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 1.19 2007/02/25 23:17:13 corecode 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 struct lwp *lp = ONLY_LWP_IN_PROC(p);
863 td->td_switch = cpu_heavy_switch;
865 KKASSERT(td->td_mpcount == 1);
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)
876 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
880 * We pre-allocate all page table pages for kernel virtual memory so
881 * this routine will only be called if KVM has been exhausted.
884 pmap_growkernel(vm_offset_t addr)
886 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
888 if (addr > virtual_end - SEG_SIZE)
889 panic("KVM exhausted");
890 kernel_vm_end = addr;
894 * The modification bit is not tracked for any pages in this range. XXX
895 * such pages in this maps should always use pmap_k*() functions and not
898 * XXX User and kernel address spaces are independant for virtual kernels,
899 * this function only applies to the kernel pmap.
902 pmap_track_modified(pmap_t pmap, vm_offset_t va)
904 if (pmap != &kernel_pmap)
906 if ((va < clean_sva) || (va >= clean_eva))
912 /************************************************************************
913 * Procedures supporting managed page table pages *
914 ************************************************************************
916 * These procedures are used to track managed page table pages. These pages
917 * use the page table page's vm_page_t to track PTEs in the page. The
918 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
920 * This allows the system to throw away page table pages for user processes
921 * at will and reinstantiate them on demand.
925 * This routine works like vm_page_lookup() but also blocks as long as the
926 * page is busy. This routine does not busy the page it returns.
928 * Unless the caller is managing objects whos pages are in a known state,
929 * the call should be made with a critical section held so the page's object
930 * association remains valid on return.
933 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
938 m = vm_page_lookup(object, pindex);
939 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
945 * This routine unholds page table pages, and if the hold count
946 * drops to zero, then it decrements the wire count.
949 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
951 pmap_inval_flush(info);
952 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
955 if (m->hold_count == 0) {
957 * unmap the page table page
959 pmap->pm_pdir[m->pindex] = 0;
960 --pmap->pm_stats.resident_count;
961 inval_ptbase_pagedir(pmap, m->pindex);
963 if (pmap->pm_ptphint == m)
964 pmap->pm_ptphint = NULL;
967 * If the page is finally unwired, simply free it.
970 if (m->wire_count == 0) {
973 vm_page_free_zero(m);
974 --vmstats.v_wire_count;
982 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
985 if (m->hold_count == 0)
986 return _pmap_unwire_pte_hold(pmap, m, info);
992 * After removing a page table entry, this routine is used to
993 * conditionally free the page, and manage the hold/wire counts.
996 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
997 pmap_inval_info_t info)
1003 * page table pages in the kernel_pmap are not managed.
1005 if (pmap == &kernel_pmap)
1007 ptepindex = (va >> PDRSHIFT);
1008 if (pmap->pm_ptphint &&
1009 (pmap->pm_ptphint->pindex == ptepindex)) {
1010 mpte = pmap->pm_ptphint;
1012 pmap_inval_flush(info);
1013 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1014 pmap->pm_ptphint = mpte;
1017 return pmap_unwire_pte_hold(pmap, mpte, info);
1021 * Attempt to release and free an vm_page in a pmap. Returns 1 on success,
1022 * 0 on failure (if the procedure had to sleep).
1025 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1027 vpte_t *pde = pmap->pm_pdir;
1029 * This code optimizes the case of freeing non-busy
1030 * page-table pages. Those pages are zero now, and
1031 * might as well be placed directly into the zero queue.
1033 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1037 pmap->pm_stats.resident_count--;
1039 if (p->hold_count) {
1040 panic("pmap_release: freeing held page table page");
1043 * Page directory pages need to have the kernel stuff cleared, so
1044 * they can go into the zero queue also.
1046 * In virtual kernels there is no 'kernel stuff'. For the moment
1047 * I just make sure the whole thing has been zero'd even though
1048 * it should already be completely zero'd.
1050 * pmaps for vkernels do not self-map because they do not share
1051 * their address space with the vkernel. Clearing of pde[] thus
1052 * only applies to page table pages and not to the page directory
1055 if (p->pindex == pmap->pm_pdindex) {
1056 bzero(pde, VPTE_PAGETABLE_SIZE);
1057 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1063 * Clear the matching hint
1065 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1066 pmap->pm_ptphint = NULL;
1069 * And throw the page away. The page is completely zero'd out so
1070 * optimize the free call.
1073 vmstats.v_wire_count--;
1074 vm_page_free_zero(p);
1079 * This routine is called if the page table page is not mapped in the page
1082 * The routine is broken up into two parts for readability.
1085 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1091 * Find or fabricate a new pagetable page
1093 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1094 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1096 KASSERT(m->queue == PQ_NONE,
1097 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1099 if (m->wire_count == 0)
1100 vmstats.v_wire_count++;
1104 * Increment the hold count for the page table page
1105 * (denoting a new mapping.)
1110 * Map the pagetable page into the process address space, if
1111 * it isn't already there.
1113 pmap->pm_stats.resident_count++;
1115 ptepa = VM_PAGE_TO_PHYS(m);
1116 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
1120 * We are likely about to access this page table page, so set the
1121 * page table hint to reduce overhead.
1123 pmap->pm_ptphint = m;
1126 * Try to use the new mapping, but if we cannot, then
1127 * do it with the routine that maps the page explicitly.
1129 if ((m->flags & PG_ZERO) == 0)
1130 pmap_zero_page(ptepa);
1132 m->valid = VM_PAGE_BITS_ALL;
1133 vm_page_flag_clear(m, PG_ZERO);
1134 vm_page_flag_set(m, PG_MAPPED);
1141 * Determine the page table page required to access the VA in the pmap
1142 * and allocate it if necessary. Return a held vm_page_t for the page.
1144 * Only used with user pmaps.
1147 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1154 * Calculate pagetable page index
1156 ptepindex = va >> PDRSHIFT;
1159 * Get the page directory entry
1161 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1164 * This supports switching from a 4MB page to a
1167 if (ptepa & VPTE_PS) {
1168 pmap->pm_pdir[ptepindex] = 0;
1175 * If the page table page is mapped, we just increment the
1176 * hold count, and activate it.
1180 * In order to get the page table page, try the
1183 if (pmap->pm_ptphint &&
1184 (pmap->pm_ptphint->pindex == ptepindex)) {
1185 m = pmap->pm_ptphint;
1187 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1188 pmap->pm_ptphint = m;
1194 * Here if the pte page isn't mapped, or if it has been deallocated.
1196 return _pmap_allocpte(pmap, ptepindex);
1199 /************************************************************************
1200 * Managed pages in pmaps *
1201 ************************************************************************
1203 * All pages entered into user pmaps and some pages entered into the kernel
1204 * pmap are managed, meaning that pmap_protect() and other related management
1205 * functions work on these pages.
1209 * free the pv_entry back to the free list. This function may be
1210 * called from an interrupt.
1212 static __inline void
1213 free_pv_entry(pv_entry_t pv)
1220 * get a new pv_entry, allocating a block from the system
1221 * when needed. This function may be called from an interrupt.
1227 if (pv_entry_high_water &&
1228 (pv_entry_count > pv_entry_high_water) &&
1229 (pmap_pagedaemon_waken == 0)) {
1230 pmap_pagedaemon_waken = 1;
1231 wakeup (&vm_pages_needed);
1233 return zalloc(&pvzone);
1237 * This routine is very drastic, but can save the system
1245 static int warningdone=0;
1247 if (pmap_pagedaemon_waken == 0)
1250 if (warningdone < 5) {
1251 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1255 for(i = 0; i < vm_page_array_size; i++) {
1256 m = &vm_page_array[i];
1257 if (m->wire_count || m->hold_count || m->busy ||
1258 (m->flags & PG_BUSY))
1262 pmap_pagedaemon_waken = 0;
1266 * If it is the first entry on the list, it is actually
1267 * in the header and we must copy the following entry up
1268 * to the header. Otherwise we must search the list for
1269 * the entry. In either case we free the now unused entry.
1272 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1273 vm_offset_t va, pmap_inval_info_t info)
1279 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1280 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1281 if (pmap == pv->pv_pmap && va == pv->pv_va)
1285 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1286 if (va == pv->pv_va)
1292 * Note that pv_ptem is NULL if the page table page itself is not
1293 * managed, even if the page being removed IS managed.
1297 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1298 m->md.pv_list_count--;
1299 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1300 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1301 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1302 ++pmap->pm_generation;
1303 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1311 * Create a pv entry for page at pa for (pmap, va). If the page table page
1312 * holding the VA is managed, mpte will be non-NULL.
1315 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1320 pv = get_pv_entry();
1325 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1326 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1327 m->md.pv_list_count++;
1333 * pmap_remove_pte: do the things to unmap a page in a process
1336 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va,
1337 pmap_inval_info_t info)
1342 oldpte = loadandclear(ptq);
1343 pmap_inval_add(info, pmap, va); /* See NOTE: PMAP_INVAL_ADD */
1344 if (oldpte & VPTE_WIRED)
1345 --pmap->pm_stats.wired_count;
1346 KKASSERT(pmap->pm_stats.wired_count >= 0);
1350 * Machines that don't support invlpg, also don't support
1351 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1354 if (oldpte & VPTE_G)
1355 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1357 pmap->pm_stats.resident_count -= 1;
1358 if (oldpte & VPTE_MANAGED) {
1359 m = PHYS_TO_VM_PAGE(oldpte);
1360 if (oldpte & VPTE_M) {
1361 #if defined(PMAP_DIAGNOSTIC)
1362 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1364 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1368 if (pmap_track_modified(pmap, va))
1371 if (oldpte & VPTE_A)
1372 vm_page_flag_set(m, PG_REFERENCED);
1373 return pmap_remove_entry(pmap, m, va, info);
1375 return pmap_unuse_pt(pmap, va, NULL, info);
1384 * Remove a single page from a process address space.
1386 * This function may not be called from an interrupt if the pmap is
1390 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1395 * if there is no pte for this address, just skip it!!! Otherwise
1396 * get a local va for mappings for this pmap and remove the entry.
1398 if (*pmap_pde(pmap, va) != 0) {
1399 ptq = get_ptbase(pmap, va);
1401 pmap_remove_pte(pmap, ptq, va, info);
1409 * Remove the given range of addresses from the specified map.
1411 * It is assumed that the start and end are properly
1412 * rounded to the page size.
1414 * This function may not be called from an interrupt if the pmap is
1418 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1422 vm_offset_t ptpaddr;
1423 vm_pindex_t sindex, eindex;
1424 struct pmap_inval_info info;
1429 KKASSERT(pmap->pm_stats.resident_count >= 0);
1430 if (pmap->pm_stats.resident_count == 0)
1433 pmap_inval_init(&info);
1436 * special handling of removing one page. a very
1437 * common operation and easy to short circuit some
1440 if (((sva + PAGE_SIZE) == eva) &&
1441 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1442 pmap_remove_page(pmap, sva, &info);
1443 pmap_inval_flush(&info);
1448 * Get a local virtual address for the mappings that are being
1451 * XXX this is really messy because the kernel pmap is not relative
1454 sindex = (sva >> PAGE_SHIFT);
1455 eindex = (eva >> PAGE_SHIFT);
1457 for (; sindex < eindex; sindex = pdnxt) {
1461 * Calculate index for next page table.
1463 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1464 if (pmap->pm_stats.resident_count == 0)
1467 pdirindex = sindex / NPDEPG;
1468 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1469 pmap->pm_pdir[pdirindex] = 0;
1470 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1471 inval_ptbase_pagedir(pmap, pdirindex);
1476 * Weed out invalid mappings. Note: we assume that the page
1477 * directory table is always allocated, and in kernel virtual.
1483 * Limit our scan to either the end of the va represented
1484 * by the current page table page, or to the end of the
1485 * range being removed.
1491 * NOTE: pmap_remove_pte() can block.
1493 for (; sindex != pdnxt; sindex++) {
1496 ptbase = get_ptbase(pmap, sindex << PAGE_SHIFT);
1499 va = i386_ptob(sindex);
1500 if (pmap_remove_pte(pmap, ptbase, va, &info))
1504 pmap_inval_flush(&info);
1510 * Removes this physical page from all physical maps in which it resides.
1511 * Reflects back modify bits to the pager.
1513 * This routine may not be called from an interrupt.
1516 pmap_remove_all(vm_page_t m)
1518 struct pmap_inval_info info;
1522 #if defined(PMAP_DIAGNOSTIC)
1524 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1527 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1528 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1532 pmap_inval_init(&info);
1534 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1535 pv->pv_pmap->pm_stats.resident_count--;
1537 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1538 KKASSERT(pte != NULL);
1540 tpte = loadandclear(pte);
1541 /* See NOTE: PMAP_INVAL_ADD */
1542 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1543 if (tpte & VPTE_WIRED)
1544 --pv->pv_pmap->pm_stats.wired_count;
1545 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
1548 vm_page_flag_set(m, PG_REFERENCED);
1551 * Update the vm_page_t clean and reference bits.
1553 if (tpte & VPTE_M) {
1554 #if defined(PMAP_DIAGNOSTIC)
1555 if (pmap_nw_modified((pt_entry_t) tpte)) {
1557 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1561 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
1564 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1565 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1566 ++pv->pv_pmap->pm_generation;
1567 m->md.pv_list_count--;
1568 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1572 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1574 pmap_inval_flush(&info);
1580 * Set the physical protection on the specified range of this map
1583 * This function may not be called from an interrupt if the map is
1584 * not the kernel_pmap.
1587 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1590 vm_offset_t pdnxt, ptpaddr;
1591 vm_pindex_t sindex, eindex;
1593 pmap_inval_info info;
1598 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1599 pmap_remove(pmap, sva, eva);
1603 if (prot & VM_PROT_WRITE)
1606 pmap_inval_init(&info);
1608 ptbase = get_ptbase(pmap, sva);
1610 sindex = (sva >> PAGE_SHIFT);
1611 eindex = (eva >> PAGE_SHIFT);
1614 for (; sindex < eindex; sindex = pdnxt) {
1618 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1620 pdirindex = sindex / NPDEPG;
1621 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1622 pmap->pm_pdir[pdirindex] &= ~(VPTE_M|VPTE_W);
1623 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1624 inval_ptbase_pagedir(pmap, pdirindex);
1629 * Weed out invalid mappings. Note: we assume that the page
1630 * directory table is always allocated, and in kernel virtual.
1635 if (pdnxt > eindex) {
1639 for (; sindex != pdnxt; sindex++) {
1643 pbits = ptbase[sindex - sbase];
1645 if (pbits & VPTE_MANAGED) {
1647 if (pbits & VPTE_A) {
1648 m = PHYS_TO_VM_PAGE(pbits);
1649 vm_page_flag_set(m, PG_REFERENCED);
1652 if (pbits & VPTE_M) {
1653 if (pmap_track_modified(pmap, i386_ptob(sindex))) {
1655 m = PHYS_TO_VM_PAGE(pbits);
1664 if (pbits != ptbase[sindex - sbase]) {
1665 ptbase[sindex - sbase] = pbits;
1666 /* See NOTE: PMAP_INVAL_ADD */
1667 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1671 pmap_inval_flush(&info);
1675 * Enter a managed page into a pmap. If the page is not wired related pmap
1676 * data can be destroyed at any time for later demand-operation.
1678 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1679 * specified protection, and wire the mapping if requested.
1681 * NOTE: This routine may not lazy-evaluate or lose information. The
1682 * page must actually be inserted into the given map NOW.
1684 * NOTE: When entering a page at a KVA address, the pmap must be the
1688 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1694 vm_offset_t origpte, newpte;
1696 pmap_inval_info info;
1704 * Get the page table page. The kernel_pmap's page table pages
1705 * are preallocated and have no associated vm_page_t.
1707 if (pmap == &kernel_pmap)
1710 mpte = pmap_allocpte(pmap, va);
1712 pmap_inval_init(&info);
1713 pte = pmap_pte(pmap, va);
1716 * Page Directory table entry not valid, we need a new PT page
1717 * and pmap_allocpte() didn't give us one. Oops!
1720 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1724 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1726 opa = origpte & VPTE_FRAME;
1728 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);
1731 if (origpte & VPTE_PS)
1732 panic("pmap_enter: attempted pmap_enter on 4MB page");
1735 * Mapping has not changed, must be protection or wiring change.
1737 if (origpte && (opa == pa)) {
1739 * Wiring change, just update stats. We don't worry about
1740 * wiring PT pages as they remain resident as long as there
1741 * are valid mappings in them. Hence, if a user page is wired,
1742 * the PT page will be also.
1744 if (wired && ((origpte & VPTE_WIRED) == 0))
1745 ++pmap->pm_stats.wired_count;
1746 else if (!wired && (origpte & VPTE_WIRED))
1747 --pmap->pm_stats.wired_count;
1748 KKASSERT(pmap->pm_stats.wired_count >= 0);
1750 #if defined(PMAP_DIAGNOSTIC)
1751 if (pmap_nw_modified((pt_entry_t) origpte)) {
1753 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1759 * Remove the extra pte reference. Note that we cannot
1760 * optimize the RO->RW case because we have adjusted the
1761 * wiring count above and may need to adjust the wiring
1768 * We might be turning off write access to the page,
1769 * so we go ahead and sense modify status.
1771 if (origpte & VPTE_MANAGED) {
1772 if ((origpte & VPTE_M) && pmap_track_modified(pmap, va)) {
1774 om = PHYS_TO_VM_PAGE(opa);
1782 * Mapping has changed, invalidate old range and fall through to
1783 * handle validating new mapping.
1787 err = pmap_remove_pte(pmap, pte, va, &info);
1789 panic("pmap_enter: pte vanished, va: 0x%x", va);
1793 * Enter on the PV list if part of our managed memory. Note that we
1794 * raise IPL while manipulating pv_table since pmap_enter can be
1795 * called at interrupt time.
1797 if (pmap_initialized &&
1798 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1799 pmap_insert_entry(pmap, va, mpte, m);
1804 * Increment counters
1806 pmap->pm_stats.resident_count++;
1808 pmap->pm_stats.wired_count++;
1812 * Now validate mapping with desired protection/wiring.
1814 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1817 newpte |= VPTE_WIRED;
1821 * if the mapping or permission bits are different, we need
1822 * to update the pte.
1824 if ((origpte & ~(VPTE_M|VPTE_A)) != newpte) {
1825 *pte = newpte | VPTE_A;
1826 /* See NOTE: PMAP_INVAL_ADD */
1827 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1829 pmap_inval_flush(&info);
1833 * This is a quick version of pmap_enter(). It is used only under the
1834 * following conditions:
1836 * (1) The pmap is not the kernel_pmap
1837 * (2) The page is not to be wired into the map
1838 * (3) The page is to mapped read-only in the pmap (initially that is)
1839 * (4) The calling procedure is responsible for flushing the TLB
1840 * (5) The page is always managed
1841 * (6) There is no prior mapping at the VA
1845 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
1849 pmap_inval_info info;
1853 KKASSERT(pmap != &kernel_pmap);
1854 pmap_inval_init(&info);
1856 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1859 * Instantiate the page table page if required
1863 * Calculate pagetable page index
1865 ptepindex = va >> PDRSHIFT;
1866 if (mpte && (mpte->pindex == ptepindex)) {
1871 * Get the page directory entry
1873 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1876 * If the page table page is mapped, we just increment
1877 * the hold count, and activate it.
1880 if (ptepa & VPTE_PS)
1881 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1882 if (pmap->pm_ptphint &&
1883 (pmap->pm_ptphint->pindex == ptepindex)) {
1884 mpte = pmap->pm_ptphint;
1886 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1887 pmap->pm_ptphint = mpte;
1893 mpte = _pmap_allocpte(pmap, ptepindex);
1898 * Ok, now that the page table page has been validated, get the pte.
1899 * If the pte is already mapped undo mpte's hold_count and
1902 pte = pmap_pte(pmap, va);
1905 pmap_unwire_pte_hold(pmap, mpte, &info);
1910 * Enter on the PV list if part of our managed memory. Note that we
1911 * raise IPL while manipulating pv_table since pmap_enter can be
1912 * called at interrupt time.
1914 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
1915 pmap_insert_entry(pmap, va, mpte, m);
1918 * Increment counters
1920 pmap->pm_stats.resident_count++;
1922 pa = VM_PAGE_TO_PHYS(m);
1925 * Now validate mapping with RO protection
1927 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1928 *pte = pa | VPTE_V | VPTE_U;
1930 *pte = pa | VPTE_V | VPTE_U | VPTE_MANAGED;
1936 * Extract the physical address for the translation at the specified
1937 * virtual address in the pmap.
1940 pmap_extract(pmap_t pmap, vm_offset_t va)
1945 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
1946 if (pte & VPTE_PS) {
1947 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
1948 rtval |= va & SEG_MASK;
1950 pte = *get_ptbase(pmap, va);
1951 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
1958 #define MAX_INIT_PT (96)
1961 * This routine preloads the ptes for a given object into the specified pmap.
1962 * This eliminates the blast of soft faults on process startup and
1963 * immediately after an mmap.
1965 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
1968 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
1969 vm_object_t object, vm_pindex_t pindex,
1970 vm_size_t size, int limit)
1972 struct rb_vm_page_scan_info info;
1976 * We can't preinit if read access isn't set or there is no pmap
1979 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
1983 * We can't preinit if the pmap is not the current pmap
1985 if (curproc == NULL || pmap != vmspace_pmap(curproc->p_vmspace))
1988 psize = size >> PAGE_SHIFT;
1990 if ((object->type != OBJT_VNODE) ||
1991 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
1992 (object->resident_page_count > MAX_INIT_PT))) {
1996 if (psize + pindex > object->size) {
1997 if (object->size < pindex)
1999 psize = object->size - pindex;
2006 * Use a red-black scan to traverse the requested range and load
2007 * any valid pages found into the pmap.
2009 * We cannot safely scan the object's memq unless we are in a
2010 * critical section since interrupts can remove pages from objects.
2012 info.start_pindex = pindex;
2013 info.end_pindex = pindex + psize - 1;
2020 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2021 pmap_object_init_pt_callback, &info);
2027 pmap_object_init_pt_callback(vm_page_t p, void *data)
2029 struct rb_vm_page_scan_info *info = data;
2030 vm_pindex_t rel_index;
2032 * don't allow an madvise to blow away our really
2033 * free pages allocating pv entries.
2035 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2036 vmstats.v_free_count < vmstats.v_free_reserved) {
2039 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2040 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2041 if ((p->queue - p->pc) == PQ_CACHE)
2042 vm_page_deactivate(p);
2044 rel_index = p->pindex - info->start_pindex;
2045 info->mpte = pmap_enter_quick(info->pmap,
2046 info->addr + i386_ptob(rel_index),
2048 vm_page_flag_set(p, PG_MAPPED);
2055 * pmap_prefault provides a quick way of clustering pagefaults into a
2056 * processes address space. It is a "cousin" of pmap_object_init_pt,
2057 * except it runs at page fault time instead of mmap time.
2061 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2063 static int pmap_prefault_pageorder[] = {
2064 -PAGE_SIZE, PAGE_SIZE,
2065 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2066 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2067 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2071 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2081 * We do not currently prefault mappings that use virtual page
2082 * tables. We do not prefault foreign pmaps.
2084 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2086 if (curproc == NULL || (pmap != vmspace_pmap(curproc->p_vmspace)))
2089 object = entry->object.vm_object;
2091 starta = addra - PFBAK * PAGE_SIZE;
2092 if (starta < entry->start)
2093 starta = entry->start;
2094 else if (starta > addra)
2098 * critical section protection is required to maintain the
2099 * page/object association, interrupts can free pages and remove
2100 * them from their objects.
2104 for (i = 0; i < PAGEORDER_SIZE; i++) {
2105 vm_object_t lobject;
2108 addr = addra + pmap_prefault_pageorder[i];
2109 if (addr > addra + (PFFOR * PAGE_SIZE))
2112 if (addr < starta || addr >= entry->end)
2116 * Make sure the page table page already exists
2118 if ((*pmap_pde(pmap, addr)) == NULL)
2122 * Get a pointer to the pte and make sure that no valid page
2125 pte = get_ptbase(pmap, addr);
2130 * Get the page to be mapped
2132 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2135 for (m = vm_page_lookup(lobject, pindex);
2136 (!m && (lobject->type == OBJT_DEFAULT) &&
2137 (lobject->backing_object));
2138 lobject = lobject->backing_object
2140 if (lobject->backing_object_offset & PAGE_MASK)
2142 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2143 m = vm_page_lookup(lobject->backing_object, pindex);
2147 * give-up when a page is not in memory
2153 * If everything meets the requirements for pmap_enter_quick(),
2154 * then enter the page.
2157 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2159 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2161 if ((m->queue - m->pc) == PQ_CACHE) {
2162 vm_page_deactivate(m);
2165 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2166 vm_page_flag_set(m, PG_MAPPED);
2174 * Routine: pmap_change_wiring
2175 * Function: Change the wiring attribute for a map/virtual-address
2177 * In/out conditions:
2178 * The mapping must already exist in the pmap.
2181 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2188 pte = get_ptbase(pmap, va);
2190 if (wired && (*pte & VPTE_WIRED) == 0)
2191 ++pmap->pm_stats.wired_count;
2192 else if (!wired && (*pte & VPTE_WIRED))
2193 --pmap->pm_stats.wired_count;
2194 KKASSERT(pmap->pm_stats.wired_count >= 0);
2197 * Wiring is not a hardware characteristic so there is no need to
2198 * invalidate TLB. However, in an SMP environment we must use
2199 * a locked bus cycle to update the pte (if we are not using
2200 * the pmap_inval_*() API that is)... it's ok to do this for simple
2204 atomic_set_int(pte, VPTE_WIRED);
2206 atomic_clear_int(pte, VPTE_WIRED);
2210 * Copy the range specified by src_addr/len
2211 * from the source map to the range dst_addr/len
2212 * in the destination map.
2214 * This routine is only advisory and need not do anything.
2217 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2218 vm_size_t len, vm_offset_t src_addr)
2220 pmap_inval_info info;
2222 vm_offset_t end_addr = src_addr + len;
2228 if (dst_addr != src_addr)
2230 if (dst_pmap->pm_pdir == NULL)
2232 if (src_pmap->pm_pdir == NULL)
2235 src_frame = get_ptbase1(src_pmap, src_addr);
2236 dst_frame = get_ptbase2(dst_pmap, src_addr);
2238 pmap_inval_init(&info);
2241 pmap_inval_add(&info, dst_pmap, -1);
2242 pmap_inval_add(&info, src_pmap, -1);
2246 * critical section protection is required to maintain the page/object
2247 * association, interrupts can free pages and remove them from
2251 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2252 vpte_t *src_pte, *dst_pte;
2253 vm_page_t dstmpte, srcmpte;
2254 vm_offset_t srcptepaddr;
2257 if (addr >= VM_MAX_USER_ADDRESS)
2258 panic("pmap_copy: invalid to pmap_copy page tables\n");
2261 * Don't let optional prefaulting of pages make us go
2262 * way below the low water mark of free pages or way
2263 * above high water mark of used pv entries.
2265 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2266 pv_entry_count > pv_entry_high_water)
2269 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2270 ptepindex = addr >> PDRSHIFT;
2272 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2273 if (srcptepaddr == 0)
2276 if (srcptepaddr & VPTE_PS) {
2277 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2278 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2279 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2284 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2285 if ((srcmpte == NULL) ||
2286 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2289 if (pdnxt > end_addr)
2292 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2293 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2294 while (addr < pdnxt) {
2298 * we only virtual copy managed pages
2300 if ((ptetemp & VPTE_MANAGED) != 0) {
2302 * We have to check after allocpte for the
2303 * pte still being around... allocpte can
2306 dstmpte = pmap_allocpte(dst_pmap, addr);
2307 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2309 * Clear the modified and accessed
2310 * (referenced) bits during the copy.
2312 * We do not have to clear the write
2313 * bit to force a fault-on-modify
2314 * because the real kernel's target
2315 * pmap is empty and will fault anyway.
2317 m = PHYS_TO_VM_PAGE(ptetemp);
2318 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2319 dst_pmap->pm_stats.resident_count++;
2320 pmap_insert_entry(dst_pmap, addr,
2323 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2325 if (dstmpte->hold_count >= srcmpte->hold_count)
2334 pmap_inval_flush(&info);
2340 * Zero the specified PA by mapping the page into KVM and clearing its
2343 * This function may be called from an interrupt and no locking is
2347 pmap_zero_page(vm_paddr_t phys)
2349 struct mdglobaldata *gd = mdcpu;
2353 panic("pmap_zero_page: CMAP3 busy");
2354 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2355 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2357 bzero(gd->gd_CADDR3, PAGE_SIZE);
2363 * pmap_page_assertzero:
2365 * Assert that a page is empty, panic if it isn't.
2368 pmap_page_assertzero(vm_paddr_t phys)
2370 struct mdglobaldata *gd = mdcpu;
2375 panic("pmap_zero_page: CMAP3 busy");
2376 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2377 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2378 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2379 for (i = 0; i < PAGE_SIZE; i += 4) {
2380 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2381 panic("pmap_page_assertzero() @ %p not zero!\n",
2382 (void *)gd->gd_CADDR3);
2392 * Zero part of a physical page by mapping it into memory and clearing
2393 * its contents with bzero.
2395 * off and size may not cover an area beyond a single hardware page.
2398 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2400 struct mdglobaldata *gd = mdcpu;
2404 panic("pmap_zero_page: CMAP3 busy");
2405 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2406 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2407 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2409 bzero((char *)gd->gd_CADDR3 + off, size);
2417 * Copy the physical page from the source PA to the target PA.
2418 * This function may be called from an interrupt. No locking
2422 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2424 struct mdglobaldata *gd = mdcpu;
2427 if (*(int *) gd->gd_CMAP1)
2428 panic("pmap_copy_page: CMAP1 busy");
2429 if (*(int *) gd->gd_CMAP2)
2430 panic("pmap_copy_page: CMAP2 busy");
2432 *(int *) gd->gd_CMAP1 = VPTE_V | VPTE_R | (src & PG_FRAME) | VPTE_A;
2433 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2435 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2436 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2438 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2440 *(int *) gd->gd_CMAP1 = 0;
2441 *(int *) gd->gd_CMAP2 = 0;
2446 * pmap_copy_page_frag:
2448 * Copy the physical page from the source PA to the target PA.
2449 * This function may be called from an interrupt. No locking
2453 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2455 struct mdglobaldata *gd = mdcpu;
2458 if (*(int *) gd->gd_CMAP1)
2459 panic("pmap_copy_page: CMAP1 busy");
2460 if (*(int *) gd->gd_CMAP2)
2461 panic("pmap_copy_page: CMAP2 busy");
2463 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2464 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2466 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2467 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2469 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2470 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2473 *(int *) gd->gd_CMAP1 = 0;
2474 *(int *) gd->gd_CMAP2 = 0;
2479 * Returns true if the pmap's pv is one of the first
2480 * 16 pvs linked to from this page. This count may
2481 * be changed upwards or downwards in the future; it
2482 * is only necessary that true be returned for a small
2483 * subset of pmaps for proper page aging.
2486 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2491 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2496 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2497 if (pv->pv_pmap == pmap) {
2510 * Remove all pages from specified address space
2511 * this aids process exit speeds. Also, this code
2512 * is special cased for current process only, but
2513 * can have the more generic (and slightly slower)
2514 * mode enabled. This is much faster than pmap_remove
2515 * in the case of running down an entire address space.
2518 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2523 pmap_inval_info info;
2525 int32_t save_generation;
2527 if (curproc && pmap == vmspace_pmap(curproc->p_vmspace))
2532 pmap_inval_init(&info);
2534 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2535 if (pv->pv_va >= eva || pv->pv_va < sva) {
2536 npv = TAILQ_NEXT(pv, pv_plist);
2540 KKASSERT(pmap == pv->pv_pmap);
2542 pte = pmap_pte(pmap, pv->pv_va);
2546 * We cannot remove wired pages from a process' mapping
2549 if (tpte & VPTE_WIRED) {
2550 npv = TAILQ_NEXT(pv, pv_plist);
2554 /* See NOTE: PMAP_INVAL_ADD */
2555 pmap_inval_add(&info, pmap, pv->pv_va);
2557 m = PHYS_TO_VM_PAGE(tpte);
2559 KASSERT(m < &vm_page_array[vm_page_array_size],
2560 ("pmap_remove_pages: bad tpte %x", tpte));
2562 pmap->pm_stats.resident_count--;
2565 * Update the vm_page_t clean and reference bits.
2567 if (tpte & VPTE_M) {
2571 npv = TAILQ_NEXT(pv, pv_plist);
2572 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2573 save_generation = ++pmap->pm_generation;
2575 m->md.pv_list_count--;
2576 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2577 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2578 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2581 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2585 * Restart the scan if we blocked during the unuse or free
2586 * calls and other removals were made.
2588 if (save_generation != pmap->pm_generation) {
2589 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2590 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2593 pmap_inval_flush(&info);
2598 * pmap_testbit tests bits in pte's
2599 * note that the testbit/changebit routines are inline,
2600 * and a lot of things compile-time evaluate.
2603 pmap_testbit(vm_page_t m, int bit)
2608 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2611 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2616 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2618 * if the bit being tested is the modified bit, then
2619 * mark clean_map and ptes as never
2622 if (bit & (VPTE_A|VPTE_M)) {
2623 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2627 #if defined(PMAP_DIAGNOSTIC)
2629 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2633 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2644 * This routine is used to clear bits in ptes. Certain bits require special
2645 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2647 static __inline void
2648 pmap_clearbit(vm_page_t m, int bit)
2650 struct pmap_inval_info info;
2655 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2658 pmap_inval_init(&info);
2662 * Loop over all current mappings setting/clearing as appropos If
2663 * setting RO do we need to clear the VAC?
2665 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2667 * don't write protect pager mappings
2669 if (bit == VPTE_W) {
2670 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2674 #if defined(PMAP_DIAGNOSTIC)
2676 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2682 * Careful here. We can use a locked bus instruction to
2683 * clear VPTE_A or VPTE_M safely but we need to synchronize
2684 * with the target cpus when we mess with VPTE_W.
2686 * On virtual kernels we must force a new fault-on-write
2687 * in the real kernel if we clear the Modify bit ourselves,
2688 * otherwise the real kernel will not get a new fault and
2689 * will never set our Modify bit again.
2691 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2692 if (bit & (VPTE_W|VPTE_M))
2693 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2697 if (bit == VPTE_W) {
2698 if (pbits & VPTE_M) {
2701 atomic_clear_int(pte, VPTE_M|VPTE_W);
2702 } else if (bit == VPTE_M) {
2704 * We do not have to make the page read-only
2705 * when clearing the Modify bit. The real
2706 * kernel will make the real PTE read-only
2707 * or otherwise detect the write and set
2708 * our VPTE_M again simply by us invalidating
2709 * the real kernel VA for the pmap (as we did
2710 * above). This allows the real kernel to
2711 * handle the write fault without forwarding
2714 atomic_clear_int(pte, VPTE_M);
2716 atomic_clear_int(pte, bit);
2720 pmap_inval_flush(&info);
2725 * pmap_page_protect:
2727 * Lower the permission for all mappings to a given page.
2730 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2732 if ((prot & VM_PROT_WRITE) == 0) {
2733 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2734 pmap_clearbit(m, VPTE_W);
2742 pmap_phys_address(int ppn)
2744 return (i386_ptob(ppn));
2748 * pmap_ts_referenced:
2750 * Return a count of reference bits for a page, clearing those bits.
2751 * It is not necessary for every reference bit to be cleared, but it
2752 * is necessary that 0 only be returned when there are truly no
2753 * reference bits set.
2755 * XXX: The exact number of bits to check and clear is a matter that
2756 * should be tested and standardized at some point in the future for
2757 * optimal aging of shared pages.
2760 pmap_ts_referenced(vm_page_t m)
2762 pv_entry_t pv, pvf, pvn;
2766 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2771 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2776 pvn = TAILQ_NEXT(pv, pv_list);
2778 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2780 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2782 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2785 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2787 if (pte && (*pte & VPTE_A)) {
2789 atomic_clear_int(pte, VPTE_A);
2791 atomic_clear_int_nonlocked(pte, VPTE_A);
2798 } while ((pv = pvn) != NULL && pv != pvf);
2808 * Return whether or not the specified physical page was modified
2809 * in any physical maps.
2812 pmap_is_modified(vm_page_t m)
2814 return pmap_testbit(m, VPTE_M);
2818 * Clear the modify bits on the specified physical page.
2821 pmap_clear_modify(vm_page_t m)
2823 pmap_clearbit(m, VPTE_M);
2827 * pmap_clear_reference:
2829 * Clear the reference bit on the specified physical page.
2832 pmap_clear_reference(vm_page_t m)
2834 pmap_clearbit(m, VPTE_A);
2838 * Miscellaneous support routines follow
2842 i386_protection_init(void)
2846 kp = protection_codes;
2847 for (prot = 0; prot < 8; prot++) {
2848 if (prot & VM_PROT_READ)
2850 if (prot & VM_PROT_WRITE)
2852 if (prot & VM_PROT_EXECUTE)
2859 * Map a set of physical memory pages into the kernel virtual
2860 * address space. Return a pointer to where it is mapped. This
2861 * routine is intended to be used for mapping device memory,
2864 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2868 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2870 vm_offset_t va, tmpva, offset;
2873 offset = pa & PAGE_MASK;
2874 size = roundup(offset + size, PAGE_SIZE);
2876 va = kmem_alloc_nofault(&kernel_map, size);
2878 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2880 pa = pa & VPTE_FRAME;
2881 for (tmpva = va; size > 0;) {
2882 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2883 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2891 return ((void *)(va + offset));
2895 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2897 vm_offset_t base, offset;
2899 base = va & VPTE_FRAME;
2900 offset = va & PAGE_MASK;
2901 size = roundup(offset + size, PAGE_SIZE);
2902 pmap_qremove(va, size >> PAGE_SHIFT);
2903 kmem_free(&kernel_map, base, size);
2907 * perform the pmap work for mincore
2910 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2916 ptep = pmap_pte(pmap, addr);
2921 if ((pte = *ptep) != 0) {
2924 val = MINCORE_INCORE;
2925 if ((pte & VPTE_MANAGED) == 0)
2928 pa = pte & VPTE_FRAME;
2930 m = PHYS_TO_VM_PAGE(pa);
2936 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2938 * Modified by someone
2940 else if (m->dirty || pmap_is_modified(m))
2941 val |= MINCORE_MODIFIED_OTHER;
2946 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2949 * Referenced by someone
2951 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
2952 val |= MINCORE_REFERENCED_OTHER;
2953 vm_page_flag_set(m, PG_REFERENCED);
2960 pmap_activate(struct proc *p)
2964 pmap = vmspace_pmap(p->p_vmspace);
2966 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2968 pmap->pm_active |= 1;
2970 #if defined(SWTCH_OPTIM_STATS)
2974 KKASSERT((p == curproc));
2976 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
2977 load_cr3(curthread->td_pcb->pcb_cr3);
2982 pmap_deactivate(struct proc *p)
2986 pmap = vmspace_pmap(p->p_vmspace);
2988 atomic_clear_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2990 pmap->pm_active &= ~1;
2993 * XXX - note we do not adjust %cr3. The caller is expected to
2994 * activate a new pmap or do a thread-exit.
2999 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3002 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3006 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3013 static void pads (pmap_t pm);
3014 void pmap_pvdump (vm_paddr_t pa);
3016 /* print address space of pmap*/
3024 if (pm == &kernel_pmap)
3026 for (i = 0; i < 1024; i++)
3028 for (j = 0; j < 1024; j++) {
3029 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3030 if (pm == &kernel_pmap && va < KERNBASE)
3032 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3034 ptep = pmap_pte(pm, va);
3035 if (ptep && (*ptep & VPTE_V)) {
3037 (void *)va, (unsigned)*ptep);
3044 pmap_pvdump(vm_paddr_t pa)
3049 kprintf("pa %08llx", (long long)pa);
3050 m = PHYS_TO_VM_PAGE(pa);
3051 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3053 kprintf(" -> pmap %p, va %x, flags %x",
3054 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3056 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);