vm_page.c revision 112390
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 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 37 * $FreeBSD: head/sys/vm/vm_page.c 112390 2003-03-19 00:34:43Z mux $ 38 */ 39 40/* 41 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 42 * All rights reserved. 43 * 44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 45 * 46 * Permission to use, copy, modify and distribute this software and 47 * its documentation is hereby granted, provided that both the copyright 48 * notice and this permission notice appear in all copies of the 49 * software, derivative works or modified versions, and any portions 50 * thereof, and that both notices appear in supporting documentation. 51 * 52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 55 * 56 * Carnegie Mellon requests users of this software to return to 57 * 58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 59 * School of Computer Science 60 * Carnegie Mellon University 61 * Pittsburgh PA 15213-3890 62 * 63 * any improvements or extensions that they make and grant Carnegie the 64 * rights to redistribute these changes. 65 */ 66 67/* 68 * GENERAL RULES ON VM_PAGE MANIPULATION 69 * 70 * - a pageq mutex is required when adding or removing a page from a 71 * page queue (vm_page_queue[]), regardless of other mutexes or the 72 * busy state of a page. 73 * 74 * - a hash chain mutex is required when associating or disassociating 75 * a page from the VM PAGE CACHE hash table (vm_page_buckets), 76 * regardless of other mutexes or the busy state of a page. 77 * 78 * - either a hash chain mutex OR a busied page is required in order 79 * to modify the page flags. A hash chain mutex must be obtained in 80 * order to busy a page. A page's flags cannot be modified by a 81 * hash chain mutex if the page is marked busy. 82 * 83 * - The object memq mutex is held when inserting or removing 84 * pages from an object (vm_page_insert() or vm_page_remove()). This 85 * is different from the object's main mutex. 86 * 87 * Generally speaking, you have to be aware of side effects when running 88 * vm_page ops. A vm_page_lookup() will return with the hash chain 89 * locked, whether it was able to lookup the page or not. vm_page_free(), 90 * vm_page_cache(), vm_page_activate(), and a number of other routines 91 * will release the hash chain mutex for you. Intermediate manipulation 92 * routines such as vm_page_flag_set() expect the hash chain to be held 93 * on entry and the hash chain will remain held on return. 94 * 95 * pageq scanning can only occur with the pageq in question locked. 96 * We have a known bottleneck with the active queue, but the cache 97 * and free queues are actually arrays already. 98 */ 99 100/* 101 * Resident memory management module. 102 */ 103 104#include <sys/param.h> 105#include <sys/systm.h> 106#include <sys/lock.h> 107#include <sys/malloc.h> 108#include <sys/mutex.h> 109#include <sys/proc.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 129struct mtx vm_page_queue_mtx; 130struct mtx vm_page_queue_free_mtx; 131 132vm_page_t vm_page_array = 0; 133int vm_page_array_size = 0; 134long first_page = 0; 135int vm_page_zero_count = 0; 136 137/* 138 * vm_set_page_size: 139 * 140 * Sets the page size, perhaps based upon the memory 141 * size. Must be called before any use of page-size 142 * dependent functions. 143 */ 144void 145vm_set_page_size(void) 146{ 147 if (cnt.v_page_size == 0) 148 cnt.v_page_size = PAGE_SIZE; 149 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 150 panic("vm_set_page_size: page size not a power of two"); 151} 152 153/* 154 * vm_page_startup: 155 * 156 * Initializes the resident memory module. 157 * 158 * Allocates memory for the page cells, and 159 * for the object/offset-to-page hash table headers. 160 * Each page cell is initialized and placed on the free list. 161 */ 162vm_offset_t 163vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr) 164{ 165 vm_offset_t mapped; 166 vm_size_t npages, page_range; 167 vm_offset_t new_end; 168 int i; 169 vm_offset_t pa; 170 int nblocks; 171 vm_offset_t last_pa; 172 173 /* the biggest memory array is the second group of pages */ 174 vm_offset_t end; 175 vm_offset_t biggestone, biggestsize; 176 177 vm_offset_t total; 178 vm_size_t bootpages; 179 180 total = 0; 181 biggestsize = 0; 182 biggestone = 0; 183 nblocks = 0; 184 vaddr = round_page(vaddr); 185 186 for (i = 0; phys_avail[i + 1]; i += 2) { 187 phys_avail[i] = round_page(phys_avail[i]); 188 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 189 } 190 191 for (i = 0; phys_avail[i + 1]; i += 2) { 192 vm_size_t size = phys_avail[i + 1] - phys_avail[i]; 193 194 if (size > biggestsize) { 195 biggestone = i; 196 biggestsize = size; 197 } 198 ++nblocks; 199 total += size; 200 } 201 202 end = phys_avail[biggestone+1]; 203 204 /* 205 * Initialize the locks. 206 */ 207 mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF); 208 mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 209 MTX_SPIN); 210 211 /* 212 * Initialize the queue headers for the free queue, the active queue 213 * and the inactive queue. 214 */ 215 vm_pageq_init(); 216 217 /* 218 * Allocate memory for use when boot strapping the kernel memory 219 * allocator. 220 */ 221 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE; 222 new_end = end - bootpages; 223 new_end = trunc_page(new_end); 224 mapped = pmap_map(&vaddr, new_end, end, 225 VM_PROT_READ | VM_PROT_WRITE); 226 bzero((caddr_t) mapped, end - new_end); 227 uma_startup((caddr_t)mapped); 228 229 /* 230 * Compute the number of pages of memory that will be available for 231 * use (taking into account the overhead of a page structure per 232 * page). 233 */ 234 first_page = phys_avail[0] / PAGE_SIZE; 235 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 236 npages = (total - (page_range * sizeof(struct vm_page)) - 237 (end - new_end)) / PAGE_SIZE; 238 end = new_end; 239 240 /* 241 * Initialize the mem entry structures now, and put them in the free 242 * queue. 243 */ 244 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 245 mapped = pmap_map(&vaddr, new_end, end, 246 VM_PROT_READ | VM_PROT_WRITE); 247 vm_page_array = (vm_page_t) mapped; 248 phys_avail[biggestone + 1] = new_end; 249 250 /* 251 * Clear all of the page structures 252 */ 253 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 254 vm_page_array_size = page_range; 255 256 /* 257 * Construct the free queue(s) in descending order (by physical 258 * address) so that the first 16MB of physical memory is allocated 259 * last rather than first. On large-memory machines, this avoids 260 * the exhaustion of low physical memory before isa_dmainit has run. 261 */ 262 cnt.v_page_count = 0; 263 cnt.v_free_count = 0; 264 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 265 pa = phys_avail[i]; 266 last_pa = phys_avail[i + 1]; 267 while (pa < last_pa && npages-- > 0) { 268 vm_pageq_add_new_page(pa); 269 pa += PAGE_SIZE; 270 } 271 } 272 return (vaddr); 273} 274 275void 276vm_page_flag_set(vm_page_t m, unsigned short bits) 277{ 278 279 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 280 m->flags |= bits; 281} 282 283void 284vm_page_flag_clear(vm_page_t m, unsigned short bits) 285{ 286 287 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 288 m->flags &= ~bits; 289} 290 291void 292vm_page_busy(vm_page_t m) 293{ 294 KASSERT((m->flags & PG_BUSY) == 0, 295 ("vm_page_busy: page already busy!!!")); 296 vm_page_flag_set(m, PG_BUSY); 297} 298 299/* 300 * vm_page_flash: 301 * 302 * wakeup anyone waiting for the page. 303 */ 304void 305vm_page_flash(vm_page_t m) 306{ 307 if (m->flags & PG_WANTED) { 308 vm_page_flag_clear(m, PG_WANTED); 309 wakeup(m); 310 } 311} 312 313/* 314 * vm_page_wakeup: 315 * 316 * clear the PG_BUSY flag and wakeup anyone waiting for the 317 * page. 318 * 319 */ 320void 321vm_page_wakeup(vm_page_t m) 322{ 323 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 324 vm_page_flag_clear(m, PG_BUSY); 325 vm_page_flash(m); 326} 327 328void 329vm_page_io_start(vm_page_t m) 330{ 331 332 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 333 m->busy++; 334} 335 336void 337vm_page_io_finish(vm_page_t m) 338{ 339 340 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 341 m->busy--; 342 if (m->busy == 0) 343 vm_page_flash(m); 344} 345 346/* 347 * Keep page from being freed by the page daemon 348 * much of the same effect as wiring, except much lower 349 * overhead and should be used only for *very* temporary 350 * holding ("wiring"). 351 */ 352void 353vm_page_hold(vm_page_t mem) 354{ 355 356 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 357 mem->hold_count++; 358} 359 360void 361vm_page_unhold(vm_page_t mem) 362{ 363 364 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 365 --mem->hold_count; 366 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 367 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 368 vm_page_free_toq(mem); 369} 370 371/* 372 * vm_page_copy: 373 * 374 * Copy one page to another 375 */ 376void 377vm_page_copy(vm_page_t src_m, vm_page_t dest_m) 378{ 379 pmap_copy_page(src_m, dest_m); 380 dest_m->valid = VM_PAGE_BITS_ALL; 381} 382 383/* 384 * vm_page_free: 385 * 386 * Free a page 387 * 388 * The clearing of PG_ZERO is a temporary safety until the code can be 389 * reviewed to determine that PG_ZERO is being properly cleared on 390 * write faults or maps. PG_ZERO was previously cleared in 391 * vm_page_alloc(). 392 */ 393void 394vm_page_free(vm_page_t m) 395{ 396 vm_page_flag_clear(m, PG_ZERO); 397 vm_page_free_toq(m); 398 vm_page_zero_idle_wakeup(); 399} 400 401/* 402 * vm_page_free_zero: 403 * 404 * Free a page to the zerod-pages queue 405 */ 406void 407vm_page_free_zero(vm_page_t m) 408{ 409 vm_page_flag_set(m, PG_ZERO); 410 vm_page_free_toq(m); 411} 412 413/* 414 * vm_page_sleep_if_busy: 415 * 416 * Sleep and release the page queues lock if PG_BUSY is set or, 417 * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the 418 * thread slept and the page queues lock was released. 419 * Otherwise, retains the page queues lock and returns FALSE. 420 */ 421int 422vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg) 423{ 424 int is_object_locked; 425 426 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 427 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 428 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 429 /* 430 * Remove mtx_owned() after vm_object locking is finished. 431 */ 432 if ((is_object_locked = m->object != NULL && 433 mtx_owned(&m->object->mtx))) 434 mtx_unlock(&m->object->mtx); 435 msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0); 436 if (is_object_locked) 437 mtx_lock(&m->object->mtx); 438 return (TRUE); 439 } 440 return (FALSE); 441} 442 443/* 444 * vm_page_dirty: 445 * 446 * make page all dirty 447 */ 448void 449vm_page_dirty(vm_page_t m) 450{ 451 KASSERT(m->queue - m->pc != PQ_CACHE, 452 ("vm_page_dirty: page in cache!")); 453 KASSERT(m->queue - m->pc != PQ_FREE, 454 ("vm_page_dirty: page is free!")); 455 m->dirty = VM_PAGE_BITS_ALL; 456} 457 458/* 459 * vm_page_splay: 460 * 461 * Implements Sleator and Tarjan's top-down splay algorithm. Returns 462 * the vm_page containing the given pindex. If, however, that 463 * pindex is not found in the vm_object, returns a vm_page that is 464 * adjacent to the pindex, coming before or after it. 465 */ 466vm_page_t 467vm_page_splay(vm_pindex_t pindex, vm_page_t root) 468{ 469 struct vm_page dummy; 470 vm_page_t lefttreemax, righttreemin, y; 471 472 if (root == NULL) 473 return (root); 474 lefttreemax = righttreemin = &dummy; 475 for (;; root = y) { 476 if (pindex < root->pindex) { 477 if ((y = root->left) == NULL) 478 break; 479 if (pindex < y->pindex) { 480 /* Rotate right. */ 481 root->left = y->right; 482 y->right = root; 483 root = y; 484 if ((y = root->left) == NULL) 485 break; 486 } 487 /* Link into the new root's right tree. */ 488 righttreemin->left = root; 489 righttreemin = root; 490 } else if (pindex > root->pindex) { 491 if ((y = root->right) == NULL) 492 break; 493 if (pindex > y->pindex) { 494 /* Rotate left. */ 495 root->right = y->left; 496 y->left = root; 497 root = y; 498 if ((y = root->right) == NULL) 499 break; 500 } 501 /* Link into the new root's left tree. */ 502 lefttreemax->right = root; 503 lefttreemax = root; 504 } else 505 break; 506 } 507 /* Assemble the new root. */ 508 lefttreemax->right = root->left; 509 righttreemin->left = root->right; 510 root->left = dummy.right; 511 root->right = dummy.left; 512 return (root); 513} 514 515/* 516 * vm_page_insert: [ internal use only ] 517 * 518 * Inserts the given mem entry into the object and object list. 519 * 520 * The pagetables are not updated but will presumably fault the page 521 * in if necessary, or if a kernel page the caller will at some point 522 * enter the page into the kernel's pmap. We are not allowed to block 523 * here so we *can't* do this anyway. 524 * 525 * The object and page must be locked, and must be splhigh. 526 * This routine may not block. 527 */ 528void 529vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 530{ 531 vm_page_t root; 532 533 if (m->object != NULL) 534 panic("vm_page_insert: already inserted"); 535 536 /* 537 * Record the object/offset pair in this page 538 */ 539 m->object = object; 540 m->pindex = pindex; 541 542 mtx_assert(object == kmem_object ? &object->mtx : &Giant, MA_OWNED); 543 /* 544 * Now link into the object's ordered list of backed pages. 545 */ 546 root = object->root; 547 if (root == NULL) { 548 m->left = NULL; 549 m->right = NULL; 550 TAILQ_INSERT_TAIL(&object->memq, m, listq); 551 } else { 552 root = vm_page_splay(pindex, root); 553 if (pindex < root->pindex) { 554 m->left = root->left; 555 m->right = root; 556 root->left = NULL; 557 TAILQ_INSERT_BEFORE(root, m, listq); 558 } else { 559 m->right = root->right; 560 m->left = root; 561 root->right = NULL; 562 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 563 } 564 } 565 object->root = m; 566 object->generation++; 567 568 /* 569 * show that the object has one more resident page. 570 */ 571 object->resident_page_count++; 572 573 /* 574 * Since we are inserting a new and possibly dirty page, 575 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 576 */ 577 if (m->flags & PG_WRITEABLE) 578 vm_object_set_writeable_dirty(object); 579} 580 581/* 582 * vm_page_remove: 583 * NOTE: used by device pager as well -wfj 584 * 585 * Removes the given mem entry from the object/offset-page 586 * table and the object page list, but do not invalidate/terminate 587 * the backing store. 588 * 589 * The object and page must be locked, and at splhigh. 590 * The underlying pmap entry (if any) is NOT removed here. 591 * This routine may not block. 592 */ 593void 594vm_page_remove(vm_page_t m) 595{ 596 vm_object_t object; 597 vm_page_t root; 598 599 GIANT_REQUIRED; 600 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 601 if (m->object == NULL) 602 return; 603 604 if ((m->flags & PG_BUSY) == 0) { 605 panic("vm_page_remove: page not busy"); 606 } 607 608 /* 609 * Basically destroy the page. 610 */ 611 vm_page_wakeup(m); 612 613 object = m->object; 614 615 /* 616 * Now remove from the object's list of backed pages. 617 */ 618 if (m != object->root) 619 vm_page_splay(m->pindex, object->root); 620 if (m->left == NULL) 621 root = m->right; 622 else { 623 root = vm_page_splay(m->pindex, m->left); 624 root->right = m->right; 625 } 626 object->root = root; 627 TAILQ_REMOVE(&object->memq, m, listq); 628 629 /* 630 * And show that the object has one fewer resident page. 631 */ 632 object->resident_page_count--; 633 object->generation++; 634 635 m->object = NULL; 636} 637 638/* 639 * vm_page_lookup: 640 * 641 * Returns the page associated with the object/offset 642 * pair specified; if none is found, NULL is returned. 643 * 644 * The object must be locked. 645 * This routine may not block. 646 * This is a critical path routine 647 */ 648vm_page_t 649vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 650{ 651 vm_page_t m; 652 653 mtx_assert(object == kmem_object ? &object->mtx : &Giant, MA_OWNED); 654 m = vm_page_splay(pindex, object->root); 655 if ((object->root = m) != NULL && m->pindex != pindex) 656 m = NULL; 657 return (m); 658} 659 660/* 661 * vm_page_rename: 662 * 663 * Move the given memory entry from its 664 * current object to the specified target object/offset. 665 * 666 * The object must be locked. 667 * This routine may not block. 668 * 669 * Note: this routine will raise itself to splvm(), the caller need not. 670 * 671 * Note: swap associated with the page must be invalidated by the move. We 672 * have to do this for several reasons: (1) we aren't freeing the 673 * page, (2) we are dirtying the page, (3) the VM system is probably 674 * moving the page from object A to B, and will then later move 675 * the backing store from A to B and we can't have a conflict. 676 * 677 * Note: we *always* dirty the page. It is necessary both for the 678 * fact that we moved it, and because we may be invalidating 679 * swap. If the page is on the cache, we have to deactivate it 680 * or vm_page_dirty() will panic. Dirty pages are not allowed 681 * on the cache. 682 */ 683void 684vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 685{ 686 int s; 687 688 s = splvm(); 689 vm_page_remove(m); 690 vm_page_insert(m, new_object, new_pindex); 691 if (m->queue - m->pc == PQ_CACHE) 692 vm_page_deactivate(m); 693 vm_page_dirty(m); 694 splx(s); 695} 696 697/* 698 * vm_page_select_cache: 699 * 700 * Find a page on the cache queue with color optimization. As pages 701 * might be found, but not applicable, they are deactivated. This 702 * keeps us from using potentially busy cached pages. 703 * 704 * This routine must be called at splvm(). 705 * This routine may not block. 706 */ 707static vm_page_t 708vm_page_select_cache(vm_pindex_t color) 709{ 710 vm_page_t m; 711 712 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 713 while (TRUE) { 714 m = vm_pageq_find(PQ_CACHE, color & PQ_L2_MASK, FALSE); 715 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 716 m->hold_count || m->wire_count)) { 717 vm_page_deactivate(m); 718 continue; 719 } 720 return m; 721 } 722} 723 724/* 725 * vm_page_select_free: 726 * 727 * Find a free or zero page, with specified preference. 728 * 729 * This routine must be called at splvm(). 730 * This routine may not block. 731 */ 732static __inline vm_page_t 733vm_page_select_free(vm_pindex_t color, boolean_t prefer_zero) 734{ 735 vm_page_t m; 736 737 m = vm_pageq_find(PQ_FREE, color & PQ_L2_MASK, prefer_zero); 738 return (m); 739} 740 741/* 742 * vm_page_alloc: 743 * 744 * Allocate and return a memory cell associated 745 * with this VM object/offset pair. 746 * 747 * page_req classes: 748 * VM_ALLOC_NORMAL normal process request 749 * VM_ALLOC_SYSTEM system *really* needs a page 750 * VM_ALLOC_INTERRUPT interrupt time request 751 * VM_ALLOC_ZERO zero page 752 * 753 * This routine may not block. 754 * 755 * Additional special handling is required when called from an 756 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 757 * the page cache in this case. 758 */ 759vm_page_t 760vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 761{ 762 vm_page_t m = NULL; 763 vm_pindex_t color; 764 int flags, page_req, s; 765 766 page_req = req & VM_ALLOC_CLASS_MASK; 767 768 if ((req & VM_ALLOC_NOOBJ) == 0) { 769 KASSERT(object != NULL, 770 ("vm_page_alloc: NULL object.")); 771 mtx_assert(object == kmem_object ? &object->mtx : &Giant, 772 MA_OWNED); 773 KASSERT(!vm_page_lookup(object, pindex), 774 ("vm_page_alloc: page already allocated")); 775 color = pindex + object->pg_color; 776 } else 777 color = pindex; 778 779 /* 780 * The pager is allowed to eat deeper into the free page list. 781 */ 782 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 783 page_req = VM_ALLOC_SYSTEM; 784 }; 785 786 s = splvm(); 787loop: 788 mtx_lock_spin(&vm_page_queue_free_mtx); 789 if (cnt.v_free_count > cnt.v_free_reserved || 790 (page_req == VM_ALLOC_SYSTEM && 791 cnt.v_cache_count == 0 && 792 cnt.v_free_count > cnt.v_interrupt_free_min) || 793 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 794 /* 795 * Allocate from the free queue if the number of free pages 796 * exceeds the minimum for the request class. 797 */ 798 m = vm_page_select_free(color, (req & VM_ALLOC_ZERO) != 0); 799 } else if (page_req != VM_ALLOC_INTERRUPT) { 800 mtx_unlock_spin(&vm_page_queue_free_mtx); 801 /* 802 * Allocatable from cache (non-interrupt only). On success, 803 * we must free the page and try again, thus ensuring that 804 * cnt.v_*_free_min counters are replenished. 805 */ 806 vm_page_lock_queues(); 807 if ((m = vm_page_select_cache(color)) == NULL) { 808 vm_page_unlock_queues(); 809 splx(s); 810#if defined(DIAGNOSTIC) 811 if (cnt.v_cache_count > 0) 812 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 813#endif 814 atomic_add_int(&vm_pageout_deficit, 1); 815 pagedaemon_wakeup(); 816 return (NULL); 817 } 818 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 819 vm_page_busy(m); 820 pmap_remove_all(m); 821 vm_page_free(m); 822 vm_page_unlock_queues(); 823 goto loop; 824 } else { 825 /* 826 * Not allocatable from cache from interrupt, give up. 827 */ 828 mtx_unlock_spin(&vm_page_queue_free_mtx); 829 splx(s); 830 atomic_add_int(&vm_pageout_deficit, 1); 831 pagedaemon_wakeup(); 832 return (NULL); 833 } 834 835 /* 836 * At this point we had better have found a good page. 837 */ 838 839 KASSERT( 840 m != NULL, 841 ("vm_page_alloc(): missing page on free queue\n") 842 ); 843 844 /* 845 * Remove from free queue 846 */ 847 848 vm_pageq_remove_nowakeup(m); 849 850 /* 851 * Initialize structure. Only the PG_ZERO flag is inherited. 852 */ 853 flags = PG_BUSY; 854 if (m->flags & PG_ZERO) { 855 vm_page_zero_count--; 856 if (req & VM_ALLOC_ZERO) 857 flags = PG_ZERO | PG_BUSY; 858 } 859 m->flags = flags; 860 if (req & VM_ALLOC_WIRED) { 861 atomic_add_int(&cnt.v_wire_count, 1); 862 m->wire_count = 1; 863 } else 864 m->wire_count = 0; 865 m->hold_count = 0; 866 m->act_count = 0; 867 m->busy = 0; 868 m->valid = 0; 869 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 870 mtx_unlock_spin(&vm_page_queue_free_mtx); 871 872 /* 873 * vm_page_insert() is safe prior to the splx(). Note also that 874 * inserting a page here does not insert it into the pmap (which 875 * could cause us to block allocating memory). We cannot block 876 * anywhere. 877 */ 878 if ((req & VM_ALLOC_NOOBJ) == 0) 879 vm_page_insert(m, object, pindex); 880 881 /* 882 * Don't wakeup too often - wakeup the pageout daemon when 883 * we would be nearly out of memory. 884 */ 885 if (vm_paging_needed()) 886 pagedaemon_wakeup(); 887 888 splx(s); 889 return (m); 890} 891 892/* 893 * vm_wait: (also see VM_WAIT macro) 894 * 895 * Block until free pages are available for allocation 896 * - Called in various places before memory allocations. 897 */ 898void 899vm_wait(void) 900{ 901 int s; 902 903 s = splvm(); 904 vm_page_lock_queues(); 905 if (curproc == pageproc) { 906 vm_pageout_pages_needed = 1; 907 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx, 908 PDROP | PSWP, "VMWait", 0); 909 } else { 910 if (!vm_pages_needed) { 911 vm_pages_needed = 1; 912 wakeup(&vm_pages_needed); 913 } 914 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM, 915 "vmwait", 0); 916 } 917 splx(s); 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 */ 930void 931vm_waitpfault(void) 932{ 933 int s; 934 935 s = splvm(); 936 vm_page_lock_queues(); 937 if (!vm_pages_needed) { 938 vm_pages_needed = 1; 939 wakeup(&vm_pages_needed); 940 } 941 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER, 942 "pfault", 0); 943 splx(s); 944} 945 946/* 947 * vm_page_activate: 948 * 949 * Put the specified page on the active list (if appropriate). 950 * Ensure that act_count is at least ACT_INIT but do not otherwise 951 * mess with it. 952 * 953 * The page queues must be locked. 954 * This routine may not block. 955 */ 956void 957vm_page_activate(vm_page_t m) 958{ 959 int s; 960 961 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 962 s = splvm(); 963 if (m->queue != PQ_ACTIVE) { 964 if ((m->queue - m->pc) == PQ_CACHE) 965 cnt.v_reactivated++; 966 vm_pageq_remove(m); 967 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 968 if (m->act_count < ACT_INIT) 969 m->act_count = ACT_INIT; 970 vm_pageq_enqueue(PQ_ACTIVE, m); 971 } 972 } else { 973 if (m->act_count < ACT_INIT) 974 m->act_count = ACT_INIT; 975 } 976 splx(s); 977} 978 979/* 980 * vm_page_free_wakeup: 981 * 982 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 983 * routine is called when a page has been added to the cache or free 984 * queues. 985 * 986 * This routine may not block. 987 * This routine must be called at splvm() 988 */ 989static __inline void 990vm_page_free_wakeup(void) 991{ 992 993 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 994 /* 995 * if pageout daemon needs pages, then tell it that there are 996 * some free. 997 */ 998 if (vm_pageout_pages_needed && 999 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1000 wakeup(&vm_pageout_pages_needed); 1001 vm_pageout_pages_needed = 0; 1002 } 1003 /* 1004 * wakeup processes that are waiting on memory if we hit a 1005 * high water mark. And wakeup scheduler process if we have 1006 * lots of memory. this process will swapin processes. 1007 */ 1008 if (vm_pages_needed && !vm_page_count_min()) { 1009 vm_pages_needed = 0; 1010 wakeup(&cnt.v_free_count); 1011 } 1012} 1013 1014/* 1015 * vm_page_free_toq: 1016 * 1017 * Returns the given page to the PQ_FREE list, 1018 * disassociating it with any VM object. 1019 * 1020 * Object and page must be locked prior to entry. 1021 * This routine may not block. 1022 */ 1023 1024void 1025vm_page_free_toq(vm_page_t m) 1026{ 1027 int s; 1028 struct vpgqueues *pq; 1029 vm_object_t object = m->object; 1030 1031 GIANT_REQUIRED; 1032 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1033 s = splvm(); 1034 cnt.v_tfree++; 1035 1036 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1037 printf( 1038 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1039 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1040 m->hold_count); 1041 if ((m->queue - m->pc) == PQ_FREE) 1042 panic("vm_page_free: freeing free page"); 1043 else 1044 panic("vm_page_free: freeing busy page"); 1045 } 1046 1047 /* 1048 * unqueue, then remove page. Note that we cannot destroy 1049 * the page here because we do not want to call the pager's 1050 * callback routine until after we've put the page on the 1051 * appropriate free queue. 1052 */ 1053 vm_pageq_remove_nowakeup(m); 1054 vm_page_remove(m); 1055 1056 /* 1057 * If fictitious remove object association and 1058 * return, otherwise delay object association removal. 1059 */ 1060 if ((m->flags & PG_FICTITIOUS) != 0) { 1061 splx(s); 1062 return; 1063 } 1064 1065 m->valid = 0; 1066 vm_page_undirty(m); 1067 1068 if (m->wire_count != 0) { 1069 if (m->wire_count > 1) { 1070 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1071 m->wire_count, (long)m->pindex); 1072 } 1073 panic("vm_page_free: freeing wired page\n"); 1074 } 1075 1076 /* 1077 * If we've exhausted the object's resident pages we want to free 1078 * it up. 1079 */ 1080 if (object && 1081 (object->type == OBJT_VNODE) && 1082 ((object->flags & OBJ_DEAD) == 0) 1083 ) { 1084 struct vnode *vp = (struct vnode *)object->handle; 1085 1086 if (vp) { 1087 VI_LOCK(vp); 1088 if (VSHOULDFREE(vp)) 1089 vfree(vp); 1090 VI_UNLOCK(vp); 1091 } 1092 } 1093 1094 /* 1095 * Clear the UNMANAGED flag when freeing an unmanaged page. 1096 */ 1097 if (m->flags & PG_UNMANAGED) { 1098 m->flags &= ~PG_UNMANAGED; 1099 } else { 1100#ifdef __alpha__ 1101 pmap_page_is_free(m); 1102#endif 1103 } 1104 1105 if (m->hold_count != 0) { 1106 m->flags &= ~PG_ZERO; 1107 m->queue = PQ_HOLD; 1108 } else 1109 m->queue = PQ_FREE + m->pc; 1110 pq = &vm_page_queues[m->queue]; 1111 mtx_lock_spin(&vm_page_queue_free_mtx); 1112 pq->lcnt++; 1113 ++(*pq->cnt); 1114 1115 /* 1116 * Put zero'd pages on the end ( where we look for zero'd pages 1117 * first ) and non-zerod pages at the head. 1118 */ 1119 if (m->flags & PG_ZERO) { 1120 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1121 ++vm_page_zero_count; 1122 } else { 1123 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1124 } 1125 mtx_unlock_spin(&vm_page_queue_free_mtx); 1126 vm_page_free_wakeup(); 1127 splx(s); 1128} 1129 1130/* 1131 * vm_page_unmanage: 1132 * 1133 * Prevent PV management from being done on the page. The page is 1134 * removed from the paging queues as if it were wired, and as a 1135 * consequence of no longer being managed the pageout daemon will not 1136 * touch it (since there is no way to locate the pte mappings for the 1137 * page). madvise() calls that mess with the pmap will also no longer 1138 * operate on the page. 1139 * 1140 * Beyond that the page is still reasonably 'normal'. Freeing the page 1141 * will clear the flag. 1142 * 1143 * This routine is used by OBJT_PHYS objects - objects using unswappable 1144 * physical memory as backing store rather then swap-backed memory and 1145 * will eventually be extended to support 4MB unmanaged physical 1146 * mappings. 1147 */ 1148void 1149vm_page_unmanage(vm_page_t m) 1150{ 1151 int s; 1152 1153 s = splvm(); 1154 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1155 if ((m->flags & PG_UNMANAGED) == 0) { 1156 if (m->wire_count == 0) 1157 vm_pageq_remove(m); 1158 } 1159 vm_page_flag_set(m, PG_UNMANAGED); 1160 splx(s); 1161} 1162 1163/* 1164 * vm_page_wire: 1165 * 1166 * Mark this page as wired down by yet 1167 * another map, removing it from paging queues 1168 * as necessary. 1169 * 1170 * The page queues must be locked. 1171 * This routine may not block. 1172 */ 1173void 1174vm_page_wire(vm_page_t m) 1175{ 1176 int s; 1177 1178 /* 1179 * Only bump the wire statistics if the page is not already wired, 1180 * and only unqueue the page if it is on some queue (if it is unmanaged 1181 * it is already off the queues). 1182 */ 1183 s = splvm(); 1184 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1185 if (m->wire_count == 0) { 1186 if ((m->flags & PG_UNMANAGED) == 0) 1187 vm_pageq_remove(m); 1188 atomic_add_int(&cnt.v_wire_count, 1); 1189 } 1190 m->wire_count++; 1191 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1192 splx(s); 1193} 1194 1195/* 1196 * vm_page_unwire: 1197 * 1198 * Release one wiring of this page, potentially 1199 * enabling it to be paged again. 1200 * 1201 * Many pages placed on the inactive queue should actually go 1202 * into the cache, but it is difficult to figure out which. What 1203 * we do instead, if the inactive target is well met, is to put 1204 * clean pages at the head of the inactive queue instead of the tail. 1205 * This will cause them to be moved to the cache more quickly and 1206 * if not actively re-referenced, freed more quickly. If we just 1207 * stick these pages at the end of the inactive queue, heavy filesystem 1208 * meta-data accesses can cause an unnecessary paging load on memory bound 1209 * processes. This optimization causes one-time-use metadata to be 1210 * reused more quickly. 1211 * 1212 * BUT, if we are in a low-memory situation we have no choice but to 1213 * put clean pages on the cache queue. 1214 * 1215 * A number of routines use vm_page_unwire() to guarantee that the page 1216 * will go into either the inactive or active queues, and will NEVER 1217 * be placed in the cache - for example, just after dirtying a page. 1218 * dirty pages in the cache are not allowed. 1219 * 1220 * The page queues must be locked. 1221 * This routine may not block. 1222 */ 1223void 1224vm_page_unwire(vm_page_t m, int activate) 1225{ 1226 int s; 1227 1228 s = splvm(); 1229 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1230 if (m->wire_count > 0) { 1231 m->wire_count--; 1232 if (m->wire_count == 0) { 1233 atomic_subtract_int(&cnt.v_wire_count, 1); 1234 if (m->flags & PG_UNMANAGED) { 1235 ; 1236 } else if (activate) 1237 vm_pageq_enqueue(PQ_ACTIVE, m); 1238 else { 1239 vm_page_flag_clear(m, PG_WINATCFLS); 1240 vm_pageq_enqueue(PQ_INACTIVE, m); 1241 } 1242 } 1243 } else { 1244 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1245 } 1246 splx(s); 1247} 1248 1249 1250/* 1251 * Move the specified page to the inactive queue. If the page has 1252 * any associated swap, the swap is deallocated. 1253 * 1254 * Normally athead is 0 resulting in LRU operation. athead is set 1255 * to 1 if we want this page to be 'as if it were placed in the cache', 1256 * except without unmapping it from the process address space. 1257 * 1258 * This routine may not block. 1259 */ 1260static __inline void 1261_vm_page_deactivate(vm_page_t m, int athead) 1262{ 1263 int s; 1264 1265 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1266 /* 1267 * Ignore if already inactive. 1268 */ 1269 if (m->queue == PQ_INACTIVE) 1270 return; 1271 1272 s = splvm(); 1273 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1274 if ((m->queue - m->pc) == PQ_CACHE) 1275 cnt.v_reactivated++; 1276 vm_page_flag_clear(m, PG_WINATCFLS); 1277 vm_pageq_remove(m); 1278 if (athead) 1279 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1280 else 1281 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1282 m->queue = PQ_INACTIVE; 1283 vm_page_queues[PQ_INACTIVE].lcnt++; 1284 cnt.v_inactive_count++; 1285 } 1286 splx(s); 1287} 1288 1289void 1290vm_page_deactivate(vm_page_t m) 1291{ 1292 _vm_page_deactivate(m, 0); 1293} 1294 1295/* 1296 * vm_page_try_to_cache: 1297 * 1298 * Returns 0 on failure, 1 on success 1299 */ 1300int 1301vm_page_try_to_cache(vm_page_t m) 1302{ 1303 1304 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1305 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1306 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1307 return (0); 1308 } 1309 vm_page_test_dirty(m); 1310 if (m->dirty) 1311 return (0); 1312 vm_page_cache(m); 1313 return (1); 1314} 1315 1316/* 1317 * vm_page_try_to_free() 1318 * 1319 * Attempt to free the page. If we cannot free it, we do nothing. 1320 * 1 is returned on success, 0 on failure. 1321 */ 1322int 1323vm_page_try_to_free(vm_page_t m) 1324{ 1325 1326 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1327 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1328 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1329 return (0); 1330 } 1331 vm_page_test_dirty(m); 1332 if (m->dirty) 1333 return (0); 1334 vm_page_busy(m); 1335 pmap_remove_all(m); 1336 vm_page_free(m); 1337 return (1); 1338} 1339 1340/* 1341 * vm_page_cache 1342 * 1343 * Put the specified page onto the page cache queue (if appropriate). 1344 * 1345 * This routine may not block. 1346 */ 1347void 1348vm_page_cache(vm_page_t m) 1349{ 1350 int s; 1351 1352 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1353 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1354 printf("vm_page_cache: attempting to cache busy page\n"); 1355 return; 1356 } 1357 if ((m->queue - m->pc) == PQ_CACHE) 1358 return; 1359 1360 /* 1361 * Remove all pmaps and indicate that the page is not 1362 * writeable or mapped. 1363 */ 1364 pmap_remove_all(m); 1365 if (m->dirty != 0) { 1366 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1367 (long)m->pindex); 1368 } 1369 s = splvm(); 1370 vm_pageq_remove_nowakeup(m); 1371 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1372 vm_page_free_wakeup(); 1373 splx(s); 1374} 1375 1376/* 1377 * vm_page_dontneed 1378 * 1379 * Cache, deactivate, or do nothing as appropriate. This routine 1380 * is typically used by madvise() MADV_DONTNEED. 1381 * 1382 * Generally speaking we want to move the page into the cache so 1383 * it gets reused quickly. However, this can result in a silly syndrome 1384 * due to the page recycling too quickly. Small objects will not be 1385 * fully cached. On the otherhand, if we move the page to the inactive 1386 * queue we wind up with a problem whereby very large objects 1387 * unnecessarily blow away our inactive and cache queues. 1388 * 1389 * The solution is to move the pages based on a fixed weighting. We 1390 * either leave them alone, deactivate them, or move them to the cache, 1391 * where moving them to the cache has the highest weighting. 1392 * By forcing some pages into other queues we eventually force the 1393 * system to balance the queues, potentially recovering other unrelated 1394 * space from active. The idea is to not force this to happen too 1395 * often. 1396 */ 1397void 1398vm_page_dontneed(vm_page_t m) 1399{ 1400 static int dnweight; 1401 int dnw; 1402 int head; 1403 1404 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1405 dnw = ++dnweight; 1406 1407 /* 1408 * occassionally leave the page alone 1409 */ 1410 if ((dnw & 0x01F0) == 0 || 1411 m->queue == PQ_INACTIVE || 1412 m->queue - m->pc == PQ_CACHE 1413 ) { 1414 if (m->act_count >= ACT_INIT) 1415 --m->act_count; 1416 return; 1417 } 1418 1419 if (m->dirty == 0) 1420 vm_page_test_dirty(m); 1421 1422 if (m->dirty || (dnw & 0x0070) == 0) { 1423 /* 1424 * Deactivate the page 3 times out of 32. 1425 */ 1426 head = 0; 1427 } else { 1428 /* 1429 * Cache the page 28 times out of every 32. Note that 1430 * the page is deactivated instead of cached, but placed 1431 * at the head of the queue instead of the tail. 1432 */ 1433 head = 1; 1434 } 1435 _vm_page_deactivate(m, head); 1436} 1437 1438/* 1439 * Grab a page, waiting until we are waken up due to the page 1440 * changing state. We keep on waiting, if the page continues 1441 * to be in the object. If the page doesn't exist, allocate it. 1442 * 1443 * This routine may block. 1444 */ 1445vm_page_t 1446vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1447{ 1448 vm_page_t m; 1449 int s, generation; 1450 1451 GIANT_REQUIRED; 1452retrylookup: 1453 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1454 vm_page_lock_queues(); 1455 if (m->busy || (m->flags & PG_BUSY)) { 1456 generation = object->generation; 1457 1458 s = splvm(); 1459 while ((object->generation == generation) && 1460 (m->busy || (m->flags & PG_BUSY))) { 1461 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1462 msleep(m, &vm_page_queue_mtx, PVM, "pgrbwt", 0); 1463 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1464 vm_page_unlock_queues(); 1465 splx(s); 1466 return NULL; 1467 } 1468 } 1469 vm_page_unlock_queues(); 1470 splx(s); 1471 goto retrylookup; 1472 } else { 1473 if (allocflags & VM_ALLOC_WIRED) 1474 vm_page_wire(m); 1475 vm_page_busy(m); 1476 vm_page_unlock_queues(); 1477 return m; 1478 } 1479 } 1480 1481 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1482 if (m == NULL) { 1483 VM_WAIT; 1484 if ((allocflags & VM_ALLOC_RETRY) == 0) 1485 return NULL; 1486 goto retrylookup; 1487 } 1488 1489 return m; 1490} 1491 1492/* 1493 * Mapping function for valid bits or for dirty bits in 1494 * a page. May not block. 1495 * 1496 * Inputs are required to range within a page. 1497 */ 1498__inline int 1499vm_page_bits(int base, int size) 1500{ 1501 int first_bit; 1502 int last_bit; 1503 1504 KASSERT( 1505 base + size <= PAGE_SIZE, 1506 ("vm_page_bits: illegal base/size %d/%d", base, size) 1507 ); 1508 1509 if (size == 0) /* handle degenerate case */ 1510 return (0); 1511 1512 first_bit = base >> DEV_BSHIFT; 1513 last_bit = (base + size - 1) >> DEV_BSHIFT; 1514 1515 return ((2 << last_bit) - (1 << first_bit)); 1516} 1517 1518/* 1519 * vm_page_set_validclean: 1520 * 1521 * Sets portions of a page valid and clean. The arguments are expected 1522 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1523 * of any partial chunks touched by the range. The invalid portion of 1524 * such chunks will be zero'd. 1525 * 1526 * This routine may not block. 1527 * 1528 * (base + size) must be less then or equal to PAGE_SIZE. 1529 */ 1530void 1531vm_page_set_validclean(vm_page_t m, int base, int size) 1532{ 1533 int pagebits; 1534 int frag; 1535 int endoff; 1536 1537 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1538 if (size == 0) /* handle degenerate case */ 1539 return; 1540 1541 /* 1542 * If the base is not DEV_BSIZE aligned and the valid 1543 * bit is clear, we have to zero out a portion of the 1544 * first block. 1545 */ 1546 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1547 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1548 pmap_zero_page_area(m, frag, base - frag); 1549 1550 /* 1551 * If the ending offset is not DEV_BSIZE aligned and the 1552 * valid bit is clear, we have to zero out a portion of 1553 * the last block. 1554 */ 1555 endoff = base + size; 1556 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1557 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1558 pmap_zero_page_area(m, endoff, 1559 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1560 1561 /* 1562 * Set valid, clear dirty bits. If validating the entire 1563 * page we can safely clear the pmap modify bit. We also 1564 * use this opportunity to clear the PG_NOSYNC flag. If a process 1565 * takes a write fault on a MAP_NOSYNC memory area the flag will 1566 * be set again. 1567 * 1568 * We set valid bits inclusive of any overlap, but we can only 1569 * clear dirty bits for DEV_BSIZE chunks that are fully within 1570 * the range. 1571 */ 1572 pagebits = vm_page_bits(base, size); 1573 m->valid |= pagebits; 1574#if 0 /* NOT YET */ 1575 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1576 frag = DEV_BSIZE - frag; 1577 base += frag; 1578 size -= frag; 1579 if (size < 0) 1580 size = 0; 1581 } 1582 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1583#endif 1584 m->dirty &= ~pagebits; 1585 if (base == 0 && size == PAGE_SIZE) { 1586 pmap_clear_modify(m); 1587 vm_page_flag_clear(m, PG_NOSYNC); 1588 } 1589} 1590 1591#if 0 1592 1593void 1594vm_page_set_dirty(vm_page_t m, int base, int size) 1595{ 1596 m->dirty |= vm_page_bits(base, size); 1597} 1598 1599#endif 1600 1601void 1602vm_page_clear_dirty(vm_page_t m, int base, int size) 1603{ 1604 GIANT_REQUIRED; 1605 m->dirty &= ~vm_page_bits(base, size); 1606} 1607 1608/* 1609 * vm_page_set_invalid: 1610 * 1611 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1612 * valid and dirty bits for the effected areas are cleared. 1613 * 1614 * May not block. 1615 */ 1616void 1617vm_page_set_invalid(vm_page_t m, int base, int size) 1618{ 1619 int bits; 1620 1621 GIANT_REQUIRED; 1622 bits = vm_page_bits(base, size); 1623 m->valid &= ~bits; 1624 m->dirty &= ~bits; 1625 m->object->generation++; 1626} 1627 1628/* 1629 * vm_page_zero_invalid() 1630 * 1631 * The kernel assumes that the invalid portions of a page contain 1632 * garbage, but such pages can be mapped into memory by user code. 1633 * When this occurs, we must zero out the non-valid portions of the 1634 * page so user code sees what it expects. 1635 * 1636 * Pages are most often semi-valid when the end of a file is mapped 1637 * into memory and the file's size is not page aligned. 1638 */ 1639void 1640vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1641{ 1642 int b; 1643 int i; 1644 1645 /* 1646 * Scan the valid bits looking for invalid sections that 1647 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1648 * valid bit may be set ) have already been zerod by 1649 * vm_page_set_validclean(). 1650 */ 1651 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1652 if (i == (PAGE_SIZE / DEV_BSIZE) || 1653 (m->valid & (1 << i)) 1654 ) { 1655 if (i > b) { 1656 pmap_zero_page_area(m, 1657 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1658 } 1659 b = i + 1; 1660 } 1661 } 1662 1663 /* 1664 * setvalid is TRUE when we can safely set the zero'd areas 1665 * as being valid. We can do this if there are no cache consistancy 1666 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1667 */ 1668 if (setvalid) 1669 m->valid = VM_PAGE_BITS_ALL; 1670} 1671 1672/* 1673 * vm_page_is_valid: 1674 * 1675 * Is (partial) page valid? Note that the case where size == 0 1676 * will return FALSE in the degenerate case where the page is 1677 * entirely invalid, and TRUE otherwise. 1678 * 1679 * May not block. 1680 */ 1681int 1682vm_page_is_valid(vm_page_t m, int base, int size) 1683{ 1684 int bits = vm_page_bits(base, size); 1685 1686 if (m->valid && ((m->valid & bits) == bits)) 1687 return 1; 1688 else 1689 return 0; 1690} 1691 1692/* 1693 * update dirty bits from pmap/mmu. May not block. 1694 */ 1695void 1696vm_page_test_dirty(vm_page_t m) 1697{ 1698 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1699 vm_page_dirty(m); 1700 } 1701} 1702 1703int so_zerocp_fullpage = 0; 1704 1705void 1706vm_page_cowfault(vm_page_t m) 1707{ 1708 vm_page_t mnew; 1709 vm_object_t object; 1710 vm_pindex_t pindex; 1711 1712 object = m->object; 1713 pindex = m->pindex; 1714 vm_page_busy(m); 1715 1716 retry_alloc: 1717 vm_page_remove(m); 1718 /* 1719 * An interrupt allocation is requested because the page 1720 * queues lock is held. 1721 */ 1722 mnew = vm_page_alloc(object, pindex, VM_ALLOC_INTERRUPT); 1723 if (mnew == NULL) { 1724 vm_page_insert(m, object, pindex); 1725 vm_page_unlock_queues(); 1726 VM_WAIT; 1727 vm_page_lock_queues(); 1728 goto retry_alloc; 1729 } 1730 1731 if (m->cow == 0) { 1732 /* 1733 * check to see if we raced with an xmit complete when 1734 * waiting to allocate a page. If so, put things back 1735 * the way they were 1736 */ 1737 vm_page_busy(mnew); 1738 vm_page_free(mnew); 1739 vm_page_insert(m, object, pindex); 1740 } else { /* clear COW & copy page */ 1741 if (so_zerocp_fullpage) { 1742 mnew->valid = VM_PAGE_BITS_ALL; 1743 } else { 1744 vm_page_copy(m, mnew); 1745 } 1746 vm_page_dirty(mnew); 1747 vm_page_flag_clear(mnew, PG_BUSY); 1748 } 1749} 1750 1751void 1752vm_page_cowclear(vm_page_t m) 1753{ 1754 1755 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1756 if (m->cow) { 1757 m->cow--; 1758 /* 1759 * let vm_fault add back write permission lazily 1760 */ 1761 } 1762 /* 1763 * sf_buf_free() will free the page, so we needn't do it here 1764 */ 1765} 1766 1767void 1768vm_page_cowsetup(vm_page_t m) 1769{ 1770 1771 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1772 m->cow++; 1773 pmap_page_protect(m, VM_PROT_READ); 1774} 1775 1776#include "opt_ddb.h" 1777#ifdef DDB 1778#include <sys/kernel.h> 1779 1780#include <ddb/ddb.h> 1781 1782DB_SHOW_COMMAND(page, vm_page_print_page_info) 1783{ 1784 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1785 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1786 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1787 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1788 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1789 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1790 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1791 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1792 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1793 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1794} 1795 1796DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1797{ 1798 int i; 1799 db_printf("PQ_FREE:"); 1800 for (i = 0; i < PQ_L2_SIZE; i++) { 1801 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1802 } 1803 db_printf("\n"); 1804 1805 db_printf("PQ_CACHE:"); 1806 for (i = 0; i < PQ_L2_SIZE; i++) { 1807 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1808 } 1809 db_printf("\n"); 1810 1811 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1812 vm_page_queues[PQ_ACTIVE].lcnt, 1813 vm_page_queues[PQ_INACTIVE].lcnt); 1814} 1815#endif /* DDB */ 1816