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 1.8 2007/01/08 16:03:25 dillon Exp $
44 #include <sys/types.h>
45 #include <sys/systm.h>
46 #include <sys/kernel.h>
49 #include <sys/vkernel.h>
51 #include <sys/thread.h>
53 #include <sys/vmspace.h>
56 #include <vm/vm_page.h>
57 #include <vm/vm_extern.h>
58 #include <vm/vm_kern.h>
59 #include <vm/vm_object.h>
60 #include <vm/vm_zone.h>
61 #include <vm/vm_pageout.h>
63 #include <machine/md_var.h>
64 #include <machine/pcb.h>
65 #include <machine/pmap_inval.h>
66 #include <machine/globaldata.h>
70 struct pmap kernel_pmap;
72 static struct vm_zone pvzone;
73 static struct vm_object pvzone_obj;
74 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
75 static int pv_entry_count;
76 static int pv_entry_max;
77 static int pv_entry_high_water;
78 static int pmap_pagedaemon_waken;
79 static boolean_t pmap_initialized = FALSE;
80 static int protection_codes[8];
82 static void i386_protection_init(void);
83 static void pmap_remove_all(vm_page_t m);
84 static int pmap_release_free_page(struct pmap *pmap, vm_page_t p);
87 #ifndef PMAP_SHPGPERPROC
88 #define PMAP_SHPGPERPROC 200
91 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
93 #define pte_prot(m, p) \
94 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
100 struct pv_entry *pvinit;
102 for (i = 0; i < vm_page_array_size; i++) {
105 m = &vm_page_array[i];
106 TAILQ_INIT(&m->md.pv_list);
107 m->md.pv_list_count = 0;
110 i = vm_page_array_size;
113 pvinit = (struct pv_entry *)kmem_alloc(&kernel_map, i*sizeof(*pvinit));
114 zbootinit(&pvzone, "PV ENTRY", sizeof(*pvinit), pvinit, i);
115 pmap_initialized = TRUE;
121 int shpgperproc = PMAP_SHPGPERPROC;
123 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
124 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
125 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
126 pv_entry_high_water = 9 * (pv_entry_max / 10);
127 zinitna(&pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
131 * Bootstrap the kernel_pmap so it can be used with pmap_enter().
133 * NOTE! pm_pdir for the kernel pmap is offset so VA's translate
134 * directly into PTD indexes (PTA is also offset for the same reason).
135 * This is necessary because, for now, KVA is not mapped at address 0.
137 * Page table pages are not managed like they are in normal pmaps, so
138 * no pteobj is needed.
143 vm_pindex_t i = (vm_offset_t)KernelPTD >> PAGE_SHIFT;
145 kernel_pmap.pm_pdir = KernelPTD - (KvaStart >> SEG_SHIFT);
146 kernel_pmap.pm_pdirpte = KernelPTA[i];
147 kernel_pmap.pm_count = 1;
148 kernel_pmap.pm_active = (cpumask_t)-1;
149 TAILQ_INIT(&kernel_pmap.pm_pvlist);
150 i386_protection_init();
154 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
155 * just dummy it up so it works well enough for fork().
157 * In DragonFly, process pmaps may only be used to manipulate user address
158 * space, never kernel address space.
161 pmap_pinit0(struct pmap *pmap)
166 /************************************************************************
167 * Procedures to manage whole physical maps *
168 ************************************************************************
170 * Initialize a preallocated and zeroed pmap structure,
171 * such as one in a vmspace structure.
174 pmap_pinit(struct pmap *pmap)
180 * No need to allocate page table space yet but we do need a valid
181 * page directory table.
183 if (pmap->pm_pdir == NULL) {
185 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
189 * allocate object for the pte array and page directory
191 npages = VPTE_PAGETABLE_SIZE +
192 (VM_MAX_USER_ADDRESS / PAGE_SIZE) * sizeof(vpte_t);
193 npages = (npages + PAGE_MASK) / PAGE_SIZE;
195 if (pmap->pm_pteobj == NULL)
196 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, npages);
197 pmap->pm_pdindex = npages - 1;
200 * allocate the page directory page
202 ptdpg = vm_page_grab(pmap->pm_pteobj, pmap->pm_pdindex,
203 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
205 ptdpg->wire_count = 1;
206 ++vmstats.v_wire_count;
208 /* not usually mapped */
209 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
210 ptdpg->valid = VM_PAGE_BITS_ALL;
212 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
213 pmap->pm_pdirpte = KernelPTA[(vm_offset_t)pmap->pm_pdir >> PAGE_SHIFT];
214 if ((ptdpg->flags & PG_ZERO) == 0)
215 bzero(pmap->pm_pdir, PAGE_SIZE);
219 pmap->pm_ptphint = NULL;
220 TAILQ_INIT(&pmap->pm_pvlist);
221 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
225 * Wire in kernel global address entries. To avoid a race condition
226 * between pmap initialization and pmap_growkernel, this procedure
227 * adds the pmap to the master list (which growkernel scans to update),
228 * then copies the template.
230 * In a virtual kernel there are no kernel global address entries.
233 pmap_pinit2(struct pmap *pmap)
236 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
241 * Release all resources held by the given physical map.
243 * Should only be called if the map contains no valid mappings.
245 static int pmap_release_callback(struct vm_page *p, void *data);
248 pmap_release(struct pmap *pmap)
250 struct mdglobaldata *gd = mdcpu;
251 vm_object_t object = pmap->pm_pteobj;
252 struct rb_vm_page_scan_info info;
254 KKASSERT(pmap != &kernel_pmap);
256 #if defined(DIAGNOSTIC)
257 if (object->ref_count != 1)
258 panic("pmap_release: pteobj reference count != 1");
261 #error "Must write code to clear PTxpdir cache across all CPUs"
264 * Once we destroy the page table, the mapping becomes invalid.
265 * Rather then waste time doing a madvise
267 if (pmap->pm_pdir == gd->gd_PT1pdir) {
268 gd->gd_PT1pdir = NULL;
270 /* madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL); */
272 if (pmap->pm_pdir == gd->gd_PT2pdir) {
273 gd->gd_PT2pdir = NULL;
275 /* madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL); */
279 info.object = object;
281 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
288 info.limit = object->generation;
290 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
291 pmap_release_callback, &info);
292 if (info.error == 0 && info.mpte) {
293 if (!pmap_release_free_page(pmap, info.mpte))
297 } while (info.error);
301 pmap_release_callback(struct vm_page *p, void *data)
303 struct rb_vm_page_scan_info *info = data;
305 if (p->pindex == info->pmap->pm_pdindex) {
309 if (!pmap_release_free_page(info->pmap, p)) {
313 if (info->object->generation != info->limit) {
321 * Retire the given physical map from service. Should only be called if
322 * the map contains no valid mappings.
325 pmap_destroy(pmap_t pmap)
332 count = --pmap->pm_count;
335 panic("destroying a pmap is not yet implemented");
340 * Add a reference to the specified pmap.
343 pmap_reference(pmap_t pmap)
350 /************************************************************************
351 * VMSPACE MANAGEMENT *
352 ************************************************************************
354 * The VMSPACE management we do in our virtual kernel must be reflected
355 * in the real kernel. This is accomplished by making vmspace system
356 * calls to the real kernel.
359 cpu_vmspace_alloc(struct vmspace *vm)
364 #define LAST_EXTENT (VM_MAX_USER_ADDRESS - 0x80000000)
366 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
367 panic("vmspace_create() failed");
369 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
370 PROT_READ|PROT_WRITE,
371 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
373 if (rp == MAP_FAILED)
374 panic("vmspace_mmap: failed1");
375 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
376 PROT_READ|PROT_WRITE,
377 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
378 MemImageFd, 0x40000000);
379 if (rp == MAP_FAILED)
380 panic("vmspace_mmap: failed2");
381 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
382 PROT_READ|PROT_WRITE,
383 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
384 MemImageFd, 0x80000000);
385 if (rp == MAP_FAILED)
386 panic("vmspace_mmap: failed3");
388 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
389 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
391 panic("vmspace_mcontrol: failed1");
392 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
393 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
395 panic("vmspace_mcontrol: failed2");
396 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
397 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
399 panic("vmspace_mcontrol: failed3");
403 cpu_vmspace_free(struct vmspace *vm)
405 if (vmspace_destroy(&vm->vm_pmap) < 0)
406 panic("vmspace_destroy() failed");
409 /************************************************************************
410 * Procedures which operate directly on the kernel PMAP *
411 ************************************************************************/
414 * This maps the requested page table and gives us access to it.
417 get_ptbase(struct pmap *pmap, vm_offset_t va)
419 struct mdglobaldata *gd = mdcpu;
421 if (pmap == &kernel_pmap) {
422 KKASSERT(va >= KvaStart && va < KvaEnd);
423 return(KernelPTA + (va >> PAGE_SHIFT));
424 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
425 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
426 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
427 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
431 * Otherwise choose one or the other and map the page table
432 * in the KVA space reserved for it.
434 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
435 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
437 if ((gd->gd_PTflip = 1 - gd->gd_PTflip) == 0) {
438 gd->gd_PT1pdir = pmap->pm_pdir;
439 *gd->gd_PT1pde = pmap->pm_pdirpte;
440 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
441 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
443 gd->gd_PT2pdir = pmap->pm_pdir;
444 *gd->gd_PT2pde = pmap->pm_pdirpte;
445 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
446 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
451 get_ptbase1(struct pmap *pmap, vm_offset_t va)
453 struct mdglobaldata *gd = mdcpu;
455 if (pmap == &kernel_pmap) {
456 KKASSERT(va >= KvaStart && va < KvaEnd);
457 return(KernelPTA + (va >> PAGE_SHIFT));
458 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
459 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
461 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
462 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
463 gd->gd_PT1pdir = pmap->pm_pdir;
464 *gd->gd_PT1pde = pmap->pm_pdirpte;
465 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
466 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
470 get_ptbase2(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_PT2pdir) {
478 return(gd->gd_PT2map + (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_PT2pdir = pmap->pm_pdir;
483 *gd->gd_PT2pde = pmap->pm_pdirpte;
484 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
485 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
489 * When removing a page directory the related VA range in the self-mapping
490 * of the page table must be invalidated.
493 inval_ptbase_pagedir(pmap_t pmap, vm_pindex_t pindex)
495 struct mdglobaldata *gd = mdcpu;
499 #error "Must inval self-mappings in all gd's"
501 if (pmap == &kernel_pmap) {
502 va = (vm_offset_t)KernelPTA + (pindex << PAGE_SHIFT);
503 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
505 if (pmap->pm_pdir == gd->gd_PT1pdir) {
506 va = (vm_offset_t)gd->gd_PT1map + (pindex << PAGE_SHIFT);
507 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
509 if (pmap->pm_pdir == gd->gd_PT2pdir) {
510 va = (vm_offset_t)gd->gd_PT2map + (pindex << PAGE_SHIFT);
511 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
516 * Return a pointer to the page table entry for the specified va in the
517 * specified pmap. NULL is returned if there is no valid page table page
520 static __inline vpte_t *
521 pmap_pte(struct pmap *pmap, vm_offset_t va)
525 ptep = &pmap->pm_pdir[va >> SEG_SHIFT];
529 return (get_ptbase(pmap, va));
535 * Enter a mapping into kernel_pmap. Mappings created in this fashion
539 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
544 pmap_inval_info info;
547 KKASSERT(va >= KvaStart && va < KvaEnd);
548 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
549 ptep = KernelPTA + (va >> PAGE_SHIFT);
550 if (*ptep & VPTE_V) {
552 pmap_inval_init(&info);
553 pmap_inval_add(&info, &kernel_pmap, va);
557 pmap_inval_flush(&info);
559 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
567 pmap_kenter_sync(vm_offset_t va)
569 pmap_inval_info info;
571 pmap_inval_init(&info);
572 pmap_inval_add(&info, &kernel_pmap, va);
573 pmap_inval_flush(&info);
577 pmap_kenter_sync_quick(vm_offset_t va)
579 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
583 * XXX these need to be recoded. They are not used in any critical path.
586 pmap_kmodify_rw(vm_offset_t va)
588 *pmap_kpte(va) |= VPTE_R | VPTE_W;
589 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
593 pmap_kmodify_nc(vm_offset_t va)
596 *pmap_kpte(va) |= VPTE_N;
597 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
602 * Map a contiguous range of physical memory to a KVM
605 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
607 while (start < end) {
608 pmap_kenter(virt, start);
616 pmap_kpte(vm_offset_t va)
620 KKASSERT(va >= KvaStart && va < KvaEnd);
621 ptep = KernelPTA + (va >> PAGE_SHIFT);
626 * Enter a mapping into kernel_pmap without any SMP interactions.
628 * Mappings created in this fashion are not managed.
631 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
636 KKASSERT(va >= KvaStart && va < KvaEnd);
638 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
639 ptep = KernelPTA + (va >> PAGE_SHIFT);
640 if (*ptep & VPTE_V) {
642 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
649 * Make a temporary mapping for a physical address. This is only intended
650 * to be used for panic dumps.
653 pmap_kenter_temporary(vm_paddr_t pa, int i)
655 pmap_kenter(crashdumpmap + (i * PAGE_SIZE), pa);
656 return ((void *)crashdumpmap);
660 * Remove an unmanaged mapping created with pmap_kenter*().
663 pmap_kremove(vm_offset_t va)
667 pmap_inval_info info;
670 KKASSERT(va >= KvaStart && va < KvaEnd);
672 ptep = KernelPTA + (va >> PAGE_SHIFT);
673 if (*ptep & VPTE_V) {
675 pmap_inval_init(&info);
676 pmap_inval_add(&info, &kernel_pmap, va);
680 pmap_inval_flush(&info);
682 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
691 * Remove an unmanaged mapping created with pmap_kenter*() without
692 * going through any SMP interactions.
695 pmap_kremove_quick(vm_offset_t va)
699 KKASSERT(va >= KvaStart && va < KvaEnd);
701 ptep = KernelPTA + (va >> PAGE_SHIFT);
702 if (*ptep & VPTE_V) {
704 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
711 * Extract the physical address from the kernel_pmap that is associated
712 * with the specified virtual address.
715 pmap_kextract(vm_offset_t va)
720 KKASSERT(va >= KvaStart && va < KvaEnd);
722 ptep = KernelPTA + (va >> PAGE_SHIFT);
723 pa = (vm_paddr_t)(*ptep & VPTE_FRAME) | (va & PAGE_MASK);
728 * Map a set of unmanaged VM pages into KVM.
731 pmap_qenter(vm_offset_t va, struct vm_page **m, int count)
733 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
737 ptep = KernelPTA + (va >> PAGE_SHIFT);
739 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
740 *ptep = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
752 * Map a set of VM pages to kernel virtual memory. If a mapping changes
753 * clear the supplied mask. The caller handles any SMP interactions.
754 * The mask is used to provide the caller with hints on what SMP interactions
758 pmap_qenter2(vm_offset_t va, struct vm_page **m, int count, cpumask_t *mask)
760 cpumask_t cmask = mycpu->gd_cpumask;
762 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
767 ptep = KernelPTA + (va >> PAGE_SHIFT);
768 npte = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
772 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
773 } else if ((*mask & cmask) == 0) {
774 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
784 * Undo the effects of pmap_qenter*().
787 pmap_qremove(vm_offset_t va, int count)
789 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
793 ptep = KernelPTA + (va >> PAGE_SHIFT);
795 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
806 /************************************************************************
807 * Misc support glue called by machine independant code *
808 ************************************************************************
810 * These routines are called by machine independant code to operate on
811 * certain machine-dependant aspects of processes, threads, and pmaps.
815 * Initialize MD portions of the thread structure.
818 pmap_init_thread(thread_t td)
820 /* enforce pcb placement */
821 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
822 td->td_savefpu = &td->td_pcb->pcb_save;
823 td->td_sp = (char *)td->td_pcb - 16;
827 * Initialize MD portions of a process structure. XXX this aint MD
830 pmap_init_proc(struct proc *p, struct thread *td)
832 p->p_addr = (void *)td->td_kstack;
835 td->td_lwp = &p->p_lwp;
836 td->td_switch = cpu_heavy_switch;
838 KKASSERT(td->td_mpcount == 1);
840 bzero(p->p_addr, sizeof(*p->p_addr));
844 * Destroy the UPAGES for a process that has exited and disassociate
845 * the process from its thread.
848 pmap_dispose_proc(struct proc *p)
852 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
854 if ((td = p->p_thread) != NULL) {
863 * We pre-allocate all page table pages for kernel virtual memory so
864 * this routine will only be called if KVM has been exhausted.
867 pmap_growkernel(vm_offset_t size)
869 panic("KVM exhausted");
873 * The modification bit is not tracked for any pages in this range. XXX
874 * such pages in this maps should always use pmap_k*() functions and not
878 pmap_track_modified(vm_offset_t va)
880 if ((va < clean_sva) || (va >= clean_eva))
886 /************************************************************************
887 * Procedures supporting managed page table pages *
888 ************************************************************************
890 * These procedures are used to track managed page table pages. These pages
891 * use the page table page's vm_page_t to track PTEs in the page. The
892 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
894 * This allows the system to throw away page table pages for user processes
895 * at will and reinstantiate them on demand.
899 * This routine works like vm_page_lookup() but also blocks as long as the
900 * page is busy. This routine does not busy the page it returns.
902 * Unless the caller is managing objects whos pages are in a known state,
903 * the call should be made with a critical section held so the page's object
904 * association remains valid on return.
907 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
912 m = vm_page_lookup(object, pindex);
913 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
919 * This routine unholds page table pages, and if the hold count
920 * drops to zero, then it decrements the wire count.
923 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
925 pmap_inval_flush(info);
926 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
929 if (m->hold_count == 0) {
931 * unmap the page table page
933 pmap->pm_pdir[m->pindex] = 0;
934 --pmap->pm_stats.resident_count;
935 inval_ptbase_pagedir(pmap, m->pindex);
937 if (pmap->pm_ptphint == m)
938 pmap->pm_ptphint = NULL;
941 * If the page is finally unwired, simply free it.
944 if (m->wire_count == 0) {
947 vm_page_free_zero(m);
948 --vmstats.v_wire_count;
956 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
959 if (m->hold_count == 0)
960 return _pmap_unwire_pte_hold(pmap, m, info);
966 * After removing a page table entry, this routine is used to
967 * conditionally free the page, and manage the hold/wire counts.
970 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
971 pmap_inval_info_t info)
977 * page table pages in the kernel_pmap are not managed.
979 if (pmap == &kernel_pmap)
981 ptepindex = (va >> PDRSHIFT);
982 if (pmap->pm_ptphint &&
983 (pmap->pm_ptphint->pindex == ptepindex)) {
984 mpte = pmap->pm_ptphint;
986 pmap_inval_flush(info);
987 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
988 pmap->pm_ptphint = mpte;
991 return pmap_unwire_pte_hold(pmap, mpte, info);
995 * Attempt to release and free an vm_page in a pmap. Returns 1 on success,
996 * 0 on failure (if the procedure had to sleep).
999 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1001 vpte_t *pde = pmap->pm_pdir;
1003 * This code optimizes the case of freeing non-busy
1004 * page-table pages. Those pages are zero now, and
1005 * might as well be placed directly into the zero queue.
1007 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1013 * Remove the page table page from the processes address space.
1016 pmap->pm_stats.resident_count--;
1018 if (p->hold_count) {
1019 panic("pmap_release: freeing held page table page");
1022 * Page directory pages need to have the kernel stuff cleared, so
1023 * they can go into the zero queue also.
1025 * In virtual kernels there is no 'kernel stuff'. For the moment
1026 * I just make sure the whole thing has been zero'd even though
1027 * it should already be completely zero'd.
1029 if (p->pindex == pmap->pm_pdindex) {
1030 bzero(pde, VPTE_PAGETABLE_SIZE);
1031 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1035 * Clear the matching hint
1037 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1038 pmap->pm_ptphint = NULL;
1041 * And throw the page away. The page is completely zero'd out so
1042 * optimize the free call.
1045 vmstats.v_wire_count--;
1046 vm_page_free_zero(p);
1051 * This routine is called if the page table page is not mapped in the page
1054 * The routine is broken up into two parts for readability.
1057 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1063 * Find or fabricate a new pagetable page
1065 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1066 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1068 KASSERT(m->queue == PQ_NONE,
1069 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1071 if (m->wire_count == 0)
1072 vmstats.v_wire_count++;
1076 * Increment the hold count for the page table page
1077 * (denoting a new mapping.)
1082 * Map the pagetable page into the process address space, if
1083 * it isn't already there.
1085 pmap->pm_stats.resident_count++;
1087 ptepa = VM_PAGE_TO_PHYS(m);
1088 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
1092 * We are likely about to access this page table page, so set the
1093 * page table hint to reduce overhead.
1095 pmap->pm_ptphint = m;
1098 * Try to use the new mapping, but if we cannot, then
1099 * do it with the routine that maps the page explicitly.
1101 if ((m->flags & PG_ZERO) == 0)
1102 pmap_zero_page(ptepa);
1104 m->valid = VM_PAGE_BITS_ALL;
1105 vm_page_flag_clear(m, PG_ZERO);
1106 vm_page_flag_set(m, PG_MAPPED);
1113 * Determine the page table page required to access the VA in the pmap
1114 * and allocate it if necessary. Return a held vm_page_t for the page.
1116 * Only used with user pmaps.
1119 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1126 * Calculate pagetable page index
1128 ptepindex = va >> PDRSHIFT;
1131 * Get the page directory entry
1133 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1136 * This supports switching from a 4MB page to a
1139 if (ptepa & VPTE_PS) {
1140 pmap->pm_pdir[ptepindex] = 0;
1147 * If the page table page is mapped, we just increment the
1148 * hold count, and activate it.
1152 * In order to get the page table page, try the
1155 if (pmap->pm_ptphint &&
1156 (pmap->pm_ptphint->pindex == ptepindex)) {
1157 m = pmap->pm_ptphint;
1159 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1160 pmap->pm_ptphint = m;
1166 * Here if the pte page isn't mapped, or if it has been deallocated.
1168 return _pmap_allocpte(pmap, ptepindex);
1171 /************************************************************************
1172 * Managed pages in pmaps *
1173 ************************************************************************
1175 * All pages entered into user pmaps and some pages entered into the kernel
1176 * pmap are managed, meaning that pmap_protect() and other related management
1177 * functions work on these pages.
1181 * free the pv_entry back to the free list. This function may be
1182 * called from an interrupt.
1184 static __inline void
1185 free_pv_entry(pv_entry_t pv)
1192 * get a new pv_entry, allocating a block from the system
1193 * when needed. This function may be called from an interrupt.
1199 if (pv_entry_high_water &&
1200 (pv_entry_count > pv_entry_high_water) &&
1201 (pmap_pagedaemon_waken == 0)) {
1202 pmap_pagedaemon_waken = 1;
1203 wakeup (&vm_pages_needed);
1205 return zalloc(&pvzone);
1209 * This routine is very drastic, but can save the system
1217 static int warningdone=0;
1219 if (pmap_pagedaemon_waken == 0)
1222 if (warningdone < 5) {
1223 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1227 for(i = 0; i < vm_page_array_size; i++) {
1228 m = &vm_page_array[i];
1229 if (m->wire_count || m->hold_count || m->busy ||
1230 (m->flags & PG_BUSY))
1234 pmap_pagedaemon_waken = 0;
1238 * If it is the first entry on the list, it is actually
1239 * in the header and we must copy the following entry up
1240 * to the header. Otherwise we must search the list for
1241 * the entry. In either case we free the now unused entry.
1244 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1245 vm_offset_t va, pmap_inval_info_t info)
1251 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1252 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1253 if (pmap == pv->pv_pmap && va == pv->pv_va)
1257 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1258 if (va == pv->pv_va)
1264 * Note that pv_ptem is NULL if the page table page itself is not
1265 * managed, even if the page being removed IS managed.
1269 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1270 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1271 m->md.pv_list_count--;
1272 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1273 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1274 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1282 * Create a pv entry for page at pa for (pmap, va). If the page table page
1283 * holding the VA is managed, mpte will be non-NULL.
1286 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1291 pv = get_pv_entry();
1296 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1297 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1298 m->md.pv_list_count++;
1304 * pmap_remove_pte: do the things to unmap a page in a process
1307 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va,
1308 pmap_inval_info_t info)
1313 pmap_inval_add(info, pmap, va);
1314 oldpte = loadandclear(ptq);
1315 if (oldpte & VPTE_W)
1316 pmap->pm_stats.wired_count -= 1;
1318 * Machines that don't support invlpg, also don't support
1319 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1322 if (oldpte & VPTE_G)
1323 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1324 pmap->pm_stats.resident_count -= 1;
1325 if (oldpte & PG_MANAGED) {
1326 m = PHYS_TO_VM_PAGE(oldpte);
1327 if (oldpte & VPTE_M) {
1328 #if defined(PMAP_DIAGNOSTIC)
1329 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1331 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1335 if (pmap_track_modified(va))
1338 if (oldpte & VPTE_A)
1339 vm_page_flag_set(m, PG_REFERENCED);
1340 return pmap_remove_entry(pmap, m, va, info);
1342 return pmap_unuse_pt(pmap, va, NULL, info);
1351 * Remove a single page from a process address space.
1353 * This function may not be called from an interrupt if the pmap is
1357 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1362 * if there is no pte for this address, just skip it!!! Otherwise
1363 * get a local va for mappings for this pmap and remove the entry.
1365 if (*pmap_pde(pmap, va) != 0) {
1366 ptq = get_ptbase(pmap, va);
1368 pmap_remove_pte(pmap, ptq, va, info);
1376 * Remove the given range of addresses from the specified map.
1378 * It is assumed that the start and end are properly
1379 * rounded to the page size.
1381 * This function may not be called from an interrupt if the pmap is
1385 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1389 vm_offset_t ptpaddr;
1390 vm_pindex_t sindex, eindex;
1392 struct pmap_inval_info info;
1397 if (pmap->pm_stats.resident_count == 0)
1400 pmap_inval_init(&info);
1403 * special handling of removing one page. a very
1404 * common operation and easy to short circuit some
1407 if (((sva + PAGE_SIZE) == eva) &&
1408 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1409 pmap_remove_page(pmap, sva, &info);
1410 pmap_inval_flush(&info);
1415 * Get a local virtual address for the mappings that are being
1418 * XXX this is really messy because the kernel pmap is not relative
1421 ptbase = get_ptbase(pmap, sva);
1423 sindex = (sva >> PAGE_SHIFT);
1424 eindex = (eva >> PAGE_SHIFT);
1427 for (; sindex < eindex; sindex = pdnxt) {
1431 * Calculate index for next page table.
1433 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1434 if (pmap->pm_stats.resident_count == 0)
1437 pdirindex = sindex / NPDEPG;
1438 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1439 pmap->pm_pdir[pdirindex] = 0;
1440 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1441 inval_ptbase_pagedir(pmap, pdirindex);
1446 * Weed out invalid mappings. Note: we assume that the page
1447 * directory table is always allocated, and in kernel virtual.
1453 * Limit our scan to either the end of the va represented
1454 * by the current page table page, or to the end of the
1455 * range being removed.
1457 if (pdnxt > eindex) {
1461 for (; sindex != pdnxt; sindex++) {
1463 if (ptbase[sindex - sbase] == 0)
1465 va = i386_ptob(sindex);
1466 if (pmap_remove_pte(pmap, ptbase + sindex - sbase, va, &info))
1470 pmap_inval_flush(&info);
1476 * Removes this physical page from all physical maps in which it resides.
1477 * Reflects back modify bits to the pager.
1479 * This routine may not be called from an interrupt.
1482 pmap_remove_all(vm_page_t m)
1484 struct pmap_inval_info info;
1488 #if defined(PMAP_DIAGNOSTIC)
1490 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1493 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1494 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1498 pmap_inval_init(&info);
1500 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1501 pv->pv_pmap->pm_stats.resident_count--;
1503 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1504 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1506 tpte = loadandclear(pte);
1508 pv->pv_pmap->pm_stats.wired_count--;
1511 vm_page_flag_set(m, PG_REFERENCED);
1514 * Update the vm_page_t clean and reference bits.
1516 if (tpte & VPTE_M) {
1517 #if defined(PMAP_DIAGNOSTIC)
1518 if (pmap_nw_modified((pt_entry_t) tpte)) {
1520 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1524 if (pmap_track_modified(pv->pv_va))
1527 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1528 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1529 m->md.pv_list_count--;
1530 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1534 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1536 pmap_inval_flush(&info);
1542 * Set the physical protection on the specified range of this map
1545 * This function may not be called from an interrupt if the map is
1546 * not the kernel_pmap.
1549 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1552 vm_offset_t pdnxt, ptpaddr;
1553 vm_pindex_t sindex, eindex;
1555 pmap_inval_info info;
1560 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1561 pmap_remove(pmap, sva, eva);
1565 if (prot & VM_PROT_WRITE)
1568 pmap_inval_init(&info);
1570 ptbase = get_ptbase(pmap, sva);
1572 sindex = (sva >> PAGE_SHIFT);
1573 eindex = (eva >> PAGE_SHIFT);
1576 for (; sindex < eindex; sindex = pdnxt) {
1580 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1582 pdirindex = sindex / NPDEPG;
1583 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1584 pmap->pm_pdir[pdirindex] &= ~(VPTE_M|VPTE_W);
1585 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1586 inval_ptbase_pagedir(pmap, pdirindex);
1591 * Weed out invalid mappings. Note: we assume that the page
1592 * directory table is always allocated, and in kernel virtual.
1597 if (pdnxt > eindex) {
1601 for (; sindex != pdnxt; sindex++) {
1606 /* XXX this isn't optimal */
1607 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1608 pbits = ptbase[sindex - sbase];
1610 if (pbits & PG_MANAGED) {
1612 if (pbits & VPTE_A) {
1613 m = PHYS_TO_VM_PAGE(pbits);
1614 vm_page_flag_set(m, PG_REFERENCED);
1617 if (pbits & VPTE_M) {
1618 if (pmap_track_modified(i386_ptob(sindex))) {
1620 m = PHYS_TO_VM_PAGE(pbits);
1629 if (pbits != ptbase[sindex - sbase]) {
1630 ptbase[sindex - sbase] = pbits;
1634 pmap_inval_flush(&info);
1638 * Enter a managed page into a pmap. If the page is not wired related pmap
1639 * data can be destroyed at any time for later demand-operation.
1641 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1642 * specified protection, and wire the mapping if requested.
1644 * NOTE: This routine may not lazy-evaluate or lose information. The
1645 * page must actually be inserted into the given map NOW.
1647 * NOTE: When entering a page at a KVA address, the pmap must be the
1651 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1657 vm_offset_t origpte, newpte;
1659 pmap_inval_info info;
1667 * Get the page table page. The kernel_pmap's page table pages
1668 * are preallocated and have no associated vm_page_t.
1670 if (pmap == &kernel_pmap)
1673 mpte = pmap_allocpte(pmap, va);
1675 pmap_inval_init(&info);
1676 pte = pmap_pte(pmap, va);
1679 * Page Directory table entry not valid, we need a new PT page
1680 * and pmap_allocpte() didn't give us one. Oops!
1683 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1687 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1688 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1690 opa = origpte & VPTE_FRAME;
1692 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);
1695 if (origpte & VPTE_PS)
1696 panic("pmap_enter: attempted pmap_enter on 4MB page");
1699 * Mapping has not changed, must be protection or wiring change.
1701 if (origpte && (opa == pa)) {
1703 * Wiring change, just update stats. We don't worry about
1704 * wiring PT pages as they remain resident as long as there
1705 * are valid mappings in them. Hence, if a user page is wired,
1706 * the PT page will be also.
1708 if (wired && ((origpte & VPTE_W) == 0))
1709 pmap->pm_stats.wired_count++;
1710 else if (!wired && (origpte & VPTE_W))
1711 pmap->pm_stats.wired_count--;
1713 #if defined(PMAP_DIAGNOSTIC)
1714 if (pmap_nw_modified((pt_entry_t) origpte)) {
1716 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1722 * Remove the extra pte reference. Note that we cannot
1723 * optimize the RO->RW case because we have adjusted the
1724 * wiring count above and may need to adjust the wiring
1731 * We might be turning off write access to the page,
1732 * so we go ahead and sense modify status.
1734 if (origpte & PG_MANAGED) {
1735 if ((origpte & VPTE_M) && pmap_track_modified(va)) {
1737 om = PHYS_TO_VM_PAGE(opa);
1745 * Mapping has changed, invalidate old range and fall through to
1746 * handle validating new mapping.
1750 err = pmap_remove_pte(pmap, pte, va, &info);
1752 panic("pmap_enter: pte vanished, va: 0x%x", va);
1756 * Enter on the PV list if part of our managed memory. Note that we
1757 * raise IPL while manipulating pv_table since pmap_enter can be
1758 * called at interrupt time.
1760 if (pmap_initialized &&
1761 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1762 pmap_insert_entry(pmap, va, mpte, m);
1767 * Increment counters
1769 pmap->pm_stats.resident_count++;
1771 pmap->pm_stats.wired_count++;
1775 * Now validate mapping with desired protection/wiring.
1777 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1784 * if the mapping or permission bits are different, we need
1785 * to update the pte.
1787 if ((origpte & ~(VPTE_M|VPTE_A)) != newpte) {
1788 *pte = newpte | VPTE_A;
1790 pmap_inval_flush(&info);
1794 * This is a quick version of pmap_enter(). It is used only under the
1795 * following conditions:
1797 * (1) The pmap is not the kernel_pmap
1798 * (2) The page is not to be wired into the map
1799 * (3) The page is to mapped read-only in the pmap (initially that is)
1800 * (4) The calling procedure is responsible for flushing the TLB
1801 * (5) The page is always managed
1802 * (6) There is no prior mapping at the VA
1806 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
1810 pmap_inval_info info;
1814 KKASSERT(pmap != &kernel_pmap);
1815 pmap_inval_init(&info);
1817 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1820 * Instantiate the page table page if required
1824 * Calculate pagetable page index
1826 ptepindex = va >> PDRSHIFT;
1827 if (mpte && (mpte->pindex == ptepindex)) {
1832 * Get the page directory entry
1834 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1837 * If the page table page is mapped, we just increment
1838 * the hold count, and activate it.
1841 if (ptepa & VPTE_PS)
1842 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1843 if (pmap->pm_ptphint &&
1844 (pmap->pm_ptphint->pindex == ptepindex)) {
1845 mpte = pmap->pm_ptphint;
1847 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1848 pmap->pm_ptphint = mpte;
1854 mpte = _pmap_allocpte(pmap, ptepindex);
1859 * Ok, now that the page table page has been validated, get the pte.
1860 * If the pte is already mapped undo mpte's hold_count and
1863 pte = pmap_pte(pmap, va);
1866 pmap_unwire_pte_hold(pmap, mpte, &info);
1871 * Enter on the PV list if part of our managed memory. Note that we
1872 * raise IPL while manipulating pv_table since pmap_enter can be
1873 * called at interrupt time.
1875 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
1876 pmap_insert_entry(pmap, va, mpte, m);
1879 * Increment counters
1881 pmap->pm_stats.resident_count++;
1883 pa = VM_PAGE_TO_PHYS(m);
1886 * Now validate mapping with RO protection
1888 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1889 *pte = pa | VPTE_V | VPTE_U;
1891 *pte = pa | VPTE_V | VPTE_U | VPTE_MANAGED;
1897 pmap_extract(pmap_t pmap, vm_offset_t va)
1902 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
1903 if (pte & VPTE_PS) {
1904 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
1905 rtval |= va & SEG_MASK;
1907 pte = *get_ptbase(pmap, va);
1908 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
1915 #define MAX_INIT_PT (96)
1918 * This routine preloads the ptes for a given object into the specified pmap.
1919 * This eliminates the blast of soft faults on process startup and
1920 * immediately after an mmap.
1922 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
1925 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
1926 vm_object_t object, vm_pindex_t pindex,
1927 vm_size_t size, int limit)
1929 struct rb_vm_page_scan_info info;
1933 * We can't preinit if read access isn't set or there is no pmap
1936 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
1940 * We can't preinit if the pmap is not the current pmap
1942 if (curproc == NULL || pmap != vmspace_pmap(curproc->p_vmspace))
1945 psize = size >> PAGE_SHIFT;
1947 if ((object->type != OBJT_VNODE) ||
1948 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
1949 (object->resident_page_count > MAX_INIT_PT))) {
1953 if (psize + pindex > object->size) {
1954 if (object->size < pindex)
1956 psize = object->size - pindex;
1963 * Use a red-black scan to traverse the requested range and load
1964 * any valid pages found into the pmap.
1966 * We cannot safely scan the object's memq unless we are in a
1967 * critical section since interrupts can remove pages from objects.
1969 info.start_pindex = pindex;
1970 info.end_pindex = pindex + psize - 1;
1977 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1978 pmap_object_init_pt_callback, &info);
1984 pmap_object_init_pt_callback(vm_page_t p, void *data)
1986 struct rb_vm_page_scan_info *info = data;
1987 vm_pindex_t rel_index;
1989 * don't allow an madvise to blow away our really
1990 * free pages allocating pv entries.
1992 if ((info->limit & MAP_PREFAULT_MADVISE) &&
1993 vmstats.v_free_count < vmstats.v_free_reserved) {
1996 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
1997 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
1998 if ((p->queue - p->pc) == PQ_CACHE)
1999 vm_page_deactivate(p);
2001 rel_index = p->pindex - info->start_pindex;
2002 info->mpte = pmap_enter_quick(info->pmap,
2003 info->addr + i386_ptob(rel_index),
2005 vm_page_flag_set(p, PG_MAPPED);
2012 * pmap_prefault provides a quick way of clustering pagefaults into a
2013 * processes address space. It is a "cousin" of pmap_object_init_pt,
2014 * except it runs at page fault time instead of mmap time.
2018 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2020 static int pmap_prefault_pageorder[] = {
2021 -PAGE_SIZE, PAGE_SIZE,
2022 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2023 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2024 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2028 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2038 * We do not currently prefault mappings that use virtual page
2039 * tables. We do not prefault foreign pmaps.
2041 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2043 if (curproc == NULL || (pmap != vmspace_pmap(curproc->p_vmspace)))
2046 object = entry->object.vm_object;
2048 starta = addra - PFBAK * PAGE_SIZE;
2049 if (starta < entry->start)
2050 starta = entry->start;
2051 else if (starta > addra)
2055 * critical section protection is required to maintain the
2056 * page/object association, interrupts can free pages and remove
2057 * them from their objects.
2061 for (i = 0; i < PAGEORDER_SIZE; i++) {
2062 vm_object_t lobject;
2065 addr = addra + pmap_prefault_pageorder[i];
2066 if (addr > addra + (PFFOR * PAGE_SIZE))
2069 if (addr < starta || addr >= entry->end)
2073 * Make sure the page table page already exists
2075 if ((*pmap_pde(pmap, addr)) == NULL)
2079 * Get a pointer to the pte and make sure that no valid page
2082 pte = get_ptbase(pmap, addr);
2087 * Get the page to be mapped
2089 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2092 for (m = vm_page_lookup(lobject, pindex);
2093 (!m && (lobject->type == OBJT_DEFAULT) &&
2094 (lobject->backing_object));
2095 lobject = lobject->backing_object
2097 if (lobject->backing_object_offset & PAGE_MASK)
2099 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2100 m = vm_page_lookup(lobject->backing_object, pindex);
2104 * give-up when a page is not in memory
2110 * If everything meets the requirements for pmap_enter_quick(),
2111 * then enter the page.
2114 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2116 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2118 if ((m->queue - m->pc) == PQ_CACHE) {
2119 vm_page_deactivate(m);
2122 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2123 vm_page_flag_set(m, PG_MAPPED);
2131 * Routine: pmap_change_wiring
2132 * Function: Change the wiring attribute for a map/virtual-address
2134 * In/out conditions:
2135 * The mapping must already exist in the pmap.
2138 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2145 pte = get_ptbase(pmap, va);
2147 if (wired && (*pte & VPTE_W) == 0)
2148 pmap->pm_stats.wired_count++;
2149 else if (!wired && (*pte & VPTE_W))
2150 pmap->pm_stats.wired_count--;
2153 * Wiring is not a hardware characteristic so there is no need to
2154 * invalidate TLB. However, in an SMP environment we must use
2155 * a locked bus cycle to update the pte (if we are not using
2156 * the pmap_inval_*() API that is)... it's ok to do this for simple
2161 atomic_set_int(pte, VPTE_W);
2163 atomic_clear_int(pte, VPTE_W);
2166 atomic_set_int_nonlocked(pte, VPTE_W);
2168 atomic_clear_int_nonlocked(pte, VPTE_W);
2173 * Copy the range specified by src_addr/len
2174 * from the source map to the range dst_addr/len
2175 * in the destination map.
2177 * This routine is only advisory and need not do anything.
2180 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2181 vm_size_t len, vm_offset_t src_addr)
2183 pmap_inval_info info;
2185 vm_offset_t end_addr = src_addr + len;
2191 if (dst_addr != src_addr)
2193 if (dst_pmap->pm_pdir == NULL)
2195 if (src_pmap->pm_pdir == NULL)
2198 src_frame = get_ptbase1(src_pmap, src_addr);
2199 dst_frame = get_ptbase2(dst_pmap, src_addr);
2201 pmap_inval_init(&info);
2204 pmap_inval_add(&info, dst_pmap, -1);
2205 pmap_inval_add(&info, src_pmap, -1);
2209 * critical section protection is required to maintain the page/object
2210 * association, interrupts can free pages and remove them from
2214 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2215 vpte_t *src_pte, *dst_pte;
2216 vm_page_t dstmpte, srcmpte;
2217 vm_offset_t srcptepaddr;
2220 if (addr >= VM_MAX_USER_ADDRESS)
2221 panic("pmap_copy: invalid to pmap_copy page tables\n");
2224 * Don't let optional prefaulting of pages make us go
2225 * way below the low water mark of free pages or way
2226 * above high water mark of used pv entries.
2228 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2229 pv_entry_count > pv_entry_high_water)
2232 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2233 ptepindex = addr >> PDRSHIFT;
2235 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2236 if (srcptepaddr == 0)
2239 if (srcptepaddr & VPTE_PS) {
2240 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2241 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2242 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2247 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2248 if ((srcmpte == NULL) ||
2249 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2252 if (pdnxt > end_addr)
2255 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2256 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2257 while (addr < pdnxt) {
2261 * we only virtual copy managed pages
2263 if ((ptetemp & PG_MANAGED) != 0) {
2265 * We have to check after allocpte for the
2266 * pte still being around... allocpte can
2269 dstmpte = pmap_allocpte(dst_pmap, addr);
2270 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2272 * Clear the modified and
2273 * accessed (referenced) bits
2276 m = PHYS_TO_VM_PAGE(ptetemp);
2277 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2278 dst_pmap->pm_stats.resident_count++;
2279 pmap_insert_entry(dst_pmap, addr,
2282 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2284 if (dstmpte->hold_count >= srcmpte->hold_count)
2293 pmap_inval_flush(&info);
2299 * Zero the specified PA by mapping the page into KVM and clearing its
2302 * This function may be called from an interrupt and no locking is
2306 pmap_zero_page(vm_paddr_t phys)
2308 struct mdglobaldata *gd = mdcpu;
2312 panic("pmap_zero_page: CMAP3 busy");
2313 *gd->gd_CMAP3 = VPTE_V | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2314 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2316 bzero(gd->gd_CADDR3, PAGE_SIZE);
2322 * pmap_page_assertzero:
2324 * Assert that a page is empty, panic if it isn't.
2327 pmap_page_assertzero(vm_paddr_t phys)
2329 struct mdglobaldata *gd = mdcpu;
2334 panic("pmap_zero_page: CMAP3 busy");
2335 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2336 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2337 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2338 for (i = 0; i < PAGE_SIZE; i += 4) {
2339 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2340 panic("pmap_page_assertzero() @ %p not zero!\n",
2341 (void *)gd->gd_CADDR3);
2351 * Zero part of a physical page by mapping it into memory and clearing
2352 * its contents with bzero.
2354 * off and size may not cover an area beyond a single hardware page.
2357 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2359 struct mdglobaldata *gd = mdcpu;
2363 panic("pmap_zero_page: CMAP3 busy");
2364 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2365 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2366 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2368 bzero((char *)gd->gd_CADDR3 + off, size);
2376 * Copy the physical page from the source PA to the target PA.
2377 * This function may be called from an interrupt. No locking
2381 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2383 struct mdglobaldata *gd = mdcpu;
2386 if (*(int *) gd->gd_CMAP1)
2387 panic("pmap_copy_page: CMAP1 busy");
2388 if (*(int *) gd->gd_CMAP2)
2389 panic("pmap_copy_page: CMAP2 busy");
2391 *(int *) gd->gd_CMAP1 = VPTE_V | VPTE_R | (src & PG_FRAME) | VPTE_A;
2392 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2394 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2395 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2397 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2399 *(int *) gd->gd_CMAP1 = 0;
2400 *(int *) gd->gd_CMAP2 = 0;
2405 * pmap_copy_page_frag:
2407 * Copy the physical page from the source PA to the target PA.
2408 * This function may be called from an interrupt. No locking
2412 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2414 struct mdglobaldata *gd = mdcpu;
2417 if (*(int *) gd->gd_CMAP1)
2418 panic("pmap_copy_page: CMAP1 busy");
2419 if (*(int *) gd->gd_CMAP2)
2420 panic("pmap_copy_page: CMAP2 busy");
2422 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2423 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2425 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2426 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2428 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2429 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2432 *(int *) gd->gd_CMAP1 = 0;
2433 *(int *) gd->gd_CMAP2 = 0;
2438 * Returns true if the pmap's pv is one of the first
2439 * 16 pvs linked to from this page. This count may
2440 * be changed upwards or downwards in the future; it
2441 * is only necessary that true be returned for a small
2442 * subset of pmaps for proper page aging.
2445 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2450 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2455 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2456 if (pv->pv_pmap == pmap) {
2469 * Remove all pages from specified address space
2470 * this aids process exit speeds. Also, this code
2471 * is special cased for current process only, but
2472 * can have the more generic (and slightly slower)
2473 * mode enabled. This is much faster than pmap_remove
2474 * in the case of running down an entire address space.
2477 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2482 pmap_inval_info info;
2485 if (curproc && pmap == vmspace_pmap(curproc->p_vmspace))
2490 pmap_inval_init(&info);
2492 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2493 if (pv->pv_va >= eva || pv->pv_va < sva) {
2494 npv = TAILQ_NEXT(pv, pv_plist);
2498 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2499 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2503 * We cannot remove wired pages from a process' mapping
2506 if (tpte & VPTE_W) {
2507 npv = TAILQ_NEXT(pv, pv_plist);
2512 m = PHYS_TO_VM_PAGE(tpte);
2514 KASSERT(m < &vm_page_array[vm_page_array_size],
2515 ("pmap_remove_pages: bad tpte %x", tpte));
2517 pv->pv_pmap->pm_stats.resident_count--;
2520 * Update the vm_page_t clean and reference bits.
2522 if (tpte & VPTE_M) {
2526 npv = TAILQ_NEXT(pv, pv_plist);
2527 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2529 m->md.pv_list_count--;
2530 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2531 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2532 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2535 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2538 pmap_inval_flush(&info);
2543 * pmap_testbit tests bits in pte's
2544 * note that the testbit/changebit routines are inline,
2545 * and a lot of things compile-time evaluate.
2548 pmap_testbit(vm_page_t m, int bit)
2553 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2556 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2561 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2563 * if the bit being tested is the modified bit, then
2564 * mark clean_map and ptes as never
2567 if (bit & (VPTE_A|VPTE_M)) {
2568 if (!pmap_track_modified(pv->pv_va))
2572 #if defined(PMAP_DIAGNOSTIC)
2574 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2578 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2589 * this routine is used to modify bits in ptes
2591 static __inline void
2592 pmap_changebit(vm_page_t m, int bit, boolean_t setem)
2594 struct pmap_inval_info info;
2598 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2601 pmap_inval_init(&info);
2605 * Loop over all current mappings setting/clearing as appropos If
2606 * setting RO do we need to clear the VAC?
2608 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2610 * don't write protect pager mappings
2612 if (!setem && (bit == VPTE_W)) {
2613 if (!pmap_track_modified(pv->pv_va))
2617 #if defined(PMAP_DIAGNOSTIC)
2619 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2625 * Careful here. We can use a locked bus instruction to
2626 * clear VPTE_A or VPTE_M safely but we need to synchronize
2627 * with the target cpus when we mess with VPTE_W.
2629 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2631 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2635 atomic_set_int(pte, bit);
2637 atomic_set_int_nonlocked(pte, bit);
2640 vpte_t pbits = *pte;
2642 if (bit == VPTE_W) {
2643 if (pbits & VPTE_M) {
2647 atomic_clear_int(pte, VPTE_M|VPTE_W);
2649 atomic_clear_int_nonlocked(pte, VPTE_M|VPTE_W);
2653 atomic_clear_int(pte, bit);
2655 atomic_clear_int_nonlocked(pte, bit);
2661 pmap_inval_flush(&info);
2666 * pmap_page_protect:
2668 * Lower the permission for all mappings to a given page.
2671 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2673 if ((prot & VM_PROT_WRITE) == 0) {
2674 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2675 pmap_changebit(m, VPTE_W, FALSE);
2683 pmap_phys_address(int ppn)
2685 return (i386_ptob(ppn));
2689 * pmap_ts_referenced:
2691 * Return a count of reference bits for a page, clearing those bits.
2692 * It is not necessary for every reference bit to be cleared, but it
2693 * is necessary that 0 only be returned when there are truly no
2694 * reference bits set.
2696 * XXX: The exact number of bits to check and clear is a matter that
2697 * should be tested and standardized at some point in the future for
2698 * optimal aging of shared pages.
2701 pmap_ts_referenced(vm_page_t m)
2703 pv_entry_t pv, pvf, pvn;
2707 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2712 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2717 pvn = TAILQ_NEXT(pv, pv_list);
2719 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2721 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2723 if (!pmap_track_modified(pv->pv_va))
2726 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2728 if (pte && (*pte & VPTE_A)) {
2730 atomic_clear_int(pte, VPTE_A);
2732 atomic_clear_int_nonlocked(pte, VPTE_A);
2739 } while ((pv = pvn) != NULL && pv != pvf);
2749 * Return whether or not the specified physical page was modified
2750 * in any physical maps.
2753 pmap_is_modified(vm_page_t m)
2755 return pmap_testbit(m, VPTE_M);
2759 * Clear the modify bits on the specified physical page.
2762 pmap_clear_modify(vm_page_t m)
2764 pmap_changebit(m, VPTE_M, FALSE);
2768 * pmap_clear_reference:
2770 * Clear the reference bit on the specified physical page.
2773 pmap_clear_reference(vm_page_t m)
2775 pmap_changebit(m, VPTE_A, FALSE);
2779 * Miscellaneous support routines follow
2783 i386_protection_init(void)
2787 kp = protection_codes;
2788 for (prot = 0; prot < 8; prot++) {
2789 if (prot & VM_PROT_READ)
2791 if (prot & VM_PROT_WRITE)
2793 if (prot & VM_PROT_EXECUTE)
2800 * Map a set of physical memory pages into the kernel virtual
2801 * address space. Return a pointer to where it is mapped. This
2802 * routine is intended to be used for mapping device memory,
2805 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2809 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2811 vm_offset_t va, tmpva, offset;
2814 offset = pa & PAGE_MASK;
2815 size = roundup(offset + size, PAGE_SIZE);
2817 va = kmem_alloc_nofault(&kernel_map, size);
2819 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2821 pa = pa & VPTE_FRAME;
2822 for (tmpva = va; size > 0;) {
2823 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2824 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2832 return ((void *)(va + offset));
2836 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2838 vm_offset_t base, offset;
2840 base = va & VPTE_FRAME;
2841 offset = va & PAGE_MASK;
2842 size = roundup(offset + size, PAGE_SIZE);
2843 pmap_qremove(va, size >> PAGE_SHIFT);
2844 kmem_free(&kernel_map, base, size);
2848 * perform the pmap work for mincore
2851 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2857 ptep = pmap_pte(pmap, addr);
2862 if ((pte = *ptep) != 0) {
2865 val = MINCORE_INCORE;
2866 if ((pte & VPTE_MANAGED) == 0)
2869 pa = pte & VPTE_FRAME;
2871 m = PHYS_TO_VM_PAGE(pa);
2877 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2879 * Modified by someone
2881 else if (m->dirty || pmap_is_modified(m))
2882 val |= MINCORE_MODIFIED_OTHER;
2887 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2890 * Referenced by someone
2892 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
2893 val |= MINCORE_REFERENCED_OTHER;
2894 vm_page_flag_set(m, PG_REFERENCED);
2901 pmap_activate(struct proc *p)
2905 pmap = vmspace_pmap(p->p_vmspace);
2907 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2909 pmap->pm_active |= 1;
2911 #if defined(SWTCH_OPTIM_STATS)
2915 p->p_thread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
2916 load_cr3(p->p_thread->td_pcb->pcb_cr3);
2921 pmap_deactivate(struct proc *p)
2925 pmap = vmspace_pmap(p->p_vmspace);
2927 atomic_clear_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2929 pmap->pm_active &= ~1;
2932 * XXX - note we do not adjust %cr3. The caller is expected to
2933 * activate a new pmap or do a thread-exit.
2938 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
2941 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
2945 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
2952 static void pads (pmap_t pm);
2953 void pmap_pvdump (vm_paddr_t pa);
2955 /* print address space of pmap*/
2963 if (pm == &kernel_pmap)
2965 for (i = 0; i < 1024; i++)
2967 for (j = 0; j < 1024; j++) {
2968 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
2969 if (pm == &kernel_pmap && va < KERNBASE)
2971 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
2973 ptep = pmap_pte(pm, va);
2974 if (ptep && (*ptep & VPTE_V)) {
2976 (void *)va, (unsigned)*ptep);
2983 pmap_pvdump(vm_paddr_t pa)
2988 kprintf("pa %08llx", (long long)pa);
2989 m = PHYS_TO_VM_PAGE(pa);
2990 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2992 kprintf(" -> pmap %p, va %x, flags %x",
2993 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
2995 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);