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 #include <sys/vmmeter.h> 63 64 #include "opt_vm.h" 65 66 #include <vm/uma.h> 67 #include <vm/vm.h> 68 #include <vm/pmap.h> 69 #include <vm/vm_map.h> 70 #include <vm/vm_object.h> 71 #include <vm/vm_kern.h> 72 #include <vm/vm_extern.h> 73 #include <vm/vm_param.h> 74 #include <vm/vm_page.h> 75 #include <vm/vm_pageout.h> 76 #include <vm/vm_phys.h> 77 #include <vm/vm_pagequeue.h> 78 #include <vm/uma_int.h> 79 80 #define VMEM_OPTORDER 5 81 #define VMEM_OPTVALUE (1 << VMEM_OPTORDER) 82 #define VMEM_MAXORDER \ 83 (VMEM_OPTVALUE - 1 + sizeof(vmem_size_t) * NBBY - VMEM_OPTORDER) 84 85 #define VMEM_HASHSIZE_MIN 16 86 #define VMEM_HASHSIZE_MAX 131072 87 88 #define VMEM_QCACHE_IDX_MAX 16 89 90 #define VMEM_FITMASK (M_BESTFIT | M_FIRSTFIT | M_NEXTFIT) 91 92 #define VMEM_FLAGS (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM | \ 93 M_BESTFIT | M_FIRSTFIT | M_NEXTFIT) 94 95 #define BT_FLAGS (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM) 96 97 #define QC_NAME_MAX 16 98 99 /* 100 * Data structures private to vmem. 101 */ 102 MALLOC_DEFINE(M_VMEM, "vmem", "vmem internal structures"); 103 104 typedef struct vmem_btag bt_t; 105 106 TAILQ_HEAD(vmem_seglist, vmem_btag); 107 LIST_HEAD(vmem_freelist, vmem_btag); 108 LIST_HEAD(vmem_hashlist, vmem_btag); 109 110 struct qcache { 111 uma_zone_t qc_cache; 112 vmem_t *qc_vmem; 113 vmem_size_t qc_size; 114 char qc_name[QC_NAME_MAX]; 115 }; 116 typedef struct qcache qcache_t; 117 #define QC_POOL_TO_QCACHE(pool) ((qcache_t *)(pool->pr_qcache)) 118 119 #define VMEM_NAME_MAX 16 120 121 /* boundary tag */ 122 struct vmem_btag { 123 TAILQ_ENTRY(vmem_btag) bt_seglist; 124 union { 125 LIST_ENTRY(vmem_btag) u_freelist; /* BT_TYPE_FREE */ 126 LIST_ENTRY(vmem_btag) u_hashlist; /* BT_TYPE_BUSY */ 127 } bt_u; 128 #define bt_hashlist bt_u.u_hashlist 129 #define bt_freelist bt_u.u_freelist 130 vmem_addr_t bt_start; 131 vmem_size_t bt_size; 132 int bt_type; 133 }; 134 135 /* vmem arena */ 136 struct vmem { 137 struct mtx_padalign vm_lock; 138 struct cv vm_cv; 139 char vm_name[VMEM_NAME_MAX+1]; 140 LIST_ENTRY(vmem) vm_alllist; 141 struct vmem_hashlist vm_hash0[VMEM_HASHSIZE_MIN]; 142 struct vmem_freelist vm_freelist[VMEM_MAXORDER]; 143 struct vmem_seglist vm_seglist; 144 struct vmem_hashlist *vm_hashlist; 145 vmem_size_t vm_hashsize; 146 147 /* Constant after init */ 148 vmem_size_t vm_qcache_max; 149 vmem_size_t vm_quantum_mask; 150 vmem_size_t vm_import_quantum; 151 int vm_quantum_shift; 152 153 /* Written on alloc/free */ 154 LIST_HEAD(, vmem_btag) vm_freetags; 155 int vm_nfreetags; 156 int vm_nbusytag; 157 vmem_size_t vm_inuse; 158 vmem_size_t vm_size; 159 vmem_size_t vm_limit; 160 struct vmem_btag vm_cursor; 161 162 /* Used on import. */ 163 vmem_import_t *vm_importfn; 164 vmem_release_t *vm_releasefn; 165 void *vm_arg; 166 167 /* Space exhaustion callback. */ 168 vmem_reclaim_t *vm_reclaimfn; 169 170 /* quantum cache */ 171 qcache_t vm_qcache[VMEM_QCACHE_IDX_MAX]; 172 }; 173 174 #define BT_TYPE_SPAN 1 /* Allocated from importfn */ 175 #define BT_TYPE_SPAN_STATIC 2 /* vmem_add() or create. */ 176 #define BT_TYPE_FREE 3 /* Available space. */ 177 #define BT_TYPE_BUSY 4 /* Used space. */ 178 #define BT_TYPE_CURSOR 5 /* Cursor for nextfit allocations. */ 179 #define BT_ISSPAN_P(bt) ((bt)->bt_type <= BT_TYPE_SPAN_STATIC) 180 181 #define BT_END(bt) ((bt)->bt_start + (bt)->bt_size - 1) 182 183 #if defined(DIAGNOSTIC) 184 static int enable_vmem_check = 1; 185 SYSCTL_INT(_debug, OID_AUTO, vmem_check, CTLFLAG_RWTUN, 186 &enable_vmem_check, 0, "Enable vmem check"); 187 static void vmem_check(vmem_t *); 188 #endif 189 190 static struct callout vmem_periodic_ch; 191 static int vmem_periodic_interval; 192 static struct task vmem_periodic_wk; 193 194 static struct mtx_padalign __exclusive_cache_line vmem_list_lock; 195 static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list); 196 static uma_zone_t vmem_zone; 197 198 /* ---- misc */ 199 #define VMEM_CONDVAR_INIT(vm, wchan) cv_init(&vm->vm_cv, wchan) 200 #define VMEM_CONDVAR_DESTROY(vm) cv_destroy(&vm->vm_cv) 201 #define VMEM_CONDVAR_WAIT(vm) cv_wait(&vm->vm_cv, &vm->vm_lock) 202 #define VMEM_CONDVAR_BROADCAST(vm) cv_broadcast(&vm->vm_cv) 203 204 #define VMEM_LOCK(vm) mtx_lock(&vm->vm_lock) 205 #define VMEM_TRYLOCK(vm) mtx_trylock(&vm->vm_lock) 206 #define VMEM_UNLOCK(vm) mtx_unlock(&vm->vm_lock) 207 #define VMEM_LOCK_INIT(vm, name) mtx_init(&vm->vm_lock, (name), NULL, MTX_DEF) 208 #define VMEM_LOCK_DESTROY(vm) mtx_destroy(&vm->vm_lock) 209 #define VMEM_ASSERT_LOCKED(vm) mtx_assert(&vm->vm_lock, MA_OWNED); 210 211 #define VMEM_ALIGNUP(addr, align) (-(-(addr) & -(align))) 212 213 #define VMEM_CROSS_P(addr1, addr2, boundary) \ 214 ((((addr1) ^ (addr2)) & -(boundary)) != 0) 215 216 #define ORDER2SIZE(order) ((order) < VMEM_OPTVALUE ? ((order) + 1) : \ 217 (vmem_size_t)1 << ((order) - (VMEM_OPTVALUE - VMEM_OPTORDER - 1))) 218 #define SIZE2ORDER(size) ((size) <= VMEM_OPTVALUE ? ((size) - 1) : \ 219 (flsl(size) + (VMEM_OPTVALUE - VMEM_OPTORDER - 2))) 220 221 /* 222 * Maximum number of boundary tags that may be required to satisfy an 223 * allocation. Two may be required to import. Another two may be 224 * required to clip edges. 225 */ 226 #define BT_MAXALLOC 4 227 228 /* 229 * Max free limits the number of locally cached boundary tags. We 230 * just want to avoid hitting the zone allocator for every call. 231 */ 232 #define BT_MAXFREE (BT_MAXALLOC * 8) 233 234 /* Allocator for boundary tags. */ 235 static uma_zone_t vmem_bt_zone; 236 237 /* boot time arena storage. */ 238 static struct vmem kernel_arena_storage; 239 static struct vmem buffer_arena_storage; 240 static struct vmem transient_arena_storage; 241 /* kernel and kmem arenas are aliased for backwards KPI compat. */ 242 vmem_t *kernel_arena = &kernel_arena_storage; 243 vmem_t *kmem_arena = &kernel_arena_storage; 244 vmem_t *buffer_arena = &buffer_arena_storage; 245 vmem_t *transient_arena = &transient_arena_storage; 246 247 #ifdef DEBUG_MEMGUARD 248 static struct vmem memguard_arena_storage; 249 vmem_t *memguard_arena = &memguard_arena_storage; 250 #endif 251 252 static bool 253 bt_isbusy(bt_t *bt) 254 { 255 return (bt->bt_type == BT_TYPE_BUSY); 256 } 257 258 static bool 259 bt_isfree(bt_t *bt) 260 { 261 return (bt->bt_type == BT_TYPE_FREE); 262 } 263 264 /* 265 * Fill the vmem's boundary tag cache. We guarantee that boundary tag 266 * allocation will not fail once bt_fill() passes. To do so we cache 267 * at least the maximum possible tag allocations in the arena. 268 */ 269 static __noinline int 270 _bt_fill(vmem_t *vm, int flags) 271 { 272 bt_t *bt; 273 274 VMEM_ASSERT_LOCKED(vm); 275 276 /* 277 * Only allow the kernel arena and arenas derived from kernel arena to 278 * dip into reserve tags. They are where new tags come from. 279 */ 280 flags &= BT_FLAGS; 281 if (vm != kernel_arena && vm->vm_arg != kernel_arena) 282 flags &= ~M_USE_RESERVE; 283 284 /* 285 * Loop until we meet the reserve. To minimize the lock shuffle 286 * and prevent simultaneous fills we first try a NOWAIT regardless 287 * of the caller's flags. Specify M_NOVM so we don't recurse while 288 * holding a vmem lock. 289 */ 290 while (vm->vm_nfreetags < BT_MAXALLOC) { 291 bt = uma_zalloc(vmem_bt_zone, 292 (flags & M_USE_RESERVE) | M_NOWAIT | M_NOVM); 293 if (bt == NULL) { 294 VMEM_UNLOCK(vm); 295 bt = uma_zalloc(vmem_bt_zone, flags); 296 VMEM_LOCK(vm); 297 if (bt == NULL) 298 break; 299 } 300 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist); 301 vm->vm_nfreetags++; 302 } 303 304 if (vm->vm_nfreetags < BT_MAXALLOC) 305 return ENOMEM; 306 307 return 0; 308 } 309 310 static inline int 311 bt_fill(vmem_t *vm, int flags) 312 { 313 if (vm->vm_nfreetags >= BT_MAXALLOC) 314 return (0); 315 return (_bt_fill(vm, flags)); 316 } 317 318 /* 319 * Pop a tag off of the freetag stack. 320 */ 321 static bt_t * 322 bt_alloc(vmem_t *vm) 323 { 324 bt_t *bt; 325 326 VMEM_ASSERT_LOCKED(vm); 327 bt = LIST_FIRST(&vm->vm_freetags); 328 MPASS(bt != NULL); 329 LIST_REMOVE(bt, bt_freelist); 330 vm->vm_nfreetags--; 331 332 return bt; 333 } 334 335 /* 336 * Trim the per-vmem free list. Returns with the lock released to 337 * avoid allocator recursions. 338 */ 339 static void 340 bt_freetrim(vmem_t *vm, int freelimit) 341 { 342 LIST_HEAD(, vmem_btag) freetags; 343 bt_t *bt; 344 345 LIST_INIT(&freetags); 346 VMEM_ASSERT_LOCKED(vm); 347 while (vm->vm_nfreetags > freelimit) { 348 bt = LIST_FIRST(&vm->vm_freetags); 349 LIST_REMOVE(bt, bt_freelist); 350 vm->vm_nfreetags--; 351 LIST_INSERT_HEAD(&freetags, bt, bt_freelist); 352 } 353 VMEM_UNLOCK(vm); 354 while ((bt = LIST_FIRST(&freetags)) != NULL) { 355 LIST_REMOVE(bt, bt_freelist); 356 uma_zfree(vmem_bt_zone, bt); 357 } 358 } 359 360 static inline void 361 bt_free(vmem_t *vm, bt_t *bt) 362 { 363 364 VMEM_ASSERT_LOCKED(vm); 365 MPASS(LIST_FIRST(&vm->vm_freetags) != bt); 366 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist); 367 vm->vm_nfreetags++; 368 } 369 370 /* 371 * Hide MAXALLOC tags before dropping the arena lock to ensure that a 372 * concurrent allocation attempt does not grab them. 373 */ 374 static void 375 bt_save(vmem_t *vm) 376 { 377 KASSERT(vm->vm_nfreetags >= BT_MAXALLOC, 378 ("%s: insufficient free tags %d", __func__, vm->vm_nfreetags)); 379 vm->vm_nfreetags -= BT_MAXALLOC; 380 } 381 382 static void 383 bt_restore(vmem_t *vm) 384 { 385 vm->vm_nfreetags += BT_MAXALLOC; 386 } 387 388 /* 389 * freelist[0] ... [1, 1] 390 * freelist[1] ... [2, 2] 391 * : 392 * freelist[29] ... [30, 30] 393 * freelist[30] ... [31, 31] 394 * freelist[31] ... [32, 63] 395 * freelist[33] ... [64, 127] 396 * : 397 * freelist[n] ... [(1 << (n - 26)), (1 << (n - 25)) - 1] 398 * : 399 */ 400 401 static struct vmem_freelist * 402 bt_freehead_tofree(vmem_t *vm, vmem_size_t size) 403 { 404 const vmem_size_t qsize = size >> vm->vm_quantum_shift; 405 const int idx = SIZE2ORDER(qsize); 406 407 MPASS(size != 0 && qsize != 0); 408 MPASS((size & vm->vm_quantum_mask) == 0); 409 MPASS(idx >= 0); 410 MPASS(idx < VMEM_MAXORDER); 411 412 return &vm->vm_freelist[idx]; 413 } 414 415 /* 416 * bt_freehead_toalloc: return the freelist for the given size and allocation 417 * strategy. 418 * 419 * For M_FIRSTFIT, return the list in which any blocks are large enough 420 * for the requested size. otherwise, return the list which can have blocks 421 * large enough for the requested size. 422 */ 423 static struct vmem_freelist * 424 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, int strat) 425 { 426 const vmem_size_t qsize = size >> vm->vm_quantum_shift; 427 int idx = SIZE2ORDER(qsize); 428 429 MPASS(size != 0 && qsize != 0); 430 MPASS((size & vm->vm_quantum_mask) == 0); 431 432 if (strat == M_FIRSTFIT && ORDER2SIZE(idx) != qsize) { 433 idx++; 434 /* check too large request? */ 435 } 436 MPASS(idx >= 0); 437 MPASS(idx < VMEM_MAXORDER); 438 439 return &vm->vm_freelist[idx]; 440 } 441 442 /* ---- boundary tag hash */ 443 444 static struct vmem_hashlist * 445 bt_hashhead(vmem_t *vm, vmem_addr_t addr) 446 { 447 struct vmem_hashlist *list; 448 unsigned int hash; 449 450 hash = hash32_buf(&addr, sizeof(addr), 0); 451 list = &vm->vm_hashlist[hash % vm->vm_hashsize]; 452 453 return list; 454 } 455 456 static bt_t * 457 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr) 458 { 459 struct vmem_hashlist *list; 460 bt_t *bt; 461 462 VMEM_ASSERT_LOCKED(vm); 463 list = bt_hashhead(vm, addr); 464 LIST_FOREACH(bt, list, bt_hashlist) { 465 if (bt->bt_start == addr) { 466 break; 467 } 468 } 469 470 return bt; 471 } 472 473 static void 474 bt_rembusy(vmem_t *vm, bt_t *bt) 475 { 476 477 VMEM_ASSERT_LOCKED(vm); 478 MPASS(vm->vm_nbusytag > 0); 479 vm->vm_inuse -= bt->bt_size; 480 vm->vm_nbusytag--; 481 LIST_REMOVE(bt, bt_hashlist); 482 } 483 484 static void 485 bt_insbusy(vmem_t *vm, bt_t *bt) 486 { 487 struct vmem_hashlist *list; 488 489 VMEM_ASSERT_LOCKED(vm); 490 MPASS(bt->bt_type == BT_TYPE_BUSY); 491 492 list = bt_hashhead(vm, bt->bt_start); 493 LIST_INSERT_HEAD(list, bt, bt_hashlist); 494 vm->vm_nbusytag++; 495 vm->vm_inuse += bt->bt_size; 496 } 497 498 /* ---- boundary tag list */ 499 500 static void 501 bt_remseg(vmem_t *vm, bt_t *bt) 502 { 503 504 MPASS(bt->bt_type != BT_TYPE_CURSOR); 505 TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist); 506 bt_free(vm, bt); 507 } 508 509 static void 510 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev) 511 { 512 513 TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist); 514 } 515 516 static void 517 bt_insseg_tail(vmem_t *vm, bt_t *bt) 518 { 519 520 TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist); 521 } 522 523 static void 524 bt_remfree(vmem_t *vm, bt_t *bt) 525 { 526 527 MPASS(bt->bt_type == BT_TYPE_FREE); 528 529 LIST_REMOVE(bt, bt_freelist); 530 } 531 532 static void 533 bt_insfree(vmem_t *vm, bt_t *bt) 534 { 535 struct vmem_freelist *list; 536 537 list = bt_freehead_tofree(vm, bt->bt_size); 538 LIST_INSERT_HEAD(list, bt, bt_freelist); 539 } 540 541 /* ---- vmem internal functions */ 542 543 /* 544 * Import from the arena into the quantum cache in UMA. 545 * 546 * We use VMEM_ADDR_QCACHE_MIN instead of 0: uma_zalloc() returns 0 to indicate 547 * failure, so UMA can't be used to cache a resource with value 0. 548 */ 549 static int 550 qc_import(void *arg, void **store, int cnt, int domain, int flags) 551 { 552 qcache_t *qc; 553 vmem_addr_t addr; 554 int i; 555 556 KASSERT((flags & M_WAITOK) == 0, ("blocking allocation")); 557 558 qc = arg; 559 for (i = 0; i < cnt; i++) { 560 if (vmem_xalloc(qc->qc_vmem, qc->qc_size, 0, 0, 0, 561 VMEM_ADDR_QCACHE_MIN, VMEM_ADDR_MAX, flags, &addr) != 0) 562 break; 563 store[i] = (void *)addr; 564 } 565 return (i); 566 } 567 568 /* 569 * Release memory from the UMA cache to the arena. 570 */ 571 static void 572 qc_release(void *arg, void **store, int cnt) 573 { 574 qcache_t *qc; 575 int i; 576 577 qc = arg; 578 for (i = 0; i < cnt; i++) 579 vmem_xfree(qc->qc_vmem, (vmem_addr_t)store[i], qc->qc_size); 580 } 581 582 static void 583 qc_init(vmem_t *vm, vmem_size_t qcache_max) 584 { 585 qcache_t *qc; 586 vmem_size_t size; 587 int qcache_idx_max; 588 int i; 589 590 MPASS((qcache_max & vm->vm_quantum_mask) == 0); 591 qcache_idx_max = MIN(qcache_max >> vm->vm_quantum_shift, 592 VMEM_QCACHE_IDX_MAX); 593 vm->vm_qcache_max = qcache_idx_max << vm->vm_quantum_shift; 594 for (i = 0; i < qcache_idx_max; i++) { 595 qc = &vm->vm_qcache[i]; 596 size = (i + 1) << vm->vm_quantum_shift; 597 snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu", 598 vm->vm_name, size); 599 qc->qc_vmem = vm; 600 qc->qc_size = size; 601 qc->qc_cache = uma_zcache_create(qc->qc_name, size, 602 NULL, NULL, NULL, NULL, qc_import, qc_release, qc, 0); 603 MPASS(qc->qc_cache); 604 } 605 } 606 607 static void 608 qc_destroy(vmem_t *vm) 609 { 610 int qcache_idx_max; 611 int i; 612 613 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift; 614 for (i = 0; i < qcache_idx_max; i++) 615 uma_zdestroy(vm->vm_qcache[i].qc_cache); 616 } 617 618 static void 619 qc_drain(vmem_t *vm) 620 { 621 int qcache_idx_max; 622 int i; 623 624 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift; 625 for (i = 0; i < qcache_idx_max; i++) 626 uma_zone_reclaim(vm->vm_qcache[i].qc_cache, UMA_RECLAIM_DRAIN); 627 } 628 629 #ifndef UMA_MD_SMALL_ALLOC 630 631 static struct mtx_padalign __exclusive_cache_line vmem_bt_lock; 632 633 /* 634 * vmem_bt_alloc: Allocate a new page of boundary tags. 635 * 636 * On architectures with uma_small_alloc there is no recursion; no address 637 * space need be allocated to allocate boundary tags. For the others, we 638 * must handle recursion. Boundary tags are necessary to allocate new 639 * boundary tags. 640 * 641 * UMA guarantees that enough tags are held in reserve to allocate a new 642 * page of kva. We dip into this reserve by specifying M_USE_RESERVE only 643 * when allocating the page to hold new boundary tags. In this way the 644 * reserve is automatically filled by the allocation that uses the reserve. 645 * 646 * We still have to guarantee that the new tags are allocated atomically since 647 * many threads may try concurrently. The bt_lock provides this guarantee. 648 * We convert WAITOK allocations to NOWAIT and then handle the blocking here 649 * on failure. It's ok to return NULL for a WAITOK allocation as UMA will 650 * loop again after checking to see if we lost the race to allocate. 651 * 652 * There is a small race between vmem_bt_alloc() returning the page and the 653 * zone lock being acquired to add the page to the zone. For WAITOK 654 * allocations we just pause briefly. NOWAIT may experience a transient 655 * failure. To alleviate this we permit a small number of simultaneous 656 * fills to proceed concurrently so NOWAIT is less likely to fail unless 657 * we are really out of KVA. 658 */ 659 static void * 660 vmem_bt_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 661 int wait) 662 { 663 vmem_addr_t addr; 664 665 *pflag = UMA_SLAB_KERNEL; 666 667 /* 668 * Single thread boundary tag allocation so that the address space 669 * and memory are added in one atomic operation. 670 */ 671 mtx_lock(&vmem_bt_lock); 672 if (vmem_xalloc(vm_dom[domain].vmd_kernel_arena, bytes, 0, 0, 0, 673 VMEM_ADDR_MIN, VMEM_ADDR_MAX, 674 M_NOWAIT | M_NOVM | M_USE_RESERVE | M_BESTFIT, &addr) == 0) { 675 if (kmem_back_domain(domain, kernel_object, addr, bytes, 676 M_NOWAIT | M_USE_RESERVE) == 0) { 677 mtx_unlock(&vmem_bt_lock); 678 return ((void *)addr); 679 } 680 vmem_xfree(vm_dom[domain].vmd_kernel_arena, addr, bytes); 681 mtx_unlock(&vmem_bt_lock); 682 /* 683 * Out of memory, not address space. This may not even be 684 * possible due to M_USE_RESERVE page allocation. 685 */ 686 if (wait & M_WAITOK) 687 vm_wait_domain(domain); 688 return (NULL); 689 } 690 mtx_unlock(&vmem_bt_lock); 691 /* 692 * We're either out of address space or lost a fill race. 693 */ 694 if (wait & M_WAITOK) 695 pause("btalloc", 1); 696 697 return (NULL); 698 } 699 #endif 700 701 void 702 vmem_startup(void) 703 { 704 705 mtx_init(&vmem_list_lock, "vmem list lock", NULL, MTX_DEF); 706 vmem_zone = uma_zcreate("vmem", 707 sizeof(struct vmem), NULL, NULL, NULL, NULL, 708 UMA_ALIGN_PTR, 0); 709 #ifdef UMA_MD_SMALL_ALLOC 710 vmem_bt_zone = uma_zcreate("vmem btag", 711 sizeof(struct vmem_btag), NULL, NULL, NULL, NULL, 712 UMA_ALIGN_PTR, UMA_ZONE_VM); 713 #else 714 vmem_bt_zone = uma_zcreate("vmem btag", 715 sizeof(struct vmem_btag), NULL, NULL, NULL, NULL, 716 UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); 717 mtx_init(&vmem_bt_lock, "btag lock", NULL, MTX_DEF); 718 uma_prealloc(vmem_bt_zone, BT_MAXALLOC); 719 /* 720 * Reserve enough tags to allocate new tags. We allow multiple 721 * CPUs to attempt to allocate new tags concurrently to limit 722 * false restarts in UMA. vmem_bt_alloc() allocates from a per-domain 723 * arena, which may involve importing a range from the kernel arena, 724 * so we need to keep at least 2 * BT_MAXALLOC tags reserved. 725 */ 726 uma_zone_reserve(vmem_bt_zone, 2 * BT_MAXALLOC * mp_ncpus); 727 uma_zone_set_allocf(vmem_bt_zone, vmem_bt_alloc); 728 #endif 729 } 730 731 /* ---- rehash */ 732 733 static int 734 vmem_rehash(vmem_t *vm, vmem_size_t newhashsize) 735 { 736 bt_t *bt; 737 int i; 738 struct vmem_hashlist *newhashlist; 739 struct vmem_hashlist *oldhashlist; 740 vmem_size_t oldhashsize; 741 742 MPASS(newhashsize > 0); 743 744 newhashlist = malloc(sizeof(struct vmem_hashlist) * newhashsize, 745 M_VMEM, M_NOWAIT); 746 if (newhashlist == NULL) 747 return ENOMEM; 748 for (i = 0; i < newhashsize; i++) { 749 LIST_INIT(&newhashlist[i]); 750 } 751 752 VMEM_LOCK(vm); 753 oldhashlist = vm->vm_hashlist; 754 oldhashsize = vm->vm_hashsize; 755 vm->vm_hashlist = newhashlist; 756 vm->vm_hashsize = newhashsize; 757 if (oldhashlist == NULL) { 758 VMEM_UNLOCK(vm); 759 return 0; 760 } 761 for (i = 0; i < oldhashsize; i++) { 762 while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) { 763 bt_rembusy(vm, bt); 764 bt_insbusy(vm, bt); 765 } 766 } 767 VMEM_UNLOCK(vm); 768 769 if (oldhashlist != vm->vm_hash0) { 770 free(oldhashlist, M_VMEM); 771 } 772 773 return 0; 774 } 775 776 static void 777 vmem_periodic_kick(void *dummy) 778 { 779 780 taskqueue_enqueue(taskqueue_thread, &vmem_periodic_wk); 781 } 782 783 static void 784 vmem_periodic(void *unused, int pending) 785 { 786 vmem_t *vm; 787 vmem_size_t desired; 788 vmem_size_t current; 789 790 mtx_lock(&vmem_list_lock); 791 LIST_FOREACH(vm, &vmem_list, vm_alllist) { 792 #ifdef DIAGNOSTIC 793 /* Convenient time to verify vmem state. */ 794 if (enable_vmem_check == 1) { 795 VMEM_LOCK(vm); 796 vmem_check(vm); 797 VMEM_UNLOCK(vm); 798 } 799 #endif 800 desired = 1 << flsl(vm->vm_nbusytag); 801 desired = MIN(MAX(desired, VMEM_HASHSIZE_MIN), 802 VMEM_HASHSIZE_MAX); 803 current = vm->vm_hashsize; 804 805 /* Grow in powers of two. Shrink less aggressively. */ 806 if (desired >= current * 2 || desired * 4 <= current) 807 vmem_rehash(vm, desired); 808 809 /* 810 * Periodically wake up threads waiting for resources, 811 * so they could ask for reclamation again. 812 */ 813 VMEM_CONDVAR_BROADCAST(vm); 814 } 815 mtx_unlock(&vmem_list_lock); 816 817 callout_reset(&vmem_periodic_ch, vmem_periodic_interval, 818 vmem_periodic_kick, NULL); 819 } 820 821 static void 822 vmem_start_callout(void *unused) 823 { 824 825 TASK_INIT(&vmem_periodic_wk, 0, vmem_periodic, NULL); 826 vmem_periodic_interval = hz * 10; 827 callout_init(&vmem_periodic_ch, 1); 828 callout_reset(&vmem_periodic_ch, vmem_periodic_interval, 829 vmem_periodic_kick, NULL); 830 } 831 SYSINIT(vfs, SI_SUB_CONFIGURE, SI_ORDER_ANY, vmem_start_callout, NULL); 832 833 static void 834 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int type) 835 { 836 bt_t *btfree, *btprev, *btspan; 837 838 VMEM_ASSERT_LOCKED(vm); 839 MPASS(type == BT_TYPE_SPAN || type == BT_TYPE_SPAN_STATIC); 840 MPASS((size & vm->vm_quantum_mask) == 0); 841 842 if (vm->vm_releasefn == NULL) { 843 /* 844 * The new segment will never be released, so see if it is 845 * contiguous with respect to an existing segment. In this case 846 * a span tag is not needed, and it may be possible now or in 847 * the future to coalesce the new segment with an existing free 848 * segment. 849 */ 850 btprev = TAILQ_LAST(&vm->vm_seglist, vmem_seglist); 851 if ((!bt_isbusy(btprev) && !bt_isfree(btprev)) || 852 btprev->bt_start + btprev->bt_size != addr) 853 btprev = NULL; 854 } else { 855 btprev = NULL; 856 } 857 858 if (btprev == NULL || bt_isbusy(btprev)) { 859 if (btprev == NULL) { 860 btspan = bt_alloc(vm); 861 btspan->bt_type = type; 862 btspan->bt_start = addr; 863 btspan->bt_size = size; 864 bt_insseg_tail(vm, btspan); 865 } 866 867 btfree = bt_alloc(vm); 868 btfree->bt_type = BT_TYPE_FREE; 869 btfree->bt_start = addr; 870 btfree->bt_size = size; 871 bt_insseg_tail(vm, btfree); 872 bt_insfree(vm, btfree); 873 } else { 874 bt_remfree(vm, btprev); 875 btprev->bt_size += size; 876 bt_insfree(vm, btprev); 877 } 878 879 vm->vm_size += size; 880 } 881 882 static void 883 vmem_destroy1(vmem_t *vm) 884 { 885 bt_t *bt; 886 887 /* 888 * Drain per-cpu quantum caches. 889 */ 890 qc_destroy(vm); 891 892 /* 893 * The vmem should now only contain empty segments. 894 */ 895 VMEM_LOCK(vm); 896 MPASS(vm->vm_nbusytag == 0); 897 898 TAILQ_REMOVE(&vm->vm_seglist, &vm->vm_cursor, bt_seglist); 899 while ((bt = TAILQ_FIRST(&vm->vm_seglist)) != NULL) 900 bt_remseg(vm, bt); 901 902 if (vm->vm_hashlist != NULL && vm->vm_hashlist != vm->vm_hash0) 903 free(vm->vm_hashlist, M_VMEM); 904 905 bt_freetrim(vm, 0); 906 907 VMEM_CONDVAR_DESTROY(vm); 908 VMEM_LOCK_DESTROY(vm); 909 uma_zfree(vmem_zone, vm); 910 } 911 912 static int 913 vmem_import(vmem_t *vm, vmem_size_t size, vmem_size_t align, int flags) 914 { 915 vmem_addr_t addr; 916 int error; 917 918 if (vm->vm_importfn == NULL) 919 return (EINVAL); 920 921 /* 922 * To make sure we get a span that meets the alignment we double it 923 * and add the size to the tail. This slightly overestimates. 924 */ 925 if (align != vm->vm_quantum_mask + 1) 926 size = (align * 2) + size; 927 size = roundup(size, vm->vm_import_quantum); 928 929 if (vm->vm_limit != 0 && vm->vm_limit < vm->vm_size + size) 930 return (ENOMEM); 931 932 bt_save(vm); 933 VMEM_UNLOCK(vm); 934 error = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr); 935 VMEM_LOCK(vm); 936 bt_restore(vm); 937 if (error) 938 return (ENOMEM); 939 940 vmem_add1(vm, addr, size, BT_TYPE_SPAN); 941 942 return 0; 943 } 944 945 /* 946 * vmem_fit: check if a bt can satisfy the given restrictions. 947 * 948 * it's a caller's responsibility to ensure the region is big enough 949 * before calling us. 950 */ 951 static int 952 vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align, 953 vmem_size_t phase, vmem_size_t nocross, vmem_addr_t minaddr, 954 vmem_addr_t maxaddr, vmem_addr_t *addrp) 955 { 956 vmem_addr_t start; 957 vmem_addr_t end; 958 959 MPASS(size > 0); 960 MPASS(bt->bt_size >= size); /* caller's responsibility */ 961 962 /* 963 * XXX assumption: vmem_addr_t and vmem_size_t are 964 * unsigned integer of the same size. 965 */ 966 967 start = bt->bt_start; 968 if (start < minaddr) { 969 start = minaddr; 970 } 971 end = BT_END(bt); 972 if (end > maxaddr) 973 end = maxaddr; 974 if (start > end) 975 return (ENOMEM); 976 977 start = VMEM_ALIGNUP(start - phase, align) + phase; 978 if (start < bt->bt_start) 979 start += align; 980 if (VMEM_CROSS_P(start, start + size - 1, nocross)) { 981 MPASS(align < nocross); 982 start = VMEM_ALIGNUP(start - phase, nocross) + phase; 983 } 984 if (start <= end && end - start >= size - 1) { 985 MPASS((start & (align - 1)) == phase); 986 MPASS(!VMEM_CROSS_P(start, start + size - 1, nocross)); 987 MPASS(minaddr <= start); 988 MPASS(maxaddr == 0 || start + size - 1 <= maxaddr); 989 MPASS(bt->bt_start <= start); 990 MPASS(BT_END(bt) - start >= size - 1); 991 *addrp = start; 992 993 return (0); 994 } 995 return (ENOMEM); 996 } 997 998 /* 999 * vmem_clip: Trim the boundary tag edges to the requested start and size. 1000 */ 1001 static void 1002 vmem_clip(vmem_t *vm, bt_t *bt, vmem_addr_t start, vmem_size_t size) 1003 { 1004 bt_t *btnew; 1005 bt_t *btprev; 1006 1007 VMEM_ASSERT_LOCKED(vm); 1008 MPASS(bt->bt_type == BT_TYPE_FREE); 1009 MPASS(bt->bt_size >= size); 1010 bt_remfree(vm, bt); 1011 if (bt->bt_start != start) { 1012 btprev = bt_alloc(vm); 1013 btprev->bt_type = BT_TYPE_FREE; 1014 btprev->bt_start = bt->bt_start; 1015 btprev->bt_size = start - bt->bt_start; 1016 bt->bt_start = start; 1017 bt->bt_size -= btprev->bt_size; 1018 bt_insfree(vm, btprev); 1019 bt_insseg(vm, btprev, 1020 TAILQ_PREV(bt, vmem_seglist, bt_seglist)); 1021 } 1022 MPASS(bt->bt_start == start); 1023 if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) { 1024 /* split */ 1025 btnew = bt_alloc(vm); 1026 btnew->bt_type = BT_TYPE_BUSY; 1027 btnew->bt_start = bt->bt_start; 1028 btnew->bt_size = size; 1029 bt->bt_start = bt->bt_start + size; 1030 bt->bt_size -= size; 1031 bt_insfree(vm, bt); 1032 bt_insseg(vm, btnew, 1033 TAILQ_PREV(bt, vmem_seglist, bt_seglist)); 1034 bt_insbusy(vm, btnew); 1035 bt = btnew; 1036 } else { 1037 bt->bt_type = BT_TYPE_BUSY; 1038 bt_insbusy(vm, bt); 1039 } 1040 MPASS(bt->bt_size >= size); 1041 } 1042 1043 static int 1044 vmem_try_fetch(vmem_t *vm, const vmem_size_t size, vmem_size_t align, int flags) 1045 { 1046 vmem_size_t avail; 1047 1048 VMEM_ASSERT_LOCKED(vm); 1049 1050 /* 1051 * XXX it is possible to fail to meet xalloc constraints with the 1052 * imported region. It is up to the user to specify the 1053 * import quantum such that it can satisfy any allocation. 1054 */ 1055 if (vmem_import(vm, size, align, flags) == 0) 1056 return (1); 1057 1058 /* 1059 * Try to free some space from the quantum cache or reclaim 1060 * functions if available. 1061 */ 1062 if (vm->vm_qcache_max != 0 || vm->vm_reclaimfn != NULL) { 1063 avail = vm->vm_size - vm->vm_inuse; 1064 bt_save(vm); 1065 VMEM_UNLOCK(vm); 1066 if (vm->vm_qcache_max != 0) 1067 qc_drain(vm); 1068 if (vm->vm_reclaimfn != NULL) 1069 vm->vm_reclaimfn(vm, flags); 1070 VMEM_LOCK(vm); 1071 bt_restore(vm); 1072 /* If we were successful retry even NOWAIT. */ 1073 if (vm->vm_size - vm->vm_inuse > avail) 1074 return (1); 1075 } 1076 if ((flags & M_NOWAIT) != 0) 1077 return (0); 1078 bt_save(vm); 1079 VMEM_CONDVAR_WAIT(vm); 1080 bt_restore(vm); 1081 return (1); 1082 } 1083 1084 static int 1085 vmem_try_release(vmem_t *vm, struct vmem_btag *bt, const bool remfree) 1086 { 1087 struct vmem_btag *prev; 1088 1089 MPASS(bt->bt_type == BT_TYPE_FREE); 1090 1091 if (vm->vm_releasefn == NULL) 1092 return (0); 1093 1094 prev = TAILQ_PREV(bt, vmem_seglist, bt_seglist); 1095 MPASS(prev != NULL); 1096 MPASS(prev->bt_type != BT_TYPE_FREE); 1097 1098 if (prev->bt_type == BT_TYPE_SPAN && prev->bt_size == bt->bt_size) { 1099 vmem_addr_t spanaddr; 1100 vmem_size_t spansize; 1101 1102 MPASS(prev->bt_start == bt->bt_start); 1103 spanaddr = prev->bt_start; 1104 spansize = prev->bt_size; 1105 if (remfree) 1106 bt_remfree(vm, bt); 1107 bt_remseg(vm, bt); 1108 bt_remseg(vm, prev); 1109 vm->vm_size -= spansize; 1110 VMEM_CONDVAR_BROADCAST(vm); 1111 bt_freetrim(vm, BT_MAXFREE); 1112 vm->vm_releasefn(vm->vm_arg, spanaddr, spansize); 1113 return (1); 1114 } 1115 return (0); 1116 } 1117 1118 static int 1119 vmem_xalloc_nextfit(vmem_t *vm, const vmem_size_t size, vmem_size_t align, 1120 const vmem_size_t phase, const vmem_size_t nocross, int flags, 1121 vmem_addr_t *addrp) 1122 { 1123 struct vmem_btag *bt, *cursor, *next, *prev; 1124 int error; 1125 1126 error = ENOMEM; 1127 VMEM_LOCK(vm); 1128 1129 /* 1130 * Make sure we have enough tags to complete the operation. 1131 */ 1132 if (bt_fill(vm, flags) != 0) 1133 goto out; 1134 1135 retry: 1136 /* 1137 * Find the next free tag meeting our constraints. If one is found, 1138 * perform the allocation. 1139 */ 1140 for (cursor = &vm->vm_cursor, bt = TAILQ_NEXT(cursor, bt_seglist); 1141 bt != cursor; bt = TAILQ_NEXT(bt, bt_seglist)) { 1142 if (bt == NULL) 1143 bt = TAILQ_FIRST(&vm->vm_seglist); 1144 if (bt->bt_type == BT_TYPE_FREE && bt->bt_size >= size && 1145 (error = vmem_fit(bt, size, align, phase, nocross, 1146 VMEM_ADDR_MIN, VMEM_ADDR_MAX, addrp)) == 0) { 1147 vmem_clip(vm, bt, *addrp, size); 1148 break; 1149 } 1150 } 1151 1152 /* 1153 * Try to coalesce free segments around the cursor. If we succeed, and 1154 * have not yet satisfied the allocation request, try again with the 1155 * newly coalesced segment. 1156 */ 1157 if ((next = TAILQ_NEXT(cursor, bt_seglist)) != NULL && 1158 (prev = TAILQ_PREV(cursor, vmem_seglist, bt_seglist)) != NULL && 1159 next->bt_type == BT_TYPE_FREE && prev->bt_type == BT_TYPE_FREE && 1160 prev->bt_start + prev->bt_size == next->bt_start) { 1161 prev->bt_size += next->bt_size; 1162 bt_remfree(vm, next); 1163 bt_remseg(vm, next); 1164 1165 /* 1166 * The coalesced segment might be able to satisfy our request. 1167 * If not, we might need to release it from the arena. 1168 */ 1169 if (error == ENOMEM && prev->bt_size >= size && 1170 (error = vmem_fit(prev, size, align, phase, nocross, 1171 VMEM_ADDR_MIN, VMEM_ADDR_MAX, addrp)) == 0) { 1172 vmem_clip(vm, prev, *addrp, size); 1173 bt = prev; 1174 } else 1175 (void)vmem_try_release(vm, prev, true); 1176 } 1177 1178 /* 1179 * If the allocation was successful, advance the cursor. 1180 */ 1181 if (error == 0) { 1182 TAILQ_REMOVE(&vm->vm_seglist, cursor, bt_seglist); 1183 for (; bt != NULL && bt->bt_start < *addrp + size; 1184 bt = TAILQ_NEXT(bt, bt_seglist)) 1185 ; 1186 if (bt != NULL) 1187 TAILQ_INSERT_BEFORE(bt, cursor, bt_seglist); 1188 else 1189 TAILQ_INSERT_HEAD(&vm->vm_seglist, cursor, bt_seglist); 1190 } 1191 1192 /* 1193 * Attempt to bring additional resources into the arena. If that fails 1194 * and M_WAITOK is specified, sleep waiting for resources to be freed. 1195 */ 1196 if (error == ENOMEM && vmem_try_fetch(vm, size, align, flags)) 1197 goto retry; 1198 1199 out: 1200 VMEM_UNLOCK(vm); 1201 return (error); 1202 } 1203 1204 /* ---- vmem API */ 1205 1206 void 1207 vmem_set_import(vmem_t *vm, vmem_import_t *importfn, 1208 vmem_release_t *releasefn, void *arg, vmem_size_t import_quantum) 1209 { 1210 1211 VMEM_LOCK(vm); 1212 KASSERT(vm->vm_size == 0, ("%s: arena is non-empty", __func__)); 1213 vm->vm_importfn = importfn; 1214 vm->vm_releasefn = releasefn; 1215 vm->vm_arg = arg; 1216 vm->vm_import_quantum = import_quantum; 1217 VMEM_UNLOCK(vm); 1218 } 1219 1220 void 1221 vmem_set_limit(vmem_t *vm, vmem_size_t limit) 1222 { 1223 1224 VMEM_LOCK(vm); 1225 vm->vm_limit = limit; 1226 VMEM_UNLOCK(vm); 1227 } 1228 1229 void 1230 vmem_set_reclaim(vmem_t *vm, vmem_reclaim_t *reclaimfn) 1231 { 1232 1233 VMEM_LOCK(vm); 1234 vm->vm_reclaimfn = reclaimfn; 1235 VMEM_UNLOCK(vm); 1236 } 1237 1238 /* 1239 * vmem_init: Initializes vmem arena. 1240 */ 1241 vmem_t * 1242 vmem_init(vmem_t *vm, const char *name, vmem_addr_t base, vmem_size_t size, 1243 vmem_size_t quantum, vmem_size_t qcache_max, int flags) 1244 { 1245 int i; 1246 1247 MPASS(quantum > 0); 1248 MPASS((quantum & (quantum - 1)) == 0); 1249 1250 bzero(vm, sizeof(*vm)); 1251 1252 VMEM_CONDVAR_INIT(vm, name); 1253 VMEM_LOCK_INIT(vm, name); 1254 vm->vm_nfreetags = 0; 1255 LIST_INIT(&vm->vm_freetags); 1256 strlcpy(vm->vm_name, name, sizeof(vm->vm_name)); 1257 vm->vm_quantum_mask = quantum - 1; 1258 vm->vm_quantum_shift = flsl(quantum) - 1; 1259 vm->vm_nbusytag = 0; 1260 vm->vm_size = 0; 1261 vm->vm_limit = 0; 1262 vm->vm_inuse = 0; 1263 qc_init(vm, qcache_max); 1264 1265 TAILQ_INIT(&vm->vm_seglist); 1266 vm->vm_cursor.bt_start = vm->vm_cursor.bt_size = 0; 1267 vm->vm_cursor.bt_type = BT_TYPE_CURSOR; 1268 TAILQ_INSERT_TAIL(&vm->vm_seglist, &vm->vm_cursor, bt_seglist); 1269 1270 for (i = 0; i < VMEM_MAXORDER; i++) 1271 LIST_INIT(&vm->vm_freelist[i]); 1272 1273 memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0)); 1274 vm->vm_hashsize = VMEM_HASHSIZE_MIN; 1275 vm->vm_hashlist = vm->vm_hash0; 1276 1277 if (size != 0) { 1278 if (vmem_add(vm, base, size, flags) != 0) { 1279 vmem_destroy1(vm); 1280 return NULL; 1281 } 1282 } 1283 1284 mtx_lock(&vmem_list_lock); 1285 LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist); 1286 mtx_unlock(&vmem_list_lock); 1287 1288 return vm; 1289 } 1290 1291 /* 1292 * vmem_create: create an arena. 1293 */ 1294 vmem_t * 1295 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size, 1296 vmem_size_t quantum, vmem_size_t qcache_max, int flags) 1297 { 1298 1299 vmem_t *vm; 1300 1301 vm = uma_zalloc(vmem_zone, flags & (M_WAITOK|M_NOWAIT)); 1302 if (vm == NULL) 1303 return (NULL); 1304 if (vmem_init(vm, name, base, size, quantum, qcache_max, 1305 flags) == NULL) 1306 return (NULL); 1307 return (vm); 1308 } 1309 1310 void 1311 vmem_destroy(vmem_t *vm) 1312 { 1313 1314 mtx_lock(&vmem_list_lock); 1315 LIST_REMOVE(vm, vm_alllist); 1316 mtx_unlock(&vmem_list_lock); 1317 1318 vmem_destroy1(vm); 1319 } 1320 1321 vmem_size_t 1322 vmem_roundup_size(vmem_t *vm, vmem_size_t size) 1323 { 1324 1325 return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask; 1326 } 1327 1328 /* 1329 * vmem_alloc: allocate resource from the arena. 1330 */ 1331 int 1332 vmem_alloc(vmem_t *vm, vmem_size_t size, int flags, vmem_addr_t *addrp) 1333 { 1334 const int strat __unused = flags & VMEM_FITMASK; 1335 qcache_t *qc; 1336 1337 flags &= VMEM_FLAGS; 1338 MPASS(size > 0); 1339 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT || strat == M_NEXTFIT); 1340 if ((flags & M_NOWAIT) == 0) 1341 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_alloc"); 1342 1343 if (size <= vm->vm_qcache_max) { 1344 /* 1345 * Resource 0 cannot be cached, so avoid a blocking allocation 1346 * in qc_import() and give the vmem_xalloc() call below a chance 1347 * to return 0. 1348 */ 1349 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift]; 1350 *addrp = (vmem_addr_t)uma_zalloc(qc->qc_cache, 1351 (flags & ~M_WAITOK) | M_NOWAIT); 1352 if (__predict_true(*addrp != 0)) 1353 return (0); 1354 } 1355 1356 return (vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, 1357 flags, addrp)); 1358 } 1359 1360 int 1361 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align, 1362 const vmem_size_t phase, const vmem_size_t nocross, 1363 const vmem_addr_t minaddr, const vmem_addr_t maxaddr, int flags, 1364 vmem_addr_t *addrp) 1365 { 1366 const vmem_size_t size = vmem_roundup_size(vm, size0); 1367 struct vmem_freelist *list; 1368 struct vmem_freelist *first; 1369 struct vmem_freelist *end; 1370 bt_t *bt; 1371 int error; 1372 int strat; 1373 1374 flags &= VMEM_FLAGS; 1375 strat = flags & VMEM_FITMASK; 1376 MPASS(size0 > 0); 1377 MPASS(size > 0); 1378 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT || strat == M_NEXTFIT); 1379 MPASS((flags & (M_NOWAIT|M_WAITOK)) != (M_NOWAIT|M_WAITOK)); 1380 if ((flags & M_NOWAIT) == 0) 1381 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_xalloc"); 1382 MPASS((align & vm->vm_quantum_mask) == 0); 1383 MPASS((align & (align - 1)) == 0); 1384 MPASS((phase & vm->vm_quantum_mask) == 0); 1385 MPASS((nocross & vm->vm_quantum_mask) == 0); 1386 MPASS((nocross & (nocross - 1)) == 0); 1387 MPASS((align == 0 && phase == 0) || phase < align); 1388 MPASS(nocross == 0 || nocross >= size); 1389 MPASS(minaddr <= maxaddr); 1390 MPASS(!VMEM_CROSS_P(phase, phase + size - 1, nocross)); 1391 if (strat == M_NEXTFIT) 1392 MPASS(minaddr == VMEM_ADDR_MIN && maxaddr == VMEM_ADDR_MAX); 1393 1394 if (align == 0) 1395 align = vm->vm_quantum_mask + 1; 1396 *addrp = 0; 1397 1398 /* 1399 * Next-fit allocations don't use the freelists. 1400 */ 1401 if (strat == M_NEXTFIT) 1402 return (vmem_xalloc_nextfit(vm, size0, align, phase, nocross, 1403 flags, addrp)); 1404 1405 end = &vm->vm_freelist[VMEM_MAXORDER]; 1406 /* 1407 * choose a free block from which we allocate. 1408 */ 1409 first = bt_freehead_toalloc(vm, size, strat); 1410 VMEM_LOCK(vm); 1411 1412 /* 1413 * Make sure we have enough tags to complete the operation. 1414 */ 1415 error = bt_fill(vm, flags); 1416 if (error != 0) 1417 goto out; 1418 for (;;) { 1419 /* 1420 * Scan freelists looking for a tag that satisfies the 1421 * allocation. If we're doing BESTFIT we may encounter 1422 * sizes below the request. If we're doing FIRSTFIT we 1423 * inspect only the first element from each list. 1424 */ 1425 for (list = first; list < end; list++) { 1426 LIST_FOREACH(bt, list, bt_freelist) { 1427 if (bt->bt_size >= size) { 1428 error = vmem_fit(bt, size, align, phase, 1429 nocross, minaddr, maxaddr, addrp); 1430 if (error == 0) { 1431 vmem_clip(vm, bt, *addrp, size); 1432 goto out; 1433 } 1434 } 1435 /* FIRST skips to the next list. */ 1436 if (strat == M_FIRSTFIT) 1437 break; 1438 } 1439 } 1440 1441 /* 1442 * Retry if the fast algorithm failed. 1443 */ 1444 if (strat == M_FIRSTFIT) { 1445 strat = M_BESTFIT; 1446 first = bt_freehead_toalloc(vm, size, strat); 1447 continue; 1448 } 1449 1450 /* 1451 * Try a few measures to bring additional resources into the 1452 * arena. If all else fails, we will sleep waiting for 1453 * resources to be freed. 1454 */ 1455 if (!vmem_try_fetch(vm, size, align, flags)) { 1456 error = ENOMEM; 1457 break; 1458 } 1459 } 1460 out: 1461 VMEM_UNLOCK(vm); 1462 if (error != 0 && (flags & M_NOWAIT) == 0) 1463 panic("failed to allocate waiting allocation\n"); 1464 1465 return (error); 1466 } 1467 1468 /* 1469 * vmem_free: free the resource to the arena. 1470 */ 1471 void 1472 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size) 1473 { 1474 qcache_t *qc; 1475 MPASS(size > 0); 1476 1477 if (size <= vm->vm_qcache_max && 1478 __predict_true(addr >= VMEM_ADDR_QCACHE_MIN)) { 1479 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift]; 1480 uma_zfree(qc->qc_cache, (void *)addr); 1481 } else 1482 vmem_xfree(vm, addr, size); 1483 } 1484 1485 void 1486 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size) 1487 { 1488 bt_t *bt; 1489 bt_t *t; 1490 1491 MPASS(size > 0); 1492 1493 VMEM_LOCK(vm); 1494 bt = bt_lookupbusy(vm, addr); 1495 MPASS(bt != NULL); 1496 MPASS(bt->bt_start == addr); 1497 MPASS(bt->bt_size == vmem_roundup_size(vm, size) || 1498 bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask); 1499 MPASS(bt->bt_type == BT_TYPE_BUSY); 1500 bt_rembusy(vm, bt); 1501 bt->bt_type = BT_TYPE_FREE; 1502 1503 /* coalesce */ 1504 t = TAILQ_NEXT(bt, bt_seglist); 1505 if (t != NULL && t->bt_type == BT_TYPE_FREE) { 1506 MPASS(BT_END(bt) < t->bt_start); /* YYY */ 1507 bt->bt_size += t->bt_size; 1508 bt_remfree(vm, t); 1509 bt_remseg(vm, t); 1510 } 1511 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist); 1512 if (t != NULL && t->bt_type == BT_TYPE_FREE) { 1513 MPASS(BT_END(t) < bt->bt_start); /* YYY */ 1514 bt->bt_size += t->bt_size; 1515 bt->bt_start = t->bt_start; 1516 bt_remfree(vm, t); 1517 bt_remseg(vm, t); 1518 } 1519 1520 if (!vmem_try_release(vm, bt, false)) { 1521 bt_insfree(vm, bt); 1522 VMEM_CONDVAR_BROADCAST(vm); 1523 bt_freetrim(vm, BT_MAXFREE); 1524 } 1525 } 1526 1527 /* 1528 * vmem_add: 1529 * 1530 */ 1531 int 1532 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags) 1533 { 1534 int error; 1535 1536 flags &= VMEM_FLAGS; 1537 1538 VMEM_LOCK(vm); 1539 error = bt_fill(vm, flags); 1540 if (error == 0) 1541 vmem_add1(vm, addr, size, BT_TYPE_SPAN_STATIC); 1542 VMEM_UNLOCK(vm); 1543 1544 return (error); 1545 } 1546 1547 /* 1548 * vmem_size: information about arenas size 1549 */ 1550 vmem_size_t 1551 vmem_size(vmem_t *vm, int typemask) 1552 { 1553 int i; 1554 1555 switch (typemask) { 1556 case VMEM_ALLOC: 1557 return vm->vm_inuse; 1558 case VMEM_FREE: 1559 return vm->vm_size - vm->vm_inuse; 1560 case VMEM_FREE|VMEM_ALLOC: 1561 return vm->vm_size; 1562 case VMEM_MAXFREE: 1563 VMEM_LOCK(vm); 1564 for (i = VMEM_MAXORDER - 1; i >= 0; i--) { 1565 if (LIST_EMPTY(&vm->vm_freelist[i])) 1566 continue; 1567 VMEM_UNLOCK(vm); 1568 return ((vmem_size_t)ORDER2SIZE(i) << 1569 vm->vm_quantum_shift); 1570 } 1571 VMEM_UNLOCK(vm); 1572 return (0); 1573 default: 1574 panic("vmem_size"); 1575 } 1576 } 1577 1578 /* ---- debug */ 1579 1580 #if defined(DDB) || defined(DIAGNOSTIC) 1581 1582 static void bt_dump(const bt_t *, int (*)(const char *, ...) 1583 __printflike(1, 2)); 1584 1585 static const char * 1586 bt_type_string(int type) 1587 { 1588 1589 switch (type) { 1590 case BT_TYPE_BUSY: 1591 return "busy"; 1592 case BT_TYPE_FREE: 1593 return "free"; 1594 case BT_TYPE_SPAN: 1595 return "span"; 1596 case BT_TYPE_SPAN_STATIC: 1597 return "static span"; 1598 case BT_TYPE_CURSOR: 1599 return "cursor"; 1600 default: 1601 break; 1602 } 1603 return "BOGUS"; 1604 } 1605 1606 static void 1607 bt_dump(const bt_t *bt, int (*pr)(const char *, ...)) 1608 { 1609 1610 (*pr)("\t%p: %jx %jx, %d(%s)\n", 1611 bt, (intmax_t)bt->bt_start, (intmax_t)bt->bt_size, 1612 bt->bt_type, bt_type_string(bt->bt_type)); 1613 } 1614 1615 static void 1616 vmem_dump(const vmem_t *vm , int (*pr)(const char *, ...) __printflike(1, 2)) 1617 { 1618 const bt_t *bt; 1619 int i; 1620 1621 (*pr)("vmem %p '%s'\n", vm, vm->vm_name); 1622 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1623 bt_dump(bt, pr); 1624 } 1625 1626 for (i = 0; i < VMEM_MAXORDER; i++) { 1627 const struct vmem_freelist *fl = &vm->vm_freelist[i]; 1628 1629 if (LIST_EMPTY(fl)) { 1630 continue; 1631 } 1632 1633 (*pr)("freelist[%d]\n", i); 1634 LIST_FOREACH(bt, fl, bt_freelist) { 1635 bt_dump(bt, pr); 1636 } 1637 } 1638 } 1639 1640 #endif /* defined(DDB) || defined(DIAGNOSTIC) */ 1641 1642 #if defined(DDB) 1643 #include <ddb/ddb.h> 1644 1645 static bt_t * 1646 vmem_whatis_lookup(vmem_t *vm, vmem_addr_t addr) 1647 { 1648 bt_t *bt; 1649 1650 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1651 if (BT_ISSPAN_P(bt)) { 1652 continue; 1653 } 1654 if (bt->bt_start <= addr && addr <= BT_END(bt)) { 1655 return bt; 1656 } 1657 } 1658 1659 return NULL; 1660 } 1661 1662 void 1663 vmem_whatis(vmem_addr_t addr, int (*pr)(const char *, ...)) 1664 { 1665 vmem_t *vm; 1666 1667 LIST_FOREACH(vm, &vmem_list, vm_alllist) { 1668 bt_t *bt; 1669 1670 bt = vmem_whatis_lookup(vm, addr); 1671 if (bt == NULL) { 1672 continue; 1673 } 1674 (*pr)("%p is %p+%zu in VMEM '%s' (%s)\n", 1675 (void *)addr, (void *)bt->bt_start, 1676 (vmem_size_t)(addr - bt->bt_start), vm->vm_name, 1677 (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free"); 1678 } 1679 } 1680 1681 void 1682 vmem_printall(const char *modif, int (*pr)(const char *, ...)) 1683 { 1684 const vmem_t *vm; 1685 1686 LIST_FOREACH(vm, &vmem_list, vm_alllist) { 1687 vmem_dump(vm, pr); 1688 } 1689 } 1690 1691 void 1692 vmem_print(vmem_addr_t addr, const char *modif, int (*pr)(const char *, ...)) 1693 { 1694 const vmem_t *vm = (const void *)addr; 1695 1696 vmem_dump(vm, pr); 1697 } 1698 1699 DB_SHOW_COMMAND(vmemdump, vmemdump) 1700 { 1701 1702 if (!have_addr) { 1703 db_printf("usage: show vmemdump <addr>\n"); 1704 return; 1705 } 1706 1707 vmem_dump((const vmem_t *)addr, db_printf); 1708 } 1709 1710 DB_SHOW_ALL_COMMAND(vmemdump, vmemdumpall) 1711 { 1712 const vmem_t *vm; 1713 1714 LIST_FOREACH(vm, &vmem_list, vm_alllist) 1715 vmem_dump(vm, db_printf); 1716 } 1717 1718 DB_SHOW_COMMAND(vmem, vmem_summ) 1719 { 1720 const vmem_t *vm = (const void *)addr; 1721 const bt_t *bt; 1722 size_t ft[VMEM_MAXORDER], ut[VMEM_MAXORDER]; 1723 size_t fs[VMEM_MAXORDER], us[VMEM_MAXORDER]; 1724 int ord; 1725 1726 if (!have_addr) { 1727 db_printf("usage: show vmem <addr>\n"); 1728 return; 1729 } 1730 1731 db_printf("vmem %p '%s'\n", vm, vm->vm_name); 1732 db_printf("\tquantum:\t%zu\n", vm->vm_quantum_mask + 1); 1733 db_printf("\tsize:\t%zu\n", vm->vm_size); 1734 db_printf("\tinuse:\t%zu\n", vm->vm_inuse); 1735 db_printf("\tfree:\t%zu\n", vm->vm_size - vm->vm_inuse); 1736 db_printf("\tbusy tags:\t%d\n", vm->vm_nbusytag); 1737 db_printf("\tfree tags:\t%d\n", vm->vm_nfreetags); 1738 1739 memset(&ft, 0, sizeof(ft)); 1740 memset(&ut, 0, sizeof(ut)); 1741 memset(&fs, 0, sizeof(fs)); 1742 memset(&us, 0, sizeof(us)); 1743 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1744 ord = SIZE2ORDER(bt->bt_size >> vm->vm_quantum_shift); 1745 if (bt->bt_type == BT_TYPE_BUSY) { 1746 ut[ord]++; 1747 us[ord] += bt->bt_size; 1748 } else if (bt->bt_type == BT_TYPE_FREE) { 1749 ft[ord]++; 1750 fs[ord] += bt->bt_size; 1751 } 1752 } 1753 db_printf("\t\t\tinuse\tsize\t\tfree\tsize\n"); 1754 for (ord = 0; ord < VMEM_MAXORDER; ord++) { 1755 if (ut[ord] == 0 && ft[ord] == 0) 1756 continue; 1757 db_printf("\t%-15zu %zu\t%-15zu %zu\t%-16zu\n", 1758 ORDER2SIZE(ord) << vm->vm_quantum_shift, 1759 ut[ord], us[ord], ft[ord], fs[ord]); 1760 } 1761 } 1762 1763 DB_SHOW_ALL_COMMAND(vmem, vmem_summall) 1764 { 1765 const vmem_t *vm; 1766 1767 LIST_FOREACH(vm, &vmem_list, vm_alllist) 1768 vmem_summ((db_expr_t)vm, TRUE, count, modif); 1769 } 1770 #endif /* defined(DDB) */ 1771 1772 #define vmem_printf printf 1773 1774 #if defined(DIAGNOSTIC) 1775 1776 static bool 1777 vmem_check_sanity(vmem_t *vm) 1778 { 1779 const bt_t *bt, *bt2; 1780 1781 MPASS(vm != NULL); 1782 1783 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1784 if (bt->bt_start > BT_END(bt)) { 1785 printf("corrupted tag\n"); 1786 bt_dump(bt, vmem_printf); 1787 return false; 1788 } 1789 } 1790 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) { 1791 if (bt->bt_type == BT_TYPE_CURSOR) { 1792 if (bt->bt_start != 0 || bt->bt_size != 0) { 1793 printf("corrupted cursor\n"); 1794 return false; 1795 } 1796 continue; 1797 } 1798 TAILQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) { 1799 if (bt == bt2) { 1800 continue; 1801 } 1802 if (bt2->bt_type == BT_TYPE_CURSOR) { 1803 continue; 1804 } 1805 if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) { 1806 continue; 1807 } 1808 if (bt->bt_start <= BT_END(bt2) && 1809 bt2->bt_start <= BT_END(bt)) { 1810 printf("overwrapped tags\n"); 1811 bt_dump(bt, vmem_printf); 1812 bt_dump(bt2, vmem_printf); 1813 return false; 1814 } 1815 } 1816 } 1817 1818 return true; 1819 } 1820 1821 static void 1822 vmem_check(vmem_t *vm) 1823 { 1824 1825 if (!vmem_check_sanity(vm)) { 1826 panic("insanity vmem %p", vm); 1827 } 1828 } 1829 1830 #endif /* defined(DIAGNOSTIC) */ 1831