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