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 1.2 2007/01/05 22:18:20 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 * Page table pages are not managed like they are in normal pmaps, so
133 * no pteobj is needed.
138 vm_pindex_t i = (((vm_offset_t)KernelPTD - KvaStart) >> PAGE_SHIFT);
140 kernel_pmap.pm_pdir = KernelPTD;
141 kernel_pmap.pm_pdirpte = KernelPTA[i];
142 kernel_pmap.pm_count = 1;
143 kernel_pmap.pm_active = (cpumask_t)-1;
144 TAILQ_INIT(&kernel_pmap.pm_pvlist);
145 i386_protection_init();
149 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
150 * just dummy it up so it works well enough for fork().
152 * In DragonFly, process pmaps may only be used to manipulate user address
153 * space, never kernel address space.
156 pmap_pinit0(struct pmap *pmap)
161 /************************************************************************
162 * Procedures to manage whole physical maps *
163 ************************************************************************
165 * Initialize a preallocated and zeroed pmap structure,
166 * such as one in a vmspace structure.
169 pmap_pinit(struct pmap *pmap)
175 * No need to allocate page table space yet but we do need a valid
176 * page directory table.
178 if (pmap->pm_pdir == NULL) {
180 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
184 * allocate object for the pte array and page directory
186 npages = VPTE_PAGETABLE_SIZE +
187 (VM_MAX_USER_ADDRESS / PAGE_SIZE) * sizeof(vpte_t);
188 npages = (npages + PAGE_MASK) / PAGE_SIZE;
190 if (pmap->pm_pteobj == NULL)
191 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, npages);
192 pmap->pm_pdindex = npages - 1;
195 * allocate the page directory page
197 ptdpg = vm_page_grab(pmap->pm_pteobj, pmap->pm_pdindex,
198 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
200 ptdpg->wire_count = 1;
201 ++vmstats.v_wire_count;
203 /* not usually mapped */
204 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
205 ptdpg->valid = VM_PAGE_BITS_ALL;
207 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
208 if ((ptdpg->flags & PG_ZERO) == 0)
209 bzero(pmap->pm_pdir, PAGE_SIZE);
213 pmap->pm_ptphint = NULL;
214 TAILQ_INIT(&pmap->pm_pvlist);
215 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
219 * Wire in kernel global address entries. To avoid a race condition
220 * between pmap initialization and pmap_growkernel, this procedure
221 * adds the pmap to the master list (which growkernel scans to update),
222 * then copies the template.
224 * In a virtual kernel there are no kernel global address entries.
227 pmap_pinit2(struct pmap *pmap)
230 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
235 * Release all resources held by the given physical map.
237 * Should only be called if the map contains no valid mappings.
239 static int pmap_release_callback(struct vm_page *p, void *data);
242 pmap_release(struct pmap *pmap)
244 vm_object_t object = pmap->pm_pteobj;
245 struct rb_vm_page_scan_info info;
247 KKASSERT(pmap != &kernel_pmap);
249 #if defined(DIAGNOSTIC)
250 if (object->ref_count != 1)
251 panic("pmap_release: pteobj reference count != 1");
255 info.object = object;
257 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
264 info.limit = object->generation;
266 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
267 pmap_release_callback, &info);
268 if (info.error == 0 && info.mpte) {
269 if (!pmap_release_free_page(pmap, info.mpte))
273 } while (info.error);
277 pmap_release_callback(struct vm_page *p, void *data)
279 struct rb_vm_page_scan_info *info = data;
281 if (p->pindex == info->pmap->pm_pdindex) {
285 if (!pmap_release_free_page(info->pmap, p)) {
289 if (info->object->generation != info->limit) {
297 * Retire the given physical map from service. Should only be called if
298 * the map contains no valid mappings.
301 pmap_destroy(pmap_t pmap)
308 count = --pmap->pm_count;
311 panic("destroying a pmap is not yet implemented");
316 * Add a reference to the specified pmap.
319 pmap_reference(pmap_t pmap)
326 /************************************************************************
327 * Procedures which operate directly on the kernel PMAP *
328 ************************************************************************/
331 * This maps the requested page table and gives us access to it.
334 get_ptbase(struct pmap *pmap)
336 struct mdglobaldata *gd = mdcpu;
338 if (pmap == &kernel_pmap) {
340 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
341 return(gd->gd_PT1map);
342 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
343 return(gd->gd_PT2map);
347 * Otherwise choose one or the other and map the page table
348 * in the KVA space reserved for it.
350 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
351 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
353 if ((gd->gd_PTflip = 1 - gd->gd_PTflip) == 0) {
354 gd->gd_PT1pdir = pmap->pm_pdir;
355 *gd->gd_PT1pde = pmap->pm_pdirpte;
356 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
357 return(gd->gd_PT1map);
359 gd->gd_PT2pdir = pmap->pm_pdir;
360 *gd->gd_PT2pde = pmap->pm_pdirpte;
361 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
362 return(gd->gd_PT2map);
367 get_ptbase1(struct pmap *pmap)
369 struct mdglobaldata *gd = mdcpu;
371 if (pmap == &kernel_pmap) {
373 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
374 return(gd->gd_PT1map);
376 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
377 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
378 gd->gd_PT1pdir = pmap->pm_pdir;
379 *gd->gd_PT1pde = pmap->pm_pdirpte;
380 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
381 return(gd->gd_PT1map);
385 get_ptbase2(struct pmap *pmap)
387 struct mdglobaldata *gd = mdcpu;
389 if (pmap == &kernel_pmap) {
391 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
392 return(gd->gd_PT2map);
394 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
395 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
396 gd->gd_PT2pdir = pmap->pm_pdir;
397 *gd->gd_PT2pde = pmap->pm_pdirpte;
398 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
399 return(gd->gd_PT2map);
403 * Return a pointer to the page table entry for the specified va in the
404 * specified pmap. NULL is returned if there is no valid page table page
407 static __inline vpte_t *
408 pmap_pte(struct pmap *pmap, vm_offset_t va)
412 ptep = &pmap->pm_pdir[va >> PAGE_SHIFT];
416 return (get_ptbase(pmap) + (va >> PAGE_SHIFT));
422 * Enter a mapping into kernel_pmap. Mappings created in this fashion
426 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
431 pmap_inval_info info;
434 KKASSERT(va >= KvaStart && va < KvaEnd);
435 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
436 ptep = KernelPTA + ((va - KvaStart) >> PAGE_SHIFT);
437 if (*ptep & VPTE_V) {
439 pmap_inval_init(&info);
440 pmap_inval_add(&info, &kernel_pmap, va);
444 pmap_inval_flush(&info);
446 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
454 pmap_kenter_sync(vm_offset_t va)
456 pmap_inval_info info;
458 pmap_inval_init(&info);
459 pmap_inval_add(&info, &kernel_pmap, va);
460 pmap_inval_flush(&info);
464 pmap_kenter_sync_quick(vm_offset_t va)
466 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
470 * Map a contiguous range of physical memory to a KVM
473 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
475 while (start < end) {
476 pmap_kenter(virt, start);
484 pmap_kpte(vm_offset_t va)
488 KKASSERT(va >= KvaStart && va < KvaEnd);
489 ptep = KernelPTA + ((va - KvaStart) >> PAGE_SHIFT);
494 * Enter a mapping into kernel_pmap without any SMP interactions.
496 * Mappings created in this fashion are not managed.
499 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
504 KKASSERT(va >= KvaStart && va < KvaEnd);
506 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
507 ptep = KernelPTA + ((va - KvaStart) >> PAGE_SHIFT);
508 if (*ptep & VPTE_V) {
510 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
517 * Make a temporary mapping for a physical address. This is only intended
518 * to be used for panic dumps.
521 pmap_kenter_temporary(vm_paddr_t pa, int i)
523 pmap_kenter(crashdumpmap + (i * PAGE_SIZE), pa);
524 return ((void *)crashdumpmap);
528 * Remove an unmanaged mapping created with pmap_kenter*().
531 pmap_kremove(vm_offset_t va)
535 pmap_inval_info info;
538 KKASSERT(va >= KvaStart && va < KvaEnd);
540 ptep = KernelPTA + ((va - KvaStart) >> PAGE_SHIFT);
541 if (*ptep & VPTE_V) {
543 pmap_inval_init(&info);
544 pmap_inval_add(&info, &kernel_pmap, va);
548 pmap_inval_flush(&info);
550 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
559 * Remove an unmanaged mapping created with pmap_kenter*() without
560 * going through any SMP interactions.
563 pmap_kremove_quick(vm_offset_t va)
567 KKASSERT(va >= KvaStart && va < KvaEnd);
569 ptep = KernelPTA + ((va - KvaStart) >> PAGE_SHIFT);
570 if (*ptep & VPTE_V) {
572 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
579 * Extract the physical address from the kernel_pmap that is associated
580 * with the specified virtual address.
583 pmap_kextract(vm_offset_t va)
588 KKASSERT(va >= KvaStart && va < KvaEnd);
590 ptep = KernelPTA + ((va - KvaStart) >> PAGE_SHIFT);
591 pa = (vm_paddr_t)(*ptep & VPTE_FRAME) | (va & PAGE_MASK);
596 * Map a set of unmanaged VM pages into KVM.
599 pmap_qenter(vm_offset_t va, struct vm_page **m, int count)
601 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
605 ptep = KernelPTA + ((va - KvaStart) >> PAGE_SHIFT);
607 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
608 *ptep = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
620 * Map a set of VM pages to kernel virtual memory. If a mapping changes
621 * clear the supplied mask. The caller handles any SMP interactions.
622 * The mask is used to provide the caller with hints on what SMP interactions
626 pmap_qenter2(vm_offset_t va, struct vm_page **m, int count, cpumask_t *mask)
628 cpumask_t cmask = mycpu->gd_cpumask;
630 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
635 ptep = KernelPTA + ((va - KvaStart) >> PAGE_SHIFT);
636 npte = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
640 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
641 } else if ((*mask & cmask) == 0) {
642 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
652 * Undo the effects of pmap_qenter*().
655 pmap_qremove(vm_offset_t va, int count)
657 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
661 ptep = KernelPTA + ((va - KvaStart) >> PAGE_SHIFT);
663 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
674 /************************************************************************
675 * Misc support glue called by machine independant code *
676 ************************************************************************
678 * These routines are called by machine independant code to operate on
679 * certain machine-dependant aspects of processes, threads, and pmaps.
683 * Initialize MD portions of the thread structure.
686 pmap_init_thread(thread_t td)
688 /* enforce pcb placement */
689 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
690 td->td_savefpu = &td->td_pcb->pcb_save;
691 td->td_sp = (char *)td->td_pcb - 16;
695 * Initialize MD portions of a process structure. XXX this aint MD
698 pmap_init_proc(struct proc *p, struct thread *td)
700 p->p_addr = (void *)td->td_kstack;
703 td->td_lwp = &p->p_lwp;
704 td->td_switch = cpu_heavy_switch;
706 KKASSERT(td->td_mpcount == 1);
708 bzero(p->p_addr, sizeof(*p->p_addr));
712 * Destroy the UPAGES for a process that has exited and disassociate
713 * the process from its thread.
716 pmap_dispose_proc(struct proc *p)
720 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
722 if ((td = p->p_thread) != NULL) {
731 * We pre-allocate all page table pages for kernel virtual memory so
732 * this routine will only be called if KVM has been exhausted.
735 pmap_growkernel(vm_offset_t size)
737 panic("KVM exhausted");
741 * The modification bit is not tracked for any pages in this range. XXX
742 * such pages in this maps should always use pmap_k*() functions and not
746 pmap_track_modified(vm_offset_t va)
748 if ((va < clean_sva) || (va >= clean_eva))
754 /************************************************************************
755 * Procedures supporting managed page table pages *
756 ************************************************************************
758 * These procedures are used to track managed page table pages. These pages
759 * use the page table page's vm_page_t to track PTEs in the page. The
760 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
762 * This allows the system to throw away page table pages for user processes
763 * at will and reinstantiate them on demand.
767 * This routine works like vm_page_lookup() but also blocks as long as the
768 * page is busy. This routine does not busy the page it returns.
770 * Unless the caller is managing objects whos pages are in a known state,
771 * the call should be made with a critical section held so the page's object
772 * association remains valid on return.
775 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
780 m = vm_page_lookup(object, pindex);
781 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
787 * This routine unholds page table pages, and if the hold count
788 * drops to zero, then it decrements the wire count.
791 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
793 pmap_inval_flush(info);
794 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
797 if (m->hold_count == 0) {
799 * unmap the page table page
801 pmap_inval_add(info, pmap, -1);
802 pmap->pm_pdir[m->pindex] = 0;
803 --pmap->pm_stats.resident_count;
805 if (pmap->pm_ptphint == m)
806 pmap->pm_ptphint = NULL;
809 * If the page is finally unwired, simply free it.
812 if (m->wire_count == 0) {
815 vm_page_free_zero(m);
816 --vmstats.v_wire_count;
824 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
827 if (m->hold_count == 0)
828 return _pmap_unwire_pte_hold(pmap, m, info);
834 * After removing a page table entry, this routine is used to
835 * conditionally free the page, and manage the hold/wire counts.
838 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
839 pmap_inval_info_t info)
845 * page table pages in the kernel_pmap are not managed.
847 if (pmap == &kernel_pmap)
849 ptepindex = (va >> PDRSHIFT);
850 if (pmap->pm_ptphint &&
851 (pmap->pm_ptphint->pindex == ptepindex)) {
852 mpte = pmap->pm_ptphint;
854 pmap_inval_flush(info);
855 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
856 pmap->pm_ptphint = mpte;
859 return pmap_unwire_pte_hold(pmap, mpte, info);
863 * Attempt to release and free an vm_page in a pmap. Returns 1 on success,
864 * 0 on failure (if the procedure had to sleep).
867 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
869 vpte_t *pde = pmap->pm_pdir;
871 * This code optimizes the case of freeing non-busy
872 * page-table pages. Those pages are zero now, and
873 * might as well be placed directly into the zero queue.
875 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
881 * Remove the page table page from the processes address space.
884 pmap->pm_stats.resident_count--;
887 panic("pmap_release: freeing held page table page");
890 * Page directory pages need to have the kernel stuff cleared, so
891 * they can go into the zero queue also.
893 * In virtual kernels there is no 'kernel stuff'. For the moment
894 * I just make sure the whole thing has been zero'd even though
895 * it should already be completely zero'd.
897 if (p->pindex == pmap->pm_pdindex) {
898 bzero(pde, VPTE_PAGETABLE_SIZE);
899 pmap_kremove((vm_offset_t)pmap->pm_pdir);
903 * Clear the matching hint
905 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
906 pmap->pm_ptphint = NULL;
909 * And throw the page away. The page is completely zero'd out so
910 * optimize the free call.
913 vmstats.v_wire_count--;
914 vm_page_free_zero(p);
919 * This routine is called if the page table page is not mapped in the page
922 * The routine is broken up into two parts for readability.
925 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
931 * Find or fabricate a new pagetable page
933 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
934 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
936 KASSERT(m->queue == PQ_NONE,
937 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
939 if (m->wire_count == 0)
940 vmstats.v_wire_count++;
944 * Increment the hold count for the page table page
945 * (denoting a new mapping.)
950 * Map the pagetable page into the process address space, if
951 * it isn't already there.
953 pmap->pm_stats.resident_count++;
955 ptepa = VM_PAGE_TO_PHYS(m);
956 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
960 * We are likely about to access this page table page, so set the
961 * page table hint to reduce overhead.
963 pmap->pm_ptphint = m;
966 * Try to use the new mapping, but if we cannot, then
967 * do it with the routine that maps the page explicitly.
969 if ((m->flags & PG_ZERO) == 0)
970 pmap_zero_page(ptepa);
972 m->valid = VM_PAGE_BITS_ALL;
973 vm_page_flag_clear(m, PG_ZERO);
974 vm_page_flag_set(m, PG_MAPPED);
981 * Determine the page table page required to access the VA in the pmap
982 * and allocate it if necessary. Return a held vm_page_t for the page.
984 * Only used with user pmaps.
987 pmap_allocpte(pmap_t pmap, vm_offset_t va)
994 * Calculate pagetable page index
996 ptepindex = va >> PDRSHIFT;
999 * Get the page directory entry
1001 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1004 * This supports switching from a 4MB page to a
1007 if (ptepa & VPTE_PS) {
1008 pmap->pm_pdir[ptepindex] = 0;
1015 * If the page table page is mapped, we just increment the
1016 * hold count, and activate it.
1020 * In order to get the page table page, try the
1023 if (pmap->pm_ptphint &&
1024 (pmap->pm_ptphint->pindex == ptepindex)) {
1025 m = pmap->pm_ptphint;
1027 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1028 pmap->pm_ptphint = m;
1034 * Here if the pte page isn't mapped, or if it has been deallocated.
1036 return _pmap_allocpte(pmap, ptepindex);
1039 /************************************************************************
1040 * Managed pages in pmaps *
1041 ************************************************************************
1043 * All pages entered into user pmaps and some pages entered into the kernel
1044 * pmap are managed, meaning that pmap_protect() and other related management
1045 * functions work on these pages.
1049 * free the pv_entry back to the free list. This function may be
1050 * called from an interrupt.
1052 static __inline void
1053 free_pv_entry(pv_entry_t pv)
1060 * get a new pv_entry, allocating a block from the system
1061 * when needed. This function may be called from an interrupt.
1067 if (pv_entry_high_water &&
1068 (pv_entry_count > pv_entry_high_water) &&
1069 (pmap_pagedaemon_waken == 0)) {
1070 pmap_pagedaemon_waken = 1;
1071 wakeup (&vm_pages_needed);
1073 return zalloc(&pvzone);
1077 * This routine is very drastic, but can save the system
1085 static int warningdone=0;
1087 if (pmap_pagedaemon_waken == 0)
1090 if (warningdone < 5) {
1091 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1095 for(i = 0; i < vm_page_array_size; i++) {
1096 m = &vm_page_array[i];
1097 if (m->wire_count || m->hold_count || m->busy ||
1098 (m->flags & PG_BUSY))
1102 pmap_pagedaemon_waken = 0;
1106 * If it is the first entry on the list, it is actually
1107 * in the header and we must copy the following entry up
1108 * to the header. Otherwise we must search the list for
1109 * the entry. In either case we free the now unused entry.
1112 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1113 vm_offset_t va, pmap_inval_info_t info)
1119 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1120 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1121 if (pmap == pv->pv_pmap && va == pv->pv_va)
1125 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1126 if (va == pv->pv_va)
1132 * Note that pv_ptem is NULL if the page table page itself is not
1133 * managed, even if the page being removed IS managed.
1137 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1138 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1139 m->md.pv_list_count--;
1140 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1141 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1142 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1150 * Create a pv entry for page at pa for (pmap, va). If the page table page
1151 * holding the VA is managed, mpte will be non-NULL.
1154 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1159 pv = get_pv_entry();
1164 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1165 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1166 m->md.pv_list_count++;
1172 * pmap_remove_pte: do the things to unmap a page in a process
1175 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va,
1176 pmap_inval_info_t info)
1181 pmap_inval_add(info, pmap, va);
1182 oldpte = loadandclear(ptq);
1183 if (oldpte & VPTE_W)
1184 pmap->pm_stats.wired_count -= 1;
1186 * Machines that don't support invlpg, also don't support
1187 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1190 if (oldpte & VPTE_G)
1191 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1192 pmap->pm_stats.resident_count -= 1;
1193 if (oldpte & PG_MANAGED) {
1194 m = PHYS_TO_VM_PAGE(oldpte);
1195 if (oldpte & VPTE_M) {
1196 #if defined(PMAP_DIAGNOSTIC)
1197 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1199 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1203 if (pmap_track_modified(va))
1206 if (oldpte & VPTE_A)
1207 vm_page_flag_set(m, PG_REFERENCED);
1208 return pmap_remove_entry(pmap, m, va, info);
1210 return pmap_unuse_pt(pmap, va, NULL, info);
1219 * Remove a single page from a process address space.
1221 * This function may not be called from an interrupt if the pmap is
1225 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1230 * if there is no pte for this address, just skip it!!! Otherwise
1231 * get a local va for mappings for this pmap and remove the entry.
1233 if (*pmap_pde(pmap, va) != 0) {
1234 ptq = get_ptbase(pmap) + (va >> PAGE_SHIFT);
1236 pmap_remove_pte(pmap, ptq, va, info);
1244 * Remove the given range of addresses from the specified map.
1246 * It is assumed that the start and end are properly
1247 * rounded to the page size.
1249 * This function may not be called from an interrupt if the pmap is
1253 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1257 vm_offset_t ptpaddr;
1258 vm_offset_t sindex, eindex;
1259 struct pmap_inval_info info;
1264 if (pmap->pm_stats.resident_count == 0)
1267 pmap_inval_init(&info);
1270 * special handling of removing one page. a very
1271 * common operation and easy to short circuit some
1274 if (((sva + PAGE_SIZE) == eva) &&
1275 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1276 pmap_remove_page(pmap, sva, &info);
1277 pmap_inval_flush(&info);
1282 * Get a local virtual address for the mappings that are being
1285 ptbase = get_ptbase(pmap);
1287 sindex = (sva >> PAGE_SHIFT);
1288 eindex = (eva >> PAGE_SHIFT);
1290 for (; sindex < eindex; sindex = pdnxt) {
1294 * Calculate index for next page table.
1296 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1297 if (pmap->pm_stats.resident_count == 0)
1300 pdirindex = sindex / NPDEPG;
1301 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1302 pmap_inval_add(&info, pmap, -1);
1303 pmap->pm_pdir[pdirindex] = 0;
1304 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1309 * Weed out invalid mappings. Note: we assume that the page
1310 * directory table is always allocated, and in kernel virtual.
1316 * Limit our scan to either the end of the va represented
1317 * by the current page table page, or to the end of the
1318 * range being removed.
1320 if (pdnxt > eindex) {
1324 for (; sindex != pdnxt; sindex++) {
1326 if (ptbase[sindex] == 0)
1328 va = i386_ptob(sindex);
1329 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1333 pmap_inval_flush(&info);
1339 * Removes this physical page from all physical maps in which it resides.
1340 * Reflects back modify bits to the pager.
1342 * This routine may not be called from an interrupt.
1345 pmap_remove_all(vm_page_t m)
1347 struct pmap_inval_info info;
1351 #if defined(PMAP_DIAGNOSTIC)
1353 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1356 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1357 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1361 pmap_inval_init(&info);
1363 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1364 pv->pv_pmap->pm_stats.resident_count--;
1366 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1367 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1369 tpte = loadandclear(pte);
1371 pv->pv_pmap->pm_stats.wired_count--;
1374 vm_page_flag_set(m, PG_REFERENCED);
1377 * Update the vm_page_t clean and reference bits.
1379 if (tpte & VPTE_M) {
1380 #if defined(PMAP_DIAGNOSTIC)
1381 if (pmap_nw_modified((pt_entry_t) tpte)) {
1383 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1387 if (pmap_track_modified(pv->pv_va))
1390 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1391 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1392 m->md.pv_list_count--;
1393 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1397 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1399 pmap_inval_flush(&info);
1405 * Set the physical protection on the specified range of this map
1408 * This function may not be called from an interrupt if the map is
1409 * not the kernel_pmap.
1412 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1415 vm_offset_t pdnxt, ptpaddr;
1416 vm_pindex_t sindex, eindex;
1417 pmap_inval_info info;
1422 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1423 pmap_remove(pmap, sva, eva);
1427 if (prot & VM_PROT_WRITE)
1430 pmap_inval_init(&info);
1432 ptbase = get_ptbase(pmap);
1434 sindex = (sva >> PAGE_SHIFT);
1435 eindex = (eva >> PAGE_SHIFT);
1437 for (; sindex < eindex; sindex = pdnxt) {
1441 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1443 pdirindex = sindex / NPDEPG;
1444 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1445 pmap_inval_add(&info, pmap, -1);
1446 pmap->pm_pdir[pdirindex] &= ~(VPTE_M|VPTE_W);
1447 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1452 * Weed out invalid mappings. Note: we assume that the page
1453 * directory table is always allocated, and in kernel virtual.
1458 if (pdnxt > eindex) {
1462 for (; sindex != pdnxt; sindex++) {
1467 /* XXX this isn't optimal */
1468 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1469 pbits = ptbase[sindex];
1471 if (pbits & PG_MANAGED) {
1473 if (pbits & VPTE_A) {
1474 m = PHYS_TO_VM_PAGE(pbits);
1475 vm_page_flag_set(m, PG_REFERENCED);
1478 if (pbits & VPTE_M) {
1479 if (pmap_track_modified(i386_ptob(sindex))) {
1481 m = PHYS_TO_VM_PAGE(pbits);
1490 if (pbits != ptbase[sindex]) {
1491 ptbase[sindex] = pbits;
1495 pmap_inval_flush(&info);
1499 * Enter a managed page into a pmap. If the page is not wired related pmap
1500 * data can be destroyed at any time for later demand-operation.
1502 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1503 * specified protection, and wire the mapping if requested.
1505 * NOTE: This routine may not lazy-evaluate or lose information. The
1506 * page must actually be inserted into the given map NOW.
1508 * NOTE: When entering a page at a KVA address, the pmap must be the
1512 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1518 vm_offset_t origpte, newpte;
1520 pmap_inval_info info;
1528 * Get the page table page. The kernel_pmap's page table pages
1529 * are preallocated and have no associated vm_page_t.
1531 if (pmap == &kernel_pmap)
1534 mpte = pmap_allocpte(pmap, va);
1536 pmap_inval_init(&info);
1537 pte = pmap_pte(pmap, va);
1540 * Page Directory table entry not valid, we need a new PT page
1541 * and pmap_allocpte() didn't give us one. Oops!
1544 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1548 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1549 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1551 opa = origpte & VPTE_FRAME;
1553 if (origpte & VPTE_PS)
1554 panic("pmap_enter: attempted pmap_enter on 4MB page");
1557 * Mapping has not changed, must be protection or wiring change.
1559 if (origpte && (opa == pa)) {
1561 * Wiring change, just update stats. We don't worry about
1562 * wiring PT pages as they remain resident as long as there
1563 * are valid mappings in them. Hence, if a user page is wired,
1564 * the PT page will be also.
1566 if (wired && ((origpte & VPTE_W) == 0))
1567 pmap->pm_stats.wired_count++;
1568 else if (!wired && (origpte & VPTE_W))
1569 pmap->pm_stats.wired_count--;
1571 #if defined(PMAP_DIAGNOSTIC)
1572 if (pmap_nw_modified((pt_entry_t) origpte)) {
1574 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1580 * Remove the extra pte reference. Note that we cannot
1581 * optimize the RO->RW case because we have adjusted the
1582 * wiring count above and may need to adjust the wiring
1589 * We might be turning off write access to the page,
1590 * so we go ahead and sense modify status.
1592 if (origpte & PG_MANAGED) {
1593 if ((origpte & VPTE_M) && pmap_track_modified(va)) {
1595 om = PHYS_TO_VM_PAGE(opa);
1603 * Mapping has changed, invalidate old range and fall through to
1604 * handle validating new mapping.
1608 err = pmap_remove_pte(pmap, pte, va, &info);
1610 panic("pmap_enter: pte vanished, va: 0x%x", va);
1614 * Enter on the PV list if part of our managed memory. Note that we
1615 * raise IPL while manipulating pv_table since pmap_enter can be
1616 * called at interrupt time.
1618 if (pmap_initialized &&
1619 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1620 pmap_insert_entry(pmap, va, mpte, m);
1625 * Increment counters
1627 pmap->pm_stats.resident_count++;
1629 pmap->pm_stats.wired_count++;
1633 * Now validate mapping with desired protection/wiring.
1635 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1642 * if the mapping or permission bits are different, we need
1643 * to update the pte.
1645 if ((origpte & ~(VPTE_M|VPTE_A)) != newpte) {
1646 *pte = newpte | VPTE_A;
1648 pmap_inval_flush(&info);
1652 * This is a quick version of pmap_enter(). It is used only under the
1653 * following conditions:
1655 * (1) The pmap is not the kernel_pmap
1656 * (2) The page is not to be wired into the map
1657 * (3) The page is to mapped read-only in the pmap (initially that is)
1658 * (4) The calling procedure is responsible for flushing the TLB
1659 * (5) The page is always managed
1660 * (6) There is no prior mapping at the VA
1664 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
1668 pmap_inval_info info;
1670 KKASSERT(pmap != &kernel_pmap);
1671 pmap_inval_init(&info);
1673 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1676 * Instantiate the page table page if required
1682 * Calculate pagetable page index
1684 ptepindex = va >> PDRSHIFT;
1685 if (mpte && (mpte->pindex == ptepindex)) {
1690 * Get the page directory entry
1692 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1695 * If the page table page is mapped, we just increment
1696 * the hold count, and activate it.
1699 if (ptepa & VPTE_PS)
1700 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1701 if (pmap->pm_ptphint &&
1702 (pmap->pm_ptphint->pindex == ptepindex)) {
1703 mpte = pmap->pm_ptphint;
1705 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1706 pmap->pm_ptphint = mpte;
1712 mpte = _pmap_allocpte(pmap, ptepindex);
1717 * Ok, now that the page table page has been validated, get the pte.
1718 * If the pte is already mapped undo mpte's hold_count and
1721 pte = pmap_pte(pmap, va);
1724 pmap_unwire_pte_hold(pmap, mpte, &info);
1729 * Enter on the PV list if part of our managed memory. Note that we
1730 * raise IPL while manipulating pv_table since pmap_enter can be
1731 * called at interrupt time.
1733 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
1734 pmap_insert_entry(pmap, va, mpte, m);
1737 * Increment counters
1739 pmap->pm_stats.resident_count++;
1741 pa = VM_PAGE_TO_PHYS(m);
1744 * Now validate mapping with RO protection
1746 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
1747 *pte = pa | VPTE_V | VPTE_U;
1749 *pte = pa | VPTE_V | VPTE_U | VPTE_MANAGED;
1755 pmap_extract(pmap_t pmap, vm_offset_t va)
1760 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
1761 if (pte & VPTE_PS) {
1762 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
1763 rtval |= va & SEG_MASK;
1765 pte = *(get_ptbase(pmap) + (va >> PAGE_SHIFT));
1766 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
1773 #define MAX_INIT_PT (96)
1776 * This routine preloads the ptes for a given object into the specified pmap.
1777 * This eliminates the blast of soft faults on process startup and
1778 * immediately after an mmap.
1780 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
1783 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
1784 vm_object_t object, vm_pindex_t pindex,
1785 vm_size_t size, int limit)
1787 struct rb_vm_page_scan_info info;
1791 * We can't preinit if read access isn't set or there is no pmap
1794 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
1798 * We can't preinit if the pmap is not the current pmap
1800 if (curproc == NULL || pmap != vmspace_pmap(curproc->p_vmspace))
1803 psize = size >> PAGE_SHIFT;
1805 if ((object->type != OBJT_VNODE) ||
1806 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
1807 (object->resident_page_count > MAX_INIT_PT))) {
1811 if (psize + pindex > object->size) {
1812 if (object->size < pindex)
1814 psize = object->size - pindex;
1821 * Use a red-black scan to traverse the requested range and load
1822 * any valid pages found into the pmap.
1824 * We cannot safely scan the object's memq unless we are in a
1825 * critical section since interrupts can remove pages from objects.
1827 info.start_pindex = pindex;
1828 info.end_pindex = pindex + psize - 1;
1835 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1836 pmap_object_init_pt_callback, &info);
1842 pmap_object_init_pt_callback(vm_page_t p, void *data)
1844 struct rb_vm_page_scan_info *info = data;
1845 vm_pindex_t rel_index;
1847 * don't allow an madvise to blow away our really
1848 * free pages allocating pv entries.
1850 if ((info->limit & MAP_PREFAULT_MADVISE) &&
1851 vmstats.v_free_count < vmstats.v_free_reserved) {
1854 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
1855 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
1856 if ((p->queue - p->pc) == PQ_CACHE)
1857 vm_page_deactivate(p);
1859 rel_index = p->pindex - info->start_pindex;
1860 info->mpte = pmap_enter_quick(info->pmap,
1861 info->addr + i386_ptob(rel_index),
1863 vm_page_flag_set(p, PG_MAPPED);
1870 * pmap_prefault provides a quick way of clustering pagefaults into a
1871 * processes address space. It is a "cousin" of pmap_object_init_pt,
1872 * except it runs at page fault time instead of mmap time.
1876 #define PAGEORDER_SIZE (PFBAK+PFFOR)
1878 static int pmap_prefault_pageorder[] = {
1879 -PAGE_SIZE, PAGE_SIZE,
1880 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
1881 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
1882 -4 * PAGE_SIZE, 4 * PAGE_SIZE
1886 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
1896 * We do not currently prefault mappings that use virtual page
1897 * tables. We do not prefault foreign pmaps.
1899 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
1901 if (curproc == NULL || (pmap != vmspace_pmap(curproc->p_vmspace)))
1904 object = entry->object.vm_object;
1906 starta = addra - PFBAK * PAGE_SIZE;
1907 if (starta < entry->start)
1908 starta = entry->start;
1909 else if (starta > addra)
1913 * critical section protection is required to maintain the
1914 * page/object association, interrupts can free pages and remove
1915 * them from their objects.
1919 for (i = 0; i < PAGEORDER_SIZE; i++) {
1920 vm_object_t lobject;
1923 addr = addra + pmap_prefault_pageorder[i];
1924 if (addr > addra + (PFFOR * PAGE_SIZE))
1927 if (addr < starta || addr >= entry->end)
1931 * Make sure the page table page already exists
1933 if ((*pmap_pde(pmap, addr)) == NULL)
1937 * Get a pointer to the pte and make sure that no valid page
1940 pte = get_ptbase(pmap) + (addr >> PAGE_SHIFT);
1945 * Get the page to be mapped
1947 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1950 for (m = vm_page_lookup(lobject, pindex);
1951 (!m && (lobject->type == OBJT_DEFAULT) &&
1952 (lobject->backing_object));
1953 lobject = lobject->backing_object
1955 if (lobject->backing_object_offset & PAGE_MASK)
1957 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
1958 m = vm_page_lookup(lobject->backing_object, pindex);
1962 * give-up when a page is not in memory
1968 * If everything meets the requirements for pmap_enter_quick(),
1969 * then enter the page.
1972 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
1974 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
1976 if ((m->queue - m->pc) == PQ_CACHE) {
1977 vm_page_deactivate(m);
1980 mpte = pmap_enter_quick(pmap, addr, m, mpte);
1981 vm_page_flag_set(m, PG_MAPPED);
1989 * Routine: pmap_change_wiring
1990 * Function: Change the wiring attribute for a map/virtual-address
1992 * In/out conditions:
1993 * The mapping must already exist in the pmap.
1996 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2003 pte = get_ptbase(pmap) + (va >> PAGE_SHIFT);
2005 if (wired && (*pte & VPTE_W) == 0)
2006 pmap->pm_stats.wired_count++;
2007 else if (!wired && (*pte & VPTE_W))
2008 pmap->pm_stats.wired_count--;
2011 * Wiring is not a hardware characteristic so there is no need to
2012 * invalidate TLB. However, in an SMP environment we must use
2013 * a locked bus cycle to update the pte (if we are not using
2014 * the pmap_inval_*() API that is)... it's ok to do this for simple
2019 atomic_set_int(pte, VPTE_W);
2021 atomic_clear_int(pte, VPTE_W);
2024 atomic_set_int_nonlocked(pte, VPTE_W);
2026 atomic_clear_int_nonlocked(pte, VPTE_W);
2031 * Copy the range specified by src_addr/len
2032 * from the source map to the range dst_addr/len
2033 * in the destination map.
2035 * This routine is only advisory and need not do anything.
2038 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2039 vm_size_t len, vm_offset_t src_addr)
2041 pmap_inval_info info;
2043 vm_offset_t end_addr = src_addr + len;
2049 if (dst_addr != src_addr)
2051 if (dst_pmap->pm_pdir == NULL)
2053 if (src_pmap->pm_pdir == NULL)
2056 src_frame = get_ptbase1(src_pmap);
2057 dst_frame = get_ptbase2(dst_pmap);
2059 pmap_inval_init(&info);
2060 pmap_inval_add(&info, dst_pmap, -1);
2061 pmap_inval_add(&info, src_pmap, -1);
2064 * critical section protection is required to maintain the page/object
2065 * association, interrupts can free pages and remove them from
2069 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2070 vpte_t *src_pte, *dst_pte;
2071 vm_page_t dstmpte, srcmpte;
2072 vm_offset_t srcptepaddr;
2075 if (addr >= VM_MAX_USER_ADDRESS)
2076 panic("pmap_copy: invalid to pmap_copy page tables\n");
2079 * Don't let optional prefaulting of pages make us go
2080 * way below the low water mark of free pages or way
2081 * above high water mark of used pv entries.
2083 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2084 pv_entry_count > pv_entry_high_water)
2087 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2088 ptepindex = addr >> PDRSHIFT;
2090 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2091 if (srcptepaddr == 0)
2094 if (srcptepaddr & VPTE_PS) {
2095 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2096 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2097 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2102 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2103 if ((srcmpte == NULL) ||
2104 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2107 if (pdnxt > end_addr)
2110 src_pte = src_frame + (addr >> PAGE_SHIFT);
2111 dst_pte = dst_frame + (addr >> PAGE_SHIFT);
2112 while (addr < pdnxt) {
2116 * we only virtual copy managed pages
2118 if ((ptetemp & PG_MANAGED) != 0) {
2120 * We have to check after allocpte for the
2121 * pte still being around... allocpte can
2124 dstmpte = pmap_allocpte(dst_pmap, addr);
2125 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2127 * Clear the modified and
2128 * accessed (referenced) bits
2131 m = PHYS_TO_VM_PAGE(ptetemp);
2132 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2133 dst_pmap->pm_stats.resident_count++;
2134 pmap_insert_entry(dst_pmap, addr,
2137 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2139 if (dstmpte->hold_count >= srcmpte->hold_count)
2148 pmap_inval_flush(&info);
2154 * Zero the specified PA by mapping the page into KVM and clearing its
2157 * This function may be called from an interrupt and no locking is
2161 pmap_zero_page(vm_paddr_t phys)
2163 struct mdglobaldata *gd = mdcpu;
2166 if (*(int *)gd->gd_CMAP3)
2167 panic("pmap_zero_page: CMAP3 busy");
2168 *(int *)gd->gd_CMAP3 =
2169 VPTE_V | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2170 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2172 bzero(gd->gd_CADDR3, PAGE_SIZE);
2173 *(int *) gd->gd_CMAP3 = 0;
2178 * pmap_page_assertzero:
2180 * Assert that a page is empty, panic if it isn't.
2183 pmap_page_assertzero(vm_paddr_t phys)
2185 struct mdglobaldata *gd = mdcpu;
2189 if (*(int *)gd->gd_CMAP3)
2190 panic("pmap_zero_page: CMAP3 busy");
2191 *(int *)gd->gd_CMAP3 =
2192 VPTE_V | VPTE_R | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2193 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2194 for (i = 0; i < PAGE_SIZE; i += 4) {
2195 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2196 panic("pmap_page_assertzero() @ %p not zero!\n",
2197 (void *)gd->gd_CADDR3);
2200 *(int *) gd->gd_CMAP3 = 0;
2207 * Zero part of a physical page by mapping it into memory and clearing
2208 * its contents with bzero.
2210 * off and size may not cover an area beyond a single hardware page.
2213 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2215 struct mdglobaldata *gd = mdcpu;
2218 if (*(int *) gd->gd_CMAP3)
2219 panic("pmap_zero_page: CMAP3 busy");
2220 *(int *) gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2221 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2223 bzero((char *)gd->gd_CADDR3 + off, size);
2224 *(int *) gd->gd_CMAP3 = 0;
2231 * Copy the physical page from the source PA to the target PA.
2232 * This function may be called from an interrupt. No locking
2236 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2238 struct mdglobaldata *gd = mdcpu;
2241 if (*(int *) gd->gd_CMAP1)
2242 panic("pmap_copy_page: CMAP1 busy");
2243 if (*(int *) gd->gd_CMAP2)
2244 panic("pmap_copy_page: CMAP2 busy");
2246 *(int *) gd->gd_CMAP1 = VPTE_V | (src & PG_FRAME) | PG_A;
2247 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2249 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2250 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2252 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2254 *(int *) gd->gd_CMAP1 = 0;
2255 *(int *) gd->gd_CMAP2 = 0;
2260 * pmap_copy_page_frag:
2262 * Copy the physical page from the source PA to the target PA.
2263 * This function may be called from an interrupt. No locking
2267 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2269 struct mdglobaldata *gd = mdcpu;
2272 if (*(int *) gd->gd_CMAP1)
2273 panic("pmap_copy_page: CMAP1 busy");
2274 if (*(int *) gd->gd_CMAP2)
2275 panic("pmap_copy_page: CMAP2 busy");
2277 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2278 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2280 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2281 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2283 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2284 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2287 *(int *) gd->gd_CMAP1 = 0;
2288 *(int *) gd->gd_CMAP2 = 0;
2293 * Returns true if the pmap's pv is one of the first
2294 * 16 pvs linked to from this page. This count may
2295 * be changed upwards or downwards in the future; it
2296 * is only necessary that true be returned for a small
2297 * subset of pmaps for proper page aging.
2300 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2305 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2310 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2311 if (pv->pv_pmap == pmap) {
2324 * Remove all pages from specified address space
2325 * this aids process exit speeds. Also, this code
2326 * is special cased for current process only, but
2327 * can have the more generic (and slightly slower)
2328 * mode enabled. This is much faster than pmap_remove
2329 * in the case of running down an entire address space.
2332 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2337 pmap_inval_info info;
2340 if (curproc && pmap == vmspace_pmap(curproc->p_vmspace))
2345 pmap_inval_init(&info);
2347 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2348 if (pv->pv_va >= eva || pv->pv_va < sva) {
2349 npv = TAILQ_NEXT(pv, pv_plist);
2353 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2354 if (pmap->pm_active)
2355 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2359 * We cannot remove wired pages from a process' mapping
2362 if (tpte & VPTE_W) {
2363 npv = TAILQ_NEXT(pv, pv_plist);
2368 m = PHYS_TO_VM_PAGE(tpte);
2370 KASSERT(m < &vm_page_array[vm_page_array_size],
2371 ("pmap_remove_pages: bad tpte %x", tpte));
2373 pv->pv_pmap->pm_stats.resident_count--;
2376 * Update the vm_page_t clean and reference bits.
2378 if (tpte & VPTE_M) {
2383 npv = TAILQ_NEXT(pv, pv_plist);
2384 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2386 m->md.pv_list_count--;
2387 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2388 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2389 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2392 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2395 pmap_inval_flush(&info);
2400 * pmap_testbit tests bits in pte's
2401 * note that the testbit/changebit routines are inline,
2402 * and a lot of things compile-time evaluate.
2405 pmap_testbit(vm_page_t m, int bit)
2410 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2413 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2418 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2420 * if the bit being tested is the modified bit, then
2421 * mark clean_map and ptes as never
2424 if (bit & (VPTE_A|VPTE_M)) {
2425 if (!pmap_track_modified(pv->pv_va))
2429 #if defined(PMAP_DIAGNOSTIC)
2431 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2435 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2446 * this routine is used to modify bits in ptes
2448 static __inline void
2449 pmap_changebit(vm_page_t m, int bit, boolean_t setem)
2451 struct pmap_inval_info info;
2455 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2458 pmap_inval_init(&info);
2462 * Loop over all current mappings setting/clearing as appropos If
2463 * setting RO do we need to clear the VAC?
2465 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2467 * don't write protect pager mappings
2469 if (!setem && (bit == VPTE_W)) {
2470 if (!pmap_track_modified(pv->pv_va))
2474 #if defined(PMAP_DIAGNOSTIC)
2476 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2482 * Careful here. We can use a locked bus instruction to
2483 * clear VPTE_A or VPTE_M safely but we need to synchronize
2484 * with the target cpus when we mess with VPTE_W.
2486 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2488 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2492 atomic_set_int(pte, bit);
2494 atomic_set_int_nonlocked(pte, bit);
2497 vm_offset_t pbits = *(vm_offset_t *)pte;
2499 if (bit == VPTE_W) {
2500 if (pbits & VPTE_M) {
2504 atomic_clear_int(pte, VPTE_M|VPTE_W);
2506 atomic_clear_int_nonlocked(pte, VPTE_M|VPTE_W);
2510 atomic_clear_int(pte, bit);
2512 atomic_clear_int_nonlocked(pte, bit);
2518 pmap_inval_flush(&info);
2523 * pmap_page_protect:
2525 * Lower the permission for all mappings to a given page.
2528 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2530 if ((prot & VM_PROT_WRITE) == 0) {
2531 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2532 pmap_changebit(m, VPTE_W, FALSE);
2540 pmap_phys_address(int ppn)
2542 return (i386_ptob(ppn));
2546 * pmap_ts_referenced:
2548 * Return a count of reference bits for a page, clearing those bits.
2549 * It is not necessary for every reference bit to be cleared, but it
2550 * is necessary that 0 only be returned when there are truly no
2551 * reference bits set.
2553 * XXX: The exact number of bits to check and clear is a matter that
2554 * should be tested and standardized at some point in the future for
2555 * optimal aging of shared pages.
2558 pmap_ts_referenced(vm_page_t m)
2560 pv_entry_t pv, pvf, pvn;
2564 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2569 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2574 pvn = TAILQ_NEXT(pv, pv_list);
2576 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2578 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2580 if (!pmap_track_modified(pv->pv_va))
2583 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2585 if (pte && (*pte & VPTE_A)) {
2587 atomic_clear_int(pte, VPTE_A);
2589 atomic_clear_int_nonlocked(pte, VPTE_A);
2596 } while ((pv = pvn) != NULL && pv != pvf);
2606 * Return whether or not the specified physical page was modified
2607 * in any physical maps.
2610 pmap_is_modified(vm_page_t m)
2612 return pmap_testbit(m, VPTE_M);
2616 * Clear the modify bits on the specified physical page.
2619 pmap_clear_modify(vm_page_t m)
2621 pmap_changebit(m, VPTE_M, FALSE);
2625 * pmap_clear_reference:
2627 * Clear the reference bit on the specified physical page.
2630 pmap_clear_reference(vm_page_t m)
2632 pmap_changebit(m, VPTE_A, FALSE);
2636 * Miscellaneous support routines follow
2640 i386_protection_init(void)
2644 kp = protection_codes;
2645 for (prot = 0; prot < 8; prot++) {
2646 if (prot & VM_PROT_READ)
2648 if (prot & VM_PROT_WRITE)
2650 if (prot & VM_PROT_EXECUTE)
2657 * Map a set of physical memory pages into the kernel virtual
2658 * address space. Return a pointer to where it is mapped. This
2659 * routine is intended to be used for mapping device memory,
2662 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2666 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2668 vm_offset_t va, tmpva, offset;
2671 offset = pa & PAGE_MASK;
2672 size = roundup(offset + size, PAGE_SIZE);
2674 va = kmem_alloc_nofault(&kernel_map, size);
2676 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2678 pa = pa & VPTE_FRAME;
2679 for (tmpva = va; size > 0;) {
2680 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2681 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2689 return ((void *)(va + offset));
2693 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2695 vm_offset_t base, offset;
2697 base = va & VPTE_FRAME;
2698 offset = va & PAGE_MASK;
2699 size = roundup(offset + size, PAGE_SIZE);
2700 pmap_qremove(va, size >> PAGE_SHIFT);
2701 kmem_free(&kernel_map, base, size);
2705 * perform the pmap work for mincore
2708 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2714 ptep = pmap_pte(pmap, addr);
2719 if ((pte = *ptep) != 0) {
2722 val = MINCORE_INCORE;
2723 if ((pte & VPTE_MANAGED) == 0)
2726 pa = pte & VPTE_FRAME;
2728 m = PHYS_TO_VM_PAGE(pa);
2734 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2736 * Modified by someone
2738 else if (m->dirty || pmap_is_modified(m))
2739 val |= MINCORE_MODIFIED_OTHER;
2744 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2747 * Referenced by someone
2749 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
2750 val |= MINCORE_REFERENCED_OTHER;
2751 vm_page_flag_set(m, PG_REFERENCED);
2758 pmap_activate(struct proc *p)
2762 pmap = vmspace_pmap(p->p_vmspace);
2764 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2766 pmap->pm_active |= 1;
2768 #if defined(SWTCH_OPTIM_STATS)
2771 panic("pmap_activate"); /* XXX store vmspace id in context */
2773 p->p_thread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
2774 load_cr3(p->p_thread->td_pcb->pcb_cr3);
2779 pmap_deactivate(struct proc *p)
2783 pmap = vmspace_pmap(p->p_vmspace);
2785 atomic_clear_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
2787 pmap->pm_active &= ~1;
2790 * XXX - note we do not adjust %cr3. The caller is expected to
2791 * activate a new pmap or do a thread-exit.
2796 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
2799 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
2803 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
2810 static void pads (pmap_t pm);
2811 void pmap_pvdump (vm_paddr_t pa);
2813 /* print address space of pmap*/
2821 if (pm == &kernel_pmap)
2823 for (i = 0; i < 1024; i++)
2825 for (j = 0; j < 1024; j++) {
2826 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
2827 if (pm == &kernel_pmap && va < KERNBASE)
2829 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
2831 ptep = pmap_pte(pm, va);
2832 if (ptep && (*ptep & VPTE_V)) {
2834 (void *)va, (unsigned)*ptep);
2841 pmap_pvdump(vm_paddr_t pa)
2846 kprintf("pa %08llx", (long long)pa);
2847 m = PHYS_TO_VM_PAGE(pa);
2848 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2850 kprintf(" -> pmap %p, va %x, flags %x",
2851 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
2853 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);