2 * Copyright (c) 2006 The DragonFly Project. All rights reserved.
3 * Copyright (c) 1991 Regents of the University of California.
5 * Copyright (c) 1994 John S. Dyson
7 * Copyright (c) 1994 David Greenman
9 * Copyright (c) 2004-2006 Matthew Dillon
10 * All rights reserved.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in
20 * the documentation and/or other materials provided with the
22 * 3. Neither the name of The DragonFly Project nor the names of its
23 * contributors may be used to endorse or promote products derived
24 * from this software without specific, prior written permission.
26 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
27 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
28 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
29 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
30 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
31 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
32 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
33 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
34 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
35 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
36 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
40 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
41 * $DragonFly: src/sys/platform/vkernel/platform/pmap.c,v 1.15 2007/01/15 09:28:40 dillon Exp $
44 * NOTE: PMAP_INVAL_ADD: In pc32 this function is called prior to adjusting
45 * the PTE in the page table, because a cpu synchronization might be required.
46 * The actual invalidation is delayed until the following call or flush. In
47 * the VKERNEL build this function is called prior to adjusting the PTE and
48 * invalidates the table synchronously (not delayed), and is not SMP safe
52 #include <sys/types.h>
53 #include <sys/systm.h>
54 #include <sys/kernel.h>
57 #include <sys/vkernel.h>
59 #include <sys/thread.h>
61 #include <sys/vmspace.h>
64 #include <vm/vm_page.h>
65 #include <vm/vm_extern.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_zone.h>
69 #include <vm/vm_pageout.h>
71 #include <machine/md_var.h>
72 #include <machine/pcb.h>
73 #include <machine/pmap_inval.h>
74 #include <machine/globaldata.h>
78 struct pmap kernel_pmap;
80 static struct vm_zone pvzone;
81 static struct vm_object pvzone_obj;
82 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
83 static int pv_entry_count;
84 static int pv_entry_max;
85 static int pv_entry_high_water;
86 static int pmap_pagedaemon_waken;
87 static boolean_t pmap_initialized = FALSE;
88 static int protection_codes[8];
90 static void i386_protection_init(void);
91 static void pmap_remove_all(vm_page_t m);
92 static int pmap_release_free_page(struct pmap *pmap, vm_page_t p);
95 #ifndef PMAP_SHPGPERPROC
96 #define PMAP_SHPGPERPROC 200
99 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
101 #define pte_prot(m, p) \
102 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
108 struct pv_entry *pvinit;
110 for (i = 0; i < vm_page_array_size; i++) {
113 m = &vm_page_array[i];
114 TAILQ_INIT(&m->md.pv_list);
115 m->md.pv_list_count = 0;
118 i = vm_page_array_size;
121 pvinit = (struct pv_entry *)kmem_alloc(&kernel_map, i*sizeof(*pvinit));
122 zbootinit(&pvzone, "PV ENTRY", sizeof(*pvinit), pvinit, i);
123 pmap_initialized = TRUE;
129 int shpgperproc = PMAP_SHPGPERPROC;
131 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
132 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
133 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
134 pv_entry_high_water = 9 * (pv_entry_max / 10);
135 zinitna(&pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
139 * Bootstrap the kernel_pmap so it can be used with pmap_enter().
141 * NOTE! pm_pdir for the kernel pmap is offset so VA's translate
142 * directly into PTD indexes (PTA is also offset for the same reason).
143 * This is necessary because, for now, KVA is not mapped at address 0.
145 * Page table pages are not managed like they are in normal pmaps, so
146 * no pteobj is needed.
151 vm_pindex_t i = (vm_offset_t)KernelPTD >> PAGE_SHIFT;
153 kernel_pmap.pm_pdir = KernelPTD - (KvaStart >> SEG_SHIFT);
154 kernel_pmap.pm_pdirpte = KernelPTA[i];
155 kernel_pmap.pm_count = 1;
156 kernel_pmap.pm_active = (cpumask_t)-1;
157 TAILQ_INIT(&kernel_pmap.pm_pvlist);
158 i386_protection_init();
162 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
163 * just dummy it up so it works well enough for fork().
165 * In DragonFly, process pmaps may only be used to manipulate user address
166 * space, never kernel address space.
169 pmap_pinit0(struct pmap *pmap)
174 /************************************************************************
175 * Procedures to manage whole physical maps *
176 ************************************************************************
178 * Initialize a preallocated and zeroed pmap structure,
179 * such as one in a vmspace structure.
182 pmap_pinit(struct pmap *pmap)
188 * No need to allocate page table space yet but we do need a valid
189 * page directory table.
191 if (pmap->pm_pdir == NULL) {
193 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
197 * allocate object for the pte array and page directory
199 npages = VPTE_PAGETABLE_SIZE +
200 (VM_MAX_USER_ADDRESS / PAGE_SIZE) * sizeof(vpte_t);
201 npages = (npages + PAGE_MASK) / PAGE_SIZE;
203 if (pmap->pm_pteobj == NULL)
204 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, npages);
205 pmap->pm_pdindex = npages - 1;
208 * allocate the page directory page
210 ptdpg = vm_page_grab(pmap->pm_pteobj, pmap->pm_pdindex,
211 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
213 ptdpg->wire_count = 1;
214 ++vmstats.v_wire_count;
216 /* not usually mapped */
217 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
218 ptdpg->valid = VM_PAGE_BITS_ALL;
220 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
221 pmap->pm_pdirpte = KernelPTA[(vm_offset_t)pmap->pm_pdir >> PAGE_SHIFT];
222 if ((ptdpg->flags & PG_ZERO) == 0)
223 bzero(pmap->pm_pdir, PAGE_SIZE);
227 pmap->pm_ptphint = NULL;
228 TAILQ_INIT(&pmap->pm_pvlist);
229 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
233 * Wire in kernel global address entries. To avoid a race condition
234 * between pmap initialization and pmap_growkernel, this procedure
235 * adds the pmap to the master list (which growkernel scans to update),
236 * then copies the template.
238 * In a virtual kernel there are no kernel global address entries.
241 pmap_pinit2(struct pmap *pmap)
244 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
249 * Release all resources held by the given physical map.
251 * Should only be called if the map contains no valid mappings.
253 static int pmap_release_callback(struct vm_page *p, void *data);
256 pmap_release(struct pmap *pmap)
258 struct mdglobaldata *gd = mdcpu;
259 vm_object_t object = pmap->pm_pteobj;
260 struct rb_vm_page_scan_info info;
262 KKASSERT(pmap != &kernel_pmap);
264 #if defined(DIAGNOSTIC)
265 if (object->ref_count != 1)
266 panic("pmap_release: pteobj reference count != 1");
269 #error "Must write code to clear PTxpdir cache across all CPUs"
272 * Once we destroy the page table, the mapping becomes invalid.
273 * Rather then waste time doing a madvise
275 if (pmap->pm_pdir == gd->gd_PT1pdir) {
276 gd->gd_PT1pdir = NULL;
278 /* madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL); */
280 if (pmap->pm_pdir == gd->gd_PT2pdir) {
281 gd->gd_PT2pdir = NULL;
283 /* madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL); */
287 info.object = object;
289 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
296 info.limit = object->generation;
298 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
299 pmap_release_callback, &info);
300 if (info.error == 0 && info.mpte) {
301 if (!pmap_release_free_page(pmap, info.mpte))
305 } while (info.error);
308 * Leave the KVA reservation for pm_pdir cached for later reuse.
310 pmap->pm_pdirpte = 0;
314 pmap_release_callback(struct vm_page *p, void *data)
316 struct rb_vm_page_scan_info *info = data;
318 if (p->pindex == info->pmap->pm_pdindex) {
322 if (!pmap_release_free_page(info->pmap, p)) {
326 if (info->object->generation != info->limit) {
334 * Retire the given physical map from service. Should only be called if
335 * the map contains no valid mappings.
338 pmap_destroy(pmap_t pmap)
345 count = --pmap->pm_count;
348 panic("destroying a pmap is not yet implemented");
353 * Add a reference to the specified pmap.
356 pmap_reference(pmap_t pmap)
363 /************************************************************************
364 * VMSPACE MANAGEMENT *
365 ************************************************************************
367 * The VMSPACE management we do in our virtual kernel must be reflected
368 * in the real kernel. This is accomplished by making vmspace system
369 * calls to the real kernel.
372 cpu_vmspace_alloc(struct vmspace *vm)
377 #define LAST_EXTENT (VM_MAX_USER_ADDRESS - 0x80000000)
379 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
380 panic("vmspace_create() failed");
382 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
383 PROT_READ|PROT_WRITE,
384 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
386 if (rp == MAP_FAILED)
387 panic("vmspace_mmap: failed1");
388 vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
390 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
391 PROT_READ|PROT_WRITE,
392 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
393 MemImageFd, 0x40000000);
394 if (rp == MAP_FAILED)
395 panic("vmspace_mmap: failed2");
396 vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
398 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
399 PROT_READ|PROT_WRITE,
400 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
401 MemImageFd, 0x80000000);
402 vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
404 if (rp == MAP_FAILED)
405 panic("vmspace_mmap: failed3");
407 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
408 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
410 panic("vmspace_mcontrol: failed1");
411 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
412 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
414 panic("vmspace_mcontrol: failed2");
415 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
416 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
418 panic("vmspace_mcontrol: failed3");
422 cpu_vmspace_free(struct vmspace *vm)
424 if (vmspace_destroy(&vm->vm_pmap) < 0)
425 panic("vmspace_destroy() failed");
428 /************************************************************************
429 * Procedures which operate directly on the kernel PMAP *
430 ************************************************************************/
433 * This maps the requested page table and gives us access to it.
436 get_ptbase(struct pmap *pmap, vm_offset_t va)
438 struct mdglobaldata *gd = mdcpu;
440 if (pmap == &kernel_pmap) {
441 KKASSERT(va >= KvaStart && va < KvaEnd);
442 return(KernelPTA + (va >> PAGE_SHIFT));
443 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
444 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
445 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
446 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
450 * Otherwise choose one or the other and map the page table
451 * in the KVA space reserved for it.
453 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
454 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
456 if ((gd->gd_PTflip = 1 - gd->gd_PTflip) == 0) {
457 gd->gd_PT1pdir = pmap->pm_pdir;
458 *gd->gd_PT1pde = pmap->pm_pdirpte;
459 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
460 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
462 gd->gd_PT2pdir = pmap->pm_pdir;
463 *gd->gd_PT2pde = pmap->pm_pdirpte;
464 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
465 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
470 get_ptbase1(struct pmap *pmap, vm_offset_t va)
472 struct mdglobaldata *gd = mdcpu;
474 if (pmap == &kernel_pmap) {
475 KKASSERT(va >= KvaStart && va < KvaEnd);
476 return(KernelPTA + (va >> PAGE_SHIFT));
477 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
478 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
480 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
481 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
482 gd->gd_PT1pdir = pmap->pm_pdir;
483 *gd->gd_PT1pde = pmap->pm_pdirpte;
484 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
485 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
489 get_ptbase2(struct pmap *pmap, vm_offset_t va)
491 struct mdglobaldata *gd = mdcpu;
493 if (pmap == &kernel_pmap) {
494 KKASSERT(va >= KvaStart && va < KvaEnd);
495 return(KernelPTA + (va >> PAGE_SHIFT));
496 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
497 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
499 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
500 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
501 gd->gd_PT2pdir = pmap->pm_pdir;
502 *gd->gd_PT2pde = pmap->pm_pdirpte;
503 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
504 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
508 * When removing a page directory the related VA range in the self-mapping
509 * of the page table must be invalidated.
512 inval_ptbase_pagedir(pmap_t pmap, vm_pindex_t pindex)
514 struct mdglobaldata *gd = mdcpu;
518 #error "Must inval self-mappings in all gd's"
520 if (pmap == &kernel_pmap) {
521 va = (vm_offset_t)KernelPTA + (pindex << PAGE_SHIFT);
522 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
525 * XXX this should not strictly be needed because the page
526 * dir should alread be invalidated. test and remove
528 va = (vm_offset_t)pindex << PAGE_SHIFT;
529 vmspace_mcontrol(pmap, (void *)va, SEG_SIZE, MADV_INVAL, 0);
531 if (pmap->pm_pdir == gd->gd_PT1pdir) {
532 va = (vm_offset_t)gd->gd_PT1map + (pindex << PAGE_SHIFT);
533 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
535 if (pmap->pm_pdir == gd->gd_PT2pdir) {
536 va = (vm_offset_t)gd->gd_PT2map + (pindex << PAGE_SHIFT);
537 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
542 * Return a pointer to the page table entry for the specified va in the
543 * specified pmap. NULL is returned if there is no valid page table page
546 static __inline vpte_t *
547 pmap_pte(struct pmap *pmap, vm_offset_t va)
551 ptep = &pmap->pm_pdir[va >> SEG_SHIFT];
555 return (get_ptbase(pmap, va));
561 * Enter a mapping into kernel_pmap. Mappings created in this fashion
565 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
570 pmap_inval_info info;
573 KKASSERT(va >= KvaStart && va < KvaEnd);
574 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
575 ptep = KernelPTA + (va >> PAGE_SHIFT);
576 if (*ptep & VPTE_V) {
578 pmap_inval_init(&info);
579 pmap_inval_add(&info, &kernel_pmap, va);
583 pmap_inval_flush(&info);
585 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
593 pmap_kenter_sync(vm_offset_t va)
595 pmap_inval_info info;
597 pmap_inval_init(&info);
598 pmap_inval_add(&info, &kernel_pmap, va);
599 pmap_inval_flush(&info);
603 pmap_kenter_sync_quick(vm_offset_t va)
605 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
609 * XXX these need to be recoded. They are not used in any critical path.
612 pmap_kmodify_rw(vm_offset_t va)
614 *pmap_kpte(va) |= VPTE_R | VPTE_W;
615 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
619 pmap_kmodify_nc(vm_offset_t va)
622 *pmap_kpte(va) |= VPTE_N;
623 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
628 * Map a contiguous range of physical memory to a KVM
631 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
633 while (start < end) {
634 pmap_kenter(virt, start);
642 pmap_kpte(vm_offset_t va)
646 KKASSERT(va >= KvaStart && va < KvaEnd);
647 ptep = KernelPTA + (va >> PAGE_SHIFT);
652 * Enter a mapping into kernel_pmap without any SMP interactions.
654 * Mappings created in this fashion are not managed.
657 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
662 KKASSERT(va >= KvaStart && va < KvaEnd);
664 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
665 ptep = KernelPTA + (va >> PAGE_SHIFT);
666 if (*ptep & VPTE_V) {
668 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
675 * Make a temporary mapping for a physical address. This is only intended
676 * to be used for panic dumps.
679 pmap_kenter_temporary(vm_paddr_t pa, int i)
681 pmap_kenter(crashdumpmap + (i * PAGE_SIZE), pa);
682 return ((void *)crashdumpmap);
686 * Remove an unmanaged mapping created with pmap_kenter*().
689 pmap_kremove(vm_offset_t va)
693 pmap_inval_info info;
696 KKASSERT(va >= KvaStart && va < KvaEnd);
698 ptep = KernelPTA + (va >> PAGE_SHIFT);
699 if (*ptep & VPTE_V) {
701 pmap_inval_init(&info);
702 pmap_inval_add(&info, &kernel_pmap, va);
706 pmap_inval_flush(&info);
708 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
717 * Remove an unmanaged mapping created with pmap_kenter*() without
718 * going through any SMP interactions.
721 pmap_kremove_quick(vm_offset_t va)
725 KKASSERT(va >= KvaStart && va < KvaEnd);
727 ptep = KernelPTA + (va >> PAGE_SHIFT);
728 if (*ptep & VPTE_V) {
730 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
737 * Extract the physical address from the kernel_pmap that is associated
738 * with the specified virtual address.
741 pmap_kextract(vm_offset_t va)
746 KKASSERT(va >= KvaStart && va < KvaEnd);
748 ptep = KernelPTA + (va >> PAGE_SHIFT);
749 pa = (vm_paddr_t)(*ptep & VPTE_FRAME) | (va & PAGE_MASK);
754 * Map a set of unmanaged VM pages into KVM.
757 pmap_qenter(vm_offset_t va, struct vm_page **m, int count)
759 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
763 ptep = KernelPTA + (va >> PAGE_SHIFT);
765 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
766 *ptep = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
778 * Map a set of VM pages to kernel virtual memory. If a mapping changes
779 * clear the supplied mask. The caller handles any SMP interactions.
780 * The mask is used to provide the caller with hints on what SMP interactions
784 pmap_qenter2(vm_offset_t va, struct vm_page **m, int count, cpumask_t *mask)
786 cpumask_t cmask = mycpu->gd_cpumask;
788 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
793 ptep = KernelPTA + (va >> PAGE_SHIFT);
794 npte = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
798 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
799 } else if ((*mask & cmask) == 0) {
800 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
810 * Undo the effects of pmap_qenter*().
813 pmap_qremove(vm_offset_t va, int count)
815 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
819 ptep = KernelPTA + (va >> PAGE_SHIFT);
821 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
832 /************************************************************************
833 * Misc support glue called by machine independant code *
834 ************************************************************************
836 * These routines are called by machine independant code to operate on
837 * certain machine-dependant aspects of processes, threads, and pmaps.
841 * Initialize MD portions of the thread structure.
844 pmap_init_thread(thread_t td)
846 /* enforce pcb placement */
847 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
848 td->td_savefpu = &td->td_pcb->pcb_save;
849 td->td_sp = (char *)td->td_pcb - 16;
853 * Initialize MD portions of a process structure. XXX this aint MD
856 pmap_init_proc(struct proc *p, struct thread *td)
858 p->p_addr = (void *)td->td_kstack;
861 td->td_lwp = &p->p_lwp;
862 td->td_switch = cpu_heavy_switch;
864 KKASSERT(td->td_mpcount == 1);
866 bzero(p->p_addr, sizeof(*p->p_addr));
870 * Destroy the UPAGES for a process that has exited and disassociate
871 * the process from its thread.
874 pmap_dispose_proc(struct proc *p)
878 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
880 if ((td = p->p_thread) != NULL) {
889 * We pre-allocate all page table pages for kernel virtual memory so
890 * this routine will only be called if KVM has been exhausted.
893 pmap_growkernel(vm_offset_t addr)
895 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
897 if (addr > virtual_end - SEG_SIZE)
898 panic("KVM exhausted");
899 kernel_vm_end = addr;
903 * The modification bit is not tracked for any pages in this range. XXX
904 * such pages in this maps should always use pmap_k*() functions and not
907 * XXX User and kernel address spaces are independant for virtual kernels,
908 * this function only applies to the kernel pmap.
911 pmap_track_modified(pmap_t pmap, vm_offset_t va)
913 if (pmap != &kernel_pmap)
915 if ((va < clean_sva) || (va >= clean_eva))
921 /************************************************************************
922 * Procedures supporting managed page table pages *
923 ************************************************************************
925 * These procedures are used to track managed page table pages. These pages
926 * use the page table page's vm_page_t to track PTEs in the page. The
927 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
929 * This allows the system to throw away page table pages for user processes
930 * at will and reinstantiate them on demand.
934 * This routine works like vm_page_lookup() but also blocks as long as the
935 * page is busy. This routine does not busy the page it returns.
937 * Unless the caller is managing objects whos pages are in a known state,
938 * the call should be made with a critical section held so the page's object
939 * association remains valid on return.
942 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
947 m = vm_page_lookup(object, pindex);
948 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
954 * This routine unholds page table pages, and if the hold count
955 * drops to zero, then it decrements the wire count.
958 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
960 pmap_inval_flush(info);
961 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
964 if (m->hold_count == 0) {
966 * unmap the page table page
968 pmap->pm_pdir[m->pindex] = 0;
969 --pmap->pm_stats.resident_count;
970 inval_ptbase_pagedir(pmap, m->pindex);
972 if (pmap->pm_ptphint == m)
973 pmap->pm_ptphint = NULL;
976 * If the page is finally unwired, simply free it.
979 if (m->wire_count == 0) {
982 vm_page_free_zero(m);
983 --vmstats.v_wire_count;
991 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
994 if (m->hold_count == 0)
995 return _pmap_unwire_pte_hold(pmap, m, info);
1001 * After removing a page table entry, this routine is used to
1002 * conditionally free the page, and manage the hold/wire counts.
1005 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1006 pmap_inval_info_t info)
1012 * page table pages in the kernel_pmap are not managed.
1014 if (pmap == &kernel_pmap)
1016 ptepindex = (va >> PDRSHIFT);
1017 if (pmap->pm_ptphint &&
1018 (pmap->pm_ptphint->pindex == ptepindex)) {
1019 mpte = pmap->pm_ptphint;
1021 pmap_inval_flush(info);
1022 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1023 pmap->pm_ptphint = mpte;
1026 return pmap_unwire_pte_hold(pmap, mpte, info);
1030 * Attempt to release and free an vm_page in a pmap. Returns 1 on success,
1031 * 0 on failure (if the procedure had to sleep).
1034 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1036 vpte_t *pde = pmap->pm_pdir;
1038 * This code optimizes the case of freeing non-busy
1039 * page-table pages. Those pages are zero now, and
1040 * might as well be placed directly into the zero queue.
1042 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1046 pmap->pm_stats.resident_count--;
1048 if (p->hold_count) {
1049 panic("pmap_release: freeing held page table page");
1052 * Page directory pages need to have the kernel stuff cleared, so
1053 * they can go into the zero queue also.
1055 * In virtual kernels there is no 'kernel stuff'. For the moment
1056 * I just make sure the whole thing has been zero'd even though
1057 * it should already be completely zero'd.
1059 * pmaps for vkernels do not self-map because they do not share
1060 * their address space with the vkernel. Clearing of pde[] thus
1061 * only applies to page table pages and not to the page directory
1064 if (p->pindex == pmap->pm_pdindex) {
1065 bzero(pde, VPTE_PAGETABLE_SIZE);
1066 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1072 * Clear the matching hint
1074 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1075 pmap->pm_ptphint = NULL;
1078 * And throw the page away. The page is completely zero'd out so
1079 * optimize the free call.
1082 vmstats.v_wire_count--;
1083 vm_page_free_zero(p);
1088 * This routine is called if the page table page is not mapped in the page
1091 * The routine is broken up into two parts for readability.
1094 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1100 * Find or fabricate a new pagetable page
1102 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1103 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1105 KASSERT(m->queue == PQ_NONE,
1106 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1108 if (m->wire_count == 0)
1109 vmstats.v_wire_count++;
1113 * Increment the hold count for the page table page
1114 * (denoting a new mapping.)
1119 * Map the pagetable page into the process address space, if
1120 * it isn't already there.
1122 pmap->pm_stats.resident_count++;
1124 ptepa = VM_PAGE_TO_PHYS(m);
1125 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
1129 * We are likely about to access this page table page, so set the
1130 * page table hint to reduce overhead.
1132 pmap->pm_ptphint = m;
1135 * Try to use the new mapping, but if we cannot, then
1136 * do it with the routine that maps the page explicitly.
1138 if ((m->flags & PG_ZERO) == 0)
1139 pmap_zero_page(ptepa);
1141 m->valid = VM_PAGE_BITS_ALL;
1142 vm_page_flag_clear(m, PG_ZERO);
1143 vm_page_flag_set(m, PG_MAPPED);
1150 * Determine the page table page required to access the VA in the pmap
1151 * and allocate it if necessary. Return a held vm_page_t for the page.
1153 * Only used with user pmaps.
1156 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1163 * Calculate pagetable page index
1165 ptepindex = va >> PDRSHIFT;
1168 * Get the page directory entry
1170 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1173 * This supports switching from a 4MB page to a
1176 if (ptepa & VPTE_PS) {
1177 pmap->pm_pdir[ptepindex] = 0;
1184 * If the page table page is mapped, we just increment the
1185 * hold count, and activate it.
1189 * In order to get the page table page, try the
1192 if (pmap->pm_ptphint &&
1193 (pmap->pm_ptphint->pindex == ptepindex)) {
1194 m = pmap->pm_ptphint;
1196 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1197 pmap->pm_ptphint = m;
1203 * Here if the pte page isn't mapped, or if it has been deallocated.
1205 return _pmap_allocpte(pmap, ptepindex);
1208 /************************************************************************
1209 * Managed pages in pmaps *
1210 ************************************************************************
1212 * All pages entered into user pmaps and some pages entered into the kernel
1213 * pmap are managed, meaning that pmap_protect() and other related management
1214 * functions work on these pages.
1218 * free the pv_entry back to the free list. This function may be
1219 * called from an interrupt.
1221 static __inline void
1222 free_pv_entry(pv_entry_t pv)
1229 * get a new pv_entry, allocating a block from the system
1230 * when needed. This function may be called from an interrupt.
1236 if (pv_entry_high_water &&
1237 (pv_entry_count > pv_entry_high_water) &&
1238 (pmap_pagedaemon_waken == 0)) {
1239 pmap_pagedaemon_waken = 1;
1240 wakeup (&vm_pages_needed);
1242 return zalloc(&pvzone);
1246 * This routine is very drastic, but can save the system
1254 static int warningdone=0;
1256 if (pmap_pagedaemon_waken == 0)
1259 if (warningdone < 5) {
1260 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1264 for(i = 0; i < vm_page_array_size; i++) {
1265 m = &vm_page_array[i];
1266 if (m->wire_count || m->hold_count || m->busy ||
1267 (m->flags & PG_BUSY))
1271 pmap_pagedaemon_waken = 0;
1275 * If it is the first entry on the list, it is actually
1276 * in the header and we must copy the following entry up
1277 * to the header. Otherwise we must search the list for
1278 * the entry. In either case we free the now unused entry.
1281 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1282 vm_offset_t va, pmap_inval_info_t info)
1288 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1289 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1290 if (pmap == pv->pv_pmap && va == pv->pv_va)
1294 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1295 if (va == pv->pv_va)
1301 * Note that pv_ptem is NULL if the page table page itself is not
1302 * managed, even if the page being removed IS managed.
1306 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1307 m->md.pv_list_count--;
1308 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1309 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1310 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1311 ++pmap->pm_generation;
1312 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1320 * Create a pv entry for page at pa for (pmap, va). If the page table page
1321 * holding the VA is managed, mpte will be non-NULL.
1324 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1329 pv = get_pv_entry();
1334 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1335 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1336 m->md.pv_list_count++;
1342 * pmap_remove_pte: do the things to unmap a page in a process
1345 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va,
1346 pmap_inval_info_t info)
1351 oldpte = loadandclear(ptq);
1352 pmap_inval_add(info, pmap, va); /* See NOTE: PMAP_INVAL_ADD */
1353 if (oldpte & VPTE_WIRED)
1354 --pmap->pm_stats.wired_count;
1355 KKASSERT(pmap->pm_stats.wired_count >= 0);
1359 * Machines that don't support invlpg, also don't support
1360 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1363 if (oldpte & VPTE_G)
1364 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1366 pmap->pm_stats.resident_count -= 1;
1367 if (oldpte & VPTE_MANAGED) {
1368 m = PHYS_TO_VM_PAGE(oldpte);
1369 if (oldpte & VPTE_M) {
1370 #if defined(PMAP_DIAGNOSTIC)
1371 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1373 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1377 if (pmap_track_modified(pmap, va))
1380 if (oldpte & VPTE_A)
1381 vm_page_flag_set(m, PG_REFERENCED);
1382 return pmap_remove_entry(pmap, m, va, info);
1384 return pmap_unuse_pt(pmap, va, NULL, info);
1393 * Remove a single page from a process address space.
1395 * This function may not be called from an interrupt if the pmap is
1399 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1404 * if there is no pte for this address, just skip it!!! Otherwise
1405 * get a local va for mappings for this pmap and remove the entry.
1407 if (*pmap_pde(pmap, va) != 0) {
1408 ptq = get_ptbase(pmap, va);
1410 pmap_remove_pte(pmap, ptq, va, info);
1418 * Remove the given range of addresses from the specified map.
1420 * It is assumed that the start and end are properly
1421 * rounded to the page size.
1423 * This function may not be called from an interrupt if the pmap is
1427 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1431 vm_offset_t ptpaddr;
1432 vm_pindex_t sindex, eindex;
1433 struct pmap_inval_info info;
1438 KKASSERT(pmap->pm_stats.resident_count >= 0);
1439 if (pmap->pm_stats.resident_count == 0)
1442 pmap_inval_init(&info);
1445 * special handling of removing one page. a very
1446 * common operation and easy to short circuit some
1449 if (((sva + PAGE_SIZE) == eva) &&
1450 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1451 pmap_remove_page(pmap, sva, &info);
1452 pmap_inval_flush(&info);
1457 * Get a local virtual address for the mappings that are being
1460 * XXX this is really messy because the kernel pmap is not relative
1463 sindex = (sva >> PAGE_SHIFT);
1464 eindex = (eva >> PAGE_SHIFT);
1466 for (; sindex < eindex; sindex = pdnxt) {
1470 * Calculate index for next page table.
1472 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1473 if (pmap->pm_stats.resident_count == 0)
1476 pdirindex = sindex / NPDEPG;
1477 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1478 pmap->pm_pdir[pdirindex] = 0;
1479 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1480 inval_ptbase_pagedir(pmap, pdirindex);
1485 * Weed out invalid mappings. Note: we assume that the page
1486 * directory table is always allocated, and in kernel virtual.
1492 * Limit our scan to either the end of the va represented
1493 * by the current page table page, or to the end of the
1494 * range being removed.
1500 * NOTE: pmap_remove_pte() can block.
1502 for (; sindex != pdnxt; sindex++) {
1505 ptbase = get_ptbase(pmap, sindex << PAGE_SHIFT);
1508 va = i386_ptob(sindex);
1509 if (pmap_remove_pte(pmap, ptbase, va, &info))
1513 pmap_inval_flush(&info);
1519 * Removes this physical page from all physical maps in which it resides.
1520 * Reflects back modify bits to the pager.
1522 * This routine may not be called from an interrupt.
1525 pmap_remove_all(vm_page_t m)
1527 struct pmap_inval_info info;
1531 #if defined(PMAP_DIAGNOSTIC)
1533 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1536 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1537 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1541 pmap_inval_init(&info);
1543 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1544 pv->pv_pmap->pm_stats.resident_count--;
1546 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1547 KKASSERT(pte != NULL);
1549 tpte = loadandclear(pte);
1550 /* See NOTE: PMAP_INVAL_ADD */
1551 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1552 if (tpte & VPTE_WIRED)
1553 --pv->pv_pmap->pm_stats.wired_count;
1554 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
1557 vm_page_flag_set(m, PG_REFERENCED);
1560 * Update the vm_page_t clean and reference bits.
1562 if (tpte & VPTE_M) {
1563 #if defined(PMAP_DIAGNOSTIC)
1564 if (pmap_nw_modified((pt_entry_t) tpte)) {
1566 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1570 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
1573 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1574 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1575 ++pv->pv_pmap->pm_generation;
1576 m->md.pv_list_count--;
1577 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1581 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1583 pmap_inval_flush(&info);
1589 * Set the physical protection on the specified range of this map
1592 * This function may not be called from an interrupt if the map is
1593 * not the kernel_pmap.
1596 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1599 vm_offset_t pdnxt, ptpaddr;
1600 vm_pindex_t sindex, eindex;
1602 pmap_inval_info info;
1607 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1608 pmap_remove(pmap, sva, eva);
1612 if (prot & VM_PROT_WRITE)
1615 pmap_inval_init(&info);
1617 ptbase = get_ptbase(pmap, sva);
1619 sindex = (sva >> PAGE_SHIFT);
1620 eindex = (eva >> PAGE_SHIFT);
1623 for (; sindex < eindex; sindex = pdnxt) {
1627 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1629 pdirindex = sindex / NPDEPG;
1630 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1631 pmap->pm_pdir[pdirindex] &= ~(VPTE_M|VPTE_W);
1632 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1633 inval_ptbase_pagedir(pmap, pdirindex);
1638 * Weed out invalid mappings. Note: we assume that the page
1639 * directory table is always allocated, and in kernel virtual.
1644 if (pdnxt > eindex) {
1648 for (; sindex != pdnxt; sindex++) {
1652 pbits = ptbase[sindex - sbase];
1654 if (pbits & VPTE_MANAGED) {
1656 if (pbits & VPTE_A) {
1657 m = PHYS_TO_VM_PAGE(pbits);
1658 vm_page_flag_set(m, PG_REFERENCED);
1661 if (pbits & VPTE_M) {
1662 if (pmap_track_modified(pmap, i386_ptob(sindex))) {
1664 m = PHYS_TO_VM_PAGE(pbits);
1673 if (pbits != ptbase[sindex - sbase]) {
1674 ptbase[sindex - sbase] = pbits;
1675 /* See NOTE: PMAP_INVAL_ADD */
1676 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1680 pmap_inval_flush(&info);
1684 * Enter a managed page into a pmap. If the page is not wired related pmap
1685 * data can be destroyed at any time for later demand-operation.
1687 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1688 * specified protection, and wire the mapping if requested.
1690 * NOTE: This routine may not lazy-evaluate or lose information. The
1691 * page must actually be inserted into the given map NOW.
1693 * NOTE: When entering a page at a KVA address, the pmap must be the
1697 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1703 vm_offset_t origpte, newpte;
1705 pmap_inval_info info;
1713 * Get the page table page. The kernel_pmap's page table pages
1714 * are preallocated and have no associated vm_page_t.
1716 if (pmap == &kernel_pmap)
1719 mpte = pmap_allocpte(pmap, va);
1721 pmap_inval_init(&info);
1722 pte = pmap_pte(pmap, va);
1725 * Page Directory table entry not valid, we need a new PT page
1726 * and pmap_allocpte() didn't give us one. Oops!
1729 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1733 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1735 opa = origpte & VPTE_FRAME;
1737 printf("pmap_enter: pmap %p va %08x pa %08x PDE %08x origpte %08x\n", pmap, va, (int)pa, pmap->pm_pdir[va >> SEG_SHIFT], origpte);
1740 if (origpte & VPTE_PS)
1741 panic("pmap_enter: attempted pmap_enter on 4MB page");
1744 * Mapping has not changed, must be protection or wiring change.
1746 if (origpte && (opa == pa)) {
1748 * Wiring change, just update stats. We don't worry about
1749 * wiring PT pages as they remain resident as long as there
1750 * are valid mappings in them. Hence, if a user page is wired,
1751 * the PT page will be also.
1753 if (wired && ((origpte & VPTE_WIRED) == 0))
1754 ++pmap->pm_stats.wired_count;
1755 else if (!wired && (origpte & VPTE_WIRED))
1756 --pmap->pm_stats.wired_count;
1757 KKASSERT(pmap->pm_stats.wired_count >= 0);
1759 #if defined(PMAP_DIAGNOSTIC)
1760 if (pmap_nw_modified((pt_entry_t) origpte)) {
1762 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1768 * Remove the extra pte reference. Note that we cannot
1769 * optimize the RO->RW case because we have adjusted the
1770 * wiring count above and may need to adjust the wiring
1777 * We might be turning off write access to the page,
1778 * so we go ahead and sense modify status.
1780 if (origpte & VPTE_MANAGED) {
1781 if ((origpte & VPTE_M) && pmap_track_modified(pmap, va)) {
1783 om = PHYS_TO_VM_PAGE(opa);
1791 * Mapping has changed, invalidate old range and fall through to
1792 * handle validating new mapping.
1796 err = pmap_remove_pte(pmap, pte, va, &info);
1798 panic("pmap_enter: pte vanished, va: 0x%x", va);
1802 * Enter on the PV list if part of our managed memory. Note that we
1803 * raise IPL while manipulating pv_table since pmap_enter can be
1804 * called at interrupt time.
1806 if (pmap_initialized &&
1807 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1808 pmap_insert_entry(pmap, va, mpte, m);
1813 * Increment counters
1815 pmap->pm_stats.resident_count++;
1817 pmap->pm_stats.wired_count++;
1821 * Now validate mapping with desired protection/wiring.
1823 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1826 newpte |= VPTE_WIRED;
1830 * if the mapping or permission bits are different, we need
1831 * to update the pte.
1833 if ((origpte & ~(VPTE_M|VPTE_A)) != newpte) {
1834 *pte = newpte | VPTE_A;
1835 /* See NOTE: PMAP_INVAL_ADD */
1836 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1838 pmap_inval_flush(&info);
1842 * This is a quick version of pmap_enter(). It is used only under the
1843 * following conditions:
1845 * (1) The pmap is not the kernel_pmap
1846 * (2) The page is not to be wired into the map
1847 * (3) The page is to mapped read-only in the pmap (initially that is)
1848 * (4) The calling procedure is responsible for flushing the TLB
1849 * (5) The page is always managed
1850 * (6) There is no prior mapping at the VA
1854 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
1858 pmap_inval_info info;
1862 KKASSERT(pmap != &kernel_pmap);
1863 pmap_inval_init(&info);
1865 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1868 * Instantiate the page table page if required
1872 * Calculate pagetable page index
1874 ptepindex = va >> PDRSHIFT;
1875 if (mpte && (mpte->pindex == ptepindex)) {
1880 * Get the page directory entry
1882 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1885 * If the page table page is mapped, we just increment
1886 * the hold count, and activate it.
1889 if (ptepa & VPTE_PS)
1890 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1891 if (pmap->pm_ptphint &&
1892 (pmap->pm_ptphint->pindex == ptepindex)) {
1893 mpte = pmap->pm_ptphint;
1895 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1896 pmap->pm_ptphint = mpte;
1902 mpte = _pmap_allocpte(pmap, ptepindex);
1907 * Ok, now that the page table page has been validated, get the pte.
1908 * If the pte is already mapped undo mpte's hold_count and
1911 pte = pmap_pte(pmap, va);
1914 pmap_unwire_pte_hold(pmap, mpte, &info);
1919 * Enter on the PV list if part of our managed memory. Note that we
1920 * raise IPL while manipulating pv_table since pmap_enter can be
1921 * called at interrupt time.
1923 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
1924 pmap_insert_entry(pmap, va, mpte, m);
1927 * Increment counters
1929 pmap->pm_stats.resident_count++;
1931 pa = VM_PAGE_TO_PHYS(m);
1934 * Now validate mapping with RO protection
1936 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1937 *pte = pa | VPTE_V | VPTE_U;
1939 *pte = pa | VPTE_V | VPTE_U | VPTE_MANAGED;
1945 * Extract the physical address for the translation at the specified
1946 * virtual address in the pmap.
1949 pmap_extract(pmap_t pmap, vm_offset_t va)
1954 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
1955 if (pte & VPTE_PS) {
1956 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
1957 rtval |= va & SEG_MASK;
1959 pte = *get_ptbase(pmap, va);
1960 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
1967 #define MAX_INIT_PT (96)
1970 * This routine preloads the ptes for a given object into the specified pmap.
1971 * This eliminates the blast of soft faults on process startup and
1972 * immediately after an mmap.
1974 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
1977 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
1978 vm_object_t object, vm_pindex_t pindex,
1979 vm_size_t size, int limit)
1981 struct rb_vm_page_scan_info info;
1985 * We can't preinit if read access isn't set or there is no pmap
1988 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
1992 * We can't preinit if the pmap is not the current pmap
1994 if (curproc == NULL || pmap != vmspace_pmap(curproc->p_vmspace))
1997 psize = size >> PAGE_SHIFT;
1999 if ((object->type != OBJT_VNODE) ||
2000 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2001 (object->resident_page_count > MAX_INIT_PT))) {
2005 if (psize + pindex > object->size) {
2006 if (object->size < pindex)
2008 psize = object->size - pindex;
2015 * Use a red-black scan to traverse the requested range and load
2016 * any valid pages found into the pmap.
2018 * We cannot safely scan the object's memq unless we are in a
2019 * critical section since interrupts can remove pages from objects.
2021 info.start_pindex = pindex;
2022 info.end_pindex = pindex + psize - 1;
2029 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2030 pmap_object_init_pt_callback, &info);
2036 pmap_object_init_pt_callback(vm_page_t p, void *data)
2038 struct rb_vm_page_scan_info *info = data;
2039 vm_pindex_t rel_index;
2041 * don't allow an madvise to blow away our really
2042 * free pages allocating pv entries.
2044 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2045 vmstats.v_free_count < vmstats.v_free_reserved) {
2048 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2049 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2050 if ((p->queue - p->pc) == PQ_CACHE)
2051 vm_page_deactivate(p);
2053 rel_index = p->pindex - info->start_pindex;
2054 info->mpte = pmap_enter_quick(info->pmap,
2055 info->addr + i386_ptob(rel_index),
2057 vm_page_flag_set(p, PG_MAPPED);
2064 * pmap_prefault provides a quick way of clustering pagefaults into a
2065 * processes address space. It is a "cousin" of pmap_object_init_pt,
2066 * except it runs at page fault time instead of mmap time.
2070 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2072 static int pmap_prefault_pageorder[] = {
2073 -PAGE_SIZE, PAGE_SIZE,
2074 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2075 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2076 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2080 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2090 * We do not currently prefault mappings that use virtual page
2091 * tables. We do not prefault foreign pmaps.
2093 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2095 if (curproc == NULL || (pmap != vmspace_pmap(curproc->p_vmspace)))
2098 object = entry->object.vm_object;
2100 starta = addra - PFBAK * PAGE_SIZE;
2101 if (starta < entry->start)
2102 starta = entry->start;
2103 else if (starta > addra)
2107 * critical section protection is required to maintain the
2108 * page/object association, interrupts can free pages and remove
2109 * them from their objects.
2113 for (i = 0; i < PAGEORDER_SIZE; i++) {
2114 vm_object_t lobject;
2117 addr = addra + pmap_prefault_pageorder[i];
2118 if (addr > addra + (PFFOR * PAGE_SIZE))
2121 if (addr < starta || addr >= entry->end)
2125 * Make sure the page table page already exists
2127 if ((*pmap_pde(pmap, addr)) == NULL)
2131 * Get a pointer to the pte and make sure that no valid page
2134 pte = get_ptbase(pmap, addr);
2139 * Get the page to be mapped
2141 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2144 for (m = vm_page_lookup(lobject, pindex);
2145 (!m && (lobject->type == OBJT_DEFAULT) &&
2146 (lobject->backing_object));
2147 lobject = lobject->backing_object
2149 if (lobject->backing_object_offset & PAGE_MASK)
2151 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2152 m = vm_page_lookup(lobject->backing_object, pindex);
2156 * give-up when a page is not in memory
2162 * If everything meets the requirements for pmap_enter_quick(),
2163 * then enter the page.
2166 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2168 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2170 if ((m->queue - m->pc) == PQ_CACHE) {
2171 vm_page_deactivate(m);
2174 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2175 vm_page_flag_set(m, PG_MAPPED);
2183 * Routine: pmap_change_wiring
2184 * Function: Change the wiring attribute for a map/virtual-address
2186 * In/out conditions:
2187 * The mapping must already exist in the pmap.
2190 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2197 pte = get_ptbase(pmap, va);
2199 if (wired && (*pte & VPTE_WIRED) == 0)
2200 ++pmap->pm_stats.wired_count;
2201 else if (!wired && (*pte & VPTE_WIRED))
2202 --pmap->pm_stats.wired_count;
2203 KKASSERT(pmap->pm_stats.wired_count >= 0);
2206 * Wiring is not a hardware characteristic so there is no need to
2207 * invalidate TLB. However, in an SMP environment we must use
2208 * a locked bus cycle to update the pte (if we are not using
2209 * the pmap_inval_*() API that is)... it's ok to do this for simple
2213 atomic_set_int(pte, VPTE_WIRED);
2215 atomic_clear_int(pte, VPTE_WIRED);
2219 * Copy the range specified by src_addr/len
2220 * from the source map to the range dst_addr/len
2221 * in the destination map.
2223 * This routine is only advisory and need not do anything.
2226 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2227 vm_size_t len, vm_offset_t src_addr)
2229 pmap_inval_info info;
2231 vm_offset_t end_addr = src_addr + len;
2237 if (dst_addr != src_addr)
2239 if (dst_pmap->pm_pdir == NULL)
2241 if (src_pmap->pm_pdir == NULL)
2244 src_frame = get_ptbase1(src_pmap, src_addr);
2245 dst_frame = get_ptbase2(dst_pmap, src_addr);
2247 pmap_inval_init(&info);
2250 pmap_inval_add(&info, dst_pmap, -1);
2251 pmap_inval_add(&info, src_pmap, -1);
2255 * critical section protection is required to maintain the page/object
2256 * association, interrupts can free pages and remove them from
2260 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2261 vpte_t *src_pte, *dst_pte;
2262 vm_page_t dstmpte, srcmpte;
2263 vm_offset_t srcptepaddr;
2266 if (addr >= VM_MAX_USER_ADDRESS)
2267 panic("pmap_copy: invalid to pmap_copy page tables\n");
2270 * Don't let optional prefaulting of pages make us go
2271 * way below the low water mark of free pages or way
2272 * above high water mark of used pv entries.
2274 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2275 pv_entry_count > pv_entry_high_water)
2278 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2279 ptepindex = addr >> PDRSHIFT;
2281 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2282 if (srcptepaddr == 0)
2285 if (srcptepaddr & VPTE_PS) {
2286 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2287 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2288 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2293 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2294 if ((srcmpte == NULL) ||
2295 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2298 if (pdnxt > end_addr)
2301 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2302 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2303 while (addr < pdnxt) {
2307 * we only virtual copy managed pages
2309 if ((ptetemp & VPTE_MANAGED) != 0) {
2311 * We have to check after allocpte for the
2312 * pte still being around... allocpte can
2315 dstmpte = pmap_allocpte(dst_pmap, addr);
2316 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2318 * Clear the modified and accessed
2319 * (referenced) bits during the copy.
2321 * We do not have to clear the write
2322 * bit to force a fault-on-modify
2323 * because the real kernel's target
2324 * pmap is empty and will fault anyway.
2326 m = PHYS_TO_VM_PAGE(ptetemp);
2327 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2328 dst_pmap->pm_stats.resident_count++;
2329 pmap_insert_entry(dst_pmap, addr,
2332 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2334 if (dstmpte->hold_count >= srcmpte->hold_count)
2343 pmap_inval_flush(&info);
2349 * Zero the specified PA by mapping the page into KVM and clearing its
2352 * This function may be called from an interrupt and no locking is
2356 pmap_zero_page(vm_paddr_t phys)
2358 struct mdglobaldata *gd = mdcpu;
2362 panic("pmap_zero_page: CMAP3 busy");
2363 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2364 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2366 bzero(gd->gd_CADDR3, PAGE_SIZE);
2372 * pmap_page_assertzero:
2374 * Assert that a page is empty, panic if it isn't.
2377 pmap_page_assertzero(vm_paddr_t phys)
2379 struct mdglobaldata *gd = mdcpu;
2384 panic("pmap_zero_page: CMAP3 busy");
2385 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2386 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2387 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2388 for (i = 0; i < PAGE_SIZE; i += 4) {
2389 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2390 panic("pmap_page_assertzero() @ %p not zero!\n",
2391 (void *)gd->gd_CADDR3);
2401 * Zero part of a physical page by mapping it into memory and clearing
2402 * its contents with bzero.
2404 * off and size may not cover an area beyond a single hardware page.
2407 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2409 struct mdglobaldata *gd = mdcpu;
2413 panic("pmap_zero_page: CMAP3 busy");
2414 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2415 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2416 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2418 bzero((char *)gd->gd_CADDR3 + off, size);
2426 * Copy the physical page from the source PA to the target PA.
2427 * This function may be called from an interrupt. No locking
2431 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2433 struct mdglobaldata *gd = mdcpu;
2436 if (*(int *) gd->gd_CMAP1)
2437 panic("pmap_copy_page: CMAP1 busy");
2438 if (*(int *) gd->gd_CMAP2)
2439 panic("pmap_copy_page: CMAP2 busy");
2441 *(int *) gd->gd_CMAP1 = VPTE_V | VPTE_R | (src & PG_FRAME) | VPTE_A;
2442 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2444 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2445 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2447 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2449 *(int *) gd->gd_CMAP1 = 0;
2450 *(int *) gd->gd_CMAP2 = 0;
2455 * pmap_copy_page_frag:
2457 * Copy the physical page from the source PA to the target PA.
2458 * This function may be called from an interrupt. No locking
2462 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2464 struct mdglobaldata *gd = mdcpu;
2467 if (*(int *) gd->gd_CMAP1)
2468 panic("pmap_copy_page: CMAP1 busy");
2469 if (*(int *) gd->gd_CMAP2)
2470 panic("pmap_copy_page: CMAP2 busy");
2472 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2473 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2475 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2476 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2478 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2479 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2482 *(int *) gd->gd_CMAP1 = 0;
2483 *(int *) gd->gd_CMAP2 = 0;
2488 * Returns true if the pmap's pv is one of the first
2489 * 16 pvs linked to from this page. This count may
2490 * be changed upwards or downwards in the future; it
2491 * is only necessary that true be returned for a small
2492 * subset of pmaps for proper page aging.
2495 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2500 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2505 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2506 if (pv->pv_pmap == pmap) {
2519 * Remove all pages from specified address space
2520 * this aids process exit speeds. Also, this code
2521 * is special cased for current process only, but
2522 * can have the more generic (and slightly slower)
2523 * mode enabled. This is much faster than pmap_remove
2524 * in the case of running down an entire address space.
2527 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2532 pmap_inval_info info;
2534 int32_t save_generation;
2536 if (curproc && pmap == vmspace_pmap(curproc->p_vmspace))
2541 pmap_inval_init(&info);
2543 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2544 if (pv->pv_va >= eva || pv->pv_va < sva) {
2545 npv = TAILQ_NEXT(pv, pv_plist);
2549 KKASSERT(pmap == pv->pv_pmap);
2551 pte = pmap_pte(pmap, pv->pv_va);
2555 * We cannot remove wired pages from a process' mapping
2558 if (tpte & VPTE_WIRED) {
2559 npv = TAILQ_NEXT(pv, pv_plist);
2563 /* See NOTE: PMAP_INVAL_ADD */
2564 pmap_inval_add(&info, pmap, pv->pv_va);
2566 m = PHYS_TO_VM_PAGE(tpte);
2568 KASSERT(m < &vm_page_array[vm_page_array_size],
2569 ("pmap_remove_pages: bad tpte %x", tpte));
2571 pmap->pm_stats.resident_count--;
2574 * Update the vm_page_t clean and reference bits.
2576 if (tpte & VPTE_M) {
2580 npv = TAILQ_NEXT(pv, pv_plist);
2581 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2582 save_generation = ++pmap->pm_generation;
2584 m->md.pv_list_count--;
2585 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2586 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2587 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2590 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2594 * Restart the scan if we blocked during the unuse or free
2595 * calls and other removals were made.
2597 if (save_generation != pmap->pm_generation) {
2598 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2599 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2602 pmap_inval_flush(&info);
2607 * pmap_testbit tests bits in pte's
2608 * note that the testbit/changebit routines are inline,
2609 * and a lot of things compile-time evaluate.
2612 pmap_testbit(vm_page_t m, int bit)
2617 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2620 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2625 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2627 * if the bit being tested is the modified bit, then
2628 * mark clean_map and ptes as never
2631 if (bit & (VPTE_A|VPTE_M)) {
2632 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2636 #if defined(PMAP_DIAGNOSTIC)
2638 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2642 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2653 * This routine is used to clear bits in ptes. Certain bits require special
2654 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2656 static __inline void
2657 pmap_clearbit(vm_page_t m, int bit)
2659 struct pmap_inval_info info;
2664 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2667 pmap_inval_init(&info);
2671 * Loop over all current mappings setting/clearing as appropos If
2672 * setting RO do we need to clear the VAC?
2674 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2676 * don't write protect pager mappings
2678 if (bit == VPTE_W) {
2679 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2683 #if defined(PMAP_DIAGNOSTIC)
2685 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2691 * Careful here. We can use a locked bus instruction to
2692 * clear VPTE_A or VPTE_M safely but we need to synchronize
2693 * with the target cpus when we mess with VPTE_W.
2695 * On virtual kernels we must force a new fault-on-write
2696 * in the real kernel if we clear the Modify bit ourselves,
2697 * otherwise the real kernel will not get a new fault and
2698 * will never set our Modify bit again.
2700 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2701 if (bit & (VPTE_W|VPTE_M))
2702 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2706 if (bit == VPTE_W) {
2707 if (pbits & VPTE_M) {
2710 atomic_clear_int(pte, VPTE_M|VPTE_W);
2711 } else if (bit == VPTE_M) {
2713 * We do not have to make the page read-only
2714 * when clearing the Modify bit. The real
2715 * kernel will make the real PTE read-only
2716 * or otherwise detect the write and set
2717 * our VPTE_M again simply by us invalidating
2718 * the real kernel VA for the pmap (as we did
2719 * above). This allows the real kernel to
2720 * handle the write fault without forwarding
2723 atomic_clear_int(pte, VPTE_M);
2725 atomic_clear_int(pte, bit);
2729 pmap_inval_flush(&info);
2734 * pmap_page_protect:
2736 * Lower the permission for all mappings to a given page.
2739 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2741 if ((prot & VM_PROT_WRITE) == 0) {
2742 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2743 pmap_clearbit(m, VPTE_W);
2751 pmap_phys_address(int ppn)
2753 return (i386_ptob(ppn));
2757 * pmap_ts_referenced:
2759 * Return a count of reference bits for a page, clearing those bits.
2760 * It is not necessary for every reference bit to be cleared, but it
2761 * is necessary that 0 only be returned when there are truly no
2762 * reference bits set.
2764 * XXX: The exact number of bits to check and clear is a matter that
2765 * should be tested and standardized at some point in the future for
2766 * optimal aging of shared pages.
2769 pmap_ts_referenced(vm_page_t m)
2771 pv_entry_t pv, pvf, pvn;
2775 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2780 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2785 pvn = TAILQ_NEXT(pv, pv_list);
2787 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2789 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2791 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2794 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2796 if (pte && (*pte & VPTE_A)) {
2798 atomic_clear_int(pte, VPTE_A);
2800 atomic_clear_int_nonlocked(pte, VPTE_A);
2807 } while ((pv = pvn) != NULL && pv != pvf);
2817 * Return whether or not the specified physical page was modified
2818 * in any physical maps.
2821 pmap_is_modified(vm_page_t m)
2823 return pmap_testbit(m, VPTE_M);
2827 * Clear the modify bits on the specified physical page.
2830 pmap_clear_modify(vm_page_t m)
2832 pmap_clearbit(m, VPTE_M);
2836 * pmap_clear_reference:
2838 * Clear the reference bit on the specified physical page.
2841 pmap_clear_reference(vm_page_t m)
2843 pmap_clearbit(m, VPTE_A);
2847 * Miscellaneous support routines follow
2851 i386_protection_init(void)
2855 kp = protection_codes;
2856 for (prot = 0; prot < 8; prot++) {
2857 if (prot & VM_PROT_READ)
2859 if (prot & VM_PROT_WRITE)
2861 if (prot & VM_PROT_EXECUTE)
2868 * Map a set of physical memory pages into the kernel virtual
2869 * address space. Return a pointer to where it is mapped. This
2870 * routine is intended to be used for mapping device memory,
2873 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2877 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2879 vm_offset_t va, tmpva, offset;
2882 offset = pa & PAGE_MASK;
2883 size = roundup(offset + size, PAGE_SIZE);
2885 va = kmem_alloc_nofault(&kernel_map, size);
2887 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2889 pa = pa & VPTE_FRAME;
2890 for (tmpva = va; size > 0;) {
2891 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2892 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2900 return ((void *)(va + offset));
2904 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2906 vm_offset_t base, offset;
2908 base = va & VPTE_FRAME;
2909 offset = va & PAGE_MASK;
2910 size = roundup(offset + size, PAGE_SIZE);
2911 pmap_qremove(va, size >> PAGE_SHIFT);
2912 kmem_free(&kernel_map, base, size);
2916 * perform the pmap work for mincore
2919 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2925 ptep = pmap_pte(pmap, addr);
2930 if ((pte = *ptep) != 0) {
2933 val = MINCORE_INCORE;
2934 if ((pte & VPTE_MANAGED) == 0)
2937 pa = pte & VPTE_FRAME;
2939 m = PHYS_TO_VM_PAGE(pa);
2945 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2947 * Modified by someone
2949 else if (m->dirty || pmap_is_modified(m))
2950 val |= MINCORE_MODIFIED_OTHER;
2955 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2958 * Referenced by someone
2960 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
2961 val |= MINCORE_REFERENCED_OTHER;
2962 vm_page_flag_set(m, PG_REFERENCED);
2969 pmap_activate(struct proc *p)
2973 pmap = vmspace_pmap(p->p_vmspace);
2975 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2977 pmap->pm_active |= 1;
2979 #if defined(SWTCH_OPTIM_STATS)
2983 p->p_thread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
2984 load_cr3(p->p_thread->td_pcb->pcb_cr3);
2989 pmap_deactivate(struct proc *p)
2993 pmap = vmspace_pmap(p->p_vmspace);
2995 atomic_clear_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2997 pmap->pm_active &= ~1;
3000 * XXX - note we do not adjust %cr3. The caller is expected to
3001 * activate a new pmap or do a thread-exit.
3006 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3009 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3013 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3020 static void pads (pmap_t pm);
3021 void pmap_pvdump (vm_paddr_t pa);
3023 /* print address space of pmap*/
3031 if (pm == &kernel_pmap)
3033 for (i = 0; i < 1024; i++)
3035 for (j = 0; j < 1024; j++) {
3036 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3037 if (pm == &kernel_pmap && va < KERNBASE)
3039 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3041 ptep = pmap_pte(pm, va);
3042 if (ptep && (*ptep & VPTE_V)) {
3044 (void *)va, (unsigned)*ptep);
3051 pmap_pvdump(vm_paddr_t pa)
3056 kprintf("pa %08llx", (long long)pa);
3057 m = PHYS_TO_VM_PAGE(pa);
3058 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3060 kprintf(" -> pmap %p, va %x, flags %x",
3061 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3063 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);