1 /*- 2 * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org> 3 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org> 4 * Copyright (c) 2004-2006 Robert N. M. Watson 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice unmodified, this list of conditions, and the following 12 * disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 */ 28 29 /* 30 * uma_core.c Implementation of the Universal Memory allocator 31 * 32 * This allocator is intended to replace the multitude of similar object caches 33 * in the standard FreeBSD kernel. The intent is to be flexible as well as 34 * efficient. A primary design goal is to return unused memory to the rest of 35 * the system. This will make the system as a whole more flexible due to the 36 * ability to move memory to subsystems which most need it instead of leaving 37 * pools of reserved memory unused. 38 * 39 * The basic ideas stem from similar slab/zone based allocators whose algorithms 40 * are well known. 41 * 42 */ 43 44 /* 45 * TODO: 46 * - Improve memory usage for large allocations 47 * - Investigate cache size adjustments 48 */ 49 50 #include <sys/cdefs.h> 51 __FBSDID("$FreeBSD$"); 52 53 #include "opt_ddb.h" 54 #include "opt_param.h" 55 #include "opt_vm.h" 56 57 #include <sys/param.h> 58 #include <sys/systm.h> 59 #include <sys/bitset.h> 60 #include <sys/eventhandler.h> 61 #include <sys/kernel.h> 62 #include <sys/types.h> 63 #include <sys/queue.h> 64 #include <sys/malloc.h> 65 #include <sys/ktr.h> 66 #include <sys/lock.h> 67 #include <sys/sysctl.h> 68 #include <sys/mutex.h> 69 #include <sys/proc.h> 70 #include <sys/random.h> 71 #include <sys/rwlock.h> 72 #include <sys/sbuf.h> 73 #include <sys/sched.h> 74 #include <sys/smp.h> 75 #include <sys/taskqueue.h> 76 #include <sys/vmmeter.h> 77 78 #include <vm/vm.h> 79 #include <vm/vm_object.h> 80 #include <vm/vm_page.h> 81 #include <vm/vm_pageout.h> 82 #include <vm/vm_param.h> 83 #include <vm/vm_map.h> 84 #include <vm/vm_kern.h> 85 #include <vm/vm_extern.h> 86 #include <vm/uma.h> 87 #include <vm/uma_int.h> 88 #include <vm/uma_dbg.h> 89 90 #include <ddb/ddb.h> 91 92 #ifdef DEBUG_MEMGUARD 93 #include <vm/memguard.h> 94 #endif 95 96 /* 97 * This is the zone and keg from which all zones are spawned. The idea is that 98 * even the zone & keg heads are allocated from the allocator, so we use the 99 * bss section to bootstrap us. 100 */ 101 static struct uma_keg masterkeg; 102 static struct uma_zone masterzone_k; 103 static struct uma_zone masterzone_z; 104 static uma_zone_t kegs = &masterzone_k; 105 static uma_zone_t zones = &masterzone_z; 106 107 /* This is the zone from which all of uma_slab_t's are allocated. */ 108 static uma_zone_t slabzone; 109 110 /* 111 * The initial hash tables come out of this zone so they can be allocated 112 * prior to malloc coming up. 113 */ 114 static uma_zone_t hashzone; 115 116 /* The boot-time adjusted value for cache line alignment. */ 117 int uma_align_cache = 64 - 1; 118 119 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 120 121 /* 122 * Are we allowed to allocate buckets? 123 */ 124 static int bucketdisable = 1; 125 126 /* Linked list of all kegs in the system */ 127 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 128 129 /* Linked list of all cache-only zones in the system */ 130 static LIST_HEAD(,uma_zone) uma_cachezones = 131 LIST_HEAD_INITIALIZER(uma_cachezones); 132 133 /* This RW lock protects the keg list */ 134 static struct rwlock_padalign __exclusive_cache_line uma_rwlock; 135 136 /* 137 * Pointer and counter to pool of pages, that is preallocated at 138 * startup to bootstrap UMA. Early zones continue to use the pool 139 * until it is depleted, so allocations may happen after boot, thus 140 * we need a mutex to protect it. 141 */ 142 static char *bootmem; 143 static int boot_pages; 144 static struct mtx uma_boot_pages_mtx; 145 146 static struct sx uma_drain_lock; 147 148 /* Is the VM done starting up? */ 149 static int booted = 0; 150 #define UMA_STARTUP 1 151 #define UMA_STARTUP2 2 152 153 /* 154 * This is the handle used to schedule events that need to happen 155 * outside of the allocation fast path. 156 */ 157 static struct callout uma_callout; 158 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 159 160 /* 161 * This structure is passed as the zone ctor arg so that I don't have to create 162 * a special allocation function just for zones. 163 */ 164 struct uma_zctor_args { 165 const char *name; 166 size_t size; 167 uma_ctor ctor; 168 uma_dtor dtor; 169 uma_init uminit; 170 uma_fini fini; 171 uma_import import; 172 uma_release release; 173 void *arg; 174 uma_keg_t keg; 175 int align; 176 uint32_t flags; 177 }; 178 179 struct uma_kctor_args { 180 uma_zone_t zone; 181 size_t size; 182 uma_init uminit; 183 uma_fini fini; 184 int align; 185 uint32_t flags; 186 }; 187 188 struct uma_bucket_zone { 189 uma_zone_t ubz_zone; 190 char *ubz_name; 191 int ubz_entries; /* Number of items it can hold. */ 192 int ubz_maxsize; /* Maximum allocation size per-item. */ 193 }; 194 195 /* 196 * Compute the actual number of bucket entries to pack them in power 197 * of two sizes for more efficient space utilization. 198 */ 199 #define BUCKET_SIZE(n) \ 200 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 201 202 #define BUCKET_MAX BUCKET_SIZE(256) 203 204 struct uma_bucket_zone bucket_zones[] = { 205 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 206 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 207 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 208 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 209 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 210 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 211 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 212 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 213 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 214 { NULL, NULL, 0} 215 }; 216 217 /* 218 * Flags and enumerations to be passed to internal functions. 219 */ 220 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; 221 222 /* Prototypes.. */ 223 224 static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 225 static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 226 static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 227 static void page_free(void *, vm_size_t, uint8_t); 228 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); 229 static void cache_drain(uma_zone_t); 230 static void bucket_drain(uma_zone_t, uma_bucket_t); 231 static void bucket_cache_drain(uma_zone_t zone); 232 static int keg_ctor(void *, int, void *, int); 233 static void keg_dtor(void *, int, void *); 234 static int zone_ctor(void *, int, void *, int); 235 static void zone_dtor(void *, int, void *); 236 static int zero_init(void *, int, int); 237 static void keg_small_init(uma_keg_t keg); 238 static void keg_large_init(uma_keg_t keg); 239 static void zone_foreach(void (*zfunc)(uma_zone_t)); 240 static void zone_timeout(uma_zone_t zone); 241 static int hash_alloc(struct uma_hash *); 242 static int hash_expand(struct uma_hash *, struct uma_hash *); 243 static void hash_free(struct uma_hash *hash); 244 static void uma_timeout(void *); 245 static void uma_startup3(void); 246 static void *zone_alloc_item(uma_zone_t, void *, int); 247 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 248 static void bucket_enable(void); 249 static void bucket_init(void); 250 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 251 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 252 static void bucket_zone_drain(void); 253 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags); 254 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); 255 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); 256 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 257 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); 258 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 259 uma_fini fini, int align, uint32_t flags); 260 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags); 261 static void zone_release(uma_zone_t zone, void **bucket, int cnt); 262 static void uma_zero_item(void *item, uma_zone_t zone); 263 264 void uma_print_zone(uma_zone_t); 265 void uma_print_stats(void); 266 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 267 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 268 269 #ifdef INVARIANTS 270 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item); 271 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item); 272 #endif 273 274 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 275 276 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 277 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 278 279 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 280 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 281 282 static int zone_warnings = 1; 283 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, 284 "Warn when UMA zones becomes full"); 285 286 /* 287 * This routine checks to see whether or not it's safe to enable buckets. 288 */ 289 static void 290 bucket_enable(void) 291 { 292 bucketdisable = vm_page_count_min(); 293 } 294 295 /* 296 * Initialize bucket_zones, the array of zones of buckets of various sizes. 297 * 298 * For each zone, calculate the memory required for each bucket, consisting 299 * of the header and an array of pointers. 300 */ 301 static void 302 bucket_init(void) 303 { 304 struct uma_bucket_zone *ubz; 305 int size; 306 307 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 308 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 309 size += sizeof(void *) * ubz->ubz_entries; 310 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 311 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 312 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET); 313 } 314 } 315 316 /* 317 * Given a desired number of entries for a bucket, return the zone from which 318 * to allocate the bucket. 319 */ 320 static struct uma_bucket_zone * 321 bucket_zone_lookup(int entries) 322 { 323 struct uma_bucket_zone *ubz; 324 325 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 326 if (ubz->ubz_entries >= entries) 327 return (ubz); 328 ubz--; 329 return (ubz); 330 } 331 332 static int 333 bucket_select(int size) 334 { 335 struct uma_bucket_zone *ubz; 336 337 ubz = &bucket_zones[0]; 338 if (size > ubz->ubz_maxsize) 339 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 340 341 for (; ubz->ubz_entries != 0; ubz++) 342 if (ubz->ubz_maxsize < size) 343 break; 344 ubz--; 345 return (ubz->ubz_entries); 346 } 347 348 static uma_bucket_t 349 bucket_alloc(uma_zone_t zone, void *udata, int flags) 350 { 351 struct uma_bucket_zone *ubz; 352 uma_bucket_t bucket; 353 354 /* 355 * This is to stop us from allocating per cpu buckets while we're 356 * running out of vm.boot_pages. Otherwise, we would exhaust the 357 * boot pages. This also prevents us from allocating buckets in 358 * low memory situations. 359 */ 360 if (bucketdisable) 361 return (NULL); 362 /* 363 * To limit bucket recursion we store the original zone flags 364 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 365 * NOVM flag to persist even through deep recursions. We also 366 * store ZFLAG_BUCKET once we have recursed attempting to allocate 367 * a bucket for a bucket zone so we do not allow infinite bucket 368 * recursion. This cookie will even persist to frees of unused 369 * buckets via the allocation path or bucket allocations in the 370 * free path. 371 */ 372 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 373 udata = (void *)(uintptr_t)zone->uz_flags; 374 else { 375 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 376 return (NULL); 377 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 378 } 379 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 380 flags |= M_NOVM; 381 ubz = bucket_zone_lookup(zone->uz_count); 382 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 383 ubz++; 384 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 385 if (bucket) { 386 #ifdef INVARIANTS 387 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 388 #endif 389 bucket->ub_cnt = 0; 390 bucket->ub_entries = ubz->ubz_entries; 391 } 392 393 return (bucket); 394 } 395 396 static void 397 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 398 { 399 struct uma_bucket_zone *ubz; 400 401 KASSERT(bucket->ub_cnt == 0, 402 ("bucket_free: Freeing a non free bucket.")); 403 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 404 udata = (void *)(uintptr_t)zone->uz_flags; 405 ubz = bucket_zone_lookup(bucket->ub_entries); 406 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 407 } 408 409 static void 410 bucket_zone_drain(void) 411 { 412 struct uma_bucket_zone *ubz; 413 414 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 415 zone_drain(ubz->ubz_zone); 416 } 417 418 static void 419 zone_log_warning(uma_zone_t zone) 420 { 421 static const struct timeval warninterval = { 300, 0 }; 422 423 if (!zone_warnings || zone->uz_warning == NULL) 424 return; 425 426 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 427 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 428 } 429 430 static inline void 431 zone_maxaction(uma_zone_t zone) 432 { 433 434 if (zone->uz_maxaction.ta_func != NULL) 435 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); 436 } 437 438 static void 439 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 440 { 441 uma_klink_t klink; 442 443 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 444 kegfn(klink->kl_keg); 445 } 446 447 /* 448 * Routine called by timeout which is used to fire off some time interval 449 * based calculations. (stats, hash size, etc.) 450 * 451 * Arguments: 452 * arg Unused 453 * 454 * Returns: 455 * Nothing 456 */ 457 static void 458 uma_timeout(void *unused) 459 { 460 bucket_enable(); 461 zone_foreach(zone_timeout); 462 463 /* Reschedule this event */ 464 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 465 } 466 467 /* 468 * Routine to perform timeout driven calculations. This expands the 469 * hashes and does per cpu statistics aggregation. 470 * 471 * Returns nothing. 472 */ 473 static void 474 keg_timeout(uma_keg_t keg) 475 { 476 477 KEG_LOCK(keg); 478 /* 479 * Expand the keg hash table. 480 * 481 * This is done if the number of slabs is larger than the hash size. 482 * What I'm trying to do here is completely reduce collisions. This 483 * may be a little aggressive. Should I allow for two collisions max? 484 */ 485 if (keg->uk_flags & UMA_ZONE_HASH && 486 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 487 struct uma_hash newhash; 488 struct uma_hash oldhash; 489 int ret; 490 491 /* 492 * This is so involved because allocating and freeing 493 * while the keg lock is held will lead to deadlock. 494 * I have to do everything in stages and check for 495 * races. 496 */ 497 newhash = keg->uk_hash; 498 KEG_UNLOCK(keg); 499 ret = hash_alloc(&newhash); 500 KEG_LOCK(keg); 501 if (ret) { 502 if (hash_expand(&keg->uk_hash, &newhash)) { 503 oldhash = keg->uk_hash; 504 keg->uk_hash = newhash; 505 } else 506 oldhash = newhash; 507 508 KEG_UNLOCK(keg); 509 hash_free(&oldhash); 510 return; 511 } 512 } 513 KEG_UNLOCK(keg); 514 } 515 516 static void 517 zone_timeout(uma_zone_t zone) 518 { 519 520 zone_foreach_keg(zone, &keg_timeout); 521 } 522 523 /* 524 * Allocate and zero fill the next sized hash table from the appropriate 525 * backing store. 526 * 527 * Arguments: 528 * hash A new hash structure with the old hash size in uh_hashsize 529 * 530 * Returns: 531 * 1 on success and 0 on failure. 532 */ 533 static int 534 hash_alloc(struct uma_hash *hash) 535 { 536 int oldsize; 537 int alloc; 538 539 oldsize = hash->uh_hashsize; 540 541 /* We're just going to go to a power of two greater */ 542 if (oldsize) { 543 hash->uh_hashsize = oldsize * 2; 544 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 545 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 546 M_UMAHASH, M_NOWAIT); 547 } else { 548 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 549 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 550 M_WAITOK); 551 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 552 } 553 if (hash->uh_slab_hash) { 554 bzero(hash->uh_slab_hash, alloc); 555 hash->uh_hashmask = hash->uh_hashsize - 1; 556 return (1); 557 } 558 559 return (0); 560 } 561 562 /* 563 * Expands the hash table for HASH zones. This is done from zone_timeout 564 * to reduce collisions. This must not be done in the regular allocation 565 * path, otherwise, we can recurse on the vm while allocating pages. 566 * 567 * Arguments: 568 * oldhash The hash you want to expand 569 * newhash The hash structure for the new table 570 * 571 * Returns: 572 * Nothing 573 * 574 * Discussion: 575 */ 576 static int 577 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 578 { 579 uma_slab_t slab; 580 int hval; 581 int i; 582 583 if (!newhash->uh_slab_hash) 584 return (0); 585 586 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 587 return (0); 588 589 /* 590 * I need to investigate hash algorithms for resizing without a 591 * full rehash. 592 */ 593 594 for (i = 0; i < oldhash->uh_hashsize; i++) 595 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 596 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 597 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 598 hval = UMA_HASH(newhash, slab->us_data); 599 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 600 slab, us_hlink); 601 } 602 603 return (1); 604 } 605 606 /* 607 * Free the hash bucket to the appropriate backing store. 608 * 609 * Arguments: 610 * slab_hash The hash bucket we're freeing 611 * hashsize The number of entries in that hash bucket 612 * 613 * Returns: 614 * Nothing 615 */ 616 static void 617 hash_free(struct uma_hash *hash) 618 { 619 if (hash->uh_slab_hash == NULL) 620 return; 621 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 622 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 623 else 624 free(hash->uh_slab_hash, M_UMAHASH); 625 } 626 627 /* 628 * Frees all outstanding items in a bucket 629 * 630 * Arguments: 631 * zone The zone to free to, must be unlocked. 632 * bucket The free/alloc bucket with items, cpu queue must be locked. 633 * 634 * Returns: 635 * Nothing 636 */ 637 638 static void 639 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 640 { 641 int i; 642 643 if (bucket == NULL) 644 return; 645 646 if (zone->uz_fini) 647 for (i = 0; i < bucket->ub_cnt; i++) 648 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 649 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 650 bucket->ub_cnt = 0; 651 } 652 653 /* 654 * Drains the per cpu caches for a zone. 655 * 656 * NOTE: This may only be called while the zone is being turn down, and not 657 * during normal operation. This is necessary in order that we do not have 658 * to migrate CPUs to drain the per-CPU caches. 659 * 660 * Arguments: 661 * zone The zone to drain, must be unlocked. 662 * 663 * Returns: 664 * Nothing 665 */ 666 static void 667 cache_drain(uma_zone_t zone) 668 { 669 uma_cache_t cache; 670 int cpu; 671 672 /* 673 * XXX: It is safe to not lock the per-CPU caches, because we're 674 * tearing down the zone anyway. I.e., there will be no further use 675 * of the caches at this point. 676 * 677 * XXX: It would good to be able to assert that the zone is being 678 * torn down to prevent improper use of cache_drain(). 679 * 680 * XXX: We lock the zone before passing into bucket_cache_drain() as 681 * it is used elsewhere. Should the tear-down path be made special 682 * there in some form? 683 */ 684 CPU_FOREACH(cpu) { 685 cache = &zone->uz_cpu[cpu]; 686 bucket_drain(zone, cache->uc_allocbucket); 687 bucket_drain(zone, cache->uc_freebucket); 688 if (cache->uc_allocbucket != NULL) 689 bucket_free(zone, cache->uc_allocbucket, NULL); 690 if (cache->uc_freebucket != NULL) 691 bucket_free(zone, cache->uc_freebucket, NULL); 692 cache->uc_allocbucket = cache->uc_freebucket = NULL; 693 } 694 ZONE_LOCK(zone); 695 bucket_cache_drain(zone); 696 ZONE_UNLOCK(zone); 697 } 698 699 static void 700 cache_shrink(uma_zone_t zone) 701 { 702 703 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 704 return; 705 706 ZONE_LOCK(zone); 707 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; 708 ZONE_UNLOCK(zone); 709 } 710 711 static void 712 cache_drain_safe_cpu(uma_zone_t zone) 713 { 714 uma_cache_t cache; 715 uma_bucket_t b1, b2; 716 717 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 718 return; 719 720 b1 = b2 = NULL; 721 ZONE_LOCK(zone); 722 critical_enter(); 723 cache = &zone->uz_cpu[curcpu]; 724 if (cache->uc_allocbucket) { 725 if (cache->uc_allocbucket->ub_cnt != 0) 726 LIST_INSERT_HEAD(&zone->uz_buckets, 727 cache->uc_allocbucket, ub_link); 728 else 729 b1 = cache->uc_allocbucket; 730 cache->uc_allocbucket = NULL; 731 } 732 if (cache->uc_freebucket) { 733 if (cache->uc_freebucket->ub_cnt != 0) 734 LIST_INSERT_HEAD(&zone->uz_buckets, 735 cache->uc_freebucket, ub_link); 736 else 737 b2 = cache->uc_freebucket; 738 cache->uc_freebucket = NULL; 739 } 740 critical_exit(); 741 ZONE_UNLOCK(zone); 742 if (b1) 743 bucket_free(zone, b1, NULL); 744 if (b2) 745 bucket_free(zone, b2, NULL); 746 } 747 748 /* 749 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 750 * This is an expensive call because it needs to bind to all CPUs 751 * one by one and enter a critical section on each of them in order 752 * to safely access their cache buckets. 753 * Zone lock must not be held on call this function. 754 */ 755 static void 756 cache_drain_safe(uma_zone_t zone) 757 { 758 int cpu; 759 760 /* 761 * Polite bucket sizes shrinking was not enouth, shrink aggressively. 762 */ 763 if (zone) 764 cache_shrink(zone); 765 else 766 zone_foreach(cache_shrink); 767 768 CPU_FOREACH(cpu) { 769 thread_lock(curthread); 770 sched_bind(curthread, cpu); 771 thread_unlock(curthread); 772 773 if (zone) 774 cache_drain_safe_cpu(zone); 775 else 776 zone_foreach(cache_drain_safe_cpu); 777 } 778 thread_lock(curthread); 779 sched_unbind(curthread); 780 thread_unlock(curthread); 781 } 782 783 /* 784 * Drain the cached buckets from a zone. Expects a locked zone on entry. 785 */ 786 static void 787 bucket_cache_drain(uma_zone_t zone) 788 { 789 uma_bucket_t bucket; 790 791 /* 792 * Drain the bucket queues and free the buckets, we just keep two per 793 * cpu (alloc/free). 794 */ 795 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 796 LIST_REMOVE(bucket, ub_link); 797 ZONE_UNLOCK(zone); 798 bucket_drain(zone, bucket); 799 bucket_free(zone, bucket, NULL); 800 ZONE_LOCK(zone); 801 } 802 803 /* 804 * Shrink further bucket sizes. Price of single zone lock collision 805 * is probably lower then price of global cache drain. 806 */ 807 if (zone->uz_count > zone->uz_count_min) 808 zone->uz_count--; 809 } 810 811 static void 812 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 813 { 814 uint8_t *mem; 815 int i; 816 uint8_t flags; 817 818 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes", 819 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera); 820 821 mem = slab->us_data; 822 flags = slab->us_flags; 823 i = start; 824 if (keg->uk_fini != NULL) { 825 for (i--; i > -1; i--) 826 keg->uk_fini(slab->us_data + (keg->uk_rsize * i), 827 keg->uk_size); 828 } 829 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 830 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 831 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 832 } 833 834 /* 835 * Frees pages from a keg back to the system. This is done on demand from 836 * the pageout daemon. 837 * 838 * Returns nothing. 839 */ 840 static void 841 keg_drain(uma_keg_t keg) 842 { 843 struct slabhead freeslabs = { 0 }; 844 uma_slab_t slab, tmp; 845 846 /* 847 * We don't want to take pages from statically allocated kegs at this 848 * time 849 */ 850 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 851 return; 852 853 CTR3(KTR_UMA, "keg_drain %s(%p) free items: %u", 854 keg->uk_name, keg, keg->uk_free); 855 KEG_LOCK(keg); 856 if (keg->uk_free == 0) 857 goto finished; 858 859 LIST_FOREACH_SAFE(slab, &keg->uk_free_slab, us_link, tmp) { 860 /* We have nowhere to free these to. */ 861 if (slab->us_flags & UMA_SLAB_BOOT) 862 continue; 863 864 LIST_REMOVE(slab, us_link); 865 keg->uk_pages -= keg->uk_ppera; 866 keg->uk_free -= keg->uk_ipers; 867 868 if (keg->uk_flags & UMA_ZONE_HASH) 869 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); 870 871 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 872 } 873 finished: 874 KEG_UNLOCK(keg); 875 876 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 877 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 878 keg_free_slab(keg, slab, keg->uk_ipers); 879 } 880 } 881 882 static void 883 zone_drain_wait(uma_zone_t zone, int waitok) 884 { 885 886 /* 887 * Set draining to interlock with zone_dtor() so we can release our 888 * locks as we go. Only dtor() should do a WAITOK call since it 889 * is the only call that knows the structure will still be available 890 * when it wakes up. 891 */ 892 ZONE_LOCK(zone); 893 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 894 if (waitok == M_NOWAIT) 895 goto out; 896 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 897 } 898 zone->uz_flags |= UMA_ZFLAG_DRAINING; 899 bucket_cache_drain(zone); 900 ZONE_UNLOCK(zone); 901 /* 902 * The DRAINING flag protects us from being freed while 903 * we're running. Normally the uma_rwlock would protect us but we 904 * must be able to release and acquire the right lock for each keg. 905 */ 906 zone_foreach_keg(zone, &keg_drain); 907 ZONE_LOCK(zone); 908 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 909 wakeup(zone); 910 out: 911 ZONE_UNLOCK(zone); 912 } 913 914 void 915 zone_drain(uma_zone_t zone) 916 { 917 918 zone_drain_wait(zone, M_NOWAIT); 919 } 920 921 /* 922 * Allocate a new slab for a keg. This does not insert the slab onto a list. 923 * 924 * Arguments: 925 * wait Shall we wait? 926 * 927 * Returns: 928 * The slab that was allocated or NULL if there is no memory and the 929 * caller specified M_NOWAIT. 930 */ 931 static uma_slab_t 932 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) 933 { 934 uma_alloc allocf; 935 uma_slab_t slab; 936 uint8_t *mem; 937 uint8_t flags; 938 int i; 939 940 mtx_assert(&keg->uk_lock, MA_OWNED); 941 slab = NULL; 942 mem = NULL; 943 944 allocf = keg->uk_allocf; 945 KEG_UNLOCK(keg); 946 947 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 948 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); 949 if (slab == NULL) 950 goto out; 951 } 952 953 /* 954 * This reproduces the old vm_zone behavior of zero filling pages the 955 * first time they are added to a zone. 956 * 957 * Malloced items are zeroed in uma_zalloc. 958 */ 959 960 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 961 wait |= M_ZERO; 962 else 963 wait &= ~M_ZERO; 964 965 if (keg->uk_flags & UMA_ZONE_NODUMP) 966 wait |= M_NODUMP; 967 968 /* zone is passed for legacy reasons. */ 969 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait); 970 if (mem == NULL) { 971 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 972 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 973 slab = NULL; 974 goto out; 975 } 976 977 /* Point the slab into the allocated memory */ 978 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 979 slab = (uma_slab_t )(mem + keg->uk_pgoff); 980 981 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 982 for (i = 0; i < keg->uk_ppera; i++) 983 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 984 985 slab->us_keg = keg; 986 slab->us_data = mem; 987 slab->us_freecount = keg->uk_ipers; 988 slab->us_flags = flags; 989 BIT_FILL(SLAB_SETSIZE, &slab->us_free); 990 #ifdef INVARIANTS 991 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); 992 #endif 993 994 if (keg->uk_init != NULL) { 995 for (i = 0; i < keg->uk_ipers; i++) 996 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 997 keg->uk_size, wait) != 0) 998 break; 999 if (i != keg->uk_ipers) { 1000 keg_free_slab(keg, slab, i); 1001 slab = NULL; 1002 goto out; 1003 } 1004 } 1005 out: 1006 KEG_LOCK(keg); 1007 1008 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)", 1009 slab, keg->uk_name, keg); 1010 1011 if (slab != NULL) { 1012 if (keg->uk_flags & UMA_ZONE_HASH) 1013 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1014 1015 keg->uk_pages += keg->uk_ppera; 1016 keg->uk_free += keg->uk_ipers; 1017 } 1018 1019 return (slab); 1020 } 1021 1022 /* 1023 * This function is intended to be used early on in place of page_alloc() so 1024 * that we may use the boot time page cache to satisfy allocations before 1025 * the VM is ready. 1026 */ 1027 static void * 1028 startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) 1029 { 1030 uma_keg_t keg; 1031 void *mem; 1032 int pages; 1033 1034 keg = zone_first_keg(zone); 1035 pages = howmany(bytes, PAGE_SIZE); 1036 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n")); 1037 1038 /* 1039 * Check our small startup cache to see if it has pages remaining. 1040 */ 1041 mtx_lock(&uma_boot_pages_mtx); 1042 if (pages <= boot_pages) { 1043 mem = bootmem; 1044 boot_pages -= pages; 1045 bootmem += pages * PAGE_SIZE; 1046 mtx_unlock(&uma_boot_pages_mtx); 1047 *pflag = UMA_SLAB_BOOT; 1048 return (mem); 1049 } 1050 mtx_unlock(&uma_boot_pages_mtx); 1051 if (booted < UMA_STARTUP2) 1052 panic("UMA: Increase vm.boot_pages"); 1053 /* 1054 * Now that we've booted reset these users to their real allocator. 1055 */ 1056 #ifdef UMA_MD_SMALL_ALLOC 1057 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; 1058 #else 1059 keg->uk_allocf = page_alloc; 1060 #endif 1061 return keg->uk_allocf(zone, bytes, pflag, wait); 1062 } 1063 1064 /* 1065 * Allocates a number of pages from the system 1066 * 1067 * Arguments: 1068 * bytes The number of bytes requested 1069 * wait Shall we wait? 1070 * 1071 * Returns: 1072 * A pointer to the alloced memory or possibly 1073 * NULL if M_NOWAIT is set. 1074 */ 1075 static void * 1076 page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) 1077 { 1078 void *p; /* Returned page */ 1079 1080 *pflag = UMA_SLAB_KMEM; 1081 p = (void *) kmem_malloc(kmem_arena, bytes, wait); 1082 1083 return (p); 1084 } 1085 1086 /* 1087 * Allocates a number of pages from within an object 1088 * 1089 * Arguments: 1090 * bytes The number of bytes requested 1091 * wait Shall we wait? 1092 * 1093 * Returns: 1094 * A pointer to the alloced memory or possibly 1095 * NULL if M_NOWAIT is set. 1096 */ 1097 static void * 1098 noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait) 1099 { 1100 TAILQ_HEAD(, vm_page) alloctail; 1101 u_long npages; 1102 vm_offset_t retkva, zkva; 1103 vm_page_t p, p_next; 1104 uma_keg_t keg; 1105 1106 TAILQ_INIT(&alloctail); 1107 keg = zone_first_keg(zone); 1108 1109 npages = howmany(bytes, PAGE_SIZE); 1110 while (npages > 0) { 1111 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT | 1112 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ); 1113 if (p != NULL) { 1114 /* 1115 * Since the page does not belong to an object, its 1116 * listq is unused. 1117 */ 1118 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1119 npages--; 1120 continue; 1121 } 1122 if (wait & M_WAITOK) { 1123 VM_WAIT; 1124 continue; 1125 } 1126 1127 /* 1128 * Page allocation failed, free intermediate pages and 1129 * exit. 1130 */ 1131 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1132 vm_page_unwire(p, PQ_NONE); 1133 vm_page_free(p); 1134 } 1135 return (NULL); 1136 } 1137 *flags = UMA_SLAB_PRIV; 1138 zkva = keg->uk_kva + 1139 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1140 retkva = zkva; 1141 TAILQ_FOREACH(p, &alloctail, listq) { 1142 pmap_qenter(zkva, &p, 1); 1143 zkva += PAGE_SIZE; 1144 } 1145 1146 return ((void *)retkva); 1147 } 1148 1149 /* 1150 * Frees a number of pages to the system 1151 * 1152 * Arguments: 1153 * mem A pointer to the memory to be freed 1154 * size The size of the memory being freed 1155 * flags The original p->us_flags field 1156 * 1157 * Returns: 1158 * Nothing 1159 */ 1160 static void 1161 page_free(void *mem, vm_size_t size, uint8_t flags) 1162 { 1163 struct vmem *vmem; 1164 1165 if (flags & UMA_SLAB_KMEM) 1166 vmem = kmem_arena; 1167 else if (flags & UMA_SLAB_KERNEL) 1168 vmem = kernel_arena; 1169 else 1170 panic("UMA: page_free used with invalid flags %x", flags); 1171 1172 kmem_free(vmem, (vm_offset_t)mem, size); 1173 } 1174 1175 /* 1176 * Zero fill initializer 1177 * 1178 * Arguments/Returns follow uma_init specifications 1179 */ 1180 static int 1181 zero_init(void *mem, int size, int flags) 1182 { 1183 bzero(mem, size); 1184 return (0); 1185 } 1186 1187 /* 1188 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1189 * 1190 * Arguments 1191 * keg The zone we should initialize 1192 * 1193 * Returns 1194 * Nothing 1195 */ 1196 static void 1197 keg_small_init(uma_keg_t keg) 1198 { 1199 u_int rsize; 1200 u_int memused; 1201 u_int wastedspace; 1202 u_int shsize; 1203 u_int slabsize; 1204 1205 if (keg->uk_flags & UMA_ZONE_PCPU) { 1206 u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU; 1207 1208 slabsize = sizeof(struct pcpu); 1209 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu), 1210 PAGE_SIZE); 1211 } else { 1212 slabsize = UMA_SLAB_SIZE; 1213 keg->uk_ppera = 1; 1214 } 1215 1216 /* 1217 * Calculate the size of each allocation (rsize) according to 1218 * alignment. If the requested size is smaller than we have 1219 * allocation bits for we round it up. 1220 */ 1221 rsize = keg->uk_size; 1222 if (rsize < slabsize / SLAB_SETSIZE) 1223 rsize = slabsize / SLAB_SETSIZE; 1224 if (rsize & keg->uk_align) 1225 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1226 keg->uk_rsize = rsize; 1227 1228 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1229 keg->uk_rsize < sizeof(struct pcpu), 1230 ("%s: size %u too large", __func__, keg->uk_rsize)); 1231 1232 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1233 shsize = 0; 1234 else 1235 shsize = sizeof(struct uma_slab); 1236 1237 keg->uk_ipers = (slabsize - shsize) / rsize; 1238 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1239 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1240 1241 memused = keg->uk_ipers * rsize + shsize; 1242 wastedspace = slabsize - memused; 1243 1244 /* 1245 * We can't do OFFPAGE if we're internal or if we've been 1246 * asked to not go to the VM for buckets. If we do this we 1247 * may end up going to the VM for slabs which we do not 1248 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1249 * of UMA_ZONE_VM, which clearly forbids it. 1250 */ 1251 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1252 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1253 return; 1254 1255 /* 1256 * See if using an OFFPAGE slab will limit our waste. Only do 1257 * this if it permits more items per-slab. 1258 * 1259 * XXX We could try growing slabsize to limit max waste as well. 1260 * Historically this was not done because the VM could not 1261 * efficiently handle contiguous allocations. 1262 */ 1263 if ((wastedspace >= slabsize / UMA_MAX_WASTE) && 1264 (keg->uk_ipers < (slabsize / keg->uk_rsize))) { 1265 keg->uk_ipers = slabsize / keg->uk_rsize; 1266 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1267 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1268 CTR6(KTR_UMA, "UMA decided we need offpage slab headers for " 1269 "keg: %s(%p), calculated wastedspace = %d, " 1270 "maximum wasted space allowed = %d, " 1271 "calculated ipers = %d, " 1272 "new wasted space = %d\n", keg->uk_name, keg, wastedspace, 1273 slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1274 slabsize - keg->uk_ipers * keg->uk_rsize); 1275 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1276 } 1277 1278 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1279 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1280 keg->uk_flags |= UMA_ZONE_HASH; 1281 } 1282 1283 /* 1284 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1285 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1286 * more complicated. 1287 * 1288 * Arguments 1289 * keg The keg we should initialize 1290 * 1291 * Returns 1292 * Nothing 1293 */ 1294 static void 1295 keg_large_init(uma_keg_t keg) 1296 { 1297 u_int shsize; 1298 1299 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1300 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1301 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1302 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1303 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1304 1305 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1306 keg->uk_ipers = 1; 1307 keg->uk_rsize = keg->uk_size; 1308 1309 /* Check whether we have enough space to not do OFFPAGE. */ 1310 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) { 1311 shsize = sizeof(struct uma_slab); 1312 if (shsize & UMA_ALIGN_PTR) 1313 shsize = (shsize & ~UMA_ALIGN_PTR) + 1314 (UMA_ALIGN_PTR + 1); 1315 1316 if (PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < shsize) { 1317 /* 1318 * We can't do OFFPAGE if we're internal, in which case 1319 * we need an extra page per allocation to contain the 1320 * slab header. 1321 */ 1322 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0) 1323 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1324 else 1325 keg->uk_ppera++; 1326 } 1327 } 1328 1329 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1330 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1331 keg->uk_flags |= UMA_ZONE_HASH; 1332 } 1333 1334 static void 1335 keg_cachespread_init(uma_keg_t keg) 1336 { 1337 int alignsize; 1338 int trailer; 1339 int pages; 1340 int rsize; 1341 1342 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1343 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1344 1345 alignsize = keg->uk_align + 1; 1346 rsize = keg->uk_size; 1347 /* 1348 * We want one item to start on every align boundary in a page. To 1349 * do this we will span pages. We will also extend the item by the 1350 * size of align if it is an even multiple of align. Otherwise, it 1351 * would fall on the same boundary every time. 1352 */ 1353 if (rsize & keg->uk_align) 1354 rsize = (rsize & ~keg->uk_align) + alignsize; 1355 if ((rsize & alignsize) == 0) 1356 rsize += alignsize; 1357 trailer = rsize - keg->uk_size; 1358 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1359 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1360 keg->uk_rsize = rsize; 1361 keg->uk_ppera = pages; 1362 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1363 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1364 KASSERT(keg->uk_ipers <= SLAB_SETSIZE, 1365 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1366 keg->uk_ipers)); 1367 } 1368 1369 /* 1370 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1371 * the keg onto the global keg list. 1372 * 1373 * Arguments/Returns follow uma_ctor specifications 1374 * udata Actually uma_kctor_args 1375 */ 1376 static int 1377 keg_ctor(void *mem, int size, void *udata, int flags) 1378 { 1379 struct uma_kctor_args *arg = udata; 1380 uma_keg_t keg = mem; 1381 uma_zone_t zone; 1382 1383 bzero(keg, size); 1384 keg->uk_size = arg->size; 1385 keg->uk_init = arg->uminit; 1386 keg->uk_fini = arg->fini; 1387 keg->uk_align = arg->align; 1388 keg->uk_free = 0; 1389 keg->uk_reserve = 0; 1390 keg->uk_pages = 0; 1391 keg->uk_flags = arg->flags; 1392 keg->uk_slabzone = NULL; 1393 1394 /* 1395 * The master zone is passed to us at keg-creation time. 1396 */ 1397 zone = arg->zone; 1398 keg->uk_name = zone->uz_name; 1399 1400 if (arg->flags & UMA_ZONE_VM) 1401 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1402 1403 if (arg->flags & UMA_ZONE_ZINIT) 1404 keg->uk_init = zero_init; 1405 1406 if (arg->flags & UMA_ZONE_MALLOC) 1407 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1408 1409 if (arg->flags & UMA_ZONE_PCPU) 1410 #ifdef SMP 1411 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1412 #else 1413 keg->uk_flags &= ~UMA_ZONE_PCPU; 1414 #endif 1415 1416 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1417 keg_cachespread_init(keg); 1418 } else { 1419 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab))) 1420 keg_large_init(keg); 1421 else 1422 keg_small_init(keg); 1423 } 1424 1425 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1426 keg->uk_slabzone = slabzone; 1427 1428 /* 1429 * If we haven't booted yet we need allocations to go through the 1430 * startup cache until the vm is ready. 1431 */ 1432 if (booted < UMA_STARTUP2) 1433 keg->uk_allocf = startup_alloc; 1434 #ifdef UMA_MD_SMALL_ALLOC 1435 else if (keg->uk_ppera == 1) 1436 keg->uk_allocf = uma_small_alloc; 1437 #endif 1438 else 1439 keg->uk_allocf = page_alloc; 1440 #ifdef UMA_MD_SMALL_ALLOC 1441 if (keg->uk_ppera == 1) 1442 keg->uk_freef = uma_small_free; 1443 else 1444 #endif 1445 keg->uk_freef = page_free; 1446 1447 /* 1448 * Initialize keg's lock 1449 */ 1450 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1451 1452 /* 1453 * If we're putting the slab header in the actual page we need to 1454 * figure out where in each page it goes. This calculates a right 1455 * justified offset into the memory on an ALIGN_PTR boundary. 1456 */ 1457 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1458 u_int totsize; 1459 1460 /* Size of the slab struct and free list */ 1461 totsize = sizeof(struct uma_slab); 1462 1463 if (totsize & UMA_ALIGN_PTR) 1464 totsize = (totsize & ~UMA_ALIGN_PTR) + 1465 (UMA_ALIGN_PTR + 1); 1466 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1467 1468 /* 1469 * The only way the following is possible is if with our 1470 * UMA_ALIGN_PTR adjustments we are now bigger than 1471 * UMA_SLAB_SIZE. I haven't checked whether this is 1472 * mathematically possible for all cases, so we make 1473 * sure here anyway. 1474 */ 1475 totsize = keg->uk_pgoff + sizeof(struct uma_slab); 1476 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1477 printf("zone %s ipers %d rsize %d size %d\n", 1478 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1479 keg->uk_size); 1480 panic("UMA slab won't fit."); 1481 } 1482 } 1483 1484 if (keg->uk_flags & UMA_ZONE_HASH) 1485 hash_alloc(&keg->uk_hash); 1486 1487 CTR5(KTR_UMA, "keg_ctor %p zone %s(%p) out %d free %d\n", 1488 keg, zone->uz_name, zone, 1489 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, 1490 keg->uk_free); 1491 1492 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1493 1494 rw_wlock(&uma_rwlock); 1495 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1496 rw_wunlock(&uma_rwlock); 1497 return (0); 1498 } 1499 1500 /* 1501 * Zone header ctor. This initializes all fields, locks, etc. 1502 * 1503 * Arguments/Returns follow uma_ctor specifications 1504 * udata Actually uma_zctor_args 1505 */ 1506 static int 1507 zone_ctor(void *mem, int size, void *udata, int flags) 1508 { 1509 struct uma_zctor_args *arg = udata; 1510 uma_zone_t zone = mem; 1511 uma_zone_t z; 1512 uma_keg_t keg; 1513 1514 bzero(zone, size); 1515 zone->uz_name = arg->name; 1516 zone->uz_ctor = arg->ctor; 1517 zone->uz_dtor = arg->dtor; 1518 zone->uz_slab = zone_fetch_slab; 1519 zone->uz_init = NULL; 1520 zone->uz_fini = NULL; 1521 zone->uz_allocs = 0; 1522 zone->uz_frees = 0; 1523 zone->uz_fails = 0; 1524 zone->uz_sleeps = 0; 1525 zone->uz_count = 0; 1526 zone->uz_count_min = 0; 1527 zone->uz_flags = 0; 1528 zone->uz_warning = NULL; 1529 timevalclear(&zone->uz_ratecheck); 1530 keg = arg->keg; 1531 1532 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 1533 1534 /* 1535 * This is a pure cache zone, no kegs. 1536 */ 1537 if (arg->import) { 1538 if (arg->flags & UMA_ZONE_VM) 1539 arg->flags |= UMA_ZFLAG_CACHEONLY; 1540 zone->uz_flags = arg->flags; 1541 zone->uz_size = arg->size; 1542 zone->uz_import = arg->import; 1543 zone->uz_release = arg->release; 1544 zone->uz_arg = arg->arg; 1545 zone->uz_lockptr = &zone->uz_lock; 1546 rw_wlock(&uma_rwlock); 1547 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 1548 rw_wunlock(&uma_rwlock); 1549 goto out; 1550 } 1551 1552 /* 1553 * Use the regular zone/keg/slab allocator. 1554 */ 1555 zone->uz_import = (uma_import)zone_import; 1556 zone->uz_release = (uma_release)zone_release; 1557 zone->uz_arg = zone; 1558 1559 if (arg->flags & UMA_ZONE_SECONDARY) { 1560 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1561 zone->uz_init = arg->uminit; 1562 zone->uz_fini = arg->fini; 1563 zone->uz_lockptr = &keg->uk_lock; 1564 zone->uz_flags |= UMA_ZONE_SECONDARY; 1565 rw_wlock(&uma_rwlock); 1566 ZONE_LOCK(zone); 1567 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1568 if (LIST_NEXT(z, uz_link) == NULL) { 1569 LIST_INSERT_AFTER(z, zone, uz_link); 1570 break; 1571 } 1572 } 1573 ZONE_UNLOCK(zone); 1574 rw_wunlock(&uma_rwlock); 1575 } else if (keg == NULL) { 1576 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1577 arg->align, arg->flags)) == NULL) 1578 return (ENOMEM); 1579 } else { 1580 struct uma_kctor_args karg; 1581 int error; 1582 1583 /* We should only be here from uma_startup() */ 1584 karg.size = arg->size; 1585 karg.uminit = arg->uminit; 1586 karg.fini = arg->fini; 1587 karg.align = arg->align; 1588 karg.flags = arg->flags; 1589 karg.zone = zone; 1590 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1591 flags); 1592 if (error) 1593 return (error); 1594 } 1595 1596 /* 1597 * Link in the first keg. 1598 */ 1599 zone->uz_klink.kl_keg = keg; 1600 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1601 zone->uz_lockptr = &keg->uk_lock; 1602 zone->uz_size = keg->uk_size; 1603 zone->uz_flags |= (keg->uk_flags & 1604 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1605 1606 /* 1607 * Some internal zones don't have room allocated for the per cpu 1608 * caches. If we're internal, bail out here. 1609 */ 1610 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1611 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1612 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1613 return (0); 1614 } 1615 1616 out: 1617 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0) 1618 zone->uz_count = bucket_select(zone->uz_size); 1619 else 1620 zone->uz_count = BUCKET_MAX; 1621 zone->uz_count_min = zone->uz_count; 1622 1623 return (0); 1624 } 1625 1626 /* 1627 * Keg header dtor. This frees all data, destroys locks, frees the hash 1628 * table and removes the keg from the global list. 1629 * 1630 * Arguments/Returns follow uma_dtor specifications 1631 * udata unused 1632 */ 1633 static void 1634 keg_dtor(void *arg, int size, void *udata) 1635 { 1636 uma_keg_t keg; 1637 1638 keg = (uma_keg_t)arg; 1639 KEG_LOCK(keg); 1640 if (keg->uk_free != 0) { 1641 printf("Freed UMA keg (%s) was not empty (%d items). " 1642 " Lost %d pages of memory.\n", 1643 keg->uk_name ? keg->uk_name : "", 1644 keg->uk_free, keg->uk_pages); 1645 } 1646 KEG_UNLOCK(keg); 1647 1648 hash_free(&keg->uk_hash); 1649 1650 KEG_LOCK_FINI(keg); 1651 } 1652 1653 /* 1654 * Zone header dtor. 1655 * 1656 * Arguments/Returns follow uma_dtor specifications 1657 * udata unused 1658 */ 1659 static void 1660 zone_dtor(void *arg, int size, void *udata) 1661 { 1662 uma_klink_t klink; 1663 uma_zone_t zone; 1664 uma_keg_t keg; 1665 1666 zone = (uma_zone_t)arg; 1667 keg = zone_first_keg(zone); 1668 1669 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1670 cache_drain(zone); 1671 1672 rw_wlock(&uma_rwlock); 1673 LIST_REMOVE(zone, uz_link); 1674 rw_wunlock(&uma_rwlock); 1675 /* 1676 * XXX there are some races here where 1677 * the zone can be drained but zone lock 1678 * released and then refilled before we 1679 * remove it... we dont care for now 1680 */ 1681 zone_drain_wait(zone, M_WAITOK); 1682 /* 1683 * Unlink all of our kegs. 1684 */ 1685 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1686 klink->kl_keg = NULL; 1687 LIST_REMOVE(klink, kl_link); 1688 if (klink == &zone->uz_klink) 1689 continue; 1690 free(klink, M_TEMP); 1691 } 1692 /* 1693 * We only destroy kegs from non secondary zones. 1694 */ 1695 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1696 rw_wlock(&uma_rwlock); 1697 LIST_REMOVE(keg, uk_link); 1698 rw_wunlock(&uma_rwlock); 1699 zone_free_item(kegs, keg, NULL, SKIP_NONE); 1700 } 1701 ZONE_LOCK_FINI(zone); 1702 } 1703 1704 /* 1705 * Traverses every zone in the system and calls a callback 1706 * 1707 * Arguments: 1708 * zfunc A pointer to a function which accepts a zone 1709 * as an argument. 1710 * 1711 * Returns: 1712 * Nothing 1713 */ 1714 static void 1715 zone_foreach(void (*zfunc)(uma_zone_t)) 1716 { 1717 uma_keg_t keg; 1718 uma_zone_t zone; 1719 1720 rw_rlock(&uma_rwlock); 1721 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1722 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1723 zfunc(zone); 1724 } 1725 rw_runlock(&uma_rwlock); 1726 } 1727 1728 /* Public functions */ 1729 /* See uma.h */ 1730 void 1731 uma_startup(void *mem, int npages) 1732 { 1733 struct uma_zctor_args args; 1734 1735 rw_init(&uma_rwlock, "UMA lock"); 1736 1737 /* "manually" create the initial zone */ 1738 memset(&args, 0, sizeof(args)); 1739 args.name = "UMA Kegs"; 1740 args.size = sizeof(struct uma_keg); 1741 args.ctor = keg_ctor; 1742 args.dtor = keg_dtor; 1743 args.uminit = zero_init; 1744 args.fini = NULL; 1745 args.keg = &masterkeg; 1746 args.align = 32 - 1; 1747 args.flags = UMA_ZFLAG_INTERNAL; 1748 /* The initial zone has no Per cpu queues so it's smaller */ 1749 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); 1750 1751 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); 1752 bootmem = mem; 1753 boot_pages = npages; 1754 1755 args.name = "UMA Zones"; 1756 args.size = sizeof(struct uma_zone) + 1757 (sizeof(struct uma_cache) * (mp_maxid + 1)); 1758 args.ctor = zone_ctor; 1759 args.dtor = zone_dtor; 1760 args.uminit = zero_init; 1761 args.fini = NULL; 1762 args.keg = NULL; 1763 args.align = 32 - 1; 1764 args.flags = UMA_ZFLAG_INTERNAL; 1765 /* The initial zone has no Per cpu queues so it's smaller */ 1766 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); 1767 1768 /* Now make a zone for slab headers */ 1769 slabzone = uma_zcreate("UMA Slabs", 1770 sizeof(struct uma_slab), 1771 NULL, NULL, NULL, NULL, 1772 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1773 1774 hashzone = uma_zcreate("UMA Hash", 1775 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 1776 NULL, NULL, NULL, NULL, 1777 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1778 1779 bucket_init(); 1780 1781 booted = UMA_STARTUP; 1782 } 1783 1784 /* see uma.h */ 1785 void 1786 uma_startup2(void) 1787 { 1788 booted = UMA_STARTUP2; 1789 bucket_enable(); 1790 sx_init(&uma_drain_lock, "umadrain"); 1791 } 1792 1793 /* 1794 * Initialize our callout handle 1795 * 1796 */ 1797 1798 static void 1799 uma_startup3(void) 1800 { 1801 1802 callout_init(&uma_callout, 1); 1803 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1804 } 1805 1806 static uma_keg_t 1807 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 1808 int align, uint32_t flags) 1809 { 1810 struct uma_kctor_args args; 1811 1812 args.size = size; 1813 args.uminit = uminit; 1814 args.fini = fini; 1815 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 1816 args.flags = flags; 1817 args.zone = zone; 1818 return (zone_alloc_item(kegs, &args, M_WAITOK)); 1819 } 1820 1821 /* See uma.h */ 1822 void 1823 uma_set_align(int align) 1824 { 1825 1826 if (align != UMA_ALIGN_CACHE) 1827 uma_align_cache = align; 1828 } 1829 1830 /* See uma.h */ 1831 uma_zone_t 1832 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 1833 uma_init uminit, uma_fini fini, int align, uint32_t flags) 1834 1835 { 1836 struct uma_zctor_args args; 1837 uma_zone_t res; 1838 bool locked; 1839 1840 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"", 1841 align, name)); 1842 1843 /* This stuff is essential for the zone ctor */ 1844 memset(&args, 0, sizeof(args)); 1845 args.name = name; 1846 args.size = size; 1847 args.ctor = ctor; 1848 args.dtor = dtor; 1849 args.uminit = uminit; 1850 args.fini = fini; 1851 #ifdef INVARIANTS 1852 /* 1853 * If a zone is being created with an empty constructor and 1854 * destructor, pass UMA constructor/destructor which checks for 1855 * memory use after free. 1856 */ 1857 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) && 1858 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) { 1859 args.ctor = trash_ctor; 1860 args.dtor = trash_dtor; 1861 args.uminit = trash_init; 1862 args.fini = trash_fini; 1863 } 1864 #endif 1865 args.align = align; 1866 args.flags = flags; 1867 args.keg = NULL; 1868 1869 if (booted < UMA_STARTUP2) { 1870 locked = false; 1871 } else { 1872 sx_slock(&uma_drain_lock); 1873 locked = true; 1874 } 1875 res = zone_alloc_item(zones, &args, M_WAITOK); 1876 if (locked) 1877 sx_sunlock(&uma_drain_lock); 1878 return (res); 1879 } 1880 1881 /* See uma.h */ 1882 uma_zone_t 1883 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 1884 uma_init zinit, uma_fini zfini, uma_zone_t master) 1885 { 1886 struct uma_zctor_args args; 1887 uma_keg_t keg; 1888 uma_zone_t res; 1889 bool locked; 1890 1891 keg = zone_first_keg(master); 1892 memset(&args, 0, sizeof(args)); 1893 args.name = name; 1894 args.size = keg->uk_size; 1895 args.ctor = ctor; 1896 args.dtor = dtor; 1897 args.uminit = zinit; 1898 args.fini = zfini; 1899 args.align = keg->uk_align; 1900 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 1901 args.keg = keg; 1902 1903 if (booted < UMA_STARTUP2) { 1904 locked = false; 1905 } else { 1906 sx_slock(&uma_drain_lock); 1907 locked = true; 1908 } 1909 /* XXX Attaches only one keg of potentially many. */ 1910 res = zone_alloc_item(zones, &args, M_WAITOK); 1911 if (locked) 1912 sx_sunlock(&uma_drain_lock); 1913 return (res); 1914 } 1915 1916 /* See uma.h */ 1917 uma_zone_t 1918 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 1919 uma_init zinit, uma_fini zfini, uma_import zimport, 1920 uma_release zrelease, void *arg, int flags) 1921 { 1922 struct uma_zctor_args args; 1923 1924 memset(&args, 0, sizeof(args)); 1925 args.name = name; 1926 args.size = size; 1927 args.ctor = ctor; 1928 args.dtor = dtor; 1929 args.uminit = zinit; 1930 args.fini = zfini; 1931 args.import = zimport; 1932 args.release = zrelease; 1933 args.arg = arg; 1934 args.align = 0; 1935 args.flags = flags; 1936 1937 return (zone_alloc_item(zones, &args, M_WAITOK)); 1938 } 1939 1940 static void 1941 zone_lock_pair(uma_zone_t a, uma_zone_t b) 1942 { 1943 if (a < b) { 1944 ZONE_LOCK(a); 1945 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); 1946 } else { 1947 ZONE_LOCK(b); 1948 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); 1949 } 1950 } 1951 1952 static void 1953 zone_unlock_pair(uma_zone_t a, uma_zone_t b) 1954 { 1955 1956 ZONE_UNLOCK(a); 1957 ZONE_UNLOCK(b); 1958 } 1959 1960 int 1961 uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 1962 { 1963 uma_klink_t klink; 1964 uma_klink_t kl; 1965 int error; 1966 1967 error = 0; 1968 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 1969 1970 zone_lock_pair(zone, master); 1971 /* 1972 * zone must use vtoslab() to resolve objects and must already be 1973 * a secondary. 1974 */ 1975 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 1976 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 1977 error = EINVAL; 1978 goto out; 1979 } 1980 /* 1981 * The new master must also use vtoslab(). 1982 */ 1983 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 1984 error = EINVAL; 1985 goto out; 1986 } 1987 1988 /* 1989 * The underlying object must be the same size. rsize 1990 * may be different. 1991 */ 1992 if (master->uz_size != zone->uz_size) { 1993 error = E2BIG; 1994 goto out; 1995 } 1996 /* 1997 * Put it at the end of the list. 1998 */ 1999 klink->kl_keg = zone_first_keg(master); 2000 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 2001 if (LIST_NEXT(kl, kl_link) == NULL) { 2002 LIST_INSERT_AFTER(kl, klink, kl_link); 2003 break; 2004 } 2005 } 2006 klink = NULL; 2007 zone->uz_flags |= UMA_ZFLAG_MULTI; 2008 zone->uz_slab = zone_fetch_slab_multi; 2009 2010 out: 2011 zone_unlock_pair(zone, master); 2012 if (klink != NULL) 2013 free(klink, M_TEMP); 2014 2015 return (error); 2016 } 2017 2018 2019 /* See uma.h */ 2020 void 2021 uma_zdestroy(uma_zone_t zone) 2022 { 2023 2024 sx_slock(&uma_drain_lock); 2025 zone_free_item(zones, zone, NULL, SKIP_NONE); 2026 sx_sunlock(&uma_drain_lock); 2027 } 2028 2029 /* See uma.h */ 2030 void * 2031 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2032 { 2033 void *item; 2034 uma_cache_t cache; 2035 uma_bucket_t bucket; 2036 int lockfail; 2037 int cpu; 2038 2039 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2040 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA); 2041 2042 /* This is the fast path allocation */ 2043 CTR4(KTR_UMA, "uma_zalloc_arg thread %x zone %s(%p) flags %d", 2044 curthread, zone->uz_name, zone, flags); 2045 2046 if (flags & M_WAITOK) { 2047 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2048 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2049 } 2050 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2051 ("uma_zalloc_arg: called with spinlock or critical section held")); 2052 2053 #ifdef DEBUG_MEMGUARD 2054 if (memguard_cmp_zone(zone)) { 2055 item = memguard_alloc(zone->uz_size, flags); 2056 if (item != NULL) { 2057 if (zone->uz_init != NULL && 2058 zone->uz_init(item, zone->uz_size, flags) != 0) 2059 return (NULL); 2060 if (zone->uz_ctor != NULL && 2061 zone->uz_ctor(item, zone->uz_size, udata, 2062 flags) != 0) { 2063 zone->uz_fini(item, zone->uz_size); 2064 return (NULL); 2065 } 2066 return (item); 2067 } 2068 /* This is unfortunate but should not be fatal. */ 2069 } 2070 #endif 2071 /* 2072 * If possible, allocate from the per-CPU cache. There are two 2073 * requirements for safe access to the per-CPU cache: (1) the thread 2074 * accessing the cache must not be preempted or yield during access, 2075 * and (2) the thread must not migrate CPUs without switching which 2076 * cache it accesses. We rely on a critical section to prevent 2077 * preemption and migration. We release the critical section in 2078 * order to acquire the zone mutex if we are unable to allocate from 2079 * the current cache; when we re-acquire the critical section, we 2080 * must detect and handle migration if it has occurred. 2081 */ 2082 critical_enter(); 2083 cpu = curcpu; 2084 cache = &zone->uz_cpu[cpu]; 2085 2086 zalloc_start: 2087 bucket = cache->uc_allocbucket; 2088 if (bucket != NULL && bucket->ub_cnt > 0) { 2089 bucket->ub_cnt--; 2090 item = bucket->ub_bucket[bucket->ub_cnt]; 2091 #ifdef INVARIANTS 2092 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2093 #endif 2094 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2095 cache->uc_allocs++; 2096 critical_exit(); 2097 if (zone->uz_ctor != NULL && 2098 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2099 atomic_add_long(&zone->uz_fails, 1); 2100 zone_free_item(zone, item, udata, SKIP_DTOR); 2101 return (NULL); 2102 } 2103 #ifdef INVARIANTS 2104 uma_dbg_alloc(zone, NULL, item); 2105 #endif 2106 if (flags & M_ZERO) 2107 uma_zero_item(item, zone); 2108 return (item); 2109 } 2110 2111 /* 2112 * We have run out of items in our alloc bucket. 2113 * See if we can switch with our free bucket. 2114 */ 2115 bucket = cache->uc_freebucket; 2116 if (bucket != NULL && bucket->ub_cnt > 0) { 2117 CTR2(KTR_UMA, 2118 "uma_zalloc: zone %s(%p) swapping empty with alloc", 2119 zone->uz_name, zone); 2120 cache->uc_freebucket = cache->uc_allocbucket; 2121 cache->uc_allocbucket = bucket; 2122 goto zalloc_start; 2123 } 2124 2125 /* 2126 * Discard any empty allocation bucket while we hold no locks. 2127 */ 2128 bucket = cache->uc_allocbucket; 2129 cache->uc_allocbucket = NULL; 2130 critical_exit(); 2131 if (bucket != NULL) 2132 bucket_free(zone, bucket, udata); 2133 2134 /* Short-circuit for zones without buckets and low memory. */ 2135 if (zone->uz_count == 0 || bucketdisable) 2136 goto zalloc_item; 2137 2138 /* 2139 * Attempt to retrieve the item from the per-CPU cache has failed, so 2140 * we must go back to the zone. This requires the zone lock, so we 2141 * must drop the critical section, then re-acquire it when we go back 2142 * to the cache. Since the critical section is released, we may be 2143 * preempted or migrate. As such, make sure not to maintain any 2144 * thread-local state specific to the cache from prior to releasing 2145 * the critical section. 2146 */ 2147 lockfail = 0; 2148 if (ZONE_TRYLOCK(zone) == 0) { 2149 /* Record contention to size the buckets. */ 2150 ZONE_LOCK(zone); 2151 lockfail = 1; 2152 } 2153 critical_enter(); 2154 cpu = curcpu; 2155 cache = &zone->uz_cpu[cpu]; 2156 2157 /* 2158 * Since we have locked the zone we may as well send back our stats. 2159 */ 2160 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2161 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2162 cache->uc_allocs = 0; 2163 cache->uc_frees = 0; 2164 2165 /* See if we lost the race to fill the cache. */ 2166 if (cache->uc_allocbucket != NULL) { 2167 ZONE_UNLOCK(zone); 2168 goto zalloc_start; 2169 } 2170 2171 /* 2172 * Check the zone's cache of buckets. 2173 */ 2174 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 2175 KASSERT(bucket->ub_cnt != 0, 2176 ("uma_zalloc_arg: Returning an empty bucket.")); 2177 2178 LIST_REMOVE(bucket, ub_link); 2179 cache->uc_allocbucket = bucket; 2180 ZONE_UNLOCK(zone); 2181 goto zalloc_start; 2182 } 2183 /* We are no longer associated with this CPU. */ 2184 critical_exit(); 2185 2186 /* 2187 * We bump the uz count when the cache size is insufficient to 2188 * handle the working set. 2189 */ 2190 if (lockfail && zone->uz_count < BUCKET_MAX) 2191 zone->uz_count++; 2192 ZONE_UNLOCK(zone); 2193 2194 /* 2195 * Now lets just fill a bucket and put it on the free list. If that 2196 * works we'll restart the allocation from the beginning and it 2197 * will use the just filled bucket. 2198 */ 2199 bucket = zone_alloc_bucket(zone, udata, flags); 2200 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p", 2201 zone->uz_name, zone, bucket); 2202 if (bucket != NULL) { 2203 ZONE_LOCK(zone); 2204 critical_enter(); 2205 cpu = curcpu; 2206 cache = &zone->uz_cpu[cpu]; 2207 /* 2208 * See if we lost the race or were migrated. Cache the 2209 * initialized bucket to make this less likely or claim 2210 * the memory directly. 2211 */ 2212 if (cache->uc_allocbucket == NULL) 2213 cache->uc_allocbucket = bucket; 2214 else 2215 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2216 ZONE_UNLOCK(zone); 2217 goto zalloc_start; 2218 } 2219 2220 /* 2221 * We may not be able to get a bucket so return an actual item. 2222 */ 2223 zalloc_item: 2224 item = zone_alloc_item(zone, udata, flags); 2225 2226 return (item); 2227 } 2228 2229 static uma_slab_t 2230 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) 2231 { 2232 uma_slab_t slab; 2233 int reserve; 2234 2235 mtx_assert(&keg->uk_lock, MA_OWNED); 2236 slab = NULL; 2237 reserve = 0; 2238 if ((flags & M_USE_RESERVE) == 0) 2239 reserve = keg->uk_reserve; 2240 2241 for (;;) { 2242 /* 2243 * Find a slab with some space. Prefer slabs that are partially 2244 * used over those that are totally full. This helps to reduce 2245 * fragmentation. 2246 */ 2247 if (keg->uk_free > reserve) { 2248 if (!LIST_EMPTY(&keg->uk_part_slab)) { 2249 slab = LIST_FIRST(&keg->uk_part_slab); 2250 } else { 2251 slab = LIST_FIRST(&keg->uk_free_slab); 2252 LIST_REMOVE(slab, us_link); 2253 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, 2254 us_link); 2255 } 2256 MPASS(slab->us_keg == keg); 2257 return (slab); 2258 } 2259 2260 /* 2261 * M_NOVM means don't ask at all! 2262 */ 2263 if (flags & M_NOVM) 2264 break; 2265 2266 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2267 keg->uk_flags |= UMA_ZFLAG_FULL; 2268 /* 2269 * If this is not a multi-zone, set the FULL bit. 2270 * Otherwise slab_multi() takes care of it. 2271 */ 2272 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2273 zone->uz_flags |= UMA_ZFLAG_FULL; 2274 zone_log_warning(zone); 2275 zone_maxaction(zone); 2276 } 2277 if (flags & M_NOWAIT) 2278 break; 2279 zone->uz_sleeps++; 2280 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2281 continue; 2282 } 2283 slab = keg_alloc_slab(keg, zone, flags); 2284 /* 2285 * If we got a slab here it's safe to mark it partially used 2286 * and return. We assume that the caller is going to remove 2287 * at least one item. 2288 */ 2289 if (slab) { 2290 MPASS(slab->us_keg == keg); 2291 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2292 return (slab); 2293 } 2294 /* 2295 * We might not have been able to get a slab but another cpu 2296 * could have while we were unlocked. Check again before we 2297 * fail. 2298 */ 2299 flags |= M_NOVM; 2300 } 2301 return (slab); 2302 } 2303 2304 static uma_slab_t 2305 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) 2306 { 2307 uma_slab_t slab; 2308 2309 if (keg == NULL) { 2310 keg = zone_first_keg(zone); 2311 KEG_LOCK(keg); 2312 } 2313 2314 for (;;) { 2315 slab = keg_fetch_slab(keg, zone, flags); 2316 if (slab) 2317 return (slab); 2318 if (flags & (M_NOWAIT | M_NOVM)) 2319 break; 2320 } 2321 KEG_UNLOCK(keg); 2322 return (NULL); 2323 } 2324 2325 /* 2326 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2327 * with the keg locked. On NULL no lock is held. 2328 * 2329 * The last pointer is used to seed the search. It is not required. 2330 */ 2331 static uma_slab_t 2332 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) 2333 { 2334 uma_klink_t klink; 2335 uma_slab_t slab; 2336 uma_keg_t keg; 2337 int flags; 2338 int empty; 2339 int full; 2340 2341 /* 2342 * Don't wait on the first pass. This will skip limit tests 2343 * as well. We don't want to block if we can find a provider 2344 * without blocking. 2345 */ 2346 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2347 /* 2348 * Use the last slab allocated as a hint for where to start 2349 * the search. 2350 */ 2351 if (last != NULL) { 2352 slab = keg_fetch_slab(last, zone, flags); 2353 if (slab) 2354 return (slab); 2355 KEG_UNLOCK(last); 2356 } 2357 /* 2358 * Loop until we have a slab incase of transient failures 2359 * while M_WAITOK is specified. I'm not sure this is 100% 2360 * required but we've done it for so long now. 2361 */ 2362 for (;;) { 2363 empty = 0; 2364 full = 0; 2365 /* 2366 * Search the available kegs for slabs. Be careful to hold the 2367 * correct lock while calling into the keg layer. 2368 */ 2369 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2370 keg = klink->kl_keg; 2371 KEG_LOCK(keg); 2372 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2373 slab = keg_fetch_slab(keg, zone, flags); 2374 if (slab) 2375 return (slab); 2376 } 2377 if (keg->uk_flags & UMA_ZFLAG_FULL) 2378 full++; 2379 else 2380 empty++; 2381 KEG_UNLOCK(keg); 2382 } 2383 if (rflags & (M_NOWAIT | M_NOVM)) 2384 break; 2385 flags = rflags; 2386 /* 2387 * All kegs are full. XXX We can't atomically check all kegs 2388 * and sleep so just sleep for a short period and retry. 2389 */ 2390 if (full && !empty) { 2391 ZONE_LOCK(zone); 2392 zone->uz_flags |= UMA_ZFLAG_FULL; 2393 zone->uz_sleeps++; 2394 zone_log_warning(zone); 2395 zone_maxaction(zone); 2396 msleep(zone, zone->uz_lockptr, PVM, 2397 "zonelimit", hz/100); 2398 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2399 ZONE_UNLOCK(zone); 2400 continue; 2401 } 2402 } 2403 return (NULL); 2404 } 2405 2406 static void * 2407 slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 2408 { 2409 void *item; 2410 uint8_t freei; 2411 2412 MPASS(keg == slab->us_keg); 2413 mtx_assert(&keg->uk_lock, MA_OWNED); 2414 2415 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; 2416 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); 2417 item = slab->us_data + (keg->uk_rsize * freei); 2418 slab->us_freecount--; 2419 keg->uk_free--; 2420 2421 /* Move this slab to the full list */ 2422 if (slab->us_freecount == 0) { 2423 LIST_REMOVE(slab, us_link); 2424 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); 2425 } 2426 2427 return (item); 2428 } 2429 2430 static int 2431 zone_import(uma_zone_t zone, void **bucket, int max, int flags) 2432 { 2433 uma_slab_t slab; 2434 uma_keg_t keg; 2435 int i; 2436 2437 slab = NULL; 2438 keg = NULL; 2439 /* Try to keep the buckets totally full */ 2440 for (i = 0; i < max; ) { 2441 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL) 2442 break; 2443 keg = slab->us_keg; 2444 while (slab->us_freecount && i < max) { 2445 bucket[i++] = slab_alloc_item(keg, slab); 2446 if (keg->uk_free <= keg->uk_reserve) 2447 break; 2448 } 2449 /* Don't grab more than one slab at a time. */ 2450 flags &= ~M_WAITOK; 2451 flags |= M_NOWAIT; 2452 } 2453 if (slab != NULL) 2454 KEG_UNLOCK(keg); 2455 2456 return i; 2457 } 2458 2459 static uma_bucket_t 2460 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags) 2461 { 2462 uma_bucket_t bucket; 2463 int max; 2464 2465 /* Don't wait for buckets, preserve caller's NOVM setting. */ 2466 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 2467 if (bucket == NULL) 2468 return (NULL); 2469 2470 max = MIN(bucket->ub_entries, zone->uz_count); 2471 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 2472 max, flags); 2473 2474 /* 2475 * Initialize the memory if necessary. 2476 */ 2477 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 2478 int i; 2479 2480 for (i = 0; i < bucket->ub_cnt; i++) 2481 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2482 flags) != 0) 2483 break; 2484 /* 2485 * If we couldn't initialize the whole bucket, put the 2486 * rest back onto the freelist. 2487 */ 2488 if (i != bucket->ub_cnt) { 2489 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 2490 bucket->ub_cnt - i); 2491 #ifdef INVARIANTS 2492 bzero(&bucket->ub_bucket[i], 2493 sizeof(void *) * (bucket->ub_cnt - i)); 2494 #endif 2495 bucket->ub_cnt = i; 2496 } 2497 } 2498 2499 if (bucket->ub_cnt == 0) { 2500 bucket_free(zone, bucket, udata); 2501 atomic_add_long(&zone->uz_fails, 1); 2502 return (NULL); 2503 } 2504 2505 return (bucket); 2506 } 2507 2508 /* 2509 * Allocates a single item from a zone. 2510 * 2511 * Arguments 2512 * zone The zone to alloc for. 2513 * udata The data to be passed to the constructor. 2514 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2515 * 2516 * Returns 2517 * NULL if there is no memory and M_NOWAIT is set 2518 * An item if successful 2519 */ 2520 2521 static void * 2522 zone_alloc_item(uma_zone_t zone, void *udata, int flags) 2523 { 2524 void *item; 2525 2526 item = NULL; 2527 2528 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1) 2529 goto fail; 2530 atomic_add_long(&zone->uz_allocs, 1); 2531 2532 /* 2533 * We have to call both the zone's init (not the keg's init) 2534 * and the zone's ctor. This is because the item is going from 2535 * a keg slab directly to the user, and the user is expecting it 2536 * to be both zone-init'd as well as zone-ctor'd. 2537 */ 2538 if (zone->uz_init != NULL) { 2539 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 2540 zone_free_item(zone, item, udata, SKIP_FINI); 2541 goto fail; 2542 } 2543 } 2544 if (zone->uz_ctor != NULL) { 2545 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2546 zone_free_item(zone, item, udata, SKIP_DTOR); 2547 goto fail; 2548 } 2549 } 2550 #ifdef INVARIANTS 2551 uma_dbg_alloc(zone, NULL, item); 2552 #endif 2553 if (flags & M_ZERO) 2554 uma_zero_item(item, zone); 2555 2556 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item, 2557 zone->uz_name, zone); 2558 2559 return (item); 2560 2561 fail: 2562 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)", 2563 zone->uz_name, zone); 2564 atomic_add_long(&zone->uz_fails, 1); 2565 return (NULL); 2566 } 2567 2568 /* See uma.h */ 2569 void 2570 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 2571 { 2572 uma_cache_t cache; 2573 uma_bucket_t bucket; 2574 int lockfail; 2575 int cpu; 2576 2577 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2578 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA); 2579 2580 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 2581 zone->uz_name); 2582 2583 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), 2584 ("uma_zfree_arg: called with spinlock or critical section held")); 2585 2586 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 2587 if (item == NULL) 2588 return; 2589 #ifdef DEBUG_MEMGUARD 2590 if (is_memguard_addr(item)) { 2591 if (zone->uz_dtor != NULL) 2592 zone->uz_dtor(item, zone->uz_size, udata); 2593 if (zone->uz_fini != NULL) 2594 zone->uz_fini(item, zone->uz_size); 2595 memguard_free(item); 2596 return; 2597 } 2598 #endif 2599 #ifdef INVARIANTS 2600 if (zone->uz_flags & UMA_ZONE_MALLOC) 2601 uma_dbg_free(zone, udata, item); 2602 else 2603 uma_dbg_free(zone, NULL, item); 2604 #endif 2605 if (zone->uz_dtor != NULL) 2606 zone->uz_dtor(item, zone->uz_size, udata); 2607 2608 /* 2609 * The race here is acceptable. If we miss it we'll just have to wait 2610 * a little longer for the limits to be reset. 2611 */ 2612 if (zone->uz_flags & UMA_ZFLAG_FULL) 2613 goto zfree_item; 2614 2615 /* 2616 * If possible, free to the per-CPU cache. There are two 2617 * requirements for safe access to the per-CPU cache: (1) the thread 2618 * accessing the cache must not be preempted or yield during access, 2619 * and (2) the thread must not migrate CPUs without switching which 2620 * cache it accesses. We rely on a critical section to prevent 2621 * preemption and migration. We release the critical section in 2622 * order to acquire the zone mutex if we are unable to free to the 2623 * current cache; when we re-acquire the critical section, we must 2624 * detect and handle migration if it has occurred. 2625 */ 2626 zfree_restart: 2627 critical_enter(); 2628 cpu = curcpu; 2629 cache = &zone->uz_cpu[cpu]; 2630 2631 zfree_start: 2632 /* 2633 * Try to free into the allocbucket first to give LIFO ordering 2634 * for cache-hot datastructures. Spill over into the freebucket 2635 * if necessary. Alloc will swap them if one runs dry. 2636 */ 2637 bucket = cache->uc_allocbucket; 2638 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 2639 bucket = cache->uc_freebucket; 2640 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2641 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 2642 ("uma_zfree: Freeing to non free bucket index.")); 2643 bucket->ub_bucket[bucket->ub_cnt] = item; 2644 bucket->ub_cnt++; 2645 cache->uc_frees++; 2646 critical_exit(); 2647 return; 2648 } 2649 2650 /* 2651 * We must go back the zone, which requires acquiring the zone lock, 2652 * which in turn means we must release and re-acquire the critical 2653 * section. Since the critical section is released, we may be 2654 * preempted or migrate. As such, make sure not to maintain any 2655 * thread-local state specific to the cache from prior to releasing 2656 * the critical section. 2657 */ 2658 critical_exit(); 2659 if (zone->uz_count == 0 || bucketdisable) 2660 goto zfree_item; 2661 2662 lockfail = 0; 2663 if (ZONE_TRYLOCK(zone) == 0) { 2664 /* Record contention to size the buckets. */ 2665 ZONE_LOCK(zone); 2666 lockfail = 1; 2667 } 2668 critical_enter(); 2669 cpu = curcpu; 2670 cache = &zone->uz_cpu[cpu]; 2671 2672 /* 2673 * Since we have locked the zone we may as well send back our stats. 2674 */ 2675 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2676 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2677 cache->uc_allocs = 0; 2678 cache->uc_frees = 0; 2679 2680 bucket = cache->uc_freebucket; 2681 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2682 ZONE_UNLOCK(zone); 2683 goto zfree_start; 2684 } 2685 cache->uc_freebucket = NULL; 2686 /* We are no longer associated with this CPU. */ 2687 critical_exit(); 2688 2689 /* Can we throw this on the zone full list? */ 2690 if (bucket != NULL) { 2691 CTR3(KTR_UMA, 2692 "uma_zfree: zone %s(%p) putting bucket %p on free list", 2693 zone->uz_name, zone, bucket); 2694 /* ub_cnt is pointing to the last free item */ 2695 KASSERT(bucket->ub_cnt != 0, 2696 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 2697 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2698 } 2699 2700 /* 2701 * We bump the uz count when the cache size is insufficient to 2702 * handle the working set. 2703 */ 2704 if (lockfail && zone->uz_count < BUCKET_MAX) 2705 zone->uz_count++; 2706 ZONE_UNLOCK(zone); 2707 2708 bucket = bucket_alloc(zone, udata, M_NOWAIT); 2709 CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p", 2710 zone->uz_name, zone, bucket); 2711 if (bucket) { 2712 critical_enter(); 2713 cpu = curcpu; 2714 cache = &zone->uz_cpu[cpu]; 2715 if (cache->uc_freebucket == NULL) { 2716 cache->uc_freebucket = bucket; 2717 goto zfree_start; 2718 } 2719 /* 2720 * We lost the race, start over. We have to drop our 2721 * critical section to free the bucket. 2722 */ 2723 critical_exit(); 2724 bucket_free(zone, bucket, udata); 2725 goto zfree_restart; 2726 } 2727 2728 /* 2729 * If nothing else caught this, we'll just do an internal free. 2730 */ 2731 zfree_item: 2732 zone_free_item(zone, item, udata, SKIP_DTOR); 2733 2734 return; 2735 } 2736 2737 static void 2738 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) 2739 { 2740 uint8_t freei; 2741 2742 mtx_assert(&keg->uk_lock, MA_OWNED); 2743 MPASS(keg == slab->us_keg); 2744 2745 /* Do we need to remove from any lists? */ 2746 if (slab->us_freecount+1 == keg->uk_ipers) { 2747 LIST_REMOVE(slab, us_link); 2748 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 2749 } else if (slab->us_freecount == 0) { 2750 LIST_REMOVE(slab, us_link); 2751 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2752 } 2753 2754 /* Slab management. */ 2755 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 2756 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); 2757 slab->us_freecount++; 2758 2759 /* Keg statistics. */ 2760 keg->uk_free++; 2761 } 2762 2763 static void 2764 zone_release(uma_zone_t zone, void **bucket, int cnt) 2765 { 2766 void *item; 2767 uma_slab_t slab; 2768 uma_keg_t keg; 2769 uint8_t *mem; 2770 int clearfull; 2771 int i; 2772 2773 clearfull = 0; 2774 keg = zone_first_keg(zone); 2775 KEG_LOCK(keg); 2776 for (i = 0; i < cnt; i++) { 2777 item = bucket[i]; 2778 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 2779 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 2780 if (zone->uz_flags & UMA_ZONE_HASH) { 2781 slab = hash_sfind(&keg->uk_hash, mem); 2782 } else { 2783 mem += keg->uk_pgoff; 2784 slab = (uma_slab_t)mem; 2785 } 2786 } else { 2787 slab = vtoslab((vm_offset_t)item); 2788 if (slab->us_keg != keg) { 2789 KEG_UNLOCK(keg); 2790 keg = slab->us_keg; 2791 KEG_LOCK(keg); 2792 } 2793 } 2794 slab_free_item(keg, slab, item); 2795 if (keg->uk_flags & UMA_ZFLAG_FULL) { 2796 if (keg->uk_pages < keg->uk_maxpages) { 2797 keg->uk_flags &= ~UMA_ZFLAG_FULL; 2798 clearfull = 1; 2799 } 2800 2801 /* 2802 * We can handle one more allocation. Since we're 2803 * clearing ZFLAG_FULL, wake up all procs blocked 2804 * on pages. This should be uncommon, so keeping this 2805 * simple for now (rather than adding count of blocked 2806 * threads etc). 2807 */ 2808 wakeup(keg); 2809 } 2810 } 2811 KEG_UNLOCK(keg); 2812 if (clearfull) { 2813 ZONE_LOCK(zone); 2814 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2815 wakeup(zone); 2816 ZONE_UNLOCK(zone); 2817 } 2818 2819 } 2820 2821 /* 2822 * Frees a single item to any zone. 2823 * 2824 * Arguments: 2825 * zone The zone to free to 2826 * item The item we're freeing 2827 * udata User supplied data for the dtor 2828 * skip Skip dtors and finis 2829 */ 2830 static void 2831 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 2832 { 2833 2834 #ifdef INVARIANTS 2835 if (skip == SKIP_NONE) { 2836 if (zone->uz_flags & UMA_ZONE_MALLOC) 2837 uma_dbg_free(zone, udata, item); 2838 else 2839 uma_dbg_free(zone, NULL, item); 2840 } 2841 #endif 2842 if (skip < SKIP_DTOR && zone->uz_dtor) 2843 zone->uz_dtor(item, zone->uz_size, udata); 2844 2845 if (skip < SKIP_FINI && zone->uz_fini) 2846 zone->uz_fini(item, zone->uz_size); 2847 2848 atomic_add_long(&zone->uz_frees, 1); 2849 zone->uz_release(zone->uz_arg, &item, 1); 2850 } 2851 2852 /* See uma.h */ 2853 int 2854 uma_zone_set_max(uma_zone_t zone, int nitems) 2855 { 2856 uma_keg_t keg; 2857 2858 keg = zone_first_keg(zone); 2859 if (keg == NULL) 2860 return (0); 2861 KEG_LOCK(keg); 2862 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 2863 if (keg->uk_maxpages * keg->uk_ipers < nitems) 2864 keg->uk_maxpages += keg->uk_ppera; 2865 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers; 2866 KEG_UNLOCK(keg); 2867 2868 return (nitems); 2869 } 2870 2871 /* See uma.h */ 2872 int 2873 uma_zone_get_max(uma_zone_t zone) 2874 { 2875 int nitems; 2876 uma_keg_t keg; 2877 2878 keg = zone_first_keg(zone); 2879 if (keg == NULL) 2880 return (0); 2881 KEG_LOCK(keg); 2882 nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers; 2883 KEG_UNLOCK(keg); 2884 2885 return (nitems); 2886 } 2887 2888 /* See uma.h */ 2889 void 2890 uma_zone_set_warning(uma_zone_t zone, const char *warning) 2891 { 2892 2893 ZONE_LOCK(zone); 2894 zone->uz_warning = warning; 2895 ZONE_UNLOCK(zone); 2896 } 2897 2898 /* See uma.h */ 2899 void 2900 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) 2901 { 2902 2903 ZONE_LOCK(zone); 2904 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); 2905 ZONE_UNLOCK(zone); 2906 } 2907 2908 /* See uma.h */ 2909 int 2910 uma_zone_get_cur(uma_zone_t zone) 2911 { 2912 int64_t nitems; 2913 u_int i; 2914 2915 ZONE_LOCK(zone); 2916 nitems = zone->uz_allocs - zone->uz_frees; 2917 CPU_FOREACH(i) { 2918 /* 2919 * See the comment in sysctl_vm_zone_stats() regarding the 2920 * safety of accessing the per-cpu caches. With the zone lock 2921 * held, it is safe, but can potentially result in stale data. 2922 */ 2923 nitems += zone->uz_cpu[i].uc_allocs - 2924 zone->uz_cpu[i].uc_frees; 2925 } 2926 ZONE_UNLOCK(zone); 2927 2928 return (nitems < 0 ? 0 : nitems); 2929 } 2930 2931 /* See uma.h */ 2932 void 2933 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 2934 { 2935 uma_keg_t keg; 2936 2937 keg = zone_first_keg(zone); 2938 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 2939 KEG_LOCK(keg); 2940 KASSERT(keg->uk_pages == 0, 2941 ("uma_zone_set_init on non-empty keg")); 2942 keg->uk_init = uminit; 2943 KEG_UNLOCK(keg); 2944 } 2945 2946 /* See uma.h */ 2947 void 2948 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 2949 { 2950 uma_keg_t keg; 2951 2952 keg = zone_first_keg(zone); 2953 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type")); 2954 KEG_LOCK(keg); 2955 KASSERT(keg->uk_pages == 0, 2956 ("uma_zone_set_fini on non-empty keg")); 2957 keg->uk_fini = fini; 2958 KEG_UNLOCK(keg); 2959 } 2960 2961 /* See uma.h */ 2962 void 2963 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 2964 { 2965 2966 ZONE_LOCK(zone); 2967 KASSERT(zone_first_keg(zone)->uk_pages == 0, 2968 ("uma_zone_set_zinit on non-empty keg")); 2969 zone->uz_init = zinit; 2970 ZONE_UNLOCK(zone); 2971 } 2972 2973 /* See uma.h */ 2974 void 2975 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 2976 { 2977 2978 ZONE_LOCK(zone); 2979 KASSERT(zone_first_keg(zone)->uk_pages == 0, 2980 ("uma_zone_set_zfini on non-empty keg")); 2981 zone->uz_fini = zfini; 2982 ZONE_UNLOCK(zone); 2983 } 2984 2985 /* See uma.h */ 2986 /* XXX uk_freef is not actually used with the zone locked */ 2987 void 2988 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 2989 { 2990 uma_keg_t keg; 2991 2992 keg = zone_first_keg(zone); 2993 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type")); 2994 KEG_LOCK(keg); 2995 keg->uk_freef = freef; 2996 KEG_UNLOCK(keg); 2997 } 2998 2999 /* See uma.h */ 3000 /* XXX uk_allocf is not actually used with the zone locked */ 3001 void 3002 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3003 { 3004 uma_keg_t keg; 3005 3006 keg = zone_first_keg(zone); 3007 KEG_LOCK(keg); 3008 keg->uk_allocf = allocf; 3009 KEG_UNLOCK(keg); 3010 } 3011 3012 /* See uma.h */ 3013 void 3014 uma_zone_reserve(uma_zone_t zone, int items) 3015 { 3016 uma_keg_t keg; 3017 3018 keg = zone_first_keg(zone); 3019 if (keg == NULL) 3020 return; 3021 KEG_LOCK(keg); 3022 keg->uk_reserve = items; 3023 KEG_UNLOCK(keg); 3024 3025 return; 3026 } 3027 3028 /* See uma.h */ 3029 int 3030 uma_zone_reserve_kva(uma_zone_t zone, int count) 3031 { 3032 uma_keg_t keg; 3033 vm_offset_t kva; 3034 u_int pages; 3035 3036 keg = zone_first_keg(zone); 3037 if (keg == NULL) 3038 return (0); 3039 pages = count / keg->uk_ipers; 3040 3041 if (pages * keg->uk_ipers < count) 3042 pages++; 3043 pages *= keg->uk_ppera; 3044 3045 #ifdef UMA_MD_SMALL_ALLOC 3046 if (keg->uk_ppera > 1) { 3047 #else 3048 if (1) { 3049 #endif 3050 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE); 3051 if (kva == 0) 3052 return (0); 3053 } else 3054 kva = 0; 3055 KEG_LOCK(keg); 3056 keg->uk_kva = kva; 3057 keg->uk_offset = 0; 3058 keg->uk_maxpages = pages; 3059 #ifdef UMA_MD_SMALL_ALLOC 3060 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3061 #else 3062 keg->uk_allocf = noobj_alloc; 3063 #endif 3064 keg->uk_flags |= UMA_ZONE_NOFREE; 3065 KEG_UNLOCK(keg); 3066 3067 return (1); 3068 } 3069 3070 /* See uma.h */ 3071 void 3072 uma_prealloc(uma_zone_t zone, int items) 3073 { 3074 int slabs; 3075 uma_slab_t slab; 3076 uma_keg_t keg; 3077 3078 keg = zone_first_keg(zone); 3079 if (keg == NULL) 3080 return; 3081 KEG_LOCK(keg); 3082 slabs = items / keg->uk_ipers; 3083 if (slabs * keg->uk_ipers < items) 3084 slabs++; 3085 while (slabs > 0) { 3086 slab = keg_alloc_slab(keg, zone, M_WAITOK); 3087 if (slab == NULL) 3088 break; 3089 MPASS(slab->us_keg == keg); 3090 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 3091 slabs--; 3092 } 3093 KEG_UNLOCK(keg); 3094 } 3095 3096 /* See uma.h */ 3097 static void 3098 uma_reclaim_locked(bool kmem_danger) 3099 { 3100 3101 CTR0(KTR_UMA, "UMA: vm asked us to release pages!"); 3102 sx_assert(&uma_drain_lock, SA_XLOCKED); 3103 bucket_enable(); 3104 zone_foreach(zone_drain); 3105 if (vm_page_count_min() || kmem_danger) { 3106 cache_drain_safe(NULL); 3107 zone_foreach(zone_drain); 3108 } 3109 /* 3110 * Some slabs may have been freed but this zone will be visited early 3111 * we visit again so that we can free pages that are empty once other 3112 * zones are drained. We have to do the same for buckets. 3113 */ 3114 zone_drain(slabzone); 3115 bucket_zone_drain(); 3116 } 3117 3118 void 3119 uma_reclaim(void) 3120 { 3121 3122 sx_xlock(&uma_drain_lock); 3123 uma_reclaim_locked(false); 3124 sx_xunlock(&uma_drain_lock); 3125 } 3126 3127 static int uma_reclaim_needed; 3128 3129 void 3130 uma_reclaim_wakeup(void) 3131 { 3132 3133 uma_reclaim_needed = 1; 3134 wakeup(&uma_reclaim_needed); 3135 } 3136 3137 void 3138 uma_reclaim_worker(void *arg __unused) 3139 { 3140 3141 sx_xlock(&uma_drain_lock); 3142 for (;;) { 3143 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM, 3144 "umarcl", 0); 3145 if (uma_reclaim_needed) { 3146 uma_reclaim_needed = 0; 3147 sx_xunlock(&uma_drain_lock); 3148 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); 3149 sx_xlock(&uma_drain_lock); 3150 uma_reclaim_locked(true); 3151 } 3152 } 3153 } 3154 3155 /* See uma.h */ 3156 int 3157 uma_zone_exhausted(uma_zone_t zone) 3158 { 3159 int full; 3160 3161 ZONE_LOCK(zone); 3162 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3163 ZONE_UNLOCK(zone); 3164 return (full); 3165 } 3166 3167 int 3168 uma_zone_exhausted_nolock(uma_zone_t zone) 3169 { 3170 return (zone->uz_flags & UMA_ZFLAG_FULL); 3171 } 3172 3173 void * 3174 uma_large_malloc(vm_size_t size, int wait) 3175 { 3176 void *mem; 3177 uma_slab_t slab; 3178 uint8_t flags; 3179 3180 slab = zone_alloc_item(slabzone, NULL, wait); 3181 if (slab == NULL) 3182 return (NULL); 3183 mem = page_alloc(NULL, size, &flags, wait); 3184 if (mem) { 3185 vsetslab((vm_offset_t)mem, slab); 3186 slab->us_data = mem; 3187 slab->us_flags = flags | UMA_SLAB_MALLOC; 3188 slab->us_size = size; 3189 } else { 3190 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3191 } 3192 3193 return (mem); 3194 } 3195 3196 void 3197 uma_large_free(uma_slab_t slab) 3198 { 3199 3200 page_free(slab->us_data, slab->us_size, slab->us_flags); 3201 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3202 } 3203 3204 static void 3205 uma_zero_item(void *item, uma_zone_t zone) 3206 { 3207 int i; 3208 3209 if (zone->uz_flags & UMA_ZONE_PCPU) { 3210 CPU_FOREACH(i) 3211 bzero(zpcpu_get_cpu(item, i), zone->uz_size); 3212 } else 3213 bzero(item, zone->uz_size); 3214 } 3215 3216 void 3217 uma_print_stats(void) 3218 { 3219 zone_foreach(uma_print_zone); 3220 } 3221 3222 static void 3223 slab_print(uma_slab_t slab) 3224 { 3225 printf("slab: keg %p, data %p, freecount %d\n", 3226 slab->us_keg, slab->us_data, slab->us_freecount); 3227 } 3228 3229 static void 3230 cache_print(uma_cache_t cache) 3231 { 3232 printf("alloc: %p(%d), free: %p(%d)\n", 3233 cache->uc_allocbucket, 3234 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3235 cache->uc_freebucket, 3236 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3237 } 3238 3239 static void 3240 uma_print_keg(uma_keg_t keg) 3241 { 3242 uma_slab_t slab; 3243 3244 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3245 "out %d free %d limit %d\n", 3246 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3247 keg->uk_ipers, keg->uk_ppera, 3248 (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, 3249 keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3250 printf("Part slabs:\n"); 3251 LIST_FOREACH(slab, &keg->uk_part_slab, us_link) 3252 slab_print(slab); 3253 printf("Free slabs:\n"); 3254 LIST_FOREACH(slab, &keg->uk_free_slab, us_link) 3255 slab_print(slab); 3256 printf("Full slabs:\n"); 3257 LIST_FOREACH(slab, &keg->uk_full_slab, us_link) 3258 slab_print(slab); 3259 } 3260 3261 void 3262 uma_print_zone(uma_zone_t zone) 3263 { 3264 uma_cache_t cache; 3265 uma_klink_t kl; 3266 int i; 3267 3268 printf("zone: %s(%p) size %d flags %#x\n", 3269 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3270 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3271 uma_print_keg(kl->kl_keg); 3272 CPU_FOREACH(i) { 3273 cache = &zone->uz_cpu[i]; 3274 printf("CPU %d Cache:\n", i); 3275 cache_print(cache); 3276 } 3277 } 3278 3279 #ifdef DDB 3280 /* 3281 * Generate statistics across both the zone and its per-cpu cache's. Return 3282 * desired statistics if the pointer is non-NULL for that statistic. 3283 * 3284 * Note: does not update the zone statistics, as it can't safely clear the 3285 * per-CPU cache statistic. 3286 * 3287 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3288 * safe from off-CPU; we should modify the caches to track this information 3289 * directly so that we don't have to. 3290 */ 3291 static void 3292 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3293 uint64_t *freesp, uint64_t *sleepsp) 3294 { 3295 uma_cache_t cache; 3296 uint64_t allocs, frees, sleeps; 3297 int cachefree, cpu; 3298 3299 allocs = frees = sleeps = 0; 3300 cachefree = 0; 3301 CPU_FOREACH(cpu) { 3302 cache = &z->uz_cpu[cpu]; 3303 if (cache->uc_allocbucket != NULL) 3304 cachefree += cache->uc_allocbucket->ub_cnt; 3305 if (cache->uc_freebucket != NULL) 3306 cachefree += cache->uc_freebucket->ub_cnt; 3307 allocs += cache->uc_allocs; 3308 frees += cache->uc_frees; 3309 } 3310 allocs += z->uz_allocs; 3311 frees += z->uz_frees; 3312 sleeps += z->uz_sleeps; 3313 if (cachefreep != NULL) 3314 *cachefreep = cachefree; 3315 if (allocsp != NULL) 3316 *allocsp = allocs; 3317 if (freesp != NULL) 3318 *freesp = frees; 3319 if (sleepsp != NULL) 3320 *sleepsp = sleeps; 3321 } 3322 #endif /* DDB */ 3323 3324 static int 3325 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3326 { 3327 uma_keg_t kz; 3328 uma_zone_t z; 3329 int count; 3330 3331 count = 0; 3332 rw_rlock(&uma_rwlock); 3333 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3334 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3335 count++; 3336 } 3337 rw_runlock(&uma_rwlock); 3338 return (sysctl_handle_int(oidp, &count, 0, req)); 3339 } 3340 3341 static int 3342 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3343 { 3344 struct uma_stream_header ush; 3345 struct uma_type_header uth; 3346 struct uma_percpu_stat ups; 3347 uma_bucket_t bucket; 3348 struct sbuf sbuf; 3349 uma_cache_t cache; 3350 uma_klink_t kl; 3351 uma_keg_t kz; 3352 uma_zone_t z; 3353 uma_keg_t k; 3354 int count, error, i; 3355 3356 error = sysctl_wire_old_buffer(req, 0); 3357 if (error != 0) 3358 return (error); 3359 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3360 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); 3361 3362 count = 0; 3363 rw_rlock(&uma_rwlock); 3364 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3365 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3366 count++; 3367 } 3368 3369 /* 3370 * Insert stream header. 3371 */ 3372 bzero(&ush, sizeof(ush)); 3373 ush.ush_version = UMA_STREAM_VERSION; 3374 ush.ush_maxcpus = (mp_maxid + 1); 3375 ush.ush_count = count; 3376 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3377 3378 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3379 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3380 bzero(&uth, sizeof(uth)); 3381 ZONE_LOCK(z); 3382 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3383 uth.uth_align = kz->uk_align; 3384 uth.uth_size = kz->uk_size; 3385 uth.uth_rsize = kz->uk_rsize; 3386 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3387 k = kl->kl_keg; 3388 uth.uth_maxpages += k->uk_maxpages; 3389 uth.uth_pages += k->uk_pages; 3390 uth.uth_keg_free += k->uk_free; 3391 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3392 * k->uk_ipers; 3393 } 3394 3395 /* 3396 * A zone is secondary is it is not the first entry 3397 * on the keg's zone list. 3398 */ 3399 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3400 (LIST_FIRST(&kz->uk_zones) != z)) 3401 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3402 3403 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3404 uth.uth_zone_free += bucket->ub_cnt; 3405 uth.uth_allocs = z->uz_allocs; 3406 uth.uth_frees = z->uz_frees; 3407 uth.uth_fails = z->uz_fails; 3408 uth.uth_sleeps = z->uz_sleeps; 3409 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 3410 /* 3411 * While it is not normally safe to access the cache 3412 * bucket pointers while not on the CPU that owns the 3413 * cache, we only allow the pointers to be exchanged 3414 * without the zone lock held, not invalidated, so 3415 * accept the possible race associated with bucket 3416 * exchange during monitoring. 3417 */ 3418 for (i = 0; i < (mp_maxid + 1); i++) { 3419 bzero(&ups, sizeof(ups)); 3420 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) 3421 goto skip; 3422 if (CPU_ABSENT(i)) 3423 goto skip; 3424 cache = &z->uz_cpu[i]; 3425 if (cache->uc_allocbucket != NULL) 3426 ups.ups_cache_free += 3427 cache->uc_allocbucket->ub_cnt; 3428 if (cache->uc_freebucket != NULL) 3429 ups.ups_cache_free += 3430 cache->uc_freebucket->ub_cnt; 3431 ups.ups_allocs = cache->uc_allocs; 3432 ups.ups_frees = cache->uc_frees; 3433 skip: 3434 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups)); 3435 } 3436 ZONE_UNLOCK(z); 3437 } 3438 } 3439 rw_runlock(&uma_rwlock); 3440 error = sbuf_finish(&sbuf); 3441 sbuf_delete(&sbuf); 3442 return (error); 3443 } 3444 3445 int 3446 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 3447 { 3448 uma_zone_t zone = *(uma_zone_t *)arg1; 3449 int error, max; 3450 3451 max = uma_zone_get_max(zone); 3452 error = sysctl_handle_int(oidp, &max, 0, req); 3453 if (error || !req->newptr) 3454 return (error); 3455 3456 uma_zone_set_max(zone, max); 3457 3458 return (0); 3459 } 3460 3461 int 3462 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 3463 { 3464 uma_zone_t zone = *(uma_zone_t *)arg1; 3465 int cur; 3466 3467 cur = uma_zone_get_cur(zone); 3468 return (sysctl_handle_int(oidp, &cur, 0, req)); 3469 } 3470 3471 #ifdef INVARIANTS 3472 static uma_slab_t 3473 uma_dbg_getslab(uma_zone_t zone, void *item) 3474 { 3475 uma_slab_t slab; 3476 uma_keg_t keg; 3477 uint8_t *mem; 3478 3479 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 3480 if (zone->uz_flags & UMA_ZONE_VTOSLAB) { 3481 slab = vtoslab((vm_offset_t)mem); 3482 } else { 3483 /* 3484 * It is safe to return the slab here even though the 3485 * zone is unlocked because the item's allocation state 3486 * essentially holds a reference. 3487 */ 3488 ZONE_LOCK(zone); 3489 keg = LIST_FIRST(&zone->uz_kegs)->kl_keg; 3490 if (keg->uk_flags & UMA_ZONE_HASH) 3491 slab = hash_sfind(&keg->uk_hash, mem); 3492 else 3493 slab = (uma_slab_t)(mem + keg->uk_pgoff); 3494 ZONE_UNLOCK(zone); 3495 } 3496 3497 return (slab); 3498 } 3499 3500 /* 3501 * Set up the slab's freei data such that uma_dbg_free can function. 3502 * 3503 */ 3504 static void 3505 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) 3506 { 3507 uma_keg_t keg; 3508 int freei; 3509 3510 if (zone_first_keg(zone) == NULL) 3511 return; 3512 if (slab == NULL) { 3513 slab = uma_dbg_getslab(zone, item); 3514 if (slab == NULL) 3515 panic("uma: item %p did not belong to zone %s\n", 3516 item, zone->uz_name); 3517 } 3518 keg = slab->us_keg; 3519 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3520 3521 if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) 3522 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n", 3523 item, zone, zone->uz_name, slab, freei); 3524 BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); 3525 3526 return; 3527 } 3528 3529 /* 3530 * Verifies freed addresses. Checks for alignment, valid slab membership 3531 * and duplicate frees. 3532 * 3533 */ 3534 static void 3535 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) 3536 { 3537 uma_keg_t keg; 3538 int freei; 3539 3540 if (zone_first_keg(zone) == NULL) 3541 return; 3542 if (slab == NULL) { 3543 slab = uma_dbg_getslab(zone, item); 3544 if (slab == NULL) 3545 panic("uma: Freed item %p did not belong to zone %s\n", 3546 item, zone->uz_name); 3547 } 3548 keg = slab->us_keg; 3549 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3550 3551 if (freei >= keg->uk_ipers) 3552 panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n", 3553 item, zone, zone->uz_name, slab, freei); 3554 3555 if (((freei * keg->uk_rsize) + slab->us_data) != item) 3556 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n", 3557 item, zone, zone->uz_name, slab, freei); 3558 3559 if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) 3560 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n", 3561 item, zone, zone->uz_name, slab, freei); 3562 3563 BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); 3564 } 3565 #endif /* INVARIANTS */ 3566 3567 #ifdef DDB 3568 DB_SHOW_COMMAND(uma, db_show_uma) 3569 { 3570 uint64_t allocs, frees, sleeps; 3571 uma_bucket_t bucket; 3572 uma_keg_t kz; 3573 uma_zone_t z; 3574 int cachefree; 3575 3576 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used", 3577 "Free", "Requests", "Sleeps", "Bucket"); 3578 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3579 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3580 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 3581 allocs = z->uz_allocs; 3582 frees = z->uz_frees; 3583 sleeps = z->uz_sleeps; 3584 cachefree = 0; 3585 } else 3586 uma_zone_sumstat(z, &cachefree, &allocs, 3587 &frees, &sleeps); 3588 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 3589 (LIST_FIRST(&kz->uk_zones) != z))) 3590 cachefree += kz->uk_free; 3591 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3592 cachefree += bucket->ub_cnt; 3593 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n", 3594 z->uz_name, (uintmax_t)kz->uk_size, 3595 (intmax_t)(allocs - frees), cachefree, 3596 (uintmax_t)allocs, sleeps, z->uz_count); 3597 if (db_pager_quit) 3598 return; 3599 } 3600 } 3601 } 3602 3603 DB_SHOW_COMMAND(umacache, db_show_umacache) 3604 { 3605 uint64_t allocs, frees; 3606 uma_bucket_t bucket; 3607 uma_zone_t z; 3608 int cachefree; 3609 3610 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 3611 "Requests", "Bucket"); 3612 LIST_FOREACH(z, &uma_cachezones, uz_link) { 3613 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL); 3614 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3615 cachefree += bucket->ub_cnt; 3616 db_printf("%18s %8ju %8jd %8d %12ju %8u\n", 3617 z->uz_name, (uintmax_t)z->uz_size, 3618 (intmax_t)(allocs - frees), cachefree, 3619 (uintmax_t)allocs, z->uz_count); 3620 if (db_pager_quit) 3621 return; 3622 } 3623 } 3624 #endif /* DDB */ 3625