1 /* $NetBSD: subr_vmem.c,v 1.108 2022/05/31 08:43:16 andvar Exp $ */ 2 3 /*- 4 * Copyright (c)2006,2007,2008,2009 YAMAMOTO Takashi, 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, 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 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 */ 28 29 /* 30 * reference: 31 * - Magazines and Vmem: Extending the Slab Allocator 32 * to Many CPUs and Arbitrary Resources 33 * http://www.usenix.org/event/usenix01/bonwick.html 34 * 35 * locking & the boundary tag pool: 36 * - A pool(9) is used for vmem boundary tags 37 * - During a pool get call the global vmem_btag_refill_lock is taken, 38 * to serialize access to the allocation reserve, but no other 39 * vmem arena locks. 40 * - During pool_put calls no vmem mutexes are locked. 41 * - pool_drain doesn't hold the pool's mutex while releasing memory to 42 * its backing therefore no interference with any vmem mutexes. 43 * - The boundary tag pool is forced to put page headers into pool pages 44 * (PR_PHINPAGE) and not off page to avoid pool recursion. 45 * (due to sizeof(bt_t) it should be the case anyway) 46 */ 47 48 #include <sys/cdefs.h> 49 __KERNEL_RCSID(0, "$NetBSD: subr_vmem.c,v 1.108 2022/05/31 08:43:16 andvar Exp $"); 50 51 #if defined(_KERNEL) && defined(_KERNEL_OPT) 52 #include "opt_ddb.h" 53 #endif /* defined(_KERNEL) && defined(_KERNEL_OPT) */ 54 55 #include <sys/param.h> 56 #include <sys/hash.h> 57 #include <sys/queue.h> 58 #include <sys/bitops.h> 59 60 #if defined(_KERNEL) 61 #include <sys/systm.h> 62 #include <sys/kernel.h> /* hz */ 63 #include <sys/callout.h> 64 #include <sys/kmem.h> 65 #include <sys/pool.h> 66 #include <sys/vmem.h> 67 #include <sys/vmem_impl.h> 68 #include <sys/workqueue.h> 69 #include <sys/atomic.h> 70 #include <uvm/uvm.h> 71 #include <uvm/uvm_extern.h> 72 #include <uvm/uvm_km.h> 73 #include <uvm/uvm_page.h> 74 #include <uvm/uvm_pdaemon.h> 75 #else /* defined(_KERNEL) */ 76 #include <stdio.h> 77 #include <errno.h> 78 #include <assert.h> 79 #include <stdlib.h> 80 #include <string.h> 81 #include "../sys/vmem.h" 82 #include "../sys/vmem_impl.h" 83 #endif /* defined(_KERNEL) */ 84 85 86 #if defined(_KERNEL) 87 #include <sys/evcnt.h> 88 #define VMEM_EVCNT_DEFINE(name) \ 89 struct evcnt vmem_evcnt_##name = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, \ 90 "vmem", #name); \ 91 EVCNT_ATTACH_STATIC(vmem_evcnt_##name); 92 #define VMEM_EVCNT_INCR(ev) vmem_evcnt_##ev.ev_count++ 93 #define VMEM_EVCNT_DECR(ev) vmem_evcnt_##ev.ev_count-- 94 95 VMEM_EVCNT_DEFINE(static_bt_count) 96 VMEM_EVCNT_DEFINE(static_bt_inuse) 97 98 #define VMEM_CONDVAR_INIT(vm, wchan) cv_init(&vm->vm_cv, wchan) 99 #define VMEM_CONDVAR_DESTROY(vm) cv_destroy(&vm->vm_cv) 100 #define VMEM_CONDVAR_WAIT(vm) cv_wait(&vm->vm_cv, &vm->vm_lock) 101 #define VMEM_CONDVAR_BROADCAST(vm) cv_broadcast(&vm->vm_cv) 102 103 #else /* defined(_KERNEL) */ 104 105 #define VMEM_EVCNT_INCR(ev) /* nothing */ 106 #define VMEM_EVCNT_DECR(ev) /* nothing */ 107 108 #define VMEM_CONDVAR_INIT(vm, wchan) /* nothing */ 109 #define VMEM_CONDVAR_DESTROY(vm) /* nothing */ 110 #define VMEM_CONDVAR_WAIT(vm) /* nothing */ 111 #define VMEM_CONDVAR_BROADCAST(vm) /* nothing */ 112 113 #define UNITTEST 114 #define KASSERT(a) assert(a) 115 #define mutex_init(a, b, c) /* nothing */ 116 #define mutex_destroy(a) /* nothing */ 117 #define mutex_enter(a) /* nothing */ 118 #define mutex_tryenter(a) true 119 #define mutex_exit(a) /* nothing */ 120 #define mutex_owned(a) /* nothing */ 121 #define ASSERT_SLEEPABLE() /* nothing */ 122 #define panic(...) printf(__VA_ARGS__); abort() 123 #endif /* defined(_KERNEL) */ 124 125 #if defined(VMEM_SANITY) 126 static void vmem_check(vmem_t *); 127 #else /* defined(VMEM_SANITY) */ 128 #define vmem_check(vm) /* nothing */ 129 #endif /* defined(VMEM_SANITY) */ 130 131 #define VMEM_HASHSIZE_MIN 1 /* XXX */ 132 #define VMEM_HASHSIZE_MAX 65536 /* XXX */ 133 #define VMEM_HASHSIZE_INIT 1 134 135 #define VM_FITMASK (VM_BESTFIT | VM_INSTANTFIT) 136 137 #if defined(_KERNEL) 138 static bool vmem_bootstrapped = false; 139 static kmutex_t vmem_list_lock; 140 static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list); 141 #endif /* defined(_KERNEL) */ 142 143 /* ---- misc */ 144 145 #define VMEM_LOCK(vm) mutex_enter(&vm->vm_lock) 146 #define VMEM_TRYLOCK(vm) mutex_tryenter(&vm->vm_lock) 147 #define VMEM_UNLOCK(vm) mutex_exit(&vm->vm_lock) 148 #define VMEM_LOCK_INIT(vm, ipl) mutex_init(&vm->vm_lock, MUTEX_DEFAULT, ipl) 149 #define VMEM_LOCK_DESTROY(vm) mutex_destroy(&vm->vm_lock) 150 #define VMEM_ASSERT_LOCKED(vm) KASSERT(mutex_owned(&vm->vm_lock)) 151 152 #define VMEM_ALIGNUP(addr, align) \ 153 (-(-(addr) & -(align))) 154 155 #define VMEM_CROSS_P(addr1, addr2, boundary) \ 156 ((((addr1) ^ (addr2)) & -(boundary)) != 0) 157 158 #define ORDER2SIZE(order) ((vmem_size_t)1 << (order)) 159 #define SIZE2ORDER(size) ((int)ilog2(size)) 160 161 #if !defined(_KERNEL) 162 #define xmalloc(sz, flags) malloc(sz) 163 #define xfree(p, sz) free(p) 164 #define bt_alloc(vm, flags) malloc(sizeof(bt_t)) 165 #define bt_free(vm, bt) free(bt) 166 #else /* defined(_KERNEL) */ 167 168 #define xmalloc(sz, flags) \ 169 kmem_alloc(sz, ((flags) & VM_SLEEP) ? KM_SLEEP : KM_NOSLEEP); 170 #define xfree(p, sz) kmem_free(p, sz); 171 172 /* 173 * BT_RESERVE calculation: 174 * we allocate memory for boundary tags with vmem; therefore we have 175 * to keep a reserve of bts used to allocated memory for bts. 176 * This reserve is 4 for each arena involved in allocating vmems memory. 177 * BT_MAXFREE: don't cache excessive counts of bts in arenas 178 */ 179 #define STATIC_BT_COUNT 200 180 #define BT_MINRESERVE 4 181 #define BT_MAXFREE 64 182 183 static struct vmem_btag static_bts[STATIC_BT_COUNT]; 184 static int static_bt_count = STATIC_BT_COUNT; 185 186 static struct vmem kmem_va_meta_arena_store; 187 vmem_t *kmem_va_meta_arena; 188 static struct vmem kmem_meta_arena_store; 189 vmem_t *kmem_meta_arena = NULL; 190 191 static kmutex_t vmem_btag_refill_lock; 192 static kmutex_t vmem_btag_lock; 193 static LIST_HEAD(, vmem_btag) vmem_btag_freelist; 194 static size_t vmem_btag_freelist_count = 0; 195 static struct pool vmem_btag_pool; 196 197 static void vmem_xfree_bt(vmem_t *, bt_t *); 198 199 static void 200 vmem_kick_pdaemon(void) 201 { 202 #if defined(_KERNEL) 203 uvm_kick_pdaemon(); 204 #endif 205 } 206 207 /* ---- boundary tag */ 208 209 static int bt_refill(vmem_t *vm); 210 static int bt_refill_locked(vmem_t *vm); 211 212 static void * 213 pool_page_alloc_vmem_meta(struct pool *pp, int flags) 214 { 215 const vm_flag_t vflags = (flags & PR_WAITOK) ? VM_SLEEP: VM_NOSLEEP; 216 vmem_addr_t va; 217 int ret; 218 219 ret = vmem_alloc(kmem_meta_arena, pp->pr_alloc->pa_pagesz, 220 (vflags & ~VM_FITMASK) | VM_INSTANTFIT | VM_POPULATING, &va); 221 222 return ret ? NULL : (void *)va; 223 } 224 225 static void 226 pool_page_free_vmem_meta(struct pool *pp, void *v) 227 { 228 229 vmem_free(kmem_meta_arena, (vmem_addr_t)v, pp->pr_alloc->pa_pagesz); 230 } 231 232 /* allocator for vmem-pool metadata */ 233 struct pool_allocator pool_allocator_vmem_meta = { 234 .pa_alloc = pool_page_alloc_vmem_meta, 235 .pa_free = pool_page_free_vmem_meta, 236 .pa_pagesz = 0 237 }; 238 239 static int 240 bt_refill_locked(vmem_t *vm) 241 { 242 bt_t *bt; 243 244 VMEM_ASSERT_LOCKED(vm); 245 246 if (vm->vm_nfreetags > BT_MINRESERVE) { 247 return 0; 248 } 249 250 mutex_enter(&vmem_btag_lock); 251 while (!LIST_EMPTY(&vmem_btag_freelist) && 252 vm->vm_nfreetags <= BT_MINRESERVE) { 253 bt = LIST_FIRST(&vmem_btag_freelist); 254 LIST_REMOVE(bt, bt_freelist); 255 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist); 256 vm->vm_nfreetags++; 257 vmem_btag_freelist_count--; 258 VMEM_EVCNT_INCR(static_bt_inuse); 259 } 260 mutex_exit(&vmem_btag_lock); 261 262 while (vm->vm_nfreetags <= BT_MINRESERVE) { 263 VMEM_UNLOCK(vm); 264 mutex_enter(&vmem_btag_refill_lock); 265 bt = pool_get(&vmem_btag_pool, PR_NOWAIT); 266 mutex_exit(&vmem_btag_refill_lock); 267 VMEM_LOCK(vm); 268 if (bt == NULL) 269 break; 270 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist); 271 vm->vm_nfreetags++; 272 } 273 274 if (vm->vm_nfreetags <= BT_MINRESERVE) { 275 return ENOMEM; 276 } 277 278 if (kmem_meta_arena != NULL) { 279 VMEM_UNLOCK(vm); 280 (void)bt_refill(kmem_arena); 281 (void)bt_refill(kmem_va_meta_arena); 282 (void)bt_refill(kmem_meta_arena); 283 VMEM_LOCK(vm); 284 } 285 286 return 0; 287 } 288 289 static int 290 bt_refill(vmem_t *vm) 291 { 292 int rv; 293 294 VMEM_LOCK(vm); 295 rv = bt_refill_locked(vm); 296 VMEM_UNLOCK(vm); 297 return rv; 298 } 299 300 static bt_t * 301 bt_alloc(vmem_t *vm, vm_flag_t flags) 302 { 303 bt_t *bt; 304 305 VMEM_ASSERT_LOCKED(vm); 306 307 while (vm->vm_nfreetags <= BT_MINRESERVE && (flags & VM_POPULATING) == 0) { 308 if (bt_refill_locked(vm)) { 309 if ((flags & VM_NOSLEEP) != 0) { 310 return NULL; 311 } 312 313 /* 314 * It would be nice to wait for something specific here 315 * but there are multiple ways that a retry could 316 * succeed and we can't wait for multiple things 317 * simultaneously. So we'll just sleep for an arbitrary 318 * short period of time and retry regardless. 319 * This should be a very rare case. 320 */ 321 322 vmem_kick_pdaemon(); 323 kpause("btalloc", false, 1, &vm->vm_lock); 324 } 325 } 326 bt = LIST_FIRST(&vm->vm_freetags); 327 LIST_REMOVE(bt, bt_freelist); 328 vm->vm_nfreetags--; 329 330 return bt; 331 } 332 333 static void 334 bt_free(vmem_t *vm, bt_t *bt) 335 { 336 337 VMEM_ASSERT_LOCKED(vm); 338 339 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist); 340 vm->vm_nfreetags++; 341 } 342 343 static void 344 bt_freetrim(vmem_t *vm, int freelimit) 345 { 346 bt_t *t; 347 LIST_HEAD(, vmem_btag) tofree; 348 349 VMEM_ASSERT_LOCKED(vm); 350 351 LIST_INIT(&tofree); 352 353 while (vm->vm_nfreetags > freelimit) { 354 bt_t *bt = LIST_FIRST(&vm->vm_freetags); 355 LIST_REMOVE(bt, bt_freelist); 356 vm->vm_nfreetags--; 357 if (bt >= static_bts 358 && bt < &static_bts[STATIC_BT_COUNT]) { 359 mutex_enter(&vmem_btag_lock); 360 LIST_INSERT_HEAD(&vmem_btag_freelist, bt, bt_freelist); 361 vmem_btag_freelist_count++; 362 mutex_exit(&vmem_btag_lock); 363 VMEM_EVCNT_DECR(static_bt_inuse); 364 } else { 365 LIST_INSERT_HEAD(&tofree, bt, bt_freelist); 366 } 367 } 368 369 VMEM_UNLOCK(vm); 370 while (!LIST_EMPTY(&tofree)) { 371 t = LIST_FIRST(&tofree); 372 LIST_REMOVE(t, bt_freelist); 373 pool_put(&vmem_btag_pool, t); 374 } 375 } 376 #endif /* defined(_KERNEL) */ 377 378 /* 379 * freelist[0] ... [1, 1] 380 * freelist[1] ... [2, 3] 381 * freelist[2] ... [4, 7] 382 * freelist[3] ... [8, 15] 383 * : 384 * freelist[n] ... [(1 << n), (1 << (n + 1)) - 1] 385 * : 386 */ 387 388 static struct vmem_freelist * 389 bt_freehead_tofree(vmem_t *vm, vmem_size_t size) 390 { 391 const vmem_size_t qsize = size >> vm->vm_quantum_shift; 392 const int idx = SIZE2ORDER(qsize); 393 394 KASSERT(size != 0 && qsize != 0); 395 KASSERT((size & vm->vm_quantum_mask) == 0); 396 KASSERT(idx >= 0); 397 KASSERT(idx < VMEM_MAXORDER); 398 399 return &vm->vm_freelist[idx]; 400 } 401 402 /* 403 * bt_freehead_toalloc: return the freelist for the given size and allocation 404 * strategy. 405 * 406 * for VM_INSTANTFIT, return the list in which any blocks are large enough 407 * for the requested size. otherwise, return the list which can have blocks 408 * large enough for the requested size. 409 */ 410 411 static struct vmem_freelist * 412 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, vm_flag_t strat) 413 { 414 const vmem_size_t qsize = size >> vm->vm_quantum_shift; 415 int idx = SIZE2ORDER(qsize); 416 417 KASSERT(size != 0 && qsize != 0); 418 KASSERT((size & vm->vm_quantum_mask) == 0); 419 420 if (strat == VM_INSTANTFIT && ORDER2SIZE(idx) != qsize) { 421 idx++; 422 /* check too large request? */ 423 } 424 KASSERT(idx >= 0); 425 KASSERT(idx < VMEM_MAXORDER); 426 427 return &vm->vm_freelist[idx]; 428 } 429 430 /* ---- boundary tag hash */ 431 432 static struct vmem_hashlist * 433 bt_hashhead(vmem_t *vm, vmem_addr_t addr) 434 { 435 struct vmem_hashlist *list; 436 unsigned int hash; 437 438 hash = hash32_buf(&addr, sizeof(addr), HASH32_BUF_INIT); 439 list = &vm->vm_hashlist[hash & vm->vm_hashmask]; 440 441 return list; 442 } 443 444 static bt_t * 445 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr) 446 { 447 struct vmem_hashlist *list; 448 bt_t *bt; 449 450 list = bt_hashhead(vm, addr); 451 LIST_FOREACH(bt, list, bt_hashlist) { 452 if (bt->bt_start == addr) { 453 break; 454 } 455 } 456 457 return bt; 458 } 459 460 static void 461 bt_rembusy(vmem_t *vm, bt_t *bt) 462 { 463 464 KASSERT(vm->vm_nbusytag > 0); 465 vm->vm_inuse -= bt->bt_size; 466 vm->vm_nbusytag--; 467 LIST_REMOVE(bt, bt_hashlist); 468 } 469 470 static void 471 bt_insbusy(vmem_t *vm, bt_t *bt) 472 { 473 struct vmem_hashlist *list; 474 475 KASSERT(bt->bt_type == BT_TYPE_BUSY); 476 477 list = bt_hashhead(vm, bt->bt_start); 478 LIST_INSERT_HEAD(list, bt, bt_hashlist); 479 if (++vm->vm_nbusytag > vm->vm_maxbusytag) { 480 vm->vm_maxbusytag = vm->vm_nbusytag; 481 } 482 vm->vm_inuse += bt->bt_size; 483 } 484 485 /* ---- boundary tag list */ 486 487 static void 488 bt_remseg(vmem_t *vm, bt_t *bt) 489 { 490 491 TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist); 492 } 493 494 static void 495 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev) 496 { 497 498 TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist); 499 } 500 501 static void 502 bt_insseg_tail(vmem_t *vm, bt_t *bt) 503 { 504 505 TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist); 506 } 507 508 static void 509 bt_remfree(vmem_t *vm, bt_t *bt) 510 { 511 512 KASSERT(bt->bt_type == BT_TYPE_FREE); 513 514 LIST_REMOVE(bt, bt_freelist); 515 } 516 517 static void 518 bt_insfree(vmem_t *vm, bt_t *bt) 519 { 520 struct vmem_freelist *list; 521 522 list = bt_freehead_tofree(vm, bt->bt_size); 523 LIST_INSERT_HEAD(list, bt, bt_freelist); 524 } 525 526 /* ---- vmem internal functions */ 527 528 #if defined(QCACHE) 529 static inline vm_flag_t 530 prf_to_vmf(int prflags) 531 { 532 vm_flag_t vmflags; 533 534 KASSERT((prflags & ~(PR_LIMITFAIL | PR_WAITOK | PR_NOWAIT)) == 0); 535 if ((prflags & PR_WAITOK) != 0) { 536 vmflags = VM_SLEEP; 537 } else { 538 vmflags = VM_NOSLEEP; 539 } 540 return vmflags; 541 } 542 543 static inline int 544 vmf_to_prf(vm_flag_t vmflags) 545 { 546 int prflags; 547 548 if ((vmflags & VM_SLEEP) != 0) { 549 prflags = PR_WAITOK; 550 } else { 551 prflags = PR_NOWAIT; 552 } 553 return prflags; 554 } 555 556 static size_t 557 qc_poolpage_size(size_t qcache_max) 558 { 559 int i; 560 561 for (i = 0; ORDER2SIZE(i) <= qcache_max * 3; i++) { 562 /* nothing */ 563 } 564 return ORDER2SIZE(i); 565 } 566 567 static void * 568 qc_poolpage_alloc(struct pool *pool, int prflags) 569 { 570 qcache_t *qc = QC_POOL_TO_QCACHE(pool); 571 vmem_t *vm = qc->qc_vmem; 572 vmem_addr_t addr; 573 574 if (vmem_alloc(vm, pool->pr_alloc->pa_pagesz, 575 prf_to_vmf(prflags) | VM_INSTANTFIT, &addr) != 0) 576 return NULL; 577 return (void *)addr; 578 } 579 580 static void 581 qc_poolpage_free(struct pool *pool, void *addr) 582 { 583 qcache_t *qc = QC_POOL_TO_QCACHE(pool); 584 vmem_t *vm = qc->qc_vmem; 585 586 vmem_free(vm, (vmem_addr_t)addr, pool->pr_alloc->pa_pagesz); 587 } 588 589 static void 590 qc_init(vmem_t *vm, size_t qcache_max, int ipl) 591 { 592 qcache_t *prevqc; 593 struct pool_allocator *pa; 594 int qcache_idx_max; 595 int i; 596 597 KASSERT((qcache_max & vm->vm_quantum_mask) == 0); 598 if (qcache_max > (VMEM_QCACHE_IDX_MAX << vm->vm_quantum_shift)) { 599 qcache_max = VMEM_QCACHE_IDX_MAX << vm->vm_quantum_shift; 600 } 601 vm->vm_qcache_max = qcache_max; 602 pa = &vm->vm_qcache_allocator; 603 memset(pa, 0, sizeof(*pa)); 604 pa->pa_alloc = qc_poolpage_alloc; 605 pa->pa_free = qc_poolpage_free; 606 pa->pa_pagesz = qc_poolpage_size(qcache_max); 607 608 qcache_idx_max = qcache_max >> vm->vm_quantum_shift; 609 prevqc = NULL; 610 for (i = qcache_idx_max; i > 0; i--) { 611 qcache_t *qc = &vm->vm_qcache_store[i - 1]; 612 size_t size = i << vm->vm_quantum_shift; 613 pool_cache_t pc; 614 615 qc->qc_vmem = vm; 616 snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu", 617 vm->vm_name, size); 618 619 pc = pool_cache_init(size, 620 ORDER2SIZE(vm->vm_quantum_shift), 0, 621 PR_NOALIGN | PR_NOTOUCH | PR_RECURSIVE /* XXX */, 622 qc->qc_name, pa, ipl, NULL, NULL, NULL); 623 624 KASSERT(pc); 625 626 qc->qc_cache = pc; 627 KASSERT(qc->qc_cache != NULL); /* XXX */ 628 if (prevqc != NULL && 629 qc->qc_cache->pc_pool.pr_itemsperpage == 630 prevqc->qc_cache->pc_pool.pr_itemsperpage) { 631 pool_cache_destroy(qc->qc_cache); 632 vm->vm_qcache[i - 1] = prevqc; 633 continue; 634 } 635 qc->qc_cache->pc_pool.pr_qcache = qc; 636 vm->vm_qcache[i - 1] = qc; 637 prevqc = qc; 638 } 639 } 640 641 static void 642 qc_destroy(vmem_t *vm) 643 { 644 const qcache_t *prevqc; 645 int i; 646 int qcache_idx_max; 647 648 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift; 649 prevqc = NULL; 650 for (i = 0; i < qcache_idx_max; i++) { 651 qcache_t *qc = vm->vm_qcache[i]; 652 653 if (prevqc == qc) { 654 continue; 655 } 656 pool_cache_destroy(qc->qc_cache); 657 prevqc = qc; 658 } 659 } 660 #endif 661 662 #if defined(_KERNEL) 663 static void 664 vmem_bootstrap(void) 665 { 666 667 mutex_init(&vmem_list_lock, MUTEX_DEFAULT, IPL_NONE); 668 mutex_init(&vmem_btag_lock, MUTEX_DEFAULT, IPL_VM); 669 mutex_init(&vmem_btag_refill_lock, MUTEX_DEFAULT, IPL_VM); 670 671 while (static_bt_count-- > 0) { 672 bt_t *bt = &static_bts[static_bt_count]; 673 LIST_INSERT_HEAD(&vmem_btag_freelist, bt, bt_freelist); 674 VMEM_EVCNT_INCR(static_bt_count); 675 vmem_btag_freelist_count++; 676 } 677 vmem_bootstrapped = TRUE; 678 } 679 680 void 681 vmem_subsystem_init(vmem_t *vm) 682 { 683 684 kmem_va_meta_arena = vmem_init(&kmem_va_meta_arena_store, "vmem-va", 685 0, 0, PAGE_SIZE, vmem_alloc, vmem_free, vm, 686 0, VM_NOSLEEP | VM_BOOTSTRAP | VM_LARGEIMPORT, 687 IPL_VM); 688 689 kmem_meta_arena = vmem_init(&kmem_meta_arena_store, "vmem-meta", 690 0, 0, PAGE_SIZE, 691 uvm_km_kmem_alloc, uvm_km_kmem_free, kmem_va_meta_arena, 692 0, VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM); 693 694 pool_init(&vmem_btag_pool, sizeof(bt_t), coherency_unit, 0, 695 PR_PHINPAGE, "vmembt", &pool_allocator_vmem_meta, IPL_VM); 696 } 697 #endif /* defined(_KERNEL) */ 698 699 static int 700 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, vm_flag_t flags, 701 int spanbttype) 702 { 703 bt_t *btspan; 704 bt_t *btfree; 705 706 VMEM_ASSERT_LOCKED(vm); 707 KASSERT((flags & (VM_SLEEP|VM_NOSLEEP)) != 0); 708 KASSERT((~flags & (VM_SLEEP|VM_NOSLEEP)) != 0); 709 KASSERT(spanbttype == BT_TYPE_SPAN || 710 spanbttype == BT_TYPE_SPAN_STATIC); 711 712 btspan = bt_alloc(vm, flags); 713 if (btspan == NULL) { 714 return ENOMEM; 715 } 716 btfree = bt_alloc(vm, flags); 717 if (btfree == NULL) { 718 bt_free(vm, btspan); 719 return ENOMEM; 720 } 721 722 btspan->bt_type = spanbttype; 723 btspan->bt_start = addr; 724 btspan->bt_size = size; 725 726 btfree->bt_type = BT_TYPE_FREE; 727 btfree->bt_start = addr; 728 btfree->bt_size = size; 729 730 bt_insseg_tail(vm, btspan); 731 bt_insseg(vm, btfree, btspan); 732 bt_insfree(vm, btfree); 733 vm->vm_size += size; 734 735 return 0; 736 } 737 738 static void 739 vmem_destroy1(vmem_t *vm) 740 { 741 742 #if defined(QCACHE) 743 qc_destroy(vm); 744 #endif /* defined(QCACHE) */ 745 VMEM_LOCK(vm); 746 747 for (int i = 0; i < vm->vm_hashsize; i++) { 748 bt_t *bt; 749 750 while ((bt = LIST_FIRST(&vm->vm_hashlist[i])) != NULL) { 751 KASSERT(bt->bt_type == BT_TYPE_SPAN_STATIC); 752 LIST_REMOVE(bt, bt_hashlist); 753 bt_free(vm, bt); 754 } 755 } 756 757 /* bt_freetrim() drops the lock. */ 758 bt_freetrim(vm, 0); 759 if (vm->vm_hashlist != &vm->vm_hash0) { 760 xfree(vm->vm_hashlist, 761 sizeof(struct vmem_hashlist) * vm->vm_hashsize); 762 } 763 764 VMEM_CONDVAR_DESTROY(vm); 765 VMEM_LOCK_DESTROY(vm); 766 xfree(vm, sizeof(*vm)); 767 } 768 769 static int 770 vmem_import(vmem_t *vm, vmem_size_t size, vm_flag_t flags) 771 { 772 vmem_addr_t addr; 773 int rc; 774 775 VMEM_ASSERT_LOCKED(vm); 776 777 if (vm->vm_importfn == NULL) { 778 return EINVAL; 779 } 780 781 if (vm->vm_flags & VM_LARGEIMPORT) { 782 size *= 16; 783 } 784 785 VMEM_UNLOCK(vm); 786 if (vm->vm_flags & VM_XIMPORT) { 787 rc = __FPTRCAST(vmem_ximport_t *, vm->vm_importfn)(vm->vm_arg, 788 size, &size, flags, &addr); 789 } else { 790 rc = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr); 791 } 792 VMEM_LOCK(vm); 793 794 if (rc) { 795 return ENOMEM; 796 } 797 798 if (vmem_add1(vm, addr, size, flags, BT_TYPE_SPAN) != 0) { 799 VMEM_UNLOCK(vm); 800 (*vm->vm_releasefn)(vm->vm_arg, addr, size); 801 VMEM_LOCK(vm); 802 return ENOMEM; 803 } 804 805 return 0; 806 } 807 808 static int 809 vmem_rehash(vmem_t *vm, size_t newhashsize, vm_flag_t flags) 810 { 811 bt_t *bt; 812 int i; 813 struct vmem_hashlist *newhashlist; 814 struct vmem_hashlist *oldhashlist; 815 size_t oldhashsize; 816 817 KASSERT(newhashsize > 0); 818 819 /* Round hash size up to a power of 2. */ 820 newhashsize = 1 << (ilog2(newhashsize) + 1); 821 822 newhashlist = 823 xmalloc(sizeof(struct vmem_hashlist) * newhashsize, flags); 824 if (newhashlist == NULL) { 825 return ENOMEM; 826 } 827 for (i = 0; i < newhashsize; i++) { 828 LIST_INIT(&newhashlist[i]); 829 } 830 831 VMEM_LOCK(vm); 832 /* Decay back to a small hash slowly. */ 833 if (vm->vm_maxbusytag >= 2) { 834 vm->vm_maxbusytag = vm->vm_maxbusytag / 2 - 1; 835 if (vm->vm_nbusytag > vm->vm_maxbusytag) { 836 vm->vm_maxbusytag = vm->vm_nbusytag; 837 } 838 } else { 839 vm->vm_maxbusytag = vm->vm_nbusytag; 840 } 841 oldhashlist = vm->vm_hashlist; 842 oldhashsize = vm->vm_hashsize; 843 vm->vm_hashlist = newhashlist; 844 vm->vm_hashsize = newhashsize; 845 vm->vm_hashmask = newhashsize - 1; 846 if (oldhashlist == NULL) { 847 VMEM_UNLOCK(vm); 848 return 0; 849 } 850 for (i = 0; i < oldhashsize; i++) { 851 while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) { 852 bt_rembusy(vm, bt); /* XXX */ 853 bt_insbusy(vm, bt); 854 } 855 } 856 VMEM_UNLOCK(vm); 857 858 if (oldhashlist != &vm->vm_hash0) { 859 xfree(oldhashlist, 860 sizeof(struct vmem_hashlist) * oldhashsize); 861 } 862 863 return 0; 864 } 865 866 /* 867 * vmem_fit: check if a bt can satisfy the given restrictions. 868 * 869 * it's a caller's responsibility to ensure the region is big enough 870 * before calling us. 871 */ 872 873 static int 874 vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align, 875 vmem_size_t phase, vmem_size_t nocross, 876 vmem_addr_t minaddr, vmem_addr_t maxaddr, vmem_addr_t *addrp) 877 { 878 vmem_addr_t start; 879 vmem_addr_t end; 880 881 KASSERT(size > 0); 882 KASSERT(bt->bt_size >= size); /* caller's responsibility */ 883 884 /* 885 * XXX assumption: vmem_addr_t and vmem_size_t are 886 * unsigned integer of the same size. 887 */ 888 889 start = bt->bt_start; 890 if (start < minaddr) { 891 start = minaddr; 892 } 893 end = BT_END(bt); 894 if (end > maxaddr) { 895 end = maxaddr; 896 } 897 if (start > end) { 898 return ENOMEM; 899 } 900 901 start = VMEM_ALIGNUP(start - phase, align) + phase; 902 if (start < bt->bt_start) { 903 start += align; 904 } 905 if (VMEM_CROSS_P(start, start + size - 1, nocross)) { 906 KASSERT(align < nocross); 907 start = VMEM_ALIGNUP(start - phase, nocross) + phase; 908 } 909 if (start <= end && end - start >= size - 1) { 910 KASSERT((start & (align - 1)) == phase); 911 KASSERT(!VMEM_CROSS_P(start, start + size - 1, nocross)); 912 KASSERT(minaddr <= start); 913 KASSERT(maxaddr == 0 || start + size - 1 <= maxaddr); 914 KASSERT(bt->bt_start <= start); 915 KASSERT(BT_END(bt) - start >= size - 1); 916 *addrp = start; 917 return 0; 918 } 919 return ENOMEM; 920 } 921 922 /* ---- vmem API */ 923 924 /* 925 * vmem_init: creates a vmem arena. 926 */ 927 928 vmem_t * 929 vmem_init(vmem_t *vm, const char *name, 930 vmem_addr_t base, vmem_size_t size, vmem_size_t quantum, 931 vmem_import_t *importfn, vmem_release_t *releasefn, 932 vmem_t *arg, vmem_size_t qcache_max, vm_flag_t flags, int ipl) 933 { 934 int i; 935 936 KASSERT((flags & (VM_SLEEP|VM_NOSLEEP)) != 0); 937 KASSERT((~flags & (VM_SLEEP|VM_NOSLEEP)) != 0); 938 KASSERT(quantum > 0); 939 940 #if defined(_KERNEL) 941 /* XXX: SMP, we get called early... */ 942 if (!vmem_bootstrapped) { 943 vmem_bootstrap(); 944 } 945 #endif /* defined(_KERNEL) */ 946 947 if (vm == NULL) { 948 vm = xmalloc(sizeof(*vm), flags); 949 } 950 if (vm == NULL) { 951 return NULL; 952 } 953 954 VMEM_CONDVAR_INIT(vm, "vmem"); 955 VMEM_LOCK_INIT(vm, ipl); 956 vm->vm_flags = flags; 957 vm->vm_nfreetags = 0; 958 LIST_INIT(&vm->vm_freetags); 959 strlcpy(vm->vm_name, name, sizeof(vm->vm_name)); 960 vm->vm_quantum_mask = quantum - 1; 961 vm->vm_quantum_shift = SIZE2ORDER(quantum); 962 KASSERT(ORDER2SIZE(vm->vm_quantum_shift) == quantum); 963 vm->vm_importfn = importfn; 964 vm->vm_releasefn = releasefn; 965 vm->vm_arg = arg; 966 vm->vm_nbusytag = 0; 967 vm->vm_maxbusytag = 0; 968 vm->vm_size = 0; 969 vm->vm_inuse = 0; 970 #if defined(QCACHE) 971 qc_init(vm, qcache_max, ipl); 972 #endif /* defined(QCACHE) */ 973 974 TAILQ_INIT(&vm->vm_seglist); 975 for (i = 0; i < VMEM_MAXORDER; i++) { 976 LIST_INIT(&vm->vm_freelist[i]); 977 } 978 memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0)); 979 vm->vm_hashsize = 1; 980 vm->vm_hashmask = vm->vm_hashsize - 1; 981 vm->vm_hashlist = &vm->vm_hash0; 982 983 if (size != 0) { 984 if (vmem_add(vm, base, size, flags) != 0) { 985 vmem_destroy1(vm); 986 return NULL; 987 } 988 } 989 990 #if defined(_KERNEL) 991 if (flags & VM_BOOTSTRAP) { 992 bt_refill(vm); 993 } 994 995 mutex_enter(&vmem_list_lock); 996 LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist); 997 mutex_exit(&vmem_list_lock); 998 #endif /* defined(_KERNEL) */ 999 1000 return vm; 1001 } 1002 1003 1004 1005 /* 1006 * vmem_create: create an arena. 1007 * 1008 * => must not be called from interrupt context. 1009 */ 1010 1011 vmem_t * 1012 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size, 1013 vmem_size_t quantum, vmem_import_t *importfn, vmem_release_t *releasefn, 1014 vmem_t *source, vmem_size_t qcache_max, vm_flag_t flags, int ipl) 1015 { 1016 1017 KASSERT((flags & (VM_XIMPORT)) == 0); 1018 1019 return vmem_init(NULL, name, base, size, quantum, 1020 importfn, releasefn, source, qcache_max, flags, ipl); 1021 } 1022 1023 /* 1024 * vmem_xcreate: create an arena takes alternative import func. 1025 * 1026 * => must not be called from interrupt context. 1027 */ 1028 1029 vmem_t * 1030 vmem_xcreate(const char *name, vmem_addr_t base, vmem_size_t size, 1031 vmem_size_t quantum, vmem_ximport_t *importfn, vmem_release_t *releasefn, 1032 vmem_t *source, vmem_size_t qcache_max, vm_flag_t flags, int ipl) 1033 { 1034 1035 KASSERT((flags & (VM_XIMPORT)) == 0); 1036 1037 return vmem_init(NULL, name, base, size, quantum, 1038 __FPTRCAST(vmem_import_t *, importfn), releasefn, source, 1039 qcache_max, flags | VM_XIMPORT, ipl); 1040 } 1041 1042 void 1043 vmem_destroy(vmem_t *vm) 1044 { 1045 1046 #if defined(_KERNEL) 1047 mutex_enter(&vmem_list_lock); 1048 LIST_REMOVE(vm, vm_alllist); 1049 mutex_exit(&vmem_list_lock); 1050 #endif /* defined(_KERNEL) */ 1051 1052 vmem_destroy1(vm); 1053 } 1054 1055 vmem_size_t 1056 vmem_roundup_size(vmem_t *vm, vmem_size_t size) 1057 { 1058 1059 return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask; 1060 } 1061 1062 /* 1063 * vmem_alloc: allocate resource from the arena. 1064 */ 1065 1066 int 1067 vmem_alloc(vmem_t *vm, vmem_size_t size, vm_flag_t flags, vmem_addr_t *addrp) 1068 { 1069 const vm_flag_t strat __diagused = flags & VM_FITMASK; 1070 int error; 1071 1072 KASSERT((flags & (VM_SLEEP|VM_NOSLEEP)) != 0); 1073 KASSERT((~flags & (VM_SLEEP|VM_NOSLEEP)) != 0); 1074 1075 KASSERT(size > 0); 1076 KASSERT(strat == VM_BESTFIT || strat == VM_INSTANTFIT); 1077 if ((flags & VM_SLEEP) != 0) { 1078 ASSERT_SLEEPABLE(); 1079 } 1080 1081 #if defined(QCACHE) 1082 if (size <= vm->vm_qcache_max) { 1083 void *p; 1084 int qidx = (size + vm->vm_quantum_mask) >> vm->vm_quantum_shift; 1085 qcache_t *qc = vm->vm_qcache[qidx - 1]; 1086 1087 p = pool_cache_get(qc->qc_cache, vmf_to_prf(flags)); 1088 if (addrp != NULL) 1089 *addrp = (vmem_addr_t)p; 1090 error = (p == NULL) ? ENOMEM : 0; 1091 goto out; 1092 } 1093 #endif /* defined(QCACHE) */ 1094 1095 error = vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, 1096 flags, addrp); 1097 out: 1098 KASSERTMSG(error || addrp == NULL || 1099 (*addrp & vm->vm_quantum_mask) == 0, 1100 "vmem %s mask=0x%jx addr=0x%jx", 1101 vm->vm_name, (uintmax_t)vm->vm_quantum_mask, (uintmax_t)*addrp); 1102 KASSERT(error == 0 || (flags & VM_SLEEP) == 0); 1103 return error; 1104 } 1105 1106 int 1107 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align, 1108 const vmem_size_t phase, const vmem_size_t nocross, 1109 const vmem_addr_t minaddr, const vmem_addr_t maxaddr, const vm_flag_t flags, 1110 vmem_addr_t *addrp) 1111 { 1112 struct vmem_freelist *list; 1113 struct vmem_freelist *first; 1114 struct vmem_freelist *end; 1115 bt_t *bt; 1116 bt_t *btnew; 1117 bt_t *btnew2; 1118 const vmem_size_t size = vmem_roundup_size(vm, size0); 1119 vm_flag_t strat = flags & VM_FITMASK; 1120 vmem_addr_t start; 1121 int rc; 1122 1123 KASSERT(size0 > 0); 1124 KASSERT(size > 0); 1125 KASSERT(strat == VM_BESTFIT || strat == VM_INSTANTFIT); 1126 if ((flags & VM_SLEEP) != 0) { 1127 ASSERT_SLEEPABLE(); 1128 } 1129 KASSERT((align & vm->vm_quantum_mask) == 0); 1130 KASSERT((align & (align - 1)) == 0); 1131 KASSERT((phase & vm->vm_quantum_mask) == 0); 1132 KASSERT((nocross & vm->vm_quantum_mask) == 0); 1133 KASSERT((nocross & (nocross - 1)) == 0); 1134 KASSERT((align == 0 && phase == 0) || phase < align); 1135 KASSERT(nocross == 0 || nocross >= size); 1136 KASSERT(minaddr <= maxaddr); 1137 KASSERT(!VMEM_CROSS_P(phase, phase + size - 1, nocross)); 1138 1139 if (align == 0) { 1140 align = vm->vm_quantum_mask + 1; 1141 } 1142 1143 /* 1144 * allocate boundary tags before acquiring the vmem lock. 1145 */ 1146 VMEM_LOCK(vm); 1147 btnew = bt_alloc(vm, flags); 1148 if (btnew == NULL) { 1149 VMEM_UNLOCK(vm); 1150 return ENOMEM; 1151 } 1152 btnew2 = bt_alloc(vm, flags); /* XXX not necessary if no restrictions */ 1153 if (btnew2 == NULL) { 1154 bt_free(vm, btnew); 1155 VMEM_UNLOCK(vm); 1156 return ENOMEM; 1157 } 1158 1159 /* 1160 * choose a free block from which we allocate. 1161 */ 1162 retry_strat: 1163 first = bt_freehead_toalloc(vm, size, strat); 1164 end = &vm->vm_freelist[VMEM_MAXORDER]; 1165 retry: 1166 bt = NULL; 1167 vmem_check(vm); 1168 if (strat == VM_INSTANTFIT) { 1169 /* 1170 * just choose the first block which satisfies our restrictions. 1171 * 1172 * note that we don't need to check the size of the blocks 1173 * because any blocks found on these list should be larger than 1174 * the given size. 1175 */ 1176 for (list = first; list < end; list++) { 1177 bt = LIST_FIRST(list); 1178 if (bt != NULL) { 1179 rc = vmem_fit(bt, size, align, phase, 1180 nocross, minaddr, maxaddr, &start); 1181 if (rc == 0) { 1182 goto gotit; 1183 } 1184 /* 1185 * don't bother to follow the bt_freelist link 1186 * here. the list can be very long and we are 1187 * told to run fast. blocks from the later free 1188 * lists are larger and have better chances to 1189 * satisfy our restrictions. 1190 */ 1191 } 1192 } 1193 } else { /* VM_BESTFIT */ 1194 /* 1195 * we assume that, for space efficiency, it's better to 1196 * allocate from a smaller block. thus we will start searching 1197 * from the lower-order list than VM_INSTANTFIT. 1198 * however, don't bother to find the smallest block in a free 1199 * list because the list can be very long. we can revisit it 1200 * if/when it turns out to be a problem. 1201 * 1202 * note that the 'first' list can contain blocks smaller than 1203 * the requested size. thus we need to check bt_size. 1204 */ 1205 for (list = first; list < end; list++) { 1206 LIST_FOREACH(bt, list, bt_freelist) { 1207 if (bt->bt_size >= size) { 1208 rc = vmem_fit(bt, size, align, phase, 1209 nocross, minaddr, maxaddr, &start); 1210 if (rc == 0) { 1211 goto gotit; 1212 } 1213 } 1214 } 1215 } 1216 } 1217 #if 1 1218 if (strat == VM_INSTANTFIT) { 1219 strat = VM_BESTFIT; 1220 goto retry_strat; 1221 } 1222 #endif 1223 if (align != vm->vm_quantum_mask + 1 || phase != 0 || nocross != 0) { 1224 1225 /* 1226 * XXX should try to import a region large enough to 1227 * satisfy restrictions? 1228 */ 1229 1230 goto fail; 1231 } 1232 /* XXX eeek, minaddr & maxaddr not respected */ 1233 if (vmem_import(vm, size, flags) == 0) { 1234 goto retry; 1235 } 1236 /* XXX */ 1237 1238 if ((flags & VM_SLEEP) != 0) { 1239 vmem_kick_pdaemon(); 1240 VMEM_CONDVAR_WAIT(vm); 1241 goto retry; 1242 } 1243 fail: 1244 bt_free(vm, btnew); 1245 bt_free(vm, btnew2); 1246 VMEM_UNLOCK(vm); 1247 return ENOMEM; 1248 1249 gotit: 1250 KASSERT(bt->bt_type == BT_TYPE_FREE); 1251 KASSERT(bt->bt_size >= size); 1252 bt_remfree(vm, bt); 1253 vmem_check(vm); 1254 if (bt->bt_start != start) { 1255 btnew2->bt_type = BT_TYPE_FREE; 1256 btnew2->bt_start = bt->bt_start; 1257 btnew2->bt_size = start - bt->bt_start; 1258 bt->bt_start = start; 1259 bt->bt_size -= btnew2->bt_size; 1260 bt_insfree(vm, btnew2); 1261 bt_insseg(vm, btnew2, TAILQ_PREV(bt, vmem_seglist, bt_seglist)); 1262 btnew2 = NULL; 1263 vmem_check(vm); 1264 } 1265 KASSERT(bt->bt_start == start); 1266 if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) { 1267 /* split */ 1268 btnew->bt_type = BT_TYPE_BUSY; 1269 btnew->bt_start = bt->bt_start; 1270 btnew->bt_size = size; 1271 bt->bt_start = bt->bt_start + size; 1272 bt->bt_size -= size; 1273 bt_insfree(vm, bt); 1274 bt_insseg(vm, btnew, TAILQ_PREV(bt, vmem_seglist, bt_seglist)); 1275 bt_insbusy(vm, btnew); 1276 vmem_check(vm); 1277 } else { 1278 bt->bt_type = BT_TYPE_BUSY; 1279 bt_insbusy(vm, bt); 1280 vmem_check(vm); 1281 bt_free(vm, btnew); 1282 btnew = bt; 1283 } 1284 if (btnew2 != NULL) { 1285 bt_free(vm, btnew2); 1286 } 1287 KASSERT(btnew->bt_size >= size); 1288 btnew->bt_type = BT_TYPE_BUSY; 1289 if (addrp != NULL) 1290 *addrp = btnew->bt_start; 1291 VMEM_UNLOCK(vm); 1292 KASSERTMSG(addrp == NULL || 1293 (*addrp & vm->vm_quantum_mask) == 0, 1294 "vmem %s mask=0x%jx addr=0x%jx", 1295 vm->vm_name, (uintmax_t)vm->vm_quantum_mask, (uintmax_t)*addrp); 1296 return 0; 1297 } 1298 1299 /* 1300 * vmem_free: free the resource to the arena. 1301 */ 1302 1303 void 1304 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size) 1305 { 1306 1307 KASSERT(size > 0); 1308 KASSERTMSG((addr & vm->vm_quantum_mask) == 0, 1309 "vmem %s mask=0x%jx addr=0x%jx", 1310 vm->vm_name, (uintmax_t)vm->vm_quantum_mask, (uintmax_t)addr); 1311 1312 #if defined(QCACHE) 1313 if (size <= vm->vm_qcache_max) { 1314 int qidx = (size + vm->vm_quantum_mask) >> vm->vm_quantum_shift; 1315 qcache_t *qc = vm->vm_qcache[qidx - 1]; 1316 1317 pool_cache_put(qc->qc_cache, (void *)addr); 1318 return; 1319 } 1320 #endif /* defined(QCACHE) */ 1321 1322 vmem_xfree(vm, addr, size); 1323 } 1324 1325 void 1326 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size) 1327 { 1328 bt_t *bt; 1329 1330 KASSERT(size > 0); 1331 KASSERTMSG((addr & vm->vm_quantum_mask) == 0, 1332 "vmem %s mask=0x%jx addr=0x%jx", 1333 vm->vm_name, (uintmax_t)vm->vm_quantum_mask, (uintmax_t)addr); 1334 1335 VMEM_LOCK(vm); 1336 1337 bt = bt_lookupbusy(vm, addr); 1338 KASSERTMSG(bt != NULL, "vmem %s addr 0x%jx size 0x%jx", 1339 vm->vm_name, (uintmax_t)addr, (uintmax_t)size); 1340 KASSERT(bt->bt_start == addr); 1341 KASSERT(bt->bt_size == vmem_roundup_size(vm, size) || 1342 bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask); 1343 1344 /* vmem_xfree_bt() drops the lock. */ 1345 vmem_xfree_bt(vm, bt); 1346 } 1347 1348 void 1349 vmem_xfreeall(vmem_t *vm) 1350 { 1351 bt_t *bt; 1352 1353 /* This can't be used if the arena has a quantum cache. */ 1354 KASSERT(vm->vm_qcache_max == 0); 1355 1356 for (;;) { 1357 VMEM_LOCK(vm); 1358 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1359 if (bt->bt_type == BT_TYPE_BUSY) 1360 break; 1361 } 1362 if (bt != NULL) { 1363 /* vmem_xfree_bt() drops the lock. */ 1364 vmem_xfree_bt(vm, bt); 1365 } else { 1366 VMEM_UNLOCK(vm); 1367 return; 1368 } 1369 } 1370 } 1371 1372 static void 1373 vmem_xfree_bt(vmem_t *vm, bt_t *bt) 1374 { 1375 bt_t *t; 1376 1377 VMEM_ASSERT_LOCKED(vm); 1378 1379 KASSERT(bt->bt_type == BT_TYPE_BUSY); 1380 bt_rembusy(vm, bt); 1381 bt->bt_type = BT_TYPE_FREE; 1382 1383 /* coalesce */ 1384 t = TAILQ_NEXT(bt, bt_seglist); 1385 if (t != NULL && t->bt_type == BT_TYPE_FREE) { 1386 KASSERT(BT_END(bt) < t->bt_start); /* YYY */ 1387 bt_remfree(vm, t); 1388 bt_remseg(vm, t); 1389 bt->bt_size += t->bt_size; 1390 bt_free(vm, t); 1391 } 1392 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist); 1393 if (t != NULL && t->bt_type == BT_TYPE_FREE) { 1394 KASSERT(BT_END(t) < bt->bt_start); /* YYY */ 1395 bt_remfree(vm, t); 1396 bt_remseg(vm, t); 1397 bt->bt_size += t->bt_size; 1398 bt->bt_start = t->bt_start; 1399 bt_free(vm, t); 1400 } 1401 1402 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist); 1403 KASSERT(t != NULL); 1404 KASSERT(BT_ISSPAN_P(t) || t->bt_type == BT_TYPE_BUSY); 1405 if (vm->vm_releasefn != NULL && t->bt_type == BT_TYPE_SPAN && 1406 t->bt_size == bt->bt_size) { 1407 vmem_addr_t spanaddr; 1408 vmem_size_t spansize; 1409 1410 KASSERT(t->bt_start == bt->bt_start); 1411 spanaddr = bt->bt_start; 1412 spansize = bt->bt_size; 1413 bt_remseg(vm, bt); 1414 bt_free(vm, bt); 1415 bt_remseg(vm, t); 1416 bt_free(vm, t); 1417 vm->vm_size -= spansize; 1418 VMEM_CONDVAR_BROADCAST(vm); 1419 /* bt_freetrim() drops the lock. */ 1420 bt_freetrim(vm, BT_MAXFREE); 1421 (*vm->vm_releasefn)(vm->vm_arg, spanaddr, spansize); 1422 } else { 1423 bt_insfree(vm, bt); 1424 VMEM_CONDVAR_BROADCAST(vm); 1425 /* bt_freetrim() drops the lock. */ 1426 bt_freetrim(vm, BT_MAXFREE); 1427 } 1428 } 1429 1430 /* 1431 * vmem_add: 1432 * 1433 * => caller must ensure appropriate spl, 1434 * if the arena can be accessed from interrupt context. 1435 */ 1436 1437 int 1438 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, vm_flag_t flags) 1439 { 1440 int rv; 1441 1442 VMEM_LOCK(vm); 1443 rv = vmem_add1(vm, addr, size, flags, BT_TYPE_SPAN_STATIC); 1444 VMEM_UNLOCK(vm); 1445 1446 return rv; 1447 } 1448 1449 /* 1450 * vmem_size: information about arenas size 1451 * 1452 * => return free/allocated size in arena 1453 */ 1454 vmem_size_t 1455 vmem_size(vmem_t *vm, int typemask) 1456 { 1457 1458 switch (typemask) { 1459 case VMEM_ALLOC: 1460 return vm->vm_inuse; 1461 case VMEM_FREE: 1462 return vm->vm_size - vm->vm_inuse; 1463 case VMEM_FREE|VMEM_ALLOC: 1464 return vm->vm_size; 1465 default: 1466 panic("vmem_size"); 1467 } 1468 } 1469 1470 /* ---- rehash */ 1471 1472 #if defined(_KERNEL) 1473 static struct callout vmem_rehash_ch; 1474 static int vmem_rehash_interval; 1475 static struct workqueue *vmem_rehash_wq; 1476 static struct work vmem_rehash_wk; 1477 1478 static void 1479 vmem_rehash_all(struct work *wk, void *dummy) 1480 { 1481 vmem_t *vm; 1482 1483 KASSERT(wk == &vmem_rehash_wk); 1484 mutex_enter(&vmem_list_lock); 1485 LIST_FOREACH(vm, &vmem_list, vm_alllist) { 1486 size_t desired; 1487 size_t current; 1488 1489 desired = atomic_load_relaxed(&vm->vm_maxbusytag); 1490 current = atomic_load_relaxed(&vm->vm_hashsize); 1491 1492 if (desired > VMEM_HASHSIZE_MAX) { 1493 desired = VMEM_HASHSIZE_MAX; 1494 } else if (desired < VMEM_HASHSIZE_MIN) { 1495 desired = VMEM_HASHSIZE_MIN; 1496 } 1497 if (desired > current * 2 || desired * 2 < current) { 1498 vmem_rehash(vm, desired, VM_NOSLEEP); 1499 } 1500 } 1501 mutex_exit(&vmem_list_lock); 1502 1503 callout_schedule(&vmem_rehash_ch, vmem_rehash_interval); 1504 } 1505 1506 static void 1507 vmem_rehash_all_kick(void *dummy) 1508 { 1509 1510 workqueue_enqueue(vmem_rehash_wq, &vmem_rehash_wk, NULL); 1511 } 1512 1513 void 1514 vmem_rehash_start(void) 1515 { 1516 int error; 1517 1518 error = workqueue_create(&vmem_rehash_wq, "vmem_rehash", 1519 vmem_rehash_all, NULL, PRI_VM, IPL_SOFTCLOCK, WQ_MPSAFE); 1520 if (error) { 1521 panic("%s: workqueue_create %d\n", __func__, error); 1522 } 1523 callout_init(&vmem_rehash_ch, CALLOUT_MPSAFE); 1524 callout_setfunc(&vmem_rehash_ch, vmem_rehash_all_kick, NULL); 1525 1526 vmem_rehash_interval = hz * 10; 1527 callout_schedule(&vmem_rehash_ch, vmem_rehash_interval); 1528 } 1529 #endif /* defined(_KERNEL) */ 1530 1531 /* ---- debug */ 1532 1533 #if defined(DDB) || defined(UNITTEST) || defined(VMEM_SANITY) 1534 1535 static void bt_dump(const bt_t *, void (*)(const char *, ...) 1536 __printflike(1, 2)); 1537 1538 static const char * 1539 bt_type_string(int type) 1540 { 1541 static const char * const table[] = { 1542 [BT_TYPE_BUSY] = "busy", 1543 [BT_TYPE_FREE] = "free", 1544 [BT_TYPE_SPAN] = "span", 1545 [BT_TYPE_SPAN_STATIC] = "static span", 1546 }; 1547 1548 if (type >= __arraycount(table)) { 1549 return "BOGUS"; 1550 } 1551 return table[type]; 1552 } 1553 1554 static void 1555 bt_dump(const bt_t *bt, void (*pr)(const char *, ...)) 1556 { 1557 1558 (*pr)("\t%p: %" PRIu64 ", %" PRIu64 ", %d(%s)\n", 1559 bt, (uint64_t)bt->bt_start, (uint64_t)bt->bt_size, 1560 bt->bt_type, bt_type_string(bt->bt_type)); 1561 } 1562 1563 static void 1564 vmem_dump(const vmem_t *vm , void (*pr)(const char *, ...) __printflike(1, 2)) 1565 { 1566 const bt_t *bt; 1567 int i; 1568 1569 (*pr)("vmem %p '%s'\n", vm, vm->vm_name); 1570 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1571 bt_dump(bt, pr); 1572 } 1573 1574 for (i = 0; i < VMEM_MAXORDER; i++) { 1575 const struct vmem_freelist *fl = &vm->vm_freelist[i]; 1576 1577 if (LIST_EMPTY(fl)) { 1578 continue; 1579 } 1580 1581 (*pr)("freelist[%d]\n", i); 1582 LIST_FOREACH(bt, fl, bt_freelist) { 1583 bt_dump(bt, pr); 1584 } 1585 } 1586 } 1587 1588 #endif /* defined(DDB) || defined(UNITTEST) || defined(VMEM_SANITY) */ 1589 1590 #if defined(DDB) 1591 static bt_t * 1592 vmem_whatis_lookup(vmem_t *vm, uintptr_t addr) 1593 { 1594 bt_t *bt; 1595 1596 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1597 if (BT_ISSPAN_P(bt)) { 1598 continue; 1599 } 1600 if (bt->bt_start <= addr && addr <= BT_END(bt)) { 1601 return bt; 1602 } 1603 } 1604 1605 return NULL; 1606 } 1607 1608 void 1609 vmem_whatis(uintptr_t addr, void (*pr)(const char *, ...)) 1610 { 1611 vmem_t *vm; 1612 1613 LIST_FOREACH(vm, &vmem_list, vm_alllist) { 1614 bt_t *bt; 1615 1616 bt = vmem_whatis_lookup(vm, addr); 1617 if (bt == NULL) { 1618 continue; 1619 } 1620 (*pr)("%p is %p+%zu in VMEM '%s' (%s)\n", 1621 (void *)addr, (void *)bt->bt_start, 1622 (size_t)(addr - bt->bt_start), vm->vm_name, 1623 (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free"); 1624 } 1625 } 1626 1627 void 1628 vmem_printall(const char *modif, void (*pr)(const char *, ...)) 1629 { 1630 const vmem_t *vm; 1631 1632 LIST_FOREACH(vm, &vmem_list, vm_alllist) { 1633 vmem_dump(vm, pr); 1634 } 1635 } 1636 1637 void 1638 vmem_print(uintptr_t addr, const char *modif, void (*pr)(const char *, ...)) 1639 { 1640 const vmem_t *vm = (const void *)addr; 1641 1642 vmem_dump(vm, pr); 1643 } 1644 #endif /* defined(DDB) */ 1645 1646 #if defined(_KERNEL) 1647 #define vmem_printf printf 1648 #else 1649 #include <stdio.h> 1650 #include <stdarg.h> 1651 1652 static void 1653 vmem_printf(const char *fmt, ...) 1654 { 1655 va_list ap; 1656 va_start(ap, fmt); 1657 vprintf(fmt, ap); 1658 va_end(ap); 1659 } 1660 #endif 1661 1662 #if defined(VMEM_SANITY) 1663 1664 static bool 1665 vmem_check_sanity(vmem_t *vm) 1666 { 1667 const bt_t *bt, *bt2; 1668 1669 KASSERT(vm != NULL); 1670 1671 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1672 if (bt->bt_start > BT_END(bt)) { 1673 printf("corrupted tag\n"); 1674 bt_dump(bt, vmem_printf); 1675 return false; 1676 } 1677 } 1678 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1679 TAILQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) { 1680 if (bt == bt2) { 1681 continue; 1682 } 1683 if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) { 1684 continue; 1685 } 1686 if (bt->bt_start <= BT_END(bt2) && 1687 bt2->bt_start <= BT_END(bt)) { 1688 printf("overwrapped tags\n"); 1689 bt_dump(bt, vmem_printf); 1690 bt_dump(bt2, vmem_printf); 1691 return false; 1692 } 1693 } 1694 } 1695 1696 return true; 1697 } 1698 1699 static void 1700 vmem_check(vmem_t *vm) 1701 { 1702 1703 if (!vmem_check_sanity(vm)) { 1704 panic("insanity vmem %p", vm); 1705 } 1706 } 1707 1708 #endif /* defined(VMEM_SANITY) */ 1709 1710 #if defined(UNITTEST) 1711 int 1712 main(void) 1713 { 1714 int rc; 1715 vmem_t *vm; 1716 vmem_addr_t p; 1717 struct reg { 1718 vmem_addr_t p; 1719 vmem_size_t sz; 1720 bool x; 1721 } *reg = NULL; 1722 int nreg = 0; 1723 int nalloc = 0; 1724 int nfree = 0; 1725 vmem_size_t total = 0; 1726 #if 1 1727 vm_flag_t strat = VM_INSTANTFIT; 1728 #else 1729 vm_flag_t strat = VM_BESTFIT; 1730 #endif 1731 1732 vm = vmem_create("test", 0, 0, 1, NULL, NULL, NULL, 0, VM_SLEEP, 1733 #ifdef _KERNEL 1734 IPL_NONE 1735 #else 1736 0 1737 #endif 1738 ); 1739 if (vm == NULL) { 1740 printf("vmem_create\n"); 1741 exit(EXIT_FAILURE); 1742 } 1743 vmem_dump(vm, vmem_printf); 1744 1745 rc = vmem_add(vm, 0, 50, VM_SLEEP); 1746 assert(rc == 0); 1747 rc = vmem_add(vm, 100, 200, VM_SLEEP); 1748 assert(rc == 0); 1749 rc = vmem_add(vm, 2000, 1, VM_SLEEP); 1750 assert(rc == 0); 1751 rc = vmem_add(vm, 40000, 65536, VM_SLEEP); 1752 assert(rc == 0); 1753 rc = vmem_add(vm, 10000, 10000, VM_SLEEP); 1754 assert(rc == 0); 1755 rc = vmem_add(vm, 500, 1000, VM_SLEEP); 1756 assert(rc == 0); 1757 rc = vmem_add(vm, 0xffffff00, 0x100, VM_SLEEP); 1758 assert(rc == 0); 1759 rc = vmem_xalloc(vm, 0x101, 0, 0, 0, 1760 0xffffff00, 0xffffffff, strat|VM_SLEEP, &p); 1761 assert(rc != 0); 1762 rc = vmem_xalloc(vm, 50, 0, 0, 0, 0, 49, strat|VM_SLEEP, &p); 1763 assert(rc == 0 && p == 0); 1764 vmem_xfree(vm, p, 50); 1765 rc = vmem_xalloc(vm, 25, 0, 0, 0, 0, 24, strat|VM_SLEEP, &p); 1766 assert(rc == 0 && p == 0); 1767 rc = vmem_xalloc(vm, 0x100, 0, 0, 0, 1768 0xffffff01, 0xffffffff, strat|VM_SLEEP, &p); 1769 assert(rc != 0); 1770 rc = vmem_xalloc(vm, 0x100, 0, 0, 0, 1771 0xffffff00, 0xfffffffe, strat|VM_SLEEP, &p); 1772 assert(rc != 0); 1773 rc = vmem_xalloc(vm, 0x100, 0, 0, 0, 1774 0xffffff00, 0xffffffff, strat|VM_SLEEP, &p); 1775 assert(rc == 0); 1776 vmem_dump(vm, vmem_printf); 1777 for (;;) { 1778 struct reg *r; 1779 int t = rand() % 100; 1780 1781 if (t > 45) { 1782 /* alloc */ 1783 vmem_size_t sz = rand() % 500 + 1; 1784 bool x; 1785 vmem_size_t align, phase, nocross; 1786 vmem_addr_t minaddr, maxaddr; 1787 1788 if (t > 70) { 1789 x = true; 1790 /* XXX */ 1791 align = 1 << (rand() % 15); 1792 phase = rand() % 65536; 1793 nocross = 1 << (rand() % 15); 1794 if (align <= phase) { 1795 phase = 0; 1796 } 1797 if (VMEM_CROSS_P(phase, phase + sz - 1, 1798 nocross)) { 1799 nocross = 0; 1800 } 1801 do { 1802 minaddr = rand() % 50000; 1803 maxaddr = rand() % 70000; 1804 } while (minaddr > maxaddr); 1805 printf("=== xalloc %" PRIu64 1806 " align=%" PRIu64 ", phase=%" PRIu64 1807 ", nocross=%" PRIu64 ", min=%" PRIu64 1808 ", max=%" PRIu64 "\n", 1809 (uint64_t)sz, 1810 (uint64_t)align, 1811 (uint64_t)phase, 1812 (uint64_t)nocross, 1813 (uint64_t)minaddr, 1814 (uint64_t)maxaddr); 1815 rc = vmem_xalloc(vm, sz, align, phase, nocross, 1816 minaddr, maxaddr, strat|VM_SLEEP, &p); 1817 } else { 1818 x = false; 1819 printf("=== alloc %" PRIu64 "\n", (uint64_t)sz); 1820 rc = vmem_alloc(vm, sz, strat|VM_SLEEP, &p); 1821 } 1822 printf("-> %" PRIu64 "\n", (uint64_t)p); 1823 vmem_dump(vm, vmem_printf); 1824 if (rc != 0) { 1825 if (x) { 1826 continue; 1827 } 1828 break; 1829 } 1830 nreg++; 1831 reg = realloc(reg, sizeof(*reg) * nreg); 1832 r = ®[nreg - 1]; 1833 r->p = p; 1834 r->sz = sz; 1835 r->x = x; 1836 total += sz; 1837 nalloc++; 1838 } else if (nreg != 0) { 1839 /* free */ 1840 r = ®[rand() % nreg]; 1841 printf("=== free %" PRIu64 ", %" PRIu64 "\n", 1842 (uint64_t)r->p, (uint64_t)r->sz); 1843 if (r->x) { 1844 vmem_xfree(vm, r->p, r->sz); 1845 } else { 1846 vmem_free(vm, r->p, r->sz); 1847 } 1848 total -= r->sz; 1849 vmem_dump(vm, vmem_printf); 1850 *r = reg[nreg - 1]; 1851 nreg--; 1852 nfree++; 1853 } 1854 printf("total=%" PRIu64 "\n", (uint64_t)total); 1855 } 1856 fprintf(stderr, "total=%" PRIu64 ", nalloc=%d, nfree=%d\n", 1857 (uint64_t)total, nalloc, nfree); 1858 exit(EXIT_SUCCESS); 1859 } 1860 #endif /* defined(UNITTEST) */ 1861