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