1 /* 2 * Copyright (c) 2005 Jeffrey M. Hsu. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Jeffrey M. Hsu. 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, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. Neither the name of The DragonFly Project nor the names of its 16 * contributors may be used to endorse or promote products derived 17 * from this software without specific, prior written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 21 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 22 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 23 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 24 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 25 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 26 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 27 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 28 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 29 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * $DragonFly: src/sys/kern/kern_objcache.c,v 1.23 2008/10/26 04:29:19 sephe Exp $ 33 */ 34 35 #include <sys/param.h> 36 #include <sys/kernel.h> 37 #include <sys/systm.h> 38 #include <sys/callout.h> 39 #include <sys/globaldata.h> 40 #include <sys/malloc.h> 41 #include <sys/queue.h> 42 #include <sys/objcache.h> 43 #include <sys/spinlock.h> 44 #include <sys/thread.h> 45 #include <sys/thread2.h> 46 #include <sys/spinlock2.h> 47 48 static MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache"); 49 static MALLOC_DEFINE(M_OBJMAG, "objcache magazine", "Object Cache Magazine"); 50 51 #define INITIAL_MAG_CAPACITY 64 52 53 struct magazine { 54 int rounds; 55 int capacity; 56 int cleaning; 57 SLIST_ENTRY(magazine) nextmagazine; 58 void *objects[]; 59 }; 60 61 SLIST_HEAD(magazinelist, magazine); 62 63 /* 64 * per-cluster cache of magazines 65 * 66 * All fields in this structure are protected by the spinlock. 67 */ 68 struct magazinedepot { 69 /* 70 * The per-cpu object caches only exchanges completely full or 71 * completely empty magazines with the depot layer, so only have 72 * to cache these two types of magazines. 73 */ 74 struct magazinelist fullmagazines; 75 struct magazinelist emptymagazines; 76 int magcapacity; 77 78 /* protect this structure */ 79 struct spinlock spin; 80 81 /* magazines not yet allocated towards limit */ 82 int unallocated_objects; 83 84 /* infrequently used fields */ 85 int waiting; /* waiting for another cpu to 86 * return a full magazine to 87 * the depot */ 88 int contested; /* depot contention count */ 89 }; 90 91 /* 92 * per-cpu object cache 93 * All fields in this structure are protected by crit_enter(). 94 */ 95 struct percpu_objcache { 96 struct magazine *loaded_magazine; /* active magazine */ 97 struct magazine *previous_magazine; /* backup magazine */ 98 99 /* statistics */ 100 int gets_cumulative; /* total calls to get */ 101 int gets_null; /* objcache_get returned NULL */ 102 int puts_cumulative; /* total calls to put */ 103 int puts_othercluster; /* returned to other cluster */ 104 105 /* infrequently used fields */ 106 int waiting; /* waiting for a thread on this cpu to 107 * return an obj to the per-cpu cache */ 108 }; 109 110 /* only until we have NUMA cluster topology information XXX */ 111 #define MAXCLUSTERS 1 112 #define myclusterid 0 113 #define CLUSTER_OF(obj) 0 114 115 /* 116 * Two-level object cache consisting of NUMA cluster-level depots of 117 * fully loaded or completely empty magazines and cpu-level caches of 118 * individual objects. 119 */ 120 struct objcache { 121 char *name; 122 123 /* object constructor and destructor from blank storage */ 124 objcache_ctor_fn *ctor; 125 objcache_dtor_fn *dtor; 126 void *privdata; 127 128 /* interface to underlying allocator */ 129 objcache_alloc_fn *alloc; 130 objcache_free_fn *free; 131 void *allocator_args; 132 133 LIST_ENTRY(objcache) oc_next; 134 int exhausted; /* oops */ 135 136 /* NUMA-cluster level caches */ 137 struct magazinedepot depot[MAXCLUSTERS]; 138 139 struct percpu_objcache cache_percpu[]; /* per-cpu caches */ 140 }; 141 142 static struct spinlock objcachelist_spin; 143 static LIST_HEAD(objcachelist, objcache) allobjcaches; 144 145 static struct magazine * 146 mag_alloc(int capacity) 147 { 148 struct magazine *mag; 149 150 mag = kmalloc(__offsetof(struct magazine, objects[capacity]), 151 M_OBJMAG, M_INTWAIT | M_ZERO); 152 mag->capacity = capacity; 153 mag->rounds = 0; 154 mag->cleaning = 0; 155 return (mag); 156 } 157 158 /* 159 * Utility routine for objects that don't require any de-construction. 160 */ 161 162 static void 163 null_dtor(void *obj, void *privdata) 164 { 165 /* do nothing */ 166 } 167 168 static boolean_t 169 null_ctor(void *obj, void *privdata, int ocflags) 170 { 171 return TRUE; 172 } 173 174 /* 175 * Create an object cache. 176 */ 177 struct objcache * 178 objcache_create(const char *name, int *cluster_limit0, int mag_capacity, 179 objcache_ctor_fn *ctor, objcache_dtor_fn *dtor, void *privdata, 180 objcache_alloc_fn *alloc, objcache_free_fn *free, 181 void *allocator_args) 182 { 183 struct objcache *oc; 184 struct magazinedepot *depot; 185 int cpuid; 186 int need; 187 int factor; 188 int nmagdepot; 189 int i; 190 int cluster_limit; 191 192 if (cluster_limit0 == NULL) 193 cluster_limit = 0; 194 else 195 cluster_limit = *cluster_limit0; 196 197 /* allocate object cache structure */ 198 oc = kmalloc(__offsetof(struct objcache, cache_percpu[ncpus]), 199 M_OBJCACHE, M_WAITOK | M_ZERO); 200 oc->name = kstrdup(name, M_TEMP); 201 oc->ctor = ctor ? ctor : null_ctor; 202 oc->dtor = dtor ? dtor : null_dtor; 203 oc->privdata = privdata; 204 oc->free = free; 205 oc->allocator_args = allocator_args; 206 207 /* initialize depots */ 208 depot = &oc->depot[0]; 209 210 spin_init(&depot->spin); 211 SLIST_INIT(&depot->fullmagazines); 212 SLIST_INIT(&depot->emptymagazines); 213 214 if (mag_capacity == 0) 215 mag_capacity = INITIAL_MAG_CAPACITY; 216 217 /* 218 * The cluster_limit must be sufficient to have three magazines per 219 * cpu. If we have a lot of cpus the mag_capacity might just be 220 * too big, reduce it if necessary. 221 * 222 * Each cpu can hold up to two magazines, with the remainder in the 223 * depot. If many objects are allocated fewer magazines are 224 * available. We have to make sure that each cpu has access to 225 * free objects until the object cache hits 75% of its limit. 226 */ 227 if (cluster_limit == 0) { 228 depot->unallocated_objects = -1; 229 } else { 230 factor = 8; 231 need = mag_capacity * ncpus * factor; 232 if (cluster_limit < need && mag_capacity > 16) { 233 kprintf("objcache(%s): too small for ncpus" 234 ", adjusting mag_capacity %d->", 235 name, mag_capacity); 236 while (need > cluster_limit && mag_capacity > 16) { 237 mag_capacity >>= 1; 238 need = mag_capacity * ncpus * factor; 239 } 240 kprintf("%d\n", mag_capacity); 241 } 242 if (cluster_limit < need) { 243 kprintf("objcache(%s): too small for ncpus" 244 ", adjusting cluster_limit %d->%d\n", 245 name, cluster_limit, need); 246 cluster_limit = need; 247 } 248 depot->unallocated_objects = cluster_limit; 249 } 250 depot->magcapacity = mag_capacity; 251 oc->alloc = alloc; 252 253 /* initialize per-cpu caches */ 254 for (cpuid = 0; cpuid < ncpus; cpuid++) { 255 struct percpu_objcache *cache_percpu = &oc->cache_percpu[cpuid]; 256 257 cache_percpu->loaded_magazine = mag_alloc(mag_capacity); 258 cache_percpu->previous_magazine = mag_alloc(mag_capacity); 259 } 260 261 /* compute initial number of empty magazines in depot */ 262 nmagdepot = 0; 263 if (cluster_limit > 0) { 264 /* max number of magazines in depot */ 265 nmagdepot = (cluster_limit - ncpus * 2 * mag_capacity) / 266 mag_capacity; 267 268 /* retain at most 50% of the limit */ 269 nmagdepot /= 2; 270 } 271 /* bound result to acceptable range */ 272 if (nmagdepot < 2) 273 nmagdepot = 2; 274 if (nmagdepot > 1000) 275 nmagdepot = 1000; 276 277 /* put empty magazines in depot */ 278 for (i = 0; i < nmagdepot; i++) { 279 struct magazine *mag = mag_alloc(mag_capacity); 280 SLIST_INSERT_HEAD(&depot->emptymagazines, mag, nextmagazine); 281 } 282 283 spin_lock(&objcachelist_spin); 284 LIST_INSERT_HEAD(&allobjcaches, oc, oc_next); 285 spin_unlock(&objcachelist_spin); 286 287 if (cluster_limit0 != NULL) 288 *cluster_limit0 = cluster_limit; 289 return (oc); 290 } 291 292 struct objcache * 293 objcache_create_simple(malloc_type_t mtype, size_t objsize) 294 { 295 struct objcache_malloc_args *margs; 296 struct objcache *oc; 297 298 margs = kmalloc(sizeof(*margs), M_OBJCACHE, M_WAITOK|M_ZERO); 299 margs->objsize = objsize; 300 margs->mtype = mtype; 301 oc = objcache_create(mtype->ks_shortdesc, NULL, 0, 302 NULL, NULL, NULL, 303 objcache_malloc_alloc, objcache_malloc_free, 304 margs); 305 return (oc); 306 } 307 308 struct objcache * 309 objcache_create_mbacked(malloc_type_t mtype, size_t objsize, 310 int *cluster_limit, int mag_capacity, 311 objcache_ctor_fn *ctor, objcache_dtor_fn *dtor, 312 void *privdata) 313 { 314 struct objcache_malloc_args *margs; 315 struct objcache *oc; 316 317 margs = kmalloc(sizeof(*margs), M_OBJCACHE, M_WAITOK|M_ZERO); 318 margs->objsize = objsize; 319 margs->mtype = mtype; 320 oc = objcache_create(mtype->ks_shortdesc, 321 cluster_limit, mag_capacity, 322 ctor, dtor, privdata, 323 objcache_malloc_alloc, objcache_malloc_free, 324 margs); 325 return(oc); 326 } 327 328 329 #define MAGAZINE_EMPTY(mag) (mag->rounds == 0) 330 #define MAGAZINE_NOTEMPTY(mag) (mag->rounds != 0) 331 #define MAGAZINE_FULL(mag) (mag->rounds == mag->capacity) 332 333 #define swap(x, y) ({ struct magazine *t = x; x = y; y = t; }) 334 335 /* 336 * Get an object from the object cache. 337 * 338 * WARNING! ocflags are only used when we have to go to the underlying 339 * allocator, so we cannot depend on flags such as M_ZERO. 340 */ 341 void * 342 objcache_get(struct objcache *oc, int ocflags) 343 { 344 struct percpu_objcache *cpucache = &oc->cache_percpu[mycpuid]; 345 struct magazine *loadedmag; 346 struct magazine *emptymag; 347 void *obj; 348 struct magazinedepot *depot; 349 350 KKASSERT((ocflags & M_ZERO) == 0); 351 crit_enter(); 352 ++cpucache->gets_cumulative; 353 354 retry: 355 /* 356 * Loaded magazine has an object. This is the hot path. 357 * It is lock-free and uses a critical section to block 358 * out interrupt handlers on the same processor. 359 */ 360 loadedmag = cpucache->loaded_magazine; 361 if (MAGAZINE_NOTEMPTY(loadedmag)) { 362 obj = loadedmag->objects[--loadedmag->rounds]; 363 crit_exit(); 364 return (obj); 365 } 366 367 /* Previous magazine has an object. */ 368 if (MAGAZINE_NOTEMPTY(cpucache->previous_magazine)) { 369 KKASSERT(cpucache->previous_magazine->cleaning + 370 cpucache->loaded_magazine->cleaning == 0); 371 swap(cpucache->loaded_magazine, cpucache->previous_magazine); 372 loadedmag = cpucache->loaded_magazine; 373 obj = loadedmag->objects[--loadedmag->rounds]; 374 crit_exit(); 375 return (obj); 376 } 377 378 /* 379 * Both magazines empty. Get a full magazine from the depot and 380 * move one of the empty ones to the depot. 381 * 382 * Obtain the depot spinlock. 383 * 384 * NOTE: Beyond this point, M_* flags are handled via oc->alloc() 385 */ 386 depot = &oc->depot[myclusterid]; 387 spin_lock(&depot->spin); 388 389 /* 390 * Recheck the cpucache after obtaining the depot spinlock. This 391 * shouldn't be necessary now but don't take any chances. 392 */ 393 if (MAGAZINE_NOTEMPTY(cpucache->loaded_magazine) || 394 MAGAZINE_NOTEMPTY(cpucache->previous_magazine) 395 ) { 396 spin_unlock(&depot->spin); 397 goto retry; 398 } 399 400 /* Check if depot has a full magazine. */ 401 if (!SLIST_EMPTY(&depot->fullmagazines)) { 402 emptymag = cpucache->previous_magazine; 403 cpucache->previous_magazine = cpucache->loaded_magazine; 404 cpucache->loaded_magazine = SLIST_FIRST(&depot->fullmagazines); 405 SLIST_REMOVE_HEAD(&depot->fullmagazines, nextmagazine); 406 407 /* 408 * Return emptymag to the depot. 409 */ 410 KKASSERT(MAGAZINE_EMPTY(emptymag)); 411 SLIST_INSERT_HEAD(&depot->emptymagazines, 412 emptymag, nextmagazine); 413 spin_unlock(&depot->spin); 414 goto retry; 415 } 416 417 /* 418 * The depot does not have any non-empty magazines. If we have 419 * not hit our object limit we can allocate a new object using 420 * the back-end allocator. 421 * 422 * note: unallocated_objects can be initialized to -1, which has 423 * the effect of removing any allocation limits. 424 */ 425 if (depot->unallocated_objects) { 426 --depot->unallocated_objects; 427 spin_unlock(&depot->spin); 428 crit_exit(); 429 430 obj = oc->alloc(oc->allocator_args, ocflags); 431 if (obj) { 432 if (oc->ctor(obj, oc->privdata, ocflags)) 433 return (obj); 434 oc->free(obj, oc->allocator_args); 435 obj = NULL; 436 } 437 if (obj == NULL) { 438 spin_lock(&depot->spin); 439 ++depot->unallocated_objects; 440 spin_unlock(&depot->spin); 441 if (depot->waiting) 442 wakeup(depot); 443 444 crit_enter(); 445 /* 446 * makes debugging easier when gets_cumulative does 447 * not include gets_null. 448 */ 449 ++cpucache->gets_null; 450 --cpucache->gets_cumulative; 451 crit_exit(); 452 } 453 return(obj); 454 } 455 if (oc->exhausted == 0) { 456 kprintf("Warning, objcache(%s): Exhausted!\n", oc->name); 457 oc->exhausted = 1; 458 } 459 460 /* 461 * Otherwise block if allowed to. 462 */ 463 if ((ocflags & (M_WAITOK|M_NULLOK)) == M_WAITOK) { 464 ++cpucache->waiting; 465 ++depot->waiting; 466 ssleep(depot, &depot->spin, 0, "objcache_get", 0); 467 --cpucache->waiting; 468 --depot->waiting; 469 spin_unlock(&depot->spin); 470 goto retry; 471 } 472 473 /* 474 * Otherwise fail 475 */ 476 ++cpucache->gets_null; 477 --cpucache->gets_cumulative; 478 crit_exit(); 479 spin_unlock(&depot->spin); 480 return (NULL); 481 } 482 483 /* 484 * Wrapper for malloc allocation routines. 485 */ 486 void * 487 objcache_malloc_alloc(void *allocator_args, int ocflags) 488 { 489 struct objcache_malloc_args *alloc_args = allocator_args; 490 491 return (kmalloc(alloc_args->objsize, alloc_args->mtype, 492 ocflags & OC_MFLAGS)); 493 } 494 495 void 496 objcache_malloc_free(void *obj, void *allocator_args) 497 { 498 struct objcache_malloc_args *alloc_args = allocator_args; 499 500 kfree(obj, alloc_args->mtype); 501 } 502 503 /* 504 * Wrapper for allocation policies that pre-allocate at initialization time 505 * and don't do run-time allocation. 506 */ 507 void * 508 objcache_nop_alloc(void *allocator_args, int ocflags) 509 { 510 return (NULL); 511 } 512 513 void 514 objcache_nop_free(void *obj, void *allocator_args) 515 { 516 } 517 518 /* 519 * Return an object to the object cache. 520 */ 521 void 522 objcache_put(struct objcache *oc, void *obj) 523 { 524 struct percpu_objcache *cpucache = &oc->cache_percpu[mycpuid]; 525 struct magazine *loadedmag; 526 struct magazinedepot *depot; 527 528 crit_enter(); 529 ++cpucache->puts_cumulative; 530 531 if (CLUSTER_OF(obj) != myclusterid) { 532 #ifdef notyet 533 /* use lazy IPI to send object to owning cluster XXX todo */ 534 ++cpucache->puts_othercluster; 535 crit_exit(); 536 return; 537 #endif 538 } 539 540 retry: 541 /* 542 * Free slot available in loaded magazine. This is the hot path. 543 * It is lock-free and uses a critical section to block out interrupt 544 * handlers on the same processor. 545 */ 546 loadedmag = cpucache->loaded_magazine; 547 if (!MAGAZINE_FULL(loadedmag)) { 548 loadedmag->objects[loadedmag->rounds++] = obj; 549 if (cpucache->waiting) 550 wakeup_mycpu(&oc->depot[myclusterid]); 551 crit_exit(); 552 return; 553 } 554 555 /* 556 * Current magazine full, but previous magazine has room. XXX 557 */ 558 if (!MAGAZINE_FULL(cpucache->previous_magazine)) { 559 KKASSERT(cpucache->previous_magazine->cleaning + 560 cpucache->loaded_magazine->cleaning == 0); 561 swap(cpucache->loaded_magazine, cpucache->previous_magazine); 562 loadedmag = cpucache->loaded_magazine; 563 loadedmag->objects[loadedmag->rounds++] = obj; 564 if (cpucache->waiting) 565 wakeup_mycpu(&oc->depot[myclusterid]); 566 crit_exit(); 567 return; 568 } 569 570 /* 571 * Both magazines full. Get an empty magazine from the depot and 572 * move a full loaded magazine to the depot. Even though the 573 * magazine may wind up with space available after we block on 574 * the spinlock, we still cycle it through to avoid the non-optimal 575 * corner-case. 576 * 577 * Obtain the depot spinlock. 578 */ 579 depot = &oc->depot[myclusterid]; 580 spin_lock(&depot->spin); 581 582 /* 583 * If an empty magazine is available in the depot, cycle it 584 * through and retry. 585 */ 586 if (!SLIST_EMPTY(&depot->emptymagazines)) { 587 KKASSERT(cpucache->previous_magazine->cleaning + 588 cpucache->loaded_magazine->cleaning == 0); 589 loadedmag = cpucache->previous_magazine; 590 cpucache->previous_magazine = cpucache->loaded_magazine; 591 cpucache->loaded_magazine = SLIST_FIRST(&depot->emptymagazines); 592 SLIST_REMOVE_HEAD(&depot->emptymagazines, nextmagazine); 593 594 /* 595 * Return loadedmag to the depot. Due to blocking it may 596 * not be entirely full and could even be empty. 597 */ 598 if (MAGAZINE_EMPTY(loadedmag)) { 599 SLIST_INSERT_HEAD(&depot->emptymagazines, 600 loadedmag, nextmagazine); 601 spin_unlock(&depot->spin); 602 } else { 603 SLIST_INSERT_HEAD(&depot->fullmagazines, 604 loadedmag, nextmagazine); 605 spin_unlock(&depot->spin); 606 if (depot->waiting) 607 wakeup(depot); 608 } 609 goto retry; 610 } 611 612 /* 613 * An empty mag is not available. This is a corner case which can 614 * occur due to cpus holding partially full magazines. Do not try 615 * to allocate a mag, just free the object. 616 */ 617 ++depot->unallocated_objects; 618 spin_unlock(&depot->spin); 619 if (depot->waiting) 620 wakeup(depot); 621 crit_exit(); 622 oc->dtor(obj, oc->privdata); 623 oc->free(obj, oc->allocator_args); 624 } 625 626 /* 627 * The object is being put back into the cache, but the caller has 628 * indicated that the object is not in any shape to be reused and should 629 * be dtor'd immediately. 630 */ 631 void 632 objcache_dtor(struct objcache *oc, void *obj) 633 { 634 struct magazinedepot *depot; 635 636 depot = &oc->depot[myclusterid]; 637 spin_lock(&depot->spin); 638 ++depot->unallocated_objects; 639 spin_unlock(&depot->spin); 640 if (depot->waiting) 641 wakeup(depot); 642 oc->dtor(obj, oc->privdata); 643 oc->free(obj, oc->allocator_args); 644 } 645 646 /* 647 * Deallocate all objects in a magazine and free the magazine if requested. 648 * The magazine must already be disassociated from the depot. 649 * 650 * Must be called with a critical section held when called with a per-cpu 651 * magazine. The magazine may be indirectly modified during the loop. 652 * 653 * The number of objects freed is returned. 654 */ 655 static int 656 mag_purge(struct objcache *oc, struct magazine *mag, int freeit) 657 { 658 int count; 659 void *obj; 660 661 count = 0; 662 ++mag->cleaning; 663 while (mag->rounds) { 664 obj = mag->objects[--mag->rounds]; 665 oc->dtor(obj, oc->privdata); /* MAY BLOCK */ 666 oc->free(obj, oc->allocator_args); /* MAY BLOCK */ 667 ++count; 668 669 /* 670 * Cycle for interrupts 671 */ 672 if ((count & 15) == 0) { 673 crit_exit(); 674 crit_enter(); 675 } 676 } 677 --mag->cleaning; 678 if (freeit) 679 kfree(mag, M_OBJMAG); 680 return(count); 681 } 682 683 /* 684 * Disassociate zero or more magazines from a magazine list associated with 685 * the depot, update the depot, and move the magazines to a temporary 686 * list. 687 * 688 * The caller must check the depot for waiters and wake it up, typically 689 * after disposing of the magazines this function loads onto the temporary 690 * list. 691 */ 692 static void 693 maglist_disassociate(struct magazinedepot *depot, struct magazinelist *maglist, 694 struct magazinelist *tmplist, boolean_t purgeall) 695 { 696 struct magazine *mag; 697 698 while ((mag = SLIST_FIRST(maglist)) != NULL) { 699 SLIST_REMOVE_HEAD(maglist, nextmagazine); 700 SLIST_INSERT_HEAD(tmplist, mag, nextmagazine); 701 depot->unallocated_objects += mag->rounds; 702 } 703 } 704 705 /* 706 * Deallocate all magazines and their contents from the passed temporary 707 * list. The magazines have already been accounted for by their depots. 708 * 709 * The total number of rounds freed is returned. This number is typically 710 * only used to determine whether a wakeup on the depot is needed or not. 711 */ 712 static int 713 maglist_purge(struct objcache *oc, struct magazinelist *maglist) 714 { 715 struct magazine *mag; 716 int count = 0; 717 718 /* 719 * can't use SLIST_FOREACH because blocking releases the depot 720 * spinlock 721 */ 722 crit_enter(); 723 while ((mag = SLIST_FIRST(maglist)) != NULL) { 724 SLIST_REMOVE_HEAD(maglist, nextmagazine); 725 count += mag_purge(oc, mag, TRUE); 726 } 727 crit_exit(); 728 return(count); 729 } 730 731 /* 732 * De-allocates all magazines on the full and empty magazine lists. 733 * 734 * Because this routine is called with a spinlock held, the magazines 735 * can only be disassociated and moved to a temporary list, not freed. 736 * 737 * The caller is responsible for freeing the magazines. 738 */ 739 static void 740 depot_disassociate(struct magazinedepot *depot, struct magazinelist *tmplist) 741 { 742 maglist_disassociate(depot, &depot->fullmagazines, tmplist, TRUE); 743 maglist_disassociate(depot, &depot->emptymagazines, tmplist, TRUE); 744 } 745 746 #ifdef notneeded 747 void 748 objcache_reclaim(struct objcache *oc) 749 { 750 struct percpu_objcache *cache_percpu = &oc->cache_percpu[myclusterid]; 751 struct magazinedepot *depot = &oc->depot[myclusterid]; 752 struct magazinelist tmplist; 753 int count; 754 755 SLIST_INIT(&tmplist); 756 crit_enter(); 757 count = mag_purge(oc, cache_percpu->loaded_magazine, FALSE); 758 count += mag_purge(oc, cache_percpu->previous_magazine, FALSE); 759 crit_exit(); 760 761 spin_lock(&depot->spin); 762 depot->unallocated_objects += count; 763 depot_disassociate(depot, &tmplist); 764 spin_unlock(&depot->spin); 765 count += maglist_purge(oc, &tmplist); 766 if (count && depot->waiting) 767 wakeup(depot); 768 } 769 #endif 770 771 /* 772 * Try to free up some memory. Return as soon as some free memory is found. 773 * For each object cache on the reclaim list, first try the current per-cpu 774 * cache, then the full magazine depot. 775 */ 776 boolean_t 777 objcache_reclaimlist(struct objcache *oclist[], int nlist, int ocflags) 778 { 779 struct objcache *oc; 780 struct percpu_objcache *cpucache; 781 struct magazinedepot *depot; 782 struct magazinelist tmplist; 783 int i, count; 784 785 SLIST_INIT(&tmplist); 786 787 for (i = 0; i < nlist; i++) { 788 oc = oclist[i]; 789 cpucache = &oc->cache_percpu[mycpuid]; 790 depot = &oc->depot[myclusterid]; 791 792 crit_enter(); 793 count = mag_purge(oc, cpucache->loaded_magazine, FALSE); 794 if (count == 0) 795 count += mag_purge(oc, cpucache->previous_magazine, FALSE); 796 crit_exit(); 797 if (count > 0) { 798 spin_lock(&depot->spin); 799 depot->unallocated_objects += count; 800 spin_unlock(&depot->spin); 801 if (depot->waiting) 802 wakeup(depot); 803 return (TRUE); 804 } 805 spin_lock(&depot->spin); 806 maglist_disassociate(depot, &depot->fullmagazines, 807 &tmplist, FALSE); 808 spin_unlock(&depot->spin); 809 count = maglist_purge(oc, &tmplist); 810 if (count > 0) { 811 if (depot->waiting) 812 wakeup(depot); 813 return (TRUE); 814 } 815 } 816 return (FALSE); 817 } 818 819 /* 820 * Destroy an object cache. Must have no existing references. 821 */ 822 void 823 objcache_destroy(struct objcache *oc) 824 { 825 struct percpu_objcache *cache_percpu; 826 struct magazinedepot *depot; 827 int clusterid, cpuid; 828 struct magazinelist tmplist; 829 830 spin_lock(&objcachelist_spin); 831 LIST_REMOVE(oc, oc_next); 832 spin_unlock(&objcachelist_spin); 833 834 SLIST_INIT(&tmplist); 835 for (clusterid = 0; clusterid < MAXCLUSTERS; clusterid++) { 836 depot = &oc->depot[clusterid]; 837 spin_lock(&depot->spin); 838 depot_disassociate(depot, &tmplist); 839 spin_unlock(&depot->spin); 840 } 841 maglist_purge(oc, &tmplist); 842 843 for (cpuid = 0; cpuid < ncpus; cpuid++) { 844 cache_percpu = &oc->cache_percpu[cpuid]; 845 846 crit_enter(); 847 mag_purge(oc, cache_percpu->loaded_magazine, TRUE); 848 mag_purge(oc, cache_percpu->previous_magazine, TRUE); 849 crit_exit(); 850 cache_percpu->loaded_magazine = NULL; 851 cache_percpu->previous_magazine = NULL; 852 /* don't bother adjusting depot->unallocated_objects */ 853 } 854 855 kfree(oc->name, M_TEMP); 856 kfree(oc, M_OBJCACHE); 857 } 858 859 #if 0 860 /* 861 * Populate the per-cluster depot with elements from a linear block 862 * of memory. Must be called for individually for each cluster. 863 * Populated depots should not be destroyed. 864 */ 865 void 866 objcache_populate_linear(struct objcache *oc, void *base, int nelts, int size) 867 { 868 char *p = base; 869 char *end = (char *)base + (nelts * size); 870 struct magazinedepot *depot = &oc->depot[myclusterid]; 871 struct magazine *emptymag = mag_alloc(depot->magcapcity); 872 873 while (p < end) { 874 emptymag->objects[emptymag->rounds++] = p; 875 if (MAGAZINE_FULL(emptymag)) { 876 spin_lock_wr(&depot->spin); 877 SLIST_INSERT_HEAD(&depot->fullmagazines, emptymag, 878 nextmagazine); 879 depot->unallocated_objects += emptymag->rounds; 880 spin_unlock_wr(&depot->spin); 881 if (depot->waiting) 882 wakeup(depot); 883 emptymag = mag_alloc(depot->magcapacity); 884 } 885 p += size; 886 } 887 if (MAGAZINE_EMPTY(emptymag)) { 888 crit_enter(); 889 mag_purge(oc, emptymag, TRUE); 890 crit_exit(); 891 } else { 892 spin_lock_wr(&depot->spin); 893 SLIST_INSERT_HEAD(&depot->fullmagazines, emptymag, 894 nextmagazine); 895 depot->unallocated_objects += emptymag->rounds; 896 spin_unlock_wr(&depot->spin); 897 if (depot->waiting) 898 wakeup(depot); 899 emptymag = mag_alloc(depot->magcapacity); 900 } 901 } 902 #endif 903 904 #if 0 905 /* 906 * Check depot contention once a minute. 907 * 2 contested locks per second allowed. 908 */ 909 static int objcache_rebalance_period; 910 static const int objcache_contention_rate = 120; 911 static struct callout objcache_callout; 912 913 #define MAXMAGSIZE 512 914 915 /* 916 * Check depot contention and increase magazine size if necessary. 917 */ 918 static void 919 objcache_timer(void *dummy) 920 { 921 struct objcache *oc; 922 struct magazinedepot *depot; 923 struct magazinelist tmplist; 924 925 XXX we need to detect when an objcache is destroyed out from under 926 us XXX 927 928 SLIST_INIT(&tmplist); 929 930 spin_lock_wr(&objcachelist_spin); 931 LIST_FOREACH(oc, &allobjcaches, oc_next) { 932 depot = &oc->depot[myclusterid]; 933 if (depot->magcapacity < MAXMAGSIZE) { 934 if (depot->contested > objcache_contention_rate) { 935 spin_lock_wr(&depot->spin); 936 depot_disassociate(depot, &tmplist); 937 depot->magcapacity *= 2; 938 spin_unlock_wr(&depot->spin); 939 kprintf("objcache_timer: increasing cache %s" 940 " magsize to %d, contested %d times\n", 941 oc->name, depot->magcapacity, 942 depot->contested); 943 } 944 depot->contested = 0; 945 } 946 spin_unlock_wr(&objcachelist_spin); 947 if (maglist_purge(oc, &tmplist) > 0 && depot->waiting) 948 wakeup(depot); 949 spin_lock_wr(&objcachelist_spin); 950 } 951 spin_unlock_wr(&objcachelist_spin); 952 953 callout_reset(&objcache_callout, objcache_rebalance_period, 954 objcache_timer, NULL); 955 } 956 957 #endif 958 959 static void 960 objcache_init(void) 961 { 962 spin_init(&objcachelist_spin); 963 #if 0 964 callout_init(&objcache_callout); 965 objcache_rebalance_period = 60 * hz; 966 callout_reset(&objcache_callout, objcache_rebalance_period, 967 objcache_timer, NULL); 968 #endif 969 } 970 SYSINIT(objcache, SI_BOOT2_OBJCACHE, SI_ORDER_FIRST, objcache_init, 0); 971