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 * The clearing of PG_ZERO is a temporary safety until the code can be 480 * reviewed to determine that PG_ZERO is being properly cleared on 481 * write faults or maps. PG_ZERO was previously cleared in 482 * vm_page_alloc(). 483 */ 484 void 485 vm_page_free(vm_page_t m) 486 { 487 vm_page_flag_clear(m, PG_ZERO); 488 vm_page_free_toq(m); 489 vm_page_zero_idle_wakeup(); 490 } 491 492 /* 493 * vm_page_free_zero: 494 * 495 * Free a page to the zerod-pages queue 496 */ 497 void 498 vm_page_free_zero(vm_page_t m) 499 { 500 vm_page_flag_set(m, PG_ZERO); 501 vm_page_free_toq(m); 502 } 503 504 /* 505 * vm_page_sleep: 506 * 507 * Sleep and release the page queues lock. 508 * 509 * The object containing the given page must be locked. 510 */ 511 void 512 vm_page_sleep(vm_page_t m, const char *msg) 513 { 514 515 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 516 if (!mtx_owned(&vm_page_queue_mtx)) 517 vm_page_lock_queues(); 518 vm_page_flag_set(m, PG_REFERENCED); 519 vm_page_unlock_queues(); 520 521 /* 522 * It's possible that while we sleep, the page will get 523 * unbusied and freed. If we are holding the object 524 * lock, we will assume we hold a reference to the object 525 * such that even if m->object changes, we can re-lock 526 * it. 527 */ 528 m->oflags |= VPO_WANTED; 529 msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0); 530 } 531 532 /* 533 * vm_page_dirty: 534 * 535 * make page all dirty 536 */ 537 void 538 vm_page_dirty(vm_page_t m) 539 { 540 KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_CACHE, 541 ("vm_page_dirty: page in cache!")); 542 KASSERT(VM_PAGE_GETKNOWNQUEUE1(m) != PQ_FREE, 543 ("vm_page_dirty: page is free!")); 544 m->dirty = VM_PAGE_BITS_ALL; 545 } 546 547 /* 548 * vm_page_splay: 549 * 550 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 551 * the vm_page containing the given pindex. If, however, that 552 * pindex is not found in the vm_object, returns a vm_page that is 553 * adjacent to the pindex, coming before or after it. 554 */ 555 vm_page_t 556 vm_page_splay(vm_pindex_t pindex, vm_page_t root) 557 { 558 struct vm_page dummy; 559 vm_page_t lefttreemax, righttreemin, y; 560 561 if (root == NULL) 562 return (root); 563 lefttreemax = righttreemin = &dummy; 564 for (;; root = y) { 565 if (pindex < root->pindex) { 566 if ((y = root->left) == NULL) 567 break; 568 if (pindex < y->pindex) { 569 /* Rotate right. */ 570 root->left = y->right; 571 y->right = root; 572 root = y; 573 if ((y = root->left) == NULL) 574 break; 575 } 576 /* Link into the new root's right tree. */ 577 righttreemin->left = root; 578 righttreemin = root; 579 } else if (pindex > root->pindex) { 580 if ((y = root->right) == NULL) 581 break; 582 if (pindex > y->pindex) { 583 /* Rotate left. */ 584 root->right = y->left; 585 y->left = root; 586 root = y; 587 if ((y = root->right) == NULL) 588 break; 589 } 590 /* Link into the new root's left tree. */ 591 lefttreemax->right = root; 592 lefttreemax = root; 593 } else 594 break; 595 } 596 /* Assemble the new root. */ 597 lefttreemax->right = root->left; 598 righttreemin->left = root->right; 599 root->left = dummy.right; 600 root->right = dummy.left; 601 return (root); 602 } 603 604 /* 605 * vm_page_insert: [ internal use only ] 606 * 607 * Inserts the given mem entry into the object and object list. 608 * 609 * The pagetables are not updated but will presumably fault the page 610 * in if necessary, or if a kernel page the caller will at some point 611 * enter the page into the kernel's pmap. We are not allowed to block 612 * here so we *can't* do this anyway. 613 * 614 * The object and page must be locked. 615 * This routine may not block. 616 */ 617 void 618 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 619 { 620 vm_page_t root; 621 622 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 623 if (m->object != NULL) 624 panic("vm_page_insert: page already inserted"); 625 626 /* 627 * Record the object/offset pair in this page 628 */ 629 m->object = object; 630 m->pindex = pindex; 631 632 /* 633 * Now link into the object's ordered list of backed pages. 634 */ 635 root = object->root; 636 if (root == NULL) { 637 m->left = NULL; 638 m->right = NULL; 639 TAILQ_INSERT_TAIL(&object->memq, m, listq); 640 } else { 641 root = vm_page_splay(pindex, root); 642 if (pindex < root->pindex) { 643 m->left = root->left; 644 m->right = root; 645 root->left = NULL; 646 TAILQ_INSERT_BEFORE(root, m, listq); 647 } else if (pindex == root->pindex) 648 panic("vm_page_insert: offset already allocated"); 649 else { 650 m->right = root->right; 651 m->left = root; 652 root->right = NULL; 653 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 654 } 655 } 656 object->root = m; 657 object->generation++; 658 659 /* 660 * show that the object has one more resident page. 661 */ 662 object->resident_page_count++; 663 /* 664 * Hold the vnode until the last page is released. 665 */ 666 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 667 vhold((struct vnode *)object->handle); 668 669 /* 670 * Since we are inserting a new and possibly dirty page, 671 * update the object's OBJ_MIGHTBEDIRTY flag. 672 */ 673 if (m->flags & PG_WRITEABLE) 674 vm_object_set_writeable_dirty(object); 675 } 676 677 /* 678 * vm_page_remove: 679 * NOTE: used by device pager as well -wfj 680 * 681 * Removes the given mem entry from the object/offset-page 682 * table and the object page list, but do not invalidate/terminate 683 * the backing store. 684 * 685 * The object and page must be locked. 686 * The underlying pmap entry (if any) is NOT removed here. 687 * This routine may not block. 688 */ 689 void 690 vm_page_remove(vm_page_t m) 691 { 692 vm_object_t object; 693 vm_page_t root; 694 695 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 696 if ((object = m->object) == NULL) 697 return; 698 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 699 if (m->oflags & VPO_BUSY) { 700 m->oflags &= ~VPO_BUSY; 701 vm_page_flash(m); 702 } 703 704 /* 705 * Now remove from the object's list of backed pages. 706 */ 707 if (m != object->root) 708 vm_page_splay(m->pindex, object->root); 709 if (m->left == NULL) 710 root = m->right; 711 else { 712 root = vm_page_splay(m->pindex, m->left); 713 root->right = m->right; 714 } 715 object->root = root; 716 TAILQ_REMOVE(&object->memq, m, listq); 717 718 /* 719 * And show that the object has one fewer resident page. 720 */ 721 object->resident_page_count--; 722 object->generation++; 723 /* 724 * The vnode may now be recycled. 725 */ 726 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 727 vdrop((struct vnode *)object->handle); 728 729 m->object = NULL; 730 } 731 732 /* 733 * vm_page_lookup: 734 * 735 * Returns the page associated with the object/offset 736 * pair specified; if none is found, NULL is returned. 737 * 738 * The object must be locked. 739 * This routine may not block. 740 * This is a critical path routine 741 */ 742 vm_page_t 743 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 744 { 745 vm_page_t m; 746 747 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 748 if ((m = object->root) != NULL && m->pindex != pindex) { 749 m = vm_page_splay(pindex, m); 750 if ((object->root = m)->pindex != pindex) 751 m = NULL; 752 } 753 return (m); 754 } 755 756 /* 757 * vm_page_rename: 758 * 759 * Move the given memory entry from its 760 * current object to the specified target object/offset. 761 * 762 * The object must be locked. 763 * This routine may not block. 764 * 765 * Note: swap associated with the page must be invalidated by the move. We 766 * have to do this for several reasons: (1) we aren't freeing the 767 * page, (2) we are dirtying the page, (3) the VM system is probably 768 * moving the page from object A to B, and will then later move 769 * the backing store from A to B and we can't have a conflict. 770 * 771 * Note: we *always* dirty the page. It is necessary both for the 772 * fact that we moved it, and because we may be invalidating 773 * swap. If the page is on the cache, we have to deactivate it 774 * or vm_page_dirty() will panic. Dirty pages are not allowed 775 * on the cache. 776 */ 777 void 778 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 779 { 780 781 vm_page_remove(m); 782 vm_page_insert(m, new_object, new_pindex); 783 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 784 vm_page_deactivate(m); 785 vm_page_dirty(m); 786 } 787 788 /* 789 * vm_page_select_cache: 790 * 791 * Move a page of the given color from the cache queue to the free 792 * queue. As pages might be found, but are not applicable, they are 793 * deactivated. 794 * 795 * This routine may not block. 796 */ 797 vm_page_t 798 vm_page_select_cache(int color) 799 { 800 vm_object_t object; 801 vm_page_t m; 802 boolean_t was_trylocked; 803 804 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 805 while ((m = vm_pageq_find(PQ_CACHE, color, FALSE)) != NULL) { 806 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 807 KASSERT(!pmap_page_is_mapped(m), 808 ("Found mapped cache page %p", m)); 809 KASSERT((m->flags & PG_UNMANAGED) == 0, 810 ("Found unmanaged cache page %p", m)); 811 KASSERT(m->wire_count == 0, ("Found wired cache page %p", m)); 812 if (m->hold_count == 0 && (object = m->object, 813 (was_trylocked = VM_OBJECT_TRYLOCK(object)) || 814 VM_OBJECT_LOCKED(object))) { 815 KASSERT((m->oflags & VPO_BUSY) == 0 && m->busy == 0, 816 ("Found busy cache page %p", m)); 817 vm_page_free(m); 818 if (was_trylocked) 819 VM_OBJECT_UNLOCK(object); 820 break; 821 } 822 vm_page_deactivate(m); 823 } 824 return (m); 825 } 826 827 /* 828 * vm_page_alloc: 829 * 830 * Allocate and return a memory cell associated 831 * with this VM object/offset pair. 832 * 833 * page_req classes: 834 * VM_ALLOC_NORMAL normal process request 835 * VM_ALLOC_SYSTEM system *really* needs a page 836 * VM_ALLOC_INTERRUPT interrupt time request 837 * VM_ALLOC_ZERO zero page 838 * 839 * This routine may not block. 840 * 841 * Additional special handling is required when called from an 842 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 843 * the page cache in this case. 844 */ 845 vm_page_t 846 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 847 { 848 vm_page_t m = NULL; 849 int color, flags, page_req; 850 851 page_req = req & VM_ALLOC_CLASS_MASK; 852 KASSERT(curthread->td_intr_nesting_level == 0 || 853 page_req == VM_ALLOC_INTERRUPT, 854 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 855 856 if ((req & VM_ALLOC_NOOBJ) == 0) { 857 KASSERT(object != NULL, 858 ("vm_page_alloc: NULL object.")); 859 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 860 color = (pindex + object->pg_color) & PQ_COLORMASK; 861 } else 862 color = pindex & PQ_COLORMASK; 863 864 /* 865 * The pager is allowed to eat deeper into the free page list. 866 */ 867 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 868 page_req = VM_ALLOC_SYSTEM; 869 }; 870 871 loop: 872 mtx_lock(&vm_page_queue_free_mtx); 873 if (cnt.v_free_count > cnt.v_free_reserved || 874 (page_req == VM_ALLOC_SYSTEM && 875 cnt.v_cache_count == 0 && 876 cnt.v_free_count > cnt.v_interrupt_free_min) || 877 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 878 /* 879 * Allocate from the free queue if the number of free pages 880 * exceeds the minimum for the request class. 881 */ 882 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 883 } else if (page_req != VM_ALLOC_INTERRUPT) { 884 mtx_unlock(&vm_page_queue_free_mtx); 885 /* 886 * Allocatable from cache (non-interrupt only). On success, 887 * we must free the page and try again, thus ensuring that 888 * cnt.v_*_free_min counters are replenished. 889 */ 890 vm_page_lock_queues(); 891 if ((m = vm_page_select_cache(color)) == NULL) { 892 KASSERT(cnt.v_cache_count == 0, 893 ("vm_page_alloc: cache queue is missing %d pages", 894 cnt.v_cache_count)); 895 vm_page_unlock_queues(); 896 atomic_add_int(&vm_pageout_deficit, 1); 897 pagedaemon_wakeup(); 898 899 if (page_req != VM_ALLOC_SYSTEM) 900 return (NULL); 901 902 mtx_lock(&vm_page_queue_free_mtx); 903 if (cnt.v_free_count <= cnt.v_interrupt_free_min) { 904 mtx_unlock(&vm_page_queue_free_mtx); 905 return (NULL); 906 } 907 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 908 } else { 909 vm_page_unlock_queues(); 910 goto loop; 911 } 912 } else { 913 /* 914 * Not allocatable from cache from interrupt, give up. 915 */ 916 mtx_unlock(&vm_page_queue_free_mtx); 917 atomic_add_int(&vm_pageout_deficit, 1); 918 pagedaemon_wakeup(); 919 return (NULL); 920 } 921 922 /* 923 * At this point we had better have found a good page. 924 */ 925 926 KASSERT( 927 m != NULL, 928 ("vm_page_alloc(): missing page on free queue") 929 ); 930 931 /* 932 * Remove from free queue 933 */ 934 vm_pageq_remove_nowakeup(m); 935 936 /* 937 * Initialize structure. Only the PG_ZERO flag is inherited. 938 */ 939 flags = 0; 940 if (m->flags & PG_ZERO) { 941 vm_page_zero_count--; 942 if (req & VM_ALLOC_ZERO) 943 flags = PG_ZERO; 944 } 945 m->flags = flags; 946 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 947 m->oflags = 0; 948 else 949 m->oflags = VPO_BUSY; 950 if (req & VM_ALLOC_WIRED) { 951 atomic_add_int(&cnt.v_wire_count, 1); 952 m->wire_count = 1; 953 } else 954 m->wire_count = 0; 955 m->hold_count = 0; 956 m->act_count = 0; 957 m->busy = 0; 958 m->valid = 0; 959 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 960 mtx_unlock(&vm_page_queue_free_mtx); 961 962 if ((req & VM_ALLOC_NOOBJ) == 0) 963 vm_page_insert(m, object, pindex); 964 else 965 m->pindex = pindex; 966 967 /* 968 * Don't wakeup too often - wakeup the pageout daemon when 969 * we would be nearly out of memory. 970 */ 971 if (vm_paging_needed()) 972 pagedaemon_wakeup(); 973 974 return (m); 975 } 976 977 /* 978 * vm_wait: (also see VM_WAIT macro) 979 * 980 * Block until free pages are available for allocation 981 * - Called in various places before memory allocations. 982 */ 983 void 984 vm_wait(void) 985 { 986 987 vm_page_lock_queues(); 988 if (curproc == pageproc) { 989 vm_pageout_pages_needed = 1; 990 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx, 991 PDROP | PSWP, "VMWait", 0); 992 } else { 993 if (!vm_pages_needed) { 994 vm_pages_needed = 1; 995 wakeup(&vm_pages_needed); 996 } 997 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM, 998 "vmwait", 0); 999 } 1000 } 1001 1002 /* 1003 * vm_waitpfault: (also see VM_WAITPFAULT macro) 1004 * 1005 * Block until free pages are available for allocation 1006 * - Called only in vm_fault so that processes page faulting 1007 * can be easily tracked. 1008 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 1009 * processes will be able to grab memory first. Do not change 1010 * this balance without careful testing first. 1011 */ 1012 void 1013 vm_waitpfault(void) 1014 { 1015 1016 vm_page_lock_queues(); 1017 if (!vm_pages_needed) { 1018 vm_pages_needed = 1; 1019 wakeup(&vm_pages_needed); 1020 } 1021 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER, 1022 "pfault", 0); 1023 } 1024 1025 /* 1026 * vm_page_activate: 1027 * 1028 * Put the specified page on the active list (if appropriate). 1029 * Ensure that act_count is at least ACT_INIT but do not otherwise 1030 * mess with it. 1031 * 1032 * The page queues must be locked. 1033 * This routine may not block. 1034 */ 1035 void 1036 vm_page_activate(vm_page_t m) 1037 { 1038 1039 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1040 if (VM_PAGE_GETKNOWNQUEUE2(m) != PQ_ACTIVE) { 1041 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1042 cnt.v_reactivated++; 1043 vm_pageq_remove(m); 1044 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1045 if (m->act_count < ACT_INIT) 1046 m->act_count = ACT_INIT; 1047 vm_pageq_enqueue(PQ_ACTIVE, m); 1048 } 1049 } else { 1050 if (m->act_count < ACT_INIT) 1051 m->act_count = ACT_INIT; 1052 } 1053 } 1054 1055 /* 1056 * vm_page_free_wakeup: 1057 * 1058 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1059 * routine is called when a page has been added to the cache or free 1060 * queues. 1061 * 1062 * The page queues must be locked. 1063 * This routine may not block. 1064 */ 1065 static inline void 1066 vm_page_free_wakeup(void) 1067 { 1068 1069 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1070 /* 1071 * if pageout daemon needs pages, then tell it that there are 1072 * some free. 1073 */ 1074 if (vm_pageout_pages_needed && 1075 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1076 wakeup(&vm_pageout_pages_needed); 1077 vm_pageout_pages_needed = 0; 1078 } 1079 /* 1080 * wakeup processes that are waiting on memory if we hit a 1081 * high water mark. And wakeup scheduler process if we have 1082 * lots of memory. this process will swapin processes. 1083 */ 1084 if (vm_pages_needed && !vm_page_count_min()) { 1085 vm_pages_needed = 0; 1086 wakeup(&cnt.v_free_count); 1087 } 1088 } 1089 1090 /* 1091 * vm_page_free_toq: 1092 * 1093 * Returns the given page to the PQ_FREE list, 1094 * disassociating it with any VM object. 1095 * 1096 * Object and page must be locked prior to entry. 1097 * This routine may not block. 1098 */ 1099 1100 void 1101 vm_page_free_toq(vm_page_t m) 1102 { 1103 struct vpgqueues *pq; 1104 1105 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1106 KASSERT(!pmap_page_is_mapped(m), 1107 ("vm_page_free_toq: freeing mapped page %p", m)); 1108 cnt.v_tfree++; 1109 1110 if (m->busy || VM_PAGE_INQUEUE1(m, PQ_FREE)) { 1111 printf( 1112 "vm_page_free: pindex(%lu), busy(%d), VPO_BUSY(%d), hold(%d)\n", 1113 (u_long)m->pindex, m->busy, (m->oflags & VPO_BUSY) ? 1 : 0, 1114 m->hold_count); 1115 if (VM_PAGE_INQUEUE1(m, PQ_FREE)) 1116 panic("vm_page_free: freeing free page"); 1117 else 1118 panic("vm_page_free: freeing busy page"); 1119 } 1120 1121 /* 1122 * unqueue, then remove page. Note that we cannot destroy 1123 * the page here because we do not want to call the pager's 1124 * callback routine until after we've put the page on the 1125 * appropriate free queue. 1126 */ 1127 vm_pageq_remove_nowakeup(m); 1128 vm_page_remove(m); 1129 1130 /* 1131 * If fictitious remove object association and 1132 * return, otherwise delay object association removal. 1133 */ 1134 if ((m->flags & PG_FICTITIOUS) != 0) { 1135 return; 1136 } 1137 1138 m->valid = 0; 1139 vm_page_undirty(m); 1140 1141 if (m->wire_count != 0) { 1142 if (m->wire_count > 1) { 1143 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1144 m->wire_count, (long)m->pindex); 1145 } 1146 panic("vm_page_free: freeing wired page"); 1147 } 1148 if (m->hold_count != 0) { 1149 m->flags &= ~PG_ZERO; 1150 VM_PAGE_SETQUEUE2(m, PQ_HOLD); 1151 } else 1152 VM_PAGE_SETQUEUE1(m, PQ_FREE); 1153 pq = &vm_page_queues[VM_PAGE_GETQUEUE(m)]; 1154 mtx_lock(&vm_page_queue_free_mtx); 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 } 1168 mtx_unlock(&vm_page_queue_free_mtx); 1169 vm_page_free_wakeup(); 1170 } 1171 1172 /* 1173 * vm_page_unmanage: 1174 * 1175 * Prevent PV management from being done on the page. The page is 1176 * removed from the paging queues as if it were wired, and as a 1177 * consequence of no longer being managed the pageout daemon will not 1178 * touch it (since there is no way to locate the pte mappings for the 1179 * page). madvise() calls that mess with the pmap will also no longer 1180 * operate on the page. 1181 * 1182 * Beyond that the page is still reasonably 'normal'. Freeing the page 1183 * will clear the flag. 1184 * 1185 * This routine is used by OBJT_PHYS objects - objects using unswappable 1186 * physical memory as backing store rather then swap-backed memory and 1187 * will eventually be extended to support 4MB unmanaged physical 1188 * mappings. 1189 */ 1190 void 1191 vm_page_unmanage(vm_page_t m) 1192 { 1193 1194 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1195 if ((m->flags & PG_UNMANAGED) == 0) { 1196 if (m->wire_count == 0) 1197 vm_pageq_remove(m); 1198 } 1199 vm_page_flag_set(m, PG_UNMANAGED); 1200 } 1201 1202 /* 1203 * vm_page_wire: 1204 * 1205 * Mark this page as wired down by yet 1206 * another map, removing it from paging queues 1207 * as necessary. 1208 * 1209 * The page queues must be locked. 1210 * This routine may not block. 1211 */ 1212 void 1213 vm_page_wire(vm_page_t m) 1214 { 1215 1216 /* 1217 * Only bump the wire statistics if the page is not already wired, 1218 * and only unqueue the page if it is on some queue (if it is unmanaged 1219 * it is already off the queues). 1220 */ 1221 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1222 if (m->flags & PG_FICTITIOUS) 1223 return; 1224 if (m->wire_count == 0) { 1225 if ((m->flags & PG_UNMANAGED) == 0) 1226 vm_pageq_remove(m); 1227 atomic_add_int(&cnt.v_wire_count, 1); 1228 } 1229 m->wire_count++; 1230 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1231 } 1232 1233 /* 1234 * vm_page_unwire: 1235 * 1236 * Release one wiring of this page, potentially 1237 * enabling it to be paged again. 1238 * 1239 * Many pages placed on the inactive queue should actually go 1240 * into the cache, but it is difficult to figure out which. What 1241 * we do instead, if the inactive target is well met, is to put 1242 * clean pages at the head of the inactive queue instead of the tail. 1243 * This will cause them to be moved to the cache more quickly and 1244 * if not actively re-referenced, freed more quickly. If we just 1245 * stick these pages at the end of the inactive queue, heavy filesystem 1246 * meta-data accesses can cause an unnecessary paging load on memory bound 1247 * processes. This optimization causes one-time-use metadata to be 1248 * reused more quickly. 1249 * 1250 * BUT, if we are in a low-memory situation we have no choice but to 1251 * put clean pages on the cache queue. 1252 * 1253 * A number of routines use vm_page_unwire() to guarantee that the page 1254 * will go into either the inactive or active queues, and will NEVER 1255 * be placed in the cache - for example, just after dirtying a page. 1256 * dirty pages in the cache are not allowed. 1257 * 1258 * The page queues must be locked. 1259 * This routine may not block. 1260 */ 1261 void 1262 vm_page_unwire(vm_page_t m, int activate) 1263 { 1264 1265 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1266 if (m->flags & PG_FICTITIOUS) 1267 return; 1268 if (m->wire_count > 0) { 1269 m->wire_count--; 1270 if (m->wire_count == 0) { 1271 atomic_subtract_int(&cnt.v_wire_count, 1); 1272 if (m->flags & PG_UNMANAGED) { 1273 ; 1274 } else if (activate) 1275 vm_pageq_enqueue(PQ_ACTIVE, m); 1276 else { 1277 vm_page_flag_clear(m, PG_WINATCFLS); 1278 vm_pageq_enqueue(PQ_INACTIVE, m); 1279 } 1280 } 1281 } else { 1282 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1283 } 1284 } 1285 1286 1287 /* 1288 * Move the specified page to the inactive queue. If the page has 1289 * any associated swap, the swap is deallocated. 1290 * 1291 * Normally athead is 0 resulting in LRU operation. athead is set 1292 * to 1 if we want this page to be 'as if it were placed in the cache', 1293 * except without unmapping it from the process address space. 1294 * 1295 * This routine may not block. 1296 */ 1297 static inline void 1298 _vm_page_deactivate(vm_page_t m, int athead) 1299 { 1300 1301 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1302 1303 /* 1304 * Ignore if already inactive. 1305 */ 1306 if (VM_PAGE_INQUEUE2(m, PQ_INACTIVE)) 1307 return; 1308 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1309 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1310 cnt.v_reactivated++; 1311 vm_page_flag_clear(m, PG_WINATCFLS); 1312 vm_pageq_remove(m); 1313 if (athead) 1314 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1315 else 1316 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1317 VM_PAGE_SETQUEUE2(m, PQ_INACTIVE); 1318 vm_page_queues[PQ_INACTIVE].lcnt++; 1319 cnt.v_inactive_count++; 1320 } 1321 } 1322 1323 void 1324 vm_page_deactivate(vm_page_t m) 1325 { 1326 _vm_page_deactivate(m, 0); 1327 } 1328 1329 /* 1330 * vm_page_try_to_cache: 1331 * 1332 * Returns 0 on failure, 1 on success 1333 */ 1334 int 1335 vm_page_try_to_cache(vm_page_t m) 1336 { 1337 1338 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1339 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1340 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1341 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) { 1342 return (0); 1343 } 1344 pmap_remove_all(m); 1345 if (m->dirty) 1346 return (0); 1347 vm_page_cache(m); 1348 return (1); 1349 } 1350 1351 /* 1352 * vm_page_try_to_free() 1353 * 1354 * Attempt to free the page. If we cannot free it, we do nothing. 1355 * 1 is returned on success, 0 on failure. 1356 */ 1357 int 1358 vm_page_try_to_free(vm_page_t m) 1359 { 1360 1361 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1362 if (m->object != NULL) 1363 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1364 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1365 (m->oflags & VPO_BUSY) || (m->flags & PG_UNMANAGED)) { 1366 return (0); 1367 } 1368 pmap_remove_all(m); 1369 if (m->dirty) 1370 return (0); 1371 vm_page_free(m); 1372 return (1); 1373 } 1374 1375 /* 1376 * vm_page_cache 1377 * 1378 * Put the specified page onto the page cache queue (if appropriate). 1379 * 1380 * This routine may not block. 1381 */ 1382 void 1383 vm_page_cache(vm_page_t m) 1384 { 1385 1386 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1387 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1388 if ((m->flags & PG_UNMANAGED) || (m->oflags & VPO_BUSY) || m->busy || 1389 m->hold_count || m->wire_count) { 1390 printf("vm_page_cache: attempting to cache busy page\n"); 1391 return; 1392 } 1393 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1394 return; 1395 1396 /* 1397 * Remove all pmaps and indicate that the page is not 1398 * writeable or mapped. 1399 */ 1400 pmap_remove_all(m); 1401 if (m->dirty != 0) { 1402 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1403 (long)m->pindex); 1404 } 1405 vm_pageq_remove_nowakeup(m); 1406 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1407 vm_page_free_wakeup(); 1408 } 1409 1410 /* 1411 * vm_page_dontneed 1412 * 1413 * Cache, deactivate, or do nothing as appropriate. This routine 1414 * is typically used by madvise() MADV_DONTNEED. 1415 * 1416 * Generally speaking we want to move the page into the cache so 1417 * it gets reused quickly. However, this can result in a silly syndrome 1418 * due to the page recycling too quickly. Small objects will not be 1419 * fully cached. On the otherhand, if we move the page to the inactive 1420 * queue we wind up with a problem whereby very large objects 1421 * unnecessarily blow away our inactive and cache queues. 1422 * 1423 * The solution is to move the pages based on a fixed weighting. We 1424 * either leave them alone, deactivate them, or move them to the cache, 1425 * where moving them to the cache has the highest weighting. 1426 * By forcing some pages into other queues we eventually force the 1427 * system to balance the queues, potentially recovering other unrelated 1428 * space from active. The idea is to not force this to happen too 1429 * often. 1430 */ 1431 void 1432 vm_page_dontneed(vm_page_t m) 1433 { 1434 static int dnweight; 1435 int dnw; 1436 int head; 1437 1438 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1439 dnw = ++dnweight; 1440 1441 /* 1442 * occassionally leave the page alone 1443 */ 1444 if ((dnw & 0x01F0) == 0 || 1445 VM_PAGE_INQUEUE2(m, PQ_INACTIVE) || 1446 VM_PAGE_INQUEUE1(m, PQ_CACHE) 1447 ) { 1448 if (m->act_count >= ACT_INIT) 1449 --m->act_count; 1450 return; 1451 } 1452 1453 if (m->dirty == 0 && pmap_is_modified(m)) 1454 vm_page_dirty(m); 1455 1456 if (m->dirty || (dnw & 0x0070) == 0) { 1457 /* 1458 * Deactivate the page 3 times out of 32. 1459 */ 1460 head = 0; 1461 } else { 1462 /* 1463 * Cache the page 28 times out of every 32. Note that 1464 * the page is deactivated instead of cached, but placed 1465 * at the head of the queue instead of the tail. 1466 */ 1467 head = 1; 1468 } 1469 _vm_page_deactivate(m, head); 1470 } 1471 1472 /* 1473 * Grab a page, waiting until we are waken up due to the page 1474 * changing state. We keep on waiting, if the page continues 1475 * to be in the object. If the page doesn't exist, first allocate it 1476 * and then conditionally zero it. 1477 * 1478 * This routine may block. 1479 */ 1480 vm_page_t 1481 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1482 { 1483 vm_page_t m; 1484 1485 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1486 retrylookup: 1487 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1488 if (vm_page_sleep_if_busy(m, TRUE, "pgrbwt")) { 1489 if ((allocflags & VM_ALLOC_RETRY) == 0) 1490 return (NULL); 1491 goto retrylookup; 1492 } else { 1493 if ((allocflags & VM_ALLOC_WIRED) != 0) { 1494 vm_page_lock_queues(); 1495 vm_page_wire(m); 1496 vm_page_unlock_queues(); 1497 } 1498 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 1499 vm_page_busy(m); 1500 return (m); 1501 } 1502 } 1503 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1504 if (m == NULL) { 1505 VM_OBJECT_UNLOCK(object); 1506 VM_WAIT; 1507 VM_OBJECT_LOCK(object); 1508 if ((allocflags & VM_ALLOC_RETRY) == 0) 1509 return (NULL); 1510 goto retrylookup; 1511 } 1512 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 1513 pmap_zero_page(m); 1514 return (m); 1515 } 1516 1517 /* 1518 * Mapping function for valid bits or for dirty bits in 1519 * a page. May not block. 1520 * 1521 * Inputs are required to range within a page. 1522 */ 1523 inline int 1524 vm_page_bits(int base, int size) 1525 { 1526 int first_bit; 1527 int last_bit; 1528 1529 KASSERT( 1530 base + size <= PAGE_SIZE, 1531 ("vm_page_bits: illegal base/size %d/%d", base, size) 1532 ); 1533 1534 if (size == 0) /* handle degenerate case */ 1535 return (0); 1536 1537 first_bit = base >> DEV_BSHIFT; 1538 last_bit = (base + size - 1) >> DEV_BSHIFT; 1539 1540 return ((2 << last_bit) - (1 << first_bit)); 1541 } 1542 1543 /* 1544 * vm_page_set_validclean: 1545 * 1546 * Sets portions of a page valid and clean. The arguments are expected 1547 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1548 * of any partial chunks touched by the range. The invalid portion of 1549 * such chunks will be zero'd. 1550 * 1551 * This routine may not block. 1552 * 1553 * (base + size) must be less then or equal to PAGE_SIZE. 1554 */ 1555 void 1556 vm_page_set_validclean(vm_page_t m, int base, int size) 1557 { 1558 int pagebits; 1559 int frag; 1560 int endoff; 1561 1562 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1563 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1564 if (size == 0) /* handle degenerate case */ 1565 return; 1566 1567 /* 1568 * If the base is not DEV_BSIZE aligned and the valid 1569 * bit is clear, we have to zero out a portion of the 1570 * first block. 1571 */ 1572 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1573 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1574 pmap_zero_page_area(m, frag, base - frag); 1575 1576 /* 1577 * If the ending offset is not DEV_BSIZE aligned and the 1578 * valid bit is clear, we have to zero out a portion of 1579 * the last block. 1580 */ 1581 endoff = base + size; 1582 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1583 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1584 pmap_zero_page_area(m, endoff, 1585 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1586 1587 /* 1588 * Set valid, clear dirty bits. If validating the entire 1589 * page we can safely clear the pmap modify bit. We also 1590 * use this opportunity to clear the VPO_NOSYNC flag. If a process 1591 * takes a write fault on a MAP_NOSYNC memory area the flag will 1592 * be set again. 1593 * 1594 * We set valid bits inclusive of any overlap, but we can only 1595 * clear dirty bits for DEV_BSIZE chunks that are fully within 1596 * the range. 1597 */ 1598 pagebits = vm_page_bits(base, size); 1599 m->valid |= pagebits; 1600 #if 0 /* NOT YET */ 1601 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1602 frag = DEV_BSIZE - frag; 1603 base += frag; 1604 size -= frag; 1605 if (size < 0) 1606 size = 0; 1607 } 1608 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1609 #endif 1610 m->dirty &= ~pagebits; 1611 if (base == 0 && size == PAGE_SIZE) { 1612 pmap_clear_modify(m); 1613 m->oflags &= ~VPO_NOSYNC; 1614 } 1615 } 1616 1617 void 1618 vm_page_clear_dirty(vm_page_t m, int base, int size) 1619 { 1620 1621 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1622 m->dirty &= ~vm_page_bits(base, size); 1623 } 1624 1625 /* 1626 * vm_page_set_invalid: 1627 * 1628 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1629 * valid and dirty bits for the effected areas are cleared. 1630 * 1631 * May not block. 1632 */ 1633 void 1634 vm_page_set_invalid(vm_page_t m, int base, int size) 1635 { 1636 int bits; 1637 1638 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1639 bits = vm_page_bits(base, size); 1640 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1641 if (m->valid == VM_PAGE_BITS_ALL && bits != 0) 1642 pmap_remove_all(m); 1643 m->valid &= ~bits; 1644 m->dirty &= ~bits; 1645 m->object->generation++; 1646 } 1647 1648 /* 1649 * vm_page_zero_invalid() 1650 * 1651 * The kernel assumes that the invalid portions of a page contain 1652 * garbage, but such pages can be mapped into memory by user code. 1653 * When this occurs, we must zero out the non-valid portions of the 1654 * page so user code sees what it expects. 1655 * 1656 * Pages are most often semi-valid when the end of a file is mapped 1657 * into memory and the file's size is not page aligned. 1658 */ 1659 void 1660 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1661 { 1662 int b; 1663 int i; 1664 1665 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1666 /* 1667 * Scan the valid bits looking for invalid sections that 1668 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1669 * valid bit may be set ) have already been zerod by 1670 * vm_page_set_validclean(). 1671 */ 1672 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1673 if (i == (PAGE_SIZE / DEV_BSIZE) || 1674 (m->valid & (1 << i)) 1675 ) { 1676 if (i > b) { 1677 pmap_zero_page_area(m, 1678 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1679 } 1680 b = i + 1; 1681 } 1682 } 1683 1684 /* 1685 * setvalid is TRUE when we can safely set the zero'd areas 1686 * as being valid. We can do this if there are no cache consistancy 1687 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1688 */ 1689 if (setvalid) 1690 m->valid = VM_PAGE_BITS_ALL; 1691 } 1692 1693 /* 1694 * vm_page_is_valid: 1695 * 1696 * Is (partial) page valid? Note that the case where size == 0 1697 * will return FALSE in the degenerate case where the page is 1698 * entirely invalid, and TRUE otherwise. 1699 * 1700 * May not block. 1701 */ 1702 int 1703 vm_page_is_valid(vm_page_t m, int base, int size) 1704 { 1705 int bits = vm_page_bits(base, size); 1706 1707 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1708 if (m->valid && ((m->valid & bits) == bits)) 1709 return 1; 1710 else 1711 return 0; 1712 } 1713 1714 /* 1715 * update dirty bits from pmap/mmu. May not block. 1716 */ 1717 void 1718 vm_page_test_dirty(vm_page_t m) 1719 { 1720 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1721 vm_page_dirty(m); 1722 } 1723 } 1724 1725 int so_zerocp_fullpage = 0; 1726 1727 void 1728 vm_page_cowfault(vm_page_t m) 1729 { 1730 vm_page_t mnew; 1731 vm_object_t object; 1732 vm_pindex_t pindex; 1733 1734 object = m->object; 1735 pindex = m->pindex; 1736 1737 retry_alloc: 1738 pmap_remove_all(m); 1739 vm_page_remove(m); 1740 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY); 1741 if (mnew == NULL) { 1742 vm_page_insert(m, object, pindex); 1743 vm_page_unlock_queues(); 1744 VM_OBJECT_UNLOCK(object); 1745 VM_WAIT; 1746 VM_OBJECT_LOCK(object); 1747 vm_page_lock_queues(); 1748 goto retry_alloc; 1749 } 1750 1751 if (m->cow == 0) { 1752 /* 1753 * check to see if we raced with an xmit complete when 1754 * waiting to allocate a page. If so, put things back 1755 * the way they were 1756 */ 1757 vm_page_free(mnew); 1758 vm_page_insert(m, object, pindex); 1759 } else { /* clear COW & copy page */ 1760 if (!so_zerocp_fullpage) 1761 pmap_copy_page(m, mnew); 1762 mnew->valid = VM_PAGE_BITS_ALL; 1763 vm_page_dirty(mnew); 1764 mnew->wire_count = m->wire_count - m->cow; 1765 m->wire_count = m->cow; 1766 } 1767 } 1768 1769 void 1770 vm_page_cowclear(vm_page_t m) 1771 { 1772 1773 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1774 if (m->cow) { 1775 m->cow--; 1776 /* 1777 * let vm_fault add back write permission lazily 1778 */ 1779 } 1780 /* 1781 * sf_buf_free() will free the page, so we needn't do it here 1782 */ 1783 } 1784 1785 void 1786 vm_page_cowsetup(vm_page_t m) 1787 { 1788 1789 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1790 m->cow++; 1791 pmap_remove_write(m); 1792 } 1793 1794 #include "opt_ddb.h" 1795 #ifdef DDB 1796 #include <sys/kernel.h> 1797 1798 #include <ddb/ddb.h> 1799 1800 DB_SHOW_COMMAND(page, vm_page_print_page_info) 1801 { 1802 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1803 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1804 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1805 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1806 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1807 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1808 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1809 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1810 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1811 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1812 } 1813 1814 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1815 { 1816 int i; 1817 db_printf("PQ_FREE:"); 1818 for (i = 0; i < PQ_NUMCOLORS; i++) { 1819 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1820 } 1821 db_printf("\n"); 1822 1823 db_printf("PQ_CACHE:"); 1824 for (i = 0; i < PQ_NUMCOLORS; i++) { 1825 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1826 } 1827 db_printf("\n"); 1828 1829 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1830 vm_page_queues[PQ_ACTIVE].lcnt, 1831 vm_page_queues[PQ_INACTIVE].lcnt); 1832 } 1833 #endif /* DDB */ 1834