vm_page.c revision 128613
1264377Sdes/* 2224638Sbrooks * Copyright (c) 1991 Regents of the University of California. 357429Smarkm * All rights reserved. 457429Smarkm * 557429Smarkm * This code is derived from software contributed to Berkeley by 657429Smarkm * The Mach Operating System project at Carnegie-Mellon University. 760576Skris * 865674Skris * Redistribution and use in source and binary forms, with or without 965674Skris * modification, are permitted provided that the following conditions 1065674Skris * are met: 1165674Skris * 1. Redistributions of source code must retain the above copyright 1265674Skris * notice, this list of conditions and the following disclaimer. 1360576Skris * 2. Redistributions in binary form must reproduce the above copyright 1465674Skris * notice, this list of conditions and the following disclaimer in the 1565674Skris * documentation and/or other materials provided with the distribution. 1692559Sdes * 4. Neither the name of the University nor the names of its contributors 1765674Skris * may be used to endorse or promote products derived from this software 1865674Skris * without specific prior written permission. 1965674Skris * 2065674Skris * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 2165674Skris * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 2265674Skris * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 2365674Skris * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 2465674Skris * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 2565674Skris * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 2665674Skris * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 2765674Skris * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 2865674Skris * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 2965674Skris * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 3065674Skris * SUCH DAMAGE. 3165674Skris * 3265674Skris * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 3365674Skris */ 3465674Skris 3565674Skris/* 3665674Skris * Copyright (c) 1987, 1990 Carnegie-Mellon University. 3757429Smarkm * All rights reserved. 3857429Smarkm * 3957429Smarkm * Authors: Avadis Tevanian, Jr., Michael Wayne Young 4057429Smarkm * 41162856Sdes * Permission to use, copy, modify and distribute this software and 42162856Sdes * its documentation is hereby granted, provided that both the copyright 43162856Sdes * notice and this permission notice appear in all copies of the 44162856Sdes * software, derivative works or modified versions, and any portions 45262566Sdes * thereof, and that both notices appear in supporting documentation. 46162856Sdes * 4760576Skris * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 4876262Sgreen * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 49262566Sdes * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 5076262Sgreen * 51264377Sdes * Carnegie Mellon requests users of this software to return to 52264377Sdes * 5357429Smarkm * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 54149753Sdes * School of Computer Science 55149753Sdes * Carnegie Mellon University 5698941Sdes * Pittsburgh PA 15213-3890 57124211Sdes * 58124211Sdes * any improvements or extensions that they make and grant Carnegie the 59124211Sdes * rights to redistribute these changes. 6057429Smarkm */ 6192559Sdes 6292559Sdes/* 6392559Sdes * GENERAL RULES ON VM_PAGE MANIPULATION 6492559Sdes * 6592559Sdes * - a pageq mutex is required when adding or removing a page from a 66248619Sdes * page queue (vm_page_queue[]), regardless of other mutexes or the 67248619Sdes * busy state of a page. 68149753Sdes * 69262566Sdes * - a hash chain mutex is required when associating or disassociating 70262566Sdes * a page from the VM PAGE CACHE hash table (vm_page_buckets), 71262566Sdes * regardless of other mutexes or the busy state of a page. 7298684Sdes * 73255767Sdes * - either a hash chain mutex OR a busied page is required in order 74255767Sdes * to modify the page flags. A hash chain mutex must be obtained in 75255767Sdes * order to busy a page. A page's flags cannot be modified by a 76248619Sdes * hash chain mutex if the page is marked busy. 77248619Sdes * 78248619Sdes * - The object memq mutex is held when inserting or removing 79248619Sdes * pages from an object (vm_page_insert() or vm_page_remove()). This 8069591Sgreen * is different from the object's main mutex. 81248619Sdes * 82248619Sdes * Generally speaking, you have to be aware of side effects when running 83248619Sdes * vm_page ops. A vm_page_lookup() will return with the hash chain 84248619Sdes * locked, whether it was able to lookup the page or not. vm_page_free(), 85248619Sdes * vm_page_cache(), vm_page_activate(), and a number of other routines 86248619Sdes * will release the hash chain mutex for you. Intermediate manipulation 87248619Sdes * routines such as vm_page_flag_set() expect the hash chain to be held 88248619Sdes * on entry and the hash chain will remain held on return. 89248619Sdes * 90248619Sdes * pageq scanning can only occur with the pageq in question locked. 91248619Sdes * We have a known bottleneck with the active queue, but the cache 9298684Sdes * and free queues are actually arrays already. 93248619Sdes */ 94248619Sdes 95248619Sdes/* 96248619Sdes * Resident memory management module. 97248619Sdes */ 98248619Sdes 99248619Sdes#include <sys/cdefs.h> 100248619Sdes__FBSDID("$FreeBSD: head/sys/vm/vm_page.c 128613 2004-04-24 20:53:55Z alc $"); 101248619Sdes 10298684Sdes#include <sys/param.h> 103262566Sdes#include <sys/systm.h> 104262566Sdes#include <sys/lock.h> 105248619Sdes#include <sys/malloc.h> 10692559Sdes#include <sys/mutex.h> 10769591Sgreen#include <sys/proc.h> 10892559Sdes#include <sys/vmmeter.h> 10957429Smarkm#include <sys/vnode.h> 110262566Sdes 111255767Sdes#include <vm/vm.h> 112262566Sdes#include <vm/vm_param.h> 113255767Sdes#include <vm/vm_kern.h> 114255767Sdes#include <vm/vm_object.h> 115255767Sdes#include <vm/vm_page.h> 116255767Sdes#include <vm/vm_pageout.h> 117255767Sdes#include <vm/vm_pager.h> 118255767Sdes#include <vm/vm_extern.h> 119255767Sdes#include <vm/uma.h> 120255767Sdes#include <vm/uma_int.h> 121262566Sdes 122262566Sdes/* 123255767Sdes * Associated with page of user-allocatable memory is a 124262566Sdes * page structure. 125255767Sdes */ 126255767Sdes 127255767Sdesstruct mtx vm_page_queue_mtx; 128255767Sdesstruct mtx vm_page_queue_free_mtx; 129255767Sdes 130255767Sdesvm_page_t vm_page_array = 0; 131255767Sdesint vm_page_array_size = 0; 132255767Sdeslong first_page = 0; 13398684Sdesint vm_page_zero_count = 0; 134126277Sdes 13557429Smarkm/* 13692559Sdes * vm_set_page_size: 13757429Smarkm * 13899063Sdes * Sets the page size, perhaps based upon the memory 13998684Sdes * size. Must be called before any use of page-size 140126277Sdes * dependent functions. 14169591Sgreen */ 14292559Sdesvoid 14369591Sgreenvm_set_page_size(void) 14499063Sdes{ 14598684Sdes if (cnt.v_page_size == 0) 146262566Sdes cnt.v_page_size = PAGE_SIZE; 147262566Sdes if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 148262566Sdes panic("vm_set_page_size: page size not a power of two"); 149262566Sdes} 150262566Sdes 151262566Sdes/* 152262566Sdes * vm_page_startup: 153262566Sdes * 154248619Sdes * Initializes the resident memory module. 155248619Sdes * 156248619Sdes * Allocates memory for the page cells, and 157248619Sdes * for the object/offset-to-page hash table headers. 158248619Sdes * Each page cell is initialized and placed on the free list. 159248619Sdes */ 160248619Sdesvm_offset_t 161248619Sdesvm_page_startup(vm_offset_t vaddr) 162262566Sdes{ 163262566Sdes vm_offset_t mapped; 164262566Sdes vm_size_t npages; 165262566Sdes vm_paddr_t page_range; 166262566Sdes vm_paddr_t new_end; 167262566Sdes int i; 168248619Sdes vm_paddr_t pa; 169248619Sdes int nblocks; 170248619Sdes vm_paddr_t last_pa; 171126277Sdes 17298684Sdes /* the biggest memory array is the second group of pages */ 17398684Sdes vm_paddr_t end; 17498684Sdes vm_paddr_t biggestsize; 17569591Sgreen int biggestone; 17676262Sgreen 177192595Sdes vm_paddr_t total; 178192595Sdes vm_size_t bootpages; 179262566Sdes 180192595Sdes total = 0; 181192595Sdes biggestsize = 0; 182192595Sdes biggestone = 0; 18369591Sgreen nblocks = 0; 18457429Smarkm vaddr = round_page(vaddr); 18576262Sgreen 18692559Sdes for (i = 0; phys_avail[i + 1]; i += 2) { 18757429Smarkm phys_avail[i] = round_page(phys_avail[i]); 18869591Sgreen phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 18969591Sgreen } 19069591Sgreen 19157429Smarkm for (i = 0; phys_avail[i + 1]; i += 2) { 19257429Smarkm vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 19369591Sgreen 194255767Sdes if (size > biggestsize) { 19569591Sgreen biggestone = i; 19660576Skris biggestsize = size; 197255767Sdes } 19869591Sgreen ++nblocks; 199147005Sdes total += size; 20069591Sgreen } 20169591Sgreen 20260576Skris end = phys_avail[biggestone+1]; 20357429Smarkm 204255767Sdes /* 20569591Sgreen * Initialize the locks. 20657429Smarkm */ 207255767Sdes mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF); 20869591Sgreen mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 20969591Sgreen MTX_SPIN); 21069591Sgreen 21169591Sgreen /* 21257429Smarkm * Initialize the queue headers for the free queue, the active queue 21357429Smarkm * and the inactive queue. 21460576Skris */ 21560576Skris vm_pageq_init(); 21660576Skris 21760576Skris /* 218255767Sdes * Allocate memory for use when boot strapping the kernel memory 219137019Sdes * allocator. 22060576Skris */ 22160576Skris bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE; 22261212Skris new_end = end - bootpages; 22360576Skris new_end = trunc_page(new_end); 224137019Sdes mapped = pmap_map(&vaddr, new_end, end, 22569591Sgreen VM_PROT_READ | VM_PROT_WRITE); 22692559Sdes bzero((caddr_t) mapped, end - new_end); 22769591Sgreen uma_startup((caddr_t)mapped); 228224638Sbrooks 229224638Sbrooks /* 230224638Sbrooks * Compute the number of pages of memory that will be available for 231224638Sbrooks * use (taking into account the overhead of a page structure per 232224638Sbrooks * page). 233224638Sbrooks */ 23469591Sgreen first_page = phys_avail[0] / PAGE_SIZE; 235255767Sdes page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 23660576Skris npages = (total - (page_range * sizeof(struct vm_page)) - 23760576Skris (end - new_end)) / PAGE_SIZE; 23860576Skris end = new_end; 23969591Sgreen 240255767Sdes /* 24160576Skris * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 24260576Skris */ 24360576Skris vaddr += PAGE_SIZE; 24457429Smarkm 24557429Smarkm /* 24657429Smarkm * Initialize the mem entry structures now, and put them in the free 24757429Smarkm * queue. 24857429Smarkm */ 24957429Smarkm new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 25057429Smarkm mapped = pmap_map(&vaddr, new_end, end, 25157429Smarkm VM_PROT_READ | VM_PROT_WRITE); 252255767Sdes vm_page_array = (vm_page_t) mapped; 25361212Skris phys_avail[biggestone + 1] = new_end; 25461212Skris 255147005Sdes /* 256147005Sdes * Clear all of the page structures 257147005Sdes */ 258147005Sdes bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 25957429Smarkm vm_page_array_size = page_range; 26057429Smarkm 26169591Sgreen /* 26269591Sgreen * Construct the free queue(s) in descending order (by physical 26357429Smarkm * address) so that the first 16MB of physical memory is allocated 264255767Sdes * last rather than first. On large-memory machines, this avoids 26569591Sgreen * the exhaustion of low physical memory before isa_dmainit has run. 26657429Smarkm */ 26757429Smarkm cnt.v_page_count = 0; 26860576Skris cnt.v_free_count = 0; 269255767Sdes for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 27092559Sdes pa = phys_avail[i]; 271137019Sdes last_pa = phys_avail[i + 1]; 27257429Smarkm while (pa < last_pa && npages-- > 0) { 27392559Sdes vm_pageq_add_new_page(pa); 27498941Sdes pa += PAGE_SIZE; 27598941Sdes } 27698941Sdes } 27792559Sdes return (vaddr); 278149753Sdes} 27998941Sdes 280149753Sdesvoid 28192559Sdesvm_page_flag_set(vm_page_t m, unsigned short bits) 28292559Sdes{ 28392559Sdes 28492559Sdes mtx_assert(&vm_page_queue_mtx, MA_OWNED); 28592559Sdes m->flags |= bits; 28692559Sdes} 28792559Sdes 28892559Sdesvoid 28992559Sdesvm_page_flag_clear(vm_page_t m, unsigned short bits) 29092559Sdes{ 29192559Sdes 292248619Sdes mtx_assert(&vm_page_queue_mtx, MA_OWNED); 29392559Sdes m->flags &= ~bits; 29469591Sgreen} 29569591Sgreen 29669591Sgreenvoid 297248619Sdesvm_page_busy(vm_page_t m) 29869591Sgreen{ 29969591Sgreen KASSERT((m->flags & PG_BUSY) == 0, 30069591Sgreen ("vm_page_busy: page already busy!!!")); 30192559Sdes vm_page_flag_set(m, PG_BUSY); 302262566Sdes} 303262566Sdes 304262566Sdes/* 305262566Sdes * vm_page_flash: 30692559Sdes * 30792559Sdes * wakeup anyone waiting for the page. 30898941Sdes */ 30998941Sdesvoid 31098941Sdesvm_page_flash(vm_page_t m) 31198941Sdes{ 31298941Sdes if (m->flags & PG_WANTED) { 31398941Sdes vm_page_flag_clear(m, PG_WANTED); 31498941Sdes wakeup(m); 315137019Sdes } 31698941Sdes} 31792559Sdes 318137019Sdes/* 31992559Sdes * vm_page_wakeup: 32092559Sdes * 321248619Sdes * clear the PG_BUSY flag and wakeup anyone waiting for the 322248619Sdes * page. 323248619Sdes * 324248619Sdes */ 325248619Sdesvoid 32692559Sdesvm_page_wakeup(vm_page_t m) 327149753Sdes{ 328113911Sdes KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 32992559Sdes vm_page_flag_clear(m, PG_BUSY); 33092559Sdes vm_page_flash(m); 33192559Sdes} 33292559Sdes 33392559Sdesvoid 33492559Sdesvm_page_io_start(vm_page_t m) 33592559Sdes{ 33698941Sdes 337149753Sdes mtx_assert(&vm_page_queue_mtx, MA_OWNED); 338149753Sdes m->busy++; 339149753Sdes} 340149753Sdes 341149753Sdesvoid 342149753Sdesvm_page_io_finish(vm_page_t m) 343149753Sdes{ 344264377Sdes 345255767Sdes mtx_assert(&vm_page_queue_mtx, MA_OWNED); 346255767Sdes m->busy--; 347149753Sdes if (m->busy == 0) 34857429Smarkm vm_page_flash(m); 34957429Smarkm} 350248619Sdes 351248619Sdes/* 352248619Sdes * Keep page from being freed by the page daemon 353248619Sdes * much of the same effect as wiring, except much lower 354248619Sdes * overhead and should be used only for *very* temporary 355248619Sdes * holding ("wiring"). 356248619Sdes */ 357248619Sdesvoid 358248619Sdesvm_page_hold(vm_page_t mem) 359262566Sdes{ 360262566Sdes 361248619Sdes mtx_assert(&vm_page_queue_mtx, MA_OWNED); 362262566Sdes mem->hold_count++; 363262566Sdes} 364248619Sdes 36560576Skrisvoid 366262566Sdesvm_page_unhold(vm_page_t mem) 367262566Sdes{ 368262566Sdes 369248619Sdes mtx_assert(&vm_page_queue_mtx, MA_OWNED); 370248619Sdes --mem->hold_count; 371248619Sdes KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 372248619Sdes if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 373248619Sdes vm_page_free_toq(mem); 374248619Sdes} 375248619Sdes 376248619Sdes/* 377248619Sdes * vm_page_free: 378248619Sdes * 379248619Sdes * Free a page 380248619Sdes * 381248619Sdes * The clearing of PG_ZERO is a temporary safety until the code can be 382248619Sdes * reviewed to determine that PG_ZERO is being properly cleared on 383248619Sdes * write faults or maps. PG_ZERO was previously cleared in 384248619Sdes * vm_page_alloc(). 385248619Sdes */ 386248619Sdesvoid 387248619Sdesvm_page_free(vm_page_t m) 388248619Sdes{ 389248619Sdes vm_page_flag_clear(m, PG_ZERO); 39069591Sgreen vm_page_free_toq(m); 391248619Sdes vm_page_zero_idle_wakeup(); 392248619Sdes} 393248619Sdes 394248619Sdes/* 395248619Sdes * vm_page_free_zero: 396248619Sdes * 397248619Sdes * Free a page to the zerod-pages queue 398248619Sdes */ 399248619Sdesvoid 400248619Sdesvm_page_free_zero(vm_page_t m) 401262566Sdes{ 402248619Sdes vm_page_flag_set(m, PG_ZERO); 403248619Sdes vm_page_free_toq(m); 404248619Sdes} 405248619Sdes 406248619Sdes/* 407248619Sdes * vm_page_sleep_if_busy: 408262566Sdes * 40960576Skris * Sleep and release the page queues lock if PG_BUSY is set or, 41060576Skris * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the 411262566Sdes * thread slept and the page queues lock was released. 412262566Sdes * Otherwise, retains the page queues lock and returns FALSE. 413262566Sdes */ 414262566Sdesint 415262566Sdesvm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg) 416262566Sdes{ 417262566Sdes int is_object_locked; 418262566Sdes 419262566Sdes mtx_assert(&vm_page_queue_mtx, MA_OWNED); 420262566Sdes if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 421262566Sdes vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 422262566Sdes /* 423262566Sdes * Remove mtx_owned() after vm_object locking is finished. 424262566Sdes */ 42560576Skris if ((is_object_locked = m->object != NULL && 42692559Sdes mtx_owned(&m->object->mtx))) 42757429Smarkm mtx_unlock(&m->object->mtx); 428262566Sdes msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0); 429264377Sdes if (is_object_locked) 430262566Sdes mtx_lock(&m->object->mtx); 43192559Sdes return (TRUE); 43257429Smarkm } 43357429Smarkm return (FALSE); 43469591Sgreen} 43569591Sgreen 43669591Sgreen/* 43769591Sgreen * vm_page_dirty: 43857429Smarkm * 43960576Skris * make page all dirty 440255767Sdes */ 441137019Sdesvoid 44257429Smarkmvm_page_dirty(vm_page_t m) 44376262Sgreen{ 44457429Smarkm KASSERT(m->queue - m->pc != PQ_CACHE, 445264377Sdes ("vm_page_dirty: page in cache!")); 446264377Sdes KASSERT(m->queue - m->pc != PQ_FREE, 447264377Sdes ("vm_page_dirty: page is free!")); 44857429Smarkm m->dirty = VM_PAGE_BITS_ALL; 449137019Sdes} 45057429Smarkm 451264377Sdes/* 45257429Smarkm * vm_page_splay: 45392559Sdes * 45492559Sdes * Implements Sleator and Tarjan's top-down splay algorithm. Returns 45598684Sdes * the vm_page containing the given pindex. If, however, that 45698684Sdes * pindex is not found in the vm_object, returns a vm_page that is 45798684Sdes * adjacent to the pindex, coming before or after it. 45898684Sdes */ 45998684Sdesvm_page_t 46098684Sdesvm_page_splay(vm_pindex_t pindex, vm_page_t root) 461126277Sdes{ 46298684Sdes struct vm_page dummy; 463255767Sdes vm_page_t lefttreemax, righttreemin, y; 46498684Sdes 46598684Sdes if (root == NULL) 46698684Sdes return (root); 46798684Sdes lefttreemax = righttreemin = &dummy; 468262566Sdes for (;; root = y) { 469262566Sdes if (pindex < root->pindex) { 47098684Sdes if ((y = root->left) == NULL) 47198684Sdes break; 47298684Sdes if (pindex < y->pindex) { 47398684Sdes /* Rotate right. */ 47498684Sdes root->left = y->right; 47598684Sdes y->right = root; 47698684Sdes root = y; 47798684Sdes if ((y = root->left) == NULL) 478255767Sdes break; 47998684Sdes } 48098684Sdes /* Link into the new root's right tree. */ 481262566Sdes righttreemin->left = root; 482262566Sdes righttreemin = root; 483262566Sdes } else if (pindex > root->pindex) { 484262566Sdes if ((y = root->right) == NULL) 485262566Sdes break; 486262566Sdes if (pindex > y->pindex) { 48798684Sdes /* Rotate left. */ 488224638Sbrooks root->right = y->left; 489224638Sbrooks y->left = root; 490224638Sbrooks root = y; 49198684Sdes if ((y = root->right) == NULL) 49298684Sdes break; 49398684Sdes } 49498684Sdes /* Link into the new root's left tree. */ 495149753Sdes lefttreemax->right = root; 49698684Sdes lefttreemax = root; 497149753Sdes } else 49898684Sdes break; 49998684Sdes } 500157019Sdes /* Assemble the new root. */ 501124211Sdes lefttreemax->right = root->left; 502124211Sdes righttreemin->left = root->right; 503124211Sdes root->left = dummy.right; 50498684Sdes root->right = dummy.left; 505248619Sdes return (root); 506124211Sdes} 507124211Sdes 508124211Sdes/* 509248619Sdes * vm_page_insert: [ internal use only ] 510248619Sdes * 51198684Sdes * Inserts the given mem entry into the object and object list. 512124211Sdes * 513124211Sdes * The pagetables are not updated but will presumably fault the page 514124211Sdes * in if necessary, or if a kernel page the caller will at some point 51598684Sdes * enter the page into the kernel's pmap. We are not allowed to block 51698684Sdes * here so we *can't* do this anyway. 51798684Sdes * 51898684Sdes * The object and page must be locked, and must be splhigh. 51998684Sdes * This routine may not block. 52098684Sdes */ 52198684Sdesvoid 52298684Sdesvm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 523255767Sdes{ 52498684Sdes vm_page_t root; 52598684Sdes 526262566Sdes VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 527262566Sdes if (m->object != NULL) 528262566Sdes panic("vm_page_insert: page already inserted"); 52998684Sdes 530224638Sbrooks /* 531224638Sbrooks * Record the object/offset pair in this page 532224638Sbrooks */ 53398684Sdes m->object = object; 53498684Sdes m->pindex = pindex; 53598684Sdes 53698684Sdes /* 53798684Sdes * Now link into the object's ordered list of backed pages. 53898684Sdes */ 539157019Sdes root = object->root; 540124211Sdes if (root == NULL) { 541124211Sdes m->left = NULL; 542124211Sdes m->right = NULL; 54398684Sdes TAILQ_INSERT_TAIL(&object->memq, m, listq); 544248619Sdes } else { 545124211Sdes root = vm_page_splay(pindex, root); 546124211Sdes if (pindex < root->pindex) { 547124211Sdes m->left = root->left; 548248619Sdes m->right = root; 549248619Sdes root->left = NULL; 55098684Sdes TAILQ_INSERT_BEFORE(root, m, listq); 551124211Sdes } else if (pindex == root->pindex) 552124211Sdes panic("vm_page_insert: offset already allocated"); 553124211Sdes else { 55498684Sdes m->right = root->right; 55598684Sdes m->left = root; 55698684Sdes root->right = NULL; 55798684Sdes TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 55898684Sdes } 55998684Sdes } 560126277Sdes object->root = m; 56198684Sdes object->generation++; 562255767Sdes 56398684Sdes /* 56498684Sdes * show that the object has one more resident page. 565248619Sdes */ 56698684Sdes object->resident_page_count++; 56798684Sdes 56898684Sdes /* 56998684Sdes * Since we are inserting a new and possibly dirty page, 57098684Sdes * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 57198684Sdes */ 57298684Sdes if (m->flags & PG_WRITEABLE) 57398684Sdes vm_object_set_writeable_dirty(object); 57498684Sdes} 57598684Sdes 57698684Sdes/* 577255767Sdes * vm_page_remove: 57898684Sdes * NOTE: used by device pager as well -wfj 57998684Sdes * 580248619Sdes * Removes the given mem entry from the object/offset-page 58198684Sdes * table and the object page list, but do not invalidate/terminate 58298684Sdes * the backing store. 58398684Sdes * 58498684Sdes * The object and page must be locked, and at splhigh. 585 * The underlying pmap entry (if any) is NOT removed here. 586 * This routine may not block. 587 */ 588void 589vm_page_remove(vm_page_t m) 590{ 591 vm_object_t object; 592 vm_page_t root; 593 594 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 595 if (m->object == NULL) 596 return; 597#ifndef __alpha__ 598 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 599#endif 600 if ((m->flags & PG_BUSY) == 0) { 601 panic("vm_page_remove: page not busy"); 602 } 603 604 /* 605 * Basically destroy the page. 606 */ 607 vm_page_wakeup(m); 608 609 object = m->object; 610 611 /* 612 * Now remove from the object's list of backed pages. 613 */ 614 if (m != object->root) 615 vm_page_splay(m->pindex, object->root); 616 if (m->left == NULL) 617 root = m->right; 618 else { 619 root = vm_page_splay(m->pindex, m->left); 620 root->right = m->right; 621 } 622 object->root = root; 623 TAILQ_REMOVE(&object->memq, m, listq); 624 625 /* 626 * And show that the object has one fewer resident page. 627 */ 628 object->resident_page_count--; 629 object->generation++; 630 631 m->object = NULL; 632} 633 634/* 635 * vm_page_lookup: 636 * 637 * Returns the page associated with the object/offset 638 * pair specified; if none is found, NULL is returned. 639 * 640 * The object must be locked. 641 * This routine may not block. 642 * This is a critical path routine 643 */ 644vm_page_t 645vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 646{ 647 vm_page_t m; 648 649 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 650 if ((m = object->root) != NULL && m->pindex != pindex) { 651 m = vm_page_splay(pindex, m); 652 if ((object->root = m)->pindex != pindex) 653 m = NULL; 654 } 655 return (m); 656} 657 658/* 659 * vm_page_rename: 660 * 661 * Move the given memory entry from its 662 * current object to the specified target object/offset. 663 * 664 * The object must be locked. 665 * This routine may not block. 666 * 667 * Note: this routine will raise itself to splvm(), the caller need not. 668 * 669 * Note: swap associated with the page must be invalidated by the move. We 670 * have to do this for several reasons: (1) we aren't freeing the 671 * page, (2) we are dirtying the page, (3) the VM system is probably 672 * moving the page from object A to B, and will then later move 673 * the backing store from A to B and we can't have a conflict. 674 * 675 * Note: we *always* dirty the page. It is necessary both for the 676 * fact that we moved it, and because we may be invalidating 677 * swap. If the page is on the cache, we have to deactivate it 678 * or vm_page_dirty() will panic. Dirty pages are not allowed 679 * on the cache. 680 */ 681void 682vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 683{ 684 int s; 685 686 s = splvm(); 687 vm_page_remove(m); 688 vm_page_insert(m, new_object, new_pindex); 689 if (m->queue - m->pc == PQ_CACHE) 690 vm_page_deactivate(m); 691 vm_page_dirty(m); 692 splx(s); 693} 694 695/* 696 * vm_page_select_cache: 697 * 698 * Find a page on the cache queue with color optimization. As pages 699 * might be found, but not applicable, they are deactivated. This 700 * keeps us from using potentially busy cached pages. 701 * 702 * This routine must be called at splvm(). 703 * This routine may not block. 704 */ 705vm_page_t 706vm_page_select_cache(int color) 707{ 708 vm_page_t m; 709 710 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 711 while (TRUE) { 712 m = vm_pageq_find(PQ_CACHE, color, FALSE); 713 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 714 m->hold_count || m->wire_count || 715 (!VM_OBJECT_TRYLOCK(m->object) && 716 !VM_OBJECT_LOCKED(m->object)))) { 717 vm_page_deactivate(m); 718 continue; 719 } 720 return m; 721 } 722} 723 724/* 725 * vm_page_alloc: 726 * 727 * Allocate and return a memory cell associated 728 * with this VM object/offset pair. 729 * 730 * page_req classes: 731 * VM_ALLOC_NORMAL normal process request 732 * VM_ALLOC_SYSTEM system *really* needs a page 733 * VM_ALLOC_INTERRUPT interrupt time request 734 * VM_ALLOC_ZERO zero page 735 * 736 * This routine may not block. 737 * 738 * Additional special handling is required when called from an 739 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 740 * the page cache in this case. 741 */ 742vm_page_t 743vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 744{ 745 vm_object_t m_object; 746 vm_page_t m = NULL; 747 int color, flags, page_req, s; 748 749 page_req = req & VM_ALLOC_CLASS_MASK; 750 751 if ((req & VM_ALLOC_NOOBJ) == 0) { 752 KASSERT(object != NULL, 753 ("vm_page_alloc: NULL object.")); 754 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 755 color = (pindex + object->pg_color) & PQ_L2_MASK; 756 } else 757 color = pindex & PQ_L2_MASK; 758 759 /* 760 * The pager is allowed to eat deeper into the free page list. 761 */ 762 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 763 page_req = VM_ALLOC_SYSTEM; 764 }; 765 766 s = splvm(); 767loop: 768 mtx_lock_spin(&vm_page_queue_free_mtx); 769 if (cnt.v_free_count > cnt.v_free_reserved || 770 (page_req == VM_ALLOC_SYSTEM && 771 cnt.v_cache_count == 0 && 772 cnt.v_free_count > cnt.v_interrupt_free_min) || 773 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 774 /* 775 * Allocate from the free queue if the number of free pages 776 * exceeds the minimum for the request class. 777 */ 778 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 779 } else if (page_req != VM_ALLOC_INTERRUPT) { 780 mtx_unlock_spin(&vm_page_queue_free_mtx); 781 /* 782 * Allocatable from cache (non-interrupt only). On success, 783 * we must free the page and try again, thus ensuring that 784 * cnt.v_*_free_min counters are replenished. 785 */ 786 vm_page_lock_queues(); 787 if ((m = vm_page_select_cache(color)) == NULL) { 788 vm_page_unlock_queues(); 789 splx(s); 790#if defined(DIAGNOSTIC) 791 if (cnt.v_cache_count > 0) 792 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 793#endif 794 atomic_add_int(&vm_pageout_deficit, 1); 795 pagedaemon_wakeup(); 796 return (NULL); 797 } 798 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 799 m_object = m->object; 800 VM_OBJECT_LOCK_ASSERT(m_object, MA_OWNED); 801 vm_page_busy(m); 802 pmap_remove_all(m); 803 vm_page_free(m); 804 vm_page_unlock_queues(); 805 if (m_object != object) 806 VM_OBJECT_UNLOCK(m_object); 807 goto loop; 808 } else { 809 /* 810 * Not allocatable from cache from interrupt, give up. 811 */ 812 mtx_unlock_spin(&vm_page_queue_free_mtx); 813 splx(s); 814 atomic_add_int(&vm_pageout_deficit, 1); 815 pagedaemon_wakeup(); 816 return (NULL); 817 } 818 819 /* 820 * At this point we had better have found a good page. 821 */ 822 823 KASSERT( 824 m != NULL, 825 ("vm_page_alloc(): missing page on free queue\n") 826 ); 827 828 /* 829 * Remove from free queue 830 */ 831 832 vm_pageq_remove_nowakeup(m); 833 834 /* 835 * Initialize structure. Only the PG_ZERO flag is inherited. 836 */ 837 flags = PG_BUSY; 838 if (m->flags & PG_ZERO) { 839 vm_page_zero_count--; 840 if (req & VM_ALLOC_ZERO) 841 flags = PG_ZERO | PG_BUSY; 842 } 843 m->flags = flags; 844 if (req & VM_ALLOC_WIRED) { 845 atomic_add_int(&cnt.v_wire_count, 1); 846 m->wire_count = 1; 847 } else 848 m->wire_count = 0; 849 m->hold_count = 0; 850 m->act_count = 0; 851 m->busy = 0; 852 m->valid = 0; 853 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 854 mtx_unlock_spin(&vm_page_queue_free_mtx); 855 856 /* 857 * vm_page_insert() is safe prior to the splx(). Note also that 858 * inserting a page here does not insert it into the pmap (which 859 * could cause us to block allocating memory). We cannot block 860 * anywhere. 861 */ 862 if ((req & VM_ALLOC_NOOBJ) == 0) 863 vm_page_insert(m, object, pindex); 864 else 865 m->pindex = pindex; 866 867 /* 868 * Don't wakeup too often - wakeup the pageout daemon when 869 * we would be nearly out of memory. 870 */ 871 if (vm_paging_needed()) 872 pagedaemon_wakeup(); 873 874 splx(s); 875 return (m); 876} 877 878/* 879 * vm_wait: (also see VM_WAIT macro) 880 * 881 * Block until free pages are available for allocation 882 * - Called in various places before memory allocations. 883 */ 884void 885vm_wait(void) 886{ 887 int s; 888 889 s = splvm(); 890 vm_page_lock_queues(); 891 if (curproc == pageproc) { 892 vm_pageout_pages_needed = 1; 893 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx, 894 PDROP | PSWP, "VMWait", 0); 895 } else { 896 if (!vm_pages_needed) { 897 vm_pages_needed = 1; 898 wakeup(&vm_pages_needed); 899 } 900 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM, 901 "vmwait", 0); 902 } 903 splx(s); 904} 905 906/* 907 * vm_waitpfault: (also see VM_WAITPFAULT macro) 908 * 909 * Block until free pages are available for allocation 910 * - Called only in vm_fault so that processes page faulting 911 * can be easily tracked. 912 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 913 * processes will be able to grab memory first. Do not change 914 * this balance without careful testing first. 915 */ 916void 917vm_waitpfault(void) 918{ 919 int s; 920 921 s = splvm(); 922 vm_page_lock_queues(); 923 if (!vm_pages_needed) { 924 vm_pages_needed = 1; 925 wakeup(&vm_pages_needed); 926 } 927 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER, 928 "pfault", 0); 929 splx(s); 930} 931 932/* 933 * vm_page_activate: 934 * 935 * Put the specified page on the active list (if appropriate). 936 * Ensure that act_count is at least ACT_INIT but do not otherwise 937 * mess with it. 938 * 939 * The page queues must be locked. 940 * This routine may not block. 941 */ 942void 943vm_page_activate(vm_page_t m) 944{ 945 int s; 946 947 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 948 s = splvm(); 949 if (m->queue != PQ_ACTIVE) { 950 if ((m->queue - m->pc) == PQ_CACHE) 951 cnt.v_reactivated++; 952 vm_pageq_remove(m); 953 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 954 if (m->act_count < ACT_INIT) 955 m->act_count = ACT_INIT; 956 vm_pageq_enqueue(PQ_ACTIVE, m); 957 } 958 } else { 959 if (m->act_count < ACT_INIT) 960 m->act_count = ACT_INIT; 961 } 962 splx(s); 963} 964 965/* 966 * vm_page_free_wakeup: 967 * 968 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 969 * routine is called when a page has been added to the cache or free 970 * queues. 971 * 972 * This routine may not block. 973 * This routine must be called at splvm() 974 */ 975static __inline void 976vm_page_free_wakeup(void) 977{ 978 979 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 980 /* 981 * if pageout daemon needs pages, then tell it that there are 982 * some free. 983 */ 984 if (vm_pageout_pages_needed && 985 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 986 wakeup(&vm_pageout_pages_needed); 987 vm_pageout_pages_needed = 0; 988 } 989 /* 990 * wakeup processes that are waiting on memory if we hit a 991 * high water mark. And wakeup scheduler process if we have 992 * lots of memory. this process will swapin processes. 993 */ 994 if (vm_pages_needed && !vm_page_count_min()) { 995 vm_pages_needed = 0; 996 wakeup(&cnt.v_free_count); 997 } 998} 999 1000/* 1001 * vm_page_free_toq: 1002 * 1003 * Returns the given page to the PQ_FREE list, 1004 * disassociating it with any VM object. 1005 * 1006 * Object and page must be locked prior to entry. 1007 * This routine may not block. 1008 */ 1009 1010void 1011vm_page_free_toq(vm_page_t m) 1012{ 1013 int s; 1014 struct vpgqueues *pq; 1015 vm_object_t object = m->object; 1016 1017 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1018 s = splvm(); 1019 cnt.v_tfree++; 1020 1021 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1022 printf( 1023 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1024 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1025 m->hold_count); 1026 if ((m->queue - m->pc) == PQ_FREE) 1027 panic("vm_page_free: freeing free page"); 1028 else 1029 panic("vm_page_free: freeing busy page"); 1030 } 1031 1032 /* 1033 * unqueue, then remove page. Note that we cannot destroy 1034 * the page here because we do not want to call the pager's 1035 * callback routine until after we've put the page on the 1036 * appropriate free queue. 1037 */ 1038 vm_pageq_remove_nowakeup(m); 1039 vm_page_remove(m); 1040 1041 /* 1042 * If fictitious remove object association and 1043 * return, otherwise delay object association removal. 1044 */ 1045 if ((m->flags & PG_FICTITIOUS) != 0) { 1046 splx(s); 1047 return; 1048 } 1049 1050 m->valid = 0; 1051 vm_page_undirty(m); 1052 1053 if (m->wire_count != 0) { 1054 if (m->wire_count > 1) { 1055 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1056 m->wire_count, (long)m->pindex); 1057 } 1058 panic("vm_page_free: freeing wired page\n"); 1059 } 1060 1061 /* 1062 * If we've exhausted the object's resident pages we want to free 1063 * it up. 1064 */ 1065 if (object && 1066 (object->type == OBJT_VNODE) && 1067 ((object->flags & OBJ_DEAD) == 0) 1068 ) { 1069 struct vnode *vp = (struct vnode *)object->handle; 1070 1071 if (vp) { 1072 VI_LOCK(vp); 1073 if (VSHOULDFREE(vp)) 1074 vfree(vp); 1075 VI_UNLOCK(vp); 1076 } 1077 } 1078 1079 /* 1080 * Clear the UNMANAGED flag when freeing an unmanaged page. 1081 */ 1082 if (m->flags & PG_UNMANAGED) { 1083 m->flags &= ~PG_UNMANAGED; 1084 } 1085 1086 if (m->hold_count != 0) { 1087 m->flags &= ~PG_ZERO; 1088 m->queue = PQ_HOLD; 1089 } else 1090 m->queue = PQ_FREE + m->pc; 1091 pq = &vm_page_queues[m->queue]; 1092 mtx_lock_spin(&vm_page_queue_free_mtx); 1093 pq->lcnt++; 1094 ++(*pq->cnt); 1095 1096 /* 1097 * Put zero'd pages on the end ( where we look for zero'd pages 1098 * first ) and non-zerod pages at the head. 1099 */ 1100 if (m->flags & PG_ZERO) { 1101 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1102 ++vm_page_zero_count; 1103 } else { 1104 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1105 } 1106 mtx_unlock_spin(&vm_page_queue_free_mtx); 1107 vm_page_free_wakeup(); 1108 splx(s); 1109} 1110 1111/* 1112 * vm_page_unmanage: 1113 * 1114 * Prevent PV management from being done on the page. The page is 1115 * removed from the paging queues as if it were wired, and as a 1116 * consequence of no longer being managed the pageout daemon will not 1117 * touch it (since there is no way to locate the pte mappings for the 1118 * page). madvise() calls that mess with the pmap will also no longer 1119 * operate on the page. 1120 * 1121 * Beyond that the page is still reasonably 'normal'. Freeing the page 1122 * will clear the flag. 1123 * 1124 * This routine is used by OBJT_PHYS objects - objects using unswappable 1125 * physical memory as backing store rather then swap-backed memory and 1126 * will eventually be extended to support 4MB unmanaged physical 1127 * mappings. 1128 */ 1129void 1130vm_page_unmanage(vm_page_t m) 1131{ 1132 int s; 1133 1134 s = splvm(); 1135 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1136 if ((m->flags & PG_UNMANAGED) == 0) { 1137 if (m->wire_count == 0) 1138 vm_pageq_remove(m); 1139 } 1140 vm_page_flag_set(m, PG_UNMANAGED); 1141 splx(s); 1142} 1143 1144/* 1145 * vm_page_wire: 1146 * 1147 * Mark this page as wired down by yet 1148 * another map, removing it from paging queues 1149 * as necessary. 1150 * 1151 * The page queues must be locked. 1152 * This routine may not block. 1153 */ 1154void 1155vm_page_wire(vm_page_t m) 1156{ 1157 int s; 1158 1159 /* 1160 * Only bump the wire statistics if the page is not already wired, 1161 * and only unqueue the page if it is on some queue (if it is unmanaged 1162 * it is already off the queues). 1163 */ 1164 s = splvm(); 1165 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1166 if (m->wire_count == 0) { 1167 if ((m->flags & PG_UNMANAGED) == 0) 1168 vm_pageq_remove(m); 1169 atomic_add_int(&cnt.v_wire_count, 1); 1170 } 1171 m->wire_count++; 1172 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1173 splx(s); 1174} 1175 1176/* 1177 * vm_page_unwire: 1178 * 1179 * Release one wiring of this page, potentially 1180 * enabling it to be paged again. 1181 * 1182 * Many pages placed on the inactive queue should actually go 1183 * into the cache, but it is difficult to figure out which. What 1184 * we do instead, if the inactive target is well met, is to put 1185 * clean pages at the head of the inactive queue instead of the tail. 1186 * This will cause them to be moved to the cache more quickly and 1187 * if not actively re-referenced, freed more quickly. If we just 1188 * stick these pages at the end of the inactive queue, heavy filesystem 1189 * meta-data accesses can cause an unnecessary paging load on memory bound 1190 * processes. This optimization causes one-time-use metadata to be 1191 * reused more quickly. 1192 * 1193 * BUT, if we are in a low-memory situation we have no choice but to 1194 * put clean pages on the cache queue. 1195 * 1196 * A number of routines use vm_page_unwire() to guarantee that the page 1197 * will go into either the inactive or active queues, and will NEVER 1198 * be placed in the cache - for example, just after dirtying a page. 1199 * dirty pages in the cache are not allowed. 1200 * 1201 * The page queues must be locked. 1202 * This routine may not block. 1203 */ 1204void 1205vm_page_unwire(vm_page_t m, int activate) 1206{ 1207 int s; 1208 1209 s = splvm(); 1210 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1211 if (m->wire_count > 0) { 1212 m->wire_count--; 1213 if (m->wire_count == 0) { 1214 atomic_subtract_int(&cnt.v_wire_count, 1); 1215 if (m->flags & PG_UNMANAGED) { 1216 ; 1217 } else if (activate) 1218 vm_pageq_enqueue(PQ_ACTIVE, m); 1219 else { 1220 vm_page_flag_clear(m, PG_WINATCFLS); 1221 vm_pageq_enqueue(PQ_INACTIVE, m); 1222 } 1223 } 1224 } else { 1225 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1226 } 1227 splx(s); 1228} 1229 1230 1231/* 1232 * Move the specified page to the inactive queue. If the page has 1233 * any associated swap, the swap is deallocated. 1234 * 1235 * Normally athead is 0 resulting in LRU operation. athead is set 1236 * to 1 if we want this page to be 'as if it were placed in the cache', 1237 * except without unmapping it from the process address space. 1238 * 1239 * This routine may not block. 1240 */ 1241static __inline void 1242_vm_page_deactivate(vm_page_t m, int athead) 1243{ 1244 int s; 1245 1246 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1247 /* 1248 * Ignore if already inactive. 1249 */ 1250 if (m->queue == PQ_INACTIVE) 1251 return; 1252 1253 s = splvm(); 1254 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1255 if ((m->queue - m->pc) == PQ_CACHE) 1256 cnt.v_reactivated++; 1257 vm_page_flag_clear(m, PG_WINATCFLS); 1258 vm_pageq_remove(m); 1259 if (athead) 1260 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1261 else 1262 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1263 m->queue = PQ_INACTIVE; 1264 vm_page_queues[PQ_INACTIVE].lcnt++; 1265 cnt.v_inactive_count++; 1266 } 1267 splx(s); 1268} 1269 1270void 1271vm_page_deactivate(vm_page_t m) 1272{ 1273 _vm_page_deactivate(m, 0); 1274} 1275 1276/* 1277 * vm_page_try_to_cache: 1278 * 1279 * Returns 0 on failure, 1 on success 1280 */ 1281int 1282vm_page_try_to_cache(vm_page_t m) 1283{ 1284 1285 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1286 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1287 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1288 return (0); 1289 } 1290 pmap_remove_all(m); 1291 if (m->dirty) 1292 return (0); 1293 vm_page_cache(m); 1294 return (1); 1295} 1296 1297/* 1298 * vm_page_try_to_free() 1299 * 1300 * Attempt to free the page. If we cannot free it, we do nothing. 1301 * 1 is returned on success, 0 on failure. 1302 */ 1303int 1304vm_page_try_to_free(vm_page_t m) 1305{ 1306 1307 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1308 if (m->object != NULL) 1309 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1310 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1311 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1312 return (0); 1313 } 1314 pmap_remove_all(m); 1315 if (m->dirty) 1316 return (0); 1317 vm_page_busy(m); 1318 vm_page_free(m); 1319 return (1); 1320} 1321 1322/* 1323 * vm_page_cache 1324 * 1325 * Put the specified page onto the page cache queue (if appropriate). 1326 * 1327 * This routine may not block. 1328 */ 1329void 1330vm_page_cache(vm_page_t m) 1331{ 1332 int s; 1333 1334 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1335 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1336 m->hold_count || m->wire_count) { 1337 printf("vm_page_cache: attempting to cache busy page\n"); 1338 return; 1339 } 1340 if ((m->queue - m->pc) == PQ_CACHE) 1341 return; 1342 1343 /* 1344 * Remove all pmaps and indicate that the page is not 1345 * writeable or mapped. 1346 */ 1347 pmap_remove_all(m); 1348 if (m->dirty != 0) { 1349 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1350 (long)m->pindex); 1351 } 1352 s = splvm(); 1353 vm_pageq_remove_nowakeup(m); 1354 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1355 vm_page_free_wakeup(); 1356 splx(s); 1357} 1358 1359/* 1360 * vm_page_dontneed 1361 * 1362 * Cache, deactivate, or do nothing as appropriate. This routine 1363 * is typically used by madvise() MADV_DONTNEED. 1364 * 1365 * Generally speaking we want to move the page into the cache so 1366 * it gets reused quickly. However, this can result in a silly syndrome 1367 * due to the page recycling too quickly. Small objects will not be 1368 * fully cached. On the otherhand, if we move the page to the inactive 1369 * queue we wind up with a problem whereby very large objects 1370 * unnecessarily blow away our inactive and cache queues. 1371 * 1372 * The solution is to move the pages based on a fixed weighting. We 1373 * either leave them alone, deactivate them, or move them to the cache, 1374 * where moving them to the cache has the highest weighting. 1375 * By forcing some pages into other queues we eventually force the 1376 * system to balance the queues, potentially recovering other unrelated 1377 * space from active. The idea is to not force this to happen too 1378 * often. 1379 */ 1380void 1381vm_page_dontneed(vm_page_t m) 1382{ 1383 static int dnweight; 1384 int dnw; 1385 int head; 1386 1387 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1388 dnw = ++dnweight; 1389 1390 /* 1391 * occassionally leave the page alone 1392 */ 1393 if ((dnw & 0x01F0) == 0 || 1394 m->queue == PQ_INACTIVE || 1395 m->queue - m->pc == PQ_CACHE 1396 ) { 1397 if (m->act_count >= ACT_INIT) 1398 --m->act_count; 1399 return; 1400 } 1401 1402 if (m->dirty == 0 && pmap_is_modified(m)) 1403 vm_page_dirty(m); 1404 1405 if (m->dirty || (dnw & 0x0070) == 0) { 1406 /* 1407 * Deactivate the page 3 times out of 32. 1408 */ 1409 head = 0; 1410 } else { 1411 /* 1412 * Cache the page 28 times out of every 32. Note that 1413 * the page is deactivated instead of cached, but placed 1414 * at the head of the queue instead of the tail. 1415 */ 1416 head = 1; 1417 } 1418 _vm_page_deactivate(m, head); 1419} 1420 1421/* 1422 * Grab a page, waiting until we are waken up due to the page 1423 * changing state. We keep on waiting, if the page continues 1424 * to be in the object. If the page doesn't exist, allocate it. 1425 * 1426 * This routine may block. 1427 */ 1428vm_page_t 1429vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1430{ 1431 vm_page_t m; 1432 1433 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1434retrylookup: 1435 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1436 vm_page_lock_queues(); 1437 if (m->busy || (m->flags & PG_BUSY)) { 1438 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1439 VM_OBJECT_UNLOCK(object); 1440 msleep(m, &vm_page_queue_mtx, PDROP | PVM, "pgrbwt", 0); 1441 VM_OBJECT_LOCK(object); 1442 if ((allocflags & VM_ALLOC_RETRY) == 0) 1443 return (NULL); 1444 goto retrylookup; 1445 } else { 1446 if (allocflags & VM_ALLOC_WIRED) 1447 vm_page_wire(m); 1448 vm_page_busy(m); 1449 vm_page_unlock_queues(); 1450 return m; 1451 } 1452 } 1453 1454 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1455 if (m == NULL) { 1456 VM_OBJECT_UNLOCK(object); 1457 VM_WAIT; 1458 VM_OBJECT_LOCK(object); 1459 if ((allocflags & VM_ALLOC_RETRY) == 0) 1460 return NULL; 1461 goto retrylookup; 1462 } 1463 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 1464 pmap_zero_page(m); 1465 1466 return m; 1467} 1468 1469/* 1470 * Mapping function for valid bits or for dirty bits in 1471 * a page. May not block. 1472 * 1473 * Inputs are required to range within a page. 1474 */ 1475__inline int 1476vm_page_bits(int base, int size) 1477{ 1478 int first_bit; 1479 int last_bit; 1480 1481 KASSERT( 1482 base + size <= PAGE_SIZE, 1483 ("vm_page_bits: illegal base/size %d/%d", base, size) 1484 ); 1485 1486 if (size == 0) /* handle degenerate case */ 1487 return (0); 1488 1489 first_bit = base >> DEV_BSHIFT; 1490 last_bit = (base + size - 1) >> DEV_BSHIFT; 1491 1492 return ((2 << last_bit) - (1 << first_bit)); 1493} 1494 1495/* 1496 * vm_page_set_validclean: 1497 * 1498 * Sets portions of a page valid and clean. The arguments are expected 1499 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1500 * of any partial chunks touched by the range. The invalid portion of 1501 * such chunks will be zero'd. 1502 * 1503 * This routine may not block. 1504 * 1505 * (base + size) must be less then or equal to PAGE_SIZE. 1506 */ 1507void 1508vm_page_set_validclean(vm_page_t m, int base, int size) 1509{ 1510 int pagebits; 1511 int frag; 1512 int endoff; 1513 1514 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1515 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1516 if (size == 0) /* handle degenerate case */ 1517 return; 1518 1519 /* 1520 * If the base is not DEV_BSIZE aligned and the valid 1521 * bit is clear, we have to zero out a portion of the 1522 * first block. 1523 */ 1524 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1525 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1526 pmap_zero_page_area(m, frag, base - frag); 1527 1528 /* 1529 * If the ending offset is not DEV_BSIZE aligned and the 1530 * valid bit is clear, we have to zero out a portion of 1531 * the last block. 1532 */ 1533 endoff = base + size; 1534 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1535 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1536 pmap_zero_page_area(m, endoff, 1537 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1538 1539 /* 1540 * Set valid, clear dirty bits. If validating the entire 1541 * page we can safely clear the pmap modify bit. We also 1542 * use this opportunity to clear the PG_NOSYNC flag. If a process 1543 * takes a write fault on a MAP_NOSYNC memory area the flag will 1544 * be set again. 1545 * 1546 * We set valid bits inclusive of any overlap, but we can only 1547 * clear dirty bits for DEV_BSIZE chunks that are fully within 1548 * the range. 1549 */ 1550 pagebits = vm_page_bits(base, size); 1551 m->valid |= pagebits; 1552#if 0 /* NOT YET */ 1553 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1554 frag = DEV_BSIZE - frag; 1555 base += frag; 1556 size -= frag; 1557 if (size < 0) 1558 size = 0; 1559 } 1560 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1561#endif 1562 m->dirty &= ~pagebits; 1563 if (base == 0 && size == PAGE_SIZE) { 1564 pmap_clear_modify(m); 1565 vm_page_flag_clear(m, PG_NOSYNC); 1566 } 1567} 1568 1569void 1570vm_page_clear_dirty(vm_page_t m, int base, int size) 1571{ 1572 1573 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1574 m->dirty &= ~vm_page_bits(base, size); 1575} 1576 1577/* 1578 * vm_page_set_invalid: 1579 * 1580 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1581 * valid and dirty bits for the effected areas are cleared. 1582 * 1583 * May not block. 1584 */ 1585void 1586vm_page_set_invalid(vm_page_t m, int base, int size) 1587{ 1588 int bits; 1589 1590 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1591 bits = vm_page_bits(base, size); 1592 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1593 m->valid &= ~bits; 1594 m->dirty &= ~bits; 1595 m->object->generation++; 1596} 1597 1598/* 1599 * vm_page_zero_invalid() 1600 * 1601 * The kernel assumes that the invalid portions of a page contain 1602 * garbage, but such pages can be mapped into memory by user code. 1603 * When this occurs, we must zero out the non-valid portions of the 1604 * page so user code sees what it expects. 1605 * 1606 * Pages are most often semi-valid when the end of a file is mapped 1607 * into memory and the file's size is not page aligned. 1608 */ 1609void 1610vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1611{ 1612 int b; 1613 int i; 1614 1615 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1616 /* 1617 * Scan the valid bits looking for invalid sections that 1618 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1619 * valid bit may be set ) have already been zerod by 1620 * vm_page_set_validclean(). 1621 */ 1622 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1623 if (i == (PAGE_SIZE / DEV_BSIZE) || 1624 (m->valid & (1 << i)) 1625 ) { 1626 if (i > b) { 1627 pmap_zero_page_area(m, 1628 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1629 } 1630 b = i + 1; 1631 } 1632 } 1633 1634 /* 1635 * setvalid is TRUE when we can safely set the zero'd areas 1636 * as being valid. We can do this if there are no cache consistancy 1637 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1638 */ 1639 if (setvalid) 1640 m->valid = VM_PAGE_BITS_ALL; 1641} 1642 1643/* 1644 * vm_page_is_valid: 1645 * 1646 * Is (partial) page valid? Note that the case where size == 0 1647 * will return FALSE in the degenerate case where the page is 1648 * entirely invalid, and TRUE otherwise. 1649 * 1650 * May not block. 1651 */ 1652int 1653vm_page_is_valid(vm_page_t m, int base, int size) 1654{ 1655 int bits = vm_page_bits(base, size); 1656 1657 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1658 if (m->valid && ((m->valid & bits) == bits)) 1659 return 1; 1660 else 1661 return 0; 1662} 1663 1664/* 1665 * update dirty bits from pmap/mmu. May not block. 1666 */ 1667void 1668vm_page_test_dirty(vm_page_t m) 1669{ 1670 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1671 vm_page_dirty(m); 1672 } 1673} 1674 1675int so_zerocp_fullpage = 0; 1676 1677void 1678vm_page_cowfault(vm_page_t m) 1679{ 1680 vm_page_t mnew; 1681 vm_object_t object; 1682 vm_pindex_t pindex; 1683 1684 object = m->object; 1685 pindex = m->pindex; 1686 vm_page_busy(m); 1687 1688 retry_alloc: 1689 vm_page_remove(m); 1690 /* 1691 * An interrupt allocation is requested because the page 1692 * queues lock is held. 1693 */ 1694 mnew = vm_page_alloc(object, pindex, VM_ALLOC_INTERRUPT); 1695 if (mnew == NULL) { 1696 vm_page_insert(m, object, pindex); 1697 vm_page_unlock_queues(); 1698 VM_OBJECT_UNLOCK(object); 1699 VM_WAIT; 1700 VM_OBJECT_LOCK(object); 1701 vm_page_lock_queues(); 1702 goto retry_alloc; 1703 } 1704 1705 if (m->cow == 0) { 1706 /* 1707 * check to see if we raced with an xmit complete when 1708 * waiting to allocate a page. If so, put things back 1709 * the way they were 1710 */ 1711 vm_page_busy(mnew); 1712 vm_page_free(mnew); 1713 vm_page_insert(m, object, pindex); 1714 } else { /* clear COW & copy page */ 1715 if (!so_zerocp_fullpage) 1716 pmap_copy_page(m, mnew); 1717 mnew->valid = VM_PAGE_BITS_ALL; 1718 vm_page_dirty(mnew); 1719 vm_page_flag_clear(mnew, PG_BUSY); 1720 } 1721} 1722 1723void 1724vm_page_cowclear(vm_page_t m) 1725{ 1726 1727 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1728 if (m->cow) { 1729 m->cow--; 1730 /* 1731 * let vm_fault add back write permission lazily 1732 */ 1733 } 1734 /* 1735 * sf_buf_free() will free the page, so we needn't do it here 1736 */ 1737} 1738 1739void 1740vm_page_cowsetup(vm_page_t m) 1741{ 1742 1743 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1744 m->cow++; 1745 pmap_page_protect(m, VM_PROT_READ); 1746} 1747 1748#include "opt_ddb.h" 1749#ifdef DDB 1750#include <sys/kernel.h> 1751 1752#include <ddb/ddb.h> 1753 1754DB_SHOW_COMMAND(page, vm_page_print_page_info) 1755{ 1756 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1757 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1758 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1759 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1760 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1761 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1762 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1763 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1764 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1765 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1766} 1767 1768DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1769{ 1770 int i; 1771 db_printf("PQ_FREE:"); 1772 for (i = 0; i < PQ_L2_SIZE; i++) { 1773 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1774 } 1775 db_printf("\n"); 1776 1777 db_printf("PQ_CACHE:"); 1778 for (i = 0; i < PQ_L2_SIZE; i++) { 1779 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1780 } 1781 db_printf("\n"); 1782 1783 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1784 vm_page_queues[PQ_ACTIVE].lcnt, 1785 vm_page_queues[PQ_INACTIVE].lcnt); 1786} 1787#endif /* DDB */ 1788