4 * Copyright (c) 2006 The DragonFly Project. All rights reserved.
5 * Copyright (c) 1991 Regents of the University of California.
7 * Copyright (c) 1994 John S. Dyson
9 * Copyright (c) 1994 David Greenman
10 * All rights reserved.
11 * Copyright (c) 2004-2006 Matthew Dillon
12 * All rights reserved.
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in
22 * the documentation and/or other materials provided with the
24 * 3. Neither the name of The DragonFly Project nor the names of its
25 * contributors may be used to endorse or promote products derived
26 * from this software without specific, prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
31 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
32 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
33 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
34 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
35 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
36 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
37 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
38 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
42 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
45 * NOTE: PMAP_INVAL_ADD: In pc32 this function is called prior to adjusting
46 * the PTE in the page table, because a cpu synchronization might be required.
47 * The actual invalidation is delayed until the following call or flush. In
48 * the VKERNEL build this function is called prior to adjusting the PTE and
49 * invalidates the table synchronously (not delayed), and is not SMP safe
53 #include <sys/types.h>
54 #include <sys/systm.h>
55 #include <sys/kernel.h>
58 #include <sys/vkernel.h>
60 #include <sys/thread.h>
62 #include <sys/vmspace.h>
65 #include <vm/vm_page.h>
66 #include <vm/vm_extern.h>
67 #include <vm/vm_kern.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_zone.h>
70 #include <vm/vm_pageout.h>
72 #include <machine/md_var.h>
73 #include <machine/pcb.h>
74 #include <machine/pmap_inval.h>
75 #include <machine/globaldata.h>
77 #include <sys/sysref2.h>
78 #include <sys/spinlock2.h>
79 #include <vm/vm_page2.h>
83 struct pmap kernel_pmap;
85 static struct vm_zone pvzone;
86 static struct vm_object pvzone_obj;
87 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
88 static int pv_entry_count;
89 static int pv_entry_max;
90 static int pv_entry_high_water;
91 static int pmap_pagedaemon_waken;
92 static boolean_t pmap_initialized = FALSE;
93 static int protection_codes[8];
95 static void i386_protection_init(void);
96 static void pmap_remove_all(vm_page_t m);
97 static int pmap_release_free_page(struct pmap *pmap, vm_page_t p);
100 #ifndef PMAP_SHPGPERPROC
101 #define PMAP_SHPGPERPROC 200
104 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
106 #define pte_prot(m, p) \
107 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
112 struct pv_entry *pvinit;
116 npages = VPTE_PAGETABLE_SIZE +
117 (VM_MAX_USER_ADDRESS / PAGE_SIZE) * sizeof(vpte_t);
118 npages = (npages + PAGE_MASK) / PAGE_SIZE;
119 kernel_pmap.pm_pteobj = vm_object_allocate(OBJT_DEFAULT, npages);
121 for (i = 0; i < vm_page_array_size; i++) {
124 m = &vm_page_array[i];
125 TAILQ_INIT(&m->md.pv_list);
126 m->md.pv_list_count = 0;
129 i = vm_page_array_size;
132 pvinit = (struct pv_entry *)kmem_alloc(&kernel_map, i*sizeof(*pvinit));
133 zbootinit(&pvzone, "PV ENTRY", sizeof(*pvinit), pvinit, i);
134 pmap_initialized = TRUE;
140 int shpgperproc = PMAP_SHPGPERPROC;
142 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
143 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
144 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
145 pv_entry_high_water = 9 * (pv_entry_max / 10);
146 zinitna(&pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
150 * Typically used to initialize a fictitious page by vm/device_pager.c
153 pmap_page_init(struct vm_page *m)
156 TAILQ_INIT(&m->md.pv_list);
160 * Bootstrap the kernel_pmap so it can be used with pmap_enter().
162 * NOTE! pm_pdir for the kernel pmap is offset so VA's translate
163 * directly into PTD indexes (PTA is also offset for the same reason).
164 * This is necessary because, for now, KVA is not mapped at address 0.
166 * Page table pages are not managed like they are in normal pmaps, so
167 * no pteobj is needed.
172 vm_pindex_t i = (vm_offset_t)KernelPTD >> PAGE_SHIFT;
175 * The kernel_pmap's pm_pteobj is used only for locking and not
178 kernel_pmap.pm_pdir = KernelPTD - (KvaStart >> SEG_SHIFT);
179 kernel_pmap.pm_pdirpte = KernelPTA[i];
180 kernel_pmap.pm_count = 1;
181 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
182 kernel_pmap.pm_pteobj = NULL; /* see pmap_init */
183 TAILQ_INIT(&kernel_pmap.pm_pvlist);
184 TAILQ_INIT(&kernel_pmap.pm_pvlist_free);
185 spin_init(&kernel_pmap.pm_spin);
186 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
187 i386_protection_init();
191 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
192 * just dummy it up so it works well enough for fork().
194 * In DragonFly, process pmaps may only be used to manipulate user address
195 * space, never kernel address space.
198 pmap_pinit0(struct pmap *pmap)
203 /************************************************************************
204 * Procedures to manage whole physical maps *
205 ************************************************************************
207 * Initialize a preallocated and zeroed pmap structure,
208 * such as one in a vmspace structure.
211 pmap_pinit(struct pmap *pmap)
217 * No need to allocate page table space yet but we do need a valid
218 * page directory table.
220 if (pmap->pm_pdir == NULL) {
222 (vpte_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
226 * allocate object for the pte array and page directory
228 npages = VPTE_PAGETABLE_SIZE +
229 (VM_MAX_USER_ADDRESS / PAGE_SIZE) * sizeof(vpte_t);
230 npages = (npages + PAGE_MASK) / PAGE_SIZE;
232 if (pmap->pm_pteobj == NULL)
233 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, npages);
234 pmap->pm_pdindex = npages - 1;
237 * allocate the page directory page
239 ptdpg = vm_page_grab(pmap->pm_pteobj, pmap->pm_pdindex,
240 VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_ZERO);
243 /* not usually mapped */
244 vm_page_flag_clear(ptdpg, PG_MAPPED);
245 vm_page_wakeup(ptdpg);
247 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
248 pmap->pm_pdirpte = KernelPTA[(vm_offset_t)pmap->pm_pdir >> PAGE_SHIFT];
252 pmap->pm_ptphint = NULL;
253 pmap->pm_cpucachemask = 0;
254 TAILQ_INIT(&pmap->pm_pvlist);
255 TAILQ_INIT(&pmap->pm_pvlist_free);
256 spin_init(&pmap->pm_spin);
257 lwkt_token_init(&pmap->pm_token, "pmap_tok");
258 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
259 pmap->pm_stats.resident_count = 1;
263 * Clean up a pmap structure so it can be physically freed
268 pmap_puninit(pmap_t pmap)
271 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
272 pmap->pm_pdir = NULL;
274 if (pmap->pm_pteobj) {
275 vm_object_deallocate(pmap->pm_pteobj);
276 pmap->pm_pteobj = NULL;
282 * Wire in kernel global address entries. To avoid a race condition
283 * between pmap initialization and pmap_growkernel, this procedure
284 * adds the pmap to the master list (which growkernel scans to update),
285 * then copies the template.
287 * In a virtual kernel there are no kernel global address entries.
292 pmap_pinit2(struct pmap *pmap)
294 spin_lock(&pmap_spin);
295 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
296 spin_unlock(&pmap_spin);
300 * Release all resources held by the given physical map.
302 * Should only be called if the map contains no valid mappings.
304 * Caller must hold pmap->pm_token
306 static int pmap_release_callback(struct vm_page *p, void *data);
309 pmap_release(struct pmap *pmap)
311 struct mdglobaldata *gd = mdcpu;
312 vm_object_t object = pmap->pm_pteobj;
313 struct rb_vm_page_scan_info info;
315 KKASSERT(pmap != &kernel_pmap);
317 #if defined(DIAGNOSTIC)
318 if (object->ref_count != 1)
319 panic("pmap_release: pteobj reference count != 1");
322 * Once we destroy the page table, the mapping becomes invalid.
323 * Don't waste time doing a madvise to invalidate the mapping, just
324 * set cpucachemask to 0.
326 if (pmap->pm_pdir == gd->gd_PT1pdir) {
327 gd->gd_PT1pdir = NULL;
329 /* madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL); */
331 if (pmap->pm_pdir == gd->gd_PT2pdir) {
332 gd->gd_PT2pdir = NULL;
334 /* madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL); */
336 if (pmap->pm_pdir == gd->gd_PT3pdir) {
337 gd->gd_PT3pdir = NULL;
339 /* madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL); */
343 info.object = object;
345 spin_lock(&pmap_spin);
346 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
347 spin_unlock(&pmap_spin);
349 vm_object_hold(object);
353 info.limit = object->generation;
355 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
356 pmap_release_callback, &info);
357 if (info.error == 0 && info.mpte) {
358 if (!pmap_release_free_page(pmap, info.mpte))
361 } while (info.error);
362 vm_object_drop(object);
365 * Leave the KVA reservation for pm_pdir cached for later reuse.
367 pmap->pm_pdirpte = 0;
368 pmap->pm_cpucachemask = 0;
372 * Callback to release a page table page backing a directory
376 pmap_release_callback(struct vm_page *p, void *data)
378 struct rb_vm_page_scan_info *info = data;
380 if (p->pindex == info->pmap->pm_pdindex) {
384 if (!pmap_release_free_page(info->pmap, p)) {
388 if (info->object->generation != info->limit) {
396 * Add a reference to the specified pmap.
401 pmap_reference(pmap_t pmap)
404 lwkt_gettoken(&vm_token);
406 lwkt_reltoken(&vm_token);
410 /************************************************************************
411 * VMSPACE MANAGEMENT *
412 ************************************************************************
414 * The VMSPACE management we do in our virtual kernel must be reflected
415 * in the real kernel. This is accomplished by making vmspace system
416 * calls to the real kernel.
419 cpu_vmspace_alloc(struct vmspace *vm)
424 #define LAST_EXTENT (VM_MAX_USER_ADDRESS - 0x80000000)
426 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
427 panic("vmspace_create() failed");
429 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
430 PROT_READ|PROT_WRITE,
431 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
433 if (rp == MAP_FAILED)
434 panic("vmspace_mmap: failed1");
435 vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
437 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
438 PROT_READ|PROT_WRITE,
439 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
440 MemImageFd, 0x40000000);
441 if (rp == MAP_FAILED)
442 panic("vmspace_mmap: failed2");
443 vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
445 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
446 PROT_READ|PROT_WRITE,
447 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
448 MemImageFd, 0x80000000);
449 vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
451 if (rp == MAP_FAILED)
452 panic("vmspace_mmap: failed3");
454 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
455 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
457 panic("vmspace_mcontrol: failed1");
458 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
459 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
461 panic("vmspace_mcontrol: failed2");
462 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
463 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
465 panic("vmspace_mcontrol: failed3");
469 cpu_vmspace_free(struct vmspace *vm)
471 if (vmspace_destroy(&vm->vm_pmap) < 0)
472 panic("vmspace_destroy() failed");
475 /************************************************************************
476 * Procedures which operate directly on the kernel PMAP *
477 ************************************************************************/
480 * This maps the requested page table and gives us access to it.
482 * This routine can be called from a potentially preempting interrupt
483 * thread or from a normal thread.
486 get_ptbase(struct pmap *pmap, vm_offset_t va)
488 struct mdglobaldata *gd = mdcpu;
490 if (pmap == &kernel_pmap) {
491 KKASSERT(va >= KvaStart && va < KvaEnd);
492 return(KernelPTA + (va >> PAGE_SHIFT));
493 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
494 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
495 *gd->gd_PT1pde = pmap->pm_pdirpte;
496 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
497 atomic_set_cpumask(&pmap->pm_cpucachemask,
500 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
501 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
502 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
503 *gd->gd_PT2pde = pmap->pm_pdirpte;
504 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
505 atomic_set_cpumask(&pmap->pm_cpucachemask,
508 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
512 * If we aren't running from a potentially preempting interrupt,
513 * load a new page table directory into the page table cache
515 if (gd->mi.gd_intr_nesting_level == 0 &&
516 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0) {
518 * Choose one or the other and map the page table
519 * in the KVA space reserved for it.
521 if ((gd->gd_PTflip = 1 - gd->gd_PTflip) == 0) {
522 gd->gd_PT1pdir = pmap->pm_pdir;
523 *gd->gd_PT1pde = pmap->pm_pdirpte;
524 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
525 atomic_set_cpumask(&pmap->pm_cpucachemask,
527 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
529 gd->gd_PT2pdir = pmap->pm_pdir;
530 *gd->gd_PT2pde = pmap->pm_pdirpte;
531 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
532 atomic_set_cpumask(&pmap->pm_cpucachemask,
534 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
539 * If we are running from a preempting interrupt use a private
540 * map. The caller must be in a critical section.
542 KKASSERT(IN_CRITICAL_SECT(curthread));
543 if (pmap->pm_pdir == gd->gd_PT3pdir) {
544 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
545 *gd->gd_PT3pde = pmap->pm_pdirpte;
546 madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL);
547 atomic_set_cpumask(&pmap->pm_cpucachemask,
551 gd->gd_PT3pdir = pmap->pm_pdir;
552 *gd->gd_PT3pde = pmap->pm_pdirpte;
553 madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL);
554 atomic_set_cpumask(&pmap->pm_cpucachemask,
557 return(gd->gd_PT3map + (va >> PAGE_SHIFT));
561 get_ptbase1(struct pmap *pmap, vm_offset_t va)
563 struct mdglobaldata *gd = mdcpu;
565 if (pmap == &kernel_pmap) {
566 KKASSERT(va >= KvaStart && va < KvaEnd);
567 return(KernelPTA + (va >> PAGE_SHIFT));
568 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
569 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
570 *gd->gd_PT1pde = pmap->pm_pdirpte;
571 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
572 atomic_set_cpumask(&pmap->pm_cpucachemask,
575 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
577 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
578 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
579 gd->gd_PT1pdir = pmap->pm_pdir;
580 *gd->gd_PT1pde = pmap->pm_pdirpte;
581 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
582 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
586 get_ptbase2(struct pmap *pmap, vm_offset_t va)
588 struct mdglobaldata *gd = mdcpu;
590 if (pmap == &kernel_pmap) {
591 KKASSERT(va >= KvaStart && va < KvaEnd);
592 return(KernelPTA + (va >> PAGE_SHIFT));
593 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
594 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
595 *gd->gd_PT2pde = pmap->pm_pdirpte;
596 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
597 atomic_set_cpumask(&pmap->pm_cpucachemask,
600 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
602 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
603 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
604 gd->gd_PT2pdir = pmap->pm_pdir;
605 *gd->gd_PT2pde = pmap->pm_pdirpte;
606 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
607 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
611 * Return a pointer to the page table entry for the specified va in the
612 * specified pmap. NULL is returned if there is no valid page table page
615 static __inline vpte_t *
616 pmap_pte(struct pmap *pmap, vm_offset_t va)
620 ptep = &pmap->pm_pdir[va >> SEG_SHIFT];
624 return (get_ptbase(pmap, va));
630 * Enter a mapping into kernel_pmap. Mappings created in this fashion
631 * are not managed. Mappings must be immediately accessible on all cpus.
633 * Call pmap_inval_pte() to invalidate the virtual pte and clean out the
634 * real pmap and handle related races before storing the new vpte.
637 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
642 KKASSERT(va >= KvaStart && va < KvaEnd);
643 npte = (vpte_t)pa | VPTE_RW | VPTE_V;
644 ptep = KernelPTA + (va >> PAGE_SHIFT);
646 pmap_inval_pte(ptep, &kernel_pmap, va);
651 * Synchronize a kvm mapping originally made for the private use on
652 * some other cpu so it can be used on all cpus.
654 * XXX add MADV_RESYNC to improve performance.
657 pmap_kenter_sync(vm_offset_t va)
659 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
663 * Synchronize a kvm mapping originally made for the private use on
664 * some other cpu so it can be used on our cpu. Turns out to be the
665 * same madvise() call, because we have to sync the real pmaps anyway.
667 * XXX add MADV_RESYNC to improve performance.
670 pmap_kenter_sync_quick(vm_offset_t va)
672 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
677 * Make a previously read-only kernel mapping R+W (not implemented by
681 pmap_kmodify_rw(vm_offset_t va)
683 *pmap_kpte(va) |= VPTE_RW;
684 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
688 * Make a kernel mapping non-cacheable (not applicable to virtual kernels)
691 pmap_kmodify_nc(vm_offset_t va)
693 *pmap_kpte(va) |= VPTE_N;
694 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
700 * Map a contiguous range of physical memory to a KVM
703 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
705 vm_offset_t sva, virt;
708 while (start < end) {
709 pmap_kenter(virt, start);
718 pmap_kpte(vm_offset_t va)
722 KKASSERT(va >= KvaStart && va < KvaEnd);
723 ptep = KernelPTA + (va >> PAGE_SHIFT);
728 * Enter an unmanaged KVA mapping for the private use of the current
729 * cpu only. pmap_kenter_sync() may be called to make the mapping usable
732 * It is illegal for the mapping to be accessed by other cpus unleess
733 * pmap_kenter_sync*() is called.
736 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
741 KKASSERT(va >= KvaStart && va < KvaEnd);
743 npte = (vpte_t)pa | VPTE_RW | VPTE_V;
744 ptep = KernelPTA + (va >> PAGE_SHIFT);
746 pmap_inval_pte_quick(ptep, &kernel_pmap, va);
751 * Make a temporary mapping for a physical address. This is only intended
752 * to be used for panic dumps.
754 * The caller is responsible for calling smp_invltlb().
757 pmap_kenter_temporary(vm_paddr_t pa, long i)
759 pmap_kenter_quick(crashdumpmap + (i * PAGE_SIZE), pa);
760 return ((void *)crashdumpmap);
764 * Remove an unmanaged mapping created with pmap_kenter*().
767 pmap_kremove(vm_offset_t va)
771 KKASSERT(va >= KvaStart && va < KvaEnd);
773 ptep = KernelPTA + (va >> PAGE_SHIFT);
775 pmap_inval_pte(ptep, &kernel_pmap, va);
780 * Remove an unmanaged mapping created with pmap_kenter*() but synchronize
781 * only with this cpu.
783 * Unfortunately because we optimize new entries by testing VPTE_V later
784 * on, we actually still have to synchronize with all the cpus. XXX maybe
785 * store a junk value and test against 0 in the other places instead?
788 pmap_kremove_quick(vm_offset_t va)
792 KKASSERT(va >= KvaStart && va < KvaEnd);
794 ptep = KernelPTA + (va >> PAGE_SHIFT);
796 pmap_inval_pte(ptep, &kernel_pmap, va); /* NOT _quick */
801 * Extract the physical address from the kernel_pmap that is associated
802 * with the specified virtual address.
805 pmap_kextract(vm_offset_t va)
810 KKASSERT(va >= KvaStart && va < KvaEnd);
812 ptep = KernelPTA + (va >> PAGE_SHIFT);
813 pa = (vm_paddr_t)(*ptep & VPTE_FRAME) | (va & PAGE_MASK);
818 * Map a set of unmanaged VM pages into KVM.
821 pmap_qenter(vm_offset_t va, struct vm_page **m, int count)
823 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
827 ptep = KernelPTA + (va >> PAGE_SHIFT);
829 pmap_inval_pte(ptep, &kernel_pmap, va);
830 *ptep = (vpte_t)(*m)->phys_addr | VPTE_RW | VPTE_V;
838 * Undo the effects of pmap_qenter*().
841 pmap_qremove(vm_offset_t va, int count)
843 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
847 ptep = KernelPTA + (va >> PAGE_SHIFT);
849 pmap_inval_pte(ptep, &kernel_pmap, va);
856 /************************************************************************
857 * Misc support glue called by machine independant code *
858 ************************************************************************
860 * These routines are called by machine independant code to operate on
861 * certain machine-dependant aspects of processes, threads, and pmaps.
865 * Initialize MD portions of the thread structure.
868 pmap_init_thread(thread_t td)
870 /* enforce pcb placement */
871 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
872 td->td_savefpu = &td->td_pcb->pcb_save;
873 td->td_sp = (char *)td->td_pcb - 16;
877 * This routine directly affects the fork perf for a process.
880 pmap_init_proc(struct proc *p)
885 * We pre-allocate all page table pages for kernel virtual memory so
886 * this routine will only be called if KVM has been exhausted.
891 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
895 addr = (kend + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
897 lwkt_gettoken(&vm_token);
898 if (addr > virtual_end - SEG_SIZE)
899 panic("KVM exhausted");
900 kernel_vm_end = addr;
901 lwkt_reltoken(&vm_token);
905 * The modification bit is not tracked for any pages in this range. XXX
906 * such pages in this maps should always use pmap_k*() functions and not
909 * XXX User and kernel address spaces are independant for virtual kernels,
910 * this function only applies to the kernel pmap.
913 pmap_track_modified(pmap_t pmap, vm_offset_t va)
915 if (pmap != &kernel_pmap)
917 if ((va < clean_sva) || (va >= clean_eva))
923 /************************************************************************
924 * Procedures supporting managed page table pages *
925 ************************************************************************
927 * These procedures are used to track managed page table pages. These pages
928 * use the page table page's vm_page_t to track PTEs in the page. The
929 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
931 * This allows the system to throw away page table pages for user processes
932 * at will and reinstantiate them on demand.
936 * This routine works like vm_page_lookup() but also blocks as long as the
937 * page is busy. This routine does not busy the page it returns.
939 * Unless the caller is managing objects whos pages are in a known state,
940 * the call should be made with a critical section held so the page's object
941 * association remains valid on return.
944 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
948 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
949 m = vm_page_lookup_busy_wait(object, pindex, FALSE, "pplookp");
955 * This routine unholds page table pages, and if the hold count
956 * drops to zero, then it decrements the wire count.
958 * We must recheck that this is the last hold reference after busy-sleeping
962 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
964 vm_page_busy_wait(m, FALSE, "pmuwpt");
965 KASSERT(m->queue == PQ_NONE,
966 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
968 if (m->hold_count == 1) {
970 * Unmap the page table page.
972 KKASSERT(pmap->pm_pdir[m->pindex] != 0);
973 pmap_inval_pde(&pmap->pm_pdir[m->pindex], pmap,
974 (vm_offset_t)m->pindex << SEG_SHIFT);
975 KKASSERT(pmap->pm_stats.resident_count > 0);
976 --pmap->pm_stats.resident_count;
978 if (pmap->pm_ptphint == m)
979 pmap->pm_ptphint = NULL;
982 * This was our last hold, the page had better be unwired
983 * after we decrement wire_count.
985 * FUTURE NOTE: shared page directory page could result in
986 * multiple wire counts.
990 KKASSERT(m->wire_count == 0);
991 atomic_add_int(&vmstats.v_wire_count, -1);
992 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
994 vm_page_free_zero(m);
997 KKASSERT(m->hold_count > 1);
1005 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
1007 KKASSERT(m->hold_count > 0);
1008 if (m->hold_count > 1) {
1012 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)
1025 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1029 * page table pages in the kernel_pmap are not managed.
1031 if (pmap == &kernel_pmap)
1033 ptepindex = (va >> PDRSHIFT);
1034 if (pmap->pm_ptphint &&
1035 (pmap->pm_ptphint->pindex == ptepindex)) {
1036 mpte = pmap->pm_ptphint;
1038 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1039 pmap->pm_ptphint = mpte;
1040 vm_page_wakeup(mpte);
1043 return pmap_unwire_pte_hold(pmap, mpte);
1047 * Attempt to release and free the vm_page backing a page directory page
1048 * in a pmap. Returns 1 on success, 0 on failure (if the procedure had
1052 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1054 vpte_t *pde = pmap->pm_pdir;
1057 * This code optimizes the case of freeing non-busy
1058 * page-table pages. Those pages are zero now, and
1059 * might as well be placed directly into the zero queue.
1061 if (vm_page_busy_try(p, FALSE)) {
1062 vm_page_sleep_busy(p, FALSE, "pmaprl");
1065 KKASSERT(pmap->pm_stats.resident_count > 0);
1066 --pmap->pm_stats.resident_count;
1068 if (p->hold_count) {
1069 panic("pmap_release: freeing held page table page");
1072 * Page directory pages need to have the kernel stuff cleared, so
1073 * they can go into the zero queue also.
1075 * In virtual kernels there is no 'kernel stuff'. For the moment
1076 * I just make sure the whole thing has been zero'd even though
1077 * it should already be completely zero'd.
1079 * pmaps for vkernels do not self-map because they do not share
1080 * their address space with the vkernel. Clearing of pde[] thus
1081 * only applies to page table pages and not to the page directory
1084 if (p->pindex == pmap->pm_pdindex) {
1085 bzero(pde, VPTE_PAGETABLE_SIZE);
1086 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1088 KKASSERT(pde[p->pindex] != 0);
1089 pmap_inval_pde(&pde[p->pindex], pmap,
1090 (vm_offset_t)p->pindex << SEG_SHIFT);
1094 * Clear the matching hint
1096 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1097 pmap->pm_ptphint = NULL;
1100 * And throw the page away. The page is completely zero'd out so
1101 * optimize the free call.
1104 atomic_add_int(&vmstats.v_wire_count, -1);
1105 vm_page_free_zero(p);
1110 * This routine is called if the page table page is not mapped in the page
1113 * The routine is broken up into two parts for readability.
1115 * It must return a held mpte and map the page directory page as required.
1116 * Because vm_page_grab() can block, we must re-check pm_pdir[ptepindex]
1119 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1125 * Find or fabricate a new pagetable page. A busied page will be
1126 * returned. This call may block.
1128 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1129 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1130 vm_page_flag_set(m, PG_MAPPED);
1132 KASSERT(m->queue == PQ_NONE,
1133 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1136 * Increment the hold count for the page we will be returning to
1142 * It is possible that someone else got in and mapped by the page
1143 * directory page while we were blocked, if so just unbusy and
1144 * return the held page.
1146 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1147 KKASSERT((ptepa & VPTE_FRAME) == VM_PAGE_TO_PHYS(m));
1154 * Map the pagetable page into the process address space, if
1155 * it isn't already there.
1157 ++pmap->pm_stats.resident_count;
1159 ptepa = VM_PAGE_TO_PHYS(m);
1160 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_RW | VPTE_V |
1164 * We are likely about to access this page table page, so set the
1165 * page table hint to reduce overhead.
1167 pmap->pm_ptphint = m;
1175 * Determine the page table page required to access the VA in the pmap
1176 * and allocate it if necessary. Return a held vm_page_t for the page.
1178 * Only used with user pmaps.
1181 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1187 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1190 * Calculate pagetable page index
1192 ptepindex = va >> PDRSHIFT;
1195 * Get the page directory entry
1197 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1200 * This supports switching from a 4MB page to a
1203 if (ptepa & VPTE_PS) {
1204 KKASSERT(pmap->pm_pdir[ptepindex] != 0);
1205 pmap_inval_pde(&pmap->pm_pdir[ptepindex], pmap,
1206 (vm_offset_t)ptepindex << SEG_SHIFT);
1211 * If the page table page is mapped, we just increment the
1212 * hold count, and activate it.
1216 * In order to get the page table page, try the
1219 if (pmap->pm_ptphint &&
1220 (pmap->pm_ptphint->pindex == ptepindex)) {
1221 m = pmap->pm_ptphint;
1223 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1224 pmap->pm_ptphint = m;
1231 * Here if the pte page isn't mapped, or if it has been deallocated.
1233 return _pmap_allocpte(pmap, ptepindex);
1236 /************************************************************************
1237 * Managed pages in pmaps *
1238 ************************************************************************
1240 * All pages entered into user pmaps and some pages entered into the kernel
1241 * pmap are managed, meaning that pmap_protect() and other related management
1242 * functions work on these pages.
1246 * free the pv_entry back to the free list. This function may be
1247 * called from an interrupt.
1249 static __inline void
1250 free_pv_entry(pv_entry_t pv)
1257 * get a new pv_entry, allocating a block from the system
1258 * when needed. This function may be called from an interrupt.
1264 if (pv_entry_high_water &&
1265 (pv_entry_count > pv_entry_high_water) &&
1266 (pmap_pagedaemon_waken == 0)) {
1267 pmap_pagedaemon_waken = 1;
1268 wakeup (&vm_pages_needed);
1270 return zalloc(&pvzone);
1274 * This routine is very drastic, but can save the system
1284 static int warningdone=0;
1286 if (pmap_pagedaemon_waken == 0)
1288 lwkt_gettoken(&vm_token);
1289 pmap_pagedaemon_waken = 0;
1291 if (warningdone < 5) {
1292 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1296 for (i = 0; i < vm_page_array_size; i++) {
1297 m = &vm_page_array[i];
1298 if (m->wire_count || m->hold_count)
1300 if (vm_page_busy_try(m, TRUE) == 0) {
1301 if (m->wire_count == 0 && m->hold_count == 0) {
1307 lwkt_reltoken(&vm_token);
1311 * If it is the first entry on the list, it is actually
1312 * in the header and we must copy the following entry up
1313 * to the header. Otherwise we must search the list for
1314 * the entry. In either case we free the now unused entry.
1316 * caller must hold vm_token
1319 pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va)
1325 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1326 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1327 if (pmap == pv->pv_pmap && va == pv->pv_va)
1331 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1332 if (va == pv->pv_va)
1338 * Note that pv_ptem is NULL if the page table page itself is not
1339 * managed, even if the page being removed IS managed.
1343 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1344 m->md.pv_list_count--;
1345 atomic_add_int(&m->object->agg_pv_list_count, -1);
1346 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1347 if (TAILQ_EMPTY(&m->md.pv_list))
1348 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1349 ++pmap->pm_generation;
1350 vm_object_hold(pmap->pm_pteobj);
1351 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1352 vm_object_drop(pmap->pm_pteobj);
1360 * Create a pv entry for page at pa for (pmap, va). If the page table page
1361 * holding the VA is managed, mpte will be non-NULL.
1364 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1369 pv = get_pv_entry();
1374 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1375 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1376 ++pmap->pm_generation;
1377 m->md.pv_list_count++;
1378 atomic_add_int(&m->object->agg_pv_list_count, 1);
1384 * pmap_remove_pte: do the things to unmap a page in a process
1387 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va)
1392 oldpte = pmap_inval_loadandclear(ptq, pmap, va);
1393 if (oldpte & VPTE_WIRED)
1394 --pmap->pm_stats.wired_count;
1395 KKASSERT(pmap->pm_stats.wired_count >= 0);
1399 * Machines that don't support invlpg, also don't support
1400 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1403 if (oldpte & VPTE_G)
1404 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1406 KKASSERT(pmap->pm_stats.resident_count > 0);
1407 --pmap->pm_stats.resident_count;
1408 if (oldpte & VPTE_MANAGED) {
1409 m = PHYS_TO_VM_PAGE(oldpte);
1410 if (oldpte & VPTE_M) {
1411 #if defined(PMAP_DIAGNOSTIC)
1412 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1414 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1418 if (pmap_track_modified(pmap, va))
1421 if (oldpte & VPTE_A)
1422 vm_page_flag_set(m, PG_REFERENCED);
1423 return pmap_remove_entry(pmap, m, va);
1425 return pmap_unuse_pt(pmap, va, NULL);
1434 * Remove a single page from a process address space.
1436 * This function may not be called from an interrupt if the pmap is
1440 pmap_remove_page(struct pmap *pmap, vm_offset_t va)
1445 * if there is no pte for this address, just skip it!!! Otherwise
1446 * get a local va for mappings for this pmap and remove the entry.
1448 if (*pmap_pde(pmap, va) != 0) {
1449 ptq = get_ptbase(pmap, va);
1451 pmap_remove_pte(pmap, ptq, va);
1457 * Remove the given range of addresses from the specified map.
1459 * It is assumed that the start and end are properly rounded to the
1462 * This function may not be called from an interrupt if the pmap is
1468 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1472 vm_offset_t ptpaddr;
1473 vm_pindex_t sindex, eindex;
1478 vm_object_hold(pmap->pm_pteobj);
1479 lwkt_gettoken(&vm_token);
1480 KKASSERT(pmap->pm_stats.resident_count >= 0);
1481 if (pmap->pm_stats.resident_count == 0) {
1482 lwkt_reltoken(&vm_token);
1483 vm_object_drop(pmap->pm_pteobj);
1488 * special handling of removing one page. a very
1489 * common operation and easy to short circuit some
1492 if (((sva + PAGE_SIZE) == eva) &&
1493 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1494 pmap_remove_page(pmap, sva);
1495 lwkt_reltoken(&vm_token);
1496 vm_object_drop(pmap->pm_pteobj);
1501 * Get a local virtual address for the mappings that are being
1504 * XXX this is really messy because the kernel pmap is not relative
1507 sindex = (sva >> PAGE_SHIFT);
1508 eindex = (eva >> PAGE_SHIFT);
1510 for (; sindex < eindex; sindex = pdnxt) {
1514 * Calculate index for next page table.
1516 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1517 if (pmap->pm_stats.resident_count == 0)
1520 pdirindex = sindex / NPDEPG;
1521 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1522 KKASSERT(pmap->pm_pdir[pdirindex] != 0);
1523 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1524 pmap_inval_pde(&pmap->pm_pdir[pdirindex], pmap,
1525 (vm_offset_t)pdirindex << SEG_SHIFT);
1530 * Weed out invalid mappings. Note: we assume that the page
1531 * directory table is always allocated, and in kernel virtual.
1537 * Limit our scan to either the end of the va represented
1538 * by the current page table page, or to the end of the
1539 * range being removed.
1545 * NOTE: pmap_remove_pte() can block.
1547 for (; sindex != pdnxt; sindex++) {
1550 ptbase = get_ptbase(pmap, sindex << PAGE_SHIFT);
1553 va = i386_ptob(sindex);
1554 if (pmap_remove_pte(pmap, ptbase, va))
1558 lwkt_reltoken(&vm_token);
1559 vm_object_drop(pmap->pm_pteobj);
1563 * Removes this physical page from all physical maps in which it resides.
1564 * Reflects back modify bits to the pager.
1566 * This routine may not be called from an interrupt.
1571 pmap_remove_all(vm_page_t m)
1576 #if defined(PMAP_DIAGNOSTIC)
1578 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1581 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1582 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1586 lwkt_gettoken(&vm_token);
1587 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1588 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1589 --pv->pv_pmap->pm_stats.resident_count;
1591 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1592 KKASSERT(pte != NULL);
1594 tpte = pmap_inval_loadandclear(pte, pv->pv_pmap, pv->pv_va);
1595 if (tpte & VPTE_WIRED)
1596 --pv->pv_pmap->pm_stats.wired_count;
1597 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
1600 vm_page_flag_set(m, PG_REFERENCED);
1603 * Update the vm_page_t clean and reference bits.
1605 if (tpte & VPTE_M) {
1606 #if defined(PMAP_DIAGNOSTIC)
1607 if (pmap_nw_modified((pt_entry_t) tpte)) {
1609 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1613 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
1616 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1617 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1618 ++pv->pv_pmap->pm_generation;
1619 m->md.pv_list_count--;
1620 atomic_add_int(&m->object->agg_pv_list_count, -1);
1621 if (TAILQ_EMPTY(&m->md.pv_list))
1622 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1623 vm_object_hold(pv->pv_pmap->pm_pteobj);
1624 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
1625 vm_object_drop(pv->pv_pmap->pm_pteobj);
1628 KKASSERT((m->flags & (PG_MAPPED | PG_WRITEABLE)) == 0);
1629 lwkt_reltoken(&vm_token);
1633 * Set the physical protection on the specified range of this map
1636 * This function may not be called from an interrupt if the map is
1637 * not the kernel_pmap.
1642 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1646 vm_offset_t pdnxt, ptpaddr;
1647 vm_pindex_t sindex, eindex;
1653 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1654 pmap_remove(pmap, sva, eva);
1658 if (prot & VM_PROT_WRITE)
1661 lwkt_gettoken(&vm_token);
1662 ptbase = get_ptbase(pmap, sva);
1664 sindex = (sva >> PAGE_SHIFT);
1665 eindex = (eva >> PAGE_SHIFT);
1668 for (; sindex < eindex; sindex = pdnxt) {
1672 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1674 pdirindex = sindex / NPDEPG;
1677 * Clear the modified and writable bits for a 4m page.
1678 * Throw away the modified bit (?)
1680 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1681 pmap_clean_pde(&pmap->pm_pdir[pdirindex], pmap,
1682 (vm_offset_t)pdirindex << SEG_SHIFT);
1683 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1688 * Weed out invalid mappings. Note: we assume that the page
1689 * directory table is always allocated, and in kernel virtual.
1694 if (pdnxt > eindex) {
1698 for (; sindex != pdnxt; sindex++) {
1703 * Clean managed pages and also check the accessed
1704 * bit. Just remove write perms for unmanaged
1705 * pages. Be careful of races, turning off write
1706 * access will force a fault rather then setting
1707 * the modified bit at an unexpected time.
1709 ptep = &ptbase[sindex - sbase];
1710 if (*ptep & VPTE_MANAGED) {
1711 pbits = pmap_clean_pte(ptep, pmap,
1714 if (pbits & VPTE_A) {
1715 m = PHYS_TO_VM_PAGE(pbits);
1716 vm_page_flag_set(m, PG_REFERENCED);
1717 atomic_clear_long(ptep, VPTE_A);
1719 if (pbits & VPTE_M) {
1720 if (pmap_track_modified(pmap, i386_ptob(sindex))) {
1722 m = PHYS_TO_VM_PAGE(pbits);
1727 pbits = pmap_setro_pte(ptep, pmap,
1732 lwkt_reltoken(&vm_token);
1736 * Enter a managed page into a pmap. If the page is not wired related pmap
1737 * data can be destroyed at any time for later demand-operation.
1739 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1740 * specified protection, and wire the mapping if requested.
1742 * NOTE: This routine may not lazy-evaluate or lose information. The
1743 * page must actually be inserted into the given map NOW.
1745 * NOTE: When entering a page at a KVA address, the pmap must be the
1751 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1752 boolean_t wired, vm_map_entry_t entry __unused)
1757 vpte_t origpte, newpte;
1765 vm_object_hold(pmap->pm_pteobj);
1766 lwkt_gettoken(&vm_token);
1769 * Get the page table page. The kernel_pmap's page table pages
1770 * are preallocated and have no associated vm_page_t.
1772 if (pmap == &kernel_pmap)
1775 mpte = pmap_allocpte(pmap, va);
1777 pte = pmap_pte(pmap, va);
1780 * Page Directory table entry not valid, we need a new PT page
1781 * and pmap_allocpte() didn't give us one. Oops!
1784 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p",
1789 * Deal with races on the original mapping (though don't worry
1790 * about VPTE_A races) by cleaning it. This will force a fault
1791 * if an attempt is made to write to the page.
1793 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1794 origpte = pmap_clean_pte(pte, pmap, va);
1795 opa = origpte & VPTE_FRAME;
1797 if (origpte & VPTE_PS)
1798 panic("pmap_enter: attempted pmap_enter on 4MB page");
1801 * Mapping has not changed, must be protection or wiring change.
1803 if (origpte && (opa == pa)) {
1805 * Wiring change, just update stats. We don't worry about
1806 * wiring PT pages as they remain resident as long as there
1807 * are valid mappings in them. Hence, if a user page is wired,
1808 * the PT page will be also.
1810 if (wired && ((origpte & VPTE_WIRED) == 0))
1811 ++pmap->pm_stats.wired_count;
1812 else if (!wired && (origpte & VPTE_WIRED))
1813 --pmap->pm_stats.wired_count;
1814 KKASSERT(pmap->pm_stats.wired_count >= 0);
1817 * Remove the extra pte reference. Note that we cannot
1818 * optimize the RO->RW case because we have adjusted the
1819 * wiring count above and may need to adjust the wiring
1826 * We might be turning off write access to the page,
1827 * so we go ahead and sense modify status.
1829 if (origpte & VPTE_MANAGED) {
1830 if ((origpte & VPTE_M) &&
1831 pmap_track_modified(pmap, va)) {
1833 om = PHYS_TO_VM_PAGE(opa);
1837 KKASSERT(m->flags & PG_MAPPED);
1842 * Mapping has changed, invalidate old range and fall through to
1843 * handle validating new mapping.
1847 err = pmap_remove_pte(pmap, pte, va);
1849 panic("pmap_enter: pte vanished, va: %p", (void *)va);
1850 pte = pmap_pte(pmap, va);
1851 origpte = pmap_clean_pte(pte, pmap, va);
1852 opa = origpte & VPTE_FRAME;
1854 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
1860 * Enter on the PV list if part of our managed memory. Note that we
1861 * raise IPL while manipulating pv_table since pmap_enter can be
1862 * called at interrupt time.
1864 if (pmap_initialized &&
1865 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1866 pmap_insert_entry(pmap, va, mpte, m);
1868 vm_page_flag_set(m, PG_MAPPED);
1872 * Increment counters
1874 ++pmap->pm_stats.resident_count;
1876 pmap->pm_stats.wired_count++;
1880 * Now validate mapping with desired protection/wiring.
1882 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1885 newpte |= VPTE_WIRED;
1886 if (pmap != &kernel_pmap)
1890 * If the mapping or permission bits are different from the
1891 * (now cleaned) original pte, an update is needed. We've
1892 * already downgraded or invalidated the page so all we have
1893 * to do now is update the bits.
1895 * XXX should we synchronize RO->RW changes to avoid another
1898 if ((origpte & ~(VPTE_RW|VPTE_M|VPTE_A)) != newpte) {
1899 *pte = newpte | VPTE_A;
1900 if (newpte & VPTE_RW)
1901 vm_page_flag_set(m, PG_WRITEABLE);
1903 KKASSERT((newpte & VPTE_MANAGED) == 0 || m->flags & PG_MAPPED);
1904 lwkt_reltoken(&vm_token);
1905 vm_object_drop(pmap->pm_pteobj);
1909 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
1911 * Currently this routine may only be used on user pmaps, not kernel_pmap.
1914 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
1922 KKASSERT(pmap != &kernel_pmap);
1924 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1927 * Calculate pagetable page (mpte), allocating it if necessary.
1929 * A held page table page (mpte), or NULL, is passed onto the
1930 * section following.
1932 ptepindex = va >> PDRSHIFT;
1934 vm_object_hold(pmap->pm_pteobj);
1935 lwkt_gettoken(&vm_token);
1939 * Get the page directory entry
1941 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1944 * If the page table page is mapped, we just increment
1945 * the hold count, and activate it.
1948 if (ptepa & VPTE_PS)
1949 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1950 if (pmap->pm_ptphint &&
1951 (pmap->pm_ptphint->pindex == ptepindex)) {
1952 mpte = pmap->pm_ptphint;
1954 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1955 pmap->pm_ptphint = mpte;
1956 vm_page_wakeup(mpte);
1961 mpte = _pmap_allocpte(pmap, ptepindex);
1963 } while (mpte == NULL);
1966 * Ok, now that the page table page has been validated, get the pte.
1967 * If the pte is already mapped undo mpte's hold_count and
1970 pte = pmap_pte(pmap, va);
1972 pmap_unwire_pte_hold(pmap, mpte);
1973 lwkt_reltoken(&vm_token);
1974 vm_object_drop(pmap->pm_pteobj);
1979 * Enter on the PV list if part of our managed memory. Note that we
1980 * raise IPL while manipulating pv_table since pmap_enter can be
1981 * called at interrupt time.
1983 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1984 pmap_insert_entry(pmap, va, mpte, m);
1985 vm_page_flag_set(m, PG_MAPPED);
1989 * Increment counters
1991 ++pmap->pm_stats.resident_count;
1993 pa = VM_PAGE_TO_PHYS(m);
1996 * Now validate mapping with RO protection
1998 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1999 *pte = (vpte_t)pa | VPTE_V | VPTE_U;
2001 *pte = (vpte_t)pa | VPTE_V | VPTE_U | VPTE_MANAGED;
2002 /*pmap_inval_add(&info, pmap, va); shouldn't be needed 0->valid */
2003 /*pmap_inval_flush(&info); don't need for vkernel */
2004 lwkt_reltoken(&vm_token);
2005 vm_object_drop(pmap->pm_pteobj);
2009 * Extract the physical address for the translation at the specified
2010 * virtual address in the pmap.
2012 * The caller must hold vm_token if non-blocking operation is desired.
2016 pmap_extract(pmap_t pmap, vm_offset_t va)
2021 lwkt_gettoken(&vm_token);
2022 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
2023 if (pte & VPTE_PS) {
2024 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
2025 rtval |= va & SEG_MASK;
2027 pte = *get_ptbase(pmap, va);
2028 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
2033 lwkt_reltoken(&vm_token);
2038 * Similar to extract but checks protections, SMP-friendly short-cut for
2039 * vm_fault_page[_quick]().
2042 pmap_fault_page_quick(pmap_t pmap __unused, vm_offset_t vaddr __unused,
2043 vm_prot_t prot __unused)
2048 #define MAX_INIT_PT (96)
2051 * This routine preloads the ptes for a given object into the specified pmap.
2052 * This eliminates the blast of soft faults on process startup and
2053 * immediately after an mmap.
2057 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2060 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2061 vm_object_t object, vm_pindex_t pindex,
2062 vm_size_t size, int limit)
2064 struct rb_vm_page_scan_info info;
2069 * We can't preinit if read access isn't set or there is no pmap
2072 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2076 * We can't preinit if the pmap is not the current pmap
2078 lp = curthread->td_lwp;
2079 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2082 psize = size >> PAGE_SHIFT;
2084 if ((object->type != OBJT_VNODE) ||
2085 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2086 (object->resident_page_count > MAX_INIT_PT))) {
2090 if (psize + pindex > object->size) {
2091 if (object->size < pindex)
2093 psize = object->size - pindex;
2100 * Use a red-black scan to traverse the requested range and load
2101 * any valid pages found into the pmap.
2103 * We cannot safely scan the object's memq unless we are in a
2104 * critical section since interrupts can remove pages from objects.
2106 info.start_pindex = pindex;
2107 info.end_pindex = pindex + psize - 1;
2113 vm_object_hold_shared(object);
2114 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2115 pmap_object_init_pt_callback, &info);
2116 vm_object_drop(object);
2120 * The caller must hold vm_token.
2124 pmap_object_init_pt_callback(vm_page_t p, void *data)
2126 struct rb_vm_page_scan_info *info = data;
2127 vm_pindex_t rel_index;
2130 * don't allow an madvise to blow away our really
2131 * free pages allocating pv entries.
2133 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2134 vmstats.v_free_count < vmstats.v_free_reserved) {
2139 * Ignore list markers and ignore pages we cannot instantly
2140 * busy (while holding the object token).
2142 if (p->flags & PG_MARKER)
2144 if (vm_page_busy_try(p, TRUE))
2146 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2147 (p->flags & PG_FICTITIOUS) == 0) {
2148 if ((p->queue - p->pc) == PQ_CACHE)
2149 vm_page_deactivate(p);
2150 rel_index = p->pindex - info->start_pindex;
2151 pmap_enter_quick(info->pmap,
2152 info->addr + i386_ptob(rel_index), p);
2159 * Return TRUE if the pmap is in shape to trivially
2160 * pre-fault the specified address.
2162 * Returns FALSE if it would be non-trivial or if a
2163 * pte is already loaded into the slot.
2168 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2173 lwkt_gettoken(&vm_token);
2174 if ((*pmap_pde(pmap, addr)) == 0) {
2177 pte = get_ptbase(pmap, addr);
2178 ret = (*pte) ? 0 : 1;
2180 lwkt_reltoken(&vm_token);
2185 * Change the wiring attribute for a map/virtual-address pair.
2186 * The mapping must already exist in the pmap.
2188 * No other requirements.
2191 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
2192 vm_map_entry_t entry __unused)
2199 lwkt_gettoken(&vm_token);
2200 pte = get_ptbase(pmap, va);
2202 if (wired && (*pte & VPTE_WIRED) == 0)
2203 ++pmap->pm_stats.wired_count;
2204 else if (!wired && (*pte & VPTE_WIRED))
2205 --pmap->pm_stats.wired_count;
2206 KKASSERT(pmap->pm_stats.wired_count >= 0);
2209 * Wiring is not a hardware characteristic so there is no need to
2210 * invalidate TLB. However, in an SMP environment we must use
2211 * a locked bus cycle to update the pte (if we are not using
2212 * the pmap_inval_*() API that is)... it's ok to do this for simple
2216 atomic_set_long(pte, VPTE_WIRED);
2218 atomic_clear_long(pte, VPTE_WIRED);
2219 lwkt_reltoken(&vm_token);
2223 * Copy the range specified by src_addr/len
2224 * from the source map to the range dst_addr/len
2225 * in the destination map.
2227 * This routine is only advisory and need not do anything.
2230 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2231 vm_size_t len, vm_offset_t src_addr)
2234 vm_offset_t end_addr = src_addr + len;
2241 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2242 * valid through blocking calls, and that's just not going to
2249 if (dst_addr != src_addr)
2251 if (dst_pmap->pm_pdir == NULL)
2253 if (src_pmap->pm_pdir == NULL)
2256 lwkt_gettoken(&vm_token);
2258 src_frame = get_ptbase1(src_pmap, src_addr);
2259 dst_frame = get_ptbase2(dst_pmap, src_addr);
2262 * critical section protection is required to maintain the page/object
2263 * association, interrupts can free pages and remove them from
2266 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2267 vpte_t *src_pte, *dst_pte;
2268 vm_page_t dstmpte, srcmpte;
2269 vm_offset_t srcptepaddr;
2272 if (addr >= VM_MAX_USER_ADDRESS)
2273 panic("pmap_copy: invalid to pmap_copy page tables");
2276 * Don't let optional prefaulting of pages make us go
2277 * way below the low water mark of free pages or way
2278 * above high water mark of used pv entries.
2280 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2281 pv_entry_count > pv_entry_high_water)
2284 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2285 ptepindex = addr >> PDRSHIFT;
2287 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2288 if (srcptepaddr == 0)
2291 if (srcptepaddr & VPTE_PS) {
2292 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2293 dst_pmap->pm_pdir[ptepindex] = (vpte_t)srcptepaddr;
2294 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2299 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2300 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2301 (srcmpte->flags & PG_BUSY)) {
2305 if (pdnxt > end_addr)
2308 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2309 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2310 while (addr < pdnxt) {
2315 * we only virtual copy managed pages
2317 if ((ptetemp & VPTE_MANAGED) != 0) {
2319 * We have to check after allocpte for the
2320 * pte still being around... allocpte can
2323 * pmap_allocpte can block, unfortunately
2324 * we have to reload the tables.
2326 dstmpte = pmap_allocpte(dst_pmap, addr);
2327 src_frame = get_ptbase1(src_pmap, src_addr);
2328 dst_frame = get_ptbase2(dst_pmap, src_addr);
2330 if ((*dst_pte == 0) && (ptetemp = *src_pte) &&
2331 (ptetemp & VPTE_MANAGED) != 0) {
2333 * Clear the modified and accessed
2334 * (referenced) bits during the copy.
2336 * We do not have to clear the write
2337 * bit to force a fault-on-modify
2338 * because the real kernel's target
2339 * pmap is empty and will fault anyway.
2341 m = PHYS_TO_VM_PAGE(ptetemp);
2342 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2343 ++dst_pmap->pm_stats.resident_count;
2344 pmap_insert_entry(dst_pmap, addr,
2346 KKASSERT(m->flags & PG_MAPPED);
2348 pmap_unwire_pte_hold(dst_pmap, dstmpte);
2350 if (dstmpte->hold_count >= srcmpte->hold_count)
2358 lwkt_reltoken(&vm_token);
2364 * Zero the specified PA by mapping the page into KVM and clearing its
2367 * This function may be called from an interrupt and no locking is
2371 pmap_zero_page(vm_paddr_t phys)
2373 struct mdglobaldata *gd = mdcpu;
2377 panic("pmap_zero_page: CMAP3 busy");
2378 *gd->gd_CMAP3 = VPTE_V | VPTE_RW | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2379 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2381 bzero(gd->gd_CADDR3, PAGE_SIZE);
2387 * pmap_page_assertzero:
2389 * Assert that a page is empty, panic if it isn't.
2392 pmap_page_assertzero(vm_paddr_t phys)
2394 struct mdglobaldata *gd = mdcpu;
2399 panic("pmap_zero_page: CMAP3 busy");
2400 *gd->gd_CMAP3 = VPTE_V | VPTE_RW |
2401 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2402 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2403 for (i = 0; i < PAGE_SIZE; i += 4) {
2404 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2405 panic("pmap_page_assertzero() @ %p not zero!",
2406 (void *)gd->gd_CADDR3);
2416 * Zero part of a physical page by mapping it into memory and clearing
2417 * its contents with bzero.
2419 * off and size may not cover an area beyond a single hardware page.
2422 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2424 struct mdglobaldata *gd = mdcpu;
2428 panic("pmap_zero_page: CMAP3 busy");
2429 *gd->gd_CMAP3 = VPTE_V | VPTE_RW |
2430 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2431 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2433 bzero((char *)gd->gd_CADDR3 + off, size);
2441 * Copy the physical page from the source PA to the target PA.
2442 * This function may be called from an interrupt. No locking
2446 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2448 struct mdglobaldata *gd = mdcpu;
2451 if (*(int *) gd->gd_CMAP1)
2452 panic("pmap_copy_page: CMAP1 busy");
2453 if (*(int *) gd->gd_CMAP2)
2454 panic("pmap_copy_page: CMAP2 busy");
2456 *(int *) gd->gd_CMAP1 = VPTE_V | (src & PG_FRAME) | VPTE_A;
2457 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_RW | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2459 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2460 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2462 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2464 *(int *) gd->gd_CMAP1 = 0;
2465 *(int *) gd->gd_CMAP2 = 0;
2470 * pmap_copy_page_frag:
2472 * Copy the physical page from the source PA to the target PA.
2473 * This function may be called from an interrupt. No locking
2477 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2479 struct mdglobaldata *gd = mdcpu;
2482 if (*(int *) gd->gd_CMAP1)
2483 panic("pmap_copy_page: CMAP1 busy");
2484 if (*(int *) gd->gd_CMAP2)
2485 panic("pmap_copy_page: CMAP2 busy");
2487 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2488 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_RW | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2490 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2491 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2493 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2494 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2497 *(int *) gd->gd_CMAP1 = 0;
2498 *(int *) gd->gd_CMAP2 = 0;
2503 * Returns true if the pmap's pv is one of the first
2504 * 16 pvs linked to from this page. This count may
2505 * be changed upwards or downwards in the future; it
2506 * is only necessary that true be returned for a small
2507 * subset of pmaps for proper page aging.
2512 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2517 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2521 lwkt_gettoken(&vm_token);
2523 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2524 if (pv->pv_pmap == pmap) {
2525 lwkt_reltoken(&vm_token);
2533 lwkt_reltoken(&vm_token);
2539 * Remove all pages from specified address space
2540 * this aids process exit speeds. Also, this code
2541 * is special cased for current process only, but
2542 * can have the more generic (and slightly slower)
2543 * mode enabled. This is much faster than pmap_remove
2544 * in the case of running down an entire address space.
2549 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2554 int32_t save_generation;
2556 if (pmap->pm_pteobj)
2557 vm_object_hold(pmap->pm_pteobj);
2558 lwkt_gettoken(&vm_token);
2559 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2560 if (pv->pv_va >= eva || pv->pv_va < sva) {
2561 npv = TAILQ_NEXT(pv, pv_plist);
2565 KKASSERT(pmap == pv->pv_pmap);
2567 pte = pmap_pte(pmap, pv->pv_va);
2570 * We cannot remove wired pages from a process' mapping
2573 if (*pte & VPTE_WIRED) {
2574 npv = TAILQ_NEXT(pv, pv_plist);
2577 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2579 m = PHYS_TO_VM_PAGE(tpte);
2581 KASSERT(m < &vm_page_array[vm_page_array_size],
2582 ("pmap_remove_pages: bad tpte %lx", tpte));
2584 KKASSERT(pmap->pm_stats.resident_count > 0);
2585 --pmap->pm_stats.resident_count;
2588 * Update the vm_page_t clean and reference bits.
2590 if (tpte & VPTE_M) {
2594 npv = TAILQ_NEXT(pv, pv_plist);
2595 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2596 save_generation = ++pmap->pm_generation;
2598 m->md.pv_list_count--;
2599 atomic_add_int(&m->object->agg_pv_list_count, -1);
2600 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2601 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2602 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2604 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2608 * Restart the scan if we blocked during the unuse or free
2609 * calls and other removals were made.
2611 if (save_generation != pmap->pm_generation) {
2612 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2613 npv = TAILQ_FIRST(&pmap->pm_pvlist);
2616 lwkt_reltoken(&vm_token);
2617 if (pmap->pm_pteobj)
2618 vm_object_drop(pmap->pm_pteobj);
2622 * pmap_testbit tests bits in active mappings of a VM page.
2624 * The caller must hold vm_token
2627 pmap_testbit(vm_page_t m, int bit)
2632 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2635 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2640 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2642 * if the bit being tested is the modified bit, then
2643 * mark clean_map and ptes as never
2646 if (bit & (VPTE_A|VPTE_M)) {
2647 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2651 #if defined(PMAP_DIAGNOSTIC)
2653 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2657 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2668 * This routine is used to clear bits in ptes. Certain bits require special
2669 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2671 * This routine is only called with certain VPTE_* bit combinations.
2673 * The caller must hold vm_token
2675 static __inline void
2676 pmap_clearbit(vm_page_t m, int bit)
2682 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2688 * Loop over all current mappings setting/clearing as appropos If
2689 * setting RO do we need to clear the VAC?
2691 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2693 * don't write protect pager mappings
2695 if (bit == VPTE_RW) {
2696 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2700 #if defined(PMAP_DIAGNOSTIC)
2702 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2708 * Careful here. We can use a locked bus instruction to
2709 * clear VPTE_A or VPTE_M safely but we need to synchronize
2710 * with the target cpus when we mess with VPTE_EW.
2712 * On virtual kernels we must force a new fault-on-write
2713 * in the real kernel if we clear the Modify bit ourselves,
2714 * otherwise the real kernel will not get a new fault and
2715 * will never set our Modify bit again.
2717 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2719 if (bit == VPTE_RW) {
2721 * We must also clear VPTE_M when clearing
2724 pbits = pmap_clean_pte(pte, pv->pv_pmap,
2728 } else if (bit == VPTE_M) {
2730 * We do not have to make the page read-only
2731 * when clearing the Modify bit. The real
2732 * kernel will make the real PTE read-only
2733 * or otherwise detect the write and set
2734 * our VPTE_M again simply by us invalidating
2735 * the real kernel VA for the pmap (as we did
2736 * above). This allows the real kernel to
2737 * handle the write fault without forwarding
2740 atomic_clear_long(pte, VPTE_M);
2741 } else if ((bit & (VPTE_RW|VPTE_M)) == (VPTE_RW|VPTE_M)) {
2743 * We've been asked to clear W & M, I guess
2744 * the caller doesn't want us to update
2745 * the dirty status of the VM page.
2747 pmap_clean_pte(pte, pv->pv_pmap, pv->pv_va);
2750 * We've been asked to clear bits that do
2751 * not interact with hardware.
2753 atomic_clear_long(pte, bit);
2761 * Lower the permission for all mappings to a given page.
2766 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2768 if ((prot & VM_PROT_WRITE) == 0) {
2769 lwkt_gettoken(&vm_token);
2770 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2771 pmap_clearbit(m, VPTE_RW);
2772 vm_page_flag_clear(m, PG_WRITEABLE);
2776 lwkt_reltoken(&vm_token);
2781 pmap_phys_address(vm_pindex_t ppn)
2783 return (i386_ptob(ppn));
2787 * Return a count of reference bits for a page, clearing those bits.
2788 * It is not necessary for every reference bit to be cleared, but it
2789 * is necessary that 0 only be returned when there are truly no
2790 * reference bits set.
2792 * XXX: The exact number of bits to check and clear is a matter that
2793 * should be tested and standardized at some point in the future for
2794 * optimal aging of shared pages.
2799 pmap_ts_referenced(vm_page_t m)
2801 pv_entry_t pv, pvf, pvn;
2805 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2809 lwkt_gettoken(&vm_token);
2811 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2816 pvn = TAILQ_NEXT(pv, pv_list);
2818 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2820 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2822 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2825 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2827 if (pte && (*pte & VPTE_A)) {
2828 atomic_clear_long(pte, VPTE_A);
2834 } while ((pv = pvn) != NULL && pv != pvf);
2836 lwkt_reltoken(&vm_token);
2843 * Return whether or not the specified physical page was modified
2844 * in any physical maps.
2849 pmap_is_modified(vm_page_t m)
2853 lwkt_gettoken(&vm_token);
2854 res = pmap_testbit(m, VPTE_M);
2855 lwkt_reltoken(&vm_token);
2860 * Clear the modify bits on the specified physical page.
2865 pmap_clear_modify(vm_page_t m)
2867 lwkt_gettoken(&vm_token);
2868 pmap_clearbit(m, VPTE_M);
2869 lwkt_reltoken(&vm_token);
2873 * Clear the reference bit on the specified physical page.
2878 pmap_clear_reference(vm_page_t m)
2880 lwkt_gettoken(&vm_token);
2881 pmap_clearbit(m, VPTE_A);
2882 lwkt_reltoken(&vm_token);
2886 * Miscellaneous support routines follow
2890 i386_protection_init(void)
2894 kp = protection_codes;
2895 for (prot = 0; prot < 8; prot++) {
2896 if (prot & VM_PROT_READ)
2898 if (prot & VM_PROT_WRITE)
2900 if (prot & VM_PROT_EXECUTE)
2909 * Map a set of physical memory pages into the kernel virtual
2910 * address space. Return a pointer to where it is mapped. This
2911 * routine is intended to be used for mapping device memory,
2914 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2918 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2920 vm_offset_t va, tmpva, offset;
2923 offset = pa & PAGE_MASK;
2924 size = roundup(offset + size, PAGE_SIZE);
2926 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
2928 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2930 pa = pa & VPTE_FRAME;
2931 for (tmpva = va; size > 0;) {
2932 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2933 *pte = pa | VPTE_RW | VPTE_V; /* | pgeflag; */
2941 return ((void *)(va + offset));
2945 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2947 vm_offset_t base, offset;
2949 base = va & VPTE_FRAME;
2950 offset = va & PAGE_MASK;
2951 size = roundup(offset + size, PAGE_SIZE);
2952 pmap_qremove(va, size >> PAGE_SHIFT);
2953 kmem_free(&kernel_map, base, size);
2959 * Sets the memory attribute for the specified page.
2962 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
2964 /* This is a vkernel, do nothing */
2968 * Change the PAT attribute on an existing kernel memory map. Caller
2969 * must ensure that the virtual memory in question is not accessed
2970 * during the adjustment.
2973 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
2975 /* This is a vkernel, do nothing */
2979 * Perform the pmap work for mincore
2984 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2990 lwkt_gettoken(&vm_token);
2992 ptep = pmap_pte(pmap, addr);
2994 lwkt_reltoken(&vm_token);
2998 if ((pte = *ptep) != 0) {
3001 val = MINCORE_INCORE;
3002 if ((pte & VPTE_MANAGED) == 0)
3005 pa = pte & VPTE_FRAME;
3007 m = PHYS_TO_VM_PAGE(pa);
3013 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3015 * Modified by someone
3017 else if (m->dirty || pmap_is_modified(m))
3018 val |= MINCORE_MODIFIED_OTHER;
3023 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3026 * Referenced by someone
3028 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3029 val |= MINCORE_REFERENCED_OTHER;
3030 vm_page_flag_set(m, PG_REFERENCED);
3034 lwkt_reltoken(&vm_token);
3039 * Caller must hold vmspace->vm_map.token for oldvm and newvm
3042 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3044 struct vmspace *oldvm;
3047 oldvm = p->p_vmspace;
3049 if (oldvm != newvm) {
3050 p->p_vmspace = newvm;
3051 KKASSERT(p->p_nthreads == 1);
3052 lp = RB_ROOT(&p->p_lwp_tree);
3053 pmap_setlwpvm(lp, newvm);
3055 sysref_get(&newvm->vm_sysref);
3056 sysref_put(&oldvm->vm_sysref);
3063 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3065 struct vmspace *oldvm;
3069 oldvm = lp->lwp_vmspace;
3071 if (oldvm != newvm) {
3072 lp->lwp_vmspace = newvm;
3073 if (curthread->td_lwp == lp) {
3074 pmap = vmspace_pmap(newvm);
3075 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3076 #if defined(SWTCH_OPTIM_STATS)
3079 pmap = vmspace_pmap(oldvm);
3080 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3088 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3091 if ((obj == NULL) || (size < NBPDR) ||
3092 ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
3096 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3101 * Used by kmalloc/kfree, page already exists at va
3104 pmap_kvtom(vm_offset_t va)
3108 KKASSERT(va >= KvaStart && va < KvaEnd);
3109 ptep = KernelPTA + (va >> PAGE_SHIFT);
3110 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
3114 pmap_object_init(vm_object_t object)
3120 pmap_object_free(vm_object_t object)