(KTR_UMA, 4013 "uma_zfree: zone %s(%p) draining cross bucket %p", 4014 zone->uz_name, zone, bucket); 4015 4016 STAILQ_INIT(&fullbuckets); 4017 4018 /* 4019 * To avoid having ndomain * ndomain buckets for sorting we have a 4020 * lock on the current crossfree bucket. A full matrix with 4021 * per-domain locking could be used if necessary. 4022 */ 4023 ZONE_CROSS_LOCK(zone); 4024 while (bucket->ub_cnt > 0) { 4025 item = bucket->ub_bucket[bucket->ub_cnt - 1]; 4026 domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item)); 4027 zdom = &zone->uz_domain[domain]; 4028 if (zdom->uzd_cross == NULL) { 4029 zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT); 4030 if (zdom->uzd_cross == NULL) 4031 break; 4032 } 4033 zdom->uzd_cross->ub_bucket[zdom->uzd_cross->ub_cnt++] = item; 4034 if (zdom->uzd_cross->ub_cnt == zdom->uzd_cross->ub_entries) { 4035 STAILQ_INSERT_HEAD(&fullbuckets, zdom->uzd_cross, 4036 ub_link); 4037 zdom->uzd_cross = NULL; 4038 } 4039 bucket->ub_cnt--; 4040 } 4041 ZONE_CROSS_UNLOCK(zone); 4042 if (!STAILQ_EMPTY(&fullbuckets)) { 4043 ZONE_LOCK(zone); 4044 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) { 4045 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 4046 bucket->ub_seq = smr_current(zone->uz_smr); 4047 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link); 4048 if (zone->uz_bkt_count >= zone->uz_bkt_max) { 4049 ZONE_UNLOCK(zone); 4050 bucket_drain(zone, b); 4051 bucket_free(zone, b, udata); 4052 ZONE_LOCK(zone); 4053 } else { 4054 domain = _vm_phys_domain( 4055 pmap_kextract( 4056 (vm_offset_t)b->ub_bucket[0])); 4057 zdom = &zone->uz_domain[domain]; 4058 zone_put_bucket(zone, zdom, b, true); 4059 } 4060 } 4061 ZONE_UNLOCK(zone); 4062 } 4063 if (bucket->ub_cnt != 0) 4064 bucket_drain(zone, bucket); 4065 bucket->ub_seq = SMR_SEQ_INVALID; 4066 bucket_free(zone, bucket, udata); 4067 } 4068 #endif 4069 4070 static void 4071 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata, 4072 int domain, int itemdomain) 4073 { 4074 uma_zone_domain_t zdom; 4075 4076 #ifdef NUMA 4077 /* 4078 * Buckets coming from the wrong domain will be entirely for the 4079 * only other domain on two domain systems. In this case we can 4080 * simply cache them. Otherwise we need to sort them back to 4081 * correct domains. 4082 */ 4083 if (domain != itemdomain && vm_ndomains > 2) { 4084 zone_free_cross(zone, bucket, udata); 4085 return; 4086 } 4087 #endif 4088 4089 /* 4090 * Attempt to save the bucket in the zone's domain bucket cache. 4091 * 4092 * We bump the uz count when the cache size is insufficient to 4093 * handle the working set. 4094 */ 4095 if (ZONE_TRYLOCK(zone) == 0) { 4096 /* Record contention to size the buckets. */ 4097 ZONE_LOCK(zone); 4098 if (zone->uz_bucket_size < zone->uz_bucket_size_max) 4099 zone->uz_bucket_size++; 4100 } 4101 4102 CTR3(KTR_UMA, 4103 "uma_zfree: zone %s(%p) putting bucket %p on free list", 4104 zone->uz_name, zone, bucket); 4105 /* ub_cnt is pointing to the last free item */ 4106 KASSERT(bucket->ub_cnt == bucket->ub_entries, 4107 ("uma_zfree: Attempting to insert partial bucket onto the full list.\n")); 4108 if (zone->uz_bkt_count >= zone->uz_bkt_max) { 4109 ZONE_UNLOCK(zone); 4110 bucket_drain(zone, bucket); 4111 bucket_free(zone, bucket, udata); 4112 } else { 4113 zdom = &zone->uz_domain[itemdomain]; 4114 zone_put_bucket(zone, zdom, bucket, true); 4115 ZONE_UNLOCK(zone); 4116 } 4117 } 4118 4119 /* 4120 * Populate a free or cross bucket for the current cpu cache. Free any 4121 * existing full bucket either to the zone cache or back to the slab layer. 4122 * 4123 * Enters and returns in a critical section. false return indicates that 4124 * we can not satisfy this free in the cache layer. true indicates that 4125 * the caller should retry. 4126 */ 4127 static __noinline bool 4128 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item, 4129 int itemdomain) 4130 { 4131 uma_cache_bucket_t cbucket; 4132 uma_bucket_t newbucket, bucket; 4133 int domain; 4134 4135 CRITICAL_ASSERT(curthread); 4136 4137 if (zone->uz_bucket_size == 0) 4138 return false; 4139 4140 cache = &zone->uz_cpu[curcpu]; 4141 newbucket = NULL; 4142 4143 /* 4144 * FIRSTTOUCH domains need to free to the correct zdom. When 4145 * enabled this is the zdom of the item. The bucket is the 4146 * cross bucket if the current domain and itemdomain do not match. 4147 */ 4148 cbucket = &cache->uc_freebucket; 4149 #ifdef NUMA 4150 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) { 4151 domain = PCPU_GET(domain); 4152 if (domain != itemdomain) { 4153 cbucket = &cache->uc_crossbucket; 4154 if (cbucket->ucb_cnt != 0) 4155 atomic_add_64(&zone->uz_xdomain, 4156 cbucket->ucb_cnt); 4157 } 4158 } else 4159 #endif 4160 itemdomain = domain = 0; 4161 bucket = cache_bucket_unload(cbucket); 4162 4163 /* We are no longer associated with this CPU. */ 4164 critical_exit(); 4165 4166 /* 4167 * Don't let SMR zones operate without a free bucket. Force 4168 * a synchronize and re-use this one. We will only degrade 4169 * to a synchronize every bucket_size items rather than every 4170 * item if we fail to allocate a bucket. 4171 */ 4172 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) { 4173 if (bucket != NULL) 4174 bucket->ub_seq = smr_advance(zone->uz_smr); 4175 newbucket = bucket_alloc(zone, udata, M_NOWAIT); 4176 if (newbucket == NULL && bucket != NULL) { 4177 bucket_drain(zone, bucket); 4178 newbucket = bucket; 4179 bucket = NULL; 4180 } 4181 } else if (!bucketdisable) 4182 newbucket = bucket_alloc(zone, udata, M_NOWAIT); 4183 4184 if (bucket != NULL) 4185 zone_free_bucket(zone, bucket, udata, domain, itemdomain); 4186 4187 critical_enter(); 4188 if ((bucket = newbucket) == NULL) 4189 return (false); 4190 cache = &zone->uz_cpu[curcpu]; 4191 #ifdef NUMA 4192 /* 4193 * Check to see if we should be populating the cross bucket. If it 4194 * is already populated we will fall through and attempt to populate 4195 * the free bucket. 4196 */ 4197 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) { 4198 domain = PCPU_GET(domain); 4199 if (domain != itemdomain && 4200 cache->uc_crossbucket.ucb_bucket == NULL) { 4201 cache_bucket_load_cross(cache, bucket); 4202 return (true); 4203 } 4204 } 4205 #endif 4206 /* 4207 * We may have lost the race to fill the bucket or switched CPUs. 4208 */ 4209 if (cache->uc_freebucket.ucb_bucket != NULL) { 4210 critical_exit(); 4211 bucket_free(zone, bucket, udata); 4212 critical_enter(); 4213 } else 4214 cache_bucket_load_free(cache, bucket); 4215 4216 return (true); 4217 } 4218 4219 void 4220 uma_zfree_domain(uma_zone_t zone, void *item, void *udata) 4221 { 4222 4223 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 4224 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 4225 4226 CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone); 4227 4228 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 4229 ("uma_zfree_domain: called with spinlock or critical section held")); 4230 4231 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 4232 if (item == NULL) 4233 return; 4234 zone_free_item(zone, item, udata, SKIP_NONE); 4235 } 4236 4237 static void 4238 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item) 4239 { 4240 uma_keg_t keg; 4241 uma_domain_t dom; 4242 int freei; 4243 4244 keg = zone->uz_keg; 4245 KEG_LOCK_ASSERT(keg, slab->us_domain); 4246 4247 /* Do we need to remove from any lists? */ 4248 dom = &keg->uk_domain[slab->us_domain]; 4249 if (slab->us_freecount+1 == keg->uk_ipers) { 4250 LIST_REMOVE(slab, us_link); 4251 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); 4252 } else if (slab->us_freecount == 0) { 4253 LIST_REMOVE(slab, us_link); 4254 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 4255 } 4256 4257 /* Slab management. */ 4258 freei = slab_item_index(slab, keg, item); 4259 BIT_SET(keg->uk_ipers, freei, &slab->us_free); 4260 slab->us_freecount++; 4261 4262 /* Keg statistics. */ 4263 dom->ud_free++; 4264 } 4265 4266 static void 4267 zone_release(void *arg, void **bucket, int cnt) 4268 { 4269 struct mtx *lock; 4270 uma_zone_t zone; 4271 uma_slab_t slab; 4272 uma_keg_t keg; 4273 uint8_t *mem; 4274 void *item; 4275 int i; 4276 4277 zone = arg; 4278 keg = zone->uz_keg; 4279 lock = NULL; 4280 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0)) 4281 lock = KEG_LOCK(keg, 0); 4282 for (i = 0; i < cnt; i++) { 4283 item = bucket[i]; 4284 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) { 4285 slab = vtoslab((vm_offset_t)item); 4286 } else { 4287 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 4288 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0) 4289 slab = hash_sfind(&keg->uk_hash, mem); 4290 else 4291 slab = (uma_slab_t)(mem + keg->uk_pgoff); 4292 } 4293 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) { 4294 if (lock != NULL) 4295 mtx_unlock(lock); 4296 lock = KEG_LOCK(keg, slab->us_domain); 4297 } 4298 slab_free_item(zone, slab, item); 4299 } 4300 if (lock != NULL) 4301 mtx_unlock(lock); 4302 } 4303 4304 /* 4305 * Frees a single item to any zone. 4306 * 4307 * Arguments: 4308 * zone The zone to free to 4309 * item The item we're freeing 4310 * udata User supplied data for the dtor 4311 * skip Skip dtors and finis 4312 */ 4313 static void 4314 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 4315 { 4316 4317 /* 4318 * If a free is sent directly to an SMR zone we have to 4319 * synchronize immediately because the item can instantly 4320 * be reallocated. This should only happen in degenerate 4321 * cases when no memory is available for per-cpu caches. 4322 */ 4323 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE) 4324 smr_synchronize(zone->uz_smr); 4325 4326 item_dtor(zone, item, zone->uz_size, udata, skip); 4327 4328 if (skip < SKIP_FINI && zone->uz_fini) 4329 zone->uz_fini(item, zone->uz_size); 4330 4331 zone->uz_release(zone->uz_arg, &item, 1); 4332 4333 if (skip & SKIP_CNT) 4334 return; 4335 4336 counter_u64_add(zone->uz_frees, 1); 4337 4338 if (zone->uz_max_items > 0) 4339 zone_free_limit(zone, 1); 4340 } 4341 4342 /* See uma.h */ 4343 int 4344 uma_zone_set_max(uma_zone_t zone, int nitems) 4345 { 4346 struct uma_bucket_zone *ubz; 4347 int count; 4348 4349 /* 4350 * XXX This can misbehave if the zone has any allocations with 4351 * no limit and a limit is imposed. There is currently no 4352 * way to clear a limit. 4353 */ 4354 ZONE_LOCK(zone); 4355 ubz = bucket_zone_max(zone, nitems); 4356 count = ubz != NULL ? ubz->ubz_entries : 0; 4357 zone->uz_bucket_size_max = zone->uz_bucket_size = count; 4358 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max) 4359 zone->uz_bucket_size_min = zone->uz_bucket_size_max; 4360 zone->uz_max_items = nitems; 4361 zone->uz_flags |= UMA_ZFLAG_LIMIT; 4362 zone_update_caches(zone); 4363 /* We may need to wake waiters. */ 4364 wakeup(&zone->uz_max_items); 4365 ZONE_UNLOCK(zone); 4366 4367 return (nitems); 4368 } 4369 4370 /* See uma.h */ 4371 void 4372 uma_zone_set_maxcache(uma_zone_t zone, int nitems) 4373 { 4374 struct uma_bucket_zone *ubz; 4375 int bpcpu; 4376 4377 ZONE_LOCK(zone); 4378 ubz = bucket_zone_max(zone, nitems); 4379 if (ubz != NULL) { 4380 bpcpu = 2; 4381 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 4382 /* Count the cross-domain bucket. */ 4383 bpcpu++; 4384 nitems -= ubz->ubz_entries * bpcpu * mp_ncpus; 4385 zone->uz_bucket_size_max = ubz->ubz_entries; 4386 } else { 4387 zone->uz_bucket_size_max = zone->uz_bucket_size = 0; 4388 } 4389 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max) 4390 zone->uz_bucket_size_min = zone->uz_bucket_size_max; 4391 zone->uz_bkt_max = nitems; 4392 ZONE_UNLOCK(zone); 4393 } 4394 4395 /* See uma.h */ 4396 int 4397 uma_zone_get_max(uma_zone_t zone) 4398 { 4399 int nitems; 4400 4401 nitems = atomic_load_64(&zone->uz_max_items); 4402 4403 return (nitems); 4404 } 4405 4406 /* See uma.h */ 4407 void 4408 uma_zone_set_warning(uma_zone_t zone, const char *warning) 4409 { 4410 4411 ZONE_ASSERT_COLD(zone); 4412 zone->uz_warning = warning; 4413 } 4414 4415 /* See uma.h */ 4416 void 4417 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) 4418 { 4419 4420 ZONE_ASSERT_COLD(zone); 4421 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); 4422 } 4423 4424 /* See uma.h */ 4425 int 4426 uma_zone_get_cur(uma_zone_t zone) 4427 { 4428 int64_t nitems; 4429 u_int i; 4430 4431 nitems = 0; 4432 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER) 4433 nitems = counter_u64_fetch(zone->uz_allocs) - 4434 counter_u64_fetch(zone->uz_frees); 4435 CPU_FOREACH(i) 4436 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) - 4437 atomic_load_64(&zone->uz_cpu[i].uc_frees); 4438 4439 return (nitems < 0 ? 0 : nitems); 4440 } 4441 4442 static uint64_t 4443 uma_zone_get_allocs(uma_zone_t zone) 4444 { 4445 uint64_t nitems; 4446 u_int i; 4447 4448 nitems = 0; 4449 if (zone->uz_allocs != EARLY_COUNTER) 4450 nitems = counter_u64_fetch(zone->uz_allocs); 4451 CPU_FOREACH(i) 4452 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs); 4453 4454 return (nitems); 4455 } 4456 4457 static uint64_t 4458 uma_zone_get_frees(uma_zone_t zone) 4459 { 4460 uint64_t nitems; 4461 u_int i; 4462 4463 nitems = 0; 4464 if (zone->uz_frees != EARLY_COUNTER) 4465 nitems = counter_u64_fetch(zone->uz_frees); 4466 CPU_FOREACH(i) 4467 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees); 4468 4469 return (nitems); 4470 } 4471 4472 #ifdef INVARIANTS 4473 /* Used only for KEG_ASSERT_COLD(). */ 4474 static uint64_t 4475 uma_keg_get_allocs(uma_keg_t keg) 4476 { 4477 uma_zone_t z; 4478 uint64_t nitems; 4479 4480 nitems = 0; 4481 LIST_FOREACH(z, &keg->uk_zones, uz_link) 4482 nitems += uma_zone_get_allocs(z); 4483 4484 return (nitems); 4485 } 4486 #endif 4487 4488 /* See uma.h */ 4489 void 4490 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 4491 { 4492 uma_keg_t keg; 4493 4494 KEG_GET(zone, keg); 4495 KEG_ASSERT_COLD(keg); 4496 keg->uk_init = uminit; 4497 } 4498 4499 /* See uma.h */ 4500 void 4501 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 4502 { 4503 uma_keg_t keg; 4504 4505 KEG_GET(zone, keg); 4506 KEG_ASSERT_COLD(keg); 4507 keg->uk_fini = fini; 4508 } 4509 4510 /* See uma.h */ 4511 void 4512 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 4513 { 4514 4515 ZONE_ASSERT_COLD(zone); 4516 zone->uz_init = zinit; 4517 } 4518 4519 /* See uma.h */ 4520 void 4521 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 4522 { 4523 4524 ZONE_ASSERT_COLD(zone); 4525 zone->uz_fini = zfini; 4526 } 4527 4528 /* See uma.h */ 4529 void 4530 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 4531 { 4532 uma_keg_t keg; 4533 4534 KEG_GET(zone, keg); 4535 KEG_ASSERT_COLD(keg); 4536 keg->uk_freef = freef; 4537 } 4538 4539 /* See uma.h */ 4540 void 4541 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 4542 { 4543 uma_keg_t keg; 4544 4545 KEG_GET(zone, keg); 4546 KEG_ASSERT_COLD(keg); 4547 keg->uk_allocf = allocf; 4548 } 4549 4550 /* See uma.h */ 4551 void 4552 uma_zone_set_smr(uma_zone_t zone, smr_t smr) 4553 { 4554 4555 ZONE_ASSERT_COLD(zone); 4556 4557 zone->uz_flags |= UMA_ZONE_SMR; 4558 zone->uz_smr = smr; 4559 zone_update_caches(zone); 4560 } 4561 4562 smr_t 4563 uma_zone_get_smr(uma_zone_t zone) 4564 { 4565 4566 return (zone->uz_smr); 4567 } 4568 4569 /* See uma.h */ 4570 void 4571 uma_zone_reserve(uma_zone_t zone, int items) 4572 { 4573 uma_keg_t keg; 4574 4575 KEG_GET(zone, keg); 4576 KEG_ASSERT_COLD(keg); 4577 keg->uk_reserve = items; 4578 } 4579 4580 /* See uma.h */ 4581 int 4582 uma_zone_reserve_kva(uma_zone_t zone, int count) 4583 { 4584 uma_keg_t keg; 4585 vm_offset_t kva; 4586 u_int pages; 4587 4588 KEG_GET(zone, keg); 4589 KEG_ASSERT_COLD(keg); 4590 ZONE_ASSERT_COLD(zone); 4591 4592 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera; 4593 4594 #ifdef UMA_MD_SMALL_ALLOC 4595 if (keg->uk_ppera > 1) { 4596 #else 4597 if (1) { 4598 #endif 4599 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE); 4600 if (kva == 0) 4601 return (0); 4602 } else 4603 kva = 0; 4604 4605 ZONE_LOCK(zone); 4606 MPASS(keg->uk_kva == 0); 4607 keg->uk_kva = kva; 4608 keg->uk_offset = 0; 4609 zone->uz_max_items = pages * keg->uk_ipers; 4610 #ifdef UMA_MD_SMALL_ALLOC 4611 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 4612 #else 4613 keg->uk_allocf = noobj_alloc; 4614 #endif 4615 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE; 4616 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE; 4617 zone_update_caches(zone); 4618 ZONE_UNLOCK(zone); 4619 4620 return (1); 4621 } 4622 4623 /* See uma.h */ 4624 void 4625 uma_prealloc(uma_zone_t zone, int items) 4626 { 4627 struct vm_domainset_iter di; 4628 uma_domain_t dom; 4629 uma_slab_t slab; 4630 uma_keg_t keg; 4631 int aflags, domain, slabs; 4632 4633 KEG_GET(zone, keg); 4634 slabs = howmany(items, keg->uk_ipers); 4635 while (slabs-- > 0) { 4636 aflags = M_NOWAIT; 4637 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 4638 &aflags); 4639 for (;;) { 4640 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK, 4641 aflags); 4642 if (slab != NULL) { 4643 dom = &keg->uk_domain[slab->us_domain]; 4644 LIST_REMOVE(slab, us_link); 4645 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, 4646 us_link); 4647 KEG_UNLOCK(keg, slab->us_domain); 4648 break; 4649 } 4650 if (vm_domainset_iter_policy(&di, &domain) != 0) 4651 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask); 4652 } 4653 } 4654 } 4655 4656 /* See uma.h */ 4657 void 4658 uma_reclaim(int req) 4659 { 4660 4661 CTR0(KTR_UMA, "UMA: vm asked us to release pages!"); 4662 sx_xlock(&uma_reclaim_lock); 4663 bucket_enable(); 4664 4665 switch (req) { 4666 case UMA_RECLAIM_TRIM: 4667 zone_foreach(zone_trim, NULL); 4668 break; 4669 case UMA_RECLAIM_DRAIN: 4670 case UMA_RECLAIM_DRAIN_CPU: 4671 zone_foreach(zone_drain, NULL); 4672 if (req == UMA_RECLAIM_DRAIN_CPU) { 4673 pcpu_cache_drain_safe(NULL); 4674 zone_foreach(zone_drain, NULL); 4675 } 4676 break; 4677 default: 4678 panic("unhandled reclamation request %d", req); 4679 } 4680 4681 /* 4682 * Some slabs may have been freed but this zone will be visited early 4683 * we visit again so that we can free pages that are empty once other 4684 * zones are drained. We have to do the same for buckets. 4685 */ 4686 zone_drain(slabzones[0], NULL); 4687 zone_drain(slabzones[1], NULL); 4688 bucket_zone_drain(); 4689 sx_xunlock(&uma_reclaim_lock); 4690 } 4691 4692 static volatile int uma_reclaim_needed; 4693 4694 void 4695 uma_reclaim_wakeup(void) 4696 { 4697 4698 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0) 4699 wakeup(uma_reclaim); 4700 } 4701 4702 void 4703 uma_reclaim_worker(void *arg __unused) 4704 { 4705 4706 for (;;) { 4707 sx_xlock(&uma_reclaim_lock); 4708 while (atomic_load_int(&uma_reclaim_needed) == 0) 4709 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl", 4710 hz); 4711 sx_xunlock(&uma_reclaim_lock); 4712 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); 4713 uma_reclaim(UMA_RECLAIM_DRAIN_CPU); 4714 atomic_store_int(&uma_reclaim_needed, 0); 4715 /* Don't fire more than once per-second. */ 4716 pause("umarclslp", hz); 4717 } 4718 } 4719 4720 /* See uma.h */ 4721 void 4722 uma_zone_reclaim(uma_zone_t zone, int req) 4723 { 4724 4725 switch (req) { 4726 case UMA_RECLAIM_TRIM: 4727 zone_trim(zone, NULL); 4728 break; 4729 case UMA_RECLAIM_DRAIN: 4730 zone_drain(zone, NULL); 4731 break; 4732 case UMA_RECLAIM_DRAIN_CPU: 4733 pcpu_cache_drain_safe(zone); 4734 zone_drain(zone, NULL); 4735 break; 4736 default: 4737 panic("unhandled reclamation request %d", req); 4738 } 4739 } 4740 4741 /* See uma.h */ 4742 int 4743 uma_zone_exhausted(uma_zone_t zone) 4744 { 4745 4746 return (atomic_load_32(&zone->uz_sleepers) > 0); 4747 } 4748 4749 unsigned long 4750 uma_limit(void) 4751 { 4752 4753 return (uma_kmem_limit); 4754 } 4755 4756 void 4757 uma_set_limit(unsigned long limit) 4758 { 4759 4760 uma_kmem_limit = limit; 4761 } 4762 4763 unsigned long 4764 uma_size(void) 4765 { 4766 4767 return (atomic_load_long(&uma_kmem_total)); 4768 } 4769 4770 long 4771 uma_avail(void) 4772 { 4773 4774 return (uma_kmem_limit - uma_size()); 4775 } 4776 4777 #ifdef DDB 4778 /* 4779 * Generate statistics across both the zone and its per-cpu cache's. Return 4780 * desired statistics if the pointer is non-NULL for that statistic. 4781 * 4782 * Note: does not update the zone statistics, as it can't safely clear the 4783 * per-CPU cache statistic. 4784 * 4785 */ 4786 static void 4787 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp, 4788 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp) 4789 { 4790 uma_cache_t cache; 4791 uint64_t allocs, frees, sleeps, xdomain; 4792 int cachefree, cpu; 4793 4794 allocs = frees = sleeps = xdomain = 0; 4795 cachefree = 0; 4796 CPU_FOREACH(cpu) { 4797 cache = &z->uz_cpu[cpu]; 4798 cachefree += cache->uc_allocbucket.ucb_cnt; 4799 cachefree += cache->uc_freebucket.ucb_cnt; 4800 xdomain += cache->uc_crossbucket.ucb_cnt; 4801 cachefree += cache->uc_crossbucket.ucb_cnt; 4802 allocs += cache->uc_allocs; 4803 frees += cache->uc_frees; 4804 } 4805 allocs += counter_u64_fetch(z->uz_allocs); 4806 frees += counter_u64_fetch(z->uz_frees); 4807 sleeps += z->uz_sleeps; 4808 xdomain += z->uz_xdomain; 4809 if (cachefreep != NULL) 4810 *cachefreep = cachefree; 4811 if (allocsp != NULL) 4812 *allocsp = allocs; 4813 if (freesp != NULL) 4814 *freesp = frees; 4815 if (sleepsp != NULL) 4816 *sleepsp = sleeps; 4817 if (xdomainp != NULL) 4818 *xdomainp = xdomain; 4819 } 4820 #endif /* DDB */ 4821 4822 static int 4823 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 4824 { 4825 uma_keg_t kz; 4826 uma_zone_t z; 4827 int count; 4828 4829 count = 0; 4830 rw_rlock(&uma_rwlock); 4831 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4832 LIST_FOREACH(z, &kz->uk_zones, uz_link) 4833 count++; 4834 } 4835 LIST_FOREACH(z, &uma_cachezones, uz_link) 4836 count++; 4837 4838 rw_runlock(&uma_rwlock); 4839 return (sysctl_handle_int(oidp, &count, 0, req)); 4840 } 4841 4842 static void 4843 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf, 4844 struct uma_percpu_stat *ups, bool internal) 4845 { 4846 uma_zone_domain_t zdom; 4847 uma_cache_t cache; 4848 int i; 4849 4850 4851 for (i = 0; i < vm_ndomains; i++) { 4852 zdom = &z->uz_domain[i]; 4853 uth->uth_zone_free += zdom->uzd_nitems; 4854 } 4855 uth->uth_allocs = counter_u64_fetch(z->uz_allocs); 4856 uth->uth_frees = counter_u64_fetch(z->uz_frees); 4857 uth->uth_fails = counter_u64_fetch(z->uz_fails); 4858 uth->uth_sleeps = z->uz_sleeps; 4859 uth->uth_xdomain = z->uz_xdomain; 4860 4861 /* 4862 * While it is not normally safe to access the cache bucket pointers 4863 * while not on the CPU that owns the cache, we only allow the pointers 4864 * to be exchanged without the zone lock held, not invalidated, so 4865 * accept the possible race associated with bucket exchange during 4866 * monitoring. Use atomic_load_ptr() to ensure that the bucket pointers 4867 * are loaded only once. 4868 */ 4869 for (i = 0; i < mp_maxid + 1; i++) { 4870 bzero(&ups[i], sizeof(*ups)); 4871 if (internal || CPU_ABSENT(i)) 4872 continue; 4873 cache = &z->uz_cpu[i]; 4874 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt; 4875 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt; 4876 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt; 4877 ups[i].ups_allocs = cache->uc_allocs; 4878 ups[i].ups_frees = cache->uc_frees; 4879 } 4880 } 4881 4882 static int 4883 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 4884 { 4885 struct uma_stream_header ush; 4886 struct uma_type_header uth; 4887 struct uma_percpu_stat *ups; 4888 struct sbuf sbuf; 4889 uma_keg_t kz; 4890 uma_zone_t z; 4891 uint64_t items; 4892 uint32_t kfree, pages; 4893 int count, error, i; 4894 4895 error = sysctl_wire_old_buffer(req, 0); 4896 if (error != 0) 4897 return (error); 4898 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 4899 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); 4900 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK); 4901 4902 count = 0; 4903 rw_rlock(&uma_rwlock); 4904 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4905 LIST_FOREACH(z, &kz->uk_zones, uz_link) 4906 count++; 4907 } 4908 4909 LIST_FOREACH(z, &uma_cachezones, uz_link) 4910 count++; 4911 4912 /* 4913 * Insert stream header. 4914 */ 4915 bzero(&ush, sizeof(ush)); 4916 ush.ush_version = UMA_STREAM_VERSION; 4917 ush.ush_maxcpus = (mp_maxid + 1); 4918 ush.ush_count = count; 4919 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 4920 4921 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4922 kfree = pages = 0; 4923 for (i = 0; i < vm_ndomains; i++) { 4924 kfree += kz->uk_domain[i].ud_free; 4925 pages += kz->uk_domain[i].ud_pages; 4926 } 4927 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 4928 bzero(&uth, sizeof(uth)); 4929 ZONE_LOCK(z); 4930 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 4931 uth.uth_align = kz->uk_align; 4932 uth.uth_size = kz->uk_size; 4933 uth.uth_rsize = kz->uk_rsize; 4934 if (z->uz_max_items > 0) { 4935 items = UZ_ITEMS_COUNT(z->uz_items); 4936 uth.uth_pages = (items / kz->uk_ipers) * 4937 kz->uk_ppera; 4938 } else 4939 uth.uth_pages = pages; 4940 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) * 4941 kz->uk_ppera; 4942 uth.uth_limit = z->uz_max_items; 4943 uth.uth_keg_free = kfree; 4944 4945 /* 4946 * A zone is secondary is it is not the first entry 4947 * on the keg's zone list. 4948 */ 4949 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 4950 (LIST_FIRST(&kz->uk_zones) != z)) 4951 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 4952 uma_vm_zone_stats(&uth, z, &sbuf, ups, 4953 kz->uk_flags & UMA_ZFLAG_INTERNAL); 4954 ZONE_UNLOCK(z); 4955 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 4956 for (i = 0; i < mp_maxid + 1; i++) 4957 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 4958 } 4959 } 4960 LIST_FOREACH(z, &uma_cachezones, uz_link) { 4961 bzero(&uth, sizeof(uth)); 4962 ZONE_LOCK(z); 4963 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 4964 uth.uth_size = z->uz_size; 4965 uma_vm_zone_stats(&uth, z, &sbuf, ups, false); 4966 ZONE_UNLOCK(z); 4967 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 4968 for (i = 0; i < mp_maxid + 1; i++) 4969 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 4970 } 4971 4972 rw_runlock(&uma_rwlock); 4973 error = sbuf_finish(&sbuf); 4974 sbuf_delete(&sbuf); 4975 free(ups, M_TEMP); 4976 return (error); 4977 } 4978 4979 int 4980 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 4981 { 4982 uma_zone_t zone = *(uma_zone_t *)arg1; 4983 int error, max; 4984 4985 max = uma_zone_get_max(zone); 4986 error = sysctl_handle_int(oidp, &max, 0, req); 4987 if (error || !req->newptr) 4988 return (error); 4989 4990 uma_zone_set_max(zone, max); 4991 4992 return (0); 4993 } 4994 4995 int 4996 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 4997 { 4998 uma_zone_t zone; 4999 int cur; 5000 5001 /* 5002 * Some callers want to add sysctls for global zones that 5003 * may not yet exist so they pass a pointer to a pointer. 5004 */ 5005 if (arg2 == 0) 5006 zone = *(uma_zone_t *)arg1; 5007 else 5008 zone = arg1; 5009 cur = uma_zone_get_cur(zone); 5010 return (sysctl_handle_int(oidp, &cur, 0, req)); 5011 } 5012 5013 static int 5014 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS) 5015 { 5016 uma_zone_t zone = arg1; 5017 uint64_t cur; 5018 5019 cur = uma_zone_get_allocs(zone); 5020 return (sysctl_handle_64(oidp, &cur, 0, req)); 5021 } 5022 5023 static int 5024 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS) 5025 { 5026 uma_zone_t zone = arg1; 5027 uint64_t cur; 5028 5029 cur = uma_zone_get_frees(zone); 5030 return (sysctl_handle_64(oidp, &cur, 0, req)); 5031 } 5032 5033 static int 5034 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS) 5035 { 5036 struct sbuf sbuf; 5037 uma_zone_t zone = arg1; 5038 int error; 5039 5040 sbuf_new_for_sysctl(&sbuf, NULL, 0, req); 5041 if (zone->uz_flags != 0) 5042 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS); 5043 else 5044 sbuf_printf(&sbuf, "0"); 5045 error = sbuf_finish(&sbuf); 5046 sbuf_delete(&sbuf); 5047 5048 return (error); 5049 } 5050 5051 static int 5052 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS) 5053 { 5054 uma_keg_t keg = arg1; 5055 int avail, effpct, total; 5056 5057 total = keg->uk_ppera * PAGE_SIZE; 5058 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0) 5059 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize; 5060 /* 5061 * We consider the client's requested size and alignment here, not the 5062 * real size determination uk_rsize, because we also adjust the real 5063 * size for internal implementation reasons (max bitset size). 5064 */ 5065 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1); 5066 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) 5067 avail *= mp_maxid + 1; 5068 effpct = 100 * avail / total; 5069 return (sysctl_handle_int(oidp, &effpct, 0, req)); 5070 } 5071 5072 static int 5073 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS) 5074 { 5075 uma_zone_t zone = arg1; 5076 uint64_t cur; 5077 5078 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items)); 5079 return (sysctl_handle_64(oidp, &cur, 0, req)); 5080 } 5081 5082 #ifdef INVARIANTS 5083 static uma_slab_t 5084 uma_dbg_getslab(uma_zone_t zone, void *item) 5085 { 5086 uma_slab_t slab; 5087 uma_keg_t keg; 5088 uint8_t *mem; 5089 5090 /* 5091 * It is safe to return the slab here even though the 5092 * zone is unlocked because the item's allocation state 5093 * essentially holds a reference. 5094 */ 5095 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 5096 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 5097 return (NULL); 5098 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB) 5099 return (vtoslab((vm_offset_t)mem)); 5100 keg = zone->uz_keg; 5101 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0) 5102 return ((uma_slab_t)(mem + keg->uk_pgoff)); 5103 KEG_LOCK(keg, 0); 5104 slab = hash_sfind(&keg->uk_hash, mem); 5105 KEG_UNLOCK(keg, 0); 5106 5107 return (slab); 5108 } 5109 5110 static bool 5111 uma_dbg_zskip(uma_zone_t zone, void *mem) 5112 { 5113 5114 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 5115 return (true); 5116 5117 return (uma_dbg_kskip(zone->uz_keg, mem)); 5118 } 5119 5120 static bool 5121 uma_dbg_kskip(uma_keg_t keg, void *mem) 5122 { 5123 uintptr_t idx; 5124 5125 if (dbg_divisor == 0) 5126 return (true); 5127 5128 if (dbg_divisor == 1) 5129 return (false); 5130 5131 idx = (uintptr_t)mem >> PAGE_SHIFT; 5132 if (keg->uk_ipers > 1) { 5133 idx *= keg->uk_ipers; 5134 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize; 5135 } 5136 5137 if ((idx / dbg_divisor) * dbg_divisor != idx) { 5138 counter_u64_add(uma_skip_cnt, 1); 5139 return (true); 5140 } 5141 counter_u64_add(uma_dbg_cnt, 1); 5142 5143 return (false); 5144 } 5145 5146 /* 5147 * Set up the slab's freei data such that uma_dbg_free can function. 5148 * 5149 */ 5150 static void 5151 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) 5152 { 5153 uma_keg_t keg; 5154 int freei; 5155 5156 if (slab == NULL) { 5157 slab = uma_dbg_getslab(zone, item); 5158 if (slab == NULL) 5159 panic("uma: item %p did not belong to zone %s\n", 5160 item, zone->uz_name); 5161 } 5162 keg = zone->uz_keg; 5163 freei = slab_item_index(slab, keg, item); 5164 5165 if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg))) 5166 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n", 5167 item, zone, zone->uz_name, slab, freei); 5168 BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)); 5169 } 5170 5171 /* 5172 * Verifies freed addresses. Checks for alignment, valid slab membership 5173 * and duplicate frees. 5174 * 5175 */ 5176 static void 5177 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) 5178 { 5179 uma_keg_t keg; 5180 int freei; 5181 5182 if (slab == NULL) { 5183 slab = uma_dbg_getslab(zone, item); 5184 if (slab == NULL) 5185 panic("uma: Freed item %p did not belong to zone %s\n", 5186 item, zone->uz_name); 5187 } 5188 keg = zone->uz_keg; 5189 freei = slab_item_index(slab, keg, item); 5190 5191 if (freei >= keg->uk_ipers) 5192 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n", 5193 item, zone, zone->uz_name, slab, freei); 5194 5195 if (slab_item(slab, keg, freei) != item) 5196 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n", 5197 item, zone, zone->uz_name, slab, freei); 5198 5199 if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg))) 5200 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n", 5201 item, zone, zone->uz_name, slab, freei); 5202 5203 BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)); 5204 } 5205 #endif /* INVARIANTS */ 5206 5207 #ifdef DDB 5208 static int64_t 5209 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used, 5210 uint64_t *sleeps, long *cachefree, uint64_t *xdomain) 5211 { 5212 uint64_t frees; 5213 int i; 5214 5215 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 5216 *allocs = counter_u64_fetch(z->uz_allocs); 5217 frees = counter_u64_fetch(z->uz_frees); 5218 *sleeps = z->uz_sleeps; 5219 *cachefree = 0; 5220 *xdomain = 0; 5221 } else 5222 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps, 5223 xdomain); 5224 for (i = 0; i < vm_ndomains; i++) { 5225 *cachefree += z->uz_domain[i].uzd_nitems; 5226 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 5227 (LIST_FIRST(&kz->uk_zones) != z))) 5228 *cachefree += kz->uk_domain[i].ud_free; 5229 } 5230 *used = *allocs - frees; 5231 return (((int64_t)*used + *cachefree) * kz->uk_size); 5232 } 5233 5234 DB_SHOW_COMMAND(uma, db_show_uma) 5235 { 5236 const char *fmt_hdr, *fmt_entry; 5237 uma_keg_t kz; 5238 uma_zone_t z; 5239 uint64_t allocs, used, sleeps, xdomain; 5240 long cachefree; 5241 /* variables for sorting */ 5242 uma_keg_t cur_keg; 5243 uma_zone_t cur_zone, last_zone; 5244 int64_t cur_size, last_size, size; 5245 int ties; 5246 5247 /* /i option produces machine-parseable CSV output */ 5248 if (modif[0] == 'i') { 5249 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n"; 5250 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n"; 5251 } else { 5252 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n"; 5253 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n"; 5254 } 5255 5256 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests", 5257 "Sleeps", "Bucket", "Total Mem", "XFree"); 5258 5259 /* Sort the zones with largest size first. */ 5260 last_zone = NULL; 5261 last_size = INT64_MAX; 5262 for (;;) { 5263 cur_zone = NULL; 5264 cur_size = -1; 5265 ties = 0; 5266 LIST_FOREACH(kz, &uma_kegs, uk_link) { 5267 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 5268 /* 5269 * In the case of size ties, print out zones 5270 * in the order they are encountered. That is, 5271 * when we encounter the most recently output 5272 * zone, we have already printed all preceding 5273 * ties, and we must print all following ties. 5274 */ 5275 if (z == last_zone) { 5276 ties = 1; 5277 continue; 5278 } 5279 size = get_uma_stats(kz, z, &allocs, &used, 5280 &sleeps, &cachefree, &xdomain); 5281 if (size > cur_size && size < last_size + ties) 5282 { 5283 cur_size = size; 5284 cur_zone = z; 5285 cur_keg = kz; 5286 } 5287 } 5288 } 5289 if (cur_zone == NULL) 5290 break; 5291 5292 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used, 5293 &sleeps, &cachefree, &xdomain); 5294 db_printf(fmt_entry, cur_zone->uz_name, 5295 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree, 5296 (uintmax_t)allocs, (uintmax_t)sleeps, 5297 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size, 5298 xdomain); 5299 5300 if (db_pager_quit) 5301 return; 5302 last_zone = cur_zone; 5303 last_size = cur_size; 5304 } 5305 } 5306 5307 DB_SHOW_COMMAND(umacache, db_show_umacache) 5308 { 5309 uma_zone_t z; 5310 uint64_t allocs, frees; 5311 long cachefree; 5312 int i; 5313 5314 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 5315 "Requests", "Bucket"); 5316 LIST_FOREACH(z, &uma_cachezones, uz_link) { 5317 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL); 5318 for (i = 0; i < vm_ndomains; i++) 5319 cachefree += z->uz_domain[i].uzd_nitems; 5320 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n", 5321 z->uz_name, (uintmax_t)z->uz_size, 5322 (intmax_t)(allocs - frees), cachefree, 5323 (uintmax_t)allocs, z->uz_bucket_size); 5324 if (db_pager_quit) 5325 return; 5326 } 5327 } 5328 #endif /* DDB */ 5329