1 /*- 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 4. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 33 */ 34 35 /*- 36 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 37 * All rights reserved. 38 * 39 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 40 * 41 * Permission to use, copy, modify and distribute this software and 42 * its documentation is hereby granted, provided that both the copyright 43 * notice and this permission notice appear in all copies of the 44 * software, derivative works or modified versions, and any portions 45 * thereof, and that both notices appear in supporting documentation. 46 * 47 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 48 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 49 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 50 * 51 * Carnegie Mellon requests users of this software to return to 52 * 53 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 54 * School of Computer Science 55 * Carnegie Mellon University 56 * Pittsburgh PA 15213-3890 57 * 58 * any improvements or extensions that they make and grant Carnegie the 59 * rights to redistribute these changes. 60 */ 61 62 /* 63 * GENERAL RULES ON VM_PAGE MANIPULATION 64 * 65 * - a pageq mutex is required when adding or removing a page from a 66 * page queue (vm_page_queue[]), regardless of other mutexes or the 67 * busy state of a page. 68 * 69 * - a hash chain mutex is required when associating or disassociating 70 * a page from the VM PAGE CACHE hash table (vm_page_buckets), 71 * regardless of other mutexes or the busy state of a page. 72 * 73 * - either a hash chain mutex OR a busied page is required in order 74 * to modify the page flags. A hash chain mutex must be obtained in 75 * order to busy a page. A page's flags cannot be modified by a 76 * hash chain mutex if the page is marked busy. 77 * 78 * - The object memq mutex is held when inserting or removing 79 * pages from an object (vm_page_insert() or vm_page_remove()). This 80 * is different from the object's main mutex. 81 * 82 * Generally speaking, you have to be aware of side effects when running 83 * vm_page ops. A vm_page_lookup() will return with the hash chain 84 * locked, whether it was able to lookup the page or not. vm_page_free(), 85 * vm_page_cache(), vm_page_activate(), and a number of other routines 86 * will release the hash chain mutex for you. Intermediate manipulation 87 * routines such as vm_page_flag_set() expect the hash chain to be held 88 * on entry and the hash chain will remain held on return. 89 * 90 * pageq scanning can only occur with the pageq in question locked. 91 * We have a known bottleneck with the active queue, but the cache 92 * and free queues are actually arrays already. 93 */ 94 95 /* 96 * Resident memory management module. 97 */ 98 99 #include <sys/cdefs.h> 100 __FBSDID("$FreeBSD$"); 101 102 #include <sys/param.h> 103 #include <sys/systm.h> 104 #include <sys/lock.h> 105 #include <sys/kernel.h> 106 #include <sys/malloc.h> 107 #include <sys/mutex.h> 108 #include <sys/proc.h> 109 #include <sys/sysctl.h> 110 #include <sys/vmmeter.h> 111 #include <sys/vnode.h> 112 113 #include <vm/vm.h> 114 #include <vm/vm_param.h> 115 #include <vm/vm_kern.h> 116 #include <vm/vm_object.h> 117 #include <vm/vm_page.h> 118 #include <vm/vm_pageout.h> 119 #include <vm/vm_pager.h> 120 #include <vm/vm_extern.h> 121 #include <vm/uma.h> 122 #include <vm/uma_int.h> 123 124 #include <machine/md_var.h> 125 126 /* 127 * Associated with page of user-allocatable memory is a 128 * page structure. 129 */ 130 131 struct mtx vm_page_queue_mtx; 132 struct mtx vm_page_queue_free_mtx; 133 134 vm_page_t vm_page_array = 0; 135 int vm_page_array_size = 0; 136 long first_page = 0; 137 int vm_page_zero_count = 0; 138 139 static int boot_pages = UMA_BOOT_PAGES; 140 TUNABLE_INT("vm.boot_pages", &boot_pages); 141 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0, 142 "number of pages allocated for bootstrapping the VM system"); 143 144 /* 145 * vm_set_page_size: 146 * 147 * Sets the page size, perhaps based upon the memory 148 * size. Must be called before any use of page-size 149 * dependent functions. 150 */ 151 void 152 vm_set_page_size(void) 153 { 154 if (cnt.v_page_size == 0) 155 cnt.v_page_size = PAGE_SIZE; 156 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 157 panic("vm_set_page_size: page size not a power of two"); 158 } 159 160 /* 161 * vm_page_blacklist_lookup: 162 * 163 * See if a physical address in this page has been listed 164 * in the blacklist tunable. Entries in the tunable are 165 * separated by spaces or commas. If an invalid integer is 166 * encountered then the rest of the string is skipped. 167 */ 168 static int 169 vm_page_blacklist_lookup(char *list, vm_paddr_t pa) 170 { 171 vm_paddr_t bad; 172 char *cp, *pos; 173 174 for (pos = list; *pos != '\0'; pos = cp) { 175 bad = strtoq(pos, &cp, 0); 176 if (*cp != '\0') { 177 if (*cp == ' ' || *cp == ',') { 178 cp++; 179 if (cp == pos) 180 continue; 181 } else 182 break; 183 } 184 if (pa == trunc_page(bad)) 185 return (1); 186 } 187 return (0); 188 } 189 190 /* 191 * vm_page_startup: 192 * 193 * Initializes the resident memory module. 194 * 195 * Allocates memory for the page cells, and 196 * for the object/offset-to-page hash table headers. 197 * Each page cell is initialized and placed on the free list. 198 */ 199 vm_offset_t 200 vm_page_startup(vm_offset_t vaddr) 201 { 202 vm_offset_t mapped; 203 vm_size_t npages; 204 vm_paddr_t page_range; 205 vm_paddr_t new_end; 206 int i; 207 vm_paddr_t pa; 208 int nblocks; 209 vm_paddr_t last_pa; 210 char *list; 211 212 /* the biggest memory array is the second group of pages */ 213 vm_paddr_t end; 214 vm_paddr_t biggestsize; 215 vm_paddr_t low_water, high_water; 216 int biggestone; 217 218 vm_paddr_t total; 219 220 total = 0; 221 biggestsize = 0; 222 biggestone = 0; 223 nblocks = 0; 224 vaddr = round_page(vaddr); 225 226 for (i = 0; phys_avail[i + 1]; i += 2) { 227 phys_avail[i] = round_page(phys_avail[i]); 228 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 229 } 230 231 low_water = phys_avail[0]; 232 high_water = phys_avail[1]; 233 234 for (i = 0; phys_avail[i + 1]; i += 2) { 235 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 236 237 if (size > biggestsize) { 238 biggestone = i; 239 biggestsize = size; 240 } 241 if (phys_avail[i] < low_water) 242 low_water = phys_avail[i]; 243 if (phys_avail[i + 1] > high_water) 244 high_water = phys_avail[i + 1]; 245 ++nblocks; 246 total += size; 247 } 248 249 end = phys_avail[biggestone+1]; 250 251 /* 252 * Initialize the locks. 253 */ 254 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF | 255 MTX_RECURSE); 256 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 257 MTX_DEF); 258 259 /* 260 * Initialize the queue headers for the free queue, the active queue 261 * and the inactive queue. 262 */ 263 vm_pageq_init(); 264 265 /* 266 * Allocate memory for use when boot strapping the kernel memory 267 * allocator. 268 */ 269 new_end = end - (boot_pages * UMA_SLAB_SIZE); 270 new_end = trunc_page(new_end); 271 mapped = pmap_map(&vaddr, new_end, end, 272 VM_PROT_READ | VM_PROT_WRITE); 273 bzero((void *)mapped, end - new_end); 274 uma_startup((void *)mapped, boot_pages); 275 276 #if defined(__amd64__) || defined(__i386__) 277 /* 278 * Allocate a bitmap to indicate that a random physical page 279 * needs to be included in a minidump. 280 * 281 * The amd64 port needs this to indicate which direct map pages 282 * need to be dumped, via calls to dump_add_page()/dump_drop_page(). 283 * 284 * However, i386 still needs this workspace internally within the 285 * minidump code. In theory, they are not needed on i386, but are 286 * included should the sf_buf code decide to use them. 287 */ 288 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE; 289 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); 290 new_end -= vm_page_dump_size; 291 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, 292 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); 293 bzero((void *)vm_page_dump, vm_page_dump_size); 294 #endif 295 /* 296 * Compute the number of pages of memory that will be available for 297 * use (taking into account the overhead of a page structure per 298 * page). 299 */ 300 first_page = low_water / PAGE_SIZE; 301 page_range = high_water / PAGE_SIZE - first_page; 302 npages = (total - (page_range * sizeof(struct vm_page)) - 303 (end - new_end)) / PAGE_SIZE; 304 end = new_end; 305 306 /* 307 * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 308 */ 309 vaddr += PAGE_SIZE; 310 311 /* 312 * Initialize the mem entry structures now, and put them in the free 313 * queue. 314 */ 315 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 316 mapped = pmap_map(&vaddr, new_end, end, 317 VM_PROT_READ | VM_PROT_WRITE); 318 vm_page_array = (vm_page_t) mapped; 319 #ifdef __amd64__ 320 /* 321 * pmap_map on amd64 comes out of the direct-map, not kvm like i386, 322 * so the pages must be tracked for a crashdump to include this data. 323 * This includes the vm_page_array and the early UMA bootstrap pages. 324 */ 325 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE) 326 dump_add_page(pa); 327 #endif 328 phys_avail[biggestone + 1] = new_end; 329 330 /* 331 * Clear all of the page structures 332 */ 333 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 334 vm_page_array_size = page_range; 335 336 /* 337 * This assertion tests the hypothesis that npages and total are 338 * redundant. XXX 339 */ 340 page_range = 0; 341 for (i = 0; phys_avail[i + 1] != 0; i += 2) 342 page_range += atop(phys_avail[i + 1] - phys_avail[i]); 343 KASSERT(page_range == npages, 344 ("vm_page_startup: inconsistent page counts")); 345 346 /* 347 * Construct the free queue(s) in descending order (by physical 348 * address) so that the first 16MB of physical memory is allocated 349 * last rather than first. On large-memory machines, this avoids 350 * the exhaustion of low physical memory before isa_dma_init has run. 351 */ 352 cnt.v_page_count = 0; 353 cnt.v_free_count = 0; 354 list = getenv("vm.blacklist"); 355 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 356 pa = phys_avail[i]; 357 last_pa = phys_avail[i + 1]; 358 while (pa < last_pa) { 359 if (list != NULL && 360 vm_page_blacklist_lookup(list, pa)) 361 printf("Skipping page with pa 0x%jx\n", 362 (uintmax_t)pa); 363 else 364 vm_pageq_add_new_page(pa); 365 pa += PAGE_SIZE; 366 } 367 } 368 freeenv(list); 369 return (vaddr); 370 } 371 372 void 373 vm_page_flag_set(vm_page_t m, unsigned short bits) 374 { 375 376 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 377 m->flags |= bits; 378 } 379 380 void 381 vm_page_flag_clear(vm_page_t m, unsigned short bits) 382 { 383 384 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 385 m->flags &= ~bits; 386 } 387 388 void 389 vm_page_busy(vm_page_t m) 390 { 391 392 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 393 KASSERT((m->oflags & VPO_BUSY) == 0, 394 ("vm_page_busy: page already busy!!!")); 395 m->oflags |= VPO_BUSY; 396 } 397 398 /* 399 * vm_page_flash: 400 * 401 * wakeup anyone waiting for the page. 402 */ 403 void 404 vm_page_flash(vm_page_t m) 405 { 406 407 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 408 if (m->oflags & VPO_WANTED) { 409 m->oflags &= ~VPO_WANTED; 410 wakeup(m); 411 } 412 } 413 414 /* 415 * vm_page_wakeup: 416 * 417 * clear the VPO_BUSY flag and wakeup anyone waiting for the 418 * page. 419 * 420 */ 421 void 422 vm_page_wakeup(vm_page_t m) 423 { 424 425 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 426 KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!")); 427 m->oflags &= ~VPO_BUSY; 428 vm_page_flash(m); 429 } 430 431 void 432 vm_page_io_start(vm_page_t m) 433 { 434 435 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 436 m->busy++; 437 } 438 439 void 440 vm_page_io_finish(vm_page_t m) 441 { 442 443 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 444 m->busy--; 445 if (m->busy == 0) 446 vm_page_flash(m); 447 } 448 449 /* 450 * Keep page from being freed by the page daemon 451 * much of the same effect as wiring, except much lower 452 * overhead and should be used only for *very* temporary 453 * holding ("wiring"). 454 */ 455 void 456 vm_page_hold(vm_page_t mem) 457 { 458 459 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 460 mem->hold_count++; 461 } 462 463 void 464 vm_page_unhold(vm_page_t mem) 465 { 466 467 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 468 --mem->hold_count; 469 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 470 if (mem->hold_count == 0 && VM_PAGE_INQUEUE2(mem, PQ_HOLD)) 471 vm_page_free_toq(mem); 472 } 473 474 /* 475 * vm_page_free: 476 * 477 * Free a page. 478 */ 479 void 480 vm_page_free(vm_page_t m) 481 { 482 483 m->flags &= ~PG_ZERO; 484 vm_page_free_toq(m); 485 } 486 487 /* 488 * vm_page_free_zero: 489 * 490 * Free a page to the zerod-pages queue 491 */ 492 void 493 vm_page_free_zero(vm_page_t m) 494 { 495 496 m->flags |= PG_ZERO; 497 vm_page_free_toq(m); 498 } 499 500 /* 501 * vm_page_sleep: 502 * 503 * Sleep and release the page queues lock. 504 * 505 * The object containing the given page must be locked. 506 */ 507 void 508 vm_page_sleep(vm_page_t m, const char *msg) 509 { 510 511 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 512 if (!mtx_owned(&vm_page_queue_mtx)) 513 vm_page_lock_queues(); 514 vm_page_flag_set(m, PG_REFERENCED); 515 vm_page_unlock_queues(); 516 517 /* 518 * It's possible that while we sleep, the page will get 519 * unbusied and freed. If we are holding the object 520 * lock, we will assume we hold a reference to the object 521 * such that even if m->object changes, we can re-lock 522 * it. 523 */ 524 m->oflags |= VPO_WANTED; 525 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0); 526 } 527 528 /* 529 * vm_page_dirty: 530 * 531 * make page all dirty 532 */ 533 void 534 vm_page_dirty(vm_page_t m) 535 { 536 KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_CACHE, 537 ("vm_page_dirty: page in cache!")); 538 KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_FREE, 539 ("vm_page_dirty: page is free!")); 540 m->dirty = VM_PAGE_BITS_ALL; 541 } 542 543 /* 544 * vm_page_splay: 545 * 546 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 547 * the vm_page containing the given pindex. If, however, that 548 * pindex is not found in the vm_object, returns a vm_page that is 549 * adjacent to the pindex, coming before or after it. 550 */ 551 vm_page_t 552 vm_page_splay(vm_pindex_t pindex, vm_page_t root) 553 { 554 struct vm_page dummy; 555 vm_page_t lefttreemax, righttreemin, y; 556 557 if (root == NULL) 558 return (root); 559 lefttreemax = righttreemin = &dummy; 560 for (;; root = y) { 561 if (pindex < root->pindex) { 562 if ((y = root->left) == NULL) 563 break; 564 if (pindex < y->pindex) { 565 /* Rotate right. */ 566 root->left = y->right; 567 y->right = root; 568 root = y; 569 if ((y = root->left) == NULL) 570 break; 571 } 572 /* Link into the new root's right tree. */ 573 righttreemin->left = root; 574 righttreemin = root; 575 } else if (pindex > root->pindex) { 576 if ((y = root->right) == NULL) 577 break; 578 if (pindex > y->pindex) { 579 /* Rotate left. */ 580 root->right = y->left; 581 y->left = root; 582 root = y; 583 if ((y = root->right) == NULL) 584 break; 585 } 586 /* Link into the new root's left tree. */ 587 lefttreemax->right = root; 588 lefttreemax = root; 589 } else 590 break; 591 } 592 /* Assemble the new root. */ 593 lefttreemax->right = root->left; 594 righttreemin->left = root->right; 595 root->left = dummy.right; 596 root->right = dummy.left; 597 return (root); 598 } 599 600 /* 601 * vm_page_insert: [ internal use only ] 602 * 603 * Inserts the given mem entry into the object and object list. 604 * 605 * The pagetables are not updated but will presumably fault the page 606 * in if necessary, or if a kernel page the caller will at some point 607 * enter the page into the kernel's pmap. We are not allowed to block 608 * here so we *can't* do this anyway. 609 * 610 * The object and page must be locked. 611 * This routine may not block. 612 */ 613 void 614 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 615 { 616 vm_page_t root; 617 618 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 619 if (m->object != NULL) 620 panic("vm_page_insert: page already inserted"); 621 622 /* 623 * Record the object/offset pair in this page 624 */ 625 m->object = object; 626 m->pindex = pindex; 627 628 /* 629 * Now link into the object's ordered list of backed pages. 630 */ 631 root = object->root; 632 if (root == NULL) { 633 m->left = NULL; 634 m->right = NULL; 635 TAILQ_INSERT_TAIL(&object->memq, m, listq); 636 } else { 637 root = vm_page_splay(pindex, root); 638 if (pindex < root->pindex) { 639 m->left = root->left; 640 m->right = root; 641 root->left = NULL; 642 TAILQ_INSERT_BEFORE(root, m, listq); 643 } else if (pindex == root->pindex) 644 panic("vm_page_insert: offset already allocated"); 645 else { 646 m->right = root->right; 647 m->left = root; 648 root->right = NULL; 649 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 650 } 651 } 652 object->root = m; 653 object->generation++; 654 655 /* 656 * show that the object has one more resident page. 657 */ 658 object->resident_page_count++; 659 /* 660 * Hold the vnode until the last page is released. 661 */ 662 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 663 vhold((struct vnode *)object->handle); 664 665 /* 666 * Since we are inserting a new and possibly dirty page, 667 * update the object's OBJ_MIGHTBEDIRTY flag. 668 */ 669 if (m->flags & PG_WRITEABLE) 670 vm_object_set_writeable_dirty(object); 671 } 672 673 /* 674 * vm_page_remove: 675 * NOTE: used by device pager as well -wfj 676 * 677 * Removes the given mem entry from the object/offset-page 678 * table and the object page list, but do not invalidate/terminate 679 * the backing store. 680 * 681 * The object and page must be locked. 682 * The underlying pmap entry (if any) is NOT removed here. 683 * This routine may not block. 684 */ 685 void 686 vm_page_remove(vm_page_t m) 687 { 688 vm_object_t object; 689 vm_page_t root; 690 691 if ((object = m->object) == NULL) 692 return; 693 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 694 if (m->oflags & VPO_BUSY) { 695 m->oflags &= ~VPO_BUSY; 696 vm_page_flash(m); 697 } 698 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 699 700 /* 701 * Now remove from the object's list of backed pages. 702 */ 703 if (m != object->root) 704 vm_page_splay(m->pindex, object->root); 705 if (m->left == NULL) 706 root = m->right; 707 else { 708 root = vm_page_splay(m->pindex, m->left); 709 root->right = m->right; 710 } 711 object->root = root; 712 TAILQ_REMOVE(&object->memq, m, listq); 713 714 /* 715 * And show that the object has one fewer resident page. 716 */ 717 object->resident_page_count--; 718 object->generation++; 719 /* 720 * The vnode may now be recycled. 721 */ 722 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 723 vdrop((struct vnode *)object->handle); 724 725 m->object = NULL; 726 } 727 728 /* 729 * vm_page_lookup: 730 * 731 * Returns the page associated with the object/offset 732 * pair specified; if none is found, NULL is returned. 733 * 734 * The object must be locked. 735 * This routine may not block. 736 * This is a critical path routine 737 */ 738 vm_page_t 739 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 740 { 741 vm_page_t m; 742 743 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 744 if ((m = object->root) != NULL && m->pindex != pindex) { 745 m = vm_page_splay(pindex, m); 746 if ((object->root = m)->pindex != pindex) 747 m = NULL; 748 } 749 return (m); 750 } 751 752 /* 753 * vm_page_rename: 754 * 755 * Move the given memory entry from its 756 * current object to the specified target object/offset. 757 * 758 * The object must be locked. 759 * This routine may not block. 760 * 761 * Note: swap associated with the page must be invalidated by the move. We 762 * have to do this for several reasons: (1) we aren't freeing the 763 * page, (2) we are dirtying the page, (3) the VM system is probably 764 * moving the page from object A to B, and will then later move 765 * the backing store from A to B and we can't have a conflict. 766 * 767 * Note: we *always* dirty the page. It is necessary both for the 768 * fact that we moved it, and because we may be invalidating 769 * swap. If the page is on the cache, we have to deactivate it 770 * or vm_page_dirty() will panic. Dirty pages are not allowed 771 * on the cache. 772 */ 773 void 774 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 775 { 776 777 vm_page_remove(m); 778 vm_page_insert(m, new_object, new_pindex); 779 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 780 vm_page_deactivate(m); 781 vm_page_dirty(m); 782 } 783 784 /* 785 * vm_page_select_cache: 786 * 787 * Move a page of the given color from the cache queue to the free 788 * queue. As pages might be found, but are not applicable, they are 789 * deactivated. 790 * 791 * This routine may not block. 792 */ 793 vm_page_t 794 vm_page_select_cache(int color) 795 { 796 vm_object_t object; 797 vm_page_t m; 798 boolean_t was_trylocked; 799 800 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 801 while ((m = vm_pageq_find(PQ_CACHE, color, FALSE)) != NULL) { 802 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 803 KASSERT(!pmap_page_is_mapped(m), 804 ("Found mapped cache page %p", m)); 805 KASSERT((m->flags & PG_UNMANAGED) == 0, 806 ("Found unmanaged cache page %p", m)); 807 KASSERT(m->wire_count == 0, ("Found wired cache page %p", m)); 808 if (m->hold_count == 0 && (object = m->object, 809 (was_trylocked = VM_OBJECT_TRYLOCK(object)) || 810 VM_OBJECT_LOCKED(object))) { 811 KASSERT((m->oflags & VPO_BUSY) == 0 && m->busy == 0, 812 ("Found busy cache page %p", m)); 813 vm_page_free(m); 814 if (was_trylocked) 815 VM_OBJECT_UNLOCK(object); 816 break; 817 } 818 vm_page_deactivate(m); 819 } 820 return (m); 821 } 822 823 /* 824 * vm_page_alloc: 825 * 826 * Allocate and return a memory cell associated 827 * with this VM object/offset pair. 828 * 829 * page_req classes: 830 * VM_ALLOC_NORMAL normal process request 831 * VM_ALLOC_SYSTEM system *really* needs a page 832 * VM_ALLOC_INTERRUPT interrupt time request 833 * VM_ALLOC_ZERO zero page 834 * 835 * This routine may not block. 836 * 837 * Additional special handling is required when called from an 838 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 839 * the page cache in this case. 840 */ 841 vm_page_t 842 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 843 { 844 vm_page_t m = NULL; 845 int color, flags, page_req; 846 847 page_req = req & VM_ALLOC_CLASS_MASK; 848 KASSERT(curthread->td_intr_nesting_level == 0 || 849 page_req == VM_ALLOC_INTERRUPT, 850 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 851 852 if ((req & VM_ALLOC_NOOBJ) == 0) { 853 KASSERT(object != NULL, 854 ("vm_page_alloc: NULL object.")); 855 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 856 color = (pindex + object->pg_color) & PQ_COLORMASK; 857 } else 858 color = pindex & PQ_COLORMASK; 859 860 /* 861 * The pager is allowed to eat deeper into the free page list. 862 */ 863 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 864 page_req = VM_ALLOC_SYSTEM; 865 }; 866 867 loop: 868 mtx_lock(&vm_page_queue_free_mtx); 869 if (cnt.v_free_count > cnt.v_free_reserved || 870 (page_req == VM_ALLOC_SYSTEM && 871 cnt.v_cache_count == 0 && 872 cnt.v_free_count > cnt.v_interrupt_free_min) || 873 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 874 /* 875 * Allocate from the free queue if the number of free pages 876 * exceeds the minimum for the request class. 877 */ 878 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 879 } else if (page_req != VM_ALLOC_INTERRUPT) { 880 mtx_unlock(&vm_page_queue_free_mtx); 881 /* 882 * Allocatable from cache (non-interrupt only). On success, 883 * we must free the page and try again, thus ensuring that 884 * cnt.v_*_free_min counters are replenished. 885 */ 886 vm_page_lock_queues(); 887 if ((m = vm_page_select_cache(color)) == NULL) { 888 KASSERT(cnt.v_cache_count == 0, 889 ("vm_page_alloc: cache queue is missing %d pages", 890 cnt.v_cache_count)); 891 vm_page_unlock_queues(); 892 atomic_add_int(&vm_pageout_deficit, 1); 893 pagedaemon_wakeup(); 894 895 if (page_req != VM_ALLOC_SYSTEM) 896 return (NULL); 897 898 mtx_lock(&vm_page_queue_free_mtx); 899 if (cnt.v_free_count <= cnt.v_interrupt_free_min) { 900 mtx_unlock(&vm_page_queue_free_mtx); 901 return (NULL); 902 } 903 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 904 } else { 905 vm_page_unlock_queues(); 906 goto loop; 907 } 908 } else { 909 /* 910 * Not allocatable from cache from interrupt, give up. 911 */ 912 mtx_unlock(&vm_page_queue_free_mtx); 913 atomic_add_int(&vm_pageout_deficit, 1); 914 pagedaemon_wakeup(); 915 return (NULL); 916 } 917 918 /* 919 * At this point we had better have found a good page. 920 */ 921 922 KASSERT( 923 m != NULL, 924 ("vm_page_alloc(): missing page on free queue") 925 ); 926 927 /* 928 * Remove from free queue 929 */ 930 vm_pageq_remove_nowakeup(m); 931 932 /* 933 * Initialize structure. Only the PG_ZERO flag is inherited. 934 */ 935 flags = 0; 936 if (m->flags & PG_ZERO) { 937 vm_page_zero_count--; 938 if (req & VM_ALLOC_ZERO) 939 flags = PG_ZERO; 940 } 941 if (object != NULL && object->type == OBJT_PHYS) 942 flags |= PG_UNMANAGED; 943 m->flags = flags; 944 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 945 m->oflags = 0; 946 else 947 m->oflags = VPO_BUSY; 948 if (req & VM_ALLOC_WIRED) { 949 atomic_add_int(&cnt.v_wire_count, 1); 950 m->wire_count = 1; 951 } else 952 m->wire_count = 0; 953 m->hold_count = 0; 954 m->act_count = 0; 955 m->busy = 0; 956 m->valid = 0; 957 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 958 mtx_unlock(&vm_page_queue_free_mtx); 959 960 if ((req & VM_ALLOC_NOOBJ) == 0) 961 vm_page_insert(m, object, pindex); 962 else 963 m->pindex = pindex; 964 965 /* 966 * Don't wakeup too often - wakeup the pageout daemon when 967 * we would be nearly out of memory. 968 */ 969 if (vm_paging_needed()) 970 pagedaemon_wakeup(); 971 972 return (m); 973 } 974 975 /* 976 * vm_wait: (also see VM_WAIT macro) 977 * 978 * Block until free pages are available for allocation 979 * - Called in various places before memory allocations. 980 */ 981 void 982 vm_wait(void) 983 { 984 985 mtx_lock(&vm_page_queue_free_mtx); 986 if (curproc == pageproc) { 987 vm_pageout_pages_needed = 1; 988 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx, 989 PDROP | PSWP, "VMWait", 0); 990 } else { 991 if (!vm_pages_needed) { 992 vm_pages_needed = 1; 993 wakeup(&vm_pages_needed); 994 } 995 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM, 996 "vmwait", 0); 997 } 998 } 999 1000 /* 1001 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1002 * 1003 * Block until free pages are available for allocation 1004 * - Called only in vm_fault so that processes page faulting 1005 * can be easily tracked. 1006 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1007 * processes will be able to grab memory first. Do not change 1008 * this balance without careful testing first. 1009 */ 1010 void 1011 vm_waitpfault(void) 1012 { 1013 1014 mtx_lock(&vm_page_queue_free_mtx); 1015 if (!vm_pages_needed) { 1016 vm_pages_needed = 1; 1017 wakeup(&vm_pages_needed); 1018 } 1019 msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER, 1020 "pfault", 0); 1021 } 1022 1023 /* 1024 * vm_page_activate: 1025 * 1026 * Put the specified page on the active list (if appropriate). 1027 * Ensure that act_count is at least ACT_INIT but do not otherwise 1028 * mess with it. 1029 * 1030 * The page queues must be locked. 1031 * This routine may not block. 1032 */ 1033 void 1034 vm_page_activate(vm_page_t m) 1035 { 1036 1037 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1038 if (VM_PAGE_GETKNOWNQUEUE2(m) != PQ_ACTIVE) { 1039 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1040 cnt.v_reactivated++; 1041 vm_pageq_remove(m); 1042 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1043 if (m->act_count < ACT_INIT) 1044 m->act_count = ACT_INIT; 1045 vm_pageq_enqueue(PQ_ACTIVE, m); 1046 } 1047 } else { 1048 if (m->act_count < ACT_INIT) 1049 m->act_count = ACT_INIT; 1050 } 1051 } 1052 1053 /* 1054 * vm_page_free_wakeup: 1055 * 1056 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1057 * routine is called when a page has been added to the cache or free 1058 * queues. 1059 * 1060 * The page queues must be locked. 1061 * This routine may not block. 1062 */ 1063 static inline void 1064 vm_page_free_wakeup(void) 1065 { 1066 1067 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); 1068 /* 1069 * if pageout daemon needs pages, then tell it that there are 1070 * some free. 1071 */ 1072 if (vm_pageout_pages_needed && 1073 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1074 wakeup(&vm_pageout_pages_needed); 1075 vm_pageout_pages_needed = 0; 1076 } 1077 /* 1078 * wakeup processes that are waiting on memory if we hit a 1079 * high water mark. And wakeup scheduler process if we have 1080 * lots of memory. this process will swapin processes. 1081 */ 1082 if (vm_pages_needed && !vm_page_count_min()) { 1083 vm_pages_needed = 0; 1084 wakeup(&cnt.v_free_count); 1085 } 1086 } 1087 1088 /* 1089 * vm_page_free_toq: 1090 * 1091 * Returns the given page to the PQ_FREE list, 1092 * disassociating it with any VM object. 1093 * 1094 * Object and page must be locked prior to entry. 1095 * This routine may not block. 1096 */ 1097 1098 void 1099 vm_page_free_toq(vm_page_t m) 1100 { 1101 struct vpgqueues *pq; 1102 1103 if (VM_PAGE_GETQUEUE(m) != PQ_NONE) 1104 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1105 KASSERT(!pmap_page_is_mapped(m), 1106 ("vm_page_free_toq: freeing mapped page %p", m)); 1107 cnt.v_tfree++; 1108 1109 if (m->busy || VM_PAGE_INQUEUE1(m, PQ_FREE)) { 1110 printf( 1111 "vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n", 1112 (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0, 1113 m->hold_count); 1114 if (VM_PAGE_INQUEUE1(m, PQ_FREE)) 1115 panic("vm_page_free: freeing free page"); 1116 else 1117 panic("vm_page_free: freeing busy page"); 1118 } 1119 1120 /* 1121 * unqueue, then remove page. Note that we cannot destroy 1122 * the page here because we do not want to call the pager's 1123 * callback routine until after we've put the page on the 1124 * appropriate free queue. 1125 */ 1126 vm_pageq_remove_nowakeup(m); 1127 vm_page_remove(m); 1128 1129 /* 1130 * If fictitious remove object association and 1131 * return, otherwise delay object association removal. 1132 */ 1133 if ((m->flags & PG_FICTITIOUS) != 0) { 1134 return; 1135 } 1136 1137 m->valid = 0; 1138 vm_page_undirty(m); 1139 1140 if (m->wire_count != 0) { 1141 if (m->wire_count > 1) { 1142 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1143 m->wire_count, (long)m->pindex); 1144 } 1145 panic("vm_page_free: freeing wired page"); 1146 } 1147 if (m->hold_count != 0) { 1148 m->flags &= ~PG_ZERO; 1149 vm_pageq_enqueue(PQ_HOLD, m); 1150 return; 1151 } 1152 VM_PAGE_SETQUEUE1(m, PQ_FREE); 1153 mtx_lock(&vm_page_queue_free_mtx); 1154 pq = &vm_page_queues[VM_PAGE_GETQUEUE(m)]; 1155 pq->lcnt++; 1156 ++(*pq->cnt); 1157 1158 /* 1159 * Put zero'd pages on the end ( where we look for zero'd pages 1160 * first ) and non-zerod pages at the head. 1161 */ 1162 if (m->flags & PG_ZERO) { 1163 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1164 ++vm_page_zero_count; 1165 } else { 1166 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1167 vm_page_zero_idle_wakeup(); 1168 } 1169 vm_page_free_wakeup(); 1170 mtx_unlock(&vm_page_queue_free_mtx); 1171 } 1172 1173 /* 1174 * vm_page_wire: 1175 * 1176 * Mark this page as wired down by yet 1177 * another map, removing it from paging queues 1178 * as necessary. 1179 * 1180 * The page queues must be locked. 1181 * This routine may not block. 1182 */ 1183 void 1184 vm_page_wire(vm_page_t m) 1185 { 1186 1187 /* 1188 * Only bump the wire statistics if the page is not already wired, 1189 * and only unqueue the page if it is on some queue (if it is unmanaged 1190 * it is already off the queues). 1191 */ 1192 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1193 if (m->flags & PG_FICTITIOUS) 1194 return; 1195 if (m->wire_count == 0) { 1196 if ((m->flags & PG_UNMANAGED) == 0) 1197 vm_pageq_remove(m); 1198 atomic_add_int(&cnt.v_wire_count, 1); 1199 } 1200 m->wire_count++; 1201 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1202 } 1203 1204 /* 1205 * vm_page_unwire: 1206 * 1207 * Release one wiring of this page, potentially 1208 * enabling it to be paged again. 1209 * 1210 * Many pages placed on the inactive queue should actually go 1211 * into the cache, but it is difficult to figure out which. What 1212 * we do instead, if the inactive target is well met, is to put 1213 * clean pages at the head of the inactive queue instead of the tail. 1214 * This will cause them to be moved to the cache more quickly and 1215 * if not actively re-referenced, freed more quickly. If we just 1216 * stick these pages at the end of the inactive queue, heavy filesystem 1217 * meta-data accesses can cause an unnecessary paging load on memory bound 1218 * processes. This optimization causes one-time-use metadata to be 1219 * reused more quickly. 1220 * 1221 * BUT, if we are in a low-memory situation we have no choice but to 1222 * put clean pages on the cache queue. 1223 * 1224 * A number of routines use vm_page_unwire() to guarantee that the page 1225 * will go into either the inactive or active queues, and will NEVER 1226 * be placed in the cache - for example, just after dirtying a page. 1227 * dirty pages in the cache are not allowed. 1228 * 1229 * The page queues must be locked. 1230 * This routine may not block. 1231 */ 1232 void 1233 vm_page_unwire(vm_page_t m, int activate) 1234 { 1235 1236 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1237 if (m->flags & PG_FICTITIOUS) 1238 return; 1239 if (m->wire_count > 0) { 1240 m->wire_count--; 1241 if (m->wire_count == 0) { 1242 atomic_subtract_int(&cnt.v_wire_count, 1); 1243 if (m->flags & PG_UNMANAGED) { 1244 ; 1245 } else if (activate) 1246 vm_pageq_enqueue(PQ_ACTIVE, m); 1247 else { 1248 vm_page_flag_clear(m, PG_WINATCFLS); 1249 vm_pageq_enqueue(PQ_INACTIVE, m); 1250 } 1251 } 1252 } else { 1253 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1254 } 1255 } 1256 1257 1258 /* 1259 * Move the specified page to the inactive queue. If the page has 1260 * any associated swap, the swap is deallocated. 1261 * 1262 * Normally athead is 0 resulting in LRU operation. athead is set 1263 * to 1 if we want this page to be 'as if it were placed in the cache', 1264 * except without unmapping it from the process address space. 1265 * 1266 * This routine may not block. 1267 */ 1268 static inline void 1269 _vm_page_deactivate(vm_page_t m, int athead) 1270 { 1271 1272 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1273 1274 /* 1275 * Ignore if already inactive. 1276 */ 1277 if (VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) 1278 return; 1279 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1280 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1281 cnt.v_reactivated++; 1282 vm_page_flag_clear(m, PG_WINATCFLS); 1283 vm_pageq_remove(m); 1284 if (athead) 1285 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1286 else 1287 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1288 VM_PAGE_SETQUEUE2(m, PQ_INACTIVE); 1289 vm_page_queues[PQ_INACTIVE].lcnt++; 1290 cnt.v_inactive_count++; 1291 } 1292 } 1293 1294 void 1295 vm_page_deactivate(vm_page_t m) 1296 { 1297 _vm_page_deactivate(m, 0); 1298 } 1299 1300 /* 1301 * vm_page_try_to_cache: 1302 * 1303 * Returns 0 on failure, 1 on success 1304 */ 1305 int 1306 vm_page_try_to_cache(vm_page_t m) 1307 { 1308 1309 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1310 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1311 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1312 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) { 1313 return (0); 1314 } 1315 pmap_remove_all(m); 1316 if (m->dirty) 1317 return (0); 1318 vm_page_cache(m); 1319 return (1); 1320 } 1321 1322 /* 1323 * vm_page_try_to_free() 1324 * 1325 * Attempt to free the page. If we cannot free it, we do nothing. 1326 * 1 is returned on success, 0 on failure. 1327 */ 1328 int 1329 vm_page_try_to_free(vm_page_t m) 1330 { 1331 1332 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1333 if (m->object != NULL) 1334 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1335 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1336 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) { 1337 return (0); 1338 } 1339 pmap_remove_all(m); 1340 if (m->dirty) 1341 return (0); 1342 vm_page_free(m); 1343 return (1); 1344 } 1345 1346 /* 1347 * vm_page_cache 1348 * 1349 * Put the specified page onto the page cache queue (if appropriate). 1350 * 1351 * This routine may not block. 1352 */ 1353 void 1354 vm_page_cache(vm_page_t m) 1355 { 1356 1357 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1358 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1359 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy || 1360 m->hold_count || m->wire_count) { 1361 printf("vm_page_cache: attempting to cache busy page\n"); 1362 return; 1363 } 1364 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1365 return; 1366 1367 /* 1368 * Remove all pmaps and indicate that the page is not 1369 * writeable or mapped. 1370 */ 1371 pmap_remove_all(m); 1372 if (m->dirty != 0) { 1373 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1374 (long)m->pindex); 1375 } 1376 vm_pageq_remove_nowakeup(m); 1377 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1378 mtx_lock(&vm_page_queue_free_mtx); 1379 vm_page_free_wakeup(); 1380 mtx_unlock(&vm_page_queue_free_mtx); 1381 } 1382 1383 /* 1384 * vm_page_dontneed 1385 * 1386 * Cache, deactivate, or do nothing as appropriate. This routine 1387 * is typically used by madvise() MADV_DONTNEED. 1388 * 1389 * Generally speaking we want to move the page into the cache so 1390 * it gets reused quickly. However, this can result in a silly syndrome 1391 * due to the page recycling too quickly. Small objects will not be 1392 * fully cached. On the otherhand, if we move the page to the inactive 1393 * queue we wind up with a problem whereby very large objects 1394 * unnecessarily blow away our inactive and cache queues. 1395 * 1396 * The solution is to move the pages based on a fixed weighting. We 1397 * either leave them alone, deactivate them, or move them to the cache, 1398 * where moving them to the cache has the highest weighting. 1399 * By forcing some pages into other queues we eventually force the 1400 * system to balance the queues, potentially recovering other unrelated 1401 * space from active. The idea is to not force this to happen too 1402 * often. 1403 */ 1404 void 1405 vm_page_dontneed(vm_page_t m) 1406 { 1407 static int dnweight; 1408 int dnw; 1409 int head; 1410 1411 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1412 dnw = ++dnweight; 1413 1414 /* 1415 * occassionally leave the page alone 1416 */ 1417 if ((dnw & 0x01F0) == 0 || 1418 VM_PAGE_INQUEUE2(m, PQ_INACTIVE) || 1419 VM_PAGE_INQUEUE1(m, PQ_CACHE) 1420 ) { 1421 if (m->act_count >= ACT_INIT) 1422 --m->act_count; 1423 return; 1424 } 1425 1426 if (m->dirty == 0 && pmap_is_modified(m)) 1427 vm_page_dirty(m); 1428 1429 if (m->dirty || (dnw & 0x0070) == 0) { 1430 /* 1431 * Deactivate the page 3 times out of 32. 1432 */ 1433 head = 0; 1434 } else { 1435 /* 1436 * Cache the page 28 times out of every 32. Note that 1437 * the page is deactivated instead of cached, but placed 1438 * at the head of the queue instead of the tail. 1439 */ 1440 head = 1; 1441 } 1442 _vm_page_deactivate(m, head); 1443 } 1444 1445 /* 1446 * Grab a page, waiting until we are waken up due to the page 1447 * changing state. We keep on waiting, if the page continues 1448 * to be in the object. If the page doesn't exist, first allocate it 1449 * and then conditionally zero it. 1450 * 1451 * This routine may block. 1452 */ 1453 vm_page_t 1454 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1455 { 1456 vm_page_t m; 1457 1458 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1459 retrylookup: 1460 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1461 if (vm_page_sleep_if_busy(m, TRUE, "pgrbwt")) { 1462 if ((allocflags & VM_ALLOC_RETRY) == 0) 1463 return (NULL); 1464 goto retrylookup; 1465 } else { 1466 if ((allocflags & VM_ALLOC_WIRED) != 0) { 1467 vm_page_lock_queues(); 1468 vm_page_wire(m); 1469 vm_page_unlock_queues(); 1470 } 1471 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 1472 vm_page_busy(m); 1473 return (m); 1474 } 1475 } 1476 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1477 if (m == NULL) { 1478 VM_OBJECT_UNLOCK(object); 1479 VM_WAIT; 1480 VM_OBJECT_LOCK(object); 1481 if ((allocflags & VM_ALLOC_RETRY) == 0) 1482 return (NULL); 1483 goto retrylookup; 1484 } 1485 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 1486 pmap_zero_page(m); 1487 return (m); 1488 } 1489 1490 /* 1491 * Mapping function for valid bits or for dirty bits in 1492 * a page. May not block. 1493 * 1494 * Inputs are required to range within a page. 1495 */ 1496 inline int 1497 vm_page_bits(int base, int size) 1498 { 1499 int first_bit; 1500 int last_bit; 1501 1502 KASSERT( 1503 base + size <= PAGE_SIZE, 1504 ("vm_page_bits: illegal base/size %d/%d", base, size) 1505 ); 1506 1507 if (size == 0) /* handle degenerate case */ 1508 return (0); 1509 1510 first_bit = base >> DEV_BSHIFT; 1511 last_bit = (base + size - 1) >> DEV_BSHIFT; 1512 1513 return ((2 << last_bit) - (1 << first_bit)); 1514 } 1515 1516 /* 1517 * vm_page_set_validclean: 1518 * 1519 * Sets portions of a page valid and clean. The arguments are expected 1520 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1521 * of any partial chunks touched by the range. The invalid portion of 1522 * such chunks will be zero'd. 1523 * 1524 * This routine may not block. 1525 * 1526 * (base + size) must be less then or equal to PAGE_SIZE. 1527 */ 1528 void 1529 vm_page_set_validclean(vm_page_t m, int base, int size) 1530 { 1531 int pagebits; 1532 int frag; 1533 int endoff; 1534 1535 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1536 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1537 if (size == 0) /* handle degenerate case */ 1538 return; 1539 1540 /* 1541 * If the base is not DEV_BSIZE aligned and the valid 1542 * bit is clear, we have to zero out a portion of the 1543 * first block. 1544 */ 1545 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1546 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1547 pmap_zero_page_area(m, frag, base - frag); 1548 1549 /* 1550 * If the ending offset is not DEV_BSIZE aligned and the 1551 * valid bit is clear, we have to zero out a portion of 1552 * the last block. 1553 */ 1554 endoff = base + size; 1555 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1556 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1557 pmap_zero_page_area(m, endoff, 1558 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1559 1560 /* 1561 * Set valid, clear dirty bits. If validating the entire 1562 * page we can safely clear the pmap modify bit. We also 1563 * use this opportunity to clear the VPO_NOSYNC flag. If a process 1564 * takes a write fault on a MAP_NOSYNC memory area the flag will 1565 * be set again. 1566 * 1567 * We set valid bits inclusive of any overlap, but we can only 1568 * clear dirty bits for DEV_BSIZE chunks that are fully within 1569 * the range. 1570 */ 1571 pagebits = vm_page_bits(base, size); 1572 m->valid |= pagebits; 1573 #if 0 /* NOT YET */ 1574 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1575 frag = DEV_BSIZE - frag; 1576 base += frag; 1577 size -= frag; 1578 if (size < 0) 1579 size = 0; 1580 } 1581 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1582 #endif 1583 m->dirty &= ~pagebits; 1584 if (base == 0 && size == PAGE_SIZE) { 1585 pmap_clear_modify(m); 1586 m->oflags &= ~VPO_NOSYNC; 1587 } 1588 } 1589 1590 void 1591 vm_page_clear_dirty(vm_page_t m, int base, int size) 1592 { 1593 1594 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1595 m->dirty &= ~vm_page_bits(base, size); 1596 } 1597 1598 /* 1599 * vm_page_set_invalid: 1600 * 1601 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1602 * valid and dirty bits for the effected areas are cleared. 1603 * 1604 * May not block. 1605 */ 1606 void 1607 vm_page_set_invalid(vm_page_t m, int base, int size) 1608 { 1609 int bits; 1610 1611 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1612 bits = vm_page_bits(base, size); 1613 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1614 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 1615 pmap_remove_all(m); 1616 m->valid &= ~bits; 1617 m->dirty &= ~bits; 1618 m->object->generation++; 1619 } 1620 1621 /* 1622 * vm_page_zero_invalid() 1623 * 1624 * The kernel assumes that the invalid portions of a page contain 1625 * garbage, but such pages can be mapped into memory by user code. 1626 * When this occurs, we must zero out the non-valid portions of the 1627 * page so user code sees what it expects. 1628 * 1629 * Pages are most often semi-valid when the end of a file is mapped 1630 * into memory and the file's size is not page aligned. 1631 */ 1632 void 1633 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1634 { 1635 int b; 1636 int i; 1637 1638 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1639 /* 1640 * Scan the valid bits looking for invalid sections that 1641 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1642 * valid bit may be set ) have already been zerod by 1643 * vm_page_set_validclean(). 1644 */ 1645 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1646 if (i == (PAGE_SIZE / DEV_BSIZE) || 1647 (m->valid & (1 << i)) 1648 ) { 1649 if (i > b) { 1650 pmap_zero_page_area(m, 1651 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1652 } 1653 b = i + 1; 1654 } 1655 } 1656 1657 /* 1658 * setvalid is TRUE when we can safely set the zero'd areas 1659 * as being valid. We can do this if there are no cache consistancy 1660 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1661 */ 1662 if (setvalid) 1663 m->valid = VM_PAGE_BITS_ALL; 1664 } 1665 1666 /* 1667 * vm_page_is_valid: 1668 * 1669 * Is (partial) page valid? Note that the case where size == 0 1670 * will return FALSE in the degenerate case where the page is 1671 * entirely invalid, and TRUE otherwise. 1672 * 1673 * May not block. 1674 */ 1675 int 1676 vm_page_is_valid(vm_page_t m, int base, int size) 1677 { 1678 int bits = vm_page_bits(base, size); 1679 1680 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1681 if (m->valid && ((m->valid & bits) == bits)) 1682 return 1; 1683 else 1684 return 0; 1685 } 1686 1687 /* 1688 * update dirty bits from pmap/mmu. May not block. 1689 */ 1690 void 1691 vm_page_test_dirty(vm_page_t m) 1692 { 1693 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1694 vm_page_dirty(m); 1695 } 1696 } 1697 1698 int so_zerocp_fullpage = 0; 1699 1700 void 1701 vm_page_cowfault(vm_page_t m) 1702 { 1703 vm_page_t mnew; 1704 vm_object_t object; 1705 vm_pindex_t pindex; 1706 1707 object = m->object; 1708 pindex = m->pindex; 1709 1710 retry_alloc: 1711 pmap_remove_all(m); 1712 vm_page_remove(m); 1713 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 1714 if (mnew == NULL) { 1715 vm_page_insert(m, object, pindex); 1716 vm_page_unlock_queues(); 1717 VM_OBJECT_UNLOCK(object); 1718 VM_WAIT; 1719 VM_OBJECT_LOCK(object); 1720 vm_page_lock_queues(); 1721 goto retry_alloc; 1722 } 1723 1724 if (m->cow == 0) { 1725 /* 1726 * check to see if we raced with an xmit complete when 1727 * waiting to allocate a page. If so, put things back 1728 * the way they were 1729 */ 1730 vm_page_free(mnew); 1731 vm_page_insert(m, object, pindex); 1732 } else { /* clear COW & copy page */ 1733 if (!so_zerocp_fullpage) 1734 pmap_copy_page(m, mnew); 1735 mnew->valid = VM_PAGE_BITS_ALL; 1736 vm_page_dirty(mnew); 1737 mnew->wire_count = m->wire_count - m->cow; 1738 m->wire_count = m->cow; 1739 } 1740 } 1741 1742 void 1743 vm_page_cowclear(vm_page_t m) 1744 { 1745 1746 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1747 if (m->cow) { 1748 m->cow--; 1749 /* 1750 * let vm_fault add back write permission lazily 1751 */ 1752 } 1753 /* 1754 * sf_buf_free() will free the page, so we needn't do it here 1755 */ 1756 } 1757 1758 void 1759 vm_page_cowsetup(vm_page_t m) 1760 { 1761 1762 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1763 m->cow++; 1764 pmap_remove_write(m); 1765 } 1766 1767 #include "opt_ddb.h" 1768 #ifdef DDB 1769 #include <sys/kernel.h> 1770 1771 #include <ddb/ddb.h> 1772 1773 DB_SHOW_COMMAND(page, vm_page_print_page_info) 1774 { 1775 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1776 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1777 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1778 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1779 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1780 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1781 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1782 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1783 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1784 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1785 } 1786 1787 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1788 { 1789 int i; 1790 db_printf("PQ_FREE:"); 1791 for (i = 0; i < PQ_NUMCOLORS; i++) { 1792 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1793 } 1794 db_printf("\n"); 1795 1796 db_printf("PQ_CACHE:"); 1797 for (i = 0; i < PQ_NUMCOLORS; i++) { 1798 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1799 } 1800 db_printf("\n"); 1801 1802 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1803 vm_page_queues[PQ_ACTIVE].lcnt, 1804 vm_page_queues[PQ_INACTIVE].lcnt); 1805 } 1806 #endif /* DDB */ 1807