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_ZONE_VM) != 0) 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 /* SMR Buckets can not be re-used until readers expire. */ 557 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && 558 bucket->ub_seq != SMR_SEQ_INVALID) { 559 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false)) 560 return (NULL); 561 bucket->ub_seq = SMR_SEQ_INVALID; 562 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR; 563 } 564 MPASS(zdom->uzd_nitems >= bucket->ub_cnt); 565 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link); 566 zdom->uzd_nitems -= bucket->ub_cnt; 567 if (zdom->uzd_imin > zdom->uzd_nitems) 568 zdom->uzd_imin = zdom->uzd_nitems; 569 zone->uz_bkt_count -= bucket->ub_cnt; 570 ZONE_UNLOCK(zone); 571 if (dtor) 572 for (i = 0; i < bucket->ub_cnt; i++) 573 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size, 574 NULL, SKIP_NONE); 575 576 return (bucket); 577 } 578 579 /* 580 * Insert a full bucket into the specified cache. The "ws" parameter indicates 581 * whether the bucket's contents should be counted as part of the zone's working 582 * set. 583 */ 584 static void 585 zone_put_bucket(uma_zone_t zone, uma_zone_domain_t zdom, uma_bucket_t bucket, 586 const bool ws) 587 { 588 589 ZONE_LOCK_ASSERT(zone); 590 KASSERT(!ws || zone->uz_bkt_count < zone->uz_bkt_max, 591 ("%s: zone %p overflow", __func__, zone)); 592 593 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link); 594 zdom->uzd_nitems += bucket->ub_cnt; 595 if (ws && zdom->uzd_imax < zdom->uzd_nitems) 596 zdom->uzd_imax = zdom->uzd_nitems; 597 zone->uz_bkt_count += bucket->ub_cnt; 598 } 599 600 /* Pops an item out of a per-cpu cache bucket. */ 601 static inline void * 602 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket) 603 { 604 void *item; 605 606 CRITICAL_ASSERT(curthread); 607 608 bucket->ucb_cnt--; 609 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt]; 610 #ifdef INVARIANTS 611 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL; 612 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 613 #endif 614 cache->uc_allocs++; 615 616 return (item); 617 } 618 619 /* Pushes an item into a per-cpu cache bucket. */ 620 static inline void 621 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item) 622 { 623 624 CRITICAL_ASSERT(curthread); 625 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL, 626 ("uma_zfree: Freeing to non free bucket index.")); 627 628 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item; 629 bucket->ucb_cnt++; 630 cache->uc_frees++; 631 } 632 633 /* 634 * Unload a UMA bucket from a per-cpu cache. 635 */ 636 static inline uma_bucket_t 637 cache_bucket_unload(uma_cache_bucket_t bucket) 638 { 639 uma_bucket_t b; 640 641 b = bucket->ucb_bucket; 642 if (b != NULL) { 643 MPASS(b->ub_entries == bucket->ucb_entries); 644 b->ub_cnt = bucket->ucb_cnt; 645 bucket->ucb_bucket = NULL; 646 bucket->ucb_entries = bucket->ucb_cnt = 0; 647 } 648 649 return (b); 650 } 651 652 static inline uma_bucket_t 653 cache_bucket_unload_alloc(uma_cache_t cache) 654 { 655 656 return (cache_bucket_unload(&cache->uc_allocbucket)); 657 } 658 659 static inline uma_bucket_t 660 cache_bucket_unload_free(uma_cache_t cache) 661 { 662 663 return (cache_bucket_unload(&cache->uc_freebucket)); 664 } 665 666 static inline uma_bucket_t 667 cache_bucket_unload_cross(uma_cache_t cache) 668 { 669 670 return (cache_bucket_unload(&cache->uc_crossbucket)); 671 } 672 673 /* 674 * Load a bucket into a per-cpu cache bucket. 675 */ 676 static inline void 677 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b) 678 { 679 680 CRITICAL_ASSERT(curthread); 681 MPASS(bucket->ucb_bucket == NULL); 682 MPASS(b->ub_seq == SMR_SEQ_INVALID); 683 684 bucket->ucb_bucket = b; 685 bucket->ucb_cnt = b->ub_cnt; 686 bucket->ucb_entries = b->ub_entries; 687 } 688 689 static inline void 690 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b) 691 { 692 693 cache_bucket_load(&cache->uc_allocbucket, b); 694 } 695 696 static inline void 697 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b) 698 { 699 700 cache_bucket_load(&cache->uc_freebucket, b); 701 } 702 703 #ifdef NUMA 704 static inline void 705 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b) 706 { 707 708 cache_bucket_load(&cache->uc_crossbucket, b); 709 } 710 #endif 711 712 /* 713 * Copy and preserve ucb_spare. 714 */ 715 static inline void 716 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2) 717 { 718 719 b1->ucb_bucket = b2->ucb_bucket; 720 b1->ucb_entries = b2->ucb_entries; 721 b1->ucb_cnt = b2->ucb_cnt; 722 } 723 724 /* 725 * Swap two cache buckets. 726 */ 727 static inline void 728 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2) 729 { 730 struct uma_cache_bucket b3; 731 732 CRITICAL_ASSERT(curthread); 733 734 cache_bucket_copy(&b3, b1); 735 cache_bucket_copy(b1, b2); 736 cache_bucket_copy(b2, &b3); 737 } 738 739 static void 740 zone_log_warning(uma_zone_t zone) 741 { 742 static const struct timeval warninterval = { 300, 0 }; 743 744 if (!zone_warnings || zone->uz_warning == NULL) 745 return; 746 747 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 748 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 749 } 750 751 static inline void 752 zone_maxaction(uma_zone_t zone) 753 { 754 755 if (zone->uz_maxaction.ta_func != NULL) 756 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); 757 } 758 759 /* 760 * Routine called by timeout which is used to fire off some time interval 761 * based calculations. (stats, hash size, etc.) 762 * 763 * Arguments: 764 * arg Unused 765 * 766 * Returns: 767 * Nothing 768 */ 769 static void 770 uma_timeout(void *unused) 771 { 772 bucket_enable(); 773 zone_foreach(zone_timeout, NULL); 774 775 /* Reschedule this event */ 776 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 777 } 778 779 /* 780 * Update the working set size estimate for the zone's bucket cache. 781 * The constants chosen here are somewhat arbitrary. With an update period of 782 * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the 783 * last 100s. 784 */ 785 static void 786 zone_domain_update_wss(uma_zone_domain_t zdom) 787 { 788 long wss; 789 790 MPASS(zdom->uzd_imax >= zdom->uzd_imin); 791 wss = zdom->uzd_imax - zdom->uzd_imin; 792 zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems; 793 zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5; 794 } 795 796 /* 797 * Routine to perform timeout driven calculations. This expands the 798 * hashes and does per cpu statistics aggregation. 799 * 800 * Returns nothing. 801 */ 802 static void 803 zone_timeout(uma_zone_t zone, void *unused) 804 { 805 uma_keg_t keg; 806 u_int slabs, pages; 807 808 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0) 809 goto update_wss; 810 811 keg = zone->uz_keg; 812 813 /* 814 * Hash zones are non-numa by definition so the first domain 815 * is the only one present. 816 */ 817 KEG_LOCK(keg, 0); 818 pages = keg->uk_domain[0].ud_pages; 819 820 /* 821 * Expand the keg hash table. 822 * 823 * This is done if the number of slabs is larger than the hash size. 824 * What I'm trying to do here is completely reduce collisions. This 825 * may be a little aggressive. Should I allow for two collisions max? 826 */ 827 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) { 828 struct uma_hash newhash; 829 struct uma_hash oldhash; 830 int ret; 831 832 /* 833 * This is so involved because allocating and freeing 834 * while the keg lock is held will lead to deadlock. 835 * I have to do everything in stages and check for 836 * races. 837 */ 838 KEG_UNLOCK(keg, 0); 839 ret = hash_alloc(&newhash, 1 << fls(slabs)); 840 KEG_LOCK(keg, 0); 841 if (ret) { 842 if (hash_expand(&keg->uk_hash, &newhash)) { 843 oldhash = keg->uk_hash; 844 keg->uk_hash = newhash; 845 } else 846 oldhash = newhash; 847 848 KEG_UNLOCK(keg, 0); 849 hash_free(&oldhash); 850 goto update_wss; 851 } 852 } 853 KEG_UNLOCK(keg, 0); 854 855 update_wss: 856 ZONE_LOCK(zone); 857 for (int i = 0; i < vm_ndomains; i++) 858 zone_domain_update_wss(&zone->uz_domain[i]); 859 ZONE_UNLOCK(zone); 860 } 861 862 /* 863 * Allocate and zero fill the next sized hash table from the appropriate 864 * backing store. 865 * 866 * Arguments: 867 * hash A new hash structure with the old hash size in uh_hashsize 868 * 869 * Returns: 870 * 1 on success and 0 on failure. 871 */ 872 static int 873 hash_alloc(struct uma_hash *hash, u_int size) 874 { 875 size_t alloc; 876 877 KASSERT(powerof2(size), ("hash size must be power of 2")); 878 if (size > UMA_HASH_SIZE_INIT) { 879 hash->uh_hashsize = size; 880 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 881 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT); 882 } else { 883 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 884 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 885 UMA_ANYDOMAIN, M_WAITOK); 886 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 887 } 888 if (hash->uh_slab_hash) { 889 bzero(hash->uh_slab_hash, alloc); 890 hash->uh_hashmask = hash->uh_hashsize - 1; 891 return (1); 892 } 893 894 return (0); 895 } 896 897 /* 898 * Expands the hash table for HASH zones. This is done from zone_timeout 899 * to reduce collisions. This must not be done in the regular allocation 900 * path, otherwise, we can recurse on the vm while allocating pages. 901 * 902 * Arguments: 903 * oldhash The hash you want to expand 904 * newhash The hash structure for the new table 905 * 906 * Returns: 907 * Nothing 908 * 909 * Discussion: 910 */ 911 static int 912 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 913 { 914 uma_hash_slab_t slab; 915 u_int hval; 916 u_int idx; 917 918 if (!newhash->uh_slab_hash) 919 return (0); 920 921 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 922 return (0); 923 924 /* 925 * I need to investigate hash algorithms for resizing without a 926 * full rehash. 927 */ 928 929 for (idx = 0; idx < oldhash->uh_hashsize; idx++) 930 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) { 931 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]); 932 LIST_REMOVE(slab, uhs_hlink); 933 hval = UMA_HASH(newhash, slab->uhs_data); 934 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 935 slab, uhs_hlink); 936 } 937 938 return (1); 939 } 940 941 /* 942 * Free the hash bucket to the appropriate backing store. 943 * 944 * Arguments: 945 * slab_hash The hash bucket we're freeing 946 * hashsize The number of entries in that hash bucket 947 * 948 * Returns: 949 * Nothing 950 */ 951 static void 952 hash_free(struct uma_hash *hash) 953 { 954 if (hash->uh_slab_hash == NULL) 955 return; 956 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 957 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 958 else 959 free(hash->uh_slab_hash, M_UMAHASH); 960 } 961 962 /* 963 * Frees all outstanding items in a bucket 964 * 965 * Arguments: 966 * zone The zone to free to, must be unlocked. 967 * bucket The free/alloc bucket with items. 968 * 969 * Returns: 970 * Nothing 971 */ 972 973 static void 974 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 975 { 976 int i; 977 978 if (bucket == NULL || bucket->ub_cnt == 0) 979 return; 980 981 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && 982 bucket->ub_seq != SMR_SEQ_INVALID) { 983 smr_wait(zone->uz_smr, bucket->ub_seq); 984 bucket->ub_seq = SMR_SEQ_INVALID; 985 for (i = 0; i < bucket->ub_cnt; i++) 986 item_dtor(zone, bucket->ub_bucket[i], 987 zone->uz_size, NULL, SKIP_NONE); 988 } 989 if (zone->uz_fini) 990 for (i = 0; i < bucket->ub_cnt; i++) 991 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 992 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 993 if (zone->uz_max_items > 0) 994 zone_free_limit(zone, bucket->ub_cnt); 995 #ifdef INVARIANTS 996 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt); 997 #endif 998 bucket->ub_cnt = 0; 999 } 1000 1001 /* 1002 * Drains the per cpu caches for a zone. 1003 * 1004 * NOTE: This may only be called while the zone is being torn down, and not 1005 * during normal operation. This is necessary in order that we do not have 1006 * to migrate CPUs to drain the per-CPU caches. 1007 * 1008 * Arguments: 1009 * zone The zone to drain, must be unlocked. 1010 * 1011 * Returns: 1012 * Nothing 1013 */ 1014 static void 1015 cache_drain(uma_zone_t zone) 1016 { 1017 uma_cache_t cache; 1018 uma_bucket_t bucket; 1019 smr_seq_t seq; 1020 int cpu; 1021 1022 /* 1023 * XXX: It is safe to not lock the per-CPU caches, because we're 1024 * tearing down the zone anyway. I.e., there will be no further use 1025 * of the caches at this point. 1026 * 1027 * XXX: It would good to be able to assert that the zone is being 1028 * torn down to prevent improper use of cache_drain(). 1029 */ 1030 seq = SMR_SEQ_INVALID; 1031 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 1032 seq = smr_current(zone->uz_smr); 1033 CPU_FOREACH(cpu) { 1034 cache = &zone->uz_cpu[cpu]; 1035 bucket = cache_bucket_unload_alloc(cache); 1036 if (bucket != NULL) { 1037 bucket_drain(zone, bucket); 1038 bucket_free(zone, bucket, NULL); 1039 } 1040 bucket = cache_bucket_unload_free(cache); 1041 if (bucket != NULL) { 1042 bucket->ub_seq = seq; 1043 bucket_drain(zone, bucket); 1044 bucket_free(zone, bucket, NULL); 1045 } 1046 bucket = cache_bucket_unload_cross(cache); 1047 if (bucket != NULL) { 1048 bucket->ub_seq = seq; 1049 bucket_drain(zone, bucket); 1050 bucket_free(zone, bucket, NULL); 1051 } 1052 } 1053 bucket_cache_reclaim(zone, true); 1054 } 1055 1056 static void 1057 cache_shrink(uma_zone_t zone, void *unused) 1058 { 1059 1060 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 1061 return; 1062 1063 ZONE_LOCK(zone); 1064 zone->uz_bucket_size = 1065 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2; 1066 ZONE_UNLOCK(zone); 1067 } 1068 1069 static void 1070 cache_drain_safe_cpu(uma_zone_t zone, void *unused) 1071 { 1072 uma_cache_t cache; 1073 uma_bucket_t b1, b2, b3; 1074 int domain; 1075 1076 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 1077 return; 1078 1079 b1 = b2 = b3 = NULL; 1080 critical_enter(); 1081 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) 1082 domain = PCPU_GET(domain); 1083 else 1084 domain = 0; 1085 cache = &zone->uz_cpu[curcpu]; 1086 b1 = cache_bucket_unload_alloc(cache); 1087 1088 /* 1089 * Don't flush SMR zone buckets. This leaves the zone without a 1090 * bucket and forces every free to synchronize(). 1091 */ 1092 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) { 1093 b2 = cache_bucket_unload_free(cache); 1094 b3 = cache_bucket_unload_cross(cache); 1095 } 1096 critical_exit(); 1097 1098 ZONE_LOCK(zone); 1099 if (b1 != NULL && b1->ub_cnt != 0) { 1100 zone_put_bucket(zone, &zone->uz_domain[domain], b1, false); 1101 b1 = NULL; 1102 } 1103 if (b2 != NULL && b2->ub_cnt != 0) { 1104 zone_put_bucket(zone, &zone->uz_domain[domain], b2, false); 1105 b2 = NULL; 1106 } 1107 ZONE_UNLOCK(zone); 1108 1109 if (b1 != NULL) 1110 bucket_free(zone, b1, NULL); 1111 if (b2 != NULL) 1112 bucket_free(zone, b2, NULL); 1113 if (b3 != NULL) { 1114 bucket_drain(zone, b3); 1115 bucket_free(zone, b3, NULL); 1116 } 1117 } 1118 1119 /* 1120 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 1121 * This is an expensive call because it needs to bind to all CPUs 1122 * one by one and enter a critical section on each of them in order 1123 * to safely access their cache buckets. 1124 * Zone lock must not be held on call this function. 1125 */ 1126 static void 1127 pcpu_cache_drain_safe(uma_zone_t zone) 1128 { 1129 int cpu; 1130 1131 /* 1132 * Polite bucket sizes shrinking was not enough, shrink aggressively. 1133 */ 1134 if (zone) 1135 cache_shrink(zone, NULL); 1136 else 1137 zone_foreach(cache_shrink, NULL); 1138 1139 CPU_FOREACH(cpu) { 1140 thread_lock(curthread); 1141 sched_bind(curthread, cpu); 1142 thread_unlock(curthread); 1143 1144 if (zone) 1145 cache_drain_safe_cpu(zone, NULL); 1146 else 1147 zone_foreach(cache_drain_safe_cpu, NULL); 1148 } 1149 thread_lock(curthread); 1150 sched_unbind(curthread); 1151 thread_unlock(curthread); 1152 } 1153 1154 /* 1155 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller 1156 * requested a drain, otherwise the per-domain caches are trimmed to either 1157 * estimated working set size. 1158 */ 1159 static void 1160 bucket_cache_reclaim(uma_zone_t zone, bool drain) 1161 { 1162 uma_zone_domain_t zdom; 1163 uma_bucket_t bucket; 1164 long target, tofree; 1165 int i; 1166 1167 for (i = 0; i < vm_ndomains; i++) { 1168 /* 1169 * The cross bucket is partially filled and not part of 1170 * the item count. Reclaim it individually here. 1171 */ 1172 zdom = &zone->uz_domain[i]; 1173 ZONE_CROSS_LOCK(zone); 1174 bucket = zdom->uzd_cross; 1175 zdom->uzd_cross = NULL; 1176 ZONE_CROSS_UNLOCK(zone); 1177 if (bucket != NULL) { 1178 bucket_drain(zone, bucket); 1179 bucket_free(zone, bucket, NULL); 1180 } 1181 1182 /* 1183 * Shrink the zone bucket size to ensure that the per-CPU caches 1184 * don't grow too large. 1185 */ 1186 ZONE_LOCK(zone); 1187 if (i == 0 && zone->uz_bucket_size > zone->uz_bucket_size_min) 1188 zone->uz_bucket_size--; 1189 1190 /* 1191 * If we were asked to drain the zone, we are done only once 1192 * this bucket cache is empty. Otherwise, we reclaim items in 1193 * excess of the zone's estimated working set size. If the 1194 * difference nitems - imin is larger than the WSS estimate, 1195 * then the estimate will grow at the end of this interval and 1196 * we ignore the historical average. 1197 */ 1198 target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems - 1199 zdom->uzd_imin); 1200 while (zdom->uzd_nitems > target) { 1201 bucket = STAILQ_FIRST(&zdom->uzd_buckets); 1202 if (bucket == NULL) 1203 break; 1204 tofree = bucket->ub_cnt; 1205 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link); 1206 zdom->uzd_nitems -= tofree; 1207 zone->uz_bkt_count -= tofree; 1208 1209 /* 1210 * Shift the bounds of the current WSS interval to avoid 1211 * perturbing the estimate. 1212 */ 1213 zdom->uzd_imax -= lmin(zdom->uzd_imax, tofree); 1214 zdom->uzd_imin -= lmin(zdom->uzd_imin, tofree); 1215 1216 ZONE_UNLOCK(zone); 1217 bucket_drain(zone, bucket); 1218 bucket_free(zone, bucket, NULL); 1219 ZONE_LOCK(zone); 1220 } 1221 ZONE_UNLOCK(zone); 1222 } 1223 } 1224 1225 static void 1226 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 1227 { 1228 uint8_t *mem; 1229 int i; 1230 uint8_t flags; 1231 1232 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes", 1233 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera); 1234 1235 mem = slab_data(slab, keg); 1236 flags = slab->us_flags; 1237 i = start; 1238 if (keg->uk_fini != NULL) { 1239 for (i--; i > -1; i--) 1240 #ifdef INVARIANTS 1241 /* 1242 * trash_fini implies that dtor was trash_dtor. trash_fini 1243 * would check that memory hasn't been modified since free, 1244 * which executed trash_dtor. 1245 * That's why we need to run uma_dbg_kskip() check here, 1246 * albeit we don't make skip check for other init/fini 1247 * invocations. 1248 */ 1249 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) || 1250 keg->uk_fini != trash_fini) 1251 #endif 1252 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size); 1253 } 1254 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) 1255 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab), 1256 NULL, SKIP_NONE); 1257 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 1258 uma_total_dec(PAGE_SIZE * keg->uk_ppera); 1259 } 1260 1261 /* 1262 * Frees pages from a keg back to the system. This is done on demand from 1263 * the pageout daemon. 1264 * 1265 * Returns nothing. 1266 */ 1267 static void 1268 keg_drain(uma_keg_t keg) 1269 { 1270 struct slabhead freeslabs; 1271 uma_domain_t dom; 1272 uma_slab_t slab, tmp; 1273 int i, n; 1274 1275 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 1276 return; 1277 1278 for (i = 0; i < vm_ndomains; i++) { 1279 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u", 1280 keg->uk_name, keg, i, dom->ud_free_items); 1281 dom = &keg->uk_domain[i]; 1282 LIST_INIT(&freeslabs); 1283 1284 KEG_LOCK(keg, i); 1285 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) { 1286 LIST_FOREACH(slab, &dom->ud_free_slab, us_link) 1287 UMA_HASH_REMOVE(&keg->uk_hash, slab); 1288 } 1289 n = dom->ud_free_slabs; 1290 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link); 1291 dom->ud_free_slabs = 0; 1292 dom->ud_free_items -= n * keg->uk_ipers; 1293 dom->ud_pages -= n * keg->uk_ppera; 1294 KEG_UNLOCK(keg, i); 1295 1296 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp) 1297 keg_free_slab(keg, slab, keg->uk_ipers); 1298 } 1299 } 1300 1301 static void 1302 zone_reclaim(uma_zone_t zone, int waitok, bool drain) 1303 { 1304 1305 /* 1306 * Set draining to interlock with zone_dtor() so we can release our 1307 * locks as we go. Only dtor() should do a WAITOK call since it 1308 * is the only call that knows the structure will still be available 1309 * when it wakes up. 1310 */ 1311 ZONE_LOCK(zone); 1312 while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) { 1313 if (waitok == M_NOWAIT) 1314 goto out; 1315 msleep(zone, &zone->uz_lock, PVM, "zonedrain", 1); 1316 } 1317 zone->uz_flags |= UMA_ZFLAG_RECLAIMING; 1318 ZONE_UNLOCK(zone); 1319 bucket_cache_reclaim(zone, drain); 1320 1321 /* 1322 * The DRAINING flag protects us from being freed while 1323 * we're running. Normally the uma_rwlock would protect us but we 1324 * must be able to release and acquire the right lock for each keg. 1325 */ 1326 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) 1327 keg_drain(zone->uz_keg); 1328 ZONE_LOCK(zone); 1329 zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING; 1330 wakeup(zone); 1331 out: 1332 ZONE_UNLOCK(zone); 1333 } 1334 1335 static void 1336 zone_drain(uma_zone_t zone, void *unused) 1337 { 1338 1339 zone_reclaim(zone, M_NOWAIT, true); 1340 } 1341 1342 static void 1343 zone_trim(uma_zone_t zone, void *unused) 1344 { 1345 1346 zone_reclaim(zone, M_NOWAIT, false); 1347 } 1348 1349 /* 1350 * Allocate a new slab for a keg and inserts it into the partial slab list. 1351 * The keg should be unlocked on entry. If the allocation succeeds it will 1352 * be locked on return. 1353 * 1354 * Arguments: 1355 * flags Wait flags for the item initialization routine 1356 * aflags Wait flags for the slab allocation 1357 * 1358 * Returns: 1359 * The slab that was allocated or NULL if there is no memory and the 1360 * caller specified M_NOWAIT. 1361 */ 1362 static uma_slab_t 1363 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags, 1364 int aflags) 1365 { 1366 uma_domain_t dom; 1367 uma_alloc allocf; 1368 uma_slab_t slab; 1369 unsigned long size; 1370 uint8_t *mem; 1371 uint8_t sflags; 1372 int i; 1373 1374 KASSERT(domain >= 0 && domain < vm_ndomains, 1375 ("keg_alloc_slab: domain %d out of range", domain)); 1376 1377 allocf = keg->uk_allocf; 1378 slab = NULL; 1379 mem = NULL; 1380 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) { 1381 uma_hash_slab_t hslab; 1382 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL, 1383 domain, aflags); 1384 if (hslab == NULL) 1385 goto fail; 1386 slab = &hslab->uhs_slab; 1387 } 1388 1389 /* 1390 * This reproduces the old vm_zone behavior of zero filling pages the 1391 * first time they are added to a zone. 1392 * 1393 * Malloced items are zeroed in uma_zalloc. 1394 */ 1395 1396 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 1397 aflags |= M_ZERO; 1398 else 1399 aflags &= ~M_ZERO; 1400 1401 if (keg->uk_flags & UMA_ZONE_NODUMP) 1402 aflags |= M_NODUMP; 1403 1404 /* zone is passed for legacy reasons. */ 1405 size = keg->uk_ppera * PAGE_SIZE; 1406 mem = allocf(zone, size, domain, &sflags, aflags); 1407 if (mem == NULL) { 1408 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) 1409 zone_free_item(slabzone(keg->uk_ipers), 1410 slab_tohashslab(slab), NULL, SKIP_NONE); 1411 goto fail; 1412 } 1413 uma_total_inc(size); 1414 1415 /* For HASH zones all pages go to the same uma_domain. */ 1416 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) 1417 domain = 0; 1418 1419 /* Point the slab into the allocated memory */ 1420 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) 1421 slab = (uma_slab_t )(mem + keg->uk_pgoff); 1422 else 1423 slab_tohashslab(slab)->uhs_data = mem; 1424 1425 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB) 1426 for (i = 0; i < keg->uk_ppera; i++) 1427 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE), 1428 zone, slab); 1429 1430 slab->us_freecount = keg->uk_ipers; 1431 slab->us_flags = sflags; 1432 slab->us_domain = domain; 1433 1434 BIT_FILL(keg->uk_ipers, &slab->us_free); 1435 #ifdef INVARIANTS 1436 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg)); 1437 #endif 1438 1439 if (keg->uk_init != NULL) { 1440 for (i = 0; i < keg->uk_ipers; i++) 1441 if (keg->uk_init(slab_item(slab, keg, i), 1442 keg->uk_size, flags) != 0) 1443 break; 1444 if (i != keg->uk_ipers) { 1445 keg_free_slab(keg, slab, i); 1446 goto fail; 1447 } 1448 } 1449 KEG_LOCK(keg, domain); 1450 1451 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)", 1452 slab, keg->uk_name, keg); 1453 1454 if (keg->uk_flags & UMA_ZFLAG_HASH) 1455 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1456 1457 /* 1458 * If we got a slab here it's safe to mark it partially used 1459 * and return. We assume that the caller is going to remove 1460 * at least one item. 1461 */ 1462 dom = &keg->uk_domain[domain]; 1463 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 1464 dom->ud_pages += keg->uk_ppera; 1465 dom->ud_free_items += keg->uk_ipers; 1466 1467 return (slab); 1468 1469 fail: 1470 return (NULL); 1471 } 1472 1473 /* 1474 * This function is intended to be used early on in place of page_alloc() so 1475 * that we may use the boot time page cache to satisfy allocations before 1476 * the VM is ready. 1477 */ 1478 static void * 1479 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1480 int wait) 1481 { 1482 vm_paddr_t pa; 1483 vm_page_t m; 1484 void *mem; 1485 int pages; 1486 int i; 1487 1488 pages = howmany(bytes, PAGE_SIZE); 1489 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__)); 1490 1491 *pflag = UMA_SLAB_BOOT; 1492 m = vm_page_alloc_contig_domain(NULL, 0, domain, 1493 malloc2vm_flags(wait) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED, pages, 1494 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT); 1495 if (m == NULL) 1496 return (NULL); 1497 1498 pa = VM_PAGE_TO_PHYS(m); 1499 for (i = 0; i < pages; i++, pa += PAGE_SIZE) { 1500 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 1501 defined(__riscv) || defined(__powerpc64__) 1502 if ((wait & M_NODUMP) == 0) 1503 dump_add_page(pa); 1504 #endif 1505 } 1506 /* Allocate KVA and indirectly advance bootmem. */ 1507 mem = (void *)pmap_map(&bootmem, m->phys_addr, 1508 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE); 1509 if ((wait & M_ZERO) != 0) 1510 bzero(mem, pages * PAGE_SIZE); 1511 1512 return (mem); 1513 } 1514 1515 static void 1516 startup_free(void *mem, vm_size_t bytes) 1517 { 1518 vm_offset_t va; 1519 vm_page_t m; 1520 1521 va = (vm_offset_t)mem; 1522 m = PHYS_TO_VM_PAGE(pmap_kextract(va)); 1523 pmap_remove(kernel_pmap, va, va + bytes); 1524 for (; bytes != 0; bytes -= PAGE_SIZE, m++) { 1525 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ 1526 defined(__riscv) || defined(__powerpc64__) 1527 dump_drop_page(VM_PAGE_TO_PHYS(m)); 1528 #endif 1529 vm_page_unwire_noq(m); 1530 vm_page_free(m); 1531 } 1532 } 1533 1534 /* 1535 * Allocates a number of pages from the system 1536 * 1537 * Arguments: 1538 * bytes The number of bytes requested 1539 * wait Shall we wait? 1540 * 1541 * Returns: 1542 * A pointer to the alloced memory or possibly 1543 * NULL if M_NOWAIT is set. 1544 */ 1545 static void * 1546 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1547 int wait) 1548 { 1549 void *p; /* Returned page */ 1550 1551 *pflag = UMA_SLAB_KERNEL; 1552 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait); 1553 1554 return (p); 1555 } 1556 1557 static void * 1558 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1559 int wait) 1560 { 1561 struct pglist alloctail; 1562 vm_offset_t addr, zkva; 1563 int cpu, flags; 1564 vm_page_t p, p_next; 1565 #ifdef NUMA 1566 struct pcpu *pc; 1567 #endif 1568 1569 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE); 1570 1571 TAILQ_INIT(&alloctail); 1572 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1573 malloc2vm_flags(wait); 1574 *pflag = UMA_SLAB_KERNEL; 1575 for (cpu = 0; cpu <= mp_maxid; cpu++) { 1576 if (CPU_ABSENT(cpu)) { 1577 p = vm_page_alloc(NULL, 0, flags); 1578 } else { 1579 #ifndef NUMA 1580 p = vm_page_alloc(NULL, 0, flags); 1581 #else 1582 pc = pcpu_find(cpu); 1583 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain))) 1584 p = NULL; 1585 else 1586 p = vm_page_alloc_domain(NULL, 0, 1587 pc->pc_domain, flags); 1588 if (__predict_false(p == NULL)) 1589 p = vm_page_alloc(NULL, 0, flags); 1590 #endif 1591 } 1592 if (__predict_false(p == NULL)) 1593 goto fail; 1594 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1595 } 1596 if ((addr = kva_alloc(bytes)) == 0) 1597 goto fail; 1598 zkva = addr; 1599 TAILQ_FOREACH(p, &alloctail, listq) { 1600 pmap_qenter(zkva, &p, 1); 1601 zkva += PAGE_SIZE; 1602 } 1603 return ((void*)addr); 1604 fail: 1605 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1606 vm_page_unwire_noq(p); 1607 vm_page_free(p); 1608 } 1609 return (NULL); 1610 } 1611 1612 /* 1613 * Allocates a number of pages from within an object 1614 * 1615 * Arguments: 1616 * bytes The number of bytes requested 1617 * wait Shall we wait? 1618 * 1619 * Returns: 1620 * A pointer to the alloced memory or possibly 1621 * NULL if M_NOWAIT is set. 1622 */ 1623 static void * 1624 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, 1625 int wait) 1626 { 1627 TAILQ_HEAD(, vm_page) alloctail; 1628 u_long npages; 1629 vm_offset_t retkva, zkva; 1630 vm_page_t p, p_next; 1631 uma_keg_t keg; 1632 1633 TAILQ_INIT(&alloctail); 1634 keg = zone->uz_keg; 1635 1636 npages = howmany(bytes, PAGE_SIZE); 1637 while (npages > 0) { 1638 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT | 1639 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1640 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK : 1641 VM_ALLOC_NOWAIT)); 1642 if (p != NULL) { 1643 /* 1644 * Since the page does not belong to an object, its 1645 * listq is unused. 1646 */ 1647 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1648 npages--; 1649 continue; 1650 } 1651 /* 1652 * Page allocation failed, free intermediate pages and 1653 * exit. 1654 */ 1655 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1656 vm_page_unwire_noq(p); 1657 vm_page_free(p); 1658 } 1659 return (NULL); 1660 } 1661 *flags = UMA_SLAB_PRIV; 1662 zkva = keg->uk_kva + 1663 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1664 retkva = zkva; 1665 TAILQ_FOREACH(p, &alloctail, listq) { 1666 pmap_qenter(zkva, &p, 1); 1667 zkva += PAGE_SIZE; 1668 } 1669 1670 return ((void *)retkva); 1671 } 1672 1673 /* 1674 * Allocate physically contiguous pages. 1675 */ 1676 static void * 1677 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1678 int wait) 1679 { 1680 1681 *pflag = UMA_SLAB_KERNEL; 1682 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain), 1683 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT)); 1684 } 1685 1686 /* 1687 * Frees a number of pages to the system 1688 * 1689 * Arguments: 1690 * mem A pointer to the memory to be freed 1691 * size The size of the memory being freed 1692 * flags The original p->us_flags field 1693 * 1694 * Returns: 1695 * Nothing 1696 */ 1697 static void 1698 page_free(void *mem, vm_size_t size, uint8_t flags) 1699 { 1700 1701 if ((flags & UMA_SLAB_BOOT) != 0) { 1702 startup_free(mem, size); 1703 return; 1704 } 1705 1706 KASSERT((flags & UMA_SLAB_KERNEL) != 0, 1707 ("UMA: page_free used with invalid flags %x", flags)); 1708 1709 kmem_free((vm_offset_t)mem, size); 1710 } 1711 1712 /* 1713 * Frees pcpu zone allocations 1714 * 1715 * Arguments: 1716 * mem A pointer to the memory to be freed 1717 * size The size of the memory being freed 1718 * flags The original p->us_flags field 1719 * 1720 * Returns: 1721 * Nothing 1722 */ 1723 static void 1724 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags) 1725 { 1726 vm_offset_t sva, curva; 1727 vm_paddr_t paddr; 1728 vm_page_t m; 1729 1730 MPASS(size == (mp_maxid+1)*PAGE_SIZE); 1731 1732 if ((flags & UMA_SLAB_BOOT) != 0) { 1733 startup_free(mem, size); 1734 return; 1735 } 1736 1737 sva = (vm_offset_t)mem; 1738 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) { 1739 paddr = pmap_kextract(curva); 1740 m = PHYS_TO_VM_PAGE(paddr); 1741 vm_page_unwire_noq(m); 1742 vm_page_free(m); 1743 } 1744 pmap_qremove(sva, size >> PAGE_SHIFT); 1745 kva_free(sva, size); 1746 } 1747 1748 1749 /* 1750 * Zero fill initializer 1751 * 1752 * Arguments/Returns follow uma_init specifications 1753 */ 1754 static int 1755 zero_init(void *mem, int size, int flags) 1756 { 1757 bzero(mem, size); 1758 return (0); 1759 } 1760 1761 #ifdef INVARIANTS 1762 struct noslabbits * 1763 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg) 1764 { 1765 1766 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers))); 1767 } 1768 #endif 1769 1770 /* 1771 * Actual size of embedded struct slab (!OFFPAGE). 1772 */ 1773 size_t 1774 slab_sizeof(int nitems) 1775 { 1776 size_t s; 1777 1778 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS; 1779 return (roundup(s, UMA_ALIGN_PTR + 1)); 1780 } 1781 1782 /* 1783 * Size of memory for embedded slabs (!OFFPAGE). 1784 */ 1785 size_t 1786 slab_space(int nitems) 1787 { 1788 return (UMA_SLAB_SIZE - slab_sizeof(nitems)); 1789 } 1790 1791 #define UMA_FIXPT_SHIFT 31 1792 #define UMA_FRAC_FIXPT(n, d) \ 1793 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d))) 1794 #define UMA_FIXPT_PCT(f) \ 1795 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT)) 1796 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100) 1797 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE) 1798 1799 /* 1800 * Compute the number of items that will fit in a slab. If hdr is true, the 1801 * item count may be limited to provide space in the slab for an inline slab 1802 * header. Otherwise, all slab space will be provided for item storage. 1803 */ 1804 static u_int 1805 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr) 1806 { 1807 u_int ipers; 1808 u_int padpi; 1809 1810 /* The padding between items is not needed after the last item. */ 1811 padpi = rsize - size; 1812 1813 if (hdr) { 1814 /* 1815 * Start with the maximum item count and remove items until 1816 * the slab header first alongside the allocatable memory. 1817 */ 1818 for (ipers = MIN(SLAB_MAX_SETSIZE, 1819 (slabsize + padpi - slab_sizeof(1)) / rsize); 1820 ipers > 0 && 1821 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize; 1822 ipers--) 1823 continue; 1824 } else { 1825 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE); 1826 } 1827 1828 return (ipers); 1829 } 1830 1831 /* 1832 * Compute the number of items that will fit in a slab for a startup zone. 1833 */ 1834 int 1835 slab_ipers(size_t size, int align) 1836 { 1837 int rsize; 1838 1839 rsize = roundup(size, align + 1); /* Assume no CACHESPREAD */ 1840 return (slab_ipers_hdr(size, rsize, UMA_SLAB_SIZE, true)); 1841 } 1842 1843 struct keg_layout_result { 1844 u_int format; 1845 u_int slabsize; 1846 u_int ipers; 1847 u_int eff; 1848 }; 1849 1850 static void 1851 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt, 1852 struct keg_layout_result *kl) 1853 { 1854 u_int total; 1855 1856 kl->format = fmt; 1857 kl->slabsize = slabsize; 1858 1859 /* Handle INTERNAL as inline with an extra page. */ 1860 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) { 1861 kl->format &= ~UMA_ZFLAG_INTERNAL; 1862 kl->slabsize += PAGE_SIZE; 1863 } 1864 1865 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize, 1866 (fmt & UMA_ZFLAG_OFFPAGE) == 0); 1867 1868 /* Account for memory used by an offpage slab header. */ 1869 total = kl->slabsize; 1870 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0) 1871 total += slabzone(kl->ipers)->uz_keg->uk_rsize; 1872 1873 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total); 1874 } 1875 1876 /* 1877 * Determine the format of a uma keg. This determines where the slab header 1878 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera. 1879 * 1880 * Arguments 1881 * keg The zone we should initialize 1882 * 1883 * Returns 1884 * Nothing 1885 */ 1886 static void 1887 keg_layout(uma_keg_t keg) 1888 { 1889 struct keg_layout_result kl = {}, kl_tmp; 1890 u_int fmts[2]; 1891 u_int alignsize; 1892 u_int nfmt; 1893 u_int pages; 1894 u_int rsize; 1895 u_int slabsize; 1896 u_int i, j; 1897 1898 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1899 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE && 1900 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0), 1901 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b", 1902 __func__, keg->uk_name, keg->uk_size, keg->uk_flags, 1903 PRINT_UMA_ZFLAGS)); 1904 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 || 1905 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0, 1906 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags, 1907 PRINT_UMA_ZFLAGS)); 1908 1909 alignsize = keg->uk_align + 1; 1910 1911 /* 1912 * Calculate the size of each allocation (rsize) according to 1913 * alignment. If the requested size is smaller than we have 1914 * allocation bits for we round it up. 1915 */ 1916 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT); 1917 rsize = roundup2(rsize, alignsize); 1918 1919 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) { 1920 /* 1921 * We want one item to start on every align boundary in a page. 1922 * To do this we will span pages. We will also extend the item 1923 * by the size of align if it is an even multiple of align. 1924 * Otherwise, it would fall on the same boundary every time. 1925 */ 1926 if ((rsize & alignsize) == 0) 1927 rsize += alignsize; 1928 slabsize = rsize * (PAGE_SIZE / alignsize); 1929 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE); 1930 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE); 1931 slabsize = round_page(slabsize); 1932 } else { 1933 /* 1934 * Start with a slab size of as many pages as it takes to 1935 * represent a single item. We will try to fit as many 1936 * additional items into the slab as possible. 1937 */ 1938 slabsize = round_page(keg->uk_size); 1939 } 1940 1941 /* Build a list of all of the available formats for this keg. */ 1942 nfmt = 0; 1943 1944 /* Evaluate an inline slab layout. */ 1945 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0) 1946 fmts[nfmt++] = 0; 1947 1948 /* TODO: vm_page-embedded slab. */ 1949 1950 /* 1951 * We can't do OFFPAGE if we're internal or if we've been 1952 * asked to not go to the VM for buckets. If we do this we 1953 * may end up going to the VM for slabs which we do not want 1954 * to do if we're UMA_ZONE_VM, which clearly forbids it. 1955 * In those cases, evaluate a pseudo-format called INTERNAL 1956 * which has an inline slab header and one extra page to 1957 * guarantee that it fits. 1958 * 1959 * Otherwise, see if using an OFFPAGE slab will improve our 1960 * efficiency. 1961 */ 1962 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0) 1963 fmts[nfmt++] = UMA_ZFLAG_INTERNAL; 1964 else 1965 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE; 1966 1967 /* 1968 * Choose a slab size and format which satisfy the minimum efficiency. 1969 * Prefer the smallest slab size that meets the constraints. 1970 * 1971 * Start with a minimum slab size, to accommodate CACHESPREAD. Then, 1972 * for small items (up to PAGE_SIZE), the iteration increment is one 1973 * page; and for large items, the increment is one item. 1974 */ 1975 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize); 1976 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u", 1977 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize, 1978 rsize, i)); 1979 for ( ; ; i++) { 1980 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) : 1981 round_page(rsize * (i - 1) + keg->uk_size); 1982 1983 for (j = 0; j < nfmt; j++) { 1984 /* Only if we have no viable format yet. */ 1985 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 && 1986 kl.ipers > 0) 1987 continue; 1988 1989 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp); 1990 if (kl_tmp.eff <= kl.eff) 1991 continue; 1992 1993 kl = kl_tmp; 1994 1995 CTR6(KTR_UMA, "keg %s layout: format %#x " 1996 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff", 1997 keg->uk_name, kl.format, kl.ipers, rsize, 1998 kl.slabsize, UMA_FIXPT_PCT(kl.eff)); 1999 2000 /* Stop when we reach the minimum efficiency. */ 2001 if (kl.eff >= UMA_MIN_EFF) 2002 break; 2003 } 2004 2005 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs || 2006 slabsize >= SLAB_MAX_SETSIZE * rsize || 2007 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0) 2008 break; 2009 } 2010 2011 pages = atop(kl.slabsize); 2012 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) 2013 pages *= mp_maxid + 1; 2014 2015 keg->uk_rsize = rsize; 2016 keg->uk_ipers = kl.ipers; 2017 keg->uk_ppera = pages; 2018 keg->uk_flags |= kl.format; 2019 2020 /* 2021 * How do we find the slab header if it is offpage or if not all item 2022 * start addresses are in the same page? We could solve the latter 2023 * case with vaddr alignment, but we don't. 2024 */ 2025 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 || 2026 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) { 2027 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0) 2028 keg->uk_flags |= UMA_ZFLAG_HASH; 2029 else 2030 keg->uk_flags |= UMA_ZFLAG_VTOSLAB; 2031 } 2032 2033 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u", 2034 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers, 2035 pages); 2036 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE, 2037 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__, 2038 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize, 2039 keg->uk_ipers, pages)); 2040 } 2041 2042 /* 2043 * Keg header ctor. This initializes all fields, locks, etc. And inserts 2044 * the keg onto the global keg list. 2045 * 2046 * Arguments/Returns follow uma_ctor specifications 2047 * udata Actually uma_kctor_args 2048 */ 2049 static int 2050 keg_ctor(void *mem, int size, void *udata, int flags) 2051 { 2052 struct uma_kctor_args *arg = udata; 2053 uma_keg_t keg = mem; 2054 uma_zone_t zone; 2055 int i; 2056 2057 bzero(keg, size); 2058 keg->uk_size = arg->size; 2059 keg->uk_init = arg->uminit; 2060 keg->uk_fini = arg->fini; 2061 keg->uk_align = arg->align; 2062 keg->uk_reserve = 0; 2063 keg->uk_flags = arg->flags; 2064 2065 /* 2066 * We use a global round-robin policy by default. Zones with 2067 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which 2068 * case the iterator is never run. 2069 */ 2070 keg->uk_dr.dr_policy = DOMAINSET_RR(); 2071 keg->uk_dr.dr_iter = 0; 2072 2073 /* 2074 * The master zone is passed to us at keg-creation time. 2075 */ 2076 zone = arg->zone; 2077 keg->uk_name = zone->uz_name; 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_items", CTLFLAG_RD, &dom->ud_free_items, 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 zone->uz_flags = arg->flags; 2465 zone->uz_size = arg->size; 2466 zone->uz_import = arg->import; 2467 zone->uz_release = arg->release; 2468 zone->uz_arg = arg->arg; 2469 rw_wlock(&uma_rwlock); 2470 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 2471 rw_wunlock(&uma_rwlock); 2472 goto out; 2473 } 2474 2475 /* 2476 * Use the regular zone/keg/slab allocator. 2477 */ 2478 zone->uz_import = zone_import; 2479 zone->uz_release = zone_release; 2480 zone->uz_arg = zone; 2481 keg = arg->keg; 2482 2483 if (arg->flags & UMA_ZONE_SECONDARY) { 2484 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 2485 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 2486 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 2487 zone->uz_init = arg->uminit; 2488 zone->uz_fini = arg->fini; 2489 zone->uz_flags |= UMA_ZONE_SECONDARY; 2490 rw_wlock(&uma_rwlock); 2491 ZONE_LOCK(zone); 2492 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 2493 if (LIST_NEXT(z, uz_link) == NULL) { 2494 LIST_INSERT_AFTER(z, zone, uz_link); 2495 break; 2496 } 2497 } 2498 ZONE_UNLOCK(zone); 2499 rw_wunlock(&uma_rwlock); 2500 } else if (keg == NULL) { 2501 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 2502 arg->align, arg->flags)) == NULL) 2503 return (ENOMEM); 2504 } else { 2505 struct uma_kctor_args karg; 2506 int error; 2507 2508 /* We should only be here from uma_startup() */ 2509 karg.size = arg->size; 2510 karg.uminit = arg->uminit; 2511 karg.fini = arg->fini; 2512 karg.align = arg->align; 2513 karg.flags = (arg->flags & ~UMA_ZONE_SMR); 2514 karg.zone = zone; 2515 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 2516 flags); 2517 if (error) 2518 return (error); 2519 } 2520 2521 /* Inherit properties from the keg. */ 2522 zone->uz_keg = keg; 2523 zone->uz_size = keg->uk_size; 2524 zone->uz_flags |= (keg->uk_flags & 2525 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 2526 2527 out: 2528 if (__predict_true(booted >= BOOT_RUNNING)) { 2529 zone_alloc_counters(zone, NULL); 2530 zone_alloc_sysctl(zone, NULL); 2531 } else { 2532 zone->uz_allocs = EARLY_COUNTER; 2533 zone->uz_frees = EARLY_COUNTER; 2534 zone->uz_fails = EARLY_COUNTER; 2535 } 2536 2537 /* Caller requests a private SMR context. */ 2538 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 2539 zone->uz_smr = smr_create(zone->uz_name); 2540 2541 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) != 2542 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET), 2543 ("Invalid zone flag combination")); 2544 if (arg->flags & UMA_ZFLAG_INTERNAL) 2545 zone->uz_bucket_size_max = zone->uz_bucket_size = 0; 2546 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0) 2547 zone->uz_bucket_size = BUCKET_MAX; 2548 else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0) 2549 zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN; 2550 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0) 2551 zone->uz_bucket_size = 0; 2552 else 2553 zone->uz_bucket_size = bucket_select(zone->uz_size); 2554 zone->uz_bucket_size_min = zone->uz_bucket_size; 2555 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL) 2556 zone->uz_flags |= UMA_ZFLAG_CTORDTOR; 2557 zone_update_caches(zone); 2558 2559 return (0); 2560 } 2561 2562 /* 2563 * Keg header dtor. This frees all data, destroys locks, frees the hash 2564 * table and removes the keg from the global list. 2565 * 2566 * Arguments/Returns follow uma_dtor specifications 2567 * udata unused 2568 */ 2569 static void 2570 keg_dtor(void *arg, int size, void *udata) 2571 { 2572 uma_keg_t keg; 2573 uint32_t free, pages; 2574 int i; 2575 2576 keg = (uma_keg_t)arg; 2577 free = pages = 0; 2578 for (i = 0; i < vm_ndomains; i++) { 2579 free += keg->uk_domain[i].ud_free_items; 2580 pages += keg->uk_domain[i].ud_pages; 2581 KEG_LOCK_FINI(keg, i); 2582 } 2583 if (pages != 0) 2584 printf("Freed UMA keg (%s) was not empty (%u items). " 2585 " Lost %u pages of memory.\n", 2586 keg->uk_name ? keg->uk_name : "", 2587 pages / keg->uk_ppera * keg->uk_ipers - free, pages); 2588 2589 hash_free(&keg->uk_hash); 2590 } 2591 2592 /* 2593 * Zone header dtor. 2594 * 2595 * Arguments/Returns follow uma_dtor specifications 2596 * udata unused 2597 */ 2598 static void 2599 zone_dtor(void *arg, int size, void *udata) 2600 { 2601 uma_zone_t zone; 2602 uma_keg_t keg; 2603 2604 zone = (uma_zone_t)arg; 2605 2606 sysctl_remove_oid(zone->uz_oid, 1, 1); 2607 2608 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 2609 cache_drain(zone); 2610 2611 rw_wlock(&uma_rwlock); 2612 LIST_REMOVE(zone, uz_link); 2613 rw_wunlock(&uma_rwlock); 2614 /* 2615 * XXX there are some races here where 2616 * the zone can be drained but zone lock 2617 * released and then refilled before we 2618 * remove it... we dont care for now 2619 */ 2620 zone_reclaim(zone, M_WAITOK, true); 2621 /* 2622 * We only destroy kegs from non secondary/non cache zones. 2623 */ 2624 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) { 2625 keg = zone->uz_keg; 2626 rw_wlock(&uma_rwlock); 2627 LIST_REMOVE(keg, uk_link); 2628 rw_wunlock(&uma_rwlock); 2629 zone_free_item(kegs, keg, NULL, SKIP_NONE); 2630 } 2631 counter_u64_free(zone->uz_allocs); 2632 counter_u64_free(zone->uz_frees); 2633 counter_u64_free(zone->uz_fails); 2634 free(zone->uz_ctlname, M_UMA); 2635 ZONE_LOCK_FINI(zone); 2636 ZONE_CROSS_LOCK_FINI(zone); 2637 } 2638 2639 static void 2640 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg) 2641 { 2642 uma_keg_t keg; 2643 uma_zone_t zone; 2644 2645 LIST_FOREACH(keg, &uma_kegs, uk_link) { 2646 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 2647 zfunc(zone, arg); 2648 } 2649 LIST_FOREACH(zone, &uma_cachezones, uz_link) 2650 zfunc(zone, arg); 2651 } 2652 2653 /* 2654 * Traverses every zone in the system and calls a callback 2655 * 2656 * Arguments: 2657 * zfunc A pointer to a function which accepts a zone 2658 * as an argument. 2659 * 2660 * Returns: 2661 * Nothing 2662 */ 2663 static void 2664 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg) 2665 { 2666 2667 rw_rlock(&uma_rwlock); 2668 zone_foreach_unlocked(zfunc, arg); 2669 rw_runlock(&uma_rwlock); 2670 } 2671 2672 /* 2673 * Initialize the kernel memory allocator. This is done after pages can be 2674 * allocated but before general KVA is available. 2675 */ 2676 void 2677 uma_startup1(vm_offset_t virtual_avail) 2678 { 2679 struct uma_zctor_args args; 2680 size_t ksize, zsize, size; 2681 uma_keg_t masterkeg; 2682 uintptr_t m; 2683 uint8_t pflag; 2684 2685 bootstart = bootmem = virtual_avail; 2686 2687 rw_init(&uma_rwlock, "UMA lock"); 2688 sx_init(&uma_reclaim_lock, "umareclaim"); 2689 2690 ksize = sizeof(struct uma_keg) + 2691 (sizeof(struct uma_domain) * vm_ndomains); 2692 ksize = roundup(ksize, UMA_SUPER_ALIGN); 2693 zsize = sizeof(struct uma_zone) + 2694 (sizeof(struct uma_cache) * (mp_maxid + 1)) + 2695 (sizeof(struct uma_zone_domain) * vm_ndomains); 2696 zsize = roundup(zsize, UMA_SUPER_ALIGN); 2697 2698 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */ 2699 size = (zsize * 2) + ksize; 2700 m = (uintptr_t)startup_alloc(NULL, size, 0, &pflag, M_NOWAIT | M_ZERO); 2701 zones = (uma_zone_t)m; 2702 m += zsize; 2703 kegs = (uma_zone_t)m; 2704 m += zsize; 2705 masterkeg = (uma_keg_t)m; 2706 2707 /* "manually" create the initial zone */ 2708 memset(&args, 0, sizeof(args)); 2709 args.name = "UMA Kegs"; 2710 args.size = ksize; 2711 args.ctor = keg_ctor; 2712 args.dtor = keg_dtor; 2713 args.uminit = zero_init; 2714 args.fini = NULL; 2715 args.keg = masterkeg; 2716 args.align = UMA_SUPER_ALIGN - 1; 2717 args.flags = UMA_ZFLAG_INTERNAL; 2718 zone_ctor(kegs, zsize, &args, M_WAITOK); 2719 2720 args.name = "UMA Zones"; 2721 args.size = zsize; 2722 args.ctor = zone_ctor; 2723 args.dtor = zone_dtor; 2724 args.uminit = zero_init; 2725 args.fini = NULL; 2726 args.keg = NULL; 2727 args.align = UMA_SUPER_ALIGN - 1; 2728 args.flags = UMA_ZFLAG_INTERNAL; 2729 zone_ctor(zones, zsize, &args, M_WAITOK); 2730 2731 /* Now make zones for slab headers */ 2732 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE, 2733 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2734 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE, 2735 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2736 2737 hashzone = uma_zcreate("UMA Hash", 2738 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 2739 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2740 2741 bucket_init(); 2742 smr_init(); 2743 } 2744 2745 #ifndef UMA_MD_SMALL_ALLOC 2746 extern void vm_radix_reserve_kva(void); 2747 #endif 2748 2749 /* 2750 * Advertise the availability of normal kva allocations and switch to 2751 * the default back-end allocator. Marks the KVA we consumed on startup 2752 * as used in the map. 2753 */ 2754 void 2755 uma_startup2(void) 2756 { 2757 2758 if (bootstart != bootmem) { 2759 vm_map_lock(kernel_map); 2760 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem, 2761 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 2762 vm_map_unlock(kernel_map); 2763 } 2764 2765 #ifndef UMA_MD_SMALL_ALLOC 2766 /* Set up radix zone to use noobj_alloc. */ 2767 vm_radix_reserve_kva(); 2768 #endif 2769 2770 booted = BOOT_KVA; 2771 zone_foreach_unlocked(zone_kva_available, NULL); 2772 bucket_enable(); 2773 } 2774 2775 /* 2776 * Finish our initialization steps. 2777 */ 2778 static void 2779 uma_startup3(void) 2780 { 2781 2782 #ifdef INVARIANTS 2783 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor); 2784 uma_dbg_cnt = counter_u64_alloc(M_WAITOK); 2785 uma_skip_cnt = counter_u64_alloc(M_WAITOK); 2786 #endif 2787 zone_foreach_unlocked(zone_alloc_counters, NULL); 2788 zone_foreach_unlocked(zone_alloc_sysctl, NULL); 2789 callout_init(&uma_callout, 1); 2790 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 2791 booted = BOOT_RUNNING; 2792 2793 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL, 2794 EVENTHANDLER_PRI_FIRST); 2795 } 2796 2797 static void 2798 uma_shutdown(void) 2799 { 2800 2801 booted = BOOT_SHUTDOWN; 2802 } 2803 2804 static uma_keg_t 2805 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 2806 int align, uint32_t flags) 2807 { 2808 struct uma_kctor_args args; 2809 2810 args.size = size; 2811 args.uminit = uminit; 2812 args.fini = fini; 2813 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 2814 args.flags = flags; 2815 args.zone = zone; 2816 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK)); 2817 } 2818 2819 /* Public functions */ 2820 /* See uma.h */ 2821 void 2822 uma_set_align(int align) 2823 { 2824 2825 if (align != UMA_ALIGN_CACHE) 2826 uma_align_cache = align; 2827 } 2828 2829 /* See uma.h */ 2830 uma_zone_t 2831 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 2832 uma_init uminit, uma_fini fini, int align, uint32_t flags) 2833 2834 { 2835 struct uma_zctor_args args; 2836 uma_zone_t res; 2837 2838 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"", 2839 align, name)); 2840 2841 /* This stuff is essential for the zone ctor */ 2842 memset(&args, 0, sizeof(args)); 2843 args.name = name; 2844 args.size = size; 2845 args.ctor = ctor; 2846 args.dtor = dtor; 2847 args.uminit = uminit; 2848 args.fini = fini; 2849 #ifdef INVARIANTS 2850 /* 2851 * Inject procedures which check for memory use after free if we are 2852 * allowed to scramble the memory while it is not allocated. This 2853 * requires that: UMA is actually able to access the memory, no init 2854 * or fini procedures, no dependency on the initial value of the 2855 * memory, and no (legitimate) use of the memory after free. Note, 2856 * the ctor and dtor do not need to be empty. 2857 */ 2858 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH | 2859 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) { 2860 args.uminit = trash_init; 2861 args.fini = trash_fini; 2862 } 2863 #endif 2864 args.align = align; 2865 args.flags = flags; 2866 args.keg = NULL; 2867 2868 sx_slock(&uma_reclaim_lock); 2869 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2870 sx_sunlock(&uma_reclaim_lock); 2871 2872 return (res); 2873 } 2874 2875 /* See uma.h */ 2876 uma_zone_t 2877 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 2878 uma_init zinit, uma_fini zfini, uma_zone_t master) 2879 { 2880 struct uma_zctor_args args; 2881 uma_keg_t keg; 2882 uma_zone_t res; 2883 2884 keg = master->uz_keg; 2885 memset(&args, 0, sizeof(args)); 2886 args.name = name; 2887 args.size = keg->uk_size; 2888 args.ctor = ctor; 2889 args.dtor = dtor; 2890 args.uminit = zinit; 2891 args.fini = zfini; 2892 args.align = keg->uk_align; 2893 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 2894 args.keg = keg; 2895 2896 sx_slock(&uma_reclaim_lock); 2897 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2898 sx_sunlock(&uma_reclaim_lock); 2899 2900 return (res); 2901 } 2902 2903 /* See uma.h */ 2904 uma_zone_t 2905 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 2906 uma_init zinit, uma_fini zfini, uma_import zimport, 2907 uma_release zrelease, void *arg, int flags) 2908 { 2909 struct uma_zctor_args args; 2910 2911 memset(&args, 0, sizeof(args)); 2912 args.name = name; 2913 args.size = size; 2914 args.ctor = ctor; 2915 args.dtor = dtor; 2916 args.uminit = zinit; 2917 args.fini = zfini; 2918 args.import = zimport; 2919 args.release = zrelease; 2920 args.arg = arg; 2921 args.align = 0; 2922 args.flags = flags | UMA_ZFLAG_CACHE; 2923 2924 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK)); 2925 } 2926 2927 /* See uma.h */ 2928 void 2929 uma_zdestroy(uma_zone_t zone) 2930 { 2931 2932 /* 2933 * Large slabs are expensive to reclaim, so don't bother doing 2934 * unnecessary work if we're shutting down. 2935 */ 2936 if (booted == BOOT_SHUTDOWN && 2937 zone->uz_fini == NULL && zone->uz_release == zone_release) 2938 return; 2939 sx_slock(&uma_reclaim_lock); 2940 zone_free_item(zones, zone, NULL, SKIP_NONE); 2941 sx_sunlock(&uma_reclaim_lock); 2942 } 2943 2944 void 2945 uma_zwait(uma_zone_t zone) 2946 { 2947 void *item; 2948 2949 item = uma_zalloc_arg(zone, NULL, M_WAITOK); 2950 uma_zfree(zone, item); 2951 } 2952 2953 void * 2954 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags) 2955 { 2956 void *item, *pcpu_item; 2957 #ifdef SMP 2958 int i; 2959 2960 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2961 #endif 2962 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO); 2963 if (item == NULL) 2964 return (NULL); 2965 pcpu_item = zpcpu_base_to_offset(item); 2966 if (flags & M_ZERO) { 2967 #ifdef SMP 2968 for (i = 0; i <= mp_maxid; i++) 2969 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size); 2970 #else 2971 bzero(item, zone->uz_size); 2972 #endif 2973 } 2974 return (pcpu_item); 2975 } 2976 2977 /* 2978 * A stub while both regular and pcpu cases are identical. 2979 */ 2980 void 2981 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata) 2982 { 2983 void *item; 2984 2985 #ifdef SMP 2986 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2987 #endif 2988 item = zpcpu_offset_to_base(pcpu_item); 2989 uma_zfree_arg(zone, item, udata); 2990 } 2991 2992 static inline void * 2993 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags, 2994 void *item) 2995 { 2996 #ifdef INVARIANTS 2997 bool skipdbg; 2998 2999 skipdbg = uma_dbg_zskip(zone, item); 3000 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 && 3001 zone->uz_ctor != trash_ctor) 3002 trash_ctor(item, size, udata, flags); 3003 #endif 3004 /* Check flags before loading ctor pointer. */ 3005 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) && 3006 __predict_false(zone->uz_ctor != NULL) && 3007 zone->uz_ctor(item, size, udata, flags) != 0) { 3008 counter_u64_add(zone->uz_fails, 1); 3009 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT); 3010 return (NULL); 3011 } 3012 #ifdef INVARIANTS 3013 if (!skipdbg) 3014 uma_dbg_alloc(zone, NULL, item); 3015 #endif 3016 if (__predict_false(flags & M_ZERO)) 3017 return (memset(item, 0, size)); 3018 3019 return (item); 3020 } 3021 3022 static inline void 3023 item_dtor(uma_zone_t zone, void *item, int size, void *udata, 3024 enum zfreeskip skip) 3025 { 3026 #ifdef INVARIANTS 3027 bool skipdbg; 3028 3029 skipdbg = uma_dbg_zskip(zone, item); 3030 if (skip == SKIP_NONE && !skipdbg) { 3031 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0) 3032 uma_dbg_free(zone, udata, item); 3033 else 3034 uma_dbg_free(zone, NULL, item); 3035 } 3036 #endif 3037 if (__predict_true(skip < SKIP_DTOR)) { 3038 if (zone->uz_dtor != NULL) 3039 zone->uz_dtor(item, size, udata); 3040 #ifdef INVARIANTS 3041 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 && 3042 zone->uz_dtor != trash_dtor) 3043 trash_dtor(item, size, udata); 3044 #endif 3045 } 3046 } 3047 3048 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS) 3049 #define UMA_ZALLOC_DEBUG 3050 static int 3051 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags) 3052 { 3053 int error; 3054 3055 error = 0; 3056 #ifdef WITNESS 3057 if (flags & M_WAITOK) { 3058 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 3059 "uma_zalloc_debug: zone \"%s\"", zone->uz_name); 3060 } 3061 #endif 3062 3063 #ifdef INVARIANTS 3064 KASSERT((flags & M_EXEC) == 0, 3065 ("uma_zalloc_debug: called with M_EXEC")); 3066 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3067 ("uma_zalloc_debug: called within spinlock or critical section")); 3068 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0, 3069 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO")); 3070 #endif 3071 3072 #ifdef DEBUG_MEMGUARD 3073 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) { 3074 void *item; 3075 item = memguard_alloc(zone->uz_size, flags); 3076 if (item != NULL) { 3077 error = EJUSTRETURN; 3078 if (zone->uz_init != NULL && 3079 zone->uz_init(item, zone->uz_size, flags) != 0) { 3080 *itemp = NULL; 3081 return (error); 3082 } 3083 if (zone->uz_ctor != NULL && 3084 zone->uz_ctor(item, zone->uz_size, udata, 3085 flags) != 0) { 3086 counter_u64_add(zone->uz_fails, 1); 3087 zone->uz_fini(item, zone->uz_size); 3088 *itemp = NULL; 3089 return (error); 3090 } 3091 *itemp = item; 3092 return (error); 3093 } 3094 /* This is unfortunate but should not be fatal. */ 3095 } 3096 #endif 3097 return (error); 3098 } 3099 3100 static int 3101 uma_zfree_debug(uma_zone_t zone, void *item, void *udata) 3102 { 3103 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3104 ("uma_zfree_debug: called with spinlock or critical section held")); 3105 3106 #ifdef DEBUG_MEMGUARD 3107 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) { 3108 if (zone->uz_dtor != NULL) 3109 zone->uz_dtor(item, zone->uz_size, udata); 3110 if (zone->uz_fini != NULL) 3111 zone->uz_fini(item, zone->uz_size); 3112 memguard_free(item); 3113 return (EJUSTRETURN); 3114 } 3115 #endif 3116 return (0); 3117 } 3118 #endif 3119 3120 static inline void * 3121 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket, 3122 void *udata, int flags) 3123 { 3124 void *item; 3125 int size, uz_flags; 3126 3127 item = cache_bucket_pop(cache, bucket); 3128 size = cache_uz_size(cache); 3129 uz_flags = cache_uz_flags(cache); 3130 critical_exit(); 3131 return (item_ctor(zone, uz_flags, size, udata, flags, item)); 3132 } 3133 3134 static __noinline void * 3135 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags) 3136 { 3137 uma_cache_bucket_t bucket; 3138 int domain; 3139 3140 while (cache_alloc(zone, cache, udata, flags)) { 3141 cache = &zone->uz_cpu[curcpu]; 3142 bucket = &cache->uc_allocbucket; 3143 if (__predict_false(bucket->ucb_cnt == 0)) 3144 continue; 3145 return (cache_alloc_item(zone, cache, bucket, udata, flags)); 3146 } 3147 critical_exit(); 3148 3149 /* 3150 * We can not get a bucket so try to return a single item. 3151 */ 3152 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) 3153 domain = PCPU_GET(domain); 3154 else 3155 domain = UMA_ANYDOMAIN; 3156 return (zone_alloc_item(zone, udata, domain, flags)); 3157 } 3158 3159 /* See uma.h */ 3160 void * 3161 uma_zalloc_smr(uma_zone_t zone, int flags) 3162 { 3163 uma_cache_bucket_t bucket; 3164 uma_cache_t cache; 3165 3166 #ifdef UMA_ZALLOC_DEBUG 3167 void *item; 3168 3169 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0, 3170 ("uma_zalloc_arg: called with non-SMR zone.\n")); 3171 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN) 3172 return (item); 3173 #endif 3174 3175 critical_enter(); 3176 cache = &zone->uz_cpu[curcpu]; 3177 bucket = &cache->uc_allocbucket; 3178 if (__predict_false(bucket->ucb_cnt == 0)) 3179 return (cache_alloc_retry(zone, cache, NULL, flags)); 3180 return (cache_alloc_item(zone, cache, bucket, NULL, flags)); 3181 } 3182 3183 /* See uma.h */ 3184 void * 3185 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 3186 { 3187 uma_cache_bucket_t bucket; 3188 uma_cache_t cache; 3189 3190 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3191 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3192 3193 /* This is the fast path allocation */ 3194 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name, 3195 zone, flags); 3196 3197 #ifdef UMA_ZALLOC_DEBUG 3198 void *item; 3199 3200 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0, 3201 ("uma_zalloc_arg: called with SMR zone.\n")); 3202 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN) 3203 return (item); 3204 #endif 3205 3206 /* 3207 * If possible, allocate from the per-CPU cache. There are two 3208 * requirements for safe access to the per-CPU cache: (1) the thread 3209 * accessing the cache must not be preempted or yield during access, 3210 * and (2) the thread must not migrate CPUs without switching which 3211 * cache it accesses. We rely on a critical section to prevent 3212 * preemption and migration. We release the critical section in 3213 * order to acquire the zone mutex if we are unable to allocate from 3214 * the current cache; when we re-acquire the critical section, we 3215 * must detect and handle migration if it has occurred. 3216 */ 3217 critical_enter(); 3218 cache = &zone->uz_cpu[curcpu]; 3219 bucket = &cache->uc_allocbucket; 3220 if (__predict_false(bucket->ucb_cnt == 0)) 3221 return (cache_alloc_retry(zone, cache, udata, flags)); 3222 return (cache_alloc_item(zone, cache, bucket, udata, flags)); 3223 } 3224 3225 /* 3226 * Replenish an alloc bucket and possibly restore an old one. Called in 3227 * a critical section. Returns in a critical section. 3228 * 3229 * A false return value indicates an allocation failure. 3230 * A true return value indicates success and the caller should retry. 3231 */ 3232 static __noinline bool 3233 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags) 3234 { 3235 uma_zone_domain_t zdom; 3236 uma_bucket_t bucket; 3237 int domain; 3238 bool lockfail; 3239 3240 CRITICAL_ASSERT(curthread); 3241 3242 /* 3243 * If we have run out of items in our alloc bucket see 3244 * if we can switch with the free bucket. 3245 * 3246 * SMR Zones can't re-use the free bucket until the sequence has 3247 * expired. 3248 */ 3249 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && 3250 cache->uc_freebucket.ucb_cnt != 0) { 3251 cache_bucket_swap(&cache->uc_freebucket, 3252 &cache->uc_allocbucket); 3253 return (true); 3254 } 3255 3256 /* 3257 * Discard any empty allocation bucket while we hold no locks. 3258 */ 3259 bucket = cache_bucket_unload_alloc(cache); 3260 critical_exit(); 3261 if (bucket != NULL) 3262 bucket_free(zone, bucket, udata); 3263 3264 /* Short-circuit for zones without buckets and low memory. */ 3265 if (zone->uz_bucket_size == 0 || bucketdisable) { 3266 critical_enter(); 3267 return (false); 3268 } 3269 3270 /* 3271 * Attempt to retrieve the item from the per-CPU cache has failed, so 3272 * we must go back to the zone. This requires the zone lock, so we 3273 * must drop the critical section, then re-acquire it when we go back 3274 * to the cache. Since the critical section is released, we may be 3275 * preempted or migrate. As such, make sure not to maintain any 3276 * thread-local state specific to the cache from prior to releasing 3277 * the critical section. 3278 */ 3279 lockfail = 0; 3280 if (ZONE_TRYLOCK(zone) == 0) { 3281 /* Record contention to size the buckets. */ 3282 ZONE_LOCK(zone); 3283 lockfail = 1; 3284 } 3285 3286 /* See if we lost the race to fill the cache. */ 3287 critical_enter(); 3288 cache = &zone->uz_cpu[curcpu]; 3289 if (cache->uc_allocbucket.ucb_bucket != NULL) { 3290 ZONE_UNLOCK(zone); 3291 return (true); 3292 } 3293 3294 /* 3295 * Check the zone's cache of buckets. 3296 */ 3297 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH) { 3298 domain = PCPU_GET(domain); 3299 zdom = &zone->uz_domain[domain]; 3300 } else { 3301 domain = UMA_ANYDOMAIN; 3302 zdom = &zone->uz_domain[0]; 3303 } 3304 3305 if ((bucket = zone_fetch_bucket(zone, zdom)) != NULL) { 3306 KASSERT(bucket->ub_cnt != 0, 3307 ("uma_zalloc_arg: Returning an empty bucket.")); 3308 cache_bucket_load_alloc(cache, bucket); 3309 return (true); 3310 } 3311 /* We are no longer associated with this CPU. */ 3312 critical_exit(); 3313 3314 /* 3315 * We bump the uz count when the cache size is insufficient to 3316 * handle the working set. 3317 */ 3318 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max) 3319 zone->uz_bucket_size++; 3320 ZONE_UNLOCK(zone); 3321 3322 /* 3323 * Fill a bucket and attempt to use it as the alloc bucket. 3324 */ 3325 bucket = zone_alloc_bucket(zone, udata, domain, flags); 3326 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p", 3327 zone->uz_name, zone, bucket); 3328 if (bucket == NULL) { 3329 critical_enter(); 3330 return (false); 3331 } 3332 3333 /* 3334 * See if we lost the race or were migrated. Cache the 3335 * initialized bucket to make this less likely or claim 3336 * the memory directly. 3337 */ 3338 ZONE_LOCK(zone); 3339 critical_enter(); 3340 cache = &zone->uz_cpu[curcpu]; 3341 if (cache->uc_allocbucket.ucb_bucket == NULL && 3342 ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0 || 3343 domain == PCPU_GET(domain))) { 3344 cache_bucket_load_alloc(cache, bucket); 3345 zdom->uzd_imax += bucket->ub_cnt; 3346 } else if (zone->uz_bkt_count >= zone->uz_bkt_max) { 3347 critical_exit(); 3348 ZONE_UNLOCK(zone); 3349 bucket_drain(zone, bucket); 3350 bucket_free(zone, bucket, udata); 3351 critical_enter(); 3352 return (true); 3353 } else 3354 zone_put_bucket(zone, zdom, bucket, false); 3355 ZONE_UNLOCK(zone); 3356 return (true); 3357 } 3358 3359 void * 3360 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags) 3361 { 3362 3363 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3364 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3365 3366 /* This is the fast path allocation */ 3367 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d", 3368 zone->uz_name, zone, domain, flags); 3369 3370 if (flags & M_WAITOK) { 3371 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 3372 "uma_zalloc_domain: zone \"%s\"", zone->uz_name); 3373 } 3374 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3375 ("uma_zalloc_domain: called with spinlock or critical section held")); 3376 3377 return (zone_alloc_item(zone, udata, domain, flags)); 3378 } 3379 3380 /* 3381 * Find a slab with some space. Prefer slabs that are partially used over those 3382 * that are totally full. This helps to reduce fragmentation. 3383 * 3384 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check 3385 * only 'domain'. 3386 */ 3387 static uma_slab_t 3388 keg_first_slab(uma_keg_t keg, int domain, bool rr) 3389 { 3390 uma_domain_t dom; 3391 uma_slab_t slab; 3392 int start; 3393 3394 KASSERT(domain >= 0 && domain < vm_ndomains, 3395 ("keg_first_slab: domain %d out of range", domain)); 3396 KEG_LOCK_ASSERT(keg, domain); 3397 3398 slab = NULL; 3399 start = domain; 3400 do { 3401 dom = &keg->uk_domain[domain]; 3402 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL) 3403 return (slab); 3404 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) { 3405 LIST_REMOVE(slab, us_link); 3406 dom->ud_free_slabs--; 3407 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 3408 return (slab); 3409 } 3410 if (rr) 3411 domain = (domain + 1) % vm_ndomains; 3412 } while (domain != start); 3413 3414 return (NULL); 3415 } 3416 3417 /* 3418 * Fetch an existing slab from a free or partial list. Returns with the 3419 * keg domain lock held if a slab was found or unlocked if not. 3420 */ 3421 static uma_slab_t 3422 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags) 3423 { 3424 uma_slab_t slab; 3425 uint32_t reserve; 3426 3427 /* HASH has a single free list. */ 3428 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) 3429 domain = 0; 3430 3431 KEG_LOCK(keg, domain); 3432 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve; 3433 if (keg->uk_domain[domain].ud_free_items <= reserve || 3434 (slab = keg_first_slab(keg, domain, rr)) == NULL) { 3435 KEG_UNLOCK(keg, domain); 3436 return (NULL); 3437 } 3438 return (slab); 3439 } 3440 3441 static uma_slab_t 3442 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags) 3443 { 3444 struct vm_domainset_iter di; 3445 uma_slab_t slab; 3446 int aflags, domain; 3447 bool rr; 3448 3449 restart: 3450 /* 3451 * Use the keg's policy if upper layers haven't already specified a 3452 * domain (as happens with first-touch zones). 3453 * 3454 * To avoid races we run the iterator with the keg lock held, but that 3455 * means that we cannot allow the vm_domainset layer to sleep. Thus, 3456 * clear M_WAITOK and handle low memory conditions locally. 3457 */ 3458 rr = rdomain == UMA_ANYDOMAIN; 3459 if (rr) { 3460 aflags = (flags & ~M_WAITOK) | M_NOWAIT; 3461 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 3462 &aflags); 3463 } else { 3464 aflags = flags; 3465 domain = rdomain; 3466 } 3467 3468 for (;;) { 3469 slab = keg_fetch_free_slab(keg, domain, rr, flags); 3470 if (slab != NULL) 3471 return (slab); 3472 3473 /* 3474 * M_NOVM means don't ask at all! 3475 */ 3476 if (flags & M_NOVM) 3477 break; 3478 3479 slab = keg_alloc_slab(keg, zone, domain, flags, aflags); 3480 if (slab != NULL) 3481 return (slab); 3482 if (!rr && (flags & M_WAITOK) == 0) 3483 break; 3484 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) { 3485 if ((flags & M_WAITOK) != 0) { 3486 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask); 3487 goto restart; 3488 } 3489 break; 3490 } 3491 } 3492 3493 /* 3494 * We might not have been able to get a slab but another cpu 3495 * could have while we were unlocked. Check again before we 3496 * fail. 3497 */ 3498 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL) 3499 return (slab); 3500 3501 return (NULL); 3502 } 3503 3504 static void * 3505 slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 3506 { 3507 uma_domain_t dom; 3508 void *item; 3509 int freei; 3510 3511 KEG_LOCK_ASSERT(keg, slab->us_domain); 3512 3513 dom = &keg->uk_domain[slab->us_domain]; 3514 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1; 3515 BIT_CLR(keg->uk_ipers, freei, &slab->us_free); 3516 item = slab_item(slab, keg, freei); 3517 slab->us_freecount--; 3518 dom->ud_free_items--; 3519 3520 /* 3521 * Move this slab to the full list. It must be on the partial list, so 3522 * we do not need to update the free slab count. In particular, 3523 * keg_fetch_slab() always returns slabs on the partial list. 3524 */ 3525 if (slab->us_freecount == 0) { 3526 LIST_REMOVE(slab, us_link); 3527 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link); 3528 } 3529 3530 return (item); 3531 } 3532 3533 static int 3534 zone_import(void *arg, void **bucket, int max, int domain, int flags) 3535 { 3536 uma_domain_t dom; 3537 uma_zone_t zone; 3538 uma_slab_t slab; 3539 uma_keg_t keg; 3540 #ifdef NUMA 3541 int stripe; 3542 #endif 3543 int i; 3544 3545 zone = arg; 3546 slab = NULL; 3547 keg = zone->uz_keg; 3548 /* Try to keep the buckets totally full */ 3549 for (i = 0; i < max; ) { 3550 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL) 3551 break; 3552 #ifdef NUMA 3553 stripe = howmany(max, vm_ndomains); 3554 #endif 3555 dom = &keg->uk_domain[slab->us_domain]; 3556 while (slab->us_freecount && i < max) { 3557 bucket[i++] = slab_alloc_item(keg, slab); 3558 if (dom->ud_free_items <= keg->uk_reserve) 3559 break; 3560 #ifdef NUMA 3561 /* 3562 * If the zone is striped we pick a new slab for every 3563 * N allocations. Eliminating this conditional will 3564 * instead pick a new domain for each bucket rather 3565 * than stripe within each bucket. The current option 3566 * produces more fragmentation and requires more cpu 3567 * time but yields better distribution. 3568 */ 3569 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 && 3570 vm_ndomains > 1 && --stripe == 0) 3571 break; 3572 #endif 3573 } 3574 KEG_UNLOCK(keg, slab->us_domain); 3575 /* Don't block if we allocated any successfully. */ 3576 flags &= ~M_WAITOK; 3577 flags |= M_NOWAIT; 3578 } 3579 3580 return i; 3581 } 3582 3583 static int 3584 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags) 3585 { 3586 uint64_t old, new, total, max; 3587 3588 /* 3589 * The hard case. We're going to sleep because there were existing 3590 * sleepers or because we ran out of items. This routine enforces 3591 * fairness by keeping fifo order. 3592 * 3593 * First release our ill gotten gains and make some noise. 3594 */ 3595 for (;;) { 3596 zone_free_limit(zone, count); 3597 zone_log_warning(zone); 3598 zone_maxaction(zone); 3599 if (flags & M_NOWAIT) 3600 return (0); 3601 3602 /* 3603 * We need to allocate an item or set ourself as a sleeper 3604 * while the sleepq lock is held to avoid wakeup races. This 3605 * is essentially a home rolled semaphore. 3606 */ 3607 sleepq_lock(&zone->uz_max_items); 3608 old = zone->uz_items; 3609 do { 3610 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX); 3611 /* Cache the max since we will evaluate twice. */ 3612 max = zone->uz_max_items; 3613 if (UZ_ITEMS_SLEEPERS(old) != 0 || 3614 UZ_ITEMS_COUNT(old) >= max) 3615 new = old + UZ_ITEMS_SLEEPER; 3616 else 3617 new = old + MIN(count, max - old); 3618 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0); 3619 3620 /* We may have successfully allocated under the sleepq lock. */ 3621 if (UZ_ITEMS_SLEEPERS(new) == 0) { 3622 sleepq_release(&zone->uz_max_items); 3623 return (new - old); 3624 } 3625 3626 /* 3627 * This is in a different cacheline from uz_items so that we 3628 * don't constantly invalidate the fastpath cacheline when we 3629 * adjust item counts. This could be limited to toggling on 3630 * transitions. 3631 */ 3632 atomic_add_32(&zone->uz_sleepers, 1); 3633 atomic_add_64(&zone->uz_sleeps, 1); 3634 3635 /* 3636 * We have added ourselves as a sleeper. The sleepq lock 3637 * protects us from wakeup races. Sleep now and then retry. 3638 */ 3639 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0); 3640 sleepq_wait(&zone->uz_max_items, PVM); 3641 3642 /* 3643 * After wakeup, remove ourselves as a sleeper and try 3644 * again. We no longer have the sleepq lock for protection. 3645 * 3646 * Subract ourselves as a sleeper while attempting to add 3647 * our count. 3648 */ 3649 atomic_subtract_32(&zone->uz_sleepers, 1); 3650 old = atomic_fetchadd_64(&zone->uz_items, 3651 -(UZ_ITEMS_SLEEPER - count)); 3652 /* We're no longer a sleeper. */ 3653 old -= UZ_ITEMS_SLEEPER; 3654 3655 /* 3656 * If we're still at the limit, restart. Notably do not 3657 * block on other sleepers. Cache the max value to protect 3658 * against changes via sysctl. 3659 */ 3660 total = UZ_ITEMS_COUNT(old); 3661 max = zone->uz_max_items; 3662 if (total >= max) 3663 continue; 3664 /* Truncate if necessary, otherwise wake other sleepers. */ 3665 if (total + count > max) { 3666 zone_free_limit(zone, total + count - max); 3667 count = max - total; 3668 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0) 3669 wakeup_one(&zone->uz_max_items); 3670 3671 return (count); 3672 } 3673 } 3674 3675 /* 3676 * Allocate 'count' items from our max_items limit. Returns the number 3677 * available. If M_NOWAIT is not specified it will sleep until at least 3678 * one item can be allocated. 3679 */ 3680 static int 3681 zone_alloc_limit(uma_zone_t zone, int count, int flags) 3682 { 3683 uint64_t old; 3684 uint64_t max; 3685 3686 max = zone->uz_max_items; 3687 MPASS(max > 0); 3688 3689 /* 3690 * We expect normal allocations to succeed with a simple 3691 * fetchadd. 3692 */ 3693 old = atomic_fetchadd_64(&zone->uz_items, count); 3694 if (__predict_true(old + count <= max)) 3695 return (count); 3696 3697 /* 3698 * If we had some items and no sleepers just return the 3699 * truncated value. We have to release the excess space 3700 * though because that may wake sleepers who weren't woken 3701 * because we were temporarily over the limit. 3702 */ 3703 if (old < max) { 3704 zone_free_limit(zone, (old + count) - max); 3705 return (max - old); 3706 } 3707 return (zone_alloc_limit_hard(zone, count, flags)); 3708 } 3709 3710 /* 3711 * Free a number of items back to the limit. 3712 */ 3713 static void 3714 zone_free_limit(uma_zone_t zone, int count) 3715 { 3716 uint64_t old; 3717 3718 MPASS(count > 0); 3719 3720 /* 3721 * In the common case we either have no sleepers or 3722 * are still over the limit and can just return. 3723 */ 3724 old = atomic_fetchadd_64(&zone->uz_items, -count); 3725 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 || 3726 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items)) 3727 return; 3728 3729 /* 3730 * Moderate the rate of wakeups. Sleepers will continue 3731 * to generate wakeups if necessary. 3732 */ 3733 wakeup_one(&zone->uz_max_items); 3734 } 3735 3736 static uma_bucket_t 3737 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags) 3738 { 3739 uma_bucket_t bucket; 3740 int maxbucket, cnt; 3741 3742 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name, 3743 zone, domain); 3744 3745 /* Avoid allocs targeting empty domains. */ 3746 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain)) 3747 domain = UMA_ANYDOMAIN; 3748 3749 if (zone->uz_max_items > 0) 3750 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size, 3751 M_NOWAIT); 3752 else 3753 maxbucket = zone->uz_bucket_size; 3754 if (maxbucket == 0) 3755 return (false); 3756 3757 /* Don't wait for buckets, preserve caller's NOVM setting. */ 3758 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 3759 if (bucket == NULL) { 3760 cnt = 0; 3761 goto out; 3762 } 3763 3764 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 3765 MIN(maxbucket, bucket->ub_entries), domain, flags); 3766 3767 /* 3768 * Initialize the memory if necessary. 3769 */ 3770 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 3771 int i; 3772 3773 for (i = 0; i < bucket->ub_cnt; i++) 3774 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 3775 flags) != 0) 3776 break; 3777 /* 3778 * If we couldn't initialize the whole bucket, put the 3779 * rest back onto the freelist. 3780 */ 3781 if (i != bucket->ub_cnt) { 3782 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 3783 bucket->ub_cnt - i); 3784 #ifdef INVARIANTS 3785 bzero(&bucket->ub_bucket[i], 3786 sizeof(void *) * (bucket->ub_cnt - i)); 3787 #endif 3788 bucket->ub_cnt = i; 3789 } 3790 } 3791 3792 cnt = bucket->ub_cnt; 3793 if (bucket->ub_cnt == 0) { 3794 bucket_free(zone, bucket, udata); 3795 counter_u64_add(zone->uz_fails, 1); 3796 bucket = NULL; 3797 } 3798 out: 3799 if (zone->uz_max_items > 0 && cnt < maxbucket) 3800 zone_free_limit(zone, maxbucket - cnt); 3801 3802 return (bucket); 3803 } 3804 3805 /* 3806 * Allocates a single item from a zone. 3807 * 3808 * Arguments 3809 * zone The zone to alloc for. 3810 * udata The data to be passed to the constructor. 3811 * domain The domain to allocate from or UMA_ANYDOMAIN. 3812 * flags M_WAITOK, M_NOWAIT, M_ZERO. 3813 * 3814 * Returns 3815 * NULL if there is no memory and M_NOWAIT is set 3816 * An item if successful 3817 */ 3818 3819 static void * 3820 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags) 3821 { 3822 void *item; 3823 3824 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) 3825 return (NULL); 3826 3827 /* Avoid allocs targeting empty domains. */ 3828 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain)) 3829 domain = UMA_ANYDOMAIN; 3830 3831 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1) 3832 goto fail_cnt; 3833 3834 /* 3835 * We have to call both the zone's init (not the keg's init) 3836 * and the zone's ctor. This is because the item is going from 3837 * a keg slab directly to the user, and the user is expecting it 3838 * to be both zone-init'd as well as zone-ctor'd. 3839 */ 3840 if (zone->uz_init != NULL) { 3841 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 3842 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT); 3843 goto fail_cnt; 3844 } 3845 } 3846 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags, 3847 item); 3848 if (item == NULL) 3849 goto fail; 3850 3851 counter_u64_add(zone->uz_allocs, 1); 3852 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item, 3853 zone->uz_name, zone); 3854 3855 return (item); 3856 3857 fail_cnt: 3858 counter_u64_add(zone->uz_fails, 1); 3859 fail: 3860 if (zone->uz_max_items > 0) 3861 zone_free_limit(zone, 1); 3862 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)", 3863 zone->uz_name, zone); 3864 3865 return (NULL); 3866 } 3867 3868 /* See uma.h */ 3869 void 3870 uma_zfree_smr(uma_zone_t zone, void *item) 3871 { 3872 uma_cache_t cache; 3873 uma_cache_bucket_t bucket; 3874 int domain, itemdomain, uz_flags; 3875 3876 #ifdef UMA_ZALLOC_DEBUG 3877 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0, 3878 ("uma_zfree_smr: called with non-SMR zone.\n")); 3879 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer.")); 3880 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN) 3881 return; 3882 #endif 3883 cache = &zone->uz_cpu[curcpu]; 3884 uz_flags = cache_uz_flags(cache); 3885 domain = itemdomain = 0; 3886 #ifdef NUMA 3887 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 3888 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item)); 3889 #endif 3890 critical_enter(); 3891 do { 3892 cache = &zone->uz_cpu[curcpu]; 3893 /* SMR Zones must free to the free bucket. */ 3894 bucket = &cache->uc_freebucket; 3895 #ifdef NUMA 3896 domain = PCPU_GET(domain); 3897 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && 3898 domain != itemdomain) { 3899 bucket = &cache->uc_crossbucket; 3900 } 3901 #endif 3902 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) { 3903 cache_bucket_push(cache, bucket, item); 3904 critical_exit(); 3905 return; 3906 } 3907 } while (cache_free(zone, cache, NULL, item, itemdomain)); 3908 critical_exit(); 3909 3910 /* 3911 * If nothing else caught this, we'll just do an internal free. 3912 */ 3913 zone_free_item(zone, item, NULL, SKIP_NONE); 3914 } 3915 3916 /* See uma.h */ 3917 void 3918 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 3919 { 3920 uma_cache_t cache; 3921 uma_cache_bucket_t bucket; 3922 int domain, itemdomain, uz_flags; 3923 3924 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3925 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3926 3927 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone); 3928 3929 #ifdef UMA_ZALLOC_DEBUG 3930 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0, 3931 ("uma_zfree_arg: called with SMR zone.\n")); 3932 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN) 3933 return; 3934 #endif 3935 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 3936 if (item == NULL) 3937 return; 3938 3939 /* 3940 * We are accessing the per-cpu cache without a critical section to 3941 * fetch size and flags. This is acceptable, if we are preempted we 3942 * will simply read another cpu's line. 3943 */ 3944 cache = &zone->uz_cpu[curcpu]; 3945 uz_flags = cache_uz_flags(cache); 3946 if (UMA_ALWAYS_CTORDTOR || 3947 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0)) 3948 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE); 3949 3950 /* 3951 * The race here is acceptable. If we miss it we'll just have to wait 3952 * a little longer for the limits to be reset. 3953 */ 3954 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) { 3955 if (zone->uz_sleepers > 0) 3956 goto zfree_item; 3957 } 3958 3959 /* 3960 * If possible, free to the per-CPU cache. There are two 3961 * requirements for safe access to the per-CPU cache: (1) the thread 3962 * accessing the cache must not be preempted or yield during access, 3963 * and (2) the thread must not migrate CPUs without switching which 3964 * cache it accesses. We rely on a critical section to prevent 3965 * preemption and migration. We release the critical section in 3966 * order to acquire the zone mutex if we are unable to free to the 3967 * current cache; when we re-acquire the critical section, we must 3968 * detect and handle migration if it has occurred. 3969 */ 3970 domain = itemdomain = 0; 3971 #ifdef NUMA 3972 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 3973 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item)); 3974 #endif 3975 critical_enter(); 3976 do { 3977 cache = &zone->uz_cpu[curcpu]; 3978 /* 3979 * Try to free into the allocbucket first to give LIFO 3980 * ordering for cache-hot datastructures. Spill over 3981 * into the freebucket if necessary. Alloc will swap 3982 * them if one runs dry. 3983 */ 3984 bucket = &cache->uc_allocbucket; 3985 #ifdef NUMA 3986 domain = PCPU_GET(domain); 3987 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && 3988 domain != itemdomain) { 3989 bucket = &cache->uc_crossbucket; 3990 } else 3991 #endif 3992 if (bucket->ucb_cnt >= bucket->ucb_entries) 3993 bucket = &cache->uc_freebucket; 3994 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) { 3995 cache_bucket_push(cache, bucket, item); 3996 critical_exit(); 3997 return; 3998 } 3999 } while (cache_free(zone, cache, udata, item, itemdomain)); 4000 critical_exit(); 4001 4002 /* 4003 * If nothing else caught this, we'll just do an internal free. 4004 */ 4005 zfree_item: 4006 zone_free_item(zone, item, udata, SKIP_DTOR); 4007 } 4008 4009 #ifdef NUMA 4010 /* 4011 * sort crossdomain free buckets to domain correct buckets and cache 4012 * them. 4013 */ 4014 static void 4015 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata) 4016 { 4017 struct uma_bucketlist fullbuckets; 4018 uma_zone_domain_t zdom; 4019 uma_bucket_t b; 4020 smr_seq_t seq; 4021 void *item; 4022 int domain; 4023 4024 CTR3(KTR_UMA, 4025 "uma_zfree: zone %s(%p) draining cross bucket %p", 4026 zone->uz_name, zone, bucket); 4027 4028 STAILQ_INIT(&fullbuckets); 4029 4030 /* 4031 * To avoid having ndomain * ndomain buckets for sorting we have a 4032 * lock on the current crossfree bucket. A full matrix with 4033 * per-domain locking could be used if necessary. 4034 */ 4035 ZONE_CROSS_LOCK(zone); 4036 4037 /* 4038 * It is possible for buckets to arrive here out of order so we fetch 4039 * the current smr seq rather than accepting the bucket's. 4040 */ 4041 seq = SMR_SEQ_INVALID; 4042 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) 4043 seq = smr_current(zone->uz_smr); 4044 while (bucket->ub_cnt > 0) { 4045 item = bucket->ub_bucket[bucket->ub_cnt - 1]; 4046 domain = _vm_phys_domain(pmap_kextract((vm_offset_t)item)); 4047 zdom = &zone->uz_domain[domain]; 4048 if (zdom->uzd_cross == NULL) { 4049 zdom->uzd_cross = bucket_alloc(zone, udata, M_NOWAIT); 4050 if (zdom->uzd_cross == NULL) 4051 break; 4052 } 4053 b = zdom->uzd_cross; 4054 b->ub_bucket[b->ub_cnt++] = item; 4055 b->ub_seq = seq; 4056 if (b->ub_cnt == b->ub_entries) { 4057 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link); 4058 zdom->uzd_cross = NULL; 4059 } 4060 bucket->ub_cnt--; 4061 } 4062 ZONE_CROSS_UNLOCK(zone); 4063 if (!STAILQ_EMPTY(&fullbuckets)) { 4064 ZONE_LOCK(zone); 4065 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) { 4066 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link); 4067 if (zone->uz_bkt_count >= zone->uz_bkt_max) { 4068 ZONE_UNLOCK(zone); 4069 bucket_drain(zone, b); 4070 bucket_free(zone, b, udata); 4071 ZONE_LOCK(zone); 4072 } else { 4073 domain = _vm_phys_domain( 4074 pmap_kextract( 4075 (vm_offset_t)b->ub_bucket[0])); 4076 zdom = &zone->uz_domain[domain]; 4077 zone_put_bucket(zone, zdom, b, true); 4078 } 4079 } 4080 ZONE_UNLOCK(zone); 4081 } 4082 if (bucket->ub_cnt != 0) 4083 bucket_drain(zone, bucket); 4084 bucket->ub_seq = SMR_SEQ_INVALID; 4085 bucket_free(zone, bucket, udata); 4086 } 4087 #endif 4088 4089 static void 4090 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata, 4091 int domain, int itemdomain) 4092 { 4093 uma_zone_domain_t zdom; 4094 4095 #ifdef NUMA 4096 /* 4097 * Buckets coming from the wrong domain will be entirely for the 4098 * only other domain on two domain systems. In this case we can 4099 * simply cache them. Otherwise we need to sort them back to 4100 * correct domains. 4101 */ 4102 if (domain != itemdomain && vm_ndomains > 2) { 4103 zone_free_cross(zone, bucket, udata); 4104 return; 4105 } 4106 #endif 4107 4108 /* 4109 * Attempt to save the bucket in the zone's domain bucket cache. 4110 * 4111 * We bump the uz count when the cache size is insufficient to 4112 * handle the working set. 4113 */ 4114 if (ZONE_TRYLOCK(zone) == 0) { 4115 /* Record contention to size the buckets. */ 4116 ZONE_LOCK(zone); 4117 if (zone->uz_bucket_size < zone->uz_bucket_size_max) 4118 zone->uz_bucket_size++; 4119 } 4120 4121 CTR3(KTR_UMA, 4122 "uma_zfree: zone %s(%p) putting bucket %p on free list", 4123 zone->uz_name, zone, bucket); 4124 /* ub_cnt is pointing to the last free item */ 4125 KASSERT(bucket->ub_cnt == bucket->ub_entries, 4126 ("uma_zfree: Attempting to insert partial bucket onto the full list.\n")); 4127 if (zone->uz_bkt_count >= zone->uz_bkt_max) { 4128 ZONE_UNLOCK(zone); 4129 bucket_drain(zone, bucket); 4130 bucket_free(zone, bucket, udata); 4131 } else { 4132 zdom = &zone->uz_domain[itemdomain]; 4133 zone_put_bucket(zone, zdom, bucket, true); 4134 ZONE_UNLOCK(zone); 4135 } 4136 } 4137 4138 /* 4139 * Populate a free or cross bucket for the current cpu cache. Free any 4140 * existing full bucket either to the zone cache or back to the slab layer. 4141 * 4142 * Enters and returns in a critical section. false return indicates that 4143 * we can not satisfy this free in the cache layer. true indicates that 4144 * the caller should retry. 4145 */ 4146 static __noinline bool 4147 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item, 4148 int itemdomain) 4149 { 4150 uma_cache_bucket_t cbucket; 4151 uma_bucket_t newbucket, bucket; 4152 int domain; 4153 4154 CRITICAL_ASSERT(curthread); 4155 4156 if (zone->uz_bucket_size == 0) 4157 return false; 4158 4159 cache = &zone->uz_cpu[curcpu]; 4160 newbucket = NULL; 4161 4162 /* 4163 * FIRSTTOUCH domains need to free to the correct zdom. When 4164 * enabled this is the zdom of the item. The bucket is the 4165 * cross bucket if the current domain and itemdomain do not match. 4166 */ 4167 cbucket = &cache->uc_freebucket; 4168 #ifdef NUMA 4169 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) { 4170 domain = PCPU_GET(domain); 4171 if (domain != itemdomain) { 4172 cbucket = &cache->uc_crossbucket; 4173 if (cbucket->ucb_cnt != 0) 4174 atomic_add_64(&zone->uz_xdomain, 4175 cbucket->ucb_cnt); 4176 } 4177 } else 4178 #endif 4179 itemdomain = domain = 0; 4180 bucket = cache_bucket_unload(cbucket); 4181 4182 /* We are no longer associated with this CPU. */ 4183 critical_exit(); 4184 4185 /* 4186 * Don't let SMR zones operate without a free bucket. Force 4187 * a synchronize and re-use this one. We will only degrade 4188 * to a synchronize every bucket_size items rather than every 4189 * item if we fail to allocate a bucket. 4190 */ 4191 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) { 4192 if (bucket != NULL) 4193 bucket->ub_seq = smr_advance(zone->uz_smr); 4194 newbucket = bucket_alloc(zone, udata, M_NOWAIT); 4195 if (newbucket == NULL && bucket != NULL) { 4196 bucket_drain(zone, bucket); 4197 newbucket = bucket; 4198 bucket = NULL; 4199 } 4200 } else if (!bucketdisable) 4201 newbucket = bucket_alloc(zone, udata, M_NOWAIT); 4202 4203 if (bucket != NULL) 4204 zone_free_bucket(zone, bucket, udata, domain, itemdomain); 4205 4206 critical_enter(); 4207 if ((bucket = newbucket) == NULL) 4208 return (false); 4209 cache = &zone->uz_cpu[curcpu]; 4210 #ifdef NUMA 4211 /* 4212 * Check to see if we should be populating the cross bucket. If it 4213 * is already populated we will fall through and attempt to populate 4214 * the free bucket. 4215 */ 4216 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) { 4217 domain = PCPU_GET(domain); 4218 if (domain != itemdomain && 4219 cache->uc_crossbucket.ucb_bucket == NULL) { 4220 cache_bucket_load_cross(cache, bucket); 4221 return (true); 4222 } 4223 } 4224 #endif 4225 /* 4226 * We may have lost the race to fill the bucket or switched CPUs. 4227 */ 4228 if (cache->uc_freebucket.ucb_bucket != NULL) { 4229 critical_exit(); 4230 bucket_free(zone, bucket, udata); 4231 critical_enter(); 4232 } else 4233 cache_bucket_load_free(cache, bucket); 4234 4235 return (true); 4236 } 4237 4238 void 4239 uma_zfree_domain(uma_zone_t zone, void *item, void *udata) 4240 { 4241 4242 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 4243 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 4244 4245 CTR2(KTR_UMA, "uma_zfree_domain zone %s(%p)", zone->uz_name, zone); 4246 4247 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 4248 ("uma_zfree_domain: called with spinlock or critical section held")); 4249 4250 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 4251 if (item == NULL) 4252 return; 4253 zone_free_item(zone, item, udata, SKIP_NONE); 4254 } 4255 4256 static void 4257 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item) 4258 { 4259 uma_keg_t keg; 4260 uma_domain_t dom; 4261 int freei; 4262 4263 keg = zone->uz_keg; 4264 KEG_LOCK_ASSERT(keg, slab->us_domain); 4265 4266 /* Do we need to remove from any lists? */ 4267 dom = &keg->uk_domain[slab->us_domain]; 4268 if (slab->us_freecount + 1 == keg->uk_ipers) { 4269 LIST_REMOVE(slab, us_link); 4270 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); 4271 dom->ud_free_slabs++; 4272 } else if (slab->us_freecount == 0) { 4273 LIST_REMOVE(slab, us_link); 4274 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 4275 } 4276 4277 /* Slab management. */ 4278 freei = slab_item_index(slab, keg, item); 4279 BIT_SET(keg->uk_ipers, freei, &slab->us_free); 4280 slab->us_freecount++; 4281 4282 /* Keg statistics. */ 4283 dom->ud_free_items++; 4284 } 4285 4286 static void 4287 zone_release(void *arg, void **bucket, int cnt) 4288 { 4289 struct mtx *lock; 4290 uma_zone_t zone; 4291 uma_slab_t slab; 4292 uma_keg_t keg; 4293 uint8_t *mem; 4294 void *item; 4295 int i; 4296 4297 zone = arg; 4298 keg = zone->uz_keg; 4299 lock = NULL; 4300 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0)) 4301 lock = KEG_LOCK(keg, 0); 4302 for (i = 0; i < cnt; i++) { 4303 item = bucket[i]; 4304 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) { 4305 slab = vtoslab((vm_offset_t)item); 4306 } else { 4307 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 4308 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0) 4309 slab = hash_sfind(&keg->uk_hash, mem); 4310 else 4311 slab = (uma_slab_t)(mem + keg->uk_pgoff); 4312 } 4313 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) { 4314 if (lock != NULL) 4315 mtx_unlock(lock); 4316 lock = KEG_LOCK(keg, slab->us_domain); 4317 } 4318 slab_free_item(zone, slab, item); 4319 } 4320 if (lock != NULL) 4321 mtx_unlock(lock); 4322 } 4323 4324 /* 4325 * Frees a single item to any zone. 4326 * 4327 * Arguments: 4328 * zone The zone to free to 4329 * item The item we're freeing 4330 * udata User supplied data for the dtor 4331 * skip Skip dtors and finis 4332 */ 4333 static __noinline void 4334 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 4335 { 4336 4337 /* 4338 * If a free is sent directly to an SMR zone we have to 4339 * synchronize immediately because the item can instantly 4340 * be reallocated. This should only happen in degenerate 4341 * cases when no memory is available for per-cpu caches. 4342 */ 4343 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE) 4344 smr_synchronize(zone->uz_smr); 4345 4346 item_dtor(zone, item, zone->uz_size, udata, skip); 4347 4348 if (skip < SKIP_FINI && zone->uz_fini) 4349 zone->uz_fini(item, zone->uz_size); 4350 4351 zone->uz_release(zone->uz_arg, &item, 1); 4352 4353 if (skip & SKIP_CNT) 4354 return; 4355 4356 counter_u64_add(zone->uz_frees, 1); 4357 4358 if (zone->uz_max_items > 0) 4359 zone_free_limit(zone, 1); 4360 } 4361 4362 /* See uma.h */ 4363 int 4364 uma_zone_set_max(uma_zone_t zone, int nitems) 4365 { 4366 struct uma_bucket_zone *ubz; 4367 int count; 4368 4369 /* 4370 * XXX This can misbehave if the zone has any allocations with 4371 * no limit and a limit is imposed. There is currently no 4372 * way to clear a limit. 4373 */ 4374 ZONE_LOCK(zone); 4375 ubz = bucket_zone_max(zone, nitems); 4376 count = ubz != NULL ? ubz->ubz_entries : 0; 4377 zone->uz_bucket_size_max = zone->uz_bucket_size = count; 4378 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max) 4379 zone->uz_bucket_size_min = zone->uz_bucket_size_max; 4380 zone->uz_max_items = nitems; 4381 zone->uz_flags |= UMA_ZFLAG_LIMIT; 4382 zone_update_caches(zone); 4383 /* We may need to wake waiters. */ 4384 wakeup(&zone->uz_max_items); 4385 ZONE_UNLOCK(zone); 4386 4387 return (nitems); 4388 } 4389 4390 /* See uma.h */ 4391 void 4392 uma_zone_set_maxcache(uma_zone_t zone, int nitems) 4393 { 4394 struct uma_bucket_zone *ubz; 4395 int bpcpu; 4396 4397 ZONE_LOCK(zone); 4398 ubz = bucket_zone_max(zone, nitems); 4399 if (ubz != NULL) { 4400 bpcpu = 2; 4401 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0) 4402 /* Count the cross-domain bucket. */ 4403 bpcpu++; 4404 nitems -= ubz->ubz_entries * bpcpu * mp_ncpus; 4405 zone->uz_bucket_size_max = ubz->ubz_entries; 4406 } else { 4407 zone->uz_bucket_size_max = zone->uz_bucket_size = 0; 4408 } 4409 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max) 4410 zone->uz_bucket_size_min = zone->uz_bucket_size_max; 4411 zone->uz_bkt_max = nitems; 4412 ZONE_UNLOCK(zone); 4413 } 4414 4415 /* See uma.h */ 4416 int 4417 uma_zone_get_max(uma_zone_t zone) 4418 { 4419 int nitems; 4420 4421 nitems = atomic_load_64(&zone->uz_max_items); 4422 4423 return (nitems); 4424 } 4425 4426 /* See uma.h */ 4427 void 4428 uma_zone_set_warning(uma_zone_t zone, const char *warning) 4429 { 4430 4431 ZONE_ASSERT_COLD(zone); 4432 zone->uz_warning = warning; 4433 } 4434 4435 /* See uma.h */ 4436 void 4437 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) 4438 { 4439 4440 ZONE_ASSERT_COLD(zone); 4441 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); 4442 } 4443 4444 /* See uma.h */ 4445 int 4446 uma_zone_get_cur(uma_zone_t zone) 4447 { 4448 int64_t nitems; 4449 u_int i; 4450 4451 nitems = 0; 4452 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER) 4453 nitems = counter_u64_fetch(zone->uz_allocs) - 4454 counter_u64_fetch(zone->uz_frees); 4455 CPU_FOREACH(i) 4456 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) - 4457 atomic_load_64(&zone->uz_cpu[i].uc_frees); 4458 4459 return (nitems < 0 ? 0 : nitems); 4460 } 4461 4462 static uint64_t 4463 uma_zone_get_allocs(uma_zone_t zone) 4464 { 4465 uint64_t nitems; 4466 u_int i; 4467 4468 nitems = 0; 4469 if (zone->uz_allocs != EARLY_COUNTER) 4470 nitems = counter_u64_fetch(zone->uz_allocs); 4471 CPU_FOREACH(i) 4472 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs); 4473 4474 return (nitems); 4475 } 4476 4477 static uint64_t 4478 uma_zone_get_frees(uma_zone_t zone) 4479 { 4480 uint64_t nitems; 4481 u_int i; 4482 4483 nitems = 0; 4484 if (zone->uz_frees != EARLY_COUNTER) 4485 nitems = counter_u64_fetch(zone->uz_frees); 4486 CPU_FOREACH(i) 4487 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees); 4488 4489 return (nitems); 4490 } 4491 4492 #ifdef INVARIANTS 4493 /* Used only for KEG_ASSERT_COLD(). */ 4494 static uint64_t 4495 uma_keg_get_allocs(uma_keg_t keg) 4496 { 4497 uma_zone_t z; 4498 uint64_t nitems; 4499 4500 nitems = 0; 4501 LIST_FOREACH(z, &keg->uk_zones, uz_link) 4502 nitems += uma_zone_get_allocs(z); 4503 4504 return (nitems); 4505 } 4506 #endif 4507 4508 /* See uma.h */ 4509 void 4510 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 4511 { 4512 uma_keg_t keg; 4513 4514 KEG_GET(zone, keg); 4515 KEG_ASSERT_COLD(keg); 4516 keg->uk_init = uminit; 4517 } 4518 4519 /* See uma.h */ 4520 void 4521 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 4522 { 4523 uma_keg_t keg; 4524 4525 KEG_GET(zone, keg); 4526 KEG_ASSERT_COLD(keg); 4527 keg->uk_fini = fini; 4528 } 4529 4530 /* See uma.h */ 4531 void 4532 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 4533 { 4534 4535 ZONE_ASSERT_COLD(zone); 4536 zone->uz_init = zinit; 4537 } 4538 4539 /* See uma.h */ 4540 void 4541 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 4542 { 4543 4544 ZONE_ASSERT_COLD(zone); 4545 zone->uz_fini = zfini; 4546 } 4547 4548 /* See uma.h */ 4549 void 4550 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 4551 { 4552 uma_keg_t keg; 4553 4554 KEG_GET(zone, keg); 4555 KEG_ASSERT_COLD(keg); 4556 keg->uk_freef = freef; 4557 } 4558 4559 /* See uma.h */ 4560 void 4561 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 4562 { 4563 uma_keg_t keg; 4564 4565 KEG_GET(zone, keg); 4566 KEG_ASSERT_COLD(keg); 4567 keg->uk_allocf = allocf; 4568 } 4569 4570 /* See uma.h */ 4571 void 4572 uma_zone_set_smr(uma_zone_t zone, smr_t smr) 4573 { 4574 4575 ZONE_ASSERT_COLD(zone); 4576 4577 zone->uz_flags |= UMA_ZONE_SMR; 4578 zone->uz_smr = smr; 4579 zone_update_caches(zone); 4580 } 4581 4582 smr_t 4583 uma_zone_get_smr(uma_zone_t zone) 4584 { 4585 4586 return (zone->uz_smr); 4587 } 4588 4589 /* See uma.h */ 4590 void 4591 uma_zone_reserve(uma_zone_t zone, int items) 4592 { 4593 uma_keg_t keg; 4594 4595 KEG_GET(zone, keg); 4596 KEG_ASSERT_COLD(keg); 4597 keg->uk_reserve = items; 4598 } 4599 4600 /* See uma.h */ 4601 int 4602 uma_zone_reserve_kva(uma_zone_t zone, int count) 4603 { 4604 uma_keg_t keg; 4605 vm_offset_t kva; 4606 u_int pages; 4607 4608 KEG_GET(zone, keg); 4609 KEG_ASSERT_COLD(keg); 4610 ZONE_ASSERT_COLD(zone); 4611 4612 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera; 4613 4614 #ifdef UMA_MD_SMALL_ALLOC 4615 if (keg->uk_ppera > 1) { 4616 #else 4617 if (1) { 4618 #endif 4619 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE); 4620 if (kva == 0) 4621 return (0); 4622 } else 4623 kva = 0; 4624 4625 ZONE_LOCK(zone); 4626 MPASS(keg->uk_kva == 0); 4627 keg->uk_kva = kva; 4628 keg->uk_offset = 0; 4629 zone->uz_max_items = pages * keg->uk_ipers; 4630 #ifdef UMA_MD_SMALL_ALLOC 4631 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 4632 #else 4633 keg->uk_allocf = noobj_alloc; 4634 #endif 4635 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE; 4636 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE; 4637 zone_update_caches(zone); 4638 ZONE_UNLOCK(zone); 4639 4640 return (1); 4641 } 4642 4643 /* See uma.h */ 4644 void 4645 uma_prealloc(uma_zone_t zone, int items) 4646 { 4647 struct vm_domainset_iter di; 4648 uma_domain_t dom; 4649 uma_slab_t slab; 4650 uma_keg_t keg; 4651 int aflags, domain, slabs; 4652 4653 KEG_GET(zone, keg); 4654 slabs = howmany(items, keg->uk_ipers); 4655 while (slabs-- > 0) { 4656 aflags = M_NOWAIT; 4657 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 4658 &aflags); 4659 for (;;) { 4660 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK, 4661 aflags); 4662 if (slab != NULL) { 4663 dom = &keg->uk_domain[slab->us_domain]; 4664 /* 4665 * keg_alloc_slab() always returns a slab on the 4666 * partial list. 4667 */ 4668 LIST_REMOVE(slab, us_link); 4669 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, 4670 us_link); 4671 dom->ud_free_slabs++; 4672 KEG_UNLOCK(keg, slab->us_domain); 4673 break; 4674 } 4675 if (vm_domainset_iter_policy(&di, &domain) != 0) 4676 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask); 4677 } 4678 } 4679 } 4680 4681 /* See uma.h */ 4682 void 4683 uma_reclaim(int req) 4684 { 4685 4686 CTR0(KTR_UMA, "UMA: vm asked us to release pages!"); 4687 sx_xlock(&uma_reclaim_lock); 4688 bucket_enable(); 4689 4690 switch (req) { 4691 case UMA_RECLAIM_TRIM: 4692 zone_foreach(zone_trim, NULL); 4693 break; 4694 case UMA_RECLAIM_DRAIN: 4695 case UMA_RECLAIM_DRAIN_CPU: 4696 zone_foreach(zone_drain, NULL); 4697 if (req == UMA_RECLAIM_DRAIN_CPU) { 4698 pcpu_cache_drain_safe(NULL); 4699 zone_foreach(zone_drain, NULL); 4700 } 4701 break; 4702 default: 4703 panic("unhandled reclamation request %d", req); 4704 } 4705 4706 /* 4707 * Some slabs may have been freed but this zone will be visited early 4708 * we visit again so that we can free pages that are empty once other 4709 * zones are drained. We have to do the same for buckets. 4710 */ 4711 zone_drain(slabzones[0], NULL); 4712 zone_drain(slabzones[1], NULL); 4713 bucket_zone_drain(); 4714 sx_xunlock(&uma_reclaim_lock); 4715 } 4716 4717 static volatile int uma_reclaim_needed; 4718 4719 void 4720 uma_reclaim_wakeup(void) 4721 { 4722 4723 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0) 4724 wakeup(uma_reclaim); 4725 } 4726 4727 void 4728 uma_reclaim_worker(void *arg __unused) 4729 { 4730 4731 for (;;) { 4732 sx_xlock(&uma_reclaim_lock); 4733 while (atomic_load_int(&uma_reclaim_needed) == 0) 4734 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl", 4735 hz); 4736 sx_xunlock(&uma_reclaim_lock); 4737 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); 4738 uma_reclaim(UMA_RECLAIM_DRAIN_CPU); 4739 atomic_store_int(&uma_reclaim_needed, 0); 4740 /* Don't fire more than once per-second. */ 4741 pause("umarclslp", hz); 4742 } 4743 } 4744 4745 /* See uma.h */ 4746 void 4747 uma_zone_reclaim(uma_zone_t zone, int req) 4748 { 4749 4750 switch (req) { 4751 case UMA_RECLAIM_TRIM: 4752 zone_trim(zone, NULL); 4753 break; 4754 case UMA_RECLAIM_DRAIN: 4755 zone_drain(zone, NULL); 4756 break; 4757 case UMA_RECLAIM_DRAIN_CPU: 4758 pcpu_cache_drain_safe(zone); 4759 zone_drain(zone, NULL); 4760 break; 4761 default: 4762 panic("unhandled reclamation request %d", req); 4763 } 4764 } 4765 4766 /* See uma.h */ 4767 int 4768 uma_zone_exhausted(uma_zone_t zone) 4769 { 4770 4771 return (atomic_load_32(&zone->uz_sleepers) > 0); 4772 } 4773 4774 unsigned long 4775 uma_limit(void) 4776 { 4777 4778 return (uma_kmem_limit); 4779 } 4780 4781 void 4782 uma_set_limit(unsigned long limit) 4783 { 4784 4785 uma_kmem_limit = limit; 4786 } 4787 4788 unsigned long 4789 uma_size(void) 4790 { 4791 4792 return (atomic_load_long(&uma_kmem_total)); 4793 } 4794 4795 long 4796 uma_avail(void) 4797 { 4798 4799 return (uma_kmem_limit - uma_size()); 4800 } 4801 4802 #ifdef DDB 4803 /* 4804 * Generate statistics across both the zone and its per-cpu cache's. Return 4805 * desired statistics if the pointer is non-NULL for that statistic. 4806 * 4807 * Note: does not update the zone statistics, as it can't safely clear the 4808 * per-CPU cache statistic. 4809 * 4810 */ 4811 static void 4812 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp, 4813 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp) 4814 { 4815 uma_cache_t cache; 4816 uint64_t allocs, frees, sleeps, xdomain; 4817 int cachefree, cpu; 4818 4819 allocs = frees = sleeps = xdomain = 0; 4820 cachefree = 0; 4821 CPU_FOREACH(cpu) { 4822 cache = &z->uz_cpu[cpu]; 4823 cachefree += cache->uc_allocbucket.ucb_cnt; 4824 cachefree += cache->uc_freebucket.ucb_cnt; 4825 xdomain += cache->uc_crossbucket.ucb_cnt; 4826 cachefree += cache->uc_crossbucket.ucb_cnt; 4827 allocs += cache->uc_allocs; 4828 frees += cache->uc_frees; 4829 } 4830 allocs += counter_u64_fetch(z->uz_allocs); 4831 frees += counter_u64_fetch(z->uz_frees); 4832 sleeps += z->uz_sleeps; 4833 xdomain += z->uz_xdomain; 4834 if (cachefreep != NULL) 4835 *cachefreep = cachefree; 4836 if (allocsp != NULL) 4837 *allocsp = allocs; 4838 if (freesp != NULL) 4839 *freesp = frees; 4840 if (sleepsp != NULL) 4841 *sleepsp = sleeps; 4842 if (xdomainp != NULL) 4843 *xdomainp = xdomain; 4844 } 4845 #endif /* DDB */ 4846 4847 static int 4848 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 4849 { 4850 uma_keg_t kz; 4851 uma_zone_t z; 4852 int count; 4853 4854 count = 0; 4855 rw_rlock(&uma_rwlock); 4856 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4857 LIST_FOREACH(z, &kz->uk_zones, uz_link) 4858 count++; 4859 } 4860 LIST_FOREACH(z, &uma_cachezones, uz_link) 4861 count++; 4862 4863 rw_runlock(&uma_rwlock); 4864 return (sysctl_handle_int(oidp, &count, 0, req)); 4865 } 4866 4867 static void 4868 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf, 4869 struct uma_percpu_stat *ups, bool internal) 4870 { 4871 uma_zone_domain_t zdom; 4872 uma_cache_t cache; 4873 int i; 4874 4875 4876 for (i = 0; i < vm_ndomains; i++) { 4877 zdom = &z->uz_domain[i]; 4878 uth->uth_zone_free += zdom->uzd_nitems; 4879 } 4880 uth->uth_allocs = counter_u64_fetch(z->uz_allocs); 4881 uth->uth_frees = counter_u64_fetch(z->uz_frees); 4882 uth->uth_fails = counter_u64_fetch(z->uz_fails); 4883 uth->uth_sleeps = z->uz_sleeps; 4884 uth->uth_xdomain = z->uz_xdomain; 4885 4886 /* 4887 * While it is not normally safe to access the cache bucket pointers 4888 * while not on the CPU that owns the cache, we only allow the pointers 4889 * to be exchanged without the zone lock held, not invalidated, so 4890 * accept the possible race associated with bucket exchange during 4891 * monitoring. Use atomic_load_ptr() to ensure that the bucket pointers 4892 * are loaded only once. 4893 */ 4894 for (i = 0; i < mp_maxid + 1; i++) { 4895 bzero(&ups[i], sizeof(*ups)); 4896 if (internal || CPU_ABSENT(i)) 4897 continue; 4898 cache = &z->uz_cpu[i]; 4899 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt; 4900 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt; 4901 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt; 4902 ups[i].ups_allocs = cache->uc_allocs; 4903 ups[i].ups_frees = cache->uc_frees; 4904 } 4905 } 4906 4907 static int 4908 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 4909 { 4910 struct uma_stream_header ush; 4911 struct uma_type_header uth; 4912 struct uma_percpu_stat *ups; 4913 struct sbuf sbuf; 4914 uma_keg_t kz; 4915 uma_zone_t z; 4916 uint64_t items; 4917 uint32_t kfree, pages; 4918 int count, error, i; 4919 4920 error = sysctl_wire_old_buffer(req, 0); 4921 if (error != 0) 4922 return (error); 4923 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 4924 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); 4925 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK); 4926 4927 count = 0; 4928 rw_rlock(&uma_rwlock); 4929 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4930 LIST_FOREACH(z, &kz->uk_zones, uz_link) 4931 count++; 4932 } 4933 4934 LIST_FOREACH(z, &uma_cachezones, uz_link) 4935 count++; 4936 4937 /* 4938 * Insert stream header. 4939 */ 4940 bzero(&ush, sizeof(ush)); 4941 ush.ush_version = UMA_STREAM_VERSION; 4942 ush.ush_maxcpus = (mp_maxid + 1); 4943 ush.ush_count = count; 4944 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 4945 4946 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4947 kfree = pages = 0; 4948 for (i = 0; i < vm_ndomains; i++) { 4949 kfree += kz->uk_domain[i].ud_free_items; 4950 pages += kz->uk_domain[i].ud_pages; 4951 } 4952 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 4953 bzero(&uth, sizeof(uth)); 4954 ZONE_LOCK(z); 4955 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 4956 uth.uth_align = kz->uk_align; 4957 uth.uth_size = kz->uk_size; 4958 uth.uth_rsize = kz->uk_rsize; 4959 if (z->uz_max_items > 0) { 4960 items = UZ_ITEMS_COUNT(z->uz_items); 4961 uth.uth_pages = (items / kz->uk_ipers) * 4962 kz->uk_ppera; 4963 } else 4964 uth.uth_pages = pages; 4965 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) * 4966 kz->uk_ppera; 4967 uth.uth_limit = z->uz_max_items; 4968 uth.uth_keg_free = kfree; 4969 4970 /* 4971 * A zone is secondary is it is not the first entry 4972 * on the keg's zone list. 4973 */ 4974 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 4975 (LIST_FIRST(&kz->uk_zones) != z)) 4976 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 4977 uma_vm_zone_stats(&uth, z, &sbuf, ups, 4978 kz->uk_flags & UMA_ZFLAG_INTERNAL); 4979 ZONE_UNLOCK(z); 4980 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 4981 for (i = 0; i < mp_maxid + 1; i++) 4982 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 4983 } 4984 } 4985 LIST_FOREACH(z, &uma_cachezones, uz_link) { 4986 bzero(&uth, sizeof(uth)); 4987 ZONE_LOCK(z); 4988 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 4989 uth.uth_size = z->uz_size; 4990 uma_vm_zone_stats(&uth, z, &sbuf, ups, false); 4991 ZONE_UNLOCK(z); 4992 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 4993 for (i = 0; i < mp_maxid + 1; i++) 4994 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 4995 } 4996 4997 rw_runlock(&uma_rwlock); 4998 error = sbuf_finish(&sbuf); 4999 sbuf_delete(&sbuf); 5000 free(ups, M_TEMP); 5001 return (error); 5002 } 5003 5004 int 5005 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 5006 { 5007 uma_zone_t zone = *(uma_zone_t *)arg1; 5008 int error, max; 5009 5010 max = uma_zone_get_max(zone); 5011 error = sysctl_handle_int(oidp, &max, 0, req); 5012 if (error || !req->newptr) 5013 return (error); 5014 5015 uma_zone_set_max(zone, max); 5016 5017 return (0); 5018 } 5019 5020 int 5021 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 5022 { 5023 uma_zone_t zone; 5024 int cur; 5025 5026 /* 5027 * Some callers want to add sysctls for global zones that 5028 * may not yet exist so they pass a pointer to a pointer. 5029 */ 5030 if (arg2 == 0) 5031 zone = *(uma_zone_t *)arg1; 5032 else 5033 zone = arg1; 5034 cur = uma_zone_get_cur(zone); 5035 return (sysctl_handle_int(oidp, &cur, 0, req)); 5036 } 5037 5038 static int 5039 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS) 5040 { 5041 uma_zone_t zone = arg1; 5042 uint64_t cur; 5043 5044 cur = uma_zone_get_allocs(zone); 5045 return (sysctl_handle_64(oidp, &cur, 0, req)); 5046 } 5047 5048 static int 5049 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS) 5050 { 5051 uma_zone_t zone = arg1; 5052 uint64_t cur; 5053 5054 cur = uma_zone_get_frees(zone); 5055 return (sysctl_handle_64(oidp, &cur, 0, req)); 5056 } 5057 5058 static int 5059 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS) 5060 { 5061 struct sbuf sbuf; 5062 uma_zone_t zone = arg1; 5063 int error; 5064 5065 sbuf_new_for_sysctl(&sbuf, NULL, 0, req); 5066 if (zone->uz_flags != 0) 5067 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS); 5068 else 5069 sbuf_printf(&sbuf, "0"); 5070 error = sbuf_finish(&sbuf); 5071 sbuf_delete(&sbuf); 5072 5073 return (error); 5074 } 5075 5076 static int 5077 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS) 5078 { 5079 uma_keg_t keg = arg1; 5080 int avail, effpct, total; 5081 5082 total = keg->uk_ppera * PAGE_SIZE; 5083 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0) 5084 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize; 5085 /* 5086 * We consider the client's requested size and alignment here, not the 5087 * real size determination uk_rsize, because we also adjust the real 5088 * size for internal implementation reasons (max bitset size). 5089 */ 5090 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1); 5091 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0) 5092 avail *= mp_maxid + 1; 5093 effpct = 100 * avail / total; 5094 return (sysctl_handle_int(oidp, &effpct, 0, req)); 5095 } 5096 5097 static int 5098 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS) 5099 { 5100 uma_zone_t zone = arg1; 5101 uint64_t cur; 5102 5103 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items)); 5104 return (sysctl_handle_64(oidp, &cur, 0, req)); 5105 } 5106 5107 #ifdef INVARIANTS 5108 static uma_slab_t 5109 uma_dbg_getslab(uma_zone_t zone, void *item) 5110 { 5111 uma_slab_t slab; 5112 uma_keg_t keg; 5113 uint8_t *mem; 5114 5115 /* 5116 * It is safe to return the slab here even though the 5117 * zone is unlocked because the item's allocation state 5118 * essentially holds a reference. 5119 */ 5120 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 5121 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 5122 return (NULL); 5123 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB) 5124 return (vtoslab((vm_offset_t)mem)); 5125 keg = zone->uz_keg; 5126 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0) 5127 return ((uma_slab_t)(mem + keg->uk_pgoff)); 5128 KEG_LOCK(keg, 0); 5129 slab = hash_sfind(&keg->uk_hash, mem); 5130 KEG_UNLOCK(keg, 0); 5131 5132 return (slab); 5133 } 5134 5135 static bool 5136 uma_dbg_zskip(uma_zone_t zone, void *mem) 5137 { 5138 5139 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0) 5140 return (true); 5141 5142 return (uma_dbg_kskip(zone->uz_keg, mem)); 5143 } 5144 5145 static bool 5146 uma_dbg_kskip(uma_keg_t keg, void *mem) 5147 { 5148 uintptr_t idx; 5149 5150 if (dbg_divisor == 0) 5151 return (true); 5152 5153 if (dbg_divisor == 1) 5154 return (false); 5155 5156 idx = (uintptr_t)mem >> PAGE_SHIFT; 5157 if (keg->uk_ipers > 1) { 5158 idx *= keg->uk_ipers; 5159 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize; 5160 } 5161 5162 if ((idx / dbg_divisor) * dbg_divisor != idx) { 5163 counter_u64_add(uma_skip_cnt, 1); 5164 return (true); 5165 } 5166 counter_u64_add(uma_dbg_cnt, 1); 5167 5168 return (false); 5169 } 5170 5171 /* 5172 * Set up the slab's freei data such that uma_dbg_free can function. 5173 * 5174 */ 5175 static void 5176 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) 5177 { 5178 uma_keg_t keg; 5179 int freei; 5180 5181 if (slab == NULL) { 5182 slab = uma_dbg_getslab(zone, item); 5183 if (slab == NULL) 5184 panic("uma: item %p did not belong to zone %s\n", 5185 item, zone->uz_name); 5186 } 5187 keg = zone->uz_keg; 5188 freei = slab_item_index(slab, keg, item); 5189 5190 if (BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg))) 5191 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n", 5192 item, zone, zone->uz_name, slab, freei); 5193 BIT_SET_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)); 5194 } 5195 5196 /* 5197 * Verifies freed addresses. Checks for alignment, valid slab membership 5198 * and duplicate frees. 5199 * 5200 */ 5201 static void 5202 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) 5203 { 5204 uma_keg_t keg; 5205 int freei; 5206 5207 if (slab == NULL) { 5208 slab = uma_dbg_getslab(zone, item); 5209 if (slab == NULL) 5210 panic("uma: Freed item %p did not belong to zone %s\n", 5211 item, zone->uz_name); 5212 } 5213 keg = zone->uz_keg; 5214 freei = slab_item_index(slab, keg, item); 5215 5216 if (freei >= keg->uk_ipers) 5217 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n", 5218 item, zone, zone->uz_name, slab, freei); 5219 5220 if (slab_item(slab, keg, freei) != item) 5221 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n", 5222 item, zone, zone->uz_name, slab, freei); 5223 5224 if (!BIT_ISSET(keg->uk_ipers, freei, slab_dbg_bits(slab, keg))) 5225 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n", 5226 item, zone, zone->uz_name, slab, freei); 5227 5228 BIT_CLR_ATOMIC(keg->uk_ipers, freei, slab_dbg_bits(slab, keg)); 5229 } 5230 #endif /* INVARIANTS */ 5231 5232 #ifdef DDB 5233 static int64_t 5234 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used, 5235 uint64_t *sleeps, long *cachefree, uint64_t *xdomain) 5236 { 5237 uint64_t frees; 5238 int i; 5239 5240 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 5241 *allocs = counter_u64_fetch(z->uz_allocs); 5242 frees = counter_u64_fetch(z->uz_frees); 5243 *sleeps = z->uz_sleeps; 5244 *cachefree = 0; 5245 *xdomain = 0; 5246 } else 5247 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps, 5248 xdomain); 5249 for (i = 0; i < vm_ndomains; i++) { 5250 *cachefree += z->uz_domain[i].uzd_nitems; 5251 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 5252 (LIST_FIRST(&kz->uk_zones) != z))) 5253 *cachefree += kz->uk_domain[i].ud_free_items; 5254 } 5255 *used = *allocs - frees; 5256 return (((int64_t)*used + *cachefree) * kz->uk_size); 5257 } 5258 5259 DB_SHOW_COMMAND(uma, db_show_uma) 5260 { 5261 const char *fmt_hdr, *fmt_entry; 5262 uma_keg_t kz; 5263 uma_zone_t z; 5264 uint64_t allocs, used, sleeps, xdomain; 5265 long cachefree; 5266 /* variables for sorting */ 5267 uma_keg_t cur_keg; 5268 uma_zone_t cur_zone, last_zone; 5269 int64_t cur_size, last_size, size; 5270 int ties; 5271 5272 /* /i option produces machine-parseable CSV output */ 5273 if (modif[0] == 'i') { 5274 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n"; 5275 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n"; 5276 } else { 5277 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n"; 5278 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n"; 5279 } 5280 5281 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests", 5282 "Sleeps", "Bucket", "Total Mem", "XFree"); 5283 5284 /* Sort the zones with largest size first. */ 5285 last_zone = NULL; 5286 last_size = INT64_MAX; 5287 for (;;) { 5288 cur_zone = NULL; 5289 cur_size = -1; 5290 ties = 0; 5291 LIST_FOREACH(kz, &uma_kegs, uk_link) { 5292 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 5293 /* 5294 * In the case of size ties, print out zones 5295 * in the order they are encountered. That is, 5296 * when we encounter the most recently output 5297 * zone, we have already printed all preceding 5298 * ties, and we must print all following ties. 5299 */ 5300 if (z == last_zone) { 5301 ties = 1; 5302 continue; 5303 } 5304 size = get_uma_stats(kz, z, &allocs, &used, 5305 &sleeps, &cachefree, &xdomain); 5306 if (size > cur_size && size < last_size + ties) 5307 { 5308 cur_size = size; 5309 cur_zone = z; 5310 cur_keg = kz; 5311 } 5312 } 5313 } 5314 if (cur_zone == NULL) 5315 break; 5316 5317 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used, 5318 &sleeps, &cachefree, &xdomain); 5319 db_printf(fmt_entry, cur_zone->uz_name, 5320 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree, 5321 (uintmax_t)allocs, (uintmax_t)sleeps, 5322 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size, 5323 xdomain); 5324 5325 if (db_pager_quit) 5326 return; 5327 last_zone = cur_zone; 5328 last_size = cur_size; 5329 } 5330 } 5331 5332 DB_SHOW_COMMAND(umacache, db_show_umacache) 5333 { 5334 uma_zone_t z; 5335 uint64_t allocs, frees; 5336 long cachefree; 5337 int i; 5338 5339 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 5340 "Requests", "Bucket"); 5341 LIST_FOREACH(z, &uma_cachezones, uz_link) { 5342 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL); 5343 for (i = 0; i < vm_ndomains; i++) 5344 cachefree += z->uz_domain[i].uzd_nitems; 5345 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n", 5346 z->uz_name, (uintmax_t)z->uz_size, 5347 (intmax_t)(allocs - frees), cachefree, 5348 (uintmax_t)allocs, z->uz_bucket_size); 5349 if (db_pager_quit) 5350 return; 5351 } 5352 } 5353 #endif /* DDB */ 5354