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/smp.h> 79 #include <sys/taskqueue.h> 80 #include <sys/vmmeter.h> 81 82 #include <vm/vm.h> 83 #include <vm/vm_domainset.h> 84 #include <vm/vm_object.h> 85 #include <vm/vm_page.h> 86 #include <vm/vm_pageout.h> 87 #include <vm/vm_param.h> 88 #include <vm/vm_phys.h> 89 #include <vm/vm_pagequeue.h> 90 #include <vm/vm_map.h> 91 #include <vm/vm_kern.h> 92 #include <vm/vm_extern.h> 93 #include <vm/uma.h> 94 #include <vm/uma_int.h> 95 #include <vm/uma_dbg.h> 96 97 #include <ddb/ddb.h> 98 99 #ifdef DEBUG_MEMGUARD 100 #include <vm/memguard.h> 101 #endif 102 103 /* 104 * This is the zone and keg from which all zones are spawned. 105 */ 106 static uma_zone_t kegs; 107 static uma_zone_t zones; 108 109 /* This is the zone from which all offpage uma_slab_ts are allocated. */ 110 static uma_zone_t slabzone; 111 112 /* 113 * The initial hash tables come out of this zone so they can be allocated 114 * prior to malloc coming up. 115 */ 116 static uma_zone_t hashzone; 117 118 /* The boot-time adjusted value for cache line alignment. */ 119 int uma_align_cache = 64 - 1; 120 121 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 122 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc"); 123 124 /* 125 * Are we allowed to allocate buckets? 126 */ 127 static int bucketdisable = 1; 128 129 /* Linked list of all kegs in the system */ 130 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 131 132 /* Linked list of all cache-only zones in the system */ 133 static LIST_HEAD(,uma_zone) uma_cachezones = 134 LIST_HEAD_INITIALIZER(uma_cachezones); 135 136 /* This RW lock protects the keg list */ 137 static struct rwlock_padalign __exclusive_cache_line uma_rwlock; 138 139 /* 140 * Pointer and counter to pool of pages, that is preallocated at 141 * startup to bootstrap UMA. 142 */ 143 static char *bootmem; 144 static int boot_pages; 145 146 static struct sx uma_reclaim_lock; 147 148 /* 149 * kmem soft limit, initialized by uma_set_limit(). Ensure that early 150 * allocations don't trigger a wakeup of the reclaim thread. 151 */ 152 unsigned long uma_kmem_limit = LONG_MAX; 153 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0, 154 "UMA kernel memory soft limit"); 155 unsigned long uma_kmem_total; 156 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0, 157 "UMA kernel memory usage"); 158 159 /* Is the VM done starting up? */ 160 static enum { BOOT_COLD = 0, BOOT_STRAPPED, BOOT_PAGEALLOC, BOOT_BUCKETS, 161 BOOT_RUNNING } booted = BOOT_COLD; 162 163 /* 164 * This is the handle used to schedule events that need to happen 165 * outside of the allocation fast path. 166 */ 167 static struct callout uma_callout; 168 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 169 170 /* 171 * This structure is passed as the zone ctor arg so that I don't have to create 172 * a special allocation function just for zones. 173 */ 174 struct uma_zctor_args { 175 const char *name; 176 size_t size; 177 uma_ctor ctor; 178 uma_dtor dtor; 179 uma_init uminit; 180 uma_fini fini; 181 uma_import import; 182 uma_release release; 183 void *arg; 184 uma_keg_t keg; 185 int align; 186 uint32_t flags; 187 }; 188 189 struct uma_kctor_args { 190 uma_zone_t zone; 191 size_t size; 192 uma_init uminit; 193 uma_fini fini; 194 int align; 195 uint32_t flags; 196 }; 197 198 struct uma_bucket_zone { 199 uma_zone_t ubz_zone; 200 char *ubz_name; 201 int ubz_entries; /* Number of items it can hold. */ 202 int ubz_maxsize; /* Maximum allocation size per-item. */ 203 }; 204 205 /* 206 * Compute the actual number of bucket entries to pack them in power 207 * of two sizes for more efficient space utilization. 208 */ 209 #define BUCKET_SIZE(n) \ 210 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 211 212 #define BUCKET_MAX BUCKET_SIZE(256) 213 #define BUCKET_MIN BUCKET_SIZE(4) 214 215 struct uma_bucket_zone bucket_zones[] = { 216 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 217 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 218 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 219 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 220 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 221 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 222 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 223 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 224 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 225 { NULL, NULL, 0} 226 }; 227 228 /* 229 * Flags and enumerations to be passed to internal functions. 230 */ 231 enum zfreeskip { 232 SKIP_NONE = 0, 233 SKIP_CNT = 0x00000001, 234 SKIP_DTOR = 0x00010000, 235 SKIP_FINI = 0x00020000, 236 }; 237 238 /* Prototypes.. */ 239 240 int uma_startup_count(int); 241 void uma_startup(void *, int); 242 void uma_startup1(void); 243 void uma_startup2(void); 244 245 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 246 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 247 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 248 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); 249 static void page_free(void *, vm_size_t, uint8_t); 250 static void pcpu_page_free(void *, vm_size_t, uint8_t); 251 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int); 252 static void cache_drain(uma_zone_t); 253 static void bucket_drain(uma_zone_t, uma_bucket_t); 254 static void bucket_cache_reclaim(uma_zone_t zone, bool); 255 static int keg_ctor(void *, int, void *, int); 256 static void keg_dtor(void *, int, void *); 257 static int zone_ctor(void *, int, void *, int); 258 static void zone_dtor(void *, int, void *); 259 static int zero_init(void *, int, int); 260 static void keg_small_init(uma_keg_t keg); 261 static void keg_large_init(uma_keg_t keg); 262 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *); 263 static void zone_timeout(uma_zone_t zone, void *); 264 static int hash_alloc(struct uma_hash *, u_int); 265 static int hash_expand(struct uma_hash *, struct uma_hash *); 266 static void hash_free(struct uma_hash *hash); 267 static void uma_timeout(void *); 268 static void uma_startup3(void); 269 static void *zone_alloc_item(uma_zone_t, void *, int, int); 270 static void *zone_alloc_item_locked(uma_zone_t, void *, int, int); 271 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 272 static void bucket_enable(void); 273 static void bucket_init(void); 274 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 275 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 276 static void bucket_zone_drain(void); 277 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int); 278 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 279 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item); 280 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 281 uma_fini fini, int align, uint32_t flags); 282 static int zone_import(void *, void **, int, int, int); 283 static void zone_release(void *, void **, int); 284 static void uma_zero_item(void *, uma_zone_t); 285 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int); 286 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int); 287 288 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 289 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 290 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS); 291 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS); 292 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS); 293 294 #ifdef INVARIANTS 295 static bool uma_dbg_kskip(uma_keg_t keg, void *mem); 296 static bool uma_dbg_zskip(uma_zone_t zone, void *mem); 297 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item); 298 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item); 299 300 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0, 301 "Memory allocation debugging"); 302 303 static u_int dbg_divisor = 1; 304 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor, 305 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0, 306 "Debug & thrash every this item in memory allocator"); 307 308 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER; 309 static counter_u64_t uma_skip_cnt = EARLY_COUNTER; 310 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD, 311 &uma_dbg_cnt, "memory items debugged"); 312 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD, 313 &uma_skip_cnt, "memory items skipped, not debugged"); 314 #endif 315 316 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 317 318 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW, 0, "Universal Memory Allocator"); 319 320 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 321 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 322 323 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 324 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 325 326 static int zone_warnings = 1; 327 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, 328 "Warn when UMA zones becomes full"); 329 330 /* 331 * This routine checks to see whether or not it's safe to enable buckets. 332 */ 333 static void 334 bucket_enable(void) 335 { 336 bucketdisable = vm_page_count_min(); 337 } 338 339 /* 340 * Initialize bucket_zones, the array of zones of buckets of various sizes. 341 * 342 * For each zone, calculate the memory required for each bucket, consisting 343 * of the header and an array of pointers. 344 */ 345 static void 346 bucket_init(void) 347 { 348 struct uma_bucket_zone *ubz; 349 int size; 350 351 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 352 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 353 size += sizeof(void *) * ubz->ubz_entries; 354 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 355 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 356 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET | UMA_ZONE_NUMA); 357 } 358 } 359 360 /* 361 * Given a desired number of entries for a bucket, return the zone from which 362 * to allocate the bucket. 363 */ 364 static struct uma_bucket_zone * 365 bucket_zone_lookup(int entries) 366 { 367 struct uma_bucket_zone *ubz; 368 369 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 370 if (ubz->ubz_entries >= entries) 371 return (ubz); 372 ubz--; 373 return (ubz); 374 } 375 376 static struct uma_bucket_zone * 377 bucket_zone_max(uma_zone_t zone, int nitems) 378 { 379 struct uma_bucket_zone *ubz; 380 int bpcpu; 381 382 bpcpu = 2; 383 #ifdef UMA_XDOMAIN 384 if ((zone->uz_flags & UMA_ZONE_NUMA) != 0) 385 /* Count the cross-domain bucket. */ 386 bpcpu++; 387 #endif 388 389 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 390 if (ubz->ubz_entries * bpcpu * mp_ncpus > nitems) 391 break; 392 if (ubz == &bucket_zones[0]) 393 ubz = NULL; 394 else 395 ubz--; 396 return (ubz); 397 } 398 399 static int 400 bucket_select(int size) 401 { 402 struct uma_bucket_zone *ubz; 403 404 ubz = &bucket_zones[0]; 405 if (size > ubz->ubz_maxsize) 406 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 407 408 for (; ubz->ubz_entries != 0; ubz++) 409 if (ubz->ubz_maxsize < size) 410 break; 411 ubz--; 412 return (ubz->ubz_entries); 413 } 414 415 static uma_bucket_t 416 bucket_alloc(uma_zone_t zone, void *udata, int flags) 417 { 418 struct uma_bucket_zone *ubz; 419 uma_bucket_t bucket; 420 421 /* 422 * This is to stop us from allocating per cpu buckets while we're 423 * running out of vm.boot_pages. Otherwise, we would exhaust the 424 * boot pages. This also prevents us from allocating buckets in 425 * low memory situations. 426 */ 427 if (bucketdisable) 428 return (NULL); 429 /* 430 * To limit bucket recursion we store the original zone flags 431 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 432 * NOVM flag to persist even through deep recursions. We also 433 * store ZFLAG_BUCKET once we have recursed attempting to allocate 434 * a bucket for a bucket zone so we do not allow infinite bucket 435 * recursion. This cookie will even persist to frees of unused 436 * buckets via the allocation path or bucket allocations in the 437 * free path. 438 */ 439 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 440 udata = (void *)(uintptr_t)zone->uz_flags; 441 else { 442 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 443 return (NULL); 444 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 445 } 446 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 447 flags |= M_NOVM; 448 ubz = bucket_zone_lookup(zone->uz_bucket_size); 449 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 450 ubz++; 451 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 452 if (bucket) { 453 #ifdef INVARIANTS 454 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 455 #endif 456 bucket->ub_cnt = 0; 457 bucket->ub_entries = ubz->ubz_entries; 458 } 459 460 return (bucket); 461 } 462 463 static void 464 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 465 { 466 struct uma_bucket_zone *ubz; 467 468 KASSERT(bucket->ub_cnt == 0, 469 ("bucket_free: Freeing a non free bucket.")); 470 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 471 udata = (void *)(uintptr_t)zone->uz_flags; 472 ubz = bucket_zone_lookup(bucket->ub_entries); 473 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 474 } 475 476 static void 477 bucket_zone_drain(void) 478 { 479 struct uma_bucket_zone *ubz; 480 481 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 482 uma_zone_reclaim(ubz->ubz_zone, UMA_RECLAIM_DRAIN); 483 } 484 485 /* 486 * Attempt to satisfy an allocation by retrieving a full bucket from one of the 487 * zone's caches. 488 */ 489 static uma_bucket_t 490 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom) 491 { 492 uma_bucket_t bucket; 493 494 ZONE_LOCK_ASSERT(zone); 495 496 if ((bucket = TAILQ_FIRST(&zdom->uzd_buckets)) != NULL) { 497 MPASS(zdom->uzd_nitems >= bucket->ub_cnt); 498 TAILQ_REMOVE(&zdom->uzd_buckets, bucket, ub_link); 499 zdom->uzd_nitems -= bucket->ub_cnt; 500 if (zdom->uzd_imin > zdom->uzd_nitems) 501 zdom->uzd_imin = zdom->uzd_nitems; 502 zone->uz_bkt_count -= bucket->ub_cnt; 503 } 504 return (bucket); 505 } 506 507 /* 508 * Insert a full bucket into the specified cache. The "ws" parameter indicates 509 * whether the bucket's contents should be counted as part of the zone's working 510 * set. 511 */ 512 static void 513 zone_put_bucket(uma_zone_t zone, uma_zone_domain_t zdom, uma_bucket_t bucket, 514 const bool ws) 515 { 516 517 ZONE_LOCK_ASSERT(zone); 518 KASSERT(!ws || zone->uz_bkt_count < zone->uz_bkt_max, 519 ("%s: zone %p overflow", __func__, zone)); 520 521 if (ws) 522 TAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link); 523 else 524 TAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link); 525 zdom->uzd_nitems += bucket->ub_cnt; 526 if (ws && zdom->uzd_imax < zdom->uzd_nitems) 527 zdom->uzd_imax = zdom->uzd_nitems; 528 zone->uz_bkt_count += bucket->ub_cnt; 529 } 530 531 static void 532 zone_log_warning(uma_zone_t zone) 533 { 534 static const struct timeval warninterval = { 300, 0 }; 535 536 if (!zone_warnings || zone->uz_warning == NULL) 537 return; 538 539 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 540 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 541 } 542 543 static inline void 544 zone_maxaction(uma_zone_t zone) 545 { 546 547 if (zone->uz_maxaction.ta_func != NULL) 548 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); 549 } 550 551 /* 552 * Routine called by timeout which is used to fire off some time interval 553 * based calculations. (stats, hash size, etc.) 554 * 555 * Arguments: 556 * arg Unused 557 * 558 * Returns: 559 * Nothing 560 */ 561 static void 562 uma_timeout(void *unused) 563 { 564 bucket_enable(); 565 zone_foreach(zone_timeout, NULL); 566 567 /* Reschedule this event */ 568 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 569 } 570 571 /* 572 * Update the working set size estimate for the zone's bucket cache. 573 * The constants chosen here are somewhat arbitrary. With an update period of 574 * 20s (UMA_TIMEOUT), this estimate is dominated by zone activity over the 575 * last 100s. 576 */ 577 static void 578 zone_domain_update_wss(uma_zone_domain_t zdom) 579 { 580 long wss; 581 582 MPASS(zdom->uzd_imax >= zdom->uzd_imin); 583 wss = zdom->uzd_imax - zdom->uzd_imin; 584 zdom->uzd_imax = zdom->uzd_imin = zdom->uzd_nitems; 585 zdom->uzd_wss = (4 * wss + zdom->uzd_wss) / 5; 586 } 587 588 /* 589 * Routine to perform timeout driven calculations. This expands the 590 * hashes and does per cpu statistics aggregation. 591 * 592 * Returns nothing. 593 */ 594 static void 595 zone_timeout(uma_zone_t zone, void *unused) 596 { 597 uma_keg_t keg; 598 u_int slabs; 599 600 if ((zone->uz_flags & UMA_ZONE_HASH) == 0) 601 goto update_wss; 602 603 keg = zone->uz_keg; 604 KEG_LOCK(keg); 605 /* 606 * Expand the keg hash table. 607 * 608 * This is done if the number of slabs is larger than the hash size. 609 * What I'm trying to do here is completely reduce collisions. This 610 * may be a little aggressive. Should I allow for two collisions max? 611 */ 612 if (keg->uk_flags & UMA_ZONE_HASH && 613 (slabs = keg->uk_pages / keg->uk_ppera) > 614 keg->uk_hash.uh_hashsize) { 615 struct uma_hash newhash; 616 struct uma_hash oldhash; 617 int ret; 618 619 /* 620 * This is so involved because allocating and freeing 621 * while the keg lock is held will lead to deadlock. 622 * I have to do everything in stages and check for 623 * races. 624 */ 625 KEG_UNLOCK(keg); 626 ret = hash_alloc(&newhash, 1 << fls(slabs)); 627 KEG_LOCK(keg); 628 if (ret) { 629 if (hash_expand(&keg->uk_hash, &newhash)) { 630 oldhash = keg->uk_hash; 631 keg->uk_hash = newhash; 632 } else 633 oldhash = newhash; 634 635 KEG_UNLOCK(keg); 636 hash_free(&oldhash); 637 return; 638 } 639 } 640 KEG_UNLOCK(keg); 641 642 update_wss: 643 ZONE_LOCK(zone); 644 for (int i = 0; i < vm_ndomains; i++) 645 zone_domain_update_wss(&zone->uz_domain[i]); 646 ZONE_UNLOCK(zone); 647 } 648 649 /* 650 * Allocate and zero fill the next sized hash table from the appropriate 651 * backing store. 652 * 653 * Arguments: 654 * hash A new hash structure with the old hash size in uh_hashsize 655 * 656 * Returns: 657 * 1 on success and 0 on failure. 658 */ 659 static int 660 hash_alloc(struct uma_hash *hash, u_int size) 661 { 662 size_t alloc; 663 664 KASSERT(powerof2(size), ("hash size must be power of 2")); 665 if (size > UMA_HASH_SIZE_INIT) { 666 hash->uh_hashsize = size; 667 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 668 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT); 669 } else { 670 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 671 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 672 UMA_ANYDOMAIN, M_WAITOK); 673 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 674 } 675 if (hash->uh_slab_hash) { 676 bzero(hash->uh_slab_hash, alloc); 677 hash->uh_hashmask = hash->uh_hashsize - 1; 678 return (1); 679 } 680 681 return (0); 682 } 683 684 /* 685 * Expands the hash table for HASH zones. This is done from zone_timeout 686 * to reduce collisions. This must not be done in the regular allocation 687 * path, otherwise, we can recurse on the vm while allocating pages. 688 * 689 * Arguments: 690 * oldhash The hash you want to expand 691 * newhash The hash structure for the new table 692 * 693 * Returns: 694 * Nothing 695 * 696 * Discussion: 697 */ 698 static int 699 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 700 { 701 uma_hash_slab_t slab; 702 u_int hval; 703 u_int idx; 704 705 if (!newhash->uh_slab_hash) 706 return (0); 707 708 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 709 return (0); 710 711 /* 712 * I need to investigate hash algorithms for resizing without a 713 * full rehash. 714 */ 715 716 for (idx = 0; idx < oldhash->uh_hashsize; idx++) 717 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) { 718 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]); 719 LIST_REMOVE(slab, uhs_hlink); 720 hval = UMA_HASH(newhash, slab->uhs_data); 721 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 722 slab, uhs_hlink); 723 } 724 725 return (1); 726 } 727 728 /* 729 * Free the hash bucket to the appropriate backing store. 730 * 731 * Arguments: 732 * slab_hash The hash bucket we're freeing 733 * hashsize The number of entries in that hash bucket 734 * 735 * Returns: 736 * Nothing 737 */ 738 static void 739 hash_free(struct uma_hash *hash) 740 { 741 if (hash->uh_slab_hash == NULL) 742 return; 743 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 744 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 745 else 746 free(hash->uh_slab_hash, M_UMAHASH); 747 } 748 749 /* 750 * Frees all outstanding items in a bucket 751 * 752 * Arguments: 753 * zone The zone to free to, must be unlocked. 754 * bucket The free/alloc bucket with items, cpu queue must be locked. 755 * 756 * Returns: 757 * Nothing 758 */ 759 760 static void 761 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 762 { 763 int i; 764 765 if (bucket == NULL) 766 return; 767 768 if (zone->uz_fini) 769 for (i = 0; i < bucket->ub_cnt; i++) 770 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 771 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 772 if (zone->uz_max_items > 0) { 773 ZONE_LOCK(zone); 774 zone->uz_items -= bucket->ub_cnt; 775 if (zone->uz_sleepers && zone->uz_items < zone->uz_max_items) 776 wakeup_one(zone); 777 ZONE_UNLOCK(zone); 778 } 779 bucket->ub_cnt = 0; 780 } 781 782 /* 783 * Drains the per cpu caches for a zone. 784 * 785 * NOTE: This may only be called while the zone is being turn down, and not 786 * during normal operation. This is necessary in order that we do not have 787 * to migrate CPUs to drain the per-CPU caches. 788 * 789 * Arguments: 790 * zone The zone to drain, must be unlocked. 791 * 792 * Returns: 793 * Nothing 794 */ 795 static void 796 cache_drain(uma_zone_t zone) 797 { 798 uma_cache_t cache; 799 int cpu; 800 801 /* 802 * XXX: It is safe to not lock the per-CPU caches, because we're 803 * tearing down the zone anyway. I.e., there will be no further use 804 * of the caches at this point. 805 * 806 * XXX: It would good to be able to assert that the zone is being 807 * torn down to prevent improper use of cache_drain(). 808 * 809 * XXX: We lock the zone before passing into bucket_cache_reclaim() as 810 * it is used elsewhere. Should the tear-down path be made special 811 * there in some form? 812 */ 813 CPU_FOREACH(cpu) { 814 cache = &zone->uz_cpu[cpu]; 815 bucket_drain(zone, cache->uc_allocbucket); 816 if (cache->uc_allocbucket != NULL) 817 bucket_free(zone, cache->uc_allocbucket, NULL); 818 cache->uc_allocbucket = NULL; 819 bucket_drain(zone, cache->uc_freebucket); 820 if (cache->uc_freebucket != NULL) 821 bucket_free(zone, cache->uc_freebucket, NULL); 822 cache->uc_freebucket = NULL; 823 bucket_drain(zone, cache->uc_crossbucket); 824 if (cache->uc_crossbucket != NULL) 825 bucket_free(zone, cache->uc_crossbucket, NULL); 826 cache->uc_crossbucket = NULL; 827 } 828 ZONE_LOCK(zone); 829 bucket_cache_reclaim(zone, true); 830 ZONE_UNLOCK(zone); 831 } 832 833 static void 834 cache_shrink(uma_zone_t zone, void *unused) 835 { 836 837 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 838 return; 839 840 ZONE_LOCK(zone); 841 zone->uz_bucket_size = 842 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2; 843 ZONE_UNLOCK(zone); 844 } 845 846 static void 847 cache_drain_safe_cpu(uma_zone_t zone, void *unused) 848 { 849 uma_cache_t cache; 850 uma_bucket_t b1, b2, b3; 851 int domain; 852 853 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 854 return; 855 856 b1 = b2 = b3 = NULL; 857 ZONE_LOCK(zone); 858 critical_enter(); 859 if (zone->uz_flags & UMA_ZONE_NUMA) 860 domain = PCPU_GET(domain); 861 else 862 domain = 0; 863 cache = &zone->uz_cpu[curcpu]; 864 if (cache->uc_allocbucket) { 865 if (cache->uc_allocbucket->ub_cnt != 0) 866 zone_put_bucket(zone, &zone->uz_domain[domain], 867 cache->uc_allocbucket, false); 868 else 869 b1 = cache->uc_allocbucket; 870 cache->uc_allocbucket = NULL; 871 } 872 if (cache->uc_freebucket) { 873 if (cache->uc_freebucket->ub_cnt != 0) 874 zone_put_bucket(zone, &zone->uz_domain[domain], 875 cache->uc_freebucket, false); 876 else 877 b2 = cache->uc_freebucket; 878 cache->uc_freebucket = NULL; 879 } 880 b3 = cache->uc_crossbucket; 881 cache->uc_crossbucket = NULL; 882 critical_exit(); 883 ZONE_UNLOCK(zone); 884 if (b1) 885 bucket_free(zone, b1, NULL); 886 if (b2) 887 bucket_free(zone, b2, NULL); 888 if (b3) { 889 bucket_drain(zone, b3); 890 bucket_free(zone, b3, NULL); 891 } 892 } 893 894 /* 895 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 896 * This is an expensive call because it needs to bind to all CPUs 897 * one by one and enter a critical section on each of them in order 898 * to safely access their cache buckets. 899 * Zone lock must not be held on call this function. 900 */ 901 static void 902 pcpu_cache_drain_safe(uma_zone_t zone) 903 { 904 int cpu; 905 906 /* 907 * Polite bucket sizes shrinking was not enouth, shrink aggressively. 908 */ 909 if (zone) 910 cache_shrink(zone, NULL); 911 else 912 zone_foreach(cache_shrink, NULL); 913 914 CPU_FOREACH(cpu) { 915 thread_lock(curthread); 916 sched_bind(curthread, cpu); 917 thread_unlock(curthread); 918 919 if (zone) 920 cache_drain_safe_cpu(zone, NULL); 921 else 922 zone_foreach(cache_drain_safe_cpu, NULL); 923 } 924 thread_lock(curthread); 925 sched_unbind(curthread); 926 thread_unlock(curthread); 927 } 928 929 /* 930 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller 931 * requested a drain, otherwise the per-domain caches are trimmed to either 932 * estimated working set size. 933 */ 934 static void 935 bucket_cache_reclaim(uma_zone_t zone, bool drain) 936 { 937 uma_zone_domain_t zdom; 938 uma_bucket_t bucket; 939 long target, tofree; 940 int i; 941 942 for (i = 0; i < vm_ndomains; i++) { 943 zdom = &zone->uz_domain[i]; 944 945 /* 946 * If we were asked to drain the zone, we are done only once 947 * this bucket cache is empty. Otherwise, we reclaim items in 948 * excess of the zone's estimated working set size. If the 949 * difference nitems - imin is larger than the WSS estimate, 950 * then the estimate will grow at the end of this interval and 951 * we ignore the historical average. 952 */ 953 target = drain ? 0 : lmax(zdom->uzd_wss, zdom->uzd_nitems - 954 zdom->uzd_imin); 955 while (zdom->uzd_nitems > target) { 956 bucket = TAILQ_LAST(&zdom->uzd_buckets, uma_bucketlist); 957 if (bucket == NULL) 958 break; 959 tofree = bucket->ub_cnt; 960 TAILQ_REMOVE(&zdom->uzd_buckets, bucket, ub_link); 961 zdom->uzd_nitems -= tofree; 962 963 /* 964 * Shift the bounds of the current WSS interval to avoid 965 * perturbing the estimate. 966 */ 967 zdom->uzd_imax -= lmin(zdom->uzd_imax, tofree); 968 zdom->uzd_imin -= lmin(zdom->uzd_imin, tofree); 969 970 ZONE_UNLOCK(zone); 971 bucket_drain(zone, bucket); 972 bucket_free(zone, bucket, NULL); 973 ZONE_LOCK(zone); 974 } 975 } 976 977 /* 978 * Shrink the zone bucket size to ensure that the per-CPU caches 979 * don't grow too large. 980 */ 981 if (zone->uz_bucket_size > zone->uz_bucket_size_min) 982 zone->uz_bucket_size--; 983 } 984 985 static void 986 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 987 { 988 uint8_t *mem; 989 int i; 990 uint8_t flags; 991 992 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes", 993 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera); 994 995 mem = slab_data(slab, keg); 996 flags = slab->us_flags; 997 i = start; 998 if (keg->uk_fini != NULL) { 999 for (i--; i > -1; i--) 1000 #ifdef INVARIANTS 1001 /* 1002 * trash_fini implies that dtor was trash_dtor. trash_fini 1003 * would check that memory hasn't been modified since free, 1004 * which executed trash_dtor. 1005 * That's why we need to run uma_dbg_kskip() check here, 1006 * albeit we don't make skip check for other init/fini 1007 * invocations. 1008 */ 1009 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) || 1010 keg->uk_fini != trash_fini) 1011 #endif 1012 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size); 1013 } 1014 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1015 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 1016 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 1017 uma_total_dec(PAGE_SIZE * keg->uk_ppera); 1018 } 1019 1020 /* 1021 * Frees pages from a keg back to the system. This is done on demand from 1022 * the pageout daemon. 1023 * 1024 * Returns nothing. 1025 */ 1026 static void 1027 keg_drain(uma_keg_t keg) 1028 { 1029 struct slabhead freeslabs = { 0 }; 1030 uma_domain_t dom; 1031 uma_slab_t slab, tmp; 1032 int i; 1033 1034 /* 1035 * We don't want to take pages from statically allocated kegs at this 1036 * time 1037 */ 1038 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 1039 return; 1040 1041 CTR3(KTR_UMA, "keg_drain %s(%p) free items: %u", 1042 keg->uk_name, keg, keg->uk_free); 1043 KEG_LOCK(keg); 1044 if (keg->uk_free == 0) 1045 goto finished; 1046 1047 for (i = 0; i < vm_ndomains; i++) { 1048 dom = &keg->uk_domain[i]; 1049 LIST_FOREACH_SAFE(slab, &dom->ud_free_slab, us_link, tmp) { 1050 /* We have nowhere to free these to. */ 1051 if (slab->us_flags & UMA_SLAB_BOOT) 1052 continue; 1053 1054 LIST_REMOVE(slab, us_link); 1055 keg->uk_pages -= keg->uk_ppera; 1056 keg->uk_free -= keg->uk_ipers; 1057 1058 if (keg->uk_flags & UMA_ZONE_HASH) 1059 UMA_HASH_REMOVE(&keg->uk_hash, slab); 1060 1061 LIST_INSERT_HEAD(&freeslabs, slab, us_link); 1062 } 1063 } 1064 1065 finished: 1066 KEG_UNLOCK(keg); 1067 1068 while ((slab = LIST_FIRST(&freeslabs)) != NULL) { 1069 LIST_REMOVE(slab, us_link); 1070 keg_free_slab(keg, slab, keg->uk_ipers); 1071 } 1072 } 1073 1074 static void 1075 zone_reclaim(uma_zone_t zone, int waitok, bool drain) 1076 { 1077 1078 /* 1079 * Set draining to interlock with zone_dtor() so we can release our 1080 * locks as we go. Only dtor() should do a WAITOK call since it 1081 * is the only call that knows the structure will still be available 1082 * when it wakes up. 1083 */ 1084 ZONE_LOCK(zone); 1085 while (zone->uz_flags & UMA_ZFLAG_RECLAIMING) { 1086 if (waitok == M_NOWAIT) 1087 goto out; 1088 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 1089 } 1090 zone->uz_flags |= UMA_ZFLAG_RECLAIMING; 1091 bucket_cache_reclaim(zone, drain); 1092 ZONE_UNLOCK(zone); 1093 1094 /* 1095 * The DRAINING flag protects us from being freed while 1096 * we're running. Normally the uma_rwlock would protect us but we 1097 * must be able to release and acquire the right lock for each keg. 1098 */ 1099 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) 1100 keg_drain(zone->uz_keg); 1101 ZONE_LOCK(zone); 1102 zone->uz_flags &= ~UMA_ZFLAG_RECLAIMING; 1103 wakeup(zone); 1104 out: 1105 ZONE_UNLOCK(zone); 1106 } 1107 1108 static void 1109 zone_drain(uma_zone_t zone, void *unused) 1110 { 1111 1112 zone_reclaim(zone, M_NOWAIT, true); 1113 } 1114 1115 static void 1116 zone_trim(uma_zone_t zone, void *unused) 1117 { 1118 1119 zone_reclaim(zone, M_NOWAIT, false); 1120 } 1121 1122 /* 1123 * Allocate a new slab for a keg. This does not insert the slab onto a list. 1124 * If the allocation was successful, the keg lock will be held upon return, 1125 * otherwise the keg will be left unlocked. 1126 * 1127 * Arguments: 1128 * flags Wait flags for the item initialization routine 1129 * aflags Wait flags for the slab allocation 1130 * 1131 * Returns: 1132 * The slab that was allocated or NULL if there is no memory and the 1133 * caller specified M_NOWAIT. 1134 */ 1135 static uma_slab_t 1136 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags, 1137 int aflags) 1138 { 1139 uma_alloc allocf; 1140 uma_slab_t slab; 1141 unsigned long size; 1142 uint8_t *mem; 1143 uint8_t sflags; 1144 int i; 1145 1146 KASSERT(domain >= 0 && domain < vm_ndomains, 1147 ("keg_alloc_slab: domain %d out of range", domain)); 1148 KEG_LOCK_ASSERT(keg); 1149 MPASS(zone->uz_lockptr == &keg->uk_lock); 1150 1151 allocf = keg->uk_allocf; 1152 KEG_UNLOCK(keg); 1153 1154 slab = NULL; 1155 mem = NULL; 1156 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1157 slab = zone_alloc_item(keg->uk_slabzone, NULL, domain, aflags); 1158 if (slab == NULL) 1159 goto out; 1160 } 1161 1162 /* 1163 * This reproduces the old vm_zone behavior of zero filling pages the 1164 * first time they are added to a zone. 1165 * 1166 * Malloced items are zeroed in uma_zalloc. 1167 */ 1168 1169 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 1170 aflags |= M_ZERO; 1171 else 1172 aflags &= ~M_ZERO; 1173 1174 if (keg->uk_flags & UMA_ZONE_NODUMP) 1175 aflags |= M_NODUMP; 1176 1177 /* zone is passed for legacy reasons. */ 1178 size = keg->uk_ppera * PAGE_SIZE; 1179 mem = allocf(zone, size, domain, &sflags, aflags); 1180 if (mem == NULL) { 1181 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1182 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 1183 slab = NULL; 1184 goto out; 1185 } 1186 uma_total_inc(size); 1187 1188 /* Point the slab into the allocated memory */ 1189 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 1190 slab = (uma_slab_t )(mem + keg->uk_pgoff); 1191 else 1192 ((uma_hash_slab_t)slab)->uhs_data = mem; 1193 1194 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 1195 for (i = 0; i < keg->uk_ppera; i++) 1196 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE), 1197 zone, slab); 1198 1199 slab->us_freecount = keg->uk_ipers; 1200 slab->us_flags = sflags; 1201 slab->us_domain = domain; 1202 BIT_FILL(keg->uk_ipers, &slab->us_free); 1203 #ifdef INVARIANTS 1204 BIT_ZERO(SLAB_MAX_SETSIZE, &slab->us_debugfree); 1205 #endif 1206 1207 if (keg->uk_init != NULL) { 1208 for (i = 0; i < keg->uk_ipers; i++) 1209 if (keg->uk_init(slab_item(slab, keg, i), 1210 keg->uk_size, flags) != 0) 1211 break; 1212 if (i != keg->uk_ipers) { 1213 keg_free_slab(keg, slab, i); 1214 slab = NULL; 1215 goto out; 1216 } 1217 } 1218 KEG_LOCK(keg); 1219 1220 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)", 1221 slab, keg->uk_name, keg); 1222 1223 if (keg->uk_flags & UMA_ZONE_HASH) 1224 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1225 1226 keg->uk_pages += keg->uk_ppera; 1227 keg->uk_free += keg->uk_ipers; 1228 1229 out: 1230 return (slab); 1231 } 1232 1233 /* 1234 * This function is intended to be used early on in place of page_alloc() so 1235 * that we may use the boot time page cache to satisfy allocations before 1236 * the VM is ready. 1237 */ 1238 static void * 1239 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1240 int wait) 1241 { 1242 uma_keg_t keg; 1243 void *mem; 1244 int pages; 1245 1246 keg = zone->uz_keg; 1247 /* 1248 * If we are in BOOT_BUCKETS or higher, than switch to real 1249 * allocator. Zones with page sized slabs switch at BOOT_PAGEALLOC. 1250 */ 1251 switch (booted) { 1252 case BOOT_COLD: 1253 case BOOT_STRAPPED: 1254 break; 1255 case BOOT_PAGEALLOC: 1256 if (keg->uk_ppera > 1) 1257 break; 1258 case BOOT_BUCKETS: 1259 case BOOT_RUNNING: 1260 #ifdef UMA_MD_SMALL_ALLOC 1261 keg->uk_allocf = (keg->uk_ppera > 1) ? 1262 page_alloc : uma_small_alloc; 1263 #else 1264 keg->uk_allocf = page_alloc; 1265 #endif 1266 return keg->uk_allocf(zone, bytes, domain, pflag, wait); 1267 } 1268 1269 /* 1270 * Check our small startup cache to see if it has pages remaining. 1271 */ 1272 pages = howmany(bytes, PAGE_SIZE); 1273 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__)); 1274 if (pages > boot_pages) 1275 panic("UMA zone \"%s\": Increase vm.boot_pages", zone->uz_name); 1276 #ifdef DIAGNOSTIC 1277 printf("%s from \"%s\", %d boot pages left\n", __func__, zone->uz_name, 1278 boot_pages); 1279 #endif 1280 mem = bootmem; 1281 boot_pages -= pages; 1282 bootmem += pages * PAGE_SIZE; 1283 *pflag = UMA_SLAB_BOOT; 1284 1285 return (mem); 1286 } 1287 1288 /* 1289 * Allocates a number of pages from the system 1290 * 1291 * Arguments: 1292 * bytes The number of bytes requested 1293 * wait Shall we wait? 1294 * 1295 * Returns: 1296 * A pointer to the alloced memory or possibly 1297 * NULL if M_NOWAIT is set. 1298 */ 1299 static void * 1300 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1301 int wait) 1302 { 1303 void *p; /* Returned page */ 1304 1305 *pflag = UMA_SLAB_KERNEL; 1306 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait); 1307 1308 return (p); 1309 } 1310 1311 static void * 1312 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 1313 int wait) 1314 { 1315 struct pglist alloctail; 1316 vm_offset_t addr, zkva; 1317 int cpu, flags; 1318 vm_page_t p, p_next; 1319 #ifdef NUMA 1320 struct pcpu *pc; 1321 #endif 1322 1323 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE); 1324 1325 TAILQ_INIT(&alloctail); 1326 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1327 malloc2vm_flags(wait); 1328 *pflag = UMA_SLAB_KERNEL; 1329 for (cpu = 0; cpu <= mp_maxid; cpu++) { 1330 if (CPU_ABSENT(cpu)) { 1331 p = vm_page_alloc(NULL, 0, flags); 1332 } else { 1333 #ifndef NUMA 1334 p = vm_page_alloc(NULL, 0, flags); 1335 #else 1336 pc = pcpu_find(cpu); 1337 p = vm_page_alloc_domain(NULL, 0, pc->pc_domain, flags); 1338 if (__predict_false(p == NULL)) 1339 p = vm_page_alloc(NULL, 0, flags); 1340 #endif 1341 } 1342 if (__predict_false(p == NULL)) 1343 goto fail; 1344 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1345 } 1346 if ((addr = kva_alloc(bytes)) == 0) 1347 goto fail; 1348 zkva = addr; 1349 TAILQ_FOREACH(p, &alloctail, listq) { 1350 pmap_qenter(zkva, &p, 1); 1351 zkva += PAGE_SIZE; 1352 } 1353 return ((void*)addr); 1354 fail: 1355 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1356 vm_page_unwire_noq(p); 1357 vm_page_free(p); 1358 } 1359 return (NULL); 1360 } 1361 1362 /* 1363 * Allocates a number of pages from within an object 1364 * 1365 * Arguments: 1366 * bytes The number of bytes requested 1367 * wait Shall we wait? 1368 * 1369 * Returns: 1370 * A pointer to the alloced memory or possibly 1371 * NULL if M_NOWAIT is set. 1372 */ 1373 static void * 1374 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, 1375 int wait) 1376 { 1377 TAILQ_HEAD(, vm_page) alloctail; 1378 u_long npages; 1379 vm_offset_t retkva, zkva; 1380 vm_page_t p, p_next; 1381 uma_keg_t keg; 1382 1383 TAILQ_INIT(&alloctail); 1384 keg = zone->uz_keg; 1385 1386 npages = howmany(bytes, PAGE_SIZE); 1387 while (npages > 0) { 1388 p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT | 1389 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | 1390 ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK : 1391 VM_ALLOC_NOWAIT)); 1392 if (p != NULL) { 1393 /* 1394 * Since the page does not belong to an object, its 1395 * listq is unused. 1396 */ 1397 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1398 npages--; 1399 continue; 1400 } 1401 /* 1402 * Page allocation failed, free intermediate pages and 1403 * exit. 1404 */ 1405 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1406 vm_page_unwire_noq(p); 1407 vm_page_free(p); 1408 } 1409 return (NULL); 1410 } 1411 *flags = UMA_SLAB_PRIV; 1412 zkva = keg->uk_kva + 1413 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1414 retkva = zkva; 1415 TAILQ_FOREACH(p, &alloctail, listq) { 1416 pmap_qenter(zkva, &p, 1); 1417 zkva += PAGE_SIZE; 1418 } 1419 1420 return ((void *)retkva); 1421 } 1422 1423 /* 1424 * Frees a number of pages to the system 1425 * 1426 * Arguments: 1427 * mem A pointer to the memory to be freed 1428 * size The size of the memory being freed 1429 * flags The original p->us_flags field 1430 * 1431 * Returns: 1432 * Nothing 1433 */ 1434 static void 1435 page_free(void *mem, vm_size_t size, uint8_t flags) 1436 { 1437 1438 if ((flags & UMA_SLAB_KERNEL) == 0) 1439 panic("UMA: page_free used with invalid flags %x", flags); 1440 1441 kmem_free((vm_offset_t)mem, size); 1442 } 1443 1444 /* 1445 * Frees pcpu zone allocations 1446 * 1447 * Arguments: 1448 * mem A pointer to the memory to be freed 1449 * size The size of the memory being freed 1450 * flags The original p->us_flags field 1451 * 1452 * Returns: 1453 * Nothing 1454 */ 1455 static void 1456 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags) 1457 { 1458 vm_offset_t sva, curva; 1459 vm_paddr_t paddr; 1460 vm_page_t m; 1461 1462 MPASS(size == (mp_maxid+1)*PAGE_SIZE); 1463 sva = (vm_offset_t)mem; 1464 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) { 1465 paddr = pmap_kextract(curva); 1466 m = PHYS_TO_VM_PAGE(paddr); 1467 vm_page_unwire_noq(m); 1468 vm_page_free(m); 1469 } 1470 pmap_qremove(sva, size >> PAGE_SHIFT); 1471 kva_free(sva, size); 1472 } 1473 1474 1475 /* 1476 * Zero fill initializer 1477 * 1478 * Arguments/Returns follow uma_init specifications 1479 */ 1480 static int 1481 zero_init(void *mem, int size, int flags) 1482 { 1483 bzero(mem, size); 1484 return (0); 1485 } 1486 1487 /* 1488 * Actual size of embedded struct slab (!OFFPAGE). 1489 */ 1490 size_t 1491 slab_sizeof(int nitems) 1492 { 1493 size_t s; 1494 1495 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems); 1496 return (roundup(s, UMA_ALIGN_PTR + 1)); 1497 } 1498 1499 /* 1500 * Size of memory for embedded slabs (!OFFPAGE). 1501 */ 1502 size_t 1503 slab_space(int nitems) 1504 { 1505 return (UMA_SLAB_SIZE - slab_sizeof(nitems)); 1506 } 1507 1508 /* 1509 * Compute the number of items that will fit in an embedded (!OFFPAGE) slab 1510 * with a given size and alignment. 1511 */ 1512 int 1513 slab_ipers(size_t size, int align) 1514 { 1515 int rsize; 1516 int nitems; 1517 1518 /* 1519 * Compute the ideal number of items that will fit in a page and 1520 * then compute the actual number based on a bitset nitems wide. 1521 */ 1522 rsize = roundup(size, align + 1); 1523 nitems = UMA_SLAB_SIZE / rsize; 1524 return (slab_space(nitems) / rsize); 1525 } 1526 1527 /* 1528 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1529 * 1530 * Arguments 1531 * keg The zone we should initialize 1532 * 1533 * Returns 1534 * Nothing 1535 */ 1536 static void 1537 keg_small_init(uma_keg_t keg) 1538 { 1539 u_int rsize; 1540 u_int memused; 1541 u_int wastedspace; 1542 u_int shsize; 1543 u_int slabsize; 1544 1545 if (keg->uk_flags & UMA_ZONE_PCPU) { 1546 u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU; 1547 1548 slabsize = UMA_PCPU_ALLOC_SIZE; 1549 keg->uk_ppera = ncpus; 1550 } else { 1551 slabsize = UMA_SLAB_SIZE; 1552 keg->uk_ppera = 1; 1553 } 1554 1555 /* 1556 * Calculate the size of each allocation (rsize) according to 1557 * alignment. If the requested size is smaller than we have 1558 * allocation bits for we round it up. 1559 */ 1560 rsize = keg->uk_size; 1561 if (rsize < slabsize / SLAB_MAX_SETSIZE) 1562 rsize = slabsize / SLAB_MAX_SETSIZE; 1563 if (rsize & keg->uk_align) 1564 rsize = roundup(rsize, keg->uk_align + 1); 1565 keg->uk_rsize = rsize; 1566 1567 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1568 keg->uk_rsize < UMA_PCPU_ALLOC_SIZE, 1569 ("%s: size %u too large", __func__, keg->uk_rsize)); 1570 1571 /* 1572 * Use a pessimistic bit count for shsize. It may be possible to 1573 * squeeze one more item in for very particular sizes if we were 1574 * to loop and reduce the bitsize if there is waste. 1575 */ 1576 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1577 shsize = 0; 1578 else 1579 shsize = slab_sizeof(slabsize / rsize); 1580 1581 if (rsize <= slabsize - shsize) 1582 keg->uk_ipers = (slabsize - shsize) / rsize; 1583 else { 1584 /* Handle special case when we have 1 item per slab, so 1585 * alignment requirement can be relaxed. */ 1586 KASSERT(keg->uk_size <= slabsize - shsize, 1587 ("%s: size %u greater than slab", __func__, keg->uk_size)); 1588 keg->uk_ipers = 1; 1589 } 1590 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE, 1591 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1592 1593 memused = keg->uk_ipers * rsize + shsize; 1594 wastedspace = slabsize - memused; 1595 1596 /* 1597 * We can't do OFFPAGE if we're internal or if we've been 1598 * asked to not go to the VM for buckets. If we do this we 1599 * may end up going to the VM for slabs which we do not 1600 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1601 * of UMA_ZONE_VM, which clearly forbids it. 1602 */ 1603 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1604 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1605 return; 1606 1607 /* 1608 * See if using an OFFPAGE slab will limit our waste. Only do 1609 * this if it permits more items per-slab. 1610 * 1611 * XXX We could try growing slabsize to limit max waste as well. 1612 * Historically this was not done because the VM could not 1613 * efficiently handle contiguous allocations. 1614 */ 1615 if ((wastedspace >= slabsize / UMA_MAX_WASTE) && 1616 (keg->uk_ipers < (slabsize / keg->uk_rsize))) { 1617 keg->uk_ipers = slabsize / keg->uk_rsize; 1618 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE, 1619 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1620 CTR6(KTR_UMA, "UMA decided we need offpage slab headers for " 1621 "keg: %s(%p), calculated wastedspace = %d, " 1622 "maximum wasted space allowed = %d, " 1623 "calculated ipers = %d, " 1624 "new wasted space = %d\n", keg->uk_name, keg, wastedspace, 1625 slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1626 slabsize - keg->uk_ipers * keg->uk_rsize); 1627 /* 1628 * If we had access to memory to embed a slab header we 1629 * also have a page structure to use vtoslab() instead of 1630 * hash to find slabs. If the zone was explicitly created 1631 * OFFPAGE we can't necessarily touch the memory. 1632 */ 1633 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) 1634 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1635 } 1636 1637 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1638 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1639 keg->uk_flags |= UMA_ZONE_HASH; 1640 } 1641 1642 /* 1643 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1644 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1645 * more complicated. 1646 * 1647 * Arguments 1648 * keg The keg we should initialize 1649 * 1650 * Returns 1651 * Nothing 1652 */ 1653 static void 1654 keg_large_init(uma_keg_t keg) 1655 { 1656 1657 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1658 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1659 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1660 1661 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1662 keg->uk_ipers = 1; 1663 keg->uk_rsize = keg->uk_size; 1664 1665 /* Check whether we have enough space to not do OFFPAGE. */ 1666 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0 && 1667 PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < 1668 slab_sizeof(SLAB_MIN_SETSIZE)) { 1669 /* 1670 * We can't do OFFPAGE if we're internal, in which case 1671 * we need an extra page per allocation to contain the 1672 * slab header. 1673 */ 1674 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0) 1675 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1676 else 1677 keg->uk_ppera++; 1678 } 1679 1680 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1681 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1682 keg->uk_flags |= UMA_ZONE_HASH; 1683 } 1684 1685 static void 1686 keg_cachespread_init(uma_keg_t keg) 1687 { 1688 int alignsize; 1689 int trailer; 1690 int pages; 1691 int rsize; 1692 1693 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1694 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1695 1696 alignsize = keg->uk_align + 1; 1697 rsize = keg->uk_size; 1698 /* 1699 * We want one item to start on every align boundary in a page. To 1700 * do this we will span pages. We will also extend the item by the 1701 * size of align if it is an even multiple of align. Otherwise, it 1702 * would fall on the same boundary every time. 1703 */ 1704 if (rsize & keg->uk_align) 1705 rsize = (rsize & ~keg->uk_align) + alignsize; 1706 if ((rsize & alignsize) == 0) 1707 rsize += alignsize; 1708 trailer = rsize - keg->uk_size; 1709 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1710 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1711 keg->uk_rsize = rsize; 1712 keg->uk_ppera = pages; 1713 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1714 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1715 KASSERT(keg->uk_ipers <= SLAB_MAX_SETSIZE, 1716 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1717 keg->uk_ipers)); 1718 } 1719 1720 /* 1721 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1722 * the keg onto the global keg list. 1723 * 1724 * Arguments/Returns follow uma_ctor specifications 1725 * udata Actually uma_kctor_args 1726 */ 1727 static int 1728 keg_ctor(void *mem, int size, void *udata, int flags) 1729 { 1730 struct uma_kctor_args *arg = udata; 1731 uma_keg_t keg = mem; 1732 uma_zone_t zone; 1733 1734 bzero(keg, size); 1735 keg->uk_size = arg->size; 1736 keg->uk_init = arg->uminit; 1737 keg->uk_fini = arg->fini; 1738 keg->uk_align = arg->align; 1739 keg->uk_free = 0; 1740 keg->uk_reserve = 0; 1741 keg->uk_pages = 0; 1742 keg->uk_flags = arg->flags; 1743 keg->uk_slabzone = NULL; 1744 1745 /* 1746 * We use a global round-robin policy by default. Zones with 1747 * UMA_ZONE_NUMA set will use first-touch instead, in which case the 1748 * iterator is never run. 1749 */ 1750 keg->uk_dr.dr_policy = DOMAINSET_RR(); 1751 keg->uk_dr.dr_iter = 0; 1752 1753 /* 1754 * The master zone is passed to us at keg-creation time. 1755 */ 1756 zone = arg->zone; 1757 keg->uk_name = zone->uz_name; 1758 1759 if (arg->flags & UMA_ZONE_VM) 1760 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1761 1762 if (arg->flags & UMA_ZONE_ZINIT) 1763 keg->uk_init = zero_init; 1764 1765 if (arg->flags & UMA_ZONE_MALLOC) 1766 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1767 1768 if (arg->flags & UMA_ZONE_PCPU) 1769 #ifdef SMP 1770 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1771 #else 1772 keg->uk_flags &= ~UMA_ZONE_PCPU; 1773 #endif 1774 1775 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1776 keg_cachespread_init(keg); 1777 } else { 1778 if (keg->uk_size > slab_space(SLAB_MIN_SETSIZE)) 1779 keg_large_init(keg); 1780 else 1781 keg_small_init(keg); 1782 } 1783 1784 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1785 keg->uk_slabzone = slabzone; 1786 1787 /* 1788 * If we haven't booted yet we need allocations to go through the 1789 * startup cache until the vm is ready. 1790 */ 1791 if (booted < BOOT_PAGEALLOC) 1792 keg->uk_allocf = startup_alloc; 1793 #ifdef UMA_MD_SMALL_ALLOC 1794 else if (keg->uk_ppera == 1) 1795 keg->uk_allocf = uma_small_alloc; 1796 #endif 1797 else if (keg->uk_flags & UMA_ZONE_PCPU) 1798 keg->uk_allocf = pcpu_page_alloc; 1799 else 1800 keg->uk_allocf = page_alloc; 1801 #ifdef UMA_MD_SMALL_ALLOC 1802 if (keg->uk_ppera == 1) 1803 keg->uk_freef = uma_small_free; 1804 else 1805 #endif 1806 if (keg->uk_flags & UMA_ZONE_PCPU) 1807 keg->uk_freef = pcpu_page_free; 1808 else 1809 keg->uk_freef = page_free; 1810 1811 /* 1812 * Initialize keg's lock 1813 */ 1814 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1815 1816 /* 1817 * If we're putting the slab header in the actual page we need to 1818 * figure out where in each page it goes. See slab_sizeof 1819 * definition. 1820 */ 1821 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1822 size_t shsize; 1823 1824 shsize = slab_sizeof(keg->uk_ipers); 1825 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize; 1826 /* 1827 * The only way the following is possible is if with our 1828 * UMA_ALIGN_PTR adjustments we are now bigger than 1829 * UMA_SLAB_SIZE. I haven't checked whether this is 1830 * mathematically possible for all cases, so we make 1831 * sure here anyway. 1832 */ 1833 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera, 1834 ("zone %s ipers %d rsize %d size %d slab won't fit", 1835 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size)); 1836 } 1837 1838 if (keg->uk_flags & UMA_ZONE_HASH) 1839 hash_alloc(&keg->uk_hash, 0); 1840 1841 CTR5(KTR_UMA, "keg_ctor %p zone %s(%p) out %d free %d\n", 1842 keg, zone->uz_name, zone, 1843 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, 1844 keg->uk_free); 1845 1846 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1847 1848 rw_wlock(&uma_rwlock); 1849 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1850 rw_wunlock(&uma_rwlock); 1851 return (0); 1852 } 1853 1854 static void 1855 zone_alloc_counters(uma_zone_t zone, void *unused) 1856 { 1857 1858 zone->uz_allocs = counter_u64_alloc(M_WAITOK); 1859 zone->uz_frees = counter_u64_alloc(M_WAITOK); 1860 zone->uz_fails = counter_u64_alloc(M_WAITOK); 1861 } 1862 1863 static void 1864 zone_alloc_sysctl(uma_zone_t zone, void *unused) 1865 { 1866 uma_zone_domain_t zdom; 1867 uma_keg_t keg; 1868 struct sysctl_oid *oid, *domainoid; 1869 int domains, i, cnt; 1870 static const char *nokeg = "cache zone"; 1871 char *c; 1872 1873 /* 1874 * Make a sysctl safe copy of the zone name by removing 1875 * any special characters and handling dups by appending 1876 * an index. 1877 */ 1878 if (zone->uz_namecnt != 0) { 1879 /* Count the number of decimal digits and '_' separator. */ 1880 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++) 1881 cnt /= 10; 1882 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1, 1883 M_UMA, M_WAITOK); 1884 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name, 1885 zone->uz_namecnt); 1886 } else 1887 zone->uz_ctlname = strdup(zone->uz_name, M_UMA); 1888 for (c = zone->uz_ctlname; *c != '\0'; c++) 1889 if (strchr("./\\ -", *c) != NULL) 1890 *c = '_'; 1891 1892 /* 1893 * Basic parameters at the root. 1894 */ 1895 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma), 1896 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD, NULL, ""); 1897 oid = zone->uz_oid; 1898 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1899 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size"); 1900 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1901 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE, 1902 zone, 0, sysctl_handle_uma_zone_flags, "A", 1903 "Allocator configuration flags"); 1904 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1905 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0, 1906 "Desired per-cpu cache size"); 1907 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1908 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0, 1909 "Maximum allowed per-cpu cache size"); 1910 1911 /* 1912 * keg if present. 1913 */ 1914 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 1915 "keg", CTLFLAG_RD, NULL, ""); 1916 keg = zone->uz_keg; 1917 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) { 1918 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1919 "name", CTLFLAG_RD, keg->uk_name, "Keg name"); 1920 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1921 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0, 1922 "Real object size with alignment"); 1923 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1924 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0, 1925 "pages per-slab allocation"); 1926 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1927 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0, 1928 "items available per-slab"); 1929 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1930 "align", CTLFLAG_RD, &keg->uk_align, 0, 1931 "item alignment mask"); 1932 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1933 "pages", CTLFLAG_RD, &keg->uk_pages, 0, 1934 "Total pages currently allocated from VM"); 1935 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1936 "free", CTLFLAG_RD, &keg->uk_free, 0, 1937 "items free in the slab layer"); 1938 } else 1939 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1940 "name", CTLFLAG_RD, nokeg, "Keg name"); 1941 1942 /* 1943 * Information about zone limits. 1944 */ 1945 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 1946 "limit", CTLFLAG_RD, NULL, ""); 1947 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1948 "items", CTLFLAG_RD, &zone->uz_items, 0, 1949 "current number of cached items"); 1950 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1951 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0, 1952 "Maximum number of cached items"); 1953 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1954 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0, 1955 "Number of threads sleeping at limit"); 1956 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1957 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0, 1958 "Total zone limit sleeps"); 1959 1960 /* 1961 * Per-domain information. 1962 */ 1963 if ((zone->uz_flags & UMA_ZONE_NUMA) != 0) 1964 domains = vm_ndomains; 1965 else 1966 domains = 1; 1967 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), 1968 OID_AUTO, "domain", CTLFLAG_RD, NULL, ""); 1969 for (i = 0; i < domains; i++) { 1970 zdom = &zone->uz_domain[i]; 1971 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid), 1972 OID_AUTO, VM_DOMAIN(i)->vmd_name, CTLFLAG_RD, NULL, ""); 1973 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1974 "nitems", CTLFLAG_RD, &zdom->uzd_nitems, 1975 "number of items in this domain"); 1976 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1977 "imax", CTLFLAG_RD, &zdom->uzd_imax, 1978 "maximum item count in this period"); 1979 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1980 "imin", CTLFLAG_RD, &zdom->uzd_imin, 1981 "minimum item count in this period"); 1982 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1983 "wss", CTLFLAG_RD, &zdom->uzd_wss, 1984 "Working set size"); 1985 } 1986 1987 /* 1988 * General statistics. 1989 */ 1990 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO, 1991 "stats", CTLFLAG_RD, NULL, ""); 1992 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1993 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE, 1994 zone, 1, sysctl_handle_uma_zone_cur, "I", 1995 "Current number of allocated items"); 1996 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 1997 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 1998 zone, 0, sysctl_handle_uma_zone_allocs, "QU", 1999 "Total allocation calls"); 2000 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2001 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE, 2002 zone, 0, sysctl_handle_uma_zone_frees, "QU", 2003 "Total free calls"); 2004 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2005 "fails", CTLFLAG_RD, &zone->uz_fails, 2006 "Number of allocation failures"); 2007 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, 2008 "xdomain", CTLFLAG_RD, &zone->uz_xdomain, 0, 2009 "Free calls from the wrong domain"); 2010 } 2011 2012 struct uma_zone_count { 2013 const char *name; 2014 int count; 2015 }; 2016 2017 static void 2018 zone_count(uma_zone_t zone, void *arg) 2019 { 2020 struct uma_zone_count *cnt; 2021 2022 cnt = arg; 2023 /* 2024 * Some zones are rapidly created with identical names and 2025 * destroyed out of order. This can lead to gaps in the count. 2026 * Use one greater than the maximum observed for this name. 2027 */ 2028 if (strcmp(zone->uz_name, cnt->name) == 0) 2029 cnt->count = MAX(cnt->count, 2030 zone->uz_namecnt + 1); 2031 } 2032 2033 /* 2034 * Zone header ctor. This initializes all fields, locks, etc. 2035 * 2036 * Arguments/Returns follow uma_ctor specifications 2037 * udata Actually uma_zctor_args 2038 */ 2039 static int 2040 zone_ctor(void *mem, int size, void *udata, int flags) 2041 { 2042 struct uma_zone_count cnt; 2043 struct uma_zctor_args *arg = udata; 2044 uma_zone_t zone = mem; 2045 uma_zone_t z; 2046 uma_keg_t keg; 2047 int i; 2048 2049 bzero(zone, size); 2050 zone->uz_name = arg->name; 2051 zone->uz_ctor = arg->ctor; 2052 zone->uz_dtor = arg->dtor; 2053 zone->uz_init = NULL; 2054 zone->uz_fini = NULL; 2055 zone->uz_sleeps = 0; 2056 zone->uz_xdomain = 0; 2057 zone->uz_bucket_size = 0; 2058 zone->uz_bucket_size_min = 0; 2059 zone->uz_bucket_size_max = BUCKET_MAX; 2060 zone->uz_flags = 0; 2061 zone->uz_warning = NULL; 2062 /* The domain structures follow the cpu structures. */ 2063 zone->uz_domain = (struct uma_zone_domain *)&zone->uz_cpu[mp_ncpus]; 2064 zone->uz_bkt_max = ULONG_MAX; 2065 timevalclear(&zone->uz_ratecheck); 2066 2067 /* Count the number of duplicate names. */ 2068 cnt.name = arg->name; 2069 cnt.count = 0; 2070 zone_foreach(zone_count, &cnt); 2071 zone->uz_namecnt = cnt.count; 2072 2073 for (i = 0; i < vm_ndomains; i++) 2074 TAILQ_INIT(&zone->uz_domain[i].uzd_buckets); 2075 2076 #ifdef INVARIANTS 2077 if (arg->uminit == trash_init && arg->fini == trash_fini) 2078 zone->uz_flags |= UMA_ZFLAG_TRASH; 2079 #endif 2080 2081 /* 2082 * This is a pure cache zone, no kegs. 2083 */ 2084 if (arg->import) { 2085 if (arg->flags & UMA_ZONE_VM) 2086 arg->flags |= UMA_ZFLAG_CACHEONLY; 2087 zone->uz_flags = arg->flags; 2088 zone->uz_size = arg->size; 2089 zone->uz_import = arg->import; 2090 zone->uz_release = arg->release; 2091 zone->uz_arg = arg->arg; 2092 zone->uz_lockptr = &zone->uz_lock; 2093 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 2094 rw_wlock(&uma_rwlock); 2095 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 2096 rw_wunlock(&uma_rwlock); 2097 goto out; 2098 } 2099 2100 /* 2101 * Use the regular zone/keg/slab allocator. 2102 */ 2103 zone->uz_import = zone_import; 2104 zone->uz_release = zone_release; 2105 zone->uz_arg = zone; 2106 keg = arg->keg; 2107 2108 if (arg->flags & UMA_ZONE_SECONDARY) { 2109 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 2110 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 2111 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 2112 zone->uz_init = arg->uminit; 2113 zone->uz_fini = arg->fini; 2114 zone->uz_lockptr = &keg->uk_lock; 2115 zone->uz_flags |= UMA_ZONE_SECONDARY; 2116 rw_wlock(&uma_rwlock); 2117 ZONE_LOCK(zone); 2118 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 2119 if (LIST_NEXT(z, uz_link) == NULL) { 2120 LIST_INSERT_AFTER(z, zone, uz_link); 2121 break; 2122 } 2123 } 2124 ZONE_UNLOCK(zone); 2125 rw_wunlock(&uma_rwlock); 2126 } else if (keg == NULL) { 2127 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 2128 arg->align, arg->flags)) == NULL) 2129 return (ENOMEM); 2130 } else { 2131 struct uma_kctor_args karg; 2132 int error; 2133 2134 /* We should only be here from uma_startup() */ 2135 karg.size = arg->size; 2136 karg.uminit = arg->uminit; 2137 karg.fini = arg->fini; 2138 karg.align = arg->align; 2139 karg.flags = arg->flags; 2140 karg.zone = zone; 2141 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 2142 flags); 2143 if (error) 2144 return (error); 2145 } 2146 2147 /* Inherit properties from the keg. */ 2148 zone->uz_keg = keg; 2149 zone->uz_size = keg->uk_size; 2150 zone->uz_flags |= (keg->uk_flags & 2151 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 2152 2153 out: 2154 if (__predict_true(booted == BOOT_RUNNING)) { 2155 zone_alloc_counters(zone, NULL); 2156 zone_alloc_sysctl(zone, NULL); 2157 } else { 2158 zone->uz_allocs = EARLY_COUNTER; 2159 zone->uz_frees = EARLY_COUNTER; 2160 zone->uz_fails = EARLY_COUNTER; 2161 } 2162 2163 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) != 2164 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET), 2165 ("Invalid zone flag combination")); 2166 if (arg->flags & UMA_ZFLAG_INTERNAL) 2167 zone->uz_bucket_size_max = zone->uz_bucket_size = 0; 2168 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0) 2169 zone->uz_bucket_size = BUCKET_MAX; 2170 else if ((arg->flags & UMA_ZONE_MINBUCKET) != 0) 2171 zone->uz_bucket_size_max = zone->uz_bucket_size = BUCKET_MIN; 2172 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0) 2173 zone->uz_bucket_size = 0; 2174 else 2175 zone->uz_bucket_size = bucket_select(zone->uz_size); 2176 zone->uz_bucket_size_min = zone->uz_bucket_size; 2177 2178 return (0); 2179 } 2180 2181 /* 2182 * Keg header dtor. This frees all data, destroys locks, frees the hash 2183 * table and removes the keg from the global list. 2184 * 2185 * Arguments/Returns follow uma_dtor specifications 2186 * udata unused 2187 */ 2188 static void 2189 keg_dtor(void *arg, int size, void *udata) 2190 { 2191 uma_keg_t keg; 2192 2193 keg = (uma_keg_t)arg; 2194 KEG_LOCK(keg); 2195 if (keg->uk_free != 0) { 2196 printf("Freed UMA keg (%s) was not empty (%d items). " 2197 " Lost %d pages of memory.\n", 2198 keg->uk_name ? keg->uk_name : "", 2199 keg->uk_free, keg->uk_pages); 2200 } 2201 KEG_UNLOCK(keg); 2202 2203 hash_free(&keg->uk_hash); 2204 2205 KEG_LOCK_FINI(keg); 2206 } 2207 2208 /* 2209 * Zone header dtor. 2210 * 2211 * Arguments/Returns follow uma_dtor specifications 2212 * udata unused 2213 */ 2214 static void 2215 zone_dtor(void *arg, int size, void *udata) 2216 { 2217 uma_zone_t zone; 2218 uma_keg_t keg; 2219 2220 zone = (uma_zone_t)arg; 2221 2222 sysctl_remove_oid(zone->uz_oid, 1, 1); 2223 2224 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 2225 cache_drain(zone); 2226 2227 rw_wlock(&uma_rwlock); 2228 LIST_REMOVE(zone, uz_link); 2229 rw_wunlock(&uma_rwlock); 2230 /* 2231 * XXX there are some races here where 2232 * the zone can be drained but zone lock 2233 * released and then refilled before we 2234 * remove it... we dont care for now 2235 */ 2236 zone_reclaim(zone, M_WAITOK, true); 2237 /* 2238 * We only destroy kegs from non secondary/non cache zones. 2239 */ 2240 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) { 2241 keg = zone->uz_keg; 2242 rw_wlock(&uma_rwlock); 2243 LIST_REMOVE(keg, uk_link); 2244 rw_wunlock(&uma_rwlock); 2245 zone_free_item(kegs, keg, NULL, SKIP_NONE); 2246 } 2247 counter_u64_free(zone->uz_allocs); 2248 counter_u64_free(zone->uz_frees); 2249 counter_u64_free(zone->uz_fails); 2250 free(zone->uz_ctlname, M_UMA); 2251 if (zone->uz_lockptr == &zone->uz_lock) 2252 ZONE_LOCK_FINI(zone); 2253 } 2254 2255 /* 2256 * Traverses every zone in the system and calls a callback 2257 * 2258 * Arguments: 2259 * zfunc A pointer to a function which accepts a zone 2260 * as an argument. 2261 * 2262 * Returns: 2263 * Nothing 2264 */ 2265 static void 2266 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg) 2267 { 2268 uma_keg_t keg; 2269 uma_zone_t zone; 2270 2271 /* 2272 * Before BOOT_RUNNING we are guaranteed to be single 2273 * threaded, so locking isn't needed. Startup functions 2274 * are allowed to use M_WAITOK. 2275 */ 2276 if (__predict_true(booted == BOOT_RUNNING)) 2277 rw_rlock(&uma_rwlock); 2278 LIST_FOREACH(keg, &uma_kegs, uk_link) { 2279 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 2280 zfunc(zone, arg); 2281 } 2282 LIST_FOREACH(zone, &uma_cachezones, uz_link) 2283 zfunc(zone, arg); 2284 if (__predict_true(booted == BOOT_RUNNING)) 2285 rw_runlock(&uma_rwlock); 2286 } 2287 2288 /* 2289 * Count how many pages do we need to bootstrap. VM supplies 2290 * its need in early zones in the argument, we add up our zones, 2291 * which consist of: UMA Slabs, UMA Hash and 9 Bucket zones. The 2292 * zone of zones and zone of kegs are accounted separately. 2293 */ 2294 #define UMA_BOOT_ZONES 11 2295 /* Zone of zones and zone of kegs have arbitrary alignment. */ 2296 #define UMA_BOOT_ALIGN 32 2297 static int zsize, ksize; 2298 int 2299 uma_startup_count(int vm_zones) 2300 { 2301 int zones, pages; 2302 size_t space, size; 2303 2304 ksize = sizeof(struct uma_keg) + 2305 (sizeof(struct uma_domain) * vm_ndomains); 2306 zsize = sizeof(struct uma_zone) + 2307 (sizeof(struct uma_cache) * (mp_maxid + 1)) + 2308 (sizeof(struct uma_zone_domain) * vm_ndomains); 2309 2310 /* 2311 * Memory for the zone of kegs and its keg, 2312 * and for zone of zones. 2313 */ 2314 pages = howmany(roundup(zsize, CACHE_LINE_SIZE) * 2 + 2315 roundup(ksize, CACHE_LINE_SIZE), PAGE_SIZE); 2316 2317 #ifdef UMA_MD_SMALL_ALLOC 2318 zones = UMA_BOOT_ZONES; 2319 #else 2320 zones = UMA_BOOT_ZONES + vm_zones; 2321 vm_zones = 0; 2322 #endif 2323 size = slab_sizeof(SLAB_MAX_SETSIZE); 2324 space = slab_space(SLAB_MAX_SETSIZE); 2325 2326 /* Memory for the rest of startup zones, UMA and VM, ... */ 2327 if (zsize > space) { 2328 /* See keg_large_init(). */ 2329 u_int ppera; 2330 2331 ppera = howmany(roundup2(zsize, UMA_BOOT_ALIGN), PAGE_SIZE); 2332 if (PAGE_SIZE * ppera - roundup2(zsize, UMA_BOOT_ALIGN) < size) 2333 ppera++; 2334 pages += (zones + vm_zones) * ppera; 2335 } else if (roundup2(zsize, UMA_BOOT_ALIGN) > space) 2336 /* See keg_small_init() special case for uk_ppera = 1. */ 2337 pages += zones; 2338 else 2339 pages += howmany(zones, 2340 space / roundup2(zsize, UMA_BOOT_ALIGN)); 2341 2342 /* ... and their kegs. Note that zone of zones allocates a keg! */ 2343 pages += howmany(zones + 1, 2344 space / roundup2(ksize, UMA_BOOT_ALIGN)); 2345 2346 return (pages); 2347 } 2348 2349 void 2350 uma_startup(void *mem, int npages) 2351 { 2352 struct uma_zctor_args args; 2353 uma_keg_t masterkeg; 2354 uintptr_t m; 2355 2356 #ifdef DIAGNOSTIC 2357 printf("Entering %s with %d boot pages configured\n", __func__, npages); 2358 #endif 2359 2360 rw_init(&uma_rwlock, "UMA lock"); 2361 2362 /* Use bootpages memory for the zone of zones and zone of kegs. */ 2363 m = (uintptr_t)mem; 2364 zones = (uma_zone_t)m; 2365 m += roundup(zsize, CACHE_LINE_SIZE); 2366 kegs = (uma_zone_t)m; 2367 m += roundup(zsize, CACHE_LINE_SIZE); 2368 masterkeg = (uma_keg_t)m; 2369 m += roundup(ksize, CACHE_LINE_SIZE); 2370 m = roundup(m, PAGE_SIZE); 2371 npages -= (m - (uintptr_t)mem) / PAGE_SIZE; 2372 mem = (void *)m; 2373 2374 /* "manually" create the initial zone */ 2375 memset(&args, 0, sizeof(args)); 2376 args.name = "UMA Kegs"; 2377 args.size = ksize; 2378 args.ctor = keg_ctor; 2379 args.dtor = keg_dtor; 2380 args.uminit = zero_init; 2381 args.fini = NULL; 2382 args.keg = masterkeg; 2383 args.align = UMA_BOOT_ALIGN - 1; 2384 args.flags = UMA_ZFLAG_INTERNAL; 2385 zone_ctor(kegs, zsize, &args, M_WAITOK); 2386 2387 bootmem = mem; 2388 boot_pages = npages; 2389 2390 args.name = "UMA Zones"; 2391 args.size = zsize; 2392 args.ctor = zone_ctor; 2393 args.dtor = zone_dtor; 2394 args.uminit = zero_init; 2395 args.fini = NULL; 2396 args.keg = NULL; 2397 args.align = UMA_BOOT_ALIGN - 1; 2398 args.flags = UMA_ZFLAG_INTERNAL; 2399 zone_ctor(zones, zsize, &args, M_WAITOK); 2400 2401 /* Now make a zone for slab headers */ 2402 slabzone = uma_zcreate("UMA Slabs", sizeof(struct uma_hash_slab), 2403 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2404 2405 hashzone = uma_zcreate("UMA Hash", 2406 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 2407 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 2408 2409 bucket_init(); 2410 2411 booted = BOOT_STRAPPED; 2412 } 2413 2414 void 2415 uma_startup1(void) 2416 { 2417 2418 #ifdef DIAGNOSTIC 2419 printf("Entering %s with %d boot pages left\n", __func__, boot_pages); 2420 #endif 2421 booted = BOOT_PAGEALLOC; 2422 } 2423 2424 void 2425 uma_startup2(void) 2426 { 2427 2428 #ifdef DIAGNOSTIC 2429 printf("Entering %s with %d boot pages left\n", __func__, boot_pages); 2430 #endif 2431 booted = BOOT_BUCKETS; 2432 sx_init(&uma_reclaim_lock, "umareclaim"); 2433 bucket_enable(); 2434 } 2435 2436 /* 2437 * Initialize our callout handle 2438 * 2439 */ 2440 static void 2441 uma_startup3(void) 2442 { 2443 2444 #ifdef INVARIANTS 2445 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor); 2446 uma_dbg_cnt = counter_u64_alloc(M_WAITOK); 2447 uma_skip_cnt = counter_u64_alloc(M_WAITOK); 2448 #endif 2449 zone_foreach(zone_alloc_counters, NULL); 2450 zone_foreach(zone_alloc_sysctl, NULL); 2451 callout_init(&uma_callout, 1); 2452 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 2453 booted = BOOT_RUNNING; 2454 } 2455 2456 static uma_keg_t 2457 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 2458 int align, uint32_t flags) 2459 { 2460 struct uma_kctor_args args; 2461 2462 args.size = size; 2463 args.uminit = uminit; 2464 args.fini = fini; 2465 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 2466 args.flags = flags; 2467 args.zone = zone; 2468 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK)); 2469 } 2470 2471 /* Public functions */ 2472 /* See uma.h */ 2473 void 2474 uma_set_align(int align) 2475 { 2476 2477 if (align != UMA_ALIGN_CACHE) 2478 uma_align_cache = align; 2479 } 2480 2481 /* See uma.h */ 2482 uma_zone_t 2483 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 2484 uma_init uminit, uma_fini fini, int align, uint32_t flags) 2485 2486 { 2487 struct uma_zctor_args args; 2488 uma_zone_t res; 2489 bool locked; 2490 2491 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"", 2492 align, name)); 2493 2494 /* Sets all zones to a first-touch domain policy. */ 2495 #ifdef UMA_FIRSTTOUCH 2496 flags |= UMA_ZONE_NUMA; 2497 #endif 2498 2499 /* This stuff is essential for the zone ctor */ 2500 memset(&args, 0, sizeof(args)); 2501 args.name = name; 2502 args.size = size; 2503 args.ctor = ctor; 2504 args.dtor = dtor; 2505 args.uminit = uminit; 2506 args.fini = fini; 2507 #ifdef INVARIANTS 2508 /* 2509 * Inject procedures which check for memory use after free if we are 2510 * allowed to scramble the memory while it is not allocated. This 2511 * requires that: UMA is actually able to access the memory, no init 2512 * or fini procedures, no dependency on the initial value of the 2513 * memory, and no (legitimate) use of the memory after free. Note, 2514 * the ctor and dtor do not need to be empty. 2515 * 2516 * XXX UMA_ZONE_OFFPAGE. 2517 */ 2518 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) && 2519 uminit == NULL && fini == NULL) { 2520 args.uminit = trash_init; 2521 args.fini = trash_fini; 2522 } 2523 #endif 2524 args.align = align; 2525 args.flags = flags; 2526 args.keg = NULL; 2527 2528 if (booted < BOOT_BUCKETS) { 2529 locked = false; 2530 } else { 2531 sx_slock(&uma_reclaim_lock); 2532 locked = true; 2533 } 2534 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2535 if (locked) 2536 sx_sunlock(&uma_reclaim_lock); 2537 return (res); 2538 } 2539 2540 /* See uma.h */ 2541 uma_zone_t 2542 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 2543 uma_init zinit, uma_fini zfini, uma_zone_t master) 2544 { 2545 struct uma_zctor_args args; 2546 uma_keg_t keg; 2547 uma_zone_t res; 2548 bool locked; 2549 2550 keg = master->uz_keg; 2551 memset(&args, 0, sizeof(args)); 2552 args.name = name; 2553 args.size = keg->uk_size; 2554 args.ctor = ctor; 2555 args.dtor = dtor; 2556 args.uminit = zinit; 2557 args.fini = zfini; 2558 args.align = keg->uk_align; 2559 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 2560 args.keg = keg; 2561 2562 if (booted < BOOT_BUCKETS) { 2563 locked = false; 2564 } else { 2565 sx_slock(&uma_reclaim_lock); 2566 locked = true; 2567 } 2568 /* XXX Attaches only one keg of potentially many. */ 2569 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); 2570 if (locked) 2571 sx_sunlock(&uma_reclaim_lock); 2572 return (res); 2573 } 2574 2575 /* See uma.h */ 2576 uma_zone_t 2577 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 2578 uma_init zinit, uma_fini zfini, uma_import zimport, 2579 uma_release zrelease, void *arg, int flags) 2580 { 2581 struct uma_zctor_args args; 2582 2583 memset(&args, 0, sizeof(args)); 2584 args.name = name; 2585 args.size = size; 2586 args.ctor = ctor; 2587 args.dtor = dtor; 2588 args.uminit = zinit; 2589 args.fini = zfini; 2590 args.import = zimport; 2591 args.release = zrelease; 2592 args.arg = arg; 2593 args.align = 0; 2594 args.flags = flags | UMA_ZFLAG_CACHE; 2595 2596 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK)); 2597 } 2598 2599 /* See uma.h */ 2600 void 2601 uma_zdestroy(uma_zone_t zone) 2602 { 2603 2604 sx_slock(&uma_reclaim_lock); 2605 zone_free_item(zones, zone, NULL, SKIP_NONE); 2606 sx_sunlock(&uma_reclaim_lock); 2607 } 2608 2609 void 2610 uma_zwait(uma_zone_t zone) 2611 { 2612 void *item; 2613 2614 item = uma_zalloc_arg(zone, NULL, M_WAITOK); 2615 uma_zfree(zone, item); 2616 } 2617 2618 void * 2619 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags) 2620 { 2621 void *item; 2622 #ifdef SMP 2623 int i; 2624 2625 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2626 #endif 2627 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO); 2628 if (item != NULL && (flags & M_ZERO)) { 2629 #ifdef SMP 2630 for (i = 0; i <= mp_maxid; i++) 2631 bzero(zpcpu_get_cpu(item, i), zone->uz_size); 2632 #else 2633 bzero(item, zone->uz_size); 2634 #endif 2635 } 2636 return (item); 2637 } 2638 2639 /* 2640 * A stub while both regular and pcpu cases are identical. 2641 */ 2642 void 2643 uma_zfree_pcpu_arg(uma_zone_t zone, void *item, void *udata) 2644 { 2645 2646 #ifdef SMP 2647 MPASS(zone->uz_flags & UMA_ZONE_PCPU); 2648 #endif 2649 uma_zfree_arg(zone, item, udata); 2650 } 2651 2652 static inline void * 2653 bucket_pop(uma_zone_t zone, uma_cache_t cache, uma_bucket_t bucket) 2654 { 2655 void *item; 2656 2657 bucket->ub_cnt--; 2658 item = bucket->ub_bucket[bucket->ub_cnt]; 2659 #ifdef INVARIANTS 2660 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2661 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2662 #endif 2663 cache->uc_allocs++; 2664 2665 return (item); 2666 } 2667 2668 static inline void 2669 bucket_push(uma_zone_t zone, uma_cache_t cache, uma_bucket_t bucket, 2670 void *item) 2671 { 2672 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 2673 ("uma_zfree: Freeing to non free bucket index.")); 2674 bucket->ub_bucket[bucket->ub_cnt] = item; 2675 bucket->ub_cnt++; 2676 cache->uc_frees++; 2677 } 2678 2679 static void * 2680 item_ctor(uma_zone_t zone, void *udata, int flags, void *item) 2681 { 2682 #ifdef INVARIANTS 2683 bool skipdbg; 2684 2685 skipdbg = uma_dbg_zskip(zone, item); 2686 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 && 2687 zone->uz_ctor != trash_ctor) 2688 trash_ctor(item, zone->uz_size, udata, flags); 2689 #endif 2690 if (__predict_false(zone->uz_ctor != NULL) && 2691 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2692 counter_u64_add(zone->uz_fails, 1); 2693 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT); 2694 return (NULL); 2695 } 2696 #ifdef INVARIANTS 2697 if (!skipdbg) 2698 uma_dbg_alloc(zone, NULL, item); 2699 #endif 2700 if (flags & M_ZERO) 2701 uma_zero_item(item, zone); 2702 2703 return (item); 2704 } 2705 2706 static inline void 2707 item_dtor(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 2708 { 2709 #ifdef INVARIANTS 2710 bool skipdbg; 2711 2712 skipdbg = uma_dbg_zskip(zone, item); 2713 if (skip == SKIP_NONE && !skipdbg) { 2714 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0) 2715 uma_dbg_free(zone, udata, item); 2716 else 2717 uma_dbg_free(zone, NULL, item); 2718 } 2719 #endif 2720 if (skip < SKIP_DTOR) { 2721 if (zone->uz_dtor != NULL) 2722 zone->uz_dtor(item, zone->uz_size, udata); 2723 #ifdef INVARIANTS 2724 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 && 2725 zone->uz_dtor != trash_dtor) 2726 trash_dtor(item, zone->uz_size, udata); 2727 #endif 2728 } 2729 } 2730 2731 /* See uma.h */ 2732 void * 2733 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2734 { 2735 uma_bucket_t bucket; 2736 uma_cache_t cache; 2737 void *item; 2738 int cpu, domain; 2739 2740 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2741 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 2742 2743 /* This is the fast path allocation */ 2744 CTR4(KTR_UMA, "uma_zalloc_arg thread %x zone %s(%p) flags %d", 2745 curthread, zone->uz_name, zone, flags); 2746 2747 if (flags & M_WAITOK) { 2748 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2749 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2750 } 2751 KASSERT((flags & M_EXEC) == 0, ("uma_zalloc_arg: called with M_EXEC")); 2752 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2753 ("uma_zalloc_arg: called with spinlock or critical section held")); 2754 if (zone->uz_flags & UMA_ZONE_PCPU) 2755 KASSERT((flags & M_ZERO) == 0, ("allocating from a pcpu zone " 2756 "with M_ZERO passed")); 2757 2758 #ifdef DEBUG_MEMGUARD 2759 if (memguard_cmp_zone(zone)) { 2760 item = memguard_alloc(zone->uz_size, flags); 2761 if (item != NULL) { 2762 if (zone->uz_init != NULL && 2763 zone->uz_init(item, zone->uz_size, flags) != 0) 2764 return (NULL); 2765 if (zone->uz_ctor != NULL && 2766 zone->uz_ctor(item, zone->uz_size, udata, 2767 flags) != 0) { 2768 counter_u64_add(zone->uz_fails, 1); 2769 zone->uz_fini(item, zone->uz_size); 2770 return (NULL); 2771 } 2772 return (item); 2773 } 2774 /* This is unfortunate but should not be fatal. */ 2775 } 2776 #endif 2777 /* 2778 * If possible, allocate from the per-CPU cache. There are two 2779 * requirements for safe access to the per-CPU cache: (1) the thread 2780 * accessing the cache must not be preempted or yield during access, 2781 * and (2) the thread must not migrate CPUs without switching which 2782 * cache it accesses. We rely on a critical section to prevent 2783 * preemption and migration. We release the critical section in 2784 * order to acquire the zone mutex if we are unable to allocate from 2785 * the current cache; when we re-acquire the critical section, we 2786 * must detect and handle migration if it has occurred. 2787 */ 2788 critical_enter(); 2789 do { 2790 cpu = curcpu; 2791 cache = &zone->uz_cpu[cpu]; 2792 bucket = cache->uc_allocbucket; 2793 if (__predict_true(bucket != NULL && bucket->ub_cnt != 0)) { 2794 item = bucket_pop(zone, cache, bucket); 2795 critical_exit(); 2796 return (item_ctor(zone, udata, flags, item)); 2797 } 2798 } while (cache_alloc(zone, cache, udata, flags)); 2799 critical_exit(); 2800 2801 /* 2802 * We can not get a bucket so try to return a single item. 2803 */ 2804 if (zone->uz_flags & UMA_ZONE_NUMA) 2805 domain = PCPU_GET(domain); 2806 else 2807 domain = UMA_ANYDOMAIN; 2808 return (zone_alloc_item_locked(zone, udata, domain, flags)); 2809 } 2810 2811 /* 2812 * Replenish an alloc bucket and possibly restore an old one. Called in 2813 * a critical section. Returns in a critical section. 2814 * 2815 * A false return value indicates failure and returns with the zone lock 2816 * held. A true return value indicates success and the caller should retry. 2817 */ 2818 static __noinline bool 2819 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags) 2820 { 2821 uma_zone_domain_t zdom; 2822 uma_bucket_t bucket; 2823 int cpu, domain; 2824 bool lockfail; 2825 2826 CRITICAL_ASSERT(curthread); 2827 2828 /* 2829 * If we have run out of items in our alloc bucket see 2830 * if we can switch with the free bucket. 2831 */ 2832 bucket = cache->uc_freebucket; 2833 if (bucket != NULL && bucket->ub_cnt != 0) { 2834 cache->uc_freebucket = cache->uc_allocbucket; 2835 cache->uc_allocbucket = bucket; 2836 return (true); 2837 } 2838 2839 /* 2840 * Discard any empty allocation bucket while we hold no locks. 2841 */ 2842 bucket = cache->uc_allocbucket; 2843 cache->uc_allocbucket = NULL; 2844 critical_exit(); 2845 if (bucket != NULL) 2846 bucket_free(zone, bucket, udata); 2847 2848 /* 2849 * Attempt to retrieve the item from the per-CPU cache has failed, so 2850 * we must go back to the zone. This requires the zone lock, so we 2851 * must drop the critical section, then re-acquire it when we go back 2852 * to the cache. Since the critical section is released, we may be 2853 * preempted or migrate. As such, make sure not to maintain any 2854 * thread-local state specific to the cache from prior to releasing 2855 * the critical section. 2856 */ 2857 lockfail = 0; 2858 if (ZONE_TRYLOCK(zone) == 0) { 2859 /* Record contention to size the buckets. */ 2860 ZONE_LOCK(zone); 2861 lockfail = 1; 2862 } 2863 2864 critical_enter(); 2865 /* Short-circuit for zones without buckets and low memory. */ 2866 if (zone->uz_bucket_size == 0 || bucketdisable) 2867 return (false); 2868 2869 cpu = curcpu; 2870 cache = &zone->uz_cpu[cpu]; 2871 2872 /* See if we lost the race to fill the cache. */ 2873 if (cache->uc_allocbucket != NULL) { 2874 ZONE_UNLOCK(zone); 2875 return (true); 2876 } 2877 2878 /* 2879 * Check the zone's cache of buckets. 2880 */ 2881 if (zone->uz_flags & UMA_ZONE_NUMA) { 2882 domain = PCPU_GET(domain); 2883 zdom = &zone->uz_domain[domain]; 2884 } else { 2885 domain = UMA_ANYDOMAIN; 2886 zdom = &zone->uz_domain[0]; 2887 } 2888 2889 if ((bucket = zone_fetch_bucket(zone, zdom)) != NULL) { 2890 ZONE_UNLOCK(zone); 2891 KASSERT(bucket->ub_cnt != 0, 2892 ("uma_zalloc_arg: Returning an empty bucket.")); 2893 cache->uc_allocbucket = bucket; 2894 return (true); 2895 } 2896 /* We are no longer associated with this CPU. */ 2897 critical_exit(); 2898 2899 /* 2900 * We bump the uz count when the cache size is insufficient to 2901 * handle the working set. 2902 */ 2903 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max) 2904 zone->uz_bucket_size++; 2905 2906 /* 2907 * Fill a bucket and attempt to use it as the alloc bucket. 2908 */ 2909 bucket = zone_alloc_bucket(zone, udata, domain, flags); 2910 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p", 2911 zone->uz_name, zone, bucket); 2912 critical_enter(); 2913 if (bucket == NULL) 2914 return (false); 2915 2916 /* 2917 * See if we lost the race or were migrated. Cache the 2918 * initialized bucket to make this less likely or claim 2919 * the memory directly. 2920 */ 2921 cpu = curcpu; 2922 cache = &zone->uz_cpu[cpu]; 2923 if (cache->uc_allocbucket == NULL && 2924 ((zone->uz_flags & UMA_ZONE_NUMA) == 0 || 2925 domain == PCPU_GET(domain))) { 2926 cache->uc_allocbucket = bucket; 2927 zdom->uzd_imax += bucket->ub_cnt; 2928 } else if (zone->uz_bkt_count >= zone->uz_bkt_max) { 2929 critical_exit(); 2930 ZONE_UNLOCK(zone); 2931 bucket_drain(zone, bucket); 2932 bucket_free(zone, bucket, udata); 2933 critical_enter(); 2934 return (true); 2935 } else 2936 zone_put_bucket(zone, zdom, bucket, false); 2937 ZONE_UNLOCK(zone); 2938 return (true); 2939 } 2940 2941 void * 2942 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags) 2943 { 2944 2945 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2946 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 2947 2948 /* This is the fast path allocation */ 2949 CTR5(KTR_UMA, 2950 "uma_zalloc_domain thread %x zone %s(%p) domain %d flags %d", 2951 curthread, zone->uz_name, zone, domain, flags); 2952 2953 if (flags & M_WAITOK) { 2954 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2955 "uma_zalloc_domain: zone \"%s\"", zone->uz_name); 2956 } 2957 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2958 ("uma_zalloc_domain: called with spinlock or critical section held")); 2959 2960 return (zone_alloc_item(zone, udata, domain, flags)); 2961 } 2962 2963 /* 2964 * Find a slab with some space. Prefer slabs that are partially used over those 2965 * that are totally full. This helps to reduce fragmentation. 2966 * 2967 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check 2968 * only 'domain'. 2969 */ 2970 static uma_slab_t 2971 keg_first_slab(uma_keg_t keg, int domain, bool rr) 2972 { 2973 uma_domain_t dom; 2974 uma_slab_t slab; 2975 int start; 2976 2977 KASSERT(domain >= 0 && domain < vm_ndomains, 2978 ("keg_first_slab: domain %d out of range", domain)); 2979 KEG_LOCK_ASSERT(keg); 2980 2981 slab = NULL; 2982 start = domain; 2983 do { 2984 dom = &keg->uk_domain[domain]; 2985 if (!LIST_EMPTY(&dom->ud_part_slab)) 2986 return (LIST_FIRST(&dom->ud_part_slab)); 2987 if (!LIST_EMPTY(&dom->ud_free_slab)) { 2988 slab = LIST_FIRST(&dom->ud_free_slab); 2989 LIST_REMOVE(slab, us_link); 2990 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 2991 return (slab); 2992 } 2993 if (rr) 2994 domain = (domain + 1) % vm_ndomains; 2995 } while (domain != start); 2996 2997 return (NULL); 2998 } 2999 3000 static uma_slab_t 3001 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags) 3002 { 3003 uint32_t reserve; 3004 3005 KEG_LOCK_ASSERT(keg); 3006 3007 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve; 3008 if (keg->uk_free <= reserve) 3009 return (NULL); 3010 return (keg_first_slab(keg, domain, rr)); 3011 } 3012 3013 static uma_slab_t 3014 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags) 3015 { 3016 struct vm_domainset_iter di; 3017 uma_domain_t dom; 3018 uma_slab_t slab; 3019 int aflags, domain; 3020 bool rr; 3021 3022 restart: 3023 KEG_LOCK_ASSERT(keg); 3024 3025 /* 3026 * Use the keg's policy if upper layers haven't already specified a 3027 * domain (as happens with first-touch zones). 3028 * 3029 * To avoid races we run the iterator with the keg lock held, but that 3030 * means that we cannot allow the vm_domainset layer to sleep. Thus, 3031 * clear M_WAITOK and handle low memory conditions locally. 3032 */ 3033 rr = rdomain == UMA_ANYDOMAIN; 3034 if (rr) { 3035 aflags = (flags & ~M_WAITOK) | M_NOWAIT; 3036 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 3037 &aflags); 3038 } else { 3039 aflags = flags; 3040 domain = rdomain; 3041 } 3042 3043 for (;;) { 3044 slab = keg_fetch_free_slab(keg, domain, rr, flags); 3045 if (slab != NULL) 3046 return (slab); 3047 3048 /* 3049 * M_NOVM means don't ask at all! 3050 */ 3051 if (flags & M_NOVM) 3052 break; 3053 3054 KASSERT(zone->uz_max_items == 0 || 3055 zone->uz_items <= zone->uz_max_items, 3056 ("%s: zone %p overflow", __func__, zone)); 3057 3058 slab = keg_alloc_slab(keg, zone, domain, flags, aflags); 3059 /* 3060 * If we got a slab here it's safe to mark it partially used 3061 * and return. We assume that the caller is going to remove 3062 * at least one item. 3063 */ 3064 if (slab) { 3065 dom = &keg->uk_domain[slab->us_domain]; 3066 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 3067 return (slab); 3068 } 3069 KEG_LOCK(keg); 3070 if (rr && vm_domainset_iter_policy(&di, &domain) != 0) { 3071 if ((flags & M_WAITOK) != 0) { 3072 KEG_UNLOCK(keg); 3073 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask); 3074 KEG_LOCK(keg); 3075 goto restart; 3076 } 3077 break; 3078 } 3079 } 3080 3081 /* 3082 * We might not have been able to get a slab but another cpu 3083 * could have while we were unlocked. Check again before we 3084 * fail. 3085 */ 3086 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL) { 3087 return (slab); 3088 } 3089 return (NULL); 3090 } 3091 3092 static void * 3093 slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 3094 { 3095 uma_domain_t dom; 3096 void *item; 3097 uint8_t freei; 3098 3099 KEG_LOCK_ASSERT(keg); 3100 3101 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1; 3102 BIT_CLR(keg->uk_ipers, freei, &slab->us_free); 3103 item = slab_item(slab, keg, freei); 3104 slab->us_freecount--; 3105 keg->uk_free--; 3106 3107 /* Move this slab to the full list */ 3108 if (slab->us_freecount == 0) { 3109 LIST_REMOVE(slab, us_link); 3110 dom = &keg->uk_domain[slab->us_domain]; 3111 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link); 3112 } 3113 3114 return (item); 3115 } 3116 3117 static int 3118 zone_import(void *arg, void **bucket, int max, int domain, int flags) 3119 { 3120 uma_zone_t zone; 3121 uma_slab_t slab; 3122 uma_keg_t keg; 3123 #ifdef NUMA 3124 int stripe; 3125 #endif 3126 int i; 3127 3128 zone = arg; 3129 slab = NULL; 3130 keg = zone->uz_keg; 3131 KEG_LOCK(keg); 3132 /* Try to keep the buckets totally full */ 3133 for (i = 0; i < max; ) { 3134 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL) 3135 break; 3136 #ifdef NUMA 3137 stripe = howmany(max, vm_ndomains); 3138 #endif 3139 while (slab->us_freecount && i < max) { 3140 bucket[i++] = slab_alloc_item(keg, slab); 3141 if (keg->uk_free <= keg->uk_reserve) 3142 break; 3143 #ifdef NUMA 3144 /* 3145 * If the zone is striped we pick a new slab for every 3146 * N allocations. Eliminating this conditional will 3147 * instead pick a new domain for each bucket rather 3148 * than stripe within each bucket. The current option 3149 * produces more fragmentation and requires more cpu 3150 * time but yields better distribution. 3151 */ 3152 if ((zone->uz_flags & UMA_ZONE_NUMA) == 0 && 3153 vm_ndomains > 1 && --stripe == 0) 3154 break; 3155 #endif 3156 } 3157 /* Don't block if we allocated any successfully. */ 3158 flags &= ~M_WAITOK; 3159 flags |= M_NOWAIT; 3160 } 3161 KEG_UNLOCK(keg); 3162 3163 return i; 3164 } 3165 3166 static uma_bucket_t 3167 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags) 3168 { 3169 uma_bucket_t bucket; 3170 int maxbucket, cnt; 3171 3172 CTR1(KTR_UMA, "zone_alloc:_bucket domain %d)", domain); 3173 3174 /* Avoid allocs targeting empty domains. */ 3175 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain)) 3176 domain = UMA_ANYDOMAIN; 3177 3178 if (zone->uz_max_items > 0) { 3179 if (zone->uz_items >= zone->uz_max_items) 3180 return (false); 3181 maxbucket = MIN(zone->uz_bucket_size, 3182 zone->uz_max_items - zone->uz_items); 3183 zone->uz_items += maxbucket; 3184 } else 3185 maxbucket = zone->uz_bucket_size; 3186 ZONE_UNLOCK(zone); 3187 3188 /* Don't wait for buckets, preserve caller's NOVM setting. */ 3189 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 3190 if (bucket == NULL) { 3191 cnt = 0; 3192 goto out; 3193 } 3194 3195 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 3196 MIN(maxbucket, bucket->ub_entries), domain, flags); 3197 3198 /* 3199 * Initialize the memory if necessary. 3200 */ 3201 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 3202 int i; 3203 3204 for (i = 0; i < bucket->ub_cnt; i++) 3205 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 3206 flags) != 0) 3207 break; 3208 /* 3209 * If we couldn't initialize the whole bucket, put the 3210 * rest back onto the freelist. 3211 */ 3212 if (i != bucket->ub_cnt) { 3213 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 3214 bucket->ub_cnt - i); 3215 #ifdef INVARIANTS 3216 bzero(&bucket->ub_bucket[i], 3217 sizeof(void *) * (bucket->ub_cnt - i)); 3218 #endif 3219 bucket->ub_cnt = i; 3220 } 3221 } 3222 3223 cnt = bucket->ub_cnt; 3224 if (bucket->ub_cnt == 0) { 3225 bucket_free(zone, bucket, udata); 3226 counter_u64_add(zone->uz_fails, 1); 3227 bucket = NULL; 3228 } 3229 out: 3230 ZONE_LOCK(zone); 3231 if (zone->uz_max_items > 0 && cnt < maxbucket) { 3232 MPASS(zone->uz_items >= maxbucket - cnt); 3233 zone->uz_items -= maxbucket - cnt; 3234 if (zone->uz_sleepers > 0 && 3235 (cnt == 0 ? zone->uz_items + 1 : zone->uz_items) < 3236 zone->uz_max_items) 3237 wakeup_one(zone); 3238 } 3239 3240 return (bucket); 3241 } 3242 3243 /* 3244 * Allocates a single item from a zone. 3245 * 3246 * Arguments 3247 * zone The zone to alloc for. 3248 * udata The data to be passed to the constructor. 3249 * domain The domain to allocate from or UMA_ANYDOMAIN. 3250 * flags M_WAITOK, M_NOWAIT, M_ZERO. 3251 * 3252 * Returns 3253 * NULL if there is no memory and M_NOWAIT is set 3254 * An item if successful 3255 */ 3256 3257 static void * 3258 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags) 3259 { 3260 3261 ZONE_LOCK(zone); 3262 return (zone_alloc_item_locked(zone, udata, domain, flags)); 3263 } 3264 3265 /* 3266 * Returns with zone unlocked. 3267 */ 3268 static void * 3269 zone_alloc_item_locked(uma_zone_t zone, void *udata, int domain, int flags) 3270 { 3271 void *item; 3272 3273 ZONE_LOCK_ASSERT(zone); 3274 3275 if (zone->uz_max_items > 0) { 3276 if (zone->uz_items >= zone->uz_max_items) { 3277 zone_log_warning(zone); 3278 zone_maxaction(zone); 3279 if (flags & M_NOWAIT) { 3280 ZONE_UNLOCK(zone); 3281 return (NULL); 3282 } 3283 zone->uz_sleeps++; 3284 zone->uz_sleepers++; 3285 while (zone->uz_items >= zone->uz_max_items) 3286 mtx_sleep(zone, zone->uz_lockptr, PVM, 3287 "zonelimit", 0); 3288 zone->uz_sleepers--; 3289 if (zone->uz_sleepers > 0 && 3290 zone->uz_items + 1 < zone->uz_max_items) 3291 wakeup_one(zone); 3292 } 3293 zone->uz_items++; 3294 } 3295 ZONE_UNLOCK(zone); 3296 3297 /* Avoid allocs targeting empty domains. */ 3298 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain)) 3299 domain = UMA_ANYDOMAIN; 3300 3301 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1) 3302 goto fail_cnt; 3303 3304 /* 3305 * We have to call both the zone's init (not the keg's init) 3306 * and the zone's ctor. This is because the item is going from 3307 * a keg slab directly to the user, and the user is expecting it 3308 * to be both zone-init'd as well as zone-ctor'd. 3309 */ 3310 if (zone->uz_init != NULL) { 3311 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 3312 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT); 3313 goto fail_cnt; 3314 } 3315 } 3316 item = item_ctor(zone, udata, flags, item); 3317 if (item == NULL) 3318 goto fail; 3319 3320 counter_u64_add(zone->uz_allocs, 1); 3321 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item, 3322 zone->uz_name, zone); 3323 3324 return (item); 3325 3326 fail_cnt: 3327 counter_u64_add(zone->uz_fails, 1); 3328 fail: 3329 if (zone->uz_max_items > 0) { 3330 ZONE_LOCK(zone); 3331 /* XXX Decrement without wakeup */ 3332 zone->uz_items--; 3333 ZONE_UNLOCK(zone); 3334 } 3335 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)", 3336 zone->uz_name, zone); 3337 return (NULL); 3338 } 3339 3340 /* See uma.h */ 3341 void 3342 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 3343 { 3344 uma_cache_t cache; 3345 uma_bucket_t bucket; 3346 int cpu, domain, itemdomain; 3347 3348 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3349 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3350 3351 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 3352 zone->uz_name); 3353 3354 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3355 ("uma_zfree_arg: called with spinlock or critical section held")); 3356 3357 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 3358 if (item == NULL) 3359 return; 3360 #ifdef DEBUG_MEMGUARD 3361 if (is_memguard_addr(item)) { 3362 if (zone->uz_dtor != NULL) 3363 zone->uz_dtor(item, zone->uz_size, udata); 3364 if (zone->uz_fini != NULL) 3365 zone->uz_fini(item, zone->uz_size); 3366 memguard_free(item); 3367 return; 3368 } 3369 #endif 3370 item_dtor(zone, item, udata, SKIP_NONE); 3371 3372 /* 3373 * The race here is acceptable. If we miss it we'll just have to wait 3374 * a little longer for the limits to be reset. 3375 */ 3376 if (zone->uz_sleepers > 0) 3377 goto zfree_item; 3378 3379 /* 3380 * If possible, free to the per-CPU cache. There are two 3381 * requirements for safe access to the per-CPU cache: (1) the thread 3382 * accessing the cache must not be preempted or yield during access, 3383 * and (2) the thread must not migrate CPUs without switching which 3384 * cache it accesses. We rely on a critical section to prevent 3385 * preemption and migration. We release the critical section in 3386 * order to acquire the zone mutex if we are unable to free to the 3387 * current cache; when we re-acquire the critical section, we must 3388 * detect and handle migration if it has occurred. 3389 */ 3390 domain = itemdomain = 0; 3391 critical_enter(); 3392 do { 3393 cpu = curcpu; 3394 cache = &zone->uz_cpu[cpu]; 3395 bucket = cache->uc_allocbucket; 3396 #ifdef UMA_XDOMAIN 3397 if ((zone->uz_flags & UMA_ZONE_NUMA) != 0) { 3398 itemdomain = _vm_phys_domain(pmap_kextract((vm_offset_t)item)); 3399 domain = PCPU_GET(domain); 3400 } 3401 if ((zone->uz_flags & UMA_ZONE_NUMA) != 0 && 3402 domain != itemdomain) { 3403 bucket = cache->uc_crossbucket; 3404 } else 3405 #endif 3406 3407 /* 3408 * Try to free into the allocbucket first to give LIFO ordering 3409 * for cache-hot datastructures. Spill over into the freebucket 3410 * if necessary. Alloc will swap them if one runs dry. 3411 */ 3412 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 3413 bucket = cache->uc_freebucket; 3414 if (__predict_true(bucket != NULL && 3415 bucket->ub_cnt < bucket->ub_entries)) { 3416 bucket_push(zone, cache, bucket, item); 3417 critical_exit(); 3418 return; 3419 } 3420 } while (cache_free(zone, cache, udata, item, itemdomain)); 3421 critical_exit(); 3422 3423 /* 3424 * If nothing else caught this, we'll just do an internal free. 3425 */ 3426 zfree_item: 3427 zone_free_item(zone, item, udata, SKIP_DTOR); 3428 } 3429 3430 static void 3431 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata, 3432 int domain, int itemdomain) 3433 { 3434 uma_zone_domain_t zdom; 3435 3436 #ifdef UMA_XDOMAIN 3437 /* 3438 * Buckets coming from the wrong domain will be entirely for the 3439 * only other domain on two domain systems. In this case we can 3440 * simply cache them. Otherwise we need to sort them back to 3441 * correct domains by freeing the contents to the slab layer. 3442 */ 3443 if (domain != itemdomain && vm_ndomains > 2) { 3444 CTR3(KTR_UMA, 3445 "uma_zfree: zone %s(%p) draining cross bucket %p", 3446 zone->uz_name, zone, bucket); 3447 bucket_drain(zone, bucket); 3448 bucket_free(zone, bucket, udata); 3449 return; 3450 } 3451 #endif 3452 /* 3453 * Attempt to save the bucket in the zone's domain bucket cache. 3454 * 3455 * We bump the uz count when the cache size is insufficient to 3456 * handle the working set. 3457 */ 3458 if (ZONE_TRYLOCK(zone) == 0) { 3459 /* Record contention to size the buckets. */ 3460 ZONE_LOCK(zone); 3461 if (zone->uz_bucket_size < zone->uz_bucket_size_max) 3462 zone->uz_bucket_size++; 3463 } 3464 3465 CTR3(KTR_UMA, 3466 "uma_zfree: zone %s(%p) putting bucket %p on free list", 3467 zone->uz_name, zone, bucket); 3468 /* ub_cnt is pointing to the last free item */ 3469 KASSERT(bucket->ub_cnt == bucket->ub_entries, 3470 ("uma_zfree: Attempting to insert partial bucket onto the full list.\n")); 3471 if (zone->uz_bkt_count >= zone->uz_bkt_max) { 3472 ZONE_UNLOCK(zone); 3473 bucket_drain(zone, bucket); 3474 bucket_free(zone, bucket, udata); 3475 } else { 3476 zdom = &zone->uz_domain[itemdomain]; 3477 zone_put_bucket(zone, zdom, bucket, true); 3478 ZONE_UNLOCK(zone); 3479 } 3480 } 3481 3482 /* 3483 * Populate a free or cross bucket for the current cpu cache. Free any 3484 * existing full bucket either to the zone cache or back to the slab layer. 3485 * 3486 * Enters and returns in a critical section. false return indicates that 3487 * we can not satisfy this free in the cache layer. true indicates that 3488 * the caller should retry. 3489 */ 3490 static __noinline bool 3491 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item, 3492 int itemdomain) 3493 { 3494 uma_bucket_t bucket; 3495 int cpu, domain; 3496 3497 CRITICAL_ASSERT(curthread); 3498 3499 if (zone->uz_bucket_size == 0 || bucketdisable) 3500 return false; 3501 3502 cpu = curcpu; 3503 cache = &zone->uz_cpu[cpu]; 3504 3505 /* 3506 * NUMA domains need to free to the correct zdom. When XDOMAIN 3507 * is enabled this is the zdom of the item and the bucket may be 3508 * the cross bucket if they do not match. 3509 */ 3510 if ((zone->uz_flags & UMA_ZONE_NUMA) != 0) 3511 #ifdef UMA_XDOMAIN 3512 domain = PCPU_GET(domain); 3513 #else 3514 itemdomain = domain = PCPU_GET(domain); 3515 #endif 3516 else 3517 itemdomain = domain = 0; 3518 #ifdef UMA_XDOMAIN 3519 if (domain != itemdomain) { 3520 bucket = cache->uc_crossbucket; 3521 cache->uc_crossbucket = NULL; 3522 if (bucket != NULL) 3523 atomic_add_64(&zone->uz_xdomain, bucket->ub_cnt); 3524 } else 3525 #endif 3526 { 3527 bucket = cache->uc_freebucket; 3528 cache->uc_freebucket = NULL; 3529 } 3530 3531 3532 /* We are no longer associated with this CPU. */ 3533 critical_exit(); 3534 3535 if (bucket != NULL) 3536 zone_free_bucket(zone, bucket, udata, domain, itemdomain); 3537 3538 bucket = bucket_alloc(zone, udata, M_NOWAIT); 3539 CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p", 3540 zone->uz_name, zone, bucket); 3541 critical_enter(); 3542 if (bucket == NULL) 3543 return (false); 3544 cpu = curcpu; 3545 cache = &zone->uz_cpu[cpu]; 3546 #ifdef UMA_XDOMAIN 3547 /* 3548 * Check to see if we should be populating the cross bucket. If it 3549 * is already populated we will fall through and attempt to populate 3550 * the free bucket. 3551 */ 3552 if ((zone->uz_flags & UMA_ZONE_NUMA) != 0) { 3553 domain = PCPU_GET(domain); 3554 if (domain != itemdomain && cache->uc_crossbucket == NULL) { 3555 cache->uc_crossbucket = bucket; 3556 return (true); 3557 } 3558 } 3559 #endif 3560 /* 3561 * We may have lost the race to fill the bucket or switched CPUs. 3562 */ 3563 if (cache->uc_freebucket != NULL) { 3564 critical_exit(); 3565 bucket_free(zone, bucket, udata); 3566 critical_enter(); 3567 } else 3568 cache->uc_freebucket = bucket; 3569 3570 return (true); 3571 } 3572 3573 void 3574 uma_zfree_domain(uma_zone_t zone, void *item, void *udata) 3575 { 3576 3577 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 3578 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); 3579 3580 CTR2(KTR_UMA, "uma_zfree_domain thread %x zone %s", curthread, 3581 zone->uz_name); 3582 3583 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 3584 ("uma_zfree_domain: called with spinlock or critical section held")); 3585 3586 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 3587 if (item == NULL) 3588 return; 3589 zone_free_item(zone, item, udata, SKIP_NONE); 3590 } 3591 3592 static void 3593 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item) 3594 { 3595 uma_keg_t keg; 3596 uma_domain_t dom; 3597 uint8_t freei; 3598 3599 keg = zone->uz_keg; 3600 MPASS(zone->uz_lockptr == &keg->uk_lock); 3601 KEG_LOCK_ASSERT(keg); 3602 3603 dom = &keg->uk_domain[slab->us_domain]; 3604 3605 /* Do we need to remove from any lists? */ 3606 if (slab->us_freecount+1 == keg->uk_ipers) { 3607 LIST_REMOVE(slab, us_link); 3608 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); 3609 } else if (slab->us_freecount == 0) { 3610 LIST_REMOVE(slab, us_link); 3611 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); 3612 } 3613 3614 /* Slab management. */ 3615 freei = slab_item_index(slab, keg, item); 3616 BIT_SET(keg->uk_ipers, freei, &slab->us_free); 3617 slab->us_freecount++; 3618 3619 /* Keg statistics. */ 3620 keg->uk_free++; 3621 } 3622 3623 static void 3624 zone_release(void *arg, void **bucket, int cnt) 3625 { 3626 uma_zone_t zone; 3627 void *item; 3628 uma_slab_t slab; 3629 uma_keg_t keg; 3630 uint8_t *mem; 3631 int i; 3632 3633 zone = arg; 3634 keg = zone->uz_keg; 3635 KEG_LOCK(keg); 3636 for (i = 0; i < cnt; i++) { 3637 item = bucket[i]; 3638 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 3639 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 3640 if (zone->uz_flags & UMA_ZONE_HASH) { 3641 slab = hash_sfind(&keg->uk_hash, mem); 3642 } else { 3643 mem += keg->uk_pgoff; 3644 slab = (uma_slab_t)mem; 3645 } 3646 } else 3647 slab = vtoslab((vm_offset_t)item); 3648 slab_free_item(zone, slab, item); 3649 } 3650 KEG_UNLOCK(keg); 3651 } 3652 3653 /* 3654 * Frees a single item to any zone. 3655 * 3656 * Arguments: 3657 * zone The zone to free to 3658 * item The item we're freeing 3659 * udata User supplied data for the dtor 3660 * skip Skip dtors and finis 3661 */ 3662 static void 3663 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 3664 { 3665 3666 item_dtor(zone, item, udata, skip); 3667 3668 if (skip < SKIP_FINI && zone->uz_fini) 3669 zone->uz_fini(item, zone->uz_size); 3670 3671 zone->uz_release(zone->uz_arg, &item, 1); 3672 3673 if (skip & SKIP_CNT) 3674 return; 3675 3676 counter_u64_add(zone->uz_frees, 1); 3677 3678 if (zone->uz_max_items > 0) { 3679 ZONE_LOCK(zone); 3680 zone->uz_items--; 3681 if (zone->uz_sleepers > 0 && 3682 zone->uz_items < zone->uz_max_items) 3683 wakeup_one(zone); 3684 ZONE_UNLOCK(zone); 3685 } 3686 } 3687 3688 /* See uma.h */ 3689 int 3690 uma_zone_set_max(uma_zone_t zone, int nitems) 3691 { 3692 struct uma_bucket_zone *ubz; 3693 int count; 3694 3695 ZONE_LOCK(zone); 3696 ubz = bucket_zone_max(zone, nitems); 3697 count = ubz != NULL ? ubz->ubz_entries : 0; 3698 zone->uz_bucket_size_max = zone->uz_bucket_size = count; 3699 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max) 3700 zone->uz_bucket_size_min = zone->uz_bucket_size_max; 3701 zone->uz_max_items = nitems; 3702 ZONE_UNLOCK(zone); 3703 3704 return (nitems); 3705 } 3706 3707 /* See uma.h */ 3708 void 3709 uma_zone_set_maxcache(uma_zone_t zone, int nitems) 3710 { 3711 struct uma_bucket_zone *ubz; 3712 int bpcpu; 3713 3714 ZONE_LOCK(zone); 3715 ubz = bucket_zone_max(zone, nitems); 3716 if (ubz != NULL) { 3717 bpcpu = 2; 3718 #ifdef UMA_XDOMAIN 3719 if ((zone->uz_flags & UMA_ZONE_NUMA) != 0) 3720 /* Count the cross-domain bucket. */ 3721 bpcpu++; 3722 #endif 3723 nitems -= ubz->ubz_entries * bpcpu * mp_ncpus; 3724 zone->uz_bucket_size_max = ubz->ubz_entries; 3725 } else { 3726 zone->uz_bucket_size_max = zone->uz_bucket_size = 0; 3727 } 3728 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max) 3729 zone->uz_bucket_size_min = zone->uz_bucket_size_max; 3730 zone->uz_bkt_max = nitems; 3731 ZONE_UNLOCK(zone); 3732 } 3733 3734 /* See uma.h */ 3735 int 3736 uma_zone_get_max(uma_zone_t zone) 3737 { 3738 int nitems; 3739 3740 ZONE_LOCK(zone); 3741 nitems = zone->uz_max_items; 3742 ZONE_UNLOCK(zone); 3743 3744 return (nitems); 3745 } 3746 3747 /* See uma.h */ 3748 void 3749 uma_zone_set_warning(uma_zone_t zone, const char *warning) 3750 { 3751 3752 ZONE_LOCK(zone); 3753 zone->uz_warning = warning; 3754 ZONE_UNLOCK(zone); 3755 } 3756 3757 /* See uma.h */ 3758 void 3759 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) 3760 { 3761 3762 ZONE_LOCK(zone); 3763 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); 3764 ZONE_UNLOCK(zone); 3765 } 3766 3767 /* See uma.h */ 3768 int 3769 uma_zone_get_cur(uma_zone_t zone) 3770 { 3771 int64_t nitems; 3772 u_int i; 3773 3774 ZONE_LOCK(zone); 3775 nitems = counter_u64_fetch(zone->uz_allocs) - 3776 counter_u64_fetch(zone->uz_frees); 3777 if ((zone->uz_flags & UMA_ZFLAG_INTERNAL) == 0) { 3778 CPU_FOREACH(i) { 3779 /* 3780 * See the comment in uma_vm_zone_stats() regarding 3781 * the safety of accessing the per-cpu caches. With 3782 * the zone lock held, it is safe, but can potentially 3783 * result in stale data. 3784 */ 3785 nitems += zone->uz_cpu[i].uc_allocs - 3786 zone->uz_cpu[i].uc_frees; 3787 } 3788 } 3789 ZONE_UNLOCK(zone); 3790 3791 return (nitems < 0 ? 0 : nitems); 3792 } 3793 3794 static uint64_t 3795 uma_zone_get_allocs(uma_zone_t zone) 3796 { 3797 uint64_t nitems; 3798 u_int i; 3799 3800 ZONE_LOCK(zone); 3801 nitems = counter_u64_fetch(zone->uz_allocs); 3802 if ((zone->uz_flags & UMA_ZFLAG_INTERNAL) == 0) { 3803 CPU_FOREACH(i) { 3804 /* 3805 * See the comment in uma_vm_zone_stats() regarding 3806 * the safety of accessing the per-cpu caches. With 3807 * the zone lock held, it is safe, but can potentially 3808 * result in stale data. 3809 */ 3810 nitems += zone->uz_cpu[i].uc_allocs; 3811 } 3812 } 3813 ZONE_UNLOCK(zone); 3814 3815 return (nitems); 3816 } 3817 3818 static uint64_t 3819 uma_zone_get_frees(uma_zone_t zone) 3820 { 3821 uint64_t nitems; 3822 u_int i; 3823 3824 ZONE_LOCK(zone); 3825 nitems = counter_u64_fetch(zone->uz_frees); 3826 if ((zone->uz_flags & UMA_ZFLAG_INTERNAL) == 0) { 3827 CPU_FOREACH(i) { 3828 /* 3829 * See the comment in uma_vm_zone_stats() regarding 3830 * the safety of accessing the per-cpu caches. With 3831 * the zone lock held, it is safe, but can potentially 3832 * result in stale data. 3833 */ 3834 nitems += zone->uz_cpu[i].uc_frees; 3835 } 3836 } 3837 ZONE_UNLOCK(zone); 3838 3839 return (nitems); 3840 } 3841 3842 /* See uma.h */ 3843 void 3844 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 3845 { 3846 uma_keg_t keg; 3847 3848 KEG_GET(zone, keg); 3849 KEG_LOCK(keg); 3850 KASSERT(keg->uk_pages == 0, 3851 ("uma_zone_set_init on non-empty keg")); 3852 keg->uk_init = uminit; 3853 KEG_UNLOCK(keg); 3854 } 3855 3856 /* See uma.h */ 3857 void 3858 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 3859 { 3860 uma_keg_t keg; 3861 3862 KEG_GET(zone, keg); 3863 KEG_LOCK(keg); 3864 KASSERT(keg->uk_pages == 0, 3865 ("uma_zone_set_fini on non-empty keg")); 3866 keg->uk_fini = fini; 3867 KEG_UNLOCK(keg); 3868 } 3869 3870 /* See uma.h */ 3871 void 3872 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 3873 { 3874 3875 ZONE_LOCK(zone); 3876 KASSERT(zone->uz_keg->uk_pages == 0, 3877 ("uma_zone_set_zinit on non-empty keg")); 3878 zone->uz_init = zinit; 3879 ZONE_UNLOCK(zone); 3880 } 3881 3882 /* See uma.h */ 3883 void 3884 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 3885 { 3886 3887 ZONE_LOCK(zone); 3888 KASSERT(zone->uz_keg->uk_pages == 0, 3889 ("uma_zone_set_zfini on non-empty keg")); 3890 zone->uz_fini = zfini; 3891 ZONE_UNLOCK(zone); 3892 } 3893 3894 /* See uma.h */ 3895 /* XXX uk_freef is not actually used with the zone locked */ 3896 void 3897 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 3898 { 3899 uma_keg_t keg; 3900 3901 KEG_GET(zone, keg); 3902 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type")); 3903 KEG_LOCK(keg); 3904 keg->uk_freef = freef; 3905 KEG_UNLOCK(keg); 3906 } 3907 3908 /* See uma.h */ 3909 /* XXX uk_allocf is not actually used with the zone locked */ 3910 void 3911 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3912 { 3913 uma_keg_t keg; 3914 3915 KEG_GET(zone, keg); 3916 KEG_LOCK(keg); 3917 keg->uk_allocf = allocf; 3918 KEG_UNLOCK(keg); 3919 } 3920 3921 /* See uma.h */ 3922 void 3923 uma_zone_reserve(uma_zone_t zone, int items) 3924 { 3925 uma_keg_t keg; 3926 3927 KEG_GET(zone, keg); 3928 KEG_LOCK(keg); 3929 keg->uk_reserve = items; 3930 KEG_UNLOCK(keg); 3931 } 3932 3933 /* See uma.h */ 3934 int 3935 uma_zone_reserve_kva(uma_zone_t zone, int count) 3936 { 3937 uma_keg_t keg; 3938 vm_offset_t kva; 3939 u_int pages; 3940 3941 KEG_GET(zone, keg); 3942 3943 pages = count / keg->uk_ipers; 3944 if (pages * keg->uk_ipers < count) 3945 pages++; 3946 pages *= keg->uk_ppera; 3947 3948 #ifdef UMA_MD_SMALL_ALLOC 3949 if (keg->uk_ppera > 1) { 3950 #else 3951 if (1) { 3952 #endif 3953 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE); 3954 if (kva == 0) 3955 return (0); 3956 } else 3957 kva = 0; 3958 3959 ZONE_LOCK(zone); 3960 MPASS(keg->uk_kva == 0); 3961 keg->uk_kva = kva; 3962 keg->uk_offset = 0; 3963 zone->uz_max_items = pages * keg->uk_ipers; 3964 #ifdef UMA_MD_SMALL_ALLOC 3965 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3966 #else 3967 keg->uk_allocf = noobj_alloc; 3968 #endif 3969 keg->uk_flags |= UMA_ZONE_NOFREE; 3970 ZONE_UNLOCK(zone); 3971 3972 return (1); 3973 } 3974 3975 /* See uma.h */ 3976 void 3977 uma_prealloc(uma_zone_t zone, int items) 3978 { 3979 struct vm_domainset_iter di; 3980 uma_domain_t dom; 3981 uma_slab_t slab; 3982 uma_keg_t keg; 3983 int aflags, domain, slabs; 3984 3985 KEG_GET(zone, keg); 3986 KEG_LOCK(keg); 3987 slabs = items / keg->uk_ipers; 3988 if (slabs * keg->uk_ipers < items) 3989 slabs++; 3990 while (slabs-- > 0) { 3991 aflags = M_NOWAIT; 3992 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, 3993 &aflags); 3994 for (;;) { 3995 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK, 3996 aflags); 3997 if (slab != NULL) { 3998 dom = &keg->uk_domain[slab->us_domain]; 3999 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, 4000 us_link); 4001 break; 4002 } 4003 KEG_LOCK(keg); 4004 if (vm_domainset_iter_policy(&di, &domain) != 0) { 4005 KEG_UNLOCK(keg); 4006 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask); 4007 KEG_LOCK(keg); 4008 } 4009 } 4010 } 4011 KEG_UNLOCK(keg); 4012 } 4013 4014 /* See uma.h */ 4015 void 4016 uma_reclaim(int req) 4017 { 4018 4019 CTR0(KTR_UMA, "UMA: vm asked us to release pages!"); 4020 sx_xlock(&uma_reclaim_lock); 4021 bucket_enable(); 4022 4023 switch (req) { 4024 case UMA_RECLAIM_TRIM: 4025 zone_foreach(zone_trim, NULL); 4026 break; 4027 case UMA_RECLAIM_DRAIN: 4028 case UMA_RECLAIM_DRAIN_CPU: 4029 zone_foreach(zone_drain, NULL); 4030 if (req == UMA_RECLAIM_DRAIN_CPU) { 4031 pcpu_cache_drain_safe(NULL); 4032 zone_foreach(zone_drain, NULL); 4033 } 4034 break; 4035 default: 4036 panic("unhandled reclamation request %d", req); 4037 } 4038 4039 /* 4040 * Some slabs may have been freed but this zone will be visited early 4041 * we visit again so that we can free pages that are empty once other 4042 * zones are drained. We have to do the same for buckets. 4043 */ 4044 zone_drain(slabzone, NULL); 4045 bucket_zone_drain(); 4046 sx_xunlock(&uma_reclaim_lock); 4047 } 4048 4049 static volatile int uma_reclaim_needed; 4050 4051 void 4052 uma_reclaim_wakeup(void) 4053 { 4054 4055 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0) 4056 wakeup(uma_reclaim); 4057 } 4058 4059 void 4060 uma_reclaim_worker(void *arg __unused) 4061 { 4062 4063 for (;;) { 4064 sx_xlock(&uma_reclaim_lock); 4065 while (atomic_load_int(&uma_reclaim_needed) == 0) 4066 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl", 4067 hz); 4068 sx_xunlock(&uma_reclaim_lock); 4069 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); 4070 uma_reclaim(UMA_RECLAIM_DRAIN_CPU); 4071 atomic_store_int(&uma_reclaim_needed, 0); 4072 /* Don't fire more than once per-second. */ 4073 pause("umarclslp", hz); 4074 } 4075 } 4076 4077 /* See uma.h */ 4078 void 4079 uma_zone_reclaim(uma_zone_t zone, int req) 4080 { 4081 4082 switch (req) { 4083 case UMA_RECLAIM_TRIM: 4084 zone_trim(zone, NULL); 4085 break; 4086 case UMA_RECLAIM_DRAIN: 4087 zone_drain(zone, NULL); 4088 break; 4089 case UMA_RECLAIM_DRAIN_CPU: 4090 pcpu_cache_drain_safe(zone); 4091 zone_drain(zone, NULL); 4092 break; 4093 default: 4094 panic("unhandled reclamation request %d", req); 4095 } 4096 } 4097 4098 /* See uma.h */ 4099 int 4100 uma_zone_exhausted(uma_zone_t zone) 4101 { 4102 int full; 4103 4104 ZONE_LOCK(zone); 4105 full = zone->uz_sleepers > 0; 4106 ZONE_UNLOCK(zone); 4107 return (full); 4108 } 4109 4110 int 4111 uma_zone_exhausted_nolock(uma_zone_t zone) 4112 { 4113 return (zone->uz_sleepers > 0); 4114 } 4115 4116 static void 4117 uma_zero_item(void *item, uma_zone_t zone) 4118 { 4119 4120 bzero(item, zone->uz_size); 4121 } 4122 4123 unsigned long 4124 uma_limit(void) 4125 { 4126 4127 return (uma_kmem_limit); 4128 } 4129 4130 void 4131 uma_set_limit(unsigned long limit) 4132 { 4133 4134 uma_kmem_limit = limit; 4135 } 4136 4137 unsigned long 4138 uma_size(void) 4139 { 4140 4141 return (atomic_load_long(&uma_kmem_total)); 4142 } 4143 4144 long 4145 uma_avail(void) 4146 { 4147 4148 return (uma_kmem_limit - uma_size()); 4149 } 4150 4151 #ifdef DDB 4152 /* 4153 * Generate statistics across both the zone and its per-cpu cache's. Return 4154 * desired statistics if the pointer is non-NULL for that statistic. 4155 * 4156 * Note: does not update the zone statistics, as it can't safely clear the 4157 * per-CPU cache statistic. 4158 * 4159 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 4160 * safe from off-CPU; we should modify the caches to track this information 4161 * directly so that we don't have to. 4162 */ 4163 static void 4164 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp, 4165 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp) 4166 { 4167 uma_cache_t cache; 4168 uint64_t allocs, frees, sleeps, xdomain; 4169 int cachefree, cpu; 4170 4171 allocs = frees = sleeps = xdomain = 0; 4172 cachefree = 0; 4173 CPU_FOREACH(cpu) { 4174 cache = &z->uz_cpu[cpu]; 4175 if (cache->uc_allocbucket != NULL) 4176 cachefree += cache->uc_allocbucket->ub_cnt; 4177 if (cache->uc_freebucket != NULL) 4178 cachefree += cache->uc_freebucket->ub_cnt; 4179 if (cache->uc_crossbucket != NULL) { 4180 xdomain += cache->uc_crossbucket->ub_cnt; 4181 cachefree += cache->uc_crossbucket->ub_cnt; 4182 } 4183 allocs += cache->uc_allocs; 4184 frees += cache->uc_frees; 4185 } 4186 allocs += counter_u64_fetch(z->uz_allocs); 4187 frees += counter_u64_fetch(z->uz_frees); 4188 sleeps += z->uz_sleeps; 4189 xdomain += z->uz_xdomain; 4190 if (cachefreep != NULL) 4191 *cachefreep = cachefree; 4192 if (allocsp != NULL) 4193 *allocsp = allocs; 4194 if (freesp != NULL) 4195 *freesp = frees; 4196 if (sleepsp != NULL) 4197 *sleepsp = sleeps; 4198 if (xdomainp != NULL) 4199 *xdomainp = xdomain; 4200 } 4201 #endif /* DDB */ 4202 4203 static int 4204 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 4205 { 4206 uma_keg_t kz; 4207 uma_zone_t z; 4208 int count; 4209 4210 count = 0; 4211 rw_rlock(&uma_rwlock); 4212 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4213 LIST_FOREACH(z, &kz->uk_zones, uz_link) 4214 count++; 4215 } 4216 LIST_FOREACH(z, &uma_cachezones, uz_link) 4217 count++; 4218 4219 rw_runlock(&uma_rwlock); 4220 return (sysctl_handle_int(oidp, &count, 0, req)); 4221 } 4222 4223 static void 4224 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf, 4225 struct uma_percpu_stat *ups, bool internal) 4226 { 4227 uma_zone_domain_t zdom; 4228 uma_bucket_t bucket; 4229 uma_cache_t cache; 4230 int i; 4231 4232 4233 for (i = 0; i < vm_ndomains; i++) { 4234 zdom = &z->uz_domain[i]; 4235 uth->uth_zone_free += zdom->uzd_nitems; 4236 } 4237 uth->uth_allocs = counter_u64_fetch(z->uz_allocs); 4238 uth->uth_frees = counter_u64_fetch(z->uz_frees); 4239 uth->uth_fails = counter_u64_fetch(z->uz_fails); 4240 uth->uth_sleeps = z->uz_sleeps; 4241 uth->uth_xdomain = z->uz_xdomain; 4242 4243 /* 4244 * While it is not normally safe to access the cache bucket pointers 4245 * while not on the CPU that owns the cache, we only allow the pointers 4246 * to be exchanged without the zone lock held, not invalidated, so 4247 * accept the possible race associated with bucket exchange during 4248 * monitoring. Use atomic_load_ptr() to ensure that the bucket pointers 4249 * are loaded only once. 4250 */ 4251 for (i = 0; i < mp_maxid + 1; i++) { 4252 bzero(&ups[i], sizeof(*ups)); 4253 if (internal || CPU_ABSENT(i)) 4254 continue; 4255 cache = &z->uz_cpu[i]; 4256 bucket = (uma_bucket_t)atomic_load_ptr(&cache->uc_allocbucket); 4257 if (bucket != NULL) 4258 ups[i].ups_cache_free += bucket->ub_cnt; 4259 bucket = (uma_bucket_t)atomic_load_ptr(&cache->uc_freebucket); 4260 if (bucket != NULL) 4261 ups[i].ups_cache_free += bucket->ub_cnt; 4262 bucket = (uma_bucket_t)atomic_load_ptr(&cache->uc_crossbucket); 4263 if (bucket != NULL) 4264 ups[i].ups_cache_free += bucket->ub_cnt; 4265 ups[i].ups_allocs = cache->uc_allocs; 4266 ups[i].ups_frees = cache->uc_frees; 4267 } 4268 } 4269 4270 static int 4271 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 4272 { 4273 struct uma_stream_header ush; 4274 struct uma_type_header uth; 4275 struct uma_percpu_stat *ups; 4276 struct sbuf sbuf; 4277 uma_keg_t kz; 4278 uma_zone_t z; 4279 int count, error, i; 4280 4281 error = sysctl_wire_old_buffer(req, 0); 4282 if (error != 0) 4283 return (error); 4284 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 4285 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); 4286 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK); 4287 4288 count = 0; 4289 rw_rlock(&uma_rwlock); 4290 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4291 LIST_FOREACH(z, &kz->uk_zones, uz_link) 4292 count++; 4293 } 4294 4295 LIST_FOREACH(z, &uma_cachezones, uz_link) 4296 count++; 4297 4298 /* 4299 * Insert stream header. 4300 */ 4301 bzero(&ush, sizeof(ush)); 4302 ush.ush_version = UMA_STREAM_VERSION; 4303 ush.ush_maxcpus = (mp_maxid + 1); 4304 ush.ush_count = count; 4305 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 4306 4307 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4308 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 4309 bzero(&uth, sizeof(uth)); 4310 ZONE_LOCK(z); 4311 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 4312 uth.uth_align = kz->uk_align; 4313 uth.uth_size = kz->uk_size; 4314 uth.uth_rsize = kz->uk_rsize; 4315 if (z->uz_max_items > 0) 4316 uth.uth_pages = (z->uz_items / kz->uk_ipers) * 4317 kz->uk_ppera; 4318 else 4319 uth.uth_pages = kz->uk_pages; 4320 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) * 4321 kz->uk_ppera; 4322 uth.uth_limit = z->uz_max_items; 4323 uth.uth_keg_free = z->uz_keg->uk_free; 4324 4325 /* 4326 * A zone is secondary is it is not the first entry 4327 * on the keg's zone list. 4328 */ 4329 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 4330 (LIST_FIRST(&kz->uk_zones) != z)) 4331 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 4332 uma_vm_zone_stats(&uth, z, &sbuf, ups, 4333 kz->uk_flags & UMA_ZFLAG_INTERNAL); 4334 ZONE_UNLOCK(z); 4335 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 4336 for (i = 0; i < mp_maxid + 1; i++) 4337 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 4338 } 4339 } 4340 LIST_FOREACH(z, &uma_cachezones, uz_link) { 4341 bzero(&uth, sizeof(uth)); 4342 ZONE_LOCK(z); 4343 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 4344 uth.uth_size = z->uz_size; 4345 uma_vm_zone_stats(&uth, z, &sbuf, ups, false); 4346 ZONE_UNLOCK(z); 4347 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 4348 for (i = 0; i < mp_maxid + 1; i++) 4349 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); 4350 } 4351 4352 rw_runlock(&uma_rwlock); 4353 error = sbuf_finish(&sbuf); 4354 sbuf_delete(&sbuf); 4355 free(ups, M_TEMP); 4356 return (error); 4357 } 4358 4359 int 4360 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 4361 { 4362 uma_zone_t zone = *(uma_zone_t *)arg1; 4363 int error, max; 4364 4365 max = uma_zone_get_max(zone); 4366 error = sysctl_handle_int(oidp, &max, 0, req); 4367 if (error || !req->newptr) 4368 return (error); 4369 4370 uma_zone_set_max(zone, max); 4371 4372 return (0); 4373 } 4374 4375 int 4376 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 4377 { 4378 uma_zone_t zone; 4379 int cur; 4380 4381 /* 4382 * Some callers want to add sysctls for global zones that 4383 * may not yet exist so they pass a pointer to a pointer. 4384 */ 4385 if (arg2 == 0) 4386 zone = *(uma_zone_t *)arg1; 4387 else 4388 zone = arg1; 4389 cur = uma_zone_get_cur(zone); 4390 return (sysctl_handle_int(oidp, &cur, 0, req)); 4391 } 4392 4393 static int 4394 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS) 4395 { 4396 uma_zone_t zone = arg1; 4397 uint64_t cur; 4398 4399 cur = uma_zone_get_allocs(zone); 4400 return (sysctl_handle_64(oidp, &cur, 0, req)); 4401 } 4402 4403 static int 4404 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS) 4405 { 4406 uma_zone_t zone = arg1; 4407 uint64_t cur; 4408 4409 cur = uma_zone_get_frees(zone); 4410 return (sysctl_handle_64(oidp, &cur, 0, req)); 4411 } 4412 4413 static int 4414 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS) 4415 { 4416 struct sbuf sbuf; 4417 uma_zone_t zone = arg1; 4418 int error; 4419 4420 sbuf_new_for_sysctl(&sbuf, NULL, 0, req); 4421 if (zone->uz_flags != 0) 4422 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS); 4423 else 4424 sbuf_printf(&sbuf, "0"); 4425 error = sbuf_finish(&sbuf); 4426 sbuf_delete(&sbuf); 4427 4428 return (error); 4429 } 4430 4431 #ifdef INVARIANTS 4432 static uma_slab_t 4433 uma_dbg_getslab(uma_zone_t zone, void *item) 4434 { 4435 uma_slab_t slab; 4436 uma_keg_t keg; 4437 uint8_t *mem; 4438 4439 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 4440 if (zone->uz_flags & UMA_ZONE_VTOSLAB) { 4441 slab = vtoslab((vm_offset_t)mem); 4442 } else { 4443 /* 4444 * It is safe to return the slab here even though the 4445 * zone is unlocked because the item's allocation state 4446 * essentially holds a reference. 4447 */ 4448 if (zone->uz_lockptr == &zone->uz_lock) 4449 return (NULL); 4450 ZONE_LOCK(zone); 4451 keg = zone->uz_keg; 4452 if (keg->uk_flags & UMA_ZONE_HASH) 4453 slab = hash_sfind(&keg->uk_hash, mem); 4454 else 4455 slab = (uma_slab_t)(mem + keg->uk_pgoff); 4456 ZONE_UNLOCK(zone); 4457 } 4458 4459 return (slab); 4460 } 4461 4462 static bool 4463 uma_dbg_zskip(uma_zone_t zone, void *mem) 4464 { 4465 4466 if (zone->uz_lockptr == &zone->uz_lock) 4467 return (true); 4468 4469 return (uma_dbg_kskip(zone->uz_keg, mem)); 4470 } 4471 4472 static bool 4473 uma_dbg_kskip(uma_keg_t keg, void *mem) 4474 { 4475 uintptr_t idx; 4476 4477 if (dbg_divisor == 0) 4478 return (true); 4479 4480 if (dbg_divisor == 1) 4481 return (false); 4482 4483 idx = (uintptr_t)mem >> PAGE_SHIFT; 4484 if (keg->uk_ipers > 1) { 4485 idx *= keg->uk_ipers; 4486 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize; 4487 } 4488 4489 if ((idx / dbg_divisor) * dbg_divisor != idx) { 4490 counter_u64_add(uma_skip_cnt, 1); 4491 return (true); 4492 } 4493 counter_u64_add(uma_dbg_cnt, 1); 4494 4495 return (false); 4496 } 4497 4498 /* 4499 * Set up the slab's freei data such that uma_dbg_free can function. 4500 * 4501 */ 4502 static void 4503 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) 4504 { 4505 uma_keg_t keg; 4506 int freei; 4507 4508 if (slab == NULL) { 4509 slab = uma_dbg_getslab(zone, item); 4510 if (slab == NULL) 4511 panic("uma: item %p did not belong to zone %s\n", 4512 item, zone->uz_name); 4513 } 4514 keg = zone->uz_keg; 4515 freei = slab_item_index(slab, keg, item); 4516 4517 if (BIT_ISSET(SLAB_MAX_SETSIZE, freei, &slab->us_debugfree)) 4518 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n", 4519 item, zone, zone->uz_name, slab, freei); 4520 BIT_SET_ATOMIC(SLAB_MAX_SETSIZE, freei, &slab->us_debugfree); 4521 4522 return; 4523 } 4524 4525 /* 4526 * Verifies freed addresses. Checks for alignment, valid slab membership 4527 * and duplicate frees. 4528 * 4529 */ 4530 static void 4531 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) 4532 { 4533 uma_keg_t keg; 4534 int freei; 4535 4536 if (slab == NULL) { 4537 slab = uma_dbg_getslab(zone, item); 4538 if (slab == NULL) 4539 panic("uma: Freed item %p did not belong to zone %s\n", 4540 item, zone->uz_name); 4541 } 4542 keg = zone->uz_keg; 4543 freei = slab_item_index(slab, keg, item); 4544 4545 if (freei >= keg->uk_ipers) 4546 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n", 4547 item, zone, zone->uz_name, slab, freei); 4548 4549 if (slab_item(slab, keg, freei) != item) 4550 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n", 4551 item, zone, zone->uz_name, slab, freei); 4552 4553 if (!BIT_ISSET(SLAB_MAX_SETSIZE, freei, &slab->us_debugfree)) 4554 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n", 4555 item, zone, zone->uz_name, slab, freei); 4556 4557 BIT_CLR_ATOMIC(SLAB_MAX_SETSIZE, freei, &slab->us_debugfree); 4558 } 4559 #endif /* INVARIANTS */ 4560 4561 #ifdef DDB 4562 static int64_t 4563 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used, 4564 uint64_t *sleeps, long *cachefree, uint64_t *xdomain) 4565 { 4566 uint64_t frees; 4567 int i; 4568 4569 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 4570 *allocs = counter_u64_fetch(z->uz_allocs); 4571 frees = counter_u64_fetch(z->uz_frees); 4572 *sleeps = z->uz_sleeps; 4573 *cachefree = 0; 4574 *xdomain = 0; 4575 } else 4576 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps, 4577 xdomain); 4578 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 4579 (LIST_FIRST(&kz->uk_zones) != z))) 4580 *cachefree += kz->uk_free; 4581 for (i = 0; i < vm_ndomains; i++) 4582 *cachefree += z->uz_domain[i].uzd_nitems; 4583 *used = *allocs - frees; 4584 return (((int64_t)*used + *cachefree) * kz->uk_size); 4585 } 4586 4587 DB_SHOW_COMMAND(uma, db_show_uma) 4588 { 4589 const char *fmt_hdr, *fmt_entry; 4590 uma_keg_t kz; 4591 uma_zone_t z; 4592 uint64_t allocs, used, sleeps, xdomain; 4593 long cachefree; 4594 /* variables for sorting */ 4595 uma_keg_t cur_keg; 4596 uma_zone_t cur_zone, last_zone; 4597 int64_t cur_size, last_size, size; 4598 int ties; 4599 4600 /* /i option produces machine-parseable CSV output */ 4601 if (modif[0] == 'i') { 4602 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n"; 4603 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n"; 4604 } else { 4605 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n"; 4606 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n"; 4607 } 4608 4609 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests", 4610 "Sleeps", "Bucket", "Total Mem", "XFree"); 4611 4612 /* Sort the zones with largest size first. */ 4613 last_zone = NULL; 4614 last_size = INT64_MAX; 4615 for (;;) { 4616 cur_zone = NULL; 4617 cur_size = -1; 4618 ties = 0; 4619 LIST_FOREACH(kz, &uma_kegs, uk_link) { 4620 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 4621 /* 4622 * In the case of size ties, print out zones 4623 * in the order they are encountered. That is, 4624 * when we encounter the most recently output 4625 * zone, we have already printed all preceding 4626 * ties, and we must print all following ties. 4627 */ 4628 if (z == last_zone) { 4629 ties = 1; 4630 continue; 4631 } 4632 size = get_uma_stats(kz, z, &allocs, &used, 4633 &sleeps, &cachefree, &xdomain); 4634 if (size > cur_size && size < last_size + ties) 4635 { 4636 cur_size = size; 4637 cur_zone = z; 4638 cur_keg = kz; 4639 } 4640 } 4641 } 4642 if (cur_zone == NULL) 4643 break; 4644 4645 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used, 4646 &sleeps, &cachefree, &xdomain); 4647 db_printf(fmt_entry, cur_zone->uz_name, 4648 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree, 4649 (uintmax_t)allocs, (uintmax_t)sleeps, 4650 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size, 4651 xdomain); 4652 4653 if (db_pager_quit) 4654 return; 4655 last_zone = cur_zone; 4656 last_size = cur_size; 4657 } 4658 } 4659 4660 DB_SHOW_COMMAND(umacache, db_show_umacache) 4661 { 4662 uma_zone_t z; 4663 uint64_t allocs, frees; 4664 long cachefree; 4665 int i; 4666 4667 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 4668 "Requests", "Bucket"); 4669 LIST_FOREACH(z, &uma_cachezones, uz_link) { 4670 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL); 4671 for (i = 0; i < vm_ndomains; i++) 4672 cachefree += z->uz_domain[i].uzd_nitems; 4673 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n", 4674 z->uz_name, (uintmax_t)z->uz_size, 4675 (intmax_t)(allocs - frees), cachefree, 4676 (uintmax_t)allocs, z->uz_bucket_size); 4677 if (db_pager_quit) 4678 return; 4679 } 4680 } 4681 #endif /* DDB */ 4682