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 1.26 2007/08/15 03:15:07 dillon Exp $
44 * NOTE: PMAP_INVAL_ADD: In pc32 this function is called prior to adjusting
45 * the PTE in the page table, because a cpu synchronization might be required.
46 * The actual invalidation is delayed until the following call or flush. In
47 * the VKERNEL build this function is called prior to adjusting the PTE and
48 * invalidates the table synchronously (not delayed), and is not SMP safe
52 #include <sys/types.h>
53 #include <sys/systm.h>
54 #include <sys/kernel.h>
57 #include <sys/vkernel.h>
59 #include <sys/thread.h>
61 #include <sys/vmspace.h>
64 #include <vm/vm_page.h>
65 #include <vm/vm_extern.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_zone.h>
69 #include <vm/vm_pageout.h>
71 #include <machine/md_var.h>
72 #include <machine/pcb.h>
73 #include <machine/pmap_inval.h>
74 #include <machine/globaldata.h>
76 #include <sys/sysref2.h>
80 struct pmap kernel_pmap;
82 static struct vm_zone pvzone;
83 static struct vm_object pvzone_obj;
84 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
85 static int pv_entry_count;
86 static int pv_entry_max;
87 static int pv_entry_high_water;
88 static int pmap_pagedaemon_waken;
89 static boolean_t pmap_initialized = FALSE;
90 static int protection_codes[8];
92 static void i386_protection_init(void);
93 static void pmap_remove_all(vm_page_t m);
94 static int pmap_release_free_page(struct pmap *pmap, vm_page_t p);
97 #ifndef PMAP_SHPGPERPROC
98 #define PMAP_SHPGPERPROC 200
101 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
103 #define pte_prot(m, p) \
104 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
110 struct pv_entry *pvinit;
112 for (i = 0; i < vm_page_array_size; i++) {
115 m = &vm_page_array[i];
116 TAILQ_INIT(&m->md.pv_list);
117 m->md.pv_list_count = 0;
120 i = vm_page_array_size;
123 pvinit = (struct pv_entry *)kmem_alloc(&kernel_map, i*sizeof(*pvinit));
124 zbootinit(&pvzone, "PV ENTRY", sizeof(*pvinit), pvinit, i);
125 pmap_initialized = TRUE;
131 int shpgperproc = PMAP_SHPGPERPROC;
133 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
134 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
135 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
136 pv_entry_high_water = 9 * (pv_entry_max / 10);
137 zinitna(&pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
141 * Bootstrap the kernel_pmap so it can be used with pmap_enter().
143 * NOTE! pm_pdir for the kernel pmap is offset so VA's translate
144 * directly into PTD indexes (PTA is also offset for the same reason).
145 * This is necessary because, for now, KVA is not mapped at address 0.
147 * Page table pages are not managed like they are in normal pmaps, so
148 * no pteobj is needed.
153 vm_pindex_t i = (vm_offset_t)KernelPTD >> PAGE_SHIFT;
155 kernel_pmap.pm_pdir = KernelPTD - (KvaStart >> SEG_SHIFT);
156 kernel_pmap.pm_pdirpte = KernelPTA[i];
157 kernel_pmap.pm_count = 1;
158 kernel_pmap.pm_active = (cpumask_t)-1;
159 TAILQ_INIT(&kernel_pmap.pm_pvlist);
160 i386_protection_init();
164 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
165 * just dummy it up so it works well enough for fork().
167 * In DragonFly, process pmaps may only be used to manipulate user address
168 * space, never kernel address space.
171 pmap_pinit0(struct pmap *pmap)
176 /************************************************************************
177 * Procedures to manage whole physical maps *
178 ************************************************************************
180 * Initialize a preallocated and zeroed pmap structure,
181 * such as one in a vmspace structure.
184 pmap_pinit(struct pmap *pmap)
190 * No need to allocate page table space yet but we do need a valid
191 * page directory table.
193 if (pmap->pm_pdir == NULL) {
195 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
199 * allocate object for the pte array and page directory
201 npages = VPTE_PAGETABLE_SIZE +
202 (VM_MAX_USER_ADDRESS / PAGE_SIZE) * sizeof(vpte_t);
203 npages = (npages + PAGE_MASK) / PAGE_SIZE;
205 if (pmap->pm_pteobj == NULL)
206 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, npages);
207 pmap->pm_pdindex = npages - 1;
210 * allocate the page directory page
212 ptdpg = vm_page_grab(pmap->pm_pteobj, pmap->pm_pdindex,
213 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
215 ptdpg->wire_count = 1;
216 ++vmstats.v_wire_count;
218 /* not usually mapped */
219 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
220 ptdpg->valid = VM_PAGE_BITS_ALL;
222 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
223 pmap->pm_pdirpte = KernelPTA[(vm_offset_t)pmap->pm_pdir >> PAGE_SHIFT];
224 if ((ptdpg->flags & PG_ZERO) == 0)
225 bzero(pmap->pm_pdir, PAGE_SIZE);
229 pmap->pm_ptphint = NULL;
230 pmap->pm_cpucachemask = 0;
231 TAILQ_INIT(&pmap->pm_pvlist);
232 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
236 * Clean up a pmap structure so it can be physically freed
239 pmap_puninit(pmap_t pmap)
242 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
243 pmap->pm_pdir = NULL;
245 if (pmap->pm_pteobj) {
246 vm_object_deallocate(pmap->pm_pteobj);
247 pmap->pm_pteobj = NULL;
253 * Wire in kernel global address entries. To avoid a race condition
254 * between pmap initialization and pmap_growkernel, this procedure
255 * adds the pmap to the master list (which growkernel scans to update),
256 * then copies the template.
258 * In a virtual kernel there are no kernel global address entries.
261 pmap_pinit2(struct pmap *pmap)
264 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
269 * Release all resources held by the given physical map.
271 * Should only be called if the map contains no valid mappings.
273 static int pmap_release_callback(struct vm_page *p, void *data);
276 pmap_release(struct pmap *pmap)
278 struct mdglobaldata *gd = mdcpu;
279 vm_object_t object = pmap->pm_pteobj;
280 struct rb_vm_page_scan_info info;
282 KKASSERT(pmap != &kernel_pmap);
284 #if defined(DIAGNOSTIC)
285 if (object->ref_count != 1)
286 panic("pmap_release: pteobj reference count != 1");
289 * Once we destroy the page table, the mapping becomes invalid.
290 * Don't waste time doing a madvise to invalidate the mapping, just
291 * set cpucachemask to 0.
293 if (pmap->pm_pdir == gd->gd_PT1pdir) {
294 gd->gd_PT1pdir = NULL;
296 /* madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL); */
298 if (pmap->pm_pdir == gd->gd_PT2pdir) {
299 gd->gd_PT2pdir = NULL;
301 /* madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL); */
305 info.object = object;
307 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
314 info.limit = object->generation;
316 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
317 pmap_release_callback, &info);
318 if (info.error == 0 && info.mpte) {
319 if (!pmap_release_free_page(pmap, info.mpte))
323 } while (info.error);
326 * Leave the KVA reservation for pm_pdir cached for later reuse.
328 pmap->pm_pdirpte = 0;
329 pmap->pm_cpucachemask = 0;
333 pmap_release_callback(struct vm_page *p, void *data)
335 struct rb_vm_page_scan_info *info = data;
337 if (p->pindex == info->pmap->pm_pdindex) {
341 if (!pmap_release_free_page(info->pmap, p)) {
345 if (info->object->generation != info->limit) {
353 * Retire the given physical map from service. Should only be called if
354 * the map contains no valid mappings.
357 pmap_destroy(pmap_t pmap)
364 count = --pmap->pm_count;
367 panic("destroying a pmap is not yet implemented");
372 * Add a reference to the specified pmap.
375 pmap_reference(pmap_t pmap)
382 /************************************************************************
383 * VMSPACE MANAGEMENT *
384 ************************************************************************
386 * The VMSPACE management we do in our virtual kernel must be reflected
387 * in the real kernel. This is accomplished by making vmspace system
388 * calls to the real kernel.
391 cpu_vmspace_alloc(struct vmspace *vm)
396 #define LAST_EXTENT (VM_MAX_USER_ADDRESS - 0x80000000)
398 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
399 panic("vmspace_create() failed");
401 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
402 PROT_READ|PROT_WRITE,
403 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
405 if (rp == MAP_FAILED)
406 panic("vmspace_mmap: failed1");
407 vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
409 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
410 PROT_READ|PROT_WRITE,
411 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
412 MemImageFd, 0x40000000);
413 if (rp == MAP_FAILED)
414 panic("vmspace_mmap: failed2");
415 vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
417 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
418 PROT_READ|PROT_WRITE,
419 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
420 MemImageFd, 0x80000000);
421 vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
423 if (rp == MAP_FAILED)
424 panic("vmspace_mmap: failed3");
426 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
427 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
429 panic("vmspace_mcontrol: failed1");
430 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
431 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
433 panic("vmspace_mcontrol: failed2");
434 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
435 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
437 panic("vmspace_mcontrol: failed3");
441 cpu_vmspace_free(struct vmspace *vm)
443 if (vmspace_destroy(&vm->vm_pmap) < 0)
444 panic("vmspace_destroy() failed");
447 /************************************************************************
448 * Procedures which operate directly on the kernel PMAP *
449 ************************************************************************/
452 * This maps the requested page table and gives us access to it.
455 get_ptbase(struct pmap *pmap, vm_offset_t va)
457 struct mdglobaldata *gd = mdcpu;
459 if (pmap == &kernel_pmap) {
460 KKASSERT(va >= KvaStart && va < KvaEnd);
461 return(KernelPTA + (va >> PAGE_SHIFT));
462 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
463 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
464 *gd->gd_PT1pde = pmap->pm_pdirpte;
465 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
466 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
468 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
469 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
470 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
471 *gd->gd_PT2pde = pmap->pm_pdirpte;
472 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
473 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
475 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
479 * Otherwise choose one or the other and map the page table
480 * in the KVA space reserved for it.
482 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
483 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
485 if ((gd->gd_PTflip = 1 - gd->gd_PTflip) == 0) {
486 gd->gd_PT1pdir = pmap->pm_pdir;
487 *gd->gd_PT1pde = pmap->pm_pdirpte;
488 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
489 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
490 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
492 gd->gd_PT2pdir = pmap->pm_pdir;
493 *gd->gd_PT2pde = pmap->pm_pdirpte;
494 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
495 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
496 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
501 get_ptbase1(struct pmap *pmap, vm_offset_t va)
503 struct mdglobaldata *gd = mdcpu;
505 if (pmap == &kernel_pmap) {
506 KKASSERT(va >= KvaStart && va < KvaEnd);
507 return(KernelPTA + (va >> PAGE_SHIFT));
508 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
509 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
510 *gd->gd_PT1pde = pmap->pm_pdirpte;
511 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
512 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
514 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
516 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
517 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
518 gd->gd_PT1pdir = pmap->pm_pdir;
519 *gd->gd_PT1pde = pmap->pm_pdirpte;
520 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
521 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
525 get_ptbase2(struct pmap *pmap, vm_offset_t va)
527 struct mdglobaldata *gd = mdcpu;
529 if (pmap == &kernel_pmap) {
530 KKASSERT(va >= KvaStart && va < KvaEnd);
531 return(KernelPTA + (va >> PAGE_SHIFT));
532 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
533 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
534 *gd->gd_PT2pde = pmap->pm_pdirpte;
535 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
536 atomic_set_int(&pmap->pm_cpucachemask, gd->mi.gd_cpumask);
538 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
540 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
541 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
542 gd->gd_PT2pdir = pmap->pm_pdir;
543 *gd->gd_PT2pde = pmap->pm_pdirpte;
544 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
545 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
549 * When removing a page directory the related VA range in the self-mapping
550 * of the page table must be invalidated.
553 inval_ptbase_pagedir(pmap_t pmap, vm_pindex_t pindex)
555 struct mdglobaldata *gd = mdcpu;
558 if (pmap == &kernel_pmap) {
559 va = (vm_offset_t)KernelPTA + (pindex << PAGE_SHIFT);
560 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
563 * XXX this should not strictly be needed because the page
564 * dir should alread be invalidated. test and remove
566 va = (vm_offset_t)pindex << PAGE_SHIFT;
567 vmspace_mcontrol(pmap, (void *)va, SEG_SIZE, MADV_INVAL, 0);
571 * Do a selective invalidation if we have a valid cache of this
574 if (pmap->pm_cpucachemask & gd->mi.gd_cpumask) {
575 if (pmap->pm_pdir == gd->gd_PT1pdir) {
576 va = (vm_offset_t)gd->gd_PT1map +
577 (pindex << PAGE_SHIFT);
578 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
580 if (pmap->pm_pdir == gd->gd_PT2pdir) {
581 va = (vm_offset_t)gd->gd_PT2map +
582 (pindex << PAGE_SHIFT);
583 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
588 * Invalidate any other cpu's cache mappings of this page table,
591 atomic_clear_int(&pmap->pm_cpucachemask, gd->gd_other_cpus);
595 * Return a pointer to the page table entry for the specified va in the
596 * specified pmap. NULL is returned if there is no valid page table page
599 static __inline vpte_t *
600 pmap_pte(struct pmap *pmap, vm_offset_t va)
604 ptep = &pmap->pm_pdir[va >> SEG_SHIFT];
608 return (get_ptbase(pmap, va));
614 * Enter a mapping into kernel_pmap. Mappings created in this fashion
615 * are not managed. Mappings must be immediately accessible on all cpus.
617 * Call pmap_inval_pte() to invalidate the virtual pte and clean out the
618 * real pmap and handle related races before storing the new vpte.
621 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
626 KKASSERT(va >= KvaStart && va < KvaEnd);
627 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
628 ptep = KernelPTA + (va >> PAGE_SHIFT);
630 pmap_inval_pte(ptep, &kernel_pmap, va);
635 * Synchronize a kvm mapping originally made for the private use on
636 * some other cpu so it can be used on all cpus.
638 * XXX add MADV_RESYNC to improve performance.
641 pmap_kenter_sync(vm_offset_t va)
643 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
647 * Synchronize a kvm mapping originally made for the private use on
648 * some other cpu so it can be used on our cpu. Turns out to be the
649 * same madvise() call, because we have to sync the real pmaps anyway.
651 * XXX add MADV_RESYNC to improve performance.
654 pmap_kenter_sync_quick(vm_offset_t va)
656 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
661 * Make a previously read-only kernel mapping R+W (not implemented by
665 pmap_kmodify_rw(vm_offset_t va)
667 *pmap_kpte(va) |= VPTE_R | VPTE_W;
668 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
672 * Make a kernel mapping non-cacheable (not applicable to virtual kernels)
675 pmap_kmodify_nc(vm_offset_t va)
677 *pmap_kpte(va) |= VPTE_N;
678 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
684 * Map a contiguous range of physical memory to a KVM
687 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
689 while (start < end) {
690 pmap_kenter(virt, start);
698 pmap_kpte(vm_offset_t va)
702 KKASSERT(va >= KvaStart && va < KvaEnd);
703 ptep = KernelPTA + (va >> PAGE_SHIFT);
708 * Enter an unmanaged KVA mapping for the private use of the current
709 * cpu only. pmap_kenter_sync() may be called to make the mapping usable
712 * It is illegal for the mapping to be accessed by other cpus unleess
713 * pmap_kenter_sync*() is called.
716 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
721 KKASSERT(va >= KvaStart && va < KvaEnd);
723 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
724 ptep = KernelPTA + (va >> PAGE_SHIFT);
726 pmap_inval_pte_quick(ptep, &kernel_pmap, va);
731 * Make a temporary mapping for a physical address. This is only intended
732 * to be used for panic dumps.
735 pmap_kenter_temporary(vm_paddr_t pa, int i)
737 pmap_kenter(crashdumpmap + (i * PAGE_SIZE), pa);
738 return ((void *)crashdumpmap);
742 * Remove an unmanaged mapping created with pmap_kenter*().
745 pmap_kremove(vm_offset_t va)
749 KKASSERT(va >= KvaStart && va < KvaEnd);
751 ptep = KernelPTA + (va >> PAGE_SHIFT);
753 pmap_inval_pte(ptep, &kernel_pmap, va);
758 * Remove an unmanaged mapping created with pmap_kenter*() but synchronize
759 * only with this cpu.
761 * Unfortunately because we optimize new entries by testing VPTE_V later
762 * on, we actually still have to synchronize with all the cpus. XXX maybe
763 * store a junk value and test against 0 in the other places instead?
766 pmap_kremove_quick(vm_offset_t va)
770 KKASSERT(va >= KvaStart && va < KvaEnd);
772 ptep = KernelPTA + (va >> PAGE_SHIFT);
774 pmap_inval_pte(ptep, &kernel_pmap, va); /* NOT _quick */
779 * Extract the physical address from the kernel_pmap that is associated
780 * with the specified virtual address.
783 pmap_kextract(vm_offset_t va)
788 KKASSERT(va >= KvaStart && va < KvaEnd);
790 ptep = KernelPTA + (va >> PAGE_SHIFT);
791 pa = (vm_paddr_t)(*ptep & VPTE_FRAME) | (va & PAGE_MASK);
796 * Map a set of unmanaged VM pages into KVM.
799 pmap_qenter(vm_offset_t va, struct vm_page **m, int count)
801 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
805 ptep = KernelPTA + (va >> PAGE_SHIFT);
807 pmap_inval_pte(ptep, &kernel_pmap, va);
808 *ptep = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
816 * Map a set of VM pages to kernel virtual memory. If a mapping changes
817 * clear the supplied mask. The caller handles any SMP interactions.
818 * The mask is used to provide the caller with hints on what SMP interactions
822 pmap_qenter2(vm_offset_t va, struct vm_page **m, int count, cpumask_t *mask)
824 cpumask_t cmask = mycpu->gd_cpumask;
826 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
831 ptep = KernelPTA + (va >> PAGE_SHIFT);
832 npte = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
835 pmap_inval_pte_quick(ptep, &kernel_pmap, va);
837 } else if ((*mask & cmask) == 0) {
838 pmap_kenter_sync_quick(va);
848 * Undo the effects of pmap_qenter*().
851 pmap_qremove(vm_offset_t va, int count)
853 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
857 ptep = KernelPTA + (va >> PAGE_SHIFT);
859 pmap_inval_pte(ptep, &kernel_pmap, va);
866 /************************************************************************
867 * Misc support glue called by machine independant code *
868 ************************************************************************
870 * These routines are called by machine independant code to operate on
871 * certain machine-dependant aspects of processes, threads, and pmaps.
875 * Initialize MD portions of the thread structure.
878 pmap_init_thread(thread_t td)
880 /* enforce pcb placement */
881 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
882 td->td_savefpu = &td->td_pcb->pcb_save;
883 td->td_sp = (char *)td->td_pcb - 16;
887 * This routine directly affects the fork perf for a process.
890 pmap_init_proc(struct proc *p)
895 * Destroy the UPAGES for a process that has exited and disassociate
896 * the process from its thread.
899 pmap_dispose_proc(struct proc *p)
901 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
905 * We pre-allocate all page table pages for kernel virtual memory so
906 * this routine will only be called if KVM has been exhausted.
909 pmap_growkernel(vm_offset_t addr)
911 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
913 if (addr > virtual_end - SEG_SIZE)
914 panic("KVM exhausted");
915 kernel_vm_end = addr;
919 * The modification bit is not tracked for any pages in this range. XXX
920 * such pages in this maps should always use pmap_k*() functions and not
923 * XXX User and kernel address spaces are independant for virtual kernels,
924 * this function only applies to the kernel pmap.
927 pmap_track_modified(pmap_t pmap, vm_offset_t va)
929 if (pmap != &kernel_pmap)
931 if ((va < clean_sva) || (va >= clean_eva))
937 /************************************************************************
938 * Procedures supporting managed page table pages *
939 ************************************************************************
941 * These procedures are used to track managed page table pages. These pages
942 * use the page table page's vm_page_t to track PTEs in the page. The
943 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
945 * This allows the system to throw away page table pages for user processes
946 * at will and reinstantiate them on demand.
950 * This routine works like vm_page_lookup() but also blocks as long as the
951 * page is busy. This routine does not busy the page it returns.
953 * Unless the caller is managing objects whos pages are in a known state,
954 * the call should be made with a critical section held so the page's object
955 * association remains valid on return.
958 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
963 m = vm_page_lookup(object, pindex);
964 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
970 * This routine unholds page table pages, and if the hold count
971 * drops to zero, then it decrements the wire count.
974 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
976 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
979 if (m->hold_count == 0) {
981 * Unmap the page table page.
983 pmap_inval_pde(&pmap->pm_pdir[m->pindex], pmap,
984 (vm_offset_t)m->pindex << SEG_SHIFT);
985 pmap->pm_pdir[m->pindex] = 0;
986 --pmap->pm_stats.resident_count;
988 if (pmap->pm_ptphint == m)
989 pmap->pm_ptphint = NULL;
992 * If the page is finally unwired, simply free it.
995 if (m->wire_count == 0) {
998 vm_page_free_zero(m);
999 --vmstats.v_wire_count;
1007 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
1010 if (m->hold_count == 0)
1011 return _pmap_unwire_pte_hold(pmap, m);
1017 * After removing a page table entry, this routine is used to
1018 * conditionally free the page, and manage the hold/wire counts.
1021 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
1027 * page table pages in the kernel_pmap are not managed.
1029 if (pmap == &kernel_pmap)
1031 ptepindex = (va >> PDRSHIFT);
1032 if (pmap->pm_ptphint &&
1033 (pmap->pm_ptphint->pindex == ptepindex)) {
1034 mpte = pmap->pm_ptphint;
1036 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1037 pmap->pm_ptphint = mpte;
1040 return pmap_unwire_pte_hold(pmap, mpte);
1044 * Attempt to release and free an vm_page in a pmap. Returns 1 on success,
1045 * 0 on failure (if the procedure had to sleep).
1048 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1050 vpte_t *pde = pmap->pm_pdir;
1052 * This code optimizes the case of freeing non-busy
1053 * page-table pages. Those pages are zero now, and
1054 * might as well be placed directly into the zero queue.
1056 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1060 pmap->pm_stats.resident_count--;
1062 if (p->hold_count) {
1063 panic("pmap_release: freeing held page table page");
1066 * Page directory pages need to have the kernel stuff cleared, so
1067 * they can go into the zero queue also.
1069 * In virtual kernels there is no 'kernel stuff'. For the moment
1070 * I just make sure the whole thing has been zero'd even though
1071 * it should already be completely zero'd.
1073 * pmaps for vkernels do not self-map because they do not share
1074 * their address space with the vkernel. Clearing of pde[] thus
1075 * only applies to page table pages and not to the page directory
1078 if (p->pindex == pmap->pm_pdindex) {
1079 bzero(pde, VPTE_PAGETABLE_SIZE);
1080 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1082 pmap_inval_pde(&pde[p->pindex], pmap,
1083 (vm_offset_t)p->pindex << SEG_SHIFT);
1088 * Clear the matching hint
1090 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1091 pmap->pm_ptphint = NULL;
1094 * And throw the page away. The page is completely zero'd out so
1095 * optimize the free call.
1098 vmstats.v_wire_count--;
1099 vm_page_free_zero(p);
1104 * This routine is called if the page table page is not mapped in the page
1107 * The routine is broken up into two parts for readability.
1110 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1116 * Find or fabricate a new pagetable page
1118 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1119 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1121 KASSERT(m->queue == PQ_NONE,
1122 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1124 if (m->wire_count == 0)
1125 vmstats.v_wire_count++;
1129 * Increment the hold count for the page table page
1130 * (denoting a new mapping.)
1135 * Map the pagetable page into the process address space, if
1136 * it isn't already there.
1138 pmap->pm_stats.resident_count++;
1140 ptepa = VM_PAGE_TO_PHYS(m);
1141 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
1145 * We are likely about to access this page table page, so set the
1146 * page table hint to reduce overhead.
1148 pmap->pm_ptphint = m;
1151 * Try to use the new mapping, but if we cannot, then
1152 * do it with the routine that maps the page explicitly.
1154 if ((m->flags & PG_ZERO) == 0)
1155 pmap_zero_page(ptepa);
1157 m->valid = VM_PAGE_BITS_ALL;
1158 vm_page_flag_clear(m, PG_ZERO);
1159 vm_page_flag_set(m, PG_MAPPED);
1166 * Determine the page table page required to access the VA in the pmap
1167 * and allocate it if necessary. Return a held vm_page_t for the page.
1169 * Only used with user pmaps.
1172 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1179 * Calculate pagetable page index
1181 ptepindex = va >> PDRSHIFT;
1184 * Get the page directory entry
1186 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1189 * This supports switching from a 4MB page to a
1192 if (ptepa & VPTE_PS) {
1193 pmap_inval_pde(&pmap->pm_pdir[ptepindex], pmap,
1194 (vm_offset_t)ptepindex << SEG_SHIFT);
1195 pmap->pm_pdir[ptepindex] = 0;
1200 * If the page table page is mapped, we just increment the
1201 * hold count, and activate it.
1205 * In order to get the page table page, try the
1208 if (pmap->pm_ptphint &&
1209 (pmap->pm_ptphint->pindex == ptepindex)) {
1210 m = pmap->pm_ptphint;
1212 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1213 pmap->pm_ptphint = m;
1219 * Here if the pte page isn't mapped, or if it has been deallocated.
1221 return _pmap_allocpte(pmap, ptepindex);
1224 /************************************************************************
1225 * Managed pages in pmaps *
1226 ************************************************************************
1228 * All pages entered into user pmaps and some pages entered into the kernel
1229 * pmap are managed, meaning that pmap_protect() and other related management
1230 * functions work on these pages.
1234 * free the pv_entry back to the free list. This function may be
1235 * called from an interrupt.
1237 static __inline void
1238 free_pv_entry(pv_entry_t pv)
1245 * get a new pv_entry, allocating a block from the system
1246 * when needed. This function may be called from an interrupt.
1252 if (pv_entry_high_water &&
1253 (pv_entry_count > pv_entry_high_water) &&
1254 (pmap_pagedaemon_waken == 0)) {
1255 pmap_pagedaemon_waken = 1;
1256 wakeup (&vm_pages_needed);
1258 return zalloc(&pvzone);
1262 * This routine is very drastic, but can save the system
1270 static int warningdone=0;
1272 if (pmap_pagedaemon_waken == 0)
1275 if (warningdone < 5) {
1276 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1280 for(i = 0; i < vm_page_array_size; i++) {
1281 m = &vm_page_array[i];
1282 if (m->wire_count || m->hold_count || m->busy ||
1283 (m->flags & PG_BUSY))
1287 pmap_pagedaemon_waken = 0;
1291 * If it is the first entry on the list, it is actually
1292 * in the header and we must copy the following entry up
1293 * to the header. Otherwise we must search the list for
1294 * the entry. In either case we free the now unused entry.
1297 pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va)
1303 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1304 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1305 if (pmap == pv->pv_pmap && va == pv->pv_va)
1309 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1310 if (va == pv->pv_va)
1316 * Note that pv_ptem is NULL if the page table page itself is not
1317 * managed, even if the page being removed IS managed.
1321 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1322 m->md.pv_list_count--;
1323 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1324 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1325 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1326 ++pmap->pm_generation;
1327 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1335 * Create a pv entry for page at pa for (pmap, va). If the page table page
1336 * holding the VA is managed, mpte will be non-NULL.
1339 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1344 pv = get_pv_entry();
1349 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1350 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1351 m->md.pv_list_count++;
1357 * pmap_remove_pte: do the things to unmap a page in a process
1360 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va)
1365 oldpte = pmap_inval_loadandclear(ptq, pmap, va);
1366 if (oldpte & VPTE_WIRED)
1367 --pmap->pm_stats.wired_count;
1368 KKASSERT(pmap->pm_stats.wired_count >= 0);
1372 * Machines that don't support invlpg, also don't support
1373 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1376 if (oldpte & VPTE_G)
1377 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1379 pmap->pm_stats.resident_count -= 1;
1380 if (oldpte & VPTE_MANAGED) {
1381 m = PHYS_TO_VM_PAGE(oldpte);
1382 if (oldpte & VPTE_M) {
1383 #if defined(PMAP_DIAGNOSTIC)
1384 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1386 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1390 if (pmap_track_modified(pmap, va))
1393 if (oldpte & VPTE_A)
1394 vm_page_flag_set(m, PG_REFERENCED);
1395 return pmap_remove_entry(pmap, m, va);
1397 return pmap_unuse_pt(pmap, va, NULL);
1406 * Remove a single page from a process address space.
1408 * This function may not be called from an interrupt if the pmap is
1412 pmap_remove_page(struct pmap *pmap, vm_offset_t va)
1417 * if there is no pte for this address, just skip it!!! Otherwise
1418 * get a local va for mappings for this pmap and remove the entry.
1420 if (*pmap_pde(pmap, va) != 0) {
1421 ptq = get_ptbase(pmap, va);
1423 pmap_remove_pte(pmap, ptq, va);
1431 * Remove the given range of addresses from the specified map.
1433 * It is assumed that the start and end are properly
1434 * rounded to the page size.
1436 * This function may not be called from an interrupt if the pmap is
1440 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1444 vm_offset_t ptpaddr;
1445 vm_pindex_t sindex, eindex;
1450 KKASSERT(pmap->pm_stats.resident_count >= 0);
1451 if (pmap->pm_stats.resident_count == 0)
1455 * special handling of removing one page. a very
1456 * common operation and easy to short circuit some
1459 if (((sva + PAGE_SIZE) == eva) &&
1460 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1461 pmap_remove_page(pmap, sva);
1466 * Get a local virtual address for the mappings that are being
1469 * XXX this is really messy because the kernel pmap is not relative
1472 sindex = (sva >> PAGE_SHIFT);
1473 eindex = (eva >> PAGE_SHIFT);
1475 for (; sindex < eindex; sindex = pdnxt) {
1479 * Calculate index for next page table.
1481 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1482 if (pmap->pm_stats.resident_count == 0)
1485 pdirindex = sindex / NPDEPG;
1486 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1487 pmap->pm_pdir[pdirindex] = 0;
1488 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1489 pmap_inval_pde(&pmap->pm_pdir[pdirindex], pmap,
1490 (vm_offset_t)pdirindex << SEG_SHIFT);
1495 * Weed out invalid mappings. Note: we assume that the page
1496 * directory table is always allocated, and in kernel virtual.
1502 * Limit our scan to either the end of the va represented
1503 * by the current page table page, or to the end of the
1504 * range being removed.
1510 * NOTE: pmap_remove_pte() can block.
1512 for (; sindex != pdnxt; sindex++) {
1515 ptbase = get_ptbase(pmap, sindex << PAGE_SHIFT);
1518 va = i386_ptob(sindex);
1519 if (pmap_remove_pte(pmap, ptbase, va))
1528 * Removes this physical page from all physical maps in which it resides.
1529 * Reflects back modify bits to the pager.
1531 * This routine may not be called from an interrupt.
1534 pmap_remove_all(vm_page_t m)
1539 #if defined(PMAP_DIAGNOSTIC)
1541 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1544 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1545 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1550 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1551 pv->pv_pmap->pm_stats.resident_count--;
1553 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1554 KKASSERT(pte != NULL);
1556 tpte = pmap_inval_loadandclear(pte, pv->pv_pmap, pv->pv_va);
1557 if (tpte & VPTE_WIRED)
1558 --pv->pv_pmap->pm_stats.wired_count;
1559 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
1562 vm_page_flag_set(m, PG_REFERENCED);
1565 * Update the vm_page_t clean and reference bits.
1567 if (tpte & VPTE_M) {
1568 #if defined(PMAP_DIAGNOSTIC)
1569 if (pmap_nw_modified((pt_entry_t) tpte)) {
1571 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1575 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
1578 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1579 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1580 ++pv->pv_pmap->pm_generation;
1581 m->md.pv_list_count--;
1582 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
1586 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1593 * Set the physical protection on the specified range of this map
1596 * This function may not be called from an interrupt if the map is
1597 * not the kernel_pmap.
1600 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1604 vm_offset_t pdnxt, ptpaddr;
1605 vm_pindex_t sindex, eindex;
1611 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1612 pmap_remove(pmap, sva, eva);
1616 if (prot & VM_PROT_WRITE)
1619 ptbase = get_ptbase(pmap, sva);
1621 sindex = (sva >> PAGE_SHIFT);
1622 eindex = (eva >> PAGE_SHIFT);
1625 for (; sindex < eindex; sindex = pdnxt) {
1629 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1631 pdirindex = sindex / NPDEPG;
1634 * Clear the modified and writable bits for a 4m page.
1635 * Throw away the modified bit (?)
1637 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1638 pmap_clean_pde(&pmap->pm_pdir[pdirindex], pmap,
1639 (vm_offset_t)pdirindex << SEG_SHIFT);
1640 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1645 * Weed out invalid mappings. Note: we assume that the page
1646 * directory table is always allocated, and in kernel virtual.
1651 if (pdnxt > eindex) {
1655 for (; sindex != pdnxt; sindex++) {
1660 * Clean managed pages and also check the accessed
1661 * bit. Just remove write perms for unmanaged
1662 * pages. Be careful of races, turning off write
1663 * access will force a fault rather then setting
1664 * the modified bit at an unexpected time.
1666 ptep = &ptbase[sindex - sbase];
1667 if (*ptep & VPTE_MANAGED) {
1668 pbits = pmap_clean_pte(ptep, pmap,
1671 if (pbits & VPTE_A) {
1672 m = PHYS_TO_VM_PAGE(pbits);
1673 vm_page_flag_set(m, PG_REFERENCED);
1674 atomic_clear_int(ptep, VPTE_A);
1676 if (pbits & VPTE_M) {
1677 if (pmap_track_modified(pmap, i386_ptob(sindex))) {
1679 m = PHYS_TO_VM_PAGE(pbits);
1684 pbits = pmap_setro_pte(ptep, pmap,
1692 * Enter a managed page into a pmap. If the page is not wired related pmap
1693 * data can be destroyed at any time for later demand-operation.
1695 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1696 * specified protection, and wire the mapping if requested.
1698 * NOTE: This routine may not lazy-evaluate or lose information. The
1699 * page must actually be inserted into the given map NOW.
1701 * NOTE: When entering a page at a KVA address, the pmap must be the
1705 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1711 vm_offset_t origpte, newpte;
1720 * Get the page table page. The kernel_pmap's page table pages
1721 * are preallocated and have no associated vm_page_t.
1723 if (pmap == &kernel_pmap)
1726 mpte = pmap_allocpte(pmap, va);
1728 pte = pmap_pte(pmap, va);
1731 * Page Directory table entry not valid, we need a new PT page
1732 * and pmap_allocpte() didn't give us one. Oops!
1735 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1740 * Deal with races on the original mapping (though don't worry
1741 * about VPTE_A races) by cleaning it. This will force a fault
1742 * if an attempt is made to write to the page.
1744 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1745 origpte = pmap_clean_pte(pte, pmap, va);
1746 opa = origpte & VPTE_FRAME;
1748 if (origpte & VPTE_PS)
1749 panic("pmap_enter: attempted pmap_enter on 4MB page");
1752 * Mapping has not changed, must be protection or wiring change.
1754 if (origpte && (opa == pa)) {
1756 * Wiring change, just update stats. We don't worry about
1757 * wiring PT pages as they remain resident as long as there
1758 * are valid mappings in them. Hence, if a user page is wired,
1759 * the PT page will be also.
1761 if (wired && ((origpte & VPTE_WIRED) == 0))
1762 ++pmap->pm_stats.wired_count;
1763 else if (!wired && (origpte & VPTE_WIRED))
1764 --pmap->pm_stats.wired_count;
1765 KKASSERT(pmap->pm_stats.wired_count >= 0);
1768 * Remove the extra pte reference. Note that we cannot
1769 * optimize the RO->RW case because we have adjusted the
1770 * wiring count above and may need to adjust the wiring
1777 * We might be turning off write access to the page,
1778 * so we go ahead and sense modify status.
1780 if (origpte & VPTE_MANAGED) {
1781 if ((origpte & VPTE_M) &&
1782 pmap_track_modified(pmap, va)) {
1784 om = PHYS_TO_VM_PAGE(opa);
1792 * Mapping has changed, invalidate old range and fall through to
1793 * handle validating new mapping.
1797 err = pmap_remove_pte(pmap, pte, va);
1799 panic("pmap_enter: pte vanished, va: 0x%x", va);
1803 * Enter on the PV list if part of our managed memory. Note that we
1804 * raise IPL while manipulating pv_table since pmap_enter can be
1805 * called at interrupt time.
1807 if (pmap_initialized &&
1808 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1809 pmap_insert_entry(pmap, va, mpte, m);
1814 * Increment counters
1816 pmap->pm_stats.resident_count++;
1818 pmap->pm_stats.wired_count++;
1822 * Now validate mapping with desired protection/wiring.
1824 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1827 newpte |= VPTE_WIRED;
1831 * If the mapping or permission bits are different from the
1832 * (now cleaned) original pte, an update is needed. We've
1833 * already downgraded or invalidated the page so all we have
1834 * to do now is update the bits.
1836 * XXX should we synchronize RO->RW changes to avoid another
1839 if ((origpte & ~(VPTE_W|VPTE_M|VPTE_A)) != newpte) {
1840 *pte = newpte | VPTE_A;
1845 * This is a quick version of pmap_enter(). It is used only under the
1846 * following conditions:
1848 * (1) The pmap is not the kernel_pmap
1849 * (2) The page is not to be wired into the map
1850 * (3) The page is to mapped read-only in the pmap (initially that is)
1851 * (4) The calling procedure is responsible for flushing the TLB
1852 * (5) The page is always managed
1853 * (6) There is no prior mapping at the VA
1857 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
1864 KKASSERT(pmap != &kernel_pmap);
1866 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1869 * Instantiate the page table page if required
1873 * Calculate pagetable page index
1875 ptepindex = va >> PDRSHIFT;
1876 if (mpte && (mpte->pindex == ptepindex)) {
1881 * Get the page directory entry
1883 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1886 * If the page table page is mapped, we just increment
1887 * the hold count, and activate it.
1890 if (ptepa & VPTE_PS)
1891 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1892 if (pmap->pm_ptphint &&
1893 (pmap->pm_ptphint->pindex == ptepindex)) {
1894 mpte = pmap->pm_ptphint;
1896 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1897 pmap->pm_ptphint = mpte;
1903 mpte = _pmap_allocpte(pmap, ptepindex);
1908 * Ok, now that the page table page has been validated, get the pte.
1909 * If the pte is already mapped undo mpte's hold_count and
1912 pte = pmap_pte(pmap, va);
1915 pmap_unwire_pte_hold(pmap, mpte);
1920 * Enter on the PV list if part of our managed memory. Note that we
1921 * raise IPL while manipulating pv_table since pmap_enter can be
1922 * called at interrupt time.
1924 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
1925 pmap_insert_entry(pmap, va, mpte, m);
1928 * Increment counters
1930 pmap->pm_stats.resident_count++;
1932 pa = VM_PAGE_TO_PHYS(m);
1935 * Now validate mapping with RO protection
1937 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1938 *pte = (vpte_t)pa | VPTE_V | VPTE_U;
1940 *pte = (vpte_t)pa | VPTE_V | VPTE_U | VPTE_MANAGED;
1946 * Extract the physical address for the translation at the specified
1947 * virtual address in the pmap.
1950 pmap_extract(pmap_t pmap, vm_offset_t va)
1955 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
1956 if (pte & VPTE_PS) {
1957 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
1958 rtval |= va & SEG_MASK;
1960 pte = *get_ptbase(pmap, va);
1961 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
1968 #define MAX_INIT_PT (96)
1971 * This routine preloads the ptes for a given object into the specified pmap.
1972 * This eliminates the blast of soft faults on process startup and
1973 * immediately after an mmap.
1975 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
1978 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
1979 vm_object_t object, vm_pindex_t pindex,
1980 vm_size_t size, int limit)
1982 struct rb_vm_page_scan_info info;
1987 * We can't preinit if read access isn't set or there is no pmap
1990 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
1994 * We can't preinit if the pmap is not the current pmap
1996 lp = curthread->td_lwp;
1997 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2000 psize = size >> PAGE_SHIFT;
2002 if ((object->type != OBJT_VNODE) ||
2003 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2004 (object->resident_page_count > MAX_INIT_PT))) {
2008 if (psize + pindex > object->size) {
2009 if (object->size < pindex)
2011 psize = object->size - pindex;
2018 * Use a red-black scan to traverse the requested range and load
2019 * any valid pages found into the pmap.
2021 * We cannot safely scan the object's memq unless we are in a
2022 * critical section since interrupts can remove pages from objects.
2024 info.start_pindex = pindex;
2025 info.end_pindex = pindex + psize - 1;
2032 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2033 pmap_object_init_pt_callback, &info);
2039 pmap_object_init_pt_callback(vm_page_t p, void *data)
2041 struct rb_vm_page_scan_info *info = data;
2042 vm_pindex_t rel_index;
2044 * don't allow an madvise to blow away our really
2045 * free pages allocating pv entries.
2047 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2048 vmstats.v_free_count < vmstats.v_free_reserved) {
2051 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2052 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2053 if ((p->queue - p->pc) == PQ_CACHE)
2054 vm_page_deactivate(p);
2056 rel_index = p->pindex - info->start_pindex;
2057 info->mpte = pmap_enter_quick(info->pmap,
2058 info->addr + i386_ptob(rel_index),
2060 vm_page_flag_set(p, PG_MAPPED);
2067 * pmap_prefault provides a quick way of clustering pagefaults into a
2068 * processes address space. It is a "cousin" of pmap_object_init_pt,
2069 * except it runs at page fault time instead of mmap time.
2073 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2075 static int pmap_prefault_pageorder[] = {
2076 -PAGE_SIZE, PAGE_SIZE,
2077 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2078 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2079 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2083 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2094 * We do not currently prefault mappings that use virtual page
2095 * tables. We do not prefault foreign pmaps.
2097 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2099 lp = curthread->td_lwp;
2100 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2103 object = entry->object.vm_object;
2105 starta = addra - PFBAK * PAGE_SIZE;
2106 if (starta < entry->start)
2107 starta = entry->start;
2108 else if (starta > addra)
2112 * critical section protection is required to maintain the
2113 * page/object association, interrupts can free pages and remove
2114 * them from their objects.
2118 for (i = 0; i < PAGEORDER_SIZE; i++) {
2119 vm_object_t lobject;
2122 addr = addra + pmap_prefault_pageorder[i];
2123 if (addr > addra + (PFFOR * PAGE_SIZE))
2126 if (addr < starta || addr >= entry->end)
2130 * Make sure the page table page already exists
2132 if ((*pmap_pde(pmap, addr)) == NULL)
2136 * Get a pointer to the pte and make sure that no valid page
2139 pte = get_ptbase(pmap, addr);
2144 * Get the page to be mapped
2146 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2149 for (m = vm_page_lookup(lobject, pindex);
2150 (!m && (lobject->type == OBJT_DEFAULT) &&
2151 (lobject->backing_object));
2152 lobject = lobject->backing_object
2154 if (lobject->backing_object_offset & PAGE_MASK)
2156 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2157 m = vm_page_lookup(lobject->backing_object, pindex);
2161 * give-up when a page is not in memory
2167 * If everything meets the requirements for pmap_enter_quick(),
2168 * then enter the page.
2171 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2173 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2175 if ((m->queue - m->pc) == PQ_CACHE) {
2176 vm_page_deactivate(m);
2179 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2180 vm_page_flag_set(m, PG_MAPPED);
2188 * Routine: pmap_change_wiring
2189 * Function: Change the wiring attribute for a map/virtual-address
2191 * In/out conditions:
2192 * The mapping must already exist in the pmap.
2195 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2202 pte = get_ptbase(pmap, va);
2204 if (wired && (*pte & VPTE_WIRED) == 0)
2205 ++pmap->pm_stats.wired_count;
2206 else if (!wired && (*pte & VPTE_WIRED))
2207 --pmap->pm_stats.wired_count;
2208 KKASSERT(pmap->pm_stats.wired_count >= 0);
2211 * Wiring is not a hardware characteristic so there is no need to
2212 * invalidate TLB. However, in an SMP environment we must use
2213 * a locked bus cycle to update the pte (if we are not using
2214 * the pmap_inval_*() API that is)... it's ok to do this for simple
2218 atomic_set_int(pte, VPTE_WIRED);
2220 atomic_clear_int(pte, VPTE_WIRED);
2224 * Copy the range specified by src_addr/len
2225 * from the source map to the range dst_addr/len
2226 * in the destination map.
2228 * This routine is only advisory and need not do anything.
2231 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2232 vm_size_t len, vm_offset_t src_addr)
2235 vm_offset_t end_addr = src_addr + len;
2241 if (dst_addr != src_addr)
2243 if (dst_pmap->pm_pdir == NULL)
2245 if (src_pmap->pm_pdir == NULL)
2248 src_frame = get_ptbase1(src_pmap, src_addr);
2249 dst_frame = get_ptbase2(dst_pmap, src_addr);
2252 * critical section protection is required to maintain the page/object
2253 * association, interrupts can free pages and remove them from
2257 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2258 vpte_t *src_pte, *dst_pte;
2259 vm_page_t dstmpte, srcmpte;
2260 vm_offset_t srcptepaddr;
2263 if (addr >= VM_MAX_USER_ADDRESS)
2264 panic("pmap_copy: invalid to pmap_copy page tables\n");
2267 * Don't let optional prefaulting of pages make us go
2268 * way below the low water mark of free pages or way
2269 * above high water mark of used pv entries.
2271 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2272 pv_entry_count > pv_entry_high_water)
2275 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2276 ptepindex = addr >> PDRSHIFT;
2278 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2279 if (srcptepaddr == 0)
2282 if (srcptepaddr & VPTE_PS) {
2283 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2284 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2285 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2290 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2291 if ((srcmpte == NULL) ||
2292 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2295 if (pdnxt > end_addr)
2298 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2299 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2300 while (addr < pdnxt) {
2304 * we only virtual copy managed pages
2306 if ((ptetemp & VPTE_MANAGED) != 0) {
2308 * We have to check after allocpte for the
2309 * pte still being around... allocpte can
2312 dstmpte = pmap_allocpte(dst_pmap, addr);
2313 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2315 * Clear the modified and accessed
2316 * (referenced) bits during the copy.
2318 * We do not have to clear the write
2319 * bit to force a fault-on-modify
2320 * because the real kernel's target
2321 * pmap is empty and will fault anyway.
2323 m = PHYS_TO_VM_PAGE(ptetemp);
2324 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2325 dst_pmap->pm_stats.resident_count++;
2326 pmap_insert_entry(dst_pmap, addr,
2329 pmap_unwire_pte_hold(dst_pmap, dstmpte);
2331 if (dstmpte->hold_count >= srcmpte->hold_count)
2345 * Zero the specified PA by mapping the page into KVM and clearing its
2348 * This function may be called from an interrupt and no locking is
2352 pmap_zero_page(vm_paddr_t phys)
2354 struct mdglobaldata *gd = mdcpu;
2358 panic("pmap_zero_page: CMAP3 busy");
2359 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2360 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2362 bzero(gd->gd_CADDR3, PAGE_SIZE);
2368 * pmap_page_assertzero:
2370 * Assert that a page is empty, panic if it isn't.
2373 pmap_page_assertzero(vm_paddr_t phys)
2375 struct mdglobaldata *gd = mdcpu;
2380 panic("pmap_zero_page: CMAP3 busy");
2381 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2382 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2383 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2384 for (i = 0; i < PAGE_SIZE; i += 4) {
2385 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2386 panic("pmap_page_assertzero() @ %p not zero!\n",
2387 (void *)gd->gd_CADDR3);
2397 * Zero part of a physical page by mapping it into memory and clearing
2398 * its contents with bzero.
2400 * off and size may not cover an area beyond a single hardware page.
2403 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2405 struct mdglobaldata *gd = mdcpu;
2409 panic("pmap_zero_page: CMAP3 busy");
2410 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2411 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2412 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2414 bzero((char *)gd->gd_CADDR3 + off, size);
2422 * Copy the physical page from the source PA to the target PA.
2423 * This function may be called from an interrupt. No locking
2427 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2429 struct mdglobaldata *gd = mdcpu;
2432 if (*(int *) gd->gd_CMAP1)
2433 panic("pmap_copy_page: CMAP1 busy");
2434 if (*(int *) gd->gd_CMAP2)
2435 panic("pmap_copy_page: CMAP2 busy");
2437 *(int *) gd->gd_CMAP1 = VPTE_V | VPTE_R | (src & PG_FRAME) | VPTE_A;
2438 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2440 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2441 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2443 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2445 *(int *) gd->gd_CMAP1 = 0;
2446 *(int *) gd->gd_CMAP2 = 0;
2451 * pmap_copy_page_frag:
2453 * Copy the physical page from the source PA to the target PA.
2454 * This function may be called from an interrupt. No locking
2458 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2460 struct mdglobaldata *gd = mdcpu;
2463 if (*(int *) gd->gd_CMAP1)
2464 panic("pmap_copy_page: CMAP1 busy");
2465 if (*(int *) gd->gd_CMAP2)
2466 panic("pmap_copy_page: CMAP2 busy");
2468 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2469 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2471 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2472 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2474 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2475 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2478 *(int *) gd->gd_CMAP1 = 0;
2479 *(int *) gd->gd_CMAP2 = 0;
2484 * Returns true if the pmap's pv is one of the first
2485 * 16 pvs linked to from this page. This count may
2486 * be changed upwards or downwards in the future; it
2487 * is only necessary that true be returned for a small
2488 * subset of pmaps for proper page aging.
2491 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2496 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2501 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2502 if (pv->pv_pmap == pmap) {
2515 * Remove all pages from specified address space
2516 * this aids process exit speeds. Also, this code
2517 * is special cased for current process only, but
2518 * can have the more generic (and slightly slower)
2519 * mode enabled. This is much faster than pmap_remove
2520 * in the case of running down an entire address space.
2523 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2529 int32_t save_generation;
2532 lp = curthread->td_lwp;
2533 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2539 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2540 if (pv->pv_va >= eva || pv->pv_va < sva) {
2541 npv = TAILQ_NEXT(pv, pv_plist);
2545 KKASSERT(pmap == pv->pv_pmap);
2547 pte = pmap_pte(pmap, pv->pv_va);
2550 * We cannot remove wired pages from a process' mapping
2553 if (*pte & VPTE_WIRED) {
2554 npv = TAILQ_NEXT(pv, pv_plist);
2557 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2559 m = PHYS_TO_VM_PAGE(tpte);
2561 KASSERT(m < &vm_page_array[vm_page_array_size],
2562 ("pmap_remove_pages: bad tpte %x", tpte));
2564 pmap->pm_stats.resident_count--;
2567 * Update the vm_page_t clean and reference bits.
2569 if (tpte & VPTE_M) {
2573 npv = TAILQ_NEXT(pv, pv_plist);
2574 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2575 save_generation = ++pmap->pm_generation;
2577 m->md.pv_list_count--;
2578 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2579 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2580 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2583 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2587 * Restart the scan if we blocked during the unuse or free
2588 * calls and other removals were made.
2590 if (save_generation != pmap->pm_generation) {
2591 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2592 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2599 * pmap_testbit tests bits in active mappings of a VM page.
2602 pmap_testbit(vm_page_t m, int bit)
2607 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2610 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2615 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2617 * if the bit being tested is the modified bit, then
2618 * mark clean_map and ptes as never
2621 if (bit & (VPTE_A|VPTE_M)) {
2622 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2626 #if defined(PMAP_DIAGNOSTIC)
2628 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2632 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2643 * This routine is used to clear bits in ptes. Certain bits require special
2644 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2646 * This routine is only called with certain VPTE_* bit combinations.
2648 static __inline void
2649 pmap_clearbit(vm_page_t m, int bit)
2655 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2661 * Loop over all current mappings setting/clearing as appropos If
2662 * setting RO do we need to clear the VAC?
2664 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2666 * don't write protect pager mappings
2668 if (bit == VPTE_W) {
2669 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2673 #if defined(PMAP_DIAGNOSTIC)
2675 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2681 * Careful here. We can use a locked bus instruction to
2682 * clear VPTE_A or VPTE_M safely but we need to synchronize
2683 * with the target cpus when we mess with VPTE_W.
2685 * On virtual kernels we must force a new fault-on-write
2686 * in the real kernel if we clear the Modify bit ourselves,
2687 * otherwise the real kernel will not get a new fault and
2688 * will never set our Modify bit again.
2690 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2692 if (bit == VPTE_W) {
2694 * We must also clear VPTE_M when clearing
2697 pbits = pmap_clean_pte(pte, pv->pv_pmap,
2701 } else if (bit == VPTE_M) {
2703 * We do not have to make the page read-only
2704 * when clearing the Modify bit. The real
2705 * kernel will make the real PTE read-only
2706 * or otherwise detect the write and set
2707 * our VPTE_M again simply by us invalidating
2708 * the real kernel VA for the pmap (as we did
2709 * above). This allows the real kernel to
2710 * handle the write fault without forwarding
2713 atomic_clear_int(pte, VPTE_M);
2714 } else if ((bit & (VPTE_W|VPTE_M)) == (VPTE_W|VPTE_M)) {
2716 * We've been asked to clear W & M, I guess
2717 * the caller doesn't want us to update
2718 * the dirty status of the VM page.
2720 pmap_clean_pte(pte, pv->pv_pmap, pv->pv_va);
2723 * We've been asked to clear bits that do
2724 * not interact with hardware.
2726 atomic_clear_int(pte, bit);
2734 * pmap_page_protect:
2736 * Lower the permission for all mappings to a given page.
2739 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2741 if ((prot & VM_PROT_WRITE) == 0) {
2742 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2743 pmap_clearbit(m, VPTE_W);
2751 pmap_phys_address(int ppn)
2753 return (i386_ptob(ppn));
2757 * pmap_ts_referenced:
2759 * Return a count of reference bits for a page, clearing those bits.
2760 * It is not necessary for every reference bit to be cleared, but it
2761 * is necessary that 0 only be returned when there are truly no
2762 * reference bits set.
2764 * XXX: The exact number of bits to check and clear is a matter that
2765 * should be tested and standardized at some point in the future for
2766 * optimal aging of shared pages.
2769 pmap_ts_referenced(vm_page_t m)
2771 pv_entry_t pv, pvf, pvn;
2775 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2780 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2785 pvn = TAILQ_NEXT(pv, pv_list);
2787 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2789 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2791 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2794 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2796 if (pte && (*pte & VPTE_A)) {
2798 atomic_clear_int(pte, VPTE_A);
2800 atomic_clear_int_nonlocked(pte, VPTE_A);
2807 } while ((pv = pvn) != NULL && pv != pvf);
2817 * Return whether or not the specified physical page was modified
2818 * in any physical maps.
2821 pmap_is_modified(vm_page_t m)
2823 return pmap_testbit(m, VPTE_M);
2827 * Clear the modify bits on the specified physical page.
2830 pmap_clear_modify(vm_page_t m)
2832 pmap_clearbit(m, VPTE_M);
2836 * pmap_clear_reference:
2838 * Clear the reference bit on the specified physical page.
2841 pmap_clear_reference(vm_page_t m)
2843 pmap_clearbit(m, VPTE_A);
2847 * Miscellaneous support routines follow
2851 i386_protection_init(void)
2855 kp = protection_codes;
2856 for (prot = 0; prot < 8; prot++) {
2857 if (prot & VM_PROT_READ)
2859 if (prot & VM_PROT_WRITE)
2861 if (prot & VM_PROT_EXECUTE)
2870 * Map a set of physical memory pages into the kernel virtual
2871 * address space. Return a pointer to where it is mapped. This
2872 * routine is intended to be used for mapping device memory,
2875 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2879 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2881 vm_offset_t va, tmpva, offset;
2884 offset = pa & PAGE_MASK;
2885 size = roundup(offset + size, PAGE_SIZE);
2887 va = kmem_alloc_nofault(&kernel_map, size);
2889 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2891 pa = pa & VPTE_FRAME;
2892 for (tmpva = va; size > 0;) {
2893 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2894 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2902 return ((void *)(va + offset));
2906 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2908 vm_offset_t base, offset;
2910 base = va & VPTE_FRAME;
2911 offset = va & PAGE_MASK;
2912 size = roundup(offset + size, PAGE_SIZE);
2913 pmap_qremove(va, size >> PAGE_SHIFT);
2914 kmem_free(&kernel_map, base, size);
2920 * perform the pmap work for mincore
2923 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2929 ptep = pmap_pte(pmap, addr);
2934 if ((pte = *ptep) != 0) {
2937 val = MINCORE_INCORE;
2938 if ((pte & VPTE_MANAGED) == 0)
2941 pa = pte & VPTE_FRAME;
2943 m = PHYS_TO_VM_PAGE(pa);
2949 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2951 * Modified by someone
2953 else if (m->dirty || pmap_is_modified(m))
2954 val |= MINCORE_MODIFIED_OTHER;
2959 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2962 * Referenced by someone
2964 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
2965 val |= MINCORE_REFERENCED_OTHER;
2966 vm_page_flag_set(m, PG_REFERENCED);
2973 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
2975 struct vmspace *oldvm;
2978 oldvm = p->p_vmspace;
2980 if (oldvm != newvm) {
2981 p->p_vmspace = newvm;
2982 KKASSERT(p->p_nthreads == 1);
2983 lp = RB_ROOT(&p->p_lwp_tree);
2984 pmap_setlwpvm(lp, newvm);
2986 sysref_get(&newvm->vm_sysref);
2987 sysref_put(&oldvm->vm_sysref);
2994 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
2996 struct vmspace *oldvm;
3000 oldvm = lp->lwp_vmspace;
3002 if (oldvm != newvm) {
3003 lp->lwp_vmspace = newvm;
3004 if (curthread->td_lwp == lp) {
3005 pmap = vmspace_pmap(newvm);
3007 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3009 pmap->pm_active |= 1;
3011 #if defined(SWTCH_OPTIM_STATS)
3014 pmap = vmspace_pmap(oldvm);
3016 atomic_clear_int(&pmap->pm_active,
3017 1 << mycpu->gd_cpuid);
3019 pmap->pm_active &= ~1;
3028 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3031 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3035 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3042 static void pads (pmap_t pm);
3043 void pmap_pvdump (vm_paddr_t pa);
3045 /* print address space of pmap*/
3053 if (pm == &kernel_pmap)
3055 for (i = 0; i < 1024; i++)
3057 for (j = 0; j < 1024; j++) {
3058 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3059 if (pm == &kernel_pmap && va < KERNBASE)
3061 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3063 ptep = pmap_pte(pm, va);
3064 if (ptep && (*ptep & VPTE_V)) {
3066 (void *)va, (unsigned)*ptep);
3073 pmap_pvdump(vm_paddr_t pa)
3078 kprintf("pa %08llx", (long long)pa);
3079 m = PHYS_TO_VM_PAGE(pa);
3080 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3082 kprintf(" -> pmap %p, va %x, flags %x",
3083 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3085 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);