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