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>
81 struct pmap kernel_pmap;
83 static struct vm_zone pvzone;
84 static struct vm_object pvzone_obj;
85 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
86 static int pv_entry_count;
87 static int pv_entry_max;
88 static int pv_entry_high_water;
89 static int pmap_pagedaemon_waken;
90 static boolean_t pmap_initialized = FALSE;
91 static int protection_codes[8];
93 static void i386_protection_init(void);
94 static void pmap_remove_all(vm_page_t m);
95 static int pmap_release_free_page(struct pmap *pmap, vm_page_t p);
98 #ifndef PMAP_SHPGPERPROC
99 #define PMAP_SHPGPERPROC 200
102 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
104 #define pte_prot(m, p) \
105 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
111 struct pv_entry *pvinit;
113 for (i = 0; i < vm_page_array_size; i++) {
116 m = &vm_page_array[i];
117 TAILQ_INIT(&m->md.pv_list);
118 m->md.pv_list_count = 0;
121 i = vm_page_array_size;
124 pvinit = (struct pv_entry *)kmem_alloc(&kernel_map, i*sizeof(*pvinit));
125 zbootinit(&pvzone, "PV ENTRY", sizeof(*pvinit), pvinit, i);
126 pmap_initialized = TRUE;
132 int shpgperproc = PMAP_SHPGPERPROC;
134 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
135 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
136 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
137 pv_entry_high_water = 9 * (pv_entry_max / 10);
138 zinitna(&pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
142 * Bootstrap the kernel_pmap so it can be used with pmap_enter().
144 * NOTE! pm_pdir for the kernel pmap is offset so VA's translate
145 * directly into PTD indexes (PTA is also offset for the same reason).
146 * This is necessary because, for now, KVA is not mapped at address 0.
148 * Page table pages are not managed like they are in normal pmaps, so
149 * no pteobj is needed.
154 vm_pindex_t i = (vm_offset_t)KernelPTD >> PAGE_SHIFT;
156 kernel_pmap.pm_pdir = KernelPTD - (KvaStart >> SEG_SHIFT);
157 kernel_pmap.pm_pdirpte = KernelPTA[i];
158 kernel_pmap.pm_count = 1;
159 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
160 TAILQ_INIT(&kernel_pmap.pm_pvlist);
161 i386_protection_init();
165 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
166 * just dummy it up so it works well enough for fork().
168 * In DragonFly, process pmaps may only be used to manipulate user address
169 * space, never kernel address space.
172 pmap_pinit0(struct pmap *pmap)
177 /************************************************************************
178 * Procedures to manage whole physical maps *
179 ************************************************************************
181 * Initialize a preallocated and zeroed pmap structure,
182 * such as one in a vmspace structure.
185 pmap_pinit(struct pmap *pmap)
191 * No need to allocate page table space yet but we do need a valid
192 * page directory table.
194 if (pmap->pm_pdir == NULL) {
196 (vpte_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
200 * allocate object for the pte array and page directory
202 npages = VPTE_PAGETABLE_SIZE +
203 (VM_MAX_USER_ADDRESS / PAGE_SIZE) * sizeof(vpte_t);
204 npages = (npages + PAGE_MASK) / PAGE_SIZE;
206 if (pmap->pm_pteobj == NULL)
207 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, npages);
208 pmap->pm_pdindex = npages - 1;
211 * allocate the page directory page
213 ptdpg = vm_page_grab(pmap->pm_pteobj, pmap->pm_pdindex,
214 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
216 ptdpg->wire_count = 1;
217 ++vmstats.v_wire_count;
219 /* not usually mapped */
220 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
221 ptdpg->valid = VM_PAGE_BITS_ALL;
223 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
224 pmap->pm_pdirpte = KernelPTA[(vm_offset_t)pmap->pm_pdir >> PAGE_SHIFT];
225 if ((ptdpg->flags & PG_ZERO) == 0)
226 bzero(pmap->pm_pdir, PAGE_SIZE);
230 pmap->pm_ptphint = NULL;
231 pmap->pm_cpucachemask = 0;
232 TAILQ_INIT(&pmap->pm_pvlist);
233 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
234 pmap->pm_stats.resident_count = 1;
238 * Clean up a pmap structure so it can be physically freed
243 pmap_puninit(pmap_t pmap)
245 lwkt_gettoken(&vm_token);
247 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
248 pmap->pm_pdir = NULL;
250 if (pmap->pm_pteobj) {
251 vm_object_deallocate(pmap->pm_pteobj);
252 pmap->pm_pteobj = NULL;
254 lwkt_reltoken(&vm_token);
259 * Wire in kernel global address entries. To avoid a race condition
260 * between pmap initialization and pmap_growkernel, this procedure
261 * adds the pmap to the master list (which growkernel scans to update),
262 * then copies the template.
264 * In a virtual kernel there are no kernel global address entries.
269 pmap_pinit2(struct pmap *pmap)
272 lwkt_gettoken(&vm_token);
273 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
274 lwkt_reltoken(&vm_token);
279 * Release all resources held by the given physical map.
281 * Should only be called if the map contains no valid mappings.
285 static int pmap_release_callback(struct vm_page *p, void *data);
288 pmap_release(struct pmap *pmap)
290 struct mdglobaldata *gd = mdcpu;
291 vm_object_t object = pmap->pm_pteobj;
292 struct rb_vm_page_scan_info info;
294 KKASSERT(pmap != &kernel_pmap);
296 #if defined(DIAGNOSTIC)
297 if (object->ref_count != 1)
298 panic("pmap_release: pteobj reference count != 1");
301 * Once we destroy the page table, the mapping becomes invalid.
302 * Don't waste time doing a madvise to invalidate the mapping, just
303 * set cpucachemask to 0.
305 if (pmap->pm_pdir == gd->gd_PT1pdir) {
306 gd->gd_PT1pdir = NULL;
308 /* madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL); */
310 if (pmap->pm_pdir == gd->gd_PT2pdir) {
311 gd->gd_PT2pdir = NULL;
313 /* madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL); */
315 if (pmap->pm_pdir == gd->gd_PT3pdir) {
316 gd->gd_PT3pdir = NULL;
318 /* madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL); */
322 info.object = object;
324 lwkt_gettoken(&vm_token);
325 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
332 info.limit = object->generation;
334 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
335 pmap_release_callback, &info);
336 if (info.error == 0 && info.mpte) {
337 if (!pmap_release_free_page(pmap, info.mpte))
341 } while (info.error);
344 * Leave the KVA reservation for pm_pdir cached for later reuse.
346 pmap->pm_pdirpte = 0;
347 pmap->pm_cpucachemask = 0;
348 lwkt_reltoken(&vm_token);
352 * Callback to release a page table page backing a directory
356 pmap_release_callback(struct vm_page *p, void *data)
358 struct rb_vm_page_scan_info *info = data;
360 if (p->pindex == info->pmap->pm_pdindex) {
364 if (!pmap_release_free_page(info->pmap, p)) {
368 if (info->object->generation != info->limit) {
376 * Retire the given physical map from service. Should only be called if
377 * the map contains no valid mappings.
382 pmap_destroy(pmap_t pmap)
387 lwkt_gettoken(&vm_token);
388 if (--pmap->pm_count == 0) {
390 panic("destroying a pmap is not yet implemented");
392 lwkt_reltoken(&vm_token);
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_R | VPTE_W | 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_R | VPTE_W;
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_R | VPTE_W | 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.
755 pmap_kenter_temporary(vm_paddr_t pa, long i)
757 pmap_kenter(crashdumpmap + (i * PAGE_SIZE), pa);
758 return ((void *)crashdumpmap);
762 * Remove an unmanaged mapping created with pmap_kenter*().
765 pmap_kremove(vm_offset_t va)
769 KKASSERT(va >= KvaStart && va < KvaEnd);
771 ptep = KernelPTA + (va >> PAGE_SHIFT);
773 pmap_inval_pte(ptep, &kernel_pmap, va);
778 * Remove an unmanaged mapping created with pmap_kenter*() but synchronize
779 * only with this cpu.
781 * Unfortunately because we optimize new entries by testing VPTE_V later
782 * on, we actually still have to synchronize with all the cpus. XXX maybe
783 * store a junk value and test against 0 in the other places instead?
786 pmap_kremove_quick(vm_offset_t va)
790 KKASSERT(va >= KvaStart && va < KvaEnd);
792 ptep = KernelPTA + (va >> PAGE_SHIFT);
794 pmap_inval_pte(ptep, &kernel_pmap, va); /* NOT _quick */
799 * Extract the physical address from the kernel_pmap that is associated
800 * with the specified virtual address.
803 pmap_kextract(vm_offset_t va)
808 KKASSERT(va >= KvaStart && va < KvaEnd);
810 ptep = KernelPTA + (va >> PAGE_SHIFT);
811 pa = (vm_paddr_t)(*ptep & VPTE_FRAME) | (va & PAGE_MASK);
816 * Map a set of unmanaged VM pages into KVM.
819 pmap_qenter(vm_offset_t va, struct vm_page **m, int count)
821 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
825 ptep = KernelPTA + (va >> PAGE_SHIFT);
827 pmap_inval_pte(ptep, &kernel_pmap, va);
828 *ptep = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
836 * Undo the effects of pmap_qenter*().
839 pmap_qremove(vm_offset_t va, int count)
841 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
845 ptep = KernelPTA + (va >> PAGE_SHIFT);
847 pmap_inval_pte(ptep, &kernel_pmap, va);
854 /************************************************************************
855 * Misc support glue called by machine independant code *
856 ************************************************************************
858 * These routines are called by machine independant code to operate on
859 * certain machine-dependant aspects of processes, threads, and pmaps.
863 * Initialize MD portions of the thread structure.
866 pmap_init_thread(thread_t td)
868 /* enforce pcb placement */
869 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
870 td->td_savefpu = &td->td_pcb->pcb_save;
871 td->td_sp = (char *)td->td_pcb - 16;
875 * This routine directly affects the fork perf for a process.
878 pmap_init_proc(struct proc *p)
883 * Destroy the UPAGES for a process that has exited and disassociate
884 * the process from its thread.
887 pmap_dispose_proc(struct proc *p)
889 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
893 * We pre-allocate all page table pages for kernel virtual memory so
894 * this routine will only be called if KVM has been exhausted.
899 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
903 addr = (kend + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
905 lwkt_gettoken(&vm_token);
906 if (addr > virtual_end - SEG_SIZE)
907 panic("KVM exhausted");
908 kernel_vm_end = addr;
909 lwkt_reltoken(&vm_token);
913 * The modification bit is not tracked for any pages in this range. XXX
914 * such pages in this maps should always use pmap_k*() functions and not
917 * XXX User and kernel address spaces are independant for virtual kernels,
918 * this function only applies to the kernel pmap.
921 pmap_track_modified(pmap_t pmap, vm_offset_t va)
923 if (pmap != &kernel_pmap)
925 if ((va < clean_sva) || (va >= clean_eva))
931 /************************************************************************
932 * Procedures supporting managed page table pages *
933 ************************************************************************
935 * These procedures are used to track managed page table pages. These pages
936 * use the page table page's vm_page_t to track PTEs in the page. The
937 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
939 * This allows the system to throw away page table pages for user processes
940 * at will and reinstantiate them on demand.
944 * This routine works like vm_page_lookup() but also blocks as long as the
945 * page is busy. This routine does not busy the page it returns.
947 * Unless the caller is managing objects whos pages are in a known state,
948 * the call should be made with a critical section held so the page's object
949 * association remains valid on return.
952 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
957 m = vm_page_lookup(object, pindex);
958 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
964 * This routine unholds page table pages, and if the hold count
965 * drops to zero, then it decrements the wire count.
967 * We must recheck that this is the last hold reference after busy-sleeping
971 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
973 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
975 KASSERT(m->queue == PQ_NONE,
976 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
978 if (m->hold_count == 1) {
980 * Unmap the page table page.
983 KKASSERT(pmap->pm_pdir[m->pindex] != 0);
984 pmap_inval_pde(&pmap->pm_pdir[m->pindex], pmap,
985 (vm_offset_t)m->pindex << SEG_SHIFT);
986 KKASSERT(pmap->pm_stats.resident_count > 0);
987 --pmap->pm_stats.resident_count;
989 if (pmap->pm_ptphint == m)
990 pmap->pm_ptphint = NULL;
993 * This was our last hold, the page had better be unwired
994 * after we decrement wire_count.
996 * FUTURE NOTE: shared page directory page could result in
997 * multiple wire counts.
1001 KKASSERT(m->wire_count == 0);
1002 --vmstats.v_wire_count;
1003 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1005 vm_page_free_zero(m);
1008 KKASSERT(m->hold_count > 1);
1014 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
1016 KKASSERT(m->hold_count > 0);
1017 if (m->hold_count > 1) {
1021 return _pmap_unwire_pte_hold(pmap, m);
1026 * After removing a page table entry, this routine is used to
1027 * conditionally free the page, and manage the hold/wire counts.
1030 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
1036 * page table pages in the kernel_pmap are not managed.
1038 if (pmap == &kernel_pmap)
1040 ptepindex = (va >> PDRSHIFT);
1041 if (pmap->pm_ptphint &&
1042 (pmap->pm_ptphint->pindex == ptepindex)) {
1043 mpte = pmap->pm_ptphint;
1045 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1046 pmap->pm_ptphint = mpte;
1049 return pmap_unwire_pte_hold(pmap, mpte);
1053 * Attempt to release and free the vm_page backing a page directory page
1054 * in a pmap. Returns 1 on success, 0 on failure (if the procedure had
1058 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1060 vpte_t *pde = pmap->pm_pdir;
1063 * This code optimizes the case of freeing non-busy
1064 * page-table pages. Those pages are zero now, and
1065 * might as well be placed directly into the zero queue.
1067 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1071 KKASSERT(pmap->pm_stats.resident_count > 0);
1072 --pmap->pm_stats.resident_count;
1074 if (p->hold_count) {
1075 panic("pmap_release: freeing held page table page");
1078 * Page directory pages need to have the kernel stuff cleared, so
1079 * they can go into the zero queue also.
1081 * In virtual kernels there is no 'kernel stuff'. For the moment
1082 * I just make sure the whole thing has been zero'd even though
1083 * it should already be completely zero'd.
1085 * pmaps for vkernels do not self-map because they do not share
1086 * their address space with the vkernel. Clearing of pde[] thus
1087 * only applies to page table pages and not to the page directory
1090 if (p->pindex == pmap->pm_pdindex) {
1091 bzero(pde, VPTE_PAGETABLE_SIZE);
1092 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1094 KKASSERT(pde[p->pindex] != 0);
1095 pmap_inval_pde(&pde[p->pindex], pmap,
1096 (vm_offset_t)p->pindex << SEG_SHIFT);
1100 * Clear the matching hint
1102 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1103 pmap->pm_ptphint = NULL;
1106 * And throw the page away. The page is completely zero'd out so
1107 * optimize the free call.
1110 vmstats.v_wire_count--;
1111 vm_page_free_zero(p);
1116 * This routine is called if the page table page is not mapped in the page
1119 * The routine is broken up into two parts for readability.
1121 * It must return a held mpte and map the page directory page as required.
1122 * Because vm_page_grab() can block, we must re-check pm_pdir[ptepindex]
1125 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1131 * Find or fabricate a new pagetable page. A busied page will be
1132 * returned. This call may block.
1134 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1135 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1137 KASSERT(m->queue == PQ_NONE,
1138 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1141 * Increment the hold count for the page we will be returning to
1147 * It is possible that someone else got in and mapped by the page
1148 * directory page while we were blocked, if so just unbusy and
1149 * return the held page.
1151 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1152 KKASSERT((ptepa & VPTE_FRAME) == VM_PAGE_TO_PHYS(m));
1157 if (m->wire_count == 0)
1158 vmstats.v_wire_count++;
1162 * Map the pagetable page into the process address space, if
1163 * it isn't already there.
1165 ++pmap->pm_stats.resident_count;
1167 ptepa = VM_PAGE_TO_PHYS(m);
1168 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
1172 * We are likely about to access this page table page, so set the
1173 * page table hint to reduce overhead.
1175 pmap->pm_ptphint = m;
1178 * Try to use the new mapping, but if we cannot, then
1179 * do it with the routine that maps the page explicitly.
1181 if ((m->flags & PG_ZERO) == 0)
1182 pmap_zero_page(ptepa);
1184 m->valid = VM_PAGE_BITS_ALL;
1185 vm_page_flag_clear(m, PG_ZERO);
1186 vm_page_flag_set(m, PG_MAPPED);
1193 * Determine the page table page required to access the VA in the pmap
1194 * and allocate it if necessary. Return a held vm_page_t for the page.
1196 * Only used with user pmaps.
1199 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1206 * Calculate pagetable page index
1208 ptepindex = va >> PDRSHIFT;
1211 * Get the page directory entry
1213 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1216 * This supports switching from a 4MB page to a
1219 if (ptepa & VPTE_PS) {
1220 KKASSERT(pmap->pm_pdir[ptepindex] != 0);
1221 pmap_inval_pde(&pmap->pm_pdir[ptepindex], pmap,
1222 (vm_offset_t)ptepindex << SEG_SHIFT);
1227 * If the page table page is mapped, we just increment the
1228 * hold count, and activate it.
1232 * In order to get the page table page, try the
1235 if (pmap->pm_ptphint &&
1236 (pmap->pm_ptphint->pindex == ptepindex)) {
1237 m = pmap->pm_ptphint;
1239 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1240 pmap->pm_ptphint = m;
1246 * Here if the pte page isn't mapped, or if it has been deallocated.
1248 return _pmap_allocpte(pmap, ptepindex);
1251 /************************************************************************
1252 * Managed pages in pmaps *
1253 ************************************************************************
1255 * All pages entered into user pmaps and some pages entered into the kernel
1256 * pmap are managed, meaning that pmap_protect() and other related management
1257 * functions work on these pages.
1261 * free the pv_entry back to the free list. This function may be
1262 * called from an interrupt.
1264 static __inline void
1265 free_pv_entry(pv_entry_t pv)
1272 * get a new pv_entry, allocating a block from the system
1273 * when needed. This function may be called from an interrupt.
1279 if (pv_entry_high_water &&
1280 (pv_entry_count > pv_entry_high_water) &&
1281 (pmap_pagedaemon_waken == 0)) {
1282 pmap_pagedaemon_waken = 1;
1283 wakeup (&vm_pages_needed);
1285 return zalloc(&pvzone);
1289 * This routine is very drastic, but can save the system
1299 static int warningdone=0;
1301 if (pmap_pagedaemon_waken == 0)
1303 lwkt_gettoken(&vm_token);
1304 pmap_pagedaemon_waken = 0;
1306 if (warningdone < 5) {
1307 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1311 for(i = 0; i < vm_page_array_size; i++) {
1312 m = &vm_page_array[i];
1313 if (m->wire_count || m->hold_count || m->busy ||
1314 (m->flags & PG_BUSY))
1318 lwkt_reltoken(&vm_token);
1322 * If it is the first entry on the list, it is actually
1323 * in the header and we must copy the following entry up
1324 * to the header. Otherwise we must search the list for
1325 * the entry. In either case we free the now unused entry.
1328 pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va)
1334 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1335 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1336 if (pmap == pv->pv_pmap && va == pv->pv_va)
1340 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1341 if (va == pv->pv_va)
1347 * Note that pv_ptem is NULL if the page table page itself is not
1348 * managed, even if the page being removed IS managed.
1352 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1353 m->md.pv_list_count--;
1354 m->object->agg_pv_list_count--;
1355 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1356 if (TAILQ_EMPTY(&m->md.pv_list))
1357 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1358 ++pmap->pm_generation;
1359 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1367 * Create a pv entry for page at pa for (pmap, va). If the page table page
1368 * holding the VA is managed, mpte will be non-NULL.
1371 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1376 pv = get_pv_entry();
1381 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1382 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1383 ++pmap->pm_generation;
1384 m->md.pv_list_count++;
1385 m->object->agg_pv_list_count++;
1391 * pmap_remove_pte: do the things to unmap a page in a process
1394 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va)
1399 oldpte = pmap_inval_loadandclear(ptq, pmap, va);
1400 if (oldpte & VPTE_WIRED)
1401 --pmap->pm_stats.wired_count;
1402 KKASSERT(pmap->pm_stats.wired_count >= 0);
1406 * Machines that don't support invlpg, also don't support
1407 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1410 if (oldpte & VPTE_G)
1411 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1413 KKASSERT(pmap->pm_stats.resident_count > 0);
1414 --pmap->pm_stats.resident_count;
1415 if (oldpte & VPTE_MANAGED) {
1416 m = PHYS_TO_VM_PAGE(oldpte);
1417 if (oldpte & VPTE_M) {
1418 #if defined(PMAP_DIAGNOSTIC)
1419 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1421 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1425 if (pmap_track_modified(pmap, va))
1428 if (oldpte & VPTE_A)
1429 vm_page_flag_set(m, PG_REFERENCED);
1430 return pmap_remove_entry(pmap, m, va);
1432 return pmap_unuse_pt(pmap, va, NULL);
1441 * Remove a single page from a process address space.
1443 * This function may not be called from an interrupt if the pmap is
1447 pmap_remove_page(struct pmap *pmap, vm_offset_t va)
1452 * if there is no pte for this address, just skip it!!! Otherwise
1453 * get a local va for mappings for this pmap and remove the entry.
1455 if (*pmap_pde(pmap, va) != 0) {
1456 ptq = get_ptbase(pmap, va);
1458 pmap_remove_pte(pmap, ptq, va);
1464 * Remove the given range of addresses from the specified map.
1466 * It is assumed that the start and end are properly rounded to the
1469 * This function may not be called from an interrupt if the pmap is
1475 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1479 vm_offset_t ptpaddr;
1480 vm_pindex_t sindex, eindex;
1485 lwkt_gettoken(&vm_token);
1486 KKASSERT(pmap->pm_stats.resident_count >= 0);
1487 if (pmap->pm_stats.resident_count == 0) {
1488 lwkt_reltoken(&vm_token);
1493 * special handling of removing one page. a very
1494 * common operation and easy to short circuit some
1497 if (((sva + PAGE_SIZE) == eva) &&
1498 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1499 pmap_remove_page(pmap, sva);
1500 lwkt_reltoken(&vm_token);
1505 * Get a local virtual address for the mappings that are being
1508 * XXX this is really messy because the kernel pmap is not relative
1511 sindex = (sva >> PAGE_SHIFT);
1512 eindex = (eva >> PAGE_SHIFT);
1514 for (; sindex < eindex; sindex = pdnxt) {
1518 * Calculate index for next page table.
1520 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1521 if (pmap->pm_stats.resident_count == 0)
1524 pdirindex = sindex / NPDEPG;
1525 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1526 KKASSERT(pmap->pm_pdir[pdirindex] != 0);
1527 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1528 pmap_inval_pde(&pmap->pm_pdir[pdirindex], pmap,
1529 (vm_offset_t)pdirindex << SEG_SHIFT);
1534 * Weed out invalid mappings. Note: we assume that the page
1535 * directory table is always allocated, and in kernel virtual.
1541 * Limit our scan to either the end of the va represented
1542 * by the current page table page, or to the end of the
1543 * range being removed.
1549 * NOTE: pmap_remove_pte() can block.
1551 for (; sindex != pdnxt; sindex++) {
1554 ptbase = get_ptbase(pmap, sindex << PAGE_SHIFT);
1557 va = i386_ptob(sindex);
1558 if (pmap_remove_pte(pmap, ptbase, va))
1562 lwkt_reltoken(&vm_token);
1566 * Removes this physical page from all physical maps in which it resides.
1567 * Reflects back modify bits to the pager.
1569 * This routine may not be called from an interrupt.
1574 pmap_remove_all(vm_page_t m)
1579 #if defined(PMAP_DIAGNOSTIC)
1581 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1584 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1585 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1590 lwkt_gettoken(&vm_token);
1591 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1592 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1593 --pv->pv_pmap->pm_stats.resident_count;
1595 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1596 KKASSERT(pte != NULL);
1598 tpte = pmap_inval_loadandclear(pte, pv->pv_pmap, pv->pv_va);
1599 if (tpte & VPTE_WIRED)
1600 --pv->pv_pmap->pm_stats.wired_count;
1601 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
1604 vm_page_flag_set(m, PG_REFERENCED);
1607 * Update the vm_page_t clean and reference bits.
1609 if (tpte & VPTE_M) {
1610 #if defined(PMAP_DIAGNOSTIC)
1611 if (pmap_nw_modified((pt_entry_t) tpte)) {
1613 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1617 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
1620 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1621 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1622 ++pv->pv_pmap->pm_generation;
1623 m->md.pv_list_count--;
1624 m->object->agg_pv_list_count--;
1625 if (TAILQ_EMPTY(&m->md.pv_list))
1626 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1627 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
1630 KKASSERT((m->flags & (PG_MAPPED | PG_WRITEABLE)) == 0);
1631 lwkt_reltoken(&vm_token);
1636 * Set the physical protection on the specified range of this map
1639 * This function may not be called from an interrupt if the map is
1640 * not the kernel_pmap.
1645 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1649 vm_offset_t pdnxt, ptpaddr;
1650 vm_pindex_t sindex, eindex;
1656 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1657 pmap_remove(pmap, sva, eva);
1661 if (prot & VM_PROT_WRITE)
1664 lwkt_gettoken(&vm_token);
1665 ptbase = get_ptbase(pmap, sva);
1667 sindex = (sva >> PAGE_SHIFT);
1668 eindex = (eva >> PAGE_SHIFT);
1671 for (; sindex < eindex; sindex = pdnxt) {
1675 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1677 pdirindex = sindex / NPDEPG;
1680 * Clear the modified and writable bits for a 4m page.
1681 * Throw away the modified bit (?)
1683 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1684 pmap_clean_pde(&pmap->pm_pdir[pdirindex], pmap,
1685 (vm_offset_t)pdirindex << SEG_SHIFT);
1686 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1691 * Weed out invalid mappings. Note: we assume that the page
1692 * directory table is always allocated, and in kernel virtual.
1697 if (pdnxt > eindex) {
1701 for (; sindex != pdnxt; sindex++) {
1706 * Clean managed pages and also check the accessed
1707 * bit. Just remove write perms for unmanaged
1708 * pages. Be careful of races, turning off write
1709 * access will force a fault rather then setting
1710 * the modified bit at an unexpected time.
1712 ptep = &ptbase[sindex - sbase];
1713 if (*ptep & VPTE_MANAGED) {
1714 pbits = pmap_clean_pte(ptep, pmap,
1717 if (pbits & VPTE_A) {
1718 m = PHYS_TO_VM_PAGE(pbits);
1719 vm_page_flag_set(m, PG_REFERENCED);
1720 atomic_clear_long(ptep, VPTE_A);
1722 if (pbits & VPTE_M) {
1723 if (pmap_track_modified(pmap, i386_ptob(sindex))) {
1725 m = PHYS_TO_VM_PAGE(pbits);
1730 pbits = pmap_setro_pte(ptep, pmap,
1735 lwkt_reltoken(&vm_token);
1739 * Enter a managed page into a pmap. If the page is not wired related pmap
1740 * data can be destroyed at any time for later demand-operation.
1742 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1743 * specified protection, and wire the mapping if requested.
1745 * NOTE: This routine may not lazy-evaluate or lose information. The
1746 * page must actually be inserted into the given map NOW.
1748 * NOTE: When entering a page at a KVA address, the pmap must be the
1754 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1760 vpte_t origpte, newpte;
1768 lwkt_gettoken(&vm_token);
1771 * Get the page table page. The kernel_pmap's page table pages
1772 * are preallocated and have no associated vm_page_t.
1774 if (pmap == &kernel_pmap)
1777 mpte = pmap_allocpte(pmap, va);
1779 pte = pmap_pte(pmap, va);
1782 * Page Directory table entry not valid, we need a new PT page
1783 * and pmap_allocpte() didn't give us one. Oops!
1786 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1791 * Deal with races on the original mapping (though don't worry
1792 * about VPTE_A races) by cleaning it. This will force a fault
1793 * if an attempt is made to write to the page.
1795 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1796 origpte = pmap_clean_pte(pte, pmap, va);
1797 opa = origpte & VPTE_FRAME;
1799 if (origpte & VPTE_PS)
1800 panic("pmap_enter: attempted pmap_enter on 4MB page");
1803 * Mapping has not changed, must be protection or wiring change.
1805 if (origpte && (opa == pa)) {
1807 * Wiring change, just update stats. We don't worry about
1808 * wiring PT pages as they remain resident as long as there
1809 * are valid mappings in them. Hence, if a user page is wired,
1810 * the PT page will be also.
1812 if (wired && ((origpte & VPTE_WIRED) == 0))
1813 ++pmap->pm_stats.wired_count;
1814 else if (!wired && (origpte & VPTE_WIRED))
1815 --pmap->pm_stats.wired_count;
1816 KKASSERT(pmap->pm_stats.wired_count >= 0);
1819 * Remove the extra pte reference. Note that we cannot
1820 * optimize the RO->RW case because we have adjusted the
1821 * wiring count above and may need to adjust the wiring
1828 * We might be turning off write access to the page,
1829 * so we go ahead and sense modify status.
1831 if (origpte & VPTE_MANAGED) {
1832 if ((origpte & VPTE_M) &&
1833 pmap_track_modified(pmap, va)) {
1835 om = PHYS_TO_VM_PAGE(opa);
1839 KKASSERT(m->flags & PG_MAPPED);
1844 * Mapping has changed, invalidate old range and fall through to
1845 * handle validating new mapping.
1849 err = pmap_remove_pte(pmap, pte, va);
1851 panic("pmap_enter: pte vanished, va: %p", (void *)va);
1852 pte = pmap_pte(pmap, va);
1853 origpte = pmap_clean_pte(pte, pmap, va);
1854 opa = origpte & VPTE_FRAME;
1856 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
1862 * Enter on the PV list if part of our managed memory. Note that we
1863 * raise IPL while manipulating pv_table since pmap_enter can be
1864 * called at interrupt time.
1866 if (pmap_initialized &&
1867 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1868 pmap_insert_entry(pmap, va, mpte, m);
1870 vm_page_flag_set(m, PG_MAPPED);
1874 * Increment counters
1876 ++pmap->pm_stats.resident_count;
1878 pmap->pm_stats.wired_count++;
1882 * Now validate mapping with desired protection/wiring.
1884 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1887 newpte |= VPTE_WIRED;
1888 if (pmap != &kernel_pmap)
1892 * If the mapping or permission bits are different from the
1893 * (now cleaned) original pte, an update is needed. We've
1894 * already downgraded or invalidated the page so all we have
1895 * to do now is update the bits.
1897 * XXX should we synchronize RO->RW changes to avoid another
1900 if ((origpte & ~(VPTE_W|VPTE_M|VPTE_A)) != newpte) {
1901 *pte = newpte | VPTE_A;
1902 if (newpte & VPTE_W)
1903 vm_page_flag_set(m, PG_WRITEABLE);
1905 KKASSERT((newpte & VPTE_MANAGED) == 0 || m->flags & PG_MAPPED);
1906 lwkt_reltoken(&vm_token);
1910 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
1912 * Currently this routine may only be used on user pmaps, not kernel_pmap.
1915 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
1923 KKASSERT(pmap != &kernel_pmap);
1925 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1928 * Calculate pagetable page (mpte), allocating it if necessary.
1930 * A held page table page (mpte), or NULL, is passed onto the
1931 * section following.
1933 ptepindex = va >> PDRSHIFT;
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;
1960 mpte = _pmap_allocpte(pmap, ptepindex);
1962 } while (mpte == NULL);
1965 * Ok, now that the page table page has been validated, get the pte.
1966 * If the pte is already mapped undo mpte's hold_count and
1969 pte = pmap_pte(pmap, va);
1971 pmap_unwire_pte_hold(pmap, mpte);
1972 lwkt_reltoken(&vm_token);
1977 * Enter on the PV list if part of our managed memory. Note that we
1978 * raise IPL while manipulating pv_table since pmap_enter can be
1979 * called at interrupt time.
1981 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1982 pmap_insert_entry(pmap, va, mpte, m);
1983 vm_page_flag_set(m, PG_MAPPED);
1987 * Increment counters
1989 ++pmap->pm_stats.resident_count;
1991 pa = VM_PAGE_TO_PHYS(m);
1994 * Now validate mapping with RO protection
1996 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1997 *pte = (vpte_t)pa | VPTE_V | VPTE_U;
1999 *pte = (vpte_t)pa | VPTE_V | VPTE_U | VPTE_MANAGED;
2000 /*pmap_inval_add(&info, pmap, va); shouldn't be needed 0->valid */
2001 /*pmap_inval_flush(&info); don't need for vkernel */
2002 lwkt_reltoken(&vm_token);
2006 * Extract the physical address for the translation at the specified
2007 * virtual address in the pmap.
2009 * The caller must hold vm_token if non-blocking operation is desired.
2013 pmap_extract(pmap_t pmap, vm_offset_t va)
2018 lwkt_gettoken(&vm_token);
2019 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
2020 if (pte & VPTE_PS) {
2021 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
2022 rtval |= va & SEG_MASK;
2024 pte = *get_ptbase(pmap, va);
2025 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
2030 lwkt_reltoken(&vm_token);
2034 #define MAX_INIT_PT (96)
2037 * This routine preloads the ptes for a given object into the specified pmap.
2038 * This eliminates the blast of soft faults on process startup and
2039 * immediately after an mmap.
2043 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2046 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2047 vm_object_t object, vm_pindex_t pindex,
2048 vm_size_t size, int limit)
2050 struct rb_vm_page_scan_info info;
2055 * We can't preinit if read access isn't set or there is no pmap
2058 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2062 * We can't preinit if the pmap is not the current pmap
2064 lp = curthread->td_lwp;
2065 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2068 psize = size >> PAGE_SHIFT;
2070 if ((object->type != OBJT_VNODE) ||
2071 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2072 (object->resident_page_count > MAX_INIT_PT))) {
2076 if (psize + pindex > object->size) {
2077 if (object->size < pindex)
2079 psize = object->size - pindex;
2086 * Use a red-black scan to traverse the requested range and load
2087 * any valid pages found into the pmap.
2089 * We cannot safely scan the object's memq unless we are in a
2090 * critical section since interrupts can remove pages from objects.
2092 info.start_pindex = pindex;
2093 info.end_pindex = pindex + psize - 1;
2100 lwkt_gettoken(&vm_token);
2101 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2102 pmap_object_init_pt_callback, &info);
2103 lwkt_reltoken(&vm_token);
2108 * The caller must hold vm_token.
2112 pmap_object_init_pt_callback(vm_page_t p, void *data)
2114 struct rb_vm_page_scan_info *info = data;
2115 vm_pindex_t rel_index;
2117 * don't allow an madvise to blow away our really
2118 * free pages allocating pv entries.
2120 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2121 vmstats.v_free_count < vmstats.v_free_reserved) {
2124 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2125 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2126 if ((p->queue - p->pc) == PQ_CACHE)
2127 vm_page_deactivate(p);
2129 rel_index = p->pindex - info->start_pindex;
2130 pmap_enter_quick(info->pmap,
2131 info->addr + i386_ptob(rel_index), p);
2138 * Return TRUE if the pmap is in shape to trivially
2139 * pre-fault the specified address.
2141 * Returns FALSE if it would be non-trivial or if a
2142 * pte is already loaded into the slot.
2147 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2152 lwkt_gettoken(&vm_token);
2153 if ((*pmap_pde(pmap, addr)) == 0) {
2156 pte = get_ptbase(pmap, addr);
2157 ret = (*pte) ? 0 : 1;
2159 lwkt_reltoken(&vm_token);
2164 * Change the wiring attribute for a map/virtual-address pair.
2165 * The mapping must already exist in the pmap.
2167 * No other requirements.
2170 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2177 lwkt_gettoken(&vm_token);
2178 pte = get_ptbase(pmap, va);
2180 if (wired && (*pte & VPTE_WIRED) == 0)
2181 ++pmap->pm_stats.wired_count;
2182 else if (!wired && (*pte & VPTE_WIRED))
2183 --pmap->pm_stats.wired_count;
2184 KKASSERT(pmap->pm_stats.wired_count >= 0);
2187 * Wiring is not a hardware characteristic so there is no need to
2188 * invalidate TLB. However, in an SMP environment we must use
2189 * a locked bus cycle to update the pte (if we are not using
2190 * the pmap_inval_*() API that is)... it's ok to do this for simple
2194 atomic_set_long(pte, VPTE_WIRED);
2196 atomic_clear_long(pte, VPTE_WIRED);
2197 lwkt_reltoken(&vm_token);
2201 * Copy the range specified by src_addr/len
2202 * from the source map to the range dst_addr/len
2203 * in the destination map.
2205 * This routine is only advisory and need not do anything.
2208 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2209 vm_size_t len, vm_offset_t src_addr)
2212 vm_offset_t end_addr = src_addr + len;
2219 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2220 * valid through blocking calls, and that's just not going to
2227 if (dst_addr != src_addr)
2229 if (dst_pmap->pm_pdir == NULL)
2231 if (src_pmap->pm_pdir == NULL)
2236 src_frame = get_ptbase1(src_pmap, src_addr);
2237 dst_frame = get_ptbase2(dst_pmap, src_addr);
2240 * critical section protection is required to maintain the page/object
2241 * association, interrupts can free pages and remove them from
2244 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2245 vpte_t *src_pte, *dst_pte;
2246 vm_page_t dstmpte, srcmpte;
2247 vm_offset_t srcptepaddr;
2250 if (addr >= VM_MAX_USER_ADDRESS)
2251 panic("pmap_copy: invalid to pmap_copy page tables\n");
2254 * Don't let optional prefaulting of pages make us go
2255 * way below the low water mark of free pages or way
2256 * above high water mark of used pv entries.
2258 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2259 pv_entry_count > pv_entry_high_water)
2262 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2263 ptepindex = addr >> PDRSHIFT;
2265 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2266 if (srcptepaddr == 0)
2269 if (srcptepaddr & VPTE_PS) {
2270 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2271 dst_pmap->pm_pdir[ptepindex] = (vpte_t)srcptepaddr;
2272 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2277 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2278 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2279 (srcmpte->flags & PG_BUSY)) {
2283 if (pdnxt > end_addr)
2286 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2287 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2288 while (addr < pdnxt) {
2293 * we only virtual copy managed pages
2295 if ((ptetemp & VPTE_MANAGED) != 0) {
2297 * We have to check after allocpte for the
2298 * pte still being around... allocpte can
2301 * pmap_allocpte can block, unfortunately
2302 * we have to reload the tables.
2304 dstmpte = pmap_allocpte(dst_pmap, addr);
2305 src_frame = get_ptbase1(src_pmap, src_addr);
2306 dst_frame = get_ptbase2(dst_pmap, src_addr);
2308 if ((*dst_pte == 0) && (ptetemp = *src_pte) &&
2309 (ptetemp & VPTE_MANAGED) != 0) {
2311 * Clear the modified and accessed
2312 * (referenced) bits during the copy.
2314 * We do not have to clear the write
2315 * bit to force a fault-on-modify
2316 * because the real kernel's target
2317 * pmap is empty and will fault anyway.
2319 m = PHYS_TO_VM_PAGE(ptetemp);
2320 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2321 ++dst_pmap->pm_stats.resident_count;
2322 pmap_insert_entry(dst_pmap, addr,
2324 KKASSERT(m->flags & PG_MAPPED);
2326 pmap_unwire_pte_hold(dst_pmap, dstmpte);
2328 if (dstmpte->hold_count >= srcmpte->hold_count)
2342 * Zero the specified PA by mapping the page into KVM and clearing its
2345 * This function may be called from an interrupt and no locking is
2349 pmap_zero_page(vm_paddr_t phys)
2351 struct mdglobaldata *gd = mdcpu;
2355 panic("pmap_zero_page: CMAP3 busy");
2356 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2357 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2359 bzero(gd->gd_CADDR3, PAGE_SIZE);
2365 * pmap_page_assertzero:
2367 * Assert that a page is empty, panic if it isn't.
2370 pmap_page_assertzero(vm_paddr_t phys)
2372 struct mdglobaldata *gd = mdcpu;
2377 panic("pmap_zero_page: CMAP3 busy");
2378 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2379 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2380 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2381 for (i = 0; i < PAGE_SIZE; i += 4) {
2382 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2383 panic("pmap_page_assertzero() @ %p not zero!\n",
2384 (void *)gd->gd_CADDR3);
2394 * Zero part of a physical page by mapping it into memory and clearing
2395 * its contents with bzero.
2397 * off and size may not cover an area beyond a single hardware page.
2400 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2402 struct mdglobaldata *gd = mdcpu;
2406 panic("pmap_zero_page: CMAP3 busy");
2407 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2408 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2409 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2411 bzero((char *)gd->gd_CADDR3 + off, size);
2419 * Copy the physical page from the source PA to the target PA.
2420 * This function may be called from an interrupt. No locking
2424 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2426 struct mdglobaldata *gd = mdcpu;
2429 if (*(int *) gd->gd_CMAP1)
2430 panic("pmap_copy_page: CMAP1 busy");
2431 if (*(int *) gd->gd_CMAP2)
2432 panic("pmap_copy_page: CMAP2 busy");
2434 *(int *) gd->gd_CMAP1 = VPTE_V | VPTE_R | (src & PG_FRAME) | VPTE_A;
2435 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2437 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2438 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2440 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2442 *(int *) gd->gd_CMAP1 = 0;
2443 *(int *) gd->gd_CMAP2 = 0;
2448 * pmap_copy_page_frag:
2450 * Copy the physical page from the source PA to the target PA.
2451 * This function may be called from an interrupt. No locking
2455 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2457 struct mdglobaldata *gd = mdcpu;
2460 if (*(int *) gd->gd_CMAP1)
2461 panic("pmap_copy_page: CMAP1 busy");
2462 if (*(int *) gd->gd_CMAP2)
2463 panic("pmap_copy_page: CMAP2 busy");
2465 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2466 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2468 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2469 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2471 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2472 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2475 *(int *) gd->gd_CMAP1 = 0;
2476 *(int *) gd->gd_CMAP2 = 0;
2481 * Returns true if the pmap's pv is one of the first
2482 * 16 pvs linked to from this page. This count may
2483 * be changed upwards or downwards in the future; it
2484 * is only necessary that true be returned for a small
2485 * subset of pmaps for proper page aging.
2490 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2495 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2499 lwkt_gettoken(&vm_token);
2501 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2502 if (pv->pv_pmap == pmap) {
2503 lwkt_reltoken(&vm_token);
2511 lwkt_reltoken(&vm_token);
2517 * Remove all pages from specified address space
2518 * this aids process exit speeds. Also, this code
2519 * is special cased for current process only, but
2520 * can have the more generic (and slightly slower)
2521 * mode enabled. This is much faster than pmap_remove
2522 * in the case of running down an entire address space.
2527 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2532 int32_t save_generation;
2535 lwkt_gettoken(&vm_token);
2536 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2537 if (pv->pv_va >= eva || pv->pv_va < sva) {
2538 npv = TAILQ_NEXT(pv, pv_plist);
2542 KKASSERT(pmap == pv->pv_pmap);
2544 pte = pmap_pte(pmap, pv->pv_va);
2547 * We cannot remove wired pages from a process' mapping
2550 if (*pte & VPTE_WIRED) {
2551 npv = TAILQ_NEXT(pv, pv_plist);
2554 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2556 m = PHYS_TO_VM_PAGE(tpte);
2558 KASSERT(m < &vm_page_array[vm_page_array_size],
2559 ("pmap_remove_pages: bad tpte %lx", tpte));
2561 KKASSERT(pmap->pm_stats.resident_count > 0);
2562 --pmap->pm_stats.resident_count;
2565 * Update the vm_page_t clean and reference bits.
2567 if (tpte & VPTE_M) {
2571 npv = TAILQ_NEXT(pv, pv_plist);
2572 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2573 save_generation = ++pmap->pm_generation;
2575 m->md.pv_list_count--;
2576 m->object->agg_pv_list_count--;
2577 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2578 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2579 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2581 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2585 * Restart the scan if we blocked during the unuse or free
2586 * calls and other removals were made.
2588 if (save_generation != pmap->pm_generation) {
2589 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2590 npv = TAILQ_FIRST(&pmap->pm_pvlist);
2593 lwkt_reltoken(&vm_token);
2598 * pmap_testbit tests bits in active mappings of a VM page.
2600 * The caller must hold vm_token
2603 pmap_testbit(vm_page_t m, int bit)
2608 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2611 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2616 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2618 * if the bit being tested is the modified bit, then
2619 * mark clean_map and ptes as never
2622 if (bit & (VPTE_A|VPTE_M)) {
2623 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2627 #if defined(PMAP_DIAGNOSTIC)
2629 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2633 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2644 * This routine is used to clear bits in ptes. Certain bits require special
2645 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2647 * This routine is only called with certain VPTE_* bit combinations.
2649 * The caller must hold vm_token
2651 static __inline void
2652 pmap_clearbit(vm_page_t m, int bit)
2658 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2664 * Loop over all current mappings setting/clearing as appropos If
2665 * setting RO do we need to clear the VAC?
2667 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2669 * don't write protect pager mappings
2671 if (bit == VPTE_W) {
2672 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2676 #if defined(PMAP_DIAGNOSTIC)
2678 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2684 * Careful here. We can use a locked bus instruction to
2685 * clear VPTE_A or VPTE_M safely but we need to synchronize
2686 * with the target cpus when we mess with VPTE_W.
2688 * On virtual kernels we must force a new fault-on-write
2689 * in the real kernel if we clear the Modify bit ourselves,
2690 * otherwise the real kernel will not get a new fault and
2691 * will never set our Modify bit again.
2693 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2695 if (bit == VPTE_W) {
2697 * We must also clear VPTE_M when clearing
2700 pbits = pmap_clean_pte(pte, pv->pv_pmap,
2704 } else if (bit == VPTE_M) {
2706 * We do not have to make the page read-only
2707 * when clearing the Modify bit. The real
2708 * kernel will make the real PTE read-only
2709 * or otherwise detect the write and set
2710 * our VPTE_M again simply by us invalidating
2711 * the real kernel VA for the pmap (as we did
2712 * above). This allows the real kernel to
2713 * handle the write fault without forwarding
2716 atomic_clear_long(pte, VPTE_M);
2717 } else if ((bit & (VPTE_W|VPTE_M)) == (VPTE_W|VPTE_M)) {
2719 * We've been asked to clear W & M, I guess
2720 * the caller doesn't want us to update
2721 * the dirty status of the VM page.
2723 pmap_clean_pte(pte, pv->pv_pmap, pv->pv_va);
2726 * We've been asked to clear bits that do
2727 * not interact with hardware.
2729 atomic_clear_long(pte, bit);
2737 * Lower the permission for all mappings to a given page.
2742 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2744 if ((prot & VM_PROT_WRITE) == 0) {
2745 lwkt_gettoken(&vm_token);
2746 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2747 pmap_clearbit(m, VPTE_W);
2748 vm_page_flag_clear(m, PG_WRITEABLE);
2752 lwkt_reltoken(&vm_token);
2757 pmap_phys_address(vm_pindex_t ppn)
2759 return (i386_ptob(ppn));
2763 * Return a count of reference bits for a page, clearing those bits.
2764 * It is not necessary for every reference bit to be cleared, but it
2765 * is necessary that 0 only be returned when there are truly no
2766 * reference bits set.
2768 * XXX: The exact number of bits to check and clear is a matter that
2769 * should be tested and standardized at some point in the future for
2770 * optimal aging of shared pages.
2775 pmap_ts_referenced(vm_page_t m)
2777 pv_entry_t pv, pvf, pvn;
2781 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2785 lwkt_gettoken(&vm_token);
2787 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2792 pvn = TAILQ_NEXT(pv, pv_list);
2794 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2796 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2798 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2801 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2803 if (pte && (*pte & VPTE_A)) {
2805 atomic_clear_long(pte, VPTE_A);
2807 atomic_clear_long_nonlocked(pte, VPTE_A);
2814 } while ((pv = pvn) != NULL && pv != pvf);
2816 lwkt_reltoken(&vm_token);
2823 * Return whether or not the specified physical page was modified
2824 * in any physical maps.
2829 pmap_is_modified(vm_page_t m)
2833 lwkt_gettoken(&vm_token);
2834 res = pmap_testbit(m, VPTE_M);
2835 lwkt_reltoken(&vm_token);
2840 * Clear the modify bits on the specified physical page.
2845 pmap_clear_modify(vm_page_t m)
2847 lwkt_gettoken(&vm_token);
2848 pmap_clearbit(m, VPTE_M);
2849 lwkt_reltoken(&vm_token);
2853 * Clear the reference bit on the specified physical page.
2858 pmap_clear_reference(vm_page_t m)
2860 lwkt_gettoken(&vm_token);
2861 pmap_clearbit(m, VPTE_A);
2862 lwkt_reltoken(&vm_token);
2866 * Miscellaneous support routines follow
2870 i386_protection_init(void)
2874 kp = protection_codes;
2875 for (prot = 0; prot < 8; prot++) {
2876 if (prot & VM_PROT_READ)
2878 if (prot & VM_PROT_WRITE)
2880 if (prot & VM_PROT_EXECUTE)
2889 * Map a set of physical memory pages into the kernel virtual
2890 * address space. Return a pointer to where it is mapped. This
2891 * routine is intended to be used for mapping device memory,
2894 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2898 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2900 vm_offset_t va, tmpva, offset;
2903 offset = pa & PAGE_MASK;
2904 size = roundup(offset + size, PAGE_SIZE);
2906 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
2908 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2910 pa = pa & VPTE_FRAME;
2911 for (tmpva = va; size > 0;) {
2912 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2913 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2921 return ((void *)(va + offset));
2925 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2927 vm_offset_t base, offset;
2929 base = va & VPTE_FRAME;
2930 offset = va & PAGE_MASK;
2931 size = roundup(offset + size, PAGE_SIZE);
2932 pmap_qremove(va, size >> PAGE_SHIFT);
2933 kmem_free(&kernel_map, base, size);
2939 * Perform the pmap work for mincore
2944 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2950 lwkt_gettoken(&vm_token);
2952 ptep = pmap_pte(pmap, addr);
2954 lwkt_reltoken(&vm_token);
2958 if ((pte = *ptep) != 0) {
2961 val = MINCORE_INCORE;
2962 if ((pte & VPTE_MANAGED) == 0)
2965 pa = pte & VPTE_FRAME;
2967 m = PHYS_TO_VM_PAGE(pa);
2973 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2975 * Modified by someone
2977 else if (m->dirty || pmap_is_modified(m))
2978 val |= MINCORE_MODIFIED_OTHER;
2983 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2986 * Referenced by someone
2988 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
2989 val |= MINCORE_REFERENCED_OTHER;
2990 vm_page_flag_set(m, PG_REFERENCED);
2994 lwkt_reltoken(&vm_token);
2999 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3001 struct vmspace *oldvm;
3004 oldvm = p->p_vmspace;
3006 if (oldvm != newvm) {
3007 p->p_vmspace = newvm;
3008 KKASSERT(p->p_nthreads == 1);
3009 lp = RB_ROOT(&p->p_lwp_tree);
3010 pmap_setlwpvm(lp, newvm);
3012 sysref_get(&newvm->vm_sysref);
3013 sysref_put(&oldvm->vm_sysref);
3020 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3022 struct vmspace *oldvm;
3026 oldvm = lp->lwp_vmspace;
3028 if (oldvm != newvm) {
3029 lp->lwp_vmspace = newvm;
3030 if (curthread->td_lwp == lp) {
3031 pmap = vmspace_pmap(newvm);
3033 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3035 pmap->pm_active |= 1;
3037 #if defined(SWTCH_OPTIM_STATS)
3040 pmap = vmspace_pmap(oldvm);
3042 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3044 pmap->pm_active &= ~(cpumask_t)1;
3053 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3056 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3060 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3065 * Used by kmalloc/kfree, page already exists at va
3068 pmap_kvtom(vm_offset_t va)
3072 KKASSERT(va >= KvaStart && va < KvaEnd);
3073 ptep = KernelPTA + (va >> PAGE_SHIFT);
3074 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));