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