2 * Copyright (c) 2006 The DragonFly Project. All rights reserved.
3 * Copyright (c) 1991 Regents of the University of California.
5 * Copyright (c) 1994 John S. Dyson
7 * Copyright (c) 1994 David Greenman
9 * Copyright (c) 2004-2006 Matthew Dillon
10 * All rights reserved.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in
20 * the documentation and/or other materials provided with the
22 * 3. Neither the name of The DragonFly Project nor the names of its
23 * contributors may be used to endorse or promote products derived
24 * from this software without specific, prior written permission.
26 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
27 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
28 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
29 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
30 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
31 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
32 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
33 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
34 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
35 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
36 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
40 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
41 * $DragonFly: src/sys/platform/vkernel/platform/pmap.c,v 1.4 2007/01/06 19:40:55 dillon Exp $
44 #include <sys/types.h>
45 #include <sys/systm.h>
46 #include <sys/kernel.h>
49 #include <sys/vkernel.h>
51 #include <sys/thread.h>
55 #include <vm/vm_page.h>
56 #include <vm/vm_extern.h>
57 #include <vm/vm_kern.h>
58 #include <vm/vm_object.h>
59 #include <vm/vm_zone.h>
60 #include <vm/vm_pageout.h>
62 #include <machine/md_var.h>
63 #include <machine/pcb.h>
64 #include <machine/pmap_inval.h>
65 #include <machine/globaldata.h>
69 struct pmap kernel_pmap;
71 static struct vm_zone pvzone;
72 static struct vm_object pvzone_obj;
73 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
74 static int pv_entry_count;
75 static int pv_entry_max;
76 static int pv_entry_high_water;
77 static int pmap_pagedaemon_waken;
78 static boolean_t pmap_initialized = FALSE;
79 static int protection_codes[8];
81 static void i386_protection_init(void);
82 static void pmap_remove_all(vm_page_t m);
83 static int pmap_release_free_page(struct pmap *pmap, vm_page_t p);
86 #ifndef PMAP_SHPGPERPROC
87 #define PMAP_SHPGPERPROC 200
90 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
92 #define pte_prot(m, p) \
93 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
99 struct pv_entry *pvinit;
101 for (i = 0; i < vm_page_array_size; i++) {
104 m = &vm_page_array[i];
105 TAILQ_INIT(&m->md.pv_list);
106 m->md.pv_list_count = 0;
109 i = vm_page_array_size;
112 pvinit = (struct pv_entry *)kmem_alloc(&kernel_map, i*sizeof(*pvinit));
113 zbootinit(&pvzone, "PV ENTRY", sizeof(*pvinit), pvinit, i);
114 pmap_initialized = TRUE;
120 int shpgperproc = PMAP_SHPGPERPROC;
122 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
123 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
124 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
125 pv_entry_high_water = 9 * (pv_entry_max / 10);
126 zinitna(&pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
130 * Bootstrap the kernel_pmap so it can be used with pmap_enter().
132 * NOTE! pm_pdir for the kernel pmap is offset so VA's translate
133 * directly into PTD indexes (PTA is also offset for the same reason).
134 * This is necessary because, for now, KVA is not mapped at address 0.
136 * Page table pages are not managed like they are in normal pmaps, so
137 * no pteobj is needed.
142 vm_pindex_t i = (vm_offset_t)KernelPTD >> PAGE_SHIFT;
144 kernel_pmap.pm_pdir = KernelPTD - (KvaStart >> SEG_SHIFT);
145 kernel_pmap.pm_pdirpte = KernelPTA[i];
146 kernel_pmap.pm_count = 1;
147 kernel_pmap.pm_active = (cpumask_t)-1;
148 TAILQ_INIT(&kernel_pmap.pm_pvlist);
149 i386_protection_init();
153 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
154 * just dummy it up so it works well enough for fork().
156 * In DragonFly, process pmaps may only be used to manipulate user address
157 * space, never kernel address space.
160 pmap_pinit0(struct pmap *pmap)
165 /************************************************************************
166 * Procedures to manage whole physical maps *
167 ************************************************************************
169 * Initialize a preallocated and zeroed pmap structure,
170 * such as one in a vmspace structure.
173 pmap_pinit(struct pmap *pmap)
179 * No need to allocate page table space yet but we do need a valid
180 * page directory table.
182 if (pmap->pm_pdir == NULL) {
184 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
188 * allocate object for the pte array and page directory
190 npages = VPTE_PAGETABLE_SIZE +
191 (VM_MAX_USER_ADDRESS / PAGE_SIZE) * sizeof(vpte_t);
192 npages = (npages + PAGE_MASK) / PAGE_SIZE;
194 if (pmap->pm_pteobj == NULL)
195 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, npages);
196 pmap->pm_pdindex = npages - 1;
199 * allocate the page directory page
201 ptdpg = vm_page_grab(pmap->pm_pteobj, pmap->pm_pdindex,
202 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
204 ptdpg->wire_count = 1;
205 ++vmstats.v_wire_count;
207 /* not usually mapped */
208 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
209 ptdpg->valid = VM_PAGE_BITS_ALL;
211 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
212 if ((ptdpg->flags & PG_ZERO) == 0)
213 bzero(pmap->pm_pdir, PAGE_SIZE);
217 pmap->pm_ptphint = NULL;
218 TAILQ_INIT(&pmap->pm_pvlist);
219 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
223 * Wire in kernel global address entries. To avoid a race condition
224 * between pmap initialization and pmap_growkernel, this procedure
225 * adds the pmap to the master list (which growkernel scans to update),
226 * then copies the template.
228 * In a virtual kernel there are no kernel global address entries.
231 pmap_pinit2(struct pmap *pmap)
234 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
239 * Release all resources held by the given physical map.
241 * Should only be called if the map contains no valid mappings.
243 static int pmap_release_callback(struct vm_page *p, void *data);
246 pmap_release(struct pmap *pmap)
248 vm_object_t object = pmap->pm_pteobj;
249 struct rb_vm_page_scan_info info;
251 KKASSERT(pmap != &kernel_pmap);
253 #if defined(DIAGNOSTIC)
254 if (object->ref_count != 1)
255 panic("pmap_release: pteobj reference count != 1");
259 info.object = object;
261 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
268 info.limit = object->generation;
270 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
271 pmap_release_callback, &info);
272 if (info.error == 0 && info.mpte) {
273 if (!pmap_release_free_page(pmap, info.mpte))
277 } while (info.error);
281 pmap_release_callback(struct vm_page *p, void *data)
283 struct rb_vm_page_scan_info *info = data;
285 if (p->pindex == info->pmap->pm_pdindex) {
289 if (!pmap_release_free_page(info->pmap, p)) {
293 if (info->object->generation != info->limit) {
301 * Retire the given physical map from service. Should only be called if
302 * the map contains no valid mappings.
305 pmap_destroy(pmap_t pmap)
312 count = --pmap->pm_count;
315 panic("destroying a pmap is not yet implemented");
320 * Add a reference to the specified pmap.
323 pmap_reference(pmap_t pmap)
330 /************************************************************************
331 * Procedures which operate directly on the kernel PMAP *
332 ************************************************************************/
335 * This maps the requested page table and gives us access to it.
338 get_ptbase(struct pmap *pmap, vm_offset_t va)
340 struct mdglobaldata *gd = mdcpu;
342 if (pmap == &kernel_pmap) {
343 KKASSERT(va >= KvaStart && va < KvaEnd);
344 return(KernelPTA + (va >> PAGE_SHIFT));
345 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
346 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
347 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
348 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
352 * Otherwise choose one or the other and map the page table
353 * in the KVA space reserved for it.
355 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
356 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
358 if ((gd->gd_PTflip = 1 - gd->gd_PTflip) == 0) {
359 gd->gd_PT1pdir = pmap->pm_pdir;
360 *gd->gd_PT1pde = pmap->pm_pdirpte;
361 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
362 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
364 gd->gd_PT2pdir = pmap->pm_pdir;
365 *gd->gd_PT2pde = pmap->pm_pdirpte;
366 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
367 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
372 get_ptbase1(struct pmap *pmap, vm_offset_t va)
374 struct mdglobaldata *gd = mdcpu;
376 if (pmap == &kernel_pmap) {
377 KKASSERT(va >= KvaStart && va < KvaEnd);
378 return(KernelPTA + (va >> PAGE_SHIFT));
379 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
380 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
382 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
383 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
384 gd->gd_PT1pdir = pmap->pm_pdir;
385 *gd->gd_PT1pde = pmap->pm_pdirpte;
386 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
387 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
391 get_ptbase2(struct pmap *pmap, vm_offset_t va)
393 struct mdglobaldata *gd = mdcpu;
395 if (pmap == &kernel_pmap) {
396 KKASSERT(va >= KvaStart && va < KvaEnd);
397 return(KernelPTA + (va >> PAGE_SHIFT));
398 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
399 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
401 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
402 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
403 gd->gd_PT2pdir = pmap->pm_pdir;
404 *gd->gd_PT2pde = pmap->pm_pdirpte;
405 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
406 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
410 * Return a pointer to the page table entry for the specified va in the
411 * specified pmap. NULL is returned if there is no valid page table page
414 static __inline vpte_t *
415 pmap_pte(struct pmap *pmap, vm_offset_t va)
419 ptep = &pmap->pm_pdir[va >> SEG_SHIFT];
423 return (get_ptbase(pmap, va));
429 * Enter a mapping into kernel_pmap. Mappings created in this fashion
433 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
438 pmap_inval_info info;
441 KKASSERT(va >= KvaStart && va < KvaEnd);
442 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
443 ptep = KernelPTA + (va >> PAGE_SHIFT);
444 if (*ptep & VPTE_V) {
446 pmap_inval_init(&info);
447 pmap_inval_add(&info, &kernel_pmap, va);
451 pmap_inval_flush(&info);
453 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
461 pmap_kenter_sync(vm_offset_t va)
463 pmap_inval_info info;
465 pmap_inval_init(&info);
466 pmap_inval_add(&info, &kernel_pmap, va);
467 pmap_inval_flush(&info);
471 pmap_kenter_sync_quick(vm_offset_t va)
473 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
477 * XXX these need to be recoded. They are not used in any critical path.
480 pmap_kmodify_rw(vm_offset_t va)
482 *pmap_kpte(va) |= VPTE_R | VPTE_W;
483 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
487 pmap_kmodify_nc(vm_offset_t va)
490 *pmap_kpte(va) |= VPTE_N;
491 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
496 * Map a contiguous range of physical memory to a KVM
499 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
501 while (start < end) {
502 pmap_kenter(virt, start);
510 pmap_kpte(vm_offset_t va)
514 KKASSERT(va >= KvaStart && va < KvaEnd);
515 ptep = KernelPTA + (va >> PAGE_SHIFT);
520 * Enter a mapping into kernel_pmap without any SMP interactions.
522 * Mappings created in this fashion are not managed.
525 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
530 KKASSERT(va >= KvaStart && va < KvaEnd);
532 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
533 ptep = KernelPTA + (va >> PAGE_SHIFT);
534 if (*ptep & VPTE_V) {
536 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
543 * Make a temporary mapping for a physical address. This is only intended
544 * to be used for panic dumps.
547 pmap_kenter_temporary(vm_paddr_t pa, int i)
549 pmap_kenter(crashdumpmap + (i * PAGE_SIZE), pa);
550 return ((void *)crashdumpmap);
554 * Remove an unmanaged mapping created with pmap_kenter*().
557 pmap_kremove(vm_offset_t va)
561 pmap_inval_info info;
564 KKASSERT(va >= KvaStart && va < KvaEnd);
566 ptep = KernelPTA + (va >> PAGE_SHIFT);
567 if (*ptep & VPTE_V) {
569 pmap_inval_init(&info);
570 pmap_inval_add(&info, &kernel_pmap, va);
574 pmap_inval_flush(&info);
576 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
585 * Remove an unmanaged mapping created with pmap_kenter*() without
586 * going through any SMP interactions.
589 pmap_kremove_quick(vm_offset_t va)
593 KKASSERT(va >= KvaStart && va < KvaEnd);
595 ptep = KernelPTA + (va >> PAGE_SHIFT);
596 if (*ptep & VPTE_V) {
598 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
605 * Extract the physical address from the kernel_pmap that is associated
606 * with the specified virtual address.
609 pmap_kextract(vm_offset_t va)
614 KKASSERT(va >= KvaStart && va < KvaEnd);
616 ptep = KernelPTA + (va >> PAGE_SHIFT);
617 pa = (vm_paddr_t)(*ptep & VPTE_FRAME) | (va & PAGE_MASK);
622 * Map a set of unmanaged VM pages into KVM.
625 pmap_qenter(vm_offset_t va, struct vm_page **m, int count)
627 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
631 ptep = KernelPTA + (va >> PAGE_SHIFT);
633 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
634 *ptep = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
646 * Map a set of VM pages to kernel virtual memory. If a mapping changes
647 * clear the supplied mask. The caller handles any SMP interactions.
648 * The mask is used to provide the caller with hints on what SMP interactions
652 pmap_qenter2(vm_offset_t va, struct vm_page **m, int count, cpumask_t *mask)
654 cpumask_t cmask = mycpu->gd_cpumask;
656 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
661 ptep = KernelPTA + (va >> PAGE_SHIFT);
662 npte = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
666 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
667 } else if ((*mask & cmask) == 0) {
668 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
678 * Undo the effects of pmap_qenter*().
681 pmap_qremove(vm_offset_t va, int count)
683 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
687 ptep = KernelPTA + (va >> PAGE_SHIFT);
689 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
700 /************************************************************************
701 * Misc support glue called by machine independant code *
702 ************************************************************************
704 * These routines are called by machine independant code to operate on
705 * certain machine-dependant aspects of processes, threads, and pmaps.
709 * Initialize MD portions of the thread structure.
712 pmap_init_thread(thread_t td)
714 /* enforce pcb placement */
715 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
716 td->td_savefpu = &td->td_pcb->pcb_save;
717 td->td_sp = (char *)td->td_pcb - 16;
721 * Initialize MD portions of a process structure. XXX this aint MD
724 pmap_init_proc(struct proc *p, struct thread *td)
726 p->p_addr = (void *)td->td_kstack;
729 td->td_lwp = &p->p_lwp;
730 td->td_switch = cpu_heavy_switch;
732 KKASSERT(td->td_mpcount == 1);
734 bzero(p->p_addr, sizeof(*p->p_addr));
738 * Destroy the UPAGES for a process that has exited and disassociate
739 * the process from its thread.
742 pmap_dispose_proc(struct proc *p)
746 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
748 if ((td = p->p_thread) != NULL) {
757 * We pre-allocate all page table pages for kernel virtual memory so
758 * this routine will only be called if KVM has been exhausted.
761 pmap_growkernel(vm_offset_t size)
763 panic("KVM exhausted");
767 * The modification bit is not tracked for any pages in this range. XXX
768 * such pages in this maps should always use pmap_k*() functions and not
772 pmap_track_modified(vm_offset_t va)
774 if ((va < clean_sva) || (va >= clean_eva))
780 /************************************************************************
781 * Procedures supporting managed page table pages *
782 ************************************************************************
784 * These procedures are used to track managed page table pages. These pages
785 * use the page table page's vm_page_t to track PTEs in the page. The
786 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
788 * This allows the system to throw away page table pages for user processes
789 * at will and reinstantiate them on demand.
793 * This routine works like vm_page_lookup() but also blocks as long as the
794 * page is busy. This routine does not busy the page it returns.
796 * Unless the caller is managing objects whos pages are in a known state,
797 * the call should be made with a critical section held so the page's object
798 * association remains valid on return.
801 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
806 m = vm_page_lookup(object, pindex);
807 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
813 * This routine unholds page table pages, and if the hold count
814 * drops to zero, then it decrements the wire count.
817 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
819 pmap_inval_flush(info);
820 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
823 if (m->hold_count == 0) {
825 * unmap the page table page
827 pmap_inval_add(info, pmap, -1);
828 pmap->pm_pdir[m->pindex] = 0;
829 --pmap->pm_stats.resident_count;
831 if (pmap->pm_ptphint == m)
832 pmap->pm_ptphint = NULL;
835 * If the page is finally unwired, simply free it.
838 if (m->wire_count == 0) {
841 vm_page_free_zero(m);
842 --vmstats.v_wire_count;
850 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
853 if (m->hold_count == 0)
854 return _pmap_unwire_pte_hold(pmap, m, info);
860 * After removing a page table entry, this routine is used to
861 * conditionally free the page, and manage the hold/wire counts.
864 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
865 pmap_inval_info_t info)
871 * page table pages in the kernel_pmap are not managed.
873 if (pmap == &kernel_pmap)
875 ptepindex = (va >> PDRSHIFT);
876 if (pmap->pm_ptphint &&
877 (pmap->pm_ptphint->pindex == ptepindex)) {
878 mpte = pmap->pm_ptphint;
880 pmap_inval_flush(info);
881 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
882 pmap->pm_ptphint = mpte;
885 return pmap_unwire_pte_hold(pmap, mpte, info);
889 * Attempt to release and free an vm_page in a pmap. Returns 1 on success,
890 * 0 on failure (if the procedure had to sleep).
893 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
895 vpte_t *pde = pmap->pm_pdir;
897 * This code optimizes the case of freeing non-busy
898 * page-table pages. Those pages are zero now, and
899 * might as well be placed directly into the zero queue.
901 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
907 * Remove the page table page from the processes address space.
910 pmap->pm_stats.resident_count--;
913 panic("pmap_release: freeing held page table page");
916 * Page directory pages need to have the kernel stuff cleared, so
917 * they can go into the zero queue also.
919 * In virtual kernels there is no 'kernel stuff'. For the moment
920 * I just make sure the whole thing has been zero'd even though
921 * it should already be completely zero'd.
923 if (p->pindex == pmap->pm_pdindex) {
924 bzero(pde, VPTE_PAGETABLE_SIZE);
925 pmap_kremove((vm_offset_t)pmap->pm_pdir);
929 * Clear the matching hint
931 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
932 pmap->pm_ptphint = NULL;
935 * And throw the page away. The page is completely zero'd out so
936 * optimize the free call.
939 vmstats.v_wire_count--;
940 vm_page_free_zero(p);
945 * This routine is called if the page table page is not mapped in the page
948 * The routine is broken up into two parts for readability.
951 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
957 * Find or fabricate a new pagetable page
959 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
960 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
962 KASSERT(m->queue == PQ_NONE,
963 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
965 if (m->wire_count == 0)
966 vmstats.v_wire_count++;
970 * Increment the hold count for the page table page
971 * (denoting a new mapping.)
976 * Map the pagetable page into the process address space, if
977 * it isn't already there.
979 pmap->pm_stats.resident_count++;
981 ptepa = VM_PAGE_TO_PHYS(m);
982 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
986 * We are likely about to access this page table page, so set the
987 * page table hint to reduce overhead.
989 pmap->pm_ptphint = m;
992 * Try to use the new mapping, but if we cannot, then
993 * do it with the routine that maps the page explicitly.
995 if ((m->flags & PG_ZERO) == 0)
996 pmap_zero_page(ptepa);
998 m->valid = VM_PAGE_BITS_ALL;
999 vm_page_flag_clear(m, PG_ZERO);
1000 vm_page_flag_set(m, PG_MAPPED);
1007 * Determine the page table page required to access the VA in the pmap
1008 * and allocate it if necessary. Return a held vm_page_t for the page.
1010 * Only used with user pmaps.
1013 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1020 * Calculate pagetable page index
1022 ptepindex = va >> PDRSHIFT;
1025 * Get the page directory entry
1027 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1030 * This supports switching from a 4MB page to a
1033 if (ptepa & VPTE_PS) {
1034 pmap->pm_pdir[ptepindex] = 0;
1041 * If the page table page is mapped, we just increment the
1042 * hold count, and activate it.
1046 * In order to get the page table page, try the
1049 if (pmap->pm_ptphint &&
1050 (pmap->pm_ptphint->pindex == ptepindex)) {
1051 m = pmap->pm_ptphint;
1053 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1054 pmap->pm_ptphint = m;
1060 * Here if the pte page isn't mapped, or if it has been deallocated.
1062 return _pmap_allocpte(pmap, ptepindex);
1065 /************************************************************************
1066 * Managed pages in pmaps *
1067 ************************************************************************
1069 * All pages entered into user pmaps and some pages entered into the kernel
1070 * pmap are managed, meaning that pmap_protect() and other related management
1071 * functions work on these pages.
1075 * free the pv_entry back to the free list. This function may be
1076 * called from an interrupt.
1078 static __inline void
1079 free_pv_entry(pv_entry_t pv)
1086 * get a new pv_entry, allocating a block from the system
1087 * when needed. This function may be called from an interrupt.
1093 if (pv_entry_high_water &&
1094 (pv_entry_count > pv_entry_high_water) &&
1095 (pmap_pagedaemon_waken == 0)) {
1096 pmap_pagedaemon_waken = 1;
1097 wakeup (&vm_pages_needed);
1099 return zalloc(&pvzone);
1103 * This routine is very drastic, but can save the system
1111 static int warningdone=0;
1113 if (pmap_pagedaemon_waken == 0)
1116 if (warningdone < 5) {
1117 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1121 for(i = 0; i < vm_page_array_size; i++) {
1122 m = &vm_page_array[i];
1123 if (m->wire_count || m->hold_count || m->busy ||
1124 (m->flags & PG_BUSY))
1128 pmap_pagedaemon_waken = 0;
1132 * If it is the first entry on the list, it is actually
1133 * in the header and we must copy the following entry up
1134 * to the header. Otherwise we must search the list for
1135 * the entry. In either case we free the now unused entry.
1138 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1139 vm_offset_t va, pmap_inval_info_t info)
1145 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1146 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1147 if (pmap == pv->pv_pmap && va == pv->pv_va)
1151 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1152 if (va == pv->pv_va)
1158 * Note that pv_ptem is NULL if the page table page itself is not
1159 * managed, even if the page being removed IS managed.
1163 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1164 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1165 m->md.pv_list_count--;
1166 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1167 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1168 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1176 * Create a pv entry for page at pa for (pmap, va). If the page table page
1177 * holding the VA is managed, mpte will be non-NULL.
1180 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1185 pv = get_pv_entry();
1190 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1191 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1192 m->md.pv_list_count++;
1198 * pmap_remove_pte: do the things to unmap a page in a process
1201 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va,
1202 pmap_inval_info_t info)
1207 pmap_inval_add(info, pmap, va);
1208 oldpte = loadandclear(ptq);
1209 if (oldpte & VPTE_W)
1210 pmap->pm_stats.wired_count -= 1;
1212 * Machines that don't support invlpg, also don't support
1213 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1216 if (oldpte & VPTE_G)
1217 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1218 pmap->pm_stats.resident_count -= 1;
1219 if (oldpte & PG_MANAGED) {
1220 m = PHYS_TO_VM_PAGE(oldpte);
1221 if (oldpte & VPTE_M) {
1222 #if defined(PMAP_DIAGNOSTIC)
1223 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1225 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1229 if (pmap_track_modified(va))
1232 if (oldpte & VPTE_A)
1233 vm_page_flag_set(m, PG_REFERENCED);
1234 return pmap_remove_entry(pmap, m, va, info);
1236 return pmap_unuse_pt(pmap, va, NULL, info);
1245 * Remove a single page from a process address space.
1247 * This function may not be called from an interrupt if the pmap is
1251 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1256 * if there is no pte for this address, just skip it!!! Otherwise
1257 * get a local va for mappings for this pmap and remove the entry.
1259 if (*pmap_pde(pmap, va) != 0) {
1260 ptq = get_ptbase(pmap, va);
1262 pmap_remove_pte(pmap, ptq, va, info);
1270 * Remove the given range of addresses from the specified map.
1272 * It is assumed that the start and end are properly
1273 * rounded to the page size.
1275 * This function may not be called from an interrupt if the pmap is
1279 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1283 vm_offset_t ptpaddr;
1284 vm_pindex_t sindex, eindex;
1286 struct pmap_inval_info info;
1291 if (pmap->pm_stats.resident_count == 0)
1294 pmap_inval_init(&info);
1297 * special handling of removing one page. a very
1298 * common operation and easy to short circuit some
1301 if (((sva + PAGE_SIZE) == eva) &&
1302 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1303 pmap_remove_page(pmap, sva, &info);
1304 pmap_inval_flush(&info);
1309 * Get a local virtual address for the mappings that are being
1312 * XXX this is really messy because the kernel pmap is not relative
1315 ptbase = get_ptbase(pmap, sva);
1317 sindex = (sva >> PAGE_SHIFT);
1318 eindex = (eva >> PAGE_SHIFT);
1321 for (; sindex < eindex; sindex = pdnxt) {
1325 * Calculate index for next page table.
1327 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1328 if (pmap->pm_stats.resident_count == 0)
1331 pdirindex = sindex / NPDEPG;
1332 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1333 pmap_inval_add(&info, pmap, -1);
1334 pmap->pm_pdir[pdirindex] = 0;
1335 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1340 * Weed out invalid mappings. Note: we assume that the page
1341 * directory table is always allocated, and in kernel virtual.
1347 * Limit our scan to either the end of the va represented
1348 * by the current page table page, or to the end of the
1349 * range being removed.
1351 if (pdnxt > eindex) {
1355 for (; sindex != pdnxt; sindex++) {
1357 if (ptbase[sindex - sbase] == 0)
1359 va = i386_ptob(sindex);
1360 if (pmap_remove_pte(pmap, ptbase + sindex - sbase, va, &info))
1364 pmap_inval_flush(&info);
1370 * Removes this physical page from all physical maps in which it resides.
1371 * Reflects back modify bits to the pager.
1373 * This routine may not be called from an interrupt.
1376 pmap_remove_all(vm_page_t m)
1378 struct pmap_inval_info info;
1382 #if defined(PMAP_DIAGNOSTIC)
1384 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1387 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1388 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1392 pmap_inval_init(&info);
1394 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1395 pv->pv_pmap->pm_stats.resident_count--;
1397 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1398 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1400 tpte = loadandclear(pte);
1402 pv->pv_pmap->pm_stats.wired_count--;
1405 vm_page_flag_set(m, PG_REFERENCED);
1408 * Update the vm_page_t clean and reference bits.
1410 if (tpte & VPTE_M) {
1411 #if defined(PMAP_DIAGNOSTIC)
1412 if (pmap_nw_modified((pt_entry_t) tpte)) {
1414 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1418 if (pmap_track_modified(pv->pv_va))
1421 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1422 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1423 m->md.pv_list_count--;
1424 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1428 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1430 pmap_inval_flush(&info);
1436 * Set the physical protection on the specified range of this map
1439 * This function may not be called from an interrupt if the map is
1440 * not the kernel_pmap.
1443 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1446 vm_offset_t pdnxt, ptpaddr;
1447 vm_pindex_t sindex, eindex;
1449 pmap_inval_info info;
1454 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1455 pmap_remove(pmap, sva, eva);
1459 if (prot & VM_PROT_WRITE)
1462 pmap_inval_init(&info);
1464 ptbase = get_ptbase(pmap, sva);
1466 sindex = (sva >> PAGE_SHIFT);
1467 eindex = (eva >> PAGE_SHIFT);
1470 for (; sindex < eindex; sindex = pdnxt) {
1474 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1476 pdirindex = sindex / NPDEPG;
1477 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1478 pmap_inval_add(&info, pmap, -1);
1479 pmap->pm_pdir[pdirindex] &= ~(VPTE_M|VPTE_W);
1480 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1485 * Weed out invalid mappings. Note: we assume that the page
1486 * directory table is always allocated, and in kernel virtual.
1491 if (pdnxt > eindex) {
1495 for (; sindex != pdnxt; sindex++) {
1500 /* XXX this isn't optimal */
1501 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1502 pbits = ptbase[sindex - sbase];
1504 if (pbits & PG_MANAGED) {
1506 if (pbits & VPTE_A) {
1507 m = PHYS_TO_VM_PAGE(pbits);
1508 vm_page_flag_set(m, PG_REFERENCED);
1511 if (pbits & VPTE_M) {
1512 if (pmap_track_modified(i386_ptob(sindex))) {
1514 m = PHYS_TO_VM_PAGE(pbits);
1523 if (pbits != ptbase[sindex - sbase]) {
1524 ptbase[sindex - sbase] = pbits;
1528 pmap_inval_flush(&info);
1532 * Enter a managed page into a pmap. If the page is not wired related pmap
1533 * data can be destroyed at any time for later demand-operation.
1535 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1536 * specified protection, and wire the mapping if requested.
1538 * NOTE: This routine may not lazy-evaluate or lose information. The
1539 * page must actually be inserted into the given map NOW.
1541 * NOTE: When entering a page at a KVA address, the pmap must be the
1545 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1551 vm_offset_t origpte, newpte;
1553 pmap_inval_info info;
1561 * Get the page table page. The kernel_pmap's page table pages
1562 * are preallocated and have no associated vm_page_t.
1564 if (pmap == &kernel_pmap)
1567 mpte = pmap_allocpte(pmap, va);
1569 pmap_inval_init(&info);
1570 pte = pmap_pte(pmap, va);
1573 * Page Directory table entry not valid, we need a new PT page
1574 * and pmap_allocpte() didn't give us one. Oops!
1577 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1581 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1582 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1584 opa = origpte & VPTE_FRAME;
1586 if (origpte & VPTE_PS)
1587 panic("pmap_enter: attempted pmap_enter on 4MB page");
1590 * Mapping has not changed, must be protection or wiring change.
1592 if (origpte && (opa == pa)) {
1594 * Wiring change, just update stats. We don't worry about
1595 * wiring PT pages as they remain resident as long as there
1596 * are valid mappings in them. Hence, if a user page is wired,
1597 * the PT page will be also.
1599 if (wired && ((origpte & VPTE_W) == 0))
1600 pmap->pm_stats.wired_count++;
1601 else if (!wired && (origpte & VPTE_W))
1602 pmap->pm_stats.wired_count--;
1604 #if defined(PMAP_DIAGNOSTIC)
1605 if (pmap_nw_modified((pt_entry_t) origpte)) {
1607 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1613 * Remove the extra pte reference. Note that we cannot
1614 * optimize the RO->RW case because we have adjusted the
1615 * wiring count above and may need to adjust the wiring
1622 * We might be turning off write access to the page,
1623 * so we go ahead and sense modify status.
1625 if (origpte & PG_MANAGED) {
1626 if ((origpte & VPTE_M) && pmap_track_modified(va)) {
1628 om = PHYS_TO_VM_PAGE(opa);
1636 * Mapping has changed, invalidate old range and fall through to
1637 * handle validating new mapping.
1641 err = pmap_remove_pte(pmap, pte, va, &info);
1643 panic("pmap_enter: pte vanished, va: 0x%x", va);
1647 * Enter on the PV list if part of our managed memory. Note that we
1648 * raise IPL while manipulating pv_table since pmap_enter can be
1649 * called at interrupt time.
1651 if (pmap_initialized &&
1652 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1653 pmap_insert_entry(pmap, va, mpte, m);
1658 * Increment counters
1660 pmap->pm_stats.resident_count++;
1662 pmap->pm_stats.wired_count++;
1666 * Now validate mapping with desired protection/wiring.
1668 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1675 * if the mapping or permission bits are different, we need
1676 * to update the pte.
1678 if ((origpte & ~(VPTE_M|VPTE_A)) != newpte) {
1679 *pte = newpte | VPTE_A;
1681 pmap_inval_flush(&info);
1685 * This is a quick version of pmap_enter(). It is used only under the
1686 * following conditions:
1688 * (1) The pmap is not the kernel_pmap
1689 * (2) The page is not to be wired into the map
1690 * (3) The page is to mapped read-only in the pmap (initially that is)
1691 * (4) The calling procedure is responsible for flushing the TLB
1692 * (5) The page is always managed
1693 * (6) There is no prior mapping at the VA
1697 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
1701 pmap_inval_info info;
1703 KKASSERT(pmap != &kernel_pmap);
1704 pmap_inval_init(&info);
1706 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1709 * Instantiate the page table page if required
1715 * Calculate pagetable page index
1717 ptepindex = va >> PDRSHIFT;
1718 if (mpte && (mpte->pindex == ptepindex)) {
1723 * Get the page directory entry
1725 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1728 * If the page table page is mapped, we just increment
1729 * the hold count, and activate it.
1732 if (ptepa & VPTE_PS)
1733 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1734 if (pmap->pm_ptphint &&
1735 (pmap->pm_ptphint->pindex == ptepindex)) {
1736 mpte = pmap->pm_ptphint;
1738 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1739 pmap->pm_ptphint = mpte;
1745 mpte = _pmap_allocpte(pmap, ptepindex);
1750 * Ok, now that the page table page has been validated, get the pte.
1751 * If the pte is already mapped undo mpte's hold_count and
1754 pte = pmap_pte(pmap, va);
1757 pmap_unwire_pte_hold(pmap, mpte, &info);
1762 * Enter on the PV list if part of our managed memory. Note that we
1763 * raise IPL while manipulating pv_table since pmap_enter can be
1764 * called at interrupt time.
1766 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
1767 pmap_insert_entry(pmap, va, mpte, m);
1770 * Increment counters
1772 pmap->pm_stats.resident_count++;
1774 pa = VM_PAGE_TO_PHYS(m);
1777 * Now validate mapping with RO protection
1779 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1780 *pte = pa | VPTE_V | VPTE_U;
1782 *pte = pa | VPTE_V | VPTE_U | VPTE_MANAGED;
1788 pmap_extract(pmap_t pmap, vm_offset_t va)
1793 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
1794 if (pte & VPTE_PS) {
1795 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
1796 rtval |= va & SEG_MASK;
1798 pte = *get_ptbase(pmap, va);
1799 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
1806 #define MAX_INIT_PT (96)
1809 * This routine preloads the ptes for a given object into the specified pmap.
1810 * This eliminates the blast of soft faults on process startup and
1811 * immediately after an mmap.
1813 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
1816 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
1817 vm_object_t object, vm_pindex_t pindex,
1818 vm_size_t size, int limit)
1820 struct rb_vm_page_scan_info info;
1824 * We can't preinit if read access isn't set or there is no pmap
1827 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
1831 * We can't preinit if the pmap is not the current pmap
1833 if (curproc == NULL || pmap != vmspace_pmap(curproc->p_vmspace))
1836 psize = size >> PAGE_SHIFT;
1838 if ((object->type != OBJT_VNODE) ||
1839 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
1840 (object->resident_page_count > MAX_INIT_PT))) {
1844 if (psize + pindex > object->size) {
1845 if (object->size < pindex)
1847 psize = object->size - pindex;
1854 * Use a red-black scan to traverse the requested range and load
1855 * any valid pages found into the pmap.
1857 * We cannot safely scan the object's memq unless we are in a
1858 * critical section since interrupts can remove pages from objects.
1860 info.start_pindex = pindex;
1861 info.end_pindex = pindex + psize - 1;
1868 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1869 pmap_object_init_pt_callback, &info);
1875 pmap_object_init_pt_callback(vm_page_t p, void *data)
1877 struct rb_vm_page_scan_info *info = data;
1878 vm_pindex_t rel_index;
1880 * don't allow an madvise to blow away our really
1881 * free pages allocating pv entries.
1883 if ((info->limit & MAP_PREFAULT_MADVISE) &&
1884 vmstats.v_free_count < vmstats.v_free_reserved) {
1887 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
1888 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
1889 if ((p->queue - p->pc) == PQ_CACHE)
1890 vm_page_deactivate(p);
1892 rel_index = p->pindex - info->start_pindex;
1893 info->mpte = pmap_enter_quick(info->pmap,
1894 info->addr + i386_ptob(rel_index),
1896 vm_page_flag_set(p, PG_MAPPED);
1903 * pmap_prefault provides a quick way of clustering pagefaults into a
1904 * processes address space. It is a "cousin" of pmap_object_init_pt,
1905 * except it runs at page fault time instead of mmap time.
1909 #define PAGEORDER_SIZE (PFBAK+PFFOR)
1911 static int pmap_prefault_pageorder[] = {
1912 -PAGE_SIZE, PAGE_SIZE,
1913 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
1914 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
1915 -4 * PAGE_SIZE, 4 * PAGE_SIZE
1919 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
1929 * We do not currently prefault mappings that use virtual page
1930 * tables. We do not prefault foreign pmaps.
1932 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
1934 if (curproc == NULL || (pmap != vmspace_pmap(curproc->p_vmspace)))
1937 object = entry->object.vm_object;
1939 starta = addra - PFBAK * PAGE_SIZE;
1940 if (starta < entry->start)
1941 starta = entry->start;
1942 else if (starta > addra)
1946 * critical section protection is required to maintain the
1947 * page/object association, interrupts can free pages and remove
1948 * them from their objects.
1952 for (i = 0; i < PAGEORDER_SIZE; i++) {
1953 vm_object_t lobject;
1956 addr = addra + pmap_prefault_pageorder[i];
1957 if (addr > addra + (PFFOR * PAGE_SIZE))
1960 if (addr < starta || addr >= entry->end)
1964 * Make sure the page table page already exists
1966 if ((*pmap_pde(pmap, addr)) == NULL)
1970 * Get a pointer to the pte and make sure that no valid page
1973 pte = get_ptbase(pmap, addr);
1978 * Get the page to be mapped
1980 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1983 for (m = vm_page_lookup(lobject, pindex);
1984 (!m && (lobject->type == OBJT_DEFAULT) &&
1985 (lobject->backing_object));
1986 lobject = lobject->backing_object
1988 if (lobject->backing_object_offset & PAGE_MASK)
1990 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
1991 m = vm_page_lookup(lobject->backing_object, pindex);
1995 * give-up when a page is not in memory
2001 * If everything meets the requirements for pmap_enter_quick(),
2002 * then enter the page.
2005 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2007 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2009 if ((m->queue - m->pc) == PQ_CACHE) {
2010 vm_page_deactivate(m);
2013 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2014 vm_page_flag_set(m, PG_MAPPED);
2022 * Routine: pmap_change_wiring
2023 * Function: Change the wiring attribute for a map/virtual-address
2025 * In/out conditions:
2026 * The mapping must already exist in the pmap.
2029 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2036 pte = get_ptbase(pmap, va);
2038 if (wired && (*pte & VPTE_W) == 0)
2039 pmap->pm_stats.wired_count++;
2040 else if (!wired && (*pte & VPTE_W))
2041 pmap->pm_stats.wired_count--;
2044 * Wiring is not a hardware characteristic so there is no need to
2045 * invalidate TLB. However, in an SMP environment we must use
2046 * a locked bus cycle to update the pte (if we are not using
2047 * the pmap_inval_*() API that is)... it's ok to do this for simple
2052 atomic_set_int(pte, VPTE_W);
2054 atomic_clear_int(pte, VPTE_W);
2057 atomic_set_int_nonlocked(pte, VPTE_W);
2059 atomic_clear_int_nonlocked(pte, VPTE_W);
2064 * Copy the range specified by src_addr/len
2065 * from the source map to the range dst_addr/len
2066 * in the destination map.
2068 * This routine is only advisory and need not do anything.
2071 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2072 vm_size_t len, vm_offset_t src_addr)
2074 pmap_inval_info info;
2076 vm_offset_t end_addr = src_addr + len;
2082 if (dst_addr != src_addr)
2084 if (dst_pmap->pm_pdir == NULL)
2086 if (src_pmap->pm_pdir == NULL)
2089 src_frame = get_ptbase1(src_pmap, src_addr);
2090 dst_frame = get_ptbase2(dst_pmap, src_addr);
2092 pmap_inval_init(&info);
2093 pmap_inval_add(&info, dst_pmap, -1);
2094 pmap_inval_add(&info, src_pmap, -1);
2097 * critical section protection is required to maintain the page/object
2098 * association, interrupts can free pages and remove them from
2102 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2103 vpte_t *src_pte, *dst_pte;
2104 vm_page_t dstmpte, srcmpte;
2105 vm_offset_t srcptepaddr;
2108 if (addr >= VM_MAX_USER_ADDRESS)
2109 panic("pmap_copy: invalid to pmap_copy page tables\n");
2112 * Don't let optional prefaulting of pages make us go
2113 * way below the low water mark of free pages or way
2114 * above high water mark of used pv entries.
2116 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2117 pv_entry_count > pv_entry_high_water)
2120 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2121 ptepindex = addr >> PDRSHIFT;
2123 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2124 if (srcptepaddr == 0)
2127 if (srcptepaddr & VPTE_PS) {
2128 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2129 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2130 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2135 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2136 if ((srcmpte == NULL) ||
2137 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2140 if (pdnxt > end_addr)
2143 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2144 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2145 while (addr < pdnxt) {
2149 * we only virtual copy managed pages
2151 if ((ptetemp & PG_MANAGED) != 0) {
2153 * We have to check after allocpte for the
2154 * pte still being around... allocpte can
2157 dstmpte = pmap_allocpte(dst_pmap, addr);
2158 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2160 * Clear the modified and
2161 * accessed (referenced) bits
2164 m = PHYS_TO_VM_PAGE(ptetemp);
2165 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2166 dst_pmap->pm_stats.resident_count++;
2167 pmap_insert_entry(dst_pmap, addr,
2170 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2172 if (dstmpte->hold_count >= srcmpte->hold_count)
2181 pmap_inval_flush(&info);
2187 * Zero the specified PA by mapping the page into KVM and clearing its
2190 * This function may be called from an interrupt and no locking is
2194 pmap_zero_page(vm_paddr_t phys)
2196 struct mdglobaldata *gd = mdcpu;
2200 panic("pmap_zero_page: CMAP3 busy");
2201 *gd->gd_CMAP3 = VPTE_V | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2202 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2204 bzero(gd->gd_CADDR3, PAGE_SIZE);
2210 * pmap_page_assertzero:
2212 * Assert that a page is empty, panic if it isn't.
2215 pmap_page_assertzero(vm_paddr_t phys)
2217 struct mdglobaldata *gd = mdcpu;
2222 panic("pmap_zero_page: CMAP3 busy");
2223 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2224 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2225 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2226 for (i = 0; i < PAGE_SIZE; i += 4) {
2227 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2228 panic("pmap_page_assertzero() @ %p not zero!\n",
2229 (void *)gd->gd_CADDR3);
2239 * Zero part of a physical page by mapping it into memory and clearing
2240 * its contents with bzero.
2242 * off and size may not cover an area beyond a single hardware page.
2245 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2247 struct mdglobaldata *gd = mdcpu;
2251 panic("pmap_zero_page: CMAP3 busy");
2252 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2253 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2254 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2256 bzero((char *)gd->gd_CADDR3 + off, size);
2264 * Copy the physical page from the source PA to the target PA.
2265 * This function may be called from an interrupt. No locking
2269 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2271 struct mdglobaldata *gd = mdcpu;
2274 if (*(int *) gd->gd_CMAP1)
2275 panic("pmap_copy_page: CMAP1 busy");
2276 if (*(int *) gd->gd_CMAP2)
2277 panic("pmap_copy_page: CMAP2 busy");
2279 *(int *) gd->gd_CMAP1 = VPTE_V | (src & PG_FRAME) | PG_A;
2280 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2282 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2283 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2285 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2287 *(int *) gd->gd_CMAP1 = 0;
2288 *(int *) gd->gd_CMAP2 = 0;
2293 * pmap_copy_page_frag:
2295 * Copy the physical page from the source PA to the target PA.
2296 * This function may be called from an interrupt. No locking
2300 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2302 struct mdglobaldata *gd = mdcpu;
2305 if (*(int *) gd->gd_CMAP1)
2306 panic("pmap_copy_page: CMAP1 busy");
2307 if (*(int *) gd->gd_CMAP2)
2308 panic("pmap_copy_page: CMAP2 busy");
2310 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2311 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2313 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2314 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2316 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2317 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2320 *(int *) gd->gd_CMAP1 = 0;
2321 *(int *) gd->gd_CMAP2 = 0;
2326 * Returns true if the pmap's pv is one of the first
2327 * 16 pvs linked to from this page. This count may
2328 * be changed upwards or downwards in the future; it
2329 * is only necessary that true be returned for a small
2330 * subset of pmaps for proper page aging.
2333 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2338 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2343 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2344 if (pv->pv_pmap == pmap) {
2357 * Remove all pages from specified address space
2358 * this aids process exit speeds. Also, this code
2359 * is special cased for current process only, but
2360 * can have the more generic (and slightly slower)
2361 * mode enabled. This is much faster than pmap_remove
2362 * in the case of running down an entire address space.
2365 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2370 pmap_inval_info info;
2373 if (curproc && pmap == vmspace_pmap(curproc->p_vmspace))
2378 pmap_inval_init(&info);
2380 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2381 if (pv->pv_va >= eva || pv->pv_va < sva) {
2382 npv = TAILQ_NEXT(pv, pv_plist);
2386 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2387 if (pmap->pm_active)
2388 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2392 * We cannot remove wired pages from a process' mapping
2395 if (tpte & VPTE_W) {
2396 npv = TAILQ_NEXT(pv, pv_plist);
2401 m = PHYS_TO_VM_PAGE(tpte);
2403 KASSERT(m < &vm_page_array[vm_page_array_size],
2404 ("pmap_remove_pages: bad tpte %x", tpte));
2406 pv->pv_pmap->pm_stats.resident_count--;
2409 * Update the vm_page_t clean and reference bits.
2411 if (tpte & VPTE_M) {
2416 npv = TAILQ_NEXT(pv, pv_plist);
2417 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2419 m->md.pv_list_count--;
2420 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2421 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2422 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2425 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2428 pmap_inval_flush(&info);
2433 * pmap_testbit tests bits in pte's
2434 * note that the testbit/changebit routines are inline,
2435 * and a lot of things compile-time evaluate.
2438 pmap_testbit(vm_page_t m, int bit)
2443 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2446 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2451 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2453 * if the bit being tested is the modified bit, then
2454 * mark clean_map and ptes as never
2457 if (bit & (VPTE_A|VPTE_M)) {
2458 if (!pmap_track_modified(pv->pv_va))
2462 #if defined(PMAP_DIAGNOSTIC)
2464 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2468 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2479 * this routine is used to modify bits in ptes
2481 static __inline void
2482 pmap_changebit(vm_page_t m, int bit, boolean_t setem)
2484 struct pmap_inval_info info;
2488 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2491 pmap_inval_init(&info);
2495 * Loop over all current mappings setting/clearing as appropos If
2496 * setting RO do we need to clear the VAC?
2498 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2500 * don't write protect pager mappings
2502 if (!setem && (bit == VPTE_W)) {
2503 if (!pmap_track_modified(pv->pv_va))
2507 #if defined(PMAP_DIAGNOSTIC)
2509 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2515 * Careful here. We can use a locked bus instruction to
2516 * clear VPTE_A or VPTE_M safely but we need to synchronize
2517 * with the target cpus when we mess with VPTE_W.
2519 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2521 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2525 atomic_set_int(pte, bit);
2527 atomic_set_int_nonlocked(pte, bit);
2530 vm_offset_t pbits = *(vm_offset_t *)pte;
2532 if (bit == VPTE_W) {
2533 if (pbits & VPTE_M) {
2537 atomic_clear_int(pte, VPTE_M|VPTE_W);
2539 atomic_clear_int_nonlocked(pte, VPTE_M|VPTE_W);
2543 atomic_clear_int(pte, bit);
2545 atomic_clear_int_nonlocked(pte, bit);
2551 pmap_inval_flush(&info);
2556 * pmap_page_protect:
2558 * Lower the permission for all mappings to a given page.
2561 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2563 if ((prot & VM_PROT_WRITE) == 0) {
2564 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2565 pmap_changebit(m, VPTE_W, FALSE);
2573 pmap_phys_address(int ppn)
2575 return (i386_ptob(ppn));
2579 * pmap_ts_referenced:
2581 * Return a count of reference bits for a page, clearing those bits.
2582 * It is not necessary for every reference bit to be cleared, but it
2583 * is necessary that 0 only be returned when there are truly no
2584 * reference bits set.
2586 * XXX: The exact number of bits to check and clear is a matter that
2587 * should be tested and standardized at some point in the future for
2588 * optimal aging of shared pages.
2591 pmap_ts_referenced(vm_page_t m)
2593 pv_entry_t pv, pvf, pvn;
2597 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2602 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2607 pvn = TAILQ_NEXT(pv, pv_list);
2609 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2611 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2613 if (!pmap_track_modified(pv->pv_va))
2616 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2618 if (pte && (*pte & VPTE_A)) {
2620 atomic_clear_int(pte, VPTE_A);
2622 atomic_clear_int_nonlocked(pte, VPTE_A);
2629 } while ((pv = pvn) != NULL && pv != pvf);
2639 * Return whether or not the specified physical page was modified
2640 * in any physical maps.
2643 pmap_is_modified(vm_page_t m)
2645 return pmap_testbit(m, VPTE_M);
2649 * Clear the modify bits on the specified physical page.
2652 pmap_clear_modify(vm_page_t m)
2654 pmap_changebit(m, VPTE_M, FALSE);
2658 * pmap_clear_reference:
2660 * Clear the reference bit on the specified physical page.
2663 pmap_clear_reference(vm_page_t m)
2665 pmap_changebit(m, VPTE_A, FALSE);
2669 * Miscellaneous support routines follow
2673 i386_protection_init(void)
2677 kp = protection_codes;
2678 for (prot = 0; prot < 8; prot++) {
2679 if (prot & VM_PROT_READ)
2681 if (prot & VM_PROT_WRITE)
2683 if (prot & VM_PROT_EXECUTE)
2690 * Map a set of physical memory pages into the kernel virtual
2691 * address space. Return a pointer to where it is mapped. This
2692 * routine is intended to be used for mapping device memory,
2695 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2699 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2701 vm_offset_t va, tmpva, offset;
2704 offset = pa & PAGE_MASK;
2705 size = roundup(offset + size, PAGE_SIZE);
2707 va = kmem_alloc_nofault(&kernel_map, size);
2709 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2711 pa = pa & VPTE_FRAME;
2712 for (tmpva = va; size > 0;) {
2713 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2714 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2722 return ((void *)(va + offset));
2726 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2728 vm_offset_t base, offset;
2730 base = va & VPTE_FRAME;
2731 offset = va & PAGE_MASK;
2732 size = roundup(offset + size, PAGE_SIZE);
2733 pmap_qremove(va, size >> PAGE_SHIFT);
2734 kmem_free(&kernel_map, base, size);
2738 * perform the pmap work for mincore
2741 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2747 ptep = pmap_pte(pmap, addr);
2752 if ((pte = *ptep) != 0) {
2755 val = MINCORE_INCORE;
2756 if ((pte & VPTE_MANAGED) == 0)
2759 pa = pte & VPTE_FRAME;
2761 m = PHYS_TO_VM_PAGE(pa);
2767 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2769 * Modified by someone
2771 else if (m->dirty || pmap_is_modified(m))
2772 val |= MINCORE_MODIFIED_OTHER;
2777 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2780 * Referenced by someone
2782 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
2783 val |= MINCORE_REFERENCED_OTHER;
2784 vm_page_flag_set(m, PG_REFERENCED);
2791 pmap_activate(struct proc *p)
2795 pmap = vmspace_pmap(p->p_vmspace);
2797 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2799 pmap->pm_active |= 1;
2801 #if defined(SWTCH_OPTIM_STATS)
2804 panic("pmap_activate"); /* XXX store vmspace id in context */
2806 p->p_thread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
2807 load_cr3(p->p_thread->td_pcb->pcb_cr3);
2812 pmap_deactivate(struct proc *p)
2816 pmap = vmspace_pmap(p->p_vmspace);
2818 atomic_clear_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2820 pmap->pm_active &= ~1;
2823 * XXX - note we do not adjust %cr3. The caller is expected to
2824 * activate a new pmap or do a thread-exit.
2829 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
2832 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
2836 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
2843 static void pads (pmap_t pm);
2844 void pmap_pvdump (vm_paddr_t pa);
2846 /* print address space of pmap*/
2854 if (pm == &kernel_pmap)
2856 for (i = 0; i < 1024; i++)
2858 for (j = 0; j < 1024; j++) {
2859 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
2860 if (pm == &kernel_pmap && va < KERNBASE)
2862 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
2864 ptep = pmap_pte(pm, va);
2865 if (ptep && (*ptep & VPTE_V)) {
2867 (void *)va, (unsigned)*ptep);
2874 pmap_pvdump(vm_paddr_t pa)
2879 kprintf("pa %08llx", (long long)pa);
2880 m = PHYS_TO_VM_PAGE(pa);
2881 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2883 kprintf(" -> pmap %p, va %x, flags %x",
2884 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
2886 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);