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 1.20 2007/02/26 21:41:08 corecode Exp $
44 * NOTE: PMAP_INVAL_ADD: In pc32 this function is called prior to adjusting
45 * the PTE in the page table, because a cpu synchronization might be required.
46 * The actual invalidation is delayed until the following call or flush. In
47 * the VKERNEL build this function is called prior to adjusting the PTE and
48 * invalidates the table synchronously (not delayed), and is not SMP safe
52 #include <sys/types.h>
53 #include <sys/systm.h>
54 #include <sys/kernel.h>
57 #include <sys/vkernel.h>
59 #include <sys/thread.h>
61 #include <sys/vmspace.h>
64 #include <vm/vm_page.h>
65 #include <vm/vm_extern.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_zone.h>
69 #include <vm/vm_pageout.h>
71 #include <machine/md_var.h>
72 #include <machine/pcb.h>
73 #include <machine/pmap_inval.h>
74 #include <machine/globaldata.h>
78 struct pmap kernel_pmap;
80 static struct vm_zone pvzone;
81 static struct vm_object pvzone_obj;
82 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
83 static int pv_entry_count;
84 static int pv_entry_max;
85 static int pv_entry_high_water;
86 static int pmap_pagedaemon_waken;
87 static boolean_t pmap_initialized = FALSE;
88 static int protection_codes[8];
90 static void i386_protection_init(void);
91 static void pmap_remove_all(vm_page_t m);
92 static int pmap_release_free_page(struct pmap *pmap, vm_page_t p);
95 #ifndef PMAP_SHPGPERPROC
96 #define PMAP_SHPGPERPROC 200
99 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
101 #define pte_prot(m, p) \
102 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
108 struct pv_entry *pvinit;
110 for (i = 0; i < vm_page_array_size; i++) {
113 m = &vm_page_array[i];
114 TAILQ_INIT(&m->md.pv_list);
115 m->md.pv_list_count = 0;
118 i = vm_page_array_size;
121 pvinit = (struct pv_entry *)kmem_alloc(&kernel_map, i*sizeof(*pvinit));
122 zbootinit(&pvzone, "PV ENTRY", sizeof(*pvinit), pvinit, i);
123 pmap_initialized = TRUE;
129 int shpgperproc = PMAP_SHPGPERPROC;
131 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
132 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
133 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
134 pv_entry_high_water = 9 * (pv_entry_max / 10);
135 zinitna(&pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
139 * Bootstrap the kernel_pmap so it can be used with pmap_enter().
141 * NOTE! pm_pdir for the kernel pmap is offset so VA's translate
142 * directly into PTD indexes (PTA is also offset for the same reason).
143 * This is necessary because, for now, KVA is not mapped at address 0.
145 * Page table pages are not managed like they are in normal pmaps, so
146 * no pteobj is needed.
151 vm_pindex_t i = (vm_offset_t)KernelPTD >> PAGE_SHIFT;
153 kernel_pmap.pm_pdir = KernelPTD - (KvaStart >> SEG_SHIFT);
154 kernel_pmap.pm_pdirpte = KernelPTA[i];
155 kernel_pmap.pm_count = 1;
156 kernel_pmap.pm_active = (cpumask_t)-1;
157 TAILQ_INIT(&kernel_pmap.pm_pvlist);
158 i386_protection_init();
162 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
163 * just dummy it up so it works well enough for fork().
165 * In DragonFly, process pmaps may only be used to manipulate user address
166 * space, never kernel address space.
169 pmap_pinit0(struct pmap *pmap)
174 /************************************************************************
175 * Procedures to manage whole physical maps *
176 ************************************************************************
178 * Initialize a preallocated and zeroed pmap structure,
179 * such as one in a vmspace structure.
182 pmap_pinit(struct pmap *pmap)
188 * No need to allocate page table space yet but we do need a valid
189 * page directory table.
191 if (pmap->pm_pdir == NULL) {
193 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
197 * allocate object for the pte array and page directory
199 npages = VPTE_PAGETABLE_SIZE +
200 (VM_MAX_USER_ADDRESS / PAGE_SIZE) * sizeof(vpte_t);
201 npages = (npages + PAGE_MASK) / PAGE_SIZE;
203 if (pmap->pm_pteobj == NULL)
204 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, npages);
205 pmap->pm_pdindex = npages - 1;
208 * allocate the page directory page
210 ptdpg = vm_page_grab(pmap->pm_pteobj, pmap->pm_pdindex,
211 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
213 ptdpg->wire_count = 1;
214 ++vmstats.v_wire_count;
216 /* not usually mapped */
217 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
218 ptdpg->valid = VM_PAGE_BITS_ALL;
220 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
221 pmap->pm_pdirpte = KernelPTA[(vm_offset_t)pmap->pm_pdir >> PAGE_SHIFT];
222 if ((ptdpg->flags & PG_ZERO) == 0)
223 bzero(pmap->pm_pdir, PAGE_SIZE);
227 pmap->pm_ptphint = NULL;
228 TAILQ_INIT(&pmap->pm_pvlist);
229 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
233 * Wire in kernel global address entries. To avoid a race condition
234 * between pmap initialization and pmap_growkernel, this procedure
235 * adds the pmap to the master list (which growkernel scans to update),
236 * then copies the template.
238 * In a virtual kernel there are no kernel global address entries.
241 pmap_pinit2(struct pmap *pmap)
244 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
249 * Release all resources held by the given physical map.
251 * Should only be called if the map contains no valid mappings.
253 static int pmap_release_callback(struct vm_page *p, void *data);
256 pmap_release(struct pmap *pmap)
258 struct mdglobaldata *gd = mdcpu;
259 vm_object_t object = pmap->pm_pteobj;
260 struct rb_vm_page_scan_info info;
262 KKASSERT(pmap != &kernel_pmap);
264 #if defined(DIAGNOSTIC)
265 if (object->ref_count != 1)
266 panic("pmap_release: pteobj reference count != 1");
269 #error "Must write code to clear PTxpdir cache across all CPUs"
272 * Once we destroy the page table, the mapping becomes invalid.
273 * Rather then waste time doing a madvise
275 if (pmap->pm_pdir == gd->gd_PT1pdir) {
276 gd->gd_PT1pdir = NULL;
278 /* madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL); */
280 if (pmap->pm_pdir == gd->gd_PT2pdir) {
281 gd->gd_PT2pdir = NULL;
283 /* madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL); */
287 info.object = object;
289 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
296 info.limit = object->generation;
298 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
299 pmap_release_callback, &info);
300 if (info.error == 0 && info.mpte) {
301 if (!pmap_release_free_page(pmap, info.mpte))
305 } while (info.error);
308 * Leave the KVA reservation for pm_pdir cached for later reuse.
310 pmap->pm_pdirpte = 0;
314 pmap_release_callback(struct vm_page *p, void *data)
316 struct rb_vm_page_scan_info *info = data;
318 if (p->pindex == info->pmap->pm_pdindex) {
322 if (!pmap_release_free_page(info->pmap, p)) {
326 if (info->object->generation != info->limit) {
334 * Retire the given physical map from service. Should only be called if
335 * the map contains no valid mappings.
338 pmap_destroy(pmap_t pmap)
345 count = --pmap->pm_count;
348 panic("destroying a pmap is not yet implemented");
353 * Add a reference to the specified pmap.
356 pmap_reference(pmap_t pmap)
363 /************************************************************************
364 * VMSPACE MANAGEMENT *
365 ************************************************************************
367 * The VMSPACE management we do in our virtual kernel must be reflected
368 * in the real kernel. This is accomplished by making vmspace system
369 * calls to the real kernel.
372 cpu_vmspace_alloc(struct vmspace *vm)
377 #define LAST_EXTENT (VM_MAX_USER_ADDRESS - 0x80000000)
379 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
380 panic("vmspace_create() failed");
382 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
383 PROT_READ|PROT_WRITE,
384 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
386 if (rp == MAP_FAILED)
387 panic("vmspace_mmap: failed1");
388 vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
390 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
391 PROT_READ|PROT_WRITE,
392 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
393 MemImageFd, 0x40000000);
394 if (rp == MAP_FAILED)
395 panic("vmspace_mmap: failed2");
396 vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
398 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
399 PROT_READ|PROT_WRITE,
400 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
401 MemImageFd, 0x80000000);
402 vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
404 if (rp == MAP_FAILED)
405 panic("vmspace_mmap: failed3");
407 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
408 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
410 panic("vmspace_mcontrol: failed1");
411 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
412 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
414 panic("vmspace_mcontrol: failed2");
415 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
416 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
418 panic("vmspace_mcontrol: failed3");
422 cpu_vmspace_free(struct vmspace *vm)
424 if (vmspace_destroy(&vm->vm_pmap) < 0)
425 panic("vmspace_destroy() failed");
428 /************************************************************************
429 * Procedures which operate directly on the kernel PMAP *
430 ************************************************************************/
433 * This maps the requested page table and gives us access to it.
436 get_ptbase(struct pmap *pmap, vm_offset_t va)
438 struct mdglobaldata *gd = mdcpu;
440 if (pmap == &kernel_pmap) {
441 KKASSERT(va >= KvaStart && va < KvaEnd);
442 return(KernelPTA + (va >> PAGE_SHIFT));
443 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
444 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
445 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
446 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
450 * Otherwise choose one or the other and map the page table
451 * in the KVA space reserved for it.
453 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
454 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
456 if ((gd->gd_PTflip = 1 - gd->gd_PTflip) == 0) {
457 gd->gd_PT1pdir = pmap->pm_pdir;
458 *gd->gd_PT1pde = pmap->pm_pdirpte;
459 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
460 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
462 gd->gd_PT2pdir = pmap->pm_pdir;
463 *gd->gd_PT2pde = pmap->pm_pdirpte;
464 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
465 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
470 get_ptbase1(struct pmap *pmap, vm_offset_t va)
472 struct mdglobaldata *gd = mdcpu;
474 if (pmap == &kernel_pmap) {
475 KKASSERT(va >= KvaStart && va < KvaEnd);
476 return(KernelPTA + (va >> PAGE_SHIFT));
477 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
478 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
480 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
481 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
482 gd->gd_PT1pdir = pmap->pm_pdir;
483 *gd->gd_PT1pde = pmap->pm_pdirpte;
484 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
485 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
489 get_ptbase2(struct pmap *pmap, vm_offset_t va)
491 struct mdglobaldata *gd = mdcpu;
493 if (pmap == &kernel_pmap) {
494 KKASSERT(va >= KvaStart && va < KvaEnd);
495 return(KernelPTA + (va >> PAGE_SHIFT));
496 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
497 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
499 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
500 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
501 gd->gd_PT2pdir = pmap->pm_pdir;
502 *gd->gd_PT2pde = pmap->pm_pdirpte;
503 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
504 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
508 * When removing a page directory the related VA range in the self-mapping
509 * of the page table must be invalidated.
512 inval_ptbase_pagedir(pmap_t pmap, vm_pindex_t pindex)
514 struct mdglobaldata *gd = mdcpu;
518 #error "Must inval self-mappings in all gd's"
520 if (pmap == &kernel_pmap) {
521 va = (vm_offset_t)KernelPTA + (pindex << PAGE_SHIFT);
522 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
525 * XXX this should not strictly be needed because the page
526 * dir should alread be invalidated. test and remove
528 va = (vm_offset_t)pindex << PAGE_SHIFT;
529 vmspace_mcontrol(pmap, (void *)va, SEG_SIZE, MADV_INVAL, 0);
531 if (pmap->pm_pdir == gd->gd_PT1pdir) {
532 va = (vm_offset_t)gd->gd_PT1map + (pindex << PAGE_SHIFT);
533 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
535 if (pmap->pm_pdir == gd->gd_PT2pdir) {
536 va = (vm_offset_t)gd->gd_PT2map + (pindex << PAGE_SHIFT);
537 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
542 * Return a pointer to the page table entry for the specified va in the
543 * specified pmap. NULL is returned if there is no valid page table page
546 static __inline vpte_t *
547 pmap_pte(struct pmap *pmap, vm_offset_t va)
551 ptep = &pmap->pm_pdir[va >> SEG_SHIFT];
555 return (get_ptbase(pmap, va));
561 * Enter a mapping into kernel_pmap. Mappings created in this fashion
565 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
570 pmap_inval_info info;
573 KKASSERT(va >= KvaStart && va < KvaEnd);
574 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
575 ptep = KernelPTA + (va >> PAGE_SHIFT);
576 if (*ptep & VPTE_V) {
578 pmap_inval_init(&info);
579 pmap_inval_add(&info, &kernel_pmap, va);
583 pmap_inval_flush(&info);
585 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
593 pmap_kenter_sync(vm_offset_t va)
595 pmap_inval_info info;
597 pmap_inval_init(&info);
598 pmap_inval_add(&info, &kernel_pmap, va);
599 pmap_inval_flush(&info);
603 pmap_kenter_sync_quick(vm_offset_t va)
605 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
609 * XXX these need to be recoded. They are not used in any critical path.
612 pmap_kmodify_rw(vm_offset_t va)
614 *pmap_kpte(va) |= VPTE_R | VPTE_W;
615 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
619 pmap_kmodify_nc(vm_offset_t va)
622 *pmap_kpte(va) |= VPTE_N;
623 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
628 * Map a contiguous range of physical memory to a KVM
631 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
633 while (start < end) {
634 pmap_kenter(virt, start);
642 pmap_kpte(vm_offset_t va)
646 KKASSERT(va >= KvaStart && va < KvaEnd);
647 ptep = KernelPTA + (va >> PAGE_SHIFT);
652 * Enter a mapping into kernel_pmap without any SMP interactions.
654 * Mappings created in this fashion are not managed.
657 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
662 KKASSERT(va >= KvaStart && va < KvaEnd);
664 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
665 ptep = KernelPTA + (va >> PAGE_SHIFT);
666 if (*ptep & VPTE_V) {
668 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
675 * Make a temporary mapping for a physical address. This is only intended
676 * to be used for panic dumps.
679 pmap_kenter_temporary(vm_paddr_t pa, int i)
681 pmap_kenter(crashdumpmap + (i * PAGE_SIZE), pa);
682 return ((void *)crashdumpmap);
686 * Remove an unmanaged mapping created with pmap_kenter*().
689 pmap_kremove(vm_offset_t va)
693 pmap_inval_info info;
696 KKASSERT(va >= KvaStart && va < KvaEnd);
698 ptep = KernelPTA + (va >> PAGE_SHIFT);
699 if (*ptep & VPTE_V) {
701 pmap_inval_init(&info);
702 pmap_inval_add(&info, &kernel_pmap, va);
706 pmap_inval_flush(&info);
708 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
717 * Remove an unmanaged mapping created with pmap_kenter*() without
718 * going through any SMP interactions.
721 pmap_kremove_quick(vm_offset_t va)
725 KKASSERT(va >= KvaStart && va < KvaEnd);
727 ptep = KernelPTA + (va >> PAGE_SHIFT);
728 if (*ptep & VPTE_V) {
730 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
737 * Extract the physical address from the kernel_pmap that is associated
738 * with the specified virtual address.
741 pmap_kextract(vm_offset_t va)
746 KKASSERT(va >= KvaStart && va < KvaEnd);
748 ptep = KernelPTA + (va >> PAGE_SHIFT);
749 pa = (vm_paddr_t)(*ptep & VPTE_FRAME) | (va & PAGE_MASK);
754 * Map a set of unmanaged VM pages into KVM.
757 pmap_qenter(vm_offset_t va, struct vm_page **m, int count)
759 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
763 ptep = KernelPTA + (va >> PAGE_SHIFT);
765 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
766 *ptep = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
778 * Map a set of VM pages to kernel virtual memory. If a mapping changes
779 * clear the supplied mask. The caller handles any SMP interactions.
780 * The mask is used to provide the caller with hints on what SMP interactions
784 pmap_qenter2(vm_offset_t va, struct vm_page **m, int count, cpumask_t *mask)
786 cpumask_t cmask = mycpu->gd_cpumask;
788 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
793 ptep = KernelPTA + (va >> PAGE_SHIFT);
794 npte = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
798 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
799 } else if ((*mask & cmask) == 0) {
800 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
810 * Undo the effects of pmap_qenter*().
813 pmap_qremove(vm_offset_t va, int count)
815 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
819 ptep = KernelPTA + (va >> PAGE_SHIFT);
821 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
832 /************************************************************************
833 * Misc support glue called by machine independant code *
834 ************************************************************************
836 * These routines are called by machine independant code to operate on
837 * certain machine-dependant aspects of processes, threads, and pmaps.
841 * Initialize MD portions of the thread structure.
844 pmap_init_thread(thread_t td)
846 /* enforce pcb placement */
847 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
848 td->td_savefpu = &td->td_pcb->pcb_save;
849 td->td_sp = (char *)td->td_pcb - 16;
853 * This routine directly affects the fork perf for a process.
856 pmap_init_proc(struct proc *p)
861 * Destroy the UPAGES for a process that has exited and disassociate
862 * the process from its thread.
865 pmap_dispose_proc(struct proc *p)
867 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
871 * We pre-allocate all page table pages for kernel virtual memory so
872 * this routine will only be called if KVM has been exhausted.
875 pmap_growkernel(vm_offset_t addr)
877 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
879 if (addr > virtual_end - SEG_SIZE)
880 panic("KVM exhausted");
881 kernel_vm_end = addr;
885 * The modification bit is not tracked for any pages in this range. XXX
886 * such pages in this maps should always use pmap_k*() functions and not
889 * XXX User and kernel address spaces are independant for virtual kernels,
890 * this function only applies to the kernel pmap.
893 pmap_track_modified(pmap_t pmap, vm_offset_t va)
895 if (pmap != &kernel_pmap)
897 if ((va < clean_sva) || (va >= clean_eva))
903 /************************************************************************
904 * Procedures supporting managed page table pages *
905 ************************************************************************
907 * These procedures are used to track managed page table pages. These pages
908 * use the page table page's vm_page_t to track PTEs in the page. The
909 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
911 * This allows the system to throw away page table pages for user processes
912 * at will and reinstantiate them on demand.
916 * This routine works like vm_page_lookup() but also blocks as long as the
917 * page is busy. This routine does not busy the page it returns.
919 * Unless the caller is managing objects whos pages are in a known state,
920 * the call should be made with a critical section held so the page's object
921 * association remains valid on return.
924 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
929 m = vm_page_lookup(object, pindex);
930 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
936 * This routine unholds page table pages, and if the hold count
937 * drops to zero, then it decrements the wire count.
940 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
942 pmap_inval_flush(info);
943 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
946 if (m->hold_count == 0) {
948 * unmap the page table page
950 pmap->pm_pdir[m->pindex] = 0;
951 --pmap->pm_stats.resident_count;
952 inval_ptbase_pagedir(pmap, m->pindex);
954 if (pmap->pm_ptphint == m)
955 pmap->pm_ptphint = NULL;
958 * If the page is finally unwired, simply free it.
961 if (m->wire_count == 0) {
964 vm_page_free_zero(m);
965 --vmstats.v_wire_count;
973 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
976 if (m->hold_count == 0)
977 return _pmap_unwire_pte_hold(pmap, m, info);
983 * After removing a page table entry, this routine is used to
984 * conditionally free the page, and manage the hold/wire counts.
987 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
988 pmap_inval_info_t info)
994 * page table pages in the kernel_pmap are not managed.
996 if (pmap == &kernel_pmap)
998 ptepindex = (va >> PDRSHIFT);
999 if (pmap->pm_ptphint &&
1000 (pmap->pm_ptphint->pindex == ptepindex)) {
1001 mpte = pmap->pm_ptphint;
1003 pmap_inval_flush(info);
1004 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1005 pmap->pm_ptphint = mpte;
1008 return pmap_unwire_pte_hold(pmap, mpte, info);
1012 * Attempt to release and free an vm_page in a pmap. Returns 1 on success,
1013 * 0 on failure (if the procedure had to sleep).
1016 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1018 vpte_t *pde = pmap->pm_pdir;
1020 * This code optimizes the case of freeing non-busy
1021 * page-table pages. Those pages are zero now, and
1022 * might as well be placed directly into the zero queue.
1024 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1028 pmap->pm_stats.resident_count--;
1030 if (p->hold_count) {
1031 panic("pmap_release: freeing held page table page");
1034 * Page directory pages need to have the kernel stuff cleared, so
1035 * they can go into the zero queue also.
1037 * In virtual kernels there is no 'kernel stuff'. For the moment
1038 * I just make sure the whole thing has been zero'd even though
1039 * it should already be completely zero'd.
1041 * pmaps for vkernels do not self-map because they do not share
1042 * their address space with the vkernel. Clearing of pde[] thus
1043 * only applies to page table pages and not to the page directory
1046 if (p->pindex == pmap->pm_pdindex) {
1047 bzero(pde, VPTE_PAGETABLE_SIZE);
1048 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1054 * Clear the matching hint
1056 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1057 pmap->pm_ptphint = NULL;
1060 * And throw the page away. The page is completely zero'd out so
1061 * optimize the free call.
1064 vmstats.v_wire_count--;
1065 vm_page_free_zero(p);
1070 * This routine is called if the page table page is not mapped in the page
1073 * The routine is broken up into two parts for readability.
1076 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1082 * Find or fabricate a new pagetable page
1084 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1085 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1087 KASSERT(m->queue == PQ_NONE,
1088 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1090 if (m->wire_count == 0)
1091 vmstats.v_wire_count++;
1095 * Increment the hold count for the page table page
1096 * (denoting a new mapping.)
1101 * Map the pagetable page into the process address space, if
1102 * it isn't already there.
1104 pmap->pm_stats.resident_count++;
1106 ptepa = VM_PAGE_TO_PHYS(m);
1107 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
1111 * We are likely about to access this page table page, so set the
1112 * page table hint to reduce overhead.
1114 pmap->pm_ptphint = m;
1117 * Try to use the new mapping, but if we cannot, then
1118 * do it with the routine that maps the page explicitly.
1120 if ((m->flags & PG_ZERO) == 0)
1121 pmap_zero_page(ptepa);
1123 m->valid = VM_PAGE_BITS_ALL;
1124 vm_page_flag_clear(m, PG_ZERO);
1125 vm_page_flag_set(m, PG_MAPPED);
1132 * Determine the page table page required to access the VA in the pmap
1133 * and allocate it if necessary. Return a held vm_page_t for the page.
1135 * Only used with user pmaps.
1138 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1145 * Calculate pagetable page index
1147 ptepindex = va >> PDRSHIFT;
1150 * Get the page directory entry
1152 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1155 * This supports switching from a 4MB page to a
1158 if (ptepa & VPTE_PS) {
1159 pmap->pm_pdir[ptepindex] = 0;
1166 * If the page table page is mapped, we just increment the
1167 * hold count, and activate it.
1171 * In order to get the page table page, try the
1174 if (pmap->pm_ptphint &&
1175 (pmap->pm_ptphint->pindex == ptepindex)) {
1176 m = pmap->pm_ptphint;
1178 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1179 pmap->pm_ptphint = m;
1185 * Here if the pte page isn't mapped, or if it has been deallocated.
1187 return _pmap_allocpte(pmap, ptepindex);
1190 /************************************************************************
1191 * Managed pages in pmaps *
1192 ************************************************************************
1194 * All pages entered into user pmaps and some pages entered into the kernel
1195 * pmap are managed, meaning that pmap_protect() and other related management
1196 * functions work on these pages.
1200 * free the pv_entry back to the free list. This function may be
1201 * called from an interrupt.
1203 static __inline void
1204 free_pv_entry(pv_entry_t pv)
1211 * get a new pv_entry, allocating a block from the system
1212 * when needed. This function may be called from an interrupt.
1218 if (pv_entry_high_water &&
1219 (pv_entry_count > pv_entry_high_water) &&
1220 (pmap_pagedaemon_waken == 0)) {
1221 pmap_pagedaemon_waken = 1;
1222 wakeup (&vm_pages_needed);
1224 return zalloc(&pvzone);
1228 * This routine is very drastic, but can save the system
1236 static int warningdone=0;
1238 if (pmap_pagedaemon_waken == 0)
1241 if (warningdone < 5) {
1242 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1246 for(i = 0; i < vm_page_array_size; i++) {
1247 m = &vm_page_array[i];
1248 if (m->wire_count || m->hold_count || m->busy ||
1249 (m->flags & PG_BUSY))
1253 pmap_pagedaemon_waken = 0;
1257 * If it is the first entry on the list, it is actually
1258 * in the header and we must copy the following entry up
1259 * to the header. Otherwise we must search the list for
1260 * the entry. In either case we free the now unused entry.
1263 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1264 vm_offset_t va, pmap_inval_info_t info)
1270 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1271 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1272 if (pmap == pv->pv_pmap && va == pv->pv_va)
1276 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1277 if (va == pv->pv_va)
1283 * Note that pv_ptem is NULL if the page table page itself is not
1284 * managed, even if the page being removed IS managed.
1288 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1289 m->md.pv_list_count--;
1290 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1291 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1292 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1293 ++pmap->pm_generation;
1294 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1302 * Create a pv entry for page at pa for (pmap, va). If the page table page
1303 * holding the VA is managed, mpte will be non-NULL.
1306 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1311 pv = get_pv_entry();
1316 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1317 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1318 m->md.pv_list_count++;
1324 * pmap_remove_pte: do the things to unmap a page in a process
1327 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va,
1328 pmap_inval_info_t info)
1333 oldpte = loadandclear(ptq);
1334 pmap_inval_add(info, pmap, va); /* See NOTE: PMAP_INVAL_ADD */
1335 if (oldpte & VPTE_WIRED)
1336 --pmap->pm_stats.wired_count;
1337 KKASSERT(pmap->pm_stats.wired_count >= 0);
1341 * Machines that don't support invlpg, also don't support
1342 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1345 if (oldpte & VPTE_G)
1346 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1348 pmap->pm_stats.resident_count -= 1;
1349 if (oldpte & VPTE_MANAGED) {
1350 m = PHYS_TO_VM_PAGE(oldpte);
1351 if (oldpte & VPTE_M) {
1352 #if defined(PMAP_DIAGNOSTIC)
1353 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1355 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1359 if (pmap_track_modified(pmap, va))
1362 if (oldpte & VPTE_A)
1363 vm_page_flag_set(m, PG_REFERENCED);
1364 return pmap_remove_entry(pmap, m, va, info);
1366 return pmap_unuse_pt(pmap, va, NULL, info);
1375 * Remove a single page from a process address space.
1377 * This function may not be called from an interrupt if the pmap is
1381 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1386 * if there is no pte for this address, just skip it!!! Otherwise
1387 * get a local va for mappings for this pmap and remove the entry.
1389 if (*pmap_pde(pmap, va) != 0) {
1390 ptq = get_ptbase(pmap, va);
1392 pmap_remove_pte(pmap, ptq, va, info);
1400 * Remove the given range of addresses from the specified map.
1402 * It is assumed that the start and end are properly
1403 * rounded to the page size.
1405 * This function may not be called from an interrupt if the pmap is
1409 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1413 vm_offset_t ptpaddr;
1414 vm_pindex_t sindex, eindex;
1415 struct pmap_inval_info info;
1420 KKASSERT(pmap->pm_stats.resident_count >= 0);
1421 if (pmap->pm_stats.resident_count == 0)
1424 pmap_inval_init(&info);
1427 * special handling of removing one page. a very
1428 * common operation and easy to short circuit some
1431 if (((sva + PAGE_SIZE) == eva) &&
1432 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1433 pmap_remove_page(pmap, sva, &info);
1434 pmap_inval_flush(&info);
1439 * Get a local virtual address for the mappings that are being
1442 * XXX this is really messy because the kernel pmap is not relative
1445 sindex = (sva >> PAGE_SHIFT);
1446 eindex = (eva >> PAGE_SHIFT);
1448 for (; sindex < eindex; sindex = pdnxt) {
1452 * Calculate index for next page table.
1454 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1455 if (pmap->pm_stats.resident_count == 0)
1458 pdirindex = sindex / NPDEPG;
1459 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1460 pmap->pm_pdir[pdirindex] = 0;
1461 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1462 inval_ptbase_pagedir(pmap, pdirindex);
1467 * Weed out invalid mappings. Note: we assume that the page
1468 * directory table is always allocated, and in kernel virtual.
1474 * Limit our scan to either the end of the va represented
1475 * by the current page table page, or to the end of the
1476 * range being removed.
1482 * NOTE: pmap_remove_pte() can block.
1484 for (; sindex != pdnxt; sindex++) {
1487 ptbase = get_ptbase(pmap, sindex << PAGE_SHIFT);
1490 va = i386_ptob(sindex);
1491 if (pmap_remove_pte(pmap, ptbase, va, &info))
1495 pmap_inval_flush(&info);
1501 * Removes this physical page from all physical maps in which it resides.
1502 * Reflects back modify bits to the pager.
1504 * This routine may not be called from an interrupt.
1507 pmap_remove_all(vm_page_t m)
1509 struct pmap_inval_info info;
1513 #if defined(PMAP_DIAGNOSTIC)
1515 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1518 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1519 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1523 pmap_inval_init(&info);
1525 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1526 pv->pv_pmap->pm_stats.resident_count--;
1528 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1529 KKASSERT(pte != NULL);
1531 tpte = loadandclear(pte);
1532 /* See NOTE: PMAP_INVAL_ADD */
1533 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1534 if (tpte & VPTE_WIRED)
1535 --pv->pv_pmap->pm_stats.wired_count;
1536 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
1539 vm_page_flag_set(m, PG_REFERENCED);
1542 * Update the vm_page_t clean and reference bits.
1544 if (tpte & VPTE_M) {
1545 #if defined(PMAP_DIAGNOSTIC)
1546 if (pmap_nw_modified((pt_entry_t) tpte)) {
1548 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1552 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
1555 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1556 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1557 ++pv->pv_pmap->pm_generation;
1558 m->md.pv_list_count--;
1559 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1563 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1565 pmap_inval_flush(&info);
1571 * Set the physical protection on the specified range of this map
1574 * This function may not be called from an interrupt if the map is
1575 * not the kernel_pmap.
1578 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1581 vm_offset_t pdnxt, ptpaddr;
1582 vm_pindex_t sindex, eindex;
1584 pmap_inval_info info;
1589 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1590 pmap_remove(pmap, sva, eva);
1594 if (prot & VM_PROT_WRITE)
1597 pmap_inval_init(&info);
1599 ptbase = get_ptbase(pmap, sva);
1601 sindex = (sva >> PAGE_SHIFT);
1602 eindex = (eva >> PAGE_SHIFT);
1605 for (; sindex < eindex; sindex = pdnxt) {
1609 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1611 pdirindex = sindex / NPDEPG;
1612 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1613 pmap->pm_pdir[pdirindex] &= ~(VPTE_M|VPTE_W);
1614 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1615 inval_ptbase_pagedir(pmap, pdirindex);
1620 * Weed out invalid mappings. Note: we assume that the page
1621 * directory table is always allocated, and in kernel virtual.
1626 if (pdnxt > eindex) {
1630 for (; sindex != pdnxt; sindex++) {
1634 pbits = ptbase[sindex - sbase];
1636 if (pbits & VPTE_MANAGED) {
1638 if (pbits & VPTE_A) {
1639 m = PHYS_TO_VM_PAGE(pbits);
1640 vm_page_flag_set(m, PG_REFERENCED);
1643 if (pbits & VPTE_M) {
1644 if (pmap_track_modified(pmap, i386_ptob(sindex))) {
1646 m = PHYS_TO_VM_PAGE(pbits);
1655 if (pbits != ptbase[sindex - sbase]) {
1656 ptbase[sindex - sbase] = pbits;
1657 /* See NOTE: PMAP_INVAL_ADD */
1658 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1662 pmap_inval_flush(&info);
1666 * Enter a managed page into a pmap. If the page is not wired related pmap
1667 * data can be destroyed at any time for later demand-operation.
1669 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1670 * specified protection, and wire the mapping if requested.
1672 * NOTE: This routine may not lazy-evaluate or lose information. The
1673 * page must actually be inserted into the given map NOW.
1675 * NOTE: When entering a page at a KVA address, the pmap must be the
1679 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1685 vm_offset_t origpte, newpte;
1687 pmap_inval_info info;
1695 * Get the page table page. The kernel_pmap's page table pages
1696 * are preallocated and have no associated vm_page_t.
1698 if (pmap == &kernel_pmap)
1701 mpte = pmap_allocpte(pmap, va);
1703 pmap_inval_init(&info);
1704 pte = pmap_pte(pmap, va);
1707 * Page Directory table entry not valid, we need a new PT page
1708 * and pmap_allocpte() didn't give us one. Oops!
1711 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1715 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1717 opa = origpte & VPTE_FRAME;
1719 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);
1722 if (origpte & VPTE_PS)
1723 panic("pmap_enter: attempted pmap_enter on 4MB page");
1726 * Mapping has not changed, must be protection or wiring change.
1728 if (origpte && (opa == pa)) {
1730 * Wiring change, just update stats. We don't worry about
1731 * wiring PT pages as they remain resident as long as there
1732 * are valid mappings in them. Hence, if a user page is wired,
1733 * the PT page will be also.
1735 if (wired && ((origpte & VPTE_WIRED) == 0))
1736 ++pmap->pm_stats.wired_count;
1737 else if (!wired && (origpte & VPTE_WIRED))
1738 --pmap->pm_stats.wired_count;
1739 KKASSERT(pmap->pm_stats.wired_count >= 0);
1741 #if defined(PMAP_DIAGNOSTIC)
1742 if (pmap_nw_modified((pt_entry_t) origpte)) {
1744 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1750 * Remove the extra pte reference. Note that we cannot
1751 * optimize the RO->RW case because we have adjusted the
1752 * wiring count above and may need to adjust the wiring
1759 * We might be turning off write access to the page,
1760 * so we go ahead and sense modify status.
1762 if (origpte & VPTE_MANAGED) {
1763 if ((origpte & VPTE_M) && pmap_track_modified(pmap, va)) {
1765 om = PHYS_TO_VM_PAGE(opa);
1773 * Mapping has changed, invalidate old range and fall through to
1774 * handle validating new mapping.
1778 err = pmap_remove_pte(pmap, pte, va, &info);
1780 panic("pmap_enter: pte vanished, va: 0x%x", va);
1784 * Enter on the PV list if part of our managed memory. Note that we
1785 * raise IPL while manipulating pv_table since pmap_enter can be
1786 * called at interrupt time.
1788 if (pmap_initialized &&
1789 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1790 pmap_insert_entry(pmap, va, mpte, m);
1795 * Increment counters
1797 pmap->pm_stats.resident_count++;
1799 pmap->pm_stats.wired_count++;
1803 * Now validate mapping with desired protection/wiring.
1805 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1808 newpte |= VPTE_WIRED;
1812 * if the mapping or permission bits are different, we need
1813 * to update the pte.
1815 if ((origpte & ~(VPTE_M|VPTE_A)) != newpte) {
1816 *pte = newpte | VPTE_A;
1817 /* See NOTE: PMAP_INVAL_ADD */
1818 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1820 pmap_inval_flush(&info);
1824 * This is a quick version of pmap_enter(). It is used only under the
1825 * following conditions:
1827 * (1) The pmap is not the kernel_pmap
1828 * (2) The page is not to be wired into the map
1829 * (3) The page is to mapped read-only in the pmap (initially that is)
1830 * (4) The calling procedure is responsible for flushing the TLB
1831 * (5) The page is always managed
1832 * (6) There is no prior mapping at the VA
1836 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
1840 pmap_inval_info info;
1844 KKASSERT(pmap != &kernel_pmap);
1845 pmap_inval_init(&info);
1847 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1850 * Instantiate the page table page if required
1854 * Calculate pagetable page index
1856 ptepindex = va >> PDRSHIFT;
1857 if (mpte && (mpte->pindex == ptepindex)) {
1862 * Get the page directory entry
1864 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1867 * If the page table page is mapped, we just increment
1868 * the hold count, and activate it.
1871 if (ptepa & VPTE_PS)
1872 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1873 if (pmap->pm_ptphint &&
1874 (pmap->pm_ptphint->pindex == ptepindex)) {
1875 mpte = pmap->pm_ptphint;
1877 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1878 pmap->pm_ptphint = mpte;
1884 mpte = _pmap_allocpte(pmap, ptepindex);
1889 * Ok, now that the page table page has been validated, get the pte.
1890 * If the pte is already mapped undo mpte's hold_count and
1893 pte = pmap_pte(pmap, va);
1896 pmap_unwire_pte_hold(pmap, mpte, &info);
1901 * Enter on the PV list if part of our managed memory. Note that we
1902 * raise IPL while manipulating pv_table since pmap_enter can be
1903 * called at interrupt time.
1905 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
1906 pmap_insert_entry(pmap, va, mpte, m);
1909 * Increment counters
1911 pmap->pm_stats.resident_count++;
1913 pa = VM_PAGE_TO_PHYS(m);
1916 * Now validate mapping with RO protection
1918 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1919 *pte = pa | VPTE_V | VPTE_U;
1921 *pte = pa | VPTE_V | VPTE_U | VPTE_MANAGED;
1927 * Extract the physical address for the translation at the specified
1928 * virtual address in the pmap.
1931 pmap_extract(pmap_t pmap, vm_offset_t va)
1936 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
1937 if (pte & VPTE_PS) {
1938 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
1939 rtval |= va & SEG_MASK;
1941 pte = *get_ptbase(pmap, va);
1942 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
1949 #define MAX_INIT_PT (96)
1952 * This routine preloads the ptes for a given object into the specified pmap.
1953 * This eliminates the blast of soft faults on process startup and
1954 * immediately after an mmap.
1956 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
1959 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
1960 vm_object_t object, vm_pindex_t pindex,
1961 vm_size_t size, int limit)
1963 struct rb_vm_page_scan_info info;
1967 * We can't preinit if read access isn't set or there is no pmap
1970 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
1974 * We can't preinit if the pmap is not the current pmap
1976 if (curproc == NULL || pmap != vmspace_pmap(curproc->p_vmspace))
1979 psize = size >> PAGE_SHIFT;
1981 if ((object->type != OBJT_VNODE) ||
1982 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
1983 (object->resident_page_count > MAX_INIT_PT))) {
1987 if (psize + pindex > object->size) {
1988 if (object->size < pindex)
1990 psize = object->size - pindex;
1997 * Use a red-black scan to traverse the requested range and load
1998 * any valid pages found into the pmap.
2000 * We cannot safely scan the object's memq unless we are in a
2001 * critical section since interrupts can remove pages from objects.
2003 info.start_pindex = pindex;
2004 info.end_pindex = pindex + psize - 1;
2011 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2012 pmap_object_init_pt_callback, &info);
2018 pmap_object_init_pt_callback(vm_page_t p, void *data)
2020 struct rb_vm_page_scan_info *info = data;
2021 vm_pindex_t rel_index;
2023 * don't allow an madvise to blow away our really
2024 * free pages allocating pv entries.
2026 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2027 vmstats.v_free_count < vmstats.v_free_reserved) {
2030 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2031 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2032 if ((p->queue - p->pc) == PQ_CACHE)
2033 vm_page_deactivate(p);
2035 rel_index = p->pindex - info->start_pindex;
2036 info->mpte = pmap_enter_quick(info->pmap,
2037 info->addr + i386_ptob(rel_index),
2039 vm_page_flag_set(p, PG_MAPPED);
2046 * pmap_prefault provides a quick way of clustering pagefaults into a
2047 * processes address space. It is a "cousin" of pmap_object_init_pt,
2048 * except it runs at page fault time instead of mmap time.
2052 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2054 static int pmap_prefault_pageorder[] = {
2055 -PAGE_SIZE, PAGE_SIZE,
2056 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2057 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2058 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2062 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2072 * We do not currently prefault mappings that use virtual page
2073 * tables. We do not prefault foreign pmaps.
2075 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2077 if (curproc == NULL || (pmap != vmspace_pmap(curproc->p_vmspace)))
2080 object = entry->object.vm_object;
2082 starta = addra - PFBAK * PAGE_SIZE;
2083 if (starta < entry->start)
2084 starta = entry->start;
2085 else if (starta > addra)
2089 * critical section protection is required to maintain the
2090 * page/object association, interrupts can free pages and remove
2091 * them from their objects.
2095 for (i = 0; i < PAGEORDER_SIZE; i++) {
2096 vm_object_t lobject;
2099 addr = addra + pmap_prefault_pageorder[i];
2100 if (addr > addra + (PFFOR * PAGE_SIZE))
2103 if (addr < starta || addr >= entry->end)
2107 * Make sure the page table page already exists
2109 if ((*pmap_pde(pmap, addr)) == NULL)
2113 * Get a pointer to the pte and make sure that no valid page
2116 pte = get_ptbase(pmap, addr);
2121 * Get the page to be mapped
2123 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2126 for (m = vm_page_lookup(lobject, pindex);
2127 (!m && (lobject->type == OBJT_DEFAULT) &&
2128 (lobject->backing_object));
2129 lobject = lobject->backing_object
2131 if (lobject->backing_object_offset & PAGE_MASK)
2133 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2134 m = vm_page_lookup(lobject->backing_object, pindex);
2138 * give-up when a page is not in memory
2144 * If everything meets the requirements for pmap_enter_quick(),
2145 * then enter the page.
2148 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2150 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2152 if ((m->queue - m->pc) == PQ_CACHE) {
2153 vm_page_deactivate(m);
2156 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2157 vm_page_flag_set(m, PG_MAPPED);
2165 * Routine: pmap_change_wiring
2166 * Function: Change the wiring attribute for a map/virtual-address
2168 * In/out conditions:
2169 * The mapping must already exist in the pmap.
2172 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2179 pte = get_ptbase(pmap, va);
2181 if (wired && (*pte & VPTE_WIRED) == 0)
2182 ++pmap->pm_stats.wired_count;
2183 else if (!wired && (*pte & VPTE_WIRED))
2184 --pmap->pm_stats.wired_count;
2185 KKASSERT(pmap->pm_stats.wired_count >= 0);
2188 * Wiring is not a hardware characteristic so there is no need to
2189 * invalidate TLB. However, in an SMP environment we must use
2190 * a locked bus cycle to update the pte (if we are not using
2191 * the pmap_inval_*() API that is)... it's ok to do this for simple
2195 atomic_set_int(pte, VPTE_WIRED);
2197 atomic_clear_int(pte, VPTE_WIRED);
2201 * Copy the range specified by src_addr/len
2202 * from the source map to the range dst_addr/len
2203 * in the destination map.
2205 * This routine is only advisory and need not do anything.
2208 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2209 vm_size_t len, vm_offset_t src_addr)
2211 pmap_inval_info info;
2213 vm_offset_t end_addr = src_addr + len;
2219 if (dst_addr != src_addr)
2221 if (dst_pmap->pm_pdir == NULL)
2223 if (src_pmap->pm_pdir == NULL)
2226 src_frame = get_ptbase1(src_pmap, src_addr);
2227 dst_frame = get_ptbase2(dst_pmap, src_addr);
2229 pmap_inval_init(&info);
2232 pmap_inval_add(&info, dst_pmap, -1);
2233 pmap_inval_add(&info, src_pmap, -1);
2237 * critical section protection is required to maintain the page/object
2238 * association, interrupts can free pages and remove them from
2242 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2243 vpte_t *src_pte, *dst_pte;
2244 vm_page_t dstmpte, srcmpte;
2245 vm_offset_t srcptepaddr;
2248 if (addr >= VM_MAX_USER_ADDRESS)
2249 panic("pmap_copy: invalid to pmap_copy page tables\n");
2252 * Don't let optional prefaulting of pages make us go
2253 * way below the low water mark of free pages or way
2254 * above high water mark of used pv entries.
2256 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2257 pv_entry_count > pv_entry_high_water)
2260 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2261 ptepindex = addr >> PDRSHIFT;
2263 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2264 if (srcptepaddr == 0)
2267 if (srcptepaddr & VPTE_PS) {
2268 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2269 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2270 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2275 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2276 if ((srcmpte == NULL) ||
2277 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2280 if (pdnxt > end_addr)
2283 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2284 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2285 while (addr < pdnxt) {
2289 * we only virtual copy managed pages
2291 if ((ptetemp & VPTE_MANAGED) != 0) {
2293 * We have to check after allocpte for the
2294 * pte still being around... allocpte can
2297 dstmpte = pmap_allocpte(dst_pmap, addr);
2298 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2300 * Clear the modified and accessed
2301 * (referenced) bits during the copy.
2303 * We do not have to clear the write
2304 * bit to force a fault-on-modify
2305 * because the real kernel's target
2306 * pmap is empty and will fault anyway.
2308 m = PHYS_TO_VM_PAGE(ptetemp);
2309 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2310 dst_pmap->pm_stats.resident_count++;
2311 pmap_insert_entry(dst_pmap, addr,
2314 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2316 if (dstmpte->hold_count >= srcmpte->hold_count)
2325 pmap_inval_flush(&info);
2331 * Zero the specified PA by mapping the page into KVM and clearing its
2334 * This function may be called from an interrupt and no locking is
2338 pmap_zero_page(vm_paddr_t phys)
2340 struct mdglobaldata *gd = mdcpu;
2344 panic("pmap_zero_page: CMAP3 busy");
2345 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2346 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2348 bzero(gd->gd_CADDR3, PAGE_SIZE);
2354 * pmap_page_assertzero:
2356 * Assert that a page is empty, panic if it isn't.
2359 pmap_page_assertzero(vm_paddr_t phys)
2361 struct mdglobaldata *gd = mdcpu;
2366 panic("pmap_zero_page: CMAP3 busy");
2367 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2368 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2369 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2370 for (i = 0; i < PAGE_SIZE; i += 4) {
2371 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2372 panic("pmap_page_assertzero() @ %p not zero!\n",
2373 (void *)gd->gd_CADDR3);
2383 * Zero part of a physical page by mapping it into memory and clearing
2384 * its contents with bzero.
2386 * off and size may not cover an area beyond a single hardware page.
2389 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2391 struct mdglobaldata *gd = mdcpu;
2395 panic("pmap_zero_page: CMAP3 busy");
2396 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2397 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2398 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2400 bzero((char *)gd->gd_CADDR3 + off, size);
2408 * Copy the physical page from the source PA to the target PA.
2409 * This function may be called from an interrupt. No locking
2413 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2415 struct mdglobaldata *gd = mdcpu;
2418 if (*(int *) gd->gd_CMAP1)
2419 panic("pmap_copy_page: CMAP1 busy");
2420 if (*(int *) gd->gd_CMAP2)
2421 panic("pmap_copy_page: CMAP2 busy");
2423 *(int *) gd->gd_CMAP1 = VPTE_V | VPTE_R | (src & PG_FRAME) | VPTE_A;
2424 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2426 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2427 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2429 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2431 *(int *) gd->gd_CMAP1 = 0;
2432 *(int *) gd->gd_CMAP2 = 0;
2437 * pmap_copy_page_frag:
2439 * Copy the physical page from the source PA to the target PA.
2440 * This function may be called from an interrupt. No locking
2444 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2446 struct mdglobaldata *gd = mdcpu;
2449 if (*(int *) gd->gd_CMAP1)
2450 panic("pmap_copy_page: CMAP1 busy");
2451 if (*(int *) gd->gd_CMAP2)
2452 panic("pmap_copy_page: CMAP2 busy");
2454 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2455 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2457 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2458 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2460 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2461 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2464 *(int *) gd->gd_CMAP1 = 0;
2465 *(int *) gd->gd_CMAP2 = 0;
2470 * Returns true if the pmap's pv is one of the first
2471 * 16 pvs linked to from this page. This count may
2472 * be changed upwards or downwards in the future; it
2473 * is only necessary that true be returned for a small
2474 * subset of pmaps for proper page aging.
2477 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2482 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2487 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2488 if (pv->pv_pmap == pmap) {
2501 * Remove all pages from specified address space
2502 * this aids process exit speeds. Also, this code
2503 * is special cased for current process only, but
2504 * can have the more generic (and slightly slower)
2505 * mode enabled. This is much faster than pmap_remove
2506 * in the case of running down an entire address space.
2509 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2514 pmap_inval_info info;
2516 int32_t save_generation;
2518 if (curproc && pmap == vmspace_pmap(curproc->p_vmspace))
2523 pmap_inval_init(&info);
2525 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2526 if (pv->pv_va >= eva || pv->pv_va < sva) {
2527 npv = TAILQ_NEXT(pv, pv_plist);
2531 KKASSERT(pmap == pv->pv_pmap);
2533 pte = pmap_pte(pmap, pv->pv_va);
2537 * We cannot remove wired pages from a process' mapping
2540 if (tpte & VPTE_WIRED) {
2541 npv = TAILQ_NEXT(pv, pv_plist);
2545 /* See NOTE: PMAP_INVAL_ADD */
2546 pmap_inval_add(&info, pmap, pv->pv_va);
2548 m = PHYS_TO_VM_PAGE(tpte);
2550 KASSERT(m < &vm_page_array[vm_page_array_size],
2551 ("pmap_remove_pages: bad tpte %x", tpte));
2553 pmap->pm_stats.resident_count--;
2556 * Update the vm_page_t clean and reference bits.
2558 if (tpte & VPTE_M) {
2562 npv = TAILQ_NEXT(pv, pv_plist);
2563 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2564 save_generation = ++pmap->pm_generation;
2566 m->md.pv_list_count--;
2567 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2568 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2569 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2572 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2576 * Restart the scan if we blocked during the unuse or free
2577 * calls and other removals were made.
2579 if (save_generation != pmap->pm_generation) {
2580 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2581 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2584 pmap_inval_flush(&info);
2589 * pmap_testbit tests bits in pte's
2590 * note that the testbit/changebit routines are inline,
2591 * and a lot of things compile-time evaluate.
2594 pmap_testbit(vm_page_t m, int bit)
2599 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2602 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2607 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2609 * if the bit being tested is the modified bit, then
2610 * mark clean_map and ptes as never
2613 if (bit & (VPTE_A|VPTE_M)) {
2614 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2618 #if defined(PMAP_DIAGNOSTIC)
2620 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2624 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2635 * This routine is used to clear bits in ptes. Certain bits require special
2636 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2638 static __inline void
2639 pmap_clearbit(vm_page_t m, int bit)
2641 struct pmap_inval_info info;
2646 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2649 pmap_inval_init(&info);
2653 * Loop over all current mappings setting/clearing as appropos If
2654 * setting RO do we need to clear the VAC?
2656 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2658 * don't write protect pager mappings
2660 if (bit == VPTE_W) {
2661 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2665 #if defined(PMAP_DIAGNOSTIC)
2667 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2673 * Careful here. We can use a locked bus instruction to
2674 * clear VPTE_A or VPTE_M safely but we need to synchronize
2675 * with the target cpus when we mess with VPTE_W.
2677 * On virtual kernels we must force a new fault-on-write
2678 * in the real kernel if we clear the Modify bit ourselves,
2679 * otherwise the real kernel will not get a new fault and
2680 * will never set our Modify bit again.
2682 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2683 if (bit & (VPTE_W|VPTE_M))
2684 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2688 if (bit == VPTE_W) {
2689 if (pbits & VPTE_M) {
2692 atomic_clear_int(pte, VPTE_M|VPTE_W);
2693 } else if (bit == VPTE_M) {
2695 * We do not have to make the page read-only
2696 * when clearing the Modify bit. The real
2697 * kernel will make the real PTE read-only
2698 * or otherwise detect the write and set
2699 * our VPTE_M again simply by us invalidating
2700 * the real kernel VA for the pmap (as we did
2701 * above). This allows the real kernel to
2702 * handle the write fault without forwarding
2705 atomic_clear_int(pte, VPTE_M);
2707 atomic_clear_int(pte, bit);
2711 pmap_inval_flush(&info);
2716 * pmap_page_protect:
2718 * Lower the permission for all mappings to a given page.
2721 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2723 if ((prot & VM_PROT_WRITE) == 0) {
2724 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2725 pmap_clearbit(m, VPTE_W);
2733 pmap_phys_address(int ppn)
2735 return (i386_ptob(ppn));
2739 * pmap_ts_referenced:
2741 * Return a count of reference bits for a page, clearing those bits.
2742 * It is not necessary for every reference bit to be cleared, but it
2743 * is necessary that 0 only be returned when there are truly no
2744 * reference bits set.
2746 * XXX: The exact number of bits to check and clear is a matter that
2747 * should be tested and standardized at some point in the future for
2748 * optimal aging of shared pages.
2751 pmap_ts_referenced(vm_page_t m)
2753 pv_entry_t pv, pvf, pvn;
2757 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2762 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2767 pvn = TAILQ_NEXT(pv, pv_list);
2769 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2771 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2773 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2776 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2778 if (pte && (*pte & VPTE_A)) {
2780 atomic_clear_int(pte, VPTE_A);
2782 atomic_clear_int_nonlocked(pte, VPTE_A);
2789 } while ((pv = pvn) != NULL && pv != pvf);
2799 * Return whether or not the specified physical page was modified
2800 * in any physical maps.
2803 pmap_is_modified(vm_page_t m)
2805 return pmap_testbit(m, VPTE_M);
2809 * Clear the modify bits on the specified physical page.
2812 pmap_clear_modify(vm_page_t m)
2814 pmap_clearbit(m, VPTE_M);
2818 * pmap_clear_reference:
2820 * Clear the reference bit on the specified physical page.
2823 pmap_clear_reference(vm_page_t m)
2825 pmap_clearbit(m, VPTE_A);
2829 * Miscellaneous support routines follow
2833 i386_protection_init(void)
2837 kp = protection_codes;
2838 for (prot = 0; prot < 8; prot++) {
2839 if (prot & VM_PROT_READ)
2841 if (prot & VM_PROT_WRITE)
2843 if (prot & VM_PROT_EXECUTE)
2850 * Map a set of physical memory pages into the kernel virtual
2851 * address space. Return a pointer to where it is mapped. This
2852 * routine is intended to be used for mapping device memory,
2855 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2859 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2861 vm_offset_t va, tmpva, offset;
2864 offset = pa & PAGE_MASK;
2865 size = roundup(offset + size, PAGE_SIZE);
2867 va = kmem_alloc_nofault(&kernel_map, size);
2869 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2871 pa = pa & VPTE_FRAME;
2872 for (tmpva = va; size > 0;) {
2873 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2874 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2882 return ((void *)(va + offset));
2886 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2888 vm_offset_t base, offset;
2890 base = va & VPTE_FRAME;
2891 offset = va & PAGE_MASK;
2892 size = roundup(offset + size, PAGE_SIZE);
2893 pmap_qremove(va, size >> PAGE_SHIFT);
2894 kmem_free(&kernel_map, base, size);
2898 * perform the pmap work for mincore
2901 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2907 ptep = pmap_pte(pmap, addr);
2912 if ((pte = *ptep) != 0) {
2915 val = MINCORE_INCORE;
2916 if ((pte & VPTE_MANAGED) == 0)
2919 pa = pte & VPTE_FRAME;
2921 m = PHYS_TO_VM_PAGE(pa);
2927 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2929 * Modified by someone
2931 else if (m->dirty || pmap_is_modified(m))
2932 val |= MINCORE_MODIFIED_OTHER;
2937 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2940 * Referenced by someone
2942 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
2943 val |= MINCORE_REFERENCED_OTHER;
2944 vm_page_flag_set(m, PG_REFERENCED);
2951 pmap_activate(struct proc *p)
2955 pmap = vmspace_pmap(p->p_vmspace);
2957 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2959 pmap->pm_active |= 1;
2961 #if defined(SWTCH_OPTIM_STATS)
2965 KKASSERT((p == curproc));
2967 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
2968 load_cr3(curthread->td_pcb->pcb_cr3);
2973 pmap_deactivate(struct proc *p)
2977 pmap = vmspace_pmap(p->p_vmspace);
2979 atomic_clear_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2981 pmap->pm_active &= ~1;
2984 * XXX - note we do not adjust %cr3. The caller is expected to
2985 * activate a new pmap or do a thread-exit.
2990 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
2993 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
2997 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3004 static void pads (pmap_t pm);
3005 void pmap_pvdump (vm_paddr_t pa);
3007 /* print address space of pmap*/
3015 if (pm == &kernel_pmap)
3017 for (i = 0; i < 1024; i++)
3019 for (j = 0; j < 1024; j++) {
3020 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3021 if (pm == &kernel_pmap && va < KERNBASE)
3023 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3025 ptep = pmap_pte(pm, va);
3026 if (ptep && (*ptep & VPTE_V)) {
3028 (void *)va, (unsigned)*ptep);
3035 pmap_pvdump(vm_paddr_t pa)
3040 kprintf("pa %08llx", (long long)pa);
3041 m = PHYS_TO_VM_PAGE(pa);
3042 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3044 kprintf(" -> pmap %p, va %x, flags %x",
3045 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3047 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);