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 1.21 2007/04/29 18:25:39 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 TAILQ_INIT(&pmap->pm_pvlist);
231 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
235 * Clean up a pmap structure so it can be physically freed
238 pmap_puninit(pmap_t pmap)
241 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
242 pmap->pm_pdir = NULL;
244 if (pmap->pm_pteobj) {
245 vm_object_deallocate(pmap->pm_pteobj);
246 pmap->pm_pteobj = NULL;
252 * Wire in kernel global address entries. To avoid a race condition
253 * between pmap initialization and pmap_growkernel, this procedure
254 * adds the pmap to the master list (which growkernel scans to update),
255 * then copies the template.
257 * In a virtual kernel there are no kernel global address entries.
260 pmap_pinit2(struct pmap *pmap)
263 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
268 * Release all resources held by the given physical map.
270 * Should only be called if the map contains no valid mappings.
272 static int pmap_release_callback(struct vm_page *p, void *data);
275 pmap_release(struct pmap *pmap)
277 struct mdglobaldata *gd = mdcpu;
278 vm_object_t object = pmap->pm_pteobj;
279 struct rb_vm_page_scan_info info;
281 KKASSERT(pmap != &kernel_pmap);
283 #if defined(DIAGNOSTIC)
284 if (object->ref_count != 1)
285 panic("pmap_release: pteobj reference count != 1");
288 #error "Must write code to clear PTxpdir cache across all CPUs"
291 * Once we destroy the page table, the mapping becomes invalid.
292 * Rather then waste time doing a madvise
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); */
306 info.object = object;
308 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
315 info.limit = object->generation;
317 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
318 pmap_release_callback, &info);
319 if (info.error == 0 && info.mpte) {
320 if (!pmap_release_free_page(pmap, info.mpte))
324 } while (info.error);
327 * Leave the KVA reservation for pm_pdir cached for later reuse.
329 pmap->pm_pdirpte = 0;
333 pmap_release_callback(struct vm_page *p, void *data)
335 struct rb_vm_page_scan_info *info = data;
337 if (p->pindex == info->pmap->pm_pdindex) {
341 if (!pmap_release_free_page(info->pmap, p)) {
345 if (info->object->generation != info->limit) {
353 * Retire the given physical map from service. Should only be called if
354 * the map contains no valid mappings.
357 pmap_destroy(pmap_t pmap)
364 count = --pmap->pm_count;
367 panic("destroying a pmap is not yet implemented");
372 * Add a reference to the specified pmap.
375 pmap_reference(pmap_t pmap)
382 /************************************************************************
383 * VMSPACE MANAGEMENT *
384 ************************************************************************
386 * The VMSPACE management we do in our virtual kernel must be reflected
387 * in the real kernel. This is accomplished by making vmspace system
388 * calls to the real kernel.
391 cpu_vmspace_alloc(struct vmspace *vm)
396 #define LAST_EXTENT (VM_MAX_USER_ADDRESS - 0x80000000)
398 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
399 panic("vmspace_create() failed");
401 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
402 PROT_READ|PROT_WRITE,
403 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
405 if (rp == MAP_FAILED)
406 panic("vmspace_mmap: failed1");
407 vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
409 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
410 PROT_READ|PROT_WRITE,
411 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
412 MemImageFd, 0x40000000);
413 if (rp == MAP_FAILED)
414 panic("vmspace_mmap: failed2");
415 vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
417 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
418 PROT_READ|PROT_WRITE,
419 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
420 MemImageFd, 0x80000000);
421 vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
423 if (rp == MAP_FAILED)
424 panic("vmspace_mmap: failed3");
426 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
427 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
429 panic("vmspace_mcontrol: failed1");
430 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
431 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
433 panic("vmspace_mcontrol: failed2");
434 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
435 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
437 panic("vmspace_mcontrol: failed3");
441 cpu_vmspace_free(struct vmspace *vm)
443 if (vmspace_destroy(&vm->vm_pmap) < 0)
444 panic("vmspace_destroy() failed");
447 /************************************************************************
448 * Procedures which operate directly on the kernel PMAP *
449 ************************************************************************/
452 * This maps the requested page table and gives us access to it.
455 get_ptbase(struct pmap *pmap, vm_offset_t va)
457 struct mdglobaldata *gd = mdcpu;
459 if (pmap == &kernel_pmap) {
460 KKASSERT(va >= KvaStart && va < KvaEnd);
461 return(KernelPTA + (va >> PAGE_SHIFT));
462 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
463 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
464 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
465 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
469 * Otherwise choose one or the other and map the page table
470 * in the KVA space reserved for it.
472 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
473 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
475 if ((gd->gd_PTflip = 1 - gd->gd_PTflip) == 0) {
476 gd->gd_PT1pdir = pmap->pm_pdir;
477 *gd->gd_PT1pde = pmap->pm_pdirpte;
478 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
479 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
481 gd->gd_PT2pdir = pmap->pm_pdir;
482 *gd->gd_PT2pde = pmap->pm_pdirpte;
483 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
484 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
489 get_ptbase1(struct pmap *pmap, vm_offset_t va)
491 struct mdglobaldata *gd = mdcpu;
493 if (pmap == &kernel_pmap) {
494 KKASSERT(va >= KvaStart && va < KvaEnd);
495 return(KernelPTA + (va >> PAGE_SHIFT));
496 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
497 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
499 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
500 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
501 gd->gd_PT1pdir = pmap->pm_pdir;
502 *gd->gd_PT1pde = pmap->pm_pdirpte;
503 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
504 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
508 get_ptbase2(struct pmap *pmap, vm_offset_t va)
510 struct mdglobaldata *gd = mdcpu;
512 if (pmap == &kernel_pmap) {
513 KKASSERT(va >= KvaStart && va < KvaEnd);
514 return(KernelPTA + (va >> PAGE_SHIFT));
515 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
516 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
518 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
519 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
520 gd->gd_PT2pdir = pmap->pm_pdir;
521 *gd->gd_PT2pde = pmap->pm_pdirpte;
522 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
523 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
527 * When removing a page directory the related VA range in the self-mapping
528 * of the page table must be invalidated.
531 inval_ptbase_pagedir(pmap_t pmap, vm_pindex_t pindex)
533 struct mdglobaldata *gd = mdcpu;
537 #error "Must inval self-mappings in all gd's"
539 if (pmap == &kernel_pmap) {
540 va = (vm_offset_t)KernelPTA + (pindex << PAGE_SHIFT);
541 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
544 * XXX this should not strictly be needed because the page
545 * dir should alread be invalidated. test and remove
547 va = (vm_offset_t)pindex << PAGE_SHIFT;
548 vmspace_mcontrol(pmap, (void *)va, SEG_SIZE, MADV_INVAL, 0);
550 if (pmap->pm_pdir == gd->gd_PT1pdir) {
551 va = (vm_offset_t)gd->gd_PT1map + (pindex << PAGE_SHIFT);
552 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
554 if (pmap->pm_pdir == gd->gd_PT2pdir) {
555 va = (vm_offset_t)gd->gd_PT2map + (pindex << PAGE_SHIFT);
556 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
561 * Return a pointer to the page table entry for the specified va in the
562 * specified pmap. NULL is returned if there is no valid page table page
565 static __inline vpte_t *
566 pmap_pte(struct pmap *pmap, vm_offset_t va)
570 ptep = &pmap->pm_pdir[va >> SEG_SHIFT];
574 return (get_ptbase(pmap, va));
580 * Enter a mapping into kernel_pmap. Mappings created in this fashion
584 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
589 pmap_inval_info info;
592 KKASSERT(va >= KvaStart && va < KvaEnd);
593 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
594 ptep = KernelPTA + (va >> PAGE_SHIFT);
595 if (*ptep & VPTE_V) {
597 pmap_inval_init(&info);
598 pmap_inval_add(&info, &kernel_pmap, va);
602 pmap_inval_flush(&info);
604 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
612 pmap_kenter_sync(vm_offset_t va)
614 pmap_inval_info info;
616 pmap_inval_init(&info);
617 pmap_inval_add(&info, &kernel_pmap, va);
618 pmap_inval_flush(&info);
622 pmap_kenter_sync_quick(vm_offset_t va)
624 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
628 * XXX these need to be recoded. They are not used in any critical path.
631 pmap_kmodify_rw(vm_offset_t va)
633 *pmap_kpte(va) |= VPTE_R | VPTE_W;
634 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
638 pmap_kmodify_nc(vm_offset_t va)
641 *pmap_kpte(va) |= VPTE_N;
642 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
647 * Map a contiguous range of physical memory to a KVM
650 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
652 while (start < end) {
653 pmap_kenter(virt, start);
661 pmap_kpte(vm_offset_t va)
665 KKASSERT(va >= KvaStart && va < KvaEnd);
666 ptep = KernelPTA + (va >> PAGE_SHIFT);
671 * Enter a mapping into kernel_pmap without any SMP interactions.
673 * Mappings created in this fashion are not managed.
676 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
681 KKASSERT(va >= KvaStart && va < KvaEnd);
683 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
684 ptep = KernelPTA + (va >> PAGE_SHIFT);
685 if (*ptep & VPTE_V) {
687 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
694 * Make a temporary mapping for a physical address. This is only intended
695 * to be used for panic dumps.
698 pmap_kenter_temporary(vm_paddr_t pa, int i)
700 pmap_kenter(crashdumpmap + (i * PAGE_SIZE), pa);
701 return ((void *)crashdumpmap);
705 * Remove an unmanaged mapping created with pmap_kenter*().
708 pmap_kremove(vm_offset_t va)
712 pmap_inval_info info;
715 KKASSERT(va >= KvaStart && va < KvaEnd);
717 ptep = KernelPTA + (va >> PAGE_SHIFT);
718 if (*ptep & VPTE_V) {
720 pmap_inval_init(&info);
721 pmap_inval_add(&info, &kernel_pmap, va);
725 pmap_inval_flush(&info);
727 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
736 * Remove an unmanaged mapping created with pmap_kenter*() without
737 * going through any SMP interactions.
740 pmap_kremove_quick(vm_offset_t va)
744 KKASSERT(va >= KvaStart && va < KvaEnd);
746 ptep = KernelPTA + (va >> PAGE_SHIFT);
747 if (*ptep & VPTE_V) {
749 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
756 * Extract the physical address from the kernel_pmap that is associated
757 * with the specified virtual address.
760 pmap_kextract(vm_offset_t va)
765 KKASSERT(va >= KvaStart && va < KvaEnd);
767 ptep = KernelPTA + (va >> PAGE_SHIFT);
768 pa = (vm_paddr_t)(*ptep & VPTE_FRAME) | (va & PAGE_MASK);
773 * Map a set of unmanaged VM pages into KVM.
776 pmap_qenter(vm_offset_t va, struct vm_page **m, int count)
778 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
782 ptep = KernelPTA + (va >> PAGE_SHIFT);
784 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
785 *ptep = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
797 * Map a set of VM pages to kernel virtual memory. If a mapping changes
798 * clear the supplied mask. The caller handles any SMP interactions.
799 * The mask is used to provide the caller with hints on what SMP interactions
803 pmap_qenter2(vm_offset_t va, struct vm_page **m, int count, cpumask_t *mask)
805 cpumask_t cmask = mycpu->gd_cpumask;
807 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
812 ptep = KernelPTA + (va >> PAGE_SHIFT);
813 npte = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
817 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
818 } else if ((*mask & cmask) == 0) {
819 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
829 * Undo the effects of pmap_qenter*().
832 pmap_qremove(vm_offset_t va, int count)
834 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
838 ptep = KernelPTA + (va >> PAGE_SHIFT);
840 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
851 /************************************************************************
852 * Misc support glue called by machine independant code *
853 ************************************************************************
855 * These routines are called by machine independant code to operate on
856 * certain machine-dependant aspects of processes, threads, and pmaps.
860 * Initialize MD portions of the thread structure.
863 pmap_init_thread(thread_t td)
865 /* enforce pcb placement */
866 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
867 td->td_savefpu = &td->td_pcb->pcb_save;
868 td->td_sp = (char *)td->td_pcb - 16;
872 * This routine directly affects the fork perf for a process.
875 pmap_init_proc(struct proc *p)
880 * Destroy the UPAGES for a process that has exited and disassociate
881 * the process from its thread.
884 pmap_dispose_proc(struct proc *p)
886 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
890 * We pre-allocate all page table pages for kernel virtual memory so
891 * this routine will only be called if KVM has been exhausted.
894 pmap_growkernel(vm_offset_t addr)
896 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
898 if (addr > virtual_end - SEG_SIZE)
899 panic("KVM exhausted");
900 kernel_vm_end = addr;
904 * The modification bit is not tracked for any pages in this range. XXX
905 * such pages in this maps should always use pmap_k*() functions and not
908 * XXX User and kernel address spaces are independant for virtual kernels,
909 * this function only applies to the kernel pmap.
912 pmap_track_modified(pmap_t pmap, vm_offset_t va)
914 if (pmap != &kernel_pmap)
916 if ((va < clean_sva) || (va >= clean_eva))
922 /************************************************************************
923 * Procedures supporting managed page table pages *
924 ************************************************************************
926 * These procedures are used to track managed page table pages. These pages
927 * use the page table page's vm_page_t to track PTEs in the page. The
928 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
930 * This allows the system to throw away page table pages for user processes
931 * at will and reinstantiate them on demand.
935 * This routine works like vm_page_lookup() but also blocks as long as the
936 * page is busy. This routine does not busy the page it returns.
938 * Unless the caller is managing objects whos pages are in a known state,
939 * the call should be made with a critical section held so the page's object
940 * association remains valid on return.
943 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
948 m = vm_page_lookup(object, pindex);
949 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
955 * This routine unholds page table pages, and if the hold count
956 * drops to zero, then it decrements the wire count.
959 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
961 pmap_inval_flush(info);
962 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
965 if (m->hold_count == 0) {
967 * unmap the page table page
969 pmap->pm_pdir[m->pindex] = 0;
970 --pmap->pm_stats.resident_count;
971 inval_ptbase_pagedir(pmap, m->pindex);
973 if (pmap->pm_ptphint == m)
974 pmap->pm_ptphint = NULL;
977 * If the page is finally unwired, simply free it.
980 if (m->wire_count == 0) {
983 vm_page_free_zero(m);
984 --vmstats.v_wire_count;
992 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
995 if (m->hold_count == 0)
996 return _pmap_unwire_pte_hold(pmap, m, info);
1002 * After removing a page table entry, this routine is used to
1003 * conditionally free the page, and manage the hold/wire counts.
1006 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1007 pmap_inval_info_t info)
1013 * page table pages in the kernel_pmap are not managed.
1015 if (pmap == &kernel_pmap)
1017 ptepindex = (va >> PDRSHIFT);
1018 if (pmap->pm_ptphint &&
1019 (pmap->pm_ptphint->pindex == ptepindex)) {
1020 mpte = pmap->pm_ptphint;
1022 pmap_inval_flush(info);
1023 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1024 pmap->pm_ptphint = mpte;
1027 return pmap_unwire_pte_hold(pmap, mpte, info);
1031 * Attempt to release and free an vm_page in a pmap. Returns 1 on success,
1032 * 0 on failure (if the procedure had to sleep).
1035 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1037 vpte_t *pde = pmap->pm_pdir;
1039 * This code optimizes the case of freeing non-busy
1040 * page-table pages. Those pages are zero now, and
1041 * might as well be placed directly into the zero queue.
1043 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1047 pmap->pm_stats.resident_count--;
1049 if (p->hold_count) {
1050 panic("pmap_release: freeing held page table page");
1053 * Page directory pages need to have the kernel stuff cleared, so
1054 * they can go into the zero queue also.
1056 * In virtual kernels there is no 'kernel stuff'. For the moment
1057 * I just make sure the whole thing has been zero'd even though
1058 * it should already be completely zero'd.
1060 * pmaps for vkernels do not self-map because they do not share
1061 * their address space with the vkernel. Clearing of pde[] thus
1062 * only applies to page table pages and not to the page directory
1065 if (p->pindex == pmap->pm_pdindex) {
1066 bzero(pde, VPTE_PAGETABLE_SIZE);
1067 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1073 * Clear the matching hint
1075 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1076 pmap->pm_ptphint = NULL;
1079 * And throw the page away. The page is completely zero'd out so
1080 * optimize the free call.
1083 vmstats.v_wire_count--;
1084 vm_page_free_zero(p);
1089 * This routine is called if the page table page is not mapped in the page
1092 * The routine is broken up into two parts for readability.
1095 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1101 * Find or fabricate a new pagetable page
1103 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1104 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1106 KASSERT(m->queue == PQ_NONE,
1107 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1109 if (m->wire_count == 0)
1110 vmstats.v_wire_count++;
1114 * Increment the hold count for the page table page
1115 * (denoting a new mapping.)
1120 * Map the pagetable page into the process address space, if
1121 * it isn't already there.
1123 pmap->pm_stats.resident_count++;
1125 ptepa = VM_PAGE_TO_PHYS(m);
1126 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
1130 * We are likely about to access this page table page, so set the
1131 * page table hint to reduce overhead.
1133 pmap->pm_ptphint = m;
1136 * Try to use the new mapping, but if we cannot, then
1137 * do it with the routine that maps the page explicitly.
1139 if ((m->flags & PG_ZERO) == 0)
1140 pmap_zero_page(ptepa);
1142 m->valid = VM_PAGE_BITS_ALL;
1143 vm_page_flag_clear(m, PG_ZERO);
1144 vm_page_flag_set(m, PG_MAPPED);
1151 * Determine the page table page required to access the VA in the pmap
1152 * and allocate it if necessary. Return a held vm_page_t for the page.
1154 * Only used with user pmaps.
1157 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1164 * Calculate pagetable page index
1166 ptepindex = va >> PDRSHIFT;
1169 * Get the page directory entry
1171 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1174 * This supports switching from a 4MB page to a
1177 if (ptepa & VPTE_PS) {
1178 pmap->pm_pdir[ptepindex] = 0;
1185 * If the page table page is mapped, we just increment the
1186 * hold count, and activate it.
1190 * In order to get the page table page, try the
1193 if (pmap->pm_ptphint &&
1194 (pmap->pm_ptphint->pindex == ptepindex)) {
1195 m = pmap->pm_ptphint;
1197 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1198 pmap->pm_ptphint = m;
1204 * Here if the pte page isn't mapped, or if it has been deallocated.
1206 return _pmap_allocpte(pmap, ptepindex);
1209 /************************************************************************
1210 * Managed pages in pmaps *
1211 ************************************************************************
1213 * All pages entered into user pmaps and some pages entered into the kernel
1214 * pmap are managed, meaning that pmap_protect() and other related management
1215 * functions work on these pages.
1219 * free the pv_entry back to the free list. This function may be
1220 * called from an interrupt.
1222 static __inline void
1223 free_pv_entry(pv_entry_t pv)
1230 * get a new pv_entry, allocating a block from the system
1231 * when needed. This function may be called from an interrupt.
1237 if (pv_entry_high_water &&
1238 (pv_entry_count > pv_entry_high_water) &&
1239 (pmap_pagedaemon_waken == 0)) {
1240 pmap_pagedaemon_waken = 1;
1241 wakeup (&vm_pages_needed);
1243 return zalloc(&pvzone);
1247 * This routine is very drastic, but can save the system
1255 static int warningdone=0;
1257 if (pmap_pagedaemon_waken == 0)
1260 if (warningdone < 5) {
1261 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1265 for(i = 0; i < vm_page_array_size; i++) {
1266 m = &vm_page_array[i];
1267 if (m->wire_count || m->hold_count || m->busy ||
1268 (m->flags & PG_BUSY))
1272 pmap_pagedaemon_waken = 0;
1276 * If it is the first entry on the list, it is actually
1277 * in the header and we must copy the following entry up
1278 * to the header. Otherwise we must search the list for
1279 * the entry. In either case we free the now unused entry.
1282 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1283 vm_offset_t va, pmap_inval_info_t info)
1289 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1290 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1291 if (pmap == pv->pv_pmap && va == pv->pv_va)
1295 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1296 if (va == pv->pv_va)
1302 * Note that pv_ptem is NULL if the page table page itself is not
1303 * managed, even if the page being removed IS managed.
1307 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1308 m->md.pv_list_count--;
1309 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1310 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1311 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1312 ++pmap->pm_generation;
1313 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1321 * Create a pv entry for page at pa for (pmap, va). If the page table page
1322 * holding the VA is managed, mpte will be non-NULL.
1325 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1330 pv = get_pv_entry();
1335 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1336 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1337 m->md.pv_list_count++;
1343 * pmap_remove_pte: do the things to unmap a page in a process
1346 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va,
1347 pmap_inval_info_t info)
1352 oldpte = loadandclear(ptq);
1353 pmap_inval_add(info, pmap, va); /* See NOTE: PMAP_INVAL_ADD */
1354 if (oldpte & VPTE_WIRED)
1355 --pmap->pm_stats.wired_count;
1356 KKASSERT(pmap->pm_stats.wired_count >= 0);
1360 * Machines that don't support invlpg, also don't support
1361 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1364 if (oldpte & VPTE_G)
1365 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1367 pmap->pm_stats.resident_count -= 1;
1368 if (oldpte & VPTE_MANAGED) {
1369 m = PHYS_TO_VM_PAGE(oldpte);
1370 if (oldpte & VPTE_M) {
1371 #if defined(PMAP_DIAGNOSTIC)
1372 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1374 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1378 if (pmap_track_modified(pmap, va))
1381 if (oldpte & VPTE_A)
1382 vm_page_flag_set(m, PG_REFERENCED);
1383 return pmap_remove_entry(pmap, m, va, info);
1385 return pmap_unuse_pt(pmap, va, NULL, info);
1394 * Remove a single page from a process address space.
1396 * This function may not be called from an interrupt if the pmap is
1400 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1405 * if there is no pte for this address, just skip it!!! Otherwise
1406 * get a local va for mappings for this pmap and remove the entry.
1408 if (*pmap_pde(pmap, va) != 0) {
1409 ptq = get_ptbase(pmap, va);
1411 pmap_remove_pte(pmap, ptq, va, info);
1419 * Remove the given range of addresses from the specified map.
1421 * It is assumed that the start and end are properly
1422 * rounded to the page size.
1424 * This function may not be called from an interrupt if the pmap is
1428 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1432 vm_offset_t ptpaddr;
1433 vm_pindex_t sindex, eindex;
1434 struct pmap_inval_info info;
1439 KKASSERT(pmap->pm_stats.resident_count >= 0);
1440 if (pmap->pm_stats.resident_count == 0)
1443 pmap_inval_init(&info);
1446 * special handling of removing one page. a very
1447 * common operation and easy to short circuit some
1450 if (((sva + PAGE_SIZE) == eva) &&
1451 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1452 pmap_remove_page(pmap, sva, &info);
1453 pmap_inval_flush(&info);
1458 * Get a local virtual address for the mappings that are being
1461 * XXX this is really messy because the kernel pmap is not relative
1464 sindex = (sva >> PAGE_SHIFT);
1465 eindex = (eva >> PAGE_SHIFT);
1467 for (; sindex < eindex; sindex = pdnxt) {
1471 * Calculate index for next page table.
1473 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1474 if (pmap->pm_stats.resident_count == 0)
1477 pdirindex = sindex / NPDEPG;
1478 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1479 pmap->pm_pdir[pdirindex] = 0;
1480 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1481 inval_ptbase_pagedir(pmap, pdirindex);
1486 * Weed out invalid mappings. Note: we assume that the page
1487 * directory table is always allocated, and in kernel virtual.
1493 * Limit our scan to either the end of the va represented
1494 * by the current page table page, or to the end of the
1495 * range being removed.
1501 * NOTE: pmap_remove_pte() can block.
1503 for (; sindex != pdnxt; sindex++) {
1506 ptbase = get_ptbase(pmap, sindex << PAGE_SHIFT);
1509 va = i386_ptob(sindex);
1510 if (pmap_remove_pte(pmap, ptbase, va, &info))
1514 pmap_inval_flush(&info);
1520 * Removes this physical page from all physical maps in which it resides.
1521 * Reflects back modify bits to the pager.
1523 * This routine may not be called from an interrupt.
1526 pmap_remove_all(vm_page_t m)
1528 struct pmap_inval_info info;
1532 #if defined(PMAP_DIAGNOSTIC)
1534 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1537 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1538 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1542 pmap_inval_init(&info);
1544 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1545 pv->pv_pmap->pm_stats.resident_count--;
1547 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1548 KKASSERT(pte != NULL);
1550 tpte = loadandclear(pte);
1551 /* See NOTE: PMAP_INVAL_ADD */
1552 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1553 if (tpte & VPTE_WIRED)
1554 --pv->pv_pmap->pm_stats.wired_count;
1555 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
1558 vm_page_flag_set(m, PG_REFERENCED);
1561 * Update the vm_page_t clean and reference bits.
1563 if (tpte & VPTE_M) {
1564 #if defined(PMAP_DIAGNOSTIC)
1565 if (pmap_nw_modified((pt_entry_t) tpte)) {
1567 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1571 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
1574 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1575 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1576 ++pv->pv_pmap->pm_generation;
1577 m->md.pv_list_count--;
1578 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1582 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1584 pmap_inval_flush(&info);
1590 * Set the physical protection on the specified range of this map
1593 * This function may not be called from an interrupt if the map is
1594 * not the kernel_pmap.
1597 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1600 vm_offset_t pdnxt, ptpaddr;
1601 vm_pindex_t sindex, eindex;
1603 pmap_inval_info info;
1608 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1609 pmap_remove(pmap, sva, eva);
1613 if (prot & VM_PROT_WRITE)
1616 pmap_inval_init(&info);
1618 ptbase = get_ptbase(pmap, sva);
1620 sindex = (sva >> PAGE_SHIFT);
1621 eindex = (eva >> PAGE_SHIFT);
1624 for (; sindex < eindex; sindex = pdnxt) {
1628 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1630 pdirindex = sindex / NPDEPG;
1631 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1632 pmap->pm_pdir[pdirindex] &= ~(VPTE_M|VPTE_W);
1633 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1634 inval_ptbase_pagedir(pmap, pdirindex);
1639 * Weed out invalid mappings. Note: we assume that the page
1640 * directory table is always allocated, and in kernel virtual.
1645 if (pdnxt > eindex) {
1649 for (; sindex != pdnxt; sindex++) {
1653 pbits = ptbase[sindex - sbase];
1655 if (pbits & VPTE_MANAGED) {
1657 if (pbits & VPTE_A) {
1658 m = PHYS_TO_VM_PAGE(pbits);
1659 vm_page_flag_set(m, PG_REFERENCED);
1662 if (pbits & VPTE_M) {
1663 if (pmap_track_modified(pmap, i386_ptob(sindex))) {
1665 m = PHYS_TO_VM_PAGE(pbits);
1674 if (pbits != ptbase[sindex - sbase]) {
1675 ptbase[sindex - sbase] = pbits;
1676 /* See NOTE: PMAP_INVAL_ADD */
1677 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1681 pmap_inval_flush(&info);
1685 * Enter a managed page into a pmap. If the page is not wired related pmap
1686 * data can be destroyed at any time for later demand-operation.
1688 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1689 * specified protection, and wire the mapping if requested.
1691 * NOTE: This routine may not lazy-evaluate or lose information. The
1692 * page must actually be inserted into the given map NOW.
1694 * NOTE: When entering a page at a KVA address, the pmap must be the
1698 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1704 vm_offset_t origpte, newpte;
1706 pmap_inval_info info;
1714 * Get the page table page. The kernel_pmap's page table pages
1715 * are preallocated and have no associated vm_page_t.
1717 if (pmap == &kernel_pmap)
1720 mpte = pmap_allocpte(pmap, va);
1722 pmap_inval_init(&info);
1723 pte = pmap_pte(pmap, va);
1726 * Page Directory table entry not valid, we need a new PT page
1727 * and pmap_allocpte() didn't give us one. Oops!
1730 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1734 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1736 opa = origpte & VPTE_FRAME;
1738 printf("pmap_enter: pmap %p va %08x pa %08x PDE %08x origpte %08x\n", pmap, va, (int)pa, pmap->pm_pdir[va >> SEG_SHIFT], origpte);
1741 if (origpte & VPTE_PS)
1742 panic("pmap_enter: attempted pmap_enter on 4MB page");
1745 * Mapping has not changed, must be protection or wiring change.
1747 if (origpte && (opa == pa)) {
1749 * Wiring change, just update stats. We don't worry about
1750 * wiring PT pages as they remain resident as long as there
1751 * are valid mappings in them. Hence, if a user page is wired,
1752 * the PT page will be also.
1754 if (wired && ((origpte & VPTE_WIRED) == 0))
1755 ++pmap->pm_stats.wired_count;
1756 else if (!wired && (origpte & VPTE_WIRED))
1757 --pmap->pm_stats.wired_count;
1758 KKASSERT(pmap->pm_stats.wired_count >= 0);
1760 #if defined(PMAP_DIAGNOSTIC)
1761 if (pmap_nw_modified((pt_entry_t) origpte)) {
1763 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1769 * Remove the extra pte reference. Note that we cannot
1770 * optimize the RO->RW case because we have adjusted the
1771 * wiring count above and may need to adjust the wiring
1778 * We might be turning off write access to the page,
1779 * so we go ahead and sense modify status.
1781 if (origpte & VPTE_MANAGED) {
1782 if ((origpte & VPTE_M) && pmap_track_modified(pmap, va)) {
1784 om = PHYS_TO_VM_PAGE(opa);
1792 * Mapping has changed, invalidate old range and fall through to
1793 * handle validating new mapping.
1797 err = pmap_remove_pte(pmap, pte, va, &info);
1799 panic("pmap_enter: pte vanished, va: 0x%x", va);
1803 * Enter on the PV list if part of our managed memory. Note that we
1804 * raise IPL while manipulating pv_table since pmap_enter can be
1805 * called at interrupt time.
1807 if (pmap_initialized &&
1808 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1809 pmap_insert_entry(pmap, va, mpte, m);
1814 * Increment counters
1816 pmap->pm_stats.resident_count++;
1818 pmap->pm_stats.wired_count++;
1822 * Now validate mapping with desired protection/wiring.
1824 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1827 newpte |= VPTE_WIRED;
1831 * if the mapping or permission bits are different, we need
1832 * to update the pte.
1834 if ((origpte & ~(VPTE_M|VPTE_A)) != newpte) {
1835 *pte = newpte | VPTE_A;
1836 /* See NOTE: PMAP_INVAL_ADD */
1837 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1839 pmap_inval_flush(&info);
1843 * This is a quick version of pmap_enter(). It is used only under the
1844 * following conditions:
1846 * (1) The pmap is not the kernel_pmap
1847 * (2) The page is not to be wired into the map
1848 * (3) The page is to mapped read-only in the pmap (initially that is)
1849 * (4) The calling procedure is responsible for flushing the TLB
1850 * (5) The page is always managed
1851 * (6) There is no prior mapping at the VA
1855 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
1859 pmap_inval_info info;
1863 KKASSERT(pmap != &kernel_pmap);
1864 pmap_inval_init(&info);
1866 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1869 * Instantiate the page table page if required
1873 * Calculate pagetable page index
1875 ptepindex = va >> PDRSHIFT;
1876 if (mpte && (mpte->pindex == ptepindex)) {
1881 * Get the page directory entry
1883 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1886 * If the page table page is mapped, we just increment
1887 * the hold count, and activate it.
1890 if (ptepa & VPTE_PS)
1891 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1892 if (pmap->pm_ptphint &&
1893 (pmap->pm_ptphint->pindex == ptepindex)) {
1894 mpte = pmap->pm_ptphint;
1896 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1897 pmap->pm_ptphint = mpte;
1903 mpte = _pmap_allocpte(pmap, ptepindex);
1908 * Ok, now that the page table page has been validated, get the pte.
1909 * If the pte is already mapped undo mpte's hold_count and
1912 pte = pmap_pte(pmap, va);
1915 pmap_unwire_pte_hold(pmap, mpte, &info);
1920 * Enter on the PV list if part of our managed memory. Note that we
1921 * raise IPL while manipulating pv_table since pmap_enter can be
1922 * called at interrupt time.
1924 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
1925 pmap_insert_entry(pmap, va, mpte, m);
1928 * Increment counters
1930 pmap->pm_stats.resident_count++;
1932 pa = VM_PAGE_TO_PHYS(m);
1935 * Now validate mapping with RO protection
1937 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1938 *pte = pa | VPTE_V | VPTE_U;
1940 *pte = pa | VPTE_V | VPTE_U | VPTE_MANAGED;
1946 * Extract the physical address for the translation at the specified
1947 * virtual address in the pmap.
1950 pmap_extract(pmap_t pmap, vm_offset_t va)
1955 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
1956 if (pte & VPTE_PS) {
1957 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
1958 rtval |= va & SEG_MASK;
1960 pte = *get_ptbase(pmap, va);
1961 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
1968 #define MAX_INIT_PT (96)
1971 * This routine preloads the ptes for a given object into the specified pmap.
1972 * This eliminates the blast of soft faults on process startup and
1973 * immediately after an mmap.
1975 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
1978 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
1979 vm_object_t object, vm_pindex_t pindex,
1980 vm_size_t size, int limit)
1982 struct rb_vm_page_scan_info info;
1986 * We can't preinit if read access isn't set or there is no pmap
1989 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
1993 * We can't preinit if the pmap is not the current pmap
1995 if (curproc == NULL || pmap != vmspace_pmap(curproc->p_vmspace))
1998 psize = size >> PAGE_SHIFT;
2000 if ((object->type != OBJT_VNODE) ||
2001 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2002 (object->resident_page_count > MAX_INIT_PT))) {
2006 if (psize + pindex > object->size) {
2007 if (object->size < pindex)
2009 psize = object->size - pindex;
2016 * Use a red-black scan to traverse the requested range and load
2017 * any valid pages found into the pmap.
2019 * We cannot safely scan the object's memq unless we are in a
2020 * critical section since interrupts can remove pages from objects.
2022 info.start_pindex = pindex;
2023 info.end_pindex = pindex + psize - 1;
2030 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2031 pmap_object_init_pt_callback, &info);
2037 pmap_object_init_pt_callback(vm_page_t p, void *data)
2039 struct rb_vm_page_scan_info *info = data;
2040 vm_pindex_t rel_index;
2042 * don't allow an madvise to blow away our really
2043 * free pages allocating pv entries.
2045 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2046 vmstats.v_free_count < vmstats.v_free_reserved) {
2049 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2050 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2051 if ((p->queue - p->pc) == PQ_CACHE)
2052 vm_page_deactivate(p);
2054 rel_index = p->pindex - info->start_pindex;
2055 info->mpte = pmap_enter_quick(info->pmap,
2056 info->addr + i386_ptob(rel_index),
2058 vm_page_flag_set(p, PG_MAPPED);
2065 * pmap_prefault provides a quick way of clustering pagefaults into a
2066 * processes address space. It is a "cousin" of pmap_object_init_pt,
2067 * except it runs at page fault time instead of mmap time.
2071 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2073 static int pmap_prefault_pageorder[] = {
2074 -PAGE_SIZE, PAGE_SIZE,
2075 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2076 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2077 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2081 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2091 * We do not currently prefault mappings that use virtual page
2092 * tables. We do not prefault foreign pmaps.
2094 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2096 if (curproc == NULL || (pmap != vmspace_pmap(curproc->p_vmspace)))
2099 object = entry->object.vm_object;
2101 starta = addra - PFBAK * PAGE_SIZE;
2102 if (starta < entry->start)
2103 starta = entry->start;
2104 else if (starta > addra)
2108 * critical section protection is required to maintain the
2109 * page/object association, interrupts can free pages and remove
2110 * them from their objects.
2114 for (i = 0; i < PAGEORDER_SIZE; i++) {
2115 vm_object_t lobject;
2118 addr = addra + pmap_prefault_pageorder[i];
2119 if (addr > addra + (PFFOR * PAGE_SIZE))
2122 if (addr < starta || addr >= entry->end)
2126 * Make sure the page table page already exists
2128 if ((*pmap_pde(pmap, addr)) == NULL)
2132 * Get a pointer to the pte and make sure that no valid page
2135 pte = get_ptbase(pmap, addr);
2140 * Get the page to be mapped
2142 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2145 for (m = vm_page_lookup(lobject, pindex);
2146 (!m && (lobject->type == OBJT_DEFAULT) &&
2147 (lobject->backing_object));
2148 lobject = lobject->backing_object
2150 if (lobject->backing_object_offset & PAGE_MASK)
2152 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2153 m = vm_page_lookup(lobject->backing_object, pindex);
2157 * give-up when a page is not in memory
2163 * If everything meets the requirements for pmap_enter_quick(),
2164 * then enter the page.
2167 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2169 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2171 if ((m->queue - m->pc) == PQ_CACHE) {
2172 vm_page_deactivate(m);
2175 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2176 vm_page_flag_set(m, PG_MAPPED);
2184 * Routine: pmap_change_wiring
2185 * Function: Change the wiring attribute for a map/virtual-address
2187 * In/out conditions:
2188 * The mapping must already exist in the pmap.
2191 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2198 pte = get_ptbase(pmap, va);
2200 if (wired && (*pte & VPTE_WIRED) == 0)
2201 ++pmap->pm_stats.wired_count;
2202 else if (!wired && (*pte & VPTE_WIRED))
2203 --pmap->pm_stats.wired_count;
2204 KKASSERT(pmap->pm_stats.wired_count >= 0);
2207 * Wiring is not a hardware characteristic so there is no need to
2208 * invalidate TLB. However, in an SMP environment we must use
2209 * a locked bus cycle to update the pte (if we are not using
2210 * the pmap_inval_*() API that is)... it's ok to do this for simple
2214 atomic_set_int(pte, VPTE_WIRED);
2216 atomic_clear_int(pte, VPTE_WIRED);
2220 * Copy the range specified by src_addr/len
2221 * from the source map to the range dst_addr/len
2222 * in the destination map.
2224 * This routine is only advisory and need not do anything.
2227 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2228 vm_size_t len, vm_offset_t src_addr)
2230 pmap_inval_info info;
2232 vm_offset_t end_addr = src_addr + len;
2238 if (dst_addr != src_addr)
2240 if (dst_pmap->pm_pdir == NULL)
2242 if (src_pmap->pm_pdir == NULL)
2245 src_frame = get_ptbase1(src_pmap, src_addr);
2246 dst_frame = get_ptbase2(dst_pmap, src_addr);
2248 pmap_inval_init(&info);
2251 pmap_inval_add(&info, dst_pmap, -1);
2252 pmap_inval_add(&info, src_pmap, -1);
2256 * critical section protection is required to maintain the page/object
2257 * association, interrupts can free pages and remove them from
2261 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2262 vpte_t *src_pte, *dst_pte;
2263 vm_page_t dstmpte, srcmpte;
2264 vm_offset_t srcptepaddr;
2267 if (addr >= VM_MAX_USER_ADDRESS)
2268 panic("pmap_copy: invalid to pmap_copy page tables\n");
2271 * Don't let optional prefaulting of pages make us go
2272 * way below the low water mark of free pages or way
2273 * above high water mark of used pv entries.
2275 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2276 pv_entry_count > pv_entry_high_water)
2279 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2280 ptepindex = addr >> PDRSHIFT;
2282 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2283 if (srcptepaddr == 0)
2286 if (srcptepaddr & VPTE_PS) {
2287 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2288 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2289 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2294 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2295 if ((srcmpte == NULL) ||
2296 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2299 if (pdnxt > end_addr)
2302 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2303 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2304 while (addr < pdnxt) {
2308 * we only virtual copy managed pages
2310 if ((ptetemp & VPTE_MANAGED) != 0) {
2312 * We have to check after allocpte for the
2313 * pte still being around... allocpte can
2316 dstmpte = pmap_allocpte(dst_pmap, addr);
2317 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2319 * Clear the modified and accessed
2320 * (referenced) bits during the copy.
2322 * We do not have to clear the write
2323 * bit to force a fault-on-modify
2324 * because the real kernel's target
2325 * pmap is empty and will fault anyway.
2327 m = PHYS_TO_VM_PAGE(ptetemp);
2328 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2329 dst_pmap->pm_stats.resident_count++;
2330 pmap_insert_entry(dst_pmap, addr,
2333 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2335 if (dstmpte->hold_count >= srcmpte->hold_count)
2344 pmap_inval_flush(&info);
2350 * Zero the specified PA by mapping the page into KVM and clearing its
2353 * This function may be called from an interrupt and no locking is
2357 pmap_zero_page(vm_paddr_t phys)
2359 struct mdglobaldata *gd = mdcpu;
2363 panic("pmap_zero_page: CMAP3 busy");
2364 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2365 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2367 bzero(gd->gd_CADDR3, PAGE_SIZE);
2373 * pmap_page_assertzero:
2375 * Assert that a page is empty, panic if it isn't.
2378 pmap_page_assertzero(vm_paddr_t phys)
2380 struct mdglobaldata *gd = mdcpu;
2385 panic("pmap_zero_page: CMAP3 busy");
2386 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2387 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2388 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2389 for (i = 0; i < PAGE_SIZE; i += 4) {
2390 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2391 panic("pmap_page_assertzero() @ %p not zero!\n",
2392 (void *)gd->gd_CADDR3);
2402 * Zero part of a physical page by mapping it into memory and clearing
2403 * its contents with bzero.
2405 * off and size may not cover an area beyond a single hardware page.
2408 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2410 struct mdglobaldata *gd = mdcpu;
2414 panic("pmap_zero_page: CMAP3 busy");
2415 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2416 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2417 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2419 bzero((char *)gd->gd_CADDR3 + off, size);
2427 * Copy the physical page from the source PA to the target PA.
2428 * This function may be called from an interrupt. No locking
2432 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2434 struct mdglobaldata *gd = mdcpu;
2437 if (*(int *) gd->gd_CMAP1)
2438 panic("pmap_copy_page: CMAP1 busy");
2439 if (*(int *) gd->gd_CMAP2)
2440 panic("pmap_copy_page: CMAP2 busy");
2442 *(int *) gd->gd_CMAP1 = VPTE_V | VPTE_R | (src & PG_FRAME) | VPTE_A;
2443 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2445 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2446 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2448 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2450 *(int *) gd->gd_CMAP1 = 0;
2451 *(int *) gd->gd_CMAP2 = 0;
2456 * pmap_copy_page_frag:
2458 * Copy the physical page from the source PA to the target PA.
2459 * This function may be called from an interrupt. No locking
2463 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2465 struct mdglobaldata *gd = mdcpu;
2468 if (*(int *) gd->gd_CMAP1)
2469 panic("pmap_copy_page: CMAP1 busy");
2470 if (*(int *) gd->gd_CMAP2)
2471 panic("pmap_copy_page: CMAP2 busy");
2473 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2474 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2476 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2477 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2479 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2480 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2483 *(int *) gd->gd_CMAP1 = 0;
2484 *(int *) gd->gd_CMAP2 = 0;
2489 * Returns true if the pmap's pv is one of the first
2490 * 16 pvs linked to from this page. This count may
2491 * be changed upwards or downwards in the future; it
2492 * is only necessary that true be returned for a small
2493 * subset of pmaps for proper page aging.
2496 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2501 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2506 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2507 if (pv->pv_pmap == pmap) {
2520 * Remove all pages from specified address space
2521 * this aids process exit speeds. Also, this code
2522 * is special cased for current process only, but
2523 * can have the more generic (and slightly slower)
2524 * mode enabled. This is much faster than pmap_remove
2525 * in the case of running down an entire address space.
2528 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2533 pmap_inval_info info;
2535 int32_t save_generation;
2537 if (curproc && pmap == vmspace_pmap(curproc->p_vmspace))
2542 pmap_inval_init(&info);
2544 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2545 if (pv->pv_va >= eva || pv->pv_va < sva) {
2546 npv = TAILQ_NEXT(pv, pv_plist);
2550 KKASSERT(pmap == pv->pv_pmap);
2552 pte = pmap_pte(pmap, pv->pv_va);
2556 * We cannot remove wired pages from a process' mapping
2559 if (tpte & VPTE_WIRED) {
2560 npv = TAILQ_NEXT(pv, pv_plist);
2564 /* See NOTE: PMAP_INVAL_ADD */
2565 pmap_inval_add(&info, pmap, pv->pv_va);
2567 m = PHYS_TO_VM_PAGE(tpte);
2569 KASSERT(m < &vm_page_array[vm_page_array_size],
2570 ("pmap_remove_pages: bad tpte %x", tpte));
2572 pmap->pm_stats.resident_count--;
2575 * Update the vm_page_t clean and reference bits.
2577 if (tpte & VPTE_M) {
2581 npv = TAILQ_NEXT(pv, pv_plist);
2582 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2583 save_generation = ++pmap->pm_generation;
2585 m->md.pv_list_count--;
2586 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2587 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2588 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2591 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2595 * Restart the scan if we blocked during the unuse or free
2596 * calls and other removals were made.
2598 if (save_generation != pmap->pm_generation) {
2599 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2600 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2603 pmap_inval_flush(&info);
2608 * pmap_testbit tests bits in pte's
2609 * note that the testbit/changebit routines are inline,
2610 * and a lot of things compile-time evaluate.
2613 pmap_testbit(vm_page_t m, int bit)
2618 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2621 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2626 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2628 * if the bit being tested is the modified bit, then
2629 * mark clean_map and ptes as never
2632 if (bit & (VPTE_A|VPTE_M)) {
2633 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2637 #if defined(PMAP_DIAGNOSTIC)
2639 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2643 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2654 * This routine is used to clear bits in ptes. Certain bits require special
2655 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2657 static __inline void
2658 pmap_clearbit(vm_page_t m, int bit)
2660 struct pmap_inval_info info;
2665 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2668 pmap_inval_init(&info);
2672 * Loop over all current mappings setting/clearing as appropos If
2673 * setting RO do we need to clear the VAC?
2675 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2677 * don't write protect pager mappings
2679 if (bit == VPTE_W) {
2680 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2684 #if defined(PMAP_DIAGNOSTIC)
2686 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2692 * Careful here. We can use a locked bus instruction to
2693 * clear VPTE_A or VPTE_M safely but we need to synchronize
2694 * with the target cpus when we mess with VPTE_W.
2696 * On virtual kernels we must force a new fault-on-write
2697 * in the real kernel if we clear the Modify bit ourselves,
2698 * otherwise the real kernel will not get a new fault and
2699 * will never set our Modify bit again.
2701 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2702 if (bit & (VPTE_W|VPTE_M))
2703 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2707 if (bit == VPTE_W) {
2708 if (pbits & VPTE_M) {
2711 atomic_clear_int(pte, VPTE_M|VPTE_W);
2712 } else if (bit == VPTE_M) {
2714 * We do not have to make the page read-only
2715 * when clearing the Modify bit. The real
2716 * kernel will make the real PTE read-only
2717 * or otherwise detect the write and set
2718 * our VPTE_M again simply by us invalidating
2719 * the real kernel VA for the pmap (as we did
2720 * above). This allows the real kernel to
2721 * handle the write fault without forwarding
2724 atomic_clear_int(pte, VPTE_M);
2726 atomic_clear_int(pte, bit);
2730 pmap_inval_flush(&info);
2735 * pmap_page_protect:
2737 * Lower the permission for all mappings to a given page.
2740 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2742 if ((prot & VM_PROT_WRITE) == 0) {
2743 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2744 pmap_clearbit(m, VPTE_W);
2752 pmap_phys_address(int ppn)
2754 return (i386_ptob(ppn));
2758 * pmap_ts_referenced:
2760 * Return a count of reference bits for a page, clearing those bits.
2761 * It is not necessary for every reference bit to be cleared, but it
2762 * is necessary that 0 only be returned when there are truly no
2763 * reference bits set.
2765 * XXX: The exact number of bits to check and clear is a matter that
2766 * should be tested and standardized at some point in the future for
2767 * optimal aging of shared pages.
2770 pmap_ts_referenced(vm_page_t m)
2772 pv_entry_t pv, pvf, pvn;
2776 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2781 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2786 pvn = TAILQ_NEXT(pv, pv_list);
2788 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2790 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2792 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2795 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2797 if (pte && (*pte & VPTE_A)) {
2799 atomic_clear_int(pte, VPTE_A);
2801 atomic_clear_int_nonlocked(pte, VPTE_A);
2808 } while ((pv = pvn) != NULL && pv != pvf);
2818 * Return whether or not the specified physical page was modified
2819 * in any physical maps.
2822 pmap_is_modified(vm_page_t m)
2824 return pmap_testbit(m, VPTE_M);
2828 * Clear the modify bits on the specified physical page.
2831 pmap_clear_modify(vm_page_t m)
2833 pmap_clearbit(m, VPTE_M);
2837 * pmap_clear_reference:
2839 * Clear the reference bit on the specified physical page.
2842 pmap_clear_reference(vm_page_t m)
2844 pmap_clearbit(m, VPTE_A);
2848 * Miscellaneous support routines follow
2852 i386_protection_init(void)
2856 kp = protection_codes;
2857 for (prot = 0; prot < 8; prot++) {
2858 if (prot & VM_PROT_READ)
2860 if (prot & VM_PROT_WRITE)
2862 if (prot & VM_PROT_EXECUTE)
2869 * Map a set of physical memory pages into the kernel virtual
2870 * address space. Return a pointer to where it is mapped. This
2871 * routine is intended to be used for mapping device memory,
2874 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2878 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2880 vm_offset_t va, tmpva, offset;
2883 offset = pa & PAGE_MASK;
2884 size = roundup(offset + size, PAGE_SIZE);
2886 va = kmem_alloc_nofault(&kernel_map, size);
2888 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2890 pa = pa & VPTE_FRAME;
2891 for (tmpva = va; size > 0;) {
2892 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2893 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2901 return ((void *)(va + offset));
2905 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2907 vm_offset_t base, offset;
2909 base = va & VPTE_FRAME;
2910 offset = va & PAGE_MASK;
2911 size = roundup(offset + size, PAGE_SIZE);
2912 pmap_qremove(va, size >> PAGE_SHIFT);
2913 kmem_free(&kernel_map, base, size);
2917 * perform the pmap work for mincore
2920 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2926 ptep = pmap_pte(pmap, addr);
2931 if ((pte = *ptep) != 0) {
2934 val = MINCORE_INCORE;
2935 if ((pte & VPTE_MANAGED) == 0)
2938 pa = pte & VPTE_FRAME;
2940 m = PHYS_TO_VM_PAGE(pa);
2946 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2948 * Modified by someone
2950 else if (m->dirty || pmap_is_modified(m))
2951 val |= MINCORE_MODIFIED_OTHER;
2956 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2959 * Referenced by someone
2961 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
2962 val |= MINCORE_REFERENCED_OTHER;
2963 vm_page_flag_set(m, PG_REFERENCED);
2970 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
2973 struct vmspace *oldvm;
2975 oldvm = p->p_vmspace;
2976 if (oldvm != newvm) {
2978 p->p_vmspace = newvm;
2980 pmap = vmspace_pmap(newvm);
2982 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2984 pmap->pm_active |= 1;
2986 #if defined(SWTCH_OPTIM_STATS)
2990 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
2991 load_cr3(curthread->td_pcb->pcb_cr3);
2993 pmap = vmspace_pmap(oldvm);
2995 atomic_clear_int(&pmap->pm_active,
2996 1 << mycpu->gd_cpuid);
2998 pmap->pm_active &= ~1;
3002 sysref_get(&newvm->vm_sysref);
3003 sysref_put(&oldvm->vm_sysref);
3010 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3013 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3017 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3024 static void pads (pmap_t pm);
3025 void pmap_pvdump (vm_paddr_t pa);
3027 /* print address space of pmap*/
3035 if (pm == &kernel_pmap)
3037 for (i = 0; i < 1024; i++)
3039 for (j = 0; j < 1024; j++) {
3040 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3041 if (pm == &kernel_pmap && va < KERNBASE)
3043 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3045 ptep = pmap_pte(pm, va);
3046 if (ptep && (*ptep & VPTE_V)) {
3048 (void *)va, (unsigned)*ptep);
3055 pmap_pvdump(vm_paddr_t pa)
3060 kprintf("pa %08llx", (long long)pa);
3061 m = PHYS_TO_VM_PAGE(pa);
3062 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3064 kprintf(" -> pmap %p, va %x, flags %x",
3065 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3067 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);