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