1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c)2006,2007,2008,2009 YAMAMOTO Takashi, 5 * Copyright (c) 2013 EMC Corp. 6 * All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following 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 AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30 /* 31 * From: 32 * $NetBSD: vmem_impl.h,v 1.2 2013/01/29 21:26:24 para Exp $ 33 * $NetBSD: subr_vmem.c,v 1.83 2013/03/06 11:20:10 yamt Exp $ 34 */ 35 36 /* 37 * reference: 38 * - Magazines and Vmem: Extending the Slab Allocator 39 * to Many CPUs and Arbitrary Resources 40 * http://www.usenix.org/event/usenix01/bonwick.html 41 */ 42 43 #include <sys/cdefs.h> 44 __FBSDID("$FreeBSD$"); 45 46 #include "opt_ddb.h" 47 48 #include <sys/param.h> 49 #include <sys/systm.h> 50 #include <sys/kernel.h> 51 #include <sys/queue.h> 52 #include <sys/callout.h> 53 #include <sys/hash.h> 54 #include <sys/lock.h> 55 #include <sys/malloc.h> 56 #include <sys/mutex.h> 57 #include <sys/smp.h> 58 #include <sys/condvar.h> 59 #include <sys/sysctl.h> 60 #include <sys/taskqueue.h> 61 #include <sys/vmem.h> 62 63 #include "opt_vm.h" 64 65 #include <vm/uma.h> 66 #include <vm/vm.h> 67 #include <vm/pmap.h> 68 #include <vm/vm_map.h> 69 #include <vm/vm_object.h> 70 #include <vm/vm_kern.h> 71 #include <vm/vm_extern.h> 72 #include <vm/vm_param.h> 73 #include <vm/vm_pageout.h> 74 75 #define VMEM_OPTORDER 5 76 #define VMEM_OPTVALUE (1 << VMEM_OPTORDER) 77 #define VMEM_MAXORDER \ 78 (VMEM_OPTVALUE - 1 + sizeof(vmem_size_t) * NBBY - VMEM_OPTORDER) 79 80 #define VMEM_HASHSIZE_MIN 16 81 #define VMEM_HASHSIZE_MAX 131072 82 83 #define VMEM_QCACHE_IDX_MAX 16 84 85 #define VMEM_FITMASK (M_BESTFIT | M_FIRSTFIT) 86 87 #define VMEM_FLAGS \ 88 (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM | M_BESTFIT | M_FIRSTFIT) 89 90 #define BT_FLAGS (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM) 91 92 #define QC_NAME_MAX 16 93 94 /* 95 * Data structures private to vmem. 96 */ 97 MALLOC_DEFINE(M_VMEM, "vmem", "vmem internal structures"); 98 99 typedef struct vmem_btag bt_t; 100 101 TAILQ_HEAD(vmem_seglist, vmem_btag); 102 LIST_HEAD(vmem_freelist, vmem_btag); 103 LIST_HEAD(vmem_hashlist, vmem_btag); 104 105 struct qcache { 106 uma_zone_t qc_cache; 107 vmem_t *qc_vmem; 108 vmem_size_t qc_size; 109 char qc_name[QC_NAME_MAX]; 110 }; 111 typedef struct qcache qcache_t; 112 #define QC_POOL_TO_QCACHE(pool) ((qcache_t *)(pool->pr_qcache)) 113 114 #define VMEM_NAME_MAX 16 115 116 /* vmem arena */ 117 struct vmem { 118 struct mtx_padalign vm_lock; 119 struct cv vm_cv; 120 char vm_name[VMEM_NAME_MAX+1]; 121 LIST_ENTRY(vmem) vm_alllist; 122 struct vmem_hashlist vm_hash0[VMEM_HASHSIZE_MIN]; 123 struct vmem_freelist vm_freelist[VMEM_MAXORDER]; 124 struct vmem_seglist vm_seglist; 125 struct vmem_hashlist *vm_hashlist; 126 vmem_size_t vm_hashsize; 127 128 /* Constant after init */ 129 vmem_size_t vm_qcache_max; 130 vmem_size_t vm_quantum_mask; 131 vmem_size_t vm_import_quantum; 132 int vm_quantum_shift; 133 134 /* Written on alloc/free */ 135 LIST_HEAD(, vmem_btag) vm_freetags; 136 int vm_nfreetags; 137 int vm_nbusytag; 138 vmem_size_t vm_inuse; 139 vmem_size_t vm_size; 140 141 /* Used on import. */ 142 vmem_import_t *vm_importfn; 143 vmem_release_t *vm_releasefn; 144 void *vm_arg; 145 146 /* Space exhaustion callback. */ 147 vmem_reclaim_t *vm_reclaimfn; 148 149 /* quantum cache */ 150 qcache_t vm_qcache[VMEM_QCACHE_IDX_MAX]; 151 }; 152 153 /* boundary tag */ 154 struct vmem_btag { 155 TAILQ_ENTRY(vmem_btag) bt_seglist; 156 union { 157 LIST_ENTRY(vmem_btag) u_freelist; /* BT_TYPE_FREE */ 158 LIST_ENTRY(vmem_btag) u_hashlist; /* BT_TYPE_BUSY */ 159 } bt_u; 160 #define bt_hashlist bt_u.u_hashlist 161 #define bt_freelist bt_u.u_freelist 162 vmem_addr_t bt_start; 163 vmem_size_t bt_size; 164 int bt_type; 165 }; 166 167 #define BT_TYPE_SPAN 1 /* Allocated from importfn */ 168 #define BT_TYPE_SPAN_STATIC 2 /* vmem_add() or create. */ 169 #define BT_TYPE_FREE 3 /* Available space. */ 170 #define BT_TYPE_BUSY 4 /* Used space. */ 171 #define BT_ISSPAN_P(bt) ((bt)->bt_type <= BT_TYPE_SPAN_STATIC) 172 173 #define BT_END(bt) ((bt)->bt_start + (bt)->bt_size - 1) 174 175 #if defined(DIAGNOSTIC) 176 static int enable_vmem_check = 1; 177 SYSCTL_INT(_debug, OID_AUTO, vmem_check, CTLFLAG_RWTUN, 178 &enable_vmem_check, 0, "Enable vmem check"); 179 static void vmem_check(vmem_t *); 180 #endif 181 182 static struct callout vmem_periodic_ch; 183 static int vmem_periodic_interval; 184 static struct task vmem_periodic_wk; 185 186 static struct mtx_padalign __exclusive_cache_line vmem_list_lock; 187 static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list); 188 189 /* ---- misc */ 190 #define VMEM_CONDVAR_INIT(vm, wchan) cv_init(&vm->vm_cv, wchan) 191 #define VMEM_CONDVAR_DESTROY(vm) cv_destroy(&vm->vm_cv) 192 #define VMEM_CONDVAR_WAIT(vm) cv_wait(&vm->vm_cv, &vm->vm_lock) 193 #define VMEM_CONDVAR_BROADCAST(vm) cv_broadcast(&vm->vm_cv) 194 195 196 #define VMEM_LOCK(vm) mtx_lock(&vm->vm_lock) 197 #define VMEM_TRYLOCK(vm) mtx_trylock(&vm->vm_lock) 198 #define VMEM_UNLOCK(vm) mtx_unlock(&vm->vm_lock) 199 #define VMEM_LOCK_INIT(vm, name) mtx_init(&vm->vm_lock, (name), NULL, MTX_DEF) 200 #define VMEM_LOCK_DESTROY(vm) mtx_destroy(&vm->vm_lock) 201 #define VMEM_ASSERT_LOCKED(vm) mtx_assert(&vm->vm_lock, MA_OWNED); 202 203 #define VMEM_ALIGNUP(addr, align) (-(-(addr) & -(align))) 204 205 #define VMEM_CROSS_P(addr1, addr2, boundary) \ 206 ((((addr1) ^ (addr2)) & -(boundary)) != 0) 207 208 #define ORDER2SIZE(order) ((order) < VMEM_OPTVALUE ? ((order) + 1) : \ 209 (vmem_size_t)1 << ((order) - (VMEM_OPTVALUE - VMEM_OPTORDER - 1))) 210 #define SIZE2ORDER(size) ((size) <= VMEM_OPTVALUE ? ((size) - 1) : \ 211 (flsl(size) + (VMEM_OPTVALUE - VMEM_OPTORDER - 2))) 212 213 /* 214 * Maximum number of boundary tags that may be required to satisfy an 215 * allocation. Two may be required to import. Another two may be 216 * required to clip edges. 217 */ 218 #define BT_MAXALLOC 4 219 220 /* 221 * Max free limits the number of locally cached boundary tags. We 222 * just want to avoid hitting the zone allocator for every call. 223 */ 224 #define BT_MAXFREE (BT_MAXALLOC * 8) 225 226 /* Allocator for boundary tags. */ 227 static uma_zone_t vmem_bt_zone; 228 229 /* boot time arena storage. */ 230 static struct vmem kernel_arena_storage; 231 static struct vmem kmem_arena_storage; 232 static struct vmem buffer_arena_storage; 233 static struct vmem transient_arena_storage; 234 vmem_t *kernel_arena = &kernel_arena_storage; 235 vmem_t *kmem_arena = &kmem_arena_storage; 236 vmem_t *buffer_arena = &buffer_arena_storage; 237 vmem_t *transient_arena = &transient_arena_storage; 238 239 #ifdef DEBUG_MEMGUARD 240 static struct vmem memguard_arena_storage; 241 vmem_t *memguard_arena = &memguard_arena_storage; 242 #endif 243 244 /* 245 * Fill the vmem's boundary tag cache. We guarantee that boundary tag 246 * allocation will not fail once bt_fill() passes. To do so we cache 247 * at least the maximum possible tag allocations in the arena. 248 */ 249 static int 250 bt_fill(vmem_t *vm, int flags) 251 { 252 bt_t *bt; 253 254 VMEM_ASSERT_LOCKED(vm); 255 256 /* 257 * Only allow the kmem arena to dip into reserve tags. It is the 258 * vmem where new tags come from. 259 */ 260 flags &= BT_FLAGS; 261 if (vm != kmem_arena) 262 flags &= ~M_USE_RESERVE; 263 264 /* 265 * Loop until we meet the reserve. To minimize the lock shuffle 266 * and prevent simultaneous fills we first try a NOWAIT regardless 267 * of the caller's flags. Specify M_NOVM so we don't recurse while 268 * holding a vmem lock. 269 */ 270 while (vm->vm_nfreetags < BT_MAXALLOC) { 271 bt = uma_zalloc(vmem_bt_zone, 272 (flags & M_USE_RESERVE) | M_NOWAIT | M_NOVM); 273 if (bt == NULL) { 274 VMEM_UNLOCK(vm); 275 bt = uma_zalloc(vmem_bt_zone, flags); 276 VMEM_LOCK(vm); 277 if (bt == NULL && (flags & M_NOWAIT) != 0) 278 break; 279 } 280 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist); 281 vm->vm_nfreetags++; 282 } 283 284 if (vm->vm_nfreetags < BT_MAXALLOC) 285 return ENOMEM; 286 287 return 0; 288 } 289 290 /* 291 * Pop a tag off of the freetag stack. 292 */ 293 static bt_t * 294 bt_alloc(vmem_t *vm) 295 { 296 bt_t *bt; 297 298 VMEM_ASSERT_LOCKED(vm); 299 bt = LIST_FIRST(&vm->vm_freetags); 300 MPASS(bt != NULL); 301 LIST_REMOVE(bt, bt_freelist); 302 vm->vm_nfreetags--; 303 304 return bt; 305 } 306 307 /* 308 * Trim the per-vmem free list. Returns with the lock released to 309 * avoid allocator recursions. 310 */ 311 static void 312 bt_freetrim(vmem_t *vm, int freelimit) 313 { 314 LIST_HEAD(, vmem_btag) freetags; 315 bt_t *bt; 316 317 LIST_INIT(&freetags); 318 VMEM_ASSERT_LOCKED(vm); 319 while (vm->vm_nfreetags > freelimit) { 320 bt = LIST_FIRST(&vm->vm_freetags); 321 LIST_REMOVE(bt, bt_freelist); 322 vm->vm_nfreetags--; 323 LIST_INSERT_HEAD(&freetags, bt, bt_freelist); 324 } 325 VMEM_UNLOCK(vm); 326 while ((bt = LIST_FIRST(&freetags)) != NULL) { 327 LIST_REMOVE(bt, bt_freelist); 328 uma_zfree(vmem_bt_zone, bt); 329 } 330 } 331 332 static inline void 333 bt_free(vmem_t *vm, bt_t *bt) 334 { 335 336 VMEM_ASSERT_LOCKED(vm); 337 MPASS(LIST_FIRST(&vm->vm_freetags) != bt); 338 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist); 339 vm->vm_nfreetags++; 340 } 341 342 /* 343 * freelist[0] ... [1, 1] 344 * freelist[1] ... [2, 2] 345 * : 346 * freelist[29] ... [30, 30] 347 * freelist[30] ... [31, 31] 348 * freelist[31] ... [32, 63] 349 * freelist[33] ... [64, 127] 350 * : 351 * freelist[n] ... [(1 << (n - 26)), (1 << (n - 25)) - 1] 352 * : 353 */ 354 355 static struct vmem_freelist * 356 bt_freehead_tofree(vmem_t *vm, vmem_size_t size) 357 { 358 const vmem_size_t qsize = size >> vm->vm_quantum_shift; 359 const int idx = SIZE2ORDER(qsize); 360 361 MPASS(size != 0 && qsize != 0); 362 MPASS((size & vm->vm_quantum_mask) == 0); 363 MPASS(idx >= 0); 364 MPASS(idx < VMEM_MAXORDER); 365 366 return &vm->vm_freelist[idx]; 367 } 368 369 /* 370 * bt_freehead_toalloc: return the freelist for the given size and allocation 371 * strategy. 372 * 373 * For M_FIRSTFIT, return the list in which any blocks are large enough 374 * for the requested size. otherwise, return the list which can have blocks 375 * large enough for the requested size. 376 */ 377 static struct vmem_freelist * 378 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, int strat) 379 { 380 const vmem_size_t qsize = size >> vm->vm_quantum_shift; 381 int idx = SIZE2ORDER(qsize); 382 383 MPASS(size != 0 && qsize != 0); 384 MPASS((size & vm->vm_quantum_mask) == 0); 385 386 if (strat == M_FIRSTFIT && ORDER2SIZE(idx) != qsize) { 387 idx++; 388 /* check too large request? */ 389 } 390 MPASS(idx >= 0); 391 MPASS(idx < VMEM_MAXORDER); 392 393 return &vm->vm_freelist[idx]; 394 } 395 396 /* ---- boundary tag hash */ 397 398 static struct vmem_hashlist * 399 bt_hashhead(vmem_t *vm, vmem_addr_t addr) 400 { 401 struct vmem_hashlist *list; 402 unsigned int hash; 403 404 hash = hash32_buf(&addr, sizeof(addr), 0); 405 list = &vm->vm_hashlist[hash % vm->vm_hashsize]; 406 407 return list; 408 } 409 410 static bt_t * 411 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr) 412 { 413 struct vmem_hashlist *list; 414 bt_t *bt; 415 416 VMEM_ASSERT_LOCKED(vm); 417 list = bt_hashhead(vm, addr); 418 LIST_FOREACH(bt, list, bt_hashlist) { 419 if (bt->bt_start == addr) { 420 break; 421 } 422 } 423 424 return bt; 425 } 426 427 static void 428 bt_rembusy(vmem_t *vm, bt_t *bt) 429 { 430 431 VMEM_ASSERT_LOCKED(vm); 432 MPASS(vm->vm_nbusytag > 0); 433 vm->vm_inuse -= bt->bt_size; 434 vm->vm_nbusytag--; 435 LIST_REMOVE(bt, bt_hashlist); 436 } 437 438 static void 439 bt_insbusy(vmem_t *vm, bt_t *bt) 440 { 441 struct vmem_hashlist *list; 442 443 VMEM_ASSERT_LOCKED(vm); 444 MPASS(bt->bt_type == BT_TYPE_BUSY); 445 446 list = bt_hashhead(vm, bt->bt_start); 447 LIST_INSERT_HEAD(list, bt, bt_hashlist); 448 vm->vm_nbusytag++; 449 vm->vm_inuse += bt->bt_size; 450 } 451 452 /* ---- boundary tag list */ 453 454 static void 455 bt_remseg(vmem_t *vm, bt_t *bt) 456 { 457 458 TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist); 459 bt_free(vm, bt); 460 } 461 462 static void 463 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev) 464 { 465 466 TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist); 467 } 468 469 static void 470 bt_insseg_tail(vmem_t *vm, bt_t *bt) 471 { 472 473 TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist); 474 } 475 476 static void 477 bt_remfree(vmem_t *vm, bt_t *bt) 478 { 479 480 MPASS(bt->bt_type == BT_TYPE_FREE); 481 482 LIST_REMOVE(bt, bt_freelist); 483 } 484 485 static void 486 bt_insfree(vmem_t *vm, bt_t *bt) 487 { 488 struct vmem_freelist *list; 489 490 list = bt_freehead_tofree(vm, bt->bt_size); 491 LIST_INSERT_HEAD(list, bt, bt_freelist); 492 } 493 494 /* ---- vmem internal functions */ 495 496 /* 497 * Import from the arena into the quantum cache in UMA. 498 */ 499 static int 500 qc_import(void *arg, void **store, int cnt, int flags) 501 { 502 qcache_t *qc; 503 vmem_addr_t addr; 504 int i; 505 506 qc = arg; 507 if ((flags & VMEM_FITMASK) == 0) 508 flags |= M_BESTFIT; 509 for (i = 0; i < cnt; i++) { 510 if (vmem_xalloc(qc->qc_vmem, qc->qc_size, 0, 0, 0, 511 VMEM_ADDR_MIN, VMEM_ADDR_MAX, flags, &addr) != 0) 512 break; 513 store[i] = (void *)addr; 514 /* Only guarantee one allocation. */ 515 flags &= ~M_WAITOK; 516 flags |= M_NOWAIT; 517 } 518 return i; 519 } 520 521 /* 522 * Release memory from the UMA cache to the arena. 523 */ 524 static void 525 qc_release(void *arg, void **store, int cnt) 526 { 527 qcache_t *qc; 528 int i; 529 530 qc = arg; 531 for (i = 0; i < cnt; i++) 532 vmem_xfree(qc->qc_vmem, (vmem_addr_t)store[i], qc->qc_size); 533 } 534 535 static void 536 qc_init(vmem_t *vm, vmem_size_t qcache_max) 537 { 538 qcache_t *qc; 539 vmem_size_t size; 540 int qcache_idx_max; 541 int i; 542 543 MPASS((qcache_max & vm->vm_quantum_mask) == 0); 544 qcache_idx_max = MIN(qcache_max >> vm->vm_quantum_shift, 545 VMEM_QCACHE_IDX_MAX); 546 vm->vm_qcache_max = qcache_idx_max << vm->vm_quantum_shift; 547 for (i = 0; i < qcache_idx_max; i++) { 548 qc = &vm->vm_qcache[i]; 549 size = (i + 1) << vm->vm_quantum_shift; 550 snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu", 551 vm->vm_name, size); 552 qc->qc_vmem = vm; 553 qc->qc_size = size; 554 qc->qc_cache = uma_zcache_create(qc->qc_name, size, 555 NULL, NULL, NULL, NULL, qc_import, qc_release, qc, 556 UMA_ZONE_VM); 557 MPASS(qc->qc_cache); 558 } 559 } 560 561 static void 562 qc_destroy(vmem_t *vm) 563 { 564 int qcache_idx_max; 565 int i; 566 567 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift; 568 for (i = 0; i < qcache_idx_max; i++) 569 uma_zdestroy(vm->vm_qcache[i].qc_cache); 570 } 571 572 static void 573 qc_drain(vmem_t *vm) 574 { 575 int qcache_idx_max; 576 int i; 577 578 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift; 579 for (i = 0; i < qcache_idx_max; i++) 580 zone_drain(vm->vm_qcache[i].qc_cache); 581 } 582 583 #ifndef UMA_MD_SMALL_ALLOC 584 585 static struct mtx_padalign __exclusive_cache_line vmem_bt_lock; 586 587 /* 588 * vmem_bt_alloc: Allocate a new page of boundary tags. 589 * 590 * On architectures with uma_small_alloc there is no recursion; no address 591 * space need be allocated to allocate boundary tags. For the others, we 592 * must handle recursion. Boundary tags are necessary to allocate new 593 * boundary tags. 594 * 595 * UMA guarantees that enough tags are held in reserve to allocate a new 596 * page of kva. We dip into this reserve by specifying M_USE_RESERVE only 597 * when allocating the page to hold new boundary tags. In this way the 598 * reserve is automatically filled by the allocation that uses the reserve. 599 * 600 * We still have to guarantee that the new tags are allocated atomically since 601 * many threads may try concurrently. The bt_lock provides this guarantee. 602 * We convert WAITOK allocations to NOWAIT and then handle the blocking here 603 * on failure. It's ok to return NULL for a WAITOK allocation as UMA will 604 * loop again after checking to see if we lost the race to allocate. 605 * 606 * There is a small race between vmem_bt_alloc() returning the page and the 607 * zone lock being acquired to add the page to the zone. For WAITOK 608 * allocations we just pause briefly. NOWAIT may experience a transient 609 * failure. To alleviate this we permit a small number of simultaneous 610 * fills to proceed concurrently so NOWAIT is less likely to fail unless 611 * we are really out of KVA. 612 */ 613 static void * 614 vmem_bt_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) 615 { 616 vmem_addr_t addr; 617 618 *pflag = UMA_SLAB_KMEM; 619 620 /* 621 * Single thread boundary tag allocation so that the address space 622 * and memory are added in one atomic operation. 623 */ 624 mtx_lock(&vmem_bt_lock); 625 if (vmem_xalloc(kmem_arena, bytes, 0, 0, 0, VMEM_ADDR_MIN, 626 VMEM_ADDR_MAX, M_NOWAIT | M_NOVM | M_USE_RESERVE | M_BESTFIT, 627 &addr) == 0) { 628 if (kmem_back(kmem_object, addr, bytes, 629 M_NOWAIT | M_USE_RESERVE) == 0) { 630 mtx_unlock(&vmem_bt_lock); 631 return ((void *)addr); 632 } 633 vmem_xfree(kmem_arena, addr, bytes); 634 mtx_unlock(&vmem_bt_lock); 635 /* 636 * Out of memory, not address space. This may not even be 637 * possible due to M_USE_RESERVE page allocation. 638 */ 639 if (wait & M_WAITOK) 640 VM_WAIT; 641 return (NULL); 642 } 643 mtx_unlock(&vmem_bt_lock); 644 /* 645 * We're either out of address space or lost a fill race. 646 */ 647 if (wait & M_WAITOK) 648 pause("btalloc", 1); 649 650 return (NULL); 651 } 652 #endif 653 654 void 655 vmem_startup(void) 656 { 657 658 mtx_init(&vmem_list_lock, "vmem list lock", NULL, MTX_DEF); 659 vmem_bt_zone = uma_zcreate("vmem btag", 660 sizeof(struct vmem_btag), NULL, NULL, NULL, NULL, 661 UMA_ALIGN_PTR, UMA_ZONE_VM); 662 #ifndef UMA_MD_SMALL_ALLOC 663 mtx_init(&vmem_bt_lock, "btag lock", NULL, MTX_DEF); 664 uma_prealloc(vmem_bt_zone, BT_MAXALLOC); 665 /* 666 * Reserve enough tags to allocate new tags. We allow multiple 667 * CPUs to attempt to allocate new tags concurrently to limit 668 * false restarts in UMA. 669 */ 670 uma_zone_reserve(vmem_bt_zone, BT_MAXALLOC * (mp_ncpus + 1) / 2); 671 uma_zone_set_allocf(vmem_bt_zone, vmem_bt_alloc); 672 #endif 673 } 674 675 /* ---- rehash */ 676 677 static int 678 vmem_rehash(vmem_t *vm, vmem_size_t newhashsize) 679 { 680 bt_t *bt; 681 int i; 682 struct vmem_hashlist *newhashlist; 683 struct vmem_hashlist *oldhashlist; 684 vmem_size_t oldhashsize; 685 686 MPASS(newhashsize > 0); 687 688 newhashlist = malloc(sizeof(struct vmem_hashlist) * newhashsize, 689 M_VMEM, M_NOWAIT); 690 if (newhashlist == NULL) 691 return ENOMEM; 692 for (i = 0; i < newhashsize; i++) { 693 LIST_INIT(&newhashlist[i]); 694 } 695 696 VMEM_LOCK(vm); 697 oldhashlist = vm->vm_hashlist; 698 oldhashsize = vm->vm_hashsize; 699 vm->vm_hashlist = newhashlist; 700 vm->vm_hashsize = newhashsize; 701 if (oldhashlist == NULL) { 702 VMEM_UNLOCK(vm); 703 return 0; 704 } 705 for (i = 0; i < oldhashsize; i++) { 706 while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) { 707 bt_rembusy(vm, bt); 708 bt_insbusy(vm, bt); 709 } 710 } 711 VMEM_UNLOCK(vm); 712 713 if (oldhashlist != vm->vm_hash0) { 714 free(oldhashlist, M_VMEM); 715 } 716 717 return 0; 718 } 719 720 static void 721 vmem_periodic_kick(void *dummy) 722 { 723 724 taskqueue_enqueue(taskqueue_thread, &vmem_periodic_wk); 725 } 726 727 static void 728 vmem_periodic(void *unused, int pending) 729 { 730 vmem_t *vm; 731 vmem_size_t desired; 732 vmem_size_t current; 733 734 mtx_lock(&vmem_list_lock); 735 LIST_FOREACH(vm, &vmem_list, vm_alllist) { 736 #ifdef DIAGNOSTIC 737 /* Convenient time to verify vmem state. */ 738 if (enable_vmem_check == 1) { 739 VMEM_LOCK(vm); 740 vmem_check(vm); 741 VMEM_UNLOCK(vm); 742 } 743 #endif 744 desired = 1 << flsl(vm->vm_nbusytag); 745 desired = MIN(MAX(desired, VMEM_HASHSIZE_MIN), 746 VMEM_HASHSIZE_MAX); 747 current = vm->vm_hashsize; 748 749 /* Grow in powers of two. Shrink less aggressively. */ 750 if (desired >= current * 2 || desired * 4 <= current) 751 vmem_rehash(vm, desired); 752 753 /* 754 * Periodically wake up threads waiting for resources, 755 * so they could ask for reclamation again. 756 */ 757 VMEM_CONDVAR_BROADCAST(vm); 758 } 759 mtx_unlock(&vmem_list_lock); 760 761 callout_reset(&vmem_periodic_ch, vmem_periodic_interval, 762 vmem_periodic_kick, NULL); 763 } 764 765 static void 766 vmem_start_callout(void *unused) 767 { 768 769 TASK_INIT(&vmem_periodic_wk, 0, vmem_periodic, NULL); 770 vmem_periodic_interval = hz * 10; 771 callout_init(&vmem_periodic_ch, 1); 772 callout_reset(&vmem_periodic_ch, vmem_periodic_interval, 773 vmem_periodic_kick, NULL); 774 } 775 SYSINIT(vfs, SI_SUB_CONFIGURE, SI_ORDER_ANY, vmem_start_callout, NULL); 776 777 static void 778 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int type) 779 { 780 bt_t *btspan; 781 bt_t *btfree; 782 783 MPASS(type == BT_TYPE_SPAN || type == BT_TYPE_SPAN_STATIC); 784 MPASS((size & vm->vm_quantum_mask) == 0); 785 786 btspan = bt_alloc(vm); 787 btspan->bt_type = type; 788 btspan->bt_start = addr; 789 btspan->bt_size = size; 790 bt_insseg_tail(vm, btspan); 791 792 btfree = bt_alloc(vm); 793 btfree->bt_type = BT_TYPE_FREE; 794 btfree->bt_start = addr; 795 btfree->bt_size = size; 796 bt_insseg(vm, btfree, btspan); 797 bt_insfree(vm, btfree); 798 799 vm->vm_size += size; 800 } 801 802 static void 803 vmem_destroy1(vmem_t *vm) 804 { 805 bt_t *bt; 806 807 /* 808 * Drain per-cpu quantum caches. 809 */ 810 qc_destroy(vm); 811 812 /* 813 * The vmem should now only contain empty segments. 814 */ 815 VMEM_LOCK(vm); 816 MPASS(vm->vm_nbusytag == 0); 817 818 while ((bt = TAILQ_FIRST(&vm->vm_seglist)) != NULL) 819 bt_remseg(vm, bt); 820 821 if (vm->vm_hashlist != NULL && vm->vm_hashlist != vm->vm_hash0) 822 free(vm->vm_hashlist, M_VMEM); 823 824 bt_freetrim(vm, 0); 825 826 VMEM_CONDVAR_DESTROY(vm); 827 VMEM_LOCK_DESTROY(vm); 828 free(vm, M_VMEM); 829 } 830 831 static int 832 vmem_import(vmem_t *vm, vmem_size_t size, vmem_size_t align, int flags) 833 { 834 vmem_addr_t addr; 835 int error; 836 837 if (vm->vm_importfn == NULL) 838 return EINVAL; 839 840 /* 841 * To make sure we get a span that meets the alignment we double it 842 * and add the size to the tail. This slightly overestimates. 843 */ 844 if (align != vm->vm_quantum_mask + 1) 845 size = (align * 2) + size; 846 size = roundup(size, vm->vm_import_quantum); 847 848 /* 849 * Hide MAXALLOC tags so we're guaranteed to be able to add this 850 * span and the tag we want to allocate from it. 851 */ 852 MPASS(vm->vm_nfreetags >= BT_MAXALLOC); 853 vm->vm_nfreetags -= BT_MAXALLOC; 854 VMEM_UNLOCK(vm); 855 error = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr); 856 VMEM_LOCK(vm); 857 vm->vm_nfreetags += BT_MAXALLOC; 858 if (error) 859 return ENOMEM; 860 861 vmem_add1(vm, addr, size, BT_TYPE_SPAN); 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 static int 873 vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align, 874 vmem_size_t phase, vmem_size_t nocross, vmem_addr_t minaddr, 875 vmem_addr_t maxaddr, vmem_addr_t *addrp) 876 { 877 vmem_addr_t start; 878 vmem_addr_t end; 879 880 MPASS(size > 0); 881 MPASS(bt->bt_size >= size); /* caller's responsibility */ 882 883 /* 884 * XXX assumption: vmem_addr_t and vmem_size_t are 885 * unsigned integer of the same size. 886 */ 887 888 start = bt->bt_start; 889 if (start < minaddr) { 890 start = minaddr; 891 } 892 end = BT_END(bt); 893 if (end > maxaddr) 894 end = maxaddr; 895 if (start > end) 896 return (ENOMEM); 897 898 start = VMEM_ALIGNUP(start - phase, align) + phase; 899 if (start < bt->bt_start) 900 start += align; 901 if (VMEM_CROSS_P(start, start + size - 1, nocross)) { 902 MPASS(align < nocross); 903 start = VMEM_ALIGNUP(start - phase, nocross) + phase; 904 } 905 if (start <= end && end - start >= size - 1) { 906 MPASS((start & (align - 1)) == phase); 907 MPASS(!VMEM_CROSS_P(start, start + size - 1, nocross)); 908 MPASS(minaddr <= start); 909 MPASS(maxaddr == 0 || start + size - 1 <= maxaddr); 910 MPASS(bt->bt_start <= start); 911 MPASS(BT_END(bt) - start >= size - 1); 912 *addrp = start; 913 914 return (0); 915 } 916 return (ENOMEM); 917 } 918 919 /* 920 * vmem_clip: Trim the boundary tag edges to the requested start and size. 921 */ 922 static void 923 vmem_clip(vmem_t *vm, bt_t *bt, vmem_addr_t start, vmem_size_t size) 924 { 925 bt_t *btnew; 926 bt_t *btprev; 927 928 VMEM_ASSERT_LOCKED(vm); 929 MPASS(bt->bt_type == BT_TYPE_FREE); 930 MPASS(bt->bt_size >= size); 931 bt_remfree(vm, bt); 932 if (bt->bt_start != start) { 933 btprev = bt_alloc(vm); 934 btprev->bt_type = BT_TYPE_FREE; 935 btprev->bt_start = bt->bt_start; 936 btprev->bt_size = start - bt->bt_start; 937 bt->bt_start = start; 938 bt->bt_size -= btprev->bt_size; 939 bt_insfree(vm, btprev); 940 bt_insseg(vm, btprev, 941 TAILQ_PREV(bt, vmem_seglist, bt_seglist)); 942 } 943 MPASS(bt->bt_start == start); 944 if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) { 945 /* split */ 946 btnew = bt_alloc(vm); 947 btnew->bt_type = BT_TYPE_BUSY; 948 btnew->bt_start = bt->bt_start; 949 btnew->bt_size = size; 950 bt->bt_start = bt->bt_start + size; 951 bt->bt_size -= size; 952 bt_insfree(vm, bt); 953 bt_insseg(vm, btnew, 954 TAILQ_PREV(bt, vmem_seglist, bt_seglist)); 955 bt_insbusy(vm, btnew); 956 bt = btnew; 957 } else { 958 bt->bt_type = BT_TYPE_BUSY; 959 bt_insbusy(vm, bt); 960 } 961 MPASS(bt->bt_size >= size); 962 bt->bt_type = BT_TYPE_BUSY; 963 } 964 965 /* ---- vmem API */ 966 967 void 968 vmem_set_import(vmem_t *vm, vmem_import_t *importfn, 969 vmem_release_t *releasefn, void *arg, vmem_size_t import_quantum) 970 { 971 972 VMEM_LOCK(vm); 973 vm->vm_importfn = importfn; 974 vm->vm_releasefn = releasefn; 975 vm->vm_arg = arg; 976 vm->vm_import_quantum = import_quantum; 977 VMEM_UNLOCK(vm); 978 } 979 980 void 981 vmem_set_reclaim(vmem_t *vm, vmem_reclaim_t *reclaimfn) 982 { 983 984 VMEM_LOCK(vm); 985 vm->vm_reclaimfn = reclaimfn; 986 VMEM_UNLOCK(vm); 987 } 988 989 /* 990 * vmem_init: Initializes vmem arena. 991 */ 992 vmem_t * 993 vmem_init(vmem_t *vm, const char *name, vmem_addr_t base, vmem_size_t size, 994 vmem_size_t quantum, vmem_size_t qcache_max, int flags) 995 { 996 int i; 997 998 MPASS(quantum > 0); 999 MPASS((quantum & (quantum - 1)) == 0); 1000 1001 bzero(vm, sizeof(*vm)); 1002 1003 VMEM_CONDVAR_INIT(vm, name); 1004 VMEM_LOCK_INIT(vm, name); 1005 vm->vm_nfreetags = 0; 1006 LIST_INIT(&vm->vm_freetags); 1007 strlcpy(vm->vm_name, name, sizeof(vm->vm_name)); 1008 vm->vm_quantum_mask = quantum - 1; 1009 vm->vm_quantum_shift = flsl(quantum) - 1; 1010 vm->vm_nbusytag = 0; 1011 vm->vm_size = 0; 1012 vm->vm_inuse = 0; 1013 qc_init(vm, qcache_max); 1014 1015 TAILQ_INIT(&vm->vm_seglist); 1016 for (i = 0; i < VMEM_MAXORDER; i++) { 1017 LIST_INIT(&vm->vm_freelist[i]); 1018 } 1019 memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0)); 1020 vm->vm_hashsize = VMEM_HASHSIZE_MIN; 1021 vm->vm_hashlist = vm->vm_hash0; 1022 1023 if (size != 0) { 1024 if (vmem_add(vm, base, size, flags) != 0) { 1025 vmem_destroy1(vm); 1026 return NULL; 1027 } 1028 } 1029 1030 mtx_lock(&vmem_list_lock); 1031 LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist); 1032 mtx_unlock(&vmem_list_lock); 1033 1034 return vm; 1035 } 1036 1037 /* 1038 * vmem_create: create an arena. 1039 */ 1040 vmem_t * 1041 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size, 1042 vmem_size_t quantum, vmem_size_t qcache_max, int flags) 1043 { 1044 1045 vmem_t *vm; 1046 1047 vm = malloc(sizeof(*vm), M_VMEM, flags & (M_WAITOK|M_NOWAIT)); 1048 if (vm == NULL) 1049 return (NULL); 1050 if (vmem_init(vm, name, base, size, quantum, qcache_max, 1051 flags) == NULL) 1052 return (NULL); 1053 return (vm); 1054 } 1055 1056 void 1057 vmem_destroy(vmem_t *vm) 1058 { 1059 1060 mtx_lock(&vmem_list_lock); 1061 LIST_REMOVE(vm, vm_alllist); 1062 mtx_unlock(&vmem_list_lock); 1063 1064 vmem_destroy1(vm); 1065 } 1066 1067 vmem_size_t 1068 vmem_roundup_size(vmem_t *vm, vmem_size_t size) 1069 { 1070 1071 return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask; 1072 } 1073 1074 /* 1075 * vmem_alloc: allocate resource from the arena. 1076 */ 1077 int 1078 vmem_alloc(vmem_t *vm, vmem_size_t size, int flags, vmem_addr_t *addrp) 1079 { 1080 const int strat __unused = flags & VMEM_FITMASK; 1081 qcache_t *qc; 1082 1083 flags &= VMEM_FLAGS; 1084 MPASS(size > 0); 1085 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT); 1086 if ((flags & M_NOWAIT) == 0) 1087 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_alloc"); 1088 1089 if (size <= vm->vm_qcache_max) { 1090 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift]; 1091 *addrp = (vmem_addr_t)uma_zalloc(qc->qc_cache, flags); 1092 if (*addrp == 0) 1093 return (ENOMEM); 1094 return (0); 1095 } 1096 1097 return vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, 1098 flags, addrp); 1099 } 1100 1101 int 1102 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align, 1103 const vmem_size_t phase, const vmem_size_t nocross, 1104 const vmem_addr_t minaddr, const vmem_addr_t maxaddr, int flags, 1105 vmem_addr_t *addrp) 1106 { 1107 const vmem_size_t size = vmem_roundup_size(vm, size0); 1108 struct vmem_freelist *list; 1109 struct vmem_freelist *first; 1110 struct vmem_freelist *end; 1111 vmem_size_t avail; 1112 bt_t *bt; 1113 int error; 1114 int strat; 1115 1116 flags &= VMEM_FLAGS; 1117 strat = flags & VMEM_FITMASK; 1118 MPASS(size0 > 0); 1119 MPASS(size > 0); 1120 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT); 1121 MPASS((flags & (M_NOWAIT|M_WAITOK)) != (M_NOWAIT|M_WAITOK)); 1122 if ((flags & M_NOWAIT) == 0) 1123 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_xalloc"); 1124 MPASS((align & vm->vm_quantum_mask) == 0); 1125 MPASS((align & (align - 1)) == 0); 1126 MPASS((phase & vm->vm_quantum_mask) == 0); 1127 MPASS((nocross & vm->vm_quantum_mask) == 0); 1128 MPASS((nocross & (nocross - 1)) == 0); 1129 MPASS((align == 0 && phase == 0) || phase < align); 1130 MPASS(nocross == 0 || nocross >= size); 1131 MPASS(minaddr <= maxaddr); 1132 MPASS(!VMEM_CROSS_P(phase, phase + size - 1, nocross)); 1133 1134 if (align == 0) 1135 align = vm->vm_quantum_mask + 1; 1136 1137 *addrp = 0; 1138 end = &vm->vm_freelist[VMEM_MAXORDER]; 1139 /* 1140 * choose a free block from which we allocate. 1141 */ 1142 first = bt_freehead_toalloc(vm, size, strat); 1143 VMEM_LOCK(vm); 1144 for (;;) { 1145 /* 1146 * Make sure we have enough tags to complete the 1147 * operation. 1148 */ 1149 if (vm->vm_nfreetags < BT_MAXALLOC && 1150 bt_fill(vm, flags) != 0) { 1151 error = ENOMEM; 1152 break; 1153 } 1154 /* 1155 * Scan freelists looking for a tag that satisfies the 1156 * allocation. If we're doing BESTFIT we may encounter 1157 * sizes below the request. If we're doing FIRSTFIT we 1158 * inspect only the first element from each list. 1159 */ 1160 for (list = first; list < end; list++) { 1161 LIST_FOREACH(bt, list, bt_freelist) { 1162 if (bt->bt_size >= size) { 1163 error = vmem_fit(bt, size, align, phase, 1164 nocross, minaddr, maxaddr, addrp); 1165 if (error == 0) { 1166 vmem_clip(vm, bt, *addrp, size); 1167 goto out; 1168 } 1169 } 1170 /* FIRST skips to the next list. */ 1171 if (strat == M_FIRSTFIT) 1172 break; 1173 } 1174 } 1175 /* 1176 * Retry if the fast algorithm failed. 1177 */ 1178 if (strat == M_FIRSTFIT) { 1179 strat = M_BESTFIT; 1180 first = bt_freehead_toalloc(vm, size, strat); 1181 continue; 1182 } 1183 /* 1184 * XXX it is possible to fail to meet restrictions with the 1185 * imported region. It is up to the user to specify the 1186 * import quantum such that it can satisfy any allocation. 1187 */ 1188 if (vmem_import(vm, size, align, flags) == 0) 1189 continue; 1190 1191 /* 1192 * Try to free some space from the quantum cache or reclaim 1193 * functions if available. 1194 */ 1195 if (vm->vm_qcache_max != 0 || vm->vm_reclaimfn != NULL) { 1196 avail = vm->vm_size - vm->vm_inuse; 1197 VMEM_UNLOCK(vm); 1198 if (vm->vm_qcache_max != 0) 1199 qc_drain(vm); 1200 if (vm->vm_reclaimfn != NULL) 1201 vm->vm_reclaimfn(vm, flags); 1202 VMEM_LOCK(vm); 1203 /* If we were successful retry even NOWAIT. */ 1204 if (vm->vm_size - vm->vm_inuse > avail) 1205 continue; 1206 } 1207 if ((flags & M_NOWAIT) != 0) { 1208 error = ENOMEM; 1209 break; 1210 } 1211 VMEM_CONDVAR_WAIT(vm); 1212 } 1213 out: 1214 VMEM_UNLOCK(vm); 1215 if (error != 0 && (flags & M_NOWAIT) == 0) 1216 panic("failed to allocate waiting allocation\n"); 1217 1218 return (error); 1219 } 1220 1221 /* 1222 * vmem_free: free the resource to the arena. 1223 */ 1224 void 1225 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size) 1226 { 1227 qcache_t *qc; 1228 MPASS(size > 0); 1229 1230 if (size <= vm->vm_qcache_max) { 1231 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift]; 1232 uma_zfree(qc->qc_cache, (void *)addr); 1233 } else 1234 vmem_xfree(vm, addr, size); 1235 } 1236 1237 void 1238 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size) 1239 { 1240 bt_t *bt; 1241 bt_t *t; 1242 1243 MPASS(size > 0); 1244 1245 VMEM_LOCK(vm); 1246 bt = bt_lookupbusy(vm, addr); 1247 MPASS(bt != NULL); 1248 MPASS(bt->bt_start == addr); 1249 MPASS(bt->bt_size == vmem_roundup_size(vm, size) || 1250 bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask); 1251 MPASS(bt->bt_type == BT_TYPE_BUSY); 1252 bt_rembusy(vm, bt); 1253 bt->bt_type = BT_TYPE_FREE; 1254 1255 /* coalesce */ 1256 t = TAILQ_NEXT(bt, bt_seglist); 1257 if (t != NULL && t->bt_type == BT_TYPE_FREE) { 1258 MPASS(BT_END(bt) < t->bt_start); /* YYY */ 1259 bt->bt_size += t->bt_size; 1260 bt_remfree(vm, t); 1261 bt_remseg(vm, t); 1262 } 1263 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist); 1264 if (t != NULL && t->bt_type == BT_TYPE_FREE) { 1265 MPASS(BT_END(t) < bt->bt_start); /* YYY */ 1266 bt->bt_size += t->bt_size; 1267 bt->bt_start = t->bt_start; 1268 bt_remfree(vm, t); 1269 bt_remseg(vm, t); 1270 } 1271 1272 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist); 1273 MPASS(t != NULL); 1274 MPASS(BT_ISSPAN_P(t) || t->bt_type == BT_TYPE_BUSY); 1275 if (vm->vm_releasefn != NULL && t->bt_type == BT_TYPE_SPAN && 1276 t->bt_size == bt->bt_size) { 1277 vmem_addr_t spanaddr; 1278 vmem_size_t spansize; 1279 1280 MPASS(t->bt_start == bt->bt_start); 1281 spanaddr = bt->bt_start; 1282 spansize = bt->bt_size; 1283 bt_remseg(vm, bt); 1284 bt_remseg(vm, t); 1285 vm->vm_size -= spansize; 1286 VMEM_CONDVAR_BROADCAST(vm); 1287 bt_freetrim(vm, BT_MAXFREE); 1288 (*vm->vm_releasefn)(vm->vm_arg, spanaddr, spansize); 1289 } else { 1290 bt_insfree(vm, bt); 1291 VMEM_CONDVAR_BROADCAST(vm); 1292 bt_freetrim(vm, BT_MAXFREE); 1293 } 1294 } 1295 1296 /* 1297 * vmem_add: 1298 * 1299 */ 1300 int 1301 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags) 1302 { 1303 int error; 1304 1305 error = 0; 1306 flags &= VMEM_FLAGS; 1307 VMEM_LOCK(vm); 1308 if (vm->vm_nfreetags >= BT_MAXALLOC || bt_fill(vm, flags) == 0) 1309 vmem_add1(vm, addr, size, BT_TYPE_SPAN_STATIC); 1310 else 1311 error = ENOMEM; 1312 VMEM_UNLOCK(vm); 1313 1314 return (error); 1315 } 1316 1317 /* 1318 * vmem_size: information about arenas size 1319 */ 1320 vmem_size_t 1321 vmem_size(vmem_t *vm, int typemask) 1322 { 1323 int i; 1324 1325 switch (typemask) { 1326 case VMEM_ALLOC: 1327 return vm->vm_inuse; 1328 case VMEM_FREE: 1329 return vm->vm_size - vm->vm_inuse; 1330 case VMEM_FREE|VMEM_ALLOC: 1331 return vm->vm_size; 1332 case VMEM_MAXFREE: 1333 VMEM_LOCK(vm); 1334 for (i = VMEM_MAXORDER - 1; i >= 0; i--) { 1335 if (LIST_EMPTY(&vm->vm_freelist[i])) 1336 continue; 1337 VMEM_UNLOCK(vm); 1338 return ((vmem_size_t)ORDER2SIZE(i) << 1339 vm->vm_quantum_shift); 1340 } 1341 VMEM_UNLOCK(vm); 1342 return (0); 1343 default: 1344 panic("vmem_size"); 1345 } 1346 } 1347 1348 /* ---- debug */ 1349 1350 #if defined(DDB) || defined(DIAGNOSTIC) 1351 1352 static void bt_dump(const bt_t *, int (*)(const char *, ...) 1353 __printflike(1, 2)); 1354 1355 static const char * 1356 bt_type_string(int type) 1357 { 1358 1359 switch (type) { 1360 case BT_TYPE_BUSY: 1361 return "busy"; 1362 case BT_TYPE_FREE: 1363 return "free"; 1364 case BT_TYPE_SPAN: 1365 return "span"; 1366 case BT_TYPE_SPAN_STATIC: 1367 return "static span"; 1368 default: 1369 break; 1370 } 1371 return "BOGUS"; 1372 } 1373 1374 static void 1375 bt_dump(const bt_t *bt, int (*pr)(const char *, ...)) 1376 { 1377 1378 (*pr)("\t%p: %jx %jx, %d(%s)\n", 1379 bt, (intmax_t)bt->bt_start, (intmax_t)bt->bt_size, 1380 bt->bt_type, bt_type_string(bt->bt_type)); 1381 } 1382 1383 static void 1384 vmem_dump(const vmem_t *vm , int (*pr)(const char *, ...) __printflike(1, 2)) 1385 { 1386 const bt_t *bt; 1387 int i; 1388 1389 (*pr)("vmem %p '%s'\n", vm, vm->vm_name); 1390 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1391 bt_dump(bt, pr); 1392 } 1393 1394 for (i = 0; i < VMEM_MAXORDER; i++) { 1395 const struct vmem_freelist *fl = &vm->vm_freelist[i]; 1396 1397 if (LIST_EMPTY(fl)) { 1398 continue; 1399 } 1400 1401 (*pr)("freelist[%d]\n", i); 1402 LIST_FOREACH(bt, fl, bt_freelist) { 1403 bt_dump(bt, pr); 1404 } 1405 } 1406 } 1407 1408 #endif /* defined(DDB) || defined(DIAGNOSTIC) */ 1409 1410 #if defined(DDB) 1411 #include <ddb/ddb.h> 1412 1413 static bt_t * 1414 vmem_whatis_lookup(vmem_t *vm, vmem_addr_t addr) 1415 { 1416 bt_t *bt; 1417 1418 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1419 if (BT_ISSPAN_P(bt)) { 1420 continue; 1421 } 1422 if (bt->bt_start <= addr && addr <= BT_END(bt)) { 1423 return bt; 1424 } 1425 } 1426 1427 return NULL; 1428 } 1429 1430 void 1431 vmem_whatis(vmem_addr_t addr, int (*pr)(const char *, ...)) 1432 { 1433 vmem_t *vm; 1434 1435 LIST_FOREACH(vm, &vmem_list, vm_alllist) { 1436 bt_t *bt; 1437 1438 bt = vmem_whatis_lookup(vm, addr); 1439 if (bt == NULL) { 1440 continue; 1441 } 1442 (*pr)("%p is %p+%zu in VMEM '%s' (%s)\n", 1443 (void *)addr, (void *)bt->bt_start, 1444 (vmem_size_t)(addr - bt->bt_start), vm->vm_name, 1445 (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free"); 1446 } 1447 } 1448 1449 void 1450 vmem_printall(const char *modif, int (*pr)(const char *, ...)) 1451 { 1452 const vmem_t *vm; 1453 1454 LIST_FOREACH(vm, &vmem_list, vm_alllist) { 1455 vmem_dump(vm, pr); 1456 } 1457 } 1458 1459 void 1460 vmem_print(vmem_addr_t addr, const char *modif, int (*pr)(const char *, ...)) 1461 { 1462 const vmem_t *vm = (const void *)addr; 1463 1464 vmem_dump(vm, pr); 1465 } 1466 1467 DB_SHOW_COMMAND(vmemdump, vmemdump) 1468 { 1469 1470 if (!have_addr) { 1471 db_printf("usage: show vmemdump <addr>\n"); 1472 return; 1473 } 1474 1475 vmem_dump((const vmem_t *)addr, db_printf); 1476 } 1477 1478 DB_SHOW_ALL_COMMAND(vmemdump, vmemdumpall) 1479 { 1480 const vmem_t *vm; 1481 1482 LIST_FOREACH(vm, &vmem_list, vm_alllist) 1483 vmem_dump(vm, db_printf); 1484 } 1485 1486 DB_SHOW_COMMAND(vmem, vmem_summ) 1487 { 1488 const vmem_t *vm = (const void *)addr; 1489 const bt_t *bt; 1490 size_t ft[VMEM_MAXORDER], ut[VMEM_MAXORDER]; 1491 size_t fs[VMEM_MAXORDER], us[VMEM_MAXORDER]; 1492 int ord; 1493 1494 if (!have_addr) { 1495 db_printf("usage: show vmem <addr>\n"); 1496 return; 1497 } 1498 1499 db_printf("vmem %p '%s'\n", vm, vm->vm_name); 1500 db_printf("\tquantum:\t%zu\n", vm->vm_quantum_mask + 1); 1501 db_printf("\tsize:\t%zu\n", vm->vm_size); 1502 db_printf("\tinuse:\t%zu\n", vm->vm_inuse); 1503 db_printf("\tfree:\t%zu\n", vm->vm_size - vm->vm_inuse); 1504 db_printf("\tbusy tags:\t%d\n", vm->vm_nbusytag); 1505 db_printf("\tfree tags:\t%d\n", vm->vm_nfreetags); 1506 1507 memset(&ft, 0, sizeof(ft)); 1508 memset(&ut, 0, sizeof(ut)); 1509 memset(&fs, 0, sizeof(fs)); 1510 memset(&us, 0, sizeof(us)); 1511 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1512 ord = SIZE2ORDER(bt->bt_size >> vm->vm_quantum_shift); 1513 if (bt->bt_type == BT_TYPE_BUSY) { 1514 ut[ord]++; 1515 us[ord] += bt->bt_size; 1516 } else if (bt->bt_type == BT_TYPE_FREE) { 1517 ft[ord]++; 1518 fs[ord] += bt->bt_size; 1519 } 1520 } 1521 db_printf("\t\t\tinuse\tsize\t\tfree\tsize\n"); 1522 for (ord = 0; ord < VMEM_MAXORDER; ord++) { 1523 if (ut[ord] == 0 && ft[ord] == 0) 1524 continue; 1525 db_printf("\t%-15zu %zu\t%-15zu %zu\t%-16zu\n", 1526 ORDER2SIZE(ord) << vm->vm_quantum_shift, 1527 ut[ord], us[ord], ft[ord], fs[ord]); 1528 } 1529 } 1530 1531 DB_SHOW_ALL_COMMAND(vmem, vmem_summall) 1532 { 1533 const vmem_t *vm; 1534 1535 LIST_FOREACH(vm, &vmem_list, vm_alllist) 1536 vmem_summ((db_expr_t)vm, TRUE, count, modif); 1537 } 1538 #endif /* defined(DDB) */ 1539 1540 #define vmem_printf printf 1541 1542 #if defined(DIAGNOSTIC) 1543 1544 static bool 1545 vmem_check_sanity(vmem_t *vm) 1546 { 1547 const bt_t *bt, *bt2; 1548 1549 MPASS(vm != NULL); 1550 1551 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1552 if (bt->bt_start > BT_END(bt)) { 1553 printf("corrupted tag\n"); 1554 bt_dump(bt, vmem_printf); 1555 return false; 1556 } 1557 } 1558 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1559 TAILQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) { 1560 if (bt == bt2) { 1561 continue; 1562 } 1563 if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) { 1564 continue; 1565 } 1566 if (bt->bt_start <= BT_END(bt2) && 1567 bt2->bt_start <= BT_END(bt)) { 1568 printf("overwrapped tags\n"); 1569 bt_dump(bt, vmem_printf); 1570 bt_dump(bt2, vmem_printf); 1571 return false; 1572 } 1573 } 1574 } 1575 1576 return true; 1577 } 1578 1579 static void 1580 vmem_check(vmem_t *vm) 1581 { 1582 1583 if (!vmem_check_sanity(vm)) { 1584 panic("insanity vmem %p", vm); 1585 } 1586 } 1587 1588 #endif /* defined(DIAGNOSTIC) */ 1589