1 /*- 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. 5 * 6 * This code is derived from software contributed to Berkeley by 7 * The Mach Operating System project at Carnegie-Mellon University. 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 * 4. Neither the name of the University nor the names of its contributors 18 * may be used to endorse or promote products derived from this software 19 * without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 * 33 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 34 */ 35 36 /*- 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63 /* 64 * GENERAL RULES ON VM_PAGE MANIPULATION 65 * 66 * - A page queue lock is required when adding or removing a page from a 67 * page queue regardless of other locks or the busy state of a page. 68 * 69 * * In general, no thread besides the page daemon can acquire or 70 * hold more than one page queue lock at a time. 71 * 72 * * The page daemon can acquire and hold any pair of page queue 73 * locks in any order. 74 * 75 * - The object lock is required when inserting or removing 76 * pages from an object (vm_page_insert() or vm_page_remove()). 77 * 78 */ 79 80 /* 81 * Resident memory management module. 82 */ 83 84 #include <sys/cdefs.h> 85 __FBSDID("$FreeBSD$"); 86 87 #include "opt_vm.h" 88 89 #include <sys/param.h> 90 #include <sys/systm.h> 91 #include <sys/lock.h> 92 #include <sys/kernel.h> 93 #include <sys/limits.h> 94 #include <sys/linker.h> 95 #include <sys/malloc.h> 96 #include <sys/mman.h> 97 #include <sys/msgbuf.h> 98 #include <sys/mutex.h> 99 #include <sys/proc.h> 100 #include <sys/rwlock.h> 101 #include <sys/sbuf.h> 102 #include <sys/smp.h> 103 #include <sys/sysctl.h> 104 #include <sys/vmmeter.h> 105 #include <sys/vnode.h> 106 107 #include <vm/vm.h> 108 #include <vm/pmap.h> 109 #include <vm/vm_param.h> 110 #include <vm/vm_kern.h> 111 #include <vm/vm_object.h> 112 #include <vm/vm_page.h> 113 #include <vm/vm_pageout.h> 114 #include <vm/vm_pager.h> 115 #include <vm/vm_phys.h> 116 #include <vm/vm_radix.h> 117 #include <vm/vm_reserv.h> 118 #include <vm/vm_extern.h> 119 #include <vm/uma.h> 120 #include <vm/uma_int.h> 121 122 #include <machine/md_var.h> 123 124 /* 125 * Associated with page of user-allocatable memory is a 126 * page structure. 127 */ 128 129 struct vm_domain vm_dom[MAXMEMDOM]; 130 struct mtx_padalign vm_page_queue_free_mtx; 131 132 struct mtx_padalign pa_lock[PA_LOCK_COUNT]; 133 134 vm_page_t vm_page_array; 135 long vm_page_array_size; 136 long first_page; 137 138 static int boot_pages = UMA_BOOT_PAGES; 139 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, 140 &boot_pages, 0, 141 "number of pages allocated for bootstrapping the VM system"); 142 143 static int pa_tryrelock_restart; 144 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, 145 &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); 146 147 static TAILQ_HEAD(, vm_page) blacklist_head; 148 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); 149 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | 150 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); 151 152 /* Is the page daemon waiting for free pages? */ 153 static int vm_pageout_pages_needed; 154 155 static uma_zone_t fakepg_zone; 156 157 static void vm_page_alloc_check(vm_page_t m); 158 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); 159 static void vm_page_enqueue(uint8_t queue, vm_page_t m); 160 static void vm_page_free_wakeup(void); 161 static void vm_page_init_fakepg(void *dummy); 162 static int vm_page_insert_after(vm_page_t m, vm_object_t object, 163 vm_pindex_t pindex, vm_page_t mpred); 164 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, 165 vm_page_t mpred); 166 static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run, 167 vm_paddr_t high); 168 169 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL); 170 171 static void 172 vm_page_init_fakepg(void *dummy) 173 { 174 175 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, 176 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 177 } 178 179 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 180 #if PAGE_SIZE == 32768 181 #ifdef CTASSERT 182 CTASSERT(sizeof(u_long) >= 8); 183 #endif 184 #endif 185 186 /* 187 * Try to acquire a physical address lock while a pmap is locked. If we 188 * fail to trylock we unlock and lock the pmap directly and cache the 189 * locked pa in *locked. The caller should then restart their loop in case 190 * the virtual to physical mapping has changed. 191 */ 192 int 193 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) 194 { 195 vm_paddr_t lockpa; 196 197 lockpa = *locked; 198 *locked = pa; 199 if (lockpa) { 200 PA_LOCK_ASSERT(lockpa, MA_OWNED); 201 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) 202 return (0); 203 PA_UNLOCK(lockpa); 204 } 205 if (PA_TRYLOCK(pa)) 206 return (0); 207 PMAP_UNLOCK(pmap); 208 atomic_add_int(&pa_tryrelock_restart, 1); 209 PA_LOCK(pa); 210 PMAP_LOCK(pmap); 211 return (EAGAIN); 212 } 213 214 /* 215 * vm_set_page_size: 216 * 217 * Sets the page size, perhaps based upon the memory 218 * size. Must be called before any use of page-size 219 * dependent functions. 220 */ 221 void 222 vm_set_page_size(void) 223 { 224 if (vm_cnt.v_page_size == 0) 225 vm_cnt.v_page_size = PAGE_SIZE; 226 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) 227 panic("vm_set_page_size: page size not a power of two"); 228 } 229 230 /* 231 * vm_page_blacklist_next: 232 * 233 * Find the next entry in the provided string of blacklist 234 * addresses. Entries are separated by space, comma, or newline. 235 * If an invalid integer is encountered then the rest of the 236 * string is skipped. Updates the list pointer to the next 237 * character, or NULL if the string is exhausted or invalid. 238 */ 239 static vm_paddr_t 240 vm_page_blacklist_next(char **list, char *end) 241 { 242 vm_paddr_t bad; 243 char *cp, *pos; 244 245 if (list == NULL || *list == NULL) 246 return (0); 247 if (**list =='\0') { 248 *list = NULL; 249 return (0); 250 } 251 252 /* 253 * If there's no end pointer then the buffer is coming from 254 * the kenv and we know it's null-terminated. 255 */ 256 if (end == NULL) 257 end = *list + strlen(*list); 258 259 /* Ensure that strtoq() won't walk off the end */ 260 if (*end != '\0') { 261 if (*end == '\n' || *end == ' ' || *end == ',') 262 *end = '\0'; 263 else { 264 printf("Blacklist not terminated, skipping\n"); 265 *list = NULL; 266 return (0); 267 } 268 } 269 270 for (pos = *list; *pos != '\0'; pos = cp) { 271 bad = strtoq(pos, &cp, 0); 272 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { 273 if (bad == 0) { 274 if (++cp < end) 275 continue; 276 else 277 break; 278 } 279 } else 280 break; 281 if (*cp == '\0' || ++cp >= end) 282 *list = NULL; 283 else 284 *list = cp; 285 return (trunc_page(bad)); 286 } 287 printf("Garbage in RAM blacklist, skipping\n"); 288 *list = NULL; 289 return (0); 290 } 291 292 /* 293 * vm_page_blacklist_check: 294 * 295 * Iterate through the provided string of blacklist addresses, pulling 296 * each entry out of the physical allocator free list and putting it 297 * onto a list for reporting via the vm.page_blacklist sysctl. 298 */ 299 static void 300 vm_page_blacklist_check(char *list, char *end) 301 { 302 vm_paddr_t pa; 303 vm_page_t m; 304 char *next; 305 int ret; 306 307 next = list; 308 while (next != NULL) { 309 if ((pa = vm_page_blacklist_next(&next, end)) == 0) 310 continue; 311 m = vm_phys_paddr_to_vm_page(pa); 312 if (m == NULL) 313 continue; 314 mtx_lock(&vm_page_queue_free_mtx); 315 ret = vm_phys_unfree_page(m); 316 mtx_unlock(&vm_page_queue_free_mtx); 317 if (ret == TRUE) { 318 TAILQ_INSERT_TAIL(&blacklist_head, m, listq); 319 if (bootverbose) 320 printf("Skipping page with pa 0x%jx\n", 321 (uintmax_t)pa); 322 } 323 } 324 } 325 326 /* 327 * vm_page_blacklist_load: 328 * 329 * Search for a special module named "ram_blacklist". It'll be a 330 * plain text file provided by the user via the loader directive 331 * of the same name. 332 */ 333 static void 334 vm_page_blacklist_load(char **list, char **end) 335 { 336 void *mod; 337 u_char *ptr; 338 u_int len; 339 340 mod = NULL; 341 ptr = NULL; 342 343 mod = preload_search_by_type("ram_blacklist"); 344 if (mod != NULL) { 345 ptr = preload_fetch_addr(mod); 346 len = preload_fetch_size(mod); 347 } 348 *list = ptr; 349 if (ptr != NULL) 350 *end = ptr + len; 351 else 352 *end = NULL; 353 return; 354 } 355 356 static int 357 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) 358 { 359 vm_page_t m; 360 struct sbuf sbuf; 361 int error, first; 362 363 first = 1; 364 error = sysctl_wire_old_buffer(req, 0); 365 if (error != 0) 366 return (error); 367 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 368 TAILQ_FOREACH(m, &blacklist_head, listq) { 369 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", 370 (uintmax_t)m->phys_addr); 371 first = 0; 372 } 373 error = sbuf_finish(&sbuf); 374 sbuf_delete(&sbuf); 375 return (error); 376 } 377 378 static void 379 vm_page_domain_init(struct vm_domain *vmd) 380 { 381 struct vm_pagequeue *pq; 382 int i; 383 384 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = 385 "vm inactive pagequeue"; 386 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) = 387 &vm_cnt.v_inactive_count; 388 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = 389 "vm active pagequeue"; 390 *__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) = 391 &vm_cnt.v_active_count; 392 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = 393 "vm laundry pagequeue"; 394 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_vcnt) = 395 &vm_cnt.v_laundry_count; 396 vmd->vmd_page_count = 0; 397 vmd->vmd_free_count = 0; 398 vmd->vmd_segs = 0; 399 vmd->vmd_oom = FALSE; 400 for (i = 0; i < PQ_COUNT; i++) { 401 pq = &vmd->vmd_pagequeues[i]; 402 TAILQ_INIT(&pq->pq_pl); 403 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", 404 MTX_DEF | MTX_DUPOK); 405 } 406 } 407 408 /* 409 * vm_page_startup: 410 * 411 * Initializes the resident memory module. 412 * 413 * Allocates memory for the page cells, and 414 * for the object/offset-to-page hash table headers. 415 * Each page cell is initialized and placed on the free list. 416 */ 417 vm_offset_t 418 vm_page_startup(vm_offset_t vaddr) 419 { 420 vm_offset_t mapped; 421 vm_paddr_t page_range; 422 vm_paddr_t new_end; 423 int i; 424 vm_paddr_t pa; 425 vm_paddr_t last_pa; 426 char *list, *listend; 427 vm_paddr_t end; 428 vm_paddr_t biggestsize; 429 vm_paddr_t low_water, high_water; 430 int biggestone; 431 int pages_per_zone; 432 433 biggestsize = 0; 434 biggestone = 0; 435 vaddr = round_page(vaddr); 436 437 for (i = 0; phys_avail[i + 1]; i += 2) { 438 phys_avail[i] = round_page(phys_avail[i]); 439 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 440 } 441 442 low_water = phys_avail[0]; 443 high_water = phys_avail[1]; 444 445 for (i = 0; i < vm_phys_nsegs; i++) { 446 if (vm_phys_segs[i].start < low_water) 447 low_water = vm_phys_segs[i].start; 448 if (vm_phys_segs[i].end > high_water) 449 high_water = vm_phys_segs[i].end; 450 } 451 for (i = 0; phys_avail[i + 1]; i += 2) { 452 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 453 454 if (size > biggestsize) { 455 biggestone = i; 456 biggestsize = size; 457 } 458 if (phys_avail[i] < low_water) 459 low_water = phys_avail[i]; 460 if (phys_avail[i + 1] > high_water) 461 high_water = phys_avail[i + 1]; 462 } 463 464 end = phys_avail[biggestone+1]; 465 466 /* 467 * Initialize the page and queue locks. 468 */ 469 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF); 470 for (i = 0; i < PA_LOCK_COUNT; i++) 471 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); 472 for (i = 0; i < vm_ndomains; i++) 473 vm_page_domain_init(&vm_dom[i]); 474 475 /* 476 * Almost all of the pages needed for boot strapping UMA are used 477 * for zone structures, so if the number of CPUs results in those 478 * structures taking more than one page each, we set aside more pages 479 * in proportion to the zone structure size. 480 */ 481 pages_per_zone = howmany(sizeof(struct uma_zone) + 482 sizeof(struct uma_cache) * (mp_maxid + 1), UMA_SLAB_SIZE); 483 if (pages_per_zone > 1) { 484 /* Reserve more pages so that we don't run out. */ 485 boot_pages = UMA_BOOT_PAGES_ZONES * pages_per_zone; 486 } 487 488 /* 489 * Allocate memory for use when boot strapping the kernel memory 490 * allocator. 491 * 492 * CTFLAG_RDTUN doesn't work during the early boot process, so we must 493 * manually fetch the value. 494 */ 495 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); 496 new_end = end - (boot_pages * UMA_SLAB_SIZE); 497 new_end = trunc_page(new_end); 498 mapped = pmap_map(&vaddr, new_end, end, 499 VM_PROT_READ | VM_PROT_WRITE); 500 bzero((void *)mapped, end - new_end); 501 uma_startup((void *)mapped, boot_pages); 502 503 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ 504 defined(__i386__) || defined(__mips__) 505 /* 506 * Allocate a bitmap to indicate that a random physical page 507 * needs to be included in a minidump. 508 * 509 * The amd64 port needs this to indicate which direct map pages 510 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 511 * 512 * However, i386 still needs this workspace internally within the 513 * minidump code. In theory, they are not needed on i386, but are 514 * included should the sf_buf code decide to use them. 515 */ 516 last_pa = 0; 517 for (i = 0; dump_avail[i + 1] != 0; i += 2) 518 if (dump_avail[i + 1] > last_pa) 519 last_pa = dump_avail[i + 1]; 520 page_range = last_pa / PAGE_SIZE; 521 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 522 new_end -= vm_page_dump_size; 523 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 524 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 525 bzero((void *)vm_page_dump, vm_page_dump_size); 526 #endif 527 #ifdef __amd64__ 528 /* 529 * Request that the physical pages underlying the message buffer be 530 * included in a crash dump. Since the message buffer is accessed 531 * through the direct map, they are not automatically included. 532 */ 533 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); 534 last_pa = pa + round_page(msgbufsize); 535 while (pa < last_pa) { 536 dump_add_page(pa); 537 pa += PAGE_SIZE; 538 } 539 #endif 540 /* 541 * Compute the number of pages of memory that will be available for 542 * use (taking into account the overhead of a page structure per 543 * page). 544 */ 545 first_page = low_water / PAGE_SIZE; 546 #ifdef VM_PHYSSEG_SPARSE 547 page_range = 0; 548 for (i = 0; i < vm_phys_nsegs; i++) { 549 page_range += atop(vm_phys_segs[i].end - 550 vm_phys_segs[i].start); 551 } 552 for (i = 0; phys_avail[i + 1] != 0; i += 2) 553 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 554 #elif defined(VM_PHYSSEG_DENSE) 555 page_range = high_water / PAGE_SIZE - first_page; 556 #else 557 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 558 #endif 559 end = new_end; 560 561 /* 562 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 563 */ 564 vaddr += PAGE_SIZE; 565 566 /* 567 * Initialize the mem entry structures now, and put them in the free 568 * queue. 569 */ 570 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 571 mapped = pmap_map(&vaddr, new_end, end, 572 VM_PROT_READ | VM_PROT_WRITE); 573 vm_page_array = (vm_page_t) mapped; 574 #if VM_NRESERVLEVEL > 0 575 /* 576 * Allocate memory for the reservation management system's data 577 * structures. 578 */ 579 new_end = vm_reserv_startup(&vaddr, new_end, high_water); 580 #endif 581 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) 582 /* 583 * pmap_map on arm64, amd64, and mips can come out of the direct-map, 584 * not kvm like i386, so the pages must be tracked for a crashdump to 585 * include this data. This includes the vm_page_array and the early 586 * UMA bootstrap pages. 587 */ 588 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 589 dump_add_page(pa); 590 #endif 591 phys_avail[biggestone + 1] = new_end; 592 593 /* 594 * Add physical memory segments corresponding to the available 595 * physical pages. 596 */ 597 for (i = 0; phys_avail[i + 1] != 0; i += 2) 598 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); 599 600 /* 601 * Clear all of the page structures 602 */ 603 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 604 for (i = 0; i < page_range; i++) 605 vm_page_array[i].order = VM_NFREEORDER; 606 vm_page_array_size = page_range; 607 608 /* 609 * Initialize the physical memory allocator. 610 */ 611 vm_phys_init(); 612 613 /* 614 * Add every available physical page that is not blacklisted to 615 * the free lists. 616 */ 617 vm_cnt.v_page_count = 0; 618 vm_cnt.v_free_count = 0; 619 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 620 pa = phys_avail[i]; 621 last_pa = phys_avail[i + 1]; 622 while (pa < last_pa) { 623 vm_phys_add_page(pa); 624 pa += PAGE_SIZE; 625 } 626 } 627 628 TAILQ_INIT(&blacklist_head); 629 vm_page_blacklist_load(&list, &listend); 630 vm_page_blacklist_check(list, listend); 631 632 list = kern_getenv("vm.blacklist"); 633 vm_page_blacklist_check(list, NULL); 634 635 freeenv(list); 636 #if VM_NRESERVLEVEL > 0 637 /* 638 * Initialize the reservation management system. 639 */ 640 vm_reserv_init(); 641 #endif 642 return (vaddr); 643 } 644 645 void 646 vm_page_reference(vm_page_t m) 647 { 648 649 vm_page_aflag_set(m, PGA_REFERENCED); 650 } 651 652 /* 653 * vm_page_busy_downgrade: 654 * 655 * Downgrade an exclusive busy page into a single shared busy page. 656 */ 657 void 658 vm_page_busy_downgrade(vm_page_t m) 659 { 660 u_int x; 661 bool locked; 662 663 vm_page_assert_xbusied(m); 664 locked = mtx_owned(vm_page_lockptr(m)); 665 666 for (;;) { 667 x = m->busy_lock; 668 x &= VPB_BIT_WAITERS; 669 if (x != 0 && !locked) 670 vm_page_lock(m); 671 if (atomic_cmpset_rel_int(&m->busy_lock, 672 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1))) 673 break; 674 if (x != 0 && !locked) 675 vm_page_unlock(m); 676 } 677 if (x != 0) { 678 wakeup(m); 679 if (!locked) 680 vm_page_unlock(m); 681 } 682 } 683 684 /* 685 * vm_page_sbusied: 686 * 687 * Return a positive value if the page is shared busied, 0 otherwise. 688 */ 689 int 690 vm_page_sbusied(vm_page_t m) 691 { 692 u_int x; 693 694 x = m->busy_lock; 695 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); 696 } 697 698 /* 699 * vm_page_sunbusy: 700 * 701 * Shared unbusy a page. 702 */ 703 void 704 vm_page_sunbusy(vm_page_t m) 705 { 706 u_int x; 707 708 vm_page_assert_sbusied(m); 709 710 for (;;) { 711 x = m->busy_lock; 712 if (VPB_SHARERS(x) > 1) { 713 if (atomic_cmpset_int(&m->busy_lock, x, 714 x - VPB_ONE_SHARER)) 715 break; 716 continue; 717 } 718 if ((x & VPB_BIT_WAITERS) == 0) { 719 KASSERT(x == VPB_SHARERS_WORD(1), 720 ("vm_page_sunbusy: invalid lock state")); 721 if (atomic_cmpset_int(&m->busy_lock, 722 VPB_SHARERS_WORD(1), VPB_UNBUSIED)) 723 break; 724 continue; 725 } 726 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), 727 ("vm_page_sunbusy: invalid lock state for waiters")); 728 729 vm_page_lock(m); 730 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { 731 vm_page_unlock(m); 732 continue; 733 } 734 wakeup(m); 735 vm_page_unlock(m); 736 break; 737 } 738 } 739 740 /* 741 * vm_page_busy_sleep: 742 * 743 * Sleep and release the page lock, using the page pointer as wchan. 744 * This is used to implement the hard-path of busying mechanism. 745 * 746 * The given page must be locked. 747 * 748 * If nonshared is true, sleep only if the page is xbusy. 749 */ 750 void 751 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) 752 { 753 u_int x; 754 755 vm_page_assert_locked(m); 756 757 x = m->busy_lock; 758 if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || 759 ((x & VPB_BIT_WAITERS) == 0 && 760 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { 761 vm_page_unlock(m); 762 return; 763 } 764 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); 765 } 766 767 /* 768 * vm_page_trysbusy: 769 * 770 * Try to shared busy a page. 771 * If the operation succeeds 1 is returned otherwise 0. 772 * The operation never sleeps. 773 */ 774 int 775 vm_page_trysbusy(vm_page_t m) 776 { 777 u_int x; 778 779 for (;;) { 780 x = m->busy_lock; 781 if ((x & VPB_BIT_SHARED) == 0) 782 return (0); 783 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) 784 return (1); 785 } 786 } 787 788 static void 789 vm_page_xunbusy_locked(vm_page_t m) 790 { 791 792 vm_page_assert_xbusied(m); 793 vm_page_assert_locked(m); 794 795 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); 796 /* There is a waiter, do wakeup() instead of vm_page_flash(). */ 797 wakeup(m); 798 } 799 800 void 801 vm_page_xunbusy_maybelocked(vm_page_t m) 802 { 803 bool lockacq; 804 805 vm_page_assert_xbusied(m); 806 807 /* 808 * Fast path for unbusy. If it succeeds, we know that there 809 * are no waiters, so we do not need a wakeup. 810 */ 811 if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER, 812 VPB_UNBUSIED)) 813 return; 814 815 lockacq = !mtx_owned(vm_page_lockptr(m)); 816 if (lockacq) 817 vm_page_lock(m); 818 vm_page_xunbusy_locked(m); 819 if (lockacq) 820 vm_page_unlock(m); 821 } 822 823 /* 824 * vm_page_xunbusy_hard: 825 * 826 * Called after the first try the exclusive unbusy of a page failed. 827 * It is assumed that the waiters bit is on. 828 */ 829 void 830 vm_page_xunbusy_hard(vm_page_t m) 831 { 832 833 vm_page_assert_xbusied(m); 834 835 vm_page_lock(m); 836 vm_page_xunbusy_locked(m); 837 vm_page_unlock(m); 838 } 839 840 /* 841 * vm_page_flash: 842 * 843 * Wakeup anyone waiting for the page. 844 * The ownership bits do not change. 845 * 846 * The given page must be locked. 847 */ 848 void 849 vm_page_flash(vm_page_t m) 850 { 851 u_int x; 852 853 vm_page_lock_assert(m, MA_OWNED); 854 855 for (;;) { 856 x = m->busy_lock; 857 if ((x & VPB_BIT_WAITERS) == 0) 858 return; 859 if (atomic_cmpset_int(&m->busy_lock, x, 860 x & (~VPB_BIT_WAITERS))) 861 break; 862 } 863 wakeup(m); 864 } 865 866 /* 867 * Keep page from being freed by the page daemon 868 * much of the same effect as wiring, except much lower 869 * overhead and should be used only for *very* temporary 870 * holding ("wiring"). 871 */ 872 void 873 vm_page_hold(vm_page_t mem) 874 { 875 876 vm_page_lock_assert(mem, MA_OWNED); 877 mem->hold_count++; 878 } 879 880 void 881 vm_page_unhold(vm_page_t mem) 882 { 883 884 vm_page_lock_assert(mem, MA_OWNED); 885 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); 886 --mem->hold_count; 887 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) 888 vm_page_free_toq(mem); 889 } 890 891 /* 892 * vm_page_unhold_pages: 893 * 894 * Unhold each of the pages that is referenced by the given array. 895 */ 896 void 897 vm_page_unhold_pages(vm_page_t *ma, int count) 898 { 899 struct mtx *mtx, *new_mtx; 900 901 mtx = NULL; 902 for (; count != 0; count--) { 903 /* 904 * Avoid releasing and reacquiring the same page lock. 905 */ 906 new_mtx = vm_page_lockptr(*ma); 907 if (mtx != new_mtx) { 908 if (mtx != NULL) 909 mtx_unlock(mtx); 910 mtx = new_mtx; 911 mtx_lock(mtx); 912 } 913 vm_page_unhold(*ma); 914 ma++; 915 } 916 if (mtx != NULL) 917 mtx_unlock(mtx); 918 } 919 920 vm_page_t 921 PHYS_TO_VM_PAGE(vm_paddr_t pa) 922 { 923 vm_page_t m; 924 925 #ifdef VM_PHYSSEG_SPARSE 926 m = vm_phys_paddr_to_vm_page(pa); 927 if (m == NULL) 928 m = vm_phys_fictitious_to_vm_page(pa); 929 return (m); 930 #elif defined(VM_PHYSSEG_DENSE) 931 long pi; 932 933 pi = atop(pa); 934 if (pi >= first_page && (pi - first_page) < vm_page_array_size) { 935 m = &vm_page_array[pi - first_page]; 936 return (m); 937 } 938 return (vm_phys_fictitious_to_vm_page(pa)); 939 #else 940 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 941 #endif 942 } 943 944 /* 945 * vm_page_getfake: 946 * 947 * Create a fictitious page with the specified physical address and 948 * memory attribute. The memory attribute is the only the machine- 949 * dependent aspect of a fictitious page that must be initialized. 950 */ 951 vm_page_t 952 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) 953 { 954 vm_page_t m; 955 956 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); 957 vm_page_initfake(m, paddr, memattr); 958 return (m); 959 } 960 961 void 962 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 963 { 964 965 if ((m->flags & PG_FICTITIOUS) != 0) { 966 /* 967 * The page's memattr might have changed since the 968 * previous initialization. Update the pmap to the 969 * new memattr. 970 */ 971 goto memattr; 972 } 973 m->phys_addr = paddr; 974 m->queue = PQ_NONE; 975 /* Fictitious pages don't use "segind". */ 976 m->flags = PG_FICTITIOUS; 977 /* Fictitious pages don't use "order" or "pool". */ 978 m->oflags = VPO_UNMANAGED; 979 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 980 m->wire_count = 1; 981 pmap_page_init(m); 982 memattr: 983 pmap_page_set_memattr(m, memattr); 984 } 985 986 /* 987 * vm_page_putfake: 988 * 989 * Release a fictitious page. 990 */ 991 void 992 vm_page_putfake(vm_page_t m) 993 { 994 995 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); 996 KASSERT((m->flags & PG_FICTITIOUS) != 0, 997 ("vm_page_putfake: bad page %p", m)); 998 uma_zfree(fakepg_zone, m); 999 } 1000 1001 /* 1002 * vm_page_updatefake: 1003 * 1004 * Update the given fictitious page to the specified physical address and 1005 * memory attribute. 1006 */ 1007 void 1008 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) 1009 { 1010 1011 KASSERT((m->flags & PG_FICTITIOUS) != 0, 1012 ("vm_page_updatefake: bad page %p", m)); 1013 m->phys_addr = paddr; 1014 pmap_page_set_memattr(m, memattr); 1015 } 1016 1017 /* 1018 * vm_page_free: 1019 * 1020 * Free a page. 1021 */ 1022 void 1023 vm_page_free(vm_page_t m) 1024 { 1025 1026 m->flags &= ~PG_ZERO; 1027 vm_page_free_toq(m); 1028 } 1029 1030 /* 1031 * vm_page_free_zero: 1032 * 1033 * Free a page to the zerod-pages queue 1034 */ 1035 void 1036 vm_page_free_zero(vm_page_t m) 1037 { 1038 1039 m->flags |= PG_ZERO; 1040 vm_page_free_toq(m); 1041 } 1042 1043 /* 1044 * Unbusy and handle the page queueing for a page from a getpages request that 1045 * was optionally read ahead or behind. 1046 */ 1047 void 1048 vm_page_readahead_finish(vm_page_t m) 1049 { 1050 1051 /* We shouldn't put invalid pages on queues. */ 1052 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m)); 1053 1054 /* 1055 * Since the page is not the actually needed one, whether it should 1056 * be activated or deactivated is not obvious. Empirical results 1057 * have shown that deactivating the page is usually the best choice, 1058 * unless the page is wanted by another thread. 1059 */ 1060 vm_page_lock(m); 1061 if ((m->busy_lock & VPB_BIT_WAITERS) != 0) 1062 vm_page_activate(m); 1063 else 1064 vm_page_deactivate(m); 1065 vm_page_unlock(m); 1066 vm_page_xunbusy(m); 1067 } 1068 1069 /* 1070 * vm_page_sleep_if_busy: 1071 * 1072 * Sleep and release the page queues lock if the page is busied. 1073 * Returns TRUE if the thread slept. 1074 * 1075 * The given page must be unlocked and object containing it must 1076 * be locked. 1077 */ 1078 int 1079 vm_page_sleep_if_busy(vm_page_t m, const char *msg) 1080 { 1081 vm_object_t obj; 1082 1083 vm_page_lock_assert(m, MA_NOTOWNED); 1084 VM_OBJECT_ASSERT_WLOCKED(m->object); 1085 1086 if (vm_page_busied(m)) { 1087 /* 1088 * The page-specific object must be cached because page 1089 * identity can change during the sleep, causing the 1090 * re-lock of a different object. 1091 * It is assumed that a reference to the object is already 1092 * held by the callers. 1093 */ 1094 obj = m->object; 1095 vm_page_lock(m); 1096 VM_OBJECT_WUNLOCK(obj); 1097 vm_page_busy_sleep(m, msg, false); 1098 VM_OBJECT_WLOCK(obj); 1099 return (TRUE); 1100 } 1101 return (FALSE); 1102 } 1103 1104 /* 1105 * vm_page_dirty_KBI: [ internal use only ] 1106 * 1107 * Set all bits in the page's dirty field. 1108 * 1109 * The object containing the specified page must be locked if the 1110 * call is made from the machine-independent layer. 1111 * 1112 * See vm_page_clear_dirty_mask(). 1113 * 1114 * This function should only be called by vm_page_dirty(). 1115 */ 1116 void 1117 vm_page_dirty_KBI(vm_page_t m) 1118 { 1119 1120 /* Refer to this operation by its public name. */ 1121 KASSERT(m->valid == VM_PAGE_BITS_ALL, 1122 ("vm_page_dirty: page is invalid!")); 1123 m->dirty = VM_PAGE_BITS_ALL; 1124 } 1125 1126 /* 1127 * vm_page_insert: [ internal use only ] 1128 * 1129 * Inserts the given mem entry into the object and object list. 1130 * 1131 * The object must be locked. 1132 */ 1133 int 1134 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 1135 { 1136 vm_page_t mpred; 1137 1138 VM_OBJECT_ASSERT_WLOCKED(object); 1139 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1140 return (vm_page_insert_after(m, object, pindex, mpred)); 1141 } 1142 1143 /* 1144 * vm_page_insert_after: 1145 * 1146 * Inserts the page "m" into the specified object at offset "pindex". 1147 * 1148 * The page "mpred" must immediately precede the offset "pindex" within 1149 * the specified object. 1150 * 1151 * The object must be locked. 1152 */ 1153 static int 1154 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, 1155 vm_page_t mpred) 1156 { 1157 vm_page_t msucc; 1158 1159 VM_OBJECT_ASSERT_WLOCKED(object); 1160 KASSERT(m->object == NULL, 1161 ("vm_page_insert_after: page already inserted")); 1162 if (mpred != NULL) { 1163 KASSERT(mpred->object == object, 1164 ("vm_page_insert_after: object doesn't contain mpred")); 1165 KASSERT(mpred->pindex < pindex, 1166 ("vm_page_insert_after: mpred doesn't precede pindex")); 1167 msucc = TAILQ_NEXT(mpred, listq); 1168 } else 1169 msucc = TAILQ_FIRST(&object->memq); 1170 if (msucc != NULL) 1171 KASSERT(msucc->pindex > pindex, 1172 ("vm_page_insert_after: msucc doesn't succeed pindex")); 1173 1174 /* 1175 * Record the object/offset pair in this page 1176 */ 1177 m->object = object; 1178 m->pindex = pindex; 1179 1180 /* 1181 * Now link into the object's ordered list of backed pages. 1182 */ 1183 if (vm_radix_insert(&object->rtree, m)) { 1184 m->object = NULL; 1185 m->pindex = 0; 1186 return (1); 1187 } 1188 vm_page_insert_radixdone(m, object, mpred); 1189 return (0); 1190 } 1191 1192 /* 1193 * vm_page_insert_radixdone: 1194 * 1195 * Complete page "m" insertion into the specified object after the 1196 * radix trie hooking. 1197 * 1198 * The page "mpred" must precede the offset "m->pindex" within the 1199 * specified object. 1200 * 1201 * The object must be locked. 1202 */ 1203 static void 1204 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) 1205 { 1206 1207 VM_OBJECT_ASSERT_WLOCKED(object); 1208 KASSERT(object != NULL && m->object == object, 1209 ("vm_page_insert_radixdone: page %p has inconsistent object", m)); 1210 if (mpred != NULL) { 1211 KASSERT(mpred->object == object, 1212 ("vm_page_insert_after: object doesn't contain mpred")); 1213 KASSERT(mpred->pindex < m->pindex, 1214 ("vm_page_insert_after: mpred doesn't precede pindex")); 1215 } 1216 1217 if (mpred != NULL) 1218 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); 1219 else 1220 TAILQ_INSERT_HEAD(&object->memq, m, listq); 1221 1222 /* 1223 * Show that the object has one more resident page. 1224 */ 1225 object->resident_page_count++; 1226 1227 /* 1228 * Hold the vnode until the last page is released. 1229 */ 1230 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 1231 vhold(object->handle); 1232 1233 /* 1234 * Since we are inserting a new and possibly dirty page, 1235 * update the object's OBJ_MIGHTBEDIRTY flag. 1236 */ 1237 if (pmap_page_is_write_mapped(m)) 1238 vm_object_set_writeable_dirty(object); 1239 } 1240 1241 /* 1242 * vm_page_remove: 1243 * 1244 * Removes the specified page from its containing object, but does not 1245 * invalidate any backing storage. 1246 * 1247 * The object must be locked. The page must be locked if it is managed. 1248 */ 1249 void 1250 vm_page_remove(vm_page_t m) 1251 { 1252 vm_object_t object; 1253 vm_page_t mrem; 1254 1255 if ((m->oflags & VPO_UNMANAGED) == 0) 1256 vm_page_assert_locked(m); 1257 if ((object = m->object) == NULL) 1258 return; 1259 VM_OBJECT_ASSERT_WLOCKED(object); 1260 if (vm_page_xbusied(m)) 1261 vm_page_xunbusy_maybelocked(m); 1262 mrem = vm_radix_remove(&object->rtree, m->pindex); 1263 KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); 1264 1265 /* 1266 * Now remove from the object's list of backed pages. 1267 */ 1268 TAILQ_REMOVE(&object->memq, m, listq); 1269 1270 /* 1271 * And show that the object has one fewer resident page. 1272 */ 1273 object->resident_page_count--; 1274 1275 /* 1276 * The vnode may now be recycled. 1277 */ 1278 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 1279 vdrop(object->handle); 1280 1281 m->object = NULL; 1282 } 1283 1284 /* 1285 * vm_page_lookup: 1286 * 1287 * Returns the page associated with the object/offset 1288 * pair specified; if none is found, NULL is returned. 1289 * 1290 * The object must be locked. 1291 */ 1292 vm_page_t 1293 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 1294 { 1295 1296 VM_OBJECT_ASSERT_LOCKED(object); 1297 return (vm_radix_lookup(&object->rtree, pindex)); 1298 } 1299 1300 /* 1301 * vm_page_find_least: 1302 * 1303 * Returns the page associated with the object with least pindex 1304 * greater than or equal to the parameter pindex, or NULL. 1305 * 1306 * The object must be locked. 1307 */ 1308 vm_page_t 1309 vm_page_find_least(vm_object_t object, vm_pindex_t pindex) 1310 { 1311 vm_page_t m; 1312 1313 VM_OBJECT_ASSERT_LOCKED(object); 1314 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) 1315 m = vm_radix_lookup_ge(&object->rtree, pindex); 1316 return (m); 1317 } 1318 1319 /* 1320 * Returns the given page's successor (by pindex) within the object if it is 1321 * resident; if none is found, NULL is returned. 1322 * 1323 * The object must be locked. 1324 */ 1325 vm_page_t 1326 vm_page_next(vm_page_t m) 1327 { 1328 vm_page_t next; 1329 1330 VM_OBJECT_ASSERT_LOCKED(m->object); 1331 if ((next = TAILQ_NEXT(m, listq)) != NULL) { 1332 MPASS(next->object == m->object); 1333 if (next->pindex != m->pindex + 1) 1334 next = NULL; 1335 } 1336 return (next); 1337 } 1338 1339 /* 1340 * Returns the given page's predecessor (by pindex) within the object if it is 1341 * resident; if none is found, NULL is returned. 1342 * 1343 * The object must be locked. 1344 */ 1345 vm_page_t 1346 vm_page_prev(vm_page_t m) 1347 { 1348 vm_page_t prev; 1349 1350 VM_OBJECT_ASSERT_LOCKED(m->object); 1351 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { 1352 MPASS(prev->object == m->object); 1353 if (prev->pindex != m->pindex - 1) 1354 prev = NULL; 1355 } 1356 return (prev); 1357 } 1358 1359 /* 1360 * Uses the page mnew as a replacement for an existing page at index 1361 * pindex which must be already present in the object. 1362 * 1363 * The existing page must not be on a paging queue. 1364 */ 1365 vm_page_t 1366 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1367 { 1368 vm_page_t mold; 1369 1370 VM_OBJECT_ASSERT_WLOCKED(object); 1371 KASSERT(mnew->object == NULL, 1372 ("vm_page_replace: page already in object")); 1373 1374 /* 1375 * This function mostly follows vm_page_insert() and 1376 * vm_page_remove() without the radix, object count and vnode 1377 * dance. Double check such functions for more comments. 1378 */ 1379 1380 mnew->object = object; 1381 mnew->pindex = pindex; 1382 mold = vm_radix_replace(&object->rtree, mnew); 1383 KASSERT(mold->queue == PQ_NONE, 1384 ("vm_page_replace: mold is on a paging queue")); 1385 1386 /* Keep the resident page list in sorted order. */ 1387 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1388 TAILQ_REMOVE(&object->memq, mold, listq); 1389 1390 mold->object = NULL; 1391 vm_page_xunbusy_maybelocked(mold); 1392 1393 /* 1394 * The object's resident_page_count does not change because we have 1395 * swapped one page for another, but OBJ_MIGHTBEDIRTY. 1396 */ 1397 if (pmap_page_is_write_mapped(mnew)) 1398 vm_object_set_writeable_dirty(object); 1399 return (mold); 1400 } 1401 1402 /* 1403 * vm_page_rename: 1404 * 1405 * Move the given memory entry from its 1406 * current object to the specified target object/offset. 1407 * 1408 * Note: swap associated with the page must be invalidated by the move. We 1409 * have to do this for several reasons: (1) we aren't freeing the 1410 * page, (2) we are dirtying the page, (3) the VM system is probably 1411 * moving the page from object A to B, and will then later move 1412 * the backing store from A to B and we can't have a conflict. 1413 * 1414 * Note: we *always* dirty the page. It is necessary both for the 1415 * fact that we moved it, and because we may be invalidating 1416 * swap. 1417 * 1418 * The objects must be locked. 1419 */ 1420 int 1421 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1422 { 1423 vm_page_t mpred; 1424 vm_pindex_t opidx; 1425 1426 VM_OBJECT_ASSERT_WLOCKED(new_object); 1427 1428 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1429 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1430 ("vm_page_rename: pindex already renamed")); 1431 1432 /* 1433 * Create a custom version of vm_page_insert() which does not depend 1434 * by m_prev and can cheat on the implementation aspects of the 1435 * function. 1436 */ 1437 opidx = m->pindex; 1438 m->pindex = new_pindex; 1439 if (vm_radix_insert(&new_object->rtree, m)) { 1440 m->pindex = opidx; 1441 return (1); 1442 } 1443 1444 /* 1445 * The operation cannot fail anymore. The removal must happen before 1446 * the listq iterator is tainted. 1447 */ 1448 m->pindex = opidx; 1449 vm_page_lock(m); 1450 vm_page_remove(m); 1451 1452 /* Return back to the new pindex to complete vm_page_insert(). */ 1453 m->pindex = new_pindex; 1454 m->object = new_object; 1455 vm_page_unlock(m); 1456 vm_page_insert_radixdone(m, new_object, mpred); 1457 vm_page_dirty(m); 1458 return (0); 1459 } 1460 1461 /* 1462 * vm_page_alloc: 1463 * 1464 * Allocate and return a page that is associated with the specified 1465 * object and offset pair. By default, this page is exclusive busied. 1466 * 1467 * The caller must always specify an allocation class. 1468 * 1469 * allocation classes: 1470 * VM_ALLOC_NORMAL normal process request 1471 * VM_ALLOC_SYSTEM system *really* needs a page 1472 * VM_ALLOC_INTERRUPT interrupt time request 1473 * 1474 * optional allocation flags: 1475 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1476 * intends to allocate 1477 * VM_ALLOC_NOBUSY do not exclusive busy the page 1478 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1479 * VM_ALLOC_NOOBJ page is not associated with an object and 1480 * should not be exclusive busy 1481 * VM_ALLOC_SBUSY shared busy the allocated page 1482 * VM_ALLOC_WIRED wire the allocated page 1483 * VM_ALLOC_ZERO prefer a zeroed page 1484 * 1485 * This routine may not sleep. 1486 */ 1487 vm_page_t 1488 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1489 { 1490 vm_page_t m, mpred; 1491 int flags, req_class; 1492 1493 mpred = NULL; /* XXX: pacify gcc */ 1494 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1495 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1496 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1497 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1498 ("vm_page_alloc: inconsistent object(%p)/req(%x)", object, req)); 1499 if (object != NULL) 1500 VM_OBJECT_ASSERT_WLOCKED(object); 1501 1502 req_class = req & VM_ALLOC_CLASS_MASK; 1503 1504 /* 1505 * The page daemon is allowed to dig deeper into the free page list. 1506 */ 1507 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1508 req_class = VM_ALLOC_SYSTEM; 1509 1510 if (object != NULL) { 1511 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1512 KASSERT(mpred == NULL || mpred->pindex != pindex, 1513 ("vm_page_alloc: pindex already allocated")); 1514 } 1515 1516 /* 1517 * Allocate a page if the number of free pages exceeds the minimum 1518 * for the request class. 1519 */ 1520 mtx_lock(&vm_page_queue_free_mtx); 1521 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved || 1522 (req_class == VM_ALLOC_SYSTEM && 1523 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) || 1524 (req_class == VM_ALLOC_INTERRUPT && 1525 vm_cnt.v_free_count > 0)) { 1526 /* 1527 * Can we allocate the page from a reservation? 1528 */ 1529 #if VM_NRESERVLEVEL > 0 1530 if (object == NULL || (object->flags & (OBJ_COLORED | 1531 OBJ_FICTITIOUS)) != OBJ_COLORED || (m = 1532 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) 1533 #endif 1534 { 1535 /* 1536 * If not, allocate it from the free page queues. 1537 */ 1538 m = vm_phys_alloc_pages(object != NULL ? 1539 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1540 #if VM_NRESERVLEVEL > 0 1541 if (m == NULL && vm_reserv_reclaim_inactive()) { 1542 m = vm_phys_alloc_pages(object != NULL ? 1543 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1544 0); 1545 } 1546 #endif 1547 } 1548 } else { 1549 /* 1550 * Not allocatable, give up. 1551 */ 1552 mtx_unlock(&vm_page_queue_free_mtx); 1553 atomic_add_int(&vm_pageout_deficit, 1554 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1555 pagedaemon_wakeup(); 1556 return (NULL); 1557 } 1558 1559 /* 1560 * At this point we had better have found a good page. 1561 */ 1562 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1563 vm_phys_freecnt_adj(m, -1); 1564 mtx_unlock(&vm_page_queue_free_mtx); 1565 vm_page_alloc_check(m); 1566 1567 /* 1568 * Initialize the page. Only the PG_ZERO flag is inherited. 1569 */ 1570 flags = 0; 1571 if ((req & VM_ALLOC_ZERO) != 0) 1572 flags = PG_ZERO; 1573 flags &= m->flags; 1574 if ((req & VM_ALLOC_NODUMP) != 0) 1575 flags |= PG_NODUMP; 1576 m->flags = flags; 1577 m->aflags = 0; 1578 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1579 VPO_UNMANAGED : 0; 1580 m->busy_lock = VPB_UNBUSIED; 1581 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1582 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1583 if ((req & VM_ALLOC_SBUSY) != 0) 1584 m->busy_lock = VPB_SHARERS_WORD(1); 1585 if (req & VM_ALLOC_WIRED) { 1586 /* 1587 * The page lock is not required for wiring a page until that 1588 * page is inserted into the object. 1589 */ 1590 atomic_add_int(&vm_cnt.v_wire_count, 1); 1591 m->wire_count = 1; 1592 } 1593 m->act_count = 0; 1594 1595 if (object != NULL) { 1596 if (vm_page_insert_after(m, object, pindex, mpred)) { 1597 pagedaemon_wakeup(); 1598 if (req & VM_ALLOC_WIRED) { 1599 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 1600 m->wire_count = 0; 1601 } 1602 KASSERT(m->object == NULL, ("page %p has object", m)); 1603 m->oflags = VPO_UNMANAGED; 1604 m->busy_lock = VPB_UNBUSIED; 1605 /* Don't change PG_ZERO. */ 1606 vm_page_free_toq(m); 1607 return (NULL); 1608 } 1609 1610 /* Ignore device objects; the pager sets "memattr" for them. */ 1611 if (object->memattr != VM_MEMATTR_DEFAULT && 1612 (object->flags & OBJ_FICTITIOUS) == 0) 1613 pmap_page_set_memattr(m, object->memattr); 1614 } else 1615 m->pindex = pindex; 1616 1617 /* 1618 * Don't wakeup too often - wakeup the pageout daemon when 1619 * we would be nearly out of memory. 1620 */ 1621 if (vm_paging_needed()) 1622 pagedaemon_wakeup(); 1623 1624 return (m); 1625 } 1626 1627 /* 1628 * vm_page_alloc_contig: 1629 * 1630 * Allocate a contiguous set of physical pages of the given size "npages" 1631 * from the free lists. All of the physical pages must be at or above 1632 * the given physical address "low" and below the given physical address 1633 * "high". The given value "alignment" determines the alignment of the 1634 * first physical page in the set. If the given value "boundary" is 1635 * non-zero, then the set of physical pages cannot cross any physical 1636 * address boundary that is a multiple of that value. Both "alignment" 1637 * and "boundary" must be a power of two. 1638 * 1639 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1640 * then the memory attribute setting for the physical pages is configured 1641 * to the object's memory attribute setting. Otherwise, the memory 1642 * attribute setting for the physical pages is configured to "memattr", 1643 * overriding the object's memory attribute setting. However, if the 1644 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1645 * memory attribute setting for the physical pages cannot be configured 1646 * to VM_MEMATTR_DEFAULT. 1647 * 1648 * The specified object may not contain fictitious pages. 1649 * 1650 * The caller must always specify an allocation class. 1651 * 1652 * allocation classes: 1653 * VM_ALLOC_NORMAL normal process request 1654 * VM_ALLOC_SYSTEM system *really* needs a page 1655 * VM_ALLOC_INTERRUPT interrupt time request 1656 * 1657 * optional allocation flags: 1658 * VM_ALLOC_NOBUSY do not exclusive busy the page 1659 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1660 * VM_ALLOC_NOOBJ page is not associated with an object and 1661 * should not be exclusive busy 1662 * VM_ALLOC_SBUSY shared busy the allocated page 1663 * VM_ALLOC_WIRED wire the allocated page 1664 * VM_ALLOC_ZERO prefer a zeroed page 1665 * 1666 * This routine may not sleep. 1667 */ 1668 vm_page_t 1669 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1670 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1671 vm_paddr_t boundary, vm_memattr_t memattr) 1672 { 1673 vm_page_t m, m_ret, mpred; 1674 u_int busy_lock, flags, oflags; 1675 int req_class; 1676 1677 mpred = NULL; /* XXX: pacify gcc */ 1678 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1679 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1680 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1681 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1682 ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, 1683 req)); 1684 if (object != NULL) { 1685 VM_OBJECT_ASSERT_WLOCKED(object); 1686 KASSERT((object->flags & OBJ_FICTITIOUS) == 0, 1687 ("vm_page_alloc_contig: object %p has fictitious pages", 1688 object)); 1689 } 1690 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1691 req_class = req & VM_ALLOC_CLASS_MASK; 1692 1693 /* 1694 * The page daemon is allowed to dig deeper into the free page list. 1695 */ 1696 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1697 req_class = VM_ALLOC_SYSTEM; 1698 1699 if (object != NULL) { 1700 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1701 KASSERT(mpred == NULL || mpred->pindex != pindex, 1702 ("vm_page_alloc_contig: pindex already allocated")); 1703 } 1704 1705 /* 1706 * Can we allocate the pages without the number of free pages falling 1707 * below the lower bound for the allocation class? 1708 */ 1709 mtx_lock(&vm_page_queue_free_mtx); 1710 if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved || 1711 (req_class == VM_ALLOC_SYSTEM && 1712 vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) || 1713 (req_class == VM_ALLOC_INTERRUPT && 1714 vm_cnt.v_free_count >= npages)) { 1715 /* 1716 * Can we allocate the pages from a reservation? 1717 */ 1718 #if VM_NRESERVLEVEL > 0 1719 retry: 1720 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1721 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1722 low, high, alignment, boundary, mpred)) == NULL) 1723 #endif 1724 /* 1725 * If not, allocate them from the free page queues. 1726 */ 1727 m_ret = vm_phys_alloc_contig(npages, low, high, 1728 alignment, boundary); 1729 } else { 1730 mtx_unlock(&vm_page_queue_free_mtx); 1731 atomic_add_int(&vm_pageout_deficit, npages); 1732 pagedaemon_wakeup(); 1733 return (NULL); 1734 } 1735 if (m_ret != NULL) 1736 vm_phys_freecnt_adj(m_ret, -npages); 1737 else { 1738 #if VM_NRESERVLEVEL > 0 1739 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1740 boundary)) 1741 goto retry; 1742 #endif 1743 } 1744 mtx_unlock(&vm_page_queue_free_mtx); 1745 if (m_ret == NULL) 1746 return (NULL); 1747 for (m = m_ret; m < &m_ret[npages]; m++) 1748 vm_page_alloc_check(m); 1749 1750 /* 1751 * Initialize the pages. Only the PG_ZERO flag is inherited. 1752 */ 1753 flags = 0; 1754 if ((req & VM_ALLOC_ZERO) != 0) 1755 flags = PG_ZERO; 1756 if ((req & VM_ALLOC_NODUMP) != 0) 1757 flags |= PG_NODUMP; 1758 oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1759 VPO_UNMANAGED : 0; 1760 busy_lock = VPB_UNBUSIED; 1761 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1762 busy_lock = VPB_SINGLE_EXCLUSIVER; 1763 if ((req & VM_ALLOC_SBUSY) != 0) 1764 busy_lock = VPB_SHARERS_WORD(1); 1765 if ((req & VM_ALLOC_WIRED) != 0) 1766 atomic_add_int(&vm_cnt.v_wire_count, npages); 1767 if (object != NULL) { 1768 if (object->memattr != VM_MEMATTR_DEFAULT && 1769 memattr == VM_MEMATTR_DEFAULT) 1770 memattr = object->memattr; 1771 } 1772 for (m = m_ret; m < &m_ret[npages]; m++) { 1773 m->aflags = 0; 1774 m->flags = (m->flags | PG_NODUMP) & flags; 1775 m->busy_lock = busy_lock; 1776 if ((req & VM_ALLOC_WIRED) != 0) 1777 m->wire_count = 1; 1778 m->act_count = 0; 1779 m->oflags = oflags; 1780 if (object != NULL) { 1781 if (vm_page_insert_after(m, object, pindex, mpred)) { 1782 pagedaemon_wakeup(); 1783 if ((req & VM_ALLOC_WIRED) != 0) 1784 atomic_subtract_int( 1785 &vm_cnt.v_wire_count, npages); 1786 KASSERT(m->object == NULL, 1787 ("page %p has object", m)); 1788 mpred = m; 1789 for (m = m_ret; m < &m_ret[npages]; m++) { 1790 if (m <= mpred && 1791 (req & VM_ALLOC_WIRED) != 0) 1792 m->wire_count = 0; 1793 m->oflags = VPO_UNMANAGED; 1794 m->busy_lock = VPB_UNBUSIED; 1795 /* Don't change PG_ZERO. */ 1796 vm_page_free_toq(m); 1797 } 1798 return (NULL); 1799 } 1800 mpred = m; 1801 } else 1802 m->pindex = pindex; 1803 if (memattr != VM_MEMATTR_DEFAULT) 1804 pmap_page_set_memattr(m, memattr); 1805 pindex++; 1806 } 1807 if (vm_paging_needed()) 1808 pagedaemon_wakeup(); 1809 return (m_ret); 1810 } 1811 1812 /* 1813 * Check a page that has been freshly dequeued from a freelist. 1814 */ 1815 static void 1816 vm_page_alloc_check(vm_page_t m) 1817 { 1818 1819 KASSERT(m->object == NULL, ("page %p has object", m)); 1820 KASSERT(m->queue == PQ_NONE, 1821 ("page %p has unexpected queue %d", m, m->queue)); 1822 KASSERT(m->wire_count == 0, ("page %p is wired", m)); 1823 KASSERT(m->hold_count == 0, ("page %p is held", m)); 1824 KASSERT(!vm_page_busied(m), ("page %p is busy", m)); 1825 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 1826 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1827 ("page %p has unexpected memattr %d", 1828 m, pmap_page_get_memattr(m))); 1829 KASSERT(m->valid == 0, ("free page %p is valid", m)); 1830 } 1831 1832 /* 1833 * vm_page_alloc_freelist: 1834 * 1835 * Allocate a physical page from the specified free page list. 1836 * 1837 * The caller must always specify an allocation class. 1838 * 1839 * allocation classes: 1840 * VM_ALLOC_NORMAL normal process request 1841 * VM_ALLOC_SYSTEM system *really* needs a page 1842 * VM_ALLOC_INTERRUPT interrupt time request 1843 * 1844 * optional allocation flags: 1845 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1846 * intends to allocate 1847 * VM_ALLOC_WIRED wire the allocated page 1848 * VM_ALLOC_ZERO prefer a zeroed page 1849 * 1850 * This routine may not sleep. 1851 */ 1852 vm_page_t 1853 vm_page_alloc_freelist(int flind, int req) 1854 { 1855 vm_page_t m; 1856 u_int flags; 1857 int req_class; 1858 1859 req_class = req & VM_ALLOC_CLASS_MASK; 1860 1861 /* 1862 * The page daemon is allowed to dig deeper into the free page list. 1863 */ 1864 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1865 req_class = VM_ALLOC_SYSTEM; 1866 1867 /* 1868 * Do not allocate reserved pages unless the req has asked for it. 1869 */ 1870 mtx_lock(&vm_page_queue_free_mtx); 1871 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved || 1872 (req_class == VM_ALLOC_SYSTEM && 1873 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) || 1874 (req_class == VM_ALLOC_INTERRUPT && 1875 vm_cnt.v_free_count > 0)) 1876 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1877 else { 1878 mtx_unlock(&vm_page_queue_free_mtx); 1879 atomic_add_int(&vm_pageout_deficit, 1880 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1881 pagedaemon_wakeup(); 1882 return (NULL); 1883 } 1884 if (m == NULL) { 1885 mtx_unlock(&vm_page_queue_free_mtx); 1886 return (NULL); 1887 } 1888 vm_phys_freecnt_adj(m, -1); 1889 mtx_unlock(&vm_page_queue_free_mtx); 1890 vm_page_alloc_check(m); 1891 1892 /* 1893 * Initialize the page. Only the PG_ZERO flag is inherited. 1894 */ 1895 m->aflags = 0; 1896 flags = 0; 1897 if ((req & VM_ALLOC_ZERO) != 0) 1898 flags = PG_ZERO; 1899 m->flags &= flags; 1900 if ((req & VM_ALLOC_WIRED) != 0) { 1901 /* 1902 * The page lock is not required for wiring a page that does 1903 * not belong to an object. 1904 */ 1905 atomic_add_int(&vm_cnt.v_wire_count, 1); 1906 m->wire_count = 1; 1907 } 1908 /* Unmanaged pages don't use "act_count". */ 1909 m->oflags = VPO_UNMANAGED; 1910 if (vm_paging_needed()) 1911 pagedaemon_wakeup(); 1912 return (m); 1913 } 1914 1915 #define VPSC_ANY 0 /* No restrictions. */ 1916 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 1917 #define VPSC_NOSUPER 2 /* Skip superpages. */ 1918 1919 /* 1920 * vm_page_scan_contig: 1921 * 1922 * Scan vm_page_array[] between the specified entries "m_start" and 1923 * "m_end" for a run of contiguous physical pages that satisfy the 1924 * specified conditions, and return the lowest page in the run. The 1925 * specified "alignment" determines the alignment of the lowest physical 1926 * page in the run. If the specified "boundary" is non-zero, then the 1927 * run of physical pages cannot span a physical address that is a 1928 * multiple of "boundary". 1929 * 1930 * "m_end" is never dereferenced, so it need not point to a vm_page 1931 * structure within vm_page_array[]. 1932 * 1933 * "npages" must be greater than zero. "m_start" and "m_end" must not 1934 * span a hole (or discontiguity) in the physical address space. Both 1935 * "alignment" and "boundary" must be a power of two. 1936 */ 1937 vm_page_t 1938 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 1939 u_long alignment, vm_paddr_t boundary, int options) 1940 { 1941 struct mtx *m_mtx, *new_mtx; 1942 vm_object_t object; 1943 vm_paddr_t pa; 1944 vm_page_t m, m_run; 1945 #if VM_NRESERVLEVEL > 0 1946 int level; 1947 #endif 1948 int m_inc, order, run_ext, run_len; 1949 1950 KASSERT(npages > 0, ("npages is 0")); 1951 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 1952 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 1953 m_run = NULL; 1954 run_len = 0; 1955 m_mtx = NULL; 1956 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 1957 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 1958 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 1959 1960 /* 1961 * If the current page would be the start of a run, check its 1962 * physical address against the end, alignment, and boundary 1963 * conditions. If it doesn't satisfy these conditions, either 1964 * terminate the scan or advance to the next page that 1965 * satisfies the failed condition. 1966 */ 1967 if (run_len == 0) { 1968 KASSERT(m_run == NULL, ("m_run != NULL")); 1969 if (m + npages > m_end) 1970 break; 1971 pa = VM_PAGE_TO_PHYS(m); 1972 if ((pa & (alignment - 1)) != 0) { 1973 m_inc = atop(roundup2(pa, alignment) - pa); 1974 continue; 1975 } 1976 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 1977 boundary) != 0) { 1978 m_inc = atop(roundup2(pa, boundary) - pa); 1979 continue; 1980 } 1981 } else 1982 KASSERT(m_run != NULL, ("m_run == NULL")); 1983 1984 /* 1985 * Avoid releasing and reacquiring the same page lock. 1986 */ 1987 new_mtx = vm_page_lockptr(m); 1988 if (m_mtx != new_mtx) { 1989 if (m_mtx != NULL) 1990 mtx_unlock(m_mtx); 1991 m_mtx = new_mtx; 1992 mtx_lock(m_mtx); 1993 } 1994 m_inc = 1; 1995 retry: 1996 if (m->wire_count != 0 || m->hold_count != 0) 1997 run_ext = 0; 1998 #if VM_NRESERVLEVEL > 0 1999 else if ((level = vm_reserv_level(m)) >= 0 && 2000 (options & VPSC_NORESERV) != 0) { 2001 run_ext = 0; 2002 /* Advance to the end of the reservation. */ 2003 pa = VM_PAGE_TO_PHYS(m); 2004 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 2005 pa); 2006 } 2007 #endif 2008 else if ((object = m->object) != NULL) { 2009 /* 2010 * The page is considered eligible for relocation if 2011 * and only if it could be laundered or reclaimed by 2012 * the page daemon. 2013 */ 2014 if (!VM_OBJECT_TRYRLOCK(object)) { 2015 mtx_unlock(m_mtx); 2016 VM_OBJECT_RLOCK(object); 2017 mtx_lock(m_mtx); 2018 if (m->object != object) { 2019 /* 2020 * The page may have been freed. 2021 */ 2022 VM_OBJECT_RUNLOCK(object); 2023 goto retry; 2024 } else if (m->wire_count != 0 || 2025 m->hold_count != 0) { 2026 run_ext = 0; 2027 goto unlock; 2028 } 2029 } 2030 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2031 ("page %p is PG_UNHOLDFREE", m)); 2032 /* Don't care: PG_NODUMP, PG_ZERO. */ 2033 if (object->type != OBJT_DEFAULT && 2034 object->type != OBJT_SWAP && 2035 object->type != OBJT_VNODE) { 2036 run_ext = 0; 2037 #if VM_NRESERVLEVEL > 0 2038 } else if ((options & VPSC_NOSUPER) != 0 && 2039 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2040 run_ext = 0; 2041 /* Advance to the end of the superpage. */ 2042 pa = VM_PAGE_TO_PHYS(m); 2043 m_inc = atop(roundup2(pa + 1, 2044 vm_reserv_size(level)) - pa); 2045 #endif 2046 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2047 m->queue != PQ_NONE && !vm_page_busied(m)) { 2048 /* 2049 * The page is allocated but eligible for 2050 * relocation. Extend the current run by one 2051 * page. 2052 */ 2053 KASSERT(pmap_page_get_memattr(m) == 2054 VM_MEMATTR_DEFAULT, 2055 ("page %p has an unexpected memattr", m)); 2056 KASSERT((m->oflags & (VPO_SWAPINPROG | 2057 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2058 ("page %p has unexpected oflags", m)); 2059 /* Don't care: VPO_NOSYNC. */ 2060 run_ext = 1; 2061 } else 2062 run_ext = 0; 2063 unlock: 2064 VM_OBJECT_RUNLOCK(object); 2065 #if VM_NRESERVLEVEL > 0 2066 } else if (level >= 0) { 2067 /* 2068 * The page is reserved but not yet allocated. In 2069 * other words, it is still free. Extend the current 2070 * run by one page. 2071 */ 2072 run_ext = 1; 2073 #endif 2074 } else if ((order = m->order) < VM_NFREEORDER) { 2075 /* 2076 * The page is enqueued in the physical memory 2077 * allocator's free page queues. Moreover, it is the 2078 * first page in a power-of-two-sized run of 2079 * contiguous free pages. Add these pages to the end 2080 * of the current run, and jump ahead. 2081 */ 2082 run_ext = 1 << order; 2083 m_inc = 1 << order; 2084 } else { 2085 /* 2086 * Skip the page for one of the following reasons: (1) 2087 * It is enqueued in the physical memory allocator's 2088 * free page queues. However, it is not the first 2089 * page in a run of contiguous free pages. (This case 2090 * rarely occurs because the scan is performed in 2091 * ascending order.) (2) It is not reserved, and it is 2092 * transitioning from free to allocated. (Conversely, 2093 * the transition from allocated to free for managed 2094 * pages is blocked by the page lock.) (3) It is 2095 * allocated but not contained by an object and not 2096 * wired, e.g., allocated by Xen's balloon driver. 2097 */ 2098 run_ext = 0; 2099 } 2100 2101 /* 2102 * Extend or reset the current run of pages. 2103 */ 2104 if (run_ext > 0) { 2105 if (run_len == 0) 2106 m_run = m; 2107 run_len += run_ext; 2108 } else { 2109 if (run_len > 0) { 2110 m_run = NULL; 2111 run_len = 0; 2112 } 2113 } 2114 } 2115 if (m_mtx != NULL) 2116 mtx_unlock(m_mtx); 2117 if (run_len >= npages) 2118 return (m_run); 2119 return (NULL); 2120 } 2121 2122 /* 2123 * vm_page_reclaim_run: 2124 * 2125 * Try to relocate each of the allocated virtual pages within the 2126 * specified run of physical pages to a new physical address. Free the 2127 * physical pages underlying the relocated virtual pages. A virtual page 2128 * is relocatable if and only if it could be laundered or reclaimed by 2129 * the page daemon. Whenever possible, a virtual page is relocated to a 2130 * physical address above "high". 2131 * 2132 * Returns 0 if every physical page within the run was already free or 2133 * just freed by a successful relocation. Otherwise, returns a non-zero 2134 * value indicating why the last attempt to relocate a virtual page was 2135 * unsuccessful. 2136 * 2137 * "req_class" must be an allocation class. 2138 */ 2139 static int 2140 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run, 2141 vm_paddr_t high) 2142 { 2143 struct mtx *m_mtx, *new_mtx; 2144 struct spglist free; 2145 vm_object_t object; 2146 vm_paddr_t pa; 2147 vm_page_t m, m_end, m_new; 2148 int error, order, req; 2149 2150 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2151 ("req_class is not an allocation class")); 2152 SLIST_INIT(&free); 2153 error = 0; 2154 m = m_run; 2155 m_end = m_run + npages; 2156 m_mtx = NULL; 2157 for (; error == 0 && m < m_end; m++) { 2158 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2159 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2160 2161 /* 2162 * Avoid releasing and reacquiring the same page lock. 2163 */ 2164 new_mtx = vm_page_lockptr(m); 2165 if (m_mtx != new_mtx) { 2166 if (m_mtx != NULL) 2167 mtx_unlock(m_mtx); 2168 m_mtx = new_mtx; 2169 mtx_lock(m_mtx); 2170 } 2171 retry: 2172 if (m->wire_count != 0 || m->hold_count != 0) 2173 error = EBUSY; 2174 else if ((object = m->object) != NULL) { 2175 /* 2176 * The page is relocated if and only if it could be 2177 * laundered or reclaimed by the page daemon. 2178 */ 2179 if (!VM_OBJECT_TRYWLOCK(object)) { 2180 mtx_unlock(m_mtx); 2181 VM_OBJECT_WLOCK(object); 2182 mtx_lock(m_mtx); 2183 if (m->object != object) { 2184 /* 2185 * The page may have been freed. 2186 */ 2187 VM_OBJECT_WUNLOCK(object); 2188 goto retry; 2189 } else if (m->wire_count != 0 || 2190 m->hold_count != 0) { 2191 error = EBUSY; 2192 goto unlock; 2193 } 2194 } 2195 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2196 ("page %p is PG_UNHOLDFREE", m)); 2197 /* Don't care: PG_NODUMP, PG_ZERO. */ 2198 if (object->type != OBJT_DEFAULT && 2199 object->type != OBJT_SWAP && 2200 object->type != OBJT_VNODE) 2201 error = EINVAL; 2202 else if (object->memattr != VM_MEMATTR_DEFAULT) 2203 error = EINVAL; 2204 else if (m->queue != PQ_NONE && !vm_page_busied(m)) { 2205 KASSERT(pmap_page_get_memattr(m) == 2206 VM_MEMATTR_DEFAULT, 2207 ("page %p has an unexpected memattr", m)); 2208 KASSERT((m->oflags & (VPO_SWAPINPROG | 2209 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2210 ("page %p has unexpected oflags", m)); 2211 /* Don't care: VPO_NOSYNC. */ 2212 if (m->valid != 0) { 2213 /* 2214 * First, try to allocate a new page 2215 * that is above "high". Failing 2216 * that, try to allocate a new page 2217 * that is below "m_run". Allocate 2218 * the new page between the end of 2219 * "m_run" and "high" only as a last 2220 * resort. 2221 */ 2222 req = req_class | VM_ALLOC_NOOBJ; 2223 if ((m->flags & PG_NODUMP) != 0) 2224 req |= VM_ALLOC_NODUMP; 2225 if (trunc_page(high) != 2226 ~(vm_paddr_t)PAGE_MASK) { 2227 m_new = vm_page_alloc_contig( 2228 NULL, 0, req, 1, 2229 round_page(high), 2230 ~(vm_paddr_t)0, 2231 PAGE_SIZE, 0, 2232 VM_MEMATTR_DEFAULT); 2233 } else 2234 m_new = NULL; 2235 if (m_new == NULL) { 2236 pa = VM_PAGE_TO_PHYS(m_run); 2237 m_new = vm_page_alloc_contig( 2238 NULL, 0, req, 1, 2239 0, pa - 1, PAGE_SIZE, 0, 2240 VM_MEMATTR_DEFAULT); 2241 } 2242 if (m_new == NULL) { 2243 pa += ptoa(npages); 2244 m_new = vm_page_alloc_contig( 2245 NULL, 0, req, 1, 2246 pa, high, PAGE_SIZE, 0, 2247 VM_MEMATTR_DEFAULT); 2248 } 2249 if (m_new == NULL) { 2250 error = ENOMEM; 2251 goto unlock; 2252 } 2253 KASSERT(m_new->wire_count == 0, 2254 ("page %p is wired", m)); 2255 2256 /* 2257 * Replace "m" with the new page. For 2258 * vm_page_replace(), "m" must be busy 2259 * and dequeued. Finally, change "m" 2260 * as if vm_page_free() was called. 2261 */ 2262 if (object->ref_count != 0) 2263 pmap_remove_all(m); 2264 m_new->aflags = m->aflags; 2265 KASSERT(m_new->oflags == VPO_UNMANAGED, 2266 ("page %p is managed", m)); 2267 m_new->oflags = m->oflags & VPO_NOSYNC; 2268 pmap_copy_page(m, m_new); 2269 m_new->valid = m->valid; 2270 m_new->dirty = m->dirty; 2271 m->flags &= ~PG_ZERO; 2272 vm_page_xbusy(m); 2273 vm_page_remque(m); 2274 vm_page_replace_checked(m_new, object, 2275 m->pindex, m); 2276 m->valid = 0; 2277 vm_page_undirty(m); 2278 2279 /* 2280 * The new page must be deactivated 2281 * before the object is unlocked. 2282 */ 2283 new_mtx = vm_page_lockptr(m_new); 2284 if (m_mtx != new_mtx) { 2285 mtx_unlock(m_mtx); 2286 m_mtx = new_mtx; 2287 mtx_lock(m_mtx); 2288 } 2289 vm_page_deactivate(m_new); 2290 } else { 2291 m->flags &= ~PG_ZERO; 2292 vm_page_remque(m); 2293 vm_page_remove(m); 2294 KASSERT(m->dirty == 0, 2295 ("page %p is dirty", m)); 2296 } 2297 SLIST_INSERT_HEAD(&free, m, plinks.s.ss); 2298 } else 2299 error = EBUSY; 2300 unlock: 2301 VM_OBJECT_WUNLOCK(object); 2302 } else { 2303 mtx_lock(&vm_page_queue_free_mtx); 2304 order = m->order; 2305 if (order < VM_NFREEORDER) { 2306 /* 2307 * The page is enqueued in the physical memory 2308 * allocator's free page queues. Moreover, it 2309 * is the first page in a power-of-two-sized 2310 * run of contiguous free pages. Jump ahead 2311 * to the last page within that run, and 2312 * continue from there. 2313 */ 2314 m += (1 << order) - 1; 2315 } 2316 #if VM_NRESERVLEVEL > 0 2317 else if (vm_reserv_is_page_free(m)) 2318 order = 0; 2319 #endif 2320 mtx_unlock(&vm_page_queue_free_mtx); 2321 if (order == VM_NFREEORDER) 2322 error = EINVAL; 2323 } 2324 } 2325 if (m_mtx != NULL) 2326 mtx_unlock(m_mtx); 2327 if ((m = SLIST_FIRST(&free)) != NULL) { 2328 mtx_lock(&vm_page_queue_free_mtx); 2329 do { 2330 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2331 vm_phys_freecnt_adj(m, 1); 2332 #if VM_NRESERVLEVEL > 0 2333 if (!vm_reserv_free_page(m)) 2334 #else 2335 if (true) 2336 #endif 2337 vm_phys_free_pages(m, 0); 2338 } while ((m = SLIST_FIRST(&free)) != NULL); 2339 vm_page_free_wakeup(); 2340 mtx_unlock(&vm_page_queue_free_mtx); 2341 } 2342 return (error); 2343 } 2344 2345 #define NRUNS 16 2346 2347 CTASSERT(powerof2(NRUNS)); 2348 2349 #define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2350 2351 #define MIN_RECLAIM 8 2352 2353 /* 2354 * vm_page_reclaim_contig: 2355 * 2356 * Reclaim allocated, contiguous physical memory satisfying the specified 2357 * conditions by relocating the virtual pages using that physical memory. 2358 * Returns true if reclamation is successful and false otherwise. Since 2359 * relocation requires the allocation of physical pages, reclamation may 2360 * fail due to a shortage of free pages. When reclamation fails, callers 2361 * are expected to perform VM_WAIT before retrying a failed allocation 2362 * operation, e.g., vm_page_alloc_contig(). 2363 * 2364 * The caller must always specify an allocation class through "req". 2365 * 2366 * allocation classes: 2367 * VM_ALLOC_NORMAL normal process request 2368 * VM_ALLOC_SYSTEM system *really* needs a page 2369 * VM_ALLOC_INTERRUPT interrupt time request 2370 * 2371 * The optional allocation flags are ignored. 2372 * 2373 * "npages" must be greater than zero. Both "alignment" and "boundary" 2374 * must be a power of two. 2375 */ 2376 bool 2377 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 2378 u_long alignment, vm_paddr_t boundary) 2379 { 2380 vm_paddr_t curr_low; 2381 vm_page_t m_run, m_runs[NRUNS]; 2382 u_long count, reclaimed; 2383 int error, i, options, req_class; 2384 2385 KASSERT(npages > 0, ("npages is 0")); 2386 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2387 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2388 req_class = req & VM_ALLOC_CLASS_MASK; 2389 2390 /* 2391 * The page daemon is allowed to dig deeper into the free page list. 2392 */ 2393 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2394 req_class = VM_ALLOC_SYSTEM; 2395 2396 /* 2397 * Return if the number of free pages cannot satisfy the requested 2398 * allocation. 2399 */ 2400 count = vm_cnt.v_free_count; 2401 if (count < npages + vm_cnt.v_free_reserved || (count < npages + 2402 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2403 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2404 return (false); 2405 2406 /* 2407 * Scan up to three times, relaxing the restrictions ("options") on 2408 * the reclamation of reservations and superpages each time. 2409 */ 2410 for (options = VPSC_NORESERV;;) { 2411 /* 2412 * Find the highest runs that satisfy the given constraints 2413 * and restrictions, and record them in "m_runs". 2414 */ 2415 curr_low = low; 2416 count = 0; 2417 for (;;) { 2418 m_run = vm_phys_scan_contig(npages, curr_low, high, 2419 alignment, boundary, options); 2420 if (m_run == NULL) 2421 break; 2422 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2423 m_runs[RUN_INDEX(count)] = m_run; 2424 count++; 2425 } 2426 2427 /* 2428 * Reclaim the highest runs in LIFO (descending) order until 2429 * the number of reclaimed pages, "reclaimed", is at least 2430 * MIN_RECLAIM. Reset "reclaimed" each time because each 2431 * reclamation is idempotent, and runs will (likely) recur 2432 * from one scan to the next as restrictions are relaxed. 2433 */ 2434 reclaimed = 0; 2435 for (i = 0; count > 0 && i < NRUNS; i++) { 2436 count--; 2437 m_run = m_runs[RUN_INDEX(count)]; 2438 error = vm_page_reclaim_run(req_class, npages, m_run, 2439 high); 2440 if (error == 0) { 2441 reclaimed += npages; 2442 if (reclaimed >= MIN_RECLAIM) 2443 return (true); 2444 } 2445 } 2446 2447 /* 2448 * Either relax the restrictions on the next scan or return if 2449 * the last scan had no restrictions. 2450 */ 2451 if (options == VPSC_NORESERV) 2452 options = VPSC_NOSUPER; 2453 else if (options == VPSC_NOSUPER) 2454 options = VPSC_ANY; 2455 else if (options == VPSC_ANY) 2456 return (reclaimed != 0); 2457 } 2458 } 2459 2460 /* 2461 * vm_wait: (also see VM_WAIT macro) 2462 * 2463 * Sleep until free pages are available for allocation. 2464 * - Called in various places before memory allocations. 2465 */ 2466 void 2467 vm_wait(void) 2468 { 2469 2470 mtx_lock(&vm_page_queue_free_mtx); 2471 if (curproc == pageproc) { 2472 vm_pageout_pages_needed = 1; 2473 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 2474 PDROP | PSWP, "VMWait", 0); 2475 } else { 2476 if (__predict_false(pageproc == NULL)) 2477 panic("vm_wait in early boot"); 2478 if (!vm_pageout_wanted) { 2479 vm_pageout_wanted = true; 2480 wakeup(&vm_pageout_wanted); 2481 } 2482 vm_pages_needed = true; 2483 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 2484 "vmwait", 0); 2485 } 2486 } 2487 2488 /* 2489 * vm_waitpfault: (also see VM_WAITPFAULT macro) 2490 * 2491 * Sleep until free pages are available for allocation. 2492 * - Called only in vm_fault so that processes page faulting 2493 * can be easily tracked. 2494 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 2495 * processes will be able to grab memory first. Do not change 2496 * this balance without careful testing first. 2497 */ 2498 void 2499 vm_waitpfault(void) 2500 { 2501 2502 mtx_lock(&vm_page_queue_free_mtx); 2503 if (!vm_pageout_wanted) { 2504 vm_pageout_wanted = true; 2505 wakeup(&vm_pageout_wanted); 2506 } 2507 vm_pages_needed = true; 2508 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 2509 "pfault", 0); 2510 } 2511 2512 struct vm_pagequeue * 2513 vm_page_pagequeue(vm_page_t m) 2514 { 2515 2516 if (vm_page_in_laundry(m)) 2517 return (&vm_dom[0].vmd_pagequeues[m->queue]); 2518 else 2519 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]); 2520 } 2521 2522 /* 2523 * vm_page_dequeue: 2524 * 2525 * Remove the given page from its current page queue. 2526 * 2527 * The page must be locked. 2528 */ 2529 void 2530 vm_page_dequeue(vm_page_t m) 2531 { 2532 struct vm_pagequeue *pq; 2533 2534 vm_page_assert_locked(m); 2535 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued", 2536 m)); 2537 pq = vm_page_pagequeue(m); 2538 vm_pagequeue_lock(pq); 2539 m->queue = PQ_NONE; 2540 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2541 vm_pagequeue_cnt_dec(pq); 2542 vm_pagequeue_unlock(pq); 2543 } 2544 2545 /* 2546 * vm_page_dequeue_locked: 2547 * 2548 * Remove the given page from its current page queue. 2549 * 2550 * The page and page queue must be locked. 2551 */ 2552 void 2553 vm_page_dequeue_locked(vm_page_t m) 2554 { 2555 struct vm_pagequeue *pq; 2556 2557 vm_page_lock_assert(m, MA_OWNED); 2558 pq = vm_page_pagequeue(m); 2559 vm_pagequeue_assert_locked(pq); 2560 m->queue = PQ_NONE; 2561 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2562 vm_pagequeue_cnt_dec(pq); 2563 } 2564 2565 /* 2566 * vm_page_enqueue: 2567 * 2568 * Add the given page to the specified page queue. 2569 * 2570 * The page must be locked. 2571 */ 2572 static void 2573 vm_page_enqueue(uint8_t queue, vm_page_t m) 2574 { 2575 struct vm_pagequeue *pq; 2576 2577 vm_page_lock_assert(m, MA_OWNED); 2578 KASSERT(queue < PQ_COUNT, 2579 ("vm_page_enqueue: invalid queue %u request for page %p", 2580 queue, m)); 2581 if (queue == PQ_LAUNDRY) 2582 pq = &vm_dom[0].vmd_pagequeues[queue]; 2583 else 2584 pq = &vm_phys_domain(m)->vmd_pagequeues[queue]; 2585 vm_pagequeue_lock(pq); 2586 m->queue = queue; 2587 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2588 vm_pagequeue_cnt_inc(pq); 2589 vm_pagequeue_unlock(pq); 2590 } 2591 2592 /* 2593 * vm_page_requeue: 2594 * 2595 * Move the given page to the tail of its current page queue. 2596 * 2597 * The page must be locked. 2598 */ 2599 void 2600 vm_page_requeue(vm_page_t m) 2601 { 2602 struct vm_pagequeue *pq; 2603 2604 vm_page_lock_assert(m, MA_OWNED); 2605 KASSERT(m->queue != PQ_NONE, 2606 ("vm_page_requeue: page %p is not queued", m)); 2607 pq = vm_page_pagequeue(m); 2608 vm_pagequeue_lock(pq); 2609 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2610 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2611 vm_pagequeue_unlock(pq); 2612 } 2613 2614 /* 2615 * vm_page_requeue_locked: 2616 * 2617 * Move the given page to the tail of its current page queue. 2618 * 2619 * The page queue must be locked. 2620 */ 2621 void 2622 vm_page_requeue_locked(vm_page_t m) 2623 { 2624 struct vm_pagequeue *pq; 2625 2626 KASSERT(m->queue != PQ_NONE, 2627 ("vm_page_requeue_locked: page %p is not queued", m)); 2628 pq = vm_page_pagequeue(m); 2629 vm_pagequeue_assert_locked(pq); 2630 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2631 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2632 } 2633 2634 /* 2635 * vm_page_activate: 2636 * 2637 * Put the specified page on the active list (if appropriate). 2638 * Ensure that act_count is at least ACT_INIT but do not otherwise 2639 * mess with it. 2640 * 2641 * The page must be locked. 2642 */ 2643 void 2644 vm_page_activate(vm_page_t m) 2645 { 2646 int queue; 2647 2648 vm_page_lock_assert(m, MA_OWNED); 2649 if ((queue = m->queue) != PQ_ACTIVE) { 2650 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2651 if (m->act_count < ACT_INIT) 2652 m->act_count = ACT_INIT; 2653 if (queue != PQ_NONE) 2654 vm_page_dequeue(m); 2655 vm_page_enqueue(PQ_ACTIVE, m); 2656 } else 2657 KASSERT(queue == PQ_NONE, 2658 ("vm_page_activate: wired page %p is queued", m)); 2659 } else { 2660 if (m->act_count < ACT_INIT) 2661 m->act_count = ACT_INIT; 2662 } 2663 } 2664 2665 /* 2666 * vm_page_free_wakeup: 2667 * 2668 * Helper routine for vm_page_free_toq(). This routine is called 2669 * when a page is added to the free queues. 2670 * 2671 * The page queues must be locked. 2672 */ 2673 static inline void 2674 vm_page_free_wakeup(void) 2675 { 2676 2677 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2678 /* 2679 * if pageout daemon needs pages, then tell it that there are 2680 * some free. 2681 */ 2682 if (vm_pageout_pages_needed && 2683 vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) { 2684 wakeup(&vm_pageout_pages_needed); 2685 vm_pageout_pages_needed = 0; 2686 } 2687 /* 2688 * wakeup processes that are waiting on memory if we hit a 2689 * high water mark. And wakeup scheduler process if we have 2690 * lots of memory. this process will swapin processes. 2691 */ 2692 if (vm_pages_needed && !vm_page_count_min()) { 2693 vm_pages_needed = false; 2694 wakeup(&vm_cnt.v_free_count); 2695 } 2696 } 2697 2698 /* 2699 * vm_page_free_toq: 2700 * 2701 * Returns the given page to the free list, 2702 * disassociating it with any VM object. 2703 * 2704 * The object must be locked. The page must be locked if it is managed. 2705 */ 2706 void 2707 vm_page_free_toq(vm_page_t m) 2708 { 2709 2710 if ((m->oflags & VPO_UNMANAGED) == 0) { 2711 vm_page_lock_assert(m, MA_OWNED); 2712 KASSERT(!pmap_page_is_mapped(m), 2713 ("vm_page_free_toq: freeing mapped page %p", m)); 2714 } else 2715 KASSERT(m->queue == PQ_NONE, 2716 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2717 PCPU_INC(cnt.v_tfree); 2718 2719 if (vm_page_sbusied(m)) 2720 panic("vm_page_free: freeing busy page %p", m); 2721 2722 /* 2723 * Unqueue, then remove page. Note that we cannot destroy 2724 * the page here because we do not want to call the pager's 2725 * callback routine until after we've put the page on the 2726 * appropriate free queue. 2727 */ 2728 vm_page_remque(m); 2729 vm_page_remove(m); 2730 2731 /* 2732 * If fictitious remove object association and 2733 * return, otherwise delay object association removal. 2734 */ 2735 if ((m->flags & PG_FICTITIOUS) != 0) { 2736 return; 2737 } 2738 2739 m->valid = 0; 2740 vm_page_undirty(m); 2741 2742 if (m->wire_count != 0) 2743 panic("vm_page_free: freeing wired page %p", m); 2744 if (m->hold_count != 0) { 2745 m->flags &= ~PG_ZERO; 2746 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2747 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2748 m->flags |= PG_UNHOLDFREE; 2749 } else { 2750 /* 2751 * Restore the default memory attribute to the page. 2752 */ 2753 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2754 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2755 2756 /* 2757 * Insert the page into the physical memory allocator's free 2758 * page queues. 2759 */ 2760 mtx_lock(&vm_page_queue_free_mtx); 2761 vm_phys_freecnt_adj(m, 1); 2762 #if VM_NRESERVLEVEL > 0 2763 if (!vm_reserv_free_page(m)) 2764 #else 2765 if (TRUE) 2766 #endif 2767 vm_phys_free_pages(m, 0); 2768 vm_page_free_wakeup(); 2769 mtx_unlock(&vm_page_queue_free_mtx); 2770 } 2771 } 2772 2773 /* 2774 * vm_page_wire: 2775 * 2776 * Mark this page as wired down by yet 2777 * another map, removing it from paging queues 2778 * as necessary. 2779 * 2780 * If the page is fictitious, then its wire count must remain one. 2781 * 2782 * The page must be locked. 2783 */ 2784 void 2785 vm_page_wire(vm_page_t m) 2786 { 2787 2788 /* 2789 * Only bump the wire statistics if the page is not already wired, 2790 * and only unqueue the page if it is on some queue (if it is unmanaged 2791 * it is already off the queues). 2792 */ 2793 vm_page_lock_assert(m, MA_OWNED); 2794 if ((m->flags & PG_FICTITIOUS) != 0) { 2795 KASSERT(m->wire_count == 1, 2796 ("vm_page_wire: fictitious page %p's wire count isn't one", 2797 m)); 2798 return; 2799 } 2800 if (m->wire_count == 0) { 2801 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 2802 m->queue == PQ_NONE, 2803 ("vm_page_wire: unmanaged page %p is queued", m)); 2804 vm_page_remque(m); 2805 atomic_add_int(&vm_cnt.v_wire_count, 1); 2806 } 2807 m->wire_count++; 2808 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2809 } 2810 2811 /* 2812 * vm_page_unwire: 2813 * 2814 * Release one wiring of the specified page, potentially allowing it to be 2815 * paged out. Returns TRUE if the number of wirings transitions to zero and 2816 * FALSE otherwise. 2817 * 2818 * Only managed pages belonging to an object can be paged out. If the number 2819 * of wirings transitions to zero and the page is eligible for page out, then 2820 * the page is added to the specified paging queue (unless PQ_NONE is 2821 * specified). 2822 * 2823 * If a page is fictitious, then its wire count must always be one. 2824 * 2825 * A managed page must be locked. 2826 */ 2827 boolean_t 2828 vm_page_unwire(vm_page_t m, uint8_t queue) 2829 { 2830 2831 KASSERT(queue < PQ_COUNT || queue == PQ_NONE, 2832 ("vm_page_unwire: invalid queue %u request for page %p", 2833 queue, m)); 2834 if ((m->oflags & VPO_UNMANAGED) == 0) 2835 vm_page_assert_locked(m); 2836 if ((m->flags & PG_FICTITIOUS) != 0) { 2837 KASSERT(m->wire_count == 1, 2838 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2839 return (FALSE); 2840 } 2841 if (m->wire_count > 0) { 2842 m->wire_count--; 2843 if (m->wire_count == 0) { 2844 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 2845 if ((m->oflags & VPO_UNMANAGED) == 0 && 2846 m->object != NULL && queue != PQ_NONE) 2847 vm_page_enqueue(queue, m); 2848 return (TRUE); 2849 } else 2850 return (FALSE); 2851 } else 2852 panic("vm_page_unwire: page %p's wire count is zero", m); 2853 } 2854 2855 /* 2856 * Move the specified page to the inactive queue. 2857 * 2858 * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive 2859 * queue. However, setting "noreuse" to TRUE will accelerate the specified 2860 * page's reclamation, but it will not unmap the page from any address space. 2861 * This is implemented by inserting the page near the head of the inactive 2862 * queue, using a marker page to guide FIFO insertion ordering. 2863 * 2864 * The page must be locked. 2865 */ 2866 static inline void 2867 _vm_page_deactivate(vm_page_t m, boolean_t noreuse) 2868 { 2869 struct vm_pagequeue *pq; 2870 int queue; 2871 2872 vm_page_assert_locked(m); 2873 2874 /* 2875 * Ignore if the page is already inactive, unless it is unlikely to be 2876 * reactivated. 2877 */ 2878 if ((queue = m->queue) == PQ_INACTIVE && !noreuse) 2879 return; 2880 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2881 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 2882 /* Avoid multiple acquisitions of the inactive queue lock. */ 2883 if (queue == PQ_INACTIVE) { 2884 vm_pagequeue_lock(pq); 2885 vm_page_dequeue_locked(m); 2886 } else { 2887 if (queue != PQ_NONE) 2888 vm_page_dequeue(m); 2889 vm_pagequeue_lock(pq); 2890 } 2891 m->queue = PQ_INACTIVE; 2892 if (noreuse) 2893 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead, 2894 m, plinks.q); 2895 else 2896 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2897 vm_pagequeue_cnt_inc(pq); 2898 vm_pagequeue_unlock(pq); 2899 } 2900 } 2901 2902 /* 2903 * Move the specified page to the inactive queue. 2904 * 2905 * The page must be locked. 2906 */ 2907 void 2908 vm_page_deactivate(vm_page_t m) 2909 { 2910 2911 _vm_page_deactivate(m, FALSE); 2912 } 2913 2914 /* 2915 * Move the specified page to the inactive queue with the expectation 2916 * that it is unlikely to be reused. 2917 * 2918 * The page must be locked. 2919 */ 2920 void 2921 vm_page_deactivate_noreuse(vm_page_t m) 2922 { 2923 2924 _vm_page_deactivate(m, TRUE); 2925 } 2926 2927 /* 2928 * vm_page_launder 2929 * 2930 * Put a page in the laundry. 2931 */ 2932 void 2933 vm_page_launder(vm_page_t m) 2934 { 2935 int queue; 2936 2937 vm_page_assert_locked(m); 2938 if ((queue = m->queue) != PQ_LAUNDRY) { 2939 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2940 if (queue != PQ_NONE) 2941 vm_page_dequeue(m); 2942 vm_page_enqueue(PQ_LAUNDRY, m); 2943 } else 2944 KASSERT(queue == PQ_NONE, 2945 ("wired page %p is queued", m)); 2946 } 2947 } 2948 2949 /* 2950 * vm_page_try_to_free() 2951 * 2952 * Attempt to free the page. If we cannot free it, we do nothing. 2953 * 1 is returned on success, 0 on failure. 2954 */ 2955 int 2956 vm_page_try_to_free(vm_page_t m) 2957 { 2958 2959 vm_page_lock_assert(m, MA_OWNED); 2960 if (m->object != NULL) 2961 VM_OBJECT_ASSERT_WLOCKED(m->object); 2962 if (m->dirty || m->hold_count || m->wire_count || 2963 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2964 return (0); 2965 pmap_remove_all(m); 2966 if (m->dirty) 2967 return (0); 2968 vm_page_free(m); 2969 return (1); 2970 } 2971 2972 /* 2973 * vm_page_advise 2974 * 2975 * Deactivate or do nothing, as appropriate. 2976 * 2977 * The object and page must be locked. 2978 */ 2979 void 2980 vm_page_advise(vm_page_t m, int advice) 2981 { 2982 2983 vm_page_assert_locked(m); 2984 VM_OBJECT_ASSERT_WLOCKED(m->object); 2985 if (advice == MADV_FREE) 2986 /* 2987 * Mark the page clean. This will allow the page to be freed 2988 * without first paging it out. MADV_FREE pages are often 2989 * quickly reused by malloc(3), so we do not do anything that 2990 * would result in a page fault on a later access. 2991 */ 2992 vm_page_undirty(m); 2993 else if (advice != MADV_DONTNEED) 2994 return; 2995 2996 /* 2997 * Clear any references to the page. Otherwise, the page daemon will 2998 * immediately reactivate the page. 2999 */ 3000 vm_page_aflag_clear(m, PGA_REFERENCED); 3001 3002 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 3003 vm_page_dirty(m); 3004 3005 /* 3006 * Place clean pages near the head of the inactive queue rather than 3007 * the tail, thus defeating the queue's LRU operation and ensuring that 3008 * the page will be reused quickly. Dirty pages not already in the 3009 * laundry are moved there. 3010 */ 3011 if (m->dirty == 0) 3012 vm_page_deactivate_noreuse(m); 3013 else 3014 vm_page_launder(m); 3015 } 3016 3017 /* 3018 * Grab a page, waiting until we are waken up due to the page 3019 * changing state. We keep on waiting, if the page continues 3020 * to be in the object. If the page doesn't exist, first allocate it 3021 * and then conditionally zero it. 3022 * 3023 * This routine may sleep. 3024 * 3025 * The object must be locked on entry. The lock will, however, be released 3026 * and reacquired if the routine sleeps. 3027 */ 3028 vm_page_t 3029 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 3030 { 3031 vm_page_t m; 3032 int sleep; 3033 3034 VM_OBJECT_ASSERT_WLOCKED(object); 3035 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 3036 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 3037 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 3038 retrylookup: 3039 if ((m = vm_page_lookup(object, pindex)) != NULL) { 3040 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 3041 vm_page_xbusied(m) : vm_page_busied(m); 3042 if (sleep) { 3043 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3044 return (NULL); 3045 /* 3046 * Reference the page before unlocking and 3047 * sleeping so that the page daemon is less 3048 * likely to reclaim it. 3049 */ 3050 vm_page_aflag_set(m, PGA_REFERENCED); 3051 vm_page_lock(m); 3052 VM_OBJECT_WUNLOCK(object); 3053 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 3054 VM_ALLOC_IGN_SBUSY) != 0); 3055 VM_OBJECT_WLOCK(object); 3056 goto retrylookup; 3057 } else { 3058 if ((allocflags & VM_ALLOC_WIRED) != 0) { 3059 vm_page_lock(m); 3060 vm_page_wire(m); 3061 vm_page_unlock(m); 3062 } 3063 if ((allocflags & 3064 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 3065 vm_page_xbusy(m); 3066 if ((allocflags & VM_ALLOC_SBUSY) != 0) 3067 vm_page_sbusy(m); 3068 return (m); 3069 } 3070 } 3071 m = vm_page_alloc(object, pindex, allocflags); 3072 if (m == NULL) { 3073 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3074 return (NULL); 3075 VM_OBJECT_WUNLOCK(object); 3076 VM_WAIT; 3077 VM_OBJECT_WLOCK(object); 3078 goto retrylookup; 3079 } 3080 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 3081 pmap_zero_page(m); 3082 return (m); 3083 } 3084 3085 /* 3086 * Mapping function for valid or dirty bits in a page. 3087 * 3088 * Inputs are required to range within a page. 3089 */ 3090 vm_page_bits_t 3091 vm_page_bits(int base, int size) 3092 { 3093 int first_bit; 3094 int last_bit; 3095 3096 KASSERT( 3097 base + size <= PAGE_SIZE, 3098 ("vm_page_bits: illegal base/size %d/%d", base, size) 3099 ); 3100 3101 if (size == 0) /* handle degenerate case */ 3102 return (0); 3103 3104 first_bit = base >> DEV_BSHIFT; 3105 last_bit = (base + size - 1) >> DEV_BSHIFT; 3106 3107 return (((vm_page_bits_t)2 << last_bit) - 3108 ((vm_page_bits_t)1 << first_bit)); 3109 } 3110 3111 /* 3112 * vm_page_set_valid_range: 3113 * 3114 * Sets portions of a page valid. The arguments are expected 3115 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3116 * of any partial chunks touched by the range. The invalid portion of 3117 * such chunks will be zeroed. 3118 * 3119 * (base + size) must be less then or equal to PAGE_SIZE. 3120 */ 3121 void 3122 vm_page_set_valid_range(vm_page_t m, int base, int size) 3123 { 3124 int endoff, frag; 3125 3126 VM_OBJECT_ASSERT_WLOCKED(m->object); 3127 if (size == 0) /* handle degenerate case */ 3128 return; 3129 3130 /* 3131 * If the base is not DEV_BSIZE aligned and the valid 3132 * bit is clear, we have to zero out a portion of the 3133 * first block. 3134 */ 3135 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3136 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 3137 pmap_zero_page_area(m, frag, base - frag); 3138 3139 /* 3140 * If the ending offset is not DEV_BSIZE aligned and the 3141 * valid bit is clear, we have to zero out a portion of 3142 * the last block. 3143 */ 3144 endoff = base + size; 3145 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3146 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 3147 pmap_zero_page_area(m, endoff, 3148 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3149 3150 /* 3151 * Assert that no previously invalid block that is now being validated 3152 * is already dirty. 3153 */ 3154 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 3155 ("vm_page_set_valid_range: page %p is dirty", m)); 3156 3157 /* 3158 * Set valid bits inclusive of any overlap. 3159 */ 3160 m->valid |= vm_page_bits(base, size); 3161 } 3162 3163 /* 3164 * Clear the given bits from the specified page's dirty field. 3165 */ 3166 static __inline void 3167 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 3168 { 3169 uintptr_t addr; 3170 #if PAGE_SIZE < 16384 3171 int shift; 3172 #endif 3173 3174 /* 3175 * If the object is locked and the page is neither exclusive busy nor 3176 * write mapped, then the page's dirty field cannot possibly be 3177 * set by a concurrent pmap operation. 3178 */ 3179 VM_OBJECT_ASSERT_WLOCKED(m->object); 3180 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 3181 m->dirty &= ~pagebits; 3182 else { 3183 /* 3184 * The pmap layer can call vm_page_dirty() without 3185 * holding a distinguished lock. The combination of 3186 * the object's lock and an atomic operation suffice 3187 * to guarantee consistency of the page dirty field. 3188 * 3189 * For PAGE_SIZE == 32768 case, compiler already 3190 * properly aligns the dirty field, so no forcible 3191 * alignment is needed. Only require existence of 3192 * atomic_clear_64 when page size is 32768. 3193 */ 3194 addr = (uintptr_t)&m->dirty; 3195 #if PAGE_SIZE == 32768 3196 atomic_clear_64((uint64_t *)addr, pagebits); 3197 #elif PAGE_SIZE == 16384 3198 atomic_clear_32((uint32_t *)addr, pagebits); 3199 #else /* PAGE_SIZE <= 8192 */ 3200 /* 3201 * Use a trick to perform a 32-bit atomic on the 3202 * containing aligned word, to not depend on the existence 3203 * of atomic_clear_{8, 16}. 3204 */ 3205 shift = addr & (sizeof(uint32_t) - 1); 3206 #if BYTE_ORDER == BIG_ENDIAN 3207 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 3208 #else 3209 shift *= NBBY; 3210 #endif 3211 addr &= ~(sizeof(uint32_t) - 1); 3212 atomic_clear_32((uint32_t *)addr, pagebits << shift); 3213 #endif /* PAGE_SIZE */ 3214 } 3215 } 3216 3217 /* 3218 * vm_page_set_validclean: 3219 * 3220 * Sets portions of a page valid and clean. The arguments are expected 3221 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3222 * of any partial chunks touched by the range. The invalid portion of 3223 * such chunks will be zero'd. 3224 * 3225 * (base + size) must be less then or equal to PAGE_SIZE. 3226 */ 3227 void 3228 vm_page_set_validclean(vm_page_t m, int base, int size) 3229 { 3230 vm_page_bits_t oldvalid, pagebits; 3231 int endoff, frag; 3232 3233 VM_OBJECT_ASSERT_WLOCKED(m->object); 3234 if (size == 0) /* handle degenerate case */ 3235 return; 3236 3237 /* 3238 * If the base is not DEV_BSIZE aligned and the valid 3239 * bit is clear, we have to zero out a portion of the 3240 * first block. 3241 */ 3242 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3243 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 3244 pmap_zero_page_area(m, frag, base - frag); 3245 3246 /* 3247 * If the ending offset is not DEV_BSIZE aligned and the 3248 * valid bit is clear, we have to zero out a portion of 3249 * the last block. 3250 */ 3251 endoff = base + size; 3252 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3253 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 3254 pmap_zero_page_area(m, endoff, 3255 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3256 3257 /* 3258 * Set valid, clear dirty bits. If validating the entire 3259 * page we can safely clear the pmap modify bit. We also 3260 * use this opportunity to clear the VPO_NOSYNC flag. If a process 3261 * takes a write fault on a MAP_NOSYNC memory area the flag will 3262 * be set again. 3263 * 3264 * We set valid bits inclusive of any overlap, but we can only 3265 * clear dirty bits for DEV_BSIZE chunks that are fully within 3266 * the range. 3267 */ 3268 oldvalid = m->valid; 3269 pagebits = vm_page_bits(base, size); 3270 m->valid |= pagebits; 3271 #if 0 /* NOT YET */ 3272 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 3273 frag = DEV_BSIZE - frag; 3274 base += frag; 3275 size -= frag; 3276 if (size < 0) 3277 size = 0; 3278 } 3279 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 3280 #endif 3281 if (base == 0 && size == PAGE_SIZE) { 3282 /* 3283 * The page can only be modified within the pmap if it is 3284 * mapped, and it can only be mapped if it was previously 3285 * fully valid. 3286 */ 3287 if (oldvalid == VM_PAGE_BITS_ALL) 3288 /* 3289 * Perform the pmap_clear_modify() first. Otherwise, 3290 * a concurrent pmap operation, such as 3291 * pmap_protect(), could clear a modification in the 3292 * pmap and set the dirty field on the page before 3293 * pmap_clear_modify() had begun and after the dirty 3294 * field was cleared here. 3295 */ 3296 pmap_clear_modify(m); 3297 m->dirty = 0; 3298 m->oflags &= ~VPO_NOSYNC; 3299 } else if (oldvalid != VM_PAGE_BITS_ALL) 3300 m->dirty &= ~pagebits; 3301 else 3302 vm_page_clear_dirty_mask(m, pagebits); 3303 } 3304 3305 void 3306 vm_page_clear_dirty(vm_page_t m, int base, int size) 3307 { 3308 3309 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 3310 } 3311 3312 /* 3313 * vm_page_set_invalid: 3314 * 3315 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3316 * valid and dirty bits for the effected areas are cleared. 3317 */ 3318 void 3319 vm_page_set_invalid(vm_page_t m, int base, int size) 3320 { 3321 vm_page_bits_t bits; 3322 vm_object_t object; 3323 3324 object = m->object; 3325 VM_OBJECT_ASSERT_WLOCKED(object); 3326 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 3327 size >= object->un_pager.vnp.vnp_size) 3328 bits = VM_PAGE_BITS_ALL; 3329 else 3330 bits = vm_page_bits(base, size); 3331 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && 3332 bits != 0) 3333 pmap_remove_all(m); 3334 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 3335 !pmap_page_is_mapped(m), 3336 ("vm_page_set_invalid: page %p is mapped", m)); 3337 m->valid &= ~bits; 3338 m->dirty &= ~bits; 3339 } 3340 3341 /* 3342 * vm_page_zero_invalid() 3343 * 3344 * The kernel assumes that the invalid portions of a page contain 3345 * garbage, but such pages can be mapped into memory by user code. 3346 * When this occurs, we must zero out the non-valid portions of the 3347 * page so user code sees what it expects. 3348 * 3349 * Pages are most often semi-valid when the end of a file is mapped 3350 * into memory and the file's size is not page aligned. 3351 */ 3352 void 3353 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3354 { 3355 int b; 3356 int i; 3357 3358 VM_OBJECT_ASSERT_WLOCKED(m->object); 3359 /* 3360 * Scan the valid bits looking for invalid sections that 3361 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 3362 * valid bit may be set ) have already been zeroed by 3363 * vm_page_set_validclean(). 3364 */ 3365 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3366 if (i == (PAGE_SIZE / DEV_BSIZE) || 3367 (m->valid & ((vm_page_bits_t)1 << i))) { 3368 if (i > b) { 3369 pmap_zero_page_area(m, 3370 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3371 } 3372 b = i + 1; 3373 } 3374 } 3375 3376 /* 3377 * setvalid is TRUE when we can safely set the zero'd areas 3378 * as being valid. We can do this if there are no cache consistancy 3379 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3380 */ 3381 if (setvalid) 3382 m->valid = VM_PAGE_BITS_ALL; 3383 } 3384 3385 /* 3386 * vm_page_is_valid: 3387 * 3388 * Is (partial) page valid? Note that the case where size == 0 3389 * will return FALSE in the degenerate case where the page is 3390 * entirely invalid, and TRUE otherwise. 3391 */ 3392 int 3393 vm_page_is_valid(vm_page_t m, int base, int size) 3394 { 3395 vm_page_bits_t bits; 3396 3397 VM_OBJECT_ASSERT_LOCKED(m->object); 3398 bits = vm_page_bits(base, size); 3399 return (m->valid != 0 && (m->valid & bits) == bits); 3400 } 3401 3402 /* 3403 * vm_page_ps_is_valid: 3404 * 3405 * Returns TRUE if the entire (super)page is valid and FALSE otherwise. 3406 */ 3407 boolean_t 3408 vm_page_ps_is_valid(vm_page_t m) 3409 { 3410 int i, npages; 3411 3412 VM_OBJECT_ASSERT_LOCKED(m->object); 3413 npages = atop(pagesizes[m->psind]); 3414 3415 /* 3416 * The physically contiguous pages that make up a superpage, i.e., a 3417 * page with a page size index ("psind") greater than zero, will 3418 * occupy adjacent entries in vm_page_array[]. 3419 */ 3420 for (i = 0; i < npages; i++) { 3421 if (m[i].valid != VM_PAGE_BITS_ALL) 3422 return (FALSE); 3423 } 3424 return (TRUE); 3425 } 3426 3427 /* 3428 * Set the page's dirty bits if the page is modified. 3429 */ 3430 void 3431 vm_page_test_dirty(vm_page_t m) 3432 { 3433 3434 VM_OBJECT_ASSERT_WLOCKED(m->object); 3435 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 3436 vm_page_dirty(m); 3437 } 3438 3439 void 3440 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 3441 { 3442 3443 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 3444 } 3445 3446 void 3447 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 3448 { 3449 3450 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 3451 } 3452 3453 int 3454 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 3455 { 3456 3457 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 3458 } 3459 3460 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 3461 void 3462 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 3463 { 3464 3465 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 3466 } 3467 3468 void 3469 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 3470 { 3471 3472 mtx_assert_(vm_page_lockptr(m), a, file, line); 3473 } 3474 #endif 3475 3476 #ifdef INVARIANTS 3477 void 3478 vm_page_object_lock_assert(vm_page_t m) 3479 { 3480 3481 /* 3482 * Certain of the page's fields may only be modified by the 3483 * holder of the containing object's lock or the exclusive busy. 3484 * holder. Unfortunately, the holder of the write busy is 3485 * not recorded, and thus cannot be checked here. 3486 */ 3487 if (m->object != NULL && !vm_page_xbusied(m)) 3488 VM_OBJECT_ASSERT_WLOCKED(m->object); 3489 } 3490 3491 void 3492 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 3493 { 3494 3495 if ((bits & PGA_WRITEABLE) == 0) 3496 return; 3497 3498 /* 3499 * The PGA_WRITEABLE flag can only be set if the page is 3500 * managed, is exclusively busied or the object is locked. 3501 * Currently, this flag is only set by pmap_enter(). 3502 */ 3503 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3504 ("PGA_WRITEABLE on unmanaged page")); 3505 if (!vm_page_xbusied(m)) 3506 VM_OBJECT_ASSERT_LOCKED(m->object); 3507 } 3508 #endif 3509 3510 #include "opt_ddb.h" 3511 #ifdef DDB 3512 #include <sys/kernel.h> 3513 3514 #include <ddb/ddb.h> 3515 3516 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3517 { 3518 3519 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count); 3520 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count); 3521 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count); 3522 db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count); 3523 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count); 3524 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 3525 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 3526 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 3527 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 3528 } 3529 3530 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3531 { 3532 int dom; 3533 3534 db_printf("pq_free %d\n", vm_cnt.v_free_count); 3535 for (dom = 0; dom < vm_ndomains; dom++) { 3536 db_printf( 3537 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d\n", 3538 dom, 3539 vm_dom[dom].vmd_page_count, 3540 vm_dom[dom].vmd_free_count, 3541 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 3542 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 3543 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt); 3544 } 3545 } 3546 3547 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 3548 { 3549 vm_page_t m; 3550 boolean_t phys; 3551 3552 if (!have_addr) { 3553 db_printf("show pginfo addr\n"); 3554 return; 3555 } 3556 3557 phys = strchr(modif, 'p') != NULL; 3558 if (phys) 3559 m = PHYS_TO_VM_PAGE(addr); 3560 else 3561 m = (vm_page_t)addr; 3562 db_printf( 3563 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 3564 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 3565 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 3566 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 3567 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 3568 } 3569 #endif /* DDB */ 3570