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 pageq mutex is required when adding or removing a page from a 67 * page queue (vm_page_queue[]), regardless of other mutexes or the 68 * busy state of a page. 69 * 70 * - a hash chain mutex is required when associating or disassociating 71 * a page from the VM PAGE CACHE hash table (vm_page_buckets), 72 * regardless of other mutexes or the busy state of a page. 73 * 74 * - either a hash chain mutex OR a busied page is required in order 75 * to modify the page flags. A hash chain mutex must be obtained in 76 * order to busy a page. A page's flags cannot be modified by a 77 * hash chain mutex if the page is marked busy. 78 * 79 * - The object memq mutex is held when inserting or removing 80 * pages from an object (vm_page_insert() or vm_page_remove()). This 81 * is different from the object's main mutex. 82 * 83 * Generally speaking, you have to be aware of side effects when running 84 * vm_page ops. A vm_page_lookup() will return with the hash chain 85 * locked, whether it was able to lookup the page or not. vm_page_free(), 86 * vm_page_cache(), vm_page_activate(), and a number of other routines 87 * will release the hash chain mutex for you. Intermediate manipulation 88 * routines such as vm_page_flag_set() expect the hash chain to be held 89 * on entry and the hash chain will remain held on return. 90 * 91 * pageq scanning can only occur with the pageq in question locked. 92 * We have a known bottleneck with the active queue, but the cache 93 * and free queues are actually arrays already. 94 */ 95 96 /* 97 * Resident memory management module. 98 */ 99 100 #include <sys/cdefs.h> 101 __FBSDID("$FreeBSD$"); 102 103 #include "opt_vm.h" 104 105 #include <sys/param.h> 106 #include <sys/systm.h> 107 #include <sys/lock.h> 108 #include <sys/kernel.h> 109 #include <sys/limits.h> 110 #include <sys/malloc.h> 111 #include <sys/mutex.h> 112 #include <sys/proc.h> 113 #include <sys/sysctl.h> 114 #include <sys/vmmeter.h> 115 #include <sys/vnode.h> 116 117 #include <vm/vm.h> 118 #include <vm/vm_param.h> 119 #include <vm/vm_kern.h> 120 #include <vm/vm_object.h> 121 #include <vm/vm_page.h> 122 #include <vm/vm_pageout.h> 123 #include <vm/vm_pager.h> 124 #include <vm/vm_phys.h> 125 #include <vm/vm_reserv.h> 126 #include <vm/vm_extern.h> 127 #include <vm/uma.h> 128 #include <vm/uma_int.h> 129 130 #include <machine/md_var.h> 131 132 /* 133 * Associated with page of user-allocatable memory is a 134 * page structure. 135 */ 136 137 struct vpgqueues vm_page_queues[PQ_COUNT]; 138 struct vpglocks vm_page_queue_lock; 139 struct vpglocks vm_page_queue_free_lock; 140 141 vm_page_t vm_page_array = 0; 142 int vm_page_array_size = 0; 143 long first_page = 0; 144 int vm_page_zero_count = 0; 145 146 static int boot_pages = UMA_BOOT_PAGES; 147 TUNABLE_INT("vm.boot_pages", &boot_pages); 148 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0, 149 "number of pages allocated for bootstrapping the VM system"); 150 151 static void vm_page_enqueue(int queue, vm_page_t m); 152 153 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ 154 #if PAGE_SIZE == 32768 155 #ifdef CTASSERT 156 CTASSERT(sizeof(u_long) >= 8); 157 #endif 158 #endif 159 160 /* 161 * vm_set_page_size: 162 * 163 * Sets the page size, perhaps based upon the memory 164 * size. Must be called before any use of page-size 165 * dependent functions. 166 */ 167 void 168 vm_set_page_size(void) 169 { 170 if (cnt.v_page_size == 0) 171 cnt.v_page_size = PAGE_SIZE; 172 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 173 panic("vm_set_page_size: page size not a power of two"); 174 } 175 176 /* 177 * vm_page_blacklist_lookup: 178 * 179 * See if a physical address in this page has been listed 180 * in the blacklist tunable. Entries in the tunable are 181 * separated by spaces or commas. If an invalid integer is 182 * encountered then the rest of the string is skipped. 183 */ 184 static int 185 vm_page_blacklist_lookup(char *list, vm_paddr_t pa) 186 { 187 vm_paddr_t bad; 188 char *cp, *pos; 189 190 for (pos = list; *pos != '\0'; pos = cp) { 191 bad = strtoq(pos, &cp, 0); 192 if (*cp != '\0') { 193 if (*cp == ' ' || *cp == ',') { 194 cp++; 195 if (cp == pos) 196 continue; 197 } else 198 break; 199 } 200 if (pa == trunc_page(bad)) 201 return (1); 202 } 203 return (0); 204 } 205 206 /* 207 * vm_page_startup: 208 * 209 * Initializes the resident memory module. 210 * 211 * Allocates memory for the page cells, and 212 * for the object/offset-to-page hash table headers. 213 * Each page cell is initialized and placed on the free list. 214 */ 215 vm_offset_t 216 vm_page_startup(vm_offset_t vaddr) 217 { 218 vm_offset_t mapped; 219 vm_paddr_t page_range; 220 vm_paddr_t new_end; 221 int i; 222 vm_paddr_t pa; 223 int nblocks; 224 vm_paddr_t last_pa; 225 char *list; 226 227 /* the biggest memory array is the second group of pages */ 228 vm_paddr_t end; 229 vm_paddr_t biggestsize; 230 vm_paddr_t low_water, high_water; 231 int biggestone; 232 233 biggestsize = 0; 234 biggestone = 0; 235 nblocks = 0; 236 vaddr = round_page(vaddr); 237 238 for (i = 0; phys_avail[i + 1]; i += 2) { 239 phys_avail[i] = round_page(phys_avail[i]); 240 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 241 } 242 243 low_water = phys_avail[0]; 244 high_water = phys_avail[1]; 245 246 for (i = 0; phys_avail[i + 1]; i += 2) { 247 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 248 249 if (size > biggestsize) { 250 biggestone = i; 251 biggestsize = size; 252 } 253 if (phys_avail[i] < low_water) 254 low_water = phys_avail[i]; 255 if (phys_avail[i + 1] > high_water) 256 high_water = phys_avail[i + 1]; 257 ++nblocks; 258 } 259 260 #ifdef XEN 261 low_water = 0; 262 #endif 263 264 end = phys_avail[biggestone+1]; 265 266 /* 267 * Initialize the locks. 268 */ 269 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF | 270 MTX_RECURSE); 271 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 272 MTX_DEF); 273 274 /* 275 * Initialize the queue headers for the hold queue, the active queue, 276 * and the inactive queue. 277 */ 278 for (i = 0; i < PQ_COUNT; i++) 279 TAILQ_INIT(&vm_page_queues[i].pl); 280 vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count; 281 vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count; 282 vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count; 283 284 /* 285 * Allocate memory for use when boot strapping the kernel memory 286 * allocator. 287 */ 288 new_end = end - (boot_pages * UMA_SLAB_SIZE); 289 new_end = trunc_page(new_end); 290 mapped = pmap_map(&vaddr, new_end, end, 291 VM_PROT_READ | VM_PROT_WRITE); 292 bzero((void *)mapped, end - new_end); 293 uma_startup((void *)mapped, boot_pages); 294 295 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) 296 /* 297 * Allocate a bitmap to indicate that a random physical page 298 * needs to be included in a minidump. 299 * 300 * The amd64 port needs this to indicate which direct map pages 301 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 302 * 303 * However, i386 still needs this workspace internally within the 304 * minidump code. In theory, they are not needed on i386, but are 305 * included should the sf_buf code decide to use them. 306 */ 307 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 308 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 309 new_end -= vm_page_dump_size; 310 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 311 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 312 bzero((void *)vm_page_dump, vm_page_dump_size); 313 #endif 314 /* 315 * Compute the number of pages of memory that will be available for 316 * use (taking into account the overhead of a page structure per 317 * page). 318 */ 319 first_page = low_water / PAGE_SIZE; 320 #ifdef VM_PHYSSEG_SPARSE 321 page_range = 0; 322 for (i = 0; phys_avail[i + 1] != 0; i += 2) 323 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 324 #elif defined(VM_PHYSSEG_DENSE) 325 page_range = high_water / PAGE_SIZE - first_page; 326 #else 327 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." 328 #endif 329 end = new_end; 330 331 /* 332 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 333 */ 334 vaddr += PAGE_SIZE; 335 336 /* 337 * Initialize the mem entry structures now, and put them in the free 338 * queue. 339 */ 340 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 341 mapped = pmap_map(&vaddr, new_end, end, 342 VM_PROT_READ | VM_PROT_WRITE); 343 vm_page_array = (vm_page_t) mapped; 344 #if VM_NRESERVLEVEL > 0 345 /* 346 * Allocate memory for the reservation management system's data 347 * structures. 348 */ 349 new_end = vm_reserv_startup(&vaddr, new_end, high_water); 350 #endif 351 #ifdef __amd64__ 352 /* 353 * pmap_map on amd64 comes out of the direct-map, not kvm like i386, 354 * so the pages must be tracked for a crashdump to include this data. 355 * This includes the vm_page_array and the early UMA bootstrap pages. 356 */ 357 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 358 dump_add_page(pa); 359 #endif 360 phys_avail[biggestone + 1] = new_end; 361 362 /* 363 * Clear all of the page structures 364 */ 365 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 366 for (i = 0; i < page_range; i++) 367 vm_page_array[i].order = VM_NFREEORDER; 368 vm_page_array_size = page_range; 369 370 /* 371 * Initialize the physical memory allocator. 372 */ 373 vm_phys_init(); 374 375 /* 376 * Add every available physical page that is not blacklisted to 377 * the free lists. 378 */ 379 cnt.v_page_count = 0; 380 cnt.v_free_count = 0; 381 list = getenv("vm.blacklist"); 382 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 383 pa = phys_avail[i]; 384 last_pa = phys_avail[i + 1]; 385 while (pa < last_pa) { 386 if (list != NULL && 387 vm_page_blacklist_lookup(list, pa)) 388 printf("Skipping page with pa 0x%jx\n", 389 (uintmax_t)pa); 390 else 391 vm_phys_add_page(pa); 392 pa += PAGE_SIZE; 393 } 394 } 395 freeenv(list); 396 #if VM_NRESERVLEVEL > 0 397 /* 398 * Initialize the reservation management system. 399 */ 400 vm_reserv_init(); 401 #endif 402 return (vaddr); 403 } 404 405 void 406 vm_page_flag_set(vm_page_t m, unsigned short bits) 407 { 408 409 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 410 m->flags |= bits; 411 } 412 413 void 414 vm_page_flag_clear(vm_page_t m, unsigned short bits) 415 { 416 417 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 418 m->flags &= ~bits; 419 } 420 421 void 422 vm_page_busy(vm_page_t m) 423 { 424 425 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 426 KASSERT((m->oflags & VPO_BUSY) == 0, 427 ("vm_page_busy: page already busy!!!")); 428 m->oflags |= VPO_BUSY; 429 } 430 431 /* 432 * vm_page_flash: 433 * 434 * wakeup anyone waiting for the page. 435 */ 436 void 437 vm_page_flash(vm_page_t m) 438 { 439 440 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 441 if (m->oflags & VPO_WANTED) { 442 m->oflags &= ~VPO_WANTED; 443 wakeup(m); 444 } 445 } 446 447 /* 448 * vm_page_wakeup: 449 * 450 * clear the VPO_BUSY flag and wakeup anyone waiting for the 451 * page. 452 * 453 */ 454 void 455 vm_page_wakeup(vm_page_t m) 456 { 457 458 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 459 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!")); 460 m->oflags &= ~VPO_BUSY; 461 vm_page_flash(m); 462 } 463 464 void 465 vm_page_io_start(vm_page_t m) 466 { 467 468 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 469 m->busy++; 470 } 471 472 void 473 vm_page_io_finish(vm_page_t m) 474 { 475 476 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 477 m->busy--; 478 if (m->busy == 0) 479 vm_page_flash(m); 480 } 481 482 /* 483 * Keep page from being freed by the page daemon 484 * much of the same effect as wiring, except much lower 485 * overhead and should be used only for *very* temporary 486 * holding ("wiring"). 487 */ 488 void 489 vm_page_hold(vm_page_t mem) 490 { 491 492 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 493 mem->hold_count++; 494 } 495 496 void 497 vm_page_unhold(vm_page_t mem) 498 { 499 500 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 501 --mem->hold_count; 502 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 503 if (mem->hold_count == 0 && VM_PAGE_INQUEUE2(mem, PQ_HOLD)) 504 vm_page_free_toq(mem); 505 } 506 507 /* 508 * vm_page_free: 509 * 510 * Free a page. 511 */ 512 void 513 vm_page_free(vm_page_t m) 514 { 515 516 m->flags &= ~PG_ZERO; 517 vm_page_free_toq(m); 518 } 519 520 /* 521 * vm_page_free_zero: 522 * 523 * Free a page to the zerod-pages queue 524 */ 525 void 526 vm_page_free_zero(vm_page_t m) 527 { 528 529 m->flags |= PG_ZERO; 530 vm_page_free_toq(m); 531 } 532 533 /* 534 * vm_page_sleep: 535 * 536 * Sleep and release the page queues lock. 537 * 538 * The object containing the given page must be locked. 539 */ 540 void 541 vm_page_sleep(vm_page_t m, const char *msg) 542 { 543 544 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 545 if (!mtx_owned(&vm_page_queue_mtx)) 546 vm_page_lock_queues(); 547 vm_page_flag_set(m, PG_REFERENCED); 548 vm_page_unlock_queues(); 549 550 /* 551 * It's possible that while we sleep, the page will get 552 * unbusied and freed. If we are holding the object 553 * lock, we will assume we hold a reference to the object 554 * such that even if m->object changes, we can re-lock 555 * it. 556 */ 557 m->oflags |= VPO_WANTED; 558 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0); 559 } 560 561 /* 562 * vm_page_dirty: 563 * 564 * make page all dirty 565 */ 566 void 567 vm_page_dirty(vm_page_t m) 568 { 569 570 KASSERT((m->flags & PG_CACHED) == 0, 571 ("vm_page_dirty: page in cache!")); 572 KASSERT(!VM_PAGE_IS_FREE(m), 573 ("vm_page_dirty: page is free!")); 574 KASSERT(m->valid == VM_PAGE_BITS_ALL, 575 ("vm_page_dirty: page is invalid!")); 576 m->dirty = VM_PAGE_BITS_ALL; 577 } 578 579 /* 580 * vm_page_splay: 581 * 582 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 583 * the vm_page containing the given pindex. If, however, that 584 * pindex is not found in the vm_object, returns a vm_page that is 585 * adjacent to the pindex, coming before or after it. 586 */ 587 vm_page_t 588 vm_page_splay(vm_pindex_t pindex, vm_page_t root) 589 { 590 struct vm_page dummy; 591 vm_page_t lefttreemax, righttreemin, y; 592 593 if (root == NULL) 594 return (root); 595 lefttreemax = righttreemin = &dummy; 596 for (;; root = y) { 597 if (pindex < root->pindex) { 598 if ((y = root->left) == NULL) 599 break; 600 if (pindex < y->pindex) { 601 /* Rotate right. */ 602 root->left = y->right; 603 y->right = root; 604 root = y; 605 if ((y = root->left) == NULL) 606 break; 607 } 608 /* Link into the new root's right tree. */ 609 righttreemin->left = root; 610 righttreemin = root; 611 } else if (pindex > root->pindex) { 612 if ((y = root->right) == NULL) 613 break; 614 if (pindex > y->pindex) { 615 /* Rotate left. */ 616 root->right = y->left; 617 y->left = root; 618 root = y; 619 if ((y = root->right) == NULL) 620 break; 621 } 622 /* Link into the new root's left tree. */ 623 lefttreemax->right = root; 624 lefttreemax = root; 625 } else 626 break; 627 } 628 /* Assemble the new root. */ 629 lefttreemax->right = root->left; 630 righttreemin->left = root->right; 631 root->left = dummy.right; 632 root->right = dummy.left; 633 return (root); 634 } 635 636 /* 637 * vm_page_insert: [ internal use only ] 638 * 639 * Inserts the given mem entry into the object and object list. 640 * 641 * The pagetables are not updated but will presumably fault the page 642 * in if necessary, or if a kernel page the caller will at some point 643 * enter the page into the kernel's pmap. We are not allowed to block 644 * here so we *can't* do this anyway. 645 * 646 * The object and page must be locked. 647 * This routine may not block. 648 */ 649 void 650 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 651 { 652 vm_page_t root; 653 654 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 655 if (m->object != NULL) 656 panic("vm_page_insert: page already inserted"); 657 658 /* 659 * Record the object/offset pair in this page 660 */ 661 m->object = object; 662 m->pindex = pindex; 663 664 /* 665 * Now link into the object's ordered list of backed pages. 666 */ 667 root = object->root; 668 if (root == NULL) { 669 m->left = NULL; 670 m->right = NULL; 671 TAILQ_INSERT_TAIL(&object->memq, m, listq); 672 } else { 673 root = vm_page_splay(pindex, root); 674 if (pindex < root->pindex) { 675 m->left = root->left; 676 m->right = root; 677 root->left = NULL; 678 TAILQ_INSERT_BEFORE(root, m, listq); 679 } else if (pindex == root->pindex) 680 panic("vm_page_insert: offset already allocated"); 681 else { 682 m->right = root->right; 683 m->left = root; 684 root->right = NULL; 685 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 686 } 687 } 688 object->root = m; 689 object->generation++; 690 691 /* 692 * show that the object has one more resident page. 693 */ 694 object->resident_page_count++; 695 /* 696 * Hold the vnode until the last page is released. 697 */ 698 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 699 vhold((struct vnode *)object->handle); 700 701 /* 702 * Since we are inserting a new and possibly dirty page, 703 * update the object's OBJ_MIGHTBEDIRTY flag. 704 */ 705 if (m->flags & PG_WRITEABLE) 706 vm_object_set_writeable_dirty(object); 707 } 708 709 /* 710 * vm_page_remove: 711 * NOTE: used by device pager as well -wfj 712 * 713 * Removes the given mem entry from the object/offset-page 714 * table and the object page list, but do not invalidate/terminate 715 * the backing store. 716 * 717 * The object and page must be locked. 718 * The underlying pmap entry (if any) is NOT removed here. 719 * This routine may not block. 720 */ 721 void 722 vm_page_remove(vm_page_t m) 723 { 724 vm_object_t object; 725 vm_page_t root; 726 727 if ((object = m->object) == NULL) 728 return; 729 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 730 if (m->oflags & VPO_BUSY) { 731 m->oflags &= ~VPO_BUSY; 732 vm_page_flash(m); 733 } 734 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 735 736 /* 737 * Now remove from the object's list of backed pages. 738 */ 739 if (m != object->root) 740 vm_page_splay(m->pindex, object->root); 741 if (m->left == NULL) 742 root = m->right; 743 else { 744 root = vm_page_splay(m->pindex, m->left); 745 root->right = m->right; 746 } 747 object->root = root; 748 TAILQ_REMOVE(&object->memq, m, listq); 749 750 /* 751 * And show that the object has one fewer resident page. 752 */ 753 object->resident_page_count--; 754 object->generation++; 755 /* 756 * The vnode may now be recycled. 757 */ 758 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 759 vdrop((struct vnode *)object->handle); 760 761 m->object = NULL; 762 } 763 764 /* 765 * vm_page_lookup: 766 * 767 * Returns the page associated with the object/offset 768 * pair specified; if none is found, NULL is returned. 769 * 770 * The object must be locked. 771 * This routine may not block. 772 * This is a critical path routine 773 */ 774 vm_page_t 775 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 776 { 777 vm_page_t m; 778 779 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 780 if ((m = object->root) != NULL && m->pindex != pindex) { 781 m = vm_page_splay(pindex, m); 782 if ((object->root = m)->pindex != pindex) 783 m = NULL; 784 } 785 return (m); 786 } 787 788 /* 789 * vm_page_rename: 790 * 791 * Move the given memory entry from its 792 * current object to the specified target object/offset. 793 * 794 * The object must be locked. 795 * This routine may not block. 796 * 797 * Note: swap associated with the page must be invalidated by the move. We 798 * have to do this for several reasons: (1) we aren't freeing the 799 * page, (2) we are dirtying the page, (3) the VM system is probably 800 * moving the page from object A to B, and will then later move 801 * the backing store from A to B and we can't have a conflict. 802 * 803 * Note: we *always* dirty the page. It is necessary both for the 804 * fact that we moved it, and because we may be invalidating 805 * swap. If the page is on the cache, we have to deactivate it 806 * or vm_page_dirty() will panic. Dirty pages are not allowed 807 * on the cache. 808 */ 809 void 810 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 811 { 812 813 vm_page_remove(m); 814 vm_page_insert(m, new_object, new_pindex); 815 vm_page_dirty(m); 816 } 817 818 /* 819 * Convert all of the given object's cached pages that have a 820 * pindex within the given range into free pages. If the value 821 * zero is given for "end", then the range's upper bound is 822 * infinity. If the given object is backed by a vnode and it 823 * transitions from having one or more cached pages to none, the 824 * vnode's hold count is reduced. 825 */ 826 void 827 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end) 828 { 829 vm_page_t m, m_next; 830 boolean_t empty; 831 832 mtx_lock(&vm_page_queue_free_mtx); 833 if (__predict_false(object->cache == NULL)) { 834 mtx_unlock(&vm_page_queue_free_mtx); 835 return; 836 } 837 m = object->cache = vm_page_splay(start, object->cache); 838 if (m->pindex < start) { 839 if (m->right == NULL) 840 m = NULL; 841 else { 842 m_next = vm_page_splay(start, m->right); 843 m_next->left = m; 844 m->right = NULL; 845 m = object->cache = m_next; 846 } 847 } 848 849 /* 850 * At this point, "m" is either (1) a reference to the page 851 * with the least pindex that is greater than or equal to 852 * "start" or (2) NULL. 853 */ 854 for (; m != NULL && (m->pindex < end || end == 0); m = m_next) { 855 /* 856 * Find "m"'s successor and remove "m" from the 857 * object's cache. 858 */ 859 if (m->right == NULL) { 860 object->cache = m->left; 861 m_next = NULL; 862 } else { 863 m_next = vm_page_splay(start, m->right); 864 m_next->left = m->left; 865 object->cache = m_next; 866 } 867 /* Convert "m" to a free page. */ 868 m->object = NULL; 869 m->valid = 0; 870 /* Clear PG_CACHED and set PG_FREE. */ 871 m->flags ^= PG_CACHED | PG_FREE; 872 KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE, 873 ("vm_page_cache_free: page %p has inconsistent flags", m)); 874 cnt.v_cache_count--; 875 cnt.v_free_count++; 876 } 877 empty = object->cache == NULL; 878 mtx_unlock(&vm_page_queue_free_mtx); 879 if (object->type == OBJT_VNODE && empty) 880 vdrop(object->handle); 881 } 882 883 /* 884 * Returns the cached page that is associated with the given 885 * object and offset. If, however, none exists, returns NULL. 886 * 887 * The free page queue must be locked. 888 */ 889 static inline vm_page_t 890 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex) 891 { 892 vm_page_t m; 893 894 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 895 if ((m = object->cache) != NULL && m->pindex != pindex) { 896 m = vm_page_splay(pindex, m); 897 if ((object->cache = m)->pindex != pindex) 898 m = NULL; 899 } 900 return (m); 901 } 902 903 /* 904 * Remove the given cached page from its containing object's 905 * collection of cached pages. 906 * 907 * The free page queue must be locked. 908 */ 909 void 910 vm_page_cache_remove(vm_page_t m) 911 { 912 vm_object_t object; 913 vm_page_t root; 914 915 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 916 KASSERT((m->flags & PG_CACHED) != 0, 917 ("vm_page_cache_remove: page %p is not cached", m)); 918 object = m->object; 919 if (m != object->cache) { 920 root = vm_page_splay(m->pindex, object->cache); 921 KASSERT(root == m, 922 ("vm_page_cache_remove: page %p is not cached in object %p", 923 m, object)); 924 } 925 if (m->left == NULL) 926 root = m->right; 927 else if (m->right == NULL) 928 root = m->left; 929 else { 930 root = vm_page_splay(m->pindex, m->left); 931 root->right = m->right; 932 } 933 object->cache = root; 934 m->object = NULL; 935 cnt.v_cache_count--; 936 } 937 938 /* 939 * Transfer all of the cached pages with offset greater than or 940 * equal to 'offidxstart' from the original object's cache to the 941 * new object's cache. However, any cached pages with offset 942 * greater than or equal to the new object's size are kept in the 943 * original object. Initially, the new object's cache must be 944 * empty. Offset 'offidxstart' in the original object must 945 * correspond to offset zero in the new object. 946 * 947 * The new object must be locked. 948 */ 949 void 950 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart, 951 vm_object_t new_object) 952 { 953 vm_page_t m, m_next; 954 955 /* 956 * Insertion into an object's collection of cached pages 957 * requires the object to be locked. In contrast, removal does 958 * not. 959 */ 960 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED); 961 KASSERT(new_object->cache == NULL, 962 ("vm_page_cache_transfer: object %p has cached pages", 963 new_object)); 964 mtx_lock(&vm_page_queue_free_mtx); 965 if ((m = orig_object->cache) != NULL) { 966 /* 967 * Transfer all of the pages with offset greater than or 968 * equal to 'offidxstart' from the original object's 969 * cache to the new object's cache. 970 */ 971 m = vm_page_splay(offidxstart, m); 972 if (m->pindex < offidxstart) { 973 orig_object->cache = m; 974 new_object->cache = m->right; 975 m->right = NULL; 976 } else { 977 orig_object->cache = m->left; 978 new_object->cache = m; 979 m->left = NULL; 980 } 981 while ((m = new_object->cache) != NULL) { 982 if ((m->pindex - offidxstart) >= new_object->size) { 983 /* 984 * Return all of the cached pages with 985 * offset greater than or equal to the 986 * new object's size to the original 987 * object's cache. 988 */ 989 new_object->cache = m->left; 990 m->left = orig_object->cache; 991 orig_object->cache = m; 992 break; 993 } 994 m_next = vm_page_splay(m->pindex, m->right); 995 /* Update the page's object and offset. */ 996 m->object = new_object; 997 m->pindex -= offidxstart; 998 if (m_next == NULL) 999 break; 1000 m->right = NULL; 1001 m_next->left = m; 1002 new_object->cache = m_next; 1003 } 1004 KASSERT(new_object->cache == NULL || 1005 new_object->type == OBJT_SWAP, 1006 ("vm_page_cache_transfer: object %p's type is incompatible" 1007 " with cached pages", new_object)); 1008 } 1009 mtx_unlock(&vm_page_queue_free_mtx); 1010 } 1011 1012 /* 1013 * vm_page_alloc: 1014 * 1015 * Allocate and return a memory cell associated 1016 * with this VM object/offset pair. 1017 * 1018 * page_req classes: 1019 * VM_ALLOC_NORMAL normal process request 1020 * VM_ALLOC_SYSTEM system *really* needs a page 1021 * VM_ALLOC_INTERRUPT interrupt time request 1022 * VM_ALLOC_ZERO zero page 1023 * VM_ALLOC_WIRED wire the allocated page 1024 * VM_ALLOC_NOOBJ page is not associated with a vm object 1025 * VM_ALLOC_NOBUSY do not set the page busy 1026 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page 1027 * is cached 1028 * 1029 * This routine may not sleep. 1030 */ 1031 vm_page_t 1032 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 1033 { 1034 struct vnode *vp = NULL; 1035 vm_object_t m_object; 1036 vm_page_t m; 1037 int flags, page_req; 1038 1039 page_req = req & VM_ALLOC_CLASS_MASK; 1040 KASSERT(curthread->td_intr_nesting_level == 0 || 1041 page_req == VM_ALLOC_INTERRUPT, 1042 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 1043 1044 if ((req & VM_ALLOC_NOOBJ) == 0) { 1045 KASSERT(object != NULL, 1046 ("vm_page_alloc: NULL object.")); 1047 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1048 } 1049 1050 /* 1051 * The pager is allowed to eat deeper into the free page list. 1052 */ 1053 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 1054 page_req = VM_ALLOC_SYSTEM; 1055 }; 1056 1057 mtx_lock(&vm_page_queue_free_mtx); 1058 if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved || 1059 (page_req == VM_ALLOC_SYSTEM && 1060 cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) || 1061 (page_req == VM_ALLOC_INTERRUPT && 1062 cnt.v_free_count + cnt.v_cache_count > 0)) { 1063 /* 1064 * Allocate from the free queue if the number of free pages 1065 * exceeds the minimum for the request class. 1066 */ 1067 if (object != NULL && 1068 (m = vm_page_cache_lookup(object, pindex)) != NULL) { 1069 if ((req & VM_ALLOC_IFNOTCACHED) != 0) { 1070 mtx_unlock(&vm_page_queue_free_mtx); 1071 return (NULL); 1072 } 1073 if (vm_phys_unfree_page(m)) 1074 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); 1075 #if VM_NRESERVLEVEL > 0 1076 else if (!vm_reserv_reactivate_page(m)) 1077 #else 1078 else 1079 #endif 1080 panic("vm_page_alloc: cache page %p is missing" 1081 " from the free queue", m); 1082 } else if ((req & VM_ALLOC_IFCACHED) != 0) { 1083 mtx_unlock(&vm_page_queue_free_mtx); 1084 return (NULL); 1085 #if VM_NRESERVLEVEL > 0 1086 } else if (object == NULL || object->type == OBJT_DEVICE || 1087 (object->flags & OBJ_COLORED) == 0 || 1088 (m = vm_reserv_alloc_page(object, pindex)) == NULL) { 1089 #else 1090 } else { 1091 #endif 1092 m = vm_phys_alloc_pages(object != NULL ? 1093 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); 1094 #if VM_NRESERVLEVEL > 0 1095 if (m == NULL && vm_reserv_reclaim_inactive()) { 1096 m = vm_phys_alloc_pages(object != NULL ? 1097 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 1098 0); 1099 } 1100 #endif 1101 } 1102 } else { 1103 /* 1104 * Not allocatable, give up. 1105 */ 1106 mtx_unlock(&vm_page_queue_free_mtx); 1107 atomic_add_int(&vm_pageout_deficit, 1); 1108 pagedaemon_wakeup(); 1109 return (NULL); 1110 } 1111 1112 /* 1113 * At this point we had better have found a good page. 1114 */ 1115 1116 KASSERT(m != NULL, ("vm_page_alloc: missing page")); 1117 KASSERT(m->queue == PQ_NONE, 1118 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue)); 1119 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m)); 1120 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m)); 1121 KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m)); 1122 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m)); 1123 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, 1124 ("vm_page_alloc: page %p has unexpected memattr %d", m, 1125 pmap_page_get_memattr(m))); 1126 if ((m->flags & PG_CACHED) != 0) { 1127 KASSERT(m->valid != 0, 1128 ("vm_page_alloc: cached page %p is invalid", m)); 1129 if (m->object == object && m->pindex == pindex) 1130 cnt.v_reactivated++; 1131 else 1132 m->valid = 0; 1133 m_object = m->object; 1134 vm_page_cache_remove(m); 1135 if (m_object->type == OBJT_VNODE && m_object->cache == NULL) 1136 vp = m_object->handle; 1137 } else { 1138 KASSERT(VM_PAGE_IS_FREE(m), 1139 ("vm_page_alloc: page %p is not free", m)); 1140 KASSERT(m->valid == 0, 1141 ("vm_page_alloc: free page %p is valid", m)); 1142 cnt.v_free_count--; 1143 } 1144 1145 /* 1146 * Initialize structure. Only the PG_ZERO flag is inherited. 1147 */ 1148 flags = 0; 1149 if (m->flags & PG_ZERO) { 1150 vm_page_zero_count--; 1151 if (req & VM_ALLOC_ZERO) 1152 flags = PG_ZERO; 1153 } 1154 if (object == NULL || object->type == OBJT_PHYS) 1155 flags |= PG_UNMANAGED; 1156 m->flags = flags; 1157 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 1158 m->oflags = 0; 1159 else 1160 m->oflags = VPO_BUSY; 1161 if (req & VM_ALLOC_WIRED) { 1162 atomic_add_int(&cnt.v_wire_count, 1); 1163 m->wire_count = 1; 1164 } 1165 m->act_count = 0; 1166 mtx_unlock(&vm_page_queue_free_mtx); 1167 1168 if (object != NULL) { 1169 /* Ignore device objects; the pager sets "memattr" for them. */ 1170 if (object->memattr != VM_MEMATTR_DEFAULT && 1171 object->type != OBJT_DEVICE && object->type != OBJT_SG) 1172 pmap_page_set_memattr(m, object->memattr); 1173 vm_page_insert(m, object, pindex); 1174 } else 1175 m->pindex = pindex; 1176 1177 /* 1178 * The following call to vdrop() must come after the above call 1179 * to vm_page_insert() in case both affect the same object and 1180 * vnode. Otherwise, the affected vnode's hold count could 1181 * temporarily become zero. 1182 */ 1183 if (vp != NULL) 1184 vdrop(vp); 1185 1186 /* 1187 * Don't wakeup too often - wakeup the pageout daemon when 1188 * we would be nearly out of memory. 1189 */ 1190 if (vm_paging_needed()) 1191 pagedaemon_wakeup(); 1192 1193 return (m); 1194 } 1195 1196 /* 1197 * vm_wait: (also see VM_WAIT macro) 1198 * 1199 * Block until free pages are available for allocation 1200 * - Called in various places before memory allocations. 1201 */ 1202 void 1203 vm_wait(void) 1204 { 1205 1206 mtx_lock(&vm_page_queue_free_mtx); 1207 if (curproc == pageproc) { 1208 vm_pageout_pages_needed = 1; 1209 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 1210 PDROP | PSWP, "VMWait", 0); 1211 } else { 1212 if (!vm_pages_needed) { 1213 vm_pages_needed = 1; 1214 wakeup(&vm_pages_needed); 1215 } 1216 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 1217 "vmwait", 0); 1218 } 1219 } 1220 1221 /* 1222 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1223 * 1224 * Block until free pages are available for allocation 1225 * - Called only in vm_fault so that processes page faulting 1226 * can be easily tracked. 1227 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1228 * processes will be able to grab memory first. Do not change 1229 * this balance without careful testing first. 1230 */ 1231 void 1232 vm_waitpfault(void) 1233 { 1234 1235 mtx_lock(&vm_page_queue_free_mtx); 1236 if (!vm_pages_needed) { 1237 vm_pages_needed = 1; 1238 wakeup(&vm_pages_needed); 1239 } 1240 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 1241 "pfault", 0); 1242 } 1243 1244 /* 1245 * vm_page_requeue: 1246 * 1247 * If the given page is contained within a page queue, move it to the tail 1248 * of that queue. 1249 * 1250 * The page queues must be locked. 1251 */ 1252 void 1253 vm_page_requeue(vm_page_t m) 1254 { 1255 int queue = VM_PAGE_GETQUEUE(m); 1256 struct vpgqueues *vpq; 1257 1258 if (queue != PQ_NONE) { 1259 vpq = &vm_page_queues[queue]; 1260 TAILQ_REMOVE(&vpq->pl, m, pageq); 1261 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1262 } 1263 } 1264 1265 /* 1266 * vm_pageq_remove: 1267 * 1268 * Remove a page from its queue. 1269 * 1270 * The queue containing the given page must be locked. 1271 * This routine may not block. 1272 */ 1273 void 1274 vm_pageq_remove(vm_page_t m) 1275 { 1276 int queue = VM_PAGE_GETQUEUE(m); 1277 struct vpgqueues *pq; 1278 1279 if (queue != PQ_NONE) { 1280 VM_PAGE_SETQUEUE2(m, PQ_NONE); 1281 pq = &vm_page_queues[queue]; 1282 TAILQ_REMOVE(&pq->pl, m, pageq); 1283 (*pq->cnt)--; 1284 } 1285 } 1286 1287 /* 1288 * vm_page_enqueue: 1289 * 1290 * Add the given page to the specified queue. 1291 * 1292 * The page queues must be locked. 1293 */ 1294 static void 1295 vm_page_enqueue(int queue, vm_page_t m) 1296 { 1297 struct vpgqueues *vpq; 1298 1299 vpq = &vm_page_queues[queue]; 1300 VM_PAGE_SETQUEUE2(m, queue); 1301 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq); 1302 ++*vpq->cnt; 1303 } 1304 1305 /* 1306 * vm_page_activate: 1307 * 1308 * Put the specified page on the active list (if appropriate). 1309 * Ensure that act_count is at least ACT_INIT but do not otherwise 1310 * mess with it. 1311 * 1312 * The page queues must be locked. 1313 * This routine may not block. 1314 */ 1315 void 1316 vm_page_activate(vm_page_t m) 1317 { 1318 1319 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1320 if (VM_PAGE_GETKNOWNQUEUE2(m) != PQ_ACTIVE) { 1321 vm_pageq_remove(m); 1322 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1323 if (m->act_count < ACT_INIT) 1324 m->act_count = ACT_INIT; 1325 vm_page_enqueue(PQ_ACTIVE, m); 1326 } 1327 } else { 1328 if (m->act_count < ACT_INIT) 1329 m->act_count = ACT_INIT; 1330 } 1331 } 1332 1333 /* 1334 * vm_page_free_wakeup: 1335 * 1336 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1337 * routine is called when a page has been added to the cache or free 1338 * queues. 1339 * 1340 * The page queues must be locked. 1341 * This routine may not block. 1342 */ 1343 static inline void 1344 vm_page_free_wakeup(void) 1345 { 1346 1347 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1348 /* 1349 * if pageout daemon needs pages, then tell it that there are 1350 * some free. 1351 */ 1352 if (vm_pageout_pages_needed && 1353 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1354 wakeup(&vm_pageout_pages_needed); 1355 vm_pageout_pages_needed = 0; 1356 } 1357 /* 1358 * wakeup processes that are waiting on memory if we hit a 1359 * high water mark. And wakeup scheduler process if we have 1360 * lots of memory. this process will swapin processes. 1361 */ 1362 if (vm_pages_needed && !vm_page_count_min()) { 1363 vm_pages_needed = 0; 1364 wakeup(&cnt.v_free_count); 1365 } 1366 } 1367 1368 /* 1369 * vm_page_free_toq: 1370 * 1371 * Returns the given page to the free list, 1372 * disassociating it with any VM object. 1373 * 1374 * Object and page must be locked prior to entry. 1375 * This routine may not block. 1376 */ 1377 1378 void 1379 vm_page_free_toq(vm_page_t m) 1380 { 1381 1382 if (VM_PAGE_GETQUEUE(m) != PQ_NONE) 1383 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1384 KASSERT(!pmap_page_is_mapped(m), 1385 ("vm_page_free_toq: freeing mapped page %p", m)); 1386 PCPU_INC(cnt.v_tfree); 1387 1388 if (m->busy || VM_PAGE_IS_FREE(m)) { 1389 printf( 1390 "vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n", 1391 (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0, 1392 m->hold_count); 1393 if (VM_PAGE_IS_FREE(m)) 1394 panic("vm_page_free: freeing free page"); 1395 else 1396 panic("vm_page_free: freeing busy page"); 1397 } 1398 1399 /* 1400 * unqueue, then remove page. Note that we cannot destroy 1401 * the page here because we do not want to call the pager's 1402 * callback routine until after we've put the page on the 1403 * appropriate free queue. 1404 */ 1405 vm_pageq_remove(m); 1406 vm_page_remove(m); 1407 1408 /* 1409 * If fictitious remove object association and 1410 * return, otherwise delay object association removal. 1411 */ 1412 if ((m->flags & PG_FICTITIOUS) != 0) { 1413 return; 1414 } 1415 1416 m->valid = 0; 1417 vm_page_undirty(m); 1418 1419 if (m->wire_count != 0) { 1420 if (m->wire_count > 1) { 1421 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1422 m->wire_count, (long)m->pindex); 1423 } 1424 panic("vm_page_free: freeing wired page"); 1425 } 1426 if (m->hold_count != 0) { 1427 m->flags &= ~PG_ZERO; 1428 vm_page_enqueue(PQ_HOLD, m); 1429 } else { 1430 /* 1431 * Restore the default memory attribute to the page. 1432 */ 1433 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 1434 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 1435 1436 /* 1437 * Insert the page into the physical memory allocator's 1438 * cache/free page queues. 1439 */ 1440 mtx_lock(&vm_page_queue_free_mtx); 1441 m->flags |= PG_FREE; 1442 cnt.v_free_count++; 1443 #if VM_NRESERVLEVEL > 0 1444 if (!vm_reserv_free_page(m)) 1445 #else 1446 if (TRUE) 1447 #endif 1448 vm_phys_free_pages(m, 0); 1449 if ((m->flags & PG_ZERO) != 0) 1450 ++vm_page_zero_count; 1451 else 1452 vm_page_zero_idle_wakeup(); 1453 vm_page_free_wakeup(); 1454 mtx_unlock(&vm_page_queue_free_mtx); 1455 } 1456 } 1457 1458 /* 1459 * vm_page_wire: 1460 * 1461 * Mark this page as wired down by yet 1462 * another map, removing it from paging queues 1463 * as necessary. 1464 * 1465 * The page queues must be locked. 1466 * This routine may not block. 1467 */ 1468 void 1469 vm_page_wire(vm_page_t m) 1470 { 1471 1472 /* 1473 * Only bump the wire statistics if the page is not already wired, 1474 * and only unqueue the page if it is on some queue (if it is unmanaged 1475 * it is already off the queues). 1476 */ 1477 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1478 if (m->flags & PG_FICTITIOUS) 1479 return; 1480 if (m->wire_count == 0) { 1481 if ((m->flags & PG_UNMANAGED) == 0) 1482 vm_pageq_remove(m); 1483 atomic_add_int(&cnt.v_wire_count, 1); 1484 } 1485 m->wire_count++; 1486 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1487 } 1488 1489 /* 1490 * vm_page_unwire: 1491 * 1492 * Release one wiring of this page, potentially 1493 * enabling it to be paged again. 1494 * 1495 * Many pages placed on the inactive queue should actually go 1496 * into the cache, but it is difficult to figure out which. What 1497 * we do instead, if the inactive target is well met, is to put 1498 * clean pages at the head of the inactive queue instead of the tail. 1499 * This will cause them to be moved to the cache more quickly and 1500 * if not actively re-referenced, freed more quickly. If we just 1501 * stick these pages at the end of the inactive queue, heavy filesystem 1502 * meta-data accesses can cause an unnecessary paging load on memory bound 1503 * processes. This optimization causes one-time-use metadata to be 1504 * reused more quickly. 1505 * 1506 * BUT, if we are in a low-memory situation we have no choice but to 1507 * put clean pages on the cache queue. 1508 * 1509 * A number of routines use vm_page_unwire() to guarantee that the page 1510 * will go into either the inactive or active queues, and will NEVER 1511 * be placed in the cache - for example, just after dirtying a page. 1512 * dirty pages in the cache are not allowed. 1513 * 1514 * The page queues must be locked. 1515 * This routine may not block. 1516 */ 1517 void 1518 vm_page_unwire(vm_page_t m, int activate) 1519 { 1520 1521 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1522 if (m->flags & PG_FICTITIOUS) 1523 return; 1524 if (m->wire_count > 0) { 1525 m->wire_count--; 1526 if (m->wire_count == 0) { 1527 atomic_subtract_int(&cnt.v_wire_count, 1); 1528 if (m->flags & PG_UNMANAGED) { 1529 ; 1530 } else if (activate) 1531 vm_page_enqueue(PQ_ACTIVE, m); 1532 else { 1533 vm_page_flag_clear(m, PG_WINATCFLS); 1534 vm_page_enqueue(PQ_INACTIVE, m); 1535 } 1536 } 1537 } else { 1538 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1539 } 1540 } 1541 1542 1543 /* 1544 * Move the specified page to the inactive queue. If the page has 1545 * any associated swap, the swap is deallocated. 1546 * 1547 * Normally athead is 0 resulting in LRU operation. athead is set 1548 * to 1 if we want this page to be 'as if it were placed in the cache', 1549 * except without unmapping it from the process address space. 1550 * 1551 * This routine may not block. 1552 */ 1553 static inline void 1554 _vm_page_deactivate(vm_page_t m, int athead) 1555 { 1556 1557 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1558 1559 /* 1560 * Ignore if already inactive. 1561 */ 1562 if (VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) 1563 return; 1564 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1565 vm_page_flag_clear(m, PG_WINATCFLS); 1566 vm_pageq_remove(m); 1567 if (athead) 1568 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1569 else 1570 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1571 VM_PAGE_SETQUEUE2(m, PQ_INACTIVE); 1572 cnt.v_inactive_count++; 1573 } 1574 } 1575 1576 void 1577 vm_page_deactivate(vm_page_t m) 1578 { 1579 _vm_page_deactivate(m, 0); 1580 } 1581 1582 /* 1583 * vm_page_try_to_cache: 1584 * 1585 * Returns 0 on failure, 1 on success 1586 */ 1587 int 1588 vm_page_try_to_cache(vm_page_t m) 1589 { 1590 1591 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1592 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1593 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1594 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) { 1595 return (0); 1596 } 1597 pmap_remove_all(m); 1598 if (m->dirty) 1599 return (0); 1600 vm_page_cache(m); 1601 return (1); 1602 } 1603 1604 /* 1605 * vm_page_try_to_free() 1606 * 1607 * Attempt to free the page. If we cannot free it, we do nothing. 1608 * 1 is returned on success, 0 on failure. 1609 */ 1610 int 1611 vm_page_try_to_free(vm_page_t m) 1612 { 1613 1614 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1615 if (m->object != NULL) 1616 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1617 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1618 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) { 1619 return (0); 1620 } 1621 pmap_remove_all(m); 1622 if (m->dirty) 1623 return (0); 1624 vm_page_free(m); 1625 return (1); 1626 } 1627 1628 /* 1629 * vm_page_cache 1630 * 1631 * Put the specified page onto the page cache queue (if appropriate). 1632 * 1633 * This routine may not block. 1634 */ 1635 void 1636 vm_page_cache(vm_page_t m) 1637 { 1638 vm_object_t object; 1639 vm_page_t root; 1640 1641 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1642 object = m->object; 1643 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1644 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy || 1645 m->hold_count || m->wire_count) { 1646 panic("vm_page_cache: attempting to cache busy page"); 1647 } 1648 pmap_remove_all(m); 1649 if (m->dirty != 0) 1650 panic("vm_page_cache: page %p is dirty", m); 1651 if (m->valid == 0 || object->type == OBJT_DEFAULT || 1652 (object->type == OBJT_SWAP && 1653 !vm_pager_has_page(object, m->pindex, NULL, NULL))) { 1654 /* 1655 * Hypothesis: A cache-elgible page belonging to a 1656 * default object or swap object but without a backing 1657 * store must be zero filled. 1658 */ 1659 vm_page_free(m); 1660 return; 1661 } 1662 KASSERT((m->flags & PG_CACHED) == 0, 1663 ("vm_page_cache: page %p is already cached", m)); 1664 cnt.v_tcached++; 1665 1666 /* 1667 * Remove the page from the paging queues. 1668 */ 1669 vm_pageq_remove(m); 1670 1671 /* 1672 * Remove the page from the object's collection of resident 1673 * pages. 1674 */ 1675 if (m != object->root) 1676 vm_page_splay(m->pindex, object->root); 1677 if (m->left == NULL) 1678 root = m->right; 1679 else { 1680 root = vm_page_splay(m->pindex, m->left); 1681 root->right = m->right; 1682 } 1683 object->root = root; 1684 TAILQ_REMOVE(&object->memq, m, listq); 1685 object->resident_page_count--; 1686 object->generation++; 1687 1688 /* 1689 * Restore the default memory attribute to the page. 1690 */ 1691 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) 1692 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); 1693 1694 /* 1695 * Insert the page into the object's collection of cached pages 1696 * and the physical memory allocator's cache/free page queues. 1697 */ 1698 vm_page_flag_clear(m, PG_ZERO); 1699 mtx_lock(&vm_page_queue_free_mtx); 1700 m->flags |= PG_CACHED; 1701 cnt.v_cache_count++; 1702 root = object->cache; 1703 if (root == NULL) { 1704 m->left = NULL; 1705 m->right = NULL; 1706 } else { 1707 root = vm_page_splay(m->pindex, root); 1708 if (m->pindex < root->pindex) { 1709 m->left = root->left; 1710 m->right = root; 1711 root->left = NULL; 1712 } else if (__predict_false(m->pindex == root->pindex)) 1713 panic("vm_page_cache: offset already cached"); 1714 else { 1715 m->right = root->right; 1716 m->left = root; 1717 root->right = NULL; 1718 } 1719 } 1720 object->cache = m; 1721 #if VM_NRESERVLEVEL > 0 1722 if (!vm_reserv_free_page(m)) { 1723 #else 1724 if (TRUE) { 1725 #endif 1726 vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); 1727 vm_phys_free_pages(m, 0); 1728 } 1729 vm_page_free_wakeup(); 1730 mtx_unlock(&vm_page_queue_free_mtx); 1731 1732 /* 1733 * Increment the vnode's hold count if this is the object's only 1734 * cached page. Decrement the vnode's hold count if this was 1735 * the object's only resident page. 1736 */ 1737 if (object->type == OBJT_VNODE) { 1738 if (root == NULL && object->resident_page_count != 0) 1739 vhold(object->handle); 1740 else if (root != NULL && object->resident_page_count == 0) 1741 vdrop(object->handle); 1742 } 1743 } 1744 1745 /* 1746 * vm_page_dontneed 1747 * 1748 * Cache, deactivate, or do nothing as appropriate. This routine 1749 * is typically used by madvise() MADV_DONTNEED. 1750 * 1751 * Generally speaking we want to move the page into the cache so 1752 * it gets reused quickly. However, this can result in a silly syndrome 1753 * due to the page recycling too quickly. Small objects will not be 1754 * fully cached. On the otherhand, if we move the page to the inactive 1755 * queue we wind up with a problem whereby very large objects 1756 * unnecessarily blow away our inactive and cache queues. 1757 * 1758 * The solution is to move the pages based on a fixed weighting. We 1759 * either leave them alone, deactivate them, or move them to the cache, 1760 * where moving them to the cache has the highest weighting. 1761 * By forcing some pages into other queues we eventually force the 1762 * system to balance the queues, potentially recovering other unrelated 1763 * space from active. The idea is to not force this to happen too 1764 * often. 1765 */ 1766 void 1767 vm_page_dontneed(vm_page_t m) 1768 { 1769 static int dnweight; 1770 int dnw; 1771 int head; 1772 1773 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1774 dnw = ++dnweight; 1775 1776 /* 1777 * occassionally leave the page alone 1778 */ 1779 if ((dnw & 0x01F0) == 0 || 1780 VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) { 1781 if (m->act_count >= ACT_INIT) 1782 --m->act_count; 1783 return; 1784 } 1785 1786 /* 1787 * Clear any references to the page. Otherwise, the page daemon will 1788 * immediately reactivate the page. 1789 */ 1790 vm_page_flag_clear(m, PG_REFERENCED); 1791 pmap_clear_reference(m); 1792 1793 if (m->dirty == 0 && pmap_is_modified(m)) 1794 vm_page_dirty(m); 1795 1796 if (m->dirty || (dnw & 0x0070) == 0) { 1797 /* 1798 * Deactivate the page 3 times out of 32. 1799 */ 1800 head = 0; 1801 } else { 1802 /* 1803 * Cache the page 28 times out of every 32. Note that 1804 * the page is deactivated instead of cached, but placed 1805 * at the head of the queue instead of the tail. 1806 */ 1807 head = 1; 1808 } 1809 _vm_page_deactivate(m, head); 1810 } 1811 1812 /* 1813 * Grab a page, waiting until we are waken up due to the page 1814 * changing state. We keep on waiting, if the page continues 1815 * to be in the object. If the page doesn't exist, first allocate it 1816 * and then conditionally zero it. 1817 * 1818 * This routine may block. 1819 */ 1820 vm_page_t 1821 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1822 { 1823 vm_page_t m; 1824 1825 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1826 retrylookup: 1827 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1828 if (vm_page_sleep_if_busy(m, TRUE, "pgrbwt")) { 1829 if ((allocflags & VM_ALLOC_RETRY) == 0) 1830 return (NULL); 1831 goto retrylookup; 1832 } else { 1833 if ((allocflags & VM_ALLOC_WIRED) != 0) { 1834 vm_page_lock_queues(); 1835 vm_page_wire(m); 1836 vm_page_unlock_queues(); 1837 } 1838 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 1839 vm_page_busy(m); 1840 return (m); 1841 } 1842 } 1843 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1844 if (m == NULL) { 1845 VM_OBJECT_UNLOCK(object); 1846 VM_WAIT; 1847 VM_OBJECT_LOCK(object); 1848 if ((allocflags & VM_ALLOC_RETRY) == 0) 1849 return (NULL); 1850 goto retrylookup; 1851 } else if (m->valid != 0) 1852 return (m); 1853 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 1854 pmap_zero_page(m); 1855 return (m); 1856 } 1857 1858 /* 1859 * Mapping function for valid bits or for dirty bits in 1860 * a page. May not block. 1861 * 1862 * Inputs are required to range within a page. 1863 */ 1864 int 1865 vm_page_bits(int base, int size) 1866 { 1867 int first_bit; 1868 int last_bit; 1869 1870 KASSERT( 1871 base + size <= PAGE_SIZE, 1872 ("vm_page_bits: illegal base/size %d/%d", base, size) 1873 ); 1874 1875 if (size == 0) /* handle degenerate case */ 1876 return (0); 1877 1878 first_bit = base >> DEV_BSHIFT; 1879 last_bit = (base + size - 1) >> DEV_BSHIFT; 1880 1881 return ((2 << last_bit) - (1 << first_bit)); 1882 } 1883 1884 /* 1885 * vm_page_set_valid: 1886 * 1887 * Sets portions of a page valid. The arguments are expected 1888 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1889 * of any partial chunks touched by the range. The invalid portion of 1890 * such chunks will be zeroed. 1891 * 1892 * (base + size) must be less then or equal to PAGE_SIZE. 1893 */ 1894 void 1895 vm_page_set_valid(vm_page_t m, int base, int size) 1896 { 1897 int endoff, frag; 1898 1899 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1900 if (size == 0) /* handle degenerate case */ 1901 return; 1902 1903 /* 1904 * If the base is not DEV_BSIZE aligned and the valid 1905 * bit is clear, we have to zero out a portion of the 1906 * first block. 1907 */ 1908 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1909 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1910 pmap_zero_page_area(m, frag, base - frag); 1911 1912 /* 1913 * If the ending offset is not DEV_BSIZE aligned and the 1914 * valid bit is clear, we have to zero out a portion of 1915 * the last block. 1916 */ 1917 endoff = base + size; 1918 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1919 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1920 pmap_zero_page_area(m, endoff, 1921 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1922 1923 /* 1924 * Assert that no previously invalid block that is now being validated 1925 * is already dirty. 1926 */ 1927 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, 1928 ("vm_page_set_valid: page %p is dirty", m)); 1929 1930 /* 1931 * Set valid bits inclusive of any overlap. 1932 */ 1933 m->valid |= vm_page_bits(base, size); 1934 } 1935 1936 /* 1937 * vm_page_set_validclean: 1938 * 1939 * Sets portions of a page valid and clean. The arguments are expected 1940 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1941 * of any partial chunks touched by the range. The invalid portion of 1942 * such chunks will be zero'd. 1943 * 1944 * This routine may not block. 1945 * 1946 * (base + size) must be less then or equal to PAGE_SIZE. 1947 */ 1948 void 1949 vm_page_set_validclean(vm_page_t m, int base, int size) 1950 { 1951 int pagebits; 1952 int frag; 1953 int endoff; 1954 1955 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1956 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1957 if (size == 0) /* handle degenerate case */ 1958 return; 1959 1960 /* 1961 * If the base is not DEV_BSIZE aligned and the valid 1962 * bit is clear, we have to zero out a portion of the 1963 * first block. 1964 */ 1965 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1966 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1967 pmap_zero_page_area(m, frag, base - frag); 1968 1969 /* 1970 * If the ending offset is not DEV_BSIZE aligned and the 1971 * valid bit is clear, we have to zero out a portion of 1972 * the last block. 1973 */ 1974 endoff = base + size; 1975 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1976 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1977 pmap_zero_page_area(m, endoff, 1978 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1979 1980 /* 1981 * Set valid, clear dirty bits. If validating the entire 1982 * page we can safely clear the pmap modify bit. We also 1983 * use this opportunity to clear the VPO_NOSYNC flag. If a process 1984 * takes a write fault on a MAP_NOSYNC memory area the flag will 1985 * be set again. 1986 * 1987 * We set valid bits inclusive of any overlap, but we can only 1988 * clear dirty bits for DEV_BSIZE chunks that are fully within 1989 * the range. 1990 */ 1991 pagebits = vm_page_bits(base, size); 1992 m->valid |= pagebits; 1993 #if 0 /* NOT YET */ 1994 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1995 frag = DEV_BSIZE - frag; 1996 base += frag; 1997 size -= frag; 1998 if (size < 0) 1999 size = 0; 2000 } 2001 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 2002 #endif 2003 m->dirty &= ~pagebits; 2004 if (base == 0 && size == PAGE_SIZE) { 2005 pmap_clear_modify(m); 2006 m->oflags &= ~VPO_NOSYNC; 2007 } 2008 } 2009 2010 void 2011 vm_page_clear_dirty(vm_page_t m, int base, int size) 2012 { 2013 2014 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2015 m->dirty &= ~vm_page_bits(base, size); 2016 } 2017 2018 /* 2019 * vm_page_set_invalid: 2020 * 2021 * Invalidates DEV_BSIZE'd chunks within a page. Both the 2022 * valid and dirty bits for the effected areas are cleared. 2023 * 2024 * May not block. 2025 */ 2026 void 2027 vm_page_set_invalid(vm_page_t m, int base, int size) 2028 { 2029 int bits; 2030 2031 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2032 bits = vm_page_bits(base, size); 2033 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2034 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 2035 pmap_remove_all(m); 2036 m->valid &= ~bits; 2037 m->dirty &= ~bits; 2038 m->object->generation++; 2039 } 2040 2041 /* 2042 * vm_page_zero_invalid() 2043 * 2044 * The kernel assumes that the invalid portions of a page contain 2045 * garbage, but such pages can be mapped into memory by user code. 2046 * When this occurs, we must zero out the non-valid portions of the 2047 * page so user code sees what it expects. 2048 * 2049 * Pages are most often semi-valid when the end of a file is mapped 2050 * into memory and the file's size is not page aligned. 2051 */ 2052 void 2053 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 2054 { 2055 int b; 2056 int i; 2057 2058 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2059 /* 2060 * Scan the valid bits looking for invalid sections that 2061 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 2062 * valid bit may be set ) have already been zerod by 2063 * vm_page_set_validclean(). 2064 */ 2065 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 2066 if (i == (PAGE_SIZE / DEV_BSIZE) || 2067 (m->valid & (1 << i)) 2068 ) { 2069 if (i > b) { 2070 pmap_zero_page_area(m, 2071 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 2072 } 2073 b = i + 1; 2074 } 2075 } 2076 2077 /* 2078 * setvalid is TRUE when we can safely set the zero'd areas 2079 * as being valid. We can do this if there are no cache consistancy 2080 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 2081 */ 2082 if (setvalid) 2083 m->valid = VM_PAGE_BITS_ALL; 2084 } 2085 2086 /* 2087 * vm_page_is_valid: 2088 * 2089 * Is (partial) page valid? Note that the case where size == 0 2090 * will return FALSE in the degenerate case where the page is 2091 * entirely invalid, and TRUE otherwise. 2092 * 2093 * May not block. 2094 */ 2095 int 2096 vm_page_is_valid(vm_page_t m, int base, int size) 2097 { 2098 int bits = vm_page_bits(base, size); 2099 2100 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 2101 if (m->valid && ((m->valid & bits) == bits)) 2102 return 1; 2103 else 2104 return 0; 2105 } 2106 2107 /* 2108 * update dirty bits from pmap/mmu. May not block. 2109 */ 2110 void 2111 vm_page_test_dirty(vm_page_t m) 2112 { 2113 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 2114 vm_page_dirty(m); 2115 } 2116 } 2117 2118 int so_zerocp_fullpage = 0; 2119 2120 /* 2121 * Replace the given page with a copy. The copied page assumes 2122 * the portion of the given page's "wire_count" that is not the 2123 * responsibility of this copy-on-write mechanism. 2124 * 2125 * The object containing the given page must have a non-zero 2126 * paging-in-progress count and be locked. 2127 */ 2128 void 2129 vm_page_cowfault(vm_page_t m) 2130 { 2131 vm_page_t mnew; 2132 vm_object_t object; 2133 vm_pindex_t pindex; 2134 2135 object = m->object; 2136 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 2137 KASSERT(object->paging_in_progress != 0, 2138 ("vm_page_cowfault: object %p's paging-in-progress count is zero.", 2139 object)); 2140 pindex = m->pindex; 2141 2142 retry_alloc: 2143 pmap_remove_all(m); 2144 vm_page_remove(m); 2145 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 2146 if (mnew == NULL) { 2147 vm_page_insert(m, object, pindex); 2148 vm_page_unlock_queues(); 2149 VM_OBJECT_UNLOCK(object); 2150 VM_WAIT; 2151 VM_OBJECT_LOCK(object); 2152 if (m == vm_page_lookup(object, pindex)) { 2153 vm_page_lock_queues(); 2154 goto retry_alloc; 2155 } else { 2156 /* 2157 * Page disappeared during the wait. 2158 */ 2159 vm_page_lock_queues(); 2160 return; 2161 } 2162 } 2163 2164 if (m->cow == 0) { 2165 /* 2166 * check to see if we raced with an xmit complete when 2167 * waiting to allocate a page. If so, put things back 2168 * the way they were 2169 */ 2170 vm_page_free(mnew); 2171 vm_page_insert(m, object, pindex); 2172 } else { /* clear COW & copy page */ 2173 if (!so_zerocp_fullpage) 2174 pmap_copy_page(m, mnew); 2175 mnew->valid = VM_PAGE_BITS_ALL; 2176 vm_page_dirty(mnew); 2177 mnew->wire_count = m->wire_count - m->cow; 2178 m->wire_count = m->cow; 2179 } 2180 } 2181 2182 void 2183 vm_page_cowclear(vm_page_t m) 2184 { 2185 2186 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2187 if (m->cow) { 2188 m->cow--; 2189 /* 2190 * let vm_fault add back write permission lazily 2191 */ 2192 } 2193 /* 2194 * sf_buf_free() will free the page, so we needn't do it here 2195 */ 2196 } 2197 2198 int 2199 vm_page_cowsetup(vm_page_t m) 2200 { 2201 2202 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 2203 if (m->cow == USHRT_MAX - 1) 2204 return (EBUSY); 2205 m->cow++; 2206 pmap_remove_write(m); 2207 return (0); 2208 } 2209 2210 #include "opt_ddb.h" 2211 #ifdef DDB 2212 #include <sys/kernel.h> 2213 2214 #include <ddb/ddb.h> 2215 2216 DB_SHOW_COMMAND(page, vm_page_print_page_info) 2217 { 2218 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 2219 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 2220 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 2221 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 2222 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 2223 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 2224 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 2225 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 2226 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 2227 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 2228 } 2229 2230 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 2231 { 2232 2233 db_printf("PQ_FREE:"); 2234 db_printf(" %d", cnt.v_free_count); 2235 db_printf("\n"); 2236 2237 db_printf("PQ_CACHE:"); 2238 db_printf(" %d", cnt.v_cache_count); 2239 db_printf("\n"); 2240 2241 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 2242 *vm_page_queues[PQ_ACTIVE].cnt, 2243 *vm_page_queues[PQ_INACTIVE].cnt); 2244 } 2245 #endif /* DDB */ 2246