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 1.28 2008/04/28 07:05:08 dillon Exp $
44 * NOTE: PMAP_INVAL_ADD: In pc32 this function is called prior to adjusting
45 * the PTE in the page table, because a cpu synchronization might be required.
46 * The actual invalidation is delayed until the following call or flush. In
47 * the VKERNEL build this function is called prior to adjusting the PTE and
48 * invalidates the table synchronously (not delayed), and is not SMP safe
52 #include <sys/types.h>
53 #include <sys/systm.h>
54 #include <sys/kernel.h>
57 #include <sys/vkernel.h>
59 #include <sys/thread.h>
61 #include <sys/vmspace.h>
64 #include <vm/vm_page.h>
65 #include <vm/vm_extern.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_zone.h>
69 #include <vm/vm_pageout.h>
71 #include <machine/md_var.h>
72 #include <machine/pcb.h>
73 #include <machine/pmap_inval.h>
74 #include <machine/globaldata.h>
76 #include <sys/sysref2.h>
80 struct pmap kernel_pmap;
82 static struct vm_zone pvzone;
83 static struct vm_object pvzone_obj;
84 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
85 static int pv_entry_count;
86 static int pv_entry_max;
87 static int pv_entry_high_water;
88 static int pmap_pagedaemon_waken;
89 static boolean_t pmap_initialized = FALSE;
90 static int protection_codes[8];
92 static void i386_protection_init(void);
93 static void pmap_remove_all(vm_page_t m);
94 static int pmap_release_free_page(struct pmap *pmap, vm_page_t p);
97 #ifndef PMAP_SHPGPERPROC
98 #define PMAP_SHPGPERPROC 200
101 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
103 #define pte_prot(m, p) \
104 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
110 struct pv_entry *pvinit;
112 for (i = 0; i < vm_page_array_size; i++) {
115 m = &vm_page_array[i];
116 TAILQ_INIT(&m->md.pv_list);
117 m->md.pv_list_count = 0;
120 i = vm_page_array_size;
123 pvinit = (struct pv_entry *)kmem_alloc(&kernel_map, i*sizeof(*pvinit));
124 zbootinit(&pvzone, "PV ENTRY", sizeof(*pvinit), pvinit, i);
125 pmap_initialized = TRUE;
131 int shpgperproc = PMAP_SHPGPERPROC;
133 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
134 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
135 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
136 pv_entry_high_water = 9 * (pv_entry_max / 10);
137 zinitna(&pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
141 * Bootstrap the kernel_pmap so it can be used with pmap_enter().
143 * NOTE! pm_pdir for the kernel pmap is offset so VA's translate
144 * directly into PTD indexes (PTA is also offset for the same reason).
145 * This is necessary because, for now, KVA is not mapped at address 0.
147 * Page table pages are not managed like they are in normal pmaps, so
148 * no pteobj is needed.
153 vm_pindex_t i = (vm_offset_t)KernelPTD >> PAGE_SHIFT;
155 kernel_pmap.pm_pdir = KernelPTD - (KvaStart >> SEG_SHIFT);
156 kernel_pmap.pm_pdirpte = KernelPTA[i];
157 kernel_pmap.pm_count = 1;
158 kernel_pmap.pm_active = (cpumask_t)-1;
159 TAILQ_INIT(&kernel_pmap.pm_pvlist);
160 i386_protection_init();
164 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
165 * just dummy it up so it works well enough for fork().
167 * In DragonFly, process pmaps may only be used to manipulate user address
168 * space, never kernel address space.
171 pmap_pinit0(struct pmap *pmap)
176 /************************************************************************
177 * Procedures to manage whole physical maps *
178 ************************************************************************
180 * Initialize a preallocated and zeroed pmap structure,
181 * such as one in a vmspace structure.
184 pmap_pinit(struct pmap *pmap)
190 * No need to allocate page table space yet but we do need a valid
191 * page directory table.
193 if (pmap->pm_pdir == NULL) {
195 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
199 * allocate object for the pte array and page directory
201 npages = VPTE_PAGETABLE_SIZE +
202 (VM_MAX_USER_ADDRESS / PAGE_SIZE) * sizeof(vpte_t);
203 npages = (npages + PAGE_MASK) / PAGE_SIZE;
205 if (pmap->pm_pteobj == NULL)
206 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, npages);
207 pmap->pm_pdindex = npages - 1;
210 * allocate the page directory page
212 ptdpg = vm_page_grab(pmap->pm_pteobj, pmap->pm_pdindex,
213 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
215 ptdpg->wire_count = 1;
216 ++vmstats.v_wire_count;
218 /* not usually mapped */
219 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
220 ptdpg->valid = VM_PAGE_BITS_ALL;
222 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
223 pmap->pm_pdirpte = KernelPTA[(vm_offset_t)pmap->pm_pdir >> PAGE_SHIFT];
224 if ((ptdpg->flags & PG_ZERO) == 0)
225 bzero(pmap->pm_pdir, PAGE_SIZE);
229 pmap->pm_ptphint = NULL;
230 pmap->pm_cpucachemask = 0;
231 TAILQ_INIT(&pmap->pm_pvlist);
232 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
233 pmap->pm_stats.resident_count = 1;
237 * Clean up a pmap structure so it can be physically freed
240 pmap_puninit(pmap_t pmap)
243 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
244 pmap->pm_pdir = NULL;
246 if (pmap->pm_pteobj) {
247 vm_object_deallocate(pmap->pm_pteobj);
248 pmap->pm_pteobj = NULL;
254 * Wire in kernel global address entries. To avoid a race condition
255 * between pmap initialization and pmap_growkernel, this procedure
256 * adds the pmap to the master list (which growkernel scans to update),
257 * then copies the template.
259 * In a virtual kernel there are no kernel global address entries.
262 pmap_pinit2(struct pmap *pmap)
265 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
270 * Release all resources held by the given physical map.
272 * Should only be called if the map contains no valid mappings.
274 static int pmap_release_callback(struct vm_page *p, void *data);
277 pmap_release(struct pmap *pmap)
279 struct mdglobaldata *gd = mdcpu;
280 vm_object_t object = pmap->pm_pteobj;
281 struct rb_vm_page_scan_info info;
283 KKASSERT(pmap != &kernel_pmap);
285 #if defined(DIAGNOSTIC)
286 if (object->ref_count != 1)
287 panic("pmap_release: pteobj reference count != 1");
290 * Once we destroy the page table, the mapping becomes invalid.
291 * Don't waste time doing a madvise to invalidate the mapping, just
292 * set cpucachemask to 0.
294 if (pmap->pm_pdir == gd->gd_PT1pdir) {
295 gd->gd_PT1pdir = NULL;
297 /* madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL); */
299 if (pmap->pm_pdir == gd->gd_PT2pdir) {
300 gd->gd_PT2pdir = NULL;
302 /* madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL); */
304 if (pmap->pm_pdir == gd->gd_PT3pdir) {
305 gd->gd_PT3pdir = NULL;
307 /* madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL); */
311 info.object = object;
313 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
320 info.limit = object->generation;
322 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
323 pmap_release_callback, &info);
324 if (info.error == 0 && info.mpte) {
325 if (!pmap_release_free_page(pmap, info.mpte))
329 } while (info.error);
332 * Leave the KVA reservation for pm_pdir cached for later reuse.
334 pmap->pm_pdirpte = 0;
335 pmap->pm_cpucachemask = 0;
339 * Callback to release a page table page backing a directory
343 pmap_release_callback(struct vm_page *p, void *data)
345 struct rb_vm_page_scan_info *info = data;
347 if (p->pindex == info->pmap->pm_pdindex) {
351 if (!pmap_release_free_page(info->pmap, p)) {
355 if (info->object->generation != info->limit) {
363 * Retire the given physical map from service. Should only be called if
364 * the map contains no valid mappings.
367 pmap_destroy(pmap_t pmap)
374 count = --pmap->pm_count;
377 panic("destroying a pmap is not yet implemented");
382 * Add a reference to the specified pmap.
385 pmap_reference(pmap_t pmap)
392 /************************************************************************
393 * VMSPACE MANAGEMENT *
394 ************************************************************************
396 * The VMSPACE management we do in our virtual kernel must be reflected
397 * in the real kernel. This is accomplished by making vmspace system
398 * calls to the real kernel.
401 cpu_vmspace_alloc(struct vmspace *vm)
406 #define LAST_EXTENT (VM_MAX_USER_ADDRESS - 0x80000000)
408 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
409 panic("vmspace_create() failed");
411 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
412 PROT_READ|PROT_WRITE,
413 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
415 if (rp == MAP_FAILED)
416 panic("vmspace_mmap: failed1");
417 vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
419 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
420 PROT_READ|PROT_WRITE,
421 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
422 MemImageFd, 0x40000000);
423 if (rp == MAP_FAILED)
424 panic("vmspace_mmap: failed2");
425 vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
427 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
428 PROT_READ|PROT_WRITE,
429 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
430 MemImageFd, 0x80000000);
431 vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
433 if (rp == MAP_FAILED)
434 panic("vmspace_mmap: failed3");
436 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
437 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
439 panic("vmspace_mcontrol: failed1");
440 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
441 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
443 panic("vmspace_mcontrol: failed2");
444 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
445 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
447 panic("vmspace_mcontrol: failed3");
451 cpu_vmspace_free(struct vmspace *vm)
453 if (vmspace_destroy(&vm->vm_pmap) < 0)
454 panic("vmspace_destroy() failed");
457 /************************************************************************
458 * Procedures which operate directly on the kernel PMAP *
459 ************************************************************************/
462 * This maps the requested page table and gives us access to it.
464 * This routine can be called from a potentially preempting interrupt
465 * thread or from a normal thread.
468 get_ptbase(struct pmap *pmap, vm_offset_t va)
470 struct mdglobaldata *gd = mdcpu;
472 if (pmap == &kernel_pmap) {
473 KKASSERT(va >= KvaStart && va < KvaEnd);
474 return(KernelPTA + (va >> PAGE_SHIFT));
475 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
476 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
477 *gd->gd_PT1pde = pmap->pm_pdirpte;
478 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
479 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
481 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
482 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
483 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
484 *gd->gd_PT2pde = pmap->pm_pdirpte;
485 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
486 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
488 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
492 * If we aren't running from a potentially preempting interrupt,
493 * load a new page table directory into the page table cache
495 if (gd->mi.gd_intr_nesting_level == 0 &&
496 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0) {
498 * Choose one or the other and map the page table
499 * in the KVA space reserved for it.
501 if ((gd->gd_PTflip = 1 - gd->gd_PTflip) == 0) {
502 gd->gd_PT1pdir = pmap->pm_pdir;
503 *gd->gd_PT1pde = pmap->pm_pdirpte;
504 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
505 atomic_set_int(&pmap->pm_cpucachemask,
507 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
509 gd->gd_PT2pdir = pmap->pm_pdir;
510 *gd->gd_PT2pde = pmap->pm_pdirpte;
511 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
512 atomic_set_int(&pmap->pm_cpucachemask,
514 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
519 * If we are running from a preempting interrupt use a private
520 * map. The caller must be in a critical section.
522 KKASSERT(IN_CRITICAL_SECT(curthread));
523 if (pmap->pm_pdir == gd->gd_PT3pdir) {
524 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
525 *gd->gd_PT3pde = pmap->pm_pdirpte;
526 madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL);
527 atomic_set_int(&pmap->pm_cpucachemask,
531 gd->gd_PT3pdir = pmap->pm_pdir;
532 *gd->gd_PT3pde = pmap->pm_pdirpte;
533 madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL);
534 atomic_set_int(&pmap->pm_cpucachemask,
537 return(gd->gd_PT3map + (va >> PAGE_SHIFT));
541 get_ptbase1(struct pmap *pmap, vm_offset_t va)
543 struct mdglobaldata *gd = mdcpu;
545 if (pmap == &kernel_pmap) {
546 KKASSERT(va >= KvaStart && va < KvaEnd);
547 return(KernelPTA + (va >> PAGE_SHIFT));
548 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
549 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
550 *gd->gd_PT1pde = pmap->pm_pdirpte;
551 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
552 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
554 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
556 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
557 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
558 gd->gd_PT1pdir = pmap->pm_pdir;
559 *gd->gd_PT1pde = pmap->pm_pdirpte;
560 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
561 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
565 get_ptbase2(struct pmap *pmap, vm_offset_t va)
567 struct mdglobaldata *gd = mdcpu;
569 if (pmap == &kernel_pmap) {
570 KKASSERT(va >= KvaStart && va < KvaEnd);
571 return(KernelPTA + (va >> PAGE_SHIFT));
572 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
573 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
574 *gd->gd_PT2pde = pmap->pm_pdirpte;
575 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
576 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
578 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
580 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
581 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
582 gd->gd_PT2pdir = pmap->pm_pdir;
583 *gd->gd_PT2pde = pmap->pm_pdirpte;
584 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
585 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
589 * Return a pointer to the page table entry for the specified va in the
590 * specified pmap. NULL is returned if there is no valid page table page
593 static __inline vpte_t *
594 pmap_pte(struct pmap *pmap, vm_offset_t va)
598 ptep = &pmap->pm_pdir[va >> SEG_SHIFT];
602 return (get_ptbase(pmap, va));
608 * Enter a mapping into kernel_pmap. Mappings created in this fashion
609 * are not managed. Mappings must be immediately accessible on all cpus.
611 * Call pmap_inval_pte() to invalidate the virtual pte and clean out the
612 * real pmap and handle related races before storing the new vpte.
615 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
620 KKASSERT(va >= KvaStart && va < KvaEnd);
621 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
622 ptep = KernelPTA + (va >> PAGE_SHIFT);
624 pmap_inval_pte(ptep, &kernel_pmap, va);
629 * Synchronize a kvm mapping originally made for the private use on
630 * some other cpu so it can be used on all cpus.
632 * XXX add MADV_RESYNC to improve performance.
635 pmap_kenter_sync(vm_offset_t va)
637 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
641 * Synchronize a kvm mapping originally made for the private use on
642 * some other cpu so it can be used on our cpu. Turns out to be the
643 * same madvise() call, because we have to sync the real pmaps anyway.
645 * XXX add MADV_RESYNC to improve performance.
648 pmap_kenter_sync_quick(vm_offset_t va)
650 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
655 * Make a previously read-only kernel mapping R+W (not implemented by
659 pmap_kmodify_rw(vm_offset_t va)
661 *pmap_kpte(va) |= VPTE_R | VPTE_W;
662 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
666 * Make a kernel mapping non-cacheable (not applicable to virtual kernels)
669 pmap_kmodify_nc(vm_offset_t va)
671 *pmap_kpte(va) |= VPTE_N;
672 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
678 * Map a contiguous range of physical memory to a KVM
681 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
683 while (start < end) {
684 pmap_kenter(virt, start);
692 pmap_kpte(vm_offset_t va)
696 KKASSERT(va >= KvaStart && va < KvaEnd);
697 ptep = KernelPTA + (va >> PAGE_SHIFT);
702 * Enter an unmanaged KVA mapping for the private use of the current
703 * cpu only. pmap_kenter_sync() may be called to make the mapping usable
706 * It is illegal for the mapping to be accessed by other cpus unleess
707 * pmap_kenter_sync*() is called.
710 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
715 KKASSERT(va >= KvaStart && va < KvaEnd);
717 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
718 ptep = KernelPTA + (va >> PAGE_SHIFT);
720 pmap_inval_pte_quick(ptep, &kernel_pmap, va);
725 * Make a temporary mapping for a physical address. This is only intended
726 * to be used for panic dumps.
729 pmap_kenter_temporary(vm_paddr_t pa, int i)
731 pmap_kenter(crashdumpmap + (i * PAGE_SIZE), pa);
732 return ((void *)crashdumpmap);
736 * Remove an unmanaged mapping created with pmap_kenter*().
739 pmap_kremove(vm_offset_t va)
743 KKASSERT(va >= KvaStart && va < KvaEnd);
745 ptep = KernelPTA + (va >> PAGE_SHIFT);
747 pmap_inval_pte(ptep, &kernel_pmap, va);
752 * Remove an unmanaged mapping created with pmap_kenter*() but synchronize
753 * only with this cpu.
755 * Unfortunately because we optimize new entries by testing VPTE_V later
756 * on, we actually still have to synchronize with all the cpus. XXX maybe
757 * store a junk value and test against 0 in the other places instead?
760 pmap_kremove_quick(vm_offset_t va)
764 KKASSERT(va >= KvaStart && va < KvaEnd);
766 ptep = KernelPTA + (va >> PAGE_SHIFT);
768 pmap_inval_pte(ptep, &kernel_pmap, va); /* NOT _quick */
773 * Extract the physical address from the kernel_pmap that is associated
774 * with the specified virtual address.
777 pmap_kextract(vm_offset_t va)
782 KKASSERT(va >= KvaStart && va < KvaEnd);
784 ptep = KernelPTA + (va >> PAGE_SHIFT);
785 pa = (vm_paddr_t)(*ptep & VPTE_FRAME) | (va & PAGE_MASK);
790 * Map a set of unmanaged VM pages into KVM.
793 pmap_qenter(vm_offset_t va, struct vm_page **m, int count)
795 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
799 ptep = KernelPTA + (va >> PAGE_SHIFT);
801 pmap_inval_pte(ptep, &kernel_pmap, va);
802 *ptep = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
810 * Map a set of VM pages to kernel virtual memory. If a mapping changes
811 * clear the supplied mask. The caller handles any SMP interactions.
812 * The mask is used to provide the caller with hints on what SMP interactions
816 pmap_qenter2(vm_offset_t va, struct vm_page **m, int count, cpumask_t *mask)
818 cpumask_t cmask = mycpu->gd_cpumask;
820 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
825 ptep = KernelPTA + (va >> PAGE_SHIFT);
826 npte = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
829 pmap_inval_pte_quick(ptep, &kernel_pmap, va);
831 } else if ((*mask & cmask) == 0) {
832 pmap_kenter_sync_quick(va);
842 * Undo the effects of pmap_qenter*().
845 pmap_qremove(vm_offset_t va, int count)
847 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
851 ptep = KernelPTA + (va >> PAGE_SHIFT);
853 pmap_inval_pte(ptep, &kernel_pmap, va);
860 /************************************************************************
861 * Misc support glue called by machine independant code *
862 ************************************************************************
864 * These routines are called by machine independant code to operate on
865 * certain machine-dependant aspects of processes, threads, and pmaps.
869 * Initialize MD portions of the thread structure.
872 pmap_init_thread(thread_t td)
874 /* enforce pcb placement */
875 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
876 td->td_savefpu = &td->td_pcb->pcb_save;
877 td->td_sp = (char *)td->td_pcb - 16;
881 * This routine directly affects the fork perf for a process.
884 pmap_init_proc(struct proc *p)
889 * Destroy the UPAGES for a process that has exited and disassociate
890 * the process from its thread.
893 pmap_dispose_proc(struct proc *p)
895 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
899 * We pre-allocate all page table pages for kernel virtual memory so
900 * this routine will only be called if KVM has been exhausted.
903 pmap_growkernel(vm_offset_t addr)
905 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
907 if (addr > virtual_end - SEG_SIZE)
908 panic("KVM exhausted");
909 kernel_vm_end = addr;
913 * The modification bit is not tracked for any pages in this range. XXX
914 * such pages in this maps should always use pmap_k*() functions and not
917 * XXX User and kernel address spaces are independant for virtual kernels,
918 * this function only applies to the kernel pmap.
921 pmap_track_modified(pmap_t pmap, vm_offset_t va)
923 if (pmap != &kernel_pmap)
925 if ((va < clean_sva) || (va >= clean_eva))
931 /************************************************************************
932 * Procedures supporting managed page table pages *
933 ************************************************************************
935 * These procedures are used to track managed page table pages. These pages
936 * use the page table page's vm_page_t to track PTEs in the page. The
937 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
939 * This allows the system to throw away page table pages for user processes
940 * at will and reinstantiate them on demand.
944 * This routine works like vm_page_lookup() but also blocks as long as the
945 * page is busy. This routine does not busy the page it returns.
947 * Unless the caller is managing objects whos pages are in a known state,
948 * the call should be made with a critical section held so the page's object
949 * association remains valid on return.
952 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
957 m = vm_page_lookup(object, pindex);
958 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
964 * This routine unholds page table pages, and if the hold count
965 * drops to zero, then it decrements the wire count.
967 * We must recheck that this is the last hold reference after busy-sleeping
971 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
973 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
975 KASSERT(m->queue == PQ_NONE,
976 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
978 if (m->hold_count == 1) {
980 * Unmap the page table page.
983 KKASSERT(pmap->pm_pdir[m->pindex] != 0);
984 pmap_inval_pde(&pmap->pm_pdir[m->pindex], pmap,
985 (vm_offset_t)m->pindex << SEG_SHIFT);
986 KKASSERT(pmap->pm_stats.resident_count > 0);
987 --pmap->pm_stats.resident_count;
989 if (pmap->pm_ptphint == m)
990 pmap->pm_ptphint = NULL;
993 * This was our last hold, the page had better be unwired
994 * after we decrement wire_count.
996 * FUTURE NOTE: shared page directory page could result in
997 * multiple wire counts.
1001 KKASSERT(m->wire_count == 0);
1002 --vmstats.v_wire_count;
1004 vm_page_free_zero(m);
1012 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
1014 KKASSERT(m->hold_count > 0);
1015 if (m->hold_count > 1) {
1019 return _pmap_unwire_pte_hold(pmap, m);
1024 * After removing a page table entry, this routine is used to
1025 * conditionally free the page, and manage the hold/wire counts.
1028 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
1034 * page table pages in the kernel_pmap are not managed.
1036 if (pmap == &kernel_pmap)
1038 ptepindex = (va >> PDRSHIFT);
1039 if (pmap->pm_ptphint &&
1040 (pmap->pm_ptphint->pindex == ptepindex)) {
1041 mpte = pmap->pm_ptphint;
1043 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1044 pmap->pm_ptphint = mpte;
1047 return pmap_unwire_pte_hold(pmap, mpte);
1051 * Attempt to release and free the vm_page backing a page directory page
1052 * in a pmap. Returns 1 on success, 0 on failure (if the procedure had
1056 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1058 vpte_t *pde = pmap->pm_pdir;
1061 * This code optimizes the case of freeing non-busy
1062 * page-table pages. Those pages are zero now, and
1063 * might as well be placed directly into the zero queue.
1065 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1069 KKASSERT(pmap->pm_stats.resident_count > 0);
1070 --pmap->pm_stats.resident_count;
1072 if (p->hold_count) {
1073 panic("pmap_release: freeing held page table page");
1076 * Page directory pages need to have the kernel stuff cleared, so
1077 * they can go into the zero queue also.
1079 * In virtual kernels there is no 'kernel stuff'. For the moment
1080 * I just make sure the whole thing has been zero'd even though
1081 * it should already be completely zero'd.
1083 * pmaps for vkernels do not self-map because they do not share
1084 * their address space with the vkernel. Clearing of pde[] thus
1085 * only applies to page table pages and not to the page directory
1088 if (p->pindex == pmap->pm_pdindex) {
1089 bzero(pde, VPTE_PAGETABLE_SIZE);
1090 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1092 KKASSERT(pde[p->pindex] != 0);
1093 pmap_inval_pde(&pde[p->pindex], pmap,
1094 (vm_offset_t)p->pindex << SEG_SHIFT);
1098 * Clear the matching hint
1100 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1101 pmap->pm_ptphint = NULL;
1104 * And throw the page away. The page is completely zero'd out so
1105 * optimize the free call.
1108 vmstats.v_wire_count--;
1109 vm_page_free_zero(p);
1114 * This routine is called if the page table page is not mapped in the page
1117 * The routine is broken up into two parts for readability.
1119 * It must return a held mpte and map the page directory page as required.
1120 * Because vm_page_grab() can block, we must re-check pm_pdir[ptepindex]
1123 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1129 * Find or fabricate a new pagetable page. A busied page will be
1130 * returned. This call may block.
1132 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1133 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1135 KASSERT(m->queue == PQ_NONE,
1136 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1139 * Increment the hold count for the page we will be returning to
1145 * It is possible that someone else got in and mapped by the page
1146 * directory page while we were blocked, if so just unbusy and
1147 * return the held page.
1149 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1150 Debugger("PTEPA RACE");
1151 KKASSERT((ptepa & VPTE_FRAME) == VM_PAGE_TO_PHYS(m));
1156 if (m->wire_count == 0)
1157 vmstats.v_wire_count++;
1161 * Map the pagetable page into the process address space, if
1162 * it isn't already there.
1164 ++pmap->pm_stats.resident_count;
1166 ptepa = VM_PAGE_TO_PHYS(m);
1167 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
1171 * We are likely about to access this page table page, so set the
1172 * page table hint to reduce overhead.
1174 pmap->pm_ptphint = m;
1177 * Try to use the new mapping, but if we cannot, then
1178 * do it with the routine that maps the page explicitly.
1180 if ((m->flags & PG_ZERO) == 0)
1181 pmap_zero_page(ptepa);
1183 m->valid = VM_PAGE_BITS_ALL;
1184 vm_page_flag_clear(m, PG_ZERO);
1185 vm_page_flag_set(m, PG_MAPPED);
1192 * Determine the page table page required to access the VA in the pmap
1193 * and allocate it if necessary. Return a held vm_page_t for the page.
1195 * Only used with user pmaps.
1198 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1205 * Calculate pagetable page index
1207 ptepindex = va >> PDRSHIFT;
1210 * Get the page directory entry
1212 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1215 * This supports switching from a 4MB page to a
1218 if (ptepa & VPTE_PS) {
1219 KKASSERT(pmap->pm_pdir[ptepindex] != 0);
1220 pmap_inval_pde(&pmap->pm_pdir[ptepindex], pmap,
1221 (vm_offset_t)ptepindex << SEG_SHIFT);
1226 * If the page table page is mapped, we just increment the
1227 * hold count, and activate it.
1231 * In order to get the page table page, try the
1234 if (pmap->pm_ptphint &&
1235 (pmap->pm_ptphint->pindex == ptepindex)) {
1236 m = pmap->pm_ptphint;
1238 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1239 pmap->pm_ptphint = m;
1245 * Here if the pte page isn't mapped, or if it has been deallocated.
1247 return _pmap_allocpte(pmap, ptepindex);
1250 /************************************************************************
1251 * Managed pages in pmaps *
1252 ************************************************************************
1254 * All pages entered into user pmaps and some pages entered into the kernel
1255 * pmap are managed, meaning that pmap_protect() and other related management
1256 * functions work on these pages.
1260 * free the pv_entry back to the free list. This function may be
1261 * called from an interrupt.
1263 static __inline void
1264 free_pv_entry(pv_entry_t pv)
1271 * get a new pv_entry, allocating a block from the system
1272 * when needed. This function may be called from an interrupt.
1278 if (pv_entry_high_water &&
1279 (pv_entry_count > pv_entry_high_water) &&
1280 (pmap_pagedaemon_waken == 0)) {
1281 pmap_pagedaemon_waken = 1;
1282 wakeup (&vm_pages_needed);
1284 return zalloc(&pvzone);
1288 * This routine is very drastic, but can save the system
1296 static int warningdone=0;
1298 if (pmap_pagedaemon_waken == 0)
1301 if (warningdone < 5) {
1302 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1306 for(i = 0; i < vm_page_array_size; i++) {
1307 m = &vm_page_array[i];
1308 if (m->wire_count || m->hold_count || m->busy ||
1309 (m->flags & PG_BUSY))
1313 pmap_pagedaemon_waken = 0;
1317 * If it is the first entry on the list, it is actually
1318 * in the header and we must copy the following entry up
1319 * to the header. Otherwise we must search the list for
1320 * the entry. In either case we free the now unused entry.
1323 pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va)
1329 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1330 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1331 if (pmap == pv->pv_pmap && va == pv->pv_va)
1335 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1336 if (va == pv->pv_va)
1342 * Note that pv_ptem is NULL if the page table page itself is not
1343 * managed, even if the page being removed IS managed.
1347 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1348 m->md.pv_list_count--;
1349 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1350 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1351 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1352 ++pmap->pm_generation;
1353 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1361 * Create a pv entry for page at pa for (pmap, va). If the page table page
1362 * holding the VA is managed, mpte will be non-NULL.
1365 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1370 pv = get_pv_entry();
1375 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1376 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1377 m->md.pv_list_count++;
1383 * pmap_remove_pte: do the things to unmap a page in a process
1386 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va)
1391 oldpte = pmap_inval_loadandclear(ptq, pmap, va);
1392 if (oldpte & VPTE_WIRED)
1393 --pmap->pm_stats.wired_count;
1394 KKASSERT(pmap->pm_stats.wired_count >= 0);
1398 * Machines that don't support invlpg, also don't support
1399 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1402 if (oldpte & VPTE_G)
1403 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1405 KKASSERT(pmap->pm_stats.resident_count > 0);
1406 --pmap->pm_stats.resident_count;
1407 if (oldpte & VPTE_MANAGED) {
1408 m = PHYS_TO_VM_PAGE(oldpte);
1409 if (oldpte & VPTE_M) {
1410 #if defined(PMAP_DIAGNOSTIC)
1411 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1413 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1417 if (pmap_track_modified(pmap, va))
1420 if (oldpte & VPTE_A)
1421 vm_page_flag_set(m, PG_REFERENCED);
1422 return pmap_remove_entry(pmap, m, va);
1424 return pmap_unuse_pt(pmap, va, NULL);
1433 * Remove a single page from a process address space.
1435 * This function may not be called from an interrupt if the pmap is
1439 pmap_remove_page(struct pmap *pmap, vm_offset_t va)
1444 * if there is no pte for this address, just skip it!!! Otherwise
1445 * get a local va for mappings for this pmap and remove the entry.
1447 if (*pmap_pde(pmap, va) != 0) {
1448 ptq = get_ptbase(pmap, va);
1450 pmap_remove_pte(pmap, ptq, va);
1458 * Remove the given range of addresses from the specified map.
1460 * It is assumed that the start and end are properly
1461 * rounded to the page size.
1463 * This function may not be called from an interrupt if the pmap is
1467 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1471 vm_offset_t ptpaddr;
1472 vm_pindex_t sindex, eindex;
1477 KKASSERT(pmap->pm_stats.resident_count >= 0);
1478 if (pmap->pm_stats.resident_count == 0)
1482 * special handling of removing one page. a very
1483 * common operation and easy to short circuit some
1486 if (((sva + PAGE_SIZE) == eva) &&
1487 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1488 pmap_remove_page(pmap, sva);
1493 * Get a local virtual address for the mappings that are being
1496 * XXX this is really messy because the kernel pmap is not relative
1499 sindex = (sva >> PAGE_SHIFT);
1500 eindex = (eva >> PAGE_SHIFT);
1502 for (; sindex < eindex; sindex = pdnxt) {
1506 * Calculate index for next page table.
1508 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1509 if (pmap->pm_stats.resident_count == 0)
1512 pdirindex = sindex / NPDEPG;
1513 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1514 KKASSERT(pmap->pm_pdir[pdirindex] != 0);
1515 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1516 pmap_inval_pde(&pmap->pm_pdir[pdirindex], pmap,
1517 (vm_offset_t)pdirindex << SEG_SHIFT);
1522 * Weed out invalid mappings. Note: we assume that the page
1523 * directory table is always allocated, and in kernel virtual.
1529 * Limit our scan to either the end of the va represented
1530 * by the current page table page, or to the end of the
1531 * range being removed.
1537 * NOTE: pmap_remove_pte() can block.
1539 for (; sindex != pdnxt; sindex++) {
1542 ptbase = get_ptbase(pmap, sindex << PAGE_SHIFT);
1545 va = i386_ptob(sindex);
1546 if (pmap_remove_pte(pmap, ptbase, va))
1555 * Removes this physical page from all physical maps in which it resides.
1556 * Reflects back modify bits to the pager.
1558 * This routine may not be called from an interrupt.
1561 pmap_remove_all(vm_page_t m)
1566 #if defined(PMAP_DIAGNOSTIC)
1568 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1571 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1572 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1577 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1578 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1579 --pv->pv_pmap->pm_stats.resident_count;
1581 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1582 KKASSERT(pte != NULL);
1584 tpte = pmap_inval_loadandclear(pte, pv->pv_pmap, pv->pv_va);
1585 if (tpte & VPTE_WIRED)
1586 --pv->pv_pmap->pm_stats.wired_count;
1587 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
1590 vm_page_flag_set(m, PG_REFERENCED);
1593 * Update the vm_page_t clean and reference bits.
1595 if (tpte & VPTE_M) {
1596 #if defined(PMAP_DIAGNOSTIC)
1597 if (pmap_nw_modified((pt_entry_t) tpte)) {
1599 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1603 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
1606 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1607 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1608 ++pv->pv_pmap->pm_generation;
1609 m->md.pv_list_count--;
1610 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
1614 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1621 * Set the physical protection on the specified range of this map
1624 * This function may not be called from an interrupt if the map is
1625 * not the kernel_pmap.
1628 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1632 vm_offset_t pdnxt, ptpaddr;
1633 vm_pindex_t sindex, eindex;
1639 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1640 pmap_remove(pmap, sva, eva);
1644 if (prot & VM_PROT_WRITE)
1647 ptbase = get_ptbase(pmap, sva);
1649 sindex = (sva >> PAGE_SHIFT);
1650 eindex = (eva >> PAGE_SHIFT);
1653 for (; sindex < eindex; sindex = pdnxt) {
1657 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1659 pdirindex = sindex / NPDEPG;
1662 * Clear the modified and writable bits for a 4m page.
1663 * Throw away the modified bit (?)
1665 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1666 pmap_clean_pde(&pmap->pm_pdir[pdirindex], pmap,
1667 (vm_offset_t)pdirindex << SEG_SHIFT);
1668 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1673 * Weed out invalid mappings. Note: we assume that the page
1674 * directory table is always allocated, and in kernel virtual.
1679 if (pdnxt > eindex) {
1683 for (; sindex != pdnxt; sindex++) {
1688 * Clean managed pages and also check the accessed
1689 * bit. Just remove write perms for unmanaged
1690 * pages. Be careful of races, turning off write
1691 * access will force a fault rather then setting
1692 * the modified bit at an unexpected time.
1694 ptep = &ptbase[sindex - sbase];
1695 if (*ptep & VPTE_MANAGED) {
1696 pbits = pmap_clean_pte(ptep, pmap,
1699 if (pbits & VPTE_A) {
1700 m = PHYS_TO_VM_PAGE(pbits);
1701 vm_page_flag_set(m, PG_REFERENCED);
1702 atomic_clear_int(ptep, VPTE_A);
1704 if (pbits & VPTE_M) {
1705 if (pmap_track_modified(pmap, i386_ptob(sindex))) {
1707 m = PHYS_TO_VM_PAGE(pbits);
1712 pbits = pmap_setro_pte(ptep, pmap,
1720 * Enter a managed page into a pmap. If the page is not wired related pmap
1721 * data can be destroyed at any time for later demand-operation.
1723 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1724 * specified protection, and wire the mapping if requested.
1726 * NOTE: This routine may not lazy-evaluate or lose information. The
1727 * page must actually be inserted into the given map NOW.
1729 * NOTE: When entering a page at a KVA address, the pmap must be the
1733 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1739 vm_offset_t origpte, newpte;
1748 * Get the page table page. The kernel_pmap's page table pages
1749 * are preallocated and have no associated vm_page_t.
1751 if (pmap == &kernel_pmap)
1754 mpte = pmap_allocpte(pmap, va);
1756 pte = pmap_pte(pmap, va);
1759 * Page Directory table entry not valid, we need a new PT page
1760 * and pmap_allocpte() didn't give us one. Oops!
1763 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1768 * Deal with races on the original mapping (though don't worry
1769 * about VPTE_A races) by cleaning it. This will force a fault
1770 * if an attempt is made to write to the page.
1772 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1773 origpte = pmap_clean_pte(pte, pmap, va);
1774 opa = origpte & VPTE_FRAME;
1776 if (origpte & VPTE_PS)
1777 panic("pmap_enter: attempted pmap_enter on 4MB page");
1780 * Mapping has not changed, must be protection or wiring change.
1782 if (origpte && (opa == pa)) {
1784 * Wiring change, just update stats. We don't worry about
1785 * wiring PT pages as they remain resident as long as there
1786 * are valid mappings in them. Hence, if a user page is wired,
1787 * the PT page will be also.
1789 if (wired && ((origpte & VPTE_WIRED) == 0))
1790 ++pmap->pm_stats.wired_count;
1791 else if (!wired && (origpte & VPTE_WIRED))
1792 --pmap->pm_stats.wired_count;
1793 KKASSERT(pmap->pm_stats.wired_count >= 0);
1796 * Remove the extra pte reference. Note that we cannot
1797 * optimize the RO->RW case because we have adjusted the
1798 * wiring count above and may need to adjust the wiring
1805 * We might be turning off write access to the page,
1806 * so we go ahead and sense modify status.
1808 if (origpte & VPTE_MANAGED) {
1809 if ((origpte & VPTE_M) &&
1810 pmap_track_modified(pmap, va)) {
1812 om = PHYS_TO_VM_PAGE(opa);
1820 * Mapping has changed, invalidate old range and fall through to
1821 * handle validating new mapping.
1825 err = pmap_remove_pte(pmap, pte, va);
1827 panic("pmap_enter: pte vanished, va: 0x%x", va);
1831 * Enter on the PV list if part of our managed memory. Note that we
1832 * raise IPL while manipulating pv_table since pmap_enter can be
1833 * called at interrupt time.
1835 if (pmap_initialized &&
1836 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1837 pmap_insert_entry(pmap, va, mpte, m);
1842 * Increment counters
1844 ++pmap->pm_stats.resident_count;
1846 pmap->pm_stats.wired_count++;
1850 * Now validate mapping with desired protection/wiring.
1852 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1855 newpte |= VPTE_WIRED;
1859 * If the mapping or permission bits are different from the
1860 * (now cleaned) original pte, an update is needed. We've
1861 * already downgraded or invalidated the page so all we have
1862 * to do now is update the bits.
1864 * XXX should we synchronize RO->RW changes to avoid another
1867 if ((origpte & ~(VPTE_W|VPTE_M|VPTE_A)) != newpte) {
1868 *pte = newpte | VPTE_A;
1873 * This is a quick version of pmap_enter(). It is used only under the
1874 * following conditions:
1876 * (1) The pmap is not the kernel_pmap
1877 * (2) The page is not to be wired into the map
1878 * (3) The page is to mapped read-only in the pmap (initially that is)
1879 * (4) The calling procedure is responsible for flushing the TLB
1880 * (5) The page is always managed
1881 * (6) There is no prior mapping at the VA
1885 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
1892 KKASSERT(pmap != &kernel_pmap);
1894 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1897 * Instantiate the page table page if required
1901 * Calculate pagetable page index
1903 ptepindex = va >> PDRSHIFT;
1904 if (mpte && (mpte->pindex == ptepindex)) {
1909 * Get the page directory entry
1911 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1914 * If the page table page is mapped, we just increment
1915 * the hold count, and activate it.
1918 if (ptepa & VPTE_PS)
1919 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1920 if (pmap->pm_ptphint &&
1921 (pmap->pm_ptphint->pindex == ptepindex)) {
1922 mpte = pmap->pm_ptphint;
1924 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1925 pmap->pm_ptphint = mpte;
1931 mpte = _pmap_allocpte(pmap, ptepindex);
1936 * Ok, now that the page table page has been validated, get the pte.
1937 * If the pte is already mapped undo mpte's hold_count and
1940 pte = pmap_pte(pmap, va);
1943 pmap_unwire_pte_hold(pmap, mpte);
1948 * Enter on the PV list if part of our managed memory. Note that we
1949 * raise IPL while manipulating pv_table since pmap_enter can be
1950 * called at interrupt time.
1952 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
1953 pmap_insert_entry(pmap, va, mpte, m);
1956 * Increment counters
1958 ++pmap->pm_stats.resident_count;
1960 pa = VM_PAGE_TO_PHYS(m);
1963 * Now validate mapping with RO protection
1965 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1966 *pte = (vpte_t)pa | VPTE_V | VPTE_U;
1968 *pte = (vpte_t)pa | VPTE_V | VPTE_U | VPTE_MANAGED;
1974 * Extract the physical address for the translation at the specified
1975 * virtual address in the pmap.
1978 pmap_extract(pmap_t pmap, vm_offset_t va)
1983 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
1984 if (pte & VPTE_PS) {
1985 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
1986 rtval |= va & SEG_MASK;
1988 pte = *get_ptbase(pmap, va);
1989 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
1996 #define MAX_INIT_PT (96)
1999 * This routine preloads the ptes for a given object into the specified pmap.
2000 * This eliminates the blast of soft faults on process startup and
2001 * immediately after an mmap.
2003 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2006 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2007 vm_object_t object, vm_pindex_t pindex,
2008 vm_size_t size, int limit)
2010 struct rb_vm_page_scan_info info;
2015 * We can't preinit if read access isn't set or there is no pmap
2018 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2022 * We can't preinit if the pmap is not the current pmap
2024 lp = curthread->td_lwp;
2025 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2028 psize = size >> PAGE_SHIFT;
2030 if ((object->type != OBJT_VNODE) ||
2031 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2032 (object->resident_page_count > MAX_INIT_PT))) {
2036 if (psize + pindex > object->size) {
2037 if (object->size < pindex)
2039 psize = object->size - pindex;
2046 * Use a red-black scan to traverse the requested range and load
2047 * any valid pages found into the pmap.
2049 * We cannot safely scan the object's memq unless we are in a
2050 * critical section since interrupts can remove pages from objects.
2052 info.start_pindex = pindex;
2053 info.end_pindex = pindex + psize - 1;
2060 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2061 pmap_object_init_pt_callback, &info);
2067 pmap_object_init_pt_callback(vm_page_t p, void *data)
2069 struct rb_vm_page_scan_info *info = data;
2070 vm_pindex_t rel_index;
2072 * don't allow an madvise to blow away our really
2073 * free pages allocating pv entries.
2075 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2076 vmstats.v_free_count < vmstats.v_free_reserved) {
2079 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2080 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2081 if ((p->queue - p->pc) == PQ_CACHE)
2082 vm_page_deactivate(p);
2084 rel_index = p->pindex - info->start_pindex;
2085 info->mpte = pmap_enter_quick(info->pmap,
2086 info->addr + i386_ptob(rel_index),
2088 vm_page_flag_set(p, PG_MAPPED);
2095 * pmap_prefault provides a quick way of clustering pagefaults into a
2096 * processes address space. It is a "cousin" of pmap_object_init_pt,
2097 * except it runs at page fault time instead of mmap time.
2101 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2103 static int pmap_prefault_pageorder[] = {
2104 -PAGE_SIZE, PAGE_SIZE,
2105 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2106 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2107 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2111 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2122 * We do not currently prefault mappings that use virtual page
2123 * tables. We do not prefault foreign pmaps.
2125 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2127 lp = curthread->td_lwp;
2128 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2131 object = entry->object.vm_object;
2133 starta = addra - PFBAK * PAGE_SIZE;
2134 if (starta < entry->start)
2135 starta = entry->start;
2136 else if (starta > addra)
2140 * critical section protection is required to maintain the
2141 * page/object association, interrupts can free pages and remove
2142 * them from their objects.
2146 for (i = 0; i < PAGEORDER_SIZE; i++) {
2147 vm_object_t lobject;
2150 addr = addra + pmap_prefault_pageorder[i];
2151 if (addr > addra + (PFFOR * PAGE_SIZE))
2154 if (addr < starta || addr >= entry->end)
2158 * Make sure the page table page already exists
2160 if ((*pmap_pde(pmap, addr)) == NULL)
2164 * Get a pointer to the pte and make sure that no valid page
2167 pte = get_ptbase(pmap, addr);
2172 * Get the page to be mapped
2174 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2177 for (m = vm_page_lookup(lobject, pindex);
2178 (!m && (lobject->type == OBJT_DEFAULT) &&
2179 (lobject->backing_object));
2180 lobject = lobject->backing_object
2182 if (lobject->backing_object_offset & PAGE_MASK)
2184 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2185 m = vm_page_lookup(lobject->backing_object, pindex);
2189 * give-up when a page is not in memory
2195 * If everything meets the requirements for pmap_enter_quick(),
2196 * then enter the page.
2199 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2201 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2203 if ((m->queue - m->pc) == PQ_CACHE) {
2204 vm_page_deactivate(m);
2207 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2208 vm_page_flag_set(m, PG_MAPPED);
2216 * Routine: pmap_change_wiring
2217 * Function: Change the wiring attribute for a map/virtual-address
2219 * In/out conditions:
2220 * The mapping must already exist in the pmap.
2223 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2230 pte = get_ptbase(pmap, va);
2232 if (wired && (*pte & VPTE_WIRED) == 0)
2233 ++pmap->pm_stats.wired_count;
2234 else if (!wired && (*pte & VPTE_WIRED))
2235 --pmap->pm_stats.wired_count;
2236 KKASSERT(pmap->pm_stats.wired_count >= 0);
2239 * Wiring is not a hardware characteristic so there is no need to
2240 * invalidate TLB. However, in an SMP environment we must use
2241 * a locked bus cycle to update the pte (if we are not using
2242 * the pmap_inval_*() API that is)... it's ok to do this for simple
2246 atomic_set_int(pte, VPTE_WIRED);
2248 atomic_clear_int(pte, VPTE_WIRED);
2252 * Copy the range specified by src_addr/len
2253 * from the source map to the range dst_addr/len
2254 * in the destination map.
2256 * This routine is only advisory and need not do anything.
2259 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2260 vm_size_t len, vm_offset_t src_addr)
2263 vm_offset_t end_addr = src_addr + len;
2269 if (dst_addr != src_addr)
2271 if (dst_pmap->pm_pdir == NULL)
2273 if (src_pmap->pm_pdir == NULL)
2278 src_frame = get_ptbase1(src_pmap, src_addr);
2279 dst_frame = get_ptbase2(dst_pmap, src_addr);
2282 * critical section protection is required to maintain the page/object
2283 * association, interrupts can free pages and remove them from
2286 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2287 vpte_t *src_pte, *dst_pte;
2288 vm_page_t dstmpte, srcmpte;
2289 vm_offset_t srcptepaddr;
2292 if (addr >= VM_MAX_USER_ADDRESS)
2293 panic("pmap_copy: invalid to pmap_copy page tables\n");
2296 * Don't let optional prefaulting of pages make us go
2297 * way below the low water mark of free pages or way
2298 * above high water mark of used pv entries.
2300 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2301 pv_entry_count > pv_entry_high_water)
2304 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2305 ptepindex = addr >> PDRSHIFT;
2307 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2308 if (srcptepaddr == 0)
2311 if (srcptepaddr & VPTE_PS) {
2312 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2313 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2314 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2319 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2320 if ((srcmpte == NULL) ||
2321 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2324 if (pdnxt > end_addr)
2327 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2328 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2329 while (addr < pdnxt) {
2333 * we only virtual copy managed pages
2335 if ((ptetemp & VPTE_MANAGED) != 0) {
2337 * We have to check after allocpte for the
2338 * pte still being around... allocpte can
2341 * pmap_allocpte can block, unfortunately
2342 * we have to reload the tables.
2344 dstmpte = pmap_allocpte(dst_pmap, addr);
2345 src_frame = get_ptbase1(src_pmap, src_addr);
2346 dst_frame = get_ptbase2(dst_pmap, src_addr);
2348 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2350 * Clear the modified and accessed
2351 * (referenced) bits during the copy.
2353 * We do not have to clear the write
2354 * bit to force a fault-on-modify
2355 * because the real kernel's target
2356 * pmap is empty and will fault anyway.
2358 m = PHYS_TO_VM_PAGE(ptetemp);
2359 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2360 ++dst_pmap->pm_stats.resident_count;
2361 pmap_insert_entry(dst_pmap, addr,
2364 pmap_unwire_pte_hold(dst_pmap, dstmpte);
2366 if (dstmpte->hold_count >= srcmpte->hold_count)
2380 * Zero the specified PA by mapping the page into KVM and clearing its
2383 * This function may be called from an interrupt and no locking is
2387 pmap_zero_page(vm_paddr_t phys)
2389 struct mdglobaldata *gd = mdcpu;
2393 panic("pmap_zero_page: CMAP3 busy");
2394 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2395 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2397 bzero(gd->gd_CADDR3, PAGE_SIZE);
2403 * pmap_page_assertzero:
2405 * Assert that a page is empty, panic if it isn't.
2408 pmap_page_assertzero(vm_paddr_t phys)
2410 struct mdglobaldata *gd = mdcpu;
2415 panic("pmap_zero_page: CMAP3 busy");
2416 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2417 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2418 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2419 for (i = 0; i < PAGE_SIZE; i += 4) {
2420 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2421 panic("pmap_page_assertzero() @ %p not zero!\n",
2422 (void *)gd->gd_CADDR3);
2432 * Zero part of a physical page by mapping it into memory and clearing
2433 * its contents with bzero.
2435 * off and size may not cover an area beyond a single hardware page.
2438 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2440 struct mdglobaldata *gd = mdcpu;
2444 panic("pmap_zero_page: CMAP3 busy");
2445 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2446 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2447 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2449 bzero((char *)gd->gd_CADDR3 + off, size);
2457 * Copy the physical page from the source PA to the target PA.
2458 * This function may be called from an interrupt. No locking
2462 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2464 struct mdglobaldata *gd = mdcpu;
2467 if (*(int *) gd->gd_CMAP1)
2468 panic("pmap_copy_page: CMAP1 busy");
2469 if (*(int *) gd->gd_CMAP2)
2470 panic("pmap_copy_page: CMAP2 busy");
2472 *(int *) gd->gd_CMAP1 = VPTE_V | VPTE_R | (src & PG_FRAME) | VPTE_A;
2473 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2475 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2476 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2478 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2480 *(int *) gd->gd_CMAP1 = 0;
2481 *(int *) gd->gd_CMAP2 = 0;
2486 * pmap_copy_page_frag:
2488 * Copy the physical page from the source PA to the target PA.
2489 * This function may be called from an interrupt. No locking
2493 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2495 struct mdglobaldata *gd = mdcpu;
2498 if (*(int *) gd->gd_CMAP1)
2499 panic("pmap_copy_page: CMAP1 busy");
2500 if (*(int *) gd->gd_CMAP2)
2501 panic("pmap_copy_page: CMAP2 busy");
2503 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2504 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2506 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2507 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2509 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2510 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2513 *(int *) gd->gd_CMAP1 = 0;
2514 *(int *) gd->gd_CMAP2 = 0;
2519 * Returns true if the pmap's pv is one of the first
2520 * 16 pvs linked to from this page. This count may
2521 * be changed upwards or downwards in the future; it
2522 * is only necessary that true be returned for a small
2523 * subset of pmaps for proper page aging.
2526 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2531 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2536 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2537 if (pv->pv_pmap == pmap) {
2550 * Remove all pages from specified address space
2551 * this aids process exit speeds. Also, this code
2552 * is special cased for current process only, but
2553 * can have the more generic (and slightly slower)
2554 * mode enabled. This is much faster than pmap_remove
2555 * in the case of running down an entire address space.
2558 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2564 int32_t save_generation;
2567 lp = curthread->td_lwp;
2568 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2574 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2575 if (pv->pv_va >= eva || pv->pv_va < sva) {
2576 npv = TAILQ_NEXT(pv, pv_plist);
2580 KKASSERT(pmap == pv->pv_pmap);
2582 pte = pmap_pte(pmap, pv->pv_va);
2585 * We cannot remove wired pages from a process' mapping
2588 if (*pte & VPTE_WIRED) {
2589 npv = TAILQ_NEXT(pv, pv_plist);
2592 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2594 m = PHYS_TO_VM_PAGE(tpte);
2596 KASSERT(m < &vm_page_array[vm_page_array_size],
2597 ("pmap_remove_pages: bad tpte %x", tpte));
2599 KKASSERT(pmap->pm_stats.resident_count > 0);
2600 --pmap->pm_stats.resident_count;
2603 * Update the vm_page_t clean and reference bits.
2605 if (tpte & VPTE_M) {
2609 npv = TAILQ_NEXT(pv, pv_plist);
2610 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2611 save_generation = ++pmap->pm_generation;
2613 m->md.pv_list_count--;
2614 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2615 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2616 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2619 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2623 * Restart the scan if we blocked during the unuse or free
2624 * calls and other removals were made.
2626 if (save_generation != pmap->pm_generation) {
2627 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2628 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2635 * pmap_testbit tests bits in active mappings of a VM page.
2638 pmap_testbit(vm_page_t m, int bit)
2643 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2646 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2651 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2653 * if the bit being tested is the modified bit, then
2654 * mark clean_map and ptes as never
2657 if (bit & (VPTE_A|VPTE_M)) {
2658 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2662 #if defined(PMAP_DIAGNOSTIC)
2664 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2668 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2679 * This routine is used to clear bits in ptes. Certain bits require special
2680 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2682 * This routine is only called with certain VPTE_* bit combinations.
2684 static __inline void
2685 pmap_clearbit(vm_page_t m, int bit)
2691 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2697 * Loop over all current mappings setting/clearing as appropos If
2698 * setting RO do we need to clear the VAC?
2700 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2702 * don't write protect pager mappings
2704 if (bit == VPTE_W) {
2705 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2709 #if defined(PMAP_DIAGNOSTIC)
2711 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2717 * Careful here. We can use a locked bus instruction to
2718 * clear VPTE_A or VPTE_M safely but we need to synchronize
2719 * with the target cpus when we mess with VPTE_W.
2721 * On virtual kernels we must force a new fault-on-write
2722 * in the real kernel if we clear the Modify bit ourselves,
2723 * otherwise the real kernel will not get a new fault and
2724 * will never set our Modify bit again.
2726 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2728 if (bit == VPTE_W) {
2730 * We must also clear VPTE_M when clearing
2733 pbits = pmap_clean_pte(pte, pv->pv_pmap,
2737 } else if (bit == VPTE_M) {
2739 * We do not have to make the page read-only
2740 * when clearing the Modify bit. The real
2741 * kernel will make the real PTE read-only
2742 * or otherwise detect the write and set
2743 * our VPTE_M again simply by us invalidating
2744 * the real kernel VA for the pmap (as we did
2745 * above). This allows the real kernel to
2746 * handle the write fault without forwarding
2749 atomic_clear_int(pte, VPTE_M);
2750 } else if ((bit & (VPTE_W|VPTE_M)) == (VPTE_W|VPTE_M)) {
2752 * We've been asked to clear W & M, I guess
2753 * the caller doesn't want us to update
2754 * the dirty status of the VM page.
2756 pmap_clean_pte(pte, pv->pv_pmap, pv->pv_va);
2759 * We've been asked to clear bits that do
2760 * not interact with hardware.
2762 atomic_clear_int(pte, bit);
2770 * pmap_page_protect:
2772 * Lower the permission for all mappings to a given page.
2775 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2777 if ((prot & VM_PROT_WRITE) == 0) {
2778 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2779 pmap_clearbit(m, VPTE_W);
2787 pmap_phys_address(int ppn)
2789 return (i386_ptob(ppn));
2793 * pmap_ts_referenced:
2795 * Return a count of reference bits for a page, clearing those bits.
2796 * It is not necessary for every reference bit to be cleared, but it
2797 * is necessary that 0 only be returned when there are truly no
2798 * reference bits set.
2800 * XXX: The exact number of bits to check and clear is a matter that
2801 * should be tested and standardized at some point in the future for
2802 * optimal aging of shared pages.
2805 pmap_ts_referenced(vm_page_t m)
2807 pv_entry_t pv, pvf, pvn;
2811 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2816 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2821 pvn = TAILQ_NEXT(pv, pv_list);
2823 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2825 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2827 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2830 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2832 if (pte && (*pte & VPTE_A)) {
2834 atomic_clear_int(pte, VPTE_A);
2836 atomic_clear_int_nonlocked(pte, VPTE_A);
2843 } while ((pv = pvn) != NULL && pv != pvf);
2853 * Return whether or not the specified physical page was modified
2854 * in any physical maps.
2857 pmap_is_modified(vm_page_t m)
2859 return pmap_testbit(m, VPTE_M);
2863 * Clear the modify bits on the specified physical page.
2866 pmap_clear_modify(vm_page_t m)
2868 pmap_clearbit(m, VPTE_M);
2872 * pmap_clear_reference:
2874 * Clear the reference bit on the specified physical page.
2877 pmap_clear_reference(vm_page_t m)
2879 pmap_clearbit(m, VPTE_A);
2883 * Miscellaneous support routines follow
2887 i386_protection_init(void)
2891 kp = protection_codes;
2892 for (prot = 0; prot < 8; prot++) {
2893 if (prot & VM_PROT_READ)
2895 if (prot & VM_PROT_WRITE)
2897 if (prot & VM_PROT_EXECUTE)
2906 * Map a set of physical memory pages into the kernel virtual
2907 * address space. Return a pointer to where it is mapped. This
2908 * routine is intended to be used for mapping device memory,
2911 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2915 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2917 vm_offset_t va, tmpva, offset;
2920 offset = pa & PAGE_MASK;
2921 size = roundup(offset + size, PAGE_SIZE);
2923 va = kmem_alloc_nofault(&kernel_map, size);
2925 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2927 pa = pa & VPTE_FRAME;
2928 for (tmpva = va; size > 0;) {
2929 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2930 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2938 return ((void *)(va + offset));
2942 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2944 vm_offset_t base, offset;
2946 base = va & VPTE_FRAME;
2947 offset = va & PAGE_MASK;
2948 size = roundup(offset + size, PAGE_SIZE);
2949 pmap_qremove(va, size >> PAGE_SHIFT);
2950 kmem_free(&kernel_map, base, size);
2956 * perform the pmap work for mincore
2959 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2965 ptep = pmap_pte(pmap, addr);
2970 if ((pte = *ptep) != 0) {
2973 val = MINCORE_INCORE;
2974 if ((pte & VPTE_MANAGED) == 0)
2977 pa = pte & VPTE_FRAME;
2979 m = PHYS_TO_VM_PAGE(pa);
2985 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2987 * Modified by someone
2989 else if (m->dirty || pmap_is_modified(m))
2990 val |= MINCORE_MODIFIED_OTHER;
2995 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2998 * Referenced by someone
3000 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3001 val |= MINCORE_REFERENCED_OTHER;
3002 vm_page_flag_set(m, PG_REFERENCED);
3009 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3011 struct vmspace *oldvm;
3014 oldvm = p->p_vmspace;
3016 if (oldvm != newvm) {
3017 p->p_vmspace = newvm;
3018 KKASSERT(p->p_nthreads == 1);
3019 lp = RB_ROOT(&p->p_lwp_tree);
3020 pmap_setlwpvm(lp, newvm);
3022 sysref_get(&newvm->vm_sysref);
3023 sysref_put(&oldvm->vm_sysref);
3030 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3032 struct vmspace *oldvm;
3036 oldvm = lp->lwp_vmspace;
3038 if (oldvm != newvm) {
3039 lp->lwp_vmspace = newvm;
3040 if (curthread->td_lwp == lp) {
3041 pmap = vmspace_pmap(newvm);
3043 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3045 pmap->pm_active |= 1;
3047 #if defined(SWTCH_OPTIM_STATS)
3050 pmap = vmspace_pmap(oldvm);
3052 atomic_clear_int(&pmap->pm_active,
3053 1 << mycpu->gd_cpuid);
3055 pmap->pm_active &= ~1;
3064 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3067 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3071 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3078 static void pads (pmap_t pm);
3079 void pmap_pvdump (vm_paddr_t pa);
3081 /* print address space of pmap*/
3089 if (pm == &kernel_pmap)
3091 for (i = 0; i < 1024; i++)
3093 for (j = 0; j < 1024; j++) {
3094 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3095 if (pm == &kernel_pmap && va < KERNBASE)
3097 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3099 ptep = pmap_pte(pm, va);
3100 if (ptep && (*ptep & VPTE_V)) {
3102 (void *)va, (unsigned)*ptep);
3109 pmap_pvdump(vm_paddr_t pa)
3114 kprintf("pa %08llx", (long long)pa);
3115 m = PHYS_TO_VM_PAGE(pa);
3116 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3118 kprintf(" -> pmap %p, va %x, flags %x",
3119 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3121 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);