2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
9 * This code is derived from software contributed to Berkeley by
10 * the Systems Programming Group of the University of Utah Computer
11 * Science Department and William Jolitz of UUNET Technologies Inc.
13 * Redistribution and use in source and binary forms, with or without
14 * 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 the
20 * documentation and/or other materials provided with the distribution.
21 * 3. All advertising materials mentioning features or use of this software
22 * must display the following acknowledgement:
23 * This product includes software developed by the University of
24 * California, Berkeley and its contributors.
25 * 4. Neither the name of the University nor the names of its contributors
26 * may be used to endorse or promote products derived from this software
27 * without specific prior written permission.
29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
42 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
43 * $DragonFly: src/sys/platform/pc32/i386/pmap.c,v 1.59 2006/10/20 17:02:19 dillon Exp $
47 * Manages physical address maps.
49 * In addition to hardware address maps, this
50 * module is called upon to provide software-use-only
51 * maps which may or may not be stored in the same
52 * form as hardware maps. These pseudo-maps are
53 * used to store intermediate results from copy
54 * operations to and from address spaces.
56 * Since the information managed by this module is
57 * also stored by the logical address mapping module,
58 * this module may throw away valid virtual-to-physical
59 * mappings at almost any time. However, invalidations
60 * of virtual-to-physical mappings must be done as
63 * In order to cope with hardware architectures which
64 * make virtual-to-physical map invalidates expensive,
65 * this module may delay invalidate or reduced protection
66 * operations until such time as they are actually
67 * necessary. This module is given full information as
68 * to which processors are currently using which maps,
69 * and to when physical maps must be made correct.
72 #include "opt_disable_pse.h"
74 #include "opt_msgbuf.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
80 #include <sys/msgbuf.h>
81 #include <sys/vmmeter.h>
85 #include <vm/vm_param.h>
86 #include <sys/sysctl.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_extern.h>
93 #include <vm/vm_pageout.h>
94 #include <vm/vm_pager.h>
95 #include <vm/vm_zone.h>
98 #include <sys/thread2.h>
100 #include <machine/cputypes.h>
101 #include <machine/md_var.h>
102 #include <machine/specialreg.h>
103 #include <machine/smp.h>
104 #include <arch/apic/apicreg.h>
105 #include <machine/globaldata.h>
106 #include <machine/pmap.h>
107 #include <machine/pmap_inval.h>
109 #define PMAP_KEEP_PDIRS
110 #ifndef PMAP_SHPGPERPROC
111 #define PMAP_SHPGPERPROC 200
114 #if defined(DIAGNOSTIC)
115 #define PMAP_DIAGNOSTIC
120 #if !defined(PMAP_DIAGNOSTIC)
121 #define PMAP_INLINE __inline
127 * Get PDEs and PTEs for user/kernel address space
129 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
130 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
132 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
133 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
134 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
135 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
136 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
140 * Given a map and a machine independent protection code,
141 * convert to a vax protection code.
143 #define pte_prot(m, p) \
144 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
145 static int protection_codes[8];
147 static struct pmap kernel_pmap_store;
150 vm_paddr_t avail_start; /* PA of first available physical page */
151 vm_paddr_t avail_end; /* PA of last available physical page */
152 vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */
153 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
154 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
155 static int pgeflag; /* PG_G or-in */
156 static int pseflag; /* PG_PS or-in */
158 static vm_object_t kptobj;
161 vm_offset_t kernel_vm_end;
164 * Data for the pv entry allocation mechanism
166 static vm_zone_t pvzone;
167 static struct vm_zone pvzone_store;
168 static struct vm_object pvzone_obj;
169 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
170 static int pmap_pagedaemon_waken = 0;
171 static struct pv_entry *pvinit;
174 * All those kernel PT submaps that BSD is so fond of
176 pt_entry_t *CMAP1 = 0, *ptmmap;
177 caddr_t CADDR1 = 0, ptvmmap = 0;
178 static pt_entry_t *msgbufmap;
179 struct msgbuf *msgbufp=0;
184 static pt_entry_t *pt_crashdumpmap;
185 static caddr_t crashdumpmap;
187 extern pt_entry_t *SMPpt;
189 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
190 static unsigned * get_ptbase (pmap_t pmap);
191 static pv_entry_t get_pv_entry (void);
192 static void i386_protection_init (void);
193 static __inline void pmap_changebit (vm_page_t m, int bit, boolean_t setem);
195 static void pmap_remove_all (vm_page_t m);
196 static vm_page_t pmap_enter_quick (pmap_t pmap, vm_offset_t va,
197 vm_page_t m, vm_page_t mpte);
198 static int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
199 vm_offset_t sva, pmap_inval_info_t info);
200 static void pmap_remove_page (struct pmap *pmap,
201 vm_offset_t va, pmap_inval_info_t info);
202 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
203 vm_offset_t va, pmap_inval_info_t info);
204 static boolean_t pmap_testbit (vm_page_t m, int bit);
205 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
206 vm_page_t mpte, vm_page_t m);
208 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
210 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
211 static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
212 static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
213 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
214 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
215 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
217 static unsigned pdir4mb;
220 * Move the kernel virtual free pointer to the next
221 * 4MB. This is used to help improve performance
222 * by using a large (4MB) page for much of the kernel
223 * (.text, .data, .bss)
226 pmap_kmem_choose(vm_offset_t addr)
228 vm_offset_t newaddr = addr;
230 if (cpu_feature & CPUID_PSE) {
231 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
240 * Extract the page table entry associated with the given map/virtual
243 * This function may NOT be called from an interrupt.
245 PMAP_INLINE unsigned *
246 pmap_pte(pmap_t pmap, vm_offset_t va)
251 pdeaddr = (unsigned *) pmap_pde(pmap, va);
252 if (*pdeaddr & PG_PS)
255 return get_ptbase(pmap) + i386_btop(va);
264 * Super fast pmap_pte routine best used when scanning the pv lists.
265 * This eliminates many course-grained invltlb calls. Note that many of
266 * the pv list scans are across different pmaps and it is very wasteful
267 * to do an entire invltlb when checking a single mapping.
269 * Should only be called while in a critical section.
272 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
274 struct mdglobaldata *gd = mdcpu;
277 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
278 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
279 unsigned index = i386_btop(va);
280 /* are we current address space or kernel? */
281 if ((pmap == kernel_pmap) ||
282 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
283 return (unsigned *) PTmap + index;
285 newpf = pde & PG_FRAME;
286 if ( ((* (unsigned *) gd->gd_PMAP1) & PG_FRAME) != newpf) {
287 * (unsigned *) gd->gd_PMAP1 = newpf | PG_RW | PG_V;
288 cpu_invlpg(gd->gd_PADDR1);
290 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
297 * Bootstrap the system enough to run with virtual memory.
299 * On the i386 this is called after mapping has already been enabled
300 * and just syncs the pmap module with what has already been done.
301 * [We can't call it easily with mapping off since the kernel is not
302 * mapped with PA == VA, hence we would have to relocate every address
303 * from the linked base (virtual) address "KERNBASE" to the actual
304 * (physical) address starting relative to 0]
307 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
311 struct mdglobaldata *gd;
315 avail_start = firstaddr;
318 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too
319 * large. It should instead be correctly calculated in locore.s and
320 * not based on 'first' (which is a physical address, not a virtual
321 * address, for the start of unused physical memory). The kernel
322 * page tables are NOT double mapped and thus should not be included
323 * in this calculation.
325 virtual_avail = (vm_offset_t) KERNBASE + firstaddr;
326 virtual_avail = pmap_kmem_choose(virtual_avail);
328 virtual_end = VM_MAX_KERNEL_ADDRESS;
331 * Initialize protection array.
333 i386_protection_init();
336 * The kernel's pmap is statically allocated so we don't have to use
337 * pmap_create, which is unlikely to work correctly at this part of
338 * the boot sequence (XXX and which no longer exists).
340 kernel_pmap = &kernel_pmap_store;
342 kernel_pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
343 kernel_pmap->pm_count = 1;
344 kernel_pmap->pm_active = (cpumask_t)-1; /* don't allow deactivation */
345 TAILQ_INIT(&kernel_pmap->pm_pvlist);
349 * Reserve some special page table entries/VA space for temporary
352 #define SYSMAP(c, p, v, n) \
353 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
356 pte = (pt_entry_t *) pmap_pte(kernel_pmap, va);
359 * CMAP1/CMAP2 are used for zeroing and copying pages.
361 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
366 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
369 * ptvmmap is used for reading arbitrary physical pages via
372 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
375 * msgbufp is used to map the system message buffer.
376 * XXX msgbufmap is not used.
378 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
379 atop(round_page(MSGBUF_SIZE)))
384 for (i = 0; i < NKPT; i++)
388 * PG_G is terribly broken on SMP because we IPI invltlb's in some
389 * cases rather then invl1pg. Actually, I don't even know why it
390 * works under UP because self-referential page table mappings
395 if (cpu_feature & CPUID_PGE)
400 * Initialize the 4MB page size flag
404 * The 4MB page version of the initial
405 * kernel page mapping.
409 #if !defined(DISABLE_PSE)
410 if (cpu_feature & CPUID_PSE) {
413 * Note that we have enabled PSE mode
416 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
417 ptditmp &= ~(NBPDR - 1);
418 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
423 * Enable the PSE mode. If we are SMP we can't do this
424 * now because the APs will not be able to use it when
427 load_cr4(rcr4() | CR4_PSE);
430 * We can do the mapping here for the single processor
431 * case. We simply ignore the old page table page from
435 * For SMP, we still need 4K pages to bootstrap APs,
436 * PSE will be enabled as soon as all APs are up.
438 PTD[KPTDI] = (pd_entry_t)ptditmp;
439 kernel_pmap->pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
445 if (cpu_apic_address == 0)
446 panic("pmap_bootstrap: no local apic!");
448 /* local apic is mapped on last page */
449 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag |
450 (cpu_apic_address & PG_FRAME));
454 * We need to finish setting up the globaldata page for the BSP.
455 * locore has already populated the page table for the mdglobaldata
458 pg = MDGLOBALDATA_BASEALLOC_PAGES;
459 gd = &CPU_prvspace[0].mdglobaldata;
460 gd->gd_CMAP1 = &SMPpt[pg + 0];
461 gd->gd_CMAP2 = &SMPpt[pg + 1];
462 gd->gd_CMAP3 = &SMPpt[pg + 2];
463 gd->gd_PMAP1 = &SMPpt[pg + 3];
464 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
465 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
466 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
467 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
474 * Set 4mb pdir for mp startup
479 if (pseflag && (cpu_feature & CPUID_PSE)) {
480 load_cr4(rcr4() | CR4_PSE);
481 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
482 kernel_pmap->pm_pdir[KPTDI] =
483 PTD[KPTDI] = (pd_entry_t)pdir4mb;
491 * Initialize the pmap module.
492 * Called by vm_init, to initialize any structures that the pmap
493 * system needs to map virtual memory.
494 * pmap_init has been enhanced to support in a fairly consistant
495 * way, discontiguous physical memory.
504 * object for kernel page table pages
506 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
509 * Allocate memory for random pmap data structures. Includes the
513 for(i = 0; i < vm_page_array_size; i++) {
516 m = &vm_page_array[i];
517 TAILQ_INIT(&m->md.pv_list);
518 m->md.pv_list_count = 0;
522 * init the pv free list
524 initial_pvs = vm_page_array_size;
525 if (initial_pvs < MINPV)
527 pvzone = &pvzone_store;
528 pvinit = (struct pv_entry *) kmem_alloc(kernel_map,
529 initial_pvs * sizeof (struct pv_entry));
530 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
534 * Now it is safe to enable pv_table recording.
536 pmap_initialized = TRUE;
540 * Initialize the address space (zone) for the pv_entries. Set a
541 * high water mark so that the system can recover from excessive
542 * numbers of pv entries.
547 int shpgperproc = PMAP_SHPGPERPROC;
549 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
550 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
551 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
552 pv_entry_high_water = 9 * (pv_entry_max / 10);
553 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
557 /***************************************************
558 * Low level helper routines.....
559 ***************************************************/
561 #if defined(PMAP_DIAGNOSTIC)
564 * This code checks for non-writeable/modified pages.
565 * This should be an invalid condition.
568 pmap_nw_modified(pt_entry_t ptea)
574 if ((pte & (PG_M|PG_RW)) == PG_M)
583 * this routine defines the region(s) of memory that should
584 * not be tested for the modified bit.
586 static PMAP_INLINE int
587 pmap_track_modified(vm_offset_t va)
589 if ((va < clean_sva) || (va >= clean_eva))
596 get_ptbase(pmap_t pmap)
598 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
599 struct globaldata *gd = mycpu;
601 /* are we current address space or kernel? */
602 if (pmap == kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
603 return (unsigned *) PTmap;
606 /* otherwise, we are alternate address space */
607 KKASSERT(gd->gd_intr_nesting_level == 0 && (gd->gd_curthread->td_flags & TDF_INTTHREAD) == 0);
609 if (frame != (((unsigned) APTDpde) & PG_FRAME)) {
610 APTDpde = (pd_entry_t)(frame | PG_RW | PG_V);
611 /* The page directory is not shared between CPUs */
614 return (unsigned *) APTmap;
620 * Extract the physical page address associated with the map/VA pair.
622 * This function may not be called from an interrupt if the pmap is
626 pmap_extract(pmap_t pmap, vm_offset_t va)
629 vm_offset_t pdirindex;
631 pdirindex = va >> PDRSHIFT;
632 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
634 if ((rtval & PG_PS) != 0) {
635 rtval &= ~(NBPDR - 1);
636 rtval |= va & (NBPDR - 1);
639 pte = get_ptbase(pmap) + i386_btop(va);
640 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
647 * Extract user accessible page only, return NULL if the page is not
648 * present or if it's current state is not sufficient. Caller will
649 * generally call vm_fault() on failure and try again.
652 pmap_extract_vmpage(pmap_t pmap, vm_offset_t va, int prot)
655 vm_offset_t pdirindex;
657 pdirindex = va >> PDRSHIFT;
658 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
662 if ((rtval & PG_PS) != 0) {
663 if ((rtval & (PG_V|PG_U)) != (PG_V|PG_U))
665 if ((prot & VM_PROT_WRITE) && (rtval & PG_RW) == 0)
667 rtval &= ~(NBPDR - 1);
668 rtval |= va & (NBPDR - 1);
669 m = PHYS_TO_VM_PAGE(rtval);
671 pte = get_ptbase(pmap) + i386_btop(va);
672 if ((*pte & (PG_V|PG_U)) != (PG_V|PG_U))
674 if ((prot & VM_PROT_WRITE) && (*pte & PG_RW) == 0)
676 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
677 m = PHYS_TO_VM_PAGE(rtval);
684 /***************************************************
685 * Low level mapping routines.....
686 ***************************************************/
689 * Routine: pmap_kenter
691 * Add a wired page to the KVA
692 * NOTE! note that in order for the mapping to take effect -- you
693 * should do an invltlb after doing the pmap_kenter().
696 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
700 pmap_inval_info info;
702 pmap_inval_init(&info);
703 pmap_inval_add(&info, kernel_pmap, va);
704 npte = pa | PG_RW | PG_V | pgeflag;
705 pte = (unsigned *)vtopte(va);
707 pmap_inval_flush(&info);
711 * Routine: pmap_kenter_quick
713 * Similar to pmap_kenter(), except we only invalidate the
714 * mapping on the current CPU.
717 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
722 npte = pa | PG_RW | PG_V | pgeflag;
723 pte = (unsigned *)vtopte(va);
725 cpu_invlpg((void *)va);
729 pmap_kenter_sync(vm_offset_t va)
731 pmap_inval_info info;
733 pmap_inval_init(&info);
734 pmap_inval_add(&info, kernel_pmap, va);
735 pmap_inval_flush(&info);
739 pmap_kenter_sync_quick(vm_offset_t va)
741 cpu_invlpg((void *)va);
745 * remove a page from the kernel pagetables
748 pmap_kremove(vm_offset_t va)
751 pmap_inval_info info;
753 pmap_inval_init(&info);
754 pmap_inval_add(&info, kernel_pmap, va);
755 pte = (unsigned *)vtopte(va);
757 pmap_inval_flush(&info);
761 pmap_kremove_quick(vm_offset_t va)
764 pte = (unsigned *)vtopte(va);
766 cpu_invlpg((void *)va);
770 * Used to map a range of physical addresses into kernel
771 * virtual address space.
773 * For now, VM is already on, we only need to map the
777 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
779 while (start < end) {
780 pmap_kenter(virt, start);
789 * Add a list of wired pages to the kva
790 * this routine is only used for temporary
791 * kernel mappings that do not need to have
792 * page modification or references recorded.
793 * Note that old mappings are simply written
794 * over. The page *must* be wired.
797 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
801 end_va = va + count * PAGE_SIZE;
803 while (va < end_va) {
806 pte = (unsigned *)vtopte(va);
807 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
808 cpu_invlpg((void *)va);
813 smp_invltlb(); /* XXX */
818 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
821 cpumask_t cmask = mycpu->gd_cpumask;
823 end_va = va + count * PAGE_SIZE;
825 while (va < end_va) {
830 * Install the new PTE. If the pte changed from the prior
831 * mapping we must reset the cpu mask and invalidate the page.
832 * If the pte is the same but we have not seen it on the
833 * current cpu, invlpg the existing mapping. Otherwise the
834 * entry is optimal and no invalidation is required.
836 pte = (unsigned *)vtopte(va);
837 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
838 if (*pte != pteval) {
841 cpu_invlpg((void *)va);
842 } else if ((*mask & cmask) == 0) {
843 cpu_invlpg((void *)va);
852 * this routine jerks page mappings from the
853 * kernel -- it is meant only for temporary mappings.
856 pmap_qremove(vm_offset_t va, int count)
860 end_va = va + count*PAGE_SIZE;
862 while (va < end_va) {
865 pte = (unsigned *)vtopte(va);
867 cpu_invlpg((void *)va);
876 * This routine works like vm_page_lookup() but also blocks as long as the
877 * page is busy. This routine does not busy the page it returns.
879 * Unless the caller is managing objects whos pages are in a known state,
880 * the call should be made with a critical section held so the page's object
881 * association remains valid on return.
884 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
889 m = vm_page_lookup(object, pindex);
890 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
896 * Create a new thread and optionally associate it with a (new) process.
897 * NOTE! the new thread's cpu may not equal the current cpu.
900 pmap_init_thread(thread_t td)
902 /* enforce pcb placement */
903 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
904 td->td_savefpu = &td->td_pcb->pcb_save;
905 td->td_sp = (char *)td->td_pcb - 16;
909 * Create the UPAGES for a new process.
910 * This routine directly affects the fork perf for a process.
913 pmap_init_proc(struct proc *p, struct thread *td)
915 p->p_addr = (void *)td->td_kstack;
918 td->td_lwp = &p->p_lwp;
919 td->td_switch = cpu_heavy_switch;
921 KKASSERT(td->td_mpcount == 1);
923 bzero(p->p_addr, sizeof(*p->p_addr));
927 * Dispose the UPAGES for a process that has exited.
928 * This routine directly impacts the exit perf of a process.
931 pmap_dispose_proc(struct proc *p)
935 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
937 if ((td = p->p_thread) != NULL) {
945 /***************************************************
946 * Page table page management routines.....
947 ***************************************************/
950 * This routine unholds page table pages, and if the hold count
951 * drops to zero, then it decrements the wire count.
954 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
956 pmap_inval_flush(info);
957 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
960 if (m->hold_count == 0) {
962 * unmap the page table page
964 pmap_inval_add(info, pmap, -1);
965 pmap->pm_pdir[m->pindex] = 0;
966 --pmap->pm_stats.resident_count;
968 if (pmap->pm_ptphint == m)
969 pmap->pm_ptphint = NULL;
972 * If the page is finally unwired, simply free it.
975 if (m->wire_count == 0) {
978 vm_page_free_zero(m);
979 --vmstats.v_wire_count;
986 static PMAP_INLINE int
987 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
990 if (m->hold_count == 0)
991 return _pmap_unwire_pte_hold(pmap, m, info);
997 * After removing a page table entry, this routine is used to
998 * conditionally free the page, and manage the hold/wire counts.
1001 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1002 pmap_inval_info_t info)
1005 if (va >= UPT_MIN_ADDRESS)
1009 ptepindex = (va >> PDRSHIFT);
1010 if (pmap->pm_ptphint &&
1011 (pmap->pm_ptphint->pindex == ptepindex)) {
1012 mpte = pmap->pm_ptphint;
1014 pmap_inval_flush(info);
1015 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1016 pmap->pm_ptphint = mpte;
1020 return pmap_unwire_pte_hold(pmap, mpte, info);
1024 pmap_pinit0(struct pmap *pmap)
1027 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE);
1028 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1030 pmap->pm_active = 0;
1031 pmap->pm_ptphint = NULL;
1032 TAILQ_INIT(&pmap->pm_pvlist);
1033 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1037 * Initialize a preallocated and zeroed pmap structure,
1038 * such as one in a vmspace structure.
1041 pmap_pinit(struct pmap *pmap)
1046 * No need to allocate page table space yet but we do need a valid
1047 * page directory table.
1049 if (pmap->pm_pdir == NULL) {
1051 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE);
1055 * allocate object for the ptes
1057 if (pmap->pm_pteobj == NULL)
1058 pmap->pm_pteobj = vm_object_allocate( OBJT_DEFAULT, PTDPTDI + 1);
1061 * allocate the page directory page
1063 ptdpg = vm_page_grab( pmap->pm_pteobj, PTDPTDI,
1064 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1066 ptdpg->wire_count = 1;
1067 ++vmstats.v_wire_count;
1070 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY); /* not usually mapped*/
1071 ptdpg->valid = VM_PAGE_BITS_ALL;
1073 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1074 if ((ptdpg->flags & PG_ZERO) == 0)
1075 bzero(pmap->pm_pdir, PAGE_SIZE);
1077 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1079 /* install self-referential address mapping entry */
1080 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1081 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1084 pmap->pm_active = 0;
1085 pmap->pm_ptphint = NULL;
1086 TAILQ_INIT(&pmap->pm_pvlist);
1087 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1091 * Wire in kernel global address entries. To avoid a race condition
1092 * between pmap initialization and pmap_growkernel, this procedure
1093 * should be called after the vmspace is attached to the process
1094 * but before this pmap is activated.
1097 pmap_pinit2(struct pmap *pmap)
1099 /* XXX copies current process, does not fill in MPPTDI */
1100 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1104 * Attempt to release and free and vm_page in a pmap. Returns 1 on success,
1105 * 0 on failure (if the procedure had to sleep).
1108 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1110 unsigned *pde = (unsigned *) pmap->pm_pdir;
1112 * This code optimizes the case of freeing non-busy
1113 * page-table pages. Those pages are zero now, and
1114 * might as well be placed directly into the zero queue.
1116 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1122 * Remove the page table page from the processes address space.
1125 pmap->pm_stats.resident_count--;
1127 if (p->hold_count) {
1128 panic("pmap_release: freeing held page table page");
1131 * Page directory pages need to have the kernel
1132 * stuff cleared, so they can go into the zero queue also.
1134 if (p->pindex == PTDPTDI) {
1135 bzero(pde + KPTDI, nkpt * PTESIZE);
1138 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1141 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1142 pmap->pm_ptphint = NULL;
1145 vmstats.v_wire_count--;
1146 vm_page_free_zero(p);
1151 * this routine is called if the page table page is not
1155 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1157 vm_offset_t pteva, ptepa;
1161 * Find or fabricate a new pagetable page
1163 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1164 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1166 KASSERT(m->queue == PQ_NONE,
1167 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1169 if (m->wire_count == 0)
1170 vmstats.v_wire_count++;
1174 * Increment the hold count for the page table page
1175 * (denoting a new mapping.)
1180 * Map the pagetable page into the process address space, if
1181 * it isn't already there.
1184 pmap->pm_stats.resident_count++;
1186 ptepa = VM_PAGE_TO_PHYS(m);
1187 pmap->pm_pdir[ptepindex] =
1188 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1191 * Set the page table hint
1193 pmap->pm_ptphint = m;
1196 * Try to use the new mapping, but if we cannot, then
1197 * do it with the routine that maps the page explicitly.
1199 if ((m->flags & PG_ZERO) == 0) {
1200 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1201 (((unsigned) PTDpde) & PG_FRAME)) {
1202 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1203 bzero((caddr_t) pteva, PAGE_SIZE);
1205 pmap_zero_page(ptepa);
1209 m->valid = VM_PAGE_BITS_ALL;
1210 vm_page_flag_clear(m, PG_ZERO);
1211 vm_page_flag_set(m, PG_MAPPED);
1218 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1225 * Calculate pagetable page index
1227 ptepindex = va >> PDRSHIFT;
1230 * Get the page directory entry
1232 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1235 * This supports switching from a 4MB page to a
1238 if (ptepa & PG_PS) {
1239 pmap->pm_pdir[ptepindex] = 0;
1246 * If the page table page is mapped, we just increment the
1247 * hold count, and activate it.
1251 * In order to get the page table page, try the
1254 if (pmap->pm_ptphint &&
1255 (pmap->pm_ptphint->pindex == ptepindex)) {
1256 m = pmap->pm_ptphint;
1258 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1259 pmap->pm_ptphint = m;
1265 * Here if the pte page isn't mapped, or if it has been deallocated.
1267 return _pmap_allocpte(pmap, ptepindex);
1271 /***************************************************
1272 * Pmap allocation/deallocation routines.
1273 ***************************************************/
1276 * Release any resources held by the given physical map.
1277 * Called when a pmap initialized by pmap_pinit is being released.
1278 * Should only be called if the map contains no valid mappings.
1281 pmap_release(struct pmap *pmap)
1283 vm_page_t p,n,ptdpg;
1284 vm_object_t object = pmap->pm_pteobj;
1287 #if defined(DIAGNOSTIC)
1288 if (object->ref_count != 1)
1289 panic("pmap_release: pteobj reference count != 1");
1295 curgeneration = object->generation;
1296 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = n) {
1297 n = TAILQ_NEXT(p, listq);
1298 if (p->pindex == PTDPTDI) {
1302 if (!pmap_release_free_page(pmap, p)) {
1306 if (object->generation != curgeneration) {
1311 if (ptdpg && !pmap_release_free_page(pmap, ptdpg)) {
1319 kvm_size(SYSCTL_HANDLER_ARGS)
1321 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE;
1323 return sysctl_handle_long(oidp, &ksize, 0, req);
1325 SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD,
1326 0, 0, kvm_size, "IU", "Size of KVM");
1329 kvm_free(SYSCTL_HANDLER_ARGS)
1331 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end;
1333 return sysctl_handle_long(oidp, &kfree, 0, req);
1335 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD,
1336 0, 0, kvm_free, "IU", "Amount of KVM free");
1339 * Grow the number of kernel page table entries, if needed.
1341 struct pmap_growkernel_info {
1345 static int pmap_growkernel_callback(struct proc *p, void *data);
1348 pmap_growkernel(vm_offset_t addr)
1350 struct pmap_growkernel_info info;
1351 vm_offset_t ptppaddr;
1356 if (kernel_vm_end == 0) {
1357 kernel_vm_end = KERNBASE;
1359 while (pdir_pde(PTD, kernel_vm_end)) {
1360 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1364 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1365 while (kernel_vm_end < addr) {
1366 if (pdir_pde(PTD, kernel_vm_end)) {
1367 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1372 * This index is bogus, but out of the way
1374 nkpg = vm_page_alloc(kptobj, nkpt,
1375 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1377 panic("pmap_growkernel: no memory to grow kernel");
1380 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1381 pmap_zero_page(ptppaddr);
1382 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1383 pdir_pde(PTD, kernel_vm_end) = newpdir;
1384 *pmap_pde(kernel_pmap, kernel_vm_end) = newpdir;
1388 * vm_fork and friends copy nkpt page table pages to the high
1389 * side of a new process's pmap. This occurs after the
1390 * process has been added to allproc, so scanning the proc
1391 * list afterwords should be sufficient to fixup existing
1394 info.newpdir = newpdir;
1395 allproc_scan(pmap_growkernel_callback, &info);
1396 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1402 pmap_growkernel_callback(struct proc *p, void *data)
1404 struct pmap_growkernel_info *info = data;
1408 pmap = vmspace_pmap(p->p_vmspace);
1409 *pmap_pde(pmap, kernel_vm_end) = info->newpdir;
1415 * Retire the given physical map from service.
1416 * Should only be called if the map contains
1417 * no valid mappings.
1420 pmap_destroy(pmap_t pmap)
1427 count = --pmap->pm_count;
1430 panic("destroying a pmap is not yet implemented");
1435 * Add a reference to the specified pmap.
1438 pmap_reference(pmap_t pmap)
1445 /***************************************************
1446 * page management routines.
1447 ***************************************************/
1450 * free the pv_entry back to the free list. This function may be
1451 * called from an interrupt.
1453 static PMAP_INLINE void
1454 free_pv_entry(pv_entry_t pv)
1461 * get a new pv_entry, allocating a block from the system
1462 * when needed. This function may be called from an interrupt.
1468 if (pv_entry_high_water &&
1469 (pv_entry_count > pv_entry_high_water) &&
1470 (pmap_pagedaemon_waken == 0)) {
1471 pmap_pagedaemon_waken = 1;
1472 wakeup (&vm_pages_needed);
1474 return zalloc(pvzone);
1478 * This routine is very drastic, but can save the system
1486 static int warningdone=0;
1488 if (pmap_pagedaemon_waken == 0)
1491 if (warningdone < 5) {
1492 printf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1496 for(i = 0; i < vm_page_array_size; i++) {
1497 m = &vm_page_array[i];
1498 if (m->wire_count || m->hold_count || m->busy ||
1499 (m->flags & PG_BUSY))
1503 pmap_pagedaemon_waken = 0;
1508 * If it is the first entry on the list, it is actually
1509 * in the header and we must copy the following entry up
1510 * to the header. Otherwise we must search the list for
1511 * the entry. In either case we free the now unused entry.
1514 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1515 vm_offset_t va, pmap_inval_info_t info)
1521 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1522 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1523 if (pmap == pv->pv_pmap && va == pv->pv_va)
1527 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1528 if (va == pv->pv_va)
1535 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1536 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1537 m->md.pv_list_count--;
1538 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1539 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1540 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1548 * Create a pv entry for page at pa for
1552 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1557 pv = get_pv_entry();
1562 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1563 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1564 m->md.pv_list_count++;
1570 * pmap_remove_pte: do the things to unmap a page in a process
1573 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1574 pmap_inval_info_t info)
1579 pmap_inval_add(info, pmap, va);
1580 oldpte = loadandclear(ptq);
1582 pmap->pm_stats.wired_count -= 1;
1584 * Machines that don't support invlpg, also don't support
1585 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1589 cpu_invlpg((void *)va);
1590 pmap->pm_stats.resident_count -= 1;
1591 if (oldpte & PG_MANAGED) {
1592 m = PHYS_TO_VM_PAGE(oldpte);
1593 if (oldpte & PG_M) {
1594 #if defined(PMAP_DIAGNOSTIC)
1595 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1597 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1601 if (pmap_track_modified(va))
1605 vm_page_flag_set(m, PG_REFERENCED);
1606 return pmap_remove_entry(pmap, m, va, info);
1608 return pmap_unuse_pt(pmap, va, NULL, info);
1617 * Remove a single page from a process address space.
1619 * This function may not be called from an interrupt if the pmap is
1623 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1628 * if there is no pte for this address, just skip it!!! Otherwise
1629 * get a local va for mappings for this pmap and remove the entry.
1631 if (*pmap_pde(pmap, va) != 0) {
1632 ptq = get_ptbase(pmap) + i386_btop(va);
1634 pmap_remove_pte(pmap, ptq, va, info);
1642 * Remove the given range of addresses from the specified map.
1644 * It is assumed that the start and end are properly
1645 * rounded to the page size.
1647 * This function may not be called from an interrupt if the pmap is
1651 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1655 vm_offset_t ptpaddr;
1656 vm_offset_t sindex, eindex;
1657 struct pmap_inval_info info;
1662 if (pmap->pm_stats.resident_count == 0)
1665 pmap_inval_init(&info);
1668 * special handling of removing one page. a very
1669 * common operation and easy to short circuit some
1672 if (((sva + PAGE_SIZE) == eva) &&
1673 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1674 pmap_remove_page(pmap, sva, &info);
1675 pmap_inval_flush(&info);
1680 * Get a local virtual address for the mappings that are being
1683 ptbase = get_ptbase(pmap);
1685 sindex = i386_btop(sva);
1686 eindex = i386_btop(eva);
1688 for (; sindex < eindex; sindex = pdnxt) {
1692 * Calculate index for next page table.
1694 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1695 if (pmap->pm_stats.resident_count == 0)
1698 pdirindex = sindex / NPDEPG;
1699 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1700 pmap_inval_add(&info, pmap, -1);
1701 pmap->pm_pdir[pdirindex] = 0;
1702 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1707 * Weed out invalid mappings. Note: we assume that the page
1708 * directory table is always allocated, and in kernel virtual.
1714 * Limit our scan to either the end of the va represented
1715 * by the current page table page, or to the end of the
1716 * range being removed.
1718 if (pdnxt > eindex) {
1722 for (; sindex != pdnxt; sindex++) {
1724 if (ptbase[sindex] == 0)
1726 va = i386_ptob(sindex);
1727 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1731 pmap_inval_flush(&info);
1737 * Removes this physical page from all physical maps in which it resides.
1738 * Reflects back modify bits to the pager.
1740 * This routine may not be called from an interrupt.
1744 pmap_remove_all(vm_page_t m)
1746 struct pmap_inval_info info;
1747 unsigned *pte, tpte;
1750 #if defined(PMAP_DIAGNOSTIC)
1752 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1755 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1756 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1760 pmap_inval_init(&info);
1762 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1763 pv->pv_pmap->pm_stats.resident_count--;
1765 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1766 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1768 tpte = loadandclear(pte);
1770 pv->pv_pmap->pm_stats.wired_count--;
1773 vm_page_flag_set(m, PG_REFERENCED);
1776 * Update the vm_page_t clean and reference bits.
1779 #if defined(PMAP_DIAGNOSTIC)
1780 if (pmap_nw_modified((pt_entry_t) tpte)) {
1782 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1786 if (pmap_track_modified(pv->pv_va))
1789 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1790 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1791 m->md.pv_list_count--;
1792 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1796 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1798 pmap_inval_flush(&info);
1804 * Set the physical protection on the specified range of this map
1807 * This function may not be called from an interrupt if the map is
1808 * not the kernel_pmap.
1811 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1814 vm_offset_t pdnxt, ptpaddr;
1815 vm_pindex_t sindex, eindex;
1816 pmap_inval_info info;
1821 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1822 pmap_remove(pmap, sva, eva);
1826 if (prot & VM_PROT_WRITE)
1829 pmap_inval_init(&info);
1831 ptbase = get_ptbase(pmap);
1833 sindex = i386_btop(sva);
1834 eindex = i386_btop(eva);
1836 for (; sindex < eindex; sindex = pdnxt) {
1840 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1842 pdirindex = sindex / NPDEPG;
1843 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1844 pmap_inval_add(&info, pmap, -1);
1845 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1846 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1851 * Weed out invalid mappings. Note: we assume that the page
1852 * directory table is always allocated, and in kernel virtual.
1857 if (pdnxt > eindex) {
1861 for (; sindex != pdnxt; sindex++) {
1866 /* XXX this isn't optimal */
1867 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1868 pbits = ptbase[sindex];
1870 if (pbits & PG_MANAGED) {
1873 m = PHYS_TO_VM_PAGE(pbits);
1874 vm_page_flag_set(m, PG_REFERENCED);
1878 if (pmap_track_modified(i386_ptob(sindex))) {
1880 m = PHYS_TO_VM_PAGE(pbits);
1889 if (pbits != ptbase[sindex]) {
1890 ptbase[sindex] = pbits;
1894 pmap_inval_flush(&info);
1898 * Insert the given physical page (p) at
1899 * the specified virtual address (v) in the
1900 * target physical map with the protection requested.
1902 * If specified, the page will be wired down, meaning
1903 * that the related pte can not be reclaimed.
1905 * NB: This is the only routine which MAY NOT lazy-evaluate
1906 * or lose information. That is, this routine must actually
1907 * insert this page into the given map NOW.
1910 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1916 vm_offset_t origpte, newpte;
1918 pmap_inval_info info;
1924 #ifdef PMAP_DIAGNOSTIC
1925 if (va > VM_MAX_KERNEL_ADDRESS)
1926 panic("pmap_enter: toobig");
1927 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1928 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1933 * In the case that a page table page is not
1934 * resident, we are creating it here.
1936 if (va < UPT_MIN_ADDRESS) {
1937 mpte = pmap_allocpte(pmap, va);
1940 pmap_inval_init(&info);
1941 pte = pmap_pte(pmap, va);
1944 * Page Directory table entry not valid, we need a new PT page
1947 panic("pmap_enter: invalid page directory pdir=%x, va=0x%x\n",
1948 (unsigned) pmap->pm_pdir[PTDPTDI], va);
1951 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
1952 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1953 origpte = *(vm_offset_t *)pte;
1954 opa = origpte & PG_FRAME;
1956 if (origpte & PG_PS)
1957 panic("pmap_enter: attempted pmap_enter on 4MB page");
1960 * Mapping has not changed, must be protection or wiring change.
1962 if (origpte && (opa == pa)) {
1964 * Wiring change, just update stats. We don't worry about
1965 * wiring PT pages as they remain resident as long as there
1966 * are valid mappings in them. Hence, if a user page is wired,
1967 * the PT page will be also.
1969 if (wired && ((origpte & PG_W) == 0))
1970 pmap->pm_stats.wired_count++;
1971 else if (!wired && (origpte & PG_W))
1972 pmap->pm_stats.wired_count--;
1974 #if defined(PMAP_DIAGNOSTIC)
1975 if (pmap_nw_modified((pt_entry_t) origpte)) {
1977 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1983 * Remove the extra pte reference. Note that we cannot
1984 * optimize the RO->RW case because we have adjusted the
1985 * wiring count above and may need to adjust the wiring
1992 * We might be turning off write access to the page,
1993 * so we go ahead and sense modify status.
1995 if (origpte & PG_MANAGED) {
1996 if ((origpte & PG_M) && pmap_track_modified(va)) {
1998 om = PHYS_TO_VM_PAGE(opa);
2006 * Mapping has changed, invalidate old range and fall through to
2007 * handle validating new mapping.
2011 err = pmap_remove_pte(pmap, pte, va, &info);
2013 panic("pmap_enter: pte vanished, va: 0x%x", va);
2017 * Enter on the PV list if part of our managed memory. Note that we
2018 * raise IPL while manipulating pv_table since pmap_enter can be
2019 * called at interrupt time.
2021 if (pmap_initialized &&
2022 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2023 pmap_insert_entry(pmap, va, mpte, m);
2028 * Increment counters
2030 pmap->pm_stats.resident_count++;
2032 pmap->pm_stats.wired_count++;
2036 * Now validate mapping with desired protection/wiring.
2038 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2042 if (va < UPT_MIN_ADDRESS)
2044 if (pmap == kernel_pmap)
2048 * if the mapping or permission bits are different, we need
2049 * to update the pte.
2051 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2052 *pte = newpte | PG_A;
2054 pmap_inval_flush(&info);
2058 * this code makes some *MAJOR* assumptions:
2059 * 1. Current pmap & pmap exists.
2062 * 4. No page table pages.
2063 * 5. Tlbflush is deferred to calling procedure.
2064 * 6. Page IS managed.
2065 * but is *MUCH* faster than pmap_enter...
2069 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
2073 pmap_inval_info info;
2075 pmap_inval_init(&info);
2078 * In the case that a page table page is not
2079 * resident, we are creating it here.
2081 if (va < UPT_MIN_ADDRESS) {
2086 * Calculate pagetable page index
2088 ptepindex = va >> PDRSHIFT;
2089 if (mpte && (mpte->pindex == ptepindex)) {
2094 * Get the page directory entry
2096 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2099 * If the page table page is mapped, we just increment
2100 * the hold count, and activate it.
2104 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2105 if (pmap->pm_ptphint &&
2106 (pmap->pm_ptphint->pindex == ptepindex)) {
2107 mpte = pmap->pm_ptphint;
2109 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2110 pmap->pm_ptphint = mpte;
2116 mpte = _pmap_allocpte(pmap, ptepindex);
2124 * This call to vtopte makes the assumption that we are
2125 * entering the page into the current pmap. In order to support
2126 * quick entry into any pmap, one would likely use pmap_pte_quick.
2127 * But that isn't as quick as vtopte.
2129 pte = (unsigned *)vtopte(va);
2132 pmap_unwire_pte_hold(pmap, mpte, &info);
2137 * Enter on the PV list if part of our managed memory. Note that we
2138 * raise IPL while manipulating pv_table since pmap_enter can be
2139 * called at interrupt time.
2141 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
2142 pmap_insert_entry(pmap, va, mpte, m);
2145 * Increment counters
2147 pmap->pm_stats.resident_count++;
2149 pa = VM_PAGE_TO_PHYS(m);
2152 * Now validate mapping with RO protection
2154 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2155 *pte = pa | PG_V | PG_U;
2157 *pte = pa | PG_V | PG_U | PG_MANAGED;
2163 * Make a temporary mapping for a physical address. This is only intended
2164 * to be used for panic dumps.
2167 pmap_kenter_temporary(vm_paddr_t pa, int i)
2169 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2170 return ((void *)crashdumpmap);
2173 #define MAX_INIT_PT (96)
2176 * This routine preloads the ptes for a given object into the specified pmap.
2177 * This eliminates the blast of soft faults on process startup and
2178 * immediately after an mmap.
2181 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2182 vm_object_t object, vm_pindex_t pindex,
2183 vm_size_t size, int limit)
2190 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2195 * XXX you must be joking, entering PTE's into a user page table
2196 * without any accounting? This could result in the page table
2197 * being freed while it still contains mappings (free with PG_ZERO
2198 * assumption leading to a non-zero page being marked PG_ZERO).
2201 * This code maps large physical mmap regions into the
2202 * processor address space. Note that some shortcuts
2203 * are taken, but the code works.
2206 (object->type == OBJT_DEVICE) &&
2207 ((addr & (NBPDR - 1)) == 0) &&
2208 ((size & (NBPDR - 1)) == 0) ) {
2211 unsigned int ptepindex;
2215 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)])
2219 p = vm_page_lookup(object, pindex);
2220 if (p && vm_page_sleep_busy(p, FALSE, "init4p"))
2224 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
2229 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) {
2234 p = vm_page_lookup(object, pindex);
2238 ptepa = (vm_offset_t) VM_PAGE_TO_PHYS(p);
2239 if (ptepa & (NBPDR - 1)) {
2243 p->valid = VM_PAGE_BITS_ALL;
2245 pmap->pm_stats.resident_count += size >> PAGE_SHIFT;
2246 npdes = size >> PDRSHIFT;
2247 for (i = 0; i < npdes; i++) {
2248 pmap->pm_pdir[ptepindex] =
2249 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_PS);
2253 vm_page_flag_set(p, PG_MAPPED);
2260 psize = i386_btop(size);
2262 if ((object->type != OBJT_VNODE) ||
2263 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2264 (object->resident_page_count > MAX_INIT_PT))) {
2268 if (psize + pindex > object->size) {
2269 if (object->size < pindex)
2271 psize = object->size - pindex;
2276 * If we are processing a major portion of the object, then scan the
2279 * We cannot safely scan the object's memq unless we are in a
2280 * critical section since interrupts can remove pages from objects.
2284 if (psize > (object->resident_page_count >> 2)) {
2287 for (p = TAILQ_FIRST(&object->memq);
2288 objpgs > 0 && p != NULL;
2289 p = TAILQ_NEXT(p, listq)
2292 if (tmpidx < pindex)
2295 if (tmpidx >= psize)
2299 * don't allow an madvise to blow away our really
2300 * free pages allocating pv entries.
2302 if ((limit & MAP_PREFAULT_MADVISE) &&
2303 vmstats.v_free_count < vmstats.v_free_reserved) {
2306 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2308 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2309 if ((p->queue - p->pc) == PQ_CACHE)
2310 vm_page_deactivate(p);
2312 mpte = pmap_enter_quick(pmap,
2313 addr + i386_ptob(tmpidx), p, mpte);
2314 vm_page_flag_set(p, PG_MAPPED);
2321 * else lookup the pages one-by-one.
2323 for (tmpidx = 0; tmpidx < psize; tmpidx += 1) {
2325 * don't allow an madvise to blow away our really
2326 * free pages allocating pv entries.
2328 if ((limit & MAP_PREFAULT_MADVISE) &&
2329 vmstats.v_free_count < vmstats.v_free_reserved) {
2332 p = vm_page_lookup(object, tmpidx + pindex);
2334 ((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2336 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2337 if ((p->queue - p->pc) == PQ_CACHE)
2338 vm_page_deactivate(p);
2340 mpte = pmap_enter_quick(pmap,
2341 addr + i386_ptob(tmpidx), p, mpte);
2342 vm_page_flag_set(p, PG_MAPPED);
2351 * pmap_prefault provides a quick way of clustering pagefaults into a
2352 * processes address space. It is a "cousin" of pmap_object_init_pt,
2353 * except it runs at page fault time instead of mmap time.
2357 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2359 static int pmap_prefault_pageorder[] = {
2360 -PAGE_SIZE, PAGE_SIZE,
2361 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2362 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2363 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2367 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2377 * We do not currently prefault mappings that use virtual page
2378 * tables. We do not prefault foreign pmaps.
2380 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2382 if (!curproc || (pmap != vmspace_pmap(curproc->p_vmspace)))
2385 object = entry->object.vm_object;
2387 starta = addra - PFBAK * PAGE_SIZE;
2388 if (starta < entry->start)
2389 starta = entry->start;
2390 else if (starta > addra)
2394 * critical section protection is required to maintain the
2395 * page/object association, interrupts can free pages and remove
2396 * them from their objects.
2400 for (i = 0; i < PAGEORDER_SIZE; i++) {
2401 vm_object_t lobject;
2404 addr = addra + pmap_prefault_pageorder[i];
2405 if (addr > addra + (PFFOR * PAGE_SIZE))
2408 if (addr < starta || addr >= entry->end)
2411 if ((*pmap_pde(pmap, addr)) == NULL)
2414 pte = (unsigned *) vtopte(addr);
2418 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2421 for (m = vm_page_lookup(lobject, pindex);
2422 (!m && (lobject->type == OBJT_DEFAULT) &&
2423 (lobject->backing_object));
2424 lobject = lobject->backing_object
2426 if (lobject->backing_object_offset & PAGE_MASK)
2428 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2429 m = vm_page_lookup(lobject->backing_object, pindex);
2433 * give-up when a page is not in memory
2438 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2440 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2442 if ((m->queue - m->pc) == PQ_CACHE) {
2443 vm_page_deactivate(m);
2446 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2447 vm_page_flag_set(m, PG_MAPPED);
2455 * Routine: pmap_change_wiring
2456 * Function: Change the wiring attribute for a map/virtual-address
2458 * In/out conditions:
2459 * The mapping must already exist in the pmap.
2462 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2469 pte = pmap_pte(pmap, va);
2471 if (wired && !pmap_pte_w(pte))
2472 pmap->pm_stats.wired_count++;
2473 else if (!wired && pmap_pte_w(pte))
2474 pmap->pm_stats.wired_count--;
2477 * Wiring is not a hardware characteristic so there is no need to
2478 * invalidate TLB. However, in an SMP environment we must use
2479 * a locked bus cycle to update the pte (if we are not using
2480 * the pmap_inval_*() API that is)... it's ok to do this for simple
2485 atomic_set_int(pte, PG_W);
2487 atomic_clear_int(pte, PG_W);
2490 atomic_set_int_nonlocked(pte, PG_W);
2492 atomic_clear_int_nonlocked(pte, PG_W);
2499 * Copy the range specified by src_addr/len
2500 * from the source map to the range dst_addr/len
2501 * in the destination map.
2503 * This routine is only advisory and need not do anything.
2506 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2507 vm_size_t len, vm_offset_t src_addr)
2509 pmap_inval_info info;
2511 vm_offset_t end_addr = src_addr + len;
2513 unsigned src_frame, dst_frame;
2516 if (dst_addr != src_addr)
2519 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2520 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2524 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2525 if (dst_frame != (((unsigned) APTDpde) & PG_FRAME)) {
2526 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2527 /* The page directory is not shared between CPUs */
2530 pmap_inval_init(&info);
2531 pmap_inval_add(&info, dst_pmap, -1);
2532 pmap_inval_add(&info, src_pmap, -1);
2535 * critical section protection is required to maintain the page/object
2536 * association, interrupts can free pages and remove them from
2540 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2541 unsigned *src_pte, *dst_pte;
2542 vm_page_t dstmpte, srcmpte;
2543 vm_offset_t srcptepaddr;
2546 if (addr >= UPT_MIN_ADDRESS)
2547 panic("pmap_copy: invalid to pmap_copy page tables\n");
2550 * Don't let optional prefaulting of pages make us go
2551 * way below the low water mark of free pages or way
2552 * above high water mark of used pv entries.
2554 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2555 pv_entry_count > pv_entry_high_water)
2558 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2559 ptepindex = addr >> PDRSHIFT;
2561 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2562 if (srcptepaddr == 0)
2565 if (srcptepaddr & PG_PS) {
2566 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2567 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2568 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2573 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2574 if ((srcmpte == NULL) ||
2575 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2578 if (pdnxt > end_addr)
2581 src_pte = (unsigned *) vtopte(addr);
2582 dst_pte = (unsigned *) avtopte(addr);
2583 while (addr < pdnxt) {
2587 * we only virtual copy managed pages
2589 if ((ptetemp & PG_MANAGED) != 0) {
2591 * We have to check after allocpte for the
2592 * pte still being around... allocpte can
2595 dstmpte = pmap_allocpte(dst_pmap, addr);
2596 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2598 * Clear the modified and
2599 * accessed (referenced) bits
2602 m = PHYS_TO_VM_PAGE(ptetemp);
2603 *dst_pte = ptetemp & ~(PG_M | PG_A);
2604 dst_pmap->pm_stats.resident_count++;
2605 pmap_insert_entry(dst_pmap, addr,
2608 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2610 if (dstmpte->hold_count >= srcmpte->hold_count)
2619 pmap_inval_flush(&info);
2623 * Routine: pmap_kernel
2625 * Returns the physical map handle for the kernel.
2630 return (kernel_pmap);
2636 * Zero the specified PA by mapping the page into KVM and clearing its
2639 * This function may be called from an interrupt and no locking is
2643 pmap_zero_page(vm_paddr_t phys)
2645 struct mdglobaldata *gd = mdcpu;
2648 if (*(int *)gd->gd_CMAP3)
2649 panic("pmap_zero_page: CMAP3 busy");
2650 *(int *)gd->gd_CMAP3 =
2651 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2652 cpu_invlpg(gd->gd_CADDR3);
2654 #if defined(I686_CPU)
2655 if (cpu_class == CPUCLASS_686)
2656 i686_pagezero(gd->gd_CADDR3);
2659 bzero(gd->gd_CADDR3, PAGE_SIZE);
2660 *(int *) gd->gd_CMAP3 = 0;
2665 * pmap_page_assertzero:
2667 * Assert that a page is empty, panic if it isn't.
2670 pmap_page_assertzero(vm_paddr_t phys)
2672 struct mdglobaldata *gd = mdcpu;
2676 if (*(int *)gd->gd_CMAP3)
2677 panic("pmap_zero_page: CMAP3 busy");
2678 *(int *)gd->gd_CMAP3 =
2679 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2680 cpu_invlpg(gd->gd_CADDR3);
2681 for (i = 0; i < PAGE_SIZE; i += 4) {
2682 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2683 panic("pmap_page_assertzero() @ %p not zero!\n",
2684 (void *)gd->gd_CADDR3);
2687 *(int *) gd->gd_CMAP3 = 0;
2694 * Zero part of a physical page by mapping it into memory and clearing
2695 * its contents with bzero.
2697 * off and size may not cover an area beyond a single hardware page.
2700 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2702 struct mdglobaldata *gd = mdcpu;
2705 if (*(int *) gd->gd_CMAP3)
2706 panic("pmap_zero_page: CMAP3 busy");
2707 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2708 cpu_invlpg(gd->gd_CADDR3);
2710 #if defined(I686_CPU)
2711 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2712 i686_pagezero(gd->gd_CADDR3);
2715 bzero((char *)gd->gd_CADDR3 + off, size);
2716 *(int *) gd->gd_CMAP3 = 0;
2723 * Copy the physical page from the source PA to the target PA.
2724 * This function may be called from an interrupt. No locking
2728 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2730 struct mdglobaldata *gd = mdcpu;
2733 if (*(int *) gd->gd_CMAP1)
2734 panic("pmap_copy_page: CMAP1 busy");
2735 if (*(int *) gd->gd_CMAP2)
2736 panic("pmap_copy_page: CMAP2 busy");
2738 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2739 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2741 cpu_invlpg(gd->gd_CADDR1);
2742 cpu_invlpg(gd->gd_CADDR2);
2744 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2746 *(int *) gd->gd_CMAP1 = 0;
2747 *(int *) gd->gd_CMAP2 = 0;
2752 * pmap_copy_page_frag:
2754 * Copy the physical page from the source PA to the target PA.
2755 * This function may be called from an interrupt. No locking
2759 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2761 struct mdglobaldata *gd = mdcpu;
2764 if (*(int *) gd->gd_CMAP1)
2765 panic("pmap_copy_page: CMAP1 busy");
2766 if (*(int *) gd->gd_CMAP2)
2767 panic("pmap_copy_page: CMAP2 busy");
2769 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2770 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2772 cpu_invlpg(gd->gd_CADDR1);
2773 cpu_invlpg(gd->gd_CADDR2);
2775 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2776 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2779 *(int *) gd->gd_CMAP1 = 0;
2780 *(int *) gd->gd_CMAP2 = 0;
2785 * Returns true if the pmap's pv is one of the first
2786 * 16 pvs linked to from this page. This count may
2787 * be changed upwards or downwards in the future; it
2788 * is only necessary that true be returned for a small
2789 * subset of pmaps for proper page aging.
2792 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2797 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2802 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2803 if (pv->pv_pmap == pmap) {
2816 * Remove all pages from specified address space
2817 * this aids process exit speeds. Also, this code
2818 * is special cased for current process only, but
2819 * can have the more generic (and slightly slower)
2820 * mode enabled. This is much faster than pmap_remove
2821 * in the case of running down an entire address space.
2824 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2826 unsigned *pte, tpte;
2829 pmap_inval_info info;
2832 if (curproc && pmap == vmspace_pmap(curproc->p_vmspace))
2837 pmap_inval_init(&info);
2839 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2841 if (pv->pv_va >= eva || pv->pv_va < sva) {
2842 npv = TAILQ_NEXT(pv, pv_plist);
2847 pte = (unsigned *)vtopte(pv->pv_va);
2849 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2850 if (pmap->pm_active)
2851 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2855 * We cannot remove wired pages from a process' mapping
2859 npv = TAILQ_NEXT(pv, pv_plist);
2864 m = PHYS_TO_VM_PAGE(tpte);
2866 KASSERT(m < &vm_page_array[vm_page_array_size],
2867 ("pmap_remove_pages: bad tpte %x", tpte));
2869 pv->pv_pmap->pm_stats.resident_count--;
2872 * Update the vm_page_t clean and reference bits.
2879 npv = TAILQ_NEXT(pv, pv_plist);
2880 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2882 m->md.pv_list_count--;
2883 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2884 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2885 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2888 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2891 pmap_inval_flush(&info);
2896 * pmap_testbit tests bits in pte's
2897 * note that the testbit/changebit routines are inline,
2898 * and a lot of things compile-time evaluate.
2901 pmap_testbit(vm_page_t m, int bit)
2906 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2909 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2914 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2916 * if the bit being tested is the modified bit, then
2917 * mark clean_map and ptes as never
2920 if (bit & (PG_A|PG_M)) {
2921 if (!pmap_track_modified(pv->pv_va))
2925 #if defined(PMAP_DIAGNOSTIC)
2927 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2931 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2942 * this routine is used to modify bits in ptes
2944 static __inline void
2945 pmap_changebit(vm_page_t m, int bit, boolean_t setem)
2947 struct pmap_inval_info info;
2951 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2954 pmap_inval_init(&info);
2958 * Loop over all current mappings setting/clearing as appropos If
2959 * setting RO do we need to clear the VAC?
2961 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2963 * don't write protect pager mappings
2965 if (!setem && (bit == PG_RW)) {
2966 if (!pmap_track_modified(pv->pv_va))
2970 #if defined(PMAP_DIAGNOSTIC)
2972 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2978 * Careful here. We can use a locked bus instruction to
2979 * clear PG_A or PG_M safely but we need to synchronize
2980 * with the target cpus when we mess with PG_RW.
2982 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2984 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2988 atomic_set_int(pte, bit);
2990 atomic_set_int_nonlocked(pte, bit);
2993 vm_offset_t pbits = *(vm_offset_t *)pte;
3000 atomic_clear_int(pte, PG_M|PG_RW);
3002 atomic_clear_int_nonlocked(pte, PG_M|PG_RW);
3006 atomic_clear_int(pte, bit);
3008 atomic_clear_int_nonlocked(pte, bit);
3014 pmap_inval_flush(&info);
3019 * pmap_page_protect:
3021 * Lower the permission for all mappings to a given page.
3024 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3026 if ((prot & VM_PROT_WRITE) == 0) {
3027 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3028 pmap_changebit(m, PG_RW, FALSE);
3036 pmap_phys_address(int ppn)
3038 return (i386_ptob(ppn));
3042 * pmap_ts_referenced:
3044 * Return a count of reference bits for a page, clearing those bits.
3045 * It is not necessary for every reference bit to be cleared, but it
3046 * is necessary that 0 only be returned when there are truly no
3047 * reference bits set.
3049 * XXX: The exact number of bits to check and clear is a matter that
3050 * should be tested and standardized at some point in the future for
3051 * optimal aging of shared pages.
3054 pmap_ts_referenced(vm_page_t m)
3056 pv_entry_t pv, pvf, pvn;
3060 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3065 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3070 pvn = TAILQ_NEXT(pv, pv_list);
3072 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3074 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3076 if (!pmap_track_modified(pv->pv_va))
3079 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3081 if (pte && (*pte & PG_A)) {
3083 atomic_clear_int(pte, PG_A);
3085 atomic_clear_int_nonlocked(pte, PG_A);
3092 } while ((pv = pvn) != NULL && pv != pvf);
3102 * Return whether or not the specified physical page was modified
3103 * in any physical maps.
3106 pmap_is_modified(vm_page_t m)
3108 return pmap_testbit(m, PG_M);
3112 * Clear the modify bits on the specified physical page.
3115 pmap_clear_modify(vm_page_t m)
3117 pmap_changebit(m, PG_M, FALSE);
3121 * pmap_clear_reference:
3123 * Clear the reference bit on the specified physical page.
3126 pmap_clear_reference(vm_page_t m)
3128 pmap_changebit(m, PG_A, FALSE);
3132 * Miscellaneous support routines follow
3136 i386_protection_init(void)
3140 kp = protection_codes;
3141 for (prot = 0; prot < 8; prot++) {
3143 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3145 * Read access is also 0. There isn't any execute bit,
3146 * so just make it readable.
3148 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3149 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3150 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3153 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3154 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3155 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3156 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3164 * Map a set of physical memory pages into the kernel virtual
3165 * address space. Return a pointer to where it is mapped. This
3166 * routine is intended to be used for mapping device memory,
3169 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3173 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3175 vm_offset_t va, tmpva, offset;
3178 offset = pa & PAGE_MASK;
3179 size = roundup(offset + size, PAGE_SIZE);
3181 va = kmem_alloc_nofault(kernel_map, size);
3183 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3186 for (tmpva = va; size > 0;) {
3187 pte = (unsigned *)vtopte(tmpva);
3188 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3196 return ((void *)(va + offset));
3200 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3202 vm_offset_t base, offset;
3204 base = va & PG_FRAME;
3205 offset = va & PAGE_MASK;
3206 size = roundup(offset + size, PAGE_SIZE);
3207 pmap_qremove(va, size >> PAGE_SHIFT);
3208 kmem_free(kernel_map, base, size);
3212 * perform the pmap work for mincore
3215 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3217 unsigned *ptep, pte;
3221 ptep = pmap_pte(pmap, addr);
3226 if ((pte = *ptep) != 0) {
3229 val = MINCORE_INCORE;
3230 if ((pte & PG_MANAGED) == 0)
3233 pa = pte & PG_FRAME;
3235 m = PHYS_TO_VM_PAGE(pa);
3241 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3243 * Modified by someone
3245 else if (m->dirty || pmap_is_modified(m))
3246 val |= MINCORE_MODIFIED_OTHER;
3251 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3254 * Referenced by someone
3256 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3257 val |= MINCORE_REFERENCED_OTHER;
3258 vm_page_flag_set(m, PG_REFERENCED);
3265 pmap_activate(struct proc *p)
3269 pmap = vmspace_pmap(p->p_vmspace);
3271 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3273 pmap->pm_active |= 1;
3275 #if defined(SWTCH_OPTIM_STATS)
3278 p->p_thread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3279 load_cr3(p->p_thread->td_pcb->pcb_cr3);
3283 pmap_deactivate(struct proc *p)
3287 pmap = vmspace_pmap(p->p_vmspace);
3289 atomic_clear_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3291 pmap->pm_active &= ~1;
3294 * XXX - note we do not adjust %cr3. The caller is expected to
3295 * activate a new pmap or do a thread-exit.
3300 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3303 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3307 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3314 static void pads (pmap_t pm);
3315 void pmap_pvdump (vm_paddr_t pa);
3317 /* print address space of pmap*/
3324 if (pm == kernel_pmap)
3326 for (i = 0; i < 1024; i++)
3328 for (j = 0; j < 1024; j++) {
3329 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3330 if (pm == kernel_pmap && va < KERNBASE)
3332 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS)
3334 ptep = pmap_pte_quick(pm, va);
3335 if (pmap_pte_v(ptep))
3336 printf("%x:%x ", va, *(int *) ptep);
3342 pmap_pvdump(vm_paddr_t pa)
3347 printf("pa %08llx", (long long)pa);
3348 m = PHYS_TO_VM_PAGE(pa);
3349 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3351 printf(" -> pmap %p, va %x, flags %x",
3352 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3354 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);