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 #include "opt_vm.h" 63 64 #include <sys/param.h> 65 #include <sys/systm.h> 66 #include <sys/kernel.h> 67 #include <sys/types.h> 68 #include <sys/queue.h> 69 #include <sys/malloc.h> 70 #include <sys/ktr.h> 71 #include <sys/lock.h> 72 #include <sys/sysctl.h> 73 #include <sys/mutex.h> 74 #include <sys/proc.h> 75 #include <sys/rwlock.h> 76 #include <sys/sbuf.h> 77 #include <sys/smp.h> 78 #include <sys/vmmeter.h> 79 80 #include <vm/vm.h> 81 #include <vm/vm_object.h> 82 #include <vm/vm_page.h> 83 #include <vm/vm_pageout.h> 84 #include <vm/vm_param.h> 85 #include <vm/vm_map.h> 86 #include <vm/vm_kern.h> 87 #include <vm/vm_extern.h> 88 #include <vm/uma.h> 89 #include <vm/uma_int.h> 90 #include <vm/uma_dbg.h> 91 92 #include <ddb/ddb.h> 93 94 #ifdef DEBUG_MEMGUARD 95 #include <vm/memguard.h> 96 #endif 97 98 /* 99 * This is the zone and keg from which all zones are spawned. The idea is that 100 * even the zone & keg heads are allocated from the allocator, so we use the 101 * bss section to bootstrap us. 102 */ 103 static struct uma_keg masterkeg; 104 static struct uma_zone masterzone_k; 105 static struct uma_zone masterzone_z; 106 static uma_zone_t kegs = &masterzone_k; 107 static uma_zone_t zones = &masterzone_z; 108 109 /* This is the zone from which all of uma_slab_t's are allocated. */ 110 static uma_zone_t slabzone; 111 static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */ 112 113 /* 114 * The initial hash tables come out of this zone so they can be allocated 115 * prior to malloc coming up. 116 */ 117 static uma_zone_t hashzone; 118 119 /* The boot-time adjusted value for cache line alignment. */ 120 int uma_align_cache = 64 - 1; 121 122 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 123 124 /* 125 * Are we allowed to allocate buckets? 126 */ 127 static int bucketdisable = 1; 128 129 /* Linked list of all kegs in the system */ 130 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 131 132 /* This mutex protects the keg list */ 133 static struct mtx uma_mtx; 134 135 /* Linked list of boot time pages */ 136 static LIST_HEAD(,uma_slab) uma_boot_pages = 137 LIST_HEAD_INITIALIZER(uma_boot_pages); 138 139 /* This mutex protects the boot time pages list */ 140 static struct mtx uma_boot_pages_mtx; 141 142 /* Is the VM done starting up? */ 143 static int booted = 0; 144 #define UMA_STARTUP 1 145 #define UMA_STARTUP2 2 146 147 /* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */ 148 static u_int uma_max_ipers; 149 static u_int uma_max_ipers_ref; 150 151 /* 152 * This is the handle used to schedule events that need to happen 153 * outside of the allocation fast path. 154 */ 155 static struct callout uma_callout; 156 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 157 158 /* 159 * This structure is passed as the zone ctor arg so that I don't have to create 160 * a special allocation function just for zones. 161 */ 162 struct uma_zctor_args { 163 const char *name; 164 size_t size; 165 uma_ctor ctor; 166 uma_dtor dtor; 167 uma_init uminit; 168 uma_fini fini; 169 uma_keg_t keg; 170 int align; 171 uint32_t flags; 172 }; 173 174 struct uma_kctor_args { 175 uma_zone_t zone; 176 size_t size; 177 uma_init uminit; 178 uma_fini fini; 179 int align; 180 uint32_t flags; 181 }; 182 183 struct uma_bucket_zone { 184 uma_zone_t ubz_zone; 185 char *ubz_name; 186 int ubz_entries; 187 }; 188 189 #define BUCKET_MAX 128 190 191 struct uma_bucket_zone bucket_zones[] = { 192 { NULL, "16 Bucket", 16 }, 193 { NULL, "32 Bucket", 32 }, 194 { NULL, "64 Bucket", 64 }, 195 { NULL, "128 Bucket", 128 }, 196 { NULL, NULL, 0} 197 }; 198 199 #define BUCKET_SHIFT 4 200 #define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1) 201 202 /* 203 * bucket_size[] maps requested bucket sizes to zones that allocate a bucket 204 * of approximately the right size. 205 */ 206 static uint8_t bucket_size[BUCKET_ZONES]; 207 208 /* 209 * Flags and enumerations to be passed to internal functions. 210 */ 211 enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI }; 212 213 #define ZFREE_STATFAIL 0x00000001 /* Update zone failure statistic. */ 214 #define ZFREE_STATFREE 0x00000002 /* Update zone free statistic. */ 215 216 /* Prototypes.. */ 217 218 static void *noobj_alloc(uma_zone_t, int, uint8_t *, int); 219 static void *page_alloc(uma_zone_t, int, uint8_t *, int); 220 static void *startup_alloc(uma_zone_t, int, uint8_t *, int); 221 static void page_free(void *, int, uint8_t); 222 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); 223 static void cache_drain(uma_zone_t); 224 static void bucket_drain(uma_zone_t, uma_bucket_t); 225 static void bucket_cache_drain(uma_zone_t zone); 226 static int keg_ctor(void *, int, void *, int); 227 static void keg_dtor(void *, int, void *); 228 static int zone_ctor(void *, int, void *, int); 229 static void zone_dtor(void *, int, void *); 230 static int zero_init(void *, int, int); 231 static void keg_small_init(uma_keg_t keg); 232 static void keg_large_init(uma_keg_t keg); 233 static void zone_foreach(void (*zfunc)(uma_zone_t)); 234 static void zone_timeout(uma_zone_t zone); 235 static int hash_alloc(struct uma_hash *); 236 static int hash_expand(struct uma_hash *, struct uma_hash *); 237 static void hash_free(struct uma_hash *hash); 238 static void uma_timeout(void *); 239 static void uma_startup3(void); 240 static void *zone_alloc_item(uma_zone_t, void *, int); 241 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip, 242 int); 243 static void bucket_enable(void); 244 static void bucket_init(void); 245 static uma_bucket_t bucket_alloc(int, int); 246 static void bucket_free(uma_bucket_t); 247 static void bucket_zone_drain(void); 248 static int zone_alloc_bucket(uma_zone_t zone, int flags); 249 static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); 250 static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); 251 static void *slab_alloc_item(uma_zone_t zone, uma_slab_t slab); 252 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 253 uma_fini fini, int align, uint32_t flags); 254 static inline void zone_relock(uma_zone_t zone, uma_keg_t keg); 255 static inline void keg_relock(uma_keg_t keg, uma_zone_t zone); 256 257 void uma_print_zone(uma_zone_t); 258 void uma_print_stats(void); 259 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 260 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 261 262 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 263 264 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 265 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 266 267 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 268 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 269 270 static int zone_warnings = 1; 271 TUNABLE_INT("vm.zone_warnings", &zone_warnings); 272 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0, 273 "Warn when UMA zones becomes full"); 274 275 /* 276 * This routine checks to see whether or not it's safe to enable buckets. 277 */ 278 279 static void 280 bucket_enable(void) 281 { 282 bucketdisable = vm_page_count_min(); 283 } 284 285 /* 286 * Initialize bucket_zones, the array of zones of buckets of various sizes. 287 * 288 * For each zone, calculate the memory required for each bucket, consisting 289 * of the header and an array of pointers. Initialize bucket_size[] to point 290 * the range of appropriate bucket sizes at the zone. 291 */ 292 static void 293 bucket_init(void) 294 { 295 struct uma_bucket_zone *ubz; 296 int i; 297 int j; 298 299 for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) { 300 int size; 301 302 ubz = &bucket_zones[j]; 303 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 304 size += sizeof(void *) * ubz->ubz_entries; 305 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 306 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 307 UMA_ZFLAG_INTERNAL | UMA_ZFLAG_BUCKET); 308 for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT)) 309 bucket_size[i >> BUCKET_SHIFT] = j; 310 } 311 } 312 313 /* 314 * Given a desired number of entries for a bucket, return the zone from which 315 * to allocate the bucket. 316 */ 317 static struct uma_bucket_zone * 318 bucket_zone_lookup(int entries) 319 { 320 int idx; 321 322 idx = howmany(entries, 1 << BUCKET_SHIFT); 323 return (&bucket_zones[bucket_size[idx]]); 324 } 325 326 static uma_bucket_t 327 bucket_alloc(int entries, int bflags) 328 { 329 struct uma_bucket_zone *ubz; 330 uma_bucket_t bucket; 331 332 /* 333 * This is to stop us from allocating per cpu buckets while we're 334 * running out of vm.boot_pages. Otherwise, we would exhaust the 335 * boot pages. This also prevents us from allocating buckets in 336 * low memory situations. 337 */ 338 if (bucketdisable) 339 return (NULL); 340 341 ubz = bucket_zone_lookup(entries); 342 bucket = zone_alloc_item(ubz->ubz_zone, NULL, bflags); 343 if (bucket) { 344 #ifdef INVARIANTS 345 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 346 #endif 347 bucket->ub_cnt = 0; 348 bucket->ub_entries = ubz->ubz_entries; 349 } 350 351 return (bucket); 352 } 353 354 static void 355 bucket_free(uma_bucket_t bucket) 356 { 357 struct uma_bucket_zone *ubz; 358 359 ubz = bucket_zone_lookup(bucket->ub_entries); 360 zone_free_item(ubz->ubz_zone, bucket, NULL, SKIP_NONE, 361 ZFREE_STATFREE); 362 } 363 364 static void 365 bucket_zone_drain(void) 366 { 367 struct uma_bucket_zone *ubz; 368 369 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 370 zone_drain(ubz->ubz_zone); 371 } 372 373 static void 374 zone_log_warning(uma_zone_t zone) 375 { 376 static const struct timeval warninterval = { 300, 0 }; 377 378 if (!zone_warnings || zone->uz_warning == NULL) 379 return; 380 381 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 382 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 383 } 384 385 static inline uma_keg_t 386 zone_first_keg(uma_zone_t zone) 387 { 388 389 return (LIST_FIRST(&zone->uz_kegs)->kl_keg); 390 } 391 392 static void 393 zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 394 { 395 uma_klink_t klink; 396 397 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 398 kegfn(klink->kl_keg); 399 } 400 401 /* 402 * Routine called by timeout which is used to fire off some time interval 403 * based calculations. (stats, hash size, etc.) 404 * 405 * Arguments: 406 * arg Unused 407 * 408 * Returns: 409 * Nothing 410 */ 411 static void 412 uma_timeout(void *unused) 413 { 414 bucket_enable(); 415 zone_foreach(zone_timeout); 416 417 /* Reschedule this event */ 418 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 419 } 420 421 /* 422 * Routine to perform timeout driven calculations. This expands the 423 * hashes and does per cpu statistics aggregation. 424 * 425 * Returns nothing. 426 */ 427 static void 428 keg_timeout(uma_keg_t keg) 429 { 430 431 KEG_LOCK(keg); 432 /* 433 * Expand the keg hash table. 434 * 435 * This is done if the number of slabs is larger than the hash size. 436 * What I'm trying to do here is completely reduce collisions. This 437 * may be a little aggressive. Should I allow for two collisions max? 438 */ 439 if (keg->uk_flags & UMA_ZONE_HASH && 440 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 441 struct uma_hash newhash; 442 struct uma_hash oldhash; 443 int ret; 444 445 /* 446 * This is so involved because allocating and freeing 447 * while the keg lock is held will lead to deadlock. 448 * I have to do everything in stages and check for 449 * races. 450 */ 451 newhash = keg->uk_hash; 452 KEG_UNLOCK(keg); 453 ret = hash_alloc(&newhash); 454 KEG_LOCK(keg); 455 if (ret) { 456 if (hash_expand(&keg->uk_hash, &newhash)) { 457 oldhash = keg->uk_hash; 458 keg->uk_hash = newhash; 459 } else 460 oldhash = newhash; 461 462 KEG_UNLOCK(keg); 463 hash_free(&oldhash); 464 KEG_LOCK(keg); 465 } 466 } 467 KEG_UNLOCK(keg); 468 } 469 470 static void 471 zone_timeout(uma_zone_t zone) 472 { 473 474 zone_foreach_keg(zone, &keg_timeout); 475 } 476 477 /* 478 * Allocate and zero fill the next sized hash table from the appropriate 479 * backing store. 480 * 481 * Arguments: 482 * hash A new hash structure with the old hash size in uh_hashsize 483 * 484 * Returns: 485 * 1 on sucess and 0 on failure. 486 */ 487 static int 488 hash_alloc(struct uma_hash *hash) 489 { 490 int oldsize; 491 int alloc; 492 493 oldsize = hash->uh_hashsize; 494 495 /* We're just going to go to a power of two greater */ 496 if (oldsize) { 497 hash->uh_hashsize = oldsize * 2; 498 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 499 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 500 M_UMAHASH, M_NOWAIT); 501 } else { 502 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 503 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 504 M_WAITOK); 505 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 506 } 507 if (hash->uh_slab_hash) { 508 bzero(hash->uh_slab_hash, alloc); 509 hash->uh_hashmask = hash->uh_hashsize - 1; 510 return (1); 511 } 512 513 return (0); 514 } 515 516 /* 517 * Expands the hash table for HASH zones. This is done from zone_timeout 518 * to reduce collisions. This must not be done in the regular allocation 519 * path, otherwise, we can recurse on the vm while allocating pages. 520 * 521 * Arguments: 522 * oldhash The hash you want to expand 523 * newhash The hash structure for the new table 524 * 525 * Returns: 526 * Nothing 527 * 528 * Discussion: 529 */ 530 static int 531 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 532 { 533 uma_slab_t slab; 534 int hval; 535 int i; 536 537 if (!newhash->uh_slab_hash) 538 return (0); 539 540 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 541 return (0); 542 543 /* 544 * I need to investigate hash algorithms for resizing without a 545 * full rehash. 546 */ 547 548 for (i = 0; i < oldhash->uh_hashsize; i++) 549 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 550 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 551 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 552 hval = UMA_HASH(newhash, slab->us_data); 553 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 554 slab, us_hlink); 555 } 556 557 return (1); 558 } 559 560 /* 561 * Free the hash bucket to the appropriate backing store. 562 * 563 * Arguments: 564 * slab_hash The hash bucket we're freeing 565 * hashsize The number of entries in that hash bucket 566 * 567 * Returns: 568 * Nothing 569 */ 570 static void 571 hash_free(struct uma_hash *hash) 572 { 573 if (hash->uh_slab_hash == NULL) 574 return; 575 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 576 zone_free_item(hashzone, 577 hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE); 578 else 579 free(hash->uh_slab_hash, M_UMAHASH); 580 } 581 582 /* 583 * Frees all outstanding items in a bucket 584 * 585 * Arguments: 586 * zone The zone to free to, must be unlocked. 587 * bucket The free/alloc bucket with items, cpu queue must be locked. 588 * 589 * Returns: 590 * Nothing 591 */ 592 593 static void 594 bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 595 { 596 void *item; 597 598 if (bucket == NULL) 599 return; 600 601 while (bucket->ub_cnt > 0) { 602 bucket->ub_cnt--; 603 item = bucket->ub_bucket[bucket->ub_cnt]; 604 #ifdef INVARIANTS 605 bucket->ub_bucket[bucket->ub_cnt] = NULL; 606 KASSERT(item != NULL, 607 ("bucket_drain: botched ptr, item is NULL")); 608 #endif 609 zone_free_item(zone, item, NULL, SKIP_DTOR, 0); 610 } 611 } 612 613 /* 614 * Drains the per cpu caches for a zone. 615 * 616 * NOTE: This may only be called while the zone is being turn down, and not 617 * during normal operation. This is necessary in order that we do not have 618 * to migrate CPUs to drain the per-CPU caches. 619 * 620 * Arguments: 621 * zone The zone to drain, must be unlocked. 622 * 623 * Returns: 624 * Nothing 625 */ 626 static void 627 cache_drain(uma_zone_t zone) 628 { 629 uma_cache_t cache; 630 int cpu; 631 632 /* 633 * XXX: It is safe to not lock the per-CPU caches, because we're 634 * tearing down the zone anyway. I.e., there will be no further use 635 * of the caches at this point. 636 * 637 * XXX: It would good to be able to assert that the zone is being 638 * torn down to prevent improper use of cache_drain(). 639 * 640 * XXX: We lock the zone before passing into bucket_cache_drain() as 641 * it is used elsewhere. Should the tear-down path be made special 642 * there in some form? 643 */ 644 CPU_FOREACH(cpu) { 645 cache = &zone->uz_cpu[cpu]; 646 bucket_drain(zone, cache->uc_allocbucket); 647 bucket_drain(zone, cache->uc_freebucket); 648 if (cache->uc_allocbucket != NULL) 649 bucket_free(cache->uc_allocbucket); 650 if (cache->uc_freebucket != NULL) 651 bucket_free(cache->uc_freebucket); 652 cache->uc_allocbucket = cache->uc_freebucket = NULL; 653 } 654 ZONE_LOCK(zone); 655 bucket_cache_drain(zone); 656 ZONE_UNLOCK(zone); 657 } 658 659 /* 660 * Drain the cached buckets from a zone. Expects a locked zone on entry. 661 */ 662 static void 663 bucket_cache_drain(uma_zone_t zone) 664 { 665 uma_bucket_t bucket; 666 667 /* 668 * Drain the bucket queues and free the buckets, we just keep two per 669 * cpu (alloc/free). 670 */ 671 while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) { 672 LIST_REMOVE(bucket, ub_link); 673 ZONE_UNLOCK(zone); 674 bucket_drain(zone, bucket); 675 bucket_free(bucket); 676 ZONE_LOCK(zone); 677 } 678 679 /* Now we do the free queue.. */ 680 while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { 681 LIST_REMOVE(bucket, ub_link); 682 bucket_free(bucket); 683 } 684 } 685 686 /* 687 * Frees pages from a keg back to the system. This is done on demand from 688 * the pageout daemon. 689 * 690 * Returns nothing. 691 */ 692 static void 693 keg_drain(uma_keg_t keg) 694 { 695 struct slabhead freeslabs = { 0 }; 696 uma_slab_t slab; 697 uma_slab_t n; 698 uint8_t flags; 699 uint8_t *mem; 700 int i; 701 702 /* 703 * We don't want to take pages from statically allocated kegs at this 704 * time 705 */ 706 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 707 return; 708 709 #ifdef UMA_DEBUG 710 printf("%s free items: %u\n", keg->uk_name, keg->uk_free); 711 #endif 712 KEG_LOCK(keg); 713 if (keg->uk_free == 0) 714 goto finished; 715 716 slab = LIST_FIRST(&keg->uk_free_slab); 717 while (slab) { 718 n = LIST_NEXT(slab, us_link); 719 720 /* We have no where to free these to */ 721 if (slab->us_flags & UMA_SLAB_BOOT) { 722 slab = n; 723 continue; 724 } 725 726 LIST_REMOVE(slab, us_link); 727 keg->uk_pages -= keg->uk_ppera; 728 keg->uk_free -= keg->uk_ipers; 729 730 if (keg->uk_flags & UMA_ZONE_HASH) 731 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); 732 733 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 734 735 slab = n; 736 } 737 finished: 738 KEG_UNLOCK(keg); 739 740 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 741 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 742 if (keg->uk_fini) 743 for (i = 0; i < keg->uk_ipers; i++) 744 keg->uk_fini( 745 slab->us_data + (keg->uk_rsize * i), 746 keg->uk_size); 747 flags = slab->us_flags; 748 mem = slab->us_data; 749 750 if (keg->uk_flags & UMA_ZONE_VTOSLAB) { 751 vm_object_t obj; 752 753 if (flags & UMA_SLAB_KMEM) 754 obj = kmem_object; 755 else if (flags & UMA_SLAB_KERNEL) 756 obj = kernel_object; 757 else 758 obj = NULL; 759 for (i = 0; i < keg->uk_ppera; i++) 760 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE), 761 obj); 762 } 763 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 764 zone_free_item(keg->uk_slabzone, slab, NULL, 765 SKIP_NONE, ZFREE_STATFREE); 766 #ifdef UMA_DEBUG 767 printf("%s: Returning %d bytes.\n", 768 keg->uk_name, PAGE_SIZE * keg->uk_ppera); 769 #endif 770 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 771 } 772 } 773 774 static void 775 zone_drain_wait(uma_zone_t zone, int waitok) 776 { 777 778 /* 779 * Set draining to interlock with zone_dtor() so we can release our 780 * locks as we go. Only dtor() should do a WAITOK call since it 781 * is the only call that knows the structure will still be available 782 * when it wakes up. 783 */ 784 ZONE_LOCK(zone); 785 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 786 if (waitok == M_NOWAIT) 787 goto out; 788 mtx_unlock(&uma_mtx); 789 msleep(zone, zone->uz_lock, PVM, "zonedrain", 1); 790 mtx_lock(&uma_mtx); 791 } 792 zone->uz_flags |= UMA_ZFLAG_DRAINING; 793 bucket_cache_drain(zone); 794 ZONE_UNLOCK(zone); 795 /* 796 * The DRAINING flag protects us from being freed while 797 * we're running. Normally the uma_mtx would protect us but we 798 * must be able to release and acquire the right lock for each keg. 799 */ 800 zone_foreach_keg(zone, &keg_drain); 801 ZONE_LOCK(zone); 802 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 803 wakeup(zone); 804 out: 805 ZONE_UNLOCK(zone); 806 } 807 808 void 809 zone_drain(uma_zone_t zone) 810 { 811 812 zone_drain_wait(zone, M_NOWAIT); 813 } 814 815 /* 816 * Allocate a new slab for a keg. This does not insert the slab onto a list. 817 * 818 * Arguments: 819 * wait Shall we wait? 820 * 821 * Returns: 822 * The slab that was allocated or NULL if there is no memory and the 823 * caller specified M_NOWAIT. 824 */ 825 static uma_slab_t 826 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) 827 { 828 uma_slabrefcnt_t slabref; 829 uma_alloc allocf; 830 uma_slab_t slab; 831 uint8_t *mem; 832 uint8_t flags; 833 int i; 834 835 mtx_assert(&keg->uk_lock, MA_OWNED); 836 slab = NULL; 837 838 #ifdef UMA_DEBUG 839 printf("slab_zalloc: Allocating a new slab for %s\n", keg->uk_name); 840 #endif 841 allocf = keg->uk_allocf; 842 KEG_UNLOCK(keg); 843 844 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 845 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); 846 if (slab == NULL) { 847 KEG_LOCK(keg); 848 return NULL; 849 } 850 } 851 852 /* 853 * This reproduces the old vm_zone behavior of zero filling pages the 854 * first time they are added to a zone. 855 * 856 * Malloced items are zeroed in uma_zalloc. 857 */ 858 859 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 860 wait |= M_ZERO; 861 else 862 wait &= ~M_ZERO; 863 864 if (keg->uk_flags & UMA_ZONE_NODUMP) 865 wait |= M_NODUMP; 866 867 /* zone is passed for legacy reasons. */ 868 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait); 869 if (mem == NULL) { 870 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 871 zone_free_item(keg->uk_slabzone, slab, NULL, 872 SKIP_NONE, ZFREE_STATFREE); 873 KEG_LOCK(keg); 874 return (NULL); 875 } 876 877 /* Point the slab into the allocated memory */ 878 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 879 slab = (uma_slab_t )(mem + keg->uk_pgoff); 880 881 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 882 for (i = 0; i < keg->uk_ppera; i++) 883 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 884 885 slab->us_keg = keg; 886 slab->us_data = mem; 887 slab->us_freecount = keg->uk_ipers; 888 slab->us_firstfree = 0; 889 slab->us_flags = flags; 890 891 if (keg->uk_flags & UMA_ZONE_REFCNT) { 892 slabref = (uma_slabrefcnt_t)slab; 893 for (i = 0; i < keg->uk_ipers; i++) { 894 slabref->us_freelist[i].us_refcnt = 0; 895 slabref->us_freelist[i].us_item = i+1; 896 } 897 } else { 898 for (i = 0; i < keg->uk_ipers; i++) 899 slab->us_freelist[i].us_item = i+1; 900 } 901 902 if (keg->uk_init != NULL) { 903 for (i = 0; i < keg->uk_ipers; i++) 904 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 905 keg->uk_size, wait) != 0) 906 break; 907 if (i != keg->uk_ipers) { 908 if (keg->uk_fini != NULL) { 909 for (i--; i > -1; i--) 910 keg->uk_fini(slab->us_data + 911 (keg->uk_rsize * i), 912 keg->uk_size); 913 } 914 if (keg->uk_flags & UMA_ZONE_VTOSLAB) { 915 vm_object_t obj; 916 917 if (flags & UMA_SLAB_KMEM) 918 obj = kmem_object; 919 else if (flags & UMA_SLAB_KERNEL) 920 obj = kernel_object; 921 else 922 obj = NULL; 923 for (i = 0; i < keg->uk_ppera; i++) 924 vsetobj((vm_offset_t)mem + 925 (i * PAGE_SIZE), obj); 926 } 927 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 928 zone_free_item(keg->uk_slabzone, slab, 929 NULL, SKIP_NONE, ZFREE_STATFREE); 930 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, 931 flags); 932 KEG_LOCK(keg); 933 return (NULL); 934 } 935 } 936 KEG_LOCK(keg); 937 938 if (keg->uk_flags & UMA_ZONE_HASH) 939 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 940 941 keg->uk_pages += keg->uk_ppera; 942 keg->uk_free += keg->uk_ipers; 943 944 return (slab); 945 } 946 947 /* 948 * This function is intended to be used early on in place of page_alloc() so 949 * that we may use the boot time page cache to satisfy allocations before 950 * the VM is ready. 951 */ 952 static void * 953 startup_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait) 954 { 955 uma_keg_t keg; 956 uma_slab_t tmps; 957 int pages, check_pages; 958 959 keg = zone_first_keg(zone); 960 pages = howmany(bytes, PAGE_SIZE); 961 check_pages = pages - 1; 962 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n")); 963 964 /* 965 * Check our small startup cache to see if it has pages remaining. 966 */ 967 mtx_lock(&uma_boot_pages_mtx); 968 969 /* First check if we have enough room. */ 970 tmps = LIST_FIRST(&uma_boot_pages); 971 while (tmps != NULL && check_pages-- > 0) 972 tmps = LIST_NEXT(tmps, us_link); 973 if (tmps != NULL) { 974 /* 975 * It's ok to lose tmps references. The last one will 976 * have tmps->us_data pointing to the start address of 977 * "pages" contiguous pages of memory. 978 */ 979 while (pages-- > 0) { 980 tmps = LIST_FIRST(&uma_boot_pages); 981 LIST_REMOVE(tmps, us_link); 982 } 983 mtx_unlock(&uma_boot_pages_mtx); 984 *pflag = tmps->us_flags; 985 return (tmps->us_data); 986 } 987 mtx_unlock(&uma_boot_pages_mtx); 988 if (booted < UMA_STARTUP2) 989 panic("UMA: Increase vm.boot_pages"); 990 /* 991 * Now that we've booted reset these users to their real allocator. 992 */ 993 #ifdef UMA_MD_SMALL_ALLOC 994 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; 995 #else 996 keg->uk_allocf = page_alloc; 997 #endif 998 return keg->uk_allocf(zone, bytes, pflag, wait); 999 } 1000 1001 /* 1002 * Allocates a number of pages from the system 1003 * 1004 * Arguments: 1005 * bytes The number of bytes requested 1006 * wait Shall we wait? 1007 * 1008 * Returns: 1009 * A pointer to the alloced memory or possibly 1010 * NULL if M_NOWAIT is set. 1011 */ 1012 static void * 1013 page_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait) 1014 { 1015 void *p; /* Returned page */ 1016 1017 *pflag = UMA_SLAB_KMEM; 1018 p = (void *) kmem_malloc(kmem_map, bytes, wait); 1019 1020 return (p); 1021 } 1022 1023 /* 1024 * Allocates a number of pages from within an object 1025 * 1026 * Arguments: 1027 * bytes The number of bytes requested 1028 * wait Shall we wait? 1029 * 1030 * Returns: 1031 * A pointer to the alloced memory or possibly 1032 * NULL if M_NOWAIT is set. 1033 */ 1034 static void * 1035 noobj_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait) 1036 { 1037 TAILQ_HEAD(, vm_page) alloctail; 1038 u_long npages; 1039 vm_offset_t retkva, zkva; 1040 vm_page_t p, p_next; 1041 uma_keg_t keg; 1042 1043 TAILQ_INIT(&alloctail); 1044 keg = zone_first_keg(zone); 1045 1046 npages = howmany(bytes, PAGE_SIZE); 1047 while (npages > 0) { 1048 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT | 1049 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ); 1050 if (p != NULL) { 1051 /* 1052 * Since the page does not belong to an object, its 1053 * listq is unused. 1054 */ 1055 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1056 npages--; 1057 continue; 1058 } 1059 if (wait & M_WAITOK) { 1060 VM_WAIT; 1061 continue; 1062 } 1063 1064 /* 1065 * Page allocation failed, free intermediate pages and 1066 * exit. 1067 */ 1068 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1069 vm_page_unwire(p, 0); 1070 vm_page_free(p); 1071 } 1072 return (NULL); 1073 } 1074 *flags = UMA_SLAB_PRIV; 1075 zkva = keg->uk_kva + 1076 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1077 retkva = zkva; 1078 TAILQ_FOREACH(p, &alloctail, listq) { 1079 pmap_qenter(zkva, &p, 1); 1080 zkva += PAGE_SIZE; 1081 } 1082 1083 return ((void *)retkva); 1084 } 1085 1086 /* 1087 * Frees a number of pages to the system 1088 * 1089 * Arguments: 1090 * mem A pointer to the memory to be freed 1091 * size The size of the memory being freed 1092 * flags The original p->us_flags field 1093 * 1094 * Returns: 1095 * Nothing 1096 */ 1097 static void 1098 page_free(void *mem, int size, uint8_t flags) 1099 { 1100 vm_map_t map; 1101 1102 if (flags & UMA_SLAB_KMEM) 1103 map = kmem_map; 1104 else if (flags & UMA_SLAB_KERNEL) 1105 map = kernel_map; 1106 else 1107 panic("UMA: page_free used with invalid flags %d", flags); 1108 1109 kmem_free(map, (vm_offset_t)mem, size); 1110 } 1111 1112 /* 1113 * Zero fill initializer 1114 * 1115 * Arguments/Returns follow uma_init specifications 1116 */ 1117 static int 1118 zero_init(void *mem, int size, int flags) 1119 { 1120 bzero(mem, size); 1121 return (0); 1122 } 1123 1124 /* 1125 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1126 * 1127 * Arguments 1128 * keg The zone we should initialize 1129 * 1130 * Returns 1131 * Nothing 1132 */ 1133 static void 1134 keg_small_init(uma_keg_t keg) 1135 { 1136 u_int rsize; 1137 u_int memused; 1138 u_int wastedspace; 1139 u_int shsize; 1140 1141 if (keg->uk_flags & UMA_ZONE_PCPU) { 1142 KASSERT(mp_ncpus > 0, ("%s: ncpus %d\n", __func__, mp_ncpus)); 1143 keg->uk_slabsize = sizeof(struct pcpu); 1144 keg->uk_ppera = howmany(mp_ncpus * sizeof(struct pcpu), 1145 PAGE_SIZE); 1146 } else { 1147 keg->uk_slabsize = UMA_SLAB_SIZE; 1148 keg->uk_ppera = 1; 1149 } 1150 1151 rsize = keg->uk_size; 1152 1153 if (rsize & keg->uk_align) 1154 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1155 if (rsize < keg->uk_slabsize / 256) 1156 rsize = keg->uk_slabsize / 256; 1157 1158 keg->uk_rsize = rsize; 1159 1160 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1161 keg->uk_rsize < sizeof(struct pcpu), 1162 ("%s: size %u too large", __func__, keg->uk_rsize)); 1163 1164 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1165 shsize = 0; 1166 } else if (keg->uk_flags & UMA_ZONE_REFCNT) { 1167 rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */ 1168 shsize = sizeof(struct uma_slab_refcnt); 1169 } else { 1170 rsize += UMA_FRITM_SZ; /* Account for linkage */ 1171 shsize = sizeof(struct uma_slab); 1172 } 1173 1174 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize; 1175 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= 256, 1176 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1177 1178 memused = keg->uk_ipers * rsize + shsize; 1179 wastedspace = keg->uk_slabsize - memused; 1180 1181 /* 1182 * We can't do OFFPAGE if we're internal or if we've been 1183 * asked to not go to the VM for buckets. If we do this we 1184 * may end up going to the VM (kmem_map) for slabs which we 1185 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a 1186 * result of UMA_ZONE_VM, which clearly forbids it. 1187 */ 1188 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1189 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1190 return; 1191 1192 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) && 1193 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) { 1194 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize; 1195 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= 256, 1196 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1197 #ifdef UMA_DEBUG 1198 printf("UMA decided we need offpage slab headers for " 1199 "keg: %s, calculated wastedspace = %d, " 1200 "maximum wasted space allowed = %d, " 1201 "calculated ipers = %d, " 1202 "new wasted space = %d\n", keg->uk_name, wastedspace, 1203 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1204 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize); 1205 #endif 1206 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1207 } 1208 1209 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1210 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1211 keg->uk_flags |= UMA_ZONE_HASH; 1212 } 1213 1214 /* 1215 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1216 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1217 * more complicated. 1218 * 1219 * Arguments 1220 * keg The keg we should initialize 1221 * 1222 * Returns 1223 * Nothing 1224 */ 1225 static void 1226 keg_large_init(uma_keg_t keg) 1227 { 1228 1229 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1230 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1231 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1232 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1233 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1234 1235 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1236 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE; 1237 keg->uk_ipers = 1; 1238 keg->uk_rsize = keg->uk_size; 1239 1240 /* We can't do OFFPAGE if we're internal, bail out here. */ 1241 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) 1242 return; 1243 1244 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1245 if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1246 keg->uk_flags |= UMA_ZONE_HASH; 1247 } 1248 1249 static void 1250 keg_cachespread_init(uma_keg_t keg) 1251 { 1252 int alignsize; 1253 int trailer; 1254 int pages; 1255 int rsize; 1256 1257 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1258 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1259 1260 alignsize = keg->uk_align + 1; 1261 rsize = keg->uk_size; 1262 /* 1263 * We want one item to start on every align boundary in a page. To 1264 * do this we will span pages. We will also extend the item by the 1265 * size of align if it is an even multiple of align. Otherwise, it 1266 * would fall on the same boundary every time. 1267 */ 1268 if (rsize & keg->uk_align) 1269 rsize = (rsize & ~keg->uk_align) + alignsize; 1270 if ((rsize & alignsize) == 0) 1271 rsize += alignsize; 1272 trailer = rsize - keg->uk_size; 1273 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1274 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1275 keg->uk_rsize = rsize; 1276 keg->uk_ppera = pages; 1277 keg->uk_slabsize = UMA_SLAB_SIZE; 1278 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1279 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1280 KASSERT(keg->uk_ipers <= uma_max_ipers, 1281 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1282 keg->uk_ipers)); 1283 } 1284 1285 /* 1286 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1287 * the keg onto the global keg list. 1288 * 1289 * Arguments/Returns follow uma_ctor specifications 1290 * udata Actually uma_kctor_args 1291 */ 1292 static int 1293 keg_ctor(void *mem, int size, void *udata, int flags) 1294 { 1295 struct uma_kctor_args *arg = udata; 1296 uma_keg_t keg = mem; 1297 uma_zone_t zone; 1298 1299 bzero(keg, size); 1300 keg->uk_size = arg->size; 1301 keg->uk_init = arg->uminit; 1302 keg->uk_fini = arg->fini; 1303 keg->uk_align = arg->align; 1304 keg->uk_free = 0; 1305 keg->uk_pages = 0; 1306 keg->uk_flags = arg->flags; 1307 keg->uk_allocf = page_alloc; 1308 keg->uk_freef = page_free; 1309 keg->uk_recurse = 0; 1310 keg->uk_slabzone = NULL; 1311 1312 /* 1313 * The master zone is passed to us at keg-creation time. 1314 */ 1315 zone = arg->zone; 1316 keg->uk_name = zone->uz_name; 1317 1318 if (arg->flags & UMA_ZONE_VM) 1319 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1320 1321 if (arg->flags & UMA_ZONE_ZINIT) 1322 keg->uk_init = zero_init; 1323 1324 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC) 1325 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1326 1327 if (arg->flags & UMA_ZONE_PCPU) 1328 #ifdef SMP 1329 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1330 #else 1331 keg->uk_flags &= ~UMA_ZONE_PCPU; 1332 #endif 1333 1334 /* 1335 * The +UMA_FRITM_SZ added to uk_size is to account for the 1336 * linkage that is added to the size in keg_small_init(). If 1337 * we don't account for this here then we may end up in 1338 * keg_small_init() with a calculated 'ipers' of 0. 1339 */ 1340 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1341 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) 1342 keg_cachespread_init(keg); 1343 else if ((keg->uk_size+UMA_FRITMREF_SZ) > 1344 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt))) 1345 keg_large_init(keg); 1346 else 1347 keg_small_init(keg); 1348 } else { 1349 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) 1350 keg_cachespread_init(keg); 1351 else if ((keg->uk_size+UMA_FRITM_SZ) > 1352 (UMA_SLAB_SIZE - sizeof(struct uma_slab))) 1353 keg_large_init(keg); 1354 else 1355 keg_small_init(keg); 1356 } 1357 1358 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1359 if (keg->uk_flags & UMA_ZONE_REFCNT) 1360 keg->uk_slabzone = slabrefzone; 1361 else 1362 keg->uk_slabzone = slabzone; 1363 } 1364 1365 /* 1366 * If we haven't booted yet we need allocations to go through the 1367 * startup cache until the vm is ready. 1368 */ 1369 if (keg->uk_ppera == 1) { 1370 #ifdef UMA_MD_SMALL_ALLOC 1371 keg->uk_allocf = uma_small_alloc; 1372 keg->uk_freef = uma_small_free; 1373 1374 if (booted < UMA_STARTUP) 1375 keg->uk_allocf = startup_alloc; 1376 #else 1377 if (booted < UMA_STARTUP2) 1378 keg->uk_allocf = startup_alloc; 1379 #endif 1380 } else if (booted < UMA_STARTUP2 && 1381 (keg->uk_flags & UMA_ZFLAG_INTERNAL)) 1382 keg->uk_allocf = startup_alloc; 1383 1384 /* 1385 * Initialize keg's lock (shared among zones). 1386 */ 1387 if (arg->flags & UMA_ZONE_MTXCLASS) 1388 KEG_LOCK_INIT(keg, 1); 1389 else 1390 KEG_LOCK_INIT(keg, 0); 1391 1392 /* 1393 * If we're putting the slab header in the actual page we need to 1394 * figure out where in each page it goes. This calculates a right 1395 * justified offset into the memory on an ALIGN_PTR boundary. 1396 */ 1397 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1398 u_int totsize; 1399 1400 /* Size of the slab struct and free list */ 1401 if (keg->uk_flags & UMA_ZONE_REFCNT) 1402 totsize = sizeof(struct uma_slab_refcnt) + 1403 keg->uk_ipers * UMA_FRITMREF_SZ; 1404 else 1405 totsize = sizeof(struct uma_slab) + 1406 keg->uk_ipers * UMA_FRITM_SZ; 1407 1408 if (totsize & UMA_ALIGN_PTR) 1409 totsize = (totsize & ~UMA_ALIGN_PTR) + 1410 (UMA_ALIGN_PTR + 1); 1411 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1412 1413 if (keg->uk_flags & UMA_ZONE_REFCNT) 1414 totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt) 1415 + keg->uk_ipers * UMA_FRITMREF_SZ; 1416 else 1417 totsize = keg->uk_pgoff + sizeof(struct uma_slab) 1418 + keg->uk_ipers * UMA_FRITM_SZ; 1419 1420 /* 1421 * The only way the following is possible is if with our 1422 * UMA_ALIGN_PTR adjustments we are now bigger than 1423 * UMA_SLAB_SIZE. I haven't checked whether this is 1424 * mathematically possible for all cases, so we make 1425 * sure here anyway. 1426 */ 1427 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1428 printf("zone %s ipers %d rsize %d size %d\n", 1429 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1430 keg->uk_size); 1431 panic("UMA slab won't fit."); 1432 } 1433 } 1434 1435 if (keg->uk_flags & UMA_ZONE_HASH) 1436 hash_alloc(&keg->uk_hash); 1437 1438 #ifdef UMA_DEBUG 1439 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n", 1440 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags, 1441 keg->uk_ipers, keg->uk_ppera, 1442 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free); 1443 #endif 1444 1445 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1446 1447 mtx_lock(&uma_mtx); 1448 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1449 mtx_unlock(&uma_mtx); 1450 return (0); 1451 } 1452 1453 /* 1454 * Zone header ctor. This initializes all fields, locks, etc. 1455 * 1456 * Arguments/Returns follow uma_ctor specifications 1457 * udata Actually uma_zctor_args 1458 */ 1459 static int 1460 zone_ctor(void *mem, int size, void *udata, int flags) 1461 { 1462 struct uma_zctor_args *arg = udata; 1463 uma_zone_t zone = mem; 1464 uma_zone_t z; 1465 uma_keg_t keg; 1466 1467 bzero(zone, size); 1468 zone->uz_name = arg->name; 1469 zone->uz_ctor = arg->ctor; 1470 zone->uz_dtor = arg->dtor; 1471 zone->uz_slab = zone_fetch_slab; 1472 zone->uz_init = NULL; 1473 zone->uz_fini = NULL; 1474 zone->uz_allocs = 0; 1475 zone->uz_frees = 0; 1476 zone->uz_fails = 0; 1477 zone->uz_sleeps = 0; 1478 zone->uz_fills = zone->uz_count = 0; 1479 zone->uz_flags = 0; 1480 zone->uz_warning = NULL; 1481 timevalclear(&zone->uz_ratecheck); 1482 keg = arg->keg; 1483 1484 if (arg->flags & UMA_ZONE_SECONDARY) { 1485 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1486 zone->uz_init = arg->uminit; 1487 zone->uz_fini = arg->fini; 1488 zone->uz_lock = &keg->uk_lock; 1489 zone->uz_flags |= UMA_ZONE_SECONDARY; 1490 mtx_lock(&uma_mtx); 1491 ZONE_LOCK(zone); 1492 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1493 if (LIST_NEXT(z, uz_link) == NULL) { 1494 LIST_INSERT_AFTER(z, zone, uz_link); 1495 break; 1496 } 1497 } 1498 ZONE_UNLOCK(zone); 1499 mtx_unlock(&uma_mtx); 1500 } else if (keg == NULL) { 1501 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1502 arg->align, arg->flags)) == NULL) 1503 return (ENOMEM); 1504 } else { 1505 struct uma_kctor_args karg; 1506 int error; 1507 1508 /* We should only be here from uma_startup() */ 1509 karg.size = arg->size; 1510 karg.uminit = arg->uminit; 1511 karg.fini = arg->fini; 1512 karg.align = arg->align; 1513 karg.flags = arg->flags; 1514 karg.zone = zone; 1515 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1516 flags); 1517 if (error) 1518 return (error); 1519 } 1520 /* 1521 * Link in the first keg. 1522 */ 1523 zone->uz_klink.kl_keg = keg; 1524 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1525 zone->uz_lock = &keg->uk_lock; 1526 zone->uz_size = keg->uk_size; 1527 zone->uz_flags |= (keg->uk_flags & 1528 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1529 1530 /* 1531 * Some internal zones don't have room allocated for the per cpu 1532 * caches. If we're internal, bail out here. 1533 */ 1534 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1535 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1536 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1537 return (0); 1538 } 1539 1540 if (keg->uk_flags & UMA_ZONE_MAXBUCKET) 1541 zone->uz_count = BUCKET_MAX; 1542 else if (keg->uk_ipers <= BUCKET_MAX) 1543 zone->uz_count = keg->uk_ipers; 1544 else 1545 zone->uz_count = BUCKET_MAX; 1546 return (0); 1547 } 1548 1549 /* 1550 * Keg header dtor. This frees all data, destroys locks, frees the hash 1551 * table and removes the keg from the global list. 1552 * 1553 * Arguments/Returns follow uma_dtor specifications 1554 * udata unused 1555 */ 1556 static void 1557 keg_dtor(void *arg, int size, void *udata) 1558 { 1559 uma_keg_t keg; 1560 1561 keg = (uma_keg_t)arg; 1562 KEG_LOCK(keg); 1563 if (keg->uk_free != 0) { 1564 printf("Freed UMA keg was not empty (%d items). " 1565 " Lost %d pages of memory.\n", 1566 keg->uk_free, keg->uk_pages); 1567 } 1568 KEG_UNLOCK(keg); 1569 1570 hash_free(&keg->uk_hash); 1571 1572 KEG_LOCK_FINI(keg); 1573 } 1574 1575 /* 1576 * Zone header dtor. 1577 * 1578 * Arguments/Returns follow uma_dtor specifications 1579 * udata unused 1580 */ 1581 static void 1582 zone_dtor(void *arg, int size, void *udata) 1583 { 1584 uma_klink_t klink; 1585 uma_zone_t zone; 1586 uma_keg_t keg; 1587 1588 zone = (uma_zone_t)arg; 1589 keg = zone_first_keg(zone); 1590 1591 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1592 cache_drain(zone); 1593 1594 mtx_lock(&uma_mtx); 1595 LIST_REMOVE(zone, uz_link); 1596 mtx_unlock(&uma_mtx); 1597 /* 1598 * XXX there are some races here where 1599 * the zone can be drained but zone lock 1600 * released and then refilled before we 1601 * remove it... we dont care for now 1602 */ 1603 zone_drain_wait(zone, M_WAITOK); 1604 /* 1605 * Unlink all of our kegs. 1606 */ 1607 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1608 klink->kl_keg = NULL; 1609 LIST_REMOVE(klink, kl_link); 1610 if (klink == &zone->uz_klink) 1611 continue; 1612 free(klink, M_TEMP); 1613 } 1614 /* 1615 * We only destroy kegs from non secondary zones. 1616 */ 1617 if ((zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1618 mtx_lock(&uma_mtx); 1619 LIST_REMOVE(keg, uk_link); 1620 mtx_unlock(&uma_mtx); 1621 zone_free_item(kegs, keg, NULL, SKIP_NONE, 1622 ZFREE_STATFREE); 1623 } 1624 } 1625 1626 /* 1627 * Traverses every zone in the system and calls a callback 1628 * 1629 * Arguments: 1630 * zfunc A pointer to a function which accepts a zone 1631 * as an argument. 1632 * 1633 * Returns: 1634 * Nothing 1635 */ 1636 static void 1637 zone_foreach(void (*zfunc)(uma_zone_t)) 1638 { 1639 uma_keg_t keg; 1640 uma_zone_t zone; 1641 1642 mtx_lock(&uma_mtx); 1643 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1644 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1645 zfunc(zone); 1646 } 1647 mtx_unlock(&uma_mtx); 1648 } 1649 1650 /* Public functions */ 1651 /* See uma.h */ 1652 void 1653 uma_startup(void *bootmem, int boot_pages) 1654 { 1655 struct uma_zctor_args args; 1656 uma_slab_t slab; 1657 u_int slabsize; 1658 u_int objsize, totsize, wsize; 1659 int i; 1660 1661 #ifdef UMA_DEBUG 1662 printf("Creating uma keg headers zone and keg.\n"); 1663 #endif 1664 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF); 1665 1666 /* 1667 * Figure out the maximum number of items-per-slab we'll have if 1668 * we're using the OFFPAGE slab header to track free items, given 1669 * all possible object sizes and the maximum desired wastage 1670 * (UMA_MAX_WASTE). 1671 * 1672 * We iterate until we find an object size for 1673 * which the calculated wastage in keg_small_init() will be 1674 * enough to warrant OFFPAGE. Since wastedspace versus objsize 1675 * is an overall increasing see-saw function, we find the smallest 1676 * objsize such that the wastage is always acceptable for objects 1677 * with that objsize or smaller. Since a smaller objsize always 1678 * generates a larger possible uma_max_ipers, we use this computed 1679 * objsize to calculate the largest ipers possible. Since the 1680 * ipers calculated for OFFPAGE slab headers is always larger than 1681 * the ipers initially calculated in keg_small_init(), we use 1682 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to 1683 * obtain the maximum ipers possible for offpage slab headers. 1684 * 1685 * It should be noted that ipers versus objsize is an inversly 1686 * proportional function which drops off rather quickly so as 1687 * long as our UMA_MAX_WASTE is such that the objsize we calculate 1688 * falls into the portion of the inverse relation AFTER the steep 1689 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386). 1690 * 1691 * Note that we have 8-bits (1 byte) to use as a freelist index 1692 * inside the actual slab header itself and this is enough to 1693 * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized 1694 * object with offpage slab header would have ipers = 1695 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is 1696 * 1 greater than what our byte-integer freelist index can 1697 * accomodate, but we know that this situation never occurs as 1698 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate 1699 * that we need to go to offpage slab headers. Or, if we do, 1700 * then we trap that condition below and panic in the INVARIANTS case. 1701 */ 1702 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - 1703 (UMA_SLAB_SIZE / UMA_MAX_WASTE); 1704 totsize = wsize; 1705 objsize = UMA_SMALLEST_UNIT; 1706 while (totsize >= wsize) { 1707 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) / 1708 (objsize + UMA_FRITM_SZ); 1709 totsize *= (UMA_FRITM_SZ + objsize); 1710 objsize++; 1711 } 1712 if (objsize > UMA_SMALLEST_UNIT) 1713 objsize--; 1714 uma_max_ipers = MAX(UMA_SLAB_SIZE / objsize, 64); 1715 1716 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - 1717 (UMA_SLAB_SIZE / UMA_MAX_WASTE); 1718 totsize = wsize; 1719 objsize = UMA_SMALLEST_UNIT; 1720 while (totsize >= wsize) { 1721 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) / 1722 (objsize + UMA_FRITMREF_SZ); 1723 totsize *= (UMA_FRITMREF_SZ + objsize); 1724 objsize++; 1725 } 1726 if (objsize > UMA_SMALLEST_UNIT) 1727 objsize--; 1728 uma_max_ipers_ref = MAX(UMA_SLAB_SIZE / objsize, 64); 1729 1730 KASSERT((uma_max_ipers_ref <= 256) && (uma_max_ipers <= 256), 1731 ("uma_startup: calculated uma_max_ipers values too large!")); 1732 1733 #ifdef UMA_DEBUG 1734 printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers); 1735 printf("Calculated uma_max_ipers_ref (for OFFPAGE) is %d\n", 1736 uma_max_ipers_ref); 1737 #endif 1738 1739 /* "manually" create the initial zone */ 1740 args.name = "UMA Kegs"; 1741 args.size = sizeof(struct uma_keg); 1742 args.ctor = keg_ctor; 1743 args.dtor = keg_dtor; 1744 args.uminit = zero_init; 1745 args.fini = NULL; 1746 args.keg = &masterkeg; 1747 args.align = 32 - 1; 1748 args.flags = UMA_ZFLAG_INTERNAL; 1749 /* The initial zone has no Per cpu queues so it's smaller */ 1750 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); 1751 1752 #ifdef UMA_DEBUG 1753 printf("Filling boot free list.\n"); 1754 #endif 1755 for (i = 0; i < boot_pages; i++) { 1756 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE)); 1757 slab->us_data = (uint8_t *)slab; 1758 slab->us_flags = UMA_SLAB_BOOT; 1759 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); 1760 } 1761 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); 1762 1763 #ifdef UMA_DEBUG 1764 printf("Creating uma zone headers zone and keg.\n"); 1765 #endif 1766 args.name = "UMA Zones"; 1767 args.size = sizeof(struct uma_zone) + 1768 (sizeof(struct uma_cache) * (mp_maxid + 1)); 1769 args.ctor = zone_ctor; 1770 args.dtor = zone_dtor; 1771 args.uminit = zero_init; 1772 args.fini = NULL; 1773 args.keg = NULL; 1774 args.align = 32 - 1; 1775 args.flags = UMA_ZFLAG_INTERNAL; 1776 /* The initial zone has no Per cpu queues so it's smaller */ 1777 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); 1778 1779 #ifdef UMA_DEBUG 1780 printf("Initializing pcpu cache locks.\n"); 1781 #endif 1782 #ifdef UMA_DEBUG 1783 printf("Creating slab and hash zones.\n"); 1784 #endif 1785 1786 /* 1787 * This is the max number of free list items we'll have with 1788 * offpage slabs. 1789 */ 1790 slabsize = uma_max_ipers * UMA_FRITM_SZ; 1791 slabsize += sizeof(struct uma_slab); 1792 1793 /* Now make a zone for slab headers */ 1794 slabzone = uma_zcreate("UMA Slabs", 1795 slabsize, 1796 NULL, NULL, NULL, NULL, 1797 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1798 1799 /* 1800 * We also create a zone for the bigger slabs with reference 1801 * counts in them, to accomodate UMA_ZONE_REFCNT zones. 1802 */ 1803 slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ; 1804 slabsize += sizeof(struct uma_slab_refcnt); 1805 slabrefzone = uma_zcreate("UMA RCntSlabs", 1806 slabsize, 1807 NULL, NULL, NULL, NULL, 1808 UMA_ALIGN_PTR, 1809 UMA_ZFLAG_INTERNAL); 1810 1811 hashzone = uma_zcreate("UMA Hash", 1812 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 1813 NULL, NULL, NULL, NULL, 1814 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1815 1816 bucket_init(); 1817 1818 booted = UMA_STARTUP; 1819 1820 #ifdef UMA_DEBUG 1821 printf("UMA startup complete.\n"); 1822 #endif 1823 } 1824 1825 /* see uma.h */ 1826 void 1827 uma_startup2(void) 1828 { 1829 booted = UMA_STARTUP2; 1830 bucket_enable(); 1831 #ifdef UMA_DEBUG 1832 printf("UMA startup2 complete.\n"); 1833 #endif 1834 } 1835 1836 /* 1837 * Initialize our callout handle 1838 * 1839 */ 1840 1841 static void 1842 uma_startup3(void) 1843 { 1844 #ifdef UMA_DEBUG 1845 printf("Starting callout.\n"); 1846 #endif 1847 callout_init(&uma_callout, CALLOUT_MPSAFE); 1848 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1849 #ifdef UMA_DEBUG 1850 printf("UMA startup3 complete.\n"); 1851 #endif 1852 } 1853 1854 static uma_keg_t 1855 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 1856 int align, uint32_t flags) 1857 { 1858 struct uma_kctor_args args; 1859 1860 args.size = size; 1861 args.uminit = uminit; 1862 args.fini = fini; 1863 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 1864 args.flags = flags; 1865 args.zone = zone; 1866 return (zone_alloc_item(kegs, &args, M_WAITOK)); 1867 } 1868 1869 /* See uma.h */ 1870 void 1871 uma_set_align(int align) 1872 { 1873 1874 if (align != UMA_ALIGN_CACHE) 1875 uma_align_cache = align; 1876 } 1877 1878 /* See uma.h */ 1879 uma_zone_t 1880 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 1881 uma_init uminit, uma_fini fini, int align, uint32_t flags) 1882 1883 { 1884 struct uma_zctor_args args; 1885 1886 /* This stuff is essential for the zone ctor */ 1887 args.name = name; 1888 args.size = size; 1889 args.ctor = ctor; 1890 args.dtor = dtor; 1891 args.uminit = uminit; 1892 args.fini = fini; 1893 args.align = align; 1894 args.flags = flags; 1895 args.keg = NULL; 1896 1897 return (zone_alloc_item(zones, &args, M_WAITOK)); 1898 } 1899 1900 /* See uma.h */ 1901 uma_zone_t 1902 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 1903 uma_init zinit, uma_fini zfini, uma_zone_t master) 1904 { 1905 struct uma_zctor_args args; 1906 uma_keg_t keg; 1907 1908 keg = zone_first_keg(master); 1909 args.name = name; 1910 args.size = keg->uk_size; 1911 args.ctor = ctor; 1912 args.dtor = dtor; 1913 args.uminit = zinit; 1914 args.fini = zfini; 1915 args.align = keg->uk_align; 1916 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 1917 args.keg = keg; 1918 1919 /* XXX Attaches only one keg of potentially many. */ 1920 return (zone_alloc_item(zones, &args, M_WAITOK)); 1921 } 1922 1923 static void 1924 zone_lock_pair(uma_zone_t a, uma_zone_t b) 1925 { 1926 if (a < b) { 1927 ZONE_LOCK(a); 1928 mtx_lock_flags(b->uz_lock, MTX_DUPOK); 1929 } else { 1930 ZONE_LOCK(b); 1931 mtx_lock_flags(a->uz_lock, MTX_DUPOK); 1932 } 1933 } 1934 1935 static void 1936 zone_unlock_pair(uma_zone_t a, uma_zone_t b) 1937 { 1938 1939 ZONE_UNLOCK(a); 1940 ZONE_UNLOCK(b); 1941 } 1942 1943 int 1944 uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 1945 { 1946 uma_klink_t klink; 1947 uma_klink_t kl; 1948 int error; 1949 1950 error = 0; 1951 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 1952 1953 zone_lock_pair(zone, master); 1954 /* 1955 * zone must use vtoslab() to resolve objects and must already be 1956 * a secondary. 1957 */ 1958 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 1959 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 1960 error = EINVAL; 1961 goto out; 1962 } 1963 /* 1964 * The new master must also use vtoslab(). 1965 */ 1966 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 1967 error = EINVAL; 1968 goto out; 1969 } 1970 /* 1971 * Both must either be refcnt, or not be refcnt. 1972 */ 1973 if ((zone->uz_flags & UMA_ZONE_REFCNT) != 1974 (master->uz_flags & UMA_ZONE_REFCNT)) { 1975 error = EINVAL; 1976 goto out; 1977 } 1978 /* 1979 * The underlying object must be the same size. rsize 1980 * may be different. 1981 */ 1982 if (master->uz_size != zone->uz_size) { 1983 error = E2BIG; 1984 goto out; 1985 } 1986 /* 1987 * Put it at the end of the list. 1988 */ 1989 klink->kl_keg = zone_first_keg(master); 1990 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 1991 if (LIST_NEXT(kl, kl_link) == NULL) { 1992 LIST_INSERT_AFTER(kl, klink, kl_link); 1993 break; 1994 } 1995 } 1996 klink = NULL; 1997 zone->uz_flags |= UMA_ZFLAG_MULTI; 1998 zone->uz_slab = zone_fetch_slab_multi; 1999 2000 out: 2001 zone_unlock_pair(zone, master); 2002 if (klink != NULL) 2003 free(klink, M_TEMP); 2004 2005 return (error); 2006 } 2007 2008 2009 /* See uma.h */ 2010 void 2011 uma_zdestroy(uma_zone_t zone) 2012 { 2013 2014 zone_free_item(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE); 2015 } 2016 2017 /* See uma.h */ 2018 void * 2019 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2020 { 2021 void *item; 2022 uma_cache_t cache; 2023 uma_bucket_t bucket; 2024 int cpu; 2025 2026 /* This is the fast path allocation */ 2027 #ifdef UMA_DEBUG_ALLOC_1 2028 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); 2029 #endif 2030 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread, 2031 zone->uz_name, flags); 2032 2033 if (flags & M_WAITOK) { 2034 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2035 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2036 } 2037 #ifdef DEBUG_MEMGUARD 2038 if (memguard_cmp_zone(zone)) { 2039 item = memguard_alloc(zone->uz_size, flags); 2040 if (item != NULL) { 2041 /* 2042 * Avoid conflict with the use-after-free 2043 * protecting infrastructure from INVARIANTS. 2044 */ 2045 if (zone->uz_init != NULL && 2046 zone->uz_init != mtrash_init && 2047 zone->uz_init(item, zone->uz_size, flags) != 0) 2048 return (NULL); 2049 if (zone->uz_ctor != NULL && 2050 zone->uz_ctor != mtrash_ctor && 2051 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2052 zone->uz_fini(item, zone->uz_size); 2053 return (NULL); 2054 } 2055 return (item); 2056 } 2057 /* This is unfortunate but should not be fatal. */ 2058 } 2059 #endif 2060 /* 2061 * If possible, allocate from the per-CPU cache. There are two 2062 * requirements for safe access to the per-CPU cache: (1) the thread 2063 * accessing the cache must not be preempted or yield during access, 2064 * and (2) the thread must not migrate CPUs without switching which 2065 * cache it accesses. We rely on a critical section to prevent 2066 * preemption and migration. We release the critical section in 2067 * order to acquire the zone mutex if we are unable to allocate from 2068 * the current cache; when we re-acquire the critical section, we 2069 * must detect and handle migration if it has occurred. 2070 */ 2071 zalloc_restart: 2072 critical_enter(); 2073 cpu = curcpu; 2074 cache = &zone->uz_cpu[cpu]; 2075 2076 zalloc_start: 2077 bucket = cache->uc_allocbucket; 2078 2079 if (bucket) { 2080 if (bucket->ub_cnt > 0) { 2081 bucket->ub_cnt--; 2082 item = bucket->ub_bucket[bucket->ub_cnt]; 2083 #ifdef INVARIANTS 2084 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2085 #endif 2086 KASSERT(item != NULL, 2087 ("uma_zalloc: Bucket pointer mangled.")); 2088 cache->uc_allocs++; 2089 critical_exit(); 2090 #ifdef INVARIANTS 2091 ZONE_LOCK(zone); 2092 uma_dbg_alloc(zone, NULL, item); 2093 ZONE_UNLOCK(zone); 2094 #endif 2095 if (zone->uz_ctor != NULL) { 2096 if (zone->uz_ctor(item, zone->uz_size, 2097 udata, flags) != 0) { 2098 zone_free_item(zone, item, udata, 2099 SKIP_DTOR, ZFREE_STATFAIL | 2100 ZFREE_STATFREE); 2101 return (NULL); 2102 } 2103 } 2104 if (flags & M_ZERO) 2105 bzero(item, zone->uz_size); 2106 return (item); 2107 } else if (cache->uc_freebucket) { 2108 /* 2109 * We have run out of items in our allocbucket. 2110 * See if we can switch with our free bucket. 2111 */ 2112 if (cache->uc_freebucket->ub_cnt > 0) { 2113 #ifdef UMA_DEBUG_ALLOC 2114 printf("uma_zalloc: Swapping empty with" 2115 " alloc.\n"); 2116 #endif 2117 bucket = cache->uc_freebucket; 2118 cache->uc_freebucket = cache->uc_allocbucket; 2119 cache->uc_allocbucket = bucket; 2120 2121 goto zalloc_start; 2122 } 2123 } 2124 } 2125 /* 2126 * Attempt to retrieve the item from the per-CPU cache has failed, so 2127 * we must go back to the zone. This requires the zone lock, so we 2128 * must drop the critical section, then re-acquire it when we go back 2129 * to the cache. Since the critical section is released, we may be 2130 * preempted or migrate. As such, make sure not to maintain any 2131 * thread-local state specific to the cache from prior to releasing 2132 * the critical section. 2133 */ 2134 critical_exit(); 2135 ZONE_LOCK(zone); 2136 critical_enter(); 2137 cpu = curcpu; 2138 cache = &zone->uz_cpu[cpu]; 2139 bucket = cache->uc_allocbucket; 2140 if (bucket != NULL) { 2141 if (bucket->ub_cnt > 0) { 2142 ZONE_UNLOCK(zone); 2143 goto zalloc_start; 2144 } 2145 bucket = cache->uc_freebucket; 2146 if (bucket != NULL && bucket->ub_cnt > 0) { 2147 ZONE_UNLOCK(zone); 2148 goto zalloc_start; 2149 } 2150 } 2151 2152 /* Since we have locked the zone we may as well send back our stats */ 2153 zone->uz_allocs += cache->uc_allocs; 2154 cache->uc_allocs = 0; 2155 zone->uz_frees += cache->uc_frees; 2156 cache->uc_frees = 0; 2157 2158 /* Our old one is now a free bucket */ 2159 if (cache->uc_allocbucket) { 2160 KASSERT(cache->uc_allocbucket->ub_cnt == 0, 2161 ("uma_zalloc_arg: Freeing a non free bucket.")); 2162 LIST_INSERT_HEAD(&zone->uz_free_bucket, 2163 cache->uc_allocbucket, ub_link); 2164 cache->uc_allocbucket = NULL; 2165 } 2166 2167 /* Check the free list for a new alloc bucket */ 2168 if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) { 2169 KASSERT(bucket->ub_cnt != 0, 2170 ("uma_zalloc_arg: Returning an empty bucket.")); 2171 2172 LIST_REMOVE(bucket, ub_link); 2173 cache->uc_allocbucket = bucket; 2174 ZONE_UNLOCK(zone); 2175 goto zalloc_start; 2176 } 2177 /* We are no longer associated with this CPU. */ 2178 critical_exit(); 2179 2180 /* Bump up our uz_count so we get here less */ 2181 if (zone->uz_count < BUCKET_MAX) 2182 zone->uz_count++; 2183 2184 /* 2185 * Now lets just fill a bucket and put it on the free list. If that 2186 * works we'll restart the allocation from the begining. 2187 */ 2188 if (zone_alloc_bucket(zone, flags)) { 2189 ZONE_UNLOCK(zone); 2190 goto zalloc_restart; 2191 } 2192 ZONE_UNLOCK(zone); 2193 /* 2194 * We may not be able to get a bucket so return an actual item. 2195 */ 2196 #ifdef UMA_DEBUG 2197 printf("uma_zalloc_arg: Bucketzone returned NULL\n"); 2198 #endif 2199 2200 item = zone_alloc_item(zone, udata, flags); 2201 return (item); 2202 } 2203 2204 static uma_slab_t 2205 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) 2206 { 2207 uma_slab_t slab; 2208 2209 mtx_assert(&keg->uk_lock, MA_OWNED); 2210 slab = NULL; 2211 2212 for (;;) { 2213 /* 2214 * Find a slab with some space. Prefer slabs that are partially 2215 * used over those that are totally full. This helps to reduce 2216 * fragmentation. 2217 */ 2218 if (keg->uk_free != 0) { 2219 if (!LIST_EMPTY(&keg->uk_part_slab)) { 2220 slab = LIST_FIRST(&keg->uk_part_slab); 2221 } else { 2222 slab = LIST_FIRST(&keg->uk_free_slab); 2223 LIST_REMOVE(slab, us_link); 2224 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, 2225 us_link); 2226 } 2227 MPASS(slab->us_keg == keg); 2228 return (slab); 2229 } 2230 2231 /* 2232 * M_NOVM means don't ask at all! 2233 */ 2234 if (flags & M_NOVM) 2235 break; 2236 2237 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2238 keg->uk_flags |= UMA_ZFLAG_FULL; 2239 /* 2240 * If this is not a multi-zone, set the FULL bit. 2241 * Otherwise slab_multi() takes care of it. 2242 */ 2243 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2244 zone->uz_flags |= UMA_ZFLAG_FULL; 2245 zone_log_warning(zone); 2246 } 2247 if (flags & M_NOWAIT) 2248 break; 2249 zone->uz_sleeps++; 2250 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2251 continue; 2252 } 2253 keg->uk_recurse++; 2254 slab = keg_alloc_slab(keg, zone, flags); 2255 keg->uk_recurse--; 2256 /* 2257 * If we got a slab here it's safe to mark it partially used 2258 * and return. We assume that the caller is going to remove 2259 * at least one item. 2260 */ 2261 if (slab) { 2262 MPASS(slab->us_keg == keg); 2263 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2264 return (slab); 2265 } 2266 /* 2267 * We might not have been able to get a slab but another cpu 2268 * could have while we were unlocked. Check again before we 2269 * fail. 2270 */ 2271 flags |= M_NOVM; 2272 } 2273 return (slab); 2274 } 2275 2276 static inline void 2277 zone_relock(uma_zone_t zone, uma_keg_t keg) 2278 { 2279 if (zone->uz_lock != &keg->uk_lock) { 2280 KEG_UNLOCK(keg); 2281 ZONE_LOCK(zone); 2282 } 2283 } 2284 2285 static inline void 2286 keg_relock(uma_keg_t keg, uma_zone_t zone) 2287 { 2288 if (zone->uz_lock != &keg->uk_lock) { 2289 ZONE_UNLOCK(zone); 2290 KEG_LOCK(keg); 2291 } 2292 } 2293 2294 static uma_slab_t 2295 zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) 2296 { 2297 uma_slab_t slab; 2298 2299 if (keg == NULL) 2300 keg = zone_first_keg(zone); 2301 /* 2302 * This is to prevent us from recursively trying to allocate 2303 * buckets. The problem is that if an allocation forces us to 2304 * grab a new bucket we will call page_alloc, which will go off 2305 * and cause the vm to allocate vm_map_entries. If we need new 2306 * buckets there too we will recurse in kmem_alloc and bad 2307 * things happen. So instead we return a NULL bucket, and make 2308 * the code that allocates buckets smart enough to deal with it 2309 */ 2310 if (keg->uk_flags & UMA_ZFLAG_BUCKET && keg->uk_recurse != 0) 2311 return (NULL); 2312 2313 for (;;) { 2314 slab = keg_fetch_slab(keg, zone, flags); 2315 if (slab) 2316 return (slab); 2317 if (flags & (M_NOWAIT | M_NOVM)) 2318 break; 2319 } 2320 return (NULL); 2321 } 2322 2323 /* 2324 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2325 * with the keg locked. Caller must call zone_relock() afterwards if the 2326 * zone lock is required. On NULL the zone lock is held. 2327 * 2328 * The last pointer is used to seed the search. It is not required. 2329 */ 2330 static uma_slab_t 2331 zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) 2332 { 2333 uma_klink_t klink; 2334 uma_slab_t slab; 2335 uma_keg_t keg; 2336 int flags; 2337 int empty; 2338 int full; 2339 2340 /* 2341 * Don't wait on the first pass. This will skip limit tests 2342 * as well. We don't want to block if we can find a provider 2343 * without blocking. 2344 */ 2345 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2346 /* 2347 * Use the last slab allocated as a hint for where to start 2348 * the search. 2349 */ 2350 if (last) { 2351 slab = keg_fetch_slab(last, zone, flags); 2352 if (slab) 2353 return (slab); 2354 zone_relock(zone, last); 2355 last = NULL; 2356 } 2357 /* 2358 * Loop until we have a slab incase of transient failures 2359 * while M_WAITOK is specified. I'm not sure this is 100% 2360 * required but we've done it for so long now. 2361 */ 2362 for (;;) { 2363 empty = 0; 2364 full = 0; 2365 /* 2366 * Search the available kegs for slabs. Be careful to hold the 2367 * correct lock while calling into the keg layer. 2368 */ 2369 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2370 keg = klink->kl_keg; 2371 keg_relock(keg, zone); 2372 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2373 slab = keg_fetch_slab(keg, zone, flags); 2374 if (slab) 2375 return (slab); 2376 } 2377 if (keg->uk_flags & UMA_ZFLAG_FULL) 2378 full++; 2379 else 2380 empty++; 2381 zone_relock(zone, keg); 2382 } 2383 if (rflags & (M_NOWAIT | M_NOVM)) 2384 break; 2385 flags = rflags; 2386 /* 2387 * All kegs are full. XXX We can't atomically check all kegs 2388 * and sleep so just sleep for a short period and retry. 2389 */ 2390 if (full && !empty) { 2391 zone->uz_flags |= UMA_ZFLAG_FULL; 2392 zone->uz_sleeps++; 2393 zone_log_warning(zone); 2394 msleep(zone, zone->uz_lock, PVM, "zonelimit", hz/100); 2395 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2396 continue; 2397 } 2398 } 2399 return (NULL); 2400 } 2401 2402 static void * 2403 slab_alloc_item(uma_zone_t zone, uma_slab_t slab) 2404 { 2405 uma_keg_t keg; 2406 uma_slabrefcnt_t slabref; 2407 void *item; 2408 uint8_t freei; 2409 2410 keg = slab->us_keg; 2411 mtx_assert(&keg->uk_lock, MA_OWNED); 2412 2413 freei = slab->us_firstfree; 2414 if (keg->uk_flags & UMA_ZONE_REFCNT) { 2415 slabref = (uma_slabrefcnt_t)slab; 2416 slab->us_firstfree = slabref->us_freelist[freei].us_item; 2417 } else { 2418 slab->us_firstfree = slab->us_freelist[freei].us_item; 2419 } 2420 item = slab->us_data + (keg->uk_rsize * freei); 2421 2422 slab->us_freecount--; 2423 keg->uk_free--; 2424 #ifdef INVARIANTS 2425 uma_dbg_alloc(zone, slab, item); 2426 #endif 2427 /* Move this slab to the full list */ 2428 if (slab->us_freecount == 0) { 2429 LIST_REMOVE(slab, us_link); 2430 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); 2431 } 2432 2433 return (item); 2434 } 2435 2436 static int 2437 zone_alloc_bucket(uma_zone_t zone, int flags) 2438 { 2439 uma_bucket_t bucket; 2440 uma_slab_t slab; 2441 uma_keg_t keg; 2442 int16_t saved; 2443 int max, origflags = flags; 2444 2445 /* 2446 * Try this zone's free list first so we don't allocate extra buckets. 2447 */ 2448 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { 2449 KASSERT(bucket->ub_cnt == 0, 2450 ("zone_alloc_bucket: Bucket on free list is not empty.")); 2451 LIST_REMOVE(bucket, ub_link); 2452 } else { 2453 int bflags; 2454 2455 bflags = (flags & ~M_ZERO); 2456 if (zone->uz_flags & UMA_ZFLAG_CACHEONLY) 2457 bflags |= M_NOVM; 2458 2459 ZONE_UNLOCK(zone); 2460 bucket = bucket_alloc(zone->uz_count, bflags); 2461 ZONE_LOCK(zone); 2462 } 2463 2464 if (bucket == NULL) { 2465 return (0); 2466 } 2467 2468 #ifdef SMP 2469 /* 2470 * This code is here to limit the number of simultaneous bucket fills 2471 * for any given zone to the number of per cpu caches in this zone. This 2472 * is done so that we don't allocate more memory than we really need. 2473 */ 2474 if (zone->uz_fills >= mp_ncpus) 2475 goto done; 2476 2477 #endif 2478 zone->uz_fills++; 2479 2480 max = MIN(bucket->ub_entries, zone->uz_count); 2481 /* Try to keep the buckets totally full */ 2482 saved = bucket->ub_cnt; 2483 slab = NULL; 2484 keg = NULL; 2485 while (bucket->ub_cnt < max && 2486 (slab = zone->uz_slab(zone, keg, flags)) != NULL) { 2487 keg = slab->us_keg; 2488 while (slab->us_freecount && bucket->ub_cnt < max) { 2489 bucket->ub_bucket[bucket->ub_cnt++] = 2490 slab_alloc_item(zone, slab); 2491 } 2492 2493 /* Don't block on the next fill */ 2494 flags |= M_NOWAIT; 2495 } 2496 if (slab) 2497 zone_relock(zone, keg); 2498 2499 /* 2500 * We unlock here because we need to call the zone's init. 2501 * It should be safe to unlock because the slab dealt with 2502 * above is already on the appropriate list within the keg 2503 * and the bucket we filled is not yet on any list, so we 2504 * own it. 2505 */ 2506 if (zone->uz_init != NULL) { 2507 int i; 2508 2509 ZONE_UNLOCK(zone); 2510 for (i = saved; i < bucket->ub_cnt; i++) 2511 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2512 origflags) != 0) 2513 break; 2514 /* 2515 * If we couldn't initialize the whole bucket, put the 2516 * rest back onto the freelist. 2517 */ 2518 if (i != bucket->ub_cnt) { 2519 int j; 2520 2521 for (j = i; j < bucket->ub_cnt; j++) { 2522 zone_free_item(zone, bucket->ub_bucket[j], 2523 NULL, SKIP_FINI, 0); 2524 #ifdef INVARIANTS 2525 bucket->ub_bucket[j] = NULL; 2526 #endif 2527 } 2528 bucket->ub_cnt = i; 2529 } 2530 ZONE_LOCK(zone); 2531 } 2532 2533 zone->uz_fills--; 2534 if (bucket->ub_cnt != 0) { 2535 LIST_INSERT_HEAD(&zone->uz_full_bucket, 2536 bucket, ub_link); 2537 return (1); 2538 } 2539 #ifdef SMP 2540 done: 2541 #endif 2542 bucket_free(bucket); 2543 2544 return (0); 2545 } 2546 /* 2547 * Allocates an item for an internal zone 2548 * 2549 * Arguments 2550 * zone The zone to alloc for. 2551 * udata The data to be passed to the constructor. 2552 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2553 * 2554 * Returns 2555 * NULL if there is no memory and M_NOWAIT is set 2556 * An item if successful 2557 */ 2558 2559 static void * 2560 zone_alloc_item(uma_zone_t zone, void *udata, int flags) 2561 { 2562 uma_slab_t slab; 2563 void *item; 2564 2565 item = NULL; 2566 2567 #ifdef UMA_DEBUG_ALLOC 2568 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); 2569 #endif 2570 ZONE_LOCK(zone); 2571 2572 slab = zone->uz_slab(zone, NULL, flags); 2573 if (slab == NULL) { 2574 zone->uz_fails++; 2575 ZONE_UNLOCK(zone); 2576 return (NULL); 2577 } 2578 2579 item = slab_alloc_item(zone, slab); 2580 2581 zone_relock(zone, slab->us_keg); 2582 zone->uz_allocs++; 2583 ZONE_UNLOCK(zone); 2584 2585 /* 2586 * We have to call both the zone's init (not the keg's init) 2587 * and the zone's ctor. This is because the item is going from 2588 * a keg slab directly to the user, and the user is expecting it 2589 * to be both zone-init'd as well as zone-ctor'd. 2590 */ 2591 if (zone->uz_init != NULL) { 2592 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 2593 zone_free_item(zone, item, udata, SKIP_FINI, 2594 ZFREE_STATFAIL | ZFREE_STATFREE); 2595 return (NULL); 2596 } 2597 } 2598 if (zone->uz_ctor != NULL) { 2599 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2600 zone_free_item(zone, item, udata, SKIP_DTOR, 2601 ZFREE_STATFAIL | ZFREE_STATFREE); 2602 return (NULL); 2603 } 2604 } 2605 if (flags & M_ZERO) 2606 bzero(item, zone->uz_size); 2607 2608 return (item); 2609 } 2610 2611 /* See uma.h */ 2612 void 2613 uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 2614 { 2615 uma_cache_t cache; 2616 uma_bucket_t bucket; 2617 int bflags; 2618 int cpu; 2619 2620 #ifdef UMA_DEBUG_ALLOC_1 2621 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); 2622 #endif 2623 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 2624 zone->uz_name); 2625 2626 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 2627 if (item == NULL) 2628 return; 2629 #ifdef DEBUG_MEMGUARD 2630 if (is_memguard_addr(item)) { 2631 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor) 2632 zone->uz_dtor(item, zone->uz_size, udata); 2633 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini) 2634 zone->uz_fini(item, zone->uz_size); 2635 memguard_free(item); 2636 return; 2637 } 2638 #endif 2639 if (zone->uz_dtor) 2640 zone->uz_dtor(item, zone->uz_size, udata); 2641 2642 #ifdef INVARIANTS 2643 ZONE_LOCK(zone); 2644 if (zone->uz_flags & UMA_ZONE_MALLOC) 2645 uma_dbg_free(zone, udata, item); 2646 else 2647 uma_dbg_free(zone, NULL, item); 2648 ZONE_UNLOCK(zone); 2649 #endif 2650 /* 2651 * The race here is acceptable. If we miss it we'll just have to wait 2652 * a little longer for the limits to be reset. 2653 */ 2654 if (zone->uz_flags & UMA_ZFLAG_FULL) 2655 goto zfree_internal; 2656 2657 /* 2658 * If possible, free to the per-CPU cache. There are two 2659 * requirements for safe access to the per-CPU cache: (1) the thread 2660 * accessing the cache must not be preempted or yield during access, 2661 * and (2) the thread must not migrate CPUs without switching which 2662 * cache it accesses. We rely on a critical section to prevent 2663 * preemption and migration. We release the critical section in 2664 * order to acquire the zone mutex if we are unable to free to the 2665 * current cache; when we re-acquire the critical section, we must 2666 * detect and handle migration if it has occurred. 2667 */ 2668 zfree_restart: 2669 critical_enter(); 2670 cpu = curcpu; 2671 cache = &zone->uz_cpu[cpu]; 2672 2673 zfree_start: 2674 bucket = cache->uc_freebucket; 2675 2676 if (bucket) { 2677 /* 2678 * Do we have room in our bucket? It is OK for this uz count 2679 * check to be slightly out of sync. 2680 */ 2681 2682 if (bucket->ub_cnt < bucket->ub_entries) { 2683 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 2684 ("uma_zfree: Freeing to non free bucket index.")); 2685 bucket->ub_bucket[bucket->ub_cnt] = item; 2686 bucket->ub_cnt++; 2687 cache->uc_frees++; 2688 critical_exit(); 2689 return; 2690 } else if (cache->uc_allocbucket) { 2691 #ifdef UMA_DEBUG_ALLOC 2692 printf("uma_zfree: Swapping buckets.\n"); 2693 #endif 2694 /* 2695 * We have run out of space in our freebucket. 2696 * See if we can switch with our alloc bucket. 2697 */ 2698 if (cache->uc_allocbucket->ub_cnt < 2699 cache->uc_freebucket->ub_cnt) { 2700 bucket = cache->uc_freebucket; 2701 cache->uc_freebucket = cache->uc_allocbucket; 2702 cache->uc_allocbucket = bucket; 2703 goto zfree_start; 2704 } 2705 } 2706 } 2707 /* 2708 * We can get here for two reasons: 2709 * 2710 * 1) The buckets are NULL 2711 * 2) The alloc and free buckets are both somewhat full. 2712 * 2713 * We must go back the zone, which requires acquiring the zone lock, 2714 * which in turn means we must release and re-acquire the critical 2715 * section. Since the critical section is released, we may be 2716 * preempted or migrate. As such, make sure not to maintain any 2717 * thread-local state specific to the cache from prior to releasing 2718 * the critical section. 2719 */ 2720 critical_exit(); 2721 ZONE_LOCK(zone); 2722 critical_enter(); 2723 cpu = curcpu; 2724 cache = &zone->uz_cpu[cpu]; 2725 if (cache->uc_freebucket != NULL) { 2726 if (cache->uc_freebucket->ub_cnt < 2727 cache->uc_freebucket->ub_entries) { 2728 ZONE_UNLOCK(zone); 2729 goto zfree_start; 2730 } 2731 if (cache->uc_allocbucket != NULL && 2732 (cache->uc_allocbucket->ub_cnt < 2733 cache->uc_freebucket->ub_cnt)) { 2734 ZONE_UNLOCK(zone); 2735 goto zfree_start; 2736 } 2737 } 2738 2739 /* Since we have locked the zone we may as well send back our stats */ 2740 zone->uz_allocs += cache->uc_allocs; 2741 cache->uc_allocs = 0; 2742 zone->uz_frees += cache->uc_frees; 2743 cache->uc_frees = 0; 2744 2745 bucket = cache->uc_freebucket; 2746 cache->uc_freebucket = NULL; 2747 2748 /* Can we throw this on the zone full list? */ 2749 if (bucket != NULL) { 2750 #ifdef UMA_DEBUG_ALLOC 2751 printf("uma_zfree: Putting old bucket on the free list.\n"); 2752 #endif 2753 /* ub_cnt is pointing to the last free item */ 2754 KASSERT(bucket->ub_cnt != 0, 2755 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 2756 LIST_INSERT_HEAD(&zone->uz_full_bucket, 2757 bucket, ub_link); 2758 } 2759 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { 2760 LIST_REMOVE(bucket, ub_link); 2761 ZONE_UNLOCK(zone); 2762 cache->uc_freebucket = bucket; 2763 goto zfree_start; 2764 } 2765 /* We are no longer associated with this CPU. */ 2766 critical_exit(); 2767 2768 /* And the zone.. */ 2769 ZONE_UNLOCK(zone); 2770 2771 #ifdef UMA_DEBUG_ALLOC 2772 printf("uma_zfree: Allocating new free bucket.\n"); 2773 #endif 2774 bflags = M_NOWAIT; 2775 2776 if (zone->uz_flags & UMA_ZFLAG_CACHEONLY) 2777 bflags |= M_NOVM; 2778 bucket = bucket_alloc(zone->uz_count, bflags); 2779 if (bucket) { 2780 ZONE_LOCK(zone); 2781 LIST_INSERT_HEAD(&zone->uz_free_bucket, 2782 bucket, ub_link); 2783 ZONE_UNLOCK(zone); 2784 goto zfree_restart; 2785 } 2786 2787 /* 2788 * If nothing else caught this, we'll just do an internal free. 2789 */ 2790 zfree_internal: 2791 zone_free_item(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE); 2792 2793 return; 2794 } 2795 2796 /* 2797 * Frees an item to an INTERNAL zone or allocates a free bucket 2798 * 2799 * Arguments: 2800 * zone The zone to free to 2801 * item The item we're freeing 2802 * udata User supplied data for the dtor 2803 * skip Skip dtors and finis 2804 */ 2805 static void 2806 zone_free_item(uma_zone_t zone, void *item, void *udata, 2807 enum zfreeskip skip, int flags) 2808 { 2809 uma_slab_t slab; 2810 uma_slabrefcnt_t slabref; 2811 uma_keg_t keg; 2812 uint8_t *mem; 2813 uint8_t freei; 2814 int clearfull; 2815 2816 if (skip < SKIP_DTOR && zone->uz_dtor) 2817 zone->uz_dtor(item, zone->uz_size, udata); 2818 2819 if (skip < SKIP_FINI && zone->uz_fini) 2820 zone->uz_fini(item, zone->uz_size); 2821 2822 ZONE_LOCK(zone); 2823 2824 if (flags & ZFREE_STATFAIL) 2825 zone->uz_fails++; 2826 if (flags & ZFREE_STATFREE) 2827 zone->uz_frees++; 2828 2829 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 2830 mem = (uint8_t *)((unsigned long)item & (~UMA_SLAB_MASK)); 2831 keg = zone_first_keg(zone); /* Must only be one. */ 2832 if (zone->uz_flags & UMA_ZONE_HASH) { 2833 slab = hash_sfind(&keg->uk_hash, mem); 2834 } else { 2835 mem += keg->uk_pgoff; 2836 slab = (uma_slab_t)mem; 2837 } 2838 } else { 2839 /* This prevents redundant lookups via free(). */ 2840 if ((zone->uz_flags & UMA_ZONE_MALLOC) && udata != NULL) 2841 slab = (uma_slab_t)udata; 2842 else 2843 slab = vtoslab((vm_offset_t)item); 2844 keg = slab->us_keg; 2845 keg_relock(keg, zone); 2846 } 2847 MPASS(keg == slab->us_keg); 2848 2849 /* Do we need to remove from any lists? */ 2850 if (slab->us_freecount+1 == keg->uk_ipers) { 2851 LIST_REMOVE(slab, us_link); 2852 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 2853 } else if (slab->us_freecount == 0) { 2854 LIST_REMOVE(slab, us_link); 2855 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2856 } 2857 2858 /* Slab management stuff */ 2859 freei = ((unsigned long)item - (unsigned long)slab->us_data) 2860 / keg->uk_rsize; 2861 2862 #ifdef INVARIANTS 2863 if (!skip) 2864 uma_dbg_free(zone, slab, item); 2865 #endif 2866 2867 if (keg->uk_flags & UMA_ZONE_REFCNT) { 2868 slabref = (uma_slabrefcnt_t)slab; 2869 slabref->us_freelist[freei].us_item = slab->us_firstfree; 2870 } else { 2871 slab->us_freelist[freei].us_item = slab->us_firstfree; 2872 } 2873 slab->us_firstfree = freei; 2874 slab->us_freecount++; 2875 2876 /* Zone statistics */ 2877 keg->uk_free++; 2878 2879 clearfull = 0; 2880 if (keg->uk_flags & UMA_ZFLAG_FULL) { 2881 if (keg->uk_pages < keg->uk_maxpages) { 2882 keg->uk_flags &= ~UMA_ZFLAG_FULL; 2883 clearfull = 1; 2884 } 2885 2886 /* 2887 * We can handle one more allocation. Since we're clearing ZFLAG_FULL, 2888 * wake up all procs blocked on pages. This should be uncommon, so 2889 * keeping this simple for now (rather than adding count of blocked 2890 * threads etc). 2891 */ 2892 wakeup(keg); 2893 } 2894 if (clearfull) { 2895 zone_relock(zone, keg); 2896 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2897 wakeup(zone); 2898 ZONE_UNLOCK(zone); 2899 } else 2900 KEG_UNLOCK(keg); 2901 } 2902 2903 /* See uma.h */ 2904 int 2905 uma_zone_set_max(uma_zone_t zone, int nitems) 2906 { 2907 uma_keg_t keg; 2908 2909 ZONE_LOCK(zone); 2910 keg = zone_first_keg(zone); 2911 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 2912 if (keg->uk_maxpages * keg->uk_ipers < nitems) 2913 keg->uk_maxpages += keg->uk_ppera; 2914 nitems = keg->uk_maxpages * keg->uk_ipers; 2915 ZONE_UNLOCK(zone); 2916 2917 return (nitems); 2918 } 2919 2920 /* See uma.h */ 2921 int 2922 uma_zone_get_max(uma_zone_t zone) 2923 { 2924 int nitems; 2925 uma_keg_t keg; 2926 2927 ZONE_LOCK(zone); 2928 keg = zone_first_keg(zone); 2929 nitems = keg->uk_maxpages * keg->uk_ipers; 2930 ZONE_UNLOCK(zone); 2931 2932 return (nitems); 2933 } 2934 2935 /* See uma.h */ 2936 void 2937 uma_zone_set_warning(uma_zone_t zone, const char *warning) 2938 { 2939 2940 ZONE_LOCK(zone); 2941 zone->uz_warning = warning; 2942 ZONE_UNLOCK(zone); 2943 } 2944 2945 /* See uma.h */ 2946 int 2947 uma_zone_get_cur(uma_zone_t zone) 2948 { 2949 int64_t nitems; 2950 u_int i; 2951 2952 ZONE_LOCK(zone); 2953 nitems = zone->uz_allocs - zone->uz_frees; 2954 CPU_FOREACH(i) { 2955 /* 2956 * See the comment in sysctl_vm_zone_stats() regarding the 2957 * safety of accessing the per-cpu caches. With the zone lock 2958 * held, it is safe, but can potentially result in stale data. 2959 */ 2960 nitems += zone->uz_cpu[i].uc_allocs - 2961 zone->uz_cpu[i].uc_frees; 2962 } 2963 ZONE_UNLOCK(zone); 2964 2965 return (nitems < 0 ? 0 : nitems); 2966 } 2967 2968 /* See uma.h */ 2969 void 2970 uma_zone_set_init(uma_zone_t zone, uma_init uminit) 2971 { 2972 uma_keg_t keg; 2973 2974 ZONE_LOCK(zone); 2975 keg = zone_first_keg(zone); 2976 KASSERT(keg->uk_pages == 0, 2977 ("uma_zone_set_init on non-empty keg")); 2978 keg->uk_init = uminit; 2979 ZONE_UNLOCK(zone); 2980 } 2981 2982 /* See uma.h */ 2983 void 2984 uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 2985 { 2986 uma_keg_t keg; 2987 2988 ZONE_LOCK(zone); 2989 keg = zone_first_keg(zone); 2990 KASSERT(keg->uk_pages == 0, 2991 ("uma_zone_set_fini on non-empty keg")); 2992 keg->uk_fini = fini; 2993 ZONE_UNLOCK(zone); 2994 } 2995 2996 /* See uma.h */ 2997 void 2998 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 2999 { 3000 ZONE_LOCK(zone); 3001 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3002 ("uma_zone_set_zinit on non-empty keg")); 3003 zone->uz_init = zinit; 3004 ZONE_UNLOCK(zone); 3005 } 3006 3007 /* See uma.h */ 3008 void 3009 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 3010 { 3011 ZONE_LOCK(zone); 3012 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3013 ("uma_zone_set_zfini on non-empty keg")); 3014 zone->uz_fini = zfini; 3015 ZONE_UNLOCK(zone); 3016 } 3017 3018 /* See uma.h */ 3019 /* XXX uk_freef is not actually used with the zone locked */ 3020 void 3021 uma_zone_set_freef(uma_zone_t zone, uma_free freef) 3022 { 3023 3024 ZONE_LOCK(zone); 3025 zone_first_keg(zone)->uk_freef = freef; 3026 ZONE_UNLOCK(zone); 3027 } 3028 3029 /* See uma.h */ 3030 /* XXX uk_allocf is not actually used with the zone locked */ 3031 void 3032 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3033 { 3034 uma_keg_t keg; 3035 3036 ZONE_LOCK(zone); 3037 keg = zone_first_keg(zone); 3038 keg->uk_flags |= UMA_ZFLAG_PRIVALLOC; 3039 keg->uk_allocf = allocf; 3040 ZONE_UNLOCK(zone); 3041 } 3042 3043 /* See uma.h */ 3044 int 3045 uma_zone_reserve_kva(uma_zone_t zone, int count) 3046 { 3047 uma_keg_t keg; 3048 vm_offset_t kva; 3049 int pages; 3050 3051 keg = zone_first_keg(zone); 3052 pages = count / keg->uk_ipers; 3053 3054 if (pages * keg->uk_ipers < count) 3055 pages++; 3056 3057 #ifdef UMA_MD_SMALL_ALLOC 3058 if (keg->uk_ppera > 1) { 3059 #else 3060 if (1) { 3061 #endif 3062 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE); 3063 if (kva == 0) 3064 return (0); 3065 } else 3066 kva = 0; 3067 ZONE_LOCK(zone); 3068 keg->uk_kva = kva; 3069 keg->uk_offset = 0; 3070 keg->uk_maxpages = pages; 3071 #ifdef UMA_MD_SMALL_ALLOC 3072 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3073 #else 3074 keg->uk_allocf = noobj_alloc; 3075 #endif 3076 keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC; 3077 ZONE_UNLOCK(zone); 3078 return (1); 3079 } 3080 3081 /* See uma.h */ 3082 void 3083 uma_prealloc(uma_zone_t zone, int items) 3084 { 3085 int slabs; 3086 uma_slab_t slab; 3087 uma_keg_t keg; 3088 3089 keg = zone_first_keg(zone); 3090 ZONE_LOCK(zone); 3091 slabs = items / keg->uk_ipers; 3092 if (slabs * keg->uk_ipers < items) 3093 slabs++; 3094 while (slabs > 0) { 3095 slab = keg_alloc_slab(keg, zone, M_WAITOK); 3096 if (slab == NULL) 3097 break; 3098 MPASS(slab->us_keg == keg); 3099 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 3100 slabs--; 3101 } 3102 ZONE_UNLOCK(zone); 3103 } 3104 3105 /* See uma.h */ 3106 uint32_t * 3107 uma_find_refcnt(uma_zone_t zone, void *item) 3108 { 3109 uma_slabrefcnt_t slabref; 3110 uma_keg_t keg; 3111 uint32_t *refcnt; 3112 int idx; 3113 3114 slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item & 3115 (~UMA_SLAB_MASK)); 3116 keg = slabref->us_keg; 3117 KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT, 3118 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT")); 3119 idx = ((unsigned long)item - (unsigned long)slabref->us_data) 3120 / keg->uk_rsize; 3121 refcnt = &slabref->us_freelist[idx].us_refcnt; 3122 return refcnt; 3123 } 3124 3125 /* See uma.h */ 3126 void 3127 uma_reclaim(void) 3128 { 3129 #ifdef UMA_DEBUG 3130 printf("UMA: vm asked us to release pages!\n"); 3131 #endif 3132 bucket_enable(); 3133 zone_foreach(zone_drain); 3134 /* 3135 * Some slabs may have been freed but this zone will be visited early 3136 * we visit again so that we can free pages that are empty once other 3137 * zones are drained. We have to do the same for buckets. 3138 */ 3139 zone_drain(slabzone); 3140 zone_drain(slabrefzone); 3141 bucket_zone_drain(); 3142 } 3143 3144 /* See uma.h */ 3145 int 3146 uma_zone_exhausted(uma_zone_t zone) 3147 { 3148 int full; 3149 3150 ZONE_LOCK(zone); 3151 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3152 ZONE_UNLOCK(zone); 3153 return (full); 3154 } 3155 3156 int 3157 uma_zone_exhausted_nolock(uma_zone_t zone) 3158 { 3159 return (zone->uz_flags & UMA_ZFLAG_FULL); 3160 } 3161 3162 void * 3163 uma_large_malloc(int size, int wait) 3164 { 3165 void *mem; 3166 uma_slab_t slab; 3167 uint8_t flags; 3168 3169 slab = zone_alloc_item(slabzone, NULL, wait); 3170 if (slab == NULL) 3171 return (NULL); 3172 mem = page_alloc(NULL, size, &flags, wait); 3173 if (mem) { 3174 vsetslab((vm_offset_t)mem, slab); 3175 slab->us_data = mem; 3176 slab->us_flags = flags | UMA_SLAB_MALLOC; 3177 slab->us_size = size; 3178 } else { 3179 zone_free_item(slabzone, slab, NULL, SKIP_NONE, 3180 ZFREE_STATFAIL | ZFREE_STATFREE); 3181 } 3182 3183 return (mem); 3184 } 3185 3186 void 3187 uma_large_free(uma_slab_t slab) 3188 { 3189 vsetobj((vm_offset_t)slab->us_data, kmem_object); 3190 page_free(slab->us_data, slab->us_size, slab->us_flags); 3191 zone_free_item(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE); 3192 } 3193 3194 void 3195 uma_print_stats(void) 3196 { 3197 zone_foreach(uma_print_zone); 3198 } 3199 3200 static void 3201 slab_print(uma_slab_t slab) 3202 { 3203 printf("slab: keg %p, data %p, freecount %d, firstfree %d\n", 3204 slab->us_keg, slab->us_data, slab->us_freecount, 3205 slab->us_firstfree); 3206 } 3207 3208 static void 3209 cache_print(uma_cache_t cache) 3210 { 3211 printf("alloc: %p(%d), free: %p(%d)\n", 3212 cache->uc_allocbucket, 3213 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3214 cache->uc_freebucket, 3215 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3216 } 3217 3218 static void 3219 uma_print_keg(uma_keg_t keg) 3220 { 3221 uma_slab_t slab; 3222 3223 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3224 "out %d free %d limit %d\n", 3225 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3226 keg->uk_ipers, keg->uk_ppera, 3227 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free, 3228 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3229 printf("Part slabs:\n"); 3230 LIST_FOREACH(slab, &keg->uk_part_slab, us_link) 3231 slab_print(slab); 3232 printf("Free slabs:\n"); 3233 LIST_FOREACH(slab, &keg->uk_free_slab, us_link) 3234 slab_print(slab); 3235 printf("Full slabs:\n"); 3236 LIST_FOREACH(slab, &keg->uk_full_slab, us_link) 3237 slab_print(slab); 3238 } 3239 3240 void 3241 uma_print_zone(uma_zone_t zone) 3242 { 3243 uma_cache_t cache; 3244 uma_klink_t kl; 3245 int i; 3246 3247 printf("zone: %s(%p) size %d flags %#x\n", 3248 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3249 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3250 uma_print_keg(kl->kl_keg); 3251 CPU_FOREACH(i) { 3252 cache = &zone->uz_cpu[i]; 3253 printf("CPU %d Cache:\n", i); 3254 cache_print(cache); 3255 } 3256 } 3257 3258 #ifdef DDB 3259 /* 3260 * Generate statistics across both the zone and its per-cpu cache's. Return 3261 * desired statistics if the pointer is non-NULL for that statistic. 3262 * 3263 * Note: does not update the zone statistics, as it can't safely clear the 3264 * per-CPU cache statistic. 3265 * 3266 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3267 * safe from off-CPU; we should modify the caches to track this information 3268 * directly so that we don't have to. 3269 */ 3270 static void 3271 uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3272 uint64_t *freesp, uint64_t *sleepsp) 3273 { 3274 uma_cache_t cache; 3275 uint64_t allocs, frees, sleeps; 3276 int cachefree, cpu; 3277 3278 allocs = frees = sleeps = 0; 3279 cachefree = 0; 3280 CPU_FOREACH(cpu) { 3281 cache = &z->uz_cpu[cpu]; 3282 if (cache->uc_allocbucket != NULL) 3283 cachefree += cache->uc_allocbucket->ub_cnt; 3284 if (cache->uc_freebucket != NULL) 3285 cachefree += cache->uc_freebucket->ub_cnt; 3286 allocs += cache->uc_allocs; 3287 frees += cache->uc_frees; 3288 } 3289 allocs += z->uz_allocs; 3290 frees += z->uz_frees; 3291 sleeps += z->uz_sleeps; 3292 if (cachefreep != NULL) 3293 *cachefreep = cachefree; 3294 if (allocsp != NULL) 3295 *allocsp = allocs; 3296 if (freesp != NULL) 3297 *freesp = frees; 3298 if (sleepsp != NULL) 3299 *sleepsp = sleeps; 3300 } 3301 #endif /* DDB */ 3302 3303 static int 3304 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3305 { 3306 uma_keg_t kz; 3307 uma_zone_t z; 3308 int count; 3309 3310 count = 0; 3311 mtx_lock(&uma_mtx); 3312 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3313 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3314 count++; 3315 } 3316 mtx_unlock(&uma_mtx); 3317 return (sysctl_handle_int(oidp, &count, 0, req)); 3318 } 3319 3320 static int 3321 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3322 { 3323 struct uma_stream_header ush; 3324 struct uma_type_header uth; 3325 struct uma_percpu_stat ups; 3326 uma_bucket_t bucket; 3327 struct sbuf sbuf; 3328 uma_cache_t cache; 3329 uma_klink_t kl; 3330 uma_keg_t kz; 3331 uma_zone_t z; 3332 uma_keg_t k; 3333 int count, error, i; 3334 3335 error = sysctl_wire_old_buffer(req, 0); 3336 if (error != 0) 3337 return (error); 3338 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3339 3340 count = 0; 3341 mtx_lock(&uma_mtx); 3342 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3343 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3344 count++; 3345 } 3346 3347 /* 3348 * Insert stream header. 3349 */ 3350 bzero(&ush, sizeof(ush)); 3351 ush.ush_version = UMA_STREAM_VERSION; 3352 ush.ush_maxcpus = (mp_maxid + 1); 3353 ush.ush_count = count; 3354 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3355 3356 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3357 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3358 bzero(&uth, sizeof(uth)); 3359 ZONE_LOCK(z); 3360 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3361 uth.uth_align = kz->uk_align; 3362 uth.uth_size = kz->uk_size; 3363 uth.uth_rsize = kz->uk_rsize; 3364 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3365 k = kl->kl_keg; 3366 uth.uth_maxpages += k->uk_maxpages; 3367 uth.uth_pages += k->uk_pages; 3368 uth.uth_keg_free += k->uk_free; 3369 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3370 * k->uk_ipers; 3371 } 3372 3373 /* 3374 * A zone is secondary is it is not the first entry 3375 * on the keg's zone list. 3376 */ 3377 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3378 (LIST_FIRST(&kz->uk_zones) != z)) 3379 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3380 3381 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link) 3382 uth.uth_zone_free += bucket->ub_cnt; 3383 uth.uth_allocs = z->uz_allocs; 3384 uth.uth_frees = z->uz_frees; 3385 uth.uth_fails = z->uz_fails; 3386 uth.uth_sleeps = z->uz_sleeps; 3387 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 3388 /* 3389 * While it is not normally safe to access the cache 3390 * bucket pointers while not on the CPU that owns the 3391 * cache, we only allow the pointers to be exchanged 3392 * without the zone lock held, not invalidated, so 3393 * accept the possible race associated with bucket 3394 * exchange during monitoring. 3395 */ 3396 for (i = 0; i < (mp_maxid + 1); i++) { 3397 bzero(&ups, sizeof(ups)); 3398 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) 3399 goto skip; 3400 if (CPU_ABSENT(i)) 3401 goto skip; 3402 cache = &z->uz_cpu[i]; 3403 if (cache->uc_allocbucket != NULL) 3404 ups.ups_cache_free += 3405 cache->uc_allocbucket->ub_cnt; 3406 if (cache->uc_freebucket != NULL) 3407 ups.ups_cache_free += 3408 cache->uc_freebucket->ub_cnt; 3409 ups.ups_allocs = cache->uc_allocs; 3410 ups.ups_frees = cache->uc_frees; 3411 skip: 3412 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups)); 3413 } 3414 ZONE_UNLOCK(z); 3415 } 3416 } 3417 mtx_unlock(&uma_mtx); 3418 error = sbuf_finish(&sbuf); 3419 sbuf_delete(&sbuf); 3420 return (error); 3421 } 3422 3423 #ifdef DDB 3424 DB_SHOW_COMMAND(uma, db_show_uma) 3425 { 3426 uint64_t allocs, frees, sleeps; 3427 uma_bucket_t bucket; 3428 uma_keg_t kz; 3429 uma_zone_t z; 3430 int cachefree; 3431 3432 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 3433 "Requests", "Sleeps"); 3434 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3435 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3436 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 3437 allocs = z->uz_allocs; 3438 frees = z->uz_frees; 3439 sleeps = z->uz_sleeps; 3440 cachefree = 0; 3441 } else 3442 uma_zone_sumstat(z, &cachefree, &allocs, 3443 &frees, &sleeps); 3444 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 3445 (LIST_FIRST(&kz->uk_zones) != z))) 3446 cachefree += kz->uk_free; 3447 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link) 3448 cachefree += bucket->ub_cnt; 3449 db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name, 3450 (uintmax_t)kz->uk_size, 3451 (intmax_t)(allocs - frees), cachefree, 3452 (uintmax_t)allocs, sleeps); 3453 if (db_pager_quit) 3454 return; 3455 } 3456 } 3457 } 3458 #endif 3459