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 * effecient. 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 /* I should really use ktr.. */ 54 /* 55 #define UMA_DEBUG 1 56 #define UMA_DEBUG_ALLOC 1 57 #define UMA_DEBUG_ALLOC_1 1 58 */ 59 60 #include "opt_ddb.h" 61 #include "opt_param.h" 62 #include "opt_vm.h" 63 64 #include <sys/param.h> 65 #include <sys/systm.h> 66 #include <sys/bitset.h> 67 #include <sys/kernel.h> 68 #include <sys/types.h> 69 #include <sys/queue.h> 70 #include <sys/malloc.h> 71 #include <sys/ktr.h> 72 #include <sys/lock.h> 73 #include <sys/sysctl.h> 74 #include <sys/mutex.h> 75 #include <sys/proc.h> 76 #include <sys/random.h> 77 #include <sys/rwlock.h> 78 #include <sys/sbuf.h> 79 #include <sys/sched.h> 80 #include <sys/smp.h> 81 #include <sys/vmmeter.h> 82 83 #include <vm/vm.h> 84 #include <vm/vm_object.h> 85 #include <vm/vm_page.h> 86 #include <vm/vm_pageout.h> 87 #include <vm/vm_param.h> 88 #include <vm/vm_map.h> 89 #include <vm/vm_kern.h> 90 #include <vm/vm_extern.h> 91 #include <vm/uma.h> 92 #include <vm/uma_int.h> 93 #include <vm/uma_dbg.h> 94 95 #include <ddb/ddb.h> 96 97 #ifdef DEBUG_MEMGUARD 98 #include <vm/memguard.h> 99 #endif 100 101 /* 102 * This is the zone and keg from which all zones are spawned. The idea is that 103 * even the zone & keg heads are allocated from the allocator, so we use the 104 * bss section to bootstrap us. 105 */ 106 static struct uma_keg masterkeg; 107 static struct uma_zone masterzone_k; 108 static struct uma_zone masterzone_z; 109 static uma_zone_t kegs = &masterzone_k; 110 static uma_zone_t zones = &masterzone_z; 111 112 /* This is the zone from which all of uma_slab_t's are allocated. */ 113 static uma_zone_t slabzone; 114 static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */ 115 116 /* 117 * The initial hash tables come out of this zone so they can be allocated 118 * prior to malloc coming up. 119 */ 120 static uma_zone_t hashzone; 121 122 /* The boot-time adjusted value for cache line alignment. */ 123 int uma_align_cache = 64 - 1; 124 125 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 126 127 /* 128 * Are we allowed to allocate buckets? 129 */ 130 static int bucketdisable = 1; 131 132 /* Linked list of all kegs in the system */ 133 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 134 135 /* Linked list of all cache-only zones in the system */ 136 static LIST_HEAD(,uma_zone) uma_cachezones = 137 LIST_HEAD_INITIALIZER(uma_cachezones); 138 139 /* This RW lock protects the keg list */ 140 static struct rwlock_padalign uma_rwlock; 141 142 /* Linked list of boot time pages */ 143 static LIST_HEAD(,uma_slab) uma_boot_pages = 144 LIST_HEAD_INITIALIZER(uma_boot_pages); 145 146 /* This mutex protects the boot time pages list */ 147 static struct mtx_padalign uma_boot_pages_mtx; 148 149 static struct sx uma_drain_lock; 150 151 /* Is the VM done starting up? */ 152 static int booted = 0; 153 #define UMA_STARTUP 1 154 #define UMA_STARTUP2 2 155 156 /* 157 * Only mbuf clusters use ref zones. Just provide enough references 158 * to support the one user. New code should not use the ref facility. 159 */ 160 static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES; 161 162 /* 163 * This is the handle used to schedule events that need to happen 164 * outside of the allocation fast path. 165 */ 166 static struct callout uma_callout; 167 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 168 169 /* 170 * This structure is passed as the zone ctor arg so that I don't have to create 171 * a special allocation function just for zones. 172 */ 173 struct uma_zctor_args { 174 const char *name; 175 size_t size; 176 uma_ctor ctor; 177 uma_dtor dtor; 178 uma_init uminit; 179 uma_fini fini; 180 uma_import import; 181 uma_release release; 182 void *arg; 183 uma_keg_t keg; 184 int align; 185 uint32_t flags; 186 }; 187 188 struct uma_kctor_args { 189 uma_zone_t zone; 190 size_t size; 191 uma_init uminit; 192 uma_fini fini; 193 int align; 194 uint32_t flags; 195 }; 196 197 struct uma_bucket_zone { 198 uma_zone_t ubz_zone; 199 char *ubz_name; 200 int ubz_entries; /* Number of items it can hold. */ 201 int ubz_maxsize; /* Maximum allocation size per-item. */ 202 }; 203 204 /* 205 * Compute the actual number of bucket entries to pack them in power 206 * of two sizes for more efficient space utilization. 207 */ 208 #define BUCKET_SIZE(n) \ 209 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 210 211 #define BUCKET_MAX BUCKET_SIZE(256) 212 213 struct uma_bucket_zone bucket_zones[] = { 214 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 215 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 216 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 217 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 218 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 219 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 220 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 221 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 222 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, 223 { NULL, NULL, 0} 224 }; 225 226 /* 227 * Flags and enumerations to be passed to internal functions. 228 */ 229 enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; 230 231 /* Prototypes.. */ 232 233 static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 234 static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 235 static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int); 236 static void page_free(void *, vm_size_t, uint8_t); 237 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); 238 static void cache_drain(uma_zone_t); 239 static void bucket_drain(uma_zone_t, uma_bucket_t); 240 static void bucket_cache_drain(uma_zone_t zone); 241 static int keg_ctor(void *, int, void *, int); 242 static void keg_dtor(void *, int, void *); 243 static int zone_ctor(void *, int, void *, int); 244 static void zone_dtor(void *, int, void *); 245 static int zero_init(void *, int, int); 246 static void keg_small_init(uma_keg_t keg); 247 static void keg_large_init(uma_keg_t keg); 248 static void zone_foreach(void (*zfunc)(uma_zone_t)); 249 static void zone_timeout(uma_zone_t zone); 250 static int hash_alloc(struct uma_hash *); 251 static int hash_expand(struct uma_hash *, struct uma_hash *); 252 static void hash_free(struct uma_hash *hash); 253 static void uma_timeout(void *); 254 static void uma_startup3(void); 255 static void *zone_alloc_item(uma_zone_t, void *, int); 256 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 257 static void bucket_enable(void); 258 static void bucket_init(void); 259 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 260 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 261 static void bucket_zone_drain(void); 262 static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags); 263 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); 264 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); 265 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 266 static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); 267 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 268 uma_fini fini, int align, uint32_t flags); 269 static int zone_import(uma_zone_t zone, void **bucket, int max, int flags); 270 static void zone_release(uma_zone_t zone, void **bucket, int cnt); 271 static void uma_zero_item(void *item, uma_zone_t zone); 272 273 void uma_print_zone(uma_zone_t); 274 void uma_print_stats(void); 275 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 276 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 277 278 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 279 280 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 281 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 282 283 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 284 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 285 286 static int zone_warnings = 1; 287 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, 288 "Warn when UMA zones becomes full"); 289 290 /* 291 * This routine checks to see whether or not it's safe to enable buckets. 292 */ 293 static void 294 bucket_enable(void) 295 { 296 bucketdisable = vm_page_count_min(); 297 } 298 299 /* 300 * Initialize bucket_zones, the array of zones of buckets of various sizes. 301 * 302 * For each zone, calculate the memory required for each bucket, consisting 303 * of the header and an array of pointers. 304 */ 305 static void 306 bucket_init(void) 307 { 308 struct uma_bucket_zone *ubz; 309 int size; 310 311 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 312 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 313 size += sizeof(void *) * ubz->ubz_entries; 314 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 315 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 316 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET); 317 } 318 } 319 320 /* 321 * Given a desired number of entries for a bucket, return the zone from which 322 * to allocate the bucket. 323 */ 324 static struct uma_bucket_zone * 325 bucket_zone_lookup(int entries) 326 { 327 struct uma_bucket_zone *ubz; 328 329 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 330 if (ubz->ubz_entries >= entries) 331 return (ubz); 332 ubz--; 333 return (ubz); 334 } 335 336 static int 337 bucket_select(int size) 338 { 339 struct uma_bucket_zone *ubz; 340 341 ubz = &bucket_zones[0]; 342 if (size > ubz->ubz_maxsize) 343 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 344 345 for (; ubz->ubz_entries != 0; ubz++) 346 if (ubz->ubz_maxsize < size) 347 break; 348 ubz--; 349 return (ubz->ubz_entries); 350 } 351 352 static uma_bucket_t 353 bucket_alloc(uma_zone_t zone, void *udata, int flags) 354 { 355 struct uma_bucket_zone *ubz; 356 uma_bucket_t bucket; 357 358 /* 359 * This is to stop us from allocating per cpu buckets while we're 360 * running out of vm.boot_pages. Otherwise, we would exhaust the 361 * boot pages. This also prevents us from allocating buckets in 362 * low memory situations. 363 */ 364 if (bucketdisable) 365 return (NULL); 366 /* 367 * To limit bucket recursion we store the original zone flags 368 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 369 * NOVM flag to persist even through deep recursions. We also 370 * store ZFLAG_BUCKET once we have recursed attempting to allocate 371 * a bucket for a bucket zone so we do not allow infinite bucket 372 * recursion. This cookie will even persist to frees of unused 373 * buckets via the allocation path or bucket allocations in the 374 * free path. 375 */ 376 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 377 udata = (void *)(uintptr_t)zone->uz_flags; 378 else { 379 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 380 return (NULL); 381 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 382 } 383 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 384 flags |= M_NOVM; 385 ubz = bucket_zone_lookup(zone->uz_count); 386 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) 387 ubz++; 388 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 389 if (bucket) { 390 #ifdef INVARIANTS 391 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 392 #endif 393 bucket->ub_cnt = 0; 394 bucket->ub_entries = ubz->ubz_entries; 395 } 396 397 return (bucket); 398 } 399 400 static void 401 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 402 { 403 struct uma_bucket_zone *ubz; 404 405 KASSERT(bucket->ub_cnt == 0, 406 ("bucket_free: Freeing a non free bucket.")); 407 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 408 udata = (void *)(uintptr_t)zone->uz_flags; 409 ubz = bucket_zone_lookup(bucket->ub_entries); 410 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 411 } 412 413 static void 414 bucket_zone_drain(void) 415 { 416 struct uma_bucket_zone *ubz; 417 418 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 419 zone_drain(ubz->ubz_zone); 420 } 421 422 static void 423 zone_log_warning(uma_zone_t zone) 424 { 425 static const struct timeval warninterval = { 300, 0 }; 426 427 if (!zone_warnings || zone->uz_warning == NULL) 428 return; 429 430 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 431 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 432 } 433 434 static void 435 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 436 { 437 uma_klink_t klink; 438 439 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 440 kegfn(klink->kl_keg); 441 } 442 443 /* 444 * Routine called by timeout which is used to fire off some time interval 445 * based calculations. (stats, hash size, etc.) 446 * 447 * Arguments: 448 * arg Unused 449 * 450 * Returns: 451 * Nothing 452 */ 453 static void 454 uma_timeout(void *unused) 455 { 456 bucket_enable(); 457 zone_foreach(zone_timeout); 458 459 /* Reschedule this event */ 460 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 461 } 462 463 /* 464 * Routine to perform timeout driven calculations. This expands the 465 * hashes and does per cpu statistics aggregation. 466 * 467 * Returns nothing. 468 */ 469 static void 470 keg_timeout(uma_keg_t keg) 471 { 472 473 KEG_LOCK(keg); 474 /* 475 * Expand the keg hash table. 476 * 477 * This is done if the number of slabs is larger than the hash size. 478 * What I'm trying to do here is completely reduce collisions. This 479 * may be a little aggressive. Should I allow for two collisions max? 480 */ 481 if (keg->uk_flags & UMA_ZONE_HASH && 482 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 483 struct uma_hash newhash; 484 struct uma_hash oldhash; 485 int ret; 486 487 /* 488 * This is so involved because allocating and freeing 489 * while the keg lock is held will lead to deadlock. 490 * I have to do everything in stages and check for 491 * races. 492 */ 493 newhash = keg->uk_hash; 494 KEG_UNLOCK(keg); 495 ret = hash_alloc(&newhash); 496 KEG_LOCK(keg); 497 if (ret) { 498 if (hash_expand(&keg->uk_hash, &newhash)) { 499 oldhash = keg->uk_hash; 500 keg->uk_hash = newhash; 501 } else 502 oldhash = newhash; 503 504 KEG_UNLOCK(keg); 505 hash_free(&oldhash); 506 return; 507 } 508 } 509 KEG_UNLOCK(keg); 510 } 511 512 static void 513 zone_timeout(uma_zone_t zone) 514 { 515 516 zone_foreach_keg(zone, &keg_timeout); 517 } 518 519 /* 520 * Allocate and zero fill the next sized hash table from the appropriate 521 * backing store. 522 * 523 * Arguments: 524 * hash A new hash structure with the old hash size in uh_hashsize 525 * 526 * Returns: 527 * 1 on sucess and 0 on failure. 528 */ 529 static int 530 hash_alloc(struct uma_hash *hash) 531 { 532 int oldsize; 533 int alloc; 534 535 oldsize = hash->uh_hashsize; 536 537 /* We're just going to go to a power of two greater */ 538 if (oldsize) { 539 hash->uh_hashsize = oldsize * 2; 540 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 541 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 542 M_UMAHASH, M_NOWAIT); 543 } else { 544 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 545 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 546 M_WAITOK); 547 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 548 } 549 if (hash->uh_slab_hash) { 550 bzero(hash->uh_slab_hash, alloc); 551 hash->uh_hashmask = hash->uh_hashsize - 1; 552 return (1); 553 } 554 555 return (0); 556 } 557 558 /* 559 * Expands the hash table for HASH zones. This is done from zone_timeout 560 * to reduce collisions. This must not be done in the regular allocation 561 * path, otherwise, we can recurse on the vm while allocating pages. 562 * 563 * Arguments: 564 * oldhash The hash you want to expand 565 * newhash The hash structure for the new table 566 * 567 * Returns: 568 * Nothing 569 * 570 * Discussion: 571 */ 572 static int 573 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 574 { 575 uma_slab_t slab; 576 int hval; 577 int i; 578 579 if (!newhash->uh_slab_hash) 580 return (0); 581 582 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 583 return (0); 584 585 /* 586 * I need to investigate hash algorithms for resizing without a 587 * full rehash. 588 */ 589 590 for (i = 0; i < oldhash->uh_hashsize; i++) 591 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 592 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 593 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 594 hval = UMA_HASH(newhash, slab->us_data); 595 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 596 slab, us_hlink); 597 } 598 599 return (1); 600 } 601 602 /* 603 * Free the hash bucket to the appropriate backing store. 604 * 605 * Arguments: 606 * slab_hash The hash bucket we're freeing 607 * hashsize The number of entries in that hash bucket 608 * 609 * Returns: 610 * Nothing 611 */ 612 static void 613 hash_free(struct uma_hash *hash) 614 { 615 if (hash->uh_slab_hash == NULL) 616 return; 617 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 618 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 619 else 620 free(hash->uh_slab_hash, M_UMAHASH); 621 } 622 623 /* 624 * Frees all outstanding items in a bucket 625 * 626 * Arguments: 627 * zone The zone to free to, must be unlocked. 628 * bucket The free/alloc bucket with items, cpu queue must be locked. 629 * 630 * Returns: 631 * Nothing 632 */ 633 634 static void 635 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 636 { 637 int i; 638 639 if (bucket == NULL) 640 return; 641 642 if (zone->uz_fini) 643 for (i = 0; i < bucket->ub_cnt; i++) 644 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 645 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 646 bucket->ub_cnt = 0; 647 } 648 649 /* 650 * Drains the per cpu caches for a zone. 651 * 652 * NOTE: This may only be called while the zone is being turn down, and not 653 * during normal operation. This is necessary in order that we do not have 654 * to migrate CPUs to drain the per-CPU caches. 655 * 656 * Arguments: 657 * zone The zone to drain, must be unlocked. 658 * 659 * Returns: 660 * Nothing 661 */ 662 static void 663 cache_drain(uma_zone_t zone) 664 { 665 uma_cache_t cache; 666 int cpu; 667 668 /* 669 * XXX: It is safe to not lock the per-CPU caches, because we're 670 * tearing down the zone anyway. I.e., there will be no further use 671 * of the caches at this point. 672 * 673 * XXX: It would good to be able to assert that the zone is being 674 * torn down to prevent improper use of cache_drain(). 675 * 676 * XXX: We lock the zone before passing into bucket_cache_drain() as 677 * it is used elsewhere. Should the tear-down path be made special 678 * there in some form? 679 */ 680 CPU_FOREACH(cpu) { 681 cache = &zone->uz_cpu[cpu]; 682 bucket_drain(zone, cache->uc_allocbucket); 683 bucket_drain(zone, cache->uc_freebucket); 684 if (cache->uc_allocbucket != NULL) 685 bucket_free(zone, cache->uc_allocbucket, NULL); 686 if (cache->uc_freebucket != NULL) 687 bucket_free(zone, cache->uc_freebucket, NULL); 688 cache->uc_allocbucket = cache->uc_freebucket = NULL; 689 } 690 ZONE_LOCK(zone); 691 bucket_cache_drain(zone); 692 ZONE_UNLOCK(zone); 693 } 694 695 static void 696 cache_shrink(uma_zone_t zone) 697 { 698 699 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 700 return; 701 702 ZONE_LOCK(zone); 703 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; 704 ZONE_UNLOCK(zone); 705 } 706 707 static void 708 cache_drain_safe_cpu(uma_zone_t zone) 709 { 710 uma_cache_t cache; 711 uma_bucket_t b1, b2; 712 713 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 714 return; 715 716 b1 = b2 = NULL; 717 ZONE_LOCK(zone); 718 critical_enter(); 719 cache = &zone->uz_cpu[curcpu]; 720 if (cache->uc_allocbucket) { 721 if (cache->uc_allocbucket->ub_cnt != 0) 722 LIST_INSERT_HEAD(&zone->uz_buckets, 723 cache->uc_allocbucket, ub_link); 724 else 725 b1 = cache->uc_allocbucket; 726 cache->uc_allocbucket = NULL; 727 } 728 if (cache->uc_freebucket) { 729 if (cache->uc_freebucket->ub_cnt != 0) 730 LIST_INSERT_HEAD(&zone->uz_buckets, 731 cache->uc_freebucket, ub_link); 732 else 733 b2 = cache->uc_freebucket; 734 cache->uc_freebucket = NULL; 735 } 736 critical_exit(); 737 ZONE_UNLOCK(zone); 738 if (b1) 739 bucket_free(zone, b1, NULL); 740 if (b2) 741 bucket_free(zone, b2, NULL); 742 } 743 744 /* 745 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 746 * This is an expensive call because it needs to bind to all CPUs 747 * one by one and enter a critical section on each of them in order 748 * to safely access their cache buckets. 749 * Zone lock must not be held on call this function. 750 */ 751 static void 752 cache_drain_safe(uma_zone_t zone) 753 { 754 int cpu; 755 756 /* 757 * Polite bucket sizes shrinking was not enouth, shrink aggressively. 758 */ 759 if (zone) 760 cache_shrink(zone); 761 else 762 zone_foreach(cache_shrink); 763 764 CPU_FOREACH(cpu) { 765 thread_lock(curthread); 766 sched_bind(curthread, cpu); 767 thread_unlock(curthread); 768 769 if (zone) 770 cache_drain_safe_cpu(zone); 771 else 772 zone_foreach(cache_drain_safe_cpu); 773 } 774 thread_lock(curthread); 775 sched_unbind(curthread); 776 thread_unlock(curthread); 777 } 778 779 /* 780 * Drain the cached buckets from a zone. Expects a locked zone on entry. 781 */ 782 static void 783 bucket_cache_drain(uma_zone_t zone) 784 { 785 uma_bucket_t bucket; 786 787 /* 788 * Drain the bucket queues and free the buckets, we just keep two per 789 * cpu (alloc/free). 790 */ 791 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 792 LIST_REMOVE(bucket, ub_link); 793 ZONE_UNLOCK(zone); 794 bucket_drain(zone, bucket); 795 bucket_free(zone, bucket, NULL); 796 ZONE_LOCK(zone); 797 } 798 799 /* 800 * Shrink further bucket sizes. Price of single zone lock collision 801 * is probably lower then price of global cache drain. 802 */ 803 if (zone->uz_count > zone->uz_count_min) 804 zone->uz_count--; 805 } 806 807 static void 808 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 809 { 810 uint8_t *mem; 811 int i; 812 uint8_t flags; 813 814 mem = slab->us_data; 815 flags = slab->us_flags; 816 i = start; 817 if (keg->uk_fini != NULL) { 818 for (i--; i > -1; i--) 819 keg->uk_fini(slab->us_data + (keg->uk_rsize * i), 820 keg->uk_size); 821 } 822 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 823 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 824 #ifdef UMA_DEBUG 825 printf("%s: Returning %d bytes.\n", keg->uk_name, 826 PAGE_SIZE * keg->uk_ppera); 827 #endif 828 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 829 } 830 831 /* 832 * Frees pages from a keg back to the system. This is done on demand from 833 * the pageout daemon. 834 * 835 * Returns nothing. 836 */ 837 static void 838 keg_drain(uma_keg_t keg) 839 { 840 struct slabhead freeslabs = { 0 }; 841 uma_slab_t slab; 842 uma_slab_t n; 843 844 /* 845 * We don't want to take pages from statically allocated kegs at this 846 * time 847 */ 848 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 849 return; 850 851 #ifdef UMA_DEBUG 852 printf("%s free items: %u\n", keg->uk_name, keg->uk_free); 853 #endif 854 KEG_LOCK(keg); 855 if (keg->uk_free == 0) 856 goto finished; 857 858 slab = LIST_FIRST(&keg->uk_free_slab); 859 while (slab) { 860 n = LIST_NEXT(slab, us_link); 861 862 /* We have no where to free these to */ 863 if (slab->us_flags & UMA_SLAB_BOOT) { 864 slab = n; 865 continue; 866 } 867 868 LIST_REMOVE(slab, us_link); 869 keg->uk_pages -= keg->uk_ppera; 870 keg->uk_free -= keg->uk_ipers; 871 872 if (keg->uk_flags & UMA_ZONE_HASH) 873 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); 874 875 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 876 877 slab = n; 878 } 879 finished: 880 KEG_UNLOCK(keg); 881 882 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 883 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 884 keg_free_slab(keg, slab, keg->uk_ipers); 885 } 886 } 887 888 static void 889 zone_drain_wait(uma_zone_t zone, int waitok) 890 { 891 892 /* 893 * Set draining to interlock with zone_dtor() so we can release our 894 * locks as we go. Only dtor() should do a WAITOK call since it 895 * is the only call that knows the structure will still be available 896 * when it wakes up. 897 */ 898 ZONE_LOCK(zone); 899 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 900 if (waitok == M_NOWAIT) 901 goto out; 902 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 903 } 904 zone->uz_flags |= UMA_ZFLAG_DRAINING; 905 bucket_cache_drain(zone); 906 ZONE_UNLOCK(zone); 907 /* 908 * The DRAINING flag protects us from being freed while 909 * we're running. Normally the uma_rwlock would protect us but we 910 * must be able to release and acquire the right lock for each keg. 911 */ 912 zone_foreach_keg(zone, &keg_drain); 913 ZONE_LOCK(zone); 914 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 915 wakeup(zone); 916 out: 917 ZONE_UNLOCK(zone); 918 } 919 920 void 921 zone_drain(uma_zone_t zone) 922 { 923 924 zone_drain_wait(zone, M_NOWAIT); 925 } 926 927 /* 928 * Allocate a new slab for a keg. This does not insert the slab onto a list. 929 * 930 * Arguments: 931 * wait Shall we wait? 932 * 933 * Returns: 934 * The slab that was allocated or NULL if there is no memory and the 935 * caller specified M_NOWAIT. 936 */ 937 static uma_slab_t 938 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) 939 { 940 uma_slabrefcnt_t slabref; 941 uma_alloc allocf; 942 uma_slab_t slab; 943 uint8_t *mem; 944 uint8_t flags; 945 int i; 946 947 mtx_assert(&keg->uk_lock, MA_OWNED); 948 slab = NULL; 949 mem = NULL; 950 951 #ifdef UMA_DEBUG 952 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name); 953 #endif 954 allocf = keg->uk_allocf; 955 KEG_UNLOCK(keg); 956 957 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 958 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); 959 if (slab == NULL) 960 goto out; 961 } 962 963 /* 964 * This reproduces the old vm_zone behavior of zero filling pages the 965 * first time they are added to a zone. 966 * 967 * Malloced items are zeroed in uma_zalloc. 968 */ 969 970 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 971 wait |= M_ZERO; 972 else 973 wait &= ~M_ZERO; 974 975 if (keg->uk_flags & UMA_ZONE_NODUMP) 976 wait |= M_NODUMP; 977 978 /* zone is passed for legacy reasons. */ 979 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait); 980 if (mem == NULL) { 981 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 982 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 983 slab = NULL; 984 goto out; 985 } 986 987 /* Point the slab into the allocated memory */ 988 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 989 slab = (uma_slab_t )(mem + keg->uk_pgoff); 990 991 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 992 for (i = 0; i < keg->uk_ppera; i++) 993 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 994 995 slab->us_keg = keg; 996 slab->us_data = mem; 997 slab->us_freecount = keg->uk_ipers; 998 slab->us_flags = flags; 999 BIT_FILL(SLAB_SETSIZE, &slab->us_free); 1000 #ifdef INVARIANTS 1001 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); 1002 #endif 1003 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1004 slabref = (uma_slabrefcnt_t)slab; 1005 for (i = 0; i < keg->uk_ipers; i++) 1006 slabref->us_refcnt[i] = 0; 1007 } 1008 1009 if (keg->uk_init != NULL) { 1010 for (i = 0; i < keg->uk_ipers; i++) 1011 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 1012 keg->uk_size, wait) != 0) 1013 break; 1014 if (i != keg->uk_ipers) { 1015 keg_free_slab(keg, slab, i); 1016 slab = NULL; 1017 goto out; 1018 } 1019 } 1020 out: 1021 KEG_LOCK(keg); 1022 1023 if (slab != NULL) { 1024 if (keg->uk_flags & UMA_ZONE_HASH) 1025 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1026 1027 keg->uk_pages += keg->uk_ppera; 1028 keg->uk_free += keg->uk_ipers; 1029 } 1030 1031 return (slab); 1032 } 1033 1034 /* 1035 * This function is intended to be used early on in place of page_alloc() so 1036 * that we may use the boot time page cache to satisfy allocations before 1037 * the VM is ready. 1038 */ 1039 static void * 1040 startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) 1041 { 1042 uma_keg_t keg; 1043 uma_slab_t tmps; 1044 int pages, check_pages; 1045 1046 keg = zone_first_keg(zone); 1047 pages = howmany(bytes, PAGE_SIZE); 1048 check_pages = pages - 1; 1049 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n")); 1050 1051 /* 1052 * Check our small startup cache to see if it has pages remaining. 1053 */ 1054 mtx_lock(&uma_boot_pages_mtx); 1055 1056 /* First check if we have enough room. */ 1057 tmps = LIST_FIRST(&uma_boot_pages); 1058 while (tmps != NULL && check_pages-- > 0) 1059 tmps = LIST_NEXT(tmps, us_link); 1060 if (tmps != NULL) { 1061 /* 1062 * It's ok to lose tmps references. The last one will 1063 * have tmps->us_data pointing to the start address of 1064 * "pages" contiguous pages of memory. 1065 */ 1066 while (pages-- > 0) { 1067 tmps = LIST_FIRST(&uma_boot_pages); 1068 LIST_REMOVE(tmps, us_link); 1069 } 1070 mtx_unlock(&uma_boot_pages_mtx); 1071 *pflag = tmps->us_flags; 1072 return (tmps->us_data); 1073 } 1074 mtx_unlock(&uma_boot_pages_mtx); 1075 if (booted < UMA_STARTUP2) 1076 panic("UMA: Increase vm.boot_pages"); 1077 /* 1078 * Now that we've booted reset these users to their real allocator. 1079 */ 1080 #ifdef UMA_MD_SMALL_ALLOC 1081 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; 1082 #else 1083 keg->uk_allocf = page_alloc; 1084 #endif 1085 return keg->uk_allocf(zone, bytes, pflag, wait); 1086 } 1087 1088 /* 1089 * Allocates a number of pages from the system 1090 * 1091 * Arguments: 1092 * bytes The number of bytes requested 1093 * wait Shall we wait? 1094 * 1095 * Returns: 1096 * A pointer to the alloced memory or possibly 1097 * NULL if M_NOWAIT is set. 1098 */ 1099 static void * 1100 page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) 1101 { 1102 void *p; /* Returned page */ 1103 1104 *pflag = UMA_SLAB_KMEM; 1105 p = (void *) kmem_malloc(kmem_arena, bytes, wait); 1106 1107 return (p); 1108 } 1109 1110 /* 1111 * Allocates a number of pages from within an object 1112 * 1113 * Arguments: 1114 * bytes The number of bytes requested 1115 * wait Shall we wait? 1116 * 1117 * Returns: 1118 * A pointer to the alloced memory or possibly 1119 * NULL if M_NOWAIT is set. 1120 */ 1121 static void * 1122 noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait) 1123 { 1124 TAILQ_HEAD(, vm_page) alloctail; 1125 u_long npages; 1126 vm_offset_t retkva, zkva; 1127 vm_page_t p, p_next; 1128 uma_keg_t keg; 1129 1130 TAILQ_INIT(&alloctail); 1131 keg = zone_first_keg(zone); 1132 1133 npages = howmany(bytes, PAGE_SIZE); 1134 while (npages > 0) { 1135 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT | 1136 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ); 1137 if (p != NULL) { 1138 /* 1139 * Since the page does not belong to an object, its 1140 * listq is unused. 1141 */ 1142 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1143 npages--; 1144 continue; 1145 } 1146 if (wait & M_WAITOK) { 1147 VM_WAIT; 1148 continue; 1149 } 1150 1151 /* 1152 * Page allocation failed, free intermediate pages and 1153 * exit. 1154 */ 1155 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1156 vm_page_unwire(p, PQ_NONE); 1157 vm_page_free(p); 1158 } 1159 return (NULL); 1160 } 1161 *flags = UMA_SLAB_PRIV; 1162 zkva = keg->uk_kva + 1163 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1164 retkva = zkva; 1165 TAILQ_FOREACH(p, &alloctail, listq) { 1166 pmap_qenter(zkva, &p, 1); 1167 zkva += PAGE_SIZE; 1168 } 1169 1170 return ((void *)retkva); 1171 } 1172 1173 /* 1174 * Frees a number of pages to the system 1175 * 1176 * Arguments: 1177 * mem A pointer to the memory to be freed 1178 * size The size of the memory being freed 1179 * flags The original p->us_flags field 1180 * 1181 * Returns: 1182 * Nothing 1183 */ 1184 static void 1185 page_free(void *mem, vm_size_t size, uint8_t flags) 1186 { 1187 struct vmem *vmem; 1188 1189 if (flags & UMA_SLAB_KMEM) 1190 vmem = kmem_arena; 1191 else if (flags & UMA_SLAB_KERNEL) 1192 vmem = kernel_arena; 1193 else 1194 panic("UMA: page_free used with invalid flags %d", flags); 1195 1196 kmem_free(vmem, (vm_offset_t)mem, size); 1197 } 1198 1199 /* 1200 * Zero fill initializer 1201 * 1202 * Arguments/Returns follow uma_init specifications 1203 */ 1204 static int 1205 zero_init(void *mem, int size, int flags) 1206 { 1207 bzero(mem, size); 1208 return (0); 1209 } 1210 1211 /* 1212 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1213 * 1214 * Arguments 1215 * keg The zone we should initialize 1216 * 1217 * Returns 1218 * Nothing 1219 */ 1220 static void 1221 keg_small_init(uma_keg_t keg) 1222 { 1223 u_int rsize; 1224 u_int memused; 1225 u_int wastedspace; 1226 u_int shsize; 1227 1228 if (keg->uk_flags & UMA_ZONE_PCPU) { 1229 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU; 1230 1231 keg->uk_slabsize = sizeof(struct pcpu); 1232 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu), 1233 PAGE_SIZE); 1234 } else { 1235 keg->uk_slabsize = UMA_SLAB_SIZE; 1236 keg->uk_ppera = 1; 1237 } 1238 1239 /* 1240 * Calculate the size of each allocation (rsize) according to 1241 * alignment. If the requested size is smaller than we have 1242 * allocation bits for we round it up. 1243 */ 1244 rsize = keg->uk_size; 1245 if (rsize < keg->uk_slabsize / SLAB_SETSIZE) 1246 rsize = keg->uk_slabsize / SLAB_SETSIZE; 1247 if (rsize & keg->uk_align) 1248 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1249 keg->uk_rsize = rsize; 1250 1251 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1252 keg->uk_rsize < sizeof(struct pcpu), 1253 ("%s: size %u too large", __func__, keg->uk_rsize)); 1254 1255 if (keg->uk_flags & UMA_ZONE_REFCNT) 1256 rsize += sizeof(uint32_t); 1257 1258 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1259 shsize = 0; 1260 else 1261 shsize = sizeof(struct uma_slab); 1262 1263 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize; 1264 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1265 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1266 1267 memused = keg->uk_ipers * rsize + shsize; 1268 wastedspace = keg->uk_slabsize - memused; 1269 1270 /* 1271 * We can't do OFFPAGE if we're internal or if we've been 1272 * asked to not go to the VM for buckets. If we do this we 1273 * may end up going to the VM for slabs which we do not 1274 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1275 * of UMA_ZONE_VM, which clearly forbids it. 1276 */ 1277 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1278 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1279 return; 1280 1281 /* 1282 * See if using an OFFPAGE slab will limit our waste. Only do 1283 * this if it permits more items per-slab. 1284 * 1285 * XXX We could try growing slabsize to limit max waste as well. 1286 * Historically this was not done because the VM could not 1287 * efficiently handle contiguous allocations. 1288 */ 1289 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) && 1290 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) { 1291 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize; 1292 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1293 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1294 #ifdef UMA_DEBUG 1295 printf("UMA decided we need offpage slab headers for " 1296 "keg: %s, calculated wastedspace = %d, " 1297 "maximum wasted space allowed = %d, " 1298 "calculated ipers = %d, " 1299 "new wasted space = %d\n", keg->uk_name, wastedspace, 1300 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1301 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize); 1302 #endif 1303 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1304 } 1305 1306 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1307 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1308 keg->uk_flags |= UMA_ZONE_HASH; 1309 } 1310 1311 /* 1312 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1313 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1314 * more complicated. 1315 * 1316 * Arguments 1317 * keg The keg we should initialize 1318 * 1319 * Returns 1320 * Nothing 1321 */ 1322 static void 1323 keg_large_init(uma_keg_t keg) 1324 { 1325 u_int shsize; 1326 1327 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1328 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1329 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1330 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1331 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1332 1333 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1334 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE; 1335 keg->uk_ipers = 1; 1336 keg->uk_rsize = keg->uk_size; 1337 1338 /* We can't do OFFPAGE if we're internal, bail out here. */ 1339 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) 1340 return; 1341 1342 /* Check whether we have enough space to not do OFFPAGE. */ 1343 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) { 1344 shsize = sizeof(struct uma_slab); 1345 if (keg->uk_flags & UMA_ZONE_REFCNT) 1346 shsize += keg->uk_ipers * sizeof(uint32_t); 1347 if (shsize & UMA_ALIGN_PTR) 1348 shsize = (shsize & ~UMA_ALIGN_PTR) + 1349 (UMA_ALIGN_PTR + 1); 1350 1351 if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize) 1352 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1353 } 1354 1355 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1356 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1357 keg->uk_flags |= UMA_ZONE_HASH; 1358 } 1359 1360 static void 1361 keg_cachespread_init(uma_keg_t keg) 1362 { 1363 int alignsize; 1364 int trailer; 1365 int pages; 1366 int rsize; 1367 1368 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1369 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1370 1371 alignsize = keg->uk_align + 1; 1372 rsize = keg->uk_size; 1373 /* 1374 * We want one item to start on every align boundary in a page. To 1375 * do this we will span pages. We will also extend the item by the 1376 * size of align if it is an even multiple of align. Otherwise, it 1377 * would fall on the same boundary every time. 1378 */ 1379 if (rsize & keg->uk_align) 1380 rsize = (rsize & ~keg->uk_align) + alignsize; 1381 if ((rsize & alignsize) == 0) 1382 rsize += alignsize; 1383 trailer = rsize - keg->uk_size; 1384 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1385 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1386 keg->uk_rsize = rsize; 1387 keg->uk_ppera = pages; 1388 keg->uk_slabsize = UMA_SLAB_SIZE; 1389 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1390 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1391 KASSERT(keg->uk_ipers <= SLAB_SETSIZE, 1392 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1393 keg->uk_ipers)); 1394 } 1395 1396 /* 1397 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1398 * the keg onto the global keg list. 1399 * 1400 * Arguments/Returns follow uma_ctor specifications 1401 * udata Actually uma_kctor_args 1402 */ 1403 static int 1404 keg_ctor(void *mem, int size, void *udata, int flags) 1405 { 1406 struct uma_kctor_args *arg = udata; 1407 uma_keg_t keg = mem; 1408 uma_zone_t zone; 1409 1410 bzero(keg, size); 1411 keg->uk_size = arg->size; 1412 keg->uk_init = arg->uminit; 1413 keg->uk_fini = arg->fini; 1414 keg->uk_align = arg->align; 1415 keg->uk_free = 0; 1416 keg->uk_reserve = 0; 1417 keg->uk_pages = 0; 1418 keg->uk_flags = arg->flags; 1419 keg->uk_allocf = page_alloc; 1420 keg->uk_freef = page_free; 1421 keg->uk_slabzone = NULL; 1422 1423 /* 1424 * The master zone is passed to us at keg-creation time. 1425 */ 1426 zone = arg->zone; 1427 keg->uk_name = zone->uz_name; 1428 1429 if (arg->flags & UMA_ZONE_VM) 1430 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1431 1432 if (arg->flags & UMA_ZONE_ZINIT) 1433 keg->uk_init = zero_init; 1434 1435 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC) 1436 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1437 1438 if (arg->flags & UMA_ZONE_PCPU) 1439 #ifdef SMP 1440 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1441 #else 1442 keg->uk_flags &= ~UMA_ZONE_PCPU; 1443 #endif 1444 1445 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1446 keg_cachespread_init(keg); 1447 } else if (keg->uk_flags & UMA_ZONE_REFCNT) { 1448 if (keg->uk_size > 1449 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - 1450 sizeof(uint32_t))) 1451 keg_large_init(keg); 1452 else 1453 keg_small_init(keg); 1454 } else { 1455 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab))) 1456 keg_large_init(keg); 1457 else 1458 keg_small_init(keg); 1459 } 1460 1461 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1462 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1463 if (keg->uk_ipers > uma_max_ipers_ref) 1464 panic("Too many ref items per zone: %d > %d\n", 1465 keg->uk_ipers, uma_max_ipers_ref); 1466 keg->uk_slabzone = slabrefzone; 1467 } else 1468 keg->uk_slabzone = slabzone; 1469 } 1470 1471 /* 1472 * If we haven't booted yet we need allocations to go through the 1473 * startup cache until the vm is ready. 1474 */ 1475 if (keg->uk_ppera == 1) { 1476 #ifdef UMA_MD_SMALL_ALLOC 1477 keg->uk_allocf = uma_small_alloc; 1478 keg->uk_freef = uma_small_free; 1479 1480 if (booted < UMA_STARTUP) 1481 keg->uk_allocf = startup_alloc; 1482 #else 1483 if (booted < UMA_STARTUP2) 1484 keg->uk_allocf = startup_alloc; 1485 #endif 1486 } else if (booted < UMA_STARTUP2 && 1487 (keg->uk_flags & UMA_ZFLAG_INTERNAL)) 1488 keg->uk_allocf = startup_alloc; 1489 1490 /* 1491 * Initialize keg's lock 1492 */ 1493 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1494 1495 /* 1496 * If we're putting the slab header in the actual page we need to 1497 * figure out where in each page it goes. This calculates a right 1498 * justified offset into the memory on an ALIGN_PTR boundary. 1499 */ 1500 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1501 u_int totsize; 1502 1503 /* Size of the slab struct and free list */ 1504 totsize = sizeof(struct uma_slab); 1505 1506 /* Size of the reference counts. */ 1507 if (keg->uk_flags & UMA_ZONE_REFCNT) 1508 totsize += keg->uk_ipers * sizeof(uint32_t); 1509 1510 if (totsize & UMA_ALIGN_PTR) 1511 totsize = (totsize & ~UMA_ALIGN_PTR) + 1512 (UMA_ALIGN_PTR + 1); 1513 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1514 1515 /* 1516 * The only way the following is possible is if with our 1517 * UMA_ALIGN_PTR adjustments we are now bigger than 1518 * UMA_SLAB_SIZE. I haven't checked whether this is 1519 * mathematically possible for all cases, so we make 1520 * sure here anyway. 1521 */ 1522 totsize = keg->uk_pgoff + sizeof(struct uma_slab); 1523 if (keg->uk_flags & UMA_ZONE_REFCNT) 1524 totsize += keg->uk_ipers * sizeof(uint32_t); 1525 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1526 printf("zone %s ipers %d rsize %d size %d\n", 1527 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1528 keg->uk_size); 1529 panic("UMA slab won't fit."); 1530 } 1531 } 1532 1533 if (keg->uk_flags & UMA_ZONE_HASH) 1534 hash_alloc(&keg->uk_hash); 1535 1536 #ifdef UMA_DEBUG 1537 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n", 1538 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags, 1539 keg->uk_ipers, keg->uk_ppera, 1540 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free); 1541 #endif 1542 1543 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1544 1545 rw_wlock(&uma_rwlock); 1546 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1547 rw_wunlock(&uma_rwlock); 1548 return (0); 1549 } 1550 1551 /* 1552 * Zone header ctor. This initializes all fields, locks, etc. 1553 * 1554 * Arguments/Returns follow uma_ctor specifications 1555 * udata Actually uma_zctor_args 1556 */ 1557 static int 1558 zone_ctor(void *mem, int size, void *udata, int flags) 1559 { 1560 struct uma_zctor_args *arg = udata; 1561 uma_zone_t zone = mem; 1562 uma_zone_t z; 1563 uma_keg_t keg; 1564 1565 bzero(zone, size); 1566 zone->uz_name = arg->name; 1567 zone->uz_ctor = arg->ctor; 1568 zone->uz_dtor = arg->dtor; 1569 zone->uz_slab = zone_fetch_slab; 1570 zone->uz_init = NULL; 1571 zone->uz_fini = NULL; 1572 zone->uz_allocs = 0; 1573 zone->uz_frees = 0; 1574 zone->uz_fails = 0; 1575 zone->uz_sleeps = 0; 1576 zone->uz_count = 0; 1577 zone->uz_count_min = 0; 1578 zone->uz_flags = 0; 1579 zone->uz_warning = NULL; 1580 timevalclear(&zone->uz_ratecheck); 1581 keg = arg->keg; 1582 1583 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 1584 1585 /* 1586 * This is a pure cache zone, no kegs. 1587 */ 1588 if (arg->import) { 1589 if (arg->flags & UMA_ZONE_VM) 1590 arg->flags |= UMA_ZFLAG_CACHEONLY; 1591 zone->uz_flags = arg->flags; 1592 zone->uz_size = arg->size; 1593 zone->uz_import = arg->import; 1594 zone->uz_release = arg->release; 1595 zone->uz_arg = arg->arg; 1596 zone->uz_lockptr = &zone->uz_lock; 1597 rw_wlock(&uma_rwlock); 1598 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); 1599 rw_wunlock(&uma_rwlock); 1600 goto out; 1601 } 1602 1603 /* 1604 * Use the regular zone/keg/slab allocator. 1605 */ 1606 zone->uz_import = (uma_import)zone_import; 1607 zone->uz_release = (uma_release)zone_release; 1608 zone->uz_arg = zone; 1609 1610 if (arg->flags & UMA_ZONE_SECONDARY) { 1611 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1612 zone->uz_init = arg->uminit; 1613 zone->uz_fini = arg->fini; 1614 zone->uz_lockptr = &keg->uk_lock; 1615 zone->uz_flags |= UMA_ZONE_SECONDARY; 1616 rw_wlock(&uma_rwlock); 1617 ZONE_LOCK(zone); 1618 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1619 if (LIST_NEXT(z, uz_link) == NULL) { 1620 LIST_INSERT_AFTER(z, zone, uz_link); 1621 break; 1622 } 1623 } 1624 ZONE_UNLOCK(zone); 1625 rw_wunlock(&uma_rwlock); 1626 } else if (keg == NULL) { 1627 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1628 arg->align, arg->flags)) == NULL) 1629 return (ENOMEM); 1630 } else { 1631 struct uma_kctor_args karg; 1632 int error; 1633 1634 /* We should only be here from uma_startup() */ 1635 karg.size = arg->size; 1636 karg.uminit = arg->uminit; 1637 karg.fini = arg->fini; 1638 karg.align = arg->align; 1639 karg.flags = arg->flags; 1640 karg.zone = zone; 1641 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1642 flags); 1643 if (error) 1644 return (error); 1645 } 1646 1647 /* 1648 * Link in the first keg. 1649 */ 1650 zone->uz_klink.kl_keg = keg; 1651 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1652 zone->uz_lockptr = &keg->uk_lock; 1653 zone->uz_size = keg->uk_size; 1654 zone->uz_flags |= (keg->uk_flags & 1655 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1656 1657 /* 1658 * Some internal zones don't have room allocated for the per cpu 1659 * caches. If we're internal, bail out here. 1660 */ 1661 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1662 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1663 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1664 return (0); 1665 } 1666 1667 out: 1668 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0) 1669 zone->uz_count = bucket_select(zone->uz_size); 1670 else 1671 zone->uz_count = BUCKET_MAX; 1672 zone->uz_count_min = zone->uz_count; 1673 1674 return (0); 1675 } 1676 1677 /* 1678 * Keg header dtor. This frees all data, destroys locks, frees the hash 1679 * table and removes the keg from the global list. 1680 * 1681 * Arguments/Returns follow uma_dtor specifications 1682 * udata unused 1683 */ 1684 static void 1685 keg_dtor(void *arg, int size, void *udata) 1686 { 1687 uma_keg_t keg; 1688 1689 keg = (uma_keg_t)arg; 1690 KEG_LOCK(keg); 1691 if (keg->uk_free != 0) { 1692 printf("Freed UMA keg (%s) was not empty (%d items). " 1693 " Lost %d pages of memory.\n", 1694 keg->uk_name ? keg->uk_name : "", 1695 keg->uk_free, keg->uk_pages); 1696 } 1697 KEG_UNLOCK(keg); 1698 1699 hash_free(&keg->uk_hash); 1700 1701 KEG_LOCK_FINI(keg); 1702 } 1703 1704 /* 1705 * Zone header dtor. 1706 * 1707 * Arguments/Returns follow uma_dtor specifications 1708 * udata unused 1709 */ 1710 static void 1711 zone_dtor(void *arg, int size, void *udata) 1712 { 1713 uma_klink_t klink; 1714 uma_zone_t zone; 1715 uma_keg_t keg; 1716 1717 zone = (uma_zone_t)arg; 1718 keg = zone_first_keg(zone); 1719 1720 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1721 cache_drain(zone); 1722 1723 rw_wlock(&uma_rwlock); 1724 LIST_REMOVE(zone, uz_link); 1725 rw_wunlock(&uma_rwlock); 1726 /* 1727 * XXX there are some races here where 1728 * the zone can be drained but zone lock 1729 * released and then refilled before we 1730 * remove it... we dont care for now 1731 */ 1732 zone_drain_wait(zone, M_WAITOK); 1733 /* 1734 * Unlink all of our kegs. 1735 */ 1736 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1737 klink->kl_keg = NULL; 1738 LIST_REMOVE(klink, kl_link); 1739 if (klink == &zone->uz_klink) 1740 continue; 1741 free(klink, M_TEMP); 1742 } 1743 /* 1744 * We only destroy kegs from non secondary zones. 1745 */ 1746 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1747 rw_wlock(&uma_rwlock); 1748 LIST_REMOVE(keg, uk_link); 1749 rw_wunlock(&uma_rwlock); 1750 zone_free_item(kegs, keg, NULL, SKIP_NONE); 1751 } 1752 ZONE_LOCK_FINI(zone); 1753 } 1754 1755 /* 1756 * Traverses every zone in the system and calls a callback 1757 * 1758 * Arguments: 1759 * zfunc A pointer to a function which accepts a zone 1760 * as an argument. 1761 * 1762 * Returns: 1763 * Nothing 1764 */ 1765 static void 1766 zone_foreach(void (*zfunc)(uma_zone_t)) 1767 { 1768 uma_keg_t keg; 1769 uma_zone_t zone; 1770 1771 rw_rlock(&uma_rwlock); 1772 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1773 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1774 zfunc(zone); 1775 } 1776 rw_runlock(&uma_rwlock); 1777 } 1778 1779 /* Public functions */ 1780 /* See uma.h */ 1781 void 1782 uma_startup(void *bootmem, int boot_pages) 1783 { 1784 struct uma_zctor_args args; 1785 uma_slab_t slab; 1786 u_int slabsize; 1787 int i; 1788 1789 #ifdef UMA_DEBUG 1790 printf("Creating uma keg headers zone and keg.\n"); 1791 #endif 1792 rw_init(&uma_rwlock, "UMA lock"); 1793 1794 /* "manually" create the initial zone */ 1795 memset(&args, 0, sizeof(args)); 1796 args.name = "UMA Kegs"; 1797 args.size = sizeof(struct uma_keg); 1798 args.ctor = keg_ctor; 1799 args.dtor = keg_dtor; 1800 args.uminit = zero_init; 1801 args.fini = NULL; 1802 args.keg = &masterkeg; 1803 args.align = 32 - 1; 1804 args.flags = UMA_ZFLAG_INTERNAL; 1805 /* The initial zone has no Per cpu queues so it's smaller */ 1806 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); 1807 1808 #ifdef UMA_DEBUG 1809 printf("Filling boot free list.\n"); 1810 #endif 1811 for (i = 0; i < boot_pages; i++) { 1812 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE)); 1813 slab->us_data = (uint8_t *)slab; 1814 slab->us_flags = UMA_SLAB_BOOT; 1815 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); 1816 } 1817 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); 1818 1819 #ifdef UMA_DEBUG 1820 printf("Creating uma zone headers zone and keg.\n"); 1821 #endif 1822 args.name = "UMA Zones"; 1823 args.size = sizeof(struct uma_zone) + 1824 (sizeof(struct uma_cache) * (mp_maxid + 1)); 1825 args.ctor = zone_ctor; 1826 args.dtor = zone_dtor; 1827 args.uminit = zero_init; 1828 args.fini = NULL; 1829 args.keg = NULL; 1830 args.align = 32 - 1; 1831 args.flags = UMA_ZFLAG_INTERNAL; 1832 /* The initial zone has no Per cpu queues so it's smaller */ 1833 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); 1834 1835 #ifdef UMA_DEBUG 1836 printf("Creating slab and hash zones.\n"); 1837 #endif 1838 1839 /* Now make a zone for slab headers */ 1840 slabzone = uma_zcreate("UMA Slabs", 1841 sizeof(struct uma_slab), 1842 NULL, NULL, NULL, NULL, 1843 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1844 1845 /* 1846 * We also create a zone for the bigger slabs with reference 1847 * counts in them, to accomodate UMA_ZONE_REFCNT zones. 1848 */ 1849 slabsize = sizeof(struct uma_slab_refcnt); 1850 slabsize += uma_max_ipers_ref * sizeof(uint32_t); 1851 slabrefzone = uma_zcreate("UMA RCntSlabs", 1852 slabsize, 1853 NULL, NULL, NULL, NULL, 1854 UMA_ALIGN_PTR, 1855 UMA_ZFLAG_INTERNAL); 1856 1857 hashzone = uma_zcreate("UMA Hash", 1858 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 1859 NULL, NULL, NULL, NULL, 1860 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1861 1862 bucket_init(); 1863 1864 booted = UMA_STARTUP; 1865 1866 #ifdef UMA_DEBUG 1867 printf("UMA startup complete.\n"); 1868 #endif 1869 } 1870 1871 /* see uma.h */ 1872 void 1873 uma_startup2(void) 1874 { 1875 booted = UMA_STARTUP2; 1876 bucket_enable(); 1877 sx_init(&uma_drain_lock, "umadrain"); 1878 #ifdef UMA_DEBUG 1879 printf("UMA startup2 complete.\n"); 1880 #endif 1881 } 1882 1883 /* 1884 * Initialize our callout handle 1885 * 1886 */ 1887 1888 static void 1889 uma_startup3(void) 1890 { 1891 #ifdef UMA_DEBUG 1892 printf("Starting callout.\n"); 1893 #endif 1894 callout_init(&uma_callout, 1); 1895 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1896 #ifdef UMA_DEBUG 1897 printf("UMA startup3 complete.\n"); 1898 #endif 1899 } 1900 1901 static uma_keg_t 1902 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 1903 int align, uint32_t flags) 1904 { 1905 struct uma_kctor_args args; 1906 1907 args.size = size; 1908 args.uminit = uminit; 1909 args.fini = fini; 1910 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 1911 args.flags = flags; 1912 args.zone = zone; 1913 return (zone_alloc_item(kegs, &args, M_WAITOK)); 1914 } 1915 1916 /* See uma.h */ 1917 void 1918 uma_set_align(int align) 1919 { 1920 1921 if (align != UMA_ALIGN_CACHE) 1922 uma_align_cache = align; 1923 } 1924 1925 /* See uma.h */ 1926 uma_zone_t 1927 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 1928 uma_init uminit, uma_fini fini, int align, uint32_t flags) 1929 1930 { 1931 struct uma_zctor_args args; 1932 uma_zone_t res; 1933 bool locked; 1934 1935 /* This stuff is essential for the zone ctor */ 1936 memset(&args, 0, sizeof(args)); 1937 args.name = name; 1938 args.size = size; 1939 args.ctor = ctor; 1940 args.dtor = dtor; 1941 args.uminit = uminit; 1942 args.fini = fini; 1943 #ifdef INVARIANTS 1944 /* 1945 * If a zone is being created with an empty constructor and 1946 * destructor, pass UMA constructor/destructor which checks for 1947 * memory use after free. 1948 */ 1949 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) && 1950 ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) { 1951 args.ctor = trash_ctor; 1952 args.dtor = trash_dtor; 1953 args.uminit = trash_init; 1954 args.fini = trash_fini; 1955 } 1956 #endif 1957 args.align = align; 1958 args.flags = flags; 1959 args.keg = NULL; 1960 1961 if (booted < UMA_STARTUP2) { 1962 locked = false; 1963 } else { 1964 sx_slock(&uma_drain_lock); 1965 locked = true; 1966 } 1967 res = zone_alloc_item(zones, &args, M_WAITOK); 1968 if (locked) 1969 sx_sunlock(&uma_drain_lock); 1970 return (res); 1971 } 1972 1973 /* See uma.h */ 1974 uma_zone_t 1975 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 1976 uma_init zinit, uma_fini zfini, uma_zone_t master) 1977 { 1978 struct uma_zctor_args args; 1979 uma_keg_t keg; 1980 uma_zone_t res; 1981 bool locked; 1982 1983 keg = zone_first_keg(master); 1984 memset(&args, 0, sizeof(args)); 1985 args.name = name; 1986 args.size = keg->uk_size; 1987 args.ctor = ctor; 1988 args.dtor = dtor; 1989 args.uminit = zinit; 1990 args.fini = zfini; 1991 args.align = keg->uk_align; 1992 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 1993 args.keg = keg; 1994 1995 if (booted < UMA_STARTUP2) { 1996 locked = false; 1997 } else { 1998 sx_slock(&uma_drain_lock); 1999 locked = true; 2000 } 2001 /* XXX Attaches only one keg of potentially many. */ 2002 res = zone_alloc_item(zones, &args, M_WAITOK); 2003 if (locked) 2004 sx_sunlock(&uma_drain_lock); 2005 return (res); 2006 } 2007 2008 /* See uma.h */ 2009 uma_zone_t 2010 uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 2011 uma_init zinit, uma_fini zfini, uma_import zimport, 2012 uma_release zrelease, void *arg, int flags) 2013 { 2014 struct uma_zctor_args args; 2015 2016 memset(&args, 0, sizeof(args)); 2017 args.name = name; 2018 args.size = size; 2019 args.ctor = ctor; 2020 args.dtor = dtor; 2021 args.uminit = zinit; 2022 args.fini = zfini; 2023 args.import = zimport; 2024 args.release = zrelease; 2025 args.arg = arg; 2026 args.align = 0; 2027 args.flags = flags; 2028 2029 return (zone_alloc_item(zones, &args, M_WAITOK)); 2030 } 2031 2032 static void 2033 zone_lock_pair(uma_zone_t a, uma_zone_t b) 2034 { 2035 if (a < b) { 2036 ZONE_LOCK(a); 2037 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); 2038 } else { 2039 ZONE_LOCK(b); 2040 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); 2041 } 2042 } 2043 2044 static void 2045 zone_unlock_pair(uma_zone_t a, uma_zone_t b) 2046 { 2047 2048 ZONE_UNLOCK(a); 2049 ZONE_UNLOCK(b); 2050 } 2051 2052 int 2053 uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 2054 { 2055 uma_klink_t klink; 2056 uma_klink_t kl; 2057 int error; 2058 2059 error = 0; 2060 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 2061 2062 zone_lock_pair(zone, master); 2063 /* 2064 * zone must use vtoslab() to resolve objects and must already be 2065 * a secondary. 2066 */ 2067 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 2068 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 2069 error = EINVAL; 2070 goto out; 2071 } 2072 /* 2073 * The new master must also use vtoslab(). 2074 */ 2075 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 2076 error = EINVAL; 2077 goto out; 2078 } 2079 /* 2080 * Both must either be refcnt, or not be refcnt. 2081 */ 2082 if ((zone->uz_flags & UMA_ZONE_REFCNT) != 2083 (master->uz_flags & UMA_ZONE_REFCNT)) { 2084 error = EINVAL; 2085 goto out; 2086 } 2087 /* 2088 * The underlying object must be the same size. rsize 2089 * may be different. 2090 */ 2091 if (master->uz_size != zone->uz_size) { 2092 error = E2BIG; 2093 goto out; 2094 } 2095 /* 2096 * Put it at the end of the list. 2097 */ 2098 klink->kl_keg = zone_first_keg(master); 2099 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 2100 if (LIST_NEXT(kl, kl_link) == NULL) { 2101 LIST_INSERT_AFTER(kl, klink, kl_link); 2102 break; 2103 } 2104 } 2105 klink = NULL; 2106 zone->uz_flags |= UMA_ZFLAG_MULTI; 2107 zone->uz_slab = zone_fetch_slab_multi; 2108 2109 out: 2110 zone_unlock_pair(zone, master); 2111 if (klink != NULL) 2112 free(klink, M_TEMP); 2113 2114 return (error); 2115 } 2116 2117 2118 /* See uma.h */ 2119 void 2120 uma_zdestroy(uma_zone_t zone) 2121 { 2122 2123 sx_slock(&uma_drain_lock); 2124 zone_free_item(zones, zone, NULL, SKIP_NONE); 2125 sx_sunlock(&uma_drain_lock); 2126 } 2127 2128 /* See uma.h */ 2129 void * 2130 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2131 { 2132 void *item; 2133 uma_cache_t cache; 2134 uma_bucket_t bucket; 2135 int lockfail; 2136 int cpu; 2137 2138 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2139 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA); 2140 2141 /* This is the fast path allocation */ 2142 #ifdef UMA_DEBUG_ALLOC_1 2143 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); 2144 #endif 2145 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread, 2146 zone->uz_name, flags); 2147 2148 if (flags & M_WAITOK) { 2149 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2150 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2151 } 2152 #ifdef DEBUG_MEMGUARD 2153 if (memguard_cmp_zone(zone)) { 2154 item = memguard_alloc(zone->uz_size, flags); 2155 if (item != NULL) { 2156 /* 2157 * Avoid conflict with the use-after-free 2158 * protecting infrastructure from INVARIANTS. 2159 */ 2160 if (zone->uz_init != NULL && 2161 zone->uz_init != mtrash_init && 2162 zone->uz_init(item, zone->uz_size, flags) != 0) 2163 return (NULL); 2164 if (zone->uz_ctor != NULL && 2165 zone->uz_ctor != mtrash_ctor && 2166 zone->uz_ctor(item, zone->uz_size, udata, 2167 flags) != 0) { 2168 zone->uz_fini(item, zone->uz_size); 2169 return (NULL); 2170 } 2171 return (item); 2172 } 2173 /* This is unfortunate but should not be fatal. */ 2174 } 2175 #endif 2176 /* 2177 * If possible, allocate from the per-CPU cache. There are two 2178 * requirements for safe access to the per-CPU cache: (1) the thread 2179 * accessing the cache must not be preempted or yield during access, 2180 * and (2) the thread must not migrate CPUs without switching which 2181 * cache it accesses. We rely on a critical section to prevent 2182 * preemption and migration. We release the critical section in 2183 * order to acquire the zone mutex if we are unable to allocate from 2184 * the current cache; when we re-acquire the critical section, we 2185 * must detect and handle migration if it has occurred. 2186 */ 2187 critical_enter(); 2188 cpu = curcpu; 2189 cache = &zone->uz_cpu[cpu]; 2190 2191 zalloc_start: 2192 bucket = cache->uc_allocbucket; 2193 if (bucket != NULL && bucket->ub_cnt > 0) { 2194 bucket->ub_cnt--; 2195 item = bucket->ub_bucket[bucket->ub_cnt]; 2196 #ifdef INVARIANTS 2197 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2198 #endif 2199 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2200 cache->uc_allocs++; 2201 critical_exit(); 2202 if (zone->uz_ctor != NULL && 2203 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2204 atomic_add_long(&zone->uz_fails, 1); 2205 zone_free_item(zone, item, udata, SKIP_DTOR); 2206 return (NULL); 2207 } 2208 #ifdef INVARIANTS 2209 uma_dbg_alloc(zone, NULL, item); 2210 #endif 2211 if (flags & M_ZERO) 2212 uma_zero_item(item, zone); 2213 return (item); 2214 } 2215 2216 /* 2217 * We have run out of items in our alloc bucket. 2218 * See if we can switch with our free bucket. 2219 */ 2220 bucket = cache->uc_freebucket; 2221 if (bucket != NULL && bucket->ub_cnt > 0) { 2222 #ifdef UMA_DEBUG_ALLOC 2223 printf("uma_zalloc: Swapping empty with alloc.\n"); 2224 #endif 2225 cache->uc_freebucket = cache->uc_allocbucket; 2226 cache->uc_allocbucket = bucket; 2227 goto zalloc_start; 2228 } 2229 2230 /* 2231 * Discard any empty allocation bucket while we hold no locks. 2232 */ 2233 bucket = cache->uc_allocbucket; 2234 cache->uc_allocbucket = NULL; 2235 critical_exit(); 2236 if (bucket != NULL) 2237 bucket_free(zone, bucket, udata); 2238 2239 /* Short-circuit for zones without buckets and low memory. */ 2240 if (zone->uz_count == 0 || bucketdisable) 2241 goto zalloc_item; 2242 2243 /* 2244 * Attempt to retrieve the item from the per-CPU cache has failed, so 2245 * we must go back to the zone. This requires the zone lock, so we 2246 * must drop the critical section, then re-acquire it when we go back 2247 * to the cache. Since the critical section is released, we may be 2248 * preempted or migrate. As such, make sure not to maintain any 2249 * thread-local state specific to the cache from prior to releasing 2250 * the critical section. 2251 */ 2252 lockfail = 0; 2253 if (ZONE_TRYLOCK(zone) == 0) { 2254 /* Record contention to size the buckets. */ 2255 ZONE_LOCK(zone); 2256 lockfail = 1; 2257 } 2258 critical_enter(); 2259 cpu = curcpu; 2260 cache = &zone->uz_cpu[cpu]; 2261 2262 /* 2263 * Since we have locked the zone we may as well send back our stats. 2264 */ 2265 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2266 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2267 cache->uc_allocs = 0; 2268 cache->uc_frees = 0; 2269 2270 /* See if we lost the race to fill the cache. */ 2271 if (cache->uc_allocbucket != NULL) { 2272 ZONE_UNLOCK(zone); 2273 goto zalloc_start; 2274 } 2275 2276 /* 2277 * Check the zone's cache of buckets. 2278 */ 2279 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 2280 KASSERT(bucket->ub_cnt != 0, 2281 ("uma_zalloc_arg: Returning an empty bucket.")); 2282 2283 LIST_REMOVE(bucket, ub_link); 2284 cache->uc_allocbucket = bucket; 2285 ZONE_UNLOCK(zone); 2286 goto zalloc_start; 2287 } 2288 /* We are no longer associated with this CPU. */ 2289 critical_exit(); 2290 2291 /* 2292 * We bump the uz count when the cache size is insufficient to 2293 * handle the working set. 2294 */ 2295 if (lockfail && zone->uz_count < BUCKET_MAX) 2296 zone->uz_count++; 2297 ZONE_UNLOCK(zone); 2298 2299 /* 2300 * Now lets just fill a bucket and put it on the free list. If that 2301 * works we'll restart the allocation from the begining and it 2302 * will use the just filled bucket. 2303 */ 2304 bucket = zone_alloc_bucket(zone, udata, flags); 2305 if (bucket != NULL) { 2306 ZONE_LOCK(zone); 2307 critical_enter(); 2308 cpu = curcpu; 2309 cache = &zone->uz_cpu[cpu]; 2310 /* 2311 * See if we lost the race or were migrated. Cache the 2312 * initialized bucket to make this less likely or claim 2313 * the memory directly. 2314 */ 2315 if (cache->uc_allocbucket == NULL) 2316 cache->uc_allocbucket = bucket; 2317 else 2318 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2319 ZONE_UNLOCK(zone); 2320 goto zalloc_start; 2321 } 2322 2323 /* 2324 * We may not be able to get a bucket so return an actual item. 2325 */ 2326 #ifdef UMA_DEBUG 2327 printf("uma_zalloc_arg: Bucketzone returned NULL\n"); 2328 #endif 2329 2330 zalloc_item: 2331 item = zone_alloc_item(zone, udata, flags); 2332 2333 return (item); 2334 } 2335 2336 static uma_slab_t 2337 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) 2338 { 2339 uma_slab_t slab; 2340 int reserve; 2341 2342 mtx_assert(&keg->uk_lock, MA_OWNED); 2343 slab = NULL; 2344 reserve = 0; 2345 if ((flags & M_USE_RESERVE) == 0) 2346 reserve = keg->uk_reserve; 2347 2348 for (;;) { 2349 /* 2350 * Find a slab with some space. Prefer slabs that are partially 2351 * used over those that are totally full. This helps to reduce 2352 * fragmentation. 2353 */ 2354 if (keg->uk_free > reserve) { 2355 if (!LIST_EMPTY(&keg->uk_part_slab)) { 2356 slab = LIST_FIRST(&keg->uk_part_slab); 2357 } else { 2358 slab = LIST_FIRST(&keg->uk_free_slab); 2359 LIST_REMOVE(slab, us_link); 2360 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, 2361 us_link); 2362 } 2363 MPASS(slab->us_keg == keg); 2364 return (slab); 2365 } 2366 2367 /* 2368 * M_NOVM means don't ask at all! 2369 */ 2370 if (flags & M_NOVM) 2371 break; 2372 2373 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2374 keg->uk_flags |= UMA_ZFLAG_FULL; 2375 /* 2376 * If this is not a multi-zone, set the FULL bit. 2377 * Otherwise slab_multi() takes care of it. 2378 */ 2379 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2380 zone->uz_flags |= UMA_ZFLAG_FULL; 2381 zone_log_warning(zone); 2382 } 2383 if (flags & M_NOWAIT) 2384 break; 2385 zone->uz_sleeps++; 2386 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2387 continue; 2388 } 2389 slab = keg_alloc_slab(keg, zone, flags); 2390 /* 2391 * If we got a slab here it's safe to mark it partially used 2392 * and return. We assume that the caller is going to remove 2393 * at least one item. 2394 */ 2395 if (slab) { 2396 MPASS(slab->us_keg == keg); 2397 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2398 return (slab); 2399 } 2400 /* 2401 * We might not have been able to get a slab but another cpu 2402 * could have while we were unlocked. Check again before we 2403 * fail. 2404 */ 2405 flags |= M_NOVM; 2406 } 2407 return (slab); 2408 } 2409 2410 static uma_slab_t 2411 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) 2412 { 2413 uma_slab_t slab; 2414 2415 if (keg == NULL) { 2416 keg = zone_first_keg(zone); 2417 KEG_LOCK(keg); 2418 } 2419 2420 for (;;) { 2421 slab = keg_fetch_slab(keg, zone, flags); 2422 if (slab) 2423 return (slab); 2424 if (flags & (M_NOWAIT | M_NOVM)) 2425 break; 2426 } 2427 KEG_UNLOCK(keg); 2428 return (NULL); 2429 } 2430 2431 /* 2432 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2433 * with the keg locked. On NULL no lock is held. 2434 * 2435 * The last pointer is used to seed the search. It is not required. 2436 */ 2437 static uma_slab_t 2438 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) 2439 { 2440 uma_klink_t klink; 2441 uma_slab_t slab; 2442 uma_keg_t keg; 2443 int flags; 2444 int empty; 2445 int full; 2446 2447 /* 2448 * Don't wait on the first pass. This will skip limit tests 2449 * as well. We don't want to block if we can find a provider 2450 * without blocking. 2451 */ 2452 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2453 /* 2454 * Use the last slab allocated as a hint for where to start 2455 * the search. 2456 */ 2457 if (last != NULL) { 2458 slab = keg_fetch_slab(last, zone, flags); 2459 if (slab) 2460 return (slab); 2461 KEG_UNLOCK(last); 2462 } 2463 /* 2464 * Loop until we have a slab incase of transient failures 2465 * while M_WAITOK is specified. I'm not sure this is 100% 2466 * required but we've done it for so long now. 2467 */ 2468 for (;;) { 2469 empty = 0; 2470 full = 0; 2471 /* 2472 * Search the available kegs for slabs. Be careful to hold the 2473 * correct lock while calling into the keg layer. 2474 */ 2475 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2476 keg = klink->kl_keg; 2477 KEG_LOCK(keg); 2478 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2479 slab = keg_fetch_slab(keg, zone, flags); 2480 if (slab) 2481 return (slab); 2482 } 2483 if (keg->uk_flags & UMA_ZFLAG_FULL) 2484 full++; 2485 else 2486 empty++; 2487 KEG_UNLOCK(keg); 2488 } 2489 if (rflags & (M_NOWAIT | M_NOVM)) 2490 break; 2491 flags = rflags; 2492 /* 2493 * All kegs are full. XXX We can't atomically check all kegs 2494 * and sleep so just sleep for a short period and retry. 2495 */ 2496 if (full && !empty) { 2497 ZONE_LOCK(zone); 2498 zone->uz_flags |= UMA_ZFLAG_FULL; 2499 zone->uz_sleeps++; 2500 zone_log_warning(zone); 2501 msleep(zone, zone->uz_lockptr, PVM, 2502 "zonelimit", hz/100); 2503 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2504 ZONE_UNLOCK(zone); 2505 continue; 2506 } 2507 } 2508 return (NULL); 2509 } 2510 2511 static void * 2512 slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 2513 { 2514 void *item; 2515 uint8_t freei; 2516 2517 MPASS(keg == slab->us_keg); 2518 mtx_assert(&keg->uk_lock, MA_OWNED); 2519 2520 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; 2521 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); 2522 item = slab->us_data + (keg->uk_rsize * freei); 2523 slab->us_freecount--; 2524 keg->uk_free--; 2525 2526 /* Move this slab to the full list */ 2527 if (slab->us_freecount == 0) { 2528 LIST_REMOVE(slab, us_link); 2529 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); 2530 } 2531 2532 return (item); 2533 } 2534 2535 static int 2536 zone_import(uma_zone_t zone, void **bucket, int max, int flags) 2537 { 2538 uma_slab_t slab; 2539 uma_keg_t keg; 2540 int i; 2541 2542 slab = NULL; 2543 keg = NULL; 2544 /* Try to keep the buckets totally full */ 2545 for (i = 0; i < max; ) { 2546 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL) 2547 break; 2548 keg = slab->us_keg; 2549 while (slab->us_freecount && i < max) { 2550 bucket[i++] = slab_alloc_item(keg, slab); 2551 if (keg->uk_free <= keg->uk_reserve) 2552 break; 2553 } 2554 /* Don't grab more than one slab at a time. */ 2555 flags &= ~M_WAITOK; 2556 flags |= M_NOWAIT; 2557 } 2558 if (slab != NULL) 2559 KEG_UNLOCK(keg); 2560 2561 return i; 2562 } 2563 2564 static uma_bucket_t 2565 zone_alloc_bucket(uma_zone_t zone, void *udata, int flags) 2566 { 2567 uma_bucket_t bucket; 2568 int max; 2569 2570 /* Don't wait for buckets, preserve caller's NOVM setting. */ 2571 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 2572 if (bucket == NULL) 2573 return (NULL); 2574 2575 max = MIN(bucket->ub_entries, zone->uz_count); 2576 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 2577 max, flags); 2578 2579 /* 2580 * Initialize the memory if necessary. 2581 */ 2582 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 2583 int i; 2584 2585 for (i = 0; i < bucket->ub_cnt; i++) 2586 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2587 flags) != 0) 2588 break; 2589 /* 2590 * If we couldn't initialize the whole bucket, put the 2591 * rest back onto the freelist. 2592 */ 2593 if (i != bucket->ub_cnt) { 2594 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 2595 bucket->ub_cnt - i); 2596 #ifdef INVARIANTS 2597 bzero(&bucket->ub_bucket[i], 2598 sizeof(void *) * (bucket->ub_cnt - i)); 2599 #endif 2600 bucket->ub_cnt = i; 2601 } 2602 } 2603 2604 if (bucket->ub_cnt == 0) { 2605 bucket_free(zone, bucket, udata); 2606 atomic_add_long(&zone->uz_fails, 1); 2607 return (NULL); 2608 } 2609 2610 return (bucket); 2611 } 2612 2613 /* 2614 * Allocates a single item from a zone. 2615 * 2616 * Arguments 2617 * zone The zone to alloc for. 2618 * udata The data to be passed to the constructor. 2619 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2620 * 2621 * Returns 2622 * NULL if there is no memory and M_NOWAIT is set 2623 * An item if successful 2624 */ 2625 2626 static void * 2627 zone_alloc_item(uma_zone_t zone, void *udata, int flags) 2628 { 2629 void *item; 2630 2631 item = NULL; 2632 2633 #ifdef UMA_DEBUG_ALLOC 2634 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); 2635 #endif 2636 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1) 2637 goto fail; 2638 atomic_add_long(&zone->uz_allocs, 1); 2639 2640 /* 2641 * We have to call both the zone's init (not the keg's init) 2642 * and the zone's ctor. This is because the item is going from 2643 * a keg slab directly to the user, and the user is expecting it 2644 * to be both zone-init'd as well as zone-ctor'd. 2645 */ 2646 if (zone->uz_init != NULL) { 2647 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 2648 zone_free_item(zone, item, udata, SKIP_FINI); 2649 goto fail; 2650 } 2651 } 2652 if (zone->uz_ctor != NULL) { 2653 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2654 zone_free_item(zone, item, udata, SKIP_DTOR); 2655 goto fail; 2656 } 2657 } 2658 #ifdef INVARIANTS 2659 uma_dbg_alloc(zone, NULL, item); 2660 #endif 2661 if (flags & M_ZERO) 2662 uma_zero_item(item, zone); 2663 2664 return (item); 2665 2666 fail: 2667 atomic_add_long(&zone->uz_fails, 1); 2668 return (NULL); 2669 } 2670 2671 /* See uma.h */ 2672 void 2673 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 2674 { 2675 uma_cache_t cache; 2676 uma_bucket_t bucket; 2677 int lockfail; 2678 int cpu; 2679 2680 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ 2681 random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA); 2682 2683 #ifdef UMA_DEBUG_ALLOC_1 2684 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); 2685 #endif 2686 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 2687 zone->uz_name); 2688 2689 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 2690 if (item == NULL) 2691 return; 2692 #ifdef DEBUG_MEMGUARD 2693 if (is_memguard_addr(item)) { 2694 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor) 2695 zone->uz_dtor(item, zone->uz_size, udata); 2696 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini) 2697 zone->uz_fini(item, zone->uz_size); 2698 memguard_free(item); 2699 return; 2700 } 2701 #endif 2702 #ifdef INVARIANTS 2703 if (zone->uz_flags & UMA_ZONE_MALLOC) 2704 uma_dbg_free(zone, udata, item); 2705 else 2706 uma_dbg_free(zone, NULL, item); 2707 #endif 2708 if (zone->uz_dtor != NULL) 2709 zone->uz_dtor(item, zone->uz_size, udata); 2710 2711 /* 2712 * The race here is acceptable. If we miss it we'll just have to wait 2713 * a little longer for the limits to be reset. 2714 */ 2715 if (zone->uz_flags & UMA_ZFLAG_FULL) 2716 goto zfree_item; 2717 2718 /* 2719 * If possible, free to the per-CPU cache. There are two 2720 * requirements for safe access to the per-CPU cache: (1) the thread 2721 * accessing the cache must not be preempted or yield during access, 2722 * and (2) the thread must not migrate CPUs without switching which 2723 * cache it accesses. We rely on a critical section to prevent 2724 * preemption and migration. We release the critical section in 2725 * order to acquire the zone mutex if we are unable to free to the 2726 * current cache; when we re-acquire the critical section, we must 2727 * detect and handle migration if it has occurred. 2728 */ 2729 zfree_restart: 2730 critical_enter(); 2731 cpu = curcpu; 2732 cache = &zone->uz_cpu[cpu]; 2733 2734 zfree_start: 2735 /* 2736 * Try to free into the allocbucket first to give LIFO ordering 2737 * for cache-hot datastructures. Spill over into the freebucket 2738 * if necessary. Alloc will swap them if one runs dry. 2739 */ 2740 bucket = cache->uc_allocbucket; 2741 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 2742 bucket = cache->uc_freebucket; 2743 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2744 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 2745 ("uma_zfree: Freeing to non free bucket index.")); 2746 bucket->ub_bucket[bucket->ub_cnt] = item; 2747 bucket->ub_cnt++; 2748 cache->uc_frees++; 2749 critical_exit(); 2750 return; 2751 } 2752 2753 /* 2754 * We must go back the zone, which requires acquiring the zone lock, 2755 * which in turn means we must release and re-acquire the critical 2756 * section. Since the critical section is released, we may be 2757 * preempted or migrate. As such, make sure not to maintain any 2758 * thread-local state specific to the cache from prior to releasing 2759 * the critical section. 2760 */ 2761 critical_exit(); 2762 if (zone->uz_count == 0 || bucketdisable) 2763 goto zfree_item; 2764 2765 lockfail = 0; 2766 if (ZONE_TRYLOCK(zone) == 0) { 2767 /* Record contention to size the buckets. */ 2768 ZONE_LOCK(zone); 2769 lockfail = 1; 2770 } 2771 critical_enter(); 2772 cpu = curcpu; 2773 cache = &zone->uz_cpu[cpu]; 2774 2775 /* 2776 * Since we have locked the zone we may as well send back our stats. 2777 */ 2778 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2779 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2780 cache->uc_allocs = 0; 2781 cache->uc_frees = 0; 2782 2783 bucket = cache->uc_freebucket; 2784 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2785 ZONE_UNLOCK(zone); 2786 goto zfree_start; 2787 } 2788 cache->uc_freebucket = NULL; 2789 2790 /* Can we throw this on the zone full list? */ 2791 if (bucket != NULL) { 2792 #ifdef UMA_DEBUG_ALLOC 2793 printf("uma_zfree: Putting old bucket on the free list.\n"); 2794 #endif 2795 /* ub_cnt is pointing to the last free item */ 2796 KASSERT(bucket->ub_cnt != 0, 2797 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 2798 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2799 } 2800 2801 /* We are no longer associated with this CPU. */ 2802 critical_exit(); 2803 2804 /* 2805 * We bump the uz count when the cache size is insufficient to 2806 * handle the working set. 2807 */ 2808 if (lockfail && zone->uz_count < BUCKET_MAX) 2809 zone->uz_count++; 2810 ZONE_UNLOCK(zone); 2811 2812 #ifdef UMA_DEBUG_ALLOC 2813 printf("uma_zfree: Allocating new free bucket.\n"); 2814 #endif 2815 bucket = bucket_alloc(zone, udata, M_NOWAIT); 2816 if (bucket) { 2817 critical_enter(); 2818 cpu = curcpu; 2819 cache = &zone->uz_cpu[cpu]; 2820 if (cache->uc_freebucket == NULL) { 2821 cache->uc_freebucket = bucket; 2822 goto zfree_start; 2823 } 2824 /* 2825 * We lost the race, start over. We have to drop our 2826 * critical section to free the bucket. 2827 */ 2828 critical_exit(); 2829 bucket_free(zone, bucket, udata); 2830 goto zfree_restart; 2831 } 2832 2833 /* 2834 * If nothing else caught this, we'll just do an internal free. 2835 */ 2836 zfree_item: 2837 zone_free_item(zone, item, udata, SKIP_DTOR); 2838 2839 return; 2840 } 2841 2842 static void 2843 slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) 2844 { 2845 uint8_t freei; 2846 2847 mtx_assert(&keg->uk_lock, MA_OWNED); 2848 MPASS(keg == slab->us_keg); 2849 2850 /* Do we need to remove from any lists? */ 2851 if (slab->us_freecount+1 == keg->uk_ipers) { 2852 LIST_REMOVE(slab, us_link); 2853 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 2854 } else if (slab->us_freecount == 0) { 2855 LIST_REMOVE(slab, us_link); 2856 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2857 } 2858 2859 /* Slab management. */ 2860 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 2861 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); 2862 slab->us_freecount++; 2863 2864 /* Keg statistics. */ 2865 keg->uk_free++; 2866 } 2867 2868 static void 2869 zone_release(uma_zone_t zone, void **bucket, int cnt) 2870 { 2871 void *item; 2872 uma_slab_t slab; 2873 uma_keg_t keg; 2874 uint8_t *mem; 2875 int clearfull; 2876 int i; 2877 2878 clearfull = 0; 2879 keg = zone_first_keg(zone); 2880 KEG_LOCK(keg); 2881 for (i = 0; i < cnt; i++) { 2882 item = bucket[i]; 2883 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 2884 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 2885 if (zone->uz_flags & UMA_ZONE_HASH) { 2886 slab = hash_sfind(&keg->uk_hash, mem); 2887 } else { 2888 mem += keg->uk_pgoff; 2889 slab = (uma_slab_t)mem; 2890 } 2891 } else { 2892 slab = vtoslab((vm_offset_t)item); 2893 if (slab->us_keg != keg) { 2894 KEG_UNLOCK(keg); 2895 keg = slab->us_keg; 2896 KEG_LOCK(keg); 2897 } 2898 } 2899 slab_free_item(keg, slab, item); 2900 if (keg->uk_flags & UMA_ZFLAG_FULL) { 2901 if (keg->uk_pages < keg->uk_maxpages) { 2902 keg->uk_flags &= ~UMA_ZFLAG_FULL; 2903 clearfull = 1; 2904 } 2905 2906 /* 2907 * We can handle one more allocation. Since we're 2908 * clearing ZFLAG_FULL, wake up all procs blocked 2909 * on pages. This should be uncommon, so keeping this 2910 * simple for now (rather than adding count of blocked 2911 * threads etc). 2912 */ 2913 wakeup(keg); 2914 } 2915 } 2916 KEG_UNLOCK(keg); 2917 if (clearfull) { 2918 ZONE_LOCK(zone); 2919 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2920 wakeup(zone); 2921 ZONE_UNLOCK(zone); 2922 } 2923 2924 } 2925 2926 /* 2927 * Frees a single item to any zone. 2928 * 2929 * Arguments: 2930 * zone The zone to free to 2931 * item The item we're freeing 2932 * udata User supplied data for the dtor 2933 * skip Skip dtors and finis 2934 */ 2935 static void 2936 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 2937 { 2938 2939 #ifdef INVARIANTS 2940 if (skip == SKIP_NONE) { 2941 if (zone->uz_flags & UMA_ZONE_MALLOC) 2942 uma_dbg_free(zone, udata, item); 2943 else 2944 uma_dbg_free(zone, NULL, item); 2945 } 2946 #endif 2947 if (skip < SKIP_DTOR && zone->uz_dtor) 2948 zone->uz_dtor(item, zone->uz_size, udata); 2949 2950 if (skip < SKIP_FINI && zone->uz_fini) 2951 zone->uz_fini(item, zone->uz_size); 2952 2953 atomic_add_long(&zone->uz_frees, 1); 2954 zone->uz_release(zone->uz_arg, &item, 1); 2955 } 2956 2957 /* See uma.h */ 2958 int 2959 uma_zone_set_max(uma_zone_t zone, int nitems) 2960 { 2961 uma_keg_t keg; 2962 2963 keg = zone_first_keg(zone); 2964 if (keg == NULL) 2965 return (0); 2966 KEG_LOCK(keg); 2967 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 2968 if (keg->uk_maxpages * keg->uk_ipers < nitems) 2969 keg->uk_maxpages += keg->uk_ppera; 2970 nitems = keg->uk_maxpages * keg->uk_ipers; 2971 KEG_UNLOCK(keg); 2972 2973 return (nitems); 2974 } 2975 2976 /* See uma.h */ 2977 int 2978 uma_zone_get_max(uma_zone_t zone) 2979 { 2980 int nitems; 2981 uma_keg_t keg; 2982 2983 keg = zone_first_keg(zone); 2984 if (keg == NULL) 2985 return (0); 2986 KEG_LOCK(keg); 2987 nitems = keg->uk_maxpages * keg->uk_ipers; 2988 KEG_UNLOCK(keg); 2989 2990 return (nitems); 2991 } 2992 2993 /* See uma.h */ 2994 void 2995 uma_zone_set_warning(uma_zone_t zone, const char *warning) 2996 { 2997 2998 ZONE_LOCK(zone); 2999 zone->uz_warning = warning; 3000 ZONE_UNLOCK(zone); 3001 } 3002 3003 /* See uma.h */ 3004 int 3005 uma_zone_get_cur(uma_zone_t zone) 3006 { 3007 int64_t nitems; 3008 u_int i; 3009 3010 ZONE_LOCK(zone); 3011 nitems = zone->uz_allocs - zone->uz_frees; 3012 CPU_FOREACH(i) { 3013 /* 3014 * See the comment in sysctl_vm_zone_stats() regarding the 3015 * safety of accessing the per-cpu caches. With the zone lock 3016 * held, it is safe, but can potentially result in stale data. 3017 */ 3018 nitems += zone->uz_cpu[i].uc_allocs - 3019 zone->uz_cpu[i].uc_frees; 3020 } 3021 ZONE_UNLOCK(zone); 3022 3023 return (nitems < 0 ? 0 : nitems); 3024 } 3025 3026 /* See uma.h */ 3027 void 3028 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 3029 { 3030 uma_keg_t keg; 3031 3032 keg = zone_first_keg(zone); 3033 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 3034 KEG_LOCK(keg); 3035 KASSERT(keg->uk_pages == 0, 3036 ("uma_zone_set_init on non-empty keg")); 3037 keg->uk_init = uminit; 3038 KEG_UNLOCK(keg); 3039 } 3040 3041 /* See uma.h */ 3042 void 3043 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 3044 { 3045 uma_keg_t keg; 3046 3047 keg = zone_first_keg(zone); 3048 KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type")); 3049 KEG_LOCK(keg); 3050 KASSERT(keg->uk_pages == 0, 3051 ("uma_zone_set_fini on non-empty keg")); 3052 keg->uk_fini = fini; 3053 KEG_UNLOCK(keg); 3054 } 3055 3056 /* See uma.h */ 3057 void 3058 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 3059 { 3060 3061 ZONE_LOCK(zone); 3062 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3063 ("uma_zone_set_zinit on non-empty keg")); 3064 zone->uz_init = zinit; 3065 ZONE_UNLOCK(zone); 3066 } 3067 3068 /* See uma.h */ 3069 void 3070 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 3071 { 3072 3073 ZONE_LOCK(zone); 3074 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3075 ("uma_zone_set_zfini on non-empty keg")); 3076 zone->uz_fini = zfini; 3077 ZONE_UNLOCK(zone); 3078 } 3079 3080 /* See uma.h */ 3081 /* XXX uk_freef is not actually used with the zone locked */ 3082 void 3083 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 3084 { 3085 uma_keg_t keg; 3086 3087 keg = zone_first_keg(zone); 3088 KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type")); 3089 KEG_LOCK(keg); 3090 keg->uk_freef = freef; 3091 KEG_UNLOCK(keg); 3092 } 3093 3094 /* See uma.h */ 3095 /* XXX uk_allocf is not actually used with the zone locked */ 3096 void 3097 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3098 { 3099 uma_keg_t keg; 3100 3101 keg = zone_first_keg(zone); 3102 KEG_LOCK(keg); 3103 keg->uk_allocf = allocf; 3104 KEG_UNLOCK(keg); 3105 } 3106 3107 /* See uma.h */ 3108 void 3109 uma_zone_reserve(uma_zone_t zone, int items) 3110 { 3111 uma_keg_t keg; 3112 3113 keg = zone_first_keg(zone); 3114 if (keg == NULL) 3115 return; 3116 KEG_LOCK(keg); 3117 keg->uk_reserve = items; 3118 KEG_UNLOCK(keg); 3119 3120 return; 3121 } 3122 3123 /* See uma.h */ 3124 int 3125 uma_zone_reserve_kva(uma_zone_t zone, int count) 3126 { 3127 uma_keg_t keg; 3128 vm_offset_t kva; 3129 u_int pages; 3130 3131 keg = zone_first_keg(zone); 3132 if (keg == NULL) 3133 return (0); 3134 pages = count / keg->uk_ipers; 3135 3136 if (pages * keg->uk_ipers < count) 3137 pages++; 3138 3139 #ifdef UMA_MD_SMALL_ALLOC 3140 if (keg->uk_ppera > 1) { 3141 #else 3142 if (1) { 3143 #endif 3144 kva = kva_alloc((vm_size_t)pages * UMA_SLAB_SIZE); 3145 if (kva == 0) 3146 return (0); 3147 } else 3148 kva = 0; 3149 KEG_LOCK(keg); 3150 keg->uk_kva = kva; 3151 keg->uk_offset = 0; 3152 keg->uk_maxpages = pages; 3153 #ifdef UMA_MD_SMALL_ALLOC 3154 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3155 #else 3156 keg->uk_allocf = noobj_alloc; 3157 #endif 3158 keg->uk_flags |= UMA_ZONE_NOFREE; 3159 KEG_UNLOCK(keg); 3160 3161 return (1); 3162 } 3163 3164 /* See uma.h */ 3165 void 3166 uma_prealloc(uma_zone_t zone, int items) 3167 { 3168 int slabs; 3169 uma_slab_t slab; 3170 uma_keg_t keg; 3171 3172 keg = zone_first_keg(zone); 3173 if (keg == NULL) 3174 return; 3175 KEG_LOCK(keg); 3176 slabs = items / keg->uk_ipers; 3177 if (slabs * keg->uk_ipers < items) 3178 slabs++; 3179 while (slabs > 0) { 3180 slab = keg_alloc_slab(keg, zone, M_WAITOK); 3181 if (slab == NULL) 3182 break; 3183 MPASS(slab->us_keg == keg); 3184 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 3185 slabs--; 3186 } 3187 KEG_UNLOCK(keg); 3188 } 3189 3190 /* See uma.h */ 3191 uint32_t * 3192 uma_find_refcnt(uma_zone_t zone, void *item) 3193 { 3194 uma_slabrefcnt_t slabref; 3195 uma_slab_t slab; 3196 uma_keg_t keg; 3197 uint32_t *refcnt; 3198 int idx; 3199 3200 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK)); 3201 slabref = (uma_slabrefcnt_t)slab; 3202 keg = slab->us_keg; 3203 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT, 3204 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT")); 3205 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3206 refcnt = &slabref->us_refcnt[idx]; 3207 return refcnt; 3208 } 3209 3210 /* See uma.h */ 3211 static void 3212 uma_reclaim_locked(bool kmem_danger) 3213 { 3214 3215 #ifdef UMA_DEBUG 3216 printf("UMA: vm asked us to release pages!\n"); 3217 #endif 3218 sx_assert(&uma_drain_lock, SA_XLOCKED); 3219 bucket_enable(); 3220 zone_foreach(zone_drain); 3221 if (vm_page_count_min() || kmem_danger) { 3222 cache_drain_safe(NULL); 3223 zone_foreach(zone_drain); 3224 } 3225 /* 3226 * Some slabs may have been freed but this zone will be visited early 3227 * we visit again so that we can free pages that are empty once other 3228 * zones are drained. We have to do the same for buckets. 3229 */ 3230 zone_drain(slabzone); 3231 zone_drain(slabrefzone); 3232 bucket_zone_drain(); 3233 } 3234 3235 void 3236 uma_reclaim(void) 3237 { 3238 3239 sx_xlock(&uma_drain_lock); 3240 uma_reclaim_locked(false); 3241 sx_xunlock(&uma_drain_lock); 3242 } 3243 3244 static int uma_reclaim_needed; 3245 3246 void 3247 uma_reclaim_wakeup(void) 3248 { 3249 3250 uma_reclaim_needed = 1; 3251 wakeup(&uma_reclaim_needed); 3252 } 3253 3254 void 3255 uma_reclaim_worker(void *arg __unused) 3256 { 3257 3258 sx_xlock(&uma_drain_lock); 3259 for (;;) { 3260 sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM, 3261 "umarcl", 0); 3262 if (uma_reclaim_needed) { 3263 uma_reclaim_needed = 0; 3264 uma_reclaim_locked(true); 3265 } 3266 } 3267 } 3268 3269 /* See uma.h */ 3270 int 3271 uma_zone_exhausted(uma_zone_t zone) 3272 { 3273 int full; 3274 3275 ZONE_LOCK(zone); 3276 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3277 ZONE_UNLOCK(zone); 3278 return (full); 3279 } 3280 3281 int 3282 uma_zone_exhausted_nolock(uma_zone_t zone) 3283 { 3284 return (zone->uz_flags & UMA_ZFLAG_FULL); 3285 } 3286 3287 void * 3288 uma_large_malloc(vm_size_t size, int wait) 3289 { 3290 void *mem; 3291 uma_slab_t slab; 3292 uint8_t flags; 3293 3294 slab = zone_alloc_item(slabzone, NULL, wait); 3295 if (slab == NULL) 3296 return (NULL); 3297 mem = page_alloc(NULL, size, &flags, wait); 3298 if (mem) { 3299 vsetslab((vm_offset_t)mem, slab); 3300 slab->us_data = mem; 3301 slab->us_flags = flags | UMA_SLAB_MALLOC; 3302 slab->us_size = size; 3303 } else { 3304 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3305 } 3306 3307 return (mem); 3308 } 3309 3310 void 3311 uma_large_free(uma_slab_t slab) 3312 { 3313 3314 page_free(slab->us_data, slab->us_size, slab->us_flags); 3315 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3316 } 3317 3318 static void 3319 uma_zero_item(void *item, uma_zone_t zone) 3320 { 3321 3322 if (zone->uz_flags & UMA_ZONE_PCPU) { 3323 for (int i = 0; i < mp_ncpus; i++) 3324 bzero(zpcpu_get_cpu(item, i), zone->uz_size); 3325 } else 3326 bzero(item, zone->uz_size); 3327 } 3328 3329 void 3330 uma_print_stats(void) 3331 { 3332 zone_foreach(uma_print_zone); 3333 } 3334 3335 static void 3336 slab_print(uma_slab_t slab) 3337 { 3338 printf("slab: keg %p, data %p, freecount %d\n", 3339 slab->us_keg, slab->us_data, slab->us_freecount); 3340 } 3341 3342 static void 3343 cache_print(uma_cache_t cache) 3344 { 3345 printf("alloc: %p(%d), free: %p(%d)\n", 3346 cache->uc_allocbucket, 3347 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3348 cache->uc_freebucket, 3349 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3350 } 3351 3352 static void 3353 uma_print_keg(uma_keg_t keg) 3354 { 3355 uma_slab_t slab; 3356 3357 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3358 "out %d free %d limit %d\n", 3359 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3360 keg->uk_ipers, keg->uk_ppera, 3361 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free, 3362 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3363 printf("Part slabs:\n"); 3364 LIST_FOREACH(slab, &keg->uk_part_slab, us_link) 3365 slab_print(slab); 3366 printf("Free slabs:\n"); 3367 LIST_FOREACH(slab, &keg->uk_free_slab, us_link) 3368 slab_print(slab); 3369 printf("Full slabs:\n"); 3370 LIST_FOREACH(slab, &keg->uk_full_slab, us_link) 3371 slab_print(slab); 3372 } 3373 3374 void 3375 uma_print_zone(uma_zone_t zone) 3376 { 3377 uma_cache_t cache; 3378 uma_klink_t kl; 3379 int i; 3380 3381 printf("zone: %s(%p) size %d flags %#x\n", 3382 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3383 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3384 uma_print_keg(kl->kl_keg); 3385 CPU_FOREACH(i) { 3386 cache = &zone->uz_cpu[i]; 3387 printf("CPU %d Cache:\n", i); 3388 cache_print(cache); 3389 } 3390 } 3391 3392 #ifdef DDB 3393 /* 3394 * Generate statistics across both the zone and its per-cpu cache's. Return 3395 * desired statistics if the pointer is non-NULL for that statistic. 3396 * 3397 * Note: does not update the zone statistics, as it can't safely clear the 3398 * per-CPU cache statistic. 3399 * 3400 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3401 * safe from off-CPU; we should modify the caches to track this information 3402 * directly so that we don't have to. 3403 */ 3404 static void 3405 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3406 uint64_t *freesp, uint64_t *sleepsp) 3407 { 3408 uma_cache_t cache; 3409 uint64_t allocs, frees, sleeps; 3410 int cachefree, cpu; 3411 3412 allocs = frees = sleeps = 0; 3413 cachefree = 0; 3414 CPU_FOREACH(cpu) { 3415 cache = &z->uz_cpu[cpu]; 3416 if (cache->uc_allocbucket != NULL) 3417 cachefree += cache->uc_allocbucket->ub_cnt; 3418 if (cache->uc_freebucket != NULL) 3419 cachefree += cache->uc_freebucket->ub_cnt; 3420 allocs += cache->uc_allocs; 3421 frees += cache->uc_frees; 3422 } 3423 allocs += z->uz_allocs; 3424 frees += z->uz_frees; 3425 sleeps += z->uz_sleeps; 3426 if (cachefreep != NULL) 3427 *cachefreep = cachefree; 3428 if (allocsp != NULL) 3429 *allocsp = allocs; 3430 if (freesp != NULL) 3431 *freesp = frees; 3432 if (sleepsp != NULL) 3433 *sleepsp = sleeps; 3434 } 3435 #endif /* DDB */ 3436 3437 static int 3438 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3439 { 3440 uma_keg_t kz; 3441 uma_zone_t z; 3442 int count; 3443 3444 count = 0; 3445 rw_rlock(&uma_rwlock); 3446 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3447 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3448 count++; 3449 } 3450 rw_runlock(&uma_rwlock); 3451 return (sysctl_handle_int(oidp, &count, 0, req)); 3452 } 3453 3454 static int 3455 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3456 { 3457 struct uma_stream_header ush; 3458 struct uma_type_header uth; 3459 struct uma_percpu_stat ups; 3460 uma_bucket_t bucket; 3461 struct sbuf sbuf; 3462 uma_cache_t cache; 3463 uma_klink_t kl; 3464 uma_keg_t kz; 3465 uma_zone_t z; 3466 uma_keg_t k; 3467 int count, error, i; 3468 3469 error = sysctl_wire_old_buffer(req, 0); 3470 if (error != 0) 3471 return (error); 3472 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3473 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); 3474 3475 count = 0; 3476 rw_rlock(&uma_rwlock); 3477 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3478 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3479 count++; 3480 } 3481 3482 /* 3483 * Insert stream header. 3484 */ 3485 bzero(&ush, sizeof(ush)); 3486 ush.ush_version = UMA_STREAM_VERSION; 3487 ush.ush_maxcpus = (mp_maxid + 1); 3488 ush.ush_count = count; 3489 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3490 3491 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3492 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3493 bzero(&uth, sizeof(uth)); 3494 ZONE_LOCK(z); 3495 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3496 uth.uth_align = kz->uk_align; 3497 uth.uth_size = kz->uk_size; 3498 uth.uth_rsize = kz->uk_rsize; 3499 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3500 k = kl->kl_keg; 3501 uth.uth_maxpages += k->uk_maxpages; 3502 uth.uth_pages += k->uk_pages; 3503 uth.uth_keg_free += k->uk_free; 3504 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3505 * k->uk_ipers; 3506 } 3507 3508 /* 3509 * A zone is secondary is it is not the first entry 3510 * on the keg's zone list. 3511 */ 3512 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3513 (LIST_FIRST(&kz->uk_zones) != z)) 3514 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3515 3516 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3517 uth.uth_zone_free += bucket->ub_cnt; 3518 uth.uth_allocs = z->uz_allocs; 3519 uth.uth_frees = z->uz_frees; 3520 uth.uth_fails = z->uz_fails; 3521 uth.uth_sleeps = z->uz_sleeps; 3522 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 3523 /* 3524 * While it is not normally safe to access the cache 3525 * bucket pointers while not on the CPU that owns the 3526 * cache, we only allow the pointers to be exchanged 3527 * without the zone lock held, not invalidated, so 3528 * accept the possible race associated with bucket 3529 * exchange during monitoring. 3530 */ 3531 for (i = 0; i < (mp_maxid + 1); i++) { 3532 bzero(&ups, sizeof(ups)); 3533 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) 3534 goto skip; 3535 if (CPU_ABSENT(i)) 3536 goto skip; 3537 cache = &z->uz_cpu[i]; 3538 if (cache->uc_allocbucket != NULL) 3539 ups.ups_cache_free += 3540 cache->uc_allocbucket->ub_cnt; 3541 if (cache->uc_freebucket != NULL) 3542 ups.ups_cache_free += 3543 cache->uc_freebucket->ub_cnt; 3544 ups.ups_allocs = cache->uc_allocs; 3545 ups.ups_frees = cache->uc_frees; 3546 skip: 3547 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups)); 3548 } 3549 ZONE_UNLOCK(z); 3550 } 3551 } 3552 rw_runlock(&uma_rwlock); 3553 error = sbuf_finish(&sbuf); 3554 sbuf_delete(&sbuf); 3555 return (error); 3556 } 3557 3558 int 3559 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) 3560 { 3561 uma_zone_t zone = *(uma_zone_t *)arg1; 3562 int error, max; 3563 3564 max = uma_zone_get_max(zone); 3565 error = sysctl_handle_int(oidp, &max, 0, req); 3566 if (error || !req->newptr) 3567 return (error); 3568 3569 uma_zone_set_max(zone, max); 3570 3571 return (0); 3572 } 3573 3574 int 3575 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) 3576 { 3577 uma_zone_t zone = *(uma_zone_t *)arg1; 3578 int cur; 3579 3580 cur = uma_zone_get_cur(zone); 3581 return (sysctl_handle_int(oidp, &cur, 0, req)); 3582 } 3583 3584 #ifdef DDB 3585 DB_SHOW_COMMAND(uma, db_show_uma) 3586 { 3587 uint64_t allocs, frees, sleeps; 3588 uma_bucket_t bucket; 3589 uma_keg_t kz; 3590 uma_zone_t z; 3591 int cachefree; 3592 3593 db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used", 3594 "Free", "Requests", "Sleeps", "Bucket"); 3595 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3596 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3597 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 3598 allocs = z->uz_allocs; 3599 frees = z->uz_frees; 3600 sleeps = z->uz_sleeps; 3601 cachefree = 0; 3602 } else 3603 uma_zone_sumstat(z, &cachefree, &allocs, 3604 &frees, &sleeps); 3605 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 3606 (LIST_FIRST(&kz->uk_zones) != z))) 3607 cachefree += kz->uk_free; 3608 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3609 cachefree += bucket->ub_cnt; 3610 db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n", 3611 z->uz_name, (uintmax_t)kz->uk_size, 3612 (intmax_t)(allocs - frees), cachefree, 3613 (uintmax_t)allocs, sleeps, z->uz_count); 3614 if (db_pager_quit) 3615 return; 3616 } 3617 } 3618 } 3619 3620 DB_SHOW_COMMAND(umacache, db_show_umacache) 3621 { 3622 uint64_t allocs, frees; 3623 uma_bucket_t bucket; 3624 uma_zone_t z; 3625 int cachefree; 3626 3627 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 3628 "Requests", "Bucket"); 3629 LIST_FOREACH(z, &uma_cachezones, uz_link) { 3630 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL); 3631 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3632 cachefree += bucket->ub_cnt; 3633 db_printf("%18s %8ju %8jd %8d %12ju %8u\n", 3634 z->uz_name, (uintmax_t)z->uz_size, 3635 (intmax_t)(allocs - frees), cachefree, 3636 (uintmax_t)allocs, z->uz_count); 3637 if (db_pager_quit) 3638 return; 3639 } 3640 } 3641 #endif 3642