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 1.16 2007/02/03 17:05:59 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);
860 p->p_addr = (void *)td->td_kstack;
864 td->td_switch = cpu_heavy_switch;
866 KKASSERT(td->td_mpcount == 1);
868 bzero(p->p_addr, sizeof(*p->p_addr));
872 * Destroy the UPAGES for a process that has exited and disassociate
873 * the process from its thread.
876 pmap_dispose_proc(struct proc *p)
878 struct thread *td = NULL;
881 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
883 lp = ONLY_LWP_IN_PROC(p);
884 if (lp != NULL && (td = lp->lwp_thread) != NULL) {
885 lp->lwp_thread = NULL;
894 * We pre-allocate all page table pages for kernel virtual memory so
895 * this routine will only be called if KVM has been exhausted.
898 pmap_growkernel(vm_offset_t addr)
900 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
902 if (addr > virtual_end - SEG_SIZE)
903 panic("KVM exhausted");
904 kernel_vm_end = addr;
908 * The modification bit is not tracked for any pages in this range. XXX
909 * such pages in this maps should always use pmap_k*() functions and not
912 * XXX User and kernel address spaces are independant for virtual kernels,
913 * this function only applies to the kernel pmap.
916 pmap_track_modified(pmap_t pmap, vm_offset_t va)
918 if (pmap != &kernel_pmap)
920 if ((va < clean_sva) || (va >= clean_eva))
926 /************************************************************************
927 * Procedures supporting managed page table pages *
928 ************************************************************************
930 * These procedures are used to track managed page table pages. These pages
931 * use the page table page's vm_page_t to track PTEs in the page. The
932 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
934 * This allows the system to throw away page table pages for user processes
935 * at will and reinstantiate them on demand.
939 * This routine works like vm_page_lookup() but also blocks as long as the
940 * page is busy. This routine does not busy the page it returns.
942 * Unless the caller is managing objects whos pages are in a known state,
943 * the call should be made with a critical section held so the page's object
944 * association remains valid on return.
947 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
952 m = vm_page_lookup(object, pindex);
953 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
959 * This routine unholds page table pages, and if the hold count
960 * drops to zero, then it decrements the wire count.
963 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
965 pmap_inval_flush(info);
966 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
969 if (m->hold_count == 0) {
971 * unmap the page table page
973 pmap->pm_pdir[m->pindex] = 0;
974 --pmap->pm_stats.resident_count;
975 inval_ptbase_pagedir(pmap, m->pindex);
977 if (pmap->pm_ptphint == m)
978 pmap->pm_ptphint = NULL;
981 * If the page is finally unwired, simply free it.
984 if (m->wire_count == 0) {
987 vm_page_free_zero(m);
988 --vmstats.v_wire_count;
996 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
999 if (m->hold_count == 0)
1000 return _pmap_unwire_pte_hold(pmap, m, info);
1006 * After removing a page table entry, this routine is used to
1007 * conditionally free the page, and manage the hold/wire counts.
1010 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1011 pmap_inval_info_t info)
1017 * page table pages in the kernel_pmap are not managed.
1019 if (pmap == &kernel_pmap)
1021 ptepindex = (va >> PDRSHIFT);
1022 if (pmap->pm_ptphint &&
1023 (pmap->pm_ptphint->pindex == ptepindex)) {
1024 mpte = pmap->pm_ptphint;
1026 pmap_inval_flush(info);
1027 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1028 pmap->pm_ptphint = mpte;
1031 return pmap_unwire_pte_hold(pmap, mpte, info);
1035 * Attempt to release and free an vm_page in a pmap. Returns 1 on success,
1036 * 0 on failure (if the procedure had to sleep).
1039 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1041 vpte_t *pde = pmap->pm_pdir;
1043 * This code optimizes the case of freeing non-busy
1044 * page-table pages. Those pages are zero now, and
1045 * might as well be placed directly into the zero queue.
1047 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1051 pmap->pm_stats.resident_count--;
1053 if (p->hold_count) {
1054 panic("pmap_release: freeing held page table page");
1057 * Page directory pages need to have the kernel stuff cleared, so
1058 * they can go into the zero queue also.
1060 * In virtual kernels there is no 'kernel stuff'. For the moment
1061 * I just make sure the whole thing has been zero'd even though
1062 * it should already be completely zero'd.
1064 * pmaps for vkernels do not self-map because they do not share
1065 * their address space with the vkernel. Clearing of pde[] thus
1066 * only applies to page table pages and not to the page directory
1069 if (p->pindex == pmap->pm_pdindex) {
1070 bzero(pde, VPTE_PAGETABLE_SIZE);
1071 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1077 * Clear the matching hint
1079 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1080 pmap->pm_ptphint = NULL;
1083 * And throw the page away. The page is completely zero'd out so
1084 * optimize the free call.
1087 vmstats.v_wire_count--;
1088 vm_page_free_zero(p);
1093 * This routine is called if the page table page is not mapped in the page
1096 * The routine is broken up into two parts for readability.
1099 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1105 * Find or fabricate a new pagetable page
1107 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1108 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1110 KASSERT(m->queue == PQ_NONE,
1111 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1113 if (m->wire_count == 0)
1114 vmstats.v_wire_count++;
1118 * Increment the hold count for the page table page
1119 * (denoting a new mapping.)
1124 * Map the pagetable page into the process address space, if
1125 * it isn't already there.
1127 pmap->pm_stats.resident_count++;
1129 ptepa = VM_PAGE_TO_PHYS(m);
1130 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
1134 * We are likely about to access this page table page, so set the
1135 * page table hint to reduce overhead.
1137 pmap->pm_ptphint = m;
1140 * Try to use the new mapping, but if we cannot, then
1141 * do it with the routine that maps the page explicitly.
1143 if ((m->flags & PG_ZERO) == 0)
1144 pmap_zero_page(ptepa);
1146 m->valid = VM_PAGE_BITS_ALL;
1147 vm_page_flag_clear(m, PG_ZERO);
1148 vm_page_flag_set(m, PG_MAPPED);
1155 * Determine the page table page required to access the VA in the pmap
1156 * and allocate it if necessary. Return a held vm_page_t for the page.
1158 * Only used with user pmaps.
1161 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1168 * Calculate pagetable page index
1170 ptepindex = va >> PDRSHIFT;
1173 * Get the page directory entry
1175 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1178 * This supports switching from a 4MB page to a
1181 if (ptepa & VPTE_PS) {
1182 pmap->pm_pdir[ptepindex] = 0;
1189 * If the page table page is mapped, we just increment the
1190 * hold count, and activate it.
1194 * In order to get the page table page, try the
1197 if (pmap->pm_ptphint &&
1198 (pmap->pm_ptphint->pindex == ptepindex)) {
1199 m = pmap->pm_ptphint;
1201 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1202 pmap->pm_ptphint = m;
1208 * Here if the pte page isn't mapped, or if it has been deallocated.
1210 return _pmap_allocpte(pmap, ptepindex);
1213 /************************************************************************
1214 * Managed pages in pmaps *
1215 ************************************************************************
1217 * All pages entered into user pmaps and some pages entered into the kernel
1218 * pmap are managed, meaning that pmap_protect() and other related management
1219 * functions work on these pages.
1223 * free the pv_entry back to the free list. This function may be
1224 * called from an interrupt.
1226 static __inline void
1227 free_pv_entry(pv_entry_t pv)
1234 * get a new pv_entry, allocating a block from the system
1235 * when needed. This function may be called from an interrupt.
1241 if (pv_entry_high_water &&
1242 (pv_entry_count > pv_entry_high_water) &&
1243 (pmap_pagedaemon_waken == 0)) {
1244 pmap_pagedaemon_waken = 1;
1245 wakeup (&vm_pages_needed);
1247 return zalloc(&pvzone);
1251 * This routine is very drastic, but can save the system
1259 static int warningdone=0;
1261 if (pmap_pagedaemon_waken == 0)
1264 if (warningdone < 5) {
1265 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1269 for(i = 0; i < vm_page_array_size; i++) {
1270 m = &vm_page_array[i];
1271 if (m->wire_count || m->hold_count || m->busy ||
1272 (m->flags & PG_BUSY))
1276 pmap_pagedaemon_waken = 0;
1280 * If it is the first entry on the list, it is actually
1281 * in the header and we must copy the following entry up
1282 * to the header. Otherwise we must search the list for
1283 * the entry. In either case we free the now unused entry.
1286 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1287 vm_offset_t va, pmap_inval_info_t info)
1293 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1294 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1295 if (pmap == pv->pv_pmap && va == pv->pv_va)
1299 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1300 if (va == pv->pv_va)
1306 * Note that pv_ptem is NULL if the page table page itself is not
1307 * managed, even if the page being removed IS managed.
1311 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1312 m->md.pv_list_count--;
1313 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1314 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1315 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1316 ++pmap->pm_generation;
1317 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1325 * Create a pv entry for page at pa for (pmap, va). If the page table page
1326 * holding the VA is managed, mpte will be non-NULL.
1329 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1334 pv = get_pv_entry();
1339 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1340 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1341 m->md.pv_list_count++;
1347 * pmap_remove_pte: do the things to unmap a page in a process
1350 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va,
1351 pmap_inval_info_t info)
1356 oldpte = loadandclear(ptq);
1357 pmap_inval_add(info, pmap, va); /* See NOTE: PMAP_INVAL_ADD */
1358 if (oldpte & VPTE_WIRED)
1359 --pmap->pm_stats.wired_count;
1360 KKASSERT(pmap->pm_stats.wired_count >= 0);
1364 * Machines that don't support invlpg, also don't support
1365 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1368 if (oldpte & VPTE_G)
1369 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1371 pmap->pm_stats.resident_count -= 1;
1372 if (oldpte & VPTE_MANAGED) {
1373 m = PHYS_TO_VM_PAGE(oldpte);
1374 if (oldpte & VPTE_M) {
1375 #if defined(PMAP_DIAGNOSTIC)
1376 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1378 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1382 if (pmap_track_modified(pmap, va))
1385 if (oldpte & VPTE_A)
1386 vm_page_flag_set(m, PG_REFERENCED);
1387 return pmap_remove_entry(pmap, m, va, info);
1389 return pmap_unuse_pt(pmap, va, NULL, info);
1398 * Remove a single page from a process address space.
1400 * This function may not be called from an interrupt if the pmap is
1404 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1409 * if there is no pte for this address, just skip it!!! Otherwise
1410 * get a local va for mappings for this pmap and remove the entry.
1412 if (*pmap_pde(pmap, va) != 0) {
1413 ptq = get_ptbase(pmap, va);
1415 pmap_remove_pte(pmap, ptq, va, info);
1423 * Remove the given range of addresses from the specified map.
1425 * It is assumed that the start and end are properly
1426 * rounded to the page size.
1428 * This function may not be called from an interrupt if the pmap is
1432 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1436 vm_offset_t ptpaddr;
1437 vm_pindex_t sindex, eindex;
1438 struct pmap_inval_info info;
1443 KKASSERT(pmap->pm_stats.resident_count >= 0);
1444 if (pmap->pm_stats.resident_count == 0)
1447 pmap_inval_init(&info);
1450 * special handling of removing one page. a very
1451 * common operation and easy to short circuit some
1454 if (((sva + PAGE_SIZE) == eva) &&
1455 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1456 pmap_remove_page(pmap, sva, &info);
1457 pmap_inval_flush(&info);
1462 * Get a local virtual address for the mappings that are being
1465 * XXX this is really messy because the kernel pmap is not relative
1468 sindex = (sva >> PAGE_SHIFT);
1469 eindex = (eva >> PAGE_SHIFT);
1471 for (; sindex < eindex; sindex = pdnxt) {
1475 * Calculate index for next page table.
1477 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1478 if (pmap->pm_stats.resident_count == 0)
1481 pdirindex = sindex / NPDEPG;
1482 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1483 pmap->pm_pdir[pdirindex] = 0;
1484 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1485 inval_ptbase_pagedir(pmap, pdirindex);
1490 * Weed out invalid mappings. Note: we assume that the page
1491 * directory table is always allocated, and in kernel virtual.
1497 * Limit our scan to either the end of the va represented
1498 * by the current page table page, or to the end of the
1499 * range being removed.
1505 * NOTE: pmap_remove_pte() can block.
1507 for (; sindex != pdnxt; sindex++) {
1510 ptbase = get_ptbase(pmap, sindex << PAGE_SHIFT);
1513 va = i386_ptob(sindex);
1514 if (pmap_remove_pte(pmap, ptbase, va, &info))
1518 pmap_inval_flush(&info);
1524 * Removes this physical page from all physical maps in which it resides.
1525 * Reflects back modify bits to the pager.
1527 * This routine may not be called from an interrupt.
1530 pmap_remove_all(vm_page_t m)
1532 struct pmap_inval_info info;
1536 #if defined(PMAP_DIAGNOSTIC)
1538 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1541 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1542 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1546 pmap_inval_init(&info);
1548 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1549 pv->pv_pmap->pm_stats.resident_count--;
1551 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1552 KKASSERT(pte != NULL);
1554 tpte = loadandclear(pte);
1555 /* See NOTE: PMAP_INVAL_ADD */
1556 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1557 if (tpte & VPTE_WIRED)
1558 --pv->pv_pmap->pm_stats.wired_count;
1559 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
1562 vm_page_flag_set(m, PG_REFERENCED);
1565 * Update the vm_page_t clean and reference bits.
1567 if (tpte & VPTE_M) {
1568 #if defined(PMAP_DIAGNOSTIC)
1569 if (pmap_nw_modified((pt_entry_t) tpte)) {
1571 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1575 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
1578 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1579 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1580 ++pv->pv_pmap->pm_generation;
1581 m->md.pv_list_count--;
1582 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1586 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1588 pmap_inval_flush(&info);
1594 * Set the physical protection on the specified range of this map
1597 * This function may not be called from an interrupt if the map is
1598 * not the kernel_pmap.
1601 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1604 vm_offset_t pdnxt, ptpaddr;
1605 vm_pindex_t sindex, eindex;
1607 pmap_inval_info info;
1612 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1613 pmap_remove(pmap, sva, eva);
1617 if (prot & VM_PROT_WRITE)
1620 pmap_inval_init(&info);
1622 ptbase = get_ptbase(pmap, sva);
1624 sindex = (sva >> PAGE_SHIFT);
1625 eindex = (eva >> PAGE_SHIFT);
1628 for (; sindex < eindex; sindex = pdnxt) {
1632 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1634 pdirindex = sindex / NPDEPG;
1635 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1636 pmap->pm_pdir[pdirindex] &= ~(VPTE_M|VPTE_W);
1637 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1638 inval_ptbase_pagedir(pmap, pdirindex);
1643 * Weed out invalid mappings. Note: we assume that the page
1644 * directory table is always allocated, and in kernel virtual.
1649 if (pdnxt > eindex) {
1653 for (; sindex != pdnxt; sindex++) {
1657 pbits = ptbase[sindex - sbase];
1659 if (pbits & VPTE_MANAGED) {
1661 if (pbits & VPTE_A) {
1662 m = PHYS_TO_VM_PAGE(pbits);
1663 vm_page_flag_set(m, PG_REFERENCED);
1666 if (pbits & VPTE_M) {
1667 if (pmap_track_modified(pmap, i386_ptob(sindex))) {
1669 m = PHYS_TO_VM_PAGE(pbits);
1678 if (pbits != ptbase[sindex - sbase]) {
1679 ptbase[sindex - sbase] = pbits;
1680 /* See NOTE: PMAP_INVAL_ADD */
1681 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1685 pmap_inval_flush(&info);
1689 * Enter a managed page into a pmap. If the page is not wired related pmap
1690 * data can be destroyed at any time for later demand-operation.
1692 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1693 * specified protection, and wire the mapping if requested.
1695 * NOTE: This routine may not lazy-evaluate or lose information. The
1696 * page must actually be inserted into the given map NOW.
1698 * NOTE: When entering a page at a KVA address, the pmap must be the
1702 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1708 vm_offset_t origpte, newpte;
1710 pmap_inval_info info;
1718 * Get the page table page. The kernel_pmap's page table pages
1719 * are preallocated and have no associated vm_page_t.
1721 if (pmap == &kernel_pmap)
1724 mpte = pmap_allocpte(pmap, va);
1726 pmap_inval_init(&info);
1727 pte = pmap_pte(pmap, va);
1730 * Page Directory table entry not valid, we need a new PT page
1731 * and pmap_allocpte() didn't give us one. Oops!
1734 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1738 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1740 opa = origpte & VPTE_FRAME;
1742 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);
1745 if (origpte & VPTE_PS)
1746 panic("pmap_enter: attempted pmap_enter on 4MB page");
1749 * Mapping has not changed, must be protection or wiring change.
1751 if (origpte && (opa == pa)) {
1753 * Wiring change, just update stats. We don't worry about
1754 * wiring PT pages as they remain resident as long as there
1755 * are valid mappings in them. Hence, if a user page is wired,
1756 * the PT page will be also.
1758 if (wired && ((origpte & VPTE_WIRED) == 0))
1759 ++pmap->pm_stats.wired_count;
1760 else if (!wired && (origpte & VPTE_WIRED))
1761 --pmap->pm_stats.wired_count;
1762 KKASSERT(pmap->pm_stats.wired_count >= 0);
1764 #if defined(PMAP_DIAGNOSTIC)
1765 if (pmap_nw_modified((pt_entry_t) origpte)) {
1767 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1773 * Remove the extra pte reference. Note that we cannot
1774 * optimize the RO->RW case because we have adjusted the
1775 * wiring count above and may need to adjust the wiring
1782 * We might be turning off write access to the page,
1783 * so we go ahead and sense modify status.
1785 if (origpte & VPTE_MANAGED) {
1786 if ((origpte & VPTE_M) && pmap_track_modified(pmap, va)) {
1788 om = PHYS_TO_VM_PAGE(opa);
1796 * Mapping has changed, invalidate old range and fall through to
1797 * handle validating new mapping.
1801 err = pmap_remove_pte(pmap, pte, va, &info);
1803 panic("pmap_enter: pte vanished, va: 0x%x", va);
1807 * Enter on the PV list if part of our managed memory. Note that we
1808 * raise IPL while manipulating pv_table since pmap_enter can be
1809 * called at interrupt time.
1811 if (pmap_initialized &&
1812 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1813 pmap_insert_entry(pmap, va, mpte, m);
1818 * Increment counters
1820 pmap->pm_stats.resident_count++;
1822 pmap->pm_stats.wired_count++;
1826 * Now validate mapping with desired protection/wiring.
1828 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1831 newpte |= VPTE_WIRED;
1835 * if the mapping or permission bits are different, we need
1836 * to update the pte.
1838 if ((origpte & ~(VPTE_M|VPTE_A)) != newpte) {
1839 *pte = newpte | VPTE_A;
1840 /* See NOTE: PMAP_INVAL_ADD */
1841 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1843 pmap_inval_flush(&info);
1847 * This is a quick version of pmap_enter(). It is used only under the
1848 * following conditions:
1850 * (1) The pmap is not the kernel_pmap
1851 * (2) The page is not to be wired into the map
1852 * (3) The page is to mapped read-only in the pmap (initially that is)
1853 * (4) The calling procedure is responsible for flushing the TLB
1854 * (5) The page is always managed
1855 * (6) There is no prior mapping at the VA
1859 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
1863 pmap_inval_info info;
1867 KKASSERT(pmap != &kernel_pmap);
1868 pmap_inval_init(&info);
1870 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1873 * Instantiate the page table page if required
1877 * Calculate pagetable page index
1879 ptepindex = va >> PDRSHIFT;
1880 if (mpte && (mpte->pindex == ptepindex)) {
1885 * Get the page directory entry
1887 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1890 * If the page table page is mapped, we just increment
1891 * the hold count, and activate it.
1894 if (ptepa & VPTE_PS)
1895 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1896 if (pmap->pm_ptphint &&
1897 (pmap->pm_ptphint->pindex == ptepindex)) {
1898 mpte = pmap->pm_ptphint;
1900 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1901 pmap->pm_ptphint = mpte;
1907 mpte = _pmap_allocpte(pmap, ptepindex);
1912 * Ok, now that the page table page has been validated, get the pte.
1913 * If the pte is already mapped undo mpte's hold_count and
1916 pte = pmap_pte(pmap, va);
1919 pmap_unwire_pte_hold(pmap, mpte, &info);
1924 * Enter on the PV list if part of our managed memory. Note that we
1925 * raise IPL while manipulating pv_table since pmap_enter can be
1926 * called at interrupt time.
1928 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
1929 pmap_insert_entry(pmap, va, mpte, m);
1932 * Increment counters
1934 pmap->pm_stats.resident_count++;
1936 pa = VM_PAGE_TO_PHYS(m);
1939 * Now validate mapping with RO protection
1941 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1942 *pte = pa | VPTE_V | VPTE_U;
1944 *pte = pa | VPTE_V | VPTE_U | VPTE_MANAGED;
1950 * Extract the physical address for the translation at the specified
1951 * virtual address in the pmap.
1954 pmap_extract(pmap_t pmap, vm_offset_t va)
1959 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
1960 if (pte & VPTE_PS) {
1961 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
1962 rtval |= va & SEG_MASK;
1964 pte = *get_ptbase(pmap, va);
1965 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
1972 #define MAX_INIT_PT (96)
1975 * This routine preloads the ptes for a given object into the specified pmap.
1976 * This eliminates the blast of soft faults on process startup and
1977 * immediately after an mmap.
1979 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
1982 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
1983 vm_object_t object, vm_pindex_t pindex,
1984 vm_size_t size, int limit)
1986 struct rb_vm_page_scan_info info;
1990 * We can't preinit if read access isn't set or there is no pmap
1993 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
1997 * We can't preinit if the pmap is not the current pmap
1999 if (curproc == NULL || pmap != vmspace_pmap(curproc->p_vmspace))
2002 psize = size >> PAGE_SHIFT;
2004 if ((object->type != OBJT_VNODE) ||
2005 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2006 (object->resident_page_count > MAX_INIT_PT))) {
2010 if (psize + pindex > object->size) {
2011 if (object->size < pindex)
2013 psize = object->size - pindex;
2020 * Use a red-black scan to traverse the requested range and load
2021 * any valid pages found into the pmap.
2023 * We cannot safely scan the object's memq unless we are in a
2024 * critical section since interrupts can remove pages from objects.
2026 info.start_pindex = pindex;
2027 info.end_pindex = pindex + psize - 1;
2034 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2035 pmap_object_init_pt_callback, &info);
2041 pmap_object_init_pt_callback(vm_page_t p, void *data)
2043 struct rb_vm_page_scan_info *info = data;
2044 vm_pindex_t rel_index;
2046 * don't allow an madvise to blow away our really
2047 * free pages allocating pv entries.
2049 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2050 vmstats.v_free_count < vmstats.v_free_reserved) {
2053 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2054 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2055 if ((p->queue - p->pc) == PQ_CACHE)
2056 vm_page_deactivate(p);
2058 rel_index = p->pindex - info->start_pindex;
2059 info->mpte = pmap_enter_quick(info->pmap,
2060 info->addr + i386_ptob(rel_index),
2062 vm_page_flag_set(p, PG_MAPPED);
2069 * pmap_prefault provides a quick way of clustering pagefaults into a
2070 * processes address space. It is a "cousin" of pmap_object_init_pt,
2071 * except it runs at page fault time instead of mmap time.
2075 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2077 static int pmap_prefault_pageorder[] = {
2078 -PAGE_SIZE, PAGE_SIZE,
2079 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2080 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2081 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2085 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2095 * We do not currently prefault mappings that use virtual page
2096 * tables. We do not prefault foreign pmaps.
2098 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2100 if (curproc == NULL || (pmap != vmspace_pmap(curproc->p_vmspace)))
2103 object = entry->object.vm_object;
2105 starta = addra - PFBAK * PAGE_SIZE;
2106 if (starta < entry->start)
2107 starta = entry->start;
2108 else if (starta > addra)
2112 * critical section protection is required to maintain the
2113 * page/object association, interrupts can free pages and remove
2114 * them from their objects.
2118 for (i = 0; i < PAGEORDER_SIZE; i++) {
2119 vm_object_t lobject;
2122 addr = addra + pmap_prefault_pageorder[i];
2123 if (addr > addra + (PFFOR * PAGE_SIZE))
2126 if (addr < starta || addr >= entry->end)
2130 * Make sure the page table page already exists
2132 if ((*pmap_pde(pmap, addr)) == NULL)
2136 * Get a pointer to the pte and make sure that no valid page
2139 pte = get_ptbase(pmap, addr);
2144 * Get the page to be mapped
2146 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2149 for (m = vm_page_lookup(lobject, pindex);
2150 (!m && (lobject->type == OBJT_DEFAULT) &&
2151 (lobject->backing_object));
2152 lobject = lobject->backing_object
2154 if (lobject->backing_object_offset & PAGE_MASK)
2156 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2157 m = vm_page_lookup(lobject->backing_object, pindex);
2161 * give-up when a page is not in memory
2167 * If everything meets the requirements for pmap_enter_quick(),
2168 * then enter the page.
2171 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2173 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2175 if ((m->queue - m->pc) == PQ_CACHE) {
2176 vm_page_deactivate(m);
2179 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2180 vm_page_flag_set(m, PG_MAPPED);
2188 * Routine: pmap_change_wiring
2189 * Function: Change the wiring attribute for a map/virtual-address
2191 * In/out conditions:
2192 * The mapping must already exist in the pmap.
2195 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2202 pte = get_ptbase(pmap, va);
2204 if (wired && (*pte & VPTE_WIRED) == 0)
2205 ++pmap->pm_stats.wired_count;
2206 else if (!wired && (*pte & VPTE_WIRED))
2207 --pmap->pm_stats.wired_count;
2208 KKASSERT(pmap->pm_stats.wired_count >= 0);
2211 * Wiring is not a hardware characteristic so there is no need to
2212 * invalidate TLB. However, in an SMP environment we must use
2213 * a locked bus cycle to update the pte (if we are not using
2214 * the pmap_inval_*() API that is)... it's ok to do this for simple
2218 atomic_set_int(pte, VPTE_WIRED);
2220 atomic_clear_int(pte, VPTE_WIRED);
2224 * Copy the range specified by src_addr/len
2225 * from the source map to the range dst_addr/len
2226 * in the destination map.
2228 * This routine is only advisory and need not do anything.
2231 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2232 vm_size_t len, vm_offset_t src_addr)
2234 pmap_inval_info info;
2236 vm_offset_t end_addr = src_addr + len;
2242 if (dst_addr != src_addr)
2244 if (dst_pmap->pm_pdir == NULL)
2246 if (src_pmap->pm_pdir == NULL)
2249 src_frame = get_ptbase1(src_pmap, src_addr);
2250 dst_frame = get_ptbase2(dst_pmap, src_addr);
2252 pmap_inval_init(&info);
2255 pmap_inval_add(&info, dst_pmap, -1);
2256 pmap_inval_add(&info, src_pmap, -1);
2260 * critical section protection is required to maintain the page/object
2261 * association, interrupts can free pages and remove them from
2265 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2266 vpte_t *src_pte, *dst_pte;
2267 vm_page_t dstmpte, srcmpte;
2268 vm_offset_t srcptepaddr;
2271 if (addr >= VM_MAX_USER_ADDRESS)
2272 panic("pmap_copy: invalid to pmap_copy page tables\n");
2275 * Don't let optional prefaulting of pages make us go
2276 * way below the low water mark of free pages or way
2277 * above high water mark of used pv entries.
2279 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2280 pv_entry_count > pv_entry_high_water)
2283 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2284 ptepindex = addr >> PDRSHIFT;
2286 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2287 if (srcptepaddr == 0)
2290 if (srcptepaddr & VPTE_PS) {
2291 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2292 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2293 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2298 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2299 if ((srcmpte == NULL) ||
2300 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2303 if (pdnxt > end_addr)
2306 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2307 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2308 while (addr < pdnxt) {
2312 * we only virtual copy managed pages
2314 if ((ptetemp & VPTE_MANAGED) != 0) {
2316 * We have to check after allocpte for the
2317 * pte still being around... allocpte can
2320 dstmpte = pmap_allocpte(dst_pmap, addr);
2321 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2323 * Clear the modified and accessed
2324 * (referenced) bits during the copy.
2326 * We do not have to clear the write
2327 * bit to force a fault-on-modify
2328 * because the real kernel's target
2329 * pmap is empty and will fault anyway.
2331 m = PHYS_TO_VM_PAGE(ptetemp);
2332 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2333 dst_pmap->pm_stats.resident_count++;
2334 pmap_insert_entry(dst_pmap, addr,
2337 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2339 if (dstmpte->hold_count >= srcmpte->hold_count)
2348 pmap_inval_flush(&info);
2354 * Zero the specified PA by mapping the page into KVM and clearing its
2357 * This function may be called from an interrupt and no locking is
2361 pmap_zero_page(vm_paddr_t phys)
2363 struct mdglobaldata *gd = mdcpu;
2367 panic("pmap_zero_page: CMAP3 busy");
2368 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2369 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2371 bzero(gd->gd_CADDR3, PAGE_SIZE);
2377 * pmap_page_assertzero:
2379 * Assert that a page is empty, panic if it isn't.
2382 pmap_page_assertzero(vm_paddr_t phys)
2384 struct mdglobaldata *gd = mdcpu;
2389 panic("pmap_zero_page: CMAP3 busy");
2390 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2391 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2392 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2393 for (i = 0; i < PAGE_SIZE; i += 4) {
2394 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2395 panic("pmap_page_assertzero() @ %p not zero!\n",
2396 (void *)gd->gd_CADDR3);
2406 * Zero part of a physical page by mapping it into memory and clearing
2407 * its contents with bzero.
2409 * off and size may not cover an area beyond a single hardware page.
2412 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2414 struct mdglobaldata *gd = mdcpu;
2418 panic("pmap_zero_page: CMAP3 busy");
2419 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2420 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2421 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2423 bzero((char *)gd->gd_CADDR3 + off, size);
2431 * Copy the physical page from the source PA to the target PA.
2432 * This function may be called from an interrupt. No locking
2436 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2438 struct mdglobaldata *gd = mdcpu;
2441 if (*(int *) gd->gd_CMAP1)
2442 panic("pmap_copy_page: CMAP1 busy");
2443 if (*(int *) gd->gd_CMAP2)
2444 panic("pmap_copy_page: CMAP2 busy");
2446 *(int *) gd->gd_CMAP1 = VPTE_V | VPTE_R | (src & PG_FRAME) | VPTE_A;
2447 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2449 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2450 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2452 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2454 *(int *) gd->gd_CMAP1 = 0;
2455 *(int *) gd->gd_CMAP2 = 0;
2460 * pmap_copy_page_frag:
2462 * Copy the physical page from the source PA to the target PA.
2463 * This function may be called from an interrupt. No locking
2467 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2469 struct mdglobaldata *gd = mdcpu;
2472 if (*(int *) gd->gd_CMAP1)
2473 panic("pmap_copy_page: CMAP1 busy");
2474 if (*(int *) gd->gd_CMAP2)
2475 panic("pmap_copy_page: CMAP2 busy");
2477 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2478 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2480 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2481 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2483 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2484 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2487 *(int *) gd->gd_CMAP1 = 0;
2488 *(int *) gd->gd_CMAP2 = 0;
2493 * Returns true if the pmap's pv is one of the first
2494 * 16 pvs linked to from this page. This count may
2495 * be changed upwards or downwards in the future; it
2496 * is only necessary that true be returned for a small
2497 * subset of pmaps for proper page aging.
2500 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2505 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2510 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2511 if (pv->pv_pmap == pmap) {
2524 * Remove all pages from specified address space
2525 * this aids process exit speeds. Also, this code
2526 * is special cased for current process only, but
2527 * can have the more generic (and slightly slower)
2528 * mode enabled. This is much faster than pmap_remove
2529 * in the case of running down an entire address space.
2532 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2537 pmap_inval_info info;
2539 int32_t save_generation;
2541 if (curproc && pmap == vmspace_pmap(curproc->p_vmspace))
2546 pmap_inval_init(&info);
2548 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2549 if (pv->pv_va >= eva || pv->pv_va < sva) {
2550 npv = TAILQ_NEXT(pv, pv_plist);
2554 KKASSERT(pmap == pv->pv_pmap);
2556 pte = pmap_pte(pmap, pv->pv_va);
2560 * We cannot remove wired pages from a process' mapping
2563 if (tpte & VPTE_WIRED) {
2564 npv = TAILQ_NEXT(pv, pv_plist);
2568 /* See NOTE: PMAP_INVAL_ADD */
2569 pmap_inval_add(&info, pmap, pv->pv_va);
2571 m = PHYS_TO_VM_PAGE(tpte);
2573 KASSERT(m < &vm_page_array[vm_page_array_size],
2574 ("pmap_remove_pages: bad tpte %x", tpte));
2576 pmap->pm_stats.resident_count--;
2579 * Update the vm_page_t clean and reference bits.
2581 if (tpte & VPTE_M) {
2585 npv = TAILQ_NEXT(pv, pv_plist);
2586 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2587 save_generation = ++pmap->pm_generation;
2589 m->md.pv_list_count--;
2590 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2591 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2592 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2595 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2599 * Restart the scan if we blocked during the unuse or free
2600 * calls and other removals were made.
2602 if (save_generation != pmap->pm_generation) {
2603 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2604 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2607 pmap_inval_flush(&info);
2612 * pmap_testbit tests bits in pte's
2613 * note that the testbit/changebit routines are inline,
2614 * and a lot of things compile-time evaluate.
2617 pmap_testbit(vm_page_t m, int bit)
2622 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2625 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2630 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2632 * if the bit being tested is the modified bit, then
2633 * mark clean_map and ptes as never
2636 if (bit & (VPTE_A|VPTE_M)) {
2637 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2641 #if defined(PMAP_DIAGNOSTIC)
2643 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2647 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2658 * This routine is used to clear bits in ptes. Certain bits require special
2659 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2661 static __inline void
2662 pmap_clearbit(vm_page_t m, int bit)
2664 struct pmap_inval_info info;
2669 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2672 pmap_inval_init(&info);
2676 * Loop over all current mappings setting/clearing as appropos If
2677 * setting RO do we need to clear the VAC?
2679 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2681 * don't write protect pager mappings
2683 if (bit == VPTE_W) {
2684 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2688 #if defined(PMAP_DIAGNOSTIC)
2690 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2696 * Careful here. We can use a locked bus instruction to
2697 * clear VPTE_A or VPTE_M safely but we need to synchronize
2698 * with the target cpus when we mess with VPTE_W.
2700 * On virtual kernels we must force a new fault-on-write
2701 * in the real kernel if we clear the Modify bit ourselves,
2702 * otherwise the real kernel will not get a new fault and
2703 * will never set our Modify bit again.
2705 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2706 if (bit & (VPTE_W|VPTE_M))
2707 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2711 if (bit == VPTE_W) {
2712 if (pbits & VPTE_M) {
2715 atomic_clear_int(pte, VPTE_M|VPTE_W);
2716 } else if (bit == VPTE_M) {
2718 * We do not have to make the page read-only
2719 * when clearing the Modify bit. The real
2720 * kernel will make the real PTE read-only
2721 * or otherwise detect the write and set
2722 * our VPTE_M again simply by us invalidating
2723 * the real kernel VA for the pmap (as we did
2724 * above). This allows the real kernel to
2725 * handle the write fault without forwarding
2728 atomic_clear_int(pte, VPTE_M);
2730 atomic_clear_int(pte, bit);
2734 pmap_inval_flush(&info);
2739 * pmap_page_protect:
2741 * Lower the permission for all mappings to a given page.
2744 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2746 if ((prot & VM_PROT_WRITE) == 0) {
2747 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2748 pmap_clearbit(m, VPTE_W);
2756 pmap_phys_address(int ppn)
2758 return (i386_ptob(ppn));
2762 * pmap_ts_referenced:
2764 * Return a count of reference bits for a page, clearing those bits.
2765 * It is not necessary for every reference bit to be cleared, but it
2766 * is necessary that 0 only be returned when there are truly no
2767 * reference bits set.
2769 * XXX: The exact number of bits to check and clear is a matter that
2770 * should be tested and standardized at some point in the future for
2771 * optimal aging of shared pages.
2774 pmap_ts_referenced(vm_page_t m)
2776 pv_entry_t pv, pvf, pvn;
2780 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2785 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2790 pvn = TAILQ_NEXT(pv, pv_list);
2792 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2794 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2796 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2799 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2801 if (pte && (*pte & VPTE_A)) {
2803 atomic_clear_int(pte, VPTE_A);
2805 atomic_clear_int_nonlocked(pte, VPTE_A);
2812 } while ((pv = pvn) != NULL && pv != pvf);
2822 * Return whether or not the specified physical page was modified
2823 * in any physical maps.
2826 pmap_is_modified(vm_page_t m)
2828 return pmap_testbit(m, VPTE_M);
2832 * Clear the modify bits on the specified physical page.
2835 pmap_clear_modify(vm_page_t m)
2837 pmap_clearbit(m, VPTE_M);
2841 * pmap_clear_reference:
2843 * Clear the reference bit on the specified physical page.
2846 pmap_clear_reference(vm_page_t m)
2848 pmap_clearbit(m, VPTE_A);
2852 * Miscellaneous support routines follow
2856 i386_protection_init(void)
2860 kp = protection_codes;
2861 for (prot = 0; prot < 8; prot++) {
2862 if (prot & VM_PROT_READ)
2864 if (prot & VM_PROT_WRITE)
2866 if (prot & VM_PROT_EXECUTE)
2873 * Map a set of physical memory pages into the kernel virtual
2874 * address space. Return a pointer to where it is mapped. This
2875 * routine is intended to be used for mapping device memory,
2878 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2882 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2884 vm_offset_t va, tmpva, offset;
2887 offset = pa & PAGE_MASK;
2888 size = roundup(offset + size, PAGE_SIZE);
2890 va = kmem_alloc_nofault(&kernel_map, size);
2892 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2894 pa = pa & VPTE_FRAME;
2895 for (tmpva = va; size > 0;) {
2896 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2897 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2905 return ((void *)(va + offset));
2909 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2911 vm_offset_t base, offset;
2913 base = va & VPTE_FRAME;
2914 offset = va & PAGE_MASK;
2915 size = roundup(offset + size, PAGE_SIZE);
2916 pmap_qremove(va, size >> PAGE_SHIFT);
2917 kmem_free(&kernel_map, base, size);
2921 * perform the pmap work for mincore
2924 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2930 ptep = pmap_pte(pmap, addr);
2935 if ((pte = *ptep) != 0) {
2938 val = MINCORE_INCORE;
2939 if ((pte & VPTE_MANAGED) == 0)
2942 pa = pte & VPTE_FRAME;
2944 m = PHYS_TO_VM_PAGE(pa);
2950 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2952 * Modified by someone
2954 else if (m->dirty || pmap_is_modified(m))
2955 val |= MINCORE_MODIFIED_OTHER;
2960 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2963 * Referenced by someone
2965 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
2966 val |= MINCORE_REFERENCED_OTHER;
2967 vm_page_flag_set(m, PG_REFERENCED);
2974 pmap_activate(struct proc *p)
2978 pmap = vmspace_pmap(p->p_vmspace);
2980 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2982 pmap->pm_active |= 1;
2984 #if defined(SWTCH_OPTIM_STATS)
2988 KKASSERT((p == curproc));
2990 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
2991 load_cr3(curthread->td_pcb->pcb_cr3);
2996 pmap_deactivate(struct proc *p)
3000 pmap = vmspace_pmap(p->p_vmspace);
3002 atomic_clear_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3004 pmap->pm_active &= ~1;
3007 * XXX - note we do not adjust %cr3. The caller is expected to
3008 * activate a new pmap or do a thread-exit.
3013 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3016 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3020 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3027 static void pads (pmap_t pm);
3028 void pmap_pvdump (vm_paddr_t pa);
3030 /* print address space of pmap*/
3038 if (pm == &kernel_pmap)
3040 for (i = 0; i < 1024; i++)
3042 for (j = 0; j < 1024; j++) {
3043 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3044 if (pm == &kernel_pmap && va < KERNBASE)
3046 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3048 ptep = pmap_pte(pm, va);
3049 if (ptep && (*ptep & VPTE_V)) {
3051 (void *)va, (unsigned)*ptep);
3058 pmap_pvdump(vm_paddr_t pa)
3063 kprintf("pa %08llx", (long long)pa);
3064 m = PHYS_TO_VM_PAGE(pa);
3065 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3067 kprintf(" -> pmap %p, va %x, flags %x",
3068 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3070 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);