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 next->pindex != m->pindex + 1) 1333 next = NULL; 1334 return (next); 1335 } 1336 1337 /* 1338 * Returns the given page's predecessor (by pindex) within the object if it is 1339 * resident; if none is found, NULL is returned. 1340 * 1341 * The object must be locked. 1342 */ 1343 vm_page_t 1344 vm_page_prev(vm_page_t m) 1345 { 1346 vm_page_t prev; 1347 1348 VM_OBJECT_ASSERT_LOCKED(m->object); 1349 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL && 1350 prev->pindex != m->pindex - 1) 1351 prev = NULL; 1352 return (prev); 1353 } 1354 1355 /* 1356 * Uses the page mnew as a replacement for an existing page at index 1357 * pindex which must be already present in the object. 1358 * 1359 * The existing page must not be on a paging queue. 1360 */ 1361 vm_page_t 1362 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) 1363 { 1364 vm_page_t mold; 1365 1366 VM_OBJECT_ASSERT_WLOCKED(object); 1367 KASSERT(mnew->object == NULL, 1368 ("vm_page_replace: page already in object")); 1369 1370 /* 1371 * This function mostly follows vm_page_insert() and 1372 * vm_page_remove() without the radix, object count and vnode 1373 * dance. Double check such functions for more comments. 1374 */ 1375 1376 mnew->object = object; 1377 mnew->pindex = pindex; 1378 mold = vm_radix_replace(&object->rtree, mnew); 1379 KASSERT(mold->queue == PQ_NONE, 1380 ("vm_page_replace: mold is on a paging queue")); 1381 1382 /* Keep the resident page list in sorted order. */ 1383 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); 1384 TAILQ_REMOVE(&object->memq, mold, listq); 1385 1386 mold->object = NULL; 1387 vm_page_xunbusy_maybelocked(mold); 1388 1389 /* 1390 * The object's resident_page_count does not change because we have 1391 * swapped one page for another, but OBJ_MIGHTBEDIRTY. 1392 */ 1393 if (pmap_page_is_write_mapped(mnew)) 1394 vm_object_set_writeable_dirty(object); 1395 return (mold); 1396 } 1397 1398 /* 1399 * vm_page_rename: 1400 * 1401 * Move the given memory entry from its 1402 * current object to the specified target object/offset. 1403 * 1404 * Note: swap associated with the page must be invalidated by the move. We 1405 * have to do this for several reasons: (1) we aren't freeing the 1406 * page, (2) we are dirtying the page, (3) the VM system is probably 1407 * moving the page from object A to B, and will then later move 1408 * the backing store from A to B and we can't have a conflict. 1409 * 1410 * Note: we *always* dirty the page. It is necessary both for the 1411 * fact that we moved it, and because we may be invalidating 1412 * swap. If the page is on the cache, we have to deactivate it 1413 * or vm_page_dirty() will panic. Dirty pages are not allowed 1414 * on the cache. 1415 * 1416 * The objects must be locked. 1417 */ 1418 int 1419 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 1420 { 1421 vm_page_t mpred; 1422 vm_pindex_t opidx; 1423 1424 VM_OBJECT_ASSERT_WLOCKED(new_object); 1425 1426 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); 1427 KASSERT(mpred == NULL || mpred->pindex != new_pindex, 1428 ("vm_page_rename: pindex already renamed")); 1429 1430 /* 1431 * Create a custom version of vm_page_insert() which does not depend 1432 * by m_prev and can cheat on the implementation aspects of the 1433 * function. 1434 */ 1435 opidx = m->pindex; 1436 m->pindex = new_pindex; 1437 if (vm_radix_insert(&new_object->rtree, m)) { 1438 m->pindex = opidx; 1439 return (1); 1440 } 1441 1442 /* 1443 * The operation cannot fail anymore. The removal must happen before 1444 * the listq iterator is tainted. 1445 */ 1446 m->pindex = opidx; 1447 vm_page_lock(m); 1448 vm_page_remove(m); 1449 1450 /* Return back to the new pindex to complete vm_page_insert(). */ 1451 m->pindex = new_pindex; 1452 m->object = new_object; 1453 vm_page_unlock(m); 1454 vm_page_insert_radixdone(m, new_object, mpred); 1455 vm_page_dirty(m); 1456 return (0); 1457 } 1458 1459 /* 1460 * vm_page_alloc: 1461 * 1462 * Allocate and return a page that is associated with the specified 1463 * object and offset pair. By default, this page is exclusive busied. 1464 * 1465 * The caller must always specify an allocation class. 1466 * 1467 * allocation classes: 1468 * VM_ALLOC_NORMAL normal process request 1469 * VM_ALLOC_SYSTEM system *really* needs a page 1470 * VM_ALLOC_INTERRUPT interrupt time request 1471 * 1472 * optional allocation flags: 1473 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1474 * intends to allocate 1475 * VM_ALLOC_NOBUSY do not exclusive busy the page 1476 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1477 * VM_ALLOC_NOOBJ page is not associated with an object and 1478 * should not be exclusive busy 1479 * VM_ALLOC_SBUSY shared busy the allocated page 1480 * VM_ALLOC_WIRED wire the allocated page 1481 * VM_ALLOC_ZERO prefer a zeroed page 1482 * 1483 * This routine may not sleep. 1484 */ 1485 vm_page_t 1486 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1487 { 1488 vm_page_t m, mpred; 1489 int flags, req_class; 1490 1491 mpred = 0; /* XXX: pacify gcc */ 1492 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1493 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1494 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1495 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1496 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1497 req)); 1498 if (object != NULL) 1499 VM_OBJECT_ASSERT_WLOCKED(object); 1500 1501 req_class = req & VM_ALLOC_CLASS_MASK; 1502 1503 /* 1504 * The page daemon is allowed to dig deeper into the free page list. 1505 */ 1506 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1507 req_class = VM_ALLOC_SYSTEM; 1508 1509 if (object != NULL) { 1510 mpred = vm_radix_lookup_le(&object->rtree, pindex); 1511 KASSERT(mpred == NULL || mpred->pindex != pindex, 1512 ("vm_page_alloc: pindex already allocated")); 1513 } 1514 1515 /* 1516 * Allocate a page if the number of free pages exceeds the minimum 1517 * for the request class. 1518 */ 1519 mtx_lock(&vm_page_queue_free_mtx); 1520 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved || 1521 (req_class == VM_ALLOC_SYSTEM && 1522 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) || 1523 (req_class == VM_ALLOC_INTERRUPT && 1524 vm_cnt.v_free_count > 0)) { 1525 /* 1526 * Can we allocate the page from a reservation? 1527 */ 1528 #if VM_NRESERVLEVEL > 0 1529 if (object == NULL || (object->flags & (OBJ_COLORED | 1530 OBJ_FICTITIOUS)) != OBJ_COLORED || (m = 1531 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) 1532 #endif 1533 { 1534 /* 1535 * If not, allocate it from the free page queues. 1536 */ 1537 m = vm_phys_alloc_pages(object != NULL ? 1538 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1539 #if VM_NRESERVLEVEL > 0 1540 if (m == NULL && vm_reserv_reclaim_inactive()) { 1541 m = vm_phys_alloc_pages(object != NULL ? 1542 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1543 0); 1544 } 1545 #endif 1546 } 1547 } else { 1548 /* 1549 * Not allocatable, give up. 1550 */ 1551 mtx_unlock(&vm_page_queue_free_mtx); 1552 atomic_add_int(&vm_pageout_deficit, 1553 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1554 pagedaemon_wakeup(); 1555 return (NULL); 1556 } 1557 1558 /* 1559 * At this point we had better have found a good page. 1560 */ 1561 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1562 vm_phys_freecnt_adj(m, -1); 1563 mtx_unlock(&vm_page_queue_free_mtx); 1564 vm_page_alloc_check(m); 1565 1566 /* 1567 * Initialize the page. Only the PG_ZERO flag is inherited. 1568 */ 1569 flags = 0; 1570 if ((req & VM_ALLOC_ZERO) != 0) 1571 flags = PG_ZERO; 1572 flags &= m->flags; 1573 if ((req & VM_ALLOC_NODUMP) != 0) 1574 flags |= PG_NODUMP; 1575 m->flags = flags; 1576 m->aflags = 0; 1577 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? 1578 VPO_UNMANAGED : 0; 1579 m->busy_lock = VPB_UNBUSIED; 1580 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) 1581 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1582 if ((req & VM_ALLOC_SBUSY) != 0) 1583 m->busy_lock = VPB_SHARERS_WORD(1); 1584 if (req & VM_ALLOC_WIRED) { 1585 /* 1586 * The page lock is not required for wiring a page until that 1587 * page is inserted into the object. 1588 */ 1589 atomic_add_int(&vm_cnt.v_wire_count, 1); 1590 m->wire_count = 1; 1591 } 1592 m->act_count = 0; 1593 1594 if (object != NULL) { 1595 if (vm_page_insert_after(m, object, pindex, mpred)) { 1596 pagedaemon_wakeup(); 1597 if (req & VM_ALLOC_WIRED) { 1598 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 1599 m->wire_count = 0; 1600 } 1601 m->object = NULL; 1602 m->oflags = VPO_UNMANAGED; 1603 m->busy_lock = VPB_UNBUSIED; 1604 vm_page_free(m); 1605 return (NULL); 1606 } 1607 1608 /* Ignore device objects; the pager sets "memattr" for them. */ 1609 if (object->memattr != VM_MEMATTR_DEFAULT && 1610 (object->flags & OBJ_FICTITIOUS) == 0) 1611 pmap_page_set_memattr(m, object->memattr); 1612 } else 1613 m->pindex = pindex; 1614 1615 /* 1616 * Don't wakeup too often - wakeup the pageout daemon when 1617 * we would be nearly out of memory. 1618 */ 1619 if (vm_paging_needed()) 1620 pagedaemon_wakeup(); 1621 1622 return (m); 1623 } 1624 1625 /* 1626 * vm_page_alloc_contig: 1627 * 1628 * Allocate a contiguous set of physical pages of the given size "npages" 1629 * from the free lists. All of the physical pages must be at or above 1630 * the given physical address "low" and below the given physical address 1631 * "high". The given value "alignment" determines the alignment of the 1632 * first physical page in the set. If the given value "boundary" is 1633 * non-zero, then the set of physical pages cannot cross any physical 1634 * address boundary that is a multiple of that value. Both "alignment" 1635 * and "boundary" must be a power of two. 1636 * 1637 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, 1638 * then the memory attribute setting for the physical pages is configured 1639 * to the object's memory attribute setting. Otherwise, the memory 1640 * attribute setting for the physical pages is configured to "memattr", 1641 * overriding the object's memory attribute setting. However, if the 1642 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the 1643 * memory attribute setting for the physical pages cannot be configured 1644 * to VM_MEMATTR_DEFAULT. 1645 * 1646 * The caller must always specify an allocation class. 1647 * 1648 * allocation classes: 1649 * VM_ALLOC_NORMAL normal process request 1650 * VM_ALLOC_SYSTEM system *really* needs a page 1651 * VM_ALLOC_INTERRUPT interrupt time request 1652 * 1653 * optional allocation flags: 1654 * VM_ALLOC_NOBUSY do not exclusive busy the page 1655 * VM_ALLOC_NODUMP do not include the page in a kernel core dump 1656 * VM_ALLOC_NOOBJ page is not associated with an object and 1657 * should not be exclusive busy 1658 * VM_ALLOC_SBUSY shared busy the allocated page 1659 * VM_ALLOC_WIRED wire the allocated page 1660 * VM_ALLOC_ZERO prefer a zeroed page 1661 * 1662 * This routine may not sleep. 1663 */ 1664 vm_page_t 1665 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, 1666 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, 1667 vm_paddr_t boundary, vm_memattr_t memattr) 1668 { 1669 vm_page_t m, m_tmp, m_ret; 1670 u_int flags; 1671 int req_class; 1672 1673 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && 1674 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && 1675 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != 1676 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), 1677 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object, 1678 req)); 1679 if (object != NULL) { 1680 VM_OBJECT_ASSERT_WLOCKED(object); 1681 KASSERT(object->type == OBJT_PHYS, 1682 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS", 1683 object)); 1684 } 1685 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); 1686 req_class = req & VM_ALLOC_CLASS_MASK; 1687 1688 /* 1689 * The page daemon is allowed to dig deeper into the free page list. 1690 */ 1691 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1692 req_class = VM_ALLOC_SYSTEM; 1693 1694 mtx_lock(&vm_page_queue_free_mtx); 1695 if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved || 1696 (req_class == VM_ALLOC_SYSTEM && 1697 vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) || 1698 (req_class == VM_ALLOC_INTERRUPT && 1699 vm_cnt.v_free_count >= npages)) { 1700 #if VM_NRESERVLEVEL > 0 1701 retry: 1702 if (object == NULL || (object->flags & OBJ_COLORED) == 0 || 1703 (m_ret = vm_reserv_alloc_contig(object, pindex, npages, 1704 low, high, alignment, boundary)) == NULL) 1705 #endif 1706 m_ret = vm_phys_alloc_contig(npages, low, high, 1707 alignment, boundary); 1708 } else { 1709 mtx_unlock(&vm_page_queue_free_mtx); 1710 atomic_add_int(&vm_pageout_deficit, npages); 1711 pagedaemon_wakeup(); 1712 return (NULL); 1713 } 1714 if (m_ret != NULL) 1715 vm_phys_freecnt_adj(m_ret, -npages); 1716 else { 1717 #if VM_NRESERVLEVEL > 0 1718 if (vm_reserv_reclaim_contig(npages, low, high, alignment, 1719 boundary)) 1720 goto retry; 1721 #endif 1722 } 1723 mtx_unlock(&vm_page_queue_free_mtx); 1724 if (m_ret == NULL) 1725 return (NULL); 1726 for (m = m_ret; m < &m_ret[npages]; m++) 1727 vm_page_alloc_check(m); 1728 1729 /* 1730 * Initialize the pages. Only the PG_ZERO flag is inherited. 1731 */ 1732 flags = 0; 1733 if ((req & VM_ALLOC_ZERO) != 0) 1734 flags = PG_ZERO; 1735 if ((req & VM_ALLOC_NODUMP) != 0) 1736 flags |= PG_NODUMP; 1737 if ((req & VM_ALLOC_WIRED) != 0) 1738 atomic_add_int(&vm_cnt.v_wire_count, npages); 1739 if (object != NULL) { 1740 if (object->memattr != VM_MEMATTR_DEFAULT && 1741 memattr == VM_MEMATTR_DEFAULT) 1742 memattr = object->memattr; 1743 } 1744 for (m = m_ret; m < &m_ret[npages]; m++) { 1745 m->aflags = 0; 1746 m->flags = (m->flags | PG_NODUMP) & flags; 1747 m->busy_lock = VPB_UNBUSIED; 1748 if (object != NULL) { 1749 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 1750 m->busy_lock = VPB_SINGLE_EXCLUSIVER; 1751 if ((req & VM_ALLOC_SBUSY) != 0) 1752 m->busy_lock = VPB_SHARERS_WORD(1); 1753 } 1754 if ((req & VM_ALLOC_WIRED) != 0) 1755 m->wire_count = 1; 1756 /* Unmanaged pages don't use "act_count". */ 1757 m->oflags = VPO_UNMANAGED; 1758 if (object != NULL) { 1759 if (vm_page_insert(m, object, pindex)) { 1760 if (vm_paging_needed()) 1761 pagedaemon_wakeup(); 1762 if ((req & VM_ALLOC_WIRED) != 0) 1763 atomic_subtract_int(&vm_cnt.v_wire_count, 1764 npages); 1765 for (m_tmp = m, m = m_ret; 1766 m < &m_ret[npages]; m++) { 1767 if ((req & VM_ALLOC_WIRED) != 0) 1768 m->wire_count = 0; 1769 if (m >= m_tmp) { 1770 m->object = NULL; 1771 m->oflags |= VPO_UNMANAGED; 1772 } 1773 m->busy_lock = VPB_UNBUSIED; 1774 vm_page_free(m); 1775 } 1776 return (NULL); 1777 } 1778 } else 1779 m->pindex = pindex; 1780 if (memattr != VM_MEMATTR_DEFAULT) 1781 pmap_page_set_memattr(m, memattr); 1782 pindex++; 1783 } 1784 if (vm_paging_needed()) 1785 pagedaemon_wakeup(); 1786 return (m_ret); 1787 } 1788 1789 /* 1790 * Check a page that has been freshly dequeued from a freelist. 1791 */ 1792 static void 1793 vm_page_alloc_check(vm_page_t m) 1794 { 1795 1796 KASSERT(m->queue == PQ_NONE, 1797 ("page %p has unexpected queue %d", m, m->queue)); 1798 KASSERT(m->wire_count == 0, ("page %p is wired", m)); 1799 KASSERT(m->hold_count == 0, ("page %p is held", m)); 1800 KASSERT(!vm_page_busied(m), ("page %p is busy", m)); 1801 KASSERT(m->dirty == 0, ("page %p is dirty", m)); 1802 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1803 ("page %p has unexpected memattr %d", 1804 m, pmap_page_get_memattr(m))); 1805 KASSERT(m->valid == 0, ("free page %p is valid", m)); 1806 } 1807 1808 /* 1809 * vm_page_alloc_freelist: 1810 * 1811 * Allocate a physical page from the specified free page list. 1812 * 1813 * The caller must always specify an allocation class. 1814 * 1815 * allocation classes: 1816 * VM_ALLOC_NORMAL normal process request 1817 * VM_ALLOC_SYSTEM system *really* needs a page 1818 * VM_ALLOC_INTERRUPT interrupt time request 1819 * 1820 * optional allocation flags: 1821 * VM_ALLOC_COUNT(number) the number of additional pages that the caller 1822 * intends to allocate 1823 * VM_ALLOC_WIRED wire the allocated page 1824 * VM_ALLOC_ZERO prefer a zeroed page 1825 * 1826 * This routine may not sleep. 1827 */ 1828 vm_page_t 1829 vm_page_alloc_freelist(int flind, int req) 1830 { 1831 vm_page_t m; 1832 u_int flags; 1833 int req_class; 1834 1835 req_class = req & VM_ALLOC_CLASS_MASK; 1836 1837 /* 1838 * The page daemon is allowed to dig deeper into the free page list. 1839 */ 1840 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 1841 req_class = VM_ALLOC_SYSTEM; 1842 1843 /* 1844 * Do not allocate reserved pages unless the req has asked for it. 1845 */ 1846 mtx_lock(&vm_page_queue_free_mtx); 1847 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved || 1848 (req_class == VM_ALLOC_SYSTEM && 1849 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) || 1850 (req_class == VM_ALLOC_INTERRUPT && 1851 vm_cnt.v_free_count > 0)) 1852 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0); 1853 else { 1854 mtx_unlock(&vm_page_queue_free_mtx); 1855 atomic_add_int(&vm_pageout_deficit, 1856 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); 1857 pagedaemon_wakeup(); 1858 return (NULL); 1859 } 1860 if (m == NULL) { 1861 mtx_unlock(&vm_page_queue_free_mtx); 1862 return (NULL); 1863 } 1864 vm_phys_freecnt_adj(m, -1); 1865 mtx_unlock(&vm_page_queue_free_mtx); 1866 vm_page_alloc_check(m); 1867 1868 /* 1869 * Initialize the page. Only the PG_ZERO flag is inherited. 1870 */ 1871 m->aflags = 0; 1872 flags = 0; 1873 if ((req & VM_ALLOC_ZERO) != 0) 1874 flags = PG_ZERO; 1875 m->flags &= flags; 1876 if ((req & VM_ALLOC_WIRED) != 0) { 1877 /* 1878 * The page lock is not required for wiring a page that does 1879 * not belong to an object. 1880 */ 1881 atomic_add_int(&vm_cnt.v_wire_count, 1); 1882 m->wire_count = 1; 1883 } 1884 /* Unmanaged pages don't use "act_count". */ 1885 m->oflags = VPO_UNMANAGED; 1886 if (vm_paging_needed()) 1887 pagedaemon_wakeup(); 1888 return (m); 1889 } 1890 1891 #define VPSC_ANY 0 /* No restrictions. */ 1892 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ 1893 #define VPSC_NOSUPER 2 /* Skip superpages. */ 1894 1895 /* 1896 * vm_page_scan_contig: 1897 * 1898 * Scan vm_page_array[] between the specified entries "m_start" and 1899 * "m_end" for a run of contiguous physical pages that satisfy the 1900 * specified conditions, and return the lowest page in the run. The 1901 * specified "alignment" determines the alignment of the lowest physical 1902 * page in the run. If the specified "boundary" is non-zero, then the 1903 * run of physical pages cannot span a physical address that is a 1904 * multiple of "boundary". 1905 * 1906 * "m_end" is never dereferenced, so it need not point to a vm_page 1907 * structure within vm_page_array[]. 1908 * 1909 * "npages" must be greater than zero. "m_start" and "m_end" must not 1910 * span a hole (or discontiguity) in the physical address space. Both 1911 * "alignment" and "boundary" must be a power of two. 1912 */ 1913 vm_page_t 1914 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, 1915 u_long alignment, vm_paddr_t boundary, int options) 1916 { 1917 struct mtx *m_mtx, *new_mtx; 1918 vm_object_t object; 1919 vm_paddr_t pa; 1920 vm_page_t m, m_run; 1921 #if VM_NRESERVLEVEL > 0 1922 int level; 1923 #endif 1924 int m_inc, order, run_ext, run_len; 1925 1926 KASSERT(npages > 0, ("npages is 0")); 1927 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 1928 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 1929 m_run = NULL; 1930 run_len = 0; 1931 m_mtx = NULL; 1932 for (m = m_start; m < m_end && run_len < npages; m += m_inc) { 1933 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 1934 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 1935 1936 /* 1937 * If the current page would be the start of a run, check its 1938 * physical address against the end, alignment, and boundary 1939 * conditions. If it doesn't satisfy these conditions, either 1940 * terminate the scan or advance to the next page that 1941 * satisfies the failed condition. 1942 */ 1943 if (run_len == 0) { 1944 KASSERT(m_run == NULL, ("m_run != NULL")); 1945 if (m + npages > m_end) 1946 break; 1947 pa = VM_PAGE_TO_PHYS(m); 1948 if ((pa & (alignment - 1)) != 0) { 1949 m_inc = atop(roundup2(pa, alignment) - pa); 1950 continue; 1951 } 1952 if (rounddown2(pa ^ (pa + ptoa(npages) - 1), 1953 boundary) != 0) { 1954 m_inc = atop(roundup2(pa, boundary) - pa); 1955 continue; 1956 } 1957 } else 1958 KASSERT(m_run != NULL, ("m_run == NULL")); 1959 1960 /* 1961 * Avoid releasing and reacquiring the same page lock. 1962 */ 1963 new_mtx = vm_page_lockptr(m); 1964 if (m_mtx != new_mtx) { 1965 if (m_mtx != NULL) 1966 mtx_unlock(m_mtx); 1967 m_mtx = new_mtx; 1968 mtx_lock(m_mtx); 1969 } 1970 m_inc = 1; 1971 retry: 1972 if (m->wire_count != 0 || m->hold_count != 0) 1973 run_ext = 0; 1974 #if VM_NRESERVLEVEL > 0 1975 else if ((level = vm_reserv_level(m)) >= 0 && 1976 (options & VPSC_NORESERV) != 0) { 1977 run_ext = 0; 1978 /* Advance to the end of the reservation. */ 1979 pa = VM_PAGE_TO_PHYS(m); 1980 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - 1981 pa); 1982 } 1983 #endif 1984 else if ((object = m->object) != NULL) { 1985 /* 1986 * The page is considered eligible for relocation if 1987 * and only if it could be laundered or reclaimed by 1988 * the page daemon. 1989 */ 1990 if (!VM_OBJECT_TRYRLOCK(object)) { 1991 mtx_unlock(m_mtx); 1992 VM_OBJECT_RLOCK(object); 1993 mtx_lock(m_mtx); 1994 if (m->object != object) { 1995 /* 1996 * The page may have been freed. 1997 */ 1998 VM_OBJECT_RUNLOCK(object); 1999 goto retry; 2000 } else if (m->wire_count != 0 || 2001 m->hold_count != 0) { 2002 run_ext = 0; 2003 goto unlock; 2004 } 2005 } 2006 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2007 ("page %p is PG_UNHOLDFREE", m)); 2008 /* Don't care: PG_NODUMP, PG_ZERO. */ 2009 if (object->type != OBJT_DEFAULT && 2010 object->type != OBJT_SWAP && 2011 object->type != OBJT_VNODE) { 2012 run_ext = 0; 2013 #if VM_NRESERVLEVEL > 0 2014 } else if ((options & VPSC_NOSUPER) != 0 && 2015 (level = vm_reserv_level_iffullpop(m)) >= 0) { 2016 run_ext = 0; 2017 /* Advance to the end of the superpage. */ 2018 pa = VM_PAGE_TO_PHYS(m); 2019 m_inc = atop(roundup2(pa + 1, 2020 vm_reserv_size(level)) - pa); 2021 #endif 2022 } else if (object->memattr == VM_MEMATTR_DEFAULT && 2023 m->queue != PQ_NONE && !vm_page_busied(m)) { 2024 /* 2025 * The page is allocated but eligible for 2026 * relocation. Extend the current run by one 2027 * page. 2028 */ 2029 KASSERT(pmap_page_get_memattr(m) == 2030 VM_MEMATTR_DEFAULT, 2031 ("page %p has an unexpected memattr", m)); 2032 KASSERT((m->oflags & (VPO_SWAPINPROG | 2033 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2034 ("page %p has unexpected oflags", m)); 2035 /* Don't care: VPO_NOSYNC. */ 2036 run_ext = 1; 2037 } else 2038 run_ext = 0; 2039 unlock: 2040 VM_OBJECT_RUNLOCK(object); 2041 #if VM_NRESERVLEVEL > 0 2042 } else if (level >= 0) { 2043 /* 2044 * The page is reserved but not yet allocated. In 2045 * other words, it is still cached or free. Extend 2046 * the current run by one page. 2047 */ 2048 run_ext = 1; 2049 #endif 2050 } else if ((order = m->order) < VM_NFREEORDER) { 2051 /* 2052 * The page is enqueued in the physical memory 2053 * allocator's cache/free page queues. Moreover, it 2054 * is the first page in a power-of-two-sized run of 2055 * contiguous cache/free pages. Add these pages to 2056 * the end of the current run, and jump ahead. 2057 */ 2058 run_ext = 1 << order; 2059 m_inc = 1 << order; 2060 } else { 2061 /* 2062 * Skip the page for one of the following reasons: (1) 2063 * It is enqueued in the physical memory allocator's 2064 * cache/free page queues. However, it is not the 2065 * first page in a run of contiguous cache/free pages. 2066 * (This case rarely occurs because the scan is 2067 * performed in ascending order.) (2) It is not 2068 * reserved, and it is transitioning from free to 2069 * allocated. (Conversely, the transition from 2070 * allocated to free for managed pages is blocked by 2071 * the page lock.) (3) It is allocated but not 2072 * contained by an object and not wired, e.g., 2073 * allocated by Xen's balloon driver. 2074 */ 2075 run_ext = 0; 2076 } 2077 2078 /* 2079 * Extend or reset the current run of pages. 2080 */ 2081 if (run_ext > 0) { 2082 if (run_len == 0) 2083 m_run = m; 2084 run_len += run_ext; 2085 } else { 2086 if (run_len > 0) { 2087 m_run = NULL; 2088 run_len = 0; 2089 } 2090 } 2091 } 2092 if (m_mtx != NULL) 2093 mtx_unlock(m_mtx); 2094 if (run_len >= npages) 2095 return (m_run); 2096 return (NULL); 2097 } 2098 2099 /* 2100 * vm_page_reclaim_run: 2101 * 2102 * Try to relocate each of the allocated virtual pages within the 2103 * specified run of physical pages to a new physical address. Free the 2104 * physical pages underlying the relocated virtual pages. A virtual page 2105 * is relocatable if and only if it could be laundered or reclaimed by 2106 * the page daemon. Whenever possible, a virtual page is relocated to a 2107 * physical address above "high". 2108 * 2109 * Returns 0 if every physical page within the run was already free or 2110 * just freed by a successful relocation. Otherwise, returns a non-zero 2111 * value indicating why the last attempt to relocate a virtual page was 2112 * unsuccessful. 2113 * 2114 * "req_class" must be an allocation class. 2115 */ 2116 static int 2117 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run, 2118 vm_paddr_t high) 2119 { 2120 struct mtx *m_mtx, *new_mtx; 2121 struct spglist free; 2122 vm_object_t object; 2123 vm_paddr_t pa; 2124 vm_page_t m, m_end, m_new; 2125 int error, order, req; 2126 2127 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, 2128 ("req_class is not an allocation class")); 2129 SLIST_INIT(&free); 2130 error = 0; 2131 m = m_run; 2132 m_end = m_run + npages; 2133 m_mtx = NULL; 2134 for (; error == 0 && m < m_end; m++) { 2135 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, 2136 ("page %p is PG_FICTITIOUS or PG_MARKER", m)); 2137 2138 /* 2139 * Avoid releasing and reacquiring the same page lock. 2140 */ 2141 new_mtx = vm_page_lockptr(m); 2142 if (m_mtx != new_mtx) { 2143 if (m_mtx != NULL) 2144 mtx_unlock(m_mtx); 2145 m_mtx = new_mtx; 2146 mtx_lock(m_mtx); 2147 } 2148 retry: 2149 if (m->wire_count != 0 || m->hold_count != 0) 2150 error = EBUSY; 2151 else if ((object = m->object) != NULL) { 2152 /* 2153 * The page is relocated if and only if it could be 2154 * laundered or reclaimed by the page daemon. 2155 */ 2156 if (!VM_OBJECT_TRYWLOCK(object)) { 2157 mtx_unlock(m_mtx); 2158 VM_OBJECT_WLOCK(object); 2159 mtx_lock(m_mtx); 2160 if (m->object != object) { 2161 /* 2162 * The page may have been freed. 2163 */ 2164 VM_OBJECT_WUNLOCK(object); 2165 goto retry; 2166 } else if (m->wire_count != 0 || 2167 m->hold_count != 0) { 2168 error = EBUSY; 2169 goto unlock; 2170 } 2171 } 2172 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2173 ("page %p is PG_UNHOLDFREE", m)); 2174 /* Don't care: PG_NODUMP, PG_ZERO. */ 2175 if (object->type != OBJT_DEFAULT && 2176 object->type != OBJT_SWAP && 2177 object->type != OBJT_VNODE) 2178 error = EINVAL; 2179 else if (object->memattr != VM_MEMATTR_DEFAULT) 2180 error = EINVAL; 2181 else if (m->queue != PQ_NONE && !vm_page_busied(m)) { 2182 KASSERT(pmap_page_get_memattr(m) == 2183 VM_MEMATTR_DEFAULT, 2184 ("page %p has an unexpected memattr", m)); 2185 KASSERT((m->oflags & (VPO_SWAPINPROG | 2186 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, 2187 ("page %p has unexpected oflags", m)); 2188 /* Don't care: VPO_NOSYNC. */ 2189 if (m->valid != 0) { 2190 /* 2191 * First, try to allocate a new page 2192 * that is above "high". Failing 2193 * that, try to allocate a new page 2194 * that is below "m_run". Allocate 2195 * the new page between the end of 2196 * "m_run" and "high" only as a last 2197 * resort. 2198 */ 2199 req = req_class | VM_ALLOC_NOOBJ; 2200 if ((m->flags & PG_NODUMP) != 0) 2201 req |= VM_ALLOC_NODUMP; 2202 if (trunc_page(high) != 2203 ~(vm_paddr_t)PAGE_MASK) { 2204 m_new = vm_page_alloc_contig( 2205 NULL, 0, req, 1, 2206 round_page(high), 2207 ~(vm_paddr_t)0, 2208 PAGE_SIZE, 0, 2209 VM_MEMATTR_DEFAULT); 2210 } else 2211 m_new = NULL; 2212 if (m_new == NULL) { 2213 pa = VM_PAGE_TO_PHYS(m_run); 2214 m_new = vm_page_alloc_contig( 2215 NULL, 0, req, 1, 2216 0, pa - 1, PAGE_SIZE, 0, 2217 VM_MEMATTR_DEFAULT); 2218 } 2219 if (m_new == NULL) { 2220 pa += ptoa(npages); 2221 m_new = vm_page_alloc_contig( 2222 NULL, 0, req, 1, 2223 pa, high, PAGE_SIZE, 0, 2224 VM_MEMATTR_DEFAULT); 2225 } 2226 if (m_new == NULL) { 2227 error = ENOMEM; 2228 goto unlock; 2229 } 2230 KASSERT(m_new->wire_count == 0, 2231 ("page %p is wired", m)); 2232 2233 /* 2234 * Replace "m" with the new page. For 2235 * vm_page_replace(), "m" must be busy 2236 * and dequeued. Finally, change "m" 2237 * as if vm_page_free() was called. 2238 */ 2239 if (object->ref_count != 0) 2240 pmap_remove_all(m); 2241 m_new->aflags = m->aflags; 2242 KASSERT(m_new->oflags == VPO_UNMANAGED, 2243 ("page %p is managed", m)); 2244 m_new->oflags = m->oflags & VPO_NOSYNC; 2245 pmap_copy_page(m, m_new); 2246 m_new->valid = m->valid; 2247 m_new->dirty = m->dirty; 2248 m->flags &= ~PG_ZERO; 2249 vm_page_xbusy(m); 2250 vm_page_remque(m); 2251 vm_page_replace_checked(m_new, object, 2252 m->pindex, m); 2253 m->valid = 0; 2254 vm_page_undirty(m); 2255 2256 /* 2257 * The new page must be deactivated 2258 * before the object is unlocked. 2259 */ 2260 new_mtx = vm_page_lockptr(m_new); 2261 if (m_mtx != new_mtx) { 2262 mtx_unlock(m_mtx); 2263 m_mtx = new_mtx; 2264 mtx_lock(m_mtx); 2265 } 2266 vm_page_deactivate(m_new); 2267 } else { 2268 m->flags &= ~PG_ZERO; 2269 vm_page_remque(m); 2270 vm_page_remove(m); 2271 KASSERT(m->dirty == 0, 2272 ("page %p is dirty", m)); 2273 } 2274 SLIST_INSERT_HEAD(&free, m, plinks.s.ss); 2275 } else 2276 error = EBUSY; 2277 unlock: 2278 VM_OBJECT_WUNLOCK(object); 2279 } else { 2280 mtx_lock(&vm_page_queue_free_mtx); 2281 order = m->order; 2282 if (order < VM_NFREEORDER) { 2283 /* 2284 * The page is enqueued in the physical memory 2285 * allocator's cache/free page queues. 2286 * Moreover, it is the first page in a power- 2287 * of-two-sized run of contiguous cache/free 2288 * pages. Jump ahead to the last page within 2289 * that run, and continue from there. 2290 */ 2291 m += (1 << order) - 1; 2292 } 2293 #if VM_NRESERVLEVEL > 0 2294 else if (vm_reserv_is_page_free(m)) 2295 order = 0; 2296 #endif 2297 mtx_unlock(&vm_page_queue_free_mtx); 2298 if (order == VM_NFREEORDER) 2299 error = EINVAL; 2300 } 2301 } 2302 if (m_mtx != NULL) 2303 mtx_unlock(m_mtx); 2304 if ((m = SLIST_FIRST(&free)) != NULL) { 2305 mtx_lock(&vm_page_queue_free_mtx); 2306 do { 2307 SLIST_REMOVE_HEAD(&free, plinks.s.ss); 2308 vm_phys_freecnt_adj(m, 1); 2309 #if VM_NRESERVLEVEL > 0 2310 if (!vm_reserv_free_page(m)) 2311 #else 2312 if (true) 2313 #endif 2314 vm_phys_free_pages(m, 0); 2315 } while ((m = SLIST_FIRST(&free)) != NULL); 2316 vm_page_free_wakeup(); 2317 mtx_unlock(&vm_page_queue_free_mtx); 2318 } 2319 return (error); 2320 } 2321 2322 #define NRUNS 16 2323 2324 CTASSERT(powerof2(NRUNS)); 2325 2326 #define RUN_INDEX(count) ((count) & (NRUNS - 1)) 2327 2328 #define MIN_RECLAIM 8 2329 2330 /* 2331 * vm_page_reclaim_contig: 2332 * 2333 * Reclaim allocated, contiguous physical memory satisfying the specified 2334 * conditions by relocating the virtual pages using that physical memory. 2335 * Returns true if reclamation is successful and false otherwise. Since 2336 * relocation requires the allocation of physical pages, reclamation may 2337 * fail due to a shortage of cache/free pages. When reclamation fails, 2338 * callers are expected to perform VM_WAIT before retrying a failed 2339 * allocation operation, e.g., vm_page_alloc_contig(). 2340 * 2341 * The caller must always specify an allocation class through "req". 2342 * 2343 * allocation classes: 2344 * VM_ALLOC_NORMAL normal process request 2345 * VM_ALLOC_SYSTEM system *really* needs a page 2346 * VM_ALLOC_INTERRUPT interrupt time request 2347 * 2348 * The optional allocation flags are ignored. 2349 * 2350 * "npages" must be greater than zero. Both "alignment" and "boundary" 2351 * must be a power of two. 2352 */ 2353 bool 2354 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, 2355 u_long alignment, vm_paddr_t boundary) 2356 { 2357 vm_paddr_t curr_low; 2358 vm_page_t m_run, m_runs[NRUNS]; 2359 u_long count, reclaimed; 2360 int error, i, options, req_class; 2361 2362 KASSERT(npages > 0, ("npages is 0")); 2363 KASSERT(powerof2(alignment), ("alignment is not a power of 2")); 2364 KASSERT(powerof2(boundary), ("boundary is not a power of 2")); 2365 req_class = req & VM_ALLOC_CLASS_MASK; 2366 2367 /* 2368 * The page daemon is allowed to dig deeper into the free page list. 2369 */ 2370 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) 2371 req_class = VM_ALLOC_SYSTEM; 2372 2373 /* 2374 * Return if the number of cached and free pages cannot satisfy the 2375 * requested allocation. 2376 */ 2377 count = vm_cnt.v_free_count; 2378 if (count < npages + vm_cnt.v_free_reserved || (count < npages + 2379 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || 2380 (count < npages && req_class == VM_ALLOC_INTERRUPT)) 2381 return (false); 2382 2383 /* 2384 * Scan up to three times, relaxing the restrictions ("options") on 2385 * the reclamation of reservations and superpages each time. 2386 */ 2387 for (options = VPSC_NORESERV;;) { 2388 /* 2389 * Find the highest runs that satisfy the given constraints 2390 * and restrictions, and record them in "m_runs". 2391 */ 2392 curr_low = low; 2393 count = 0; 2394 for (;;) { 2395 m_run = vm_phys_scan_contig(npages, curr_low, high, 2396 alignment, boundary, options); 2397 if (m_run == NULL) 2398 break; 2399 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); 2400 m_runs[RUN_INDEX(count)] = m_run; 2401 count++; 2402 } 2403 2404 /* 2405 * Reclaim the highest runs in LIFO (descending) order until 2406 * the number of reclaimed pages, "reclaimed", is at least 2407 * MIN_RECLAIM. Reset "reclaimed" each time because each 2408 * reclamation is idempotent, and runs will (likely) recur 2409 * from one scan to the next as restrictions are relaxed. 2410 */ 2411 reclaimed = 0; 2412 for (i = 0; count > 0 && i < NRUNS; i++) { 2413 count--; 2414 m_run = m_runs[RUN_INDEX(count)]; 2415 error = vm_page_reclaim_run(req_class, npages, m_run, 2416 high); 2417 if (error == 0) { 2418 reclaimed += npages; 2419 if (reclaimed >= MIN_RECLAIM) 2420 return (true); 2421 } 2422 } 2423 2424 /* 2425 * Either relax the restrictions on the next scan or return if 2426 * the last scan had no restrictions. 2427 */ 2428 if (options == VPSC_NORESERV) 2429 options = VPSC_NOSUPER; 2430 else if (options == VPSC_NOSUPER) 2431 options = VPSC_ANY; 2432 else if (options == VPSC_ANY) 2433 return (reclaimed != 0); 2434 } 2435 } 2436 2437 /* 2438 * vm_wait: (also see VM_WAIT macro) 2439 * 2440 * Sleep until free pages are available for allocation. 2441 * - Called in various places before memory allocations. 2442 */ 2443 void 2444 vm_wait(void) 2445 { 2446 2447 mtx_lock(&vm_page_queue_free_mtx); 2448 if (curproc == pageproc) { 2449 vm_pageout_pages_needed = 1; 2450 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 2451 PDROP | PSWP, "VMWait", 0); 2452 } else { 2453 if (__predict_false(pageproc == NULL)) 2454 panic("vm_wait in early boot"); 2455 if (!vm_pageout_wanted) { 2456 vm_pageout_wanted = true; 2457 wakeup(&vm_pageout_wanted); 2458 } 2459 vm_pages_needed = true; 2460 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 2461 "vmwait", 0); 2462 } 2463 } 2464 2465 /* 2466 * vm_waitpfault: (also see VM_WAITPFAULT macro) 2467 * 2468 * Sleep until free pages are available for allocation. 2469 * - Called only in vm_fault so that processes page faulting 2470 * can be easily tracked. 2471 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 2472 * processes will be able to grab memory first. Do not change 2473 * this balance without careful testing first. 2474 */ 2475 void 2476 vm_waitpfault(void) 2477 { 2478 2479 mtx_lock(&vm_page_queue_free_mtx); 2480 if (!vm_pageout_wanted) { 2481 vm_pageout_wanted = true; 2482 wakeup(&vm_pageout_wanted); 2483 } 2484 vm_pages_needed = true; 2485 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 2486 "pfault", 0); 2487 } 2488 2489 struct vm_pagequeue * 2490 vm_page_pagequeue(vm_page_t m) 2491 { 2492 2493 if (vm_page_in_laundry(m)) 2494 return (&vm_dom[0].vmd_pagequeues[m->queue]); 2495 else 2496 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]); 2497 } 2498 2499 /* 2500 * vm_page_dequeue: 2501 * 2502 * Remove the given page from its current page queue. 2503 * 2504 * The page must be locked. 2505 */ 2506 void 2507 vm_page_dequeue(vm_page_t m) 2508 { 2509 struct vm_pagequeue *pq; 2510 2511 vm_page_assert_locked(m); 2512 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued", 2513 m)); 2514 pq = vm_page_pagequeue(m); 2515 vm_pagequeue_lock(pq); 2516 m->queue = PQ_NONE; 2517 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2518 vm_pagequeue_cnt_dec(pq); 2519 vm_pagequeue_unlock(pq); 2520 } 2521 2522 /* 2523 * vm_page_dequeue_locked: 2524 * 2525 * Remove the given page from its current page queue. 2526 * 2527 * The page and page queue must be locked. 2528 */ 2529 void 2530 vm_page_dequeue_locked(vm_page_t m) 2531 { 2532 struct vm_pagequeue *pq; 2533 2534 vm_page_lock_assert(m, MA_OWNED); 2535 pq = vm_page_pagequeue(m); 2536 vm_pagequeue_assert_locked(pq); 2537 m->queue = PQ_NONE; 2538 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2539 vm_pagequeue_cnt_dec(pq); 2540 } 2541 2542 /* 2543 * vm_page_enqueue: 2544 * 2545 * Add the given page to the specified page queue. 2546 * 2547 * The page must be locked. 2548 */ 2549 static void 2550 vm_page_enqueue(uint8_t queue, vm_page_t m) 2551 { 2552 struct vm_pagequeue *pq; 2553 2554 vm_page_lock_assert(m, MA_OWNED); 2555 KASSERT(queue < PQ_COUNT, 2556 ("vm_page_enqueue: invalid queue %u request for page %p", 2557 queue, m)); 2558 if (queue == PQ_LAUNDRY) 2559 pq = &vm_dom[0].vmd_pagequeues[queue]; 2560 else 2561 pq = &vm_phys_domain(m)->vmd_pagequeues[queue]; 2562 vm_pagequeue_lock(pq); 2563 m->queue = queue; 2564 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2565 vm_pagequeue_cnt_inc(pq); 2566 vm_pagequeue_unlock(pq); 2567 } 2568 2569 /* 2570 * vm_page_requeue: 2571 * 2572 * Move the given page to the tail of its current page queue. 2573 * 2574 * The page must be locked. 2575 */ 2576 void 2577 vm_page_requeue(vm_page_t m) 2578 { 2579 struct vm_pagequeue *pq; 2580 2581 vm_page_lock_assert(m, MA_OWNED); 2582 KASSERT(m->queue != PQ_NONE, 2583 ("vm_page_requeue: page %p is not queued", m)); 2584 pq = vm_page_pagequeue(m); 2585 vm_pagequeue_lock(pq); 2586 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2587 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2588 vm_pagequeue_unlock(pq); 2589 } 2590 2591 /* 2592 * vm_page_requeue_locked: 2593 * 2594 * Move the given page to the tail of its current page queue. 2595 * 2596 * The page queue must be locked. 2597 */ 2598 void 2599 vm_page_requeue_locked(vm_page_t m) 2600 { 2601 struct vm_pagequeue *pq; 2602 2603 KASSERT(m->queue != PQ_NONE, 2604 ("vm_page_requeue_locked: page %p is not queued", m)); 2605 pq = vm_page_pagequeue(m); 2606 vm_pagequeue_assert_locked(pq); 2607 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); 2608 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2609 } 2610 2611 /* 2612 * vm_page_activate: 2613 * 2614 * Put the specified page on the active list (if appropriate). 2615 * Ensure that act_count is at least ACT_INIT but do not otherwise 2616 * mess with it. 2617 * 2618 * The page must be locked. 2619 */ 2620 void 2621 vm_page_activate(vm_page_t m) 2622 { 2623 int queue; 2624 2625 vm_page_lock_assert(m, MA_OWNED); 2626 if ((queue = m->queue) != PQ_ACTIVE) { 2627 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2628 if (m->act_count < ACT_INIT) 2629 m->act_count = ACT_INIT; 2630 if (queue != PQ_NONE) 2631 vm_page_dequeue(m); 2632 vm_page_enqueue(PQ_ACTIVE, m); 2633 } else 2634 KASSERT(queue == PQ_NONE, 2635 ("vm_page_activate: wired page %p is queued", m)); 2636 } else { 2637 if (m->act_count < ACT_INIT) 2638 m->act_count = ACT_INIT; 2639 } 2640 } 2641 2642 /* 2643 * vm_page_free_wakeup: 2644 * 2645 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 2646 * routine is called when a page has been added to the cache or free 2647 * queues. 2648 * 2649 * The page queues must be locked. 2650 */ 2651 static inline void 2652 vm_page_free_wakeup(void) 2653 { 2654 2655 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 2656 /* 2657 * if pageout daemon needs pages, then tell it that there are 2658 * some free. 2659 */ 2660 if (vm_pageout_pages_needed && 2661 vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) { 2662 wakeup(&vm_pageout_pages_needed); 2663 vm_pageout_pages_needed = 0; 2664 } 2665 /* 2666 * wakeup processes that are waiting on memory if we hit a 2667 * high water mark. And wakeup scheduler process if we have 2668 * lots of memory. this process will swapin processes. 2669 */ 2670 if (vm_pages_needed && !vm_page_count_min()) { 2671 vm_pages_needed = false; 2672 wakeup(&vm_cnt.v_free_count); 2673 } 2674 } 2675 2676 /* 2677 * vm_page_free_toq: 2678 * 2679 * Returns the given page to the free list, 2680 * disassociating it with any VM object. 2681 * 2682 * The object must be locked. The page must be locked if it is managed. 2683 */ 2684 void 2685 vm_page_free_toq(vm_page_t m) 2686 { 2687 2688 if ((m->oflags & VPO_UNMANAGED) == 0) { 2689 vm_page_lock_assert(m, MA_OWNED); 2690 KASSERT(!pmap_page_is_mapped(m), 2691 ("vm_page_free_toq: freeing mapped page %p", m)); 2692 } else 2693 KASSERT(m->queue == PQ_NONE, 2694 ("vm_page_free_toq: unmanaged page %p is queued", m)); 2695 PCPU_INC(cnt.v_tfree); 2696 2697 if (vm_page_sbusied(m)) 2698 panic("vm_page_free: freeing busy page %p", m); 2699 2700 /* 2701 * Unqueue, then remove page. Note that we cannot destroy 2702 * the page here because we do not want to call the pager's 2703 * callback routine until after we've put the page on the 2704 * appropriate free queue. 2705 */ 2706 vm_page_remque(m); 2707 vm_page_remove(m); 2708 2709 /* 2710 * If fictitious remove object association and 2711 * return, otherwise delay object association removal. 2712 */ 2713 if ((m->flags & PG_FICTITIOUS) != 0) { 2714 return; 2715 } 2716 2717 m->valid = 0; 2718 vm_page_undirty(m); 2719 2720 if (m->wire_count != 0) 2721 panic("vm_page_free: freeing wired page %p", m); 2722 if (m->hold_count != 0) { 2723 m->flags &= ~PG_ZERO; 2724 KASSERT((m->flags & PG_UNHOLDFREE) == 0, 2725 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m)); 2726 m->flags |= PG_UNHOLDFREE; 2727 } else { 2728 /* 2729 * Restore the default memory attribute to the page. 2730 */ 2731 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 2732 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 2733 2734 /* 2735 * Insert the page into the physical memory allocator's 2736 * cache/free page queues. 2737 */ 2738 mtx_lock(&vm_page_queue_free_mtx); 2739 vm_phys_freecnt_adj(m, 1); 2740 #if VM_NRESERVLEVEL > 0 2741 if (!vm_reserv_free_page(m)) 2742 #else 2743 if (TRUE) 2744 #endif 2745 vm_phys_free_pages(m, 0); 2746 vm_page_free_wakeup(); 2747 mtx_unlock(&vm_page_queue_free_mtx); 2748 } 2749 } 2750 2751 /* 2752 * vm_page_wire: 2753 * 2754 * Mark this page as wired down by yet 2755 * another map, removing it from paging queues 2756 * as necessary. 2757 * 2758 * If the page is fictitious, then its wire count must remain one. 2759 * 2760 * The page must be locked. 2761 */ 2762 void 2763 vm_page_wire(vm_page_t m) 2764 { 2765 2766 /* 2767 * Only bump the wire statistics if the page is not already wired, 2768 * and only unqueue the page if it is on some queue (if it is unmanaged 2769 * it is already off the queues). 2770 */ 2771 vm_page_lock_assert(m, MA_OWNED); 2772 if ((m->flags & PG_FICTITIOUS) != 0) { 2773 KASSERT(m->wire_count == 1, 2774 ("vm_page_wire: fictitious page %p's wire count isn't one", 2775 m)); 2776 return; 2777 } 2778 if (m->wire_count == 0) { 2779 KASSERT((m->oflags & VPO_UNMANAGED) == 0 || 2780 m->queue == PQ_NONE, 2781 ("vm_page_wire: unmanaged page %p is queued", m)); 2782 vm_page_remque(m); 2783 atomic_add_int(&vm_cnt.v_wire_count, 1); 2784 } 2785 m->wire_count++; 2786 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 2787 } 2788 2789 /* 2790 * vm_page_unwire: 2791 * 2792 * Release one wiring of the specified page, potentially allowing it to be 2793 * paged out. Returns TRUE if the number of wirings transitions to zero and 2794 * FALSE otherwise. 2795 * 2796 * Only managed pages belonging to an object can be paged out. If the number 2797 * of wirings transitions to zero and the page is eligible for page out, then 2798 * the page is added to the specified paging queue (unless PQ_NONE is 2799 * specified). 2800 * 2801 * If a page is fictitious, then its wire count must always be one. 2802 * 2803 * A managed page must be locked. 2804 */ 2805 boolean_t 2806 vm_page_unwire(vm_page_t m, uint8_t queue) 2807 { 2808 2809 KASSERT(queue < PQ_COUNT || queue == PQ_NONE, 2810 ("vm_page_unwire: invalid queue %u request for page %p", 2811 queue, m)); 2812 if ((m->oflags & VPO_UNMANAGED) == 0) 2813 vm_page_assert_locked(m); 2814 if ((m->flags & PG_FICTITIOUS) != 0) { 2815 KASSERT(m->wire_count == 1, 2816 ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); 2817 return (FALSE); 2818 } 2819 if (m->wire_count > 0) { 2820 m->wire_count--; 2821 if (m->wire_count == 0) { 2822 atomic_subtract_int(&vm_cnt.v_wire_count, 1); 2823 if ((m->oflags & VPO_UNMANAGED) == 0 && 2824 m->object != NULL && queue != PQ_NONE) 2825 vm_page_enqueue(queue, m); 2826 return (TRUE); 2827 } else 2828 return (FALSE); 2829 } else 2830 panic("vm_page_unwire: page %p's wire count is zero", m); 2831 } 2832 2833 /* 2834 * Move the specified page to the inactive queue. 2835 * 2836 * Many pages placed on the inactive queue should actually go 2837 * into the cache, but it is difficult to figure out which. What 2838 * we do instead, if the inactive target is well met, is to put 2839 * clean pages at the head of the inactive queue instead of the tail. 2840 * This will cause them to be moved to the cache more quickly and 2841 * if not actively re-referenced, reclaimed more quickly. If we just 2842 * stick these pages at the end of the inactive queue, heavy filesystem 2843 * meta-data accesses can cause an unnecessary paging load on memory bound 2844 * processes. This optimization causes one-time-use metadata to be 2845 * reused more quickly. 2846 * 2847 * Normally noreuse is FALSE, resulting in LRU operation. noreuse is set 2848 * to TRUE if we want this page to be 'as if it were placed in the cache', 2849 * except without unmapping it from the process address space. In 2850 * practice this is implemented by inserting the page at the head of the 2851 * queue, using a marker page to guide FIFO insertion ordering. 2852 * 2853 * The page must be locked. 2854 */ 2855 static inline void 2856 _vm_page_deactivate(vm_page_t m, boolean_t noreuse) 2857 { 2858 struct vm_pagequeue *pq; 2859 int queue; 2860 2861 vm_page_assert_locked(m); 2862 2863 /* 2864 * Ignore if the page is already inactive, unless it is unlikely to be 2865 * reactivated. 2866 */ 2867 if ((queue = m->queue) == PQ_INACTIVE && !noreuse) 2868 return; 2869 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2870 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE]; 2871 /* Avoid multiple acquisitions of the inactive queue lock. */ 2872 if (queue == PQ_INACTIVE) { 2873 vm_pagequeue_lock(pq); 2874 vm_page_dequeue_locked(m); 2875 } else { 2876 if (queue != PQ_NONE) 2877 vm_page_dequeue(m); 2878 vm_pagequeue_lock(pq); 2879 } 2880 m->queue = PQ_INACTIVE; 2881 if (noreuse) 2882 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead, 2883 m, plinks.q); 2884 else 2885 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); 2886 vm_pagequeue_cnt_inc(pq); 2887 vm_pagequeue_unlock(pq); 2888 } 2889 } 2890 2891 /* 2892 * Move the specified page to the inactive queue. 2893 * 2894 * The page must be locked. 2895 */ 2896 void 2897 vm_page_deactivate(vm_page_t m) 2898 { 2899 2900 _vm_page_deactivate(m, FALSE); 2901 } 2902 2903 /* 2904 * Move the specified page to the inactive queue with the expectation 2905 * that it is unlikely to be reused. 2906 * 2907 * The page must be locked. 2908 */ 2909 void 2910 vm_page_deactivate_noreuse(vm_page_t m) 2911 { 2912 2913 _vm_page_deactivate(m, TRUE); 2914 } 2915 2916 /* 2917 * vm_page_launder 2918 * 2919 * Put a page in the laundry. 2920 */ 2921 void 2922 vm_page_launder(vm_page_t m) 2923 { 2924 int queue; 2925 2926 vm_page_assert_locked(m); 2927 if ((queue = m->queue) != PQ_LAUNDRY) { 2928 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) { 2929 if (queue != PQ_NONE) 2930 vm_page_dequeue(m); 2931 vm_page_enqueue(PQ_LAUNDRY, m); 2932 } else 2933 KASSERT(queue == PQ_NONE, 2934 ("wired page %p is queued", m)); 2935 } 2936 } 2937 2938 /* 2939 * vm_page_try_to_free() 2940 * 2941 * Attempt to free the page. If we cannot free it, we do nothing. 2942 * 1 is returned on success, 0 on failure. 2943 */ 2944 int 2945 vm_page_try_to_free(vm_page_t m) 2946 { 2947 2948 vm_page_lock_assert(m, MA_OWNED); 2949 if (m->object != NULL) 2950 VM_OBJECT_ASSERT_WLOCKED(m->object); 2951 if (m->dirty || m->hold_count || m->wire_count || 2952 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m)) 2953 return (0); 2954 pmap_remove_all(m); 2955 if (m->dirty) 2956 return (0); 2957 vm_page_free(m); 2958 return (1); 2959 } 2960 2961 /* 2962 * vm_page_advise 2963 * 2964 * Deactivate or do nothing, as appropriate. 2965 * 2966 * The object and page must be locked. 2967 */ 2968 void 2969 vm_page_advise(vm_page_t m, int advice) 2970 { 2971 2972 vm_page_assert_locked(m); 2973 VM_OBJECT_ASSERT_WLOCKED(m->object); 2974 if (advice == MADV_FREE) 2975 /* 2976 * Mark the page clean. This will allow the page to be freed 2977 * up by the system. However, such pages are often reused 2978 * quickly by malloc() so we do not do anything that would 2979 * cause a page fault if we can help it. 2980 * 2981 * Specifically, we do not try to actually free the page now 2982 * nor do we try to put it in the cache (which would cause a 2983 * page fault on reuse). 2984 * 2985 * But we do make the page as freeable as we can without 2986 * actually taking the step of unmapping it. 2987 */ 2988 vm_page_undirty(m); 2989 else if (advice != MADV_DONTNEED) 2990 return; 2991 2992 /* 2993 * Clear any references to the page. Otherwise, the page daemon will 2994 * immediately reactivate the page. 2995 */ 2996 vm_page_aflag_clear(m, PGA_REFERENCED); 2997 2998 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) 2999 vm_page_dirty(m); 3000 3001 /* 3002 * Place clean pages near the head of the inactive queue rather than 3003 * the tail, thus defeating the queue's LRU operation and ensuring that 3004 * the page will be reused quickly. Dirty pages not already in the 3005 * laundry are moved there. 3006 */ 3007 if (m->dirty == 0) 3008 vm_page_deactivate_noreuse(m); 3009 else 3010 vm_page_launder(m); 3011 } 3012 3013 /* 3014 * Grab a page, waiting until we are waken up due to the page 3015 * changing state. We keep on waiting, if the page continues 3016 * to be in the object. If the page doesn't exist, first allocate it 3017 * and then conditionally zero it. 3018 * 3019 * This routine may sleep. 3020 * 3021 * The object must be locked on entry. The lock will, however, be released 3022 * and reacquired if the routine sleeps. 3023 */ 3024 vm_page_t 3025 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 3026 { 3027 vm_page_t m; 3028 int sleep; 3029 3030 VM_OBJECT_ASSERT_WLOCKED(object); 3031 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || 3032 (allocflags & VM_ALLOC_IGN_SBUSY) != 0, 3033 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); 3034 retrylookup: 3035 if ((m = vm_page_lookup(object, pindex)) != NULL) { 3036 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? 3037 vm_page_xbusied(m) : vm_page_busied(m); 3038 if (sleep) { 3039 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3040 return (NULL); 3041 /* 3042 * Reference the page before unlocking and 3043 * sleeping so that the page daemon is less 3044 * likely to reclaim it. 3045 */ 3046 vm_page_aflag_set(m, PGA_REFERENCED); 3047 vm_page_lock(m); 3048 VM_OBJECT_WUNLOCK(object); 3049 vm_page_busy_sleep(m, "pgrbwt", (allocflags & 3050 VM_ALLOC_IGN_SBUSY) != 0); 3051 VM_OBJECT_WLOCK(object); 3052 goto retrylookup; 3053 } else { 3054 if ((allocflags & VM_ALLOC_WIRED) != 0) { 3055 vm_page_lock(m); 3056 vm_page_wire(m); 3057 vm_page_unlock(m); 3058 } 3059 if ((allocflags & 3060 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) 3061 vm_page_xbusy(m); 3062 if ((allocflags & VM_ALLOC_SBUSY) != 0) 3063 vm_page_sbusy(m); 3064 return (m); 3065 } 3066 } 3067 m = vm_page_alloc(object, pindex, allocflags); 3068 if (m == NULL) { 3069 if ((allocflags & VM_ALLOC_NOWAIT) != 0) 3070 return (NULL); 3071 VM_OBJECT_WUNLOCK(object); 3072 VM_WAIT; 3073 VM_OBJECT_WLOCK(object); 3074 goto retrylookup; 3075 } 3076 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 3077 pmap_zero_page(m); 3078 return (m); 3079 } 3080 3081 /* 3082 * Mapping function for valid or dirty bits in a page. 3083 * 3084 * Inputs are required to range within a page. 3085 */ 3086 vm_page_bits_t 3087 vm_page_bits(int base, int size) 3088 { 3089 int first_bit; 3090 int last_bit; 3091 3092 KASSERT( 3093 base + size <= PAGE_SIZE, 3094 ("vm_page_bits: illegal base/size %d/%d", base, size) 3095 ); 3096 3097 if (size == 0) /* handle degenerate case */ 3098 return (0); 3099 3100 first_bit = base >> DEV_BSHIFT; 3101 last_bit = (base + size - 1) >> DEV_BSHIFT; 3102 3103 return (((vm_page_bits_t)2 << last_bit) - 3104 ((vm_page_bits_t)1 << first_bit)); 3105 } 3106 3107 /* 3108 * vm_page_set_valid_range: 3109 * 3110 * Sets portions of a page valid. The arguments are expected 3111 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3112 * of any partial chunks touched by the range. The invalid portion of 3113 * such chunks will be zeroed. 3114 * 3115 * (base + size) must be less then or equal to PAGE_SIZE. 3116 */ 3117 void 3118 vm_page_set_valid_range(vm_page_t m, int base, int size) 3119 { 3120 int endoff, frag; 3121 3122 VM_OBJECT_ASSERT_WLOCKED(m->object); 3123 if (size == 0) /* handle degenerate case */ 3124 return; 3125 3126 /* 3127 * If the base is not DEV_BSIZE aligned and the valid 3128 * bit is clear, we have to zero out a portion of the 3129 * first block. 3130 */ 3131 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3132 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 3133 pmap_zero_page_area(m, frag, base - frag); 3134 3135 /* 3136 * If the ending offset is not DEV_BSIZE aligned and the 3137 * valid bit is clear, we have to zero out a portion of 3138 * the last block. 3139 */ 3140 endoff = base + size; 3141 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3142 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 3143 pmap_zero_page_area(m, endoff, 3144 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3145 3146 /* 3147 * Assert that no previously invalid block that is now being validated 3148 * is already dirty. 3149 */ 3150 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 3151 ("vm_page_set_valid_range: page %p is dirty", m)); 3152 3153 /* 3154 * Set valid bits inclusive of any overlap. 3155 */ 3156 m->valid |= vm_page_bits(base, size); 3157 } 3158 3159 /* 3160 * Clear the given bits from the specified page's dirty field. 3161 */ 3162 static __inline void 3163 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) 3164 { 3165 uintptr_t addr; 3166 #if PAGE_SIZE < 16384 3167 int shift; 3168 #endif 3169 3170 /* 3171 * If the object is locked and the page is neither exclusive busy nor 3172 * write mapped, then the page's dirty field cannot possibly be 3173 * set by a concurrent pmap operation. 3174 */ 3175 VM_OBJECT_ASSERT_WLOCKED(m->object); 3176 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) 3177 m->dirty &= ~pagebits; 3178 else { 3179 /* 3180 * The pmap layer can call vm_page_dirty() without 3181 * holding a distinguished lock. The combination of 3182 * the object's lock and an atomic operation suffice 3183 * to guarantee consistency of the page dirty field. 3184 * 3185 * For PAGE_SIZE == 32768 case, compiler already 3186 * properly aligns the dirty field, so no forcible 3187 * alignment is needed. Only require existence of 3188 * atomic_clear_64 when page size is 32768. 3189 */ 3190 addr = (uintptr_t)&m->dirty; 3191 #if PAGE_SIZE == 32768 3192 atomic_clear_64((uint64_t *)addr, pagebits); 3193 #elif PAGE_SIZE == 16384 3194 atomic_clear_32((uint32_t *)addr, pagebits); 3195 #else /* PAGE_SIZE <= 8192 */ 3196 /* 3197 * Use a trick to perform a 32-bit atomic on the 3198 * containing aligned word, to not depend on the existence 3199 * of atomic_clear_{8, 16}. 3200 */ 3201 shift = addr & (sizeof(uint32_t) - 1); 3202 #if BYTE_ORDER == BIG_ENDIAN 3203 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; 3204 #else 3205 shift *= NBBY; 3206 #endif 3207 addr &= ~(sizeof(uint32_t) - 1); 3208 atomic_clear_32((uint32_t *)addr, pagebits << shift); 3209 #endif /* PAGE_SIZE */ 3210 } 3211 } 3212 3213 /* 3214 * vm_page_set_validclean: 3215 * 3216 * Sets portions of a page valid and clean. The arguments are expected 3217 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 3218 * of any partial chunks touched by the range. The invalid portion of 3219 * such chunks will be zero'd. 3220 * 3221 * (base + size) must be less then or equal to PAGE_SIZE. 3222 */ 3223 void 3224 vm_page_set_validclean(vm_page_t m, int base, int size) 3225 { 3226 vm_page_bits_t oldvalid, pagebits; 3227 int endoff, frag; 3228 3229 VM_OBJECT_ASSERT_WLOCKED(m->object); 3230 if (size == 0) /* handle degenerate case */ 3231 return; 3232 3233 /* 3234 * If the base is not DEV_BSIZE aligned and the valid 3235 * bit is clear, we have to zero out a portion of the 3236 * first block. 3237 */ 3238 if ((frag = rounddown2(base, DEV_BSIZE)) != base && 3239 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) 3240 pmap_zero_page_area(m, frag, base - frag); 3241 3242 /* 3243 * If the ending offset is not DEV_BSIZE aligned and the 3244 * valid bit is clear, we have to zero out a portion of 3245 * the last block. 3246 */ 3247 endoff = base + size; 3248 if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && 3249 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) 3250 pmap_zero_page_area(m, endoff, 3251 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 3252 3253 /* 3254 * Set valid, clear dirty bits. If validating the entire 3255 * page we can safely clear the pmap modify bit. We also 3256 * use this opportunity to clear the VPO_NOSYNC flag. If a process 3257 * takes a write fault on a MAP_NOSYNC memory area the flag will 3258 * be set again. 3259 * 3260 * We set valid bits inclusive of any overlap, but we can only 3261 * clear dirty bits for DEV_BSIZE chunks that are fully within 3262 * the range. 3263 */ 3264 oldvalid = m->valid; 3265 pagebits = vm_page_bits(base, size); 3266 m->valid |= pagebits; 3267 #if 0 /* NOT YET */ 3268 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 3269 frag = DEV_BSIZE - frag; 3270 base += frag; 3271 size -= frag; 3272 if (size < 0) 3273 size = 0; 3274 } 3275 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 3276 #endif 3277 if (base == 0 && size == PAGE_SIZE) { 3278 /* 3279 * The page can only be modified within the pmap if it is 3280 * mapped, and it can only be mapped if it was previously 3281 * fully valid. 3282 */ 3283 if (oldvalid == VM_PAGE_BITS_ALL) 3284 /* 3285 * Perform the pmap_clear_modify() first. Otherwise, 3286 * a concurrent pmap operation, such as 3287 * pmap_protect(), could clear a modification in the 3288 * pmap and set the dirty field on the page before 3289 * pmap_clear_modify() had begun and after the dirty 3290 * field was cleared here. 3291 */ 3292 pmap_clear_modify(m); 3293 m->dirty = 0; 3294 m->oflags &= ~VPO_NOSYNC; 3295 } else if (oldvalid != VM_PAGE_BITS_ALL) 3296 m->dirty &= ~pagebits; 3297 else 3298 vm_page_clear_dirty_mask(m, pagebits); 3299 } 3300 3301 void 3302 vm_page_clear_dirty(vm_page_t m, int base, int size) 3303 { 3304 3305 vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); 3306 } 3307 3308 /* 3309 * vm_page_set_invalid: 3310 * 3311 * Invalidates DEV_BSIZE'd chunks within a page. Both the 3312 * valid and dirty bits for the effected areas are cleared. 3313 */ 3314 void 3315 vm_page_set_invalid(vm_page_t m, int base, int size) 3316 { 3317 vm_page_bits_t bits; 3318 vm_object_t object; 3319 3320 object = m->object; 3321 VM_OBJECT_ASSERT_WLOCKED(object); 3322 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + 3323 size >= object->un_pager.vnp.vnp_size) 3324 bits = VM_PAGE_BITS_ALL; 3325 else 3326 bits = vm_page_bits(base, size); 3327 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && 3328 bits != 0) 3329 pmap_remove_all(m); 3330 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || 3331 !pmap_page_is_mapped(m), 3332 ("vm_page_set_invalid: page %p is mapped", m)); 3333 m->valid &= ~bits; 3334 m->dirty &= ~bits; 3335 } 3336 3337 /* 3338 * vm_page_zero_invalid() 3339 * 3340 * The kernel assumes that the invalid portions of a page contain 3341 * garbage, but such pages can be mapped into memory by user code. 3342 * When this occurs, we must zero out the non-valid portions of the 3343 * page so user code sees what it expects. 3344 * 3345 * Pages are most often semi-valid when the end of a file is mapped 3346 * into memory and the file's size is not page aligned. 3347 */ 3348 void 3349 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 3350 { 3351 int b; 3352 int i; 3353 3354 VM_OBJECT_ASSERT_WLOCKED(m->object); 3355 /* 3356 * Scan the valid bits looking for invalid sections that 3357 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the 3358 * valid bit may be set ) have already been zeroed by 3359 * vm_page_set_validclean(). 3360 */ 3361 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 3362 if (i == (PAGE_SIZE / DEV_BSIZE) || 3363 (m->valid & ((vm_page_bits_t)1 << i))) { 3364 if (i > b) { 3365 pmap_zero_page_area(m, 3366 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 3367 } 3368 b = i + 1; 3369 } 3370 } 3371 3372 /* 3373 * setvalid is TRUE when we can safely set the zero'd areas 3374 * as being valid. We can do this if there are no cache consistancy 3375 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 3376 */ 3377 if (setvalid) 3378 m->valid = VM_PAGE_BITS_ALL; 3379 } 3380 3381 /* 3382 * vm_page_is_valid: 3383 * 3384 * Is (partial) page valid? Note that the case where size == 0 3385 * will return FALSE in the degenerate case where the page is 3386 * entirely invalid, and TRUE otherwise. 3387 */ 3388 int 3389 vm_page_is_valid(vm_page_t m, int base, int size) 3390 { 3391 vm_page_bits_t bits; 3392 3393 VM_OBJECT_ASSERT_LOCKED(m->object); 3394 bits = vm_page_bits(base, size); 3395 return (m->valid != 0 && (m->valid & bits) == bits); 3396 } 3397 3398 /* 3399 * vm_page_ps_is_valid: 3400 * 3401 * Returns TRUE if the entire (super)page is valid and FALSE otherwise. 3402 */ 3403 boolean_t 3404 vm_page_ps_is_valid(vm_page_t m) 3405 { 3406 int i, npages; 3407 3408 VM_OBJECT_ASSERT_LOCKED(m->object); 3409 npages = atop(pagesizes[m->psind]); 3410 3411 /* 3412 * The physically contiguous pages that make up a superpage, i.e., a 3413 * page with a page size index ("psind") greater than zero, will 3414 * occupy adjacent entries in vm_page_array[]. 3415 */ 3416 for (i = 0; i < npages; i++) { 3417 if (m[i].valid != VM_PAGE_BITS_ALL) 3418 return (FALSE); 3419 } 3420 return (TRUE); 3421 } 3422 3423 /* 3424 * Set the page's dirty bits if the page is modified. 3425 */ 3426 void 3427 vm_page_test_dirty(vm_page_t m) 3428 { 3429 3430 VM_OBJECT_ASSERT_WLOCKED(m->object); 3431 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) 3432 vm_page_dirty(m); 3433 } 3434 3435 void 3436 vm_page_lock_KBI(vm_page_t m, const char *file, int line) 3437 { 3438 3439 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); 3440 } 3441 3442 void 3443 vm_page_unlock_KBI(vm_page_t m, const char *file, int line) 3444 { 3445 3446 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); 3447 } 3448 3449 int 3450 vm_page_trylock_KBI(vm_page_t m, const char *file, int line) 3451 { 3452 3453 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); 3454 } 3455 3456 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) 3457 void 3458 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) 3459 { 3460 3461 vm_page_lock_assert_KBI(m, MA_OWNED, file, line); 3462 } 3463 3464 void 3465 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) 3466 { 3467 3468 mtx_assert_(vm_page_lockptr(m), a, file, line); 3469 } 3470 #endif 3471 3472 #ifdef INVARIANTS 3473 void 3474 vm_page_object_lock_assert(vm_page_t m) 3475 { 3476 3477 /* 3478 * Certain of the page's fields may only be modified by the 3479 * holder of the containing object's lock or the exclusive busy. 3480 * holder. Unfortunately, the holder of the write busy is 3481 * not recorded, and thus cannot be checked here. 3482 */ 3483 if (m->object != NULL && !vm_page_xbusied(m)) 3484 VM_OBJECT_ASSERT_WLOCKED(m->object); 3485 } 3486 3487 void 3488 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) 3489 { 3490 3491 if ((bits & PGA_WRITEABLE) == 0) 3492 return; 3493 3494 /* 3495 * The PGA_WRITEABLE flag can only be set if the page is 3496 * managed, is exclusively busied or the object is locked. 3497 * Currently, this flag is only set by pmap_enter(). 3498 */ 3499 KASSERT((m->oflags & VPO_UNMANAGED) == 0, 3500 ("PGA_WRITEABLE on unmanaged page")); 3501 if (!vm_page_xbusied(m)) 3502 VM_OBJECT_ASSERT_LOCKED(m->object); 3503 } 3504 #endif 3505 3506 #include "opt_ddb.h" 3507 #ifdef DDB 3508 #include <sys/kernel.h> 3509 3510 #include <ddb/ddb.h> 3511 3512 DB_SHOW_COMMAND(page, vm_page_print_page_info) 3513 { 3514 3515 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count); 3516 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count); 3517 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count); 3518 db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count); 3519 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count); 3520 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); 3521 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); 3522 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); 3523 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); 3524 } 3525 3526 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 3527 { 3528 int dom; 3529 3530 db_printf("pq_free %d\n", vm_cnt.v_free_count); 3531 for (dom = 0; dom < vm_ndomains; dom++) { 3532 db_printf( 3533 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d\n", 3534 dom, 3535 vm_dom[dom].vmd_page_count, 3536 vm_dom[dom].vmd_free_count, 3537 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, 3538 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, 3539 vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt); 3540 } 3541 } 3542 3543 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) 3544 { 3545 vm_page_t m; 3546 boolean_t phys; 3547 3548 if (!have_addr) { 3549 db_printf("show pginfo addr\n"); 3550 return; 3551 } 3552 3553 phys = strchr(modif, 'p') != NULL; 3554 if (phys) 3555 m = PHYS_TO_VM_PAGE(addr); 3556 else 3557 m = (vm_page_t)addr; 3558 db_printf( 3559 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" 3560 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", 3561 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, 3562 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, 3563 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); 3564 } 3565 #endif /* DDB */ 3566