vm_page.c revision 153385
155682Smarkm/*- 2233294Sstas * Copyright (c) 1991 Regents of the University of California. 3233294Sstas * All rights reserved. 4233294Sstas * 555682Smarkm * This code is derived from software contributed to Berkeley by 6233294Sstas * The Mach Operating System project at Carnegie-Mellon University. 755682Smarkm * 8233294Sstas * Redistribution and use in source and binary forms, with or without 9233294Sstas * modification, are permitted provided that the following conditions 10233294Sstas * are met: 1155682Smarkm * 1. Redistributions of source code must retain the above copyright 12233294Sstas * notice, this list of conditions and the following disclaimer. 13233294Sstas * 2. Redistributions in binary form must reproduce the above copyright 1455682Smarkm * notice, this list of conditions and the following disclaimer in the 15233294Sstas * documentation and/or other materials provided with the distribution. 16233294Sstas * 4. Neither the name of the University nor the names of its contributors 17233294Sstas * may be used to endorse or promote products derived from this software 1855682Smarkm * without specific prior written permission. 19233294Sstas * 20233294Sstas * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21233294Sstas * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22233294Sstas * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23233294Sstas * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24233294Sstas * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25233294Sstas * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26233294Sstas * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27233294Sstas * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28233294Sstas * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29233294Sstas * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30233294Sstas * SUCH DAMAGE. 31233294Sstas * 32233294Sstas * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 33233294Sstas */ 3455682Smarkm 3555682Smarkm/*- 3655682Smarkm * Copyright (c) 1987, 1990 Carnegie-Mellon University. 37233294Sstas * All rights reserved. 3855682Smarkm * 3955682Smarkm * Authors: Avadis Tevanian, Jr., Michael Wayne Young 4055682Smarkm * 4155682Smarkm * Permission to use, copy, modify and distribute this software and 42178825Sdfr * its documentation is hereby granted, provided that both the copyright 43178825Sdfr * notice and this permission notice appear in all copies of the 4455682Smarkm * software, derivative works or modified versions, and any portions 4555682Smarkm * thereof, and that both notices appear in supporting documentation. 4655682Smarkm * 4755682Smarkm * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 4855682Smarkm * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 4955682Smarkm * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 5055682Smarkm * 51178825Sdfr * Carnegie Mellon requests users of this software to return to 5255682Smarkm * 5355682Smarkm * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 54178825Sdfr * School of Computer Science 5555682Smarkm * Carnegie Mellon University 5655682Smarkm * Pittsburgh PA 15213-3890 5755682Smarkm * 5855682Smarkm * any improvements or extensions that they make and grant Carnegie the 5955682Smarkm * rights to redistribute these changes. 6072445Sassar */ 61102644Snectar 6255682Smarkm/* 6355682Smarkm * GENERAL RULES ON VM_PAGE MANIPULATION 6455682Smarkm * 6555682Smarkm * - a pageq mutex is required when adding or removing a page from a 6655682Smarkm * page queue (vm_page_queue[]), regardless of other mutexes or the 6772445Sassar * busy state of a page. 6855682Smarkm * 6955682Smarkm * - a hash chain mutex is required when associating or disassociating 7055682Smarkm * a page from the VM PAGE CACHE hash table (vm_page_buckets), 7155682Smarkm * regardless of other mutexes or the busy state of a page. 7255682Smarkm * 7355682Smarkm * - either a hash chain mutex OR a busied page is required in order 7455682Smarkm * to modify the page flags. A hash chain mutex must be obtained in 7555682Smarkm * order to busy a page. A page's flags cannot be modified by a 7655682Smarkm * hash chain mutex if the page is marked busy. 7755682Smarkm * 7855682Smarkm * - The object memq mutex is held when inserting or removing 7955682Smarkm * pages from an object (vm_page_insert() or vm_page_remove()). This 8055682Smarkm * is different from the object's main mutex. 8155682Smarkm * 8255682Smarkm * Generally speaking, you have to be aware of side effects when running 8355682Smarkm * vm_page ops. A vm_page_lookup() will return with the hash chain 8455682Smarkm * locked, whether it was able to lookup the page or not. vm_page_free(), 8555682Smarkm * vm_page_cache(), vm_page_activate(), and a number of other routines 8655682Smarkm * will release the hash chain mutex for you. Intermediate manipulation 8755682Smarkm * routines such as vm_page_flag_set() expect the hash chain to be held 8855682Smarkm * on entry and the hash chain will remain held on return. 8955682Smarkm * 9055682Smarkm * pageq scanning can only occur with the pageq in question locked. 9155682Smarkm * We have a known bottleneck with the active queue, but the cache 9255682Smarkm * and free queues are actually arrays already. 9355682Smarkm */ 9455682Smarkm 9555682Smarkm/* 9655682Smarkm * Resident memory management module. 9755682Smarkm */ 9855682Smarkm 9955682Smarkm#include <sys/cdefs.h> 10055682Smarkm__FBSDID("$FreeBSD: head/sys/vm/vm_page.c 153385 2005-12-13 19:59:09Z alc $"); 10155682Smarkm 10255682Smarkm#include <sys/param.h> 10355682Smarkm#include <sys/systm.h> 10455682Smarkm#include <sys/lock.h> 10555682Smarkm#include <sys/kernel.h> 10655682Smarkm#include <sys/malloc.h> 10755682Smarkm#include <sys/mutex.h> 10855682Smarkm#include <sys/proc.h> 10955682Smarkm#include <sys/sysctl.h> 11055682Smarkm#include <sys/vmmeter.h> 11155682Smarkm#include <sys/vnode.h> 11255682Smarkm 11355682Smarkm#include <vm/vm.h> 11455682Smarkm#include <vm/vm_param.h> 11555682Smarkm#include <vm/vm_kern.h> 11655682Smarkm#include <vm/vm_object.h> 11755682Smarkm#include <vm/vm_page.h> 118233294Sstas#include <vm/vm_pageout.h> 11955682Smarkm#include <vm/vm_pager.h> 12055682Smarkm#include <vm/vm_extern.h> 12155682Smarkm#include <vm/uma.h> 12255682Smarkm#include <vm/uma_int.h> 12355682Smarkm 12478527Sassar/* 125233294Sstas * Associated with page of user-allocatable memory is a 126233294Sstas * page structure. 12755682Smarkm */ 12878527Sassar 12955682Smarkmstruct mtx vm_page_queue_mtx; 13055682Smarkmstruct mtx vm_page_queue_free_mtx; 13155682Smarkm 132233294Sstasvm_page_t vm_page_array = 0; 133233294Sstasint vm_page_array_size = 0; 13455682Smarkmlong first_page = 0; 13555682Smarkmint vm_page_zero_count = 0; 13655682Smarkm 13755682Smarkmstatic int boot_pages = UMA_BOOT_PAGES; 13855682SmarkmTUNABLE_INT("vm.boot_pages", &boot_pages); 13955682SmarkmSYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0, 140233294Sstas "number of pages allocated for bootstrapping the VM system"); 14155682Smarkm 14255682Smarkm/* 14355682Smarkm * vm_set_page_size: 14455682Smarkm * 145178825Sdfr * Sets the page size, perhaps based upon the memory 146178825Sdfr * size. Must be called before any use of page-size 14755682Smarkm * dependent functions. 14855682Smarkm */ 14955682Smarkmvoid 15078527Sassarvm_set_page_size(void) 151233294Sstas{ 152233294Sstas if (cnt.v_page_size == 0) 15355682Smarkm cnt.v_page_size = PAGE_SIZE; 15478527Sassar if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 15555682Smarkm panic("vm_set_page_size: page size not a power of two"); 15655682Smarkm} 157178825Sdfr 158178825Sdfr/* 15955682Smarkm * vm_page_startup: 16055682Smarkm * 16155682Smarkm * Initializes the resident memory module. 16255682Smarkm * 16355682Smarkm * Allocates memory for the page cells, and 164102644Snectar * for the object/offset-to-page hash table headers. 16555682Smarkm * Each page cell is initialized and placed on the free list. 16655682Smarkm */ 16755682Smarkmvm_offset_t 168233294Sstasvm_page_startup(vm_offset_t vaddr) 169178825Sdfr{ 17055682Smarkm vm_offset_t mapped; 17155682Smarkm vm_size_t npages; 172233294Sstas vm_paddr_t page_range; 17355682Smarkm vm_paddr_t new_end; 174102644Snectar int i; 17555682Smarkm vm_paddr_t pa; 17655682Smarkm int nblocks; 17755682Smarkm vm_paddr_t last_pa; 178233294Sstas 17955682Smarkm /* the biggest memory array is the second group of pages */ 18055682Smarkm vm_paddr_t end; 18155682Smarkm vm_paddr_t biggestsize; 18255682Smarkm int biggestone; 18355682Smarkm 18455682Smarkm vm_paddr_t total; 18555682Smarkm 18655682Smarkm total = 0; 18755682Smarkm biggestsize = 0; 18855682Smarkm biggestone = 0; 18955682Smarkm nblocks = 0; 190102644Snectar vaddr = round_page(vaddr); 19155682Smarkm 19255682Smarkm for (i = 0; phys_avail[i + 1]; i += 2) { 19378527Sassar phys_avail[i] = round_page(phys_avail[i]); 194233294Sstas phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 195233294Sstas } 19655682Smarkm 19778527Sassar for (i = 0; phys_avail[i + 1]; i += 2) { 19855682Smarkm vm_paddr_t size = phys_avail[i + 1] - phys_avail[i]; 19955682Smarkm 20055682Smarkm if (size > biggestsize) { 20155682Smarkm biggestone = i; 20255682Smarkm biggestsize = size; 20355682Smarkm } 20455682Smarkm ++nblocks; 20555682Smarkm total += size; 20655682Smarkm } 20755682Smarkm 20855682Smarkm end = phys_avail[biggestone+1]; 20955682Smarkm 21055682Smarkm /* 21155682Smarkm * Initialize the locks. 212102644Snectar */ 213102644Snectar mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF | 21455682Smarkm MTX_RECURSE); 21555682Smarkm mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL, 21655682Smarkm MTX_SPIN); 21755682Smarkm 218233294Sstas /* 219178825Sdfr * Initialize the queue headers for the free queue, the active queue 22055682Smarkm * and the inactive queue. 22155682Smarkm */ 22255682Smarkm vm_pageq_init(); 22355682Smarkm 224233294Sstas /* 225233294Sstas * Allocate memory for use when boot strapping the kernel memory 22655682Smarkm * allocator. 22755682Smarkm */ 22855682Smarkm new_end = end - (boot_pages * UMA_SLAB_SIZE); 22955682Smarkm new_end = trunc_page(new_end); 230233294Sstas mapped = pmap_map(&vaddr, new_end, end, 231233294Sstas VM_PROT_READ | VM_PROT_WRITE); 232233294Sstas bzero((void *)mapped, end - new_end); 233233294Sstas uma_startup((void *)mapped, boot_pages); 234233294Sstas 235233294Sstas /* 236233294Sstas * Compute the number of pages of memory that will be available for 237233294Sstas * use (taking into account the overhead of a page structure per 238178825Sdfr * page). 23955682Smarkm */ 240178825Sdfr first_page = phys_avail[0] / PAGE_SIZE; 241178825Sdfr page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 24255682Smarkm npages = (total - (page_range * sizeof(struct vm_page)) - 24355682Smarkm (end - new_end)) / PAGE_SIZE; 244233294Sstas end = new_end; 24555682Smarkm 24655682Smarkm /* 24755682Smarkm * Reserve an unmapped guard page to trap access to vm_page_array[-1]. 24855682Smarkm */ 24955682Smarkm vaddr += PAGE_SIZE; 25055682Smarkm 25155682Smarkm /* 25255682Smarkm * Initialize the mem entry structures now, and put them in the free 25355682Smarkm * queue. 25455682Smarkm */ 255102644Snectar new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 25655682Smarkm mapped = pmap_map(&vaddr, new_end, end, 25755682Smarkm VM_PROT_READ | VM_PROT_WRITE); 25878527Sassar vm_page_array = (vm_page_t) mapped; 259233294Sstas phys_avail[biggestone + 1] = new_end; 260233294Sstas 26155682Smarkm /* 26278527Sassar * Clear all of the page structures 26355682Smarkm */ 26455682Smarkm bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 26555682Smarkm vm_page_array_size = page_range; 26655682Smarkm 26755682Smarkm /* 26855682Smarkm * Construct the free queue(s) in descending order (by physical 26955682Smarkm * address) so that the first 16MB of physical memory is allocated 27055682Smarkm * last rather than first. On large-memory machines, this avoids 27155682Smarkm * the exhaustion of low physical memory before isa_dma_init has run. 27255682Smarkm */ 273233294Sstas cnt.v_page_count = 0; 27478527Sassar cnt.v_free_count = 0; 27555682Smarkm for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 27655682Smarkm pa = phys_avail[i]; 27755682Smarkm last_pa = phys_avail[i + 1]; 27855682Smarkm while (pa < last_pa && npages-- > 0) { 27978527Sassar vm_pageq_add_new_page(pa); 28078527Sassar pa += PAGE_SIZE; 28155682Smarkm } 28255682Smarkm } 28355682Smarkm return (vaddr); 28455682Smarkm} 28555682Smarkm 28655682Smarkmvoid 28755682Smarkmvm_page_flag_set(vm_page_t m, unsigned short bits) 28855682Smarkm{ 28955682Smarkm 29055682Smarkm mtx_assert(&vm_page_queue_mtx, MA_OWNED); 29155682Smarkm m->flags |= bits; 29255682Smarkm} 29355682Smarkm 29478527Sassarvoid 295233294Sstasvm_page_flag_clear(vm_page_t m, unsigned short bits) 296233294Sstas{ 29778527Sassar 29878527Sassar mtx_assert(&vm_page_queue_mtx, MA_OWNED); 29955682Smarkm m->flags &= ~bits; 30055682Smarkm} 30155682Smarkm 30255682Smarkmvoid 30355682Smarkmvm_page_busy(vm_page_t m) 30455682Smarkm{ 305178825Sdfr 30655682Smarkm VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 30755682Smarkm KASSERT((m->flags & PG_BUSY) == 0, 30855682Smarkm ("vm_page_busy: page already busy!!!")); 30955682Smarkm vm_page_flag_set(m, PG_BUSY); 31078527Sassar} 311233294Sstas 312233294Sstas/* 31355682Smarkm * vm_page_flash: 31478527Sassar * 31555682Smarkm * wakeup anyone waiting for the page. 316233294Sstas */ 31755682Smarkmvoid 31878527Sassarvm_page_flash(vm_page_t m) 31978527Sassar{ 320233294Sstas 321233294Sstas VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 32278527Sassar if (m->flags & PG_WANTED) { 323178825Sdfr vm_page_flag_clear(m, PG_WANTED); 32478527Sassar wakeup(m); 32578527Sassar } 326233294Sstas} 32755682Smarkm 32855682Smarkm/* 32978527Sassar * vm_page_wakeup: 33055682Smarkm * 331233294Sstas * clear the PG_BUSY flag and wakeup anyone waiting for the 332233294Sstas * page. 333233294Sstas * 334178825Sdfr */ 33578527Sassarvoid 33655682Smarkmvm_page_wakeup(vm_page_t m) 33755682Smarkm{ 33855682Smarkm 33955682Smarkm VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 340178825Sdfr KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 34155682Smarkm vm_page_flag_clear(m, PG_BUSY); 342102644Snectar vm_page_flash(m); 343102644Snectar} 344102644Snectar 345102644Snectarvoid 346102644Snectarvm_page_io_start(vm_page_t m) 347102644Snectar{ 348102644Snectar 349102644Snectar VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 350102644Snectar m->busy++; 351102644Snectar} 352102644Snectar 353102644Snectarvoid 35455682Smarkmvm_page_io_finish(vm_page_t m) 35555682Smarkm{ 35655682Smarkm 357233294Sstas VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 358233294Sstas mtx_assert(&vm_page_queue_mtx, MA_OWNED); 35955682Smarkm m->busy--; 36055682Smarkm if (m->busy == 0) 36155682Smarkm vm_page_flash(m); 36255682Smarkm} 36355682Smarkm 364233294Sstas/* 36555682Smarkm * Keep page from being freed by the page daemon 36655682Smarkm * much of the same effect as wiring, except much lower 36755682Smarkm * overhead and should be used only for *very* temporary 36855682Smarkm * holding ("wiring"). 36955682Smarkm */ 37055682Smarkmvoid 37155682Smarkmvm_page_hold(vm_page_t mem) 37255682Smarkm{ 37355682Smarkm 37455682Smarkm mtx_assert(&vm_page_queue_mtx, MA_OWNED); 37555682Smarkm mem->hold_count++; 37655682Smarkm} 37755682Smarkm 37855682Smarkmvoid 37955682Smarkmvm_page_unhold(vm_page_t mem) 380233294Sstas{ 38155682Smarkm 38255682Smarkm mtx_assert(&vm_page_queue_mtx, MA_OWNED); 38355682Smarkm --mem->hold_count; 38455682Smarkm KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 385233294Sstas if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 38655682Smarkm vm_page_free_toq(mem); 38755682Smarkm} 388233294Sstas 38955682Smarkm/* 39055682Smarkm * vm_page_free: 39155682Smarkm * 39255682Smarkm * Free a page 39355682Smarkm * 394178825Sdfr * The clearing of PG_ZERO is a temporary safety until the code can be 395178825Sdfr * reviewed to determine that PG_ZERO is being properly cleared on 396178825Sdfr * write faults or maps. PG_ZERO was previously cleared in 397178825Sdfr * vm_page_alloc(). 398178825Sdfr */ 399178825Sdfrvoid 400178825Sdfrvm_page_free(vm_page_t m) 40155682Smarkm{ 40255682Smarkm vm_page_flag_clear(m, PG_ZERO); 40355682Smarkm vm_page_free_toq(m); 40455682Smarkm vm_page_zero_idle_wakeup(); 40555682Smarkm} 40655682Smarkm 407233294Sstas/* 40855682Smarkm * vm_page_free_zero: 40955682Smarkm * 41055682Smarkm * Free a page to the zerod-pages queue 41155682Smarkm */ 41255682Smarkmvoid 41355682Smarkmvm_page_free_zero(vm_page_t m) 41455682Smarkm{ 41555682Smarkm vm_page_flag_set(m, PG_ZERO); 416233294Sstas vm_page_free_toq(m); 417103423Snectar} 418103423Snectar 41955682Smarkm/* 420103423Snectar * vm_page_sleep_if_busy: 42155682Smarkm * 42255682Smarkm * Sleep and release the page queues lock if PG_BUSY is set or, 423103423Snectar * if also_m_busy is TRUE, busy is non-zero. Returns TRUE if the 424233294Sstas * thread slept and the page queues lock was released. 425103423Snectar * Otherwise, retains the page queues lock and returns FALSE. 426103423Snectar */ 427103423Snectarint 428103423Snectarvm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg) 429103423Snectar{ 430103423Snectar vm_object_t object; 431233294Sstas 432233294Sstas mtx_assert(&vm_page_queue_mtx, MA_OWNED); 433103423Snectar VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 434103423Snectar if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 435103423Snectar vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 436103423Snectar /* 437178825Sdfr * It's possible that while we sleep, the page will get 438103423Snectar * unbusied and freed. If we are holding the object 439103423Snectar * lock, we will assume we hold a reference to the object 440103423Snectar * such that even if m->object changes, we can re-lock 441103423Snectar * it. 442103423Snectar */ 44355682Smarkm object = m->object; 44455682Smarkm VM_OBJECT_UNLOCK(object); 44555682Smarkm msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0); 446233294Sstas VM_OBJECT_LOCK(object); 44755682Smarkm return (TRUE); 44855682Smarkm } 44955682Smarkm return (FALSE); 45055682Smarkm} 45155682Smarkm 45255682Smarkm/* 45355682Smarkm * vm_page_dirty: 454103423Snectar * 45555682Smarkm * make page all dirty 45655682Smarkm */ 457233294Sstasvoid 45855682Smarkmvm_page_dirty(vm_page_t m) 45955682Smarkm{ 46055682Smarkm KASSERT(m->queue - m->pc != PQ_CACHE, 46155682Smarkm ("vm_page_dirty: page in cache!")); 46255682Smarkm KASSERT(m->queue - m->pc != PQ_FREE, 46355682Smarkm ("vm_page_dirty: page is free!")); 46455682Smarkm m->dirty = VM_PAGE_BITS_ALL; 46555682Smarkm} 46655682Smarkm 46755682Smarkm/* 46855682Smarkm * vm_page_splay: 46955682Smarkm * 47055682Smarkm * Implements Sleator and Tarjan's top-down splay algorithm. Returns 47155682Smarkm * the vm_page containing the given pindex. If, however, that 47255682Smarkm * pindex is not found in the vm_object, returns a vm_page that is 47355682Smarkm * adjacent to the pindex, coming before or after it. 47455682Smarkm */ 47555682Smarkmvm_page_t 476233294Sstasvm_page_splay(vm_pindex_t pindex, vm_page_t root) 47755682Smarkm{ 47855682Smarkm struct vm_page dummy; 47955682Smarkm vm_page_t lefttreemax, righttreemin, y; 48055682Smarkm 48155682Smarkm if (root == NULL) 48255682Smarkm return (root); 48355682Smarkm lefttreemax = righttreemin = &dummy; 48455682Smarkm for (;; root = y) { 48555682Smarkm if (pindex < root->pindex) { 48655682Smarkm if ((y = root->left) == NULL) 48755682Smarkm break; 48855682Smarkm if (pindex < y->pindex) { 48955682Smarkm /* Rotate right. */ 49055682Smarkm root->left = y->right; 49155682Smarkm y->right = root; 49255682Smarkm root = y; 493233294Sstas if ((y = root->left) == NULL) 494233294Sstas break; 495233294Sstas } 496233294Sstas /* Link into the new root's right tree. */ 497233294Sstas righttreemin->left = root; 498233294Sstas righttreemin = root; 499233294Sstas } else if (pindex > root->pindex) { 500233294Sstas if ((y = root->right) == NULL) 501233294Sstas break; 502233294Sstas if (pindex > y->pindex) { 503233294Sstas /* Rotate left. */ 504233294Sstas root->right = y->left; 505233294Sstas y->left = root; 506233294Sstas root = y; 507233294Sstas if ((y = root->right) == NULL) 508233294Sstas break; 509233294Sstas } 510233294Sstas /* Link into the new root's left tree. */ 511233294Sstas lefttreemax->right = root; 512233294Sstas lefttreemax = root; 513233294Sstas } else 514233294Sstas break; 515233294Sstas } 516233294Sstas /* Assemble the new root. */ 517 lefttreemax->right = root->left; 518 righttreemin->left = root->right; 519 root->left = dummy.right; 520 root->right = dummy.left; 521 return (root); 522} 523 524/* 525 * vm_page_insert: [ internal use only ] 526 * 527 * Inserts the given mem entry into the object and object list. 528 * 529 * The pagetables are not updated but will presumably fault the page 530 * in if necessary, or if a kernel page the caller will at some point 531 * enter the page into the kernel's pmap. We are not allowed to block 532 * here so we *can't* do this anyway. 533 * 534 * The object and page must be locked. 535 * This routine may not block. 536 */ 537void 538vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 539{ 540 vm_page_t root; 541 542 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 543 if (m->object != NULL) 544 panic("vm_page_insert: page already inserted"); 545 546 /* 547 * Record the object/offset pair in this page 548 */ 549 m->object = object; 550 m->pindex = pindex; 551 552 /* 553 * Now link into the object's ordered list of backed pages. 554 */ 555 root = object->root; 556 if (root == NULL) { 557 m->left = NULL; 558 m->right = NULL; 559 TAILQ_INSERT_TAIL(&object->memq, m, listq); 560 } else { 561 root = vm_page_splay(pindex, root); 562 if (pindex < root->pindex) { 563 m->left = root->left; 564 m->right = root; 565 root->left = NULL; 566 TAILQ_INSERT_BEFORE(root, m, listq); 567 } else if (pindex == root->pindex) 568 panic("vm_page_insert: offset already allocated"); 569 else { 570 m->right = root->right; 571 m->left = root; 572 root->right = NULL; 573 TAILQ_INSERT_AFTER(&object->memq, root, m, listq); 574 } 575 } 576 object->root = m; 577 object->generation++; 578 579 /* 580 * show that the object has one more resident page. 581 */ 582 object->resident_page_count++; 583 /* 584 * Hold the vnode until the last page is released. 585 */ 586 if (object->resident_page_count == 1 && object->type == OBJT_VNODE) 587 vhold((struct vnode *)object->handle); 588 589 /* 590 * Since we are inserting a new and possibly dirty page, 591 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 592 */ 593 if (m->flags & PG_WRITEABLE) 594 vm_object_set_writeable_dirty(object); 595} 596 597/* 598 * vm_page_remove: 599 * NOTE: used by device pager as well -wfj 600 * 601 * Removes the given mem entry from the object/offset-page 602 * table and the object page list, but do not invalidate/terminate 603 * the backing store. 604 * 605 * The object and page must be locked. 606 * The underlying pmap entry (if any) is NOT removed here. 607 * This routine may not block. 608 */ 609void 610vm_page_remove(vm_page_t m) 611{ 612 vm_object_t object; 613 vm_page_t root; 614 615 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 616 if ((object = m->object) == NULL) 617 return; 618 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 619 if (m->flags & PG_BUSY) { 620 vm_page_flag_clear(m, PG_BUSY); 621 vm_page_flash(m); 622 } 623 624 /* 625 * Now remove from the object's list of backed pages. 626 */ 627 if (m != object->root) 628 vm_page_splay(m->pindex, object->root); 629 if (m->left == NULL) 630 root = m->right; 631 else { 632 root = vm_page_splay(m->pindex, m->left); 633 root->right = m->right; 634 } 635 object->root = root; 636 TAILQ_REMOVE(&object->memq, m, listq); 637 638 /* 639 * And show that the object has one fewer resident page. 640 */ 641 object->resident_page_count--; 642 object->generation++; 643 /* 644 * The vnode may now be recycled. 645 */ 646 if (object->resident_page_count == 0 && object->type == OBJT_VNODE) 647 vdrop((struct vnode *)object->handle); 648 649 m->object = NULL; 650} 651 652/* 653 * vm_page_lookup: 654 * 655 * Returns the page associated with the object/offset 656 * pair specified; if none is found, NULL is returned. 657 * 658 * The object must be locked. 659 * This routine may not block. 660 * This is a critical path routine 661 */ 662vm_page_t 663vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 664{ 665 vm_page_t m; 666 667 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 668 if ((m = object->root) != NULL && m->pindex != pindex) { 669 m = vm_page_splay(pindex, m); 670 if ((object->root = m)->pindex != pindex) 671 m = NULL; 672 } 673 return (m); 674} 675 676/* 677 * vm_page_rename: 678 * 679 * Move the given memory entry from its 680 * current object to the specified target object/offset. 681 * 682 * The object must be locked. 683 * This routine may not block. 684 * 685 * Note: swap associated with the page must be invalidated by the move. We 686 * have to do this for several reasons: (1) we aren't freeing the 687 * page, (2) we are dirtying the page, (3) the VM system is probably 688 * moving the page from object A to B, and will then later move 689 * the backing store from A to B and we can't have a conflict. 690 * 691 * Note: we *always* dirty the page. It is necessary both for the 692 * fact that we moved it, and because we may be invalidating 693 * swap. If the page is on the cache, we have to deactivate it 694 * or vm_page_dirty() will panic. Dirty pages are not allowed 695 * on the cache. 696 */ 697void 698vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 699{ 700 701 vm_page_remove(m); 702 vm_page_insert(m, new_object, new_pindex); 703 if (m->queue - m->pc == PQ_CACHE) 704 vm_page_deactivate(m); 705 vm_page_dirty(m); 706} 707 708/* 709 * vm_page_select_cache: 710 * 711 * Move a page of the given color from the cache queue to the free 712 * queue. As pages might be found, but are not applicable, they are 713 * deactivated. 714 * 715 * This routine may not block. 716 */ 717vm_page_t 718vm_page_select_cache(int color) 719{ 720 vm_object_t object; 721 vm_page_t m; 722 boolean_t was_trylocked; 723 724 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 725 while ((m = vm_pageq_find(PQ_CACHE, color, FALSE)) != NULL) { 726 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 727 KASSERT(!pmap_page_is_mapped(m), 728 ("Found mapped cache page %p", m)); 729 KASSERT((m->flags & PG_UNMANAGED) == 0, 730 ("Found unmanaged cache page %p", m)); 731 KASSERT(m->wire_count == 0, ("Found wired cache page %p", m)); 732 if (m->hold_count == 0 && (object = m->object, 733 (was_trylocked = VM_OBJECT_TRYLOCK(object)) || 734 VM_OBJECT_LOCKED(object))) { 735 KASSERT((m->flags & PG_BUSY) == 0 && m->busy == 0, 736 ("Found busy cache page %p", m)); 737 vm_page_free(m); 738 if (was_trylocked) 739 VM_OBJECT_UNLOCK(object); 740 break; 741 } 742 vm_page_deactivate(m); 743 } 744 return (m); 745} 746 747/* 748 * vm_page_alloc: 749 * 750 * Allocate and return a memory cell associated 751 * with this VM object/offset pair. 752 * 753 * page_req classes: 754 * VM_ALLOC_NORMAL normal process request 755 * VM_ALLOC_SYSTEM system *really* needs a page 756 * VM_ALLOC_INTERRUPT interrupt time request 757 * VM_ALLOC_ZERO zero page 758 * 759 * This routine may not block. 760 * 761 * Additional special handling is required when called from an 762 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 763 * the page cache in this case. 764 */ 765vm_page_t 766vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) 767{ 768 vm_page_t m = NULL; 769 int color, flags, page_req; 770 771 page_req = req & VM_ALLOC_CLASS_MASK; 772 KASSERT(curthread->td_intr_nesting_level == 0 || 773 page_req == VM_ALLOC_INTERRUPT, 774 ("vm_page_alloc(NORMAL|SYSTEM) in interrupt context")); 775 776 if ((req & VM_ALLOC_NOOBJ) == 0) { 777 KASSERT(object != NULL, 778 ("vm_page_alloc: NULL object.")); 779 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 780 color = (pindex + object->pg_color) & PQ_L2_MASK; 781 } else 782 color = pindex & PQ_L2_MASK; 783 784 /* 785 * The pager is allowed to eat deeper into the free page list. 786 */ 787 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 788 page_req = VM_ALLOC_SYSTEM; 789 }; 790 791loop: 792 mtx_lock_spin(&vm_page_queue_free_mtx); 793 if (cnt.v_free_count > cnt.v_free_reserved || 794 (page_req == VM_ALLOC_SYSTEM && 795 cnt.v_cache_count == 0 && 796 cnt.v_free_count > cnt.v_interrupt_free_min) || 797 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) { 798 /* 799 * Allocate from the free queue if the number of free pages 800 * exceeds the minimum for the request class. 801 */ 802 m = vm_pageq_find(PQ_FREE, color, (req & VM_ALLOC_ZERO) != 0); 803 } else if (page_req != VM_ALLOC_INTERRUPT) { 804 mtx_unlock_spin(&vm_page_queue_free_mtx); 805 /* 806 * Allocatable from cache (non-interrupt only). On success, 807 * we must free the page and try again, thus ensuring that 808 * cnt.v_*_free_min counters are replenished. 809 */ 810 vm_page_lock_queues(); 811 if ((m = vm_page_select_cache(color)) == NULL) { 812#if defined(DIAGNOSTIC) 813 if (cnt.v_cache_count > 0) 814 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 815#endif 816 vm_page_unlock_queues(); 817 atomic_add_int(&vm_pageout_deficit, 1); 818 pagedaemon_wakeup(); 819 return (NULL); 820 } 821 vm_page_unlock_queues(); 822 goto loop; 823 } else { 824 /* 825 * Not allocatable from cache from interrupt, give up. 826 */ 827 mtx_unlock_spin(&vm_page_queue_free_mtx); 828 atomic_add_int(&vm_pageout_deficit, 1); 829 pagedaemon_wakeup(); 830 return (NULL); 831 } 832 833 /* 834 * At this point we had better have found a good page. 835 */ 836 837 KASSERT( 838 m != NULL, 839 ("vm_page_alloc(): missing page on free queue") 840 ); 841 842 /* 843 * Remove from free queue 844 */ 845 vm_pageq_remove_nowakeup(m); 846 847 /* 848 * Initialize structure. Only the PG_ZERO flag is inherited. 849 */ 850 flags = PG_BUSY; 851 if (m->flags & PG_ZERO) { 852 vm_page_zero_count--; 853 if (req & VM_ALLOC_ZERO) 854 flags = PG_ZERO | PG_BUSY; 855 } 856 if (req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) 857 flags &= ~PG_BUSY; 858 m->flags = flags; 859 if (req & VM_ALLOC_WIRED) { 860 atomic_add_int(&cnt.v_wire_count, 1); 861 m->wire_count = 1; 862 } else 863 m->wire_count = 0; 864 m->hold_count = 0; 865 m->act_count = 0; 866 m->busy = 0; 867 m->valid = 0; 868 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 869 mtx_unlock_spin(&vm_page_queue_free_mtx); 870 871 if ((req & VM_ALLOC_NOOBJ) == 0) 872 vm_page_insert(m, object, pindex); 873 else 874 m->pindex = pindex; 875 876 /* 877 * Don't wakeup too often - wakeup the pageout daemon when 878 * we would be nearly out of memory. 879 */ 880 if (vm_paging_needed()) 881 pagedaemon_wakeup(); 882 883 return (m); 884} 885 886/* 887 * vm_wait: (also see VM_WAIT macro) 888 * 889 * Block until free pages are available for allocation 890 * - Called in various places before memory allocations. 891 */ 892void 893vm_wait(void) 894{ 895 896 vm_page_lock_queues(); 897 if (curproc == pageproc) { 898 vm_pageout_pages_needed = 1; 899 msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx, 900 PDROP | PSWP, "VMWait", 0); 901 } else { 902 if (!vm_pages_needed) { 903 vm_pages_needed = 1; 904 wakeup(&vm_pages_needed); 905 } 906 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM, 907 "vmwait", 0); 908 } 909} 910 911/* 912 * vm_waitpfault: (also see VM_WAITPFAULT macro) 913 * 914 * Block until free pages are available for allocation 915 * - Called only in vm_fault so that processes page faulting 916 * can be easily tracked. 917 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 918 * processes will be able to grab memory first. Do not change 919 * this balance without careful testing first. 920 */ 921void 922vm_waitpfault(void) 923{ 924 925 vm_page_lock_queues(); 926 if (!vm_pages_needed) { 927 vm_pages_needed = 1; 928 wakeup(&vm_pages_needed); 929 } 930 msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER, 931 "pfault", 0); 932} 933 934/* 935 * vm_page_activate: 936 * 937 * Put the specified page on the active list (if appropriate). 938 * Ensure that act_count is at least ACT_INIT but do not otherwise 939 * mess with it. 940 * 941 * The page queues must be locked. 942 * This routine may not block. 943 */ 944void 945vm_page_activate(vm_page_t m) 946{ 947 948 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 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} 963 964/* 965 * vm_page_free_wakeup: 966 * 967 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 968 * routine is called when a page has been added to the cache or free 969 * queues. 970 * 971 * The page queues must be locked. 972 * This routine may not block. 973 */ 974static __inline void 975vm_page_free_wakeup(void) 976{ 977 978 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 979 /* 980 * if pageout daemon needs pages, then tell it that there are 981 * some free. 982 */ 983 if (vm_pageout_pages_needed && 984 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 985 wakeup(&vm_pageout_pages_needed); 986 vm_pageout_pages_needed = 0; 987 } 988 /* 989 * wakeup processes that are waiting on memory if we hit a 990 * high water mark. And wakeup scheduler process if we have 991 * lots of memory. this process will swapin processes. 992 */ 993 if (vm_pages_needed && !vm_page_count_min()) { 994 vm_pages_needed = 0; 995 wakeup(&cnt.v_free_count); 996 } 997} 998 999/* 1000 * vm_page_free_toq: 1001 * 1002 * Returns the given page to the PQ_FREE list, 1003 * disassociating it with any VM object. 1004 * 1005 * Object and page must be locked prior to entry. 1006 * This routine may not block. 1007 */ 1008 1009void 1010vm_page_free_toq(vm_page_t m) 1011{ 1012 struct vpgqueues *pq; 1013 1014 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1015 KASSERT(!pmap_page_is_mapped(m), 1016 ("vm_page_free_toq: freeing mapped page %p", m)); 1017 cnt.v_tfree++; 1018 1019 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1020 printf( 1021 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1022 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1023 m->hold_count); 1024 if ((m->queue - m->pc) == PQ_FREE) 1025 panic("vm_page_free: freeing free page"); 1026 else 1027 panic("vm_page_free: freeing busy page"); 1028 } 1029 1030 /* 1031 * unqueue, then remove page. Note that we cannot destroy 1032 * the page here because we do not want to call the pager's 1033 * callback routine until after we've put the page on the 1034 * appropriate free queue. 1035 */ 1036 vm_pageq_remove_nowakeup(m); 1037 vm_page_remove(m); 1038 1039 /* 1040 * If fictitious remove object association and 1041 * return, otherwise delay object association removal. 1042 */ 1043 if ((m->flags & PG_FICTITIOUS) != 0) { 1044 return; 1045 } 1046 1047 m->valid = 0; 1048 vm_page_undirty(m); 1049 1050 if (m->wire_count != 0) { 1051 if (m->wire_count > 1) { 1052 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1053 m->wire_count, (long)m->pindex); 1054 } 1055 panic("vm_page_free: freeing wired page"); 1056 } 1057 1058 /* 1059 * Clear the UNMANAGED flag when freeing an unmanaged page. 1060 */ 1061 if (m->flags & PG_UNMANAGED) { 1062 m->flags &= ~PG_UNMANAGED; 1063 } 1064 1065 if (m->hold_count != 0) { 1066 m->flags &= ~PG_ZERO; 1067 m->queue = PQ_HOLD; 1068 } else 1069 m->queue = PQ_FREE + m->pc; 1070 pq = &vm_page_queues[m->queue]; 1071 mtx_lock_spin(&vm_page_queue_free_mtx); 1072 pq->lcnt++; 1073 ++(*pq->cnt); 1074 1075 /* 1076 * Put zero'd pages on the end ( where we look for zero'd pages 1077 * first ) and non-zerod pages at the head. 1078 */ 1079 if (m->flags & PG_ZERO) { 1080 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1081 ++vm_page_zero_count; 1082 } else { 1083 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1084 } 1085 mtx_unlock_spin(&vm_page_queue_free_mtx); 1086 vm_page_free_wakeup(); 1087} 1088 1089/* 1090 * vm_page_unmanage: 1091 * 1092 * Prevent PV management from being done on the page. The page is 1093 * removed from the paging queues as if it were wired, and as a 1094 * consequence of no longer being managed the pageout daemon will not 1095 * touch it (since there is no way to locate the pte mappings for the 1096 * page). madvise() calls that mess with the pmap will also no longer 1097 * operate on the page. 1098 * 1099 * Beyond that the page is still reasonably 'normal'. Freeing the page 1100 * will clear the flag. 1101 * 1102 * This routine is used by OBJT_PHYS objects - objects using unswappable 1103 * physical memory as backing store rather then swap-backed memory and 1104 * will eventually be extended to support 4MB unmanaged physical 1105 * mappings. 1106 */ 1107void 1108vm_page_unmanage(vm_page_t m) 1109{ 1110 1111 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1112 if ((m->flags & PG_UNMANAGED) == 0) { 1113 if (m->wire_count == 0) 1114 vm_pageq_remove(m); 1115 } 1116 vm_page_flag_set(m, PG_UNMANAGED); 1117} 1118 1119/* 1120 * vm_page_wire: 1121 * 1122 * Mark this page as wired down by yet 1123 * another map, removing it from paging queues 1124 * as necessary. 1125 * 1126 * The page queues must be locked. 1127 * This routine may not block. 1128 */ 1129void 1130vm_page_wire(vm_page_t m) 1131{ 1132 1133 /* 1134 * Only bump the wire statistics if the page is not already wired, 1135 * and only unqueue the page if it is on some queue (if it is unmanaged 1136 * it is already off the queues). 1137 */ 1138 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1139 if (m->flags & PG_FICTITIOUS) 1140 return; 1141 if (m->wire_count == 0) { 1142 if ((m->flags & PG_UNMANAGED) == 0) 1143 vm_pageq_remove(m); 1144 atomic_add_int(&cnt.v_wire_count, 1); 1145 } 1146 m->wire_count++; 1147 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1148} 1149 1150/* 1151 * vm_page_unwire: 1152 * 1153 * Release one wiring of this page, potentially 1154 * enabling it to be paged again. 1155 * 1156 * Many pages placed on the inactive queue should actually go 1157 * into the cache, but it is difficult to figure out which. What 1158 * we do instead, if the inactive target is well met, is to put 1159 * clean pages at the head of the inactive queue instead of the tail. 1160 * This will cause them to be moved to the cache more quickly and 1161 * if not actively re-referenced, freed more quickly. If we just 1162 * stick these pages at the end of the inactive queue, heavy filesystem 1163 * meta-data accesses can cause an unnecessary paging load on memory bound 1164 * processes. This optimization causes one-time-use metadata to be 1165 * reused more quickly. 1166 * 1167 * BUT, if we are in a low-memory situation we have no choice but to 1168 * put clean pages on the cache queue. 1169 * 1170 * A number of routines use vm_page_unwire() to guarantee that the page 1171 * will go into either the inactive or active queues, and will NEVER 1172 * be placed in the cache - for example, just after dirtying a page. 1173 * dirty pages in the cache are not allowed. 1174 * 1175 * The page queues must be locked. 1176 * This routine may not block. 1177 */ 1178void 1179vm_page_unwire(vm_page_t m, int activate) 1180{ 1181 1182 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1183 if (m->flags & PG_FICTITIOUS) 1184 return; 1185 if (m->wire_count > 0) { 1186 m->wire_count--; 1187 if (m->wire_count == 0) { 1188 atomic_subtract_int(&cnt.v_wire_count, 1); 1189 if (m->flags & PG_UNMANAGED) { 1190 ; 1191 } else if (activate) 1192 vm_pageq_enqueue(PQ_ACTIVE, m); 1193 else { 1194 vm_page_flag_clear(m, PG_WINATCFLS); 1195 vm_pageq_enqueue(PQ_INACTIVE, m); 1196 } 1197 } 1198 } else { 1199 panic("vm_page_unwire: invalid wire count: %d", m->wire_count); 1200 } 1201} 1202 1203 1204/* 1205 * Move the specified page to the inactive queue. If the page has 1206 * any associated swap, the swap is deallocated. 1207 * 1208 * Normally athead is 0 resulting in LRU operation. athead is set 1209 * to 1 if we want this page to be 'as if it were placed in the cache', 1210 * except without unmapping it from the process address space. 1211 * 1212 * This routine may not block. 1213 */ 1214static __inline void 1215_vm_page_deactivate(vm_page_t m, int athead) 1216{ 1217 1218 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1219 1220 /* 1221 * Ignore if already inactive. 1222 */ 1223 if (m->queue == PQ_INACTIVE) 1224 return; 1225 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1226 if ((m->queue - m->pc) == PQ_CACHE) 1227 cnt.v_reactivated++; 1228 vm_page_flag_clear(m, PG_WINATCFLS); 1229 vm_pageq_remove(m); 1230 if (athead) 1231 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1232 else 1233 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1234 m->queue = PQ_INACTIVE; 1235 vm_page_queues[PQ_INACTIVE].lcnt++; 1236 cnt.v_inactive_count++; 1237 } 1238} 1239 1240void 1241vm_page_deactivate(vm_page_t m) 1242{ 1243 _vm_page_deactivate(m, 0); 1244} 1245 1246/* 1247 * vm_page_try_to_cache: 1248 * 1249 * Returns 0 on failure, 1 on success 1250 */ 1251int 1252vm_page_try_to_cache(vm_page_t m) 1253{ 1254 1255 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1256 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1257 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1258 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1259 return (0); 1260 } 1261 pmap_remove_all(m); 1262 if (m->dirty) 1263 return (0); 1264 vm_page_cache(m); 1265 return (1); 1266} 1267 1268/* 1269 * vm_page_try_to_free() 1270 * 1271 * Attempt to free the page. If we cannot free it, we do nothing. 1272 * 1 is returned on success, 0 on failure. 1273 */ 1274int 1275vm_page_try_to_free(vm_page_t m) 1276{ 1277 1278 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1279 if (m->object != NULL) 1280 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1281 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1282 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1283 return (0); 1284 } 1285 pmap_remove_all(m); 1286 if (m->dirty) 1287 return (0); 1288 vm_page_free(m); 1289 return (1); 1290} 1291 1292/* 1293 * vm_page_cache 1294 * 1295 * Put the specified page onto the page cache queue (if appropriate). 1296 * 1297 * This routine may not block. 1298 */ 1299void 1300vm_page_cache(vm_page_t m) 1301{ 1302 1303 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1304 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1305 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 1306 m->hold_count || m->wire_count) { 1307 printf("vm_page_cache: attempting to cache busy page\n"); 1308 return; 1309 } 1310 if ((m->queue - m->pc) == PQ_CACHE) 1311 return; 1312 1313 /* 1314 * Remove all pmaps and indicate that the page is not 1315 * writeable or mapped. 1316 */ 1317 pmap_remove_all(m); 1318 if (m->dirty != 0) { 1319 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1320 (long)m->pindex); 1321 } 1322 vm_pageq_remove_nowakeup(m); 1323 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1324 vm_page_free_wakeup(); 1325} 1326 1327/* 1328 * vm_page_dontneed 1329 * 1330 * Cache, deactivate, or do nothing as appropriate. This routine 1331 * is typically used by madvise() MADV_DONTNEED. 1332 * 1333 * Generally speaking we want to move the page into the cache so 1334 * it gets reused quickly. However, this can result in a silly syndrome 1335 * due to the page recycling too quickly. Small objects will not be 1336 * fully cached. On the otherhand, if we move the page to the inactive 1337 * queue we wind up with a problem whereby very large objects 1338 * unnecessarily blow away our inactive and cache queues. 1339 * 1340 * The solution is to move the pages based on a fixed weighting. We 1341 * either leave them alone, deactivate them, or move them to the cache, 1342 * where moving them to the cache has the highest weighting. 1343 * By forcing some pages into other queues we eventually force the 1344 * system to balance the queues, potentially recovering other unrelated 1345 * space from active. The idea is to not force this to happen too 1346 * often. 1347 */ 1348void 1349vm_page_dontneed(vm_page_t m) 1350{ 1351 static int dnweight; 1352 int dnw; 1353 int head; 1354 1355 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1356 dnw = ++dnweight; 1357 1358 /* 1359 * occassionally leave the page alone 1360 */ 1361 if ((dnw & 0x01F0) == 0 || 1362 m->queue == PQ_INACTIVE || 1363 m->queue - m->pc == PQ_CACHE 1364 ) { 1365 if (m->act_count >= ACT_INIT) 1366 --m->act_count; 1367 return; 1368 } 1369 1370 if (m->dirty == 0 && pmap_is_modified(m)) 1371 vm_page_dirty(m); 1372 1373 if (m->dirty || (dnw & 0x0070) == 0) { 1374 /* 1375 * Deactivate the page 3 times out of 32. 1376 */ 1377 head = 0; 1378 } else { 1379 /* 1380 * Cache the page 28 times out of every 32. Note that 1381 * the page is deactivated instead of cached, but placed 1382 * at the head of the queue instead of the tail. 1383 */ 1384 head = 1; 1385 } 1386 _vm_page_deactivate(m, head); 1387} 1388 1389/* 1390 * Grab a page, waiting until we are waken up due to the page 1391 * changing state. We keep on waiting, if the page continues 1392 * to be in the object. If the page doesn't exist, first allocate it 1393 * and then conditionally zero it. 1394 * 1395 * This routine may block. 1396 */ 1397vm_page_t 1398vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1399{ 1400 vm_page_t m; 1401 1402 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1403retrylookup: 1404 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1405 vm_page_lock_queues(); 1406 if (m->busy || (m->flags & PG_BUSY)) { 1407 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1408 VM_OBJECT_UNLOCK(object); 1409 msleep(m, &vm_page_queue_mtx, PDROP | PVM, "pgrbwt", 0); 1410 VM_OBJECT_LOCK(object); 1411 if ((allocflags & VM_ALLOC_RETRY) == 0) 1412 return (NULL); 1413 goto retrylookup; 1414 } else { 1415 if (allocflags & VM_ALLOC_WIRED) 1416 vm_page_wire(m); 1417 if ((allocflags & VM_ALLOC_NOBUSY) == 0) 1418 vm_page_busy(m); 1419 vm_page_unlock_queues(); 1420 return (m); 1421 } 1422 } 1423 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1424 if (m == NULL) { 1425 VM_OBJECT_UNLOCK(object); 1426 VM_WAIT; 1427 VM_OBJECT_LOCK(object); 1428 if ((allocflags & VM_ALLOC_RETRY) == 0) 1429 return (NULL); 1430 goto retrylookup; 1431 } 1432 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) 1433 pmap_zero_page(m); 1434 return (m); 1435} 1436 1437/* 1438 * Mapping function for valid bits or for dirty bits in 1439 * a page. May not block. 1440 * 1441 * Inputs are required to range within a page. 1442 */ 1443__inline int 1444vm_page_bits(int base, int size) 1445{ 1446 int first_bit; 1447 int last_bit; 1448 1449 KASSERT( 1450 base + size <= PAGE_SIZE, 1451 ("vm_page_bits: illegal base/size %d/%d", base, size) 1452 ); 1453 1454 if (size == 0) /* handle degenerate case */ 1455 return (0); 1456 1457 first_bit = base >> DEV_BSHIFT; 1458 last_bit = (base + size - 1) >> DEV_BSHIFT; 1459 1460 return ((2 << last_bit) - (1 << first_bit)); 1461} 1462 1463/* 1464 * vm_page_set_validclean: 1465 * 1466 * Sets portions of a page valid and clean. The arguments are expected 1467 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1468 * of any partial chunks touched by the range. The invalid portion of 1469 * such chunks will be zero'd. 1470 * 1471 * This routine may not block. 1472 * 1473 * (base + size) must be less then or equal to PAGE_SIZE. 1474 */ 1475void 1476vm_page_set_validclean(vm_page_t m, int base, int size) 1477{ 1478 int pagebits; 1479 int frag; 1480 int endoff; 1481 1482 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1483 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1484 if (size == 0) /* handle degenerate case */ 1485 return; 1486 1487 /* 1488 * If the base is not DEV_BSIZE aligned and the valid 1489 * bit is clear, we have to zero out a portion of the 1490 * first block. 1491 */ 1492 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1493 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) 1494 pmap_zero_page_area(m, frag, base - frag); 1495 1496 /* 1497 * If the ending offset is not DEV_BSIZE aligned and the 1498 * valid bit is clear, we have to zero out a portion of 1499 * the last block. 1500 */ 1501 endoff = base + size; 1502 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1503 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) 1504 pmap_zero_page_area(m, endoff, 1505 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); 1506 1507 /* 1508 * Set valid, clear dirty bits. If validating the entire 1509 * page we can safely clear the pmap modify bit. We also 1510 * use this opportunity to clear the PG_NOSYNC flag. If a process 1511 * takes a write fault on a MAP_NOSYNC memory area the flag will 1512 * be set again. 1513 * 1514 * We set valid bits inclusive of any overlap, but we can only 1515 * clear dirty bits for DEV_BSIZE chunks that are fully within 1516 * the range. 1517 */ 1518 pagebits = vm_page_bits(base, size); 1519 m->valid |= pagebits; 1520#if 0 /* NOT YET */ 1521 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1522 frag = DEV_BSIZE - frag; 1523 base += frag; 1524 size -= frag; 1525 if (size < 0) 1526 size = 0; 1527 } 1528 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1529#endif 1530 m->dirty &= ~pagebits; 1531 if (base == 0 && size == PAGE_SIZE) { 1532 pmap_clear_modify(m); 1533 vm_page_flag_clear(m, PG_NOSYNC); 1534 } 1535} 1536 1537void 1538vm_page_clear_dirty(vm_page_t m, int base, int size) 1539{ 1540 1541 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1542 m->dirty &= ~vm_page_bits(base, size); 1543} 1544 1545/* 1546 * vm_page_set_invalid: 1547 * 1548 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1549 * valid and dirty bits for the effected areas are cleared. 1550 * 1551 * May not block. 1552 */ 1553void 1554vm_page_set_invalid(vm_page_t m, int base, int size) 1555{ 1556 int bits; 1557 1558 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1559 bits = vm_page_bits(base, size); 1560 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1561 m->valid &= ~bits; 1562 m->dirty &= ~bits; 1563 m->object->generation++; 1564} 1565 1566/* 1567 * vm_page_zero_invalid() 1568 * 1569 * The kernel assumes that the invalid portions of a page contain 1570 * garbage, but such pages can be mapped into memory by user code. 1571 * When this occurs, we must zero out the non-valid portions of the 1572 * page so user code sees what it expects. 1573 * 1574 * Pages are most often semi-valid when the end of a file is mapped 1575 * into memory and the file's size is not page aligned. 1576 */ 1577void 1578vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1579{ 1580 int b; 1581 int i; 1582 1583 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1584 /* 1585 * Scan the valid bits looking for invalid sections that 1586 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1587 * valid bit may be set ) have already been zerod by 1588 * vm_page_set_validclean(). 1589 */ 1590 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1591 if (i == (PAGE_SIZE / DEV_BSIZE) || 1592 (m->valid & (1 << i)) 1593 ) { 1594 if (i > b) { 1595 pmap_zero_page_area(m, 1596 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); 1597 } 1598 b = i + 1; 1599 } 1600 } 1601 1602 /* 1603 * setvalid is TRUE when we can safely set the zero'd areas 1604 * as being valid. We can do this if there are no cache consistancy 1605 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1606 */ 1607 if (setvalid) 1608 m->valid = VM_PAGE_BITS_ALL; 1609} 1610 1611/* 1612 * vm_page_is_valid: 1613 * 1614 * Is (partial) page valid? Note that the case where size == 0 1615 * will return FALSE in the degenerate case where the page is 1616 * entirely invalid, and TRUE otherwise. 1617 * 1618 * May not block. 1619 */ 1620int 1621vm_page_is_valid(vm_page_t m, int base, int size) 1622{ 1623 int bits = vm_page_bits(base, size); 1624 1625 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1626 if (m->valid && ((m->valid & bits) == bits)) 1627 return 1; 1628 else 1629 return 0; 1630} 1631 1632/* 1633 * update dirty bits from pmap/mmu. May not block. 1634 */ 1635void 1636vm_page_test_dirty(vm_page_t m) 1637{ 1638 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1639 vm_page_dirty(m); 1640 } 1641} 1642 1643int so_zerocp_fullpage = 0; 1644 1645void 1646vm_page_cowfault(vm_page_t m) 1647{ 1648 vm_page_t mnew; 1649 vm_object_t object; 1650 vm_pindex_t pindex; 1651 1652 object = m->object; 1653 pindex = m->pindex; 1654 1655 retry_alloc: 1656 pmap_remove_all(m); 1657 vm_page_remove(m); 1658 mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL); 1659 if (mnew == NULL) { 1660 vm_page_insert(m, object, pindex); 1661 vm_page_unlock_queues(); 1662 VM_OBJECT_UNLOCK(object); 1663 VM_WAIT; 1664 VM_OBJECT_LOCK(object); 1665 vm_page_lock_queues(); 1666 goto retry_alloc; 1667 } 1668 1669 if (m->cow == 0) { 1670 /* 1671 * check to see if we raced with an xmit complete when 1672 * waiting to allocate a page. If so, put things back 1673 * the way they were 1674 */ 1675 vm_page_free(mnew); 1676 vm_page_insert(m, object, pindex); 1677 } else { /* clear COW & copy page */ 1678 if (!so_zerocp_fullpage) 1679 pmap_copy_page(m, mnew); 1680 mnew->valid = VM_PAGE_BITS_ALL; 1681 vm_page_dirty(mnew); 1682 vm_page_flag_clear(mnew, PG_BUSY); 1683 mnew->wire_count = m->wire_count - m->cow; 1684 m->wire_count = m->cow; 1685 } 1686} 1687 1688void 1689vm_page_cowclear(vm_page_t m) 1690{ 1691 1692 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1693 if (m->cow) { 1694 m->cow--; 1695 /* 1696 * let vm_fault add back write permission lazily 1697 */ 1698 } 1699 /* 1700 * sf_buf_free() will free the page, so we needn't do it here 1701 */ 1702} 1703 1704void 1705vm_page_cowsetup(vm_page_t m) 1706{ 1707 1708 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1709 m->cow++; 1710 pmap_page_protect(m, VM_PROT_READ); 1711} 1712 1713#include "opt_ddb.h" 1714#ifdef DDB 1715#include <sys/kernel.h> 1716 1717#include <ddb/ddb.h> 1718 1719DB_SHOW_COMMAND(page, vm_page_print_page_info) 1720{ 1721 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1722 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1723 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1724 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1725 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1726 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1727 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1728 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1729 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1730 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1731} 1732 1733DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1734{ 1735 int i; 1736 db_printf("PQ_FREE:"); 1737 for (i = 0; i < PQ_L2_SIZE; i++) { 1738 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1739 } 1740 db_printf("\n"); 1741 1742 db_printf("PQ_CACHE:"); 1743 for (i = 0; i < PQ_L2_SIZE; i++) { 1744 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1745 } 1746 db_printf("\n"); 1747 1748 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1749 vm_page_queues[PQ_ACTIVE].lcnt, 1750 vm_page_queues[PQ_INACTIVE].lcnt); 1751} 1752#endif /* DDB */ 1753