vm_reserv.c revision 269072
1/*- 2 * Copyright (c) 2002-2006 Rice University 3 * Copyright (c) 2007-2008 Alan L. Cox <alc@cs.rice.edu> 4 * All rights reserved. 5 * 6 * This software was developed for the FreeBSD Project by Alan L. Cox, 7 * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 19 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 20 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 21 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 22 * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 24 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS 25 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 26 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY 28 * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 29 * POSSIBILITY OF SUCH DAMAGE. 30 */ 31 32/* 33 * Superpage reservation management module 34 * 35 * Any external functions defined by this module are only to be used by the 36 * virtual memory system. 37 */ 38 39#include <sys/cdefs.h> 40__FBSDID("$FreeBSD: stable/10/sys/vm/vm_reserv.c 269072 2014-07-24 16:29:44Z kib $"); 41 42#include "opt_vm.h" 43 44#include <sys/param.h> 45#include <sys/kernel.h> 46#include <sys/lock.h> 47#include <sys/malloc.h> 48#include <sys/mutex.h> 49#include <sys/queue.h> 50#include <sys/rwlock.h> 51#include <sys/sbuf.h> 52#include <sys/sysctl.h> 53#include <sys/systm.h> 54 55#include <vm/vm.h> 56#include <vm/vm_param.h> 57#include <vm/vm_object.h> 58#include <vm/vm_page.h> 59#include <vm/vm_phys.h> 60#include <vm/vm_radix.h> 61#include <vm/vm_reserv.h> 62 63/* 64 * The reservation system supports the speculative allocation of large physical 65 * pages ("superpages"). Speculative allocation enables the fully-automatic 66 * utilization of superpages by the virtual memory system. In other words, no 67 * programmatic directives are required to use superpages. 68 */ 69 70#if VM_NRESERVLEVEL > 0 71 72/* 73 * The number of small pages that are contained in a level 0 reservation 74 */ 75#define VM_LEVEL_0_NPAGES (1 << VM_LEVEL_0_ORDER) 76 77/* 78 * The number of bits by which a physical address is shifted to obtain the 79 * reservation number 80 */ 81#define VM_LEVEL_0_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT) 82 83/* 84 * The size of a level 0 reservation in bytes 85 */ 86#define VM_LEVEL_0_SIZE (1 << VM_LEVEL_0_SHIFT) 87 88/* 89 * Computes the index of the small page underlying the given (object, pindex) 90 * within the reservation's array of small pages. 91 */ 92#define VM_RESERV_INDEX(object, pindex) \ 93 (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1)) 94 95/* 96 * The reservation structure 97 * 98 * A reservation structure is constructed whenever a large physical page is 99 * speculatively allocated to an object. The reservation provides the small 100 * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets 101 * within that object. The reservation's "popcnt" tracks the number of these 102 * small physical pages that are in use at any given time. When and if the 103 * reservation is not fully utilized, it appears in the queue of partially- 104 * populated reservations. The reservation always appears on the containing 105 * object's list of reservations. 106 * 107 * A partially-populated reservation can be broken and reclaimed at any time. 108 */ 109struct vm_reserv { 110 TAILQ_ENTRY(vm_reserv) partpopq; 111 LIST_ENTRY(vm_reserv) objq; 112 vm_object_t object; /* containing object */ 113 vm_pindex_t pindex; /* offset within object */ 114 vm_page_t pages; /* first page of a superpage */ 115 int popcnt; /* # of pages in use */ 116 char inpartpopq; 117}; 118 119/* 120 * The reservation array 121 * 122 * This array is analoguous in function to vm_page_array. It differs in the 123 * respect that it may contain a greater number of useful reservation 124 * structures than there are (physical) superpages. These "invalid" 125 * reservation structures exist to trade-off space for time in the 126 * implementation of vm_reserv_from_page(). Invalid reservation structures are 127 * distinguishable from "valid" reservation structures by inspecting the 128 * reservation's "pages" field. Invalid reservation structures have a NULL 129 * "pages" field. 130 * 131 * vm_reserv_from_page() maps a small (physical) page to an element of this 132 * array by computing a physical reservation number from the page's physical 133 * address. The physical reservation number is used as the array index. 134 * 135 * An "active" reservation is a valid reservation structure that has a non-NULL 136 * "object" field and a non-zero "popcnt" field. In other words, every active 137 * reservation belongs to a particular object. Moreover, every active 138 * reservation has an entry in the containing object's list of reservations. 139 */ 140static vm_reserv_t vm_reserv_array; 141 142/* 143 * The partially-populated reservation queue 144 * 145 * This queue enables the fast recovery of an unused cached or free small page 146 * from a partially-populated reservation. The reservation at the head of 147 * this queue is the least-recently-changed, partially-populated reservation. 148 * 149 * Access to this queue is synchronized by the free page queue lock. 150 */ 151static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop = 152 TAILQ_HEAD_INITIALIZER(vm_rvq_partpop); 153 154static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info"); 155 156static long vm_reserv_broken; 157SYSCTL_LONG(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD, 158 &vm_reserv_broken, 0, "Cumulative number of broken reservations"); 159 160static long vm_reserv_freed; 161SYSCTL_LONG(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD, 162 &vm_reserv_freed, 0, "Cumulative number of freed reservations"); 163 164static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS); 165 166SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0, 167 sysctl_vm_reserv_partpopq, "A", "Partially-populated reservation queues"); 168 169static long vm_reserv_reclaimed; 170SYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD, 171 &vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations"); 172 173static void vm_reserv_depopulate(vm_reserv_t rv); 174static vm_reserv_t vm_reserv_from_page(vm_page_t m); 175static boolean_t vm_reserv_has_pindex(vm_reserv_t rv, 176 vm_pindex_t pindex); 177static void vm_reserv_populate(vm_reserv_t rv); 178static void vm_reserv_reclaim(vm_reserv_t rv); 179 180/* 181 * Describes the current state of the partially-populated reservation queue. 182 */ 183static int 184sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS) 185{ 186 struct sbuf sbuf; 187 vm_reserv_t rv; 188 int counter, error, level, unused_pages; 189 190 error = sysctl_wire_old_buffer(req, 0); 191 if (error != 0) 192 return (error); 193 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 194 sbuf_printf(&sbuf, "\nLEVEL SIZE NUMBER\n\n"); 195 for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) { 196 counter = 0; 197 unused_pages = 0; 198 mtx_lock(&vm_page_queue_free_mtx); 199 TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) { 200 counter++; 201 unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt; 202 } 203 mtx_unlock(&vm_page_queue_free_mtx); 204 sbuf_printf(&sbuf, "%5d: %6dK, %6d\n", level, 205 unused_pages * ((int)PAGE_SIZE / 1024), counter); 206 } 207 error = sbuf_finish(&sbuf); 208 sbuf_delete(&sbuf); 209 return (error); 210} 211 212/* 213 * Reduces the given reservation's population count. If the population count 214 * becomes zero, the reservation is destroyed. Additionally, moves the 215 * reservation to the tail of the partially-populated reservations queue if the 216 * population count is non-zero. 217 * 218 * The free page queue lock must be held. 219 */ 220static void 221vm_reserv_depopulate(vm_reserv_t rv) 222{ 223 224 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 225 KASSERT(rv->object != NULL, 226 ("vm_reserv_depopulate: reserv %p is free", rv)); 227 KASSERT(rv->popcnt > 0, 228 ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv)); 229 if (rv->inpartpopq) { 230 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq); 231 rv->inpartpopq = FALSE; 232 } else { 233 KASSERT(rv->pages->psind == 1, 234 ("vm_reserv_depopulate: reserv %p is already demoted", 235 rv)); 236 rv->pages->psind = 0; 237 } 238 rv->popcnt--; 239 if (rv->popcnt == 0) { 240 LIST_REMOVE(rv, objq); 241 rv->object = NULL; 242 vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER); 243 vm_reserv_freed++; 244 } else { 245 rv->inpartpopq = TRUE; 246 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq); 247 } 248} 249 250/* 251 * Returns the reservation to which the given page might belong. 252 */ 253static __inline vm_reserv_t 254vm_reserv_from_page(vm_page_t m) 255{ 256 257 return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]); 258} 259 260/* 261 * Returns TRUE if the given reservation contains the given page index and 262 * FALSE otherwise. 263 */ 264static __inline boolean_t 265vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex) 266{ 267 268 return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0); 269} 270 271/* 272 * Increases the given reservation's population count. Moves the reservation 273 * to the tail of the partially-populated reservation queue. 274 * 275 * The free page queue must be locked. 276 */ 277static void 278vm_reserv_populate(vm_reserv_t rv) 279{ 280 281 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 282 KASSERT(rv->object != NULL, 283 ("vm_reserv_populate: reserv %p is free", rv)); 284 KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES, 285 ("vm_reserv_populate: reserv %p is already full", rv)); 286 KASSERT(rv->pages->psind == 0, 287 ("vm_reserv_populate: reserv %p is already promoted", rv)); 288 if (rv->inpartpopq) { 289 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq); 290 rv->inpartpopq = FALSE; 291 } 292 rv->popcnt++; 293 if (rv->popcnt < VM_LEVEL_0_NPAGES) { 294 rv->inpartpopq = TRUE; 295 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq); 296 } else 297 rv->pages->psind = 1; 298} 299 300/* 301 * Allocates a contiguous set of physical pages of the given size "npages" 302 * from an existing or newly-created reservation. All of the physical pages 303 * must be at or above the given physical address "low" and below the given 304 * physical address "high". The given value "alignment" determines the 305 * alignment of the first physical page in the set. If the given value 306 * "boundary" is non-zero, then the set of physical pages cannot cross any 307 * physical address boundary that is a multiple of that value. Both 308 * "alignment" and "boundary" must be a power of two. 309 * 310 * The object and free page queue must be locked. 311 */ 312vm_page_t 313vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, u_long npages, 314 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) 315{ 316 vm_paddr_t pa, size; 317 vm_page_t m, m_ret, mpred, msucc; 318 vm_pindex_t first, leftcap, rightcap; 319 vm_reserv_t rv; 320 u_long allocpages, maxpages, minpages; 321 int i, index, n; 322 323 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 324 VM_OBJECT_ASSERT_WLOCKED(object); 325 KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0")); 326 327 /* 328 * Is a reservation fundamentally impossible? 329 */ 330 if (pindex < VM_RESERV_INDEX(object, pindex) || 331 pindex + npages > object->size) 332 return (NULL); 333 334 /* 335 * All reservations of a particular size have the same alignment. 336 * Assuming that the first page is allocated from a reservation, the 337 * least significant bits of its physical address can be determined 338 * from its offset from the beginning of the reservation and the size 339 * of the reservation. 340 * 341 * Could the specified index within a reservation of the smallest 342 * possible size satisfy the alignment and boundary requirements? 343 */ 344 pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT; 345 if ((pa & (alignment - 1)) != 0) 346 return (NULL); 347 size = npages << PAGE_SHIFT; 348 if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) 349 return (NULL); 350 351 /* 352 * Look for an existing reservation. 353 */ 354 mpred = vm_radix_lookup_le(&object->rtree, pindex); 355 if (mpred != NULL) { 356 KASSERT(mpred->pindex < pindex, 357 ("vm_reserv_alloc_contig: pindex already allocated")); 358 rv = vm_reserv_from_page(mpred); 359 if (rv->object == object && vm_reserv_has_pindex(rv, pindex)) 360 goto found; 361 msucc = TAILQ_NEXT(mpred, listq); 362 } else 363 msucc = TAILQ_FIRST(&object->memq); 364 if (msucc != NULL) { 365 KASSERT(msucc->pindex > pindex, 366 ("vm_reserv_alloc_page: pindex already allocated")); 367 rv = vm_reserv_from_page(msucc); 368 if (rv->object == object && vm_reserv_has_pindex(rv, pindex)) 369 goto found; 370 } 371 372 /* 373 * Could at least one reservation fit between the first index to the 374 * left that can be used and the first index to the right that cannot 375 * be used? 376 */ 377 first = pindex - VM_RESERV_INDEX(object, pindex); 378 if (mpred != NULL) { 379 if ((rv = vm_reserv_from_page(mpred))->object != object) 380 leftcap = mpred->pindex + 1; 381 else 382 leftcap = rv->pindex + VM_LEVEL_0_NPAGES; 383 if (leftcap > first) 384 return (NULL); 385 } 386 minpages = VM_RESERV_INDEX(object, pindex) + npages; 387 maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES); 388 allocpages = maxpages; 389 if (msucc != NULL) { 390 if ((rv = vm_reserv_from_page(msucc))->object != object) 391 rightcap = msucc->pindex; 392 else 393 rightcap = rv->pindex; 394 if (first + maxpages > rightcap) { 395 if (maxpages == VM_LEVEL_0_NPAGES) 396 return (NULL); 397 allocpages = minpages; 398 } 399 } 400 401 /* 402 * Would the last new reservation extend past the end of the object? 403 */ 404 if (first + maxpages > object->size) { 405 /* 406 * Don't allocate the last new reservation if the object is a 407 * vnode or backed by another object that is a vnode. 408 */ 409 if (object->type == OBJT_VNODE || 410 (object->backing_object != NULL && 411 object->backing_object->type == OBJT_VNODE)) { 412 if (maxpages == VM_LEVEL_0_NPAGES) 413 return (NULL); 414 allocpages = minpages; 415 } 416 /* Speculate that the object may grow. */ 417 } 418 419 /* 420 * Allocate and populate the new reservations. The alignment and 421 * boundary specified for this allocation may be different from the 422 * alignment and boundary specified for the requested pages. For 423 * instance, the specified index may not be the first page within the 424 * first new reservation. 425 */ 426 m = vm_phys_alloc_contig(allocpages, low, high, ulmax(alignment, 427 VM_LEVEL_0_SIZE), boundary > VM_LEVEL_0_SIZE ? boundary : 0); 428 if (m == NULL) 429 return (NULL); 430 m_ret = NULL; 431 index = VM_RESERV_INDEX(object, pindex); 432 do { 433 rv = vm_reserv_from_page(m); 434 KASSERT(rv->pages == m, 435 ("vm_reserv_alloc_contig: reserv %p's pages is corrupted", 436 rv)); 437 KASSERT(rv->object == NULL, 438 ("vm_reserv_alloc_contig: reserv %p isn't free", rv)); 439 LIST_INSERT_HEAD(&object->rvq, rv, objq); 440 rv->object = object; 441 rv->pindex = first; 442 KASSERT(rv->popcnt == 0, 443 ("vm_reserv_alloc_contig: reserv %p's popcnt is corrupted", 444 rv)); 445 KASSERT(!rv->inpartpopq, 446 ("vm_reserv_alloc_contig: reserv %p's inpartpopq is TRUE", 447 rv)); 448 n = ulmin(VM_LEVEL_0_NPAGES - index, npages); 449 for (i = 0; i < n; i++) 450 vm_reserv_populate(rv); 451 npages -= n; 452 if (m_ret == NULL) { 453 m_ret = &rv->pages[index]; 454 index = 0; 455 } 456 m += VM_LEVEL_0_NPAGES; 457 first += VM_LEVEL_0_NPAGES; 458 allocpages -= VM_LEVEL_0_NPAGES; 459 } while (allocpages > 0); 460 return (m_ret); 461 462 /* 463 * Found a matching reservation. 464 */ 465found: 466 index = VM_RESERV_INDEX(object, pindex); 467 /* Does the allocation fit within the reservation? */ 468 if (index + npages > VM_LEVEL_0_NPAGES) 469 return (NULL); 470 m = &rv->pages[index]; 471 pa = VM_PAGE_TO_PHYS(m); 472 if (pa < low || pa + size > high || (pa & (alignment - 1)) != 0 || 473 ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) 474 return (NULL); 475 /* Handle vm_page_rename(m, new_object, ...). */ 476 for (i = 0; i < npages; i++) 477 if ((rv->pages[index + i].flags & (PG_CACHED | PG_FREE)) == 0) 478 return (NULL); 479 for (i = 0; i < npages; i++) 480 vm_reserv_populate(rv); 481 return (m); 482} 483 484/* 485 * Allocates a page from an existing or newly-created reservation. 486 * 487 * The page "mpred" must immediately precede the offset "pindex" within the 488 * specified object. 489 * 490 * The object and free page queue must be locked. 491 */ 492vm_page_t 493vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, vm_page_t mpred) 494{ 495 vm_page_t m, msucc; 496 vm_pindex_t first, leftcap, rightcap; 497 vm_reserv_t rv; 498 499 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 500 VM_OBJECT_ASSERT_WLOCKED(object); 501 502 /* 503 * Is a reservation fundamentally impossible? 504 */ 505 if (pindex < VM_RESERV_INDEX(object, pindex) || 506 pindex >= object->size) 507 return (NULL); 508 509 /* 510 * Look for an existing reservation. 511 */ 512 if (mpred != NULL) { 513 KASSERT(mpred->object == object, 514 ("vm_reserv_alloc_page: object doesn't contain mpred")); 515 KASSERT(mpred->pindex < pindex, 516 ("vm_reserv_alloc_page: mpred doesn't precede pindex")); 517 rv = vm_reserv_from_page(mpred); 518 if (rv->object == object && vm_reserv_has_pindex(rv, pindex)) 519 goto found; 520 msucc = TAILQ_NEXT(mpred, listq); 521 } else 522 msucc = TAILQ_FIRST(&object->memq); 523 if (msucc != NULL) { 524 KASSERT(msucc->pindex > pindex, 525 ("vm_reserv_alloc_page: msucc doesn't succeed pindex")); 526 rv = vm_reserv_from_page(msucc); 527 if (rv->object == object && vm_reserv_has_pindex(rv, pindex)) 528 goto found; 529 } 530 531 /* 532 * Could a reservation fit between the first index to the left that 533 * can be used and the first index to the right that cannot be used? 534 */ 535 first = pindex - VM_RESERV_INDEX(object, pindex); 536 if (mpred != NULL) { 537 if ((rv = vm_reserv_from_page(mpred))->object != object) 538 leftcap = mpred->pindex + 1; 539 else 540 leftcap = rv->pindex + VM_LEVEL_0_NPAGES; 541 if (leftcap > first) 542 return (NULL); 543 } 544 if (msucc != NULL) { 545 if ((rv = vm_reserv_from_page(msucc))->object != object) 546 rightcap = msucc->pindex; 547 else 548 rightcap = rv->pindex; 549 if (first + VM_LEVEL_0_NPAGES > rightcap) 550 return (NULL); 551 } 552 553 /* 554 * Would a new reservation extend past the end of the object? 555 */ 556 if (first + VM_LEVEL_0_NPAGES > object->size) { 557 /* 558 * Don't allocate a new reservation if the object is a vnode or 559 * backed by another object that is a vnode. 560 */ 561 if (object->type == OBJT_VNODE || 562 (object->backing_object != NULL && 563 object->backing_object->type == OBJT_VNODE)) 564 return (NULL); 565 /* Speculate that the object may grow. */ 566 } 567 568 /* 569 * Allocate and populate the new reservation. 570 */ 571 m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER); 572 if (m == NULL) 573 return (NULL); 574 rv = vm_reserv_from_page(m); 575 KASSERT(rv->pages == m, 576 ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv)); 577 KASSERT(rv->object == NULL, 578 ("vm_reserv_alloc_page: reserv %p isn't free", rv)); 579 LIST_INSERT_HEAD(&object->rvq, rv, objq); 580 rv->object = object; 581 rv->pindex = first; 582 KASSERT(rv->popcnt == 0, 583 ("vm_reserv_alloc_page: reserv %p's popcnt is corrupted", rv)); 584 KASSERT(!rv->inpartpopq, 585 ("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE", rv)); 586 vm_reserv_populate(rv); 587 return (&rv->pages[VM_RESERV_INDEX(object, pindex)]); 588 589 /* 590 * Found a matching reservation. 591 */ 592found: 593 m = &rv->pages[VM_RESERV_INDEX(object, pindex)]; 594 /* Handle vm_page_rename(m, new_object, ...). */ 595 if ((m->flags & (PG_CACHED | PG_FREE)) == 0) 596 return (NULL); 597 vm_reserv_populate(rv); 598 return (m); 599} 600 601/* 602 * Breaks all reservations belonging to the given object. 603 */ 604void 605vm_reserv_break_all(vm_object_t object) 606{ 607 vm_reserv_t rv; 608 int i; 609 610 mtx_lock(&vm_page_queue_free_mtx); 611 while ((rv = LIST_FIRST(&object->rvq)) != NULL) { 612 KASSERT(rv->object == object, 613 ("vm_reserv_break_all: reserv %p is corrupted", rv)); 614 if (rv->inpartpopq) { 615 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq); 616 rv->inpartpopq = FALSE; 617 } 618 LIST_REMOVE(rv, objq); 619 rv->object = NULL; 620 for (i = 0; i < VM_LEVEL_0_NPAGES; i++) { 621 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) 622 vm_phys_free_pages(&rv->pages[i], 0); 623 else 624 rv->popcnt--; 625 } 626 KASSERT(rv->popcnt == 0, 627 ("vm_reserv_break_all: reserv %p's popcnt is corrupted", 628 rv)); 629 vm_reserv_broken++; 630 } 631 mtx_unlock(&vm_page_queue_free_mtx); 632} 633 634/* 635 * Frees the given page if it belongs to a reservation. Returns TRUE if the 636 * page is freed and FALSE otherwise. 637 * 638 * The free page queue lock must be held. 639 */ 640boolean_t 641vm_reserv_free_page(vm_page_t m) 642{ 643 vm_reserv_t rv; 644 645 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 646 rv = vm_reserv_from_page(m); 647 if (rv->object == NULL) 648 return (FALSE); 649 if ((m->flags & PG_CACHED) != 0 && m->pool != VM_FREEPOOL_CACHE) 650 vm_phys_set_pool(VM_FREEPOOL_CACHE, rv->pages, 651 VM_LEVEL_0_ORDER); 652 vm_reserv_depopulate(rv); 653 return (TRUE); 654} 655 656/* 657 * Initializes the reservation management system. Specifically, initializes 658 * the reservation array. 659 * 660 * Requires that vm_page_array and first_page are initialized! 661 */ 662void 663vm_reserv_init(void) 664{ 665 vm_paddr_t paddr; 666 int i; 667 668 /* 669 * Initialize the reservation array. Specifically, initialize the 670 * "pages" field for every element that has an underlying superpage. 671 */ 672 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 673 paddr = roundup2(phys_avail[i], VM_LEVEL_0_SIZE); 674 while (paddr + VM_LEVEL_0_SIZE <= phys_avail[i + 1]) { 675 vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages = 676 PHYS_TO_VM_PAGE(paddr); 677 paddr += VM_LEVEL_0_SIZE; 678 } 679 } 680} 681 682/* 683 * Returns a reservation level if the given page belongs to a fully-populated 684 * reservation and -1 otherwise. 685 */ 686int 687vm_reserv_level_iffullpop(vm_page_t m) 688{ 689 vm_reserv_t rv; 690 691 rv = vm_reserv_from_page(m); 692 return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1); 693} 694 695/* 696 * Prepare for the reactivation of a cached page. 697 * 698 * First, suppose that the given page "m" was allocated individually, i.e., not 699 * as part of a reservation, and cached. Then, suppose a reservation 700 * containing "m" is allocated by the same object. Although "m" and the 701 * reservation belong to the same object, "m"'s pindex may not match the 702 * reservation's. 703 * 704 * The free page queue must be locked. 705 */ 706boolean_t 707vm_reserv_reactivate_page(vm_page_t m) 708{ 709 vm_reserv_t rv; 710 int i, m_index; 711 712 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 713 rv = vm_reserv_from_page(m); 714 if (rv->object == NULL) 715 return (FALSE); 716 KASSERT((m->flags & PG_CACHED) != 0, 717 ("vm_reserv_uncache_page: page %p is not cached", m)); 718 if (m->object == rv->object && 719 m->pindex - rv->pindex == VM_RESERV_INDEX(m->object, m->pindex)) 720 vm_reserv_populate(rv); 721 else { 722 KASSERT(rv->inpartpopq, 723 ("vm_reserv_uncache_page: reserv %p's inpartpopq is FALSE", 724 rv)); 725 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq); 726 rv->inpartpopq = FALSE; 727 LIST_REMOVE(rv, objq); 728 rv->object = NULL; 729 /* Don't vm_phys_free_pages(m, 0). */ 730 m_index = m - rv->pages; 731 for (i = 0; i < m_index; i++) { 732 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) 733 vm_phys_free_pages(&rv->pages[i], 0); 734 else 735 rv->popcnt--; 736 } 737 for (i++; i < VM_LEVEL_0_NPAGES; i++) { 738 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) 739 vm_phys_free_pages(&rv->pages[i], 0); 740 else 741 rv->popcnt--; 742 } 743 KASSERT(rv->popcnt == 0, 744 ("vm_reserv_uncache_page: reserv %p's popcnt is corrupted", 745 rv)); 746 vm_reserv_broken++; 747 } 748 return (TRUE); 749} 750 751/* 752 * Breaks the given partially-populated reservation, releasing its cached and 753 * free pages to the physical memory allocator. 754 * 755 * The free page queue lock must be held. 756 */ 757static void 758vm_reserv_reclaim(vm_reserv_t rv) 759{ 760 int i; 761 762 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 763 KASSERT(rv->inpartpopq, 764 ("vm_reserv_reclaim: reserv %p's inpartpopq is corrupted", rv)); 765 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq); 766 rv->inpartpopq = FALSE; 767 KASSERT(rv->object != NULL, 768 ("vm_reserv_reclaim: reserv %p is free", rv)); 769 LIST_REMOVE(rv, objq); 770 rv->object = NULL; 771 for (i = 0; i < VM_LEVEL_0_NPAGES; i++) { 772 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) 773 vm_phys_free_pages(&rv->pages[i], 0); 774 else 775 rv->popcnt--; 776 } 777 KASSERT(rv->popcnt == 0, 778 ("vm_reserv_reclaim: reserv %p's popcnt is corrupted", rv)); 779 vm_reserv_reclaimed++; 780} 781 782/* 783 * Breaks the reservation at the head of the partially-populated reservation 784 * queue, releasing its cached and free pages to the physical memory 785 * allocator. Returns TRUE if a reservation is broken and FALSE otherwise. 786 * 787 * The free page queue lock must be held. 788 */ 789boolean_t 790vm_reserv_reclaim_inactive(void) 791{ 792 vm_reserv_t rv; 793 794 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 795 if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) { 796 vm_reserv_reclaim(rv); 797 return (TRUE); 798 } 799 return (FALSE); 800} 801 802/* 803 * Searches the partially-populated reservation queue for the least recently 804 * active reservation with unused pages, i.e., cached or free, that satisfy the 805 * given request for contiguous physical memory. If a satisfactory reservation 806 * is found, it is broken. Returns TRUE if a reservation is broken and FALSE 807 * otherwise. 808 * 809 * The free page queue lock must be held. 810 */ 811boolean_t 812vm_reserv_reclaim_contig(u_long npages, vm_paddr_t low, vm_paddr_t high, 813 u_long alignment, vm_paddr_t boundary) 814{ 815 vm_paddr_t pa, pa_length, size; 816 vm_reserv_t rv; 817 int i; 818 819 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 820 if (npages > VM_LEVEL_0_NPAGES - 1) 821 return (FALSE); 822 size = npages << PAGE_SHIFT; 823 TAILQ_FOREACH(rv, &vm_rvq_partpop, partpopq) { 824 pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]); 825 if (pa + PAGE_SIZE - size < low) { 826 /* this entire reservation is too low; go to next */ 827 continue; 828 } 829 pa_length = 0; 830 for (i = 0; i < VM_LEVEL_0_NPAGES; i++) 831 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) { 832 pa_length += PAGE_SIZE; 833 if (pa_length == PAGE_SIZE) { 834 pa = VM_PAGE_TO_PHYS(&rv->pages[i]); 835 if (pa + size > high) { 836 /* skip to next reservation */ 837 break; 838 } else if (pa < low || 839 (pa & (alignment - 1)) != 0 || 840 ((pa ^ (pa + size - 1)) & 841 ~(boundary - 1)) != 0) 842 pa_length = 0; 843 } 844 if (pa_length >= size) { 845 vm_reserv_reclaim(rv); 846 return (TRUE); 847 } 848 } else 849 pa_length = 0; 850 } 851 return (FALSE); 852} 853 854/* 855 * Transfers the reservation underlying the given page to a new object. 856 * 857 * The object must be locked. 858 */ 859void 860vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object, 861 vm_pindex_t old_object_offset) 862{ 863 vm_reserv_t rv; 864 865 VM_OBJECT_ASSERT_WLOCKED(new_object); 866 rv = vm_reserv_from_page(m); 867 if (rv->object == old_object) { 868 mtx_lock(&vm_page_queue_free_mtx); 869 if (rv->object == old_object) { 870 LIST_REMOVE(rv, objq); 871 LIST_INSERT_HEAD(&new_object->rvq, rv, objq); 872 rv->object = new_object; 873 rv->pindex -= old_object_offset; 874 } 875 mtx_unlock(&vm_page_queue_free_mtx); 876 } 877} 878 879/* 880 * Allocates the virtual and physical memory required by the reservation 881 * management system's data structures, in particular, the reservation array. 882 */ 883vm_paddr_t 884vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water) 885{ 886 vm_paddr_t new_end; 887 size_t size; 888 889 /* 890 * Calculate the size (in bytes) of the reservation array. Round up 891 * from "high_water" because every small page is mapped to an element 892 * in the reservation array based on its physical address. Thus, the 893 * number of elements in the reservation array can be greater than the 894 * number of superpages. 895 */ 896 size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv); 897 898 /* 899 * Allocate and map the physical memory for the reservation array. The 900 * next available virtual address is returned by reference. 901 */ 902 new_end = end - round_page(size); 903 vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end, 904 VM_PROT_READ | VM_PROT_WRITE); 905 bzero(vm_reserv_array, size); 906 907 /* 908 * Return the next available physical address. 909 */ 910 return (new_end); 911} 912 913#endif /* VM_NRESERVLEVEL > 0 */ 914