1 /*- 2 * Copyright (c) 2002-2005, 2009 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 63 #include <sys/param.h> 64 #include <sys/systm.h> 65 #include <sys/kernel.h> 66 #include <sys/types.h> 67 #include <sys/queue.h> 68 #include <sys/malloc.h> 69 #include <sys/ktr.h> 70 #include <sys/lock.h> 71 #include <sys/sysctl.h> 72 #include <sys/mutex.h> 73 #include <sys/proc.h> 74 #include <sys/sbuf.h> 75 #include <sys/smp.h> 76 #include <sys/vmmeter.h> 77 78 #include <vm/vm.h> 79 #include <vm/vm_object.h> 80 #include <vm/vm_page.h> 81 #include <vm/vm_param.h> 82 #include <vm/vm_map.h> 83 #include <vm/vm_kern.h> 84 #include <vm/vm_extern.h> 85 #include <vm/uma.h> 86 #include <vm/uma_int.h> 87 #include <vm/uma_dbg.h> 88 89 #include <machine/vmparam.h> 90 91 #include <ddb/ddb.h> 92 93 /* 94 * This is the zone and keg from which all zones are spawned. The idea is that 95 * even the zone & keg heads are allocated from the allocator, so we use the 96 * bss section to bootstrap us. 97 */ 98 static struct uma_keg masterkeg; 99 static struct uma_zone masterzone_k; 100 static struct uma_zone masterzone_z; 101 static uma_zone_t kegs = &masterzone_k; 102 static uma_zone_t zones = &masterzone_z; 103 104 /* This is the zone from which all of uma_slab_t's are allocated. */ 105 static uma_zone_t slabzone; 106 static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */ 107 108 /* 109 * The initial hash tables come out of this zone so they can be allocated 110 * prior to malloc coming up. 111 */ 112 static uma_zone_t hashzone; 113 114 /* The boot-time adjusted value for cache line alignment. */ 115 static int uma_align_cache = 64 - 1; 116 117 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 118 119 /* 120 * Are we allowed to allocate buckets? 121 */ 122 static int bucketdisable = 1; 123 124 /* Linked list of all kegs in the system */ 125 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(&uma_kegs); 126 127 /* This mutex protects the keg list */ 128 static struct mtx uma_mtx; 129 130 /* Linked list of boot time pages */ 131 static LIST_HEAD(,uma_slab) uma_boot_pages = 132 LIST_HEAD_INITIALIZER(&uma_boot_pages); 133 134 /* This mutex protects the boot time pages list */ 135 static struct mtx uma_boot_pages_mtx; 136 137 /* Is the VM done starting up? */ 138 static int booted = 0; 139 140 /* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */ 141 static u_int uma_max_ipers; 142 static u_int uma_max_ipers_ref; 143 144 /* 145 * This is the handle used to schedule events that need to happen 146 * outside of the allocation fast path. 147 */ 148 static struct callout uma_callout; 149 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 150 151 /* 152 * This structure is passed as the zone ctor arg so that I don't have to create 153 * a special allocation function just for zones. 154 */ 155 struct uma_zctor_args { 156 char *name; 157 size_t size; 158 uma_ctor ctor; 159 uma_dtor dtor; 160 uma_init uminit; 161 uma_fini fini; 162 uma_keg_t keg; 163 int align; 164 u_int32_t flags; 165 }; 166 167 struct uma_kctor_args { 168 uma_zone_t zone; 169 size_t size; 170 uma_init uminit; 171 uma_fini fini; 172 int align; 173 u_int32_t flags; 174 }; 175 176 struct uma_bucket_zone { 177 uma_zone_t ubz_zone; 178 char *ubz_name; 179 int ubz_entries; 180 }; 181 182 #define BUCKET_MAX 128 183 184 struct uma_bucket_zone bucket_zones[] = { 185 { NULL, "16 Bucket", 16 }, 186 { NULL, "32 Bucket", 32 }, 187 { NULL, "64 Bucket", 64 }, 188 { NULL, "128 Bucket", 128 }, 189 { NULL, NULL, 0} 190 }; 191 192 #define BUCKET_SHIFT 4 193 #define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1) 194 195 /* 196 * bucket_size[] maps requested bucket sizes to zones that allocate a bucket 197 * of approximately the right size. 198 */ 199 static uint8_t bucket_size[BUCKET_ZONES]; 200 201 /* 202 * Flags and enumerations to be passed to internal functions. 203 */ 204 enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI }; 205 206 #define ZFREE_STATFAIL 0x00000001 /* Update zone failure statistic. */ 207 #define ZFREE_STATFREE 0x00000002 /* Update zone free statistic. */ 208 209 /* Prototypes.. */ 210 211 static void *obj_alloc(uma_zone_t, int, u_int8_t *, int); 212 static void *page_alloc(uma_zone_t, int, u_int8_t *, int); 213 static void *startup_alloc(uma_zone_t, int, u_int8_t *, int); 214 static void page_free(void *, int, u_int8_t); 215 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); 216 static void cache_drain(uma_zone_t); 217 static void bucket_drain(uma_zone_t, uma_bucket_t); 218 static void bucket_cache_drain(uma_zone_t zone); 219 static int keg_ctor(void *, int, void *, int); 220 static void keg_dtor(void *, int, void *); 221 static int zone_ctor(void *, int, void *, int); 222 static void zone_dtor(void *, int, void *); 223 static int zero_init(void *, int, int); 224 static void keg_small_init(uma_keg_t keg); 225 static void keg_large_init(uma_keg_t keg); 226 static void zone_foreach(void (*zfunc)(uma_zone_t)); 227 static void zone_timeout(uma_zone_t zone); 228 static int hash_alloc(struct uma_hash *); 229 static int hash_expand(struct uma_hash *, struct uma_hash *); 230 static void hash_free(struct uma_hash *hash); 231 static void uma_timeout(void *); 232 static void uma_startup3(void); 233 static void *zone_alloc_item(uma_zone_t, void *, int); 234 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip, 235 int); 236 static void bucket_enable(void); 237 static void bucket_init(void); 238 static uma_bucket_t bucket_alloc(int, int); 239 static void bucket_free(uma_bucket_t); 240 static void bucket_zone_drain(void); 241 static int zone_alloc_bucket(uma_zone_t zone, int flags); 242 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); 243 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); 244 static void *slab_alloc_item(uma_zone_t zone, uma_slab_t slab); 245 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 246 uma_fini fini, int align, u_int32_t flags); 247 static inline void zone_relock(uma_zone_t zone, uma_keg_t keg); 248 static inline void keg_relock(uma_keg_t keg, uma_zone_t zone); 249 250 void uma_print_zone(uma_zone_t); 251 void uma_print_stats(void); 252 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 253 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 254 255 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 256 257 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 258 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 259 260 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 261 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 262 263 /* 264 * This routine checks to see whether or not it's safe to enable buckets. 265 */ 266 267 static void 268 bucket_enable(void) 269 { 270 if (cnt.v_free_count < cnt.v_free_min) 271 bucketdisable = 1; 272 else 273 bucketdisable = 0; 274 } 275 276 /* 277 * Initialize bucket_zones, the array of zones of buckets of various sizes. 278 * 279 * For each zone, calculate the memory required for each bucket, consisting 280 * of the header and an array of pointers. Initialize bucket_size[] to point 281 * the range of appropriate bucket sizes at the zone. 282 */ 283 static void 284 bucket_init(void) 285 { 286 struct uma_bucket_zone *ubz; 287 int i; 288 int j; 289 290 for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) { 291 int size; 292 293 ubz = &bucket_zones[j]; 294 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 295 size += sizeof(void *) * ubz->ubz_entries; 296 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 297 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 298 UMA_ZFLAG_INTERNAL | UMA_ZFLAG_BUCKET); 299 for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT)) 300 bucket_size[i >> BUCKET_SHIFT] = j; 301 } 302 } 303 304 /* 305 * Given a desired number of entries for a bucket, return the zone from which 306 * to allocate the bucket. 307 */ 308 static struct uma_bucket_zone * 309 bucket_zone_lookup(int entries) 310 { 311 int idx; 312 313 idx = howmany(entries, 1 << BUCKET_SHIFT); 314 return (&bucket_zones[bucket_size[idx]]); 315 } 316 317 static uma_bucket_t 318 bucket_alloc(int entries, int bflags) 319 { 320 struct uma_bucket_zone *ubz; 321 uma_bucket_t bucket; 322 323 /* 324 * This is to stop us from allocating per cpu buckets while we're 325 * running out of vm.boot_pages. Otherwise, we would exhaust the 326 * boot pages. This also prevents us from allocating buckets in 327 * low memory situations. 328 */ 329 if (bucketdisable) 330 return (NULL); 331 332 ubz = bucket_zone_lookup(entries); 333 bucket = zone_alloc_item(ubz->ubz_zone, NULL, bflags); 334 if (bucket) { 335 #ifdef INVARIANTS 336 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 337 #endif 338 bucket->ub_cnt = 0; 339 bucket->ub_entries = ubz->ubz_entries; 340 } 341 342 return (bucket); 343 } 344 345 static void 346 bucket_free(uma_bucket_t bucket) 347 { 348 struct uma_bucket_zone *ubz; 349 350 ubz = bucket_zone_lookup(bucket->ub_entries); 351 zone_free_item(ubz->ubz_zone, bucket, NULL, SKIP_NONE, 352 ZFREE_STATFREE); 353 } 354 355 static void 356 bucket_zone_drain(void) 357 { 358 struct uma_bucket_zone *ubz; 359 360 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 361 zone_drain(ubz->ubz_zone); 362 } 363 364 static inline uma_keg_t 365 zone_first_keg(uma_zone_t zone) 366 { 367 368 return (LIST_FIRST(&zone->uz_kegs)->kl_keg); 369 } 370 371 static void 372 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 373 { 374 uma_klink_t klink; 375 376 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 377 kegfn(klink->kl_keg); 378 } 379 380 /* 381 * Routine called by timeout which is used to fire off some time interval 382 * based calculations. (stats, hash size, etc.) 383 * 384 * Arguments: 385 * arg Unused 386 * 387 * Returns: 388 * Nothing 389 */ 390 static void 391 uma_timeout(void *unused) 392 { 393 bucket_enable(); 394 zone_foreach(zone_timeout); 395 396 /* Reschedule this event */ 397 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 398 } 399 400 /* 401 * Routine to perform timeout driven calculations. This expands the 402 * hashes and does per cpu statistics aggregation. 403 * 404 * Returns nothing. 405 */ 406 static void 407 keg_timeout(uma_keg_t keg) 408 { 409 410 KEG_LOCK(keg); 411 /* 412 * Expand the keg hash table. 413 * 414 * This is done if the number of slabs is larger than the hash size. 415 * What I'm trying to do here is completely reduce collisions. This 416 * may be a little aggressive. Should I allow for two collisions max? 417 */ 418 if (keg->uk_flags & UMA_ZONE_HASH && 419 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 420 struct uma_hash newhash; 421 struct uma_hash oldhash; 422 int ret; 423 424 /* 425 * This is so involved because allocating and freeing 426 * while the keg lock is held will lead to deadlock. 427 * I have to do everything in stages and check for 428 * races. 429 */ 430 newhash = keg->uk_hash; 431 KEG_UNLOCK(keg); 432 ret = hash_alloc(&newhash); 433 KEG_LOCK(keg); 434 if (ret) { 435 if (hash_expand(&keg->uk_hash, &newhash)) { 436 oldhash = keg->uk_hash; 437 keg->uk_hash = newhash; 438 } else 439 oldhash = newhash; 440 441 KEG_UNLOCK(keg); 442 hash_free(&oldhash); 443 KEG_LOCK(keg); 444 } 445 } 446 KEG_UNLOCK(keg); 447 } 448 449 static void 450 zone_timeout(uma_zone_t zone) 451 { 452 453 zone_foreach_keg(zone, &keg_timeout); 454 } 455 456 /* 457 * Allocate and zero fill the next sized hash table from the appropriate 458 * backing store. 459 * 460 * Arguments: 461 * hash A new hash structure with the old hash size in uh_hashsize 462 * 463 * Returns: 464 * 1 on sucess and 0 on failure. 465 */ 466 static int 467 hash_alloc(struct uma_hash *hash) 468 { 469 int oldsize; 470 int alloc; 471 472 oldsize = hash->uh_hashsize; 473 474 /* We're just going to go to a power of two greater */ 475 if (oldsize) { 476 hash->uh_hashsize = oldsize * 2; 477 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 478 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 479 M_UMAHASH, M_NOWAIT); 480 } else { 481 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 482 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 483 M_WAITOK); 484 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 485 } 486 if (hash->uh_slab_hash) { 487 bzero(hash->uh_slab_hash, alloc); 488 hash->uh_hashmask = hash->uh_hashsize - 1; 489 return (1); 490 } 491 492 return (0); 493 } 494 495 /* 496 * Expands the hash table for HASH zones. This is done from zone_timeout 497 * to reduce collisions. This must not be done in the regular allocation 498 * path, otherwise, we can recurse on the vm while allocating pages. 499 * 500 * Arguments: 501 * oldhash The hash you want to expand 502 * newhash The hash structure for the new table 503 * 504 * Returns: 505 * Nothing 506 * 507 * Discussion: 508 */ 509 static int 510 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 511 { 512 uma_slab_t slab; 513 int hval; 514 int i; 515 516 if (!newhash->uh_slab_hash) 517 return (0); 518 519 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 520 return (0); 521 522 /* 523 * I need to investigate hash algorithms for resizing without a 524 * full rehash. 525 */ 526 527 for (i = 0; i < oldhash->uh_hashsize; i++) 528 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 529 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 530 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 531 hval = UMA_HASH(newhash, slab->us_data); 532 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 533 slab, us_hlink); 534 } 535 536 return (1); 537 } 538 539 /* 540 * Free the hash bucket to the appropriate backing store. 541 * 542 * Arguments: 543 * slab_hash The hash bucket we're freeing 544 * hashsize The number of entries in that hash bucket 545 * 546 * Returns: 547 * Nothing 548 */ 549 static void 550 hash_free(struct uma_hash *hash) 551 { 552 if (hash->uh_slab_hash == NULL) 553 return; 554 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 555 zone_free_item(hashzone, 556 hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE); 557 else 558 free(hash->uh_slab_hash, M_UMAHASH); 559 } 560 561 /* 562 * Frees all outstanding items in a bucket 563 * 564 * Arguments: 565 * zone The zone to free to, must be unlocked. 566 * bucket The free/alloc bucket with items, cpu queue must be locked. 567 * 568 * Returns: 569 * Nothing 570 */ 571 572 static void 573 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 574 { 575 void *item; 576 577 if (bucket == NULL) 578 return; 579 580 while (bucket->ub_cnt > 0) { 581 bucket->ub_cnt--; 582 item = bucket->ub_bucket[bucket->ub_cnt]; 583 #ifdef INVARIANTS 584 bucket->ub_bucket[bucket->ub_cnt] = NULL; 585 KASSERT(item != NULL, 586 ("bucket_drain: botched ptr, item is NULL")); 587 #endif 588 zone_free_item(zone, item, NULL, SKIP_DTOR, 0); 589 } 590 } 591 592 /* 593 * Drains the per cpu caches for a zone. 594 * 595 * NOTE: This may only be called while the zone is being turn down, and not 596 * during normal operation. This is necessary in order that we do not have 597 * to migrate CPUs to drain the per-CPU caches. 598 * 599 * Arguments: 600 * zone The zone to drain, must be unlocked. 601 * 602 * Returns: 603 * Nothing 604 */ 605 static void 606 cache_drain(uma_zone_t zone) 607 { 608 uma_cache_t cache; 609 int cpu; 610 611 /* 612 * XXX: It is safe to not lock the per-CPU caches, because we're 613 * tearing down the zone anyway. I.e., there will be no further use 614 * of the caches at this point. 615 * 616 * XXX: It would good to be able to assert that the zone is being 617 * torn down to prevent improper use of cache_drain(). 618 * 619 * XXX: We lock the zone before passing into bucket_cache_drain() as 620 * it is used elsewhere. Should the tear-down path be made special 621 * there in some form? 622 */ 623 for (cpu = 0; cpu <= mp_maxid; cpu++) { 624 if (CPU_ABSENT(cpu)) 625 continue; 626 cache = &zone->uz_cpu[cpu]; 627 bucket_drain(zone, cache->uc_allocbucket); 628 bucket_drain(zone, cache->uc_freebucket); 629 if (cache->uc_allocbucket != NULL) 630 bucket_free(cache->uc_allocbucket); 631 if (cache->uc_freebucket != NULL) 632 bucket_free(cache->uc_freebucket); 633 cache->uc_allocbucket = cache->uc_freebucket = NULL; 634 } 635 ZONE_LOCK(zone); 636 bucket_cache_drain(zone); 637 ZONE_UNLOCK(zone); 638 } 639 640 /* 641 * Drain the cached buckets from a zone. Expects a locked zone on entry. 642 */ 643 static void 644 bucket_cache_drain(uma_zone_t zone) 645 { 646 uma_bucket_t bucket; 647 648 /* 649 * Drain the bucket queues and free the buckets, we just keep two per 650 * cpu (alloc/free). 651 */ 652 while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) { 653 LIST_REMOVE(bucket, ub_link); 654 ZONE_UNLOCK(zone); 655 bucket_drain(zone, bucket); 656 bucket_free(bucket); 657 ZONE_LOCK(zone); 658 } 659 660 /* Now we do the free queue.. */ 661 while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { 662 LIST_REMOVE(bucket, ub_link); 663 bucket_free(bucket); 664 } 665 } 666 667 /* 668 * Frees pages from a keg back to the system. This is done on demand from 669 * the pageout daemon. 670 * 671 * Returns nothing. 672 */ 673 static void 674 keg_drain(uma_keg_t keg) 675 { 676 struct slabhead freeslabs = { 0 }; 677 uma_slab_t slab; 678 uma_slab_t n; 679 u_int8_t flags; 680 u_int8_t *mem; 681 int i; 682 683 /* 684 * We don't want to take pages from statically allocated kegs at this 685 * time 686 */ 687 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 688 return; 689 690 #ifdef UMA_DEBUG 691 printf("%s free items: %u\n", keg->uk_name, keg->uk_free); 692 #endif 693 KEG_LOCK(keg); 694 if (keg->uk_free == 0) 695 goto finished; 696 697 slab = LIST_FIRST(&keg->uk_free_slab); 698 while (slab) { 699 n = LIST_NEXT(slab, us_link); 700 701 /* We have no where to free these to */ 702 if (slab->us_flags & UMA_SLAB_BOOT) { 703 slab = n; 704 continue; 705 } 706 707 LIST_REMOVE(slab, us_link); 708 keg->uk_pages -= keg->uk_ppera; 709 keg->uk_free -= keg->uk_ipers; 710 711 if (keg->uk_flags & UMA_ZONE_HASH) 712 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); 713 714 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 715 716 slab = n; 717 } 718 finished: 719 KEG_UNLOCK(keg); 720 721 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 722 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 723 if (keg->uk_fini) 724 for (i = 0; i < keg->uk_ipers; i++) 725 keg->uk_fini( 726 slab->us_data + (keg->uk_rsize * i), 727 keg->uk_size); 728 flags = slab->us_flags; 729 mem = slab->us_data; 730 731 if (keg->uk_flags & UMA_ZONE_VTOSLAB) { 732 vm_object_t obj; 733 734 if (flags & UMA_SLAB_KMEM) 735 obj = kmem_object; 736 else if (flags & UMA_SLAB_KERNEL) 737 obj = kernel_object; 738 else 739 obj = NULL; 740 for (i = 0; i < keg->uk_ppera; i++) 741 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE), 742 obj); 743 } 744 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 745 zone_free_item(keg->uk_slabzone, slab, NULL, 746 SKIP_NONE, ZFREE_STATFREE); 747 #ifdef UMA_DEBUG 748 printf("%s: Returning %d bytes.\n", 749 keg->uk_name, UMA_SLAB_SIZE * keg->uk_ppera); 750 #endif 751 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags); 752 } 753 } 754 755 static void 756 zone_drain_wait(uma_zone_t zone, int waitok) 757 { 758 759 /* 760 * Set draining to interlock with zone_dtor() so we can release our 761 * locks as we go. Only dtor() should do a WAITOK call since it 762 * is the only call that knows the structure will still be available 763 * when it wakes up. 764 */ 765 ZONE_LOCK(zone); 766 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 767 if (waitok == M_NOWAIT) 768 goto out; 769 mtx_unlock(&uma_mtx); 770 msleep(zone, zone->uz_lock, PVM, "zonedrain", 1); 771 mtx_lock(&uma_mtx); 772 } 773 zone->uz_flags |= UMA_ZFLAG_DRAINING; 774 bucket_cache_drain(zone); 775 ZONE_UNLOCK(zone); 776 /* 777 * The DRAINING flag protects us from being freed while 778 * we're running. Normally the uma_mtx would protect us but we 779 * must be able to release and acquire the right lock for each keg. 780 */ 781 zone_foreach_keg(zone, &keg_drain); 782 ZONE_LOCK(zone); 783 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 784 wakeup(zone); 785 out: 786 ZONE_UNLOCK(zone); 787 } 788 789 void 790 zone_drain(uma_zone_t zone) 791 { 792 793 zone_drain_wait(zone, M_NOWAIT); 794 } 795 796 /* 797 * Allocate a new slab for a keg. This does not insert the slab onto a list. 798 * 799 * Arguments: 800 * wait Shall we wait? 801 * 802 * Returns: 803 * The slab that was allocated or NULL if there is no memory and the 804 * caller specified M_NOWAIT. 805 */ 806 static uma_slab_t 807 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) 808 { 809 uma_slabrefcnt_t slabref; 810 uma_alloc allocf; 811 uma_slab_t slab; 812 u_int8_t *mem; 813 u_int8_t flags; 814 int i; 815 816 mtx_assert(&keg->uk_lock, MA_OWNED); 817 slab = NULL; 818 819 #ifdef UMA_DEBUG 820 printf("slab_zalloc: Allocating a new slab for %s\n", keg->uk_name); 821 #endif 822 allocf = keg->uk_allocf; 823 KEG_UNLOCK(keg); 824 825 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 826 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); 827 if (slab == NULL) { 828 KEG_LOCK(keg); 829 return NULL; 830 } 831 } 832 833 /* 834 * This reproduces the old vm_zone behavior of zero filling pages the 835 * first time they are added to a zone. 836 * 837 * Malloced items are zeroed in uma_zalloc. 838 */ 839 840 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 841 wait |= M_ZERO; 842 else 843 wait &= ~M_ZERO; 844 845 /* zone is passed for legacy reasons. */ 846 mem = allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE, &flags, wait); 847 if (mem == NULL) { 848 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 849 zone_free_item(keg->uk_slabzone, slab, NULL, 850 SKIP_NONE, ZFREE_STATFREE); 851 KEG_LOCK(keg); 852 return (NULL); 853 } 854 855 /* Point the slab into the allocated memory */ 856 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 857 slab = (uma_slab_t )(mem + keg->uk_pgoff); 858 859 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 860 for (i = 0; i < keg->uk_ppera; i++) 861 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 862 863 slab->us_keg = keg; 864 slab->us_data = mem; 865 slab->us_freecount = keg->uk_ipers; 866 slab->us_firstfree = 0; 867 slab->us_flags = flags; 868 869 if (keg->uk_flags & UMA_ZONE_REFCNT) { 870 slabref = (uma_slabrefcnt_t)slab; 871 for (i = 0; i < keg->uk_ipers; i++) { 872 slabref->us_freelist[i].us_refcnt = 0; 873 slabref->us_freelist[i].us_item = i+1; 874 } 875 } else { 876 for (i = 0; i < keg->uk_ipers; i++) 877 slab->us_freelist[i].us_item = i+1; 878 } 879 880 if (keg->uk_init != NULL) { 881 for (i = 0; i < keg->uk_ipers; i++) 882 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 883 keg->uk_size, wait) != 0) 884 break; 885 if (i != keg->uk_ipers) { 886 if (keg->uk_fini != NULL) { 887 for (i--; i > -1; i--) 888 keg->uk_fini(slab->us_data + 889 (keg->uk_rsize * i), 890 keg->uk_size); 891 } 892 if (keg->uk_flags & UMA_ZONE_VTOSLAB) { 893 vm_object_t obj; 894 895 if (flags & UMA_SLAB_KMEM) 896 obj = kmem_object; 897 else if (flags & UMA_SLAB_KERNEL) 898 obj = kernel_object; 899 else 900 obj = NULL; 901 for (i = 0; i < keg->uk_ppera; i++) 902 vsetobj((vm_offset_t)mem + 903 (i * PAGE_SIZE), obj); 904 } 905 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 906 zone_free_item(keg->uk_slabzone, slab, 907 NULL, SKIP_NONE, ZFREE_STATFREE); 908 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, 909 flags); 910 KEG_LOCK(keg); 911 return (NULL); 912 } 913 } 914 KEG_LOCK(keg); 915 916 if (keg->uk_flags & UMA_ZONE_HASH) 917 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 918 919 keg->uk_pages += keg->uk_ppera; 920 keg->uk_free += keg->uk_ipers; 921 922 return (slab); 923 } 924 925 /* 926 * This function is intended to be used early on in place of page_alloc() so 927 * that we may use the boot time page cache to satisfy allocations before 928 * the VM is ready. 929 */ 930 static void * 931 startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait) 932 { 933 uma_keg_t keg; 934 uma_slab_t tmps; 935 936 keg = zone_first_keg(zone); 937 938 /* 939 * Check our small startup cache to see if it has pages remaining. 940 */ 941 mtx_lock(&uma_boot_pages_mtx); 942 if ((tmps = LIST_FIRST(&uma_boot_pages)) != NULL) { 943 LIST_REMOVE(tmps, us_link); 944 mtx_unlock(&uma_boot_pages_mtx); 945 *pflag = tmps->us_flags; 946 return (tmps->us_data); 947 } 948 mtx_unlock(&uma_boot_pages_mtx); 949 if (booted == 0) 950 panic("UMA: Increase vm.boot_pages"); 951 /* 952 * Now that we've booted reset these users to their real allocator. 953 */ 954 #ifdef UMA_MD_SMALL_ALLOC 955 keg->uk_allocf = uma_small_alloc; 956 #else 957 keg->uk_allocf = page_alloc; 958 #endif 959 return keg->uk_allocf(zone, bytes, pflag, wait); 960 } 961 962 /* 963 * Allocates a number of pages from the system 964 * 965 * Arguments: 966 * bytes The number of bytes requested 967 * wait Shall we wait? 968 * 969 * Returns: 970 * A pointer to the alloced memory or possibly 971 * NULL if M_NOWAIT is set. 972 */ 973 static void * 974 page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait) 975 { 976 void *p; /* Returned page */ 977 978 *pflag = UMA_SLAB_KMEM; 979 p = (void *) kmem_malloc(kmem_map, bytes, wait); 980 981 return (p); 982 } 983 984 /* 985 * Allocates a number of pages from within an object 986 * 987 * Arguments: 988 * bytes The number of bytes requested 989 * wait Shall we wait? 990 * 991 * Returns: 992 * A pointer to the alloced memory or possibly 993 * NULL if M_NOWAIT is set. 994 */ 995 static void * 996 obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 997 { 998 vm_object_t object; 999 vm_offset_t retkva, zkva; 1000 vm_page_t p; 1001 int pages, startpages; 1002 uma_keg_t keg; 1003 1004 keg = zone_first_keg(zone); 1005 object = keg->uk_obj; 1006 retkva = 0; 1007 1008 /* 1009 * This looks a little weird since we're getting one page at a time. 1010 */ 1011 VM_OBJECT_LOCK(object); 1012 p = TAILQ_LAST(&object->memq, pglist); 1013 pages = p != NULL ? p->pindex + 1 : 0; 1014 startpages = pages; 1015 zkva = keg->uk_kva + pages * PAGE_SIZE; 1016 for (; bytes > 0; bytes -= PAGE_SIZE) { 1017 p = vm_page_alloc(object, pages, 1018 VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED); 1019 if (p == NULL) { 1020 if (pages != startpages) 1021 pmap_qremove(retkva, pages - startpages); 1022 while (pages != startpages) { 1023 pages--; 1024 p = TAILQ_LAST(&object->memq, pglist); 1025 vm_page_lock_queues(); 1026 vm_page_unwire(p, 0); 1027 vm_page_free(p); 1028 vm_page_unlock_queues(); 1029 } 1030 retkva = 0; 1031 goto done; 1032 } 1033 pmap_qenter(zkva, &p, 1); 1034 if (retkva == 0) 1035 retkva = zkva; 1036 zkva += PAGE_SIZE; 1037 pages += 1; 1038 } 1039 done: 1040 VM_OBJECT_UNLOCK(object); 1041 *flags = UMA_SLAB_PRIV; 1042 1043 return ((void *)retkva); 1044 } 1045 1046 /* 1047 * Frees a number of pages to the system 1048 * 1049 * Arguments: 1050 * mem A pointer to the memory to be freed 1051 * size The size of the memory being freed 1052 * flags The original p->us_flags field 1053 * 1054 * Returns: 1055 * Nothing 1056 */ 1057 static void 1058 page_free(void *mem, int size, u_int8_t flags) 1059 { 1060 vm_map_t map; 1061 1062 if (flags & UMA_SLAB_KMEM) 1063 map = kmem_map; 1064 else 1065 panic("UMA: page_free used with invalid flags %d\n", flags); 1066 1067 kmem_free(map, (vm_offset_t)mem, size); 1068 } 1069 1070 /* 1071 * Zero fill initializer 1072 * 1073 * Arguments/Returns follow uma_init specifications 1074 */ 1075 static int 1076 zero_init(void *mem, int size, int flags) 1077 { 1078 bzero(mem, size); 1079 return (0); 1080 } 1081 1082 /* 1083 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1084 * 1085 * Arguments 1086 * keg The zone we should initialize 1087 * 1088 * Returns 1089 * Nothing 1090 */ 1091 static void 1092 keg_small_init(uma_keg_t keg) 1093 { 1094 u_int rsize; 1095 u_int memused; 1096 u_int wastedspace; 1097 u_int shsize; 1098 1099 KASSERT(keg != NULL, ("Keg is null in keg_small_init")); 1100 rsize = keg->uk_size; 1101 1102 if (rsize < UMA_SMALLEST_UNIT) 1103 rsize = UMA_SMALLEST_UNIT; 1104 if (rsize & keg->uk_align) 1105 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1106 1107 keg->uk_rsize = rsize; 1108 keg->uk_ppera = 1; 1109 1110 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1111 rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */ 1112 shsize = sizeof(struct uma_slab_refcnt); 1113 } else { 1114 rsize += UMA_FRITM_SZ; /* Account for linkage */ 1115 shsize = sizeof(struct uma_slab); 1116 } 1117 1118 keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize; 1119 KASSERT(keg->uk_ipers != 0, ("keg_small_init: ipers is 0")); 1120 memused = keg->uk_ipers * rsize + shsize; 1121 wastedspace = UMA_SLAB_SIZE - memused; 1122 1123 /* 1124 * We can't do OFFPAGE if we're internal or if we've been 1125 * asked to not go to the VM for buckets. If we do this we 1126 * may end up going to the VM (kmem_map) for slabs which we 1127 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a 1128 * result of UMA_ZONE_VM, which clearly forbids it. 1129 */ 1130 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1131 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1132 return; 1133 1134 if ((wastedspace >= UMA_MAX_WASTE) && 1135 (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) { 1136 keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize; 1137 KASSERT(keg->uk_ipers <= 255, 1138 ("keg_small_init: keg->uk_ipers too high!")); 1139 #ifdef UMA_DEBUG 1140 printf("UMA decided we need offpage slab headers for " 1141 "keg: %s, calculated wastedspace = %d, " 1142 "maximum wasted space allowed = %d, " 1143 "calculated ipers = %d, " 1144 "new wasted space = %d\n", keg->uk_name, wastedspace, 1145 UMA_MAX_WASTE, keg->uk_ipers, 1146 UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize); 1147 #endif 1148 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1149 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1150 keg->uk_flags |= UMA_ZONE_HASH; 1151 } 1152 } 1153 1154 /* 1155 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1156 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1157 * more complicated. 1158 * 1159 * Arguments 1160 * keg The keg we should initialize 1161 * 1162 * Returns 1163 * Nothing 1164 */ 1165 static void 1166 keg_large_init(uma_keg_t keg) 1167 { 1168 int pages; 1169 1170 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1171 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1172 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1173 1174 pages = keg->uk_size / UMA_SLAB_SIZE; 1175 1176 /* Account for remainder */ 1177 if ((pages * UMA_SLAB_SIZE) < keg->uk_size) 1178 pages++; 1179 1180 keg->uk_ppera = pages; 1181 keg->uk_ipers = 1; 1182 1183 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1184 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1185 keg->uk_flags |= UMA_ZONE_HASH; 1186 1187 keg->uk_rsize = keg->uk_size; 1188 } 1189 1190 static void 1191 keg_cachespread_init(uma_keg_t keg) 1192 { 1193 int alignsize; 1194 int trailer; 1195 int pages; 1196 int rsize; 1197 1198 alignsize = keg->uk_align + 1; 1199 rsize = keg->uk_size; 1200 /* 1201 * We want one item to start on every align boundary in a page. To 1202 * do this we will span pages. We will also extend the item by the 1203 * size of align if it is an even multiple of align. Otherwise, it 1204 * would fall on the same boundary every time. 1205 */ 1206 if (rsize & keg->uk_align) 1207 rsize = (rsize & ~keg->uk_align) + alignsize; 1208 if ((rsize & alignsize) == 0) 1209 rsize += alignsize; 1210 trailer = rsize - keg->uk_size; 1211 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1212 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1213 keg->uk_rsize = rsize; 1214 keg->uk_ppera = pages; 1215 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1216 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1217 KASSERT(keg->uk_ipers <= uma_max_ipers, 1218 ("keg_small_init: keg->uk_ipers too high(%d) increase max_ipers", 1219 keg->uk_ipers)); 1220 } 1221 1222 /* 1223 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1224 * the keg onto the global keg list. 1225 * 1226 * Arguments/Returns follow uma_ctor specifications 1227 * udata Actually uma_kctor_args 1228 */ 1229 static int 1230 keg_ctor(void *mem, int size, void *udata, int flags) 1231 { 1232 struct uma_kctor_args *arg = udata; 1233 uma_keg_t keg = mem; 1234 uma_zone_t zone; 1235 1236 bzero(keg, size); 1237 keg->uk_size = arg->size; 1238 keg->uk_init = arg->uminit; 1239 keg->uk_fini = arg->fini; 1240 keg->uk_align = arg->align; 1241 keg->uk_free = 0; 1242 keg->uk_pages = 0; 1243 keg->uk_flags = arg->flags; 1244 keg->uk_allocf = page_alloc; 1245 keg->uk_freef = page_free; 1246 keg->uk_recurse = 0; 1247 keg->uk_slabzone = NULL; 1248 1249 /* 1250 * The master zone is passed to us at keg-creation time. 1251 */ 1252 zone = arg->zone; 1253 keg->uk_name = zone->uz_name; 1254 1255 if (arg->flags & UMA_ZONE_VM) 1256 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1257 1258 if (arg->flags & UMA_ZONE_ZINIT) 1259 keg->uk_init = zero_init; 1260 1261 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC) 1262 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1263 1264 /* 1265 * The +UMA_FRITM_SZ added to uk_size is to account for the 1266 * linkage that is added to the size in keg_small_init(). If 1267 * we don't account for this here then we may end up in 1268 * keg_small_init() with a calculated 'ipers' of 0. 1269 */ 1270 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1271 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) 1272 keg_cachespread_init(keg); 1273 else if ((keg->uk_size+UMA_FRITMREF_SZ) > 1274 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt))) 1275 keg_large_init(keg); 1276 else 1277 keg_small_init(keg); 1278 } else { 1279 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) 1280 keg_cachespread_init(keg); 1281 else if ((keg->uk_size+UMA_FRITM_SZ) > 1282 (UMA_SLAB_SIZE - sizeof(struct uma_slab))) 1283 keg_large_init(keg); 1284 else 1285 keg_small_init(keg); 1286 } 1287 1288 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1289 if (keg->uk_flags & UMA_ZONE_REFCNT) 1290 keg->uk_slabzone = slabrefzone; 1291 else 1292 keg->uk_slabzone = slabzone; 1293 } 1294 1295 /* 1296 * If we haven't booted yet we need allocations to go through the 1297 * startup cache until the vm is ready. 1298 */ 1299 if (keg->uk_ppera == 1) { 1300 #ifdef UMA_MD_SMALL_ALLOC 1301 keg->uk_allocf = uma_small_alloc; 1302 keg->uk_freef = uma_small_free; 1303 #endif 1304 if (booted == 0) 1305 keg->uk_allocf = startup_alloc; 1306 } 1307 1308 /* 1309 * Initialize keg's lock (shared among zones). 1310 */ 1311 if (arg->flags & UMA_ZONE_MTXCLASS) 1312 KEG_LOCK_INIT(keg, 1); 1313 else 1314 KEG_LOCK_INIT(keg, 0); 1315 1316 /* 1317 * If we're putting the slab header in the actual page we need to 1318 * figure out where in each page it goes. This calculates a right 1319 * justified offset into the memory on an ALIGN_PTR boundary. 1320 */ 1321 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1322 u_int totsize; 1323 1324 /* Size of the slab struct and free list */ 1325 if (keg->uk_flags & UMA_ZONE_REFCNT) 1326 totsize = sizeof(struct uma_slab_refcnt) + 1327 keg->uk_ipers * UMA_FRITMREF_SZ; 1328 else 1329 totsize = sizeof(struct uma_slab) + 1330 keg->uk_ipers * UMA_FRITM_SZ; 1331 1332 if (totsize & UMA_ALIGN_PTR) 1333 totsize = (totsize & ~UMA_ALIGN_PTR) + 1334 (UMA_ALIGN_PTR + 1); 1335 keg->uk_pgoff = UMA_SLAB_SIZE - totsize; 1336 1337 if (keg->uk_flags & UMA_ZONE_REFCNT) 1338 totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt) 1339 + keg->uk_ipers * UMA_FRITMREF_SZ; 1340 else 1341 totsize = keg->uk_pgoff + sizeof(struct uma_slab) 1342 + keg->uk_ipers * UMA_FRITM_SZ; 1343 1344 /* 1345 * The only way the following is possible is if with our 1346 * UMA_ALIGN_PTR adjustments we are now bigger than 1347 * UMA_SLAB_SIZE. I haven't checked whether this is 1348 * mathematically possible for all cases, so we make 1349 * sure here anyway. 1350 */ 1351 if (totsize > UMA_SLAB_SIZE) { 1352 printf("zone %s ipers %d rsize %d size %d\n", 1353 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1354 keg->uk_size); 1355 panic("UMA slab won't fit.\n"); 1356 } 1357 } 1358 1359 if (keg->uk_flags & UMA_ZONE_HASH) 1360 hash_alloc(&keg->uk_hash); 1361 1362 #ifdef UMA_DEBUG 1363 printf("UMA: %s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n", 1364 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags, 1365 keg->uk_ipers, keg->uk_ppera, 1366 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free); 1367 #endif 1368 1369 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1370 1371 mtx_lock(&uma_mtx); 1372 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1373 mtx_unlock(&uma_mtx); 1374 return (0); 1375 } 1376 1377 /* 1378 * Zone header ctor. This initializes all fields, locks, etc. 1379 * 1380 * Arguments/Returns follow uma_ctor specifications 1381 * udata Actually uma_zctor_args 1382 */ 1383 static int 1384 zone_ctor(void *mem, int size, void *udata, int flags) 1385 { 1386 struct uma_zctor_args *arg = udata; 1387 uma_zone_t zone = mem; 1388 uma_zone_t z; 1389 uma_keg_t keg; 1390 1391 bzero(zone, size); 1392 zone->uz_name = arg->name; 1393 zone->uz_ctor = arg->ctor; 1394 zone->uz_dtor = arg->dtor; 1395 zone->uz_slab = zone_fetch_slab; 1396 zone->uz_init = NULL; 1397 zone->uz_fini = NULL; 1398 zone->uz_allocs = 0; 1399 zone->uz_frees = 0; 1400 zone->uz_fails = 0; 1401 zone->uz_fills = zone->uz_count = 0; 1402 zone->uz_flags = 0; 1403 keg = arg->keg; 1404 1405 if (arg->flags & UMA_ZONE_SECONDARY) { 1406 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1407 zone->uz_init = arg->uminit; 1408 zone->uz_fini = arg->fini; 1409 zone->uz_lock = &keg->uk_lock; 1410 zone->uz_flags |= UMA_ZONE_SECONDARY; 1411 mtx_lock(&uma_mtx); 1412 ZONE_LOCK(zone); 1413 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1414 if (LIST_NEXT(z, uz_link) == NULL) { 1415 LIST_INSERT_AFTER(z, zone, uz_link); 1416 break; 1417 } 1418 } 1419 ZONE_UNLOCK(zone); 1420 mtx_unlock(&uma_mtx); 1421 } else if (keg == NULL) { 1422 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1423 arg->align, arg->flags)) == NULL) 1424 return (ENOMEM); 1425 } else { 1426 struct uma_kctor_args karg; 1427 int error; 1428 1429 /* We should only be here from uma_startup() */ 1430 karg.size = arg->size; 1431 karg.uminit = arg->uminit; 1432 karg.fini = arg->fini; 1433 karg.align = arg->align; 1434 karg.flags = arg->flags; 1435 karg.zone = zone; 1436 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1437 flags); 1438 if (error) 1439 return (error); 1440 } 1441 /* 1442 * Link in the first keg. 1443 */ 1444 zone->uz_klink.kl_keg = keg; 1445 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1446 zone->uz_lock = &keg->uk_lock; 1447 zone->uz_size = keg->uk_size; 1448 zone->uz_flags |= (keg->uk_flags & 1449 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1450 1451 /* 1452 * Some internal zones don't have room allocated for the per cpu 1453 * caches. If we're internal, bail out here. 1454 */ 1455 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1456 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1457 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1458 return (0); 1459 } 1460 1461 if (keg->uk_flags & UMA_ZONE_MAXBUCKET) 1462 zone->uz_count = BUCKET_MAX; 1463 else if (keg->uk_ipers <= BUCKET_MAX) 1464 zone->uz_count = keg->uk_ipers; 1465 else 1466 zone->uz_count = BUCKET_MAX; 1467 return (0); 1468 } 1469 1470 /* 1471 * Keg header dtor. This frees all data, destroys locks, frees the hash 1472 * table and removes the keg from the global list. 1473 * 1474 * Arguments/Returns follow uma_dtor specifications 1475 * udata unused 1476 */ 1477 static void 1478 keg_dtor(void *arg, int size, void *udata) 1479 { 1480 uma_keg_t keg; 1481 1482 keg = (uma_keg_t)arg; 1483 KEG_LOCK(keg); 1484 if (keg->uk_free != 0) { 1485 printf("Freed UMA keg was not empty (%d items). " 1486 " Lost %d pages of memory.\n", 1487 keg->uk_free, keg->uk_pages); 1488 } 1489 KEG_UNLOCK(keg); 1490 1491 hash_free(&keg->uk_hash); 1492 1493 KEG_LOCK_FINI(keg); 1494 } 1495 1496 /* 1497 * Zone header dtor. 1498 * 1499 * Arguments/Returns follow uma_dtor specifications 1500 * udata unused 1501 */ 1502 static void 1503 zone_dtor(void *arg, int size, void *udata) 1504 { 1505 uma_klink_t klink; 1506 uma_zone_t zone; 1507 uma_keg_t keg; 1508 1509 zone = (uma_zone_t)arg; 1510 keg = zone_first_keg(zone); 1511 1512 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1513 cache_drain(zone); 1514 1515 mtx_lock(&uma_mtx); 1516 LIST_REMOVE(zone, uz_link); 1517 mtx_unlock(&uma_mtx); 1518 /* 1519 * XXX there are some races here where 1520 * the zone can be drained but zone lock 1521 * released and then refilled before we 1522 * remove it... we dont care for now 1523 */ 1524 zone_drain_wait(zone, M_WAITOK); 1525 /* 1526 * Unlink all of our kegs. 1527 */ 1528 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1529 klink->kl_keg = NULL; 1530 LIST_REMOVE(klink, kl_link); 1531 if (klink == &zone->uz_klink) 1532 continue; 1533 free(klink, M_TEMP); 1534 } 1535 /* 1536 * We only destroy kegs from non secondary zones. 1537 */ 1538 if ((zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1539 mtx_lock(&uma_mtx); 1540 LIST_REMOVE(keg, uk_link); 1541 mtx_unlock(&uma_mtx); 1542 zone_free_item(kegs, keg, NULL, SKIP_NONE, 1543 ZFREE_STATFREE); 1544 } 1545 } 1546 1547 /* 1548 * Traverses every zone in the system and calls a callback 1549 * 1550 * Arguments: 1551 * zfunc A pointer to a function which accepts a zone 1552 * as an argument. 1553 * 1554 * Returns: 1555 * Nothing 1556 */ 1557 static void 1558 zone_foreach(void (*zfunc)(uma_zone_t)) 1559 { 1560 uma_keg_t keg; 1561 uma_zone_t zone; 1562 1563 mtx_lock(&uma_mtx); 1564 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1565 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1566 zfunc(zone); 1567 } 1568 mtx_unlock(&uma_mtx); 1569 } 1570 1571 /* Public functions */ 1572 /* See uma.h */ 1573 void 1574 uma_startup(void *bootmem, int boot_pages) 1575 { 1576 struct uma_zctor_args args; 1577 uma_slab_t slab; 1578 u_int slabsize; 1579 u_int objsize, totsize, wsize; 1580 int i; 1581 1582 #ifdef UMA_DEBUG 1583 printf("Creating uma keg headers zone and keg.\n"); 1584 #endif 1585 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF); 1586 1587 /* 1588 * Figure out the maximum number of items-per-slab we'll have if 1589 * we're using the OFFPAGE slab header to track free items, given 1590 * all possible object sizes and the maximum desired wastage 1591 * (UMA_MAX_WASTE). 1592 * 1593 * We iterate until we find an object size for 1594 * which the calculated wastage in keg_small_init() will be 1595 * enough to warrant OFFPAGE. Since wastedspace versus objsize 1596 * is an overall increasing see-saw function, we find the smallest 1597 * objsize such that the wastage is always acceptable for objects 1598 * with that objsize or smaller. Since a smaller objsize always 1599 * generates a larger possible uma_max_ipers, we use this computed 1600 * objsize to calculate the largest ipers possible. Since the 1601 * ipers calculated for OFFPAGE slab headers is always larger than 1602 * the ipers initially calculated in keg_small_init(), we use 1603 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to 1604 * obtain the maximum ipers possible for offpage slab headers. 1605 * 1606 * It should be noted that ipers versus objsize is an inversly 1607 * proportional function which drops off rather quickly so as 1608 * long as our UMA_MAX_WASTE is such that the objsize we calculate 1609 * falls into the portion of the inverse relation AFTER the steep 1610 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386). 1611 * 1612 * Note that we have 8-bits (1 byte) to use as a freelist index 1613 * inside the actual slab header itself and this is enough to 1614 * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized 1615 * object with offpage slab header would have ipers = 1616 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is 1617 * 1 greater than what our byte-integer freelist index can 1618 * accomodate, but we know that this situation never occurs as 1619 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate 1620 * that we need to go to offpage slab headers. Or, if we do, 1621 * then we trap that condition below and panic in the INVARIANTS case. 1622 */ 1623 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE; 1624 totsize = wsize; 1625 objsize = UMA_SMALLEST_UNIT; 1626 while (totsize >= wsize) { 1627 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) / 1628 (objsize + UMA_FRITM_SZ); 1629 totsize *= (UMA_FRITM_SZ + objsize); 1630 objsize++; 1631 } 1632 if (objsize > UMA_SMALLEST_UNIT) 1633 objsize--; 1634 uma_max_ipers = MAX(UMA_SLAB_SIZE / objsize, 64); 1635 1636 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE; 1637 totsize = wsize; 1638 objsize = UMA_SMALLEST_UNIT; 1639 while (totsize >= wsize) { 1640 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) / 1641 (objsize + UMA_FRITMREF_SZ); 1642 totsize *= (UMA_FRITMREF_SZ + objsize); 1643 objsize++; 1644 } 1645 if (objsize > UMA_SMALLEST_UNIT) 1646 objsize--; 1647 uma_max_ipers_ref = MAX(UMA_SLAB_SIZE / objsize, 64); 1648 1649 KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255), 1650 ("uma_startup: calculated uma_max_ipers values too large!")); 1651 1652 #ifdef UMA_DEBUG 1653 printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers); 1654 printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n", 1655 uma_max_ipers_ref); 1656 #endif 1657 1658 /* "manually" create the initial zone */ 1659 args.name = "UMA Kegs"; 1660 args.size = sizeof(struct uma_keg); 1661 args.ctor = keg_ctor; 1662 args.dtor = keg_dtor; 1663 args.uminit = zero_init; 1664 args.fini = NULL; 1665 args.keg = &masterkeg; 1666 args.align = 32 - 1; 1667 args.flags = UMA_ZFLAG_INTERNAL; 1668 /* The initial zone has no Per cpu queues so it's smaller */ 1669 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); 1670 1671 #ifdef UMA_DEBUG 1672 printf("Filling boot free list.\n"); 1673 #endif 1674 for (i = 0; i < boot_pages; i++) { 1675 slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE)); 1676 slab->us_data = (u_int8_t *)slab; 1677 slab->us_flags = UMA_SLAB_BOOT; 1678 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); 1679 } 1680 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); 1681 1682 #ifdef UMA_DEBUG 1683 printf("Creating uma zone headers zone and keg.\n"); 1684 #endif 1685 args.name = "UMA Zones"; 1686 args.size = sizeof(struct uma_zone) + 1687 (sizeof(struct uma_cache) * (mp_maxid + 1)); 1688 args.ctor = zone_ctor; 1689 args.dtor = zone_dtor; 1690 args.uminit = zero_init; 1691 args.fini = NULL; 1692 args.keg = NULL; 1693 args.align = 32 - 1; 1694 args.flags = UMA_ZFLAG_INTERNAL; 1695 /* The initial zone has no Per cpu queues so it's smaller */ 1696 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); 1697 1698 #ifdef UMA_DEBUG 1699 printf("Initializing pcpu cache locks.\n"); 1700 #endif 1701 #ifdef UMA_DEBUG 1702 printf("Creating slab and hash zones.\n"); 1703 #endif 1704 1705 /* 1706 * This is the max number of free list items we'll have with 1707 * offpage slabs. 1708 */ 1709 slabsize = uma_max_ipers * UMA_FRITM_SZ; 1710 slabsize += sizeof(struct uma_slab); 1711 1712 /* Now make a zone for slab headers */ 1713 slabzone = uma_zcreate("UMA Slabs", 1714 slabsize, 1715 NULL, NULL, NULL, NULL, 1716 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1717 1718 /* 1719 * We also create a zone for the bigger slabs with reference 1720 * counts in them, to accomodate UMA_ZONE_REFCNT zones. 1721 */ 1722 slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ; 1723 slabsize += sizeof(struct uma_slab_refcnt); 1724 slabrefzone = uma_zcreate("UMA RCntSlabs", 1725 slabsize, 1726 NULL, NULL, NULL, NULL, 1727 UMA_ALIGN_PTR, 1728 UMA_ZFLAG_INTERNAL); 1729 1730 hashzone = uma_zcreate("UMA Hash", 1731 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 1732 NULL, NULL, NULL, NULL, 1733 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1734 1735 bucket_init(); 1736 1737 #if defined(UMA_MD_SMALL_ALLOC) && !defined(UMA_MD_SMALL_ALLOC_NEEDS_VM) 1738 booted = 1; 1739 #endif 1740 1741 #ifdef UMA_DEBUG 1742 printf("UMA startup complete.\n"); 1743 #endif 1744 } 1745 1746 /* see uma.h */ 1747 void 1748 uma_startup2(void) 1749 { 1750 booted = 1; 1751 bucket_enable(); 1752 #ifdef UMA_DEBUG 1753 printf("UMA startup2 complete.\n"); 1754 #endif 1755 } 1756 1757 /* 1758 * Initialize our callout handle 1759 * 1760 */ 1761 1762 static void 1763 uma_startup3(void) 1764 { 1765 #ifdef UMA_DEBUG 1766 printf("Starting callout.\n"); 1767 #endif 1768 callout_init(&uma_callout, CALLOUT_MPSAFE); 1769 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1770 #ifdef UMA_DEBUG 1771 printf("UMA startup3 complete.\n"); 1772 #endif 1773 } 1774 1775 static uma_keg_t 1776 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 1777 int align, u_int32_t flags) 1778 { 1779 struct uma_kctor_args args; 1780 1781 args.size = size; 1782 args.uminit = uminit; 1783 args.fini = fini; 1784 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 1785 args.flags = flags; 1786 args.zone = zone; 1787 return (zone_alloc_item(kegs, &args, M_WAITOK)); 1788 } 1789 1790 /* See uma.h */ 1791 void 1792 uma_set_align(int align) 1793 { 1794 1795 if (align != UMA_ALIGN_CACHE) 1796 uma_align_cache = align; 1797 } 1798 1799 /* See uma.h */ 1800 uma_zone_t 1801 uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 1802 uma_init uminit, uma_fini fini, int align, u_int32_t flags) 1803 1804 { 1805 struct uma_zctor_args args; 1806 1807 /* This stuff is essential for the zone ctor */ 1808 args.name = name; 1809 args.size = size; 1810 args.ctor = ctor; 1811 args.dtor = dtor; 1812 args.uminit = uminit; 1813 args.fini = fini; 1814 args.align = align; 1815 args.flags = flags; 1816 args.keg = NULL; 1817 1818 return (zone_alloc_item(zones, &args, M_WAITOK)); 1819 } 1820 1821 /* See uma.h */ 1822 uma_zone_t 1823 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 1824 uma_init zinit, uma_fini zfini, uma_zone_t master) 1825 { 1826 struct uma_zctor_args args; 1827 uma_keg_t keg; 1828 1829 keg = zone_first_keg(master); 1830 args.name = name; 1831 args.size = keg->uk_size; 1832 args.ctor = ctor; 1833 args.dtor = dtor; 1834 args.uminit = zinit; 1835 args.fini = zfini; 1836 args.align = keg->uk_align; 1837 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 1838 args.keg = keg; 1839 1840 /* XXX Attaches only one keg of potentially many. */ 1841 return (zone_alloc_item(zones, &args, M_WAITOK)); 1842 } 1843 1844 static void 1845 zone_lock_pair(uma_zone_t a, uma_zone_t b) 1846 { 1847 if (a < b) { 1848 ZONE_LOCK(a); 1849 mtx_lock_flags(b->uz_lock, MTX_DUPOK); 1850 } else { 1851 ZONE_LOCK(b); 1852 mtx_lock_flags(a->uz_lock, MTX_DUPOK); 1853 } 1854 } 1855 1856 static void 1857 zone_unlock_pair(uma_zone_t a, uma_zone_t b) 1858 { 1859 1860 ZONE_UNLOCK(a); 1861 ZONE_UNLOCK(b); 1862 } 1863 1864 int 1865 uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 1866 { 1867 uma_klink_t klink; 1868 uma_klink_t kl; 1869 int error; 1870 1871 error = 0; 1872 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 1873 1874 zone_lock_pair(zone, master); 1875 /* 1876 * zone must use vtoslab() to resolve objects and must already be 1877 * a secondary. 1878 */ 1879 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 1880 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 1881 error = EINVAL; 1882 goto out; 1883 } 1884 /* 1885 * The new master must also use vtoslab(). 1886 */ 1887 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 1888 error = EINVAL; 1889 goto out; 1890 } 1891 /* 1892 * Both must either be refcnt, or not be refcnt. 1893 */ 1894 if ((zone->uz_flags & UMA_ZONE_REFCNT) != 1895 (master->uz_flags & UMA_ZONE_REFCNT)) { 1896 error = EINVAL; 1897 goto out; 1898 } 1899 /* 1900 * The underlying object must be the same size. rsize 1901 * may be different. 1902 */ 1903 if (master->uz_size != zone->uz_size) { 1904 error = E2BIG; 1905 goto out; 1906 } 1907 /* 1908 * Put it at the end of the list. 1909 */ 1910 klink->kl_keg = zone_first_keg(master); 1911 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 1912 if (LIST_NEXT(kl, kl_link) == NULL) { 1913 LIST_INSERT_AFTER(kl, klink, kl_link); 1914 break; 1915 } 1916 } 1917 klink = NULL; 1918 zone->uz_flags |= UMA_ZFLAG_MULTI; 1919 zone->uz_slab = zone_fetch_slab_multi; 1920 1921 out: 1922 zone_unlock_pair(zone, master); 1923 if (klink != NULL) 1924 free(klink, M_TEMP); 1925 1926 return (error); 1927 } 1928 1929 1930 /* See uma.h */ 1931 void 1932 uma_zdestroy(uma_zone_t zone) 1933 { 1934 1935 zone_free_item(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE); 1936 } 1937 1938 /* See uma.h */ 1939 void * 1940 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 1941 { 1942 void *item; 1943 uma_cache_t cache; 1944 uma_bucket_t bucket; 1945 int cpu; 1946 1947 /* This is the fast path allocation */ 1948 #ifdef UMA_DEBUG_ALLOC_1 1949 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); 1950 #endif 1951 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread, 1952 zone->uz_name, flags); 1953 1954 if (flags & M_WAITOK) { 1955 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 1956 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 1957 } 1958 1959 /* 1960 * If possible, allocate from the per-CPU cache. There are two 1961 * requirements for safe access to the per-CPU cache: (1) the thread 1962 * accessing the cache must not be preempted or yield during access, 1963 * and (2) the thread must not migrate CPUs without switching which 1964 * cache it accesses. We rely on a critical section to prevent 1965 * preemption and migration. We release the critical section in 1966 * order to acquire the zone mutex if we are unable to allocate from 1967 * the current cache; when we re-acquire the critical section, we 1968 * must detect and handle migration if it has occurred. 1969 */ 1970 zalloc_restart: 1971 critical_enter(); 1972 cpu = curcpu; 1973 cache = &zone->uz_cpu[cpu]; 1974 1975 zalloc_start: 1976 bucket = cache->uc_allocbucket; 1977 1978 if (bucket) { 1979 if (bucket->ub_cnt > 0) { 1980 bucket->ub_cnt--; 1981 item = bucket->ub_bucket[bucket->ub_cnt]; 1982 #ifdef INVARIANTS 1983 bucket->ub_bucket[bucket->ub_cnt] = NULL; 1984 #endif 1985 KASSERT(item != NULL, 1986 ("uma_zalloc: Bucket pointer mangled.")); 1987 cache->uc_allocs++; 1988 critical_exit(); 1989 #ifdef INVARIANTS 1990 ZONE_LOCK(zone); 1991 uma_dbg_alloc(zone, NULL, item); 1992 ZONE_UNLOCK(zone); 1993 #endif 1994 if (zone->uz_ctor != NULL) { 1995 if (zone->uz_ctor(item, zone->uz_size, 1996 udata, flags) != 0) { 1997 zone_free_item(zone, item, udata, 1998 SKIP_DTOR, ZFREE_STATFAIL | 1999 ZFREE_STATFREE); 2000 return (NULL); 2001 } 2002 } 2003 if (flags & M_ZERO) 2004 bzero(item, zone->uz_size); 2005 return (item); 2006 } else if (cache->uc_freebucket) { 2007 /* 2008 * We have run out of items in our allocbucket. 2009 * See if we can switch with our free bucket. 2010 */ 2011 if (cache->uc_freebucket->ub_cnt > 0) { 2012 #ifdef UMA_DEBUG_ALLOC 2013 printf("uma_zalloc: Swapping empty with" 2014 " alloc.\n"); 2015 #endif 2016 bucket = cache->uc_freebucket; 2017 cache->uc_freebucket = cache->uc_allocbucket; 2018 cache->uc_allocbucket = bucket; 2019 2020 goto zalloc_start; 2021 } 2022 } 2023 } 2024 /* 2025 * Attempt to retrieve the item from the per-CPU cache has failed, so 2026 * we must go back to the zone. This requires the zone lock, so we 2027 * must drop the critical section, then re-acquire it when we go back 2028 * to the cache. Since the critical section is released, we may be 2029 * preempted or migrate. As such, make sure not to maintain any 2030 * thread-local state specific to the cache from prior to releasing 2031 * the critical section. 2032 */ 2033 critical_exit(); 2034 ZONE_LOCK(zone); 2035 critical_enter(); 2036 cpu = curcpu; 2037 cache = &zone->uz_cpu[cpu]; 2038 bucket = cache->uc_allocbucket; 2039 if (bucket != NULL) { 2040 if (bucket->ub_cnt > 0) { 2041 ZONE_UNLOCK(zone); 2042 goto zalloc_start; 2043 } 2044 bucket = cache->uc_freebucket; 2045 if (bucket != NULL && bucket->ub_cnt > 0) { 2046 ZONE_UNLOCK(zone); 2047 goto zalloc_start; 2048 } 2049 } 2050 2051 /* Since we have locked the zone we may as well send back our stats */ 2052 zone->uz_allocs += cache->uc_allocs; 2053 cache->uc_allocs = 0; 2054 zone->uz_frees += cache->uc_frees; 2055 cache->uc_frees = 0; 2056 2057 /* Our old one is now a free bucket */ 2058 if (cache->uc_allocbucket) { 2059 KASSERT(cache->uc_allocbucket->ub_cnt == 0, 2060 ("uma_zalloc_arg: Freeing a non free bucket.")); 2061 LIST_INSERT_HEAD(&zone->uz_free_bucket, 2062 cache->uc_allocbucket, ub_link); 2063 cache->uc_allocbucket = NULL; 2064 } 2065 2066 /* Check the free list for a new alloc bucket */ 2067 if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) { 2068 KASSERT(bucket->ub_cnt != 0, 2069 ("uma_zalloc_arg: Returning an empty bucket.")); 2070 2071 LIST_REMOVE(bucket, ub_link); 2072 cache->uc_allocbucket = bucket; 2073 ZONE_UNLOCK(zone); 2074 goto zalloc_start; 2075 } 2076 /* We are no longer associated with this CPU. */ 2077 critical_exit(); 2078 2079 /* Bump up our uz_count so we get here less */ 2080 if (zone->uz_count < BUCKET_MAX) 2081 zone->uz_count++; 2082 2083 /* 2084 * Now lets just fill a bucket and put it on the free list. If that 2085 * works we'll restart the allocation from the begining. 2086 */ 2087 if (zone_alloc_bucket(zone, flags)) { 2088 ZONE_UNLOCK(zone); 2089 goto zalloc_restart; 2090 } 2091 ZONE_UNLOCK(zone); 2092 /* 2093 * We may not be able to get a bucket so return an actual item. 2094 */ 2095 #ifdef UMA_DEBUG 2096 printf("uma_zalloc_arg: Bucketzone returned NULL\n"); 2097 #endif 2098 2099 item = zone_alloc_item(zone, udata, flags); 2100 return (item); 2101 } 2102 2103 static uma_slab_t 2104 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) 2105 { 2106 uma_slab_t slab; 2107 2108 mtx_assert(&keg->uk_lock, MA_OWNED); 2109 slab = NULL; 2110 2111 for (;;) { 2112 /* 2113 * Find a slab with some space. Prefer slabs that are partially 2114 * used over those that are totally full. This helps to reduce 2115 * fragmentation. 2116 */ 2117 if (keg->uk_free != 0) { 2118 if (!LIST_EMPTY(&keg->uk_part_slab)) { 2119 slab = LIST_FIRST(&keg->uk_part_slab); 2120 } else { 2121 slab = LIST_FIRST(&keg->uk_free_slab); 2122 LIST_REMOVE(slab, us_link); 2123 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, 2124 us_link); 2125 } 2126 MPASS(slab->us_keg == keg); 2127 return (slab); 2128 } 2129 2130 /* 2131 * M_NOVM means don't ask at all! 2132 */ 2133 if (flags & M_NOVM) 2134 break; 2135 2136 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2137 keg->uk_flags |= UMA_ZFLAG_FULL; 2138 /* 2139 * If this is not a multi-zone, set the FULL bit. 2140 * Otherwise slab_multi() takes care of it. 2141 */ 2142 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) 2143 zone->uz_flags |= UMA_ZFLAG_FULL; 2144 if (flags & M_NOWAIT) 2145 break; 2146 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2147 continue; 2148 } 2149 keg->uk_recurse++; 2150 slab = keg_alloc_slab(keg, zone, flags); 2151 keg->uk_recurse--; 2152 /* 2153 * If we got a slab here it's safe to mark it partially used 2154 * and return. We assume that the caller is going to remove 2155 * at least one item. 2156 */ 2157 if (slab) { 2158 MPASS(slab->us_keg == keg); 2159 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2160 return (slab); 2161 } 2162 /* 2163 * We might not have been able to get a slab but another cpu 2164 * could have while we were unlocked. Check again before we 2165 * fail. 2166 */ 2167 flags |= M_NOVM; 2168 } 2169 return (slab); 2170 } 2171 2172 static inline void 2173 zone_relock(uma_zone_t zone, uma_keg_t keg) 2174 { 2175 if (zone->uz_lock != &keg->uk_lock) { 2176 KEG_UNLOCK(keg); 2177 ZONE_LOCK(zone); 2178 } 2179 } 2180 2181 static inline void 2182 keg_relock(uma_keg_t keg, uma_zone_t zone) 2183 { 2184 if (zone->uz_lock != &keg->uk_lock) { 2185 ZONE_UNLOCK(zone); 2186 KEG_LOCK(keg); 2187 } 2188 } 2189 2190 static uma_slab_t 2191 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) 2192 { 2193 uma_slab_t slab; 2194 2195 if (keg == NULL) 2196 keg = zone_first_keg(zone); 2197 /* 2198 * This is to prevent us from recursively trying to allocate 2199 * buckets. The problem is that if an allocation forces us to 2200 * grab a new bucket we will call page_alloc, which will go off 2201 * and cause the vm to allocate vm_map_entries. If we need new 2202 * buckets there too we will recurse in kmem_alloc and bad 2203 * things happen. So instead we return a NULL bucket, and make 2204 * the code that allocates buckets smart enough to deal with it 2205 */ 2206 if (keg->uk_flags & UMA_ZFLAG_BUCKET && keg->uk_recurse != 0) 2207 return (NULL); 2208 2209 for (;;) { 2210 slab = keg_fetch_slab(keg, zone, flags); 2211 if (slab) 2212 return (slab); 2213 if (flags & (M_NOWAIT | M_NOVM)) 2214 break; 2215 } 2216 return (NULL); 2217 } 2218 2219 /* 2220 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2221 * with the keg locked. Caller must call zone_relock() afterwards if the 2222 * zone lock is required. On NULL the zone lock is held. 2223 * 2224 * The last pointer is used to seed the search. It is not required. 2225 */ 2226 static uma_slab_t 2227 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) 2228 { 2229 uma_klink_t klink; 2230 uma_slab_t slab; 2231 uma_keg_t keg; 2232 int flags; 2233 int empty; 2234 int full; 2235 2236 /* 2237 * Don't wait on the first pass. This will skip limit tests 2238 * as well. We don't want to block if we can find a provider 2239 * without blocking. 2240 */ 2241 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2242 /* 2243 * Use the last slab allocated as a hint for where to start 2244 * the search. 2245 */ 2246 if (last) { 2247 slab = keg_fetch_slab(last, zone, flags); 2248 if (slab) 2249 return (slab); 2250 zone_relock(zone, last); 2251 last = NULL; 2252 } 2253 /* 2254 * Loop until we have a slab incase of transient failures 2255 * while M_WAITOK is specified. I'm not sure this is 100% 2256 * required but we've done it for so long now. 2257 */ 2258 for (;;) { 2259 empty = 0; 2260 full = 0; 2261 /* 2262 * Search the available kegs for slabs. Be careful to hold the 2263 * correct lock while calling into the keg layer. 2264 */ 2265 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2266 keg = klink->kl_keg; 2267 keg_relock(keg, zone); 2268 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2269 slab = keg_fetch_slab(keg, zone, flags); 2270 if (slab) 2271 return (slab); 2272 } 2273 if (keg->uk_flags & UMA_ZFLAG_FULL) 2274 full++; 2275 else 2276 empty++; 2277 zone_relock(zone, keg); 2278 } 2279 if (rflags & (M_NOWAIT | M_NOVM)) 2280 break; 2281 flags = rflags; 2282 /* 2283 * All kegs are full. XXX We can't atomically check all kegs 2284 * and sleep so just sleep for a short period and retry. 2285 */ 2286 if (full && !empty) { 2287 zone->uz_flags |= UMA_ZFLAG_FULL; 2288 msleep(zone, zone->uz_lock, PVM, "zonelimit", hz/100); 2289 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2290 continue; 2291 } 2292 } 2293 return (NULL); 2294 } 2295 2296 static void * 2297 slab_alloc_item(uma_zone_t zone, uma_slab_t slab) 2298 { 2299 uma_keg_t keg; 2300 uma_slabrefcnt_t slabref; 2301 void *item; 2302 u_int8_t freei; 2303 2304 keg = slab->us_keg; 2305 mtx_assert(&keg->uk_lock, MA_OWNED); 2306 2307 freei = slab->us_firstfree; 2308 if (keg->uk_flags & UMA_ZONE_REFCNT) { 2309 slabref = (uma_slabrefcnt_t)slab; 2310 slab->us_firstfree = slabref->us_freelist[freei].us_item; 2311 } else { 2312 slab->us_firstfree = slab->us_freelist[freei].us_item; 2313 } 2314 item = slab->us_data + (keg->uk_rsize * freei); 2315 2316 slab->us_freecount--; 2317 keg->uk_free--; 2318 #ifdef INVARIANTS 2319 uma_dbg_alloc(zone, slab, item); 2320 #endif 2321 /* Move this slab to the full list */ 2322 if (slab->us_freecount == 0) { 2323 LIST_REMOVE(slab, us_link); 2324 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); 2325 } 2326 2327 return (item); 2328 } 2329 2330 static int 2331 zone_alloc_bucket(uma_zone_t zone, int flags) 2332 { 2333 uma_bucket_t bucket; 2334 uma_slab_t slab; 2335 uma_keg_t keg; 2336 int16_t saved; 2337 int max, origflags = flags; 2338 2339 /* 2340 * Try this zone's free list first so we don't allocate extra buckets. 2341 */ 2342 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { 2343 KASSERT(bucket->ub_cnt == 0, 2344 ("zone_alloc_bucket: Bucket on free list is not empty.")); 2345 LIST_REMOVE(bucket, ub_link); 2346 } else { 2347 int bflags; 2348 2349 bflags = (flags & ~M_ZERO); 2350 if (zone->uz_flags & UMA_ZFLAG_CACHEONLY) 2351 bflags |= M_NOVM; 2352 2353 ZONE_UNLOCK(zone); 2354 bucket = bucket_alloc(zone->uz_count, bflags); 2355 ZONE_LOCK(zone); 2356 } 2357 2358 if (bucket == NULL) { 2359 return (0); 2360 } 2361 2362 #ifdef SMP 2363 /* 2364 * This code is here to limit the number of simultaneous bucket fills 2365 * for any given zone to the number of per cpu caches in this zone. This 2366 * is done so that we don't allocate more memory than we really need. 2367 */ 2368 if (zone->uz_fills >= mp_ncpus) 2369 goto done; 2370 2371 #endif 2372 zone->uz_fills++; 2373 2374 max = MIN(bucket->ub_entries, zone->uz_count); 2375 /* Try to keep the buckets totally full */ 2376 saved = bucket->ub_cnt; 2377 slab = NULL; 2378 keg = NULL; 2379 while (bucket->ub_cnt < max && 2380 (slab = zone->uz_slab(zone, keg, flags)) != NULL) { 2381 keg = slab->us_keg; 2382 while (slab->us_freecount && bucket->ub_cnt < max) { 2383 bucket->ub_bucket[bucket->ub_cnt++] = 2384 slab_alloc_item(zone, slab); 2385 } 2386 2387 /* Don't block on the next fill */ 2388 flags |= M_NOWAIT; 2389 } 2390 if (slab) 2391 zone_relock(zone, keg); 2392 2393 /* 2394 * We unlock here because we need to call the zone's init. 2395 * It should be safe to unlock because the slab dealt with 2396 * above is already on the appropriate list within the keg 2397 * and the bucket we filled is not yet on any list, so we 2398 * own it. 2399 */ 2400 if (zone->uz_init != NULL) { 2401 int i; 2402 2403 ZONE_UNLOCK(zone); 2404 for (i = saved; i < bucket->ub_cnt; i++) 2405 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2406 origflags) != 0) 2407 break; 2408 /* 2409 * If we couldn't initialize the whole bucket, put the 2410 * rest back onto the freelist. 2411 */ 2412 if (i != bucket->ub_cnt) { 2413 int j; 2414 2415 for (j = i; j < bucket->ub_cnt; j++) { 2416 zone_free_item(zone, bucket->ub_bucket[j], 2417 NULL, SKIP_FINI, 0); 2418 #ifdef INVARIANTS 2419 bucket->ub_bucket[j] = NULL; 2420 #endif 2421 } 2422 bucket->ub_cnt = i; 2423 } 2424 ZONE_LOCK(zone); 2425 } 2426 2427 zone->uz_fills--; 2428 if (bucket->ub_cnt != 0) { 2429 LIST_INSERT_HEAD(&zone->uz_full_bucket, 2430 bucket, ub_link); 2431 return (1); 2432 } 2433 #ifdef SMP 2434 done: 2435 #endif 2436 bucket_free(bucket); 2437 2438 return (0); 2439 } 2440 /* 2441 * Allocates an item for an internal zone 2442 * 2443 * Arguments 2444 * zone The zone to alloc for. 2445 * udata The data to be passed to the constructor. 2446 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2447 * 2448 * Returns 2449 * NULL if there is no memory and M_NOWAIT is set 2450 * An item if successful 2451 */ 2452 2453 static void * 2454 zone_alloc_item(uma_zone_t zone, void *udata, int flags) 2455 { 2456 uma_slab_t slab; 2457 void *item; 2458 2459 item = NULL; 2460 2461 #ifdef UMA_DEBUG_ALLOC 2462 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); 2463 #endif 2464 ZONE_LOCK(zone); 2465 2466 slab = zone->uz_slab(zone, NULL, flags); 2467 if (slab == NULL) { 2468 zone->uz_fails++; 2469 ZONE_UNLOCK(zone); 2470 return (NULL); 2471 } 2472 2473 item = slab_alloc_item(zone, slab); 2474 2475 zone_relock(zone, slab->us_keg); 2476 zone->uz_allocs++; 2477 ZONE_UNLOCK(zone); 2478 2479 /* 2480 * We have to call both the zone's init (not the keg's init) 2481 * and the zone's ctor. This is because the item is going from 2482 * a keg slab directly to the user, and the user is expecting it 2483 * to be both zone-init'd as well as zone-ctor'd. 2484 */ 2485 if (zone->uz_init != NULL) { 2486 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 2487 zone_free_item(zone, item, udata, SKIP_FINI, 2488 ZFREE_STATFAIL | ZFREE_STATFREE); 2489 return (NULL); 2490 } 2491 } 2492 if (zone->uz_ctor != NULL) { 2493 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2494 zone_free_item(zone, item, udata, SKIP_DTOR, 2495 ZFREE_STATFAIL | ZFREE_STATFREE); 2496 return (NULL); 2497 } 2498 } 2499 if (flags & M_ZERO) 2500 bzero(item, zone->uz_size); 2501 2502 return (item); 2503 } 2504 2505 /* See uma.h */ 2506 void 2507 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 2508 { 2509 uma_cache_t cache; 2510 uma_bucket_t bucket; 2511 int bflags; 2512 int cpu; 2513 2514 #ifdef UMA_DEBUG_ALLOC_1 2515 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); 2516 #endif 2517 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 2518 zone->uz_name); 2519 2520 if (zone->uz_dtor) 2521 zone->uz_dtor(item, zone->uz_size, udata); 2522 2523 #ifdef INVARIANTS 2524 ZONE_LOCK(zone); 2525 if (zone->uz_flags & UMA_ZONE_MALLOC) 2526 uma_dbg_free(zone, udata, item); 2527 else 2528 uma_dbg_free(zone, NULL, item); 2529 ZONE_UNLOCK(zone); 2530 #endif 2531 /* 2532 * The race here is acceptable. If we miss it we'll just have to wait 2533 * a little longer for the limits to be reset. 2534 */ 2535 if (zone->uz_flags & UMA_ZFLAG_FULL) 2536 goto zfree_internal; 2537 2538 /* 2539 * If possible, free to the per-CPU cache. There are two 2540 * requirements for safe access to the per-CPU cache: (1) the thread 2541 * accessing the cache must not be preempted or yield during access, 2542 * and (2) the thread must not migrate CPUs without switching which 2543 * cache it accesses. We rely on a critical section to prevent 2544 * preemption and migration. We release the critical section in 2545 * order to acquire the zone mutex if we are unable to free to the 2546 * current cache; when we re-acquire the critical section, we must 2547 * detect and handle migration if it has occurred. 2548 */ 2549 zfree_restart: 2550 critical_enter(); 2551 cpu = curcpu; 2552 cache = &zone->uz_cpu[cpu]; 2553 2554 zfree_start: 2555 bucket = cache->uc_freebucket; 2556 2557 if (bucket) { 2558 /* 2559 * Do we have room in our bucket? It is OK for this uz count 2560 * check to be slightly out of sync. 2561 */ 2562 2563 if (bucket->ub_cnt < bucket->ub_entries) { 2564 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 2565 ("uma_zfree: Freeing to non free bucket index.")); 2566 bucket->ub_bucket[bucket->ub_cnt] = item; 2567 bucket->ub_cnt++; 2568 cache->uc_frees++; 2569 critical_exit(); 2570 return; 2571 } else if (cache->uc_allocbucket) { 2572 #ifdef UMA_DEBUG_ALLOC 2573 printf("uma_zfree: Swapping buckets.\n"); 2574 #endif 2575 /* 2576 * We have run out of space in our freebucket. 2577 * See if we can switch with our alloc bucket. 2578 */ 2579 if (cache->uc_allocbucket->ub_cnt < 2580 cache->uc_freebucket->ub_cnt) { 2581 bucket = cache->uc_freebucket; 2582 cache->uc_freebucket = cache->uc_allocbucket; 2583 cache->uc_allocbucket = bucket; 2584 goto zfree_start; 2585 } 2586 } 2587 } 2588 /* 2589 * We can get here for two reasons: 2590 * 2591 * 1) The buckets are NULL 2592 * 2) The alloc and free buckets are both somewhat full. 2593 * 2594 * We must go back the zone, which requires acquiring the zone lock, 2595 * which in turn means we must release and re-acquire the critical 2596 * section. Since the critical section is released, we may be 2597 * preempted or migrate. As such, make sure not to maintain any 2598 * thread-local state specific to the cache from prior to releasing 2599 * the critical section. 2600 */ 2601 critical_exit(); 2602 ZONE_LOCK(zone); 2603 critical_enter(); 2604 cpu = curcpu; 2605 cache = &zone->uz_cpu[cpu]; 2606 if (cache->uc_freebucket != NULL) { 2607 if (cache->uc_freebucket->ub_cnt < 2608 cache->uc_freebucket->ub_entries) { 2609 ZONE_UNLOCK(zone); 2610 goto zfree_start; 2611 } 2612 if (cache->uc_allocbucket != NULL && 2613 (cache->uc_allocbucket->ub_cnt < 2614 cache->uc_freebucket->ub_cnt)) { 2615 ZONE_UNLOCK(zone); 2616 goto zfree_start; 2617 } 2618 } 2619 2620 /* Since we have locked the zone we may as well send back our stats */ 2621 zone->uz_allocs += cache->uc_allocs; 2622 cache->uc_allocs = 0; 2623 zone->uz_frees += cache->uc_frees; 2624 cache->uc_frees = 0; 2625 2626 bucket = cache->uc_freebucket; 2627 cache->uc_freebucket = NULL; 2628 2629 /* Can we throw this on the zone full list? */ 2630 if (bucket != NULL) { 2631 #ifdef UMA_DEBUG_ALLOC 2632 printf("uma_zfree: Putting old bucket on the free list.\n"); 2633 #endif 2634 /* ub_cnt is pointing to the last free item */ 2635 KASSERT(bucket->ub_cnt != 0, 2636 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 2637 LIST_INSERT_HEAD(&zone->uz_full_bucket, 2638 bucket, ub_link); 2639 } 2640 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { 2641 LIST_REMOVE(bucket, ub_link); 2642 ZONE_UNLOCK(zone); 2643 cache->uc_freebucket = bucket; 2644 goto zfree_start; 2645 } 2646 /* We are no longer associated with this CPU. */ 2647 critical_exit(); 2648 2649 /* And the zone.. */ 2650 ZONE_UNLOCK(zone); 2651 2652 #ifdef UMA_DEBUG_ALLOC 2653 printf("uma_zfree: Allocating new free bucket.\n"); 2654 #endif 2655 bflags = M_NOWAIT; 2656 2657 if (zone->uz_flags & UMA_ZFLAG_CACHEONLY) 2658 bflags |= M_NOVM; 2659 bucket = bucket_alloc(zone->uz_count, bflags); 2660 if (bucket) { 2661 ZONE_LOCK(zone); 2662 LIST_INSERT_HEAD(&zone->uz_free_bucket, 2663 bucket, ub_link); 2664 ZONE_UNLOCK(zone); 2665 goto zfree_restart; 2666 } 2667 2668 /* 2669 * If nothing else caught this, we'll just do an internal free. 2670 */ 2671 zfree_internal: 2672 zone_free_item(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE); 2673 2674 return; 2675 } 2676 2677 /* 2678 * Frees an item to an INTERNAL zone or allocates a free bucket 2679 * 2680 * Arguments: 2681 * zone The zone to free to 2682 * item The item we're freeing 2683 * udata User supplied data for the dtor 2684 * skip Skip dtors and finis 2685 */ 2686 static void 2687 zone_free_item(uma_zone_t zone, void *item, void *udata, 2688 enum zfreeskip skip, int flags) 2689 { 2690 uma_slab_t slab; 2691 uma_slabrefcnt_t slabref; 2692 uma_keg_t keg; 2693 u_int8_t *mem; 2694 u_int8_t freei; 2695 int clearfull; 2696 2697 if (skip < SKIP_DTOR && zone->uz_dtor) 2698 zone->uz_dtor(item, zone->uz_size, udata); 2699 2700 if (skip < SKIP_FINI && zone->uz_fini) 2701 zone->uz_fini(item, zone->uz_size); 2702 2703 ZONE_LOCK(zone); 2704 2705 if (flags & ZFREE_STATFAIL) 2706 zone->uz_fails++; 2707 if (flags & ZFREE_STATFREE) 2708 zone->uz_frees++; 2709 2710 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 2711 mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK)); 2712 keg = zone_first_keg(zone); /* Must only be one. */ 2713 if (zone->uz_flags & UMA_ZONE_HASH) { 2714 slab = hash_sfind(&keg->uk_hash, mem); 2715 } else { 2716 mem += keg->uk_pgoff; 2717 slab = (uma_slab_t)mem; 2718 } 2719 } else { 2720 /* This prevents redundant lookups via free(). */ 2721 if ((zone->uz_flags & UMA_ZONE_MALLOC) && udata != NULL) 2722 slab = (uma_slab_t)udata; 2723 else 2724 slab = vtoslab((vm_offset_t)item); 2725 keg = slab->us_keg; 2726 keg_relock(keg, zone); 2727 } 2728 MPASS(keg == slab->us_keg); 2729 2730 /* Do we need to remove from any lists? */ 2731 if (slab->us_freecount+1 == keg->uk_ipers) { 2732 LIST_REMOVE(slab, us_link); 2733 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 2734 } else if (slab->us_freecount == 0) { 2735 LIST_REMOVE(slab, us_link); 2736 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2737 } 2738 2739 /* Slab management stuff */ 2740 freei = ((unsigned long)item - (unsigned long)slab->us_data) 2741 / keg->uk_rsize; 2742 2743 #ifdef INVARIANTS 2744 if (!skip) 2745 uma_dbg_free(zone, slab, item); 2746 #endif 2747 2748 if (keg->uk_flags & UMA_ZONE_REFCNT) { 2749 slabref = (uma_slabrefcnt_t)slab; 2750 slabref->us_freelist[freei].us_item = slab->us_firstfree; 2751 } else { 2752 slab->us_freelist[freei].us_item = slab->us_firstfree; 2753 } 2754 slab->us_firstfree = freei; 2755 slab->us_freecount++; 2756 2757 /* Zone statistics */ 2758 keg->uk_free++; 2759 2760 clearfull = 0; 2761 if (keg->uk_flags & UMA_ZFLAG_FULL) { 2762 if (keg->uk_pages < keg->uk_maxpages) { 2763 keg->uk_flags &= ~UMA_ZFLAG_FULL; 2764 clearfull = 1; 2765 } 2766 2767 /* 2768 * We can handle one more allocation. Since we're clearing ZFLAG_FULL, 2769 * wake up all procs blocked on pages. This should be uncommon, so 2770 * keeping this simple for now (rather than adding count of blocked 2771 * threads etc). 2772 */ 2773 wakeup(keg); 2774 } 2775 if (clearfull) { 2776 zone_relock(zone, keg); 2777 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2778 wakeup(zone); 2779 ZONE_UNLOCK(zone); 2780 } else 2781 KEG_UNLOCK(keg); 2782 } 2783 2784 /* See uma.h */ 2785 void 2786 uma_zone_set_max(uma_zone_t zone, int nitems) 2787 { 2788 uma_keg_t keg; 2789 2790 ZONE_LOCK(zone); 2791 keg = zone_first_keg(zone); 2792 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 2793 if (keg->uk_maxpages * keg->uk_ipers < nitems) 2794 keg->uk_maxpages += keg->uk_ppera; 2795 2796 ZONE_UNLOCK(zone); 2797 } 2798 2799 /* See uma.h */ 2800 void 2801 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 2802 { 2803 uma_keg_t keg; 2804 2805 ZONE_LOCK(zone); 2806 keg = zone_first_keg(zone); 2807 KASSERT(keg->uk_pages == 0, 2808 ("uma_zone_set_init on non-empty keg")); 2809 keg->uk_init = uminit; 2810 ZONE_UNLOCK(zone); 2811 } 2812 2813 /* See uma.h */ 2814 void 2815 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 2816 { 2817 uma_keg_t keg; 2818 2819 ZONE_LOCK(zone); 2820 keg = zone_first_keg(zone); 2821 KASSERT(keg->uk_pages == 0, 2822 ("uma_zone_set_fini on non-empty keg")); 2823 keg->uk_fini = fini; 2824 ZONE_UNLOCK(zone); 2825 } 2826 2827 /* See uma.h */ 2828 void 2829 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 2830 { 2831 ZONE_LOCK(zone); 2832 KASSERT(zone_first_keg(zone)->uk_pages == 0, 2833 ("uma_zone_set_zinit on non-empty keg")); 2834 zone->uz_init = zinit; 2835 ZONE_UNLOCK(zone); 2836 } 2837 2838 /* See uma.h */ 2839 void 2840 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 2841 { 2842 ZONE_LOCK(zone); 2843 KASSERT(zone_first_keg(zone)->uk_pages == 0, 2844 ("uma_zone_set_zfini on non-empty keg")); 2845 zone->uz_fini = zfini; 2846 ZONE_UNLOCK(zone); 2847 } 2848 2849 /* See uma.h */ 2850 /* XXX uk_freef is not actually used with the zone locked */ 2851 void 2852 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 2853 { 2854 2855 ZONE_LOCK(zone); 2856 zone_first_keg(zone)->uk_freef = freef; 2857 ZONE_UNLOCK(zone); 2858 } 2859 2860 /* See uma.h */ 2861 /* XXX uk_allocf is not actually used with the zone locked */ 2862 void 2863 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 2864 { 2865 uma_keg_t keg; 2866 2867 ZONE_LOCK(zone); 2868 keg = zone_first_keg(zone); 2869 keg->uk_flags |= UMA_ZFLAG_PRIVALLOC; 2870 keg->uk_allocf = allocf; 2871 ZONE_UNLOCK(zone); 2872 } 2873 2874 /* See uma.h */ 2875 int 2876 uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count) 2877 { 2878 uma_keg_t keg; 2879 vm_offset_t kva; 2880 int pages; 2881 2882 keg = zone_first_keg(zone); 2883 pages = count / keg->uk_ipers; 2884 2885 if (pages * keg->uk_ipers < count) 2886 pages++; 2887 2888 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE); 2889 2890 if (kva == 0) 2891 return (0); 2892 if (obj == NULL) { 2893 obj = vm_object_allocate(OBJT_DEFAULT, 2894 pages); 2895 } else { 2896 VM_OBJECT_LOCK_INIT(obj, "uma object"); 2897 _vm_object_allocate(OBJT_DEFAULT, 2898 pages, obj); 2899 } 2900 ZONE_LOCK(zone); 2901 keg->uk_kva = kva; 2902 keg->uk_obj = obj; 2903 keg->uk_maxpages = pages; 2904 keg->uk_allocf = obj_alloc; 2905 keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC; 2906 ZONE_UNLOCK(zone); 2907 return (1); 2908 } 2909 2910 /* See uma.h */ 2911 void 2912 uma_prealloc(uma_zone_t zone, int items) 2913 { 2914 int slabs; 2915 uma_slab_t slab; 2916 uma_keg_t keg; 2917 2918 keg = zone_first_keg(zone); 2919 ZONE_LOCK(zone); 2920 slabs = items / keg->uk_ipers; 2921 if (slabs * keg->uk_ipers < items) 2922 slabs++; 2923 while (slabs > 0) { 2924 slab = keg_alloc_slab(keg, zone, M_WAITOK); 2925 if (slab == NULL) 2926 break; 2927 MPASS(slab->us_keg == keg); 2928 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 2929 slabs--; 2930 } 2931 ZONE_UNLOCK(zone); 2932 } 2933 2934 /* See uma.h */ 2935 u_int32_t * 2936 uma_find_refcnt(uma_zone_t zone, void *item) 2937 { 2938 uma_slabrefcnt_t slabref; 2939 uma_keg_t keg; 2940 u_int32_t *refcnt; 2941 int idx; 2942 2943 slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item & 2944 (~UMA_SLAB_MASK)); 2945 keg = slabref->us_keg; 2946 KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT, 2947 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT")); 2948 idx = ((unsigned long)item - (unsigned long)slabref->us_data) 2949 / keg->uk_rsize; 2950 refcnt = &slabref->us_freelist[idx].us_refcnt; 2951 return refcnt; 2952 } 2953 2954 /* See uma.h */ 2955 void 2956 uma_reclaim(void) 2957 { 2958 #ifdef UMA_DEBUG 2959 printf("UMA: vm asked us to release pages!\n"); 2960 #endif 2961 bucket_enable(); 2962 zone_foreach(zone_drain); 2963 /* 2964 * Some slabs may have been freed but this zone will be visited early 2965 * we visit again so that we can free pages that are empty once other 2966 * zones are drained. We have to do the same for buckets. 2967 */ 2968 zone_drain(slabzone); 2969 zone_drain(slabrefzone); 2970 bucket_zone_drain(); 2971 } 2972 2973 /* See uma.h */ 2974 int 2975 uma_zone_exhausted(uma_zone_t zone) 2976 { 2977 int full; 2978 2979 ZONE_LOCK(zone); 2980 full = (zone->uz_flags & UMA_ZFLAG_FULL); 2981 ZONE_UNLOCK(zone); 2982 return (full); 2983 } 2984 2985 int 2986 uma_zone_exhausted_nolock(uma_zone_t zone) 2987 { 2988 return (zone->uz_flags & UMA_ZFLAG_FULL); 2989 } 2990 2991 void * 2992 uma_large_malloc(int size, int wait) 2993 { 2994 void *mem; 2995 uma_slab_t slab; 2996 u_int8_t flags; 2997 2998 slab = zone_alloc_item(slabzone, NULL, wait); 2999 if (slab == NULL) 3000 return (NULL); 3001 mem = page_alloc(NULL, size, &flags, wait); 3002 if (mem) { 3003 vsetslab((vm_offset_t)mem, slab); 3004 slab->us_data = mem; 3005 slab->us_flags = flags | UMA_SLAB_MALLOC; 3006 slab->us_size = size; 3007 } else { 3008 zone_free_item(slabzone, slab, NULL, SKIP_NONE, 3009 ZFREE_STATFAIL | ZFREE_STATFREE); 3010 } 3011 3012 return (mem); 3013 } 3014 3015 void 3016 uma_large_free(uma_slab_t slab) 3017 { 3018 vsetobj((vm_offset_t)slab->us_data, kmem_object); 3019 page_free(slab->us_data, slab->us_size, slab->us_flags); 3020 zone_free_item(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE); 3021 } 3022 3023 void 3024 uma_print_stats(void) 3025 { 3026 zone_foreach(uma_print_zone); 3027 } 3028 3029 static void 3030 slab_print(uma_slab_t slab) 3031 { 3032 printf("slab: keg %p, data %p, freecount %d, firstfree %d\n", 3033 slab->us_keg, slab->us_data, slab->us_freecount, 3034 slab->us_firstfree); 3035 } 3036 3037 static void 3038 cache_print(uma_cache_t cache) 3039 { 3040 printf("alloc: %p(%d), free: %p(%d)\n", 3041 cache->uc_allocbucket, 3042 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3043 cache->uc_freebucket, 3044 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3045 } 3046 3047 static void 3048 uma_print_keg(uma_keg_t keg) 3049 { 3050 uma_slab_t slab; 3051 3052 printf("keg: %s(%p) size %d(%d) flags %d ipers %d ppera %d " 3053 "out %d free %d limit %d\n", 3054 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3055 keg->uk_ipers, keg->uk_ppera, 3056 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free, 3057 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3058 printf("Part slabs:\n"); 3059 LIST_FOREACH(slab, &keg->uk_part_slab, us_link) 3060 slab_print(slab); 3061 printf("Free slabs:\n"); 3062 LIST_FOREACH(slab, &keg->uk_free_slab, us_link) 3063 slab_print(slab); 3064 printf("Full slabs:\n"); 3065 LIST_FOREACH(slab, &keg->uk_full_slab, us_link) 3066 slab_print(slab); 3067 } 3068 3069 void 3070 uma_print_zone(uma_zone_t zone) 3071 { 3072 uma_cache_t cache; 3073 uma_klink_t kl; 3074 int i; 3075 3076 printf("zone: %s(%p) size %d flags %d\n", 3077 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3078 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3079 uma_print_keg(kl->kl_keg); 3080 for (i = 0; i <= mp_maxid; i++) { 3081 if (CPU_ABSENT(i)) 3082 continue; 3083 cache = &zone->uz_cpu[i]; 3084 printf("CPU %d Cache:\n", i); 3085 cache_print(cache); 3086 } 3087 } 3088 3089 #ifdef DDB 3090 /* 3091 * Generate statistics across both the zone and its per-cpu cache's. Return 3092 * desired statistics if the pointer is non-NULL for that statistic. 3093 * 3094 * Note: does not update the zone statistics, as it can't safely clear the 3095 * per-CPU cache statistic. 3096 * 3097 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3098 * safe from off-CPU; we should modify the caches to track this information 3099 * directly so that we don't have to. 3100 */ 3101 static void 3102 uma_zone_sumstat(uma_zone_t z, int *cachefreep, u_int64_t *allocsp, 3103 u_int64_t *freesp) 3104 { 3105 uma_cache_t cache; 3106 u_int64_t allocs, frees; 3107 int cachefree, cpu; 3108 3109 allocs = frees = 0; 3110 cachefree = 0; 3111 for (cpu = 0; cpu <= mp_maxid; cpu++) { 3112 if (CPU_ABSENT(cpu)) 3113 continue; 3114 cache = &z->uz_cpu[cpu]; 3115 if (cache->uc_allocbucket != NULL) 3116 cachefree += cache->uc_allocbucket->ub_cnt; 3117 if (cache->uc_freebucket != NULL) 3118 cachefree += cache->uc_freebucket->ub_cnt; 3119 allocs += cache->uc_allocs; 3120 frees += cache->uc_frees; 3121 } 3122 allocs += z->uz_allocs; 3123 frees += z->uz_frees; 3124 if (cachefreep != NULL) 3125 *cachefreep = cachefree; 3126 if (allocsp != NULL) 3127 *allocsp = allocs; 3128 if (freesp != NULL) 3129 *freesp = frees; 3130 } 3131 #endif /* DDB */ 3132 3133 static int 3134 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3135 { 3136 uma_keg_t kz; 3137 uma_zone_t z; 3138 int count; 3139 3140 count = 0; 3141 mtx_lock(&uma_mtx); 3142 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3143 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3144 count++; 3145 } 3146 mtx_unlock(&uma_mtx); 3147 return (sysctl_handle_int(oidp, &count, 0, req)); 3148 } 3149 3150 static int 3151 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3152 { 3153 struct uma_stream_header ush; 3154 struct uma_type_header uth; 3155 struct uma_percpu_stat ups; 3156 uma_bucket_t bucket; 3157 struct sbuf sbuf; 3158 uma_cache_t cache; 3159 uma_klink_t kl; 3160 uma_keg_t kz; 3161 uma_zone_t z; 3162 uma_keg_t k; 3163 char *buffer; 3164 int buflen, count, error, i; 3165 3166 mtx_lock(&uma_mtx); 3167 restart: 3168 mtx_assert(&uma_mtx, MA_OWNED); 3169 count = 0; 3170 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3171 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3172 count++; 3173 } 3174 mtx_unlock(&uma_mtx); 3175 3176 buflen = sizeof(ush) + count * (sizeof(uth) + sizeof(ups) * 3177 (mp_maxid + 1)) + 1; 3178 buffer = malloc(buflen, M_TEMP, M_WAITOK | M_ZERO); 3179 3180 mtx_lock(&uma_mtx); 3181 i = 0; 3182 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3183 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3184 i++; 3185 } 3186 if (i > count) { 3187 free(buffer, M_TEMP); 3188 goto restart; 3189 } 3190 count = i; 3191 3192 sbuf_new(&sbuf, buffer, buflen, SBUF_FIXEDLEN); 3193 3194 /* 3195 * Insert stream header. 3196 */ 3197 bzero(&ush, sizeof(ush)); 3198 ush.ush_version = UMA_STREAM_VERSION; 3199 ush.ush_maxcpus = (mp_maxid + 1); 3200 ush.ush_count = count; 3201 if (sbuf_bcat(&sbuf, &ush, sizeof(ush)) < 0) { 3202 mtx_unlock(&uma_mtx); 3203 error = ENOMEM; 3204 goto out; 3205 } 3206 3207 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3208 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3209 bzero(&uth, sizeof(uth)); 3210 ZONE_LOCK(z); 3211 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3212 uth.uth_align = kz->uk_align; 3213 uth.uth_size = kz->uk_size; 3214 uth.uth_rsize = kz->uk_rsize; 3215 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3216 k = kl->kl_keg; 3217 uth.uth_maxpages += k->uk_maxpages; 3218 uth.uth_pages += k->uk_pages; 3219 uth.uth_keg_free += k->uk_free; 3220 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3221 * k->uk_ipers; 3222 } 3223 3224 /* 3225 * A zone is secondary is it is not the first entry 3226 * on the keg's zone list. 3227 */ 3228 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3229 (LIST_FIRST(&kz->uk_zones) != z)) 3230 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3231 3232 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link) 3233 uth.uth_zone_free += bucket->ub_cnt; 3234 uth.uth_allocs = z->uz_allocs; 3235 uth.uth_frees = z->uz_frees; 3236 uth.uth_fails = z->uz_fails; 3237 if (sbuf_bcat(&sbuf, &uth, sizeof(uth)) < 0) { 3238 ZONE_UNLOCK(z); 3239 mtx_unlock(&uma_mtx); 3240 error = ENOMEM; 3241 goto out; 3242 } 3243 /* 3244 * While it is not normally safe to access the cache 3245 * bucket pointers while not on the CPU that owns the 3246 * cache, we only allow the pointers to be exchanged 3247 * without the zone lock held, not invalidated, so 3248 * accept the possible race associated with bucket 3249 * exchange during monitoring. 3250 */ 3251 for (i = 0; i < (mp_maxid + 1); i++) { 3252 bzero(&ups, sizeof(ups)); 3253 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) 3254 goto skip; 3255 if (CPU_ABSENT(i)) 3256 goto skip; 3257 cache = &z->uz_cpu[i]; 3258 if (cache->uc_allocbucket != NULL) 3259 ups.ups_cache_free += 3260 cache->uc_allocbucket->ub_cnt; 3261 if (cache->uc_freebucket != NULL) 3262 ups.ups_cache_free += 3263 cache->uc_freebucket->ub_cnt; 3264 ups.ups_allocs = cache->uc_allocs; 3265 ups.ups_frees = cache->uc_frees; 3266 skip: 3267 if (sbuf_bcat(&sbuf, &ups, sizeof(ups)) < 0) { 3268 ZONE_UNLOCK(z); 3269 mtx_unlock(&uma_mtx); 3270 error = ENOMEM; 3271 goto out; 3272 } 3273 } 3274 ZONE_UNLOCK(z); 3275 } 3276 } 3277 mtx_unlock(&uma_mtx); 3278 sbuf_finish(&sbuf); 3279 error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf)); 3280 out: 3281 free(buffer, M_TEMP); 3282 return (error); 3283 } 3284 3285 #ifdef DDB 3286 DB_SHOW_COMMAND(uma, db_show_uma) 3287 { 3288 u_int64_t allocs, frees; 3289 uma_bucket_t bucket; 3290 uma_keg_t kz; 3291 uma_zone_t z; 3292 int cachefree; 3293 3294 db_printf("%18s %8s %8s %8s %12s\n", "Zone", "Size", "Used", "Free", 3295 "Requests"); 3296 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3297 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3298 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 3299 allocs = z->uz_allocs; 3300 frees = z->uz_frees; 3301 cachefree = 0; 3302 } else 3303 uma_zone_sumstat(z, &cachefree, &allocs, 3304 &frees); 3305 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 3306 (LIST_FIRST(&kz->uk_zones) != z))) 3307 cachefree += kz->uk_free; 3308 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link) 3309 cachefree += bucket->ub_cnt; 3310 db_printf("%18s %8ju %8jd %8d %12ju\n", z->uz_name, 3311 (uintmax_t)kz->uk_size, 3312 (intmax_t)(allocs - frees), cachefree, 3313 (uintmax_t)allocs); 3314 } 3315 } 3316 } 3317 #endif 3318