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 $
43 * NOTE: PMAP_INVAL_ADD: In pc32 this function is called prior to adjusting
44 * the PTE in the page table, because a cpu synchronization might be required.
45 * The actual invalidation is delayed until the following call or flush. In
46 * the VKERNEL build this function is called prior to adjusting the PTE and
47 * invalidates the table synchronously (not delayed), and is not SMP safe
51 #include <sys/types.h>
52 #include <sys/systm.h>
53 #include <sys/kernel.h>
56 #include <sys/vkernel.h>
58 #include <sys/thread.h>
60 #include <sys/vmspace.h>
63 #include <vm/vm_page.h>
64 #include <vm/vm_extern.h>
65 #include <vm/vm_kern.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_zone.h>
68 #include <vm/vm_pageout.h>
70 #include <machine/md_var.h>
71 #include <machine/pcb.h>
72 #include <machine/pmap_inval.h>
73 #include <machine/globaldata.h>
75 #include <sys/sysref2.h>
79 struct pmap kernel_pmap;
81 static struct vm_zone pvzone;
82 static struct vm_object pvzone_obj;
83 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
84 static int pv_entry_count;
85 static int pv_entry_max;
86 static int pv_entry_high_water;
87 static int pmap_pagedaemon_waken;
88 static boolean_t pmap_initialized = FALSE;
89 static int protection_codes[8];
91 static void i386_protection_init(void);
92 static void pmap_remove_all(vm_page_t m);
93 static int pmap_release_free_page(struct pmap *pmap, vm_page_t p);
96 #ifndef PMAP_SHPGPERPROC
97 #define PMAP_SHPGPERPROC 200
100 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
102 #define pte_prot(m, p) \
103 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
109 struct pv_entry *pvinit;
111 for (i = 0; i < vm_page_array_size; i++) {
114 m = &vm_page_array[i];
115 TAILQ_INIT(&m->md.pv_list);
116 m->md.pv_list_count = 0;
119 i = vm_page_array_size;
122 pvinit = (struct pv_entry *)kmem_alloc(&kernel_map, i*sizeof(*pvinit));
123 zbootinit(&pvzone, "PV ENTRY", sizeof(*pvinit), pvinit, i);
124 pmap_initialized = TRUE;
130 int shpgperproc = PMAP_SHPGPERPROC;
132 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
133 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
134 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
135 pv_entry_high_water = 9 * (pv_entry_max / 10);
136 zinitna(&pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
140 * Bootstrap the kernel_pmap so it can be used with pmap_enter().
142 * NOTE! pm_pdir for the kernel pmap is offset so VA's translate
143 * directly into PTD indexes (PTA is also offset for the same reason).
144 * This is necessary because, for now, KVA is not mapped at address 0.
146 * Page table pages are not managed like they are in normal pmaps, so
147 * no pteobj is needed.
152 vm_pindex_t i = (vm_offset_t)KernelPTD >> PAGE_SHIFT;
154 kernel_pmap.pm_pdir = KernelPTD - (KvaStart >> SEG_SHIFT);
155 kernel_pmap.pm_pdirpte = KernelPTA[i];
156 kernel_pmap.pm_count = 1;
157 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
158 TAILQ_INIT(&kernel_pmap.pm_pvlist);
159 i386_protection_init();
163 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
164 * just dummy it up so it works well enough for fork().
166 * In DragonFly, process pmaps may only be used to manipulate user address
167 * space, never kernel address space.
170 pmap_pinit0(struct pmap *pmap)
175 /************************************************************************
176 * Procedures to manage whole physical maps *
177 ************************************************************************
179 * Initialize a preallocated and zeroed pmap structure,
180 * such as one in a vmspace structure.
183 pmap_pinit(struct pmap *pmap)
189 * No need to allocate page table space yet but we do need a valid
190 * page directory table.
192 if (pmap->pm_pdir == NULL) {
194 (vpte_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
198 * allocate object for the pte array and page directory
200 npages = VPTE_PAGETABLE_SIZE +
201 (VM_MAX_USER_ADDRESS / PAGE_SIZE) * sizeof(vpte_t);
202 npages = (npages + PAGE_MASK) / PAGE_SIZE;
204 if (pmap->pm_pteobj == NULL)
205 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, npages);
206 pmap->pm_pdindex = npages - 1;
209 * allocate the page directory page
211 ptdpg = vm_page_grab(pmap->pm_pteobj, pmap->pm_pdindex,
212 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
214 ptdpg->wire_count = 1;
215 ++vmstats.v_wire_count;
217 /* not usually mapped */
218 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
219 ptdpg->valid = VM_PAGE_BITS_ALL;
221 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
222 pmap->pm_pdirpte = KernelPTA[(vm_offset_t)pmap->pm_pdir >> PAGE_SHIFT];
223 if ((ptdpg->flags & PG_ZERO) == 0)
224 bzero(pmap->pm_pdir, PAGE_SIZE);
228 pmap->pm_ptphint = NULL;
229 pmap->pm_cpucachemask = 0;
230 TAILQ_INIT(&pmap->pm_pvlist);
231 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
232 pmap->pm_stats.resident_count = 1;
236 * Clean up a pmap structure so it can be physically freed
239 pmap_puninit(pmap_t pmap)
242 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
243 pmap->pm_pdir = NULL;
245 if (pmap->pm_pteobj) {
246 vm_object_deallocate(pmap->pm_pteobj);
247 pmap->pm_pteobj = NULL;
253 * Wire in kernel global address entries. To avoid a race condition
254 * between pmap initialization and pmap_growkernel, this procedure
255 * adds the pmap to the master list (which growkernel scans to update),
256 * then copies the template.
258 * In a virtual kernel there are no kernel global address entries.
261 pmap_pinit2(struct pmap *pmap)
264 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
269 * Release all resources held by the given physical map.
271 * Should only be called if the map contains no valid mappings.
273 static int pmap_release_callback(struct vm_page *p, void *data);
276 pmap_release(struct pmap *pmap)
278 struct mdglobaldata *gd = mdcpu;
279 vm_object_t object = pmap->pm_pteobj;
280 struct rb_vm_page_scan_info info;
282 KKASSERT(pmap != &kernel_pmap);
284 #if defined(DIAGNOSTIC)
285 if (object->ref_count != 1)
286 panic("pmap_release: pteobj reference count != 1");
289 * Once we destroy the page table, the mapping becomes invalid.
290 * Don't waste time doing a madvise to invalidate the mapping, just
291 * set cpucachemask to 0.
293 if (pmap->pm_pdir == gd->gd_PT1pdir) {
294 gd->gd_PT1pdir = NULL;
296 /* madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL); */
298 if (pmap->pm_pdir == gd->gd_PT2pdir) {
299 gd->gd_PT2pdir = NULL;
301 /* madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL); */
303 if (pmap->pm_pdir == gd->gd_PT3pdir) {
304 gd->gd_PT3pdir = NULL;
306 /* madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL); */
310 info.object = object;
312 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
319 info.limit = object->generation;
321 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
322 pmap_release_callback, &info);
323 if (info.error == 0 && info.mpte) {
324 if (!pmap_release_free_page(pmap, info.mpte))
328 } while (info.error);
331 * Leave the KVA reservation for pm_pdir cached for later reuse.
333 pmap->pm_pdirpte = 0;
334 pmap->pm_cpucachemask = 0;
338 * Callback to release a page table page backing a directory
342 pmap_release_callback(struct vm_page *p, void *data)
344 struct rb_vm_page_scan_info *info = data;
346 if (p->pindex == info->pmap->pm_pdindex) {
350 if (!pmap_release_free_page(info->pmap, p)) {
354 if (info->object->generation != info->limit) {
362 * Retire the given physical map from service. Should only be called if
363 * the map contains no valid mappings.
366 pmap_destroy(pmap_t pmap)
373 count = --pmap->pm_count;
376 panic("destroying a pmap is not yet implemented");
381 * Add a reference to the specified pmap.
384 pmap_reference(pmap_t pmap)
391 /************************************************************************
392 * VMSPACE MANAGEMENT *
393 ************************************************************************
395 * The VMSPACE management we do in our virtual kernel must be reflected
396 * in the real kernel. This is accomplished by making vmspace system
397 * calls to the real kernel.
400 cpu_vmspace_alloc(struct vmspace *vm)
405 #define LAST_EXTENT (VM_MAX_USER_ADDRESS - 0x80000000)
407 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
408 panic("vmspace_create() failed");
410 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
411 PROT_READ|PROT_WRITE,
412 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
414 if (rp == MAP_FAILED)
415 panic("vmspace_mmap: failed1");
416 vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
418 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
419 PROT_READ|PROT_WRITE,
420 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
421 MemImageFd, 0x40000000);
422 if (rp == MAP_FAILED)
423 panic("vmspace_mmap: failed2");
424 vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
426 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
427 PROT_READ|PROT_WRITE,
428 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
429 MemImageFd, 0x80000000);
430 vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
432 if (rp == MAP_FAILED)
433 panic("vmspace_mmap: failed3");
435 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
436 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
438 panic("vmspace_mcontrol: failed1");
439 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
440 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
442 panic("vmspace_mcontrol: failed2");
443 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
444 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
446 panic("vmspace_mcontrol: failed3");
450 cpu_vmspace_free(struct vmspace *vm)
452 if (vmspace_destroy(&vm->vm_pmap) < 0)
453 panic("vmspace_destroy() failed");
456 /************************************************************************
457 * Procedures which operate directly on the kernel PMAP *
458 ************************************************************************/
461 * This maps the requested page table and gives us access to it.
463 * This routine can be called from a potentially preempting interrupt
464 * thread or from a normal thread.
467 get_ptbase(struct pmap *pmap, vm_offset_t va)
469 struct mdglobaldata *gd = mdcpu;
471 if (pmap == &kernel_pmap) {
472 KKASSERT(va >= KvaStart && va < KvaEnd);
473 return(KernelPTA + (va >> PAGE_SHIFT));
474 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
475 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
476 *gd->gd_PT1pde = pmap->pm_pdirpte;
477 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
478 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
480 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
481 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
482 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
483 *gd->gd_PT2pde = pmap->pm_pdirpte;
484 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
485 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
487 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
491 * If we aren't running from a potentially preempting interrupt,
492 * load a new page table directory into the page table cache
494 if (gd->mi.gd_intr_nesting_level == 0 &&
495 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0) {
497 * Choose one or the other and map the page table
498 * in the KVA space reserved for it.
500 if ((gd->gd_PTflip = 1 - gd->gd_PTflip) == 0) {
501 gd->gd_PT1pdir = pmap->pm_pdir;
502 *gd->gd_PT1pde = pmap->pm_pdirpte;
503 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
504 atomic_set_int(&pmap->pm_cpucachemask,
506 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
508 gd->gd_PT2pdir = pmap->pm_pdir;
509 *gd->gd_PT2pde = pmap->pm_pdirpte;
510 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
511 atomic_set_int(&pmap->pm_cpucachemask,
513 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
518 * If we are running from a preempting interrupt use a private
519 * map. The caller must be in a critical section.
521 KKASSERT(IN_CRITICAL_SECT(curthread));
522 if (pmap->pm_pdir == gd->gd_PT3pdir) {
523 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
524 *gd->gd_PT3pde = pmap->pm_pdirpte;
525 madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL);
526 atomic_set_int(&pmap->pm_cpucachemask,
530 gd->gd_PT3pdir = pmap->pm_pdir;
531 *gd->gd_PT3pde = pmap->pm_pdirpte;
532 madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL);
533 atomic_set_int(&pmap->pm_cpucachemask,
536 return(gd->gd_PT3map + (va >> PAGE_SHIFT));
540 get_ptbase1(struct pmap *pmap, vm_offset_t va)
542 struct mdglobaldata *gd = mdcpu;
544 if (pmap == &kernel_pmap) {
545 KKASSERT(va >= KvaStart && va < KvaEnd);
546 return(KernelPTA + (va >> PAGE_SHIFT));
547 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
548 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
549 *gd->gd_PT1pde = pmap->pm_pdirpte;
550 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
551 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
553 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
555 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
556 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
557 gd->gd_PT1pdir = pmap->pm_pdir;
558 *gd->gd_PT1pde = pmap->pm_pdirpte;
559 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
560 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
564 get_ptbase2(struct pmap *pmap, vm_offset_t va)
566 struct mdglobaldata *gd = mdcpu;
568 if (pmap == &kernel_pmap) {
569 KKASSERT(va >= KvaStart && va < KvaEnd);
570 return(KernelPTA + (va >> PAGE_SHIFT));
571 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
572 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
573 *gd->gd_PT2pde = pmap->pm_pdirpte;
574 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
575 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
577 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
579 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
580 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
581 gd->gd_PT2pdir = pmap->pm_pdir;
582 *gd->gd_PT2pde = pmap->pm_pdirpte;
583 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
584 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
588 * Return a pointer to the page table entry for the specified va in the
589 * specified pmap. NULL is returned if there is no valid page table page
592 static __inline vpte_t *
593 pmap_pte(struct pmap *pmap, vm_offset_t va)
597 ptep = &pmap->pm_pdir[va >> SEG_SHIFT];
601 return (get_ptbase(pmap, va));
607 * Enter a mapping into kernel_pmap. Mappings created in this fashion
608 * are not managed. Mappings must be immediately accessible on all cpus.
610 * Call pmap_inval_pte() to invalidate the virtual pte and clean out the
611 * real pmap and handle related races before storing the new vpte.
614 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
619 KKASSERT(va >= KvaStart && va < KvaEnd);
620 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
621 ptep = KernelPTA + (va >> PAGE_SHIFT);
623 pmap_inval_pte(ptep, &kernel_pmap, va);
628 * Synchronize a kvm mapping originally made for the private use on
629 * some other cpu so it can be used on all cpus.
631 * XXX add MADV_RESYNC to improve performance.
634 pmap_kenter_sync(vm_offset_t va)
636 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
640 * Synchronize a kvm mapping originally made for the private use on
641 * some other cpu so it can be used on our cpu. Turns out to be the
642 * same madvise() call, because we have to sync the real pmaps anyway.
644 * XXX add MADV_RESYNC to improve performance.
647 pmap_kenter_sync_quick(vm_offset_t va)
649 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
654 * Make a previously read-only kernel mapping R+W (not implemented by
658 pmap_kmodify_rw(vm_offset_t va)
660 *pmap_kpte(va) |= VPTE_R | VPTE_W;
661 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
665 * Make a kernel mapping non-cacheable (not applicable to virtual kernels)
668 pmap_kmodify_nc(vm_offset_t va)
670 *pmap_kpte(va) |= VPTE_N;
671 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
677 * Map a contiguous range of physical memory to a KVM
680 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
682 vm_offset_t sva, virt;
685 while (start < end) {
686 pmap_kenter(virt, start);
695 pmap_kpte(vm_offset_t va)
699 KKASSERT(va >= KvaStart && va < KvaEnd);
700 ptep = KernelPTA + (va >> PAGE_SHIFT);
705 * Enter an unmanaged KVA mapping for the private use of the current
706 * cpu only. pmap_kenter_sync() may be called to make the mapping usable
709 * It is illegal for the mapping to be accessed by other cpus unleess
710 * pmap_kenter_sync*() is called.
713 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
718 KKASSERT(va >= KvaStart && va < KvaEnd);
720 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
721 ptep = KernelPTA + (va >> PAGE_SHIFT);
723 pmap_inval_pte_quick(ptep, &kernel_pmap, va);
728 * Make a temporary mapping for a physical address. This is only intended
729 * to be used for panic dumps.
732 pmap_kenter_temporary(vm_paddr_t pa, int i)
734 pmap_kenter(crashdumpmap + (i * PAGE_SIZE), pa);
735 return ((void *)crashdumpmap);
739 * Remove an unmanaged mapping created with pmap_kenter*().
742 pmap_kremove(vm_offset_t va)
746 KKASSERT(va >= KvaStart && va < KvaEnd);
748 ptep = KernelPTA + (va >> PAGE_SHIFT);
750 pmap_inval_pte(ptep, &kernel_pmap, va);
755 * Remove an unmanaged mapping created with pmap_kenter*() but synchronize
756 * only with this cpu.
758 * Unfortunately because we optimize new entries by testing VPTE_V later
759 * on, we actually still have to synchronize with all the cpus. XXX maybe
760 * store a junk value and test against 0 in the other places instead?
763 pmap_kremove_quick(vm_offset_t va)
767 KKASSERT(va >= KvaStart && va < KvaEnd);
769 ptep = KernelPTA + (va >> PAGE_SHIFT);
771 pmap_inval_pte(ptep, &kernel_pmap, va); /* NOT _quick */
776 * Extract the physical address from the kernel_pmap that is associated
777 * with the specified virtual address.
780 pmap_kextract(vm_offset_t va)
785 KKASSERT(va >= KvaStart && va < KvaEnd);
787 ptep = KernelPTA + (va >> PAGE_SHIFT);
788 pa = (vm_paddr_t)(*ptep & VPTE_FRAME) | (va & PAGE_MASK);
793 * Map a set of unmanaged VM pages into KVM.
796 pmap_qenter(vm_offset_t va, struct vm_page **m, int count)
798 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
802 ptep = KernelPTA + (va >> PAGE_SHIFT);
804 pmap_inval_pte(ptep, &kernel_pmap, va);
805 *ptep = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
813 * Undo the effects of pmap_qenter*().
816 pmap_qremove(vm_offset_t va, int count)
818 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
822 ptep = KernelPTA + (va >> PAGE_SHIFT);
824 pmap_inval_pte(ptep, &kernel_pmap, va);
831 /************************************************************************
832 * Misc support glue called by machine independant code *
833 ************************************************************************
835 * These routines are called by machine independant code to operate on
836 * certain machine-dependant aspects of processes, threads, and pmaps.
840 * Initialize MD portions of the thread structure.
843 pmap_init_thread(thread_t td)
845 /* enforce pcb placement */
846 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
847 td->td_savefpu = &td->td_pcb->pcb_save;
848 td->td_sp = (char *)td->td_pcb - 16;
852 * This routine directly affects the fork perf for a process.
855 pmap_init_proc(struct proc *p)
860 * Destroy the UPAGES for a process that has exited and disassociate
861 * the process from its thread.
864 pmap_dispose_proc(struct proc *p)
866 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
870 * We pre-allocate all page table pages for kernel virtual memory so
871 * this routine will only be called if KVM has been exhausted.
874 pmap_growkernel(vm_offset_t addr)
876 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
878 if (addr > virtual_end - SEG_SIZE)
879 panic("KVM exhausted");
880 kernel_vm_end = addr;
884 * The modification bit is not tracked for any pages in this range. XXX
885 * such pages in this maps should always use pmap_k*() functions and not
888 * XXX User and kernel address spaces are independant for virtual kernels,
889 * this function only applies to the kernel pmap.
892 pmap_track_modified(pmap_t pmap, vm_offset_t va)
894 if (pmap != &kernel_pmap)
896 if ((va < clean_sva) || (va >= clean_eva))
902 /************************************************************************
903 * Procedures supporting managed page table pages *
904 ************************************************************************
906 * These procedures are used to track managed page table pages. These pages
907 * use the page table page's vm_page_t to track PTEs in the page. The
908 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
910 * This allows the system to throw away page table pages for user processes
911 * at will and reinstantiate them on demand.
915 * This routine works like vm_page_lookup() but also blocks as long as the
916 * page is busy. This routine does not busy the page it returns.
918 * Unless the caller is managing objects whos pages are in a known state,
919 * the call should be made with a critical section held so the page's object
920 * association remains valid on return.
923 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
928 m = vm_page_lookup(object, pindex);
929 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
935 * This routine unholds page table pages, and if the hold count
936 * drops to zero, then it decrements the wire count.
938 * We must recheck that this is the last hold reference after busy-sleeping
942 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
944 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
946 KASSERT(m->queue == PQ_NONE,
947 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
949 if (m->hold_count == 1) {
951 * Unmap the page table page.
954 KKASSERT(pmap->pm_pdir[m->pindex] != 0);
955 pmap_inval_pde(&pmap->pm_pdir[m->pindex], pmap,
956 (vm_offset_t)m->pindex << SEG_SHIFT);
957 KKASSERT(pmap->pm_stats.resident_count > 0);
958 --pmap->pm_stats.resident_count;
960 if (pmap->pm_ptphint == m)
961 pmap->pm_ptphint = NULL;
964 * This was our last hold, the page had better be unwired
965 * after we decrement wire_count.
967 * FUTURE NOTE: shared page directory page could result in
968 * multiple wire counts.
972 KKASSERT(m->wire_count == 0);
973 --vmstats.v_wire_count;
974 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
976 vm_page_free_zero(m);
979 KKASSERT(m->hold_count > 1);
985 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
987 KKASSERT(m->hold_count > 0);
988 if (m->hold_count > 1) {
992 return _pmap_unwire_pte_hold(pmap, m);
997 * After removing a page table entry, this routine is used to
998 * conditionally free the page, and manage the hold/wire counts.
1001 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
1007 * page table pages in the kernel_pmap are not managed.
1009 if (pmap == &kernel_pmap)
1011 ptepindex = (va >> PDRSHIFT);
1012 if (pmap->pm_ptphint &&
1013 (pmap->pm_ptphint->pindex == ptepindex)) {
1014 mpte = pmap->pm_ptphint;
1016 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1017 pmap->pm_ptphint = mpte;
1020 return pmap_unwire_pte_hold(pmap, mpte);
1024 * Attempt to release and free the vm_page backing a page directory page
1025 * in a pmap. Returns 1 on success, 0 on failure (if the procedure had
1029 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1031 vpte_t *pde = pmap->pm_pdir;
1034 * This code optimizes the case of freeing non-busy
1035 * page-table pages. Those pages are zero now, and
1036 * might as well be placed directly into the zero queue.
1038 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1042 KKASSERT(pmap->pm_stats.resident_count > 0);
1043 --pmap->pm_stats.resident_count;
1045 if (p->hold_count) {
1046 panic("pmap_release: freeing held page table page");
1049 * Page directory pages need to have the kernel stuff cleared, so
1050 * they can go into the zero queue also.
1052 * In virtual kernels there is no 'kernel stuff'. For the moment
1053 * I just make sure the whole thing has been zero'd even though
1054 * it should already be completely zero'd.
1056 * pmaps for vkernels do not self-map because they do not share
1057 * their address space with the vkernel. Clearing of pde[] thus
1058 * only applies to page table pages and not to the page directory
1061 if (p->pindex == pmap->pm_pdindex) {
1062 bzero(pde, VPTE_PAGETABLE_SIZE);
1063 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1065 KKASSERT(pde[p->pindex] != 0);
1066 pmap_inval_pde(&pde[p->pindex], pmap,
1067 (vm_offset_t)p->pindex << SEG_SHIFT);
1071 * Clear the matching hint
1073 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1074 pmap->pm_ptphint = NULL;
1077 * And throw the page away. The page is completely zero'd out so
1078 * optimize the free call.
1081 vmstats.v_wire_count--;
1082 vm_page_free_zero(p);
1087 * This routine is called if the page table page is not mapped in the page
1090 * The routine is broken up into two parts for readability.
1092 * It must return a held mpte and map the page directory page as required.
1093 * Because vm_page_grab() can block, we must re-check pm_pdir[ptepindex]
1096 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1102 * Find or fabricate a new pagetable page. A busied page will be
1103 * returned. This call may block.
1105 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1106 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1108 KASSERT(m->queue == PQ_NONE,
1109 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1112 * Increment the hold count for the page we will be returning to
1118 * It is possible that someone else got in and mapped by the page
1119 * directory page while we were blocked, if so just unbusy and
1120 * return the held page.
1122 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1123 KKASSERT((ptepa & VPTE_FRAME) == VM_PAGE_TO_PHYS(m));
1128 if (m->wire_count == 0)
1129 vmstats.v_wire_count++;
1133 * Map the pagetable page into the process address space, if
1134 * it isn't already there.
1136 ++pmap->pm_stats.resident_count;
1138 ptepa = VM_PAGE_TO_PHYS(m);
1139 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
1143 * We are likely about to access this page table page, so set the
1144 * page table hint to reduce overhead.
1146 pmap->pm_ptphint = m;
1149 * Try to use the new mapping, but if we cannot, then
1150 * do it with the routine that maps the page explicitly.
1152 if ((m->flags & PG_ZERO) == 0)
1153 pmap_zero_page(ptepa);
1155 m->valid = VM_PAGE_BITS_ALL;
1156 vm_page_flag_clear(m, PG_ZERO);
1157 vm_page_flag_set(m, PG_MAPPED);
1164 * Determine the page table page required to access the VA in the pmap
1165 * and allocate it if necessary. Return a held vm_page_t for the page.
1167 * Only used with user pmaps.
1170 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1177 * Calculate pagetable page index
1179 ptepindex = va >> PDRSHIFT;
1182 * Get the page directory entry
1184 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1187 * This supports switching from a 4MB page to a
1190 if (ptepa & VPTE_PS) {
1191 KKASSERT(pmap->pm_pdir[ptepindex] != 0);
1192 pmap_inval_pde(&pmap->pm_pdir[ptepindex], pmap,
1193 (vm_offset_t)ptepindex << SEG_SHIFT);
1198 * If the page table page is mapped, we just increment the
1199 * hold count, and activate it.
1203 * In order to get the page table page, try the
1206 if (pmap->pm_ptphint &&
1207 (pmap->pm_ptphint->pindex == ptepindex)) {
1208 m = pmap->pm_ptphint;
1210 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1211 pmap->pm_ptphint = m;
1217 * Here if the pte page isn't mapped, or if it has been deallocated.
1219 return _pmap_allocpte(pmap, ptepindex);
1222 /************************************************************************
1223 * Managed pages in pmaps *
1224 ************************************************************************
1226 * All pages entered into user pmaps and some pages entered into the kernel
1227 * pmap are managed, meaning that pmap_protect() and other related management
1228 * functions work on these pages.
1232 * free the pv_entry back to the free list. This function may be
1233 * called from an interrupt.
1235 static __inline void
1236 free_pv_entry(pv_entry_t pv)
1243 * get a new pv_entry, allocating a block from the system
1244 * when needed. This function may be called from an interrupt.
1250 if (pv_entry_high_water &&
1251 (pv_entry_count > pv_entry_high_water) &&
1252 (pmap_pagedaemon_waken == 0)) {
1253 pmap_pagedaemon_waken = 1;
1254 wakeup (&vm_pages_needed);
1256 return zalloc(&pvzone);
1260 * This routine is very drastic, but can save the system
1268 static int warningdone=0;
1270 if (pmap_pagedaemon_waken == 0)
1272 pmap_pagedaemon_waken = 0;
1274 if (warningdone < 5) {
1275 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1279 for(i = 0; i < vm_page_array_size; i++) {
1280 m = &vm_page_array[i];
1281 if (m->wire_count || m->hold_count || m->busy ||
1282 (m->flags & PG_BUSY))
1289 * If it is the first entry on the list, it is actually
1290 * in the header and we must copy the following entry up
1291 * to the header. Otherwise we must search the list for
1292 * the entry. In either case we free the now unused entry.
1295 pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va)
1301 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1302 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1303 if (pmap == pv->pv_pmap && va == pv->pv_va)
1307 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1308 if (va == pv->pv_va)
1314 * Note that pv_ptem is NULL if the page table page itself is not
1315 * managed, even if the page being removed IS managed.
1319 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1320 m->md.pv_list_count--;
1321 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1322 if (TAILQ_EMPTY(&m->md.pv_list))
1323 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1324 ++pmap->pm_generation;
1325 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1333 * Create a pv entry for page at pa for (pmap, va). If the page table page
1334 * holding the VA is managed, mpte will be non-NULL.
1337 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1342 pv = get_pv_entry();
1347 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1348 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1349 ++pmap->pm_generation;
1350 m->md.pv_list_count++;
1356 * pmap_remove_pte: do the things to unmap a page in a process
1359 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va)
1364 oldpte = pmap_inval_loadandclear(ptq, pmap, va);
1365 if (oldpte & VPTE_WIRED)
1366 --pmap->pm_stats.wired_count;
1367 KKASSERT(pmap->pm_stats.wired_count >= 0);
1371 * Machines that don't support invlpg, also don't support
1372 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1375 if (oldpte & VPTE_G)
1376 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1378 KKASSERT(pmap->pm_stats.resident_count > 0);
1379 --pmap->pm_stats.resident_count;
1380 if (oldpte & VPTE_MANAGED) {
1381 m = PHYS_TO_VM_PAGE(oldpte);
1382 if (oldpte & VPTE_M) {
1383 #if defined(PMAP_DIAGNOSTIC)
1384 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1386 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1390 if (pmap_track_modified(pmap, va))
1393 if (oldpte & VPTE_A)
1394 vm_page_flag_set(m, PG_REFERENCED);
1395 return pmap_remove_entry(pmap, m, va);
1397 return pmap_unuse_pt(pmap, va, NULL);
1406 * Remove a single page from a process address space.
1408 * This function may not be called from an interrupt if the pmap is
1412 pmap_remove_page(struct pmap *pmap, vm_offset_t va)
1417 * if there is no pte for this address, just skip it!!! Otherwise
1418 * get a local va for mappings for this pmap and remove the entry.
1420 if (*pmap_pde(pmap, va) != 0) {
1421 ptq = get_ptbase(pmap, va);
1423 pmap_remove_pte(pmap, ptq, va);
1431 * Remove the given range of addresses from the specified map.
1433 * It is assumed that the start and end are properly
1434 * rounded to the page size.
1436 * This function may not be called from an interrupt if the pmap is
1440 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1444 vm_offset_t ptpaddr;
1445 vm_pindex_t sindex, eindex;
1450 KKASSERT(pmap->pm_stats.resident_count >= 0);
1451 if (pmap->pm_stats.resident_count == 0)
1455 * special handling of removing one page. a very
1456 * common operation and easy to short circuit some
1459 if (((sva + PAGE_SIZE) == eva) &&
1460 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1461 pmap_remove_page(pmap, sva);
1466 * Get a local virtual address for the mappings that are being
1469 * XXX this is really messy because the kernel pmap is not relative
1472 sindex = (sva >> PAGE_SHIFT);
1473 eindex = (eva >> PAGE_SHIFT);
1475 for (; sindex < eindex; sindex = pdnxt) {
1479 * Calculate index for next page table.
1481 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1482 if (pmap->pm_stats.resident_count == 0)
1485 pdirindex = sindex / NPDEPG;
1486 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1487 KKASSERT(pmap->pm_pdir[pdirindex] != 0);
1488 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1489 pmap_inval_pde(&pmap->pm_pdir[pdirindex], pmap,
1490 (vm_offset_t)pdirindex << SEG_SHIFT);
1495 * Weed out invalid mappings. Note: we assume that the page
1496 * directory table is always allocated, and in kernel virtual.
1502 * Limit our scan to either the end of the va represented
1503 * by the current page table page, or to the end of the
1504 * range being removed.
1510 * NOTE: pmap_remove_pte() can block.
1512 for (; sindex != pdnxt; sindex++) {
1515 ptbase = get_ptbase(pmap, sindex << PAGE_SHIFT);
1518 va = i386_ptob(sindex);
1519 if (pmap_remove_pte(pmap, ptbase, va))
1528 * Removes this physical page from all physical maps in which it resides.
1529 * Reflects back modify bits to the pager.
1531 * This routine may not be called from an interrupt.
1534 pmap_remove_all(vm_page_t m)
1539 #if defined(PMAP_DIAGNOSTIC)
1541 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1544 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1545 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1550 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1551 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1552 --pv->pv_pmap->pm_stats.resident_count;
1554 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1555 KKASSERT(pte != NULL);
1557 tpte = pmap_inval_loadandclear(pte, pv->pv_pmap, pv->pv_va);
1558 if (tpte & VPTE_WIRED)
1559 --pv->pv_pmap->pm_stats.wired_count;
1560 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
1563 vm_page_flag_set(m, PG_REFERENCED);
1566 * Update the vm_page_t clean and reference bits.
1568 if (tpte & VPTE_M) {
1569 #if defined(PMAP_DIAGNOSTIC)
1570 if (pmap_nw_modified((pt_entry_t) tpte)) {
1572 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1576 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
1579 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1580 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1581 ++pv->pv_pmap->pm_generation;
1582 m->md.pv_list_count--;
1583 if (TAILQ_EMPTY(&m->md.pv_list))
1584 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1585 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
1588 KKASSERT((m->flags & (PG_MAPPED | PG_WRITEABLE)) == 0);
1595 * Set the physical protection on the specified range of this map
1598 * This function may not be called from an interrupt if the map is
1599 * not the kernel_pmap.
1602 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1606 vm_offset_t pdnxt, ptpaddr;
1607 vm_pindex_t sindex, eindex;
1613 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1614 pmap_remove(pmap, sva, eva);
1618 if (prot & VM_PROT_WRITE)
1621 ptbase = get_ptbase(pmap, sva);
1623 sindex = (sva >> PAGE_SHIFT);
1624 eindex = (eva >> PAGE_SHIFT);
1627 for (; sindex < eindex; sindex = pdnxt) {
1631 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1633 pdirindex = sindex / NPDEPG;
1636 * Clear the modified and writable bits for a 4m page.
1637 * Throw away the modified bit (?)
1639 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1640 pmap_clean_pde(&pmap->pm_pdir[pdirindex], pmap,
1641 (vm_offset_t)pdirindex << SEG_SHIFT);
1642 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1647 * Weed out invalid mappings. Note: we assume that the page
1648 * directory table is always allocated, and in kernel virtual.
1653 if (pdnxt > eindex) {
1657 for (; sindex != pdnxt; sindex++) {
1662 * Clean managed pages and also check the accessed
1663 * bit. Just remove write perms for unmanaged
1664 * pages. Be careful of races, turning off write
1665 * access will force a fault rather then setting
1666 * the modified bit at an unexpected time.
1668 ptep = &ptbase[sindex - sbase];
1669 if (*ptep & VPTE_MANAGED) {
1670 pbits = pmap_clean_pte(ptep, pmap,
1673 if (pbits & VPTE_A) {
1674 m = PHYS_TO_VM_PAGE(pbits);
1675 vm_page_flag_set(m, PG_REFERENCED);
1676 atomic_clear_long(ptep, VPTE_A);
1678 if (pbits & VPTE_M) {
1679 if (pmap_track_modified(pmap, i386_ptob(sindex))) {
1681 m = PHYS_TO_VM_PAGE(pbits);
1686 pbits = pmap_setro_pte(ptep, pmap,
1694 * Enter a managed page into a pmap. If the page is not wired related pmap
1695 * data can be destroyed at any time for later demand-operation.
1697 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1698 * specified protection, and wire the mapping if requested.
1700 * NOTE: This routine may not lazy-evaluate or lose information. The
1701 * page must actually be inserted into the given map NOW.
1703 * NOTE: When entering a page at a KVA address, the pmap must be the
1707 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1713 vpte_t origpte, newpte;
1722 * Get the page table page. The kernel_pmap's page table pages
1723 * are preallocated and have no associated vm_page_t.
1725 if (pmap == &kernel_pmap)
1728 mpte = pmap_allocpte(pmap, va);
1730 pte = pmap_pte(pmap, va);
1733 * Page Directory table entry not valid, we need a new PT page
1734 * and pmap_allocpte() didn't give us one. Oops!
1737 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1742 * Deal with races on the original mapping (though don't worry
1743 * about VPTE_A races) by cleaning it. This will force a fault
1744 * if an attempt is made to write to the page.
1746 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1747 origpte = pmap_clean_pte(pte, pmap, va);
1748 opa = origpte & VPTE_FRAME;
1750 if (origpte & VPTE_PS)
1751 panic("pmap_enter: attempted pmap_enter on 4MB page");
1754 * Mapping has not changed, must be protection or wiring change.
1756 if (origpte && (opa == pa)) {
1758 * Wiring change, just update stats. We don't worry about
1759 * wiring PT pages as they remain resident as long as there
1760 * are valid mappings in them. Hence, if a user page is wired,
1761 * the PT page will be also.
1763 if (wired && ((origpte & VPTE_WIRED) == 0))
1764 ++pmap->pm_stats.wired_count;
1765 else if (!wired && (origpte & VPTE_WIRED))
1766 --pmap->pm_stats.wired_count;
1767 KKASSERT(pmap->pm_stats.wired_count >= 0);
1770 * Remove the extra pte reference. Note that we cannot
1771 * optimize the RO->RW case because we have adjusted the
1772 * wiring count above and may need to adjust the wiring
1779 * We might be turning off write access to the page,
1780 * so we go ahead and sense modify status.
1782 if (origpte & VPTE_MANAGED) {
1783 if ((origpte & VPTE_M) &&
1784 pmap_track_modified(pmap, va)) {
1786 om = PHYS_TO_VM_PAGE(opa);
1790 KKASSERT(m->flags & PG_MAPPED);
1795 * Mapping has changed, invalidate old range and fall through to
1796 * handle validating new mapping.
1800 err = pmap_remove_pte(pmap, pte, va);
1802 panic("pmap_enter: pte vanished, va: %p", (void *)va);
1803 pte = pmap_pte(pmap, va);
1804 origpte = pmap_clean_pte(pte, pmap, va);
1805 opa = origpte & VPTE_FRAME;
1807 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
1813 * Enter on the PV list if part of our managed memory. Note that we
1814 * raise IPL while manipulating pv_table since pmap_enter can be
1815 * called at interrupt time.
1817 if (pmap_initialized &&
1818 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1819 pmap_insert_entry(pmap, va, mpte, m);
1821 vm_page_flag_set(m, PG_MAPPED);
1825 * Increment counters
1827 ++pmap->pm_stats.resident_count;
1829 pmap->pm_stats.wired_count++;
1833 * Now validate mapping with desired protection/wiring.
1835 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1838 newpte |= VPTE_WIRED;
1839 if (pmap != &kernel_pmap)
1843 * If the mapping or permission bits are different from the
1844 * (now cleaned) original pte, an update is needed. We've
1845 * already downgraded or invalidated the page so all we have
1846 * to do now is update the bits.
1848 * XXX should we synchronize RO->RW changes to avoid another
1851 if ((origpte & ~(VPTE_W|VPTE_M|VPTE_A)) != newpte) {
1852 *pte = newpte | VPTE_A;
1853 if (newpte & VPTE_W)
1854 vm_page_flag_set(m, PG_WRITEABLE);
1856 KKASSERT((newpte & VPTE_MANAGED) == 0 || m->flags & PG_MAPPED);
1860 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
1862 * Currently this routine may only be used on user pmaps, not kernel_pmap.
1865 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
1873 KKASSERT(pmap != &kernel_pmap);
1875 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1878 * Calculate pagetable page (mpte), allocating it if necessary.
1880 * A held page table page (mpte), or NULL, is passed onto the
1881 * section following.
1883 ptepindex = va >> PDRSHIFT;
1887 * Get the page directory entry
1889 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1892 * If the page table page is mapped, we just increment
1893 * the hold count, and activate it.
1896 if (ptepa & VPTE_PS)
1897 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1898 if (pmap->pm_ptphint &&
1899 (pmap->pm_ptphint->pindex == ptepindex)) {
1900 mpte = pmap->pm_ptphint;
1902 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1903 pmap->pm_ptphint = mpte;
1908 mpte = _pmap_allocpte(pmap, ptepindex);
1910 } while (mpte == NULL);
1913 * Ok, now that the page table page has been validated, get the pte.
1914 * If the pte is already mapped undo mpte's hold_count and
1917 pte = pmap_pte(pmap, va);
1919 pmap_unwire_pte_hold(pmap, mpte);
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);
1930 vm_page_flag_set(m, PG_MAPPED);
1934 * Increment counters
1936 ++pmap->pm_stats.resident_count;
1938 pa = VM_PAGE_TO_PHYS(m);
1941 * Now validate mapping with RO protection
1943 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1944 *pte = (vpte_t)pa | VPTE_V | VPTE_U;
1946 *pte = (vpte_t)pa | VPTE_V | VPTE_U | VPTE_MANAGED;
1947 /*pmap_inval_add(&info, pmap, va); shouldn't be needed 0->valid */
1948 /*pmap_inval_flush(&info); don't need for vkernel */
1952 * Extract the physical address for the translation at the specified
1953 * virtual address in the pmap.
1956 pmap_extract(pmap_t pmap, vm_offset_t va)
1961 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
1962 if (pte & VPTE_PS) {
1963 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
1964 rtval |= va & SEG_MASK;
1966 pte = *get_ptbase(pmap, va);
1967 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
1974 #define MAX_INIT_PT (96)
1977 * This routine preloads the ptes for a given object into the specified pmap.
1978 * This eliminates the blast of soft faults on process startup and
1979 * immediately after an mmap.
1981 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
1984 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
1985 vm_object_t object, vm_pindex_t pindex,
1986 vm_size_t size, int limit)
1988 struct rb_vm_page_scan_info info;
1993 * We can't preinit if read access isn't set or there is no pmap
1996 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2000 * We can't preinit if the pmap is not the current pmap
2002 lp = curthread->td_lwp;
2003 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2006 psize = size >> PAGE_SHIFT;
2008 if ((object->type != OBJT_VNODE) ||
2009 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2010 (object->resident_page_count > MAX_INIT_PT))) {
2014 if (psize + pindex > object->size) {
2015 if (object->size < pindex)
2017 psize = object->size - pindex;
2024 * Use a red-black scan to traverse the requested range and load
2025 * any valid pages found into the pmap.
2027 * We cannot safely scan the object's memq unless we are in a
2028 * critical section since interrupts can remove pages from objects.
2030 info.start_pindex = pindex;
2031 info.end_pindex = pindex + psize - 1;
2038 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2039 pmap_object_init_pt_callback, &info);
2045 pmap_object_init_pt_callback(vm_page_t p, void *data)
2047 struct rb_vm_page_scan_info *info = data;
2048 vm_pindex_t rel_index;
2050 * don't allow an madvise to blow away our really
2051 * free pages allocating pv entries.
2053 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2054 vmstats.v_free_count < vmstats.v_free_reserved) {
2057 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2058 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2059 if ((p->queue - p->pc) == PQ_CACHE)
2060 vm_page_deactivate(p);
2062 rel_index = p->pindex - info->start_pindex;
2063 pmap_enter_quick(info->pmap,
2064 info->addr + i386_ptob(rel_index), p);
2071 * Return TRUE if the pmap is in shape to trivially
2072 * pre-fault the specified address.
2074 * Returns FALSE if it would be non-trivial or if a
2075 * pte is already loaded into the slot.
2078 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2082 if ((*pmap_pde(pmap, addr)) == 0)
2085 pte = get_ptbase(pmap, addr);
2093 * Routine: pmap_change_wiring
2094 * Function: Change the wiring attribute for a map/virtual-address
2096 * In/out conditions:
2097 * The mapping must already exist in the pmap.
2100 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2107 pte = get_ptbase(pmap, va);
2109 if (wired && (*pte & VPTE_WIRED) == 0)
2110 ++pmap->pm_stats.wired_count;
2111 else if (!wired && (*pte & VPTE_WIRED))
2112 --pmap->pm_stats.wired_count;
2113 KKASSERT(pmap->pm_stats.wired_count >= 0);
2116 * Wiring is not a hardware characteristic so there is no need to
2117 * invalidate TLB. However, in an SMP environment we must use
2118 * a locked bus cycle to update the pte (if we are not using
2119 * the pmap_inval_*() API that is)... it's ok to do this for simple
2123 atomic_set_long(pte, VPTE_WIRED);
2125 atomic_clear_long(pte, VPTE_WIRED);
2129 * Copy the range specified by src_addr/len
2130 * from the source map to the range dst_addr/len
2131 * in the destination map.
2133 * This routine is only advisory and need not do anything.
2136 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2137 vm_size_t len, vm_offset_t src_addr)
2140 vm_offset_t end_addr = src_addr + len;
2147 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2148 * valid through blocking calls, and that's just not going to
2155 if (dst_addr != src_addr)
2157 if (dst_pmap->pm_pdir == NULL)
2159 if (src_pmap->pm_pdir == NULL)
2164 src_frame = get_ptbase1(src_pmap, src_addr);
2165 dst_frame = get_ptbase2(dst_pmap, src_addr);
2168 * critical section protection is required to maintain the page/object
2169 * association, interrupts can free pages and remove them from
2172 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2173 vpte_t *src_pte, *dst_pte;
2174 vm_page_t dstmpte, srcmpte;
2175 vm_offset_t srcptepaddr;
2178 if (addr >= VM_MAX_USER_ADDRESS)
2179 panic("pmap_copy: invalid to pmap_copy page tables\n");
2182 * Don't let optional prefaulting of pages make us go
2183 * way below the low water mark of free pages or way
2184 * above high water mark of used pv entries.
2186 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2187 pv_entry_count > pv_entry_high_water)
2190 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2191 ptepindex = addr >> PDRSHIFT;
2193 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2194 if (srcptepaddr == 0)
2197 if (srcptepaddr & VPTE_PS) {
2198 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2199 dst_pmap->pm_pdir[ptepindex] = (vpte_t)srcptepaddr;
2200 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2205 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2206 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2207 (srcmpte->flags & PG_BUSY)) {
2211 if (pdnxt > end_addr)
2214 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2215 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2216 while (addr < pdnxt) {
2221 * we only virtual copy managed pages
2223 if ((ptetemp & VPTE_MANAGED) != 0) {
2225 * We have to check after allocpte for the
2226 * pte still being around... allocpte can
2229 * pmap_allocpte can block, unfortunately
2230 * we have to reload the tables.
2232 dstmpte = pmap_allocpte(dst_pmap, addr);
2233 src_frame = get_ptbase1(src_pmap, src_addr);
2234 dst_frame = get_ptbase2(dst_pmap, src_addr);
2236 if ((*dst_pte == 0) && (ptetemp = *src_pte) &&
2237 (ptetemp & VPTE_MANAGED) != 0) {
2239 * Clear the modified and accessed
2240 * (referenced) bits during the copy.
2242 * We do not have to clear the write
2243 * bit to force a fault-on-modify
2244 * because the real kernel's target
2245 * pmap is empty and will fault anyway.
2247 m = PHYS_TO_VM_PAGE(ptetemp);
2248 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2249 ++dst_pmap->pm_stats.resident_count;
2250 pmap_insert_entry(dst_pmap, addr,
2252 KKASSERT(m->flags & PG_MAPPED);
2254 pmap_unwire_pte_hold(dst_pmap, dstmpte);
2256 if (dstmpte->hold_count >= srcmpte->hold_count)
2270 * Zero the specified PA by mapping the page into KVM and clearing its
2273 * This function may be called from an interrupt and no locking is
2277 pmap_zero_page(vm_paddr_t phys)
2279 struct mdglobaldata *gd = mdcpu;
2283 panic("pmap_zero_page: CMAP3 busy");
2284 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2285 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2287 bzero(gd->gd_CADDR3, PAGE_SIZE);
2293 * pmap_page_assertzero:
2295 * Assert that a page is empty, panic if it isn't.
2298 pmap_page_assertzero(vm_paddr_t phys)
2300 struct mdglobaldata *gd = mdcpu;
2305 panic("pmap_zero_page: CMAP3 busy");
2306 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2307 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2308 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2309 for (i = 0; i < PAGE_SIZE; i += 4) {
2310 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2311 panic("pmap_page_assertzero() @ %p not zero!\n",
2312 (void *)gd->gd_CADDR3);
2322 * Zero part of a physical page by mapping it into memory and clearing
2323 * its contents with bzero.
2325 * off and size may not cover an area beyond a single hardware page.
2328 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2330 struct mdglobaldata *gd = mdcpu;
2334 panic("pmap_zero_page: CMAP3 busy");
2335 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2336 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2337 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2339 bzero((char *)gd->gd_CADDR3 + off, size);
2347 * Copy the physical page from the source PA to the target PA.
2348 * This function may be called from an interrupt. No locking
2352 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2354 struct mdglobaldata *gd = mdcpu;
2357 if (*(int *) gd->gd_CMAP1)
2358 panic("pmap_copy_page: CMAP1 busy");
2359 if (*(int *) gd->gd_CMAP2)
2360 panic("pmap_copy_page: CMAP2 busy");
2362 *(int *) gd->gd_CMAP1 = VPTE_V | VPTE_R | (src & PG_FRAME) | VPTE_A;
2363 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2365 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2366 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2368 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2370 *(int *) gd->gd_CMAP1 = 0;
2371 *(int *) gd->gd_CMAP2 = 0;
2376 * pmap_copy_page_frag:
2378 * Copy the physical page from the source PA to the target PA.
2379 * This function may be called from an interrupt. No locking
2383 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2385 struct mdglobaldata *gd = mdcpu;
2388 if (*(int *) gd->gd_CMAP1)
2389 panic("pmap_copy_page: CMAP1 busy");
2390 if (*(int *) gd->gd_CMAP2)
2391 panic("pmap_copy_page: CMAP2 busy");
2393 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2394 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2396 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2397 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2399 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2400 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2403 *(int *) gd->gd_CMAP1 = 0;
2404 *(int *) gd->gd_CMAP2 = 0;
2409 * Returns true if the pmap's pv is one of the first
2410 * 16 pvs linked to from this page. This count may
2411 * be changed upwards or downwards in the future; it
2412 * is only necessary that true be returned for a small
2413 * subset of pmaps for proper page aging.
2416 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2421 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2426 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2427 if (pv->pv_pmap == pmap) {
2440 * Remove all pages from specified address space
2441 * this aids process exit speeds. Also, this code
2442 * is special cased for current process only, but
2443 * can have the more generic (and slightly slower)
2444 * mode enabled. This is much faster than pmap_remove
2445 * in the case of running down an entire address space.
2448 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2453 int32_t save_generation;
2456 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2457 if (pv->pv_va >= eva || pv->pv_va < sva) {
2458 npv = TAILQ_NEXT(pv, pv_plist);
2462 KKASSERT(pmap == pv->pv_pmap);
2464 pte = pmap_pte(pmap, pv->pv_va);
2467 * We cannot remove wired pages from a process' mapping
2470 if (*pte & VPTE_WIRED) {
2471 npv = TAILQ_NEXT(pv, pv_plist);
2474 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2476 m = PHYS_TO_VM_PAGE(tpte);
2478 KASSERT(m < &vm_page_array[vm_page_array_size],
2479 ("pmap_remove_pages: bad tpte %lx", tpte));
2481 KKASSERT(pmap->pm_stats.resident_count > 0);
2482 --pmap->pm_stats.resident_count;
2485 * Update the vm_page_t clean and reference bits.
2487 if (tpte & VPTE_M) {
2491 npv = TAILQ_NEXT(pv, pv_plist);
2492 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2493 save_generation = ++pmap->pm_generation;
2495 m->md.pv_list_count--;
2496 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2497 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2498 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2500 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2504 * Restart the scan if we blocked during the unuse or free
2505 * calls and other removals were made.
2507 if (save_generation != pmap->pm_generation) {
2508 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2509 npv = TAILQ_FIRST(&pmap->pm_pvlist);
2516 * pmap_testbit tests bits in active mappings of a VM page.
2519 pmap_testbit(vm_page_t m, int bit)
2524 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2527 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2532 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2534 * if the bit being tested is the modified bit, then
2535 * mark clean_map and ptes as never
2538 if (bit & (VPTE_A|VPTE_M)) {
2539 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2543 #if defined(PMAP_DIAGNOSTIC)
2545 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2549 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2560 * This routine is used to clear bits in ptes. Certain bits require special
2561 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2563 * This routine is only called with certain VPTE_* bit combinations.
2565 static __inline void
2566 pmap_clearbit(vm_page_t m, int bit)
2572 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2578 * Loop over all current mappings setting/clearing as appropos If
2579 * setting RO do we need to clear the VAC?
2581 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2583 * don't write protect pager mappings
2585 if (bit == VPTE_W) {
2586 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2590 #if defined(PMAP_DIAGNOSTIC)
2592 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2598 * Careful here. We can use a locked bus instruction to
2599 * clear VPTE_A or VPTE_M safely but we need to synchronize
2600 * with the target cpus when we mess with VPTE_W.
2602 * On virtual kernels we must force a new fault-on-write
2603 * in the real kernel if we clear the Modify bit ourselves,
2604 * otherwise the real kernel will not get a new fault and
2605 * will never set our Modify bit again.
2607 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2609 if (bit == VPTE_W) {
2611 * We must also clear VPTE_M when clearing
2614 pbits = pmap_clean_pte(pte, pv->pv_pmap,
2618 } else if (bit == VPTE_M) {
2620 * We do not have to make the page read-only
2621 * when clearing the Modify bit. The real
2622 * kernel will make the real PTE read-only
2623 * or otherwise detect the write and set
2624 * our VPTE_M again simply by us invalidating
2625 * the real kernel VA for the pmap (as we did
2626 * above). This allows the real kernel to
2627 * handle the write fault without forwarding
2630 atomic_clear_long(pte, VPTE_M);
2631 } else if ((bit & (VPTE_W|VPTE_M)) == (VPTE_W|VPTE_M)) {
2633 * We've been asked to clear W & M, I guess
2634 * the caller doesn't want us to update
2635 * the dirty status of the VM page.
2637 pmap_clean_pte(pte, pv->pv_pmap, pv->pv_va);
2640 * We've been asked to clear bits that do
2641 * not interact with hardware.
2643 atomic_clear_long(pte, bit);
2651 * pmap_page_protect:
2653 * Lower the permission for all mappings to a given page.
2656 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2658 if ((prot & VM_PROT_WRITE) == 0) {
2659 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2660 pmap_clearbit(m, VPTE_W);
2661 vm_page_flag_clear(m, PG_WRITEABLE);
2669 pmap_phys_address(vm_pindex_t ppn)
2671 return (i386_ptob(ppn));
2675 * pmap_ts_referenced:
2677 * Return a count of reference bits for a page, clearing those bits.
2678 * It is not necessary for every reference bit to be cleared, but it
2679 * is necessary that 0 only be returned when there are truly no
2680 * reference bits set.
2682 * XXX: The exact number of bits to check and clear is a matter that
2683 * should be tested and standardized at some point in the future for
2684 * optimal aging of shared pages.
2687 pmap_ts_referenced(vm_page_t m)
2689 pv_entry_t pv, pvf, pvn;
2693 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2698 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2703 pvn = TAILQ_NEXT(pv, pv_list);
2705 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2707 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2709 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2712 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2714 if (pte && (*pte & VPTE_A)) {
2716 atomic_clear_long(pte, VPTE_A);
2718 atomic_clear_long_nonlocked(pte, VPTE_A);
2725 } while ((pv = pvn) != NULL && pv != pvf);
2735 * Return whether or not the specified physical page was modified
2736 * in any physical maps.
2739 pmap_is_modified(vm_page_t m)
2741 return pmap_testbit(m, VPTE_M);
2745 * Clear the modify bits on the specified physical page.
2748 pmap_clear_modify(vm_page_t m)
2750 pmap_clearbit(m, VPTE_M);
2754 * pmap_clear_reference:
2756 * Clear the reference bit on the specified physical page.
2759 pmap_clear_reference(vm_page_t m)
2761 pmap_clearbit(m, VPTE_A);
2765 * Miscellaneous support routines follow
2769 i386_protection_init(void)
2773 kp = protection_codes;
2774 for (prot = 0; prot < 8; prot++) {
2775 if (prot & VM_PROT_READ)
2777 if (prot & VM_PROT_WRITE)
2779 if (prot & VM_PROT_EXECUTE)
2788 * Map a set of physical memory pages into the kernel virtual
2789 * address space. Return a pointer to where it is mapped. This
2790 * routine is intended to be used for mapping device memory,
2793 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2797 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2799 vm_offset_t va, tmpva, offset;
2802 offset = pa & PAGE_MASK;
2803 size = roundup(offset + size, PAGE_SIZE);
2805 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
2807 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2809 pa = pa & VPTE_FRAME;
2810 for (tmpva = va; size > 0;) {
2811 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2812 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2820 return ((void *)(va + offset));
2824 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2826 vm_offset_t base, offset;
2828 base = va & VPTE_FRAME;
2829 offset = va & PAGE_MASK;
2830 size = roundup(offset + size, PAGE_SIZE);
2831 pmap_qremove(va, size >> PAGE_SHIFT);
2832 kmem_free(&kernel_map, base, size);
2838 * perform the pmap work for mincore
2841 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2847 ptep = pmap_pte(pmap, addr);
2852 if ((pte = *ptep) != 0) {
2855 val = MINCORE_INCORE;
2856 if ((pte & VPTE_MANAGED) == 0)
2859 pa = pte & VPTE_FRAME;
2861 m = PHYS_TO_VM_PAGE(pa);
2867 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2869 * Modified by someone
2871 else if (m->dirty || pmap_is_modified(m))
2872 val |= MINCORE_MODIFIED_OTHER;
2877 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2880 * Referenced by someone
2882 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
2883 val |= MINCORE_REFERENCED_OTHER;
2884 vm_page_flag_set(m, PG_REFERENCED);
2891 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
2893 struct vmspace *oldvm;
2896 oldvm = p->p_vmspace;
2898 if (oldvm != newvm) {
2899 p->p_vmspace = newvm;
2900 KKASSERT(p->p_nthreads == 1);
2901 lp = RB_ROOT(&p->p_lwp_tree);
2902 pmap_setlwpvm(lp, newvm);
2904 sysref_get(&newvm->vm_sysref);
2905 sysref_put(&oldvm->vm_sysref);
2912 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
2914 struct vmspace *oldvm;
2918 oldvm = lp->lwp_vmspace;
2920 if (oldvm != newvm) {
2921 lp->lwp_vmspace = newvm;
2922 if (curthread->td_lwp == lp) {
2923 pmap = vmspace_pmap(newvm);
2925 atomic_set_int(&pmap->pm_active, mycpu->gd_cpumask);
2927 pmap->pm_active |= 1;
2929 #if defined(SWTCH_OPTIM_STATS)
2932 pmap = vmspace_pmap(oldvm);
2934 atomic_clear_int(&pmap->pm_active, mycpu->gd_cpumask);
2936 pmap->pm_active &= ~1;
2945 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
2948 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
2952 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);